Skip to content

Commit 684d08f

Browse files
mbrkusaninmbrkusanin
committed
Use bfloat
1 parent 5c48c1f commit 684d08f

29 files changed

+1100
-369
lines changed

‎clang/include/clang/Basic/BuiltinsAMDGPU.def

Lines changed: 14 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -293,20 +293,20 @@ TARGET_BUILTIN(__builtin_amdgcn_s_wait_event_export_ready, "v", "n", "gfx11-inst
293293
// Postfix w64 indicates the builtin requires wavefront size of 64.
294294
//===----------------------------------------------------------------------===//
295295
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w32, "V8fV16hV16hV8f", "nc", "gfx11-insts")
296-
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32, "V8fV16sV16sV8f", "nc", "gfx11-insts")
296+
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32, "V8fV16yV16yV8f", "nc", "gfx11-insts")
297297
TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w32, "V16hV16hV16hV16hIb", "nc", "gfx11-insts")
298-
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32, "V16sV16sV16sV16sIb", "nc", "gfx11-insts")
298+
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32, "V16yV16yV16yV16yIb", "nc", "gfx11-insts")
299299
TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_tied_w32, "V16hV16hV16hV16hIb", "nc", "gfx11-insts")
300-
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w32, "V16sV16sV16sV16sIb", "nc", "gfx11-insts")
300+
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w32, "V16yV16yV16yV16yIb", "nc", "gfx11-insts")
301301
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32, "V8iIbV4iIbV4iV8iIb", "nc", "gfx11-insts")
302302
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32, "V8iIbV2iIbV2iV8iIb", "nc", "gfx11-insts")
303303

304304
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w64, "V4fV16hV16hV4f", "nc", "gfx11-insts")
305-
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64, "V4fV16sV16sV4f", "nc", "gfx11-insts")
305+
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64, "V4fV16yV16yV4f", "nc", "gfx11-insts")
306306
TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w64, "V8hV16hV16hV8hIb", "nc", "gfx11-insts")
307-
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64, "V8sV16sV16sV8sIb", "nc", "gfx11-insts")
307+
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64, "V8yV16yV16yV8yIb", "nc", "gfx11-insts")
308308
TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_tied_w64, "V8hV16hV16hV8hIb", "nc", "gfx11-insts")
309-
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w64, "V8sV16sV16sV8sIb", "nc", "gfx11-insts")
309+
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w64, "V8yV16yV16yV8yIb", "nc", "gfx11-insts")
310310
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64, "V4iIbV4iIbV4iV4iIb", "nc", "gfx11-insts")
311311
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64, "V4iIbV2iIbV2iV4iIb", "nc", "gfx11-insts")
312312

@@ -447,9 +447,9 @@ TARGET_BUILTIN(__builtin_amdgcn_global_load_tr_v4f16, "V4hV4h*1", "nc", "gfx12-i
447447
// builtins.
448448
//===----------------------------------------------------------------------===//
449449
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12, "V8fV8hV8hV8f", "nc", "gfx12-insts,wavefrontsize32")
450-
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12, "V8fV8sV8sV8f", "nc", "gfx12-insts,wavefrontsize32")
450+
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12, "V8fV8yV8yV8f", "nc", "gfx12-insts,wavefrontsize32")
451451
TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w32_gfx12, "V8hV8hV8hV8h", "nc", "gfx12-insts,wavefrontsize32")
452-
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32_gfx12, "V8sV8sV8sV8s", "nc", "gfx12-insts,wavefrontsize32")
452+
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32_gfx12, "V8yV8yV8yV8y", "nc", "gfx12-insts,wavefrontsize32")
453453
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12, "V8iIbV2iIbV2iV8iIb", "nc", "gfx12-insts,wavefrontsize32")
454454
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w32_gfx12, "V8iIbiIbiV8iIb", "nc", "gfx12-insts,wavefrontsize32")
455455
// These are gfx12-only, but for consistency with the other WMMA variants we're
@@ -461,9 +461,9 @@ TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf8_bf8_w32_gfx12, "V8fV2iV2iV
461461
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x32_iu4_w32_gfx12, "V8iIbV2iIbV2iV8iIb", "nc", "gfx12-insts,wavefrontsize32")
462462

