Skip to content

Commit 95ddc1a

Browse files
arsenmarsenm
authored
AMDGPU: Add v_smfmac_f32_16x16x64_bf16 for gfx950 (#117211)
1 parent e50eaa2 commit 95ddc1a

File tree

13 files changed

+430
-1
lines changed

13 files changed

+430
-1
lines changed

‎clang/include/clang/Basic/BuiltinsAMDGPU.def

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -446,6 +446,8 @@ TARGET_BUILTIN(__builtin_amdgcn_mfma_i32_32x32x32_i8, "V16iV4iV4iV16iIiIiIi", "n
446446

447447
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_f16, "V4fV8hV16hV4fiIiIi", "nc", "gfx950-insts")
448448
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x32_f16, "V16fV8hV16hV16fiIiIi", "nc", "gfx950-insts")
449+
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x64_bf16, "V4fV8yV16yV4fiIiIi", "nc", "gfx950-insts")
450+
449451
//===----------------------------------------------------------------------===//
450452
// GFX12+ only builtins.
451453
//===----------------------------------------------------------------------===//

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

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -26,6 +26,7 @@ typedef short v16s __attribute__((ext_vector_type(16)));
2626
typedef short v32s __attribute__((ext_vector_type(32)));
2727
typedef double v4d __attribute__((ext_vector_type(4)));
2828
typedef __bf16 v8bf16 __attribute__((ext_vector_type(8)));
29+
typedef __bf16 v16bf16 __attribute__((ext_vector_type(16)));
2930

3031

3132
#ifdef MFMA_GFX908_TESTS
@@ -481,4 +482,11 @@ void test_smfmac_f32_32x32x32_f16(global v16f* out, v8h a, v16h b, v16f c, int i
481482
*out = __builtin_amdgcn_smfmac_f32_32x32x32_f16(a, b, c, idx, 0, 0);
482483
}
483484

485+
// CHECK-GFX950-LABEL: @test_smfmac_f32_16x16x64_bf16
486+
// CHECK-GFX950: call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x64.bf16(<8 x bfloat> %a, <16 x bfloat> %b, <4 x float> %c, i32 %idx, i32 0, i32 0)
487+
void test_smfmac_f32_16x16x64_bf16(global v4f* out, v8bf16 a, v16bf16 b, v4f c, int idx)
488+
{
489+
*out = __builtin_amdgcn_smfmac_f32_16x16x64_bf16(a, b, c, idx, 0, 0);
490+
}
491+
484492
#endif

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

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -6,6 +6,7 @@ typedef float float16 __attribute__((ext_vector_type(16)));
66
typedef half half8 __attribute__((ext_vector_type(8)));
77
typedef half half16 __attribute__((ext_vector_type(16)));
88
typedef __bf16 bfloat8 __attribute__((ext_vector_type(8)));
9+
typedef __bf16 bfloat16 __attribute__((ext_vector_type(16)));
910
typedef int int4 __attribute__((ext_vector_type(4)));
1011
typedef int int8 __attribute__((ext_vector_type(8)));
1112
typedef int int16 __attribute__((ext_vector_type(16)));
@@ -75,3 +76,9 @@ void test_smfmac_f32_32x32x32_f16(global float16* out, half8 a, half16 b, float1
7576
*out = __builtin_amdgcn_smfmac_f32_32x32x32_f16(a, b, c, idx, d, 0); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_32x32x32_f16' must be a constant integer}}
7677
*out = __builtin_amdgcn_smfmac_f32_32x32x32_f16(a, b, c, idx, 0, d); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_32x32x32_f16' must be a constant integer}}
7778
}
79+
80+
void test_smfmac_f32_16x16x64_bf16(global float4* out, bfloat8 a, bfloat16 b, float4 c, int idx, int d)
81+
{
82+
*out = __builtin_amdgcn_smfmac_f32_16x16x64_bf16(a, b, c, idx, d, 0); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x64_bf16' must be a constant integer}}
83+
*out = __builtin_amdgcn_smfmac_f32_16x16x64_bf16(a, b, c, idx, 0, d); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x64_bf16' must be a constant integer}}
84+
}

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

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -36,6 +36,7 @@ void test(__global float4* out0, half8 a0, half8 b0, float4 c0,
3636
*out5 = __builtin_amdgcn_mfma_f32_16x16x32_bf16(a5, b5, c5, 0, 0, 0); // expected-error{{'__builtin_amdgcn_mfma_f32_16x16x32_bf16' needs target feature gfx950-insts}}
3737
*out6 = __builtin_amdgcn_smfmac_f32_16x16x64_f16(a6, b6, c6, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_16x16x64_f16' needs target feature gfx950-insts}}
3838
*out7 = __builtin_amdgcn_smfmac_f32_32x32x32_f16(a7, b7, c7, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_32x32x32_f16' needs target feature gfx950-insts}}
39+
*out8 = __builtin_amdgcn_smfmac_f32_16x16x64_bf16(a8, b8, c8, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_16x16x64_bf16' needs target feature gfx950-insts}}
3940
*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}}
4041
*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}}
4142
}

