Skip to content

Commit 8211467

Browse files
arsenmpravinjagtap
arsenm
authored and
pravinjagtap
committed
AMDGPU: Add v_smfmac_f32_16x16x128_fp8_bf8 for gfx950 (llvm#117234)
1 parent e424fee commit 8211467

File tree

12 files changed

+304
-1
lines changed

12 files changed

+304
-1
lines changed

‎clang/include/clang/Basic/BuiltinsAMDGPU.def

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -452,6 +452,7 @@ TARGET_BUILTIN(__builtin_amdgcn_smfmac_i32_16x16x128_i8, "V4iV4iV8iV4iiIiIi", "n
452452
TARGET_BUILTIN(__builtin_amdgcn_smfmac_i32_32x32x64_i8, "V16iV4iV8iV16iiIiIi", "nc", "gfx950-insts")
453453
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_bf8_bf8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
454454
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
455+
TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
455456

456457
//===----------------------------------------------------------------------===//
457458
// GFX12+ only builtins.

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

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -524,4 +524,11 @@ void test_smfmac_f32_16x16x128_bf8_fp8(global v4f* out, v4i a, v8i b, v4f c, int
524524
*out = __builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8(a, b, c, idx, 0, 0);
525525
}
526526

527+
// CHECK-GFX950-LABEL: @test_smfmac_f32_16x16x128_fp8_bf8
528+
// CHECK-GFX950: call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %a, <8 x i32> %b, <4 x float> %c, i32 %idx, i32 0, i32 0)
529+
void test_smfmac_f32_16x16x128_fp8_bf8(global v4f* out, v4i a, v8i b, v4f c, int idx)
530+
{
531+
*out = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8(a, b, c, idx, 0, 0);
532+
}
533+
527534
#endif

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

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -112,3 +112,9 @@ void test_smfmac_f32_16x16x128_bf8_fp8(global float4* out, int4 a, int8 b, float
112112
*out = __builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8(a, b, c, idx, d, 0); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8' must be a constant integer}}
113113
*out = __builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8(a, b, c, idx, 0, d); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8' must be a constant integer}}
114114
}
115+
116+
void test_smfmac_f32_16x16x128_fp8_bf8(global float4* out, int4 a, int8 b, float4 c, int idx, int d)
117+
{
118+
*out = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8(a, b, c, idx, d, 0); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8' must be a constant integer}}
119+
*out = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8(a, b, c, idx, 0, d); // expected-error{{argument to '__builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8' must be a constant integer}}
120+
}

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

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -42,6 +42,7 @@ void test(__global float4* out0, half8 a0, half8 b0, float4 c0,
4242
*out11 = __builtin_amdgcn_smfmac_i32_32x32x64_i8(a11, b11, c11, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_i32_32x32x64_i8' needs target feature gfx950-insts}}
4343
*out12 = __builtin_amdgcn_smfmac_f32_16x16x128_bf8_bf8(a12, b12, c12, 0, 0, 0); // expected-error{{'__builtin_amdgcn_smfmac_f32_16x16x128_bf8_bf8' needs target feature gfx950-insts}}
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}}
45+
*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}}
4546
*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}}
4647
*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}}
4748
}

‎llvm/include/llvm/IR/IntrinsicsAMDGPU.td

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -3187,6 +3187,7 @@ def int_amdgcn_smfmac_i32_16x16x128_i8 : AMDGPUMSmfmacIntrinsic<llvm_v4i32_ty,
31873187
def int_amdgcn_smfmac_i32_32x32x64_i8 : AMDGPUMSmfmacIntrinsic<llvm_v16i32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
31883188
def int_amdgcn_smfmac_f32_16x16x128_bf8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
31893189
def int_amdgcn_smfmac_f32_16x16x128_bf8_fp8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
3190+
def int_amdgcn_smfmac_f32_16x16x128_fp8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
31903191
}
31913192

31923193
//===----------------------------------------------------------------------===//

‎llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1083,6 +1083,7 @@ bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I) const {
10831083
case Intrinsic::amdgcn_smfmac_i32_32x32x64_i8:
10841084
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_bf8:
10851085
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
1086+
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8:
10861087
return selectSMFMACIntrin(I);
10871088
default:
10881089
return selectImpl(I, *CoverageInfo);
@@ -3525,6 +3526,9 @@ bool AMDGPUInstructionSelector::selectSMFMACIntrin(MachineInstr &MI) const {
35253526
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
35263527
Opc = AMDGPU::V_SMFMAC_F32_16X16X128_BF8_FP8_e64;
35273528
break;
3529+
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8:
3530+
Opc = AMDGPU::V_SMFMAC_F32_16X16X128_FP8_BF8_e64;
3531+
break;
35283532
default:
35293533
llvm_unreachable("unhandled smfmac intrinsic");
35303534
}

‎llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4789,7 +4789,8 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
47894789
case Intrinsic::amdgcn_smfmac_i32_16x16x128_i8:
47904790
case Intrinsic::amdgcn_smfmac_i32_32x32x64_i8:
47914791
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_bf8:
4792-
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8: {
4792+
case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
4793+
case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8: {
47934794
// vdst, srcA, srcB, srcC, idx
47944795
OpdsMapping[0] = getAGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
47954796
OpdsMapping[2] = getVGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);

‎llvm/lib/Target/AMDGPU/VOP3PInstructions.td

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1067,6 +1067,7 @@ defm V_SMFMAC_I32_16X16X128_I8 : SMFMACInst<"v_smfmac_i32_16x16x128_i8",
10671067
defm V_SMFMAC_I32_32X32X64_I8 : SMFMACInst<"v_smfmac_i32_32x32x64_i8", "I32_32X32X64_I8", int_amdgcn_smfmac_i32_32x32x64_i8>;
10681068
defm V_SMFMAC_F32_16X16X128_BF8_BF8 : SMFMACInst<"v_smfmac_f32_16x16x128_bf8_bf8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_bf8_bf8>;
10691069
defm V_SMFMAC_F32_16X16X128_BF8_FP8 : SMFMACInst<"v_smfmac_f32_16x16x128_bf8_fp8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_bf8_fp8>;
1070+
defm V_SMFMAC_F32_16X16X128_FP8_BF8 : SMFMACInst<"v_smfmac_f32_16x16x128_fp8_bf8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_fp8_bf8>;
10701071
}
10711072

10721073
def MAIInstInfoTable : GenericTable {
@@ -2174,6 +2175,7 @@ defm V_SMFMAC_I32_32X32X64_I8 : VOP3P_Real_SMFMAC <0x47, "v_smfmac_i32_32x3
21742175

21752176
defm V_SMFMAC_F32_16X16X128_BF8_BF8 : VOP3P_Real_SMFMAC <0x3b, "v_smfmac_f32_16x16x128bf8bf8">;
21762177
defm V_SMFMAC_F32_16X16X128_BF8_FP8 : VOP3P_Real_SMFMAC <0x3c, "v_smfmac_f32_16x16x128bf8fp8">;
2178+
defm V_SMFMAC_F32_16X16X128_FP8_BF8 : VOP3P_Real_SMFMAC <0x3d, "v_smfmac_f32_16x16x128fp8bf8">;
21772179

21782180
defm V_PK_FMA_F32 : VOP3P_Real_vi <0x30>;
21792181
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
@@ -386,6 +386,15 @@ define amdgpu_kernel void @smfmac_f32_16x16x128_bf8_fp8(<4 x i32> %arg0, <8 x i3
386386
ret void
387387
}
388388

