2019-05-02 00:40:49 +08:00
|
|
|
; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX6789,SI %s
|
|
|
|
|
; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX6789,GFX8910,SIVI,PRT %s
|
|
|
|
|
; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX6789,PRT %s
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; RUN: llc -march=amdgcn -mcpu=gfx900 -mattr=-enable-prt-strict-null -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GFX900,NOPRT %s
|
2019-05-02 00:40:49 +08:00
|
|
|
; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefixes=GCN,GFX10 %s
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v0, s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_1d(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.1d.v4f32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_1d_tfe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v{{[0-9]+}}, s[0:7] dmask:0xf unorm tfe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v{{[0-9]+}}, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_1d_tfe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_lwe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v{{[0-9]+}}, s[0:7] dmask:0xf unorm lwe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v{{[0-9]+}}, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm lwe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_1d_lwe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>, i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 2, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v[0:1], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v[0:1], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_2d(<8 x i32> inreg %rsrc, i32 %s, i32 %t) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.2d.v4f32.i32(i32 15, i32 %s, i32 %t, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2d_tfe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm tfe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_2d_tfe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.2d.v4f32i32.i32(i32 15, i32 %s, i32 %t, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_3d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v[0:3], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v[0:2], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_3d(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %r) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.3d.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %r, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_3d_tfe_lwe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm tfe lwe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm tfe lwe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_3d_tfe_lwe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t, i32 %r) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.3d.v4f32i32.i32(i32 15, i32 %s, i32 %t, i32 %r, <8 x i32> %rsrc, i32 3, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_cube:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v[0:3], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v[0:2], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_CUBE unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_cube(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.cube.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %slice, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_cube_lwe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm lwe da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_CUBE unorm lwe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_cube_lwe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.cube.v4f32i32.i32(i32 15, i32 %s, i32 %t, i32 %slice, <8 x i32> %rsrc, i32 2, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_1darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v[0:1], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v[0:1], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_1darray(<8 x i32> inreg %rsrc, i32 %s, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.1darray.v4f32.i32(i32 15, i32 %s, i32 %slice, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_1darray_tfe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm tfe da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D_ARRAY unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_1darray_tfe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1darray.v4f32i32.i32(i32 15, i32 %s, i32 %slice, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v[0:3], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v[0:2], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_2darray(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.2darray.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %slice, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2darray_lwe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm lwe da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_ARRAY unorm lwe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_2darray_lwe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.2darray.v4f32i32.i32(i32 15, i32 %s, i32 %t, i32 %slice, <8 x i32> %rsrc, i32 2, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2dmsaa:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v[0:3], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v[0:2], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_2dmsaa(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %fragid) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.2dmsaa.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %fragid, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2dmsaa_both:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm tfe lwe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA unorm tfe lwe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_2dmsaa_both(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t, i32 %fragid) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.2dmsaa.v4f32i32.i32(i32 15, i32 %s, i32 %t, i32 %fragid, <8 x i32> %rsrc, i32 3, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2darraymsaa:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v[0:3], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v[0:3], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_2darraymsaa(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %slice, i32 %fragid) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.2darraymsaa.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %slice, i32 %fragid, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_2darraymsaa_tfe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm tfe da{{$}}
|
|
|
|
|
