AMDGPU: Add first gfx950 mfma instructions (llvm#116312)

arsenm · Pravin Jagtap · commit 508ba412c055 · 2025-01-21T21:00:03.000+05:30
Scheduling info and hazards are wrong and TBD.
diff --git a/clang/include/clang/Basic/BuiltinsAMDGPU.def b/clang/include/clang/Basic/BuiltinsAMDGPU.def
@@ -431,6 +431,12 @@ TARGET_BUILTIN(__builtin_amdgcn_cvt_pk_fp8_f32, "iffiIb", "nc", "fp8-conversion-
 TARGET_BUILTIN(__builtin_amdgcn_cvt_sr_bf8_f32, "ifiiIi", "nc", "fp8-conversion-insts")
 TARGET_BUILTIN(__builtin_amdgcn_cvt_sr_fp8_f32, "ifiiIi", "nc", "fp8-conversion-insts")
 
+//===----------------------------------------------------------------------===//
+// GFX950 only builtins.
+//===----------------------------------------------------------------------===//
+TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_16x16x32_f16, "V4fV8hV8hV4fIiIiIi", "nc", "gfx950-insts")
+TARGET_BUILTIN(__builtin_amdgcn_mfma_f32_32x32x16_f16, "V16fV8hV8hV16fIiIiIi", "nc", "gfx950-insts")
+
 //===----------------------------------------------------------------------===//
 // GFX12+ only builtins.
 //===----------------------------------------------------------------------===//
diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn-mfma.cl
@@ -2,6 +2,7 @@
 // RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx908 -DMFMA_GFX908_TESTS -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX908
 // RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx90a -DMFMA_GFX90A_TESTS -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX90A
 // RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx940 -DMFMA_GFX940_TESTS -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX940
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx950 -DMFMA_GFX950_TESTS -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX950
 
 #pragma OPENCL EXTENSION cl_khr_fp64:enable
 
@@ -222,7 +223,7 @@ void test_mfma_f64_4x4x4f64(global double* out, double a, double b, double c)
 
 #endif // MFMA_GFX90A_TESTS
 
-#ifdef MFMA_GFX940_TESTS
+#if defined(MFMA_GFX940_TESTS) || defined(MFMA_GFX950_TESTS)
 // CHECK-GFX940-LABEL: @test_mfma_i32_16x16x32_i8
 // CHECK-GFX940: call <4 x i32> @llvm.amdgcn.mfma.i32.16x16x32.i8(i64 %a, i64 %b, <4 x i32> %c, i32 0, i32 0, i32 0)
 void test_mfma_i32_16x16x32_i8(global v4i* out, long a, long b, v4i c)
@@ -404,4 +405,24 @@ void test_smfmac_f32_32x32x32_fp8_fp8(global v16f* out, v2i a, v4i b, v16f c, in
 {
   *out = __builtin_amdgcn_smfmac_f32_32x32x32_fp8_fp8(a, b, c, idx, 0, 0);
 }
-#endif // MFMA_GFX940_TESTS
+#endif // defined(MFMA_GFX940_TESTS) || defined(MFMA_GFX950_TESTS)
+
+#ifdef MFMA_GFX950_TESTS
+
+// CHECK-GFX950-LABEL: @test_mfma_f32_16x16x32_f16(
+// CHECK-GFX950: tail call <4 x float> @llvm.amdgcn.mfma.f32.16x16x32.f16(<8 x half> %a, <8 x half> %b, <4 x float> %c, i32 1, i32 2, i32 3)
+
+v4f test_mfma_f32_16x16x32_f16(v8h a, v8h b, v4f c)
+{
+  return __builtin_amdgcn_mfma_f32_16x16x32_f16(a, b, c, 1, 2, 3);
+}
+
+// CHECK-GFX950-LABEL: @test_mfma_f32_32x32x16_f16
+// CHECK-GFX950: tail call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> %a, <8 x half> %b, <16 x float> %c, i32 1, i32 2, i32 3)
+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);
+}
+
+
+#endif
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950-param.cl
@@ -0,0 +1,21 @@
+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -triple amdgcn-- -target-cpu gfx950 -verify -S -o - %s
+
+typedef float float4 __attribute__((ext_vector_type(4)));
+typedef float float16 __attribute__((ext_vector_type(16)));
+typedef half half8 __attribute__((ext_vector_type(8)));
+
+
+void test_mfma_f32_16x16x32_f16(__global float4* out, half8 a, half8 b, float4 c, int X) {
+
+  *out = __builtin_amdgcn_mfma_f32_16x16x32_f16(a, b, c, X, 0, 0); // expected-error{{argument to '__builtin_amdgcn_mfma_f32_16x16x32_f16' must be a constant integer}}
+  *out = __builtin_amdgcn_mfma_f32_16x16x32_f16(a, b, c, 0, X, 0); // expected-error{{argument to '__builtin_amdgcn_mfma_f32_16x16x32_f16' must be a constant integer}}
+  *out = __builtin_amdgcn_mfma_f32_16x16x32_f16(a, b, c, 0, 0, X); // expected-error{{argument to '__builtin_amdgcn_mfma_f32_16x16x32_f16' must be a constant integer}}
+}
+
+
+void test_mfma_f32_32x32x16_f16(__global float16* out, half8 a, half8 b, float16 c, int X) {
+  *out = __builtin_amdgcn_mfma_f32_32x32x16_f16(a, b, c, X, 0, 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, 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}}
+}
diff --git a/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950.cl b/clang/test/SemaOpenCL/builtins-amdgcn-error-gfx950.cl
@@ -0,0 +1,12 @@
+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -triple amdgcn-- -target-cpu gfx940 -verify -S -o - %s
+
+typedef float float4 __attribute__((ext_vector_type(4)));
+typedef float float16 __attribute__((ext_vector_type(16)));
+typedef half half8 __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) {
+  *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}}
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -3139,6 +3139,15 @@ def int_amdgcn_cvt_sr_fp8_f32 : ClangBuiltin<"__builtin_amdgcn_cvt_sr_fp8_f32">,
             [llvm_float_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty],
             [IntrNoMem, ImmArg<ArgIndex<3>>]>;
 
