[OpenCL] Set half-precision Div and Sqrt accuracy #179621
Open
+62
−16
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
OpenCL spec relaxed half-precision divide to 1 ULP and sqrt to 1.5 ULP in KhronosGroup/OpenCL-Docs#1293 KhronosGroup/OpenCL-Docs#1386
llvmbot
added
clang
Contributor
There was a problem hiding this comment.
Pull request overview
This PR updates the OpenCL code generation to reflect the relaxed accuracy requirements for half-precision floating-point operations introduced in OpenCL 3.0. The changes set half-precision division to 1 ULP (Unit in the Last Place) and sqrt to 1.5 ULP, aligning with the updated OpenCL specification.
Changes:
- Modified
SetSqrtFPAccuracyandSetDivFPAccuracyfunctions to support half-precision types and apply appropriate accuracy values (1.5 ULP for sqrt, 1 ULP for division) - Updated test expectations to verify half-precision operations now have fpmath metadata attached
- Added new test cases for half-precision division and sqrt operations
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
| clang/lib/CodeGen/CGExpr.cpp | Implements half-precision support in sqrt and division accuracy functions with type-specific ULP values |
| clang/test/CodeGenOpenCL/sqrt-fpmath.cl | Updates test expectations to verify fpmath metadata is now applied to half-precision sqrt calls |
| clang/test/CodeGenOpenCL/fpmath.cl | Adds comprehensive test coverage for half-precision division and sqrt operations with fpmath metadata verification |
Member
|
@llvm/pr-subscribers-clang Author: Wenju He (wenju-he) ChangesOpenCL spec relaxed half-precision divide to 1 ULP and sqrt to 1.5 ULP in KhronosGroup/OpenCL-Docs#1293 KhronosGroup/OpenCL-Docs#1386 Full diff: https://github.com/llvm/llvm-project/pull/179621.diff 3 Files Affected:
diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index 339314ecff9cd..71a14d65c1bfe 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -6979,14 +6979,15 @@ void CodeGenFunction::SetFPAccuracy(llvm::Value *Val, float Accuracy) {
void CodeGenFunction::SetSqrtFPAccuracy(llvm::Value *Val) {
llvm::Type *EltTy = Val->getType()->getScalarType();
- if (!EltTy->isFloatTy())
+ if (!EltTy->isFloatTy() && !EltTy->isHalfTy())
return;
if ((getLangOpts().OpenCL &&
!CGM.getCodeGenOpts().OpenCLCorrectlyRoundedDivSqrt) ||
(getLangOpts().HIP && getLangOpts().CUDAIsDevice &&
!CGM.getCodeGenOpts().HIPCorrectlyRoundedDivSqrt)) {
- // OpenCL v1.1 s7.4: minimum accuracy of single precision / is 3ulp
+ // OpenCL v1.1 s7.4: minimum accuracy of single precision sqrt is 3 ulp.
+ // OpenCL v3.0 s7.4: minimum accuracy of half precision sqrt is 1.5 ulp.
//
// OpenCL v1.2 s5.6.4.2: The -cl-fp32-correctly-rounded-divide-sqrt
// build option allows an application to specify that single precision
@@ -6994,20 +6995,21 @@ void CodeGenFunction::SetSqrtFPAccuracy(llvm::Value *Val) {
// source are correctly rounded.
//
// TODO: CUDA has a prec-sqrt flag
- SetFPAccuracy(Val, 3.0f);
+ SetFPAccuracy(Val, EltTy->isFloatTy() ? 3.0f : 1.5f);
}
}
void CodeGenFunction::SetDivFPAccuracy(llvm::Value *Val) {
llvm::Type *EltTy = Val->getType()->getScalarType();
- if (!EltTy->isFloatTy())
+ if (!EltTy->isFloatTy() && !EltTy->isHalfTy())
return;
if ((getLangOpts().OpenCL &&
!CGM.getCodeGenOpts().OpenCLCorrectlyRoundedDivSqrt) ||
(getLangOpts().HIP && getLangOpts().CUDAIsDevice &&
!CGM.getCodeGenOpts().HIPCorrectlyRoundedDivSqrt)) {
- // OpenCL v1.1 s7.4: minimum accuracy of single precision / is 2.5ulp
+ // OpenCL v1.1 s7.4: minimum accuracy of single precision / is 2.5 ulp.
+ // OpenCL v3.0 s7.4: minimum accuracy of half precision / is 1 ulp.
//
// OpenCL v1.2 s5.6.4.2: The -cl-fp32-correctly-rounded-divide-sqrt
// build option allows an application to specify that single precision
@@ -7015,7 +7017,7 @@ void CodeGenFunction::SetDivFPAccuracy(llvm::Value *Val) {
// source are correctly rounded.
//
// TODO: CUDA has a prec-div flag
- SetFPAccuracy(Val, 2.5f);
+ SetFPAccuracy(Val, EltTy->isFloatTy() ? 2.5f : 1.f);
}
}
diff --git a/clang/test/CodeGenOpenCL/fpmath.cl b/clang/test/CodeGenOpenCL/fpmath.cl
index f3649d52e0091..5915496b3963d 100644
--- a/clang/test/CodeGenOpenCL/fpmath.cl
+++ b/clang/test/CodeGenOpenCL/fpmath.cl
@@ -1,8 +1,44 @@
// RUN: %clang_cc1 %s -emit-llvm -o - -triple spir-unknown-unknown | FileCheck --check-prefix=CHECK --check-prefix=NODIVOPT %s
// RUN: %clang_cc1 %s -emit-llvm -o - -triple spir-unknown-unknown -cl-fp32-correctly-rounded-divide-sqrt | FileCheck --check-prefix=CHECK --check-prefix=DIVOPT %s
-// RUN: %clang_cc1 %s -emit-llvm -o - -DNOFP64 -cl-std=CL1.2 -triple r600-unknown-unknown -target-cpu r600 -pedantic | FileCheck --check-prefix=CHECK-FLT %s
+// RUN: %clang_cc1 %s -emit-llvm -o - -DNOFP16 -DNOFP64 -cl-std=CL1.2 -triple r600-unknown-unknown -target-cpu r600 -pedantic | FileCheck --check-prefix=CHECK-FLT %s
// RUN: %clang_cc1 %s -emit-llvm -o - -DFP64 -cl-std=CL1.2 -triple spir-unknown-unknown -pedantic | FileCheck --check-prefix=CHECK-DBL %s
+#ifndef NOFP16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+typedef __attribute__(( ext_vector_type(4) )) half half4;
+
+half hpscalardiv(half a, half b) {
+ // CHECK: @hpscalardiv
+ // CHECK: fdiv{{.*}},
+ // NODIVOPT: !fpmath ![[MD_HFDIV:[0-9]+]]
+ // DIVOPT-NOT: !fpmath !{{[0-9]+}}
+ return a / b;
+}
+
+half4 hpvectordiv(half4 a, half4 b) {
+ // CHECK: @hpvectordiv
+ // CHECK: fdiv{{.*}},
+ // NODIVOPT: !fpmath ![[MD_HFDIV]]
+ // DIVOPT-NOT: !fpmath !{{[0-9]+}}
+ return a / b;
+}
+
+half elementwise_sqrt_f16(half a) {
+ // CHECK-LABEL: @elementwise_sqrt_f16
+ // NODIVOPT: call half @llvm.sqrt.f16(half %{{.+}}), !fpmath ![[MD_HSQRT:[0-9]+]]
+ // DIVOPT: call half @llvm.sqrt.f16(half %{{.+}}){{$}}
+ return __builtin_elementwise_sqrt(a);
+}
+
+half4 elementwise_sqrt_v4f16(half4 a) {
+ // CHECK-LABEL: @elementwise_sqrt_v4f16
+ // NODIVOPT: call <4 x half> @llvm.sqrt.v4f16(<4 x half> %{{.+}}), !fpmath ![[MD_HSQRT]]
+ // DIVOPT: call <4 x half> @llvm.sqrt.v4f16(<4 x half> %{{.+}}){{$}}
+ return __builtin_elementwise_sqrt(a);
+}
+
+#endif // NOFP16
+
typedef __attribute__(( ext_vector_type(4) )) float float4;
float spscalardiv(float a, float b) {
@@ -30,14 +66,14 @@ float spscalarsqrt(float a) {
float elementwise_sqrt_f32(float a) {
// CHECK-LABEL: @elementwise_sqrt_f32
- // NODIVOPT: call float @llvm.sqrt.f32(float %{{.+}}), !fpmath ![[MD_SQRT:[0-9]+]]
+ // NODIVOPT: call float @llvm.sqrt.f32(float %{{.+}}), !fpmath ![[MD_SQRT]]
// DIVOPT: call float @llvm.sqrt.f32(float %{{.+}}){{$}}
