Skip to content

AMDGPU: Builtin & codegen support for v_cvt_scalef32_pk32_{bf|f}16_{bf|fp}6 for gfx950 #117747

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Nov 27, 2024
Merged

AMDGPU: Builtin & codegen support for v_cvt_scalef32_pk32_{bf|f}16_{bf|fp}6 for gfx950 #117747

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions clang/include/clang/Basic/BuiltinsAMDGPU.def
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -588,6 +588,10 @@ TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_f16_fp4, "V2hUifIi", "nc", "fp4-
TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk_bf16_fp4, "V2yUifIi", "nc", "fp4-cvt-scale-insts")
TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk32_f32_fp6, "V32fV6Uif", "nc", "fp6bf6-cvt-scale-insts")
TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk32_f32_bf6, "V32fV6Uif", "nc", "fp6bf6-cvt-scale-insts")
TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk32_f16_fp6, "V32hV6Uif", "nc", "fp6bf6-cvt-scale-insts")
TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk32_bf16_fp6, "V32yV6Uif", "nc", "fp6bf6-cvt-scale-insts")
TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk32_f16_bf6, "V32hV6Uif", "nc", "fp6bf6-cvt-scale-insts")
TARGET_BUILTIN(__builtin_amdgcn_cvt_scalef32_pk32_bf16_bf6, "V32yV6Uif", "nc", "fp6bf6-cvt-scale-insts")

#undef BUILTIN
#undef TARGET_BUILTIN
13 changes: 10 additions & 3 deletions clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950-err.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,9 +18,12 @@ typedef float __attribute__((ext_vector_type(2))) float2;
typedef __bf16 __attribute__((ext_vector_type(2))) bfloat2;
typedef float __attribute__((ext_vector_type(32))) float32;
typedef unsigned int __attribute__((ext_vector_type(6))) uint6;
typedef half __attribute__((ext_vector_type(32))) half32;
typedef __bf16 __attribute__((ext_vector_type(32))) bfloat32;

void test(global uint* out, global uint2* out_v2u32, uint a, uint b, global half2* out_v2f16, global float* out_f32, float scale, global short2* out_v2i16, float src0, float src1,
global float2* out_v2f32, half2 src0_v2f16, bfloat2 src0_v2bf16, global bfloat2* out_v2bf16, global float32* out_v36f32, uint6 src_v6i32) {
global float2* out_v2f32, half2 src0_v2f16, bfloat2 src0_v2bf16, global bfloat2* out_v2bf16, global float32* out_v32f32, uint6 src_v6i32,
global half32 *out_v32f16, global bfloat32 *out_v32bf16) {
*out = __builtin_amdgcn_prng_b32(a); // expected-error{{'__builtin_amdgcn_prng_b32' needs target feature prng-inst}}
*out_v2u32 = __builtin_amdgcn_permlane16_swap(a, b, false, false); // expected-error{{'__builtin_amdgcn_permlane16_swap' needs target feature permlane16-swap}}
*out_v2u32 = __builtin_amdgcn_permlane32_swap(a, b, false, false); // expected-error{{'__builtin_amdgcn_permlane32_swap' needs target feature permlane32-swap}}
Expand All @@ -40,6 +43,10 @@ void test(global uint* out, global uint2* out_v2u32, uint a, uint b, global half
*out = __builtin_amdgcn_cvt_scalef32_pk_fp4_f32(*out, src0, src1, scale, 3); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_fp4_f32' needs target feature fp4-cvt-scale-insts}}
*out_v2f16 = __builtin_amdgcn_cvt_scalef32_pk_f16_fp4(a, scale, 3); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_f16_fp4' needs target feature fp4-cvt-scale-insts}}
*out_v2bf16 = __builtin_amdgcn_cvt_scalef32_pk_bf16_fp4(a, scale, 3); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk_bf16_fp4' needs target feature fp4-cvt-scale-insts}}
*out_v36f32 = __builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_f32_fp6' needs target feature fp6bf6-cvt-scale-insts}}
*out_v36f32 = __builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_f32_bf6' needs target feature fp6bf6-cvt-scale-insts}}
*out_v32f32 = __builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_f32_fp6' needs target feature fp6bf6-cvt-scale-insts}}
*out_v32f32 = __builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_f32_bf6' needs target feature fp6bf6-cvt-scale-insts}}
*out_v32f16 = __builtin_amdgcn_cvt_scalef32_pk32_f16_fp6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_f16_fp6' needs target feature fp6bf6-cvt-scale-insts}}
*out_v32f16 = __builtin_amdgcn_cvt_scalef32_pk32_f16_bf6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_f16_bf6' needs target feature fp6bf6-cvt-scale-insts}}
*out_v32bf16 = __builtin_amdgcn_cvt_scalef32_pk32_bf16_fp6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_bf16_fp6' needs target feature fp6bf6-cvt-scale-insts}}
*out_v32bf16 = __builtin_amdgcn_cvt_scalef32_pk32_bf16_bf6(src_v6i32, scale); // expected-error{{'__builtin_amdgcn_cvt_scalef32_pk32_bf16_bf6' needs target feature fp6bf6-cvt-scale-insts}}
}
52 changes: 52 additions & 0 deletions clang/test/CodeGenOpenCL/builtins-amdgcn-gfx950.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -932,3 +932,55 @@ void test_cvt_scalef32_pk_f32_fp6(global float32* out, uint6 src, float scale)
*out = __builtin_amdgcn_cvt_scalef32_pk32_f32_fp6(src, scale);
*out = __builtin_amdgcn_cvt_scalef32_pk32_f32_bf6(src, scale);
}

