[flang-commits] [flang] [flang][AMDGPU] Convert math ops to AMD GPU library calls instead of libm calls (PR #99517)

[Krzysztof Drewniak via flang-commits]

================
@@ -0,0 +1,184 @@
+!REQUIRES: amdgpu-registered-target
+!RUN: %flang_fc1 -triple amdgcn-amd-amdhsa -emit-llvm -fopenmp -fopenmp-is-target-device %s -o - | FileCheck %s
+
+subroutine omp_pow_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_pow_f32(float {{.*}}, float {{.*}})
+  y = x ** x
+end subroutine omp_pow_f32
+
+subroutine omp_pow_f64(x, y)
+!$omp declare target
+  real(8) :: x, y
+!CHECK: call double @__ocml_pow_f64(double {{.*}}, double {{.*}})
+  y = x ** x
+end subroutine omp_pow_f64
+
+subroutine omp_sin_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_sin_f32(float {{.*}})
+  y = sin(x)
+end subroutine omp_sin_f32
+
+subroutine omp_sin_f64(x, y)
+!$omp declare target
+  real(8) :: x, y
+!CHECK: call double @__ocml_sin_f64(double {{.*}})
+  y = sin(x)
+end subroutine omp_sin_f64
+
+subroutine omp_abs_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_fabs_f32(float {{.*}})
+  y = abs(x)
+end subroutine omp_abs_f32
+
+subroutine omp_abs_f64(x, y)
+!$omp declare target
+  real(8) :: x, y
+!CHECK: call double @__ocml_fabs_f64(double {{.*}})
+  y = abs(x)
+end subroutine omp_abs_f64
+
+subroutine omp_atan_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_atan_f32(float {{.*}})
+  y = atan(x)
+end subroutine omp_atan_f32
+
+subroutine omp_atan_f64(x, y)
+!$omp declare target
+  real(8) :: x, y
+!CHECK: call double @__ocml_atan_f64(double {{.*}})
+  y = atan(x)
+end subroutine omp_atan_f64
+
+subroutine omp_atan2_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_atan2_f32(float {{.*}}, float {{.*}})
+  y = atan2(x, x)
+end subroutine omp_atan2_f32
+
+subroutine omp_atan2_f64(x, y)
+!$omp declare target
+  real(8) :: x, y
+!CHECK: call double @__ocml_atan2_f64(double {{.*}}, double {{.*}})
+  y = atan2(x ,x)
+end subroutine omp_atan2_f64
+
+subroutine omp_cos_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_cos_f32(float {{.*}})
+  y = cos(x)
+end subroutine omp_cos_f32
+
+subroutine omp_cos_f64(x, y)
+!$omp declare target
+  real(8) :: x, y
+!CHECK: call double @__ocml_cos_f64(double {{.*}})
+  y = cos(x)
+end subroutine omp_cos_f64
+
+subroutine omp_erf_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_erf_f32(float {{.*}})
+  y = erf(x)
+end subroutine omp_erf_f32
+
+subroutine omp_erf_f64(x, y)
+!$omp declare target
+  real(8) :: x, y
+!CHECK: call double @__ocml_erf_f64(double {{.*}})
+  y = erf(x)
+end subroutine omp_erf_f64
+
+subroutine omp_exp_f32(x, y)
+!$omp declare target
+  real :: x, y
+!CHECK: call float @__ocml_exp_f32(float {{.*}})
----------------
krzysz00 wrote:

... Yeah, that one does seem to just wrap the intrinsic going off of an `llvm-dis`

https://github.com/llvm/llvm-project/pull/99517