463463
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_f16_w64_gfx12, "V4fV4hV4hV4f", "nc", "gfx12-insts,wavefrontsize64")
464-
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64_gfx12, "V4fV4sV4sV4f", "nc", "gfx12-insts,wavefrontsize64")
464+
TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf16_w64_gfx12, "V4fV4yV4yV4f", "nc", "gfx12-insts,wavefrontsize64")
465465
TARGET_BUILTIN(__builtin_amdgcn_wmma_f16_16x16x16_f16_w64_gfx12, "V4hV4hV4hV4h", "nc", "gfx12-insts,wavefrontsize64")
466-
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64_gfx12, "V4sV4sV4sV4s", "nc", "gfx12-insts,wavefrontsize64")
466+
TARGET_BUILTIN(__builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64_gfx12, "V4yV4yV4yV4y", "nc", "gfx12-insts,wavefrontsize64")
467467
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu8_w64_gfx12, "V4iIbiIbiV4iIb", "nc", "gfx12-insts,wavefrontsize64")
468468
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x16_iu4_w64_gfx12, "V4iIbiIbiV4iIb", "nc", "gfx12-insts,wavefrontsize64")
469469
// These are gfx12-only, but for consistency with the other WMMA variants we're
@@ -475,9 +475,9 @@ TARGET_BUILTIN(__builtin_amdgcn_wmma_f32_16x16x16_bf8_bf8_w64_gfx12, "V4fiiV4f",
475475
TARGET_BUILTIN(__builtin_amdgcn_wmma_i32_16x16x32_iu4_w64_gfx12, "V4iIbiIbiV4iIb", "nc", "gfx12-insts,wavefrontsize64")
476476

477477
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_f16_w32, "V8fV8hV16hV8fs", "nc", "gfx12-insts,wavefrontsize32")
478-
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf16_w32, "V8fV8sV16sV8fs", "nc", "gfx12-insts,wavefrontsize32")
478+
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf16_w32, "V8fV8yV16yV8fs", "nc", "gfx12-insts,wavefrontsize32")
479479
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f16_16x16x32_f16_w32, "V8hV8hV16hV8hs", "nc", "gfx12-insts,wavefrontsize32")
480-
TARGET_BUILTIN(__builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w32, "V8sV8sV16sV8ss", "nc", "gfx12-insts,wavefrontsize32")
480+
TARGET_BUILTIN(__builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w32, "V8yV8yV16yV8ys", "nc", "gfx12-insts,wavefrontsize32")
481481
TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu8_w32, "V8iIbV2iIbV4iV8isIb", "nc", "gfx12-insts,wavefrontsize32")
482482
TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu4_w32, "V8iIbiIbV2iV8isIb", "nc", "gfx12-insts,wavefrontsize32")
483483
TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x64_iu4_w32, "V8iIbV2iIbV4iV8isIb", "nc", "gfx12-insts,wavefrontsize32")
@@ -487,9 +487,9 @@ TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w32, "V8fV2iV4iV8fs"
487487
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w32, "V8fV2iV4iV8fs", "nc", "gfx12-insts,wavefrontsize32")
488488

489489
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_f16_w64, "V4fV4hV8hV4fs", "nc", "gfx12-insts,wavefrontsize64")
490-
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf16_w64, "V4fV4sV8sV4fs", "nc", "gfx12-insts,wavefrontsize64")
490+
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f32_16x16x32_bf16_w64, "V4fV4yV8yV4fs", "nc", "gfx12-insts,wavefrontsize64")
491491
TARGET_BUILTIN(__builtin_amdgcn_swmmac_f16_16x16x32_f16_w64, "V4hV4hV8hV4hs", "nc", "gfx12-insts,wavefrontsize64")
492-
TARGET_BUILTIN(__builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w64, "V4sV4sV8sV4ss", "nc", "gfx12-insts,wavefrontsize64")
492+
TARGET_BUILTIN(__builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w64, "V4yV4yV8yV4ys", "nc", "gfx12-insts,wavefrontsize64")
493493
TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu8_w64, "V4iIbiIbV2iV4isIb", "nc", "gfx12-insts,wavefrontsize64")
494494
TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x32_iu4_w64, "V4iIbiIbiV4isIb", "nc", "gfx12-insts,wavefrontsize64")
495495
TARGET_BUILTIN(__builtin_amdgcn_swmmac_i32_16x16x64_iu4_w64, "V4iIbiIbV2iV4isIb", "nc", "gfx12-insts,wavefrontsize64")