; GFX10: image_load v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA_ARRAY unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_2darraymsaa_tfe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t, i32 %slice, i32 %fragid) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.2darraymsaa.v4f32i32.i32(i32 15, i32 %s, i32 %t, i32 %slice, i32 %fragid, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_mip_1d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:3], v[0:1], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:3], v[0:1], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_mip_1d(<8 x i32> inreg %rsrc, i32 %s, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.mip.1d.v4f32.i32(i32 15, i32 %s, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_mip_1d_lwe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm lwe{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm lwe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_mip_1d_lwe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.mip.1d.v4f32i32.i32(i32 15, i32 %s, i32 %mip, <8 x i32> %rsrc, i32 2, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_mip_2d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:3], v[0:2], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_mip_2d(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.mip.2d.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; GCN-LABEL: {{^}}load_mip_2d_tfe:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v4, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v3
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf unorm tfe{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v4, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v4
|
|
|
|
|
define amdgpu_ps <4 x float> @load_mip_2d_tfe(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s, i32 %t, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.mip.2d.v4f32i32.i32(i32 15, i32 %s, i32 %t, i32 %mip, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; Make sure that error flag is returned even with dmask 0
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_V2_tfe_dmask0:
|
|
|
|
|
; GCN: v_mov_b32_e32 v1, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v2, v1
|
|
|
|
|
; PRT: image_load v[1:2], v0, s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT: image_load v[0:1], v0, s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
define amdgpu_ps float @load_1d_V2_tfe_dmask0(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<2 x float>,i32} @llvm.amdgcn.image.load.1d.v2f32i32.i32(i32 0, i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.err = extractvalue {<2 x float>, i32} %v, 1
|
|
|
|
|
%vv = bitcast i32 %v.err to float
|
|
|
|
|
ret float %vv
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_V1_tfe_dmask0:
|
|
|
|
|
; GCN: v_mov_b32_e32 v1, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v2, v1
|
|
|
|
|
; PRT: image_load v[1:2], v0, s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT: image_load v[0:1], v0, s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
define amdgpu_ps float @load_1d_V1_tfe_dmask0(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {float,i32} @llvm.amdgcn.image.load.1d.f32i32.i32(i32 0, i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.err = extractvalue {float, i32} %v, 1
|
|
|
|
|
%vv = bitcast i32 %v.err to float
|
|
|
|
|
ret float %vv
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_mip_2d_tfe_dmask0:
|
|
|
|
|
; GCN: v_mov_b32_e32 v3, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v3
|
|
|
|
|
; PRT: image_load_mip v[3:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT: image_load_mip v[2:3], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
define amdgpu_ps float @load_mip_2d_tfe_dmask0(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.mip.2d.v4f32i32.i32(i32 0, i32 %s, i32 %t, i32 %mip, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
%vv = bitcast i32 %v.err to float
|
|
|
|
|
ret float %vv
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; Do not make dmask 0 even if no result (other than tfe) is used.
|
|
|
|
|
; GCN-LABEL: {{^}}load_mip_2d_tfe_nouse:
|
|
|
|
|
; GCN: v_mov_b32_e32 v3, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v3
|
|
|
|
|
; PRT: image_load_mip v[3:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT: image_load_mip v[2:3], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
define amdgpu_ps float @load_mip_2d_tfe_nouse(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.mip.2d.v4f32i32.i32(i32 15, i32 %s, i32 %t, i32 %mip, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
%vv = bitcast i32 %v.err to float
|
|
|
|
|
ret float %vv
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_mip_2d_tfe_nouse_V2:
|
|
|
|
|
; GCN: v_mov_b32_e32 v3, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v3
|
|
|
|
|
; PRT: image_load_mip v[3:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT: image_load_mip v[2:3], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x1 unorm tfe{{$}}
|
|
|
|
|
define amdgpu_ps float @load_mip_2d_tfe_nouse_V2(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<2 x float>,i32} @llvm.amdgcn.image.load.mip.2d.v2f32i32.i32(i32 6, i32 %s, i32 %t, i32 %mip, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.err = extractvalue {<2 x float>, i32} %v, 1
|
|
|
|
|
%vv = bitcast i32 %v.err to float
|
|
|
|
|
ret float %vv
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_mip_2d_tfe_nouse_V1:
|
|
|
|
|
; GCN: v_mov_b32_e32 v3, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v3
|
|
|
|
|
; PRT: image_load_mip v[3:4], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x2 unorm tfe{{$}}
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
|
|
|
|
; NOPRT: image_load_mip v[2:3], v[{{[0-9]+:[0-9]+}}], s[0:7] dmask:0x2 unorm tfe{{$}}
|
|
|
|
|
define amdgpu_ps float @load_mip_2d_tfe_nouse_V1(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {float, i32} @llvm.amdgcn.image.load.mip.2d.f32i32.i32(i32 2, i32 %s, i32 %t, i32 %mip, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.err = extractvalue {float, i32} %v, 1
|
|
|
|
|
%vv = bitcast i32 %v.err to float
|
|
|
|
|
ret float %vv
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; Check for dmask being materially smaller than return type
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_tfe_V4_dmask3:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v3, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v2
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v{{[0-9]+}}, s[0:7] dmask:0x7 unorm tfe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v{{[0-9]+}}, s[0:7] dmask:0x7 dim:SQ_RSRC_IMG_1D unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v3, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v3
|
|
|
|
|
define amdgpu_ps <4 x float> @load_1d_tfe_V4_dmask3(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 7, i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_tfe_V4_dmask2:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v2, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v1
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:2], v{{[0-9]+}}, s[0:7] dmask:0x6 unorm tfe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:2], v{{[0-9]+}}, s[0:7] dmask:0x6 dim:SQ_RSRC_IMG_1D unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v2, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v2
|
|
|
|
|
define amdgpu_ps <4 x float> @load_1d_tfe_V4_dmask2(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 6, i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_tfe_V4_dmask1:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v1, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:1], v{{[0-9]+}}, s[0:7] dmask:0x8 unorm tfe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:1], v{{[0-9]+}}, s[0:7] dmask:0x8 dim:SQ_RSRC_IMG_1D unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v1, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v1
|
|
|
|
|
define amdgpu_ps <4 x float> @load_1d_tfe_V4_dmask1(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 8, i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<4 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<4 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <4 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_tfe_V2_dmask1:
|
|
|
|
|
; PRT: v_mov_b32_e32 v0, 0
|
|
|
|
|
; PRT-DAG: v_mov_b32_e32 v{{[0-9]+}}, v0
|
|
|
|
|
; NOPRT: v_mov_b32_e32 v1, 0
|
|
|
|
|
; NOPRT-NOT: v_mov_b32_e32 v0
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:1], v{{[0-9]+}}, s[0:7] dmask:0x8 unorm tfe{{$}}
|
|
|
|
|
; GFX10: image_load v[0:1], v{{[0-9]+}}, s[0:7] dmask:0x8 dim:SQ_RSRC_IMG_1D unorm tfe ;
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
; SIVI: buffer_store_dword v1, off, s[8:11], 0
|
|
|
|
|
; GFX900: global_store_dword v[{{[0-9]+:[0-9]+}}], v1
|
|
|
|
|
define amdgpu_ps <2 x float> @load_1d_tfe_V2_dmask1(<8 x i32> inreg %rsrc, i32 addrspace(1)* inreg %out, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call {<2 x float>,i32} @llvm.amdgcn.image.load.1d.v2f32i32.i32(i32 8, i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
|
|
|
|
|
%v.vec = extractvalue {<2 x float>, i32} %v, 0
|
|
|
|
|
%v.err = extractvalue {<2 x float>, i32} %v, 1
|
|
|
|
|
store i32 %v.err, i32 addrspace(1)* %out, align 4
|
|
|
|
|
ret <2 x float> %v.vec
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
; GCN-LABEL: {{^}}load_mip_3d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_mip_3d(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %r, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.mip.3d.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %r, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_mip_cube:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_CUBE unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_mip_cube(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %slice, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.mip.cube.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %slice, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_mip_1darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:3], v[0:2], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_mip_1darray(<8 x i32> inreg %rsrc, i32 %s, i32 %slice, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.mip.1darray.v4f32.i32(i32 15, i32 %s, i32 %slice, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_mip_2darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_load_mip v[0:3], v[0:3], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_mip_2darray(<8 x i32> inreg %rsrc, i32 %s, i32 %t, i32 %slice, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.mip.2darray.v4f32.i32(i32 15, i32 %s, i32 %t, i32 %slice, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_1d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v4, s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_1d(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_2d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v[4:5], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v[4:5], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_2d(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.2d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_3d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v[4:7], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v[4:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_3d(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %r) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.3d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %r, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_cube:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v[4:7], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v[4:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_CUBE unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_cube(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.cube.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %slice, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_1darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v[4:5], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v[4:5], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_1darray(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1darray.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %slice, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_2darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v[4:7], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v[4:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_2darray(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %slice) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.2darray.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %slice, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_2dmsaa:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v[4:7], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v[4:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_2dmsaa(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %fragid) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.2dmsaa.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %fragid, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_2darraymsaa:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v[4:7], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v[4:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_2darraymsaa(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %slice, i32 %fragid) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.2darraymsaa.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %slice, i32 %fragid, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_mip_1d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store_mip v[0:3], v[4:5], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_store_mip v[0:3], v[4:5], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_mip_1d(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.mip.1d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_mip_2d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_store_mip v[0:3], v[4:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_mip_2d(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.mip.2d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_mip_3d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_mip_3d(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %r, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.mip.3d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %r, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_mip_cube:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_CUBE unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_mip_cube(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %slice, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.mip.cube.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %slice, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_mip_1darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_store_mip v[0:3], v[4:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_mip_1darray(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %slice, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.mip.1darray.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %slice, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_mip_2darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_store_mip v[0:3], v[4:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_mip_2darray(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s, i32 %t, i32 %slice, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.mip.2darray.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, i32 %t, i32 %slice, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_1d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_1d(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.1d.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_2d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_2d(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.2d.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_3d:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_3d(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.3d.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_cube:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_CUBE unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_cube(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.cube.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_1darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_1darray(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.1darray.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_2darray:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_2darray(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.2darray.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_2dmsaa:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_2dmsaa(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.2dmsaa.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}getresinfo_2darraymsaa:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf unorm da{{$}}
|
|
|
|
|
; GFX10: image_get_resinfo v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA_ARRAY unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @getresinfo_2darraymsaa(<8 x i32> inreg %rsrc, i32 %mip) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.getresinfo.2darraymsaa.v4f32.i32(i32 15, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_V1:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v0, v0, s[0:7] dmask:0x8 unorm{{$}}
|
|
|
|
|
; GFX10: image_load v0, v0, s[0:7] dmask:0x8 dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps float @load_1d_V1(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call float @llvm.amdgcn.image.load.1d.f32.i32(i32 8, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret float %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_V2:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:1], v0, s[0:7] dmask:0x9 unorm{{$}}
|
|
|
|
|
; GFX10: image_load v[0:1], v0, s[0:7] dmask:0x9 dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <2 x float> @load_1d_V2(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <2 x float> @llvm.amdgcn.image.load.1d.v2f32.i32(i32 9, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <2 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_1d_V1:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v0, v1, s[0:7] dmask:0x2 unorm{{$}}
|
|
|
|
|
; GFX10: image_store v0, v1, s[0:7] dmask:0x2 dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_1d_V1(<8 x i32> inreg %rsrc, float %vdata, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.f32.i32(float %vdata, i32 2, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_1d_V2:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:1], v2, s[0:7] dmask:0xc unorm{{$}}
|
|
|
|
|
; GFX10: image_store v[0:1], v2, s[0:7] dmask:0xc dim:SQ_RSRC_IMG_1D unorm ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_1d_V2(<8 x i32> inreg %rsrc, <2 x float> %vdata, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.v2f32.i32(<2 x float> %vdata, i32 12, i32 %s, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_glc:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v0, s[0:7] dmask:0xf unorm glc{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm glc ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_1d_glc(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.1d.v4f32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 1)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_slc:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v0, s[0:7] dmask:0xf unorm slc{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm slc ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_1d_slc(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.1d.v4f32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 2)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}load_1d_glc_slc:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_load v[0:3], v0, s[0:7] dmask:0xf unorm glc slc{{$}}
|
|
|
|
|
; GFX10: image_load v[0:3], v0, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm glc slc ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps <4 x float> @load_1d_glc_slc(<8 x i32> inreg %rsrc, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
%v = call <4 x float> @llvm.amdgcn.image.load.1d.v4f32.i32(i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 3)
|
|
|
|
|
ret <4 x float> %v
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_1d_glc:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v4, s[0:7] dmask:0xf unorm glc{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm glc ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_1d_glc(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 1)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_1d_slc:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v4, s[0:7] dmask:0xf unorm slc{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm slc ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_1d_slc(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 2)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; GCN-LABEL: {{^}}store_1d_glc_slc:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GFX6789: image_store v[0:3], v4, s[0:7] dmask:0xf unorm glc slc{{$}}
|
|
|
|
|
; GFX10: image_store v[0:3], v4, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm glc slc ;
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
define amdgpu_ps void @store_1d_glc_slc(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %s) {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.v4f32.i32(<4 x float> %vdata, i32 15, i32 %s, <8 x i32> %rsrc, i32 0, i32 3)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Convert test cases to the dimension-aware intrinsics
Summary:
Also explicitly port over some tests in llvm.amdgcn.image.* that were
missing. Some tests are removed because they no longer apply (i.e.
explicitly testing building an address vector via insertelement).
