32 lines
876 B
Plaintext
32 lines
876 B
Plaintext
// RUN: %{ispc} --target=avx512skx-x32 --nowrap --x86-asm-syntax=intel --emit-asm -o - %s 2>&1 | FileCheck %s
|
|
// RUN: %{ispc} --target=avx512skx-x64 --nowrap --x86-asm-syntax=intel --emit-asm -o - %s 2>&1 | FileCheck %s
|
|
|
|
// REQUIRES: X86_ENABLED
|
|
// XFAIL: *
|
|
|
|
struct vec4 {
|
|
float V[4];
|
|
};
|
|
|
|
uniform vec4 vmax(const uniform vec4 &V1, const uniform vec4 &V2) {
|
|
uniform vec4 Result;
|
|
|
|
foreach (i = 0 ... 4) {
|
|
Result.V[i] = max(V1.V[i], V2.V[i]);
|
|
}
|
|
|
|
return Result;
|
|
}
|
|
|
|
// CHECK-LABEL: foo:
|
|
// CHECK: vmov{{.}}ps xmm0, xmmword ptr [r{{.*}}]
|
|
// CHECK-NEXT: vmaxps xmm0, xmm0, xmmword ptr [r{{.*}}]
|
|
// CHECK-NEXT: vmov{{.}}ps xmmword ptr [r{{.*}}], xmm0
|
|
export void foo(uniform float A[], uniform float B[]) {
|
|
|
|
uniform vec4 *uniform pA = (uniform vec4 * uniform) A;
|
|
uniform vec4 *uniform pB = (uniform vec4 * uniform) B;
|
|
|
|
*pA = vmax(*pA, *pB);
|
|
}
|