‎clang/test/CodeGenOpenCL/builtins-amdgcn-gfx12-wmma-w32.cl

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -5,7 +5,7 @@
55
typedef int v2i __attribute__((ext_vector_type(2)));
66
typedef float v8f __attribute__((ext_vector_type(8)));
77
typedef half v8h __attribute__((ext_vector_type(8)));
8-
typedef short v8s __attribute__((ext_vector_type(8)));
8+
typedef __bf16 v8bf __attribute__((ext_vector_type(8)));
99
typedef int v8i __attribute__((ext_vector_type(8)));
1010

1111
// Wave32
@@ -31,11 +31,11 @@ void test_amdgcn_wmma_f32_16x16x16_f16_w32(global v8f* out, v8h a, v8h b, v8f c)
3131

3232
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_bf16_w32(
3333
// CHECK-GFX1200-NEXT: entry:
34-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v8i16.v8f32(<8 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <8 x float> [[C:%.*]])
34+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v8bf16.v8f32(<8 x bfloat> [[A:%.*]], <8 x bfloat> [[B:%.*]], <8 x float> [[C:%.*]])
3535
// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
3636
// CHECK-GFX1200-NEXT: ret void
3737
//
38-
void test_amdgcn_wmma_f32_16x16x16_bf16_w32(global v8f* out, v8s a, v8s b, v8f c)
38+
void test_amdgcn_wmma_f32_16x16x16_bf16_w32(global v8f* out, v8bf a, v8bf b, v8f c)
3939
{
4040
*out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a, b, c);
4141
}
@@ -61,11 +61,11 @@ void test_amdgcn_wmma_f16_16x16x16_f16_w32(global v8h* out, v8h a, v8h b, v8h c)
6161

6262
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_bf16_16x16x16_bf16_w32(
6363
// CHECK-GFX1200-NEXT: entry:
64-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v8i16.v8i16(<8 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <8 x i16> [[C:%.*]], i1 false)
65-
// CHECK-GFX1200-NEXT: store <8 x i16> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
64+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x bfloat> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v8bf16.v8bf16(<8 x bfloat> [[A:%.*]], <8 x bfloat> [[B:%.*]], <8 x bfloat> [[C:%.*]], i1 false)
65+
// CHECK-GFX1200-NEXT: store <8 x bfloat> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
6666
// CHECK-GFX1200-NEXT: ret void
6767
//
68-
void test_amdgcn_wmma_bf16_16x16x16_bf16_w32(global v8s* out, v8s a, v8s b, v8s c)
68+
void test_amdgcn_wmma_bf16_16x16x16_bf16_w32(global v8bf* out, v8bf a, v8bf b, v8bf c)
6969
{
7070
*out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32_gfx12(a, b, c);
7171
}

‎clang/test/CodeGenOpenCL/builtins-amdgcn-gfx12-wmma-w64.cl

Lines changed: 6 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -4,7 +4,7 @@
44

55
typedef float v4f __attribute__((ext_vector_type(4)));
66
typedef half v4h __attribute__((ext_vector_type(4)));
7-
typedef short v4s __attribute__((ext_vector_type(4)));
7+
typedef __bf16 v4bf __attribute__((ext_vector_type(4)));
88
typedef int v4i __attribute__((ext_vector_type(4)));
99

1010
// Wave64
@@ -30,11 +30,11 @@ void test_amdgcn_wmma_f32_16x16x16_f16_w64(global v4f* out, v4h a, v4h b, v4f c)
3030

3131
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_bf16_w64(
3232
// CHECK-GFX1200-NEXT: entry:
33-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v4i16.v4f32(<4 x i16> [[A:%.*]], <4 x i16> [[B:%.*]], <4 x float> [[C:%.*]])
33+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v4bf16.v4f32(<4 x bfloat> [[A:%.*]], <4 x bfloat> [[B:%.*]], <4 x float> [[C:%.*]])
3434
// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
3535
// CHECK-GFX1200-NEXT: ret void
3636
//
37-
void test_amdgcn_wmma_f32_16x16x16_bf16_w64(global v4f* out, v4s a, v4s b, v4f c)
37+
void test_amdgcn_wmma_f32_16x16x16_bf16_w64(global v4f* out, v4bf a, v4bf b, v4f c)
3838
{
3939
*out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w64_gfx12(a, b, c);
4040
}
@@ -60,11 +60,11 @@ void test_amdgcn_wmma_f16_16x16x16_f16_w64(global v4h* out, v4h a, v4h b, v4h c)
6060

