maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/half.ll

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

642 lines
22 KiB
LLVM
Raw Normal View History

[FileCheck] Add -allow-deprecated-dag-overlap to failing llvm tests See https://reviews.llvm.org/D47106 for details. Reviewed By: probinson Differential Revision: https://reviews.llvm.org/D47171 This commit drops that patch's changes to: llvm/test/CodeGen/NVPTX/f16x2-instructions.ll llvm/test/CodeGen/NVPTX/param-load-store.ll For some reason, the dos line endings there prevent me from commiting via the monorepo. A follow-up commit (not via the monorepo) will finish the patch. llvm-svn: 336843
2018-07-12 04:25:49 +08:00
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-amd-amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -enable-var-scope -check-prefixes=GCN,SI %s
; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-amd-amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck -allow-deprecated-dag-overlap -enable-var-scope -check-prefixes=GCN,VI %s
CodeGen: soften f16 type by default instead of marking legal. Actual support for softening f16 operations is still limited, and can be added when it's needed. But Soften is much closer to being a useful thing to try than keeping it Legal when no registers can actually hold such values. Longer term, we probably want something between Soften and Promote semantics for most targets, it'll be more efficient to promote the 4 basic operations to f32 than libcall them. llvm-svn: 213372
2014-07-18 20:41:46 +08:00
[AMDGPU] Add f16 support (VI+) Differential Revision: https://reviews.llvm.org/D25975 llvm-svn: 286753
2016-11-13 15:01:11 +08:00
; half args should be promoted to float for SI and lower.
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}load_f16_arg:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; GCN: s_load_dword [[ARG:s[0-9]+]]
; GCN: v_mov_b32_e32 [[V_ARG:v[0-9]+]], [[ARG]]
; GCN: flat_store_short v{{\[[0-9]+:[0-9]+\]}}, [[V_ARG]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_f16_arg(half addrspace(1)* %out, half %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
store half %arg, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}load_v2f16_arg:
AMDGPU: Use better alignment for kernarg lowering This was just emitting loads with the ABI alignment for the raw type. The true alignment is often better, especially when an illegal vector type was scalarized. The better alignment allows using a scalar load more often. llvm-svn: 333558
2018-05-31 00:17:51 +08:00
; GCN: s_load_dword [[ARG:s[0-9]+]]
; GCN: v_mov_b32_e32 [[V_ARG:v[0-9]+]], [[ARG]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[V_ARG]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_v2f16_arg(<2 x half> addrspace(1)* %out, <2 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
store <2 x half> %arg, <2 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}load_v3f16_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: s_load_dwordx2
; GCN: s_load_dwordx2
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; GCN-NOT: {buffer|flat|global}}_load_
AMDGPU: Use better alignment for kernarg lowering This was just emitting loads with the ABI alignment for the raw type. The true alignment is often better, especially when an illegal vector type was scalarized. The better alignment allows using a scalar load more often. llvm-svn: 333558
2018-05-31 00:17:51 +08:00
AMDGPU: Make v4i16/v4f16 legal Some image loads return these, and it's awkward working around them not being legal. llvm-svn: 334835
2018-06-15 23:15:46 +08:00
; GCN-NOT: {{flat|global}}_load
; GCN-DAG: {{flat|global}}_store_dword
; GCN-DAG: {{flat|global}}_store_short
; GCN-NOT: {{flat|global}}_store
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_v3f16_arg(<3 x half> addrspace(1)* %out, <3 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
store <3 x half> %arg, <3 x half> addrspace(1)* %out
ret void
}
AMDGPU: Make v2i16/v2f16 legal on VI This usually results in better code. Fixes using inline asm with short2, and also fixes having a different ABI for function parameters between VI and gfx9. Partially cleans up the mess used for lowering of the d16 operations. Making v4f16 legal will help clean this up more, but this requires additional work. llvm-svn: 332953
2018-05-22 14:32:10 +08:00
; FIXME: Why not one load?
