AMDGPU: Define v_mfma_f32_32x32x16_bf16 for gfx950 (llvm#116679)

arsenm · pravinjagtap · commit 1e7b44d34ced · 2025-01-21T21:00:03.000+05:30
Unlike the existing gfx940 intrinsics using short/i16 in place of
bfloat, this uses the natural bfloat type.
diff --git a/clang/include/clang/Basic/BuiltinsAMDGPU.def b/clang/include/clang/Basic/BuiltinsAMDGPU.def
@@ -437,6 +437,8 @@ TARGET_BUILTIN(__builtin_amdgcn_cvt_sr_fp8_f32, "ifiiIi", "nc", "fp8-conversion-
 TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_f16, "V4fV8hV8hV4fIiIiIi", "nc", "gfx950-insts")
 TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_f16, "V16fV8hV8hV16fIiIiIi", "nc", "gfx950-insts")
 
+TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_bf16, "V16fV8yV8yV16fIiIiIi", "nc", "gfx950-insts")
+
 //===----------------------------------------------------------------------===//
 // GFX12+ only builtins.
 //===----------------------------------------------------------------------===//
diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl
@@ -24,6 +24,7 @@ typedef short  v8s   __attribute__((ext_vector_type(8)));
 typedef short  v16s  __attribute__((ext_vector_type(16)));
 typedef short  v32s  __attribute__((ext_vector_type(32)));
 typedef double v4d   __attribute__((ext_vector_type(4)));
+typedef __bf16 v8bf16   __attribute__((ext_vector_type(8)));
 
 
 #ifdef MFMA_GFX908_TESTS
@@ -424,5 +425,10 @@ v16f test_mfma_f32_32x32x16_f16(v8h a, v8h b, v16f c)
   return __builtin_amdgcn_mfma_f32_32x32x16_f16(a, b, c, 1, 2, 3);
 }
 
+// CHECK-GFX950-LABEL: @test_mfma_f32_32x32x16_bf16(
+// CHECK-GFX950: tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.bf16(<8 x bfloat> %a, <8 x bfloat> %b, <16 x float> %c, i32 1, i32 2, i32 3)
+v16f test_mfma_f32_32x32x16_bf16(v8bf16 a, v8bf16 b, v16f c) {
+  return __builtin_amdgcn_mfma_f32_32x32x16_bf16(a, b, c, 1, 2, 3);
+}
 
 #endif
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl
@@ -4,6 +4,7 @@
 typedef float float4 __attribute__((ext_vector_type(4)));
 typedef float float16 __attribute__((ext_vector_type(16)));
 typedef half half8 __attribute__((ext_vector_type(8)));
+typedef __bf16 bfloat8 __attribute__((ext_vector_type(8)));
 
 
 void test_mfma_f32_16x16x32_f16(__global float4* out, half8 a, half8 b, float4 c, int X) {
@@ -19,3 +20,9 @@ void test_mfma_f32_32x32x16_f16(__global float16* out, half8 a, half8 b, float16
   *out = __builtin_amdgcn_mfma_f32_32x32x16_f16(a, b, c, 0, X, 0); // expected-error{{argument to '__builtin_amdgcn_mfma_f32_32x32x16_f16' must be a constant integer}}
   *out = __builtin_amdgcn_mfma_f32_32x32x16_f16(a, b, c, 0, 0, X); // expected-error{{argument to '__builtin_amdgcn_mfma_f32_32x32x16_f16' must be a constant integer}}
 }
+
+void test_mfma_f32_32x32x16_bf16(__global float16* out, bfloat8 a, bfloat8 b, float16 c, int X) {
+  *out = __builtin_amdgcn_mfma_f32_32x32x16_bf16(a, b, c, X, 0, 0); // expected-error{{argument to '__builtin_amdgcn_mfma_f32_32x32x16_bf16' must be a constant integer}}
+  *out = __builtin_amdgcn_mfma_f32_32x32x16_bf16(a, b, c, 0, X, 0);  // expected-error{{argument to '__builtin_amdgcn_mfma_f32_32x32x16_bf16' must be a constant integer}}
+  *out = __builtin_amdgcn_mfma_f32_32x32x16_bf16(a, b, c, 0, 0, X);  // expected-error{{argument to '__builtin_amdgcn_mfma_f32_32x32x16_bf16' must be a constant integer}}
+}
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950.cl
@@ -4,9 +4,12 @@
 typedef float float4 __attribute__((ext_vector_type(4)));
 typedef float float16 __attribute__((ext_vector_type(16)));
 typedef half half8 __attribute__((ext_vector_type(8)));
+typedef __bf16 bfloat8 __attribute__((ext_vector_type(8)));
 
 void test(__global float4* out0, half8 a0, half8 b0, float4 c0,
-          __global float16* out1, half8 a1, half8 b1, float16 c1) {
+          __global float16* out1, half8 a1, half8 b1, float16 c1,
+          __global float16* out2, bfloat8 a2, bfloat8 b2, float16 c2) {
   *out0 = __builtin_amdgcn_mfma_f32_16x16x32_f16(a0, b0, c0, 0, 0, 0); // expected-error{{'__builtin_amdgcn_mfma_f32_16x16x32_f16' needs target feature gfx950-insts}}
   *out1 = __builtin_amdgcn_mfma_f32_32x32x16_f16(a1, b1, c1, 0, 0, 0); // expected-error{{'__builtin_amdgcn_mfma_f32_32x32x16_f16' needs target feature gfx950-insts}}
+  *out2 = __builtin_amdgcn_mfma_f32_32x32x16_bf16(a2, b2, c2, 0, 0, 0); // expected-error{{'__builtin_amdgcn_mfma_f32_32x32x16_bf16' needs target feature gfx950-insts}}
 }
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -3146,6 +3146,8 @@ def int_amdgcn_cvt_sr_fp8_f32 : ClangBuiltin<"__builtin_amdgcn_cvt_sr_fp8_f32">,
 defset list<Intrinsic> AMDGPUMFMAIntrinsics950 = {
 def int_amdgcn_mfma_f32_16x16x32_f16 : AMDGPUMfmaIntrinsic<llvm_v4f32_ty, llvm_v8f16_ty>;
 def int_amdgcn_mfma_f32_32x32x16_f16 : AMDGPUMfmaIntrinsic<llvm_v16f32_ty, llvm_v8f16_ty>;
+
+def int_amdgcn_mfma_f32_32x32x16_bf16 : AMDGPUMfmaIntrinsic<llvm_v16f32_ty, llvm_v8bf16_ty>;
 }
 
 //===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -2648,6 +2648,7 @@ def VOP_V16F32_V2I32_V4I32_I32    : VOPProfile <[v16f32, v2i32, v4i32, i32]>;
 
 def VOP_V4F32_V8F16_V8F16_V4F32   : VOPProfile <[v4f32,  v8f16, v8f16, v4f32]>;
 def VOP_V16F32_V8F16_V8F16_V16F32 : VOPProfile <[v16f32, v8f16, v8f16, v16f32]>;
+def VOP_V16F32_V8BF16_V8BF16_V16F32 : VOPProfile <[v16f32, v8bf16, v8bf16, v16f32]>;
 
 
 class Commutable_REV <string revOp, bit isOrig> {
diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -635,6 +635,9 @@ def VOPProfileMAI_F32_V8F16_X32_VCD : VOPProfileMAI<VOP_V4F32_V8F16_V8F16_V4F32,
 def VOPProfileMAI_F32_V8F16_X16     : VOPProfileMAI<VOP_V16F32_V8F16_V8F16_V16F32,     AISrc_512_f32,  ADst_512, AVSrc_128>;
 def VOPProfileMAI_F32_V8F16_X16_VCD : VOPProfileMAI<VOP_V16F32_V8F16_V8F16_V16F32,     VISrc_512_f32,  VDst_512, AVSrc_128>;
 
+def VOPProfileMAI_F32_V8BF16_X16     : VOPProfileMAI<VOP_V16F32_V8BF16_V8BF16_V16F32,     AISrc_512_f32,  ADst_512, AVSrc_128>;
+def VOPProfileMAI_F32_V8BF16_X16_VCD : VOPProfileMAI<VOP_V16F32_V8BF16_V8BF16_V16F32,     VISrc_512_f32,  VDst_512, AVSrc_128>;
+
 class MFMATable <bit is_mac, string Name> {
   bit IsMac = is_mac;
   string FMAOp = Name;
@@ -751,6 +754,7 @@ defm V_MFMA_F32_32X32X4BF16 : MAIInst<"v_mfma_f32_32x32x4bf16", "F32_V2I16_X16",
 let SubtargetPredicate = HasGFX950Insts, is_gfx940_xdl = 1 in {
 defm V_MFMA_F32_16X16X32_F16   : MAIInst<"v_mfma_f32_16x16x32f16",    "F32_V8F16_X32", int_amdgcn_mfma_f32_16x16x32_f16>;
 defm V_MFMA_F32_32X32X16_F16   : MAIInst<"v_mfma_f32_32x32x16f16",    "F32_V8F16_X16", int_amdgcn_mfma_f32_32x32x16_f16>;
+defm V_MFMA_F32_32X32X16_BF16  : MAIInst<"v_mfma_f32_32x32x16bf16",   "F32_V8BF16_X16", int_amdgcn_mfma_f32_32x32x16_bf16>;
 }
 