‎llvm/include/llvm/IR/IntrinsicsAMDGPU.td

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -3154,6 +3154,7 @@ def int_amdgcn_mfma_scale_f32_16x16x128_f8f6f4 : AMDGPUMfmaScaleIntrinsic<llvm_v
31543154
def int_amdgcn_mfma_scale_f32_32x32x64_f8f6f4 : AMDGPUMfmaScaleIntrinsic<llvm_v16f32_ty>;
31553155
def int_amdgcn_smfmac_f32_16x16x64_f16 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v8f16_ty, llvm_v16f16_ty>;
31563156
def int_amdgcn_smfmac_f32_32x32x32_f16 : AMDGPUMSmfmacIntrinsic<llvm_v16f32_ty, llvm_v8f16_ty, llvm_v16f16_ty>;
3157+
def int_amdgcn_smfmac_f32_16x16x64_bf16 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v8bf16_ty, llvm_v16bf16_ty>;
31573158
}
31583159

31593160
//===----------------------------------------------------------------------===//

‎llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1092,6 +1092,7 @@ bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I) const {
10921092
case Intrinsic::amdgcn_smfmac_f32_32x32x32_fp8_fp8:
10931093
case Intrinsic::amdgcn_smfmac_f32_16x16x64_f16:
10941094
case Intrinsic::amdgcn_smfmac_f32_32x32x32_f16:
1095+
case Intrinsic::amdgcn_smfmac_f32_16x16x64_bf16:
10951096
return selectSMFMACIntrin(I);
10961097
default:
10971098
return selectImpl(I, *CoverageInfo);
@@ -3490,6 +3491,9 @@ bool AMDGPUInstructionSelector::selectSMFMACIntrin(MachineInstr &MI) const {
34903491
case Intrinsic::amdgcn_smfmac_f32_32x32x32_f16:
34913492
Opc = AMDGPU::V_SMFMAC_F32_32X32X32_F16_e64;
34923493
break;
3494+
case Intrinsic::amdgcn_smfmac_f32_16x16x64_bf16:
3495+
Opc = AMDGPU::V_SMFMAC_F32_16X16X64_BF16_e64;
3496+
break;
34933497
default:
34943498
llvm_unreachable("unhandled smfmac intrinsic");
34953499
}

‎llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4806,7 +4806,8 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
48064806
case Intrinsic::amdgcn_smfmac_f32_32x32x32_fp8_bf8:
48074807
case Intrinsic::amdgcn_smfmac_f32_32x32x32_fp8_fp8:
48084808
case Intrinsic::amdgcn_smfmac_f32_16x16x64_f16:
4809-
case Intrinsic::amdgcn_smfmac_f32_32x32x32_f16: {
4809+
case Intrinsic::amdgcn_smfmac_f32_32x32x32_f16:
4810+
case Intrinsic::amdgcn_smfmac_f32_16x16x64_bf16: {
48104811
// vdst, srcA, srcB, srcC, idx
48114812
OpdsMapping[0] = getAGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
48124813
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
@@ -2870,6 +2870,7 @@ def VOP_V16F32_I64_I64_V16F32 : VOPProfile <[v16f32, i64, i64, v16f32]>;
28702870

28712871
def VOP_V4F32_V4F16_V8F16_I32 : VOPProfile <[v4f32, v4f16, v8f16, i32]>;
28722872
def VOP_V4F32_V8F16_V16F16_I32 : VOPProfile <[v4f32, v8f16, v16f16, i32]>;
2873+
def VOP_V4F32_V8BF16_V16BF16_I32 : VOPProfile <[v4f32, v8bf16, v16bf16, i32]>;
28732874
def VOP_V16F32_V4F16_V8F16_I32 : VOPProfile <[v16f32, v4f16, v8f16, i32]>;
28742875
def VOP_V16F32_V8F16_V16F16_I32 : VOPProfile <[v16f32, v8f16, v16f16, i32]>;
28752876
def VOP_V4F32_V4I16_V8I16_I32 : VOPProfile <[v4f32, v4i16, v8i16, i32]>;

‎llvm/lib/Target/AMDGPU/VOP3PInstructions.td

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -631,6 +631,7 @@ def VOPProfileMAI_F32_I64_X16_VCD : VOPProfileMAI<VOP_V16F32_I64_I64_V16F32,
631631
def VOPProfileSMFMAC_F32_16X16X32_F16 : VOPProfileSMFMAC<VOP_V4F32_V4F16_V8F16_I32, AVDst_128, AVSrc_64, AVSrc_128>;
632632
def VOPProfileSMFMAC_F32_16X16X64_F16 : VOPProfileSMFMAC<VOP_V4F32_V8F16_V16F16_I32, AVDst_128, AVSrc_128, AVSrc_256>;
633633
def VOPProfileSMFMAC_F32_32X32X32_F16 : VOPProfileSMFMAC<VOP_V16F32_V8F16_V16F16_I32, AVDst_512, AVSrc_128, AVSrc_256>;
634+
def VOPProfileSMFMAC_F32_16X16X64_BF16 : VOPProfileSMFMAC<VOP_V4F32_V8BF16_V16BF16_I32, AVDst_128, AVSrc_128, AVSrc_256>;
634635
def VOPProfileSMFMAC_F32_32X32X16_F16 : VOPProfileSMFMAC<VOP_V16F32_V4F16_V8F16_I32, AVDst_512, AVSrc_64, AVSrc_128>;
635636
def VOPProfileSMFMAC_F32_16X16X32_I16 : VOPProfileSMFMAC<VOP_V4F32_V4I16_V8I16_I32, AVDst_128, AVSrc_64, AVSrc_128>;
636637
def VOPProfileSMFMAC_F32_32X32X16_I16 : VOPProfileSMFMAC<VOP_V16F32_V4I16_V8I16_I32, AVDst_512, AVSrc_64, AVSrc_128>;
@@ -1047,6 +1048,7 @@ defm V_SMFMAC_F32_32X32X32_FP8_FP8 : SMFMACInst<"v_smfmac_f32_32x32x32_fp8_fp8",
10471048
let SubtargetPredicate = HasGFX950Insts in {
10481049
defm V_SMFMAC_F32_16X16X64_F16 : SMFMACInst<"v_smfmac_f32_16x16x64_f16", "F32_16X16X64_F16", int_amdgcn_smfmac_f32_16x16x64_f16>;
10491050
defm V_SMFMAC_F32_32X32X32_F16 : SMFMACInst<"v_smfmac_f32_32x32x32_f16", "F32_32X32X32_F16", int_amdgcn_smfmac_f32_32x32x32_f16>;
1051+
defm V_SMFMAC_F32_16X16X64_BF16 : SMFMACInst<"v_smfmac_f32_16x16x64_bf16", "F32_16X16X64_BF16", int_amdgcn_smfmac_f32_16x16x64_bf16>;
10501052
}
10511053

10521054
def MAIInstInfoTable : GenericTable {
@@ -2140,6 +2142,7 @@ defm V_SMFMAC_F32_32X32X32_FP8_FP8 : VOP3P_Real_SMFMAC <0x7f, "v_smfmac_f32_32x3
21402142

21412143
defm V_SMFMAC_F32_16X16X64_F16 : VOP3P_Real_SMFMAC <0x5a, "v_smfmac_f32_16x16x64f16">;
21422144
defm V_SMFMAC_F32_32X32X32_F16 : VOP3P_Real_SMFMAC <0x5b, "v_smfmac_f32_32x32x32f16">;
2145+
defm V_SMFMAC_F32_16X16X64_BF16 : VOP3P_Real_SMFMAC <0x39, "v_smfmac_f32_16x16x64bf16">;
21432146

21442147
defm V_PK_FMA_F32 : VOP3P_Real_vi <0x30>;
21452148
defm V_PK_MUL_F32 : VOP3P_Real_vi <0x31>;

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

Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -350,6 +350,16 @@ define amdgpu_kernel void @smfmac_f32_32x32x32_f16(<8 x half> %arg0, <16 x half>
350350
ret void
351351
}
352352

353+
declare <4 x float> @llvm.amdgcn.smmfmac.f32.16x16x64.bf16(<8 x bfloat>, <16 x bfloat>, <4 x float>, i32, i32 immarg, i32 immarg)
354+
355+
; CHECK: DIVERGENT: %result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x64.bf16(<8 x bfloat> %arg0, <16 x bfloat> %arg1, <4 x float> %arg2, i32 %arg3, i32 immarg 0, i32 immarg 0)
356+
define amdgpu_kernel void @smfmac_f32_16x16x64_bf16(<8 x bfloat> %arg0, <16 x bfloat> %arg1, <4 x float> %arg2, i32 %arg3, ptr addrspace(1) %out) {
357+
%result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x64.bf16(<8 x bfloat> %arg0, <16 x bfloat> %arg1, <4 x float> %arg2, i32 %arg3, i32 immarg 0, i32 immarg 0)
358+
store <4 x float> %result, ptr addrspace(1) %out
359+
ret void
360+
}
361+
362+
353363
declare i32 @llvm.amdgcn.ds.swizzle(i32, i32) #1
354364
declare i32 @llvm.amdgcn.permlane16.i32(i32, i32, i32, i32, i1, i1) #1
355365
declare i32 @llvm.amdgcn.permlanex16.i32(i32, i32, i32, i32, i1, i1) #1

0 commit comments

Comments
 (0)