AMDGPU: Use better alignment for kernarg lowering This was just emitting loads with the ABI alignment for the raw type. The true alignment is often better, especially when an illegal vector type was scalarized. The better alignment allows using a scalar load more often. llvm-svn: 333558
2018-05-31 00:17:51 +08:00
; GCN-LABEL: {{^}}load_v4f16_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN-DAG: s_load_dwordx2 s{{\[}}[[ARG0_LO:[0-9]+]]:[[ARG0_HI:[0-9]+]]{{\]}}, s{{\[[0-9]+:[0-9]+\]}}, {{0x2|0x8}}
AMDGPU: Make v4i16/v4f16 legal Some image loads return these, and it's awkward working around them not being legal. llvm-svn: 334835
2018-06-15 23:15:46 +08:00
; GCN-DAG: v_mov_b32_e32 v[[V_ARG0_LO:[0-9]+]], s[[ARG0_LO]]
; GCN-DAG: v_mov_b32_e32 v[[V_ARG0_HI:[0-9]+]], s[[ARG0_HI]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[V_ARG0_LO]]:[[V_ARG0_HI]]{{\]}}
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_v4f16_arg(<4 x half> addrspace(1)* %out, <4 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
store <4 x half> %arg, <4 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}load_v8f16_arg:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @load_v8f16_arg(<8 x half> addrspace(1)* %out, <8 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
store <8 x half> %arg, <8 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v2f16_arg:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v2f16_arg(<2 x float> addrspace(1)* %out, <2 x half> %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%fpext = fpext <2 x half> %in to <2 x float>
store <2 x float> %fpext, <2 x float> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_f16_to_f32_arg:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_f16_to_f32_arg(float addrspace(1)* %out, half %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext half %arg to float
store float %ext, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v2f16_to_v2f32_arg:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v2f16_to_v2f32_arg(<2 x float> addrspace(1)* %out, <2 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <2 x half> %arg to <2 x float>
store <2 x float> %ext, <2 x float> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v3f16_to_v3f32_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: s_load_dwordx2 s
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
[AMDGPU] Support for v3i32/v3f32 Added support for dwordx3 for most load/store types, but not DS, and not intrinsics yet. SI (gfx6) does not have dwordx3 instructions, so they are not enabled there. Some of this patch is from Matt Arsenault, also of AMD. Differential Revision: https://reviews.llvm.org/D58902 Change-Id: I913ef54f1433a7149da8d72f4af54dbb13436bd9 llvm-svn: 356659
2019-03-21 20:01:21 +08:00
; GCN: s_load_dwordx2 s
; GCN-NOT: _load
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: v_cvt_f32_f16_e32
; GCN-NOT: v_cvt_f32_f16
[AMDGPU] Support for v3i32/v3f32 Added support for dwordx3 for most load/store types, but not DS, and not intrinsics yet. SI (gfx6) does not have dwordx3 instructions, so they are not enabled there. Some of this patch is from Matt Arsenault, also of AMD. Differential Revision: https://reviews.llvm.org/D58902 Change-Id: I913ef54f1433a7149da8d72f4af54dbb13436bd9 llvm-svn: 356659
2019-03-21 20:01:21 +08:00
; GCN-DAG: _store_dwordx3
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v3f16_to_v3f32_arg(<3 x float> addrspace(1)* %out, <3 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <3 x half> %arg to <3 x float>
store <3 x float> %ext, <3 x float> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v4f16_to_v4f32_arg:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v4f16_to_v4f32_arg(<4 x float> addrspace(1)* %out, <4 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <4 x half> %arg to <4 x float>
store <4 x float> %ext, <4 x float> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v8f16_to_v8f32_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: s_load_dwordx4
AMDGPU: Split x8 and x16 vector loads instead of scalarize The one regression in the builtin tests is in the read2 test which now (again) has many extra copies, but this should be solved once the pass is replaced with a DAG combine. llvm-svn: 253974
2015-11-24 20:05:03 +08:00
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx4
; GCN: flat_store_dwordx4
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v8f16_to_v8f32_arg(<8 x float> addrspace(1)* %out, <8 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <8 x half> %arg to <8 x float>
store <8 x float> %ext, <8 x float> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_f16_to_f64_arg:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; GCN: s_load_dword [[ARG:s[0-9]+]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: v_cvt_f32_f16_e32 v[[ARG_F32:[0-9]+]], [[ARG]]
; GCN: v_cvt_f64_f32_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], v[[ARG_F32]]
; GCN: flat_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_f16_to_f64_arg(double addrspace(1)* %out, half %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext half %arg to double
store double %ext, double addrspace(1)* %out
ret void
}
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}extload_v2f16_to_v2f64_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN-DAG: s_load_dword s
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; GCN: s_lshr_b32
AMDGPU: Make v2i16/v2f16 legal on VI This usually results in better code. Fixes using inline asm with short2, and also fixes having a different ABI for function parameters between VI and gfx9. Partially cleans up the mess used for lowering of the d16 operations. Making v4f16 legal will help clean this up more, but this requires additional work. llvm-svn: 332953
2018-05-22 14:32:10 +08:00
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v2f16_to_v2f64_arg(<2 x double> addrspace(1)* %out, <2 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <2 x half> %arg to <2 x double>
store <2 x double> %ext, <2 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v3f16_to_v3f64_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: s_load_dwordx2 s
; GCN: s_load_dwordx2 s
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v3f16_to_v3f64_arg(<3 x double> addrspace(1)* %out, <3 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <3 x half> %arg to <3 x double>
store <3 x double> %ext, <3 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v4f16_to_v4f64_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: s_load_dwordx2 s
; GCN: s_load_dwordx2 s
AMDGPU: Make v2i16/v2f16 legal on VI This usually results in better code. Fixes using inline asm with short2, and also fixes having a different ABI for function parameters between VI and gfx9. Partially cleans up the mess used for lowering of the d16 operations. Making v4f16 legal will help clean this up more, but this requires additional work. llvm-svn: 332953
2018-05-22 14:32:10 +08:00
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v4f16_to_v4f64_arg(<4 x double> addrspace(1)* %out, <4 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <4 x half> %arg to <4 x double>
store <4 x double> %ext, <4 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}extload_v8f16_to_v8f64_arg:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: s_load_dwordx2 s
; GCN: s_load_dwordx4 s
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f32_f16_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN-DAG: v_cvt_f64_f32_e32
; GCN-DAG: v_cvt_f64_f32_e32
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @extload_v8f16_to_v8f64_arg(<8 x double> addrspace(1)* %out, <8 x half> %arg) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%ext = fpext <8 x half> %arg to <8 x double>
store <8 x double> %ext, <8 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_load_store_f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_ushort [[TMP:v[0-9]+]]
; GCN: flat_store_short v{{\[[0-9]+:[0-9]+\]}}, [[TMP]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_load_store_f16(half addrspace(1)* %out, half addrspace(1)* %in) #0 {
[opaque pointer type] Add textual IR support for explicit type parameter to load instruction Essentially the same as the GEP change in r230786. A similar migration script can be used to update test cases, though a few more test case improvements/changes were required this time around: (r229269-r229278) import fileinput import sys import re pat = re.compile(r"((?:=|:|^)\s*load (?:atomic )?(?:volatile )?(.*?))(| addrspace\(\d+\) *)\*($| *(?:%|@|null|undef|blockaddress|getelementptr|addrspacecast|bitcast|inttoptr|\[\[[a-zA-Z]|\{\{).*$)") for line in sys.stdin: sys.stdout.write(re.sub(pat, r"\1, \2\3*\4", line)) Reviewers: rafael, dexonsmith, grosser Differential Revision: http://reviews.llvm.org/D7649 llvm-svn: 230794
2015-02-28 05:17:42 +08:00
%val = load half, half addrspace(1)* %in
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
store half %val, half addrspace(1)* %out
CodeGen: soften f16 type by default instead of marking legal. Actual support for softening f16 operations is still limited, and can be added when it's needed. But Soften is much closer to being a useful thing to try than keeping it Legal when no registers can actually hold such values. Longer term, we probably want something between Soften and Promote semantics for most targets, it'll be more efficient to promote the 4 basic operations to f32 than libcall them. llvm-svn: 213372
2014-07-18 20:41:46 +08:00
ret void
}
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_load_store_v2f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dword [[TMP:v[0-9]+]]
; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[TMP]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_load_store_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <2 x half>, <2 x half> addrspace(1)* %in
store <2 x half> %val, <2 x half> addrspace(1)* %out
CodeGen: soften f16 type by default instead of marking legal. Actual support for softening f16 operations is still limited, and can be added when it's needed. But Soften is much closer to being a useful thing to try than keeping it Legal when no registers can actually hold such values. Longer term, we probably want something between Soften and Promote semantics for most targets, it'll be more efficient to promote the 4 basic operations to f32 than libcall them. llvm-svn: 213372
2014-07-18 20:41:46 +08:00
ret void
}
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_load_store_v4f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx2 [[TMP:v\[[0-9]+:[0-9]+\]]]
; GCN: flat_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[TMP]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_load_store_v4f16(<4 x half> addrspace(1)* %in, <4 x half> addrspace(1)* %out) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <4 x half>, <4 x half> addrspace(1)* %in
store <4 x half> %val, <4 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_load_store_v8f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx4 [[TMP:v\[[0-9]+:[0-9]+\]]]
; GCN: flat_store_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, [[TMP:v\[[0-9]+:[0-9]+\]]]
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_load_store_v8f16(<8 x half> addrspace(1)* %out, <8 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <8 x half>, <8 x half> addrspace(1)* %in
store <8 x half> %val, <8 x half> addrspace(1)* %out
CodeGen: soften f16 type by default instead of marking legal. Actual support for softening f16 operations is still limited, and can be added when it's needed. But Soften is much closer to being a useful thing to try than keeping it Legal when no registers can actually hold such values. Longer term, we probably want something between Soften and Promote semantics for most targets, it'll be more efficient to promote the 4 basic operations to f32 than libcall them. llvm-svn: 213372
2014-07-18 20:41:46 +08:00
ret void
}
R600: support fpext/fptrunc operations to and from f16. llvm-svn: 213376
2014-07-18 21:01:37 +08:00
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_extload_f16_to_f32:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_ushort [[LOAD:v[0-9]+]]
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: v_cvt_f32_f16_e32 [[CVT:v[0-9]+]], [[LOAD]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[CVT]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_f16_to_f32(float addrspace(1)* %out, half addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load half, half addrspace(1)* %in
%cvt = fpext half %val to float
store float %cvt, float addrspace(1)* %out
ret void
}
R600: support fpext/fptrunc operations to and from f16. llvm-svn: 213376
2014-07-18 21:01:37 +08:00
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_extload_v2f16_to_v2f32:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dword [[LOAD:v[0-9]+]],
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; SI-DAG: v_cvt_f32_f16_e32 v[[CVT0:[0-9]+]], [[LOAD]]
; SI-DAG: v_lshrrev_b32_e32 [[HI:v[0-9]+]], 16, [[LOAD]]
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; SI: v_cvt_f32_f16_e32 v[[CVT1:[0-9]+]], [[HI]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; VI: v_cvt_f32_f16_e32 v[[CVT0:[0-9]+]], [[LOAD]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; VI: v_cvt_f32_f16_sdwa v[[CVT1:[0-9]+]], [[LOAD]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[CVT0]]:[[CVT1]]{{\]}}
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v2f16_to_v2f32(<2 x float> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <2 x half>, <2 x half> addrspace(1)* %in
%cvt = fpext <2 x half> %val to <2 x float>
store <2 x float> %cvt, <2 x float> addrspace(1)* %out
R600: support fpext/fptrunc operations to and from f16. llvm-svn: 213376
2014-07-18 21:01:37 +08:00
ret void
}
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_extload_v3f16_to_v3f32:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v3f16_to_v3f32(<3 x float> addrspace(1)* %out, <3 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <3 x half>, <3 x half> addrspace(1)* %in
%cvt = fpext <3 x half> %val to <3 x float>
store <3 x float> %cvt, <3 x float> addrspace(1)* %out
ret void
}
R600: support fpext/fptrunc operations to and from f16. llvm-svn: 213376
2014-07-18 21:01:37 +08:00
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_extload_v4f16_to_v4f32:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v4f16_to_v4f32(<4 x float> addrspace(1)* %out, <4 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <4 x half>, <4 x half> addrspace(1)* %in
%cvt = fpext <4 x half> %val to <4 x float>
store <4 x float> %cvt, <4 x float> addrspace(1)* %out
R600: support fpext/fptrunc operations to and from f16. llvm-svn: 213376
2014-07-18 21:01:37 +08:00
ret void
}
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_extload_v8f16_to_v8f32:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v8f16_to_v8f32(<8 x float> addrspace(1)* %out, <8 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <8 x half>, <8 x half> addrspace(1)* %in
%cvt = fpext <8 x half> %val to <8 x float>
store <8 x float> %cvt, <8 x float> addrspace(1)* %out
ret void
}
R600: support fpext/fptrunc operations to and from f16. llvm-svn: 213376
2014-07-18 21:01:37 +08:00
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_extload_v16f16_to_v16f32:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx4
; GCN: flat_load_dwordx4
AMDGPU: Split x8 and x16 vector loads instead of scalarize The one regression in the builtin tests is in the read2 test which now (again) has many extra copies, but this should be solved once the pass is replaced with a DAG combine. llvm-svn: 253974
2015-11-24 20:05:03 +08:00
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx4
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
[AMDGPU] Don't cluster stores Clustering loads has caching benefits, but as far as I know there is no advantage to clustering stores on any AMDGPU subtargets. The disadvantage is that it tends to increase register pressure and restricts scheduling freedom. Differential Revision: https://reviews.llvm.org/D85530
2020-06-03 17:01:12 +08:00
; SI: v_cvt_f32_f16_e32
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; VI: v_cvt_f32_f16_e32
; VI: v_cvt_f32_f16_sdwa
AMDGPU: Split x8 and x16 vector loads instead of scalarize The one regression in the builtin tests is in the read2 test which now (again) has many extra copies, but this should be solved once the pass is replaced with a DAG combine. llvm-svn: 253974
2015-11-24 20:05:03 +08:00
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx4
; GCN: flat_store_dwordx4
; GCN: flat_store_dwordx4
AMDGPU: Split x8 and x16 vector loads instead of scalarize The one regression in the builtin tests is in the read2 test which now (again) has many extra copies, but this should be solved once the pass is replaced with a DAG combine. llvm-svn: 253974
2015-11-24 20:05:03 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v16f16_to_v16f32(<16 x float> addrspace(1)* %out, <16 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <16 x half>, <16 x half> addrspace(1)* %in
%cvt = fpext <16 x half> %val to <16 x float>
store <16 x float> %cvt, <16 x float> addrspace(1)* %out
R600: support fpext/fptrunc operations to and from f16. llvm-svn: 213376
2014-07-18 21:01:37 +08:00
ret void
}
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-LABEL: {{^}}global_extload_f16_to_f64:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_ushort [[LOAD:v[0-9]+]]
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: v_cvt_f32_f16_e32 [[CVT0:v[0-9]+]], [[LOAD]]
; GCN: v_cvt_f64_f32_e32 [[CVT1:v\[[0-9]+:[0-9]+\]]], [[CVT0]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[CVT1]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_f16_to_f64(double addrspace(1)* %out, half addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load half, half addrspace(1)* %in
%cvt = fpext half %val to double
store double %cvt, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_extload_v2f16_to_v2f64:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN-DAG: flat_load_dword [[LOAD:v[0-9]+]],
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; SI-DAG: v_lshrrev_b32_e32 [[HI:v[0-9]+]], 16, [[LOAD]]
; SI-DAG: v_cvt_f32_f16_e32 v[[CVT0:[0-9]+]], [[LOAD]]
; SI-DAG: v_cvt_f32_f16_e32 v[[CVT1:[0-9]+]], [[HI]]
; SI-DAG: v_cvt_f64_f32_e32 v{{\[}}[[CVT2_LO:[0-9]+]]:[[CVT2_HI:[0-9]+]]{{\]}}, v[[CVT0]]
; SI-DAG: v_cvt_f64_f32_e32 v{{\[}}[[CVT3_LO:[0-9]+]]:[[CVT3_HI:[0-9]+]]{{\]}}, v[[CVT1]]
; VI-DAG: v_cvt_f32_f16_sdwa v[[CVT0:[0-9]+]], [[LOAD]] dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
; VI-DAG: v_cvt_f32_f16_e32 v[[CVT1:[0-9]+]], [[LOAD]]
; VI-DAG: v_cvt_f64_f32_e32 v{{\[}}[[CVT3_LO:[0-9]+]]:[[CVT3_HI:[0-9]+]]{{\]}}, v[[CVT0]]
; VI-DAG: v_cvt_f64_f32_e32 v{{\[}}[[CVT2_LO:[0-9]+]]:[[CVT2_HI:[0-9]+]]{{\]}}, v[[CVT1]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN-DAG: flat_store_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[CVT2_LO]]:[[CVT3_HI]]{{\]}}
SelectionDAG: Support Expand of f16 extloads Currently this hits an assert that extload should always be supported, which assumes integer extloads. This moves a hack out of SI's argument lowering and is covered by existing tests. llvm-svn: 247113