 let Predicates = [isGFX90APlus] in {
@@ -1790,6 +1794,8 @@ defm V_MFMA_F64_4X4X4F64        : VOP3P_Real_MFMA_gfx90a <0x6f>;
 
 defm V_MFMA_F32_16X16X32_F16     : VOP3P_Real_MFMA_gfx950 <0x54, "v_mfma_f32_16x16x32_f16">;
 defm V_MFMA_F32_32X32X16_F16     : VOP3P_Real_MFMA_gfx950 <0x55, "v_mfma_f32_32x32x16_f16">;
+defm V_MFMA_F32_32X32X16_BF16    : VOP3P_Real_MFMA_gfx950 <0x37, "v_mfma_f32_32x32x16_bf16">;
+
 defm V_MFMA_I32_32X32X16I8       : VOP3P_Real_MFMA_gfx940 <0x56, "v_mfma_i32_32x32x16_i8">;
 defm V_MFMA_I32_16X16X32I8       : VOP3P_Real_MFMA_gfx940 <0x57, "v_mfma_i32_16x16x32_i8">;
 let SubtargetPredicate = HasXF32Insts in {
diff --git a/llvm/test/Analysis/UniformityAnalysis/AMDGPU/intrinsics.ll b/llvm/test/Analysis/UniformityAnalysis/AMDGPU/intrinsics.ll
@@ -278,6 +278,14 @@ define amdgpu_kernel void @mfma_f32_32x32x16_f16(<8 x half> %arg0, <8 x half> %a
   ret void
 }
 
+; CHECK: DIVERGENT:  %result = call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.bf16(<8 x bfloat> %arg0, <8 x bfloat> %arg1, <16 x float> %arg2, i32 immarg 0, i32 immarg 0, i32 immarg 0)
+define amdgpu_kernel void @mfma_f32_32x32x16_bf16(<8 x bfloat> %arg0, <8 x bfloat> %arg1, <16 x float> %arg2, ptr addrspace(1) %out) {
+  %result = call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.bf16(<8 x bfloat> %arg0, <8 x bfloat> %arg1, <16 x float> %arg2, i32 immarg 0, i32 immarg 0, i32 immarg 0)
+  store <16 x float> %result, ptr addrspace(1) %out
+  ret void
+}
+
+
 declare i32 @llvm.amdgcn.ds.swizzle(i32, i32) #1
 declare i32 @llvm.amdgcn.permlane16.i32(i32, i32, i32, i32, i1, i1) #1
 declare i32 @llvm.amdgcn.permlanex16.i32(i32, i32, i32, i32, i1, i1) #1
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.gfx950.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.gfx950.bf16.ll
diff --git a/llvm/test/MC/AMDGPU/mai-gfx950.s b/llvm/test/MC/AMDGPU/mai-gfx950.s
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx950_mai.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx950_mai.txt
diff --git a/llvm/test/tools/llvm-mca/AMDGPU/gfx950.s b/llvm/test/tools/llvm-mca/AMDGPU/gfx950.s

Original file line number	Diff line number	Diff line change
`@@ -3146,6 +3146,8 @@ def int_amdgcn_cvt_sr_fp8_f32 : ClangBuiltin<"__builtin_amdgcn_cvt_sr_fp8_f32">,`
`3146`	`3146`	`defset list<Intrinsic> AMDGPUMFMAIntrinsics950 = {`
`3147`	`3147`	`def int_amdgcn_mfma_f32_16x16x32_f16 : AMDGPUMfmaIntrinsic<llvm_v4f32_ty, llvm_v8f16_ty>;`
`3148`	`3148`	`def int_amdgcn_mfma_f32_32x32x16_f16 : AMDGPUMfmaIntrinsic<llvm_v16f32_ty, llvm_v8f16_ty>;`
	`3149`	`+`
	`3150`	`+def int_amdgcn_mfma_f32_32x32x16_bf16 : AMDGPUMfmaIntrinsic<llvm_v16f32_ty, llvm_v8bf16_ty>;`
`3149`	`3151`	`}`
`3150`	`3152`
`3151`	`3153`	`//===----------------------------------------------------------------------===//`