Use llvm instrinsic by default
Provide amdgpu workaround
v2: drop old amd copyrights
Reviewer: Aaron Watry
Reviewed-by: Vedran Miletić <vedran@miletic.net>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 316588
AMDGPU targets don't have insturction for it,
so it'll be expanded to C * log2 anyway.
v2: use native_log2 instead of the more precise sw implementation
v3: move to amdgpu
v4: drop old AMD copyright
Reviewer: Aaron Watry
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 316587
The implementation uses r600 sepcific intrinsics
LLVM-4 switched to _ro_t and _rw_t image types
Portions of the code can be moved back as more targets/llvm versions add image support
Reviewer: Aaron Watry
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 315341
PTX does not differentiate between read and write fences. Hence, these a
lowered to a mem_fence call. The mem_fence function compiles to the
“member.cta” instruction, which commits all outstanding reads and writes
of a thread such that these become visible to all other threads in the same
CTA (i.e., work-group). The instruction does not differentiate between
global and local memory. Hence, the flags parameter is ignored, except
for deciding whether a “member.cta” instruction should be issued at all.
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 315235
This generates a "bar.sync 0” instruction, which not only causes the
threads to wait, but does acts as a memory fence, as required by
OpenCL. The fence does not differentiate between local and global
memory. Unfortunately, there is no similar instruction which does
not include a memory fence. Hence, we cannot optimize the case
where neither CLK_LOCAL_MEM_FENCE nor CLK_GLOBAL_MEM_FENCE is
passed.
llvm-svn: 315228
for loop would only report status of the last command
v2: return '1'
call test instead of '['
Reviewer: Jeroen Ketema
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 315193
The generated llvm IR mostly identical. char/uchar case is a bit worse.
reviewer: Tom Stellard
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 314701
Broken since r314111
V2: pointed out by Jan Vesely
- Use format() instead of % formating
Patch-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Signed-off-by: Pavel Ondračka <pavel.ondracka@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 314261
Also copy/modify the unary_intrin.inc from math/ to make the
intrinsic declaration somewhat reusable.
Passes CL CTS integer_ops/test_integer_ops popcount tests for CL 1.2
Tested-by on GCN 1.0 (Pitcairn)
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 312854
v2: add vload(half) as well
make helpers amdgpu specific (NVPTX uses different private AS numbering)
use clang builtin on clang >= 6
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <tstellar@redhat.com>
llvm-svn: 312839
Add missing undefs
Make helpers amdgpu specific (NVPTX uses different numbering for private AS)
Use clang builtins on clang >= 6
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <tstellar@redhat.com>
llvm-svn: 312838
Just add the SOURCE_X.Y list to the list of sources if X.Y is the current llvm version.
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <tstellar@redhat.com>
llvm-svn: 312837
This file is only compiled for GCN which all share the same layout
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 312493
We don't have memory fences for r600 so just call group barrier directly
Make sure that barrier is called even with 0 flags
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 312492
This was added in CL 1.1
Tested with a Radeon HD 7850 (Pitcairn) using the CL CTS via:
test_conformance/relationals/test_relationals shuffle_built_in_dual_input
v2: Add half support to shuffle2
Move shuffle2 to misc/
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 312404
This was added in CL 1.1
Tested with a Radeon HD 7850 (Pitcairn) using the CL CTS via:
test_conformance/relationals/test_relationals shuffle_built_in
v2: Add half-precision support to shuffle when available.
Move to misc/ and add section 6.12.12 to clc.h
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 312403
Uses the same mechanism to enable fp16 as we use for fp64 when
processing clc.h
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 312402
Specs require using fences when barrier() is invoked:
"The barrier function will either flush any variables stored in local memory
or queue a memory fence to ensure correct ordering of memory operations to local memory."
and
"The barrier function will queue a memory fence to ensure correct ordering
of memory operations to global memory."
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Aaron Watry <awatry@gmail.com>
Tested-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 311022
I can't reproduce the error that made me add this.
Reported-by: Kim Gräsman <kim.grasman@gmail.com>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Kim Gräsman <kim.grasman@gmail.com>
Reviewed-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 310968
mostly prints and exceptions.