389+
declare <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32>, <8 x i32>, <4 x float>, i32, i32, i32)
390+
391+
; CHECK: DIVERGENT: %result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3, i32 1, i32 2)
392+
define amdgpu_kernel void @smfmac_f32_16x16x128_fp8_bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3, ptr addrspace(1) %out) {
393+
%result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3, i32 1, i32 2)
394+
store <4 x float> %result, ptr addrspace(1) %out
395+
ret void
396+
}
397+
389398
declare i32 @llvm.amdgcn.ds.swizzle(i32, i32) #1
390399
declare i32 @llvm.amdgcn.permlane16.i32(i32, i32, i32, i32, i1, i1) #1
391400
declare i32 @llvm.amdgcn.permlanex16.i32(i32, i32, i32, i32, i1, i1) #1

‎llvm/test/CodeGen/AMDGPU/llvm.amdgcn.smfmac.gfx950.ll

Lines changed: 213 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2450,4 +2450,217 @@ define <4 x float> @test_smfmac_f32_16x16x128_bf8_fp8__sgpr(<4 x i32> inreg %arg
24502450
ret <4 x float> %result
24512451
}
24522452

2453+
; --------------------------------------------------------------------
2454+
; llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8
2455+
; --------------------------------------------------------------------
2456+
2457+
declare <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32>, <8 x i32>, <4 x float>, i32, i32 immarg, i32 immarg)
2458+
2459+
define amdgpu_kernel void @test_smfmac_f32_16x16x128_fp8_bf8__vgpr(ptr addrspace(1) %arg, <4 x i32> %a, <8 x i32> %b, i32 %idx) #0 {
2460+
; SDAG-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__vgpr:
2461+
; SDAG: ; %bb.0: ; %bb
2462+
; SDAG-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
2463+
; SDAG-NEXT: v_lshlrev_b32_e32 v0, 4, v0
2464+
; SDAG-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x34
2465+
; SDAG-NEXT: v_mov_b32_e32 v16, 0
2466+
; SDAG-NEXT: s_waitcnt lgkmcnt(0)
2467+
; SDAG-NEXT: global_load_dwordx4 v[8:11], v0, s[2:3]
2468+
; SDAG-NEXT: s_load_dword s16, s[0:1], 0x64
2469+
; SDAG-NEXT: s_load_dwordx4 s[12:15], s[0:1], 0x54
2470+
; SDAG-NEXT: v_mov_b32_e32 v12, s4
2471+
; SDAG-NEXT: v_mov_b32_e32 v13, s5
2472+
; SDAG-NEXT: v_mov_b32_e32 v14, s6
2473+
; SDAG-NEXT: v_mov_b32_e32 v15, s7
2474+
; SDAG-NEXT: v_mov_b32_e32 v0, s8
2475+
; SDAG-NEXT: v_mov_b32_e32 v1, s9
2476+
; SDAG-NEXT: v_mov_b32_e32 v2, s10
2477+
; SDAG-NEXT: v_mov_b32_e32 v3, s11
2478+
; SDAG-NEXT: s_waitcnt lgkmcnt(0)
2479+
; SDAG-NEXT: v_mov_b32_e32 v4, s12
2480+
; SDAG-NEXT: v_mov_b32_e32 v5, s13
2481+
; SDAG-NEXT: v_mov_b32_e32 v6, s14
2482+
; SDAG-NEXT: v_mov_b32_e32 v7, s15
2483+
; SDAG-NEXT: v_mov_b32_e32 v17, s16
2484+
; SDAG-NEXT: s_waitcnt vmcnt(0)
2485+
; SDAG-NEXT: s_nop 0
2486+
; SDAG-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[8:11], v[12:15], v[0:7], v17 cbsz:1 abid:2
2487+
; SDAG-NEXT: s_nop 6
2488+
; SDAG-NEXT: global_store_dwordx4 v16, v[8:11], s[2:3]
2489+
; SDAG-NEXT: s_endpgm
