Use arith/math ops instead of spirv.CL.* ops in tests. (#702)

Author: charithaintc

include/imex/Conversion/XeGPUToSPIRV/XeGPUToSPIRV.h

@@ -14,6 +14,7 @@
 #ifndef IMEX_CONVERSION_XEGPUTOSPIRV_H
 #define IMEX_CONVERSION_XEGPUTOSPIRV_H
 
+#include "mlir/Dialect/SPIRV/Transforms/SPIRVConversion.h"
 #include <mlir/Dialect/SPIRV/IR/SPIRVDialect.h>
 #include <mlir/Dialect/SPIRV/IR/SPIRVOps.h>
 #include <mlir/Transforms/DialectConversion.h>
@@ -25,7 +26,9 @@ class Pass;
 
 namespace imex {
 
-template <typename SPIRVOp> std::string getVCIntrinsicName(SPIRVOp op);
+// helper to check the legal vector lengths for arith/math ops
+bool isGenericVectorTy(mlir::Type type);
+
 // XeGPU to VC Intrinsics pattern
 void populateXeGPUToVCIntrinsicsPatterns(
     mlir::SPIRVTypeConverter &typeConverter, mlir::RewritePatternSet &patterns);

lib/Conversion/GPUToSPIRV/GPUToSPIRVPass.cpp

@@ -211,6 +211,8 @@ void GPUXToSPIRVPass::runOnOperation() {
     mlir::RewritePatternSet patterns(context);
     mlir::SPIRVConversionOptions options;
     options.use64bitIndex = true;
+    // FIXME: activate fast math per operator basis.
+    options.enableFastMathMode = true;
 
     mlir::SPIRVTypeConverter typeConverter(targetAttr, options);
 
@@ -360,9 +362,24 @@ void GPUXToSPIRVPass::runOnOperation() {
           });
     }
 
+    // SPIR-V elementwise arith/math ops require special handling if the operate
+    // on large vectors. We dynamically legalize these ops based on the vector
+    // size they consume.
+    // FIXME: this is not an exhaustive list of arith/math ops that need special
+    // handling.
+    target->addDynamicallyLegalOp<mlir::spirv::CLExpOp>(
+        [&](mlir::spirv::CLExpOp op) {
+          return imex::isGenericVectorTy(op.getType());
+        });
+    target->addDynamicallyLegalOp<mlir::spirv::CLFMaxOp>(
+        [&](mlir::spirv::CLFMaxOp op) {
+          return imex::isGenericVectorTy(op.getType());
+        });
+
     //------- Upstream Conversion------------
     mlir::populateGPUToSPIRVPatterns(typeConverter, patterns);
     mlir::arith::populateArithToSPIRVPatterns(typeConverter, patterns);
+    mlir::populateMathToSPIRVPatterns(typeConverter, patterns);
     mlir::populateMemRefToSPIRVPatterns(typeConverter, patterns);
     mlir::populateFuncToSPIRVPatterns(typeConverter, patterns);
     // ---------------------------------------

lib/Conversion/XeGPUToSPIRV/XeGPUToSPIRV.cpp

@@ -363,7 +363,7 @@ class UpdateNDOffsetToVCPattern : public OpConversionPattern<UpdateNDOffsetOp> {
     auto desc = adaptor.getTensorDesc();
     for (size_t i = 0; i < offsets.size(); i++) {
       auto offset = offsets[i];
-      if (auto cst = dyn_cast<spirv::ConstantOp>(offset.getDefiningOp()))
+      if (auto cst = offset.getDefiningOp<arith::ConstantOp>())
         if (auto attr = dyn_cast<mlir::IntegerAttr>(cst.getValue());
             attr && attr.getInt() == 0)
           continue;
@@ -1538,21 +1538,13 @@ struct SPIRVElementwiseToVC : public OpConversionPattern<SPIRVOp> {
     if (!dstType)
       return failure();
 
+    // This lowering pattern is needed only for spirv ops with large vector
+    // lengths.
+    assert(
+        !imex::isGenericVectorTy(dstType) &&
+        "Vector size is considered generic and op does not require lowering to "
+        "VC intrinsic. Consider marking this op + vector length as legal.");
     auto vecSize = dstType.template dyn_cast<VectorType>().getNumElements();
-    auto hasGenericVecSize = [&]() -> bool {
-      // if the input is scalar, we keep the operation as is.
-      if (isa<spirv::ScalarType>(dstType))
-        return true;
-      // or, if the vector size is 2, 3, 4, 8, or 16, we keep the operation.
-      return vecSize == 2 || vecSize == 3 || vecSize == 4 || vecSize == 8 ||
-             vecSize == 16;
-    };
-
-    if (hasGenericVecSize()) {
-      rewriter.replaceOpWithNewOp<SPIRVOp>(op, dstType, adaptor.getOperands());
-      return success();
-    }
-
     // for larger vector lengths, "llvm.genx.exp" returns the base 2
     // exponentiation of the input. To get the base e exponentiation, we need to
     // scale the input by log2(e)
@@ -1588,6 +1580,14 @@ struct SPIRVElementwiseToVC : public OpConversionPattern<SPIRVOp> {
 };
 } // namespace
 
+bool imex::isGenericVectorTy(mlir::Type type) {
+  if (isa<spirv::ScalarType>(type))
+    return true;
+  auto vecSize = type.dyn_cast<VectorType>().getNumElements();
+  return vecSize == 2 || vecSize == 3 || vecSize == 4 || vecSize == 8 ||
+         vecSize == 16;
+}
+
 void imex::populateXeGPUToVCIntrinsicsPatterns(
     SPIRVTypeConverter &typeConverter, RewritePatternSet &patterns) {
   patterns.add<CreateNdDescToSