Skip to content

Commit 8a5c241

Browse files
arsenmarsenm
authored
AMDGPU: Add v_smfmac_f32_32x32x64_bf8_bf8 for gfx950 (#117256)
1 parent 6da8ff8 commit 8a5c241

File tree

13 files changed

+507
-1
lines changed

13 files changed

+507
-1
lines changed

‎clang/include/clang/Basic/BuiltinsAMDGPU.def

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -454,6 +454,7 @@ TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_bf8_bf8, "V4fV4iV8iV4fiIiIi
454454
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
455455
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
456456
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
457+
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8, "V16fV4iV8iV16fiIiIi", "nc", "gfx950-insts")
457458

458459
//===----------------------------------------------------------------------===//
459460
// GFX12+ only builtins.

‎clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -538,4 +538,11 @@ void test_smfmac_f32_16x16x128_fp8_fp8(global v4f* out, v4i a, v8i b, v4f c, int
538538
*out = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8(a, b, c, idx, 0, 0);
539539
}
540540

541+
// CHECK-GFX950-LABEL: @test_smfmac_f32_32x32x64_bf8_bf8
542+
// CHECK-GFX950: call <16 x float> @llvm.amdgcn.smfmac.f32.32x32x64.bf8.bf8(<4 x i32> %a, <8 x i32> %b, <16 x float> %c, i32 %idx, i32 0, i32 0)
543+
void test_smfmac_f32_32x32x64_bf8_bf8(global v16f* out, v4i a, v8i b, v16f c, int idx)
544+
{
545+
*out = __builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8(a, b, c, idx, 0, 0);
546+
}
547+
541548
#endif

‎clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -124,3 +124,9 @@ void test_smfmac_f32_16x16x128_fp8_fp8(global float4* out, int4 a, int8 b, float
124124
*out = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8(a, b, c, idx, d, 0); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8' must be a constant integer}}
125125
*out = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8(a, b, c, idx, 0, d); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8' must be a constant integer}}
126126
}
127+
128+
void test_smfmac_f32_32x32x64_bf8_bf8(global float16* out, int4 a, int8 b, float16 c, int idx, int d)
129+
{
130+
*out = __builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8(a, b, c, idx, d, 0); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8' must be a constant integer}}
131+
*out = __builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8(a, b, c, idx, 0, d); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8' must be a constant integer}}
132+
}

‎clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950.cl

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -44,6 +44,7 @@ void test(__global float4* out0, half8 a0, half8 b0, float4 c0,
4444
*out12 = __builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8(a12, b12, c12, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8' needs target feature gfx950-insts}}
4545
*out12 = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8(a12, b12, c12, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8' needs target feature gfx950-insts}}
4646
*out12 = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8(a12, b12, c12, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8' needs target feature gfx950-insts}}
47+
*out13 = __builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8(a13, b13, c13, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8' needs target feature gfx950-insts}}
4748
*out14 = __builtin_amdgcn_mfma_scale_f32_16x16x128_f8f6f4(a14, b14, c14, 0, 0, 0, d14, 0, e14); // expected-error{{'__builtin_amdgcn_mfma_scale_f32_16x16x128_f8f6f4' needs target feature gfx950-insts}}
4849
*out15 = __builtin_amdgcn_mfma_scale_f32_32x32x64_f8f6f4(a15, b15, c15, 0, 0, 0, d15, 0, e15); // expected-error{{'__builtin_amdgcn_mfma_scale_f32_32x32x64_f8f6f4' needs target feature gfx950-insts}}
4950
}

‎llvm/include/llvm/IR/IntrinsicsAMDGPU.td

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -3162,6 +3162,7 @@ def int_amdgcn_smfmac_f32_16x16x128_bf8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_
31623162
def int_amdgcn_smfmac_f32_16x16x128_bf8_fp8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
31633163
def int_amdgcn_smfmac_f32_16x16x128_fp8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
31643164
def int_amdgcn_smfmac_f32_16x16x128_fp8_fp8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
3165+
def int_amdgcn_smfmac_f32_32x32x64_bf8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v16f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
31653166
}
31663167

