[AMDGPU] Optimize _L image intrinsic to _LZ when lod is zero

[rtaylor]

Summary

Add _L to _LZ image intrinsic table mapping to table gen.
In ISelLowering check if image intrinsic has lod and if it's equal
to zero, if so remove lod and change opcode to equivalent mapped _LZ.

Change-Id: Ie24cd7e788e2195d846c7bd256151178cbb9ec71

[arsenm]

lib/Target/AMDGPU/SIISelLowering.cpp

4631

Does this need to check that it is positive zero?

[arsenm]

Should this be done in an IR pass instead?

[rtaylor]

In D49483#1166525, @arsenm wrote:

Should this be done in an IR pass instead?

This could be done in the IR but this avoids a long(er) switch statement (# of combinations) and keeps more of the image intrinsic work in the same place, which seemed per conversations the best way to go. Is there some advantage to moving this to the IR (InstCombine for example)?

lib/Target/AMDGPU/SIISelLowering.cpp

4631

I used isZero to cover both pos and neg zero.

[rtaylor]

Ping.

[nhaehnle]

I don't think doing this as an IR pass has any advantage, so this is fine.

Please add tests for gather intrinsics as well, apart from that it looks good to me.

lib/Target/AMDGPU/SIISelLowering.cpp

4631

That's an interesting question. Since LOD is clamped at 0.0 anyway, the only possible difference could come from switching between magnification and minification filters. The OpenGL spec says that the magnification filter is used when LOD is less than or equal to 0.0 (and actually even when it's <= 0.5 for nearest mipmap minification). So this test could be relaxed to check whether ConstantLod is <= 0.0.

[rtaylor]

Added gather4 test cases.
Added 'isNegative' to _L to _LZ check to include <= 0.0

[nhaehnle]

Thanks! LGTM

lib/Target/AMDGPU/MIMGInstructions.td

def MIMGDimInfoTable : GenericTable {
   let CppTypeName = "MIMGDimInfo";
   let Fields = ["Dim", "NumCoords", "NumGradients", "DA"];
   GenericEnum TypeOf_Dim = MIMGDim;
 
   let PrimaryKey = ["Dim"];
   let PrimaryKeyName = "getMIMGDimInfo";
 }
 
+class MIMGLZMapping<MIMGBaseOpcode l, MIMGBaseOpcode lz> {
+  MIMGBaseOpcode L = l;
+  MIMGBaseOpcode LZ = lz;
+}
+
+def MIMGLZMappingTable : GenericTable {
+  let FilterClass = "MIMGLZMapping";
+  let CppTypeName = "MIMGLZMappingInfo";
+  let Fields = ["L", "LZ"];
+  GenericEnum TypeOf_L = MIMGBaseOpcode;
+  GenericEnum TypeOf_LZ = MIMGBaseOpcode;
+
+  let PrimaryKey = ["L"];
+  let PrimaryKeyName = "getMIMGLZMappingInfo";
+}
+
 class mimg <bits<7> si, bits<7> vi = si> {
   field bits<7> SI = si;
   field bits<7> VI = vi;
 }
 
 class MIMG <dag outs, string dns = "">
   : InstSI <outs, (ins), "", []> {
 
def ImageDimIntrinsicTable : GenericTable {
   let PrimaryKeyName = "getImageDimIntrinsicInfo";
   let PrimaryKeyEarlyOut = 1;
 }
 
 foreach intr = !listconcat(AMDGPUImageDimIntrinsics,
                            AMDGPUImageDimAtomicIntrinsics) in {
   def : ImageDimIntrinsicInfo<intr>;
 }
+
+// L to LZ Optimization Mapping
+def : MIMGLZMapping<IMAGE_SAMPLE_L, IMAGE_SAMPLE_LZ>;
+def : MIMGLZMapping<IMAGE_SAMPLE_C_L, IMAGE_SAMPLE_C_LZ>;
+def : MIMGLZMapping<IMAGE_SAMPLE_L_O, IMAGE_SAMPLE_LZ_O>;
+def : MIMGLZMapping<IMAGE_SAMPLE_C_L_O, IMAGE_SAMPLE_C_LZ_O>;
+def : MIMGLZMapping<IMAGE_GATHER4_L, IMAGE_GATHER4_LZ>;
+def : MIMGLZMapping<IMAGE_GATHER4_C_L, IMAGE_GATHER4_C_LZ>;
+def : MIMGLZMapping<IMAGE_GATHER4_L_O, IMAGE_GATHER4_LZ_O>;
+def : MIMGLZMapping<IMAGE_GATHER4_C_L_O, IMAGE_GATHER4_C_LZ_O>;

lib/Target/AMDGPU/SIISelLowering.cpp

 SDValue SITargetLowering::lowerImage(SDValue Op,
                                      const AMDGPU::ImageDimIntrinsicInfo *Intr,
                                      SelectionDAG &DAG) const {
   SDLoc DL(Op);
   MachineFunction &MF = DAG.getMachineFunction();
   const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
       AMDGPU::getMIMGBaseOpcodeInfo(Intr->BaseOpcode);
   const AMDGPU::MIMGDimInfo *DimInfo = AMDGPU::getMIMGDimInfo(Intr->Dim);
+  const AMDGPU::MIMGLZMappingInfo *LZMappingInfo =
+      AMDGPU::getMIMGLZMappingInfo(Intr->BaseOpcode);
+  unsigned IntrOpcode = Intr->BaseOpcode;
 
