clang/test/CodeGenOpenCL/builtins-amdgcn-swmmac-w64.cl

// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 6
// REQUIRES: amdgpu-registered-target
// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1200 -target-feature +wavefrontsize64 -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX1200

typedef int    v2i   __attribute__((ext_vector_type(2)));
typedef int    v4i   __attribute__((ext_vector_type(4)));
typedef float  v4f   __attribute__((ext_vector_type(4)));
typedef half   v4h   __attribute__((ext_vector_type(4)));
typedef short  v4s   __attribute__((ext_vector_type(4)));
typedef half   v8h   __attribute__((ext_vector_type(8)));
typedef short  v8s   __attribute__((ext_vector_type(8)));

// Wave64

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_f16_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <4 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.f16.v4f32.v4f16.v8f16.i32(<4 x half> [[A]], <8 x half> [[B]], <4 x float> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8:![0-9]+]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_f32_16x16x32_f16_w64(global v4f* out, v4h a, v8h b, v4f c, int index)
{
  *out = __builtin_amdgcn_swmmac_f32_16x16x32_f16_w64(a, b, c, index);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_bf16_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <4 x i16> noundef [[A:%.*]], <8 x i16> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf16.v4f32.v4i16.v8i16.i32(<4 x i16> [[A]], <8 x i16> [[B]], <4 x float> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_f32_16x16x32_bf16_w64(global v4f* out, v4s a, v8s b, v4f c, int index)
{
  *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf16_w64(a, b, c, index);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f16_16x16x32_f16_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 8)) [[OUT:%.*]], <4 x half> noundef [[A:%.*]], <8 x half> noundef [[B:%.*]], <4 x half> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x half> @llvm.amdgcn.swmmac.f16.16x16x32.f16.v4f16.v4f16.v8f16.i32(<4 x half> [[A]], <8 x half> [[B]], <4 x half> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x half> [[TMP0]], ptr addrspace(1) [[OUT]], align 8, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_f16_16x16x32_f16_w64(global v4h* out, v4h a, v8h b, v4h c, int index)
{
  *out = __builtin_amdgcn_swmmac_f16_16x16x32_f16_w64(a, b, c, index);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_bf16_16x16x32_bf16_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 8)) [[OUT:%.*]], <4 x i16> noundef [[A:%.*]], <8 x i16> noundef [[B:%.*]], <4 x i16> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x i16> @llvm.amdgcn.swmmac.bf16.16x16x32.bf16.v4i16.v4i16.v8i16.i32(<4 x i16> [[A]], <8 x i16> [[B]], <4 x i16> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x i16> [[TMP0]], ptr addrspace(1) [[OUT]], align 8, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_bf16_16x16x32_bf16_w64(global v4s* out, v4s a, v8s b, v4s c, int index)
{
  *out = __builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w64(a, b, c, index);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_i32_16x16x32_iu8_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], i32 noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <4 x i32> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.swmmac.i32.16x16x32.iu8.v4i32.i32.v2i32.i32(i1 true, i32 [[A]], i1 true, <2 x i32> [[B]], <4 x i32> [[C]], i32 [[INDEX]], i1 true)
// CHECK-GFX1200-NEXT:    store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_i32_16x16x32_iu8_w64(global v4i* out, int a, v2i b, v4i c, int index)
{
  *out = __builtin_amdgcn_swmmac_i32_16x16x32_iu8_w64(true, a, true, b, c, index, true);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_i32_16x16x32_iu4_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], i32 noundef [[A:%.*]], i32 noundef [[B:%.*]], <4 x i32> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.swmmac.i32.16x16x32.iu4.v4i32.i32.i32.i32(i1 true, i32 [[A]], i1 true, i32 [[B]], <4 x i32> [[C]], i32 [[INDEX]], i1 true)
// CHECK-GFX1200-NEXT:    store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_i32_16x16x32_iu4_w64(global v4i* out, int a, int b, v4i c, int index)
{
  *out = __builtin_amdgcn_swmmac_i32_16x16x32_iu4_w64(true, a, true, b, c, index, true);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_i32_16x16x64_iu4_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], i32 noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <4 x i32> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.swmmac.i32.16x16x64.iu4.v4i32.i32.v2i32.i32(i1 true, i32 [[A]], i1 true, <2 x i32> [[B]], <4 x i32> [[C]], i32 [[INDEX]], i1 true)
// CHECK-GFX1200-NEXT:    store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_i32_16x16x64_iu4_w64(global v4i* out, int a, v2i b, v4i c, int index)
{
  *out = __builtin_amdgcn_swmmac_i32_16x16x64_iu4_w64(true, a, true, b, c, index, true);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], i32 noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.fp8.fp8.v4f32.i32.v2i32.i32(i32 [[A]], <2 x i32> [[B]], <4 x float> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w64(global v4f* out, int a, v2i b, v4f c, int index)
{
  *out = __builtin_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w64(a, b, c, index);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], i32 noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.fp8.bf8.v4f32.i32.v2i32.i32(i32 [[A]], <2 x i32> [[B]], <4 x float> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w64(global v4f* out, int a, v2i b, v4f c, int index)
{
  *out = __builtin_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w64(a, b, c, index);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], i32 noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf8.fp8.v4f32.i32.v2i32.i32(i32 [[A]], <2 x i32> [[B]], <4 x float> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w64(global v4f* out, int a, v2i b, v4f c, int index)
{
  *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w64(a, b, c, index);
}

// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w64(
// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], i32 noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {
// CHECK-GFX1200-NEXT:  [[ENTRY:.*:]]
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf8.bf8.v4f32.i32.v2i32.i32(i32 [[A]], <2 x i32> [[B]], <4 x float> [[C]], i32 [[INDEX]])
// CHECK-GFX1200-NEXT:    store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]
// CHECK-GFX1200-NEXT:    ret void
//
void test_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w64(global v4f* out, int a, v2i b, v4f c, int index)
{
  *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w64(a, b, c, index);
}
//.
// CHECK-GFX1200: [[META6:![0-9]+]] = !{!"omnipotent char", [[META7:![0-9]+]], i64 0}
// CHECK-GFX1200: [[META7]] = !{!"Simple C/C++ TBAA"}
// CHECK-GFX1200: [[CHAR_TBAA8]] = !{[[META6]], [[META6]], i64 0}
//.