AMDGPU: Define new sched model for gfx950

llvm/lib/Target/AMDGPU/GCNProcessors.td

@@ -204,7 +204,7 @@ def : ProcessorModel<"gfx942", SIDPGFX940FullSpeedModel,
   FeatureISAVersion9_4_2.Features
 >;
 
-def : ProcessorModel<"gfx950", SIDPGFX940FullSpeedModel,
+def : ProcessorModel<"gfx950", SIDPGFX950FullSpeedModel,
   FeatureISAVersion9_5_0.Features
 >;
 

llvm/lib/Target/AMDGPU/SISchedule.td

@@ -64,6 +64,7 @@ def Write8PassMAI  : SchedWrite;
 def Write16PassMAI : SchedWrite;
 def Write4PassDGEMM : SchedWrite;
 def Write8PassDGEMM : SchedWrite;
+def Write16PassDGEMM : SchedWrite;
 
 // Scalar float instructions
 def WriteSFPU : SchedWrite;
@@ -94,6 +95,7 @@ def SIFullSpeedModel : SISchedMachineModel;
 def SIQuarterSpeedModel : SISchedMachineModel;
 def SIDPFullSpeedModel : SISchedMachineModel;
 def SIDPGFX940FullSpeedModel : SISchedMachineModel;
+def SIDPGFX950FullSpeedModel : SISchedMachineModel;
 def GFX10SpeedModel : SISchedMachineModel;
 def GFX11SpeedModel : SISchedMachineModel;
 def GFX12SpeedModel : SISchedMachineModel;
@@ -169,6 +171,8 @@ multiclass SICommonWriteRes {
   def : HWVALUWriteRes<Write4PassDGEMM,    4>;
   let ReleaseAtCycles = [8] in
   def : HWVALUWriteRes<Write8PassDGEMM,    8>;
+  let ReleaseAtCycles = [16] in
+  def : HWVALUWriteRes<Write16PassDGEMM,  16>;
 
   let ReleaseAtCycles = [2] in
   def : HWWriteRes<Write2PassMAI,  [HWXDL], 2>;
@@ -201,6 +205,13 @@ def WriteCopy : SchedWriteVariant<[
     SchedVar<PredIsVGPR64Copy, [Write64Bit]>,
     SchedVar<NoSchedPred, [WriteSALU]>]>;
 
+// Check if any matrix inputs are interpreted as f8 in an f8f6f4 mfma
+// instruction.
+def PredIsF8_MFMA_SCALE : SchedPredicate<[{
+  TII->getNamedOperand(*MI, AMDGPU::OpName::cbsz)->getImm() <= AMDGPU::MFMAScaleFormats::FP8_E5M2 ||
+  TII->getNamedOperand(*MI, AMDGPU::OpName::blgp)->getImm() <= AMDGPU::MFMAScaleFormats::FP8_E5M2
+}]>;
+
 let SchedModel = SIFullSpeedModel in {
 
 defm : SICommonWriteRes;
@@ -299,6 +310,58 @@ def : InstRW<[Write8PassMAI,   MIMFMARead], (instregex "^V_SMFMAC_.32_32X32X")>;
 
 } // End SchedModel = SIDPGFX940FullSpeedModel
 
+
+let SchedModel = SIDPGFX950FullSpeedModel in {
+defm : SICommonWriteRes;
+
+def : HWVALUWriteRes<WriteFloatFMA,    1>;
+def : HWVALUWriteRes<WriteDouble,      1>;
+def : HWVALUWriteRes<WriteDoubleAdd,   1>;
+def : HWVALUWriteRes<WriteDoubleCvt,   1>;
+def : HWVALUWriteRes<WriteTrans64,     4>;
+def : HWVALUWriteRes<WriteIntMul,      1>;
+def : HWVALUWriteRes<Write64Bit,       1>;
+
+def : InstRW<[WriteCopy], (instrs COPY)>;
+def : InstRW<[Write64Bit], (instregex "^V_ACCVGPR_WRITE_B32_e64$")>;
+def : InstRW<[Write2PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_4X4X")>;
+
+def : InstRW<[Write4PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_16X16X8X")>;
+def : InstRW<[Write4PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_16X16X16")>;
+def : InstRW<[Write4PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_16X16X32")>;
+def : InstRW<[Write4PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_16X16X64")>;
+def : InstRW<[Write8PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_16X16X[14][FBI]")>;
+
+def : InstRW<[Write8PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_32X32X4XF")>;
+def : InstRW<[Write8PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_32X32X8")>;
+def : InstRW<[Write8PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_32X32X16")>;
+def : InstRW<[Write8PassMAI,   MIMFMARead], (instregex "^V_MFMA_.32_32X32X32_")>;
+def : InstRW<[Write16PassMAI,  MIMFMARead], (instregex "^V_MFMA_.32_32X32X[124][FBI]")>;
+
+def : InstRW<[Write4PassDGEMM, MIMFMARead], (instregex "^V_MFMA_.64_4X4X")>;
+def : InstRW<[Write16PassDGEMM, MIMFMARead], (instregex "^V_MFMA_.64_16X16X")>;
+
+def : InstRW<[Write4PassMAI,   MIMFMARead], (instregex "^V_SMFMAC_.32_16X16X")>;
+def : InstRW<[Write8PassMAI,   MIMFMARead], (instregex "^V_SMFMAC_.32_32X32X")>;
+
+
+// If either matrix format is f8, the instruction takes 2x as many
+// cycles. TODO: This isn't reflected in MCA.
+def WriteMFMAScale_16X16X128_F8F6F4 : SchedWriteVariant<[
+    SchedVar<PredIsF8_MFMA_SCALE, [Write8PassMAI]>,
+    SchedVar<NoSchedPred, [Write4PassMAI]>]>;
+def WriteMFMAScale_32X32X64_F8F6F4 : SchedWriteVariant<[
+    SchedVar<PredIsF8_MFMA_SCALE, [Write16PassMAI]>,
+    SchedVar<NoSchedPred, [Write8PassMAI]>]>;
+
+def : InstRW<[WriteMFMAScale_16X16X128_F8F6F4, MIMFMARead],
+       (instregex "^V_MFMA(_SCALE)?_.32_16X16X128_F8F6F4")>;
+def : InstRW<[WriteMFMAScale_32X32X64_F8F6F4,  MIMFMARead],
+        (instregex "^V_MFMA(_SCALE)?_.32_32X32X64_F8F6F4")>;
+
+} // End SchedModel = SIDPGFX950FullSpeedModel
+
+
 let SchedModel = GFX10SpeedModel in {
 
 // The latency values are 1 / (operations / cycle).