31673168
//===----------------------------------------------------------------------===//

‎llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1100,6 +1100,7 @@ bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I) const {
11001100
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
11011101
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8:
11021102
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8:
1103+
case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_bf8:
11031104
return selectSMFMACIntrin(I);
11041105
default:
11051106
return selectImpl(I, *CoverageInfo);
@@ -3551,6 +3552,9 @@ bool AMDGPUInstructionSelector::selectSMFMACIntrin(MachineInstr &MI) const {
35513552
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8:
35523553
Opc = AMDGPU::V_SMFMAC_F32_16X16X128_FP8_FP8_e64;
35533554
break;
3555+
case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_bf8:
3556+
Opc = AMDGPU::V_SMFMAC_F32_32X32X64_BF8_BF8_e64;
3557+
break;
35543558
default:
35553559
llvm_unreachable("unhandled smfmac intrinsic");
35563560
}

‎llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4814,7 +4814,8 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
48144814
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_bf8:
48154815
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
48164816
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8:
4817-
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8: {
4817+
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8:
4818+
case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_bf8: {
48184819
// vdst, srcA, srcB, srcC, idx
48194820
OpdsMapping[0] = getAGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
48204821
OpdsMapping[2] = getVGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);

‎llvm/lib/Target/AMDGPU/SIInstrInfo.td

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2883,6 +2883,7 @@ def VOP_V16F32_V2I32_V4I32_I32 : VOPProfile <[v16f32, v2i32, v4i32, i32]>;
28832883
def VOP_V4I32_V4I32_V8I32_I32 : VOPProfile <[v4i32, v4i32, v8i32, i32]>;
28842884
def VOP_V16I32_V4I32_V8I32_I32 : VOPProfile <[v16i32, v4i32, v8i32, i32]>;
28852885
def VOP_V4F32_V4I32_V8I32_I32 : VOPProfile <[v4f32, v4i32, v8i32, i32]>;
2886+
def VOP_V16F32_V4I32_V8I32_I32 : VOPProfile <[v16f32, v4i32, v8i32, i32]>;
28862887

28872888
def VOP_V4F32_V8F16_V8F16_V4F32 : VOPProfile <[v4f32, v8f16, v8f16, v4f32]>;
28882889
def VOP_V16F32_V8F16_V8F16_V16F32 : VOPProfile <[v16f32, v8f16, v8f16, v16f32]>;

‎llvm/lib/Target/AMDGPU/VOP3PInstructions.td

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -644,6 +644,7 @@ def VOPProfileSMFMAC_I32_16X16X128_I8 : VOPProfileSMFMAC<VOP_V4I32_V4I32_V8I32_I
644644
def VOPProfileSMFMAC_I32_32X32X64_I8 : VOPProfileSMFMAC<VOP_V16I32_V4I32_V8I32_I32, AVDst_512, AVSrc_128, AVSrc_256>;
645645

646646
def VOPProfileSMFMAC_F32_16X16X128_F8 : VOPProfileSMFMAC<VOP_V4F32_V4I32_V8I32_I32, AVDst_128, AVSrc_128, AVSrc_256>;
647+
def VOPProfileSMFMAC_F32_32X32X64_F8 : VOPProfileSMFMAC<VOP_V16F32_V4I32_V8I32_I32, AVDst_512, AVSrc_128, AVSrc_256>;
647648