This is in preparation for the eventual removal of the old-style
intrinsics.
Some additional notes:
- constant-address-space-32bit.ll: change some GCN-NEXT to GCN because
the instruction schedule was subtly altered
- insert_vector_elt.ll: the old test didn't actually test anything,
because %tmp1 was not used; remove the load, because it doesn't work
(Because of the amdgpu_ps calling convention? In any case, it's
orthogonal to what the test claims to be testing.)
Change-Id: Idfa99b6512ad139e755e82b8b89548ab08f0afcf
Reviewers: arsenm, rampitec
Subscribers: MatzeB, qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D48018
llvm-svn: 335229
2018-06-21 21:37:19 +08:00
|
|
|
; GCN-LABEL: {{^}}getresinfo_dmask0:
|
|
|
|
|
; GCN-NOT: image
|
|
|
|
|
; GCN: ; return to shader part epilog
|
|
|
|
|
define amdgpu_ps <4 x float> @getresinfo_dmask0(<8 x i32> inreg %rsrc, <4 x float> %vdata, i32 %mip) #0 {
|
|
|
|
|
main_body:
|
|
|
|
|
%r = call <4 x float> @llvm.amdgcn.image.getresinfo.1d.v4f32.i32(i32 0, i32 %mip, <8 x i32> %rsrc, i32 0, i32 0)
|
|
|
|
|
ret <4 x float> %r
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; Ideally, the register allocator would avoid the wait here
|
|
|
|
|
;
|
|
|
|
|
; GCN-LABEL: {{^}}image_store_wait:
|
2019-05-02 00:40:49 +08:00
|
|
|
; GCN: image_store v[0:3], v4, s[0:7] dmask:0xf
|
AMDGPU: Convert test cases to the dimension-aware intrinsics
Summary:
Also explicitly port over some tests in llvm.amdgcn.image.* that were
missing. Some tests are removed because they no longer apply (i.e.
explicitly testing building an address vector via insertelement).
This is in preparation for the eventual removal of the old-style
intrinsics.
Some additional notes:
- constant-address-space-32bit.ll: change some GCN-NEXT to GCN because
the instruction schedule was subtly altered
- insert_vector_elt.ll: the old test didn't actually test anything,
because %tmp1 was not used; remove the load, because it doesn't work
(Because of the amdgpu_ps calling convention? In any case, it's
orthogonal to what the test claims to be testing.)
Change-Id: Idfa99b6512ad139e755e82b8b89548ab08f0afcf
Reviewers: arsenm, rampitec
Subscribers: MatzeB, qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D48018
llvm-svn: 335229
2018-06-21 21:37:19 +08:00
|
|
|
; SI: s_waitcnt expcnt(0)
|
2019-05-02 00:40:49 +08:00
|
|
|
; GCN: image_load v[0:3], v4, s[8:15] dmask:0xf
|
AMDGPU: Convert test cases to the dimension-aware intrinsics
Summary:
Also explicitly port over some tests in llvm.amdgcn.image.* that were
missing. Some tests are removed because they no longer apply (i.e.
explicitly testing building an address vector via insertelement).
This is in preparation for the eventual removal of the old-style
intrinsics.
Some additional notes:
- constant-address-space-32bit.ll: change some GCN-NEXT to GCN because
the instruction schedule was subtly altered
- insert_vector_elt.ll: the old test didn't actually test anything,
because %tmp1 was not used; remove the load, because it doesn't work
(Because of the amdgpu_ps calling convention? In any case, it's
orthogonal to what the test claims to be testing.)
Change-Id: Idfa99b6512ad139e755e82b8b89548ab08f0afcf
Reviewers: arsenm, rampitec
Subscribers: MatzeB, qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D48018
llvm-svn: 335229
2018-06-21 21:37:19 +08:00
|
|
|
; GCN: s_waitcnt vmcnt(0)
|
2019-05-02 00:40:49 +08:00
|
|
|
; GCN: image_store v[0:3], v4, s[16:23] dmask:0xf
|
AMDGPU: Convert test cases to the dimension-aware intrinsics
Summary:
Also explicitly port over some tests in llvm.amdgcn.image.* that were
missing. Some tests are removed because they no longer apply (i.e.
explicitly testing building an address vector via insertelement).