Few behavioral changes are documented in the text
Generated Makefile is identical between python2 and python3
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 309820
also consolidate macros into one file, and rename to clcmacros.h
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 309358
Trivially define native_tan as a redirect to tan.
If there are any targets with a native implementation, we can deal with it later.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Matt Arsenault <arsenm2@gmail.com>
llvm-svn: 295920
The previous location outside the block would crash prepare-builtins
when no the builtins file accidentially not passed on the command line.
llvm-svn: 294916
Ported from the amd-builtins branch.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Matt Arsenault <Matthew.Arsenault@amd.com>
CC: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 292335
Ported from the amd-builtins branch.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Matt Arsenault <Matthew.Arsenault@amd.com>
CC: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 292334
Just use lgamma_r and ignore the value returned in the second argument
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 281565
Ported from the amd-builtins branch, which is itself based on the
Sun Microsystems implementation.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 281564
This macro is currently unused, but I plan to use it shortly.
The previous form did casts of pointers without an address space, which
doesn't work so well for CL 1.x.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 281563
clang (since r280553) allows pointer casts in function overloads,
so we need to disambiguate the second argument.
clang might be smarter about overloads in the future
see https://reviews.llvm.org/D24113, but let's be safe in libclc anyway.
llvm-svn: 280871
This should be uniqued when linking, but right now it creates
a lot of metadata spam listing the same version. This should also
probably be reporting the compiled version of the user program,
which may differ from the library. Currently the library IR files report
1.0 while 1.1/1.2 are the default for user programs.
llvm-svn: 279692
OpenCL 1.0: "Returns y if y < x, otherwise it returns x. If x *and* y
are infinite or NaN, the return values are undefined."
OpenCL 1.1+: "Returns y if y < x, otherwise it returns x. If x *or* y
are infinite or NaN, the return values are undefined."
The 1.0 version is stricter so use that one.
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 276704
Also fix get_global_id to consider offset
No idea how to add this for ptx, so they are stuck with the old get_global_id
implementation.
v2: split to a separate patch
v3: Switch R600 to use implictarg.ptr
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 276443
v2: split into 2 patches
use clang builtins for other intrinsics as well
v3: Fix warnings
Switch r600 to use implictarg.ptr
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 276442
Fixes fdim piglit on Turks
v2: use CL fmax instead of __builtin
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <tom.stellard@amd.com>
llvm-svn: 269807
The scalar float/double function bodies are a direct copy/paste,
aside from the removed (optional) code in float function body that
requires subnormals.
reviewers: jvesely
Patch by: Vedran Miletić <rivanvx@gmail.com>
llvm-svn: 268766
Based on the amd-builtin, but explicitly vectorized for all sizes (not just
float4), and includes a vectorized double implementation.
Passes piglit (float) tests on pitcairn.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 268708
The scalar float/double function bodies are a direct copy/paste
with usage of the CLC wrappers to vectorize them.
This commit also adds in the FP_ILOGB0 and FP_ILOGBNAN macros which are
equal to the results of ilogb(0.0f) and ilogb(float nan) respectively.
v2: Add FP_ILOGB0 and FP_ILOGBNAN definitions
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
v1 Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 261639
Most files remain in a common amdgpu directory.
Also switches barriers to to use convergent,
and use llvm.amdgcn.s.barrier.
This now requires 3.9/trunk to build amdgcn.
llvm-svn: 260777
Also remove definitions if provided by clang (3.7+)
This halves the size of builtin.opt.{cedar,barts}.bc
reviewer: tstellard
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 260303
Make cl_khr_fp64 define per-device.
This patch does not change the generated Makefile (for llvm 3.6, 3.7)
v2: Make the device defines per LLVM version, 'all' for all versions
reviewer: tstellard
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 260302
The float implementation is almost a direct port from the amd-builtins,
but instead of just having a scalar and float4 implementation, it has
a scalar and arbitrary width vector implementation.
The double scalar is also a direct port from AMD's builtin release.
The double vector implementation copies the logic in the float vector
implementation using the values from the double scalar version.
Both have been tested in piglit using tests sent to that project's
mailing list.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 260114
The spec says (section 6.12.3, CL version 1.2):
The macro names given in the following list must use the values
specified. The values shall all be constant expressions suitable
for use in #if preprocessing directives.
This commit addresses the second part of that statement.
