[libclc] Implement __clc_rsqrt with __ocml_rsqrt_* functions #152436
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Motivation is to upstream use of __ocml_rsqrt_ in https://github.com/intel/llvm/blob/sycl/libclc/libspirv/lib/amdgcn-amdhsa/math/rsqrt.cl llvm-diff shows vectorized calls of llvm.sqrt.v2f32 and fdiv are scalarized: Old: > %2 = tail call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %0), !fpmath !5 > %3 = fdiv contract <2 x float> splat (float 1.000000e+00), %2, !fpmath !4 !4 = !{float 2.500000e+00} !5 = !{float 3.000000e+00} New: < %2 = extractelement <2 x float> %0, i64 0 < %3 = tail call float @__ocml_rsqrt_f32(float noundef %2) < %4 = insertelement <2 x float> poison, float %3, i64 0 < %5 = extractelement <2 x float> %0, i64 1 < %6 = tail call float @__ocml_rsqrt_f32(float noundef %5) < %7 = insertelement <2 x float> %4, float %6, i64 1
|
@arsenm I'm not sure if the change is an improvement, please review. |
You can test this locally with the following command:git-clang-format --diff HEAD~1 HEAD --extensions cl -- libclc/clc/lib/amdgcn/math/clc_rsqrt.clView the diff from clang-format here.diff --git a/libclc/clc/lib/amdgcn/math/clc_rsqrt.cl b/libclc/clc/lib/amdgcn/math/clc_rsqrt.cl
index 4a9ae94b7..69164fff7 100644
--- a/libclc/clc/lib/amdgcn/math/clc_rsqrt.cl
+++ b/libclc/clc/lib/amdgcn/math/clc_rsqrt.cl
@@ -10,13 +10,17 @@
float __ocml_rsqrt_f32(float);
-_CLC_OVERLOAD _CLC_DEF float __clc_rsqrt(float x) { return __ocml_rsqrt_f32(x); }
+_CLC_OVERLOAD _CLC_DEF float __clc_rsqrt(float x) {
+ return __ocml_rsqrt_f32(x);
+}
#ifdef cl_khr_fp64
#pragma OPENCL EXTENSION cl_khr_fp64 : enable
double __ocml_rsqrt_f64(double);
-_CLC_OVERLOAD _CLC_DEF double __clc_rsqrt(double x) { return __ocml_rsqrt_f64(x); }
+_CLC_OVERLOAD _CLC_DEF double __clc_rsqrt(double x) {
+ return __ocml_rsqrt_f64(x);
+}
#endif
@@ -24,9 +28,7 @@ _CLC_OVERLOAD _CLC_DEF double __clc_rsqrt(double x) { return __ocml_rsqrt_f64(x)
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
half __ocml_rsqrt_f16(half);
-_CLC_OVERLOAD _CLC_DEF half __clc_rsqrt(half x) {
- return __ocml_rsqrt_f16(x);
-}
+_CLC_OVERLOAD _CLC_DEF half __clc_rsqrt(half x) { return __ocml_rsqrt_f16(x); }
#endif
|
arsenm
reviewed
Aug 7, 2025
thanks @arsenm. So current generic implementation is good: llvm-project/libclc/clc/lib/generic/math/clc_rsqrt.inc Lines 10 to 13 in eb0ddba Close this PR. |
uditagarwal97
pushed a commit
to intel/llvm
that referenced
this pull request
Aug 8, 2025
…#19722) Before this PR, libspirv-nvptx64--nvidiacl uses __nv_rsqrtf/isinf/copysign in these two functions, and libspirv-amdgcn--amdhsa uses __ocml_rsqrt/copysign. We can upstream use of target built-in to CLC library: Use of __nv_rsqrtf/isinf is to be upstreamed by llvm/llvm-project#150174. Then the changes to libspirv-nvptx64--nvidiacl should be minimized. Generic implementation of __clc_rsqrt is better than __ocml_rsqrt according to llvm/llvm-project#152436. Use of these target built-in scalarize llvm vector intrinsic. So it is likely better to use generic implementation. In addition, use of generic implementation aligns with OpenCL library. Delete libspirv/generic/math/floatn.inc. Delete unused clc/relational/floatn.inc and libspirv/generic/math/minmag.inc.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation is to upstream use of _ocml_rsqrt in
https://github.com/intel/llvm/blob/sycl/libclc/libspirv/lib/amdgcn-amdhsa/math/rsqrt.cl
llvm-diff shows vectorized calls of llvm.sqrt.v2f32 and fdiv are scalarized: Old:
> %2 = tail call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> %0), !fpmath !5
> %3 = fdiv contract <2 x float> splat (float 1.000000e+00), %2, !fpmath !4
!4 = !{float 2.500000e+00}
!5 = !{float 3.000000e+00}
New:
< %2 = extractelement <2 x float> %0, i64 0
< %3 = tail call float @__ocml_rsqrt_f32(float noundef %2)
< %4 = insertelement <2 x float> poison, float %3, i64 0
< %5 = extractelement <2 x float> %0, i64 1
< %6 = tail call float @__ocml_rsqrt_f32(float noundef %5)
< %7 = insertelement <2 x float> %4, float %6, i64 1