return __builtin_elementwise_sqrt(a);
}
float4 elementwise_sqrt_v4f32(float4 a) {
// CHECK-LABEL: @elementwise_sqrt_v4f32
- // NODIVOPT: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{.+}}), !fpmath ![[MD_SQRT:[0-9]+]]
+ // NODIVOPT: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{.+}}), !fpmath ![[MD_SQRT]]
// DIVOPT: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{.+}}){{$}}
return __builtin_elementwise_sqrt(a);
}
@@ -90,5 +126,7 @@ double4 elementwise_sqrt_v4f64(double4 a) {
#endif
+// NODIVOPT: ![[MD_HFDIV]] = !{float 1.000000e+00}
+// NODIVOPT: ![[MD_HSQRT]] = !{float 1.500000e+00}
// NODIVOPT: ![[MD_FDIV]] = !{float 2.500000e+00}
// NODIVOPT: ![[MD_SQRT]] = !{float 3.000000e+00}
diff --git a/clang/test/CodeGenOpenCL/sqrt-fpmath.cl b/clang/test/CodeGenOpenCL/sqrt-fpmath.cl
index d0637283a7ec1..6f4adf56930ff 100644
--- a/clang/test/CodeGenOpenCL/sqrt-fpmath.cl
+++ b/clang/test/CodeGenOpenCL/sqrt-fpmath.cl
@@ -134,46 +134,52 @@ double16 call_sqrt_v16f64(double16 x) {
}
-// Not for f16
// CHECK-LABEL: define {{.*}} half @call_sqrt_f16(
-// CHECK: call {{.*}} half @_Z4sqrtDh(half noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} half @_Z4sqrtDh(half noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH:\![0-9]+]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} half @_Z4sqrtDh(half noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half call_sqrt_f16(half x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <2 x half> @call_sqrt_v2f16(
-// CHECK: call {{.*}} <2 x half> @_Z4sqrtDv2_Dh(<2 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <2 x half> @_Z4sqrtDv2_Dh(<2 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <2 x half> @_Z4sqrtDv2_Dh(<2 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half2 call_sqrt_v2f16(half2 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <3 x half> @call_sqrt_v3f16(
-// CHECK: call {{.*}} <3 x half> @_Z4sqrtDv3_Dh(<3 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <3 x half> @_Z4sqrtDv3_Dh(<3 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <3 x half> @_Z4sqrtDv3_Dh(<3 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half3 call_sqrt_v3f16(half3 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <4 x half> @call_sqrt_v4f16(
-// CHECK: call {{.*}} <4 x half> @_Z4sqrtDv4_Dh(<4 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <4 x half> @_Z4sqrtDv4_Dh(<4 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <4 x half> @_Z4sqrtDv4_Dh(<4 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half4 call_sqrt_v4f16(half4 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <8 x half> @call_sqrt_v8f16(
-// CHECK: call {{.*}} <8 x half> @_Z4sqrtDv8_Dh(<8 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <8 x half> @_Z4sqrtDv8_Dh(<8 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <8 x half> @_Z4sqrtDv8_Dh(<8 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half8 call_sqrt_v8f16(half8 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <16 x half> @call_sqrt_v16f16(