2490+
;
2491+
; GISEL-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__vgpr:
2492+
; GISEL: ; %bb.0: ; %bb
2493+
; GISEL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x24
2494+
; GISEL-NEXT: v_lshlrev_b32_e32 v0, 4, v0
2495+
; GISEL-NEXT: s_waitcnt lgkmcnt(0)
2496+
; GISEL-NEXT: global_load_dwordx4 v[8:11], v0, s[2:3]
2497+
; GISEL-NEXT: s_load_dwordx8 s[4:11], s[0:1], 0x34
2498+
; GISEL-NEXT: s_load_dwordx4 s[12:15], s[0:1], 0x54
2499+
; GISEL-NEXT: s_load_dword s16, s[0:1], 0x64
2500+
; GISEL-NEXT: s_waitcnt lgkmcnt(0)
2501+
; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[6:7]
2502+
; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[4:5]
2503+
; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9]
2504+
; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11]
2505+
; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13]
2506+
; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15]
2507+
; GISEL-NEXT: v_mov_b32_e32 v16, s16
2508+
; GISEL-NEXT: s_waitcnt vmcnt(0)
2509+
; GISEL-NEXT: s_nop 0
2510+
; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[8:11], v[12:15], v[0:7], v16 cbsz:1 abid:2
2511+
; GISEL-NEXT: v_mov_b32_e32 v0, 0
2512+
; GISEL-NEXT: s_nop 5
2513+
; GISEL-NEXT: global_store_dwordx4 v0, v[8:11], s[2:3]
2514+
; GISEL-NEXT: s_endpgm
2515+
bb:
2516+
%id = call i32 @llvm.amdgcn.workitem.id.x()
2517+
%gep = getelementptr <4 x float>, ptr addrspace(1) %arg, i32 %id
2518+
%in.1 = load <4 x float>, ptr addrspace(1) %gep
2519+
%mai.1 = tail call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %a, <8 x i32> %b, <4 x float> %in.1, i32 %idx, i32 1, i32 2)
2520+
store <4 x float> %mai.1, ptr addrspace(1) %arg
2521+
ret void
2522+
}
2523+
2524+
define <4 x float> @test_smfmac_f32_16x16x128_fp8_bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3) {
2525+
; SDAG-LABEL: test_smfmac_f32_16x16x128_fp8_bf8:
2526+
; SDAG: ; %bb.0:
2527+
; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2528+
; SDAG-NEXT: v_accvgpr_write_b32 a0, v12
2529+
; SDAG-NEXT: v_accvgpr_write_b32 a1, v13
2530+
; SDAG-NEXT: v_accvgpr_write_b32 a2, v14
2531+
; SDAG-NEXT: v_accvgpr_write_b32 a3, v15
2532+
; SDAG-NEXT: s_nop 1
2533+
; SDAG-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 a[0:3], v[0:3], v[4:11], v16
2534+
; SDAG-NEXT: s_nop 6
2535+
; SDAG-NEXT: v_accvgpr_read_b32 v0, a0
2536+
; SDAG-NEXT: v_accvgpr_read_b32 v1, a1
2537+
; SDAG-NEXT: v_accvgpr_read_b32 v2, a2
2538+
; SDAG-NEXT: v_accvgpr_read_b32 v3, a3
2539+
; SDAG-NEXT: s_setpc_b64 s[30:31]
2540+
;
2541+
; GISEL-LABEL: test_smfmac_f32_16x16x128_fp8_bf8:
2542+
; GISEL: ; %bb.0:
2543+
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2544+
; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[12:15], v[0:3], v[4:11], v16
2545+
; GISEL-NEXT: s_nop 6
2546+
; GISEL-NEXT: v_mov_b32_e32 v0, v12
2547+
; GISEL-NEXT: v_mov_b32_e32 v1, v13
2548+
; GISEL-NEXT: v_mov_b32_e32 v2, v14
