AMDGPU: Add v_smfmac_f32_32x32x64_bf8_bf8 for gfx950 (#117256)

llvm · Nov 22, 2024 · 8a5c241 · 8a5c241
1 parent 6da8ff8
commit 8a5c241
Show file tree

Hide file tree

Showing 13 changed files with 507 additions and 1 deletion.
diff --git a/clang/include/clang/Basic/BuiltinsAMDGPU.def b/clang/include/clang/Basic/BuiltinsAMDGPU.def
@@ -454,6 +454,7 @@ TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_bf8_bf8, "V4fV4iV8iV4fiIiIi
 TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_bf8_fp8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
 TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_fp8_bf8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
 TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8, "V4fV4iV8iV4fiIiIi", "nc", "gfx950-insts")
+TARGET_BUILTIN(__builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8, "V16fV4iV8iV16fiIiIi", "nc", "gfx950-insts")
 
 //===----------------------------------------------------------------------===//
 // GFX12+ only builtins.

diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl
@@ -538,4 +538,11 @@ void test_smfmac_f32_16x16x128_fp8_fp8(global v4f* out, v4i a, v8i b, v4f c, int
   *out = __builtin_amdgcn_smfmac_f32_16x16x128_fp8_fp8(a, b, c, idx, 0, 0);
 }
 
+// CHECK-GFX950-LABEL: @test_smfmac_f32_32x32x64_bf8_bf8
+// 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)
+void test_smfmac_f32_32x32x64_bf8_bf8(global v16f* out, v4i a, v8i b, v16f c, int idx)
+{
+  *out = __builtin_amdgcn_smfmac_f32_32x32x64_bf8_bf8(a, b, c, idx, 0, 0);
+}
+
 #endif
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl
@@ -124,3 +124,9 @@ void test_smfmac_f32_16x16x128_fp8_fp8(global float4* out, int4 a, int8 b, float
   *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}}
   *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}}
 }
+
+void test_smfmac_f32_32x32x64_bf8_bf8(global float16* out, int4 a, int8 b, float16 c, int idx, int d)
+{
+  *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}}
+  *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}}
+}
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950.cl
@@ -44,6 +44,7 @@ void test(__global float4* out0, half8 a0, half8 b0, float4 c0,
   *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}}
   *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}}
   *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}}
+  *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}}
   *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}}
   *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}}
 }
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -3162,6 +3162,7 @@ def int_amdgcn_smfmac_f32_16x16x128_bf8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_
 def int_amdgcn_smfmac_f32_16x16x128_bf8_fp8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
 def int_amdgcn_smfmac_f32_16x16x128_fp8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
 def int_amdgcn_smfmac_f32_16x16x128_fp8_fp8 : AMDGPUMSmfmacIntrinsic<llvm_v4f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
+def int_amdgcn_smfmac_f32_32x32x64_bf8_bf8 : AMDGPUMSmfmacIntrinsic<llvm_v16f32_ty, llvm_v4i32_ty, llvm_v8i32_ty>;
 }
 
 //===----------------------------------------------------------------------===//

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -1100,6 +1100,7 @@ bool AMDGPUInstructionSelector::selectG_INTRINSIC(MachineInstr &I) const {
   case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
   case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8:
   case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8:
+  case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_bf8:
     return selectSMFMACIntrin(I);
   default:
     return selectImpl(I, *CoverageInfo);
@@ -3551,6 +3552,9 @@ bool AMDGPUInstructionSelector::selectSMFMACIntrin(MachineInstr &MI) const {
   case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8:
     Opc = AMDGPU::V_SMFMAC_F32_16X16X128_FP8_FP8_e64;
     break;
+  case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_bf8:
+    Opc = AMDGPU::V_SMFMAC_F32_32X32X64_BF8_BF8_e64;
+    break;
   default:
     llvm_unreachable("unhandled smfmac intrinsic");
   }

diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -4814,7 +4814,8 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
     case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_bf8:
     case Intrinsic::amdgcn_smfmac_f32_16x16x128_bf8_fp8:
     case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_bf8:
-    case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8: {
+    case Intrinsic::amdgcn_smfmac_f32_16x16x128_fp8_fp8:
+    case Intrinsic::amdgcn_smfmac_f32_32x32x64_bf8_bf8: {
       // vdst, srcA, srcB, srcC, idx
       OpdsMapping[0] = getAGPROpMapping(MI.getOperand(0).getReg(), MRI, *TRI);
       OpdsMapping[2] = getVGPROpMapping(MI.getOperand(2).getReg(), MRI, *TRI);