-// CHECK: call {{.*}} <16 x half> @_Z4sqrtDv16_Dh(<16 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <16 x half> @_Z4sqrtDv16_Dh(<16 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <16 x half> @_Z4sqrtDv16_Dh(<16 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half16 call_sqrt_v16f16(half16 x) {
return sqrt(x);
}
// DEFAULT: [[FPMATH]] = !{float 3.000000e+00}
+// DEFAULT: [[HFPMATH]] = !{float 1.500000e+00}
|
Member
|
@llvm/pr-subscribers-clang-codegen Author: Wenju He (wenju-he) ChangesOpenCL spec relaxed half-precision divide to 1 ULP and sqrt to 1.5 ULP in KhronosGroup/OpenCL-Docs#1293 KhronosGroup/OpenCL-Docs#1386 Full diff: https://github.com/llvm/llvm-project/pull/179621.diff 3 Files Affected:
diff --git a/clang/lib/CodeGen/CGExpr.cpp b/clang/lib/CodeGen/CGExpr.cpp
index 339314ecff9cd..71a14d65c1bfe 100644
--- a/clang/lib/CodeGen/CGExpr.cpp
+++ b/clang/lib/CodeGen/CGExpr.cpp
@@ -6979,14 +6979,15 @@ void CodeGenFunction::SetFPAccuracy(llvm::Value *Val, float Accuracy) {
void CodeGenFunction::SetSqrtFPAccuracy(llvm::Value *Val) {
llvm::Type *EltTy = Val->getType()->getScalarType();
- if (!EltTy->isFloatTy())
+ if (!EltTy->isFloatTy() && !EltTy->isHalfTy())
return;
if ((getLangOpts().OpenCL &&
!CGM.getCodeGenOpts().OpenCLCorrectlyRoundedDivSqrt) ||
(getLangOpts().HIP && getLangOpts().CUDAIsDevice &&
!CGM.getCodeGenOpts().HIPCorrectlyRoundedDivSqrt)) {
- // OpenCL v1.1 s7.4: minimum accuracy of single precision / is 3ulp
+ // OpenCL v1.1 s7.4: minimum accuracy of single precision sqrt is 3 ulp.
+ // OpenCL v3.0 s7.4: minimum accuracy of half precision sqrt is 1.5 ulp.
//
// OpenCL v1.2 s5.6.4.2: The -cl-fp32-correctly-rounded-divide-sqrt
// build option allows an application to specify that single precision
@@ -6994,20 +6995,21 @@ void CodeGenFunction::SetSqrtFPAccuracy(llvm::Value *Val) {
// source are correctly rounded.
//
// TODO: CUDA has a prec-sqrt flag
- SetFPAccuracy(Val, 3.0f);
+ SetFPAccuracy(Val, EltTy->isFloatTy() ? 3.0f : 1.5f);
}
}
void CodeGenFunction::SetDivFPAccuracy(llvm::Value *Val) {
llvm::Type *EltTy = Val->getType()->getScalarType();
- if (!EltTy->isFloatTy())
+ if (!EltTy->isFloatTy() && !EltTy->isHalfTy())
return;
if ((getLangOpts().OpenCL &&
!CGM.getCodeGenOpts().OpenCLCorrectlyRoundedDivSqrt) ||
(getLangOpts().HIP && getLangOpts().CUDAIsDevice &&
!CGM.getCodeGenOpts().HIPCorrectlyRoundedDivSqrt)) {
- // OpenCL v1.1 s7.4: minimum accuracy of single precision / is 2.5ulp
+ // OpenCL v1.1 s7.4: minimum accuracy of single precision / is 2.5 ulp.
+ // OpenCL v3.0 s7.4: minimum accuracy of half precision / is 1 ulp.
//
// OpenCL v1.2 s5.6.4.2: The -cl-fp32-correctly-rounded-divide-sqrt
// build option allows an application to specify that single precision
@@ -7015,7 +7017,7 @@ void CodeGenFunction::SetDivFPAccuracy(llvm::Value *Val) {
// source are correctly rounded.