   SmallVector<EVT, 2> ResultTypes(Op->value_begin(), Op->value_end());
   bool IsD16 = false;
   SDValue VData;
   int NumVDataDwords;
   unsigned AddrIdx; // Index of first address argument
   unsigned DMask;
 
SDValue SITargetLowering::lowerImage(SDValue Op,
 
   unsigned NumVAddrs = BaseOpcode->NumExtraArgs +
                        (BaseOpcode->Gradients ? DimInfo->NumGradients : 0) +
                        (BaseOpcode->Coordinates ? DimInfo->NumCoords : 0) +
                        (BaseOpcode->LodOrClampOrMip ? 1 : 0);
   SmallVector<SDValue, 4> VAddrs;
   for (unsigned i = 0; i < NumVAddrs; ++i)
     VAddrs.push_back(Op.getOperand(AddrIdx + i));
+
+  // Optimize _L to _LZ when _L is zero
+  if (LZMappingInfo) {
+    if (auto ConstantLod =
+         dyn_cast<ConstantFPSDNode>(VAddrs[NumVAddrs-1].getNode())) {
+      if (ConstantLod->isZero() || ConstantLod->isNegative()) {
+        IntrOpcode = LZMappingInfo->LZ;  // set new opcode to _lz variant of _l
+        VAddrs.pop_back();               // remove 'lod'
+      }
+    }
+  }
+
   SDValue VAddr = getBuildDwordsVector(DAG, DL, VAddrs);
 
   SDValue True = DAG.getTargetConstant(1, DL, MVT::i1);
   SDValue False = DAG.getTargetConstant(0, DL, MVT::i1);
   unsigned CtrlIdx; // Index of texfailctrl argument
   SDValue Unorm;
   if (!BaseOpcode->Sampler) {
     Unorm = True;
if (BaseOpcode->HasD16)
     Ops.push_back(IsD16 ? True : False);
   if (isa<MemSDNode>(Op))
     Ops.push_back(Op.getOperand(0)); // chain
 
   int NumVAddrDwords = VAddr.getValueType().getSizeInBits() / 32;
   int Opcode = -1;
 
   if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
-    Opcode = AMDGPU::getMIMGOpcode(Intr->BaseOpcode, AMDGPU::MIMGEncGfx8,
+    Opcode = AMDGPU::getMIMGOpcode(IntrOpcode, AMDGPU::MIMGEncGfx8,
                                    NumVDataDwords, NumVAddrDwords);
   if (Opcode == -1)
-    Opcode = AMDGPU::getMIMGOpcode(Intr->BaseOpcode, AMDGPU::MIMGEncGfx6,
+    Opcode = AMDGPU::getMIMGOpcode(IntrOpcode, AMDGPU::MIMGEncGfx6,
                                    NumVDataDwords, NumVAddrDwords);
   assert(Opcode != -1);
 
   MachineSDNode *NewNode = DAG.getMachineNode(Opcode, DL, ResultTypes, Ops);
   if (auto MemOp = dyn_cast<MemSDNode>(Op)) {
     MachineInstr::mmo_iterator MemRefs = MF.allocateMemRefsArray(1);
     *MemRefs = MemOp->getMemOperand();
     NewNode->setMemRefs(MemRefs, MemRefs + 1);

lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

 class MachineMemOperand;
 class Triple;
 
 namespace AMDGPU {
 
 #define GET_MIMGBaseOpcode_DECL
 #define GET_MIMGDim_DECL
 #define GET_MIMGEncoding_DECL
+#define GET_MIMGLZMapping_DECL
 #include "AMDGPUGenSearchableTables.inc"
 
 namespace IsaInfo {
 
 enum {
   // The closed Vulkan driver sets 96, which limits the wave count to 8 but
   // doesn't spill SGPRs as much as when 80 is set.
   FIXED_NUM_SGPRS_FOR_INIT_BUG = 96,
struct MIMGDimInfo {
   uint8_t NumCoords;
   uint8_t NumGradients;
   bool DA;
 };
 
 LLVM_READONLY
 const MIMGDimInfo *getMIMGDimInfo(unsigned Dim);
 
+struct MIMGLZMappingInfo {
+  MIMGBaseOpcode L;
+  MIMGBaseOpcode LZ;
+};
+
+LLVM_READONLY
+const MIMGLZMappingInfo *getMIMGLZMappingInfo(unsigned L);
+
 LLVM_READONLY
 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding,
                   unsigned VDataDwords, unsigned VAddrDwords);
 
 LLVM_READONLY
 int getMaskedMIMGOp(unsigned Opc, unsigned NewChannels);
 