6161
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_bf16_16x16x16_bf16_w64(
6262
// CHECK-GFX1200-NEXT: entry:
63-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v4i16.v4i16(<4 x i16> [[A:%.*]], <4 x i16> [[B:%.*]], <4 x i16> [[C:%.*]], i1 false)
64-
// CHECK-GFX1200-NEXT: store <4 x i16> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 8, !tbaa [[TBAA4]]
63+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x bfloat> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v4bf16.v4bf16(<4 x bfloat> [[A:%.*]], <4 x bfloat> [[B:%.*]], <4 x bfloat> [[C:%.*]], i1 false)
64+
// CHECK-GFX1200-NEXT: store <4 x bfloat> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 8, !tbaa [[TBAA4]]
6565
// CHECK-GFX1200-NEXT: ret void
6666
//
67-
void test_amdgcn_wmma_bf16_16x16x16_bf16_w64(global v4s* out, v4s a, v4s b, v4s c)
67+
void test_amdgcn_wmma_bf16_16x16x16_bf16_w64(global v4bf* out, v4bf a, v4bf b, v4bf c)
6868
{
6969
*out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64_gfx12(a, b, c);
7070
}

‎clang/test/CodeGenOpenCL/builtins-amdgcn-swmmac-w32.cl

Lines changed: 7 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -6,10 +6,10 @@ typedef int v2i __attribute__((ext_vector_type(2)));
66
typedef int v4i __attribute__((ext_vector_type(4)));
77
typedef float v8f __attribute__((ext_vector_type(8)));
88
typedef half v8h __attribute__((ext_vector_type(8)));
9-
typedef short v8s __attribute__((ext_vector_type(8)));
9+
typedef __bf16 v8bf __attribute__((ext_vector_type(8)));
1010
typedef int v8i __attribute__((ext_vector_type(8)));
1111
typedef half v16h __attribute__((ext_vector_type(16)));
12-
typedef short v16s __attribute__((ext_vector_type(16)));
12+
typedef __bf16 v16bf __attribute__((ext_vector_type(16)));
1313

1414
// Wave32
1515

@@ -26,11 +26,11 @@ void test_amdgcn_swmmac_f32_16x16x32_f16_w32(global v8f* out, v8h a, v16h b, v8f
2626

2727
// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_bf16_w32(
2828
// CHECK-GFX1200-NEXT: entry:
29-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf16.v8i16.v16i16.v8f32.i16(<8 x i16> [[A:%.*]], <16 x i16> [[B:%.*]], <8 x float> [[C:%.*]], i16 [[INDEX:%.*]])
29+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf16.v8bf16.v16bf16.v8f32.i16(<8 x bfloat> [[A:%.*]], <16 x bfloat> [[B:%.*]], <8 x float> [[C:%.*]], i16 [[INDEX:%.*]])
3030
// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
3131
// CHECK-GFX1200-NEXT: ret void
3232
//
33-
void test_amdgcn_swmmac_f32_16x16x32_bf16_w32(global v8f* out, v8s a, v16s b, v8f c, short index)
33+
void test_amdgcn_swmmac_f32_16x16x32_bf16_w32(global v8f* out, v8bf a, v16bf b, v8f c, short index)
3434
{
3535
*out = __builtin_amdgcn_swmmac_f32_16x16x32_bf16_w32(a, b, c, index);
3636
}
@@ -48,11 +48,11 @@ void test_amdgcn_swmmac_f16_16x16x32_f16_w32(global v8h* out, v8h a, v16h b, v8h
4848