// CHECK-LABEL: @test_cvt_scalef32_pk32_f16_fpbf6(
// CHECK-NEXT: entry:
// CHECK-NEXT: [[OUT_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
// CHECK-NEXT: [[SRC_ADDR:%.*]] = alloca <6 x i32>, align 32, addrspace(5)
// CHECK-NEXT: [[SCALE_ADDR:%.*]] = alloca float, align 4, addrspace(5)
// CHECK-NEXT: store ptr addrspace(1) [[OUT:%.*]], ptr addrspace(5) [[OUT_ADDR]], align 8
// CHECK-NEXT: store <6 x i32> [[SRC:%.*]], ptr addrspace(5) [[SRC_ADDR]], align 32
// CHECK-NEXT: store float [[SCALE:%.*]], ptr addrspace(5) [[SCALE_ADDR]], align 4
// CHECK-NEXT: [[TMP0:%.*]] = load <6 x i32>, ptr addrspace(5) [[SRC_ADDR]], align 32
// CHECK-NEXT: [[TMP1:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
// CHECK-NEXT: [[TMP2:%.*]] = call <32 x half> @llvm.amdgcn.cvt.scalef32.pk32.f16.fp6(<6 x i32> [[TMP0]], float [[TMP1]])
// CHECK-NEXT: [[TMP3:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
// CHECK-NEXT: store <32 x half> [[TMP2]], ptr addrspace(1) [[TMP3]], align 64
// CHECK-NEXT: [[TMP4:%.*]] = load <6 x i32>, ptr addrspace(5) [[SRC_ADDR]], align 32
// CHECK-NEXT: [[TMP5:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
// CHECK-NEXT: [[TMP6:%.*]] = call <32 x half> @llvm.amdgcn.cvt.scalef32.pk32.f16.bf6(<6 x i32> [[TMP4]], float [[TMP5]])
// CHECK-NEXT: [[TMP7:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
// CHECK-NEXT: store <32 x half> [[TMP6]], ptr addrspace(1) [[TMP7]], align 64
// CHECK-NEXT: ret void
//
void test_cvt_scalef32_pk32_f16_fpbf6(global half32 *out, uint6 src, float scale)
{
*out = __builtin_amdgcn_cvt_scalef32_pk32_f16_fp6(src, scale);
*out = __builtin_amdgcn_cvt_scalef32_pk32_f16_bf6(src, scale);
}