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -2883,6 +2883,7 @@ def VOP_V16F32_V2I32_V4I32_I32    : VOPProfile <[v16f32, v2i32, v4i32, i32]>;
 def VOP_V4I32_V4I32_V8I32_I32     : VOPProfile <[v4i32,  v4i32, v8i32, i32]>;
 def VOP_V16I32_V4I32_V8I32_I32    : VOPProfile <[v16i32, v4i32, v8i32, i32]>;
 def VOP_V4F32_V4I32_V8I32_I32     : VOPProfile <[v4f32, v4i32, v8i32, i32]>;
+def VOP_V16F32_V4I32_V8I32_I32    : VOPProfile <[v16f32, v4i32, v8i32, i32]>;
 
 def VOP_V4F32_V8F16_V8F16_V4F32   : VOPProfile <[v4f32,  v8f16, v8f16, v4f32]>;
 def VOP_V16F32_V8F16_V8F16_V16F32 : VOPProfile <[v16f32, v8f16, v8f16, v16f32]>;

diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -644,6 +644,7 @@ def VOPProfileSMFMAC_I32_16X16X128_I8 : VOPProfileSMFMAC<VOP_V4I32_V4I32_V8I32_I
 def VOPProfileSMFMAC_I32_32X32X64_I8 : VOPProfileSMFMAC<VOP_V16I32_V4I32_V8I32_I32, AVDst_512, AVSrc_128, AVSrc_256>;
 
 def VOPProfileSMFMAC_F32_16X16X128_F8  : VOPProfileSMFMAC<VOP_V4F32_V4I32_V8I32_I32,  AVDst_128, AVSrc_128, AVSrc_256>;
+def VOPProfileSMFMAC_F32_32X32X64_F8  : VOPProfileSMFMAC<VOP_V16F32_V4I32_V8I32_I32,  AVDst_512, AVSrc_128, AVSrc_256>;
 
 def VOPProfileMAI_F32_V8F16_X32     : VOPProfileMAI<VOP_V4F32_V8F16_V8F16_V4F32,       AISrc_128_f32,  ADst_128, AVSrc_128>;
 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
 defm V_SMFMAC_F32_16X16X128_BF8_FP8 : SMFMACInst<"v_smfmac_f32_16x16x128_bf8_fp8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_bf8_fp8>;
 defm V_SMFMAC_F32_16X16X128_FP8_BF8 : SMFMACInst<"v_smfmac_f32_16x16x128_fp8_bf8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_fp8_bf8>;
 defm V_SMFMAC_F32_16X16X128_FP8_FP8 : SMFMACInst<"v_smfmac_f32_16x16x128_fp8_fp8", "F32_16X16X128_F8", int_amdgcn_smfmac_f32_16x16x128_fp8_fp8>;
+defm V_SMFMAC_F32_32X32X64_BF8_BF8 : SMFMACInst<"v_smfmac_f32_32x32x64_bf8_bf8", "F32_32X32X64_F8", int_amdgcn_smfmac_f32_32x32x64_bf8_bf8>;
 }
 
 def MAIInstInfoTable : GenericTable {
@@ -2163,6 +2165,7 @@ defm V_SMFMAC_F32_16X16X128_BF8_BF8 : VOP3P_Real_SMFMAC <0x3b, "v_smfmac_f32_16x
 defm V_SMFMAC_F32_16X16X128_BF8_FP8 : VOP3P_Real_SMFMAC <0x3c, "v_smfmac_f32_16x16x128bf8fp8">;
 defm V_SMFMAC_F32_16X16X128_FP8_BF8 : VOP3P_Real_SMFMAC <0x3d, "v_smfmac_f32_16x16x128fp8bf8">;
 defm V_SMFMAC_F32_16X16X128_FP8_FP8 : VOP3P_Real_SMFMAC <0x43, "v_smfmac_f32_16x16x128fp8fp8">;
+defm V_SMFMAC_F32_32X32X64_BF8_BF8 : VOP3P_Real_SMFMAC <0x4b, "v_smfmac_f32_32x32x64bf8bf8">;
 
 defm V_PK_FMA_F32 : VOP3P_Real_vi <0x30>;
 defm V_PK_MUL_F32 : VOP3P_Real_vi <0x31>;

diff --git a/llvm/test/Analysis/UniformityAnalysis/AMDGPU/intrinsics.ll b/llvm/test/Analysis/UniformityAnalysis/AMDGPU/intrinsics.ll
@@ -404,6 +404,15 @@ define amdgpu_kernel void @smfmac_f32_16x16x128_fp8_fp8(<4 x i32> %arg0, <8 x i3
   ret void
 }
 
+declare <16 x float> @llvm.amdgcn.smfmac.f32.32x32x64.bf8.bf8(<4 x i32>, <8 x i32>, <16 x float>, i32, i32, i32)
+
+; 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)
+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) {
+  %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)
+  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