2015-09-09 09:12:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v2f16_to_v2f64(<2 x double> addrspace(1)* %out, <2 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <2 x half>, <2 x half> addrspace(1)* %in
%cvt = fpext <2 x half> %val to <2 x double>
store <2 x double> %cvt, <2 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_extload_v3f16_to_v3f64:
AMDGPU: Make v2i64/v2f64 legal types. They can be loaded and stored, so count them as legal. This is mostly to fix a number of common cases for load/store merging. llvm-svn: 254086
2015-11-26 03:58:34 +08:00
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; XSI: flat_load_dwordx2 [[LOAD:v\[[0-9]+:[0-9]+\]]]
AMDGPU: Add VI i16 support Patch By: Wei Ding Differential Revision: https://reviews.llvm.org/D18049 llvm-svn: 286464
2016-11-11 00:02:37 +08:00
; XSI: v_cvt_f32_f16_e32
; XSI: v_cvt_f32_f16_e32
; XSI-DAG: v_lshrrev_b32_e32 {{v[0-9]+}}, 16, {{v[0-9]+}}
; XSI: v_cvt_f32_f16_e32
; XSI-NOT: v_cvt_f32_f16
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; XVI: flat_load_dwordx2 [[LOAD:v\[[0-9]+:[0-9]+\]]]
AMDGPU: Add VI i16 support Patch By: Wei Ding Differential Revision: https://reviews.llvm.org/D18049 llvm-svn: 286464
2016-11-11 00:02:37 +08:00
; XVI: v_cvt_f32_f16_e32
; XVI: v_cvt_f32_f16_e32
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; XVI: v_cvt_f32_f16_sdwa
AMDGPU: Add VI i16 support Patch By: Wei Ding Differential Revision: https://reviews.llvm.org/D18049 llvm-svn: 286464
2016-11-11 00:02:37 +08:00
; XVI-NOT: v_cvt_f32_f16
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx2 v{{\[}}[[IN_LO:[0-9]+]]:[[IN_HI:[0-9]+]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN: v_cvt_f32_f16_e32
; GCN: v_cvt_f32_f16_e32
; SI: v_cvt_f32_f16_e32
; VI: v_cvt_f32_f16_sdwa
; GCN-NOT: v_cvt_f32_f16
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
; GCN: v_cvt_f64_f32_e32
AMDGPU: Make v2i64/v2f64 legal types. They can be loaded and stored, so count them as legal. This is mostly to fix a number of common cases for load/store merging. llvm-svn: 254086
2015-11-26 03:58:34 +08:00
; GCN-NOT: v_cvt_f64_f32_e32
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-27 03:10:00 +08:00
; GCN-DAG: flat_store_dwordx4
; GCN-DAG: flat_store_dwordx2
AMDGPU: Make v2i64/v2f64 legal types. They can be loaded and stored, so count them as legal. This is mostly to fix a number of common cases for load/store merging. llvm-svn: 254086
2015-11-26 03:58:34 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v3f16_to_v3f64(<3 x double> addrspace(1)* %out, <3 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <3 x half>, <3 x half> addrspace(1)* %in
%cvt = fpext <3 x half> %val to <3 x double>
store <3 x double> %cvt, <3 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_extload_v4f16_to_v4f64:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v4f16_to_v4f64(<4 x double> addrspace(1)* %out, <4 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <4 x half>, <4 x half> addrspace(1)* %in
%cvt = fpext <4 x half> %val to <4 x double>
store <4 x double> %cvt, <4 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_extload_v8f16_to_v8f64:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v8f16_to_v8f64(<8 x double> addrspace(1)* %out, <8 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <8 x half>, <8 x half> addrspace(1)* %in
%cvt = fpext <8 x half> %val to <8 x double>
store <8 x double> %cvt, <8 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_extload_v16f16_to_v16f64:
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_extload_v16f16_to_v16f64(<16 x double> addrspace(1)* %out, <16 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <16 x half>, <16 x half> addrspace(1)* %in
%cvt = fpext <16 x half> %val to <16 x double>
store <16 x double> %cvt, <16 x double> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_truncstore_f32_to_f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dword [[LOAD:v[0-9]+]]
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: v_cvt_f16_f32_e32 [[CVT:v[0-9]+]], [[LOAD]]
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_short v{{\[[0-9]+:[0-9]+\]}}, [[CVT]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_truncstore_f32_to_f16(half addrspace(1)* %out, float addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load float, float addrspace(1)* %in
%cvt = fptrunc float %val to half
store half %cvt, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_truncstore_v2f32_to_v2f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx2 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN-DAG: v_cvt_f16_f32_e32 [[CVT0:v[0-9]+]], v[[LO]]
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; SI-DAG: v_cvt_f16_f32_e32 [[CVT1:v[0-9]+]], v[[HI]]
; SI-DAG: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 16, [[CVT1]]