This is in preparation for the eventual removal of the old-style
intrinsics.
Some additional notes:
- constant-address-space-32bit.ll: change some GCN-NEXT to GCN because
the instruction schedule was subtly altered
- insert_vector_elt.ll: the old test didn't actually test anything,
because %tmp1 was not used; remove the load, because it doesn't work
(Because of the amdgpu_ps calling convention? In any case, it's
orthogonal to what the test claims to be testing.)
Change-Id: Idfa99b6512ad139e755e82b8b89548ab08f0afcf
Reviewers: arsenm, rampitec
Subscribers: MatzeB, qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D48018
llvm-svn: 335229
2018-06-21 21:37:19 +08:00
|
|
|
define amdgpu_ps void @image_store_wait(<8 x i32> inreg %arg, <8 x i32> inreg %arg1, <8 x i32> inreg %arg2, <4 x float> %arg3, i32 %arg4) #0 {
|
|
|
|
|
main_body:
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.v4f32.i32(<4 x float> %arg3, i32 15, i32 %arg4, <8 x i32> %arg, i32 0, i32 0)
|
|
|
|
|
%data = call <4 x float> @llvm.amdgcn.image.load.1d.v4f32.i32(i32 15, i32 %arg4, <8 x i32> %arg1, i32 0, i32 0)
|
|
|
|
|
call void @llvm.amdgcn.image.store.1d.v4f32.i32(<4 x float> %data, i32 15, i32 %arg4, <8 x i32> %arg2, i32 0, i32 0)
|
|
|
|
|
ret void
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
; SI won't merge ds memory operations, because of the signed offset bug, so
|
2019-05-02 00:40:49 +08:00
|
|
|
; we only have check lines for VI+.
|
|
|
|
|
; GFX8910-LABEL: image_load_mmo
|
|
|
|
|
; GFX8910: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0
|
|
|
|
|
; GFX8910: ds_write2_b32 v{{[0-9]+}}, [[ZERO]], [[ZERO]] offset1:4
|
AMDGPU: Convert test cases to the dimension-aware intrinsics
Summary:
Also explicitly port over some tests in llvm.amdgcn.image.* that were
missing. Some tests are removed because they no longer apply (i.e.
explicitly testing building an address vector via insertelement).
This is in preparation for the eventual removal of the old-style
intrinsics.
Some additional notes:
- constant-address-space-32bit.ll: change some GCN-NEXT to GCN because
the instruction schedule was subtly altered
- insert_vector_elt.ll: the old test didn't actually test anything,
because %tmp1 was not used; remove the load, because it doesn't work
(Because of the amdgpu_ps calling convention? In any case, it's
orthogonal to what the test claims to be testing.)
Change-Id: Idfa99b6512ad139e755e82b8b89548ab08f0afcf
Reviewers: arsenm, rampitec
Subscribers: MatzeB, qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D48018
llvm-svn: 335229
2018-06-21 21:37:19 +08:00
|
|
|
define amdgpu_ps float @image_load_mmo(<8 x i32> inreg %rsrc, float addrspace(3)* %lds, <2 x i32> %c) #0 {
|
|
|
|
|
store float 0.000000e+00, float addrspace(3)* %lds
|
|
|
|
|
%c0 = extractelement <2 x i32> %c, i32 0
|
|
|
|
|
%c1 = extractelement <2 x i32> %c, i32 1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
%tex = call float @llvm.amdgcn.image.load.2d.f32.i32(i32 1, i32 %c0, i32 %c1, <8 x i32> %rsrc, i32 0, i32 0)
|
AMDGPU: Convert test cases to the dimension-aware intrinsics
Summary:
Also explicitly port over some tests in llvm.amdgcn.image.* that were
missing. Some tests are removed because they no longer apply (i.e.
explicitly testing building an address vector via insertelement).
This is in preparation for the eventual removal of the old-style
intrinsics.
Some additional notes:
- constant-address-space-32bit.ll: change some GCN-NEXT to GCN because
the instruction schedule was subtly altered
- insert_vector_elt.ll: the old test didn't actually test anything,
because %tmp1 was not used; remove the load, because it doesn't work
(Because of the amdgpu_ps calling convention? In any case, it's
orthogonal to what the test claims to be testing.)
