[clang] dbb0418 - [HIP] Add __exp2f to the HIP math header (#236328)

via cfe-commits cfe-commits at lists.llvm.org
Mon Nov 9 18:58:04 PST 2026


Author: Keshav Vinayak Jha
Date: 2026-11-10T08:27:58+05:30
New Revision: dbb0418043e246d31aaab324dbe0c1e7097be391

URL: https://github.com/llvm/llvm-project/commit/dbb0418043e246d31aaab324dbe0c1e7097be391
DIFF: https://github.com/llvm/llvm-project/commit/dbb0418043e246d31aaab324dbe0c1e7097be391.diff

LOG: [HIP] Add __exp2f to the HIP math header (#236328)

Add `__exp2f` to `__clang_hip_math.h`, next to `__expf` and `__exp10f`.
It lowers to `llvm.amdgcn.exp2`, which is a bare `v_exp_f32`.

`exp2f` goes through `llvm.exp2`. For f32, `AMDGPU` wraps `v_exp_f32` in
denormal handling unless f32 denormals are flushed. Kernels that know
their inputs stay above -126, such as the softmax in flash-attention,
only need the instruction. `__expf` and `__exp10f` already wrap the same
builtin, so this fills in the missing member of that family.

**CUDA does not define `__exp2f`, so this is a HIP-only extension.**

Assisted-by: Claude code

Signed-off-by: Keshav Vinayak Jha <keshavvinayakjha at gmail.com>

Added: 
    

Modified: 
    clang/docs/ReleaseNotes.md
    clang/lib/Headers/__clang_hip_math.h
    clang/test/Headers/__clang_hip_math.hip

Removed: 
    


################################################################################
diff  --git a/clang/docs/ReleaseNotes.md b/clang/docs/ReleaseNotes.md
index ae351da057cc8d..9c36e1e1679457 100644
--- a/clang/docs/ReleaseNotes.md
+++ b/clang/docs/ReleaseNotes.md
@@ -1412,6 +1412,11 @@ documentation](CIR/index.md) for more information regarding ClangIR's status and
   link-time failures. `dynamic_cast<void*>` is still allowed because it does not
   use RTTI.
 
+- The HIP math header now provides `__exp2f`, a fast `exp2` intrinsic that
+  lowers to a single `v_exp_f32` without the denormal handling that `exp2f`
+  adds, so results in the denormal range may be flushed to zero. CUDA has no
+  `__exp2f`, so it is a HIP-only extension.
+
 #### CUDA Support
 
 - Fixed missing diagnostics for wrong-side functions reached through constructor

diff  --git a/clang/lib/Headers/__clang_hip_math.h b/clang/lib/Headers/__clang_hip_math.h
index c1fc112c95a4c1..2b673ed4c5b015 100644
--- a/clang/lib/Headers/__clang_hip_math.h
+++ b/clang/lib/Headers/__clang_hip_math.h
@@ -166,6 +166,9 @@ float __exp10f(float __x) {
   return __builtin_amdgcn_exp2f(__log2_10 * __x);
 }
 
+__DEVICE__
+float __exp2f(float __x) { return __builtin_amdgcn_exp2f(__x); }
+
 __DEVICE__
 float __expf(float __x) {
   const float __log2_e = 0x1.715476p+0;

diff  --git a/clang/test/Headers/__clang_hip_math.hip b/clang/test/Headers/__clang_hip_math.hip
index 8440b3fa08d37a..9a004ac06f083b 100644
--- a/clang/test/Headers/__clang_hip_math.hip
+++ b/clang/test/Headers/__clang_hip_math.hip
@@ -8335,6 +8335,40 @@ extern "C" __device__ float test___exp10f(float x) {
   return __exp10f(x);
 }
 
+// DEFAULT-LABEL: define dso_local noundef float @test___exp2f(
+// DEFAULT-SAME: float noundef [[X:%.*]]) local_unnamed_addr #[[ATTR3]] {
+// DEFAULT-NEXT:  [[ENTRY:.*:]]
+// DEFAULT-NEXT:    [[TMP0:%.*]] = tail call contract noundef float @llvm.amdgcn.exp2.f32(float [[X]])
+// DEFAULT-NEXT:    ret float [[TMP0]]
+//
+// FINITEONLY-LABEL: define dso_local noundef nofpclass(nan inf) float @test___exp2f(
+// FINITEONLY-SAME: float noundef nofpclass(nan inf) [[X:%.*]]) local_unnamed_addr #[[ATTR3]] {
+// FINITEONLY-NEXT:  [[ENTRY:.*:]]
+// FINITEONLY-NEXT:    [[TMP0:%.*]] = tail call nnan ninf contract noundef nofpclass(nan inf) float @llvm.amdgcn.exp2.f32(float nofpclass(nan inf) [[X]])
+// FINITEONLY-NEXT:    ret float [[TMP0]]
+//
+// APPROX-LABEL: define dso_local noundef float @test___exp2f(
+// APPROX-SAME: float noundef [[X:%.*]]) local_unnamed_addr #[[ATTR3]] {
+// APPROX-NEXT:  [[ENTRY:.*:]]
+// APPROX-NEXT:    [[TMP0:%.*]] = tail call contract noundef float @llvm.amdgcn.exp2.f32(float [[X]])
+// APPROX-NEXT:    ret float [[TMP0]]
+//
+// NCRDIV-LABEL: define dso_local noundef float @test___exp2f(
+// NCRDIV-SAME: float noundef [[X:%.*]]) local_unnamed_addr #[[ATTR3]] {
+// NCRDIV-NEXT:  [[ENTRY:.*:]]
+// NCRDIV-NEXT:    [[TMP0:%.*]] = tail call contract noundef float @llvm.amdgcn.exp2.f32(float [[X]])
+// NCRDIV-NEXT:    ret float [[TMP0]]
+//
+// AMDGCNSPIRV-LABEL: define spir_func noundef float @test___exp2f(
+// AMDGCNSPIRV-SAME: float noundef [[X:%.*]]) local_unnamed_addr addrspace(4) #[[ATTR3]] {
+// AMDGCNSPIRV-NEXT:  [[ENTRY:.*:]]
+// AMDGCNSPIRV-NEXT:    [[TMP0:%.*]] = tail call contract noundef addrspace(4) float @llvm.amdgcn.exp2.f32(float [[X]])
+// AMDGCNSPIRV-NEXT:    ret float [[TMP0]]
+//
+extern "C" __device__ float test___exp2f(float x) {
+  return __exp2f(x);
+}
+
 // DEFAULT-LABEL: define dso_local noundef float @test___expf(
 // DEFAULT-SAME: float noundef [[X:%.*]]) local_unnamed_addr #[[ATTR3]] {
 // DEFAULT-NEXT:  [[ENTRY:.*:]]


        


More information about the cfe-commits mailing list