AMDGPU: Allow SIShrinkInstructions to work in non-SSA Immediates can be folded as long as the immediate is a vreg. Also undo commuting instructions if it didn't fold an immediate. llvm-svn: 307575
2017-07-11 03:53:57 +08:00
; SI: v_or_b32_e32 [[PACKED:v[0-9]+]], [[CVT0]], [[SHL]]
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; VI-DAG: v_cvt_f16_f32_sdwa [[CVT1:v[0-9]+]], v[[HI]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD
AMDGPU: Allow SIShrinkInstructions to work in non-SSA Immediates can be folded as long as the immediate is a vreg. Also undo commuting instructions if it didn't fold an immediate. llvm-svn: 307575
2017-07-11 03:53:57 +08:00
; VI: v_or_b32_e32 [[PACKED:v[0-9]+]], [[CVT0]], [[CVT1]]
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN-DAG: flat_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[PACKED]]
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_truncstore_v2f32_to_v2f16(<2 x half> addrspace(1)* %out, <2 x float> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <2 x float>, <2 x float> addrspace(1)* %in
%cvt = fptrunc <2 x float> %val to <2 x half>
store <2 x half> %cvt, <2 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_truncstore_v3f32_to_v3f16:
[AMDGPU] Support for v3i32/v3f32 Added support for dwordx3 for most load/store types, but not DS, and not intrinsics yet. SI (gfx6) does not have dwordx3 instructions, so they are not enabled there. Some of this patch is from Matt Arsenault, also of AMD. Differential Revision: https://reviews.llvm.org/D58902 Change-Id: I913ef54f1433a7149da8d72f4af54dbb13436bd9 llvm-svn: 356659
2019-03-21 20:01:21 +08:00
; GCN: flat_load_dwordx3
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; GCN-DAG: v_cvt_f16_f32_e32
; SI-DAG: v_cvt_f16_f32_e32
; VI-DAG: v_cvt_f16_f32_sdwa
; GCN-DAG: v_cvt_f16_f32_e32
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_short
; GCN: flat_store_dword
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_truncstore_v3f32_to_v3f16(<3 x half> addrspace(1)* %out, <3 x float> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <3 x float>, <3 x float> addrspace(1)* %in
%cvt = fptrunc <3 x float> %val to <3 x half>
store <3 x half> %cvt, <3 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_truncstore_v4f32_to_v4f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx4
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; GCN-DAG: v_cvt_f16_f32_e32
; SI-DAG: v_cvt_f16_f32_e32
; SI-DAG: v_cvt_f16_f32_e32
; VI-DAG: v_cvt_f16_f32_sdwa
; VI-DAG: v_cvt_f16_f32_sdwa
; GCN-DAG: v_cvt_f16_f32_e32
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx2
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_truncstore_v4f32_to_v4f16(<4 x half> addrspace(1)* %out, <4 x float> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <4 x float>, <4 x float> addrspace(1)* %in
%cvt = fptrunc <4 x float> %val to <4 x half>
store <4 x half> %cvt, <4 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_truncstore_v8f32_to_v8f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx4
; GCN: flat_load_dwordx4
[AMDGPU] Resubmit SDWA peephole: enable by default Reviewers: vpykhtin, rampitec, arsenm Subscribers: qcolombet, kzhuravl, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye Differential Revision: https://reviews.llvm.org/D31671 llvm-svn: 299654
2017-04-06 23:03:28 +08:00
; SI: v_cvt_f16_f32_e32
; SI: v_cvt_f16_f32_e32
; SI: v_cvt_f16_f32_e32
; SI: v_cvt_f16_f32_e32
; SI: v_cvt_f16_f32_e32
; SI: v_cvt_f16_f32_e32
; SI: v_cvt_f16_f32_e32
; SI: v_cvt_f16_f32_e32
; VI-DAG: v_cvt_f16_f32_e32
; VI-DAG: v_cvt_f16_f32_e32
; VI-DAG: v_cvt_f16_f32_e32
; VI-DAG: v_cvt_f16_f32_e32
; VI-DAG: v_cvt_f16_f32_sdwa
; VI-DAG: v_cvt_f16_f32_sdwa
; VI-DAG: v_cvt_f16_f32_sdwa
; VI-DAG: v_cvt_f16_f32_sdwa
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_store_dwordx4
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_truncstore_v8f32_to_v8f16(<8 x half> addrspace(1)* %out, <8 x float> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <8 x float>, <8 x float> addrspace(1)* %in
%cvt = fptrunc <8 x float> %val to <8 x half>
store <8 x half> %cvt, <8 x half> addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}global_truncstore_v16f32_to_v16f16:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_dwordx4
; GCN: flat_load_dwordx4
; GCN: flat_load_dwordx4
; GCN: flat_load_dwordx4