2549+
; GISEL-NEXT: v_mov_b32_e32 v3, v15
2550+
; GISEL-NEXT: s_setpc_b64 s[30:31]
2551+
%result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3, i32 immarg 0, i32 immarg 0)
2552+
ret <4 x float> %result
2553+
}
2554+
2555+
define <4 x float> @test_smfmac_f32_16x16x128_fp8_bf8__flags0(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3) {
2556+
; SDAG-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__flags0:
2557+
; SDAG: ; %bb.0:
2558+
; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2559+
; SDAG-NEXT: v_accvgpr_write_b32 a0, v12
2560+
; SDAG-NEXT: v_accvgpr_write_b32 a1, v13
2561+
; SDAG-NEXT: v_accvgpr_write_b32 a2, v14
2562+
; SDAG-NEXT: v_accvgpr_write_b32 a3, v15
2563+
; SDAG-NEXT: s_nop 1
2564+
; SDAG-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 a[0:3], v[0:3], v[4:11], v16 cbsz:1 abid:3
2565+
; SDAG-NEXT: s_nop 6
2566+
; SDAG-NEXT: v_accvgpr_read_b32 v0, a0
2567+
; SDAG-NEXT: v_accvgpr_read_b32 v1, a1
2568+
; SDAG-NEXT: v_accvgpr_read_b32 v2, a2
2569+
; SDAG-NEXT: v_accvgpr_read_b32 v3, a3
2570+
; SDAG-NEXT: s_setpc_b64 s[30:31]
2571+
;
2572+
; GISEL-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__flags0:
2573+
; GISEL: ; %bb.0:
2574+
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2575+
; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[12:15], v[0:3], v[4:11], v16 cbsz:1 abid:3
2576+
; GISEL-NEXT: s_nop 6
2577+
; GISEL-NEXT: v_mov_b32_e32 v0, v12
2578+
; GISEL-NEXT: v_mov_b32_e32 v1, v13
2579+
; GISEL-NEXT: v_mov_b32_e32 v2, v14
2580+
; GISEL-NEXT: v_mov_b32_e32 v3, v15
2581+
; GISEL-NEXT: s_setpc_b64 s[30:31]
2582+
%result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3, i32 immarg 1, i32 immarg 3)
2583+
ret <4 x float> %result
2584+
}
2585+
2586+
define <4 x float> @test_smfmac_f32_16x16x128_fp8_bf8__flags1(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3) {
2587+
; SDAG-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__flags1:
2588+
; SDAG: ; %bb.0:
2589+
; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2590+
; SDAG-NEXT: v_accvgpr_write_b32 a0, v12
2591+
; SDAG-NEXT: v_accvgpr_write_b32 a1, v13
2592+
; SDAG-NEXT: v_accvgpr_write_b32 a2, v14
2593+
; SDAG-NEXT: v_accvgpr_write_b32 a3, v15
2594+
; SDAG-NEXT: s_nop 1
2595+
; SDAG-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 a[0:3], v[0:3], v[4:11], v16 cbsz:3 abid:1
2596+
; SDAG-NEXT: s_nop 6
2597+
; SDAG-NEXT: v_accvgpr_read_b32 v0, a0
2598+
; SDAG-NEXT: v_accvgpr_read_b32 v1, a1
2599+
; SDAG-NEXT: v_accvgpr_read_b32 v2, a2
2600+
; SDAG-NEXT: v_accvgpr_read_b32 v3, a3
2601+
; SDAG-NEXT: s_setpc_b64 s[30:31]
2602+
;
2603+
; GISEL-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__flags1:
2604+
; GISEL: ; %bb.0:
2605+
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2606+
; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[12:15], v[0:3], v[4:11], v16 cbsz:3 abid:1
2607+
; GISEL-NEXT: s_nop 6
2608+