Change-Id: Idfa99b6512ad139e755e82b8b89548ab08f0afcf
Reviewers: arsenm, rampitec
Subscribers: MatzeB, qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D48018
llvm-svn: 335229
2018-06-21 21:37:19 +08:00
|
|
|
%tmp2 = getelementptr float, float addrspace(3)* %lds, i32 4
|
|
|
|
|
store float 0.000000e+00, float addrspace(3)* %tmp2
|
|
|
|
|
ret float %tex
|
|
|
|
|
}
|
|
|
|
|
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.1d.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {float,i32} @llvm.amdgcn.image.load.1d.f32i32.i32(i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare {<2 x float>,i32} @llvm.amdgcn.image.load.1d.v2f32i32.i32(i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.2d.v4f32.i32(i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.2d.v4f32i32.i32(i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.3d.v4f32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.3d.v4f32i32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.cube.v4f32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.cube.v4f32i32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.1darray.v4f32.i32(i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.1darray.v4f32i32.i32(i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.2darray.v4f32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.2darray.v4f32i32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.2dmsaa.v4f32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.2dmsaa.v4f32i32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.2darraymsaa.v4f32.i32(i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.2darraymsaa.v4f32i32.i32(i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.mip.1d.v4f32.i32(i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.mip.2d.v4f32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
[AMDGPU] Add support for TFE/LWE in image intrinsics. 2nd try
TFE and LWE support requires extra result registers that are written in the
event of a failure in order to detect that failure case.
The specific use-case that initiated these changes is sparse texture support.
This means that if image intrinsics are used with either option turned on, the
programmer must ensure that the return type can contain all of the expected
results. This can result in redundant registers since the vector size must be a
power-of-2.
This change takes roughly 6 parts:
1. Modify the instruction defs in tablegen to add new instruction variants that
can accomodate the extra return values.
2. Updates to lowerImage in SIISelLowering.cpp to accomodate setting TFE or LWE
(where the bulk of the work for these instruction types is now done)
3. Extra verification code to catch cases where intrinsics have been used but
insufficient return registers are used.
4. Modification to the adjustWritemask optimisation to account for TFE/LWE being
enabled (requires extra registers to be maintained for error return value).
5. An extra pass to zero initialize the error value return - this is because if
the error does not occur, the register is not written and thus must be zeroed
before use. Also added a new (on by default) option to ensure ALL return values
are zero-initialized that is required for sparse texture support.
6. Disable the inst_combine optimization in the presence of tfe/lwe (later TODO
for this to re-enable and handle correctly).
There's an additional fix now to avoid a dmask=0
For an image intrinsic with tfe where all result channels except tfe
were unused, I was getting an image instruction with dmask=0 and only a
single vgpr result for tfe. That is incorrect because the hardware
assumes there is at least one vgpr result, plus the one for tfe.
Fixed by forcing dmask to 1, which gives the desired two vgpr result
with tfe in the second one.
The TFE or LWE result is returned from the intrinsics using an aggregate
type. Look in the test code provided to see how this works, but in essence IR
code to invoke the intrinsic looks as follows:
%v = call {<4 x float>,i32} @llvm.amdgcn.image.load.1d.v4f32i32.i32(i32 15,
i32 %s, <8 x i32> %rsrc, i32 1, i32 0)
%v.vec = extractvalue {<4 x float>, i32} %v, 0
%v.err = extractvalue {<4 x float>, i32} %v, 1
This re-submit of the change also includes a slight modification in
SIISelLowering.cpp to work-around a compiler bug for the powerpc_le
platform that caused a buildbot failure on a previous submission.