//
// TODO: CUDA has a prec-div flag
- SetFPAccuracy(Val, 2.5f);
+ SetFPAccuracy(Val, EltTy->isFloatTy() ? 2.5f : 1.f);
}
}
diff --git a/clang/test/CodeGenOpenCL/fpmath.cl b/clang/test/CodeGenOpenCL/fpmath.cl
index f3649d52e0091..5915496b3963d 100644
--- a/clang/test/CodeGenOpenCL/fpmath.cl
+++ b/clang/test/CodeGenOpenCL/fpmath.cl
@@ -1,8 +1,44 @@
// RUN: %clang_cc1 %s -emit-llvm -o - -triple spir-unknown-unknown | FileCheck --check-prefix=CHECK --check-prefix=NODIVOPT %s
// RUN: %clang_cc1 %s -emit-llvm -o - -triple spir-unknown-unknown -cl-fp32-correctly-rounded-divide-sqrt | FileCheck --check-prefix=CHECK --check-prefix=DIVOPT %s
-// RUN: %clang_cc1 %s -emit-llvm -o - -DNOFP64 -cl-std=CL1.2 -triple r600-unknown-unknown -target-cpu r600 -pedantic | FileCheck --check-prefix=CHECK-FLT %s
+// RUN: %clang_cc1 %s -emit-llvm -o - -DNOFP16 -DNOFP64 -cl-std=CL1.2 -triple r600-unknown-unknown -target-cpu r600 -pedantic | FileCheck --check-prefix=CHECK-FLT %s
// RUN: %clang_cc1 %s -emit-llvm -o - -DFP64 -cl-std=CL1.2 -triple spir-unknown-unknown -pedantic | FileCheck --check-prefix=CHECK-DBL %s
+#ifndef NOFP16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+typedef __attribute__(( ext_vector_type(4) )) half half4;
+
+half hpscalardiv(half a, half b) {
+ // CHECK: @hpscalardiv
+ // CHECK: fdiv{{.*}},
+ // NODIVOPT: !fpmath ![[MD_HFDIV:[0-9]+]]
+ // DIVOPT-NOT: !fpmath !{{[0-9]+}}
+ return a / b;
+}
+
+half4 hpvectordiv(half4 a, half4 b) {
+ // CHECK: @hpvectordiv
+ // CHECK: fdiv{{.*}},
+ // NODIVOPT: !fpmath ![[MD_HFDIV]]
+ // DIVOPT-NOT: !fpmath !{{[0-9]+}}
+ return a / b;
+}
+
+half elementwise_sqrt_f16(half a) {
+ // CHECK-LABEL: @elementwise_sqrt_f16
+ // NODIVOPT: call half @llvm.sqrt.f16(half %{{.+}}), !fpmath ![[MD_HSQRT:[0-9]+]]
+ // DIVOPT: call half @llvm.sqrt.f16(half %{{.+}}){{$}}
+ return __builtin_elementwise_sqrt(a);
+}
+
+half4 elementwise_sqrt_v4f16(half4 a) {
+ // CHECK-LABEL: @elementwise_sqrt_v4f16
+ // NODIVOPT: call <4 x half> @llvm.sqrt.v4f16(<4 x half> %{{.+}}), !fpmath ![[MD_HSQRT]]
+ // DIVOPT: call <4 x half> @llvm.sqrt.v4f16(<4 x half> %{{.+}}){{$}}
+ return __builtin_elementwise_sqrt(a);
+}
+
+#endif // NOFP16
+
typedef __attribute__(( ext_vector_type(4) )) float float4;
float spscalardiv(float a, float b) {
@@ -30,14 +66,14 @@ float spscalarsqrt(float a) {
float elementwise_sqrt_f32(float a) {
// CHECK-LABEL: @elementwise_sqrt_f32
- // NODIVOPT: call float @llvm.sqrt.f32(float %{{.+}}), !fpmath ![[MD_SQRT:[0-9]+]]
+ // NODIVOPT: call float @llvm.sqrt.f32(float %{{.+}}), !fpmath ![[MD_SQRT]]
// DIVOPT: call float @llvm.sqrt.f32(float %{{.+}}){{$}}
return __builtin_elementwise_sqrt(a);
}
float4 elementwise_sqrt_v4f32(float4 a) {
// CHECK-LABEL: @elementwise_sqrt_v4f32
- // NODIVOPT: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{.+}}), !fpmath ![[MD_SQRT:[0-9]+]]
+ // NODIVOPT: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{.+}}), !fpmath ![[MD_SQRT]]
// DIVOPT: call <4 x float> @llvm.sqrt.v4f32(<4 x float> %{{.+}}){{$}}
return __builtin_elementwise_sqrt(a);
}
@@ -90,5 +126,7 @@ double4 elementwise_sqrt_v4f64(double4 a) {
#endif
+// NODIVOPT: ![[MD_HFDIV]] = !{float 1.000000e+00}
+// NODIVOPT: ![[MD_HSQRT]] = !{float 1.500000e+00}