; GISEL-NEXT: v_mov_b32_e32 v0, v12
2609+
; GISEL-NEXT: v_mov_b32_e32 v1, v13
2610+
; GISEL-NEXT: v_mov_b32_e32 v2, v14
2611+
; GISEL-NEXT: v_mov_b32_e32 v3, v15
2612+
; GISEL-NEXT: s_setpc_b64 s[30:31]
2613+
%result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3, i32 immarg 3, i32 immarg 1)
2614+
ret <4 x float> %result
2615+
}
2616+
2617+
define <4 x float> @test_smfmac_f32_16x16x128_fp8_bf8__sgpr(<4 x i32> inreg %arg0, <8 x i32> inreg %arg1, <4 x float> inreg %arg2, i32 inreg %arg3) {
2618+
; SDAG-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__sgpr:
2619+
; SDAG: ; %bb.0:
2620+
; SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2621+
; SDAG-NEXT: v_mov_b32_e32 v8, s0
2622+
; SDAG-NEXT: v_mov_b32_e32 v9, s1
2623+
; SDAG-NEXT: v_mov_b32_e32 v10, s2
2624+
; SDAG-NEXT: v_mov_b32_e32 v11, s3
2625+
; SDAG-NEXT: v_mov_b32_e32 v0, s4
2626+
; SDAG-NEXT: v_mov_b32_e32 v1, s5
2627+
; SDAG-NEXT: v_mov_b32_e32 v2, s6
2628+
; SDAG-NEXT: v_mov_b32_e32 v3, s7
2629+
; SDAG-NEXT: v_mov_b32_e32 v4, s8
2630+
; SDAG-NEXT: v_mov_b32_e32 v5, s9
2631+
; SDAG-NEXT: v_mov_b32_e32 v6, s10
2632+
; SDAG-NEXT: v_mov_b32_e32 v7, s11
2633+
; SDAG-NEXT: v_accvgpr_write_b32 a0, s12
2634+
; SDAG-NEXT: v_accvgpr_write_b32 a1, s13
2635+
; SDAG-NEXT: v_accvgpr_write_b32 a2, s14
2636+
; SDAG-NEXT: v_accvgpr_write_b32 a3, s15
2637+
; SDAG-NEXT: v_mov_b32_e32 v12, s16
2638+
; SDAG-NEXT: s_nop 1
2639+
; SDAG-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 a[0:3], v[8:11], v[0:7], v12
2640+
; SDAG-NEXT: s_nop 6
2641+
; SDAG-NEXT: v_accvgpr_read_b32 v0, a0
2642+
; SDAG-NEXT: v_accvgpr_read_b32 v1, a1
2643+
; SDAG-NEXT: v_accvgpr_read_b32 v2, a2
2644+
; SDAG-NEXT: v_accvgpr_read_b32 v3, a3
2645+
; SDAG-NEXT: s_setpc_b64 s[30:31]
2646+
;
2647+
; GISEL-LABEL: test_smfmac_f32_16x16x128_fp8_bf8__sgpr:
2648+
; GISEL: ; %bb.0:
2649+
; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2650+
; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[2:3]
2651+
; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[0:1]
2652+
; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[4:5]
2653+
; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[12:13]
2654+
; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[6:7]
2655+
; GISEL-NEXT: v_mov_b64_e32 v[8:9], s[8:9]
2656+
; GISEL-NEXT: v_mov_b64_e32 v[10:11], s[10:11]
2657+
; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[14:15]
2658+
; GISEL-NEXT: v_mov_b32_e32 v16, s16
2659+
; GISEL-NEXT: s_nop 1
2660+
; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[0:3], v[12:15], v[4:11], v16
2661+
; GISEL-NEXT: s_setpc_b64 s[30:31]
2662+
%result = call <4 x float> @llvm.amdgcn.smfmac.f32.16x16x128.fp8.bf8(<4 x i32> %arg0, <8 x i32> %arg1, <4 x float> %arg2, i32 %arg3, i32 immarg 0, i32 immarg 0)
2663+
ret <4 x float> %result
2664+
}
2665+
24532666
attributes #0 = { "amdgpu-flat-work-group-size"="1,256" }

0 commit comments

Comments
 (0)