Differential revision: https://reviews.llvm.org/D48826
Change-Id: If222bc03642e76cf98059a6bef5d5bffeda38dda
Work around for ppcle compiler bug
Change-Id: Ie284cf24b2271215be1b9dc95b485fd15000e32b
llvm-svn: 351054
2019-01-14 19:55:24 +08:00
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.mip.1d.v4f32i32.i32(i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare {<4 x float>,i32} @llvm.amdgcn.image.load.mip.2d.v4f32i32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare {<2 x float>,i32} @llvm.amdgcn.image.load.mip.2d.v2f32i32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare {float,i32} @llvm.amdgcn.image.load.mip.2d.f32i32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.mip.3d.v4f32.i32(i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.mip.cube.v4f32.i32(i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.mip.1darray.v4f32.i32(i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.load.mip.2darray.v4f32.i32(i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
|
|
|
|
|
declare void @llvm.amdgcn.image.store.1d.v4f32.i32(<4 x float>, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.2d.v4f32.i32(<4 x float>, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.3d.v4f32.i32(<4 x float>, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.cube.v4f32.i32(<4 x float>, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.1darray.v4f32.i32(<4 x float>, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.2darray.v4f32.i32(<4 x float>, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.2dmsaa.v4f32.i32(<4 x float>, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.2darraymsaa.v4f32.i32(<4 x float>, i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
|
|
|
|
|
declare void @llvm.amdgcn.image.store.mip.1d.v4f32.i32(<4 x float>, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.mip.2d.v4f32.i32(<4 x float>, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.mip.3d.v4f32.i32(<4 x float>, i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.mip.cube.v4f32.i32(<4 x float>, i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.mip.1darray.v4f32.i32(<4 x float>, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.mip.2darray.v4f32.i32(<4 x float>, i32, i32, i32, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.1d.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.2d.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.3d.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.cube.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.1darray.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.2darray.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.2dmsaa.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
declare <4 x float> @llvm.amdgcn.image.getresinfo.2darraymsaa.v4f32.i32(i32, i32, <8 x i32>, i32, i32) #2
|
|
|
|
|
|
|
|
|
|
declare float @llvm.amdgcn.image.load.1d.f32.i32(i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Convert test cases to the dimension-aware intrinsics
Summary:
Also explicitly port over some tests in llvm.amdgcn.image.* that were
missing. Some tests are removed because they no longer apply (i.e.
explicitly testing building an address vector via insertelement).
This is in preparation for the eventual removal of the old-style
intrinsics.
Some additional notes:
- constant-address-space-32bit.ll: change some GCN-NEXT to GCN because
the instruction schedule was subtly altered
- insert_vector_elt.ll: the old test didn't actually test anything,
because %tmp1 was not used; remove the load, because it doesn't work
(Because of the amdgpu_ps calling convention? In any case, it's
orthogonal to what the test claims to be testing.)
Change-Id: Idfa99b6512ad139e755e82b8b89548ab08f0afcf
Reviewers: arsenm, rampitec
Subscribers: MatzeB, qcolombet, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D48018
llvm-svn: 335229
2018-06-21 21:37:19 +08:00
|
|
|
declare float @llvm.amdgcn.image.load.2d.f32.i32(i32, i32, i32, <8 x i32>, i32, i32) #1
|
AMDGPU: Dimension-aware image intrinsics
Summary:
These new image intrinsics contain the texture type as part of
their name and have each component of the address/coordinate as
individual parameters.
This is a preparatory step for implementing the A16 feature, where
coordinates are passed as half-floats or -ints, but the Z compare
value and texel offsets are still full dwords, making it difficult
or impossible to distinguish between A16 on or off in the old-style
intrinsics.
Additionally, these intrinsics pass the 'texfailpolicy' and
'cachectrl' as i32 bit fields to reduce operand clutter and allow
for future extensibility.
v2:
- gather4 supports 2darray images
- fix a bug with 1D images on SI
Change-Id: I099f309e0a394082a5901ea196c3967afb867f04
Reviewers: arsenm, rampitec, b-sumner
Subscribers: kzhuravl, wdng, yaxunl, dstuttard, tpr, llvm-commits, t-tye
Differential Revision: https://reviews.llvm.org/D44939
llvm-svn: 329166
2018-04-04 18:58:54 +08:00
|
|
|
declare <2 x float> @llvm.amdgcn.image.load.1d.v2f32.i32(i32, i32, <8 x i32>, i32, i32) #1
|
|
|
|
|
declare void @llvm.amdgcn.image.store.1d.f32.i32(float, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
declare void @llvm.amdgcn.image.store.1d.v2f32.i32(<2 x float>, i32, i32, <8 x i32>, i32, i32) #0
|
|
|
|
|
|
|
|
|
|
attributes #0 = { nounwind }
|
|
|
|
|
attributes #1 = { nounwind readonly }
|
|
|
|
|
attributes #2 = { nounwind readnone }
|