Introduce target hook for optimizing register copies Allow a target to do something other than search for copies that will avoid cross register bank copies. Implement for SI by only rewriting the most basic copies, so it should look through anything like a subregister extract. I'm not entirely satisified with this because it seems like eliminating a reg_sequence that isn't fully used should work generically for all targets without them having to override something. However, it seems to be tricky to have a simple implementation of this without rewriting to invalid kinds of subregister copies on some targets. I'm not sure if there is currently a generic way to easily check if a subregister index would be valid for the current use. The current set of TargetRegisterInfo::get*Class functions don't quite behave like I would expect (e.g. getSubClassWithSubReg returns the maximal register class rather than the minimal), so I'm not sure how to make the generic test keep searching if SrcRC:SrcSubReg is a valid replacement for DefRC:DefSubReg. Making the default implementation to check for simple copies breaks a variety of ARM and x86 tests by producing illegal subregister uses. The ARM tests are not actually changed since it should still be using the same sharesSameRegisterFile implementation, this just relaxes them to not check for specific registers. llvm-svn: 248478
2015-09-24 16:36:14 +08:00
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
; GCN-DAG: v_cvt_f16_f32_e32
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN-DAG: flat_store_dwordx4
; GCN-DAG: flat_store_dwordx4
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @global_truncstore_v16f32_to_v16f16(<16 x half> addrspace(1)* %out, <16 x float> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load <16 x float>, <16 x float> addrspace(1)* %in
%cvt = fptrunc <16 x float> %val to <16 x half>
store <16 x half> %cvt, <16 x half> addrspace(1)* %out
ret void
}
; FIXME: Unsafe math should fold conversions away
; GCN-LABEL: {{^}}fadd_f16:
; SI-DAG: v_cvt_f32_f16_e32 v{{[0-9]+}},
; SI-DAG: v_cvt_f32_f16_e32 v{{[0-9]+}},
; SI-DAG: v_cvt_f32_f16_e32 v{{[0-9]+}},
; SI-DAG: v_cvt_f32_f16_e32 v{{[0-9]+}},
; SI: v_add_f32
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @fadd_f16(half addrspace(1)* %out, half %a, half %b) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%add = fadd half %a, %b
store half %add, half addrspace(1)* %out, align 4
ret void
}
; GCN-LABEL: {{^}}fadd_v2f16:
; SI: v_add_f32
; SI: v_add_f32
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @fadd_v2f16(<2 x half> addrspace(1)* %out, <2 x half> %a, <2 x half> %b) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%add = fadd <2 x half> %a, %b
store <2 x half> %add, <2 x half> addrspace(1)* %out, align 8
ret void
}
; GCN-LABEL: {{^}}fadd_v4f16:
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @fadd_v4f16(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%b_ptr = getelementptr <4 x half>, <4 x half> addrspace(1)* %in, i32 1
%a = load <4 x half>, <4 x half> addrspace(1)* %in, align 16
%b = load <4 x half>, <4 x half> addrspace(1)* %b_ptr, align 16
%result = fadd <4 x half> %a, %b
store <4 x half> %result, <4 x half> addrspace(1)* %out, align 16
ret void
}
; GCN-LABEL: {{^}}fadd_v8f16:
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; SI: v_add_f32
; GCN: s_endpgm
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @fadd_v8f16(<8 x half> addrspace(1)* %out, <8 x half> %a, <8 x half> %b) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%add = fadd <8 x half> %a, %b
store <8 x half> %add, <8 x half> addrspace(1)* %out, align 32
ret void
}
; GCN-LABEL: {{^}}test_bitcast_from_half:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_ushort [[TMP:v[0-9]+]]
; GCN-NOT: [[TMP]]
; GCN: flat_store_short v{{\[[0-9]+:[0-9]+\]}}, [[TMP]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @test_bitcast_from_half(half addrspace(1)* %in, i16 addrspace(1)* %out) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load half, half addrspace(1)* %in
%val_int = bitcast half %val to i16
store i16 %val_int, i16 addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}test_bitcast_to_half:
AMDGPU: Switch some half using-tests to use amdhsa The default clover ABI weirdly promotes half to float, which should probably be fixed. llvm-svn: 333730
2018-06-01 15:06:03 +08:00
; GCN: flat_load_ushort [[TMP:v[0-9]+]]
; GCN: flat_store_short v{{\[[0-9]+:[0-9]+\]}}, [[TMP]]
AMDGPU: Mark all unspecified CC functions in tests as amdgpu_kernel Currently the default C calling convention functions are treated the same as compute kernels. Make this explicit so the default calling convention can be changed to a non-kernel. Converted with perl -pi -e 's/define void/define amdgpu_kernel void/' on the relevant test directories (and undoing in one place that actually wanted a non-kernel). llvm-svn: 298444
2017-03-22 05:39:51 +08:00
define amdgpu_kernel void @test_bitcast_to_half(half addrspace(1)* %out, i16 addrspace(1)* %in) #0 {
R600/SI: Fix some cases for load / store of half Mostly argument loads were producing broken zextloads from an FP type. llvm-svn: 239049
2015-06-05 00:00:27 +08:00
%val = load i16, i16 addrspace(1)* %in
%val_fp = bitcast i16 %val to half
store half %val_fp, half addrspace(1)* %out
ret void
}
attributes #0 = { nounwind }