4949
// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_bf16_16x16x32_bf16_w32(
5050
// CHECK-GFX1200-NEXT: entry:
51-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x i16> @llvm.amdgcn.swmmac.bf16.16x16x32.bf16.v8i16.v16i16.v8i16.i16(<8 x i16> [[A:%.*]], <16 x i16> [[B:%.*]], <8 x i16> [[C:%.*]], i16 [[INDEX:%.*]])
52-
// CHECK-GFX1200-NEXT: store <8 x i16> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
51+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x bfloat> @llvm.amdgcn.swmmac.bf16.16x16x32.bf16.v8bf16.v16bf16.v8bf16.i16(<8 x bfloat> [[A:%.*]], <16 x bfloat> [[B:%.*]], <8 x bfloat> [[C:%.*]], i16 [[INDEX:%.*]])
52+
// CHECK-GFX1200-NEXT: store <8 x bfloat> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
5353
// CHECK-GFX1200-NEXT: ret void
5454
//
55-
void test_amdgcn_swmmac_bf16_16x16x32_bf16_w32(global v8s* out, v8s a, v16s b, v8s c, short index)
55+
void test_amdgcn_swmmac_bf16_16x16x32_bf16_w32(global v8bf* out, v8bf a, v16bf b, v8bf c, short index)
5656
{
5757
*out = __builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w32(a, b, c, index);
5858
}

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

Lines changed: 7 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -6,9 +6,9 @@ typedef int v2i __attribute__((ext_vector_type(2)));
66
typedef int v4i __attribute__((ext_vector_type(4)));
77
typedef float v4f __attribute__((ext_vector_type(4)));
88
typedef half v4h __attribute__((ext_vector_type(4)));
9-
typedef short v4s __attribute__((ext_vector_type(4)));
9+
typedef __bf16 v4bf __attribute__((ext_vector_type(4)));
1010
typedef half v8h __attribute__((ext_vector_type(8)));
11-
typedef short v8s __attribute__((ext_vector_type(8)));
11+
typedef __bf16 v8bf __attribute__((ext_vector_type(8)));
1212

1313
// Wave64
1414

@@ -25,11 +25,11 @@ void test_amdgcn_swmmac_f32_16x16x32_f16_w64(global v4f* out, v4h a, v8h b, v4f
2525

2626
// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_bf16_w64(
2727
// CHECK-GFX1200-NEXT: entry:
28-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf16.v4i16.v8i16.v4f32.i16(<4 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]])
28+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf16.v4bf16.v8bf16.v4f32.i16(<4 x bfloat> [[A:%.*]], <8 x bfloat> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]])
2929
// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
3030
// CHECK-GFX1200-NEXT: ret void
3131
//
32-
void test_amdgcn_swmmac_f32_16x16x32_bf16_w64(global v4f* out, v4s a, v8s b, v4f c, short index)
32+
void test_amdgcn_swmmac_f32_16x16x32_bf16_w64(global v4f* out, v4bf a, v8bf b, v4f c, short index)
3333
{
3434
*out = __builtin_amdgcn_swmmac_f32_16x16x32_bf16_w64(a, b, c, index);
3535
}
@@ -47,11 +47,11 @@ void test_amdgcn_swmmac_f16_16x16x32_f16_w64(global v4h* out, v4h a, v8h b, v4h
4747

4848
// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_bf16_16x16x32_bf16_w64(
4949
// CHECK-GFX1200-NEXT: entry:
50-
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x i16> @llvm.amdgcn.swmmac.bf16.16x16x32.bf16.v4i16.v8i16.v4i16.i16(<4 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <4 x i16> [[C:%.*]], i16 [[INDEX:%.*]])
51-
// CHECK-GFX1200-NEXT: store <4 x i16> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 8, !tbaa [[TBAA4]]
50+
// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x bfloat> @llvm.amdgcn.swmmac.bf16.16x16x32.bf16.v4bf16.v8bf16.v4bf16.i16(<4 x bfloat> [[A:%.*]], <8 x bfloat> [[B:%.*]], <4 x bfloat> [[C:%.*]], i16 [[INDEX:%.*]])
51+
// CHECK-GFX1200-NEXT: store <4 x bfloat> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 8, !tbaa [[TBAA4]]
5252
// CHECK-GFX1200-NEXT: ret void
5353
//
54-
void test_amdgcn_swmmac_bf16_16x16x32_bf16_w64(global v4s* out, v4s a, v8s b, v4s c, short index)
54+
void test_amdgcn_swmmac_bf16_16x16x32_bf16_w64(global v4bf* out, v4bf a, v8bf b, v4bf c, short index)
5555
{
5656
*out = __builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w64(a, b, c, index);
5757
}

0 commit comments

Comments
 (0)