// CHECK-LABEL: @test_cvt_scalef32_pk32_bf16_fpbf6(
// CHECK-NEXT: entry:
// CHECK-NEXT: [[OUT_ADDR:%.*]] = alloca ptr addrspace(1), align 8, addrspace(5)
// CHECK-NEXT: [[SRC_ADDR:%.*]] = alloca <6 x i32>, align 32, addrspace(5)
// CHECK-NEXT: [[SCALE_ADDR:%.*]] = alloca float, align 4, addrspace(5)
// CHECK-NEXT: store ptr addrspace(1) [[OUT:%.*]], ptr addrspace(5) [[OUT_ADDR]], align 8
// CHECK-NEXT: store <6 x i32> [[SRC:%.*]], ptr addrspace(5) [[SRC_ADDR]], align 32
// CHECK-NEXT: store float [[SCALE:%.*]], ptr addrspace(5) [[SCALE_ADDR]], align 4
// CHECK-NEXT: [[TMP0:%.*]] = load <6 x i32>, ptr addrspace(5) [[SRC_ADDR]], align 32
// CHECK-NEXT: [[TMP1:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
// CHECK-NEXT: [[TMP2:%.*]] = call <32 x bfloat> @llvm.amdgcn.cvt.scalef32.pk32.bf16.fp6(<6 x i32> [[TMP0]], float [[TMP1]])
// CHECK-NEXT: [[TMP3:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
// CHECK-NEXT: store <32 x bfloat> [[TMP2]], ptr addrspace(1) [[TMP3]], align 64
// CHECK-NEXT: [[TMP4:%.*]] = load <6 x i32>, ptr addrspace(5) [[SRC_ADDR]], align 32
// CHECK-NEXT: [[TMP5:%.*]] = load float, ptr addrspace(5) [[SCALE_ADDR]], align 4
// CHECK-NEXT: [[TMP6:%.*]] = call <32 x bfloat> @llvm.amdgcn.cvt.scalef32.pk32.bf16.bf6(<6 x i32> [[TMP4]], float [[TMP5]])
// CHECK-NEXT: [[TMP7:%.*]] = load ptr addrspace(1), ptr addrspace(5) [[OUT_ADDR]], align 8
// CHECK-NEXT: store <32 x bfloat> [[TMP6]], ptr addrspace(1) [[TMP7]], align 64
// CHECK-NEXT: ret void
//
void test_cvt_scalef32_pk32_bf16_fpbf6(global bfloat32 *out, uint6 src, float scale)
{
*out = __builtin_amdgcn_cvt_scalef32_pk32_bf16_fp6(src, scale);
*out = __builtin_amdgcn_cvt_scalef32_pk32_bf16_bf6(src, scale);
}
6 changes: 6 additions & 0 deletions llvm/include/llvm/IR/IntrinsicsAMDGPU.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -700,6 +700,12 @@ def int_amdgcn_cvt_scalef32_pk_bf16_fp4: AMDGPUCvtScaleFP4FP8BF8ToF1632Intrinsic
def int_amdgcn_cvt_scalef32_pk32_f32_fp6 : AMDGPUCvtScaleF32Intrinsic<llvm_v32f32_ty, llvm_v6i32_ty, "cvt_scalef32_pk32_f32_fp6">;
def int_amdgcn_cvt_scalef32_pk32_f32_bf6 : AMDGPUCvtScaleF32Intrinsic<llvm_v32f32_ty, llvm_v6i32_ty, "cvt_scalef32_pk32_f32_bf6">;

// llvm.amdgcn.cvt.scalef32.pk32.f16.fp6 v6i32 src, float scale
def int_amdgcn_cvt_scalef32_pk32_f16_bf6 : AMDGPUCvtScaleF32Intrinsic<llvm_v32f16_ty, llvm_v6i32_ty, "cvt_scalef32_pk32_f16_bf6">;
def int_amdgcn_cvt_scalef32_pk32_bf16_bf6 : AMDGPUCvtScaleF32Intrinsic<llvm_v32bf16_ty, llvm_v6i32_ty, "cvt_scalef32_pk32_bf16_bf6">;
def int_amdgcn_cvt_scalef32_pk32_f16_fp6 : AMDGPUCvtScaleF32Intrinsic<llvm_v32f16_ty, llvm_v6i32_ty, "cvt_scalef32_pk32_f16_fp6">;
def int_amdgcn_cvt_scalef32_pk32_bf16_fp6 : AMDGPUCvtScaleF32Intrinsic<llvm_v32bf16_ty, llvm_v6i32_ty, "cvt_scalef32_pk32_bf16_fp6">;