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <tom@stellard.net>
CC: Moritz Pflanzer <moritz.pflanzer14@imperial.ac.uk>
CC: Serge Martin <edb+libclc@sigluy.net>
llvm-svn: 249445
The values for the char/short/integer/long minimums were declared with
their actual values, not the definitions from the CL spec (v1.1). As
a result, (-2147483648) was actually being treated as a long by the
compiler, not an int, which caused issues when trying to add/subtract
that value from a vector.
Update the definitions to use the values declared by the spec, and also
add explicit casts for the char/short/int minimums so that the compiler
actually treats them as shorts/chars. Without those casts, they
actually end up stored as integers, and the compiler may end up storing
the INT_MIN as a long.
The compiler can sign extend the values if it needs to convert the
char->short, short->int, or int->long
v2: Add explicit cast for INT_MIN and fix some type-o's and wrapping
in the commit message.
Reported-by: Moritz Pflanzer <moritz.pflanzer14@imperial.ac.uk>
CC: Moritz Pflanzer <moritz.pflanzer14@imperial.ac.uk>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
Signed-off-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 247661
Use the implementation was ported from the AMD builtin library rather
than LLVM Intrinsics.
This has been tested with piglit, OpenCV, and the ocl conformance tests.
llvm-svn: 243131
Passing values less than 0 to the llvm.sqrt() intrinsic results in
undefined behavior, so we need to check the input and return NaN if
is is less than 0.
v2:
- Fix build failures.
llvm-svn: 241906
Using exp2(x * M_LOG2E_F) does not give us accurate enough results for
OpenCL. If you look at the new exp implementation you'll see that
it does multiply the input by M_LOG2E_F, but it still uses the original
input in part of the calculation.
This exp implementation was ported from the AMD builtin library
and has been tested with piglit, OpenCV, and the ocl conformance tests.
llvm-svn: 237229
Not all targets support the intrinsic, so it's better to have a
generic implementation which does not use it.
This exp2 implementation was ported from the AMD builtin library
and has been tested with piglit, OpenCV, and the ocl conformance tests.
llvm-svn: 237228
This implementation was ported from the AMD builtin library
and has been tested with piglit, OpenCV, and the ocl conformance tests.
v2:
- Remove f suffix from constant in double implementations.
- Consolidate implementations using the .cl/.inc approach.
v3:
- Use __CLC_FPSIZE instead of __CLC_FP{32,64}
v4 (Jan Vesely):
- Limit to single precision.
llvm-svn: 236920
This is a generic implementation which just calls rsqrt.
Targets should override this if they want a faster implementation.
v2:
- Alphabettize SOURCES
v3 (Jan Vesely):
Limit to single precision types.
llvm-svn: 236915
Ported from AMD builtin library, passes piglit on Turks.
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 236647
The new implementation was ported from the AMD builtin library
and has been tested with piglit, OpenCV, and the ocl conformance tests.
llvm-svn: 236608
This makes it possible for runtime implementations to disable
subnormal handling at runtime.
When this flag is enabled, decisions about how to handle subnormals
in the library will be controlled by an external variable called
__CLC_SUBNORMAL_DISABLE.
Function implementations should use these new helpers for querying subnormal
support:
__clc_fp16_subnormals_supported();
__clc_fp32_subnormals_supported();
__clc_fp64_subnormals_supported();
In order for the library to link correctly with this feature,
users will be required to either:
1. Insert this variable into the module (if using the LLVM/Clang C++/C APIs).
2. Pass either subnormal_disable.bc or subnormal_use_default.bc to the
linker. These files are distributed with liblclc and installed to
$(installdir). e.g.:
llvm-link -o kernel-out.bc kernel.bc builtins-nosubnormal.bc subnormal_disable.bc
or
llvm-link -o kernel-out.bc kernel.bc builtins-nosubnormal.bc subnormal_use_default.bc
If you do not supply the --enable-runtime-subnormal then the library
behaves the same as it did before this commit.