 LLVM_READONLY

lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

struct MIMGInfo {
   uint8_t MIMGEncoding;
   uint8_t VDataDwords;
   uint8_t VAddrDwords;
 };
 
 #define GET_MIMGBaseOpcodesTable_IMPL
 #define GET_MIMGDimInfoTable_IMPL
 #define GET_MIMGInfoTable_IMPL
+#define GET_MIMGLZMappingTable_IMPL
 #include "AMDGPUGenSearchableTables.inc"
 
 int getMIMGOpcode(unsigned BaseOpcode, unsigned MIMGEncoding,
                   unsigned VDataDwords, unsigned VAddrDwords) {
   const MIMGInfo *Info = getMIMGOpcodeHelper(BaseOpcode, MIMGEncoding,
                                              VDataDwords, VAddrDwords);
   return Info ? Info->Opcode : -1;
 }

test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.ltolz.ll

+; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
+; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN %s
+
+
+; GCN-LABEL: {{^}}sample_l_1d:
+; GCN: image_sample_lz v[0:3], v0, s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_l_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %s, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.l.1d.v4f32.f32(i32 15, float %s, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}sample_l_2d:
+; GCN: image_sample_lz v[0:3], v[0:1], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_l_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.l.2d.v4f32.f32(i32 15, float %s, float %t, float -0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}sample_c_l_1d:
+; GCN: image_sample_c_lz v[0:3], v[0:1], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_c_l_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %zcompare, float %s, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.c.l.1d.v4f32.f32(i32 15, float %zcompare, float %s, float -2.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}sample_c_l_2d:
+; GCN: image_sample_c_lz v[0:3], v[0:3], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_c_l_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %zcompare, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.c.l.2d.v4f32.f32(i32 15, float %zcompare, float %s, float %t, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}sample_l_o_1d:
+; GCN: image_sample_lz_o v[0:3], v[0:1], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_l_o_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, i32 %offset, float %s, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.l.o.1d.v4f32.f32(i32 15, i32 %offset, float %s, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}sample_l_o_2d:
+; GCN: image_sample_lz_o v[0:3], v[0:3], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_l_o_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, i32 %offset, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.l.o.2d.v4f32.f32(i32 15, i32 %offset, float %s, float %t, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}sample_c_l_o_1d:
+; GCN: image_sample_c_lz_o v[0:3], v[0:3], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_c_l_o_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, i32 %offset, float %zcompare, float %s, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.c.l.o.1d.v4f32.f32(i32 15, i32 %offset, float %zcompare, float %s, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}sample_c_l_o_2d:
+; GCN: image_sample_c_lz_o v[0:3], v[0:3], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @sample_c_l_o_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, i32 %offset, float %zcompare, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.sample.c.l.o.2d.v4f32.f32(i32 15, i32 %offset, float %zcompare, float %s, float %t, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}gather4_l_2d:
+; GCN: image_gather4_lz v[0:3], v[0:1], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @gather4_l_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.gather4.l.2d.v4f32.f32(i32 15, float %s, float %t, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}gather4_c_l_2d:
+; GCN: image_gather4_c_lz v[0:3], v[0:3], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @gather4_c_l_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %zcompare, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.gather4.c.l.2d.v4f32.f32(i32 15, float %zcompare, float %s, float %t, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}gather4_l_o_2d:
+; GCN: image_gather4_lz_o v[0:3], v[0:3], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @gather4_l_o_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, i32 %offset, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.gather4.l.o.2d.v4f32.f32(i32 15, i32 %offset, float %s, float %t, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+; GCN-LABEL: {{^}}gather4_c_l_o_2d:
+; GCN: image_gather4_c_lz_o v[0:3], v[0:3], s[0:7], s[8:11] dmask:0xf{{$}}
+define amdgpu_ps <4 x float> @gather4_c_l_o_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, i32 %offset, float %zcompare, float %s, float %t, float %lod) {
+main_body:
+  %v = call <4 x float> @llvm.amdgcn.image.gather4.c.l.o.2d.v4f32.f32(i32 15, i32 %offset, float %zcompare, float %s, float %t, float 0.0, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 0, i32 0)
+  ret <4 x float> %v
+}
+
+declare <4 x float> @llvm.amdgcn.image.sample.l.1d.v4f32.f32(i32, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.sample.l.2d.v4f32.f32(i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.sample.c.l.1d.v4f32.f32(i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.sample.c.l.2d.v4f32.f32(i32, float, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.sample.l.o.1d.v4f32.f32(i32, i32, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.sample.l.o.2d.v4f32.f32(i32, i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.sample.c.l.o.1d.v4f32.f32(i32, i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.sample.c.l.o.2d.v4f32.f32(i32, i32, float, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+
+declare <4 x float> @llvm.amdgcn.image.gather4.l.2d.v4f32.f32(i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.gather4.c.l.2d.v4f32.f32(i32, float, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.gather4.l.o.2d.v4f32.f32(i32, i32, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1
+declare <4 x float> @llvm.amdgcn.image.gather4.c.l.o.2d.v4f32.f32(i32, i32, float, float, float, float, <8 x i32>, <4 x i32>, i1, i32, i32) #1