648649
def VOPProfileMAI_F32_V8F16_X32 : VOPProfileMAI<VOP_V4F32_V8F16_V8F16_V4F32, AISrc_128_f32, ADst_128, AVSrc_128>;
649650
def VOPProfileMAI_F32_V8F16_X32_VCD : VOPProfileMAI<VOP_V4F32_V8F16_V8F16_V4F32, VISrc_128_f32, VDst_128, AVSrc_128>;
@@ -1061,6 +1062,7 @@ defm V_SMFMAC_F32_16X16X128_BF8_BF8 : SMFMACInst<"v_smfmac_f32_16x16x128_bf8_bf8
10611062
defm V_SMFMAC_F32_16X16X128_BF8_FP8 : SMFMACInst<"v_smfmac_f32_16x16x128_bf8_fp8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_bf8_fp8>;
10621063
defm V_SMFMAC_F32_16X16X128_FP8_BF8 : SMFMACInst<"v_smfmac_f32_16x16x128_fp8_bf8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_fp8_bf8>;
10631064
defm V_SMFMAC_F32_16X16X128_FP8_FP8 : SMFMACInst<"v_smfmac_f32_16x16x128_fp8_fp8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_fp8_fp8>;
1065+
defm V_SMFMAC_F32_32X32X64_BF8_BF8 : SMFMACInst<"v_smfmac_f32_32x32x64_bf8_bf8", "F32_32X32X64_F8", int_amdgcn_smfmac_f32_32x32x64_bf8_bf8>;
10641066
}
10651067

10661068
def MAIInstInfoTable : GenericTable {
@@ -2163,6 +2165,7 @@ defm V_SMFMAC_F32_16X16X128_BF8_BF8 : VOP3P_Real_SMFMAC <0x3b, "v_smfmac_f32_16x
21632165
defm V_SMFMAC_F32_16X16X128_BF8_FP8 : VOP3P_Real_SMFMAC <0x3c, "v_smfmac_f32_16x16x128bf8fp8">;
21642166
defm V_SMFMAC_F32_16X16X128_FP8_BF8 : VOP3P_Real_SMFMAC <0x3d, "v_smfmac_f32_16x16x128fp8bf8">;
21652167
defm V_SMFMAC_F32_16X16X128_FP8_FP8 : VOP3P_Real_SMFMAC <0x43, "v_smfmac_f32_16x16x128fp8fp8">;
2168+
defm V_SMFMAC_F32_32X32X64_BF8_BF8 : VOP3P_Real_SMFMAC <0x4b, "v_smfmac_f32_32x32x64bf8bf8">;
21662169

21672170
defm V_PK_FMA_F32 : VOP3P_Real_vi <0x30>;
21682171
defm V_PK_MUL_F32 : VOP3P_Real_vi <0x31>;

‎llvm/test/Analysis/UniformityAnalysis/AMDGPU/intrinsics.ll

Lines changed: 9 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -404,6 +404,15 @@ define amdgpu_kernel void @smfmac_f32_16x16x128_fp8_fp8(<4 x i32> %arg0, <8 x i3
404404
ret void
405405
}
406406

407+
declare <16 x float> @llvm.amdgcn.smfmac.f32.32x32x64.bf8.bf8(<4 x i32>, <8 x i32>, <16 x float>, i32, i32, i32)
408+
409+
; CHECK: DIVERGENT: %result = call <16 x float> @llvm.amdgcn.smfmac.f32.32x32x64.bf8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 %arg3, i32 1, i32 2)
410+
define amdgpu_kernel void @smfmac_f32_32x32x64_bf8_bf8(<4 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 %arg3, ptr addrspace(1) %out) {
411+
%result = call <16 x float> @llvm.amdgcn.smfmac.f32.32x32x64.bf8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <16 x float> %arg2, i32 %arg3, i32 1, i32 2)
412+
store <16 x float> %result, ptr addrspace(1) %out
413+
ret void
414+
}
415+
407416
declare i32 @llvm.amdgcn.ds.swizzle(i32, i32) #1
408417
declare i32 @llvm.amdgcn.permlane16.i32(i32, i32, i32, i32, i1, i1) #1
409418
declare i32 @llvm.amdgcn.permlanex16.i32(i32, i32, i32, i32, i1, i1) #1

0 commit comments

Comments
 (0)