def int_amdgcn_prng_b32 : DefaultAttrsIntrinsic<
[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]
>, ClangBuiltin<"__builtin_amdgcn_prng_b32">;
Expand Down
4 changes: 4 additions & 0 deletions llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4565,6 +4565,10 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
case Intrinsic::amdgcn_cvt_scalef32_pk_bf16_fp4:
case Intrinsic::amdgcn_cvt_scalef32_pk32_f32_fp6:
case Intrinsic::amdgcn_cvt_scalef32_pk32_f32_bf6:
case Intrinsic::amdgcn_cvt_scalef32_pk32_f16_bf6:
case Intrinsic::amdgcn_cvt_scalef32_pk32_bf16_bf6:
case Intrinsic::amdgcn_cvt_scalef32_pk32_f16_fp6:
case Intrinsic::amdgcn_cvt_scalef32_pk32_bf16_fp6:
case Intrinsic::amdgcn_ashr_pk_i8_i32:
case Intrinsic::amdgcn_ashr_pk_u8_i32:
case Intrinsic::amdgcn_cvt_scalef32_2xpk16_fp6_f32:
Expand Down
8 changes: 4 additions & 4 deletions llvm/lib/Target/AMDGPU/VOP3Instructions.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1066,10 +1066,10 @@ let SubtargetPredicate = HasFP4ConversionScaleInsts, mayRaiseFPException = 0 in
let SubtargetPredicate = HasFP6BF6ConversionScaleInsts, mayRaiseFPException = 0 in {
defm V_CVT_SCALEF32_PK32_F32_FP6 : VOP3Inst<"v_cvt_scalef32_pk32_f32_fp6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32F32_V6I32_F32>, int_amdgcn_cvt_scalef32_pk32_f32_fp6>;
defm V_CVT_SCALEF32_PK32_F32_BF6 : VOP3Inst<"v_cvt_scalef32_pk32_f32_bf6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32F32_V6I32_F32>, int_amdgcn_cvt_scalef32_pk32_f32_bf6>;
defm V_CVT_SCALEF32_PK32_F16_FP6 : VOP3Inst<"v_cvt_scalef32_pk32_f16_fp6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32F16_V6I32_F32>>;
defm V_CVT_SCALEF32_PK32_BF16_FP6 : VOP3Inst<"v_cvt_scalef32_pk32_bf16_fp6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32BF16_V6I32_F32>>;
defm V_CVT_SCALEF32_PK32_F16_BF6 : VOP3Inst<"v_cvt_scalef32_pk32_f16_bf6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32F16_V6I32_F32>>;
defm V_CVT_SCALEF32_PK32_BF16_BF6 : VOP3Inst<"v_cvt_scalef32_pk32_bf16_bf6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32BF16_V6I32_F32>>;
defm V_CVT_SCALEF32_PK32_F16_FP6 : VOP3Inst<"v_cvt_scalef32_pk32_f16_fp6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32F16_V6I32_F32>, int_amdgcn_cvt_scalef32_pk32_f16_fp6>;
defm V_CVT_SCALEF32_PK32_BF16_FP6 : VOP3Inst<"v_cvt_scalef32_pk32_bf16_fp6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32BF16_V6I32_F32>, int_amdgcn_cvt_scalef32_pk32_bf16_fp6>;
defm V_CVT_SCALEF32_PK32_F16_BF6 : VOP3Inst<"v_cvt_scalef32_pk32_f16_bf6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32F16_V6I32_F32>, int_amdgcn_cvt_scalef32_pk32_f16_bf6>;
defm V_CVT_SCALEF32_PK32_BF16_BF6 : VOP3Inst<"v_cvt_scalef32_pk32_bf16_bf6", VOP3_CVT_SCALEF32_PK_F864_Profile<VOP_V32BF16_V6I32_F32>, int_amdgcn_cvt_scalef32_pk32_bf16_bf6>;
}

let SubtargetPredicate = HasF16BF16ToFP6BF6ConversionScaleInsts, mayRaiseFPException = 0 in {
Expand Down
Loading