In addition to these changes, the patch adds helper functions that
should be used when implementing library functions that need
special handling for denormals:
__clc_fp16_subnormals_supported();
__clc_fp32_subnormals_supported();
__clc_fp64_subnormals_supported();
llvm-svn: 235329
This is a generic implementation which just calls sqrt. Targets should
override this if they want a faster implementation.
v2:
- Alphabetize SOURCES
llvm-svn: 232965
This implementation was ported from the AMD builtin library
and has been tested with piglit, OpenCV, and the ocl conformance tests.
v2:
- Remove unnecessary copyright.
llvm-svn: 232964
We need to reinterpret float/double types as uint/ulong in order to
perform the bitwise operations.
This has been tested with piglit, OpenCV, and the ocl conformance tests.
v2:
- Use vector operations rather than splitting vectors into scalar
components.
Reviewed-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 231373
It has been part of the common functions since 1.0
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 231137
Ported from the libclc/amd-builtins branch
v2: Rename sincos_f_piby4 to __libclc__sincosf_piby4
Add cospi(double) implementation instead of using llvm.cos
Notes:
The sincosD_piby4.h file is mostly the same as the builtin implementation
released by AMD. The inline attribute declaration is changed, and M_PI is
used instead of a constant double. Otherwise, the only difference is that
the header explicitly enables the fp64 pragma.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jeroen Ketema <j.ketema@imperial.ac.uk>
CC: Tom Stellard <tom@stellard.net>
CC: Matt Arsenault <Matthew.Arsenault@amd.com>
llvm-svn: 230641
Some functions are implemented using hand-written LLVM IR, and
LLVM assembly format is allowed to change between versions, so we
should require a specific version of LLVM.
llvm-svn: 225041
Including a standard or system header isn't allowed in OpenCL.
The type "size_t" needs to be explicitely defined now.
v2: Use __SIZE_TYPE__ instead of unsigned int.
v3: Define ptrdiff_t and NULL.
Patch-by: Jean-Sébastien Pédron
Reviewed-by: Jeroen Ketema
Reviewed-by: Jan Vesely
llvm-svn: 222235
v2: Fix function declaration
Add range metadata to r600 implementation
v3: change prefix to AMDGPU
Reviewed-by: Tom Stellard <tom@stellard.net>
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 219793
This is a simple implementation which just copies data synchronously.
v2:
- Use size_t.
v3:
- Fix possible race condition by splitting the copy among multiple
work items.
llvm-svn: 219008
We were missing the local versions of the atom_* before
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 217911
Uses the algorithm:
tan(x) = sin(x) / sqrt(1-sin^2(x))
An alternative is:
tan(x) = sin(x) / cos(x)
Which produces more verbose bitcode and longer assembly.
Either way, the generated bitcode seems pretty nasty and a more optimized
but still precise-enough solution is welcome.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 217511
Passes the tests that were submitted to the piglit list
Tested on R600 (Pitcairn)
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jan Vesely <jan.vesely@rutgers.edu>
llvm-svn: 217509
This was previously implemented with a macro and we were using
__builtin_copysign(), which takes double inputs for the float
version of copysign().
Reviewed-and-Tested-by: Aaron Watry <awatry@gmail.com>
llvm-svn: 217045
v2: use space instead of '=' to make Mac happy
Signed-off-by: Jan Vesely <jan.vesely@rutgers.edu>
Reviewed-by: Jeroen Ketema <j.ketema@imperial.ac.uk>
llvm-svn: 216922
These were missing and caused mad24/mul24 with int3/uint3 arg type to fail
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 216321
This generates bitcode which is indistinguishable from what was
hand-written for int32 types in v[load|store]_impl.ll.
v4: Use vec2+scalar for vec3 load/stores to prevent corruption (per Tom)
v3: Also remove unused generic/lib/shared/v[load|store]_impl.ll
v2: (Per Matt Arsenault) Fix alignment issues with vector load stores
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
CC: Matt Arsenault <Matthew.Arsenault@amd.com>
CC: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 216069
These were present in CL 1.0, just not implemented yet.
v2: Use hex values and fix commit message
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Jeroen Ketema <j.ketema@imperial.ac.uk>
CC: Matt Arsenault <Matthew.Arsenault@amd.com>
llvm-svn: 213321
Vector true is -1, not 1, which means we need to use the relational unary
macro instead of the normal unary builtin one.
Signed-off-by: Aaron Watry <awatry@gmail.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
llvm-svn: 213316
Jan Vesely