// NODIVOPT: ![[MD_FDIV]] = !{float 2.500000e+00}
// NODIVOPT: ![[MD_SQRT]] = !{float 3.000000e+00}
diff --git a/clang/test/CodeGenOpenCL/sqrt-fpmath.cl b/clang/test/CodeGenOpenCL/sqrt-fpmath.cl
index d0637283a7ec1..6f4adf56930ff 100644
--- a/clang/test/CodeGenOpenCL/sqrt-fpmath.cl
+++ b/clang/test/CodeGenOpenCL/sqrt-fpmath.cl
@@ -134,46 +134,52 @@ double16 call_sqrt_v16f64(double16 x) {
}
-// Not for f16
// CHECK-LABEL: define {{.*}} half @call_sqrt_f16(
-// CHECK: call {{.*}} half @_Z4sqrtDh(half noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} half @_Z4sqrtDh(half noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH:\![0-9]+]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} half @_Z4sqrtDh(half noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half call_sqrt_f16(half x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <2 x half> @call_sqrt_v2f16(
-// CHECK: call {{.*}} <2 x half> @_Z4sqrtDv2_Dh(<2 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <2 x half> @_Z4sqrtDv2_Dh(<2 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <2 x half> @_Z4sqrtDv2_Dh(<2 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half2 call_sqrt_v2f16(half2 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <3 x half> @call_sqrt_v3f16(
-// CHECK: call {{.*}} <3 x half> @_Z4sqrtDv3_Dh(<3 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <3 x half> @_Z4sqrtDv3_Dh(<3 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <3 x half> @_Z4sqrtDv3_Dh(<3 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half3 call_sqrt_v3f16(half3 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <4 x half> @call_sqrt_v4f16(
-// CHECK: call {{.*}} <4 x half> @_Z4sqrtDv4_Dh(<4 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <4 x half> @_Z4sqrtDv4_Dh(<4 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <4 x half> @_Z4sqrtDv4_Dh(<4 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half4 call_sqrt_v4f16(half4 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <8 x half> @call_sqrt_v8f16(
-// CHECK: call {{.*}} <8 x half> @_Z4sqrtDv8_Dh(<8 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <8 x half> @_Z4sqrtDv8_Dh(<8 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <8 x half> @_Z4sqrtDv8_Dh(<8 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half8 call_sqrt_v8f16(half8 x) {
return sqrt(x);
}
// CHECK-LABEL: define {{.*}} <16 x half> @call_sqrt_v16f16(
-// CHECK: call {{.*}} <16 x half> @_Z4sqrtDv16_Dh(<16 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
+// DEFAULT: call {{.*}} <16 x half> @_Z4sqrtDv16_Dh(<16 x half> noundef %{{.+}}) #{{[0-9]+}}, !fpmath [[HFPMATH]]{{$}}
+// CORRECTLYROUNDED: call {{.*}} <16 x half> @_Z4sqrtDv16_Dh(<16 x half> noundef %{{.+}}) #{{[0-9]+$}}{{$}}
half16 call_sqrt_v16f16(half16 x) {
return sqrt(x);
}
// DEFAULT: [[FPMATH]] = !{float 3.000000e+00}
+// DEFAULT: [[HFPMATH]] = !{float 1.500000e+00}
|
arsenm
approved these changes
Feb 4, 2026
svenvh
approved these changes
Feb 4, 2026
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
OpenCL spec relaxed half-precision divide to 1 ULP and sqrt to 1.5 ULP in KhronosGroup/OpenCL-Docs#1293 KhronosGroup/OpenCL-Docs#1386
This can enable target to use hardware rcp instruction for half.