+//===----------------------------------------------------------------------===//
+// gfx950 intrinsics
+//===----------------------------------------------------------------------===//
+
+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>;
+}
+
 //===----------------------------------------------------------------------===//
 // Special Intrinsics for backend internal use only. No frontend
 // should emit calls to these.
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1980,6 +1980,14 @@ def isGFX940Plus :
   Predicate<"Subtarget->hasGFX940Insts()">,
   AssemblerPredicate<(all_of FeatureGFX940Insts)>;
 
+def isNotGFX940Plus :
+  Predicate<"!Subtarget->hasGFX940Insts()">,
+  AssemblerPredicate<(all_of (not FeatureGFX940Insts))>;
+
+def HasGFX950Insts :
+  Predicate<"Subtarget->hasGFX950Insts()">,
+  AssemblerPredicate<(all_of FeatureGFX950Insts)>;
+
 def isGFX8GFX9NotGFX940 :
   Predicate<"!Subtarget->hasGFX940Insts() &&"
             "(Subtarget->getGeneration() == AMDGPUSubtarget::VOLCANIC_ISLANDS ||"
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -4724,7 +4724,9 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
     case Intrinsic::amdgcn_mfma_f32_32x32x16_bf8_bf8:
     case Intrinsic::amdgcn_mfma_f32_32x32x16_bf8_fp8:
     case Intrinsic::amdgcn_mfma_f32_32x32x16_fp8_bf8:
-    case Intrinsic::amdgcn_mfma_f32_32x32x16_fp8_fp8: {
+    case Intrinsic::amdgcn_mfma_f32_32x32x16_fp8_fp8:
+    case Intrinsic::amdgcn_mfma_f32_16x16x32_f16:
+    case Intrinsic::amdgcn_mfma_f32_32x32x16_f16: {
       // Default for MAI intrinsics.
       // srcC can also be an immediate which can be folded later.
       // FIXME: Should we eventually add an alternative mapping with AGPR src
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td b/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
@@ -341,6 +341,8 @@ foreach intr = AMDGPUMFMAIntrinsics90A in
 def : SourceOfDivergence<intr>;
 foreach intr = AMDGPUMFMAIntrinsics940 in
 def : SourceOfDivergence<intr>;
+foreach intr = AMDGPUMFMAIntrinsics950 in
+def : SourceOfDivergence<intr>;
 foreach intr = AMDGPUWMMAIntrinsicsGFX11 in
 def : SourceOfDivergence<intr>;
 foreach intr = AMDGPUWMMAIntrinsicsGFX12 in
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -1284,6 +1284,10 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   // hasGFX90AInsts is also true.
   bool hasGFX940Insts() const { return GFX940Insts; }
 
+  // GFX950 is a derivation to GFX940. hasGFX950Insts() implies that
+  // hasGFX940Insts and hasGFX90AInsts are also true.
+  bool hasGFX950Insts() const { return GFX950Insts; }
+
   bool hasSALUFloatInsts() const { return HasSALUFloatInsts; }
 
   bool hasVGPRSingleUseHintInsts() const { return HasVGPRSingleUseHintInsts; }
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -2645,6 +2645,10 @@ def VOP_V16I32_V2I32_V4I32_I32    : VOPProfile <[v16i32, v2i32, v4i32, i32]>;
 def VOP_V4F32_V2I32_V4I32_I32     : VOPProfile <[v4f32,  v2i32, v4i32, i32]>;
 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]>;
+
+
 class Commutable_REV <string revOp, bit isOrig> {
   string RevOp = revOp;
   bit IsOrig = isOrig;
diff --git a/llvm/lib/Target/AMDGPU/VOP3PInstructions.td b/llvm/lib/Target/AMDGPU/VOP3PInstructions.td
@@ -630,6 +630,11 @@ def VOPProfileSMFMAC_I32_32X32X32_I8  : VOPProfileSMFMAC<VOP_V16I32_V2I32_V4I32_
 def VOPProfileSMFMAC_F32_16X16X64_F8  : VOPProfileSMFMAC<VOP_V4F32_V2I32_V4I32_I32,  AVDst_128, AVSrc_64, AVSrc_128>;
 def VOPProfileSMFMAC_F32_32X32X32_F8  : VOPProfileSMFMAC<VOP_V16F32_V2I32_V4I32_I32, AVDst_512, AVSrc_64, AVSrc_128>;
 
+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>;
+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>;
+
 class MFMATable <bit is_mac, string Name> {
   bit IsMac = is_mac;
   string FMAOp = Name;
@@ -743,6 +748,11 @@ defm V_MFMA_F32_32X32X4BF16 : MAIInst<"v_mfma_f32_32x32x4bf16", "F32_V2I16_X16",
 
 } // End SubtargetPredicate = HasMAIInsts
 
+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>;
+}
+
 let Predicates = [isGFX90APlus] in {
   let is_gfx940_xdl = 1 in {
   defm V_MFMA_F32_32X32X4BF16_1K  : MAIInst<"v_mfma_f32_32x32x4bf16_1k",  "F32_V4I16_X32",  int_amdgcn_mfma_f32_32x32x4bf16_1k>;
@@ -1654,6 +1664,16 @@ multiclass VOP3P_Real_MFMA_gfx940<bits<7> op, string Name = !cast<VOP3_Pseudo>(N
   }
 }
 
+multiclass VOP3P_Real_MFMA_gfx950<bits<7> op, string Name = !cast<VOP3_Pseudo>(NAME#"_e64").Mnemonic,
+                                  VOP3_Pseudo PS_ACD = !cast<VOP3_Pseudo>(NAME # "_e64"),
+                                  VOP3_Pseudo PS_VCD = !cast<VOP3_Pseudo>(NAME # "_vgprcd" # "_e64")> {
+    let SubtargetPredicate = HasGFX950Insts,
+        AssemblerPredicate = HasGFX950Insts in {
+    defm "" : VOP3P_Real_MFMA_gfx940<op, Name, PS_ACD, PS_VCD>;
+  }
+}
+
+
 multiclass VOP3P_Real_MFMA_vi<bits<7> op> {
   def _vi : VOP3P_Real<!cast<VOP3_Pseudo>(NAME#"_e64"), SIEncodingFamily.VI>,
             VOP3Pe_MAI <op, !cast<VOP3_Pseudo>(NAME#"_e64").Pfl, ?> {
@@ -1768,6 +1788,8 @@ defm V_MFMA_F32_16X16X16BF16_1K : VOP3P_Real_MFMA_gfx90a <0x67>;
 defm V_MFMA_F64_16X16X4F64      : VOP3P_Real_MFMA_gfx90a <0x6e>;
 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_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
@@ -261,6 +261,23 @@ bb:
   ret void
 }
 
+declare <4 x float> @llvm.amdgcn.mfma.f32.16x16x32.f16(<8 x half>, <8 x half>, <4 x float>, i32 immarg, i32 immarg, i32 immarg)
+declare <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half>, <8 x half>, <16 x float>, i32 immarg, i32 immarg, i32 immarg)
+
+; CHECK: DIVERGENT: %result = call <4 x float> @llvm.amdgcn.mfma.f32.16x16x32.f16(<8 x half> %arg0, <8 x half> %arg1, <4 x float> %arg2, i32 immarg 0, i32 immarg 0, i32 immarg 0)
+define amdgpu_kernel void @mfma_f32_16x16x32_f16(<8 x half> %arg0, <8 x half> %arg1, <4 x float> %arg2, ptr addrspace(1) %out) {
+  %result = call <4 x float> @llvm.amdgcn.mfma.f32.16x16x32.f16(<8 x half> %arg0, <8 x half> %arg1, <4 x float> %arg2, i32 immarg 0, i32 immarg 0, i32 immarg 0)
+  store <4 x float> %result, ptr addrspace(1) %out
+  ret void
+}
+
+; CHECK: DIVERGENT: %result = call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> %arg0, <8 x half> %arg1, <16 x float> %arg2, i32 immarg 0, i32 immarg 0, i32 immarg 0)
+define amdgpu_kernel void @mfma_f32_32x32x16_f16(<8 x half> %arg0, <8 x half> %arg1, <16 x float> %arg2, ptr addrspace(1) %out) {
+  %result = call <16 x float> @llvm.amdgcn.mfma.f32.32x32x16.f16(<8 x half> %arg0, <8 x half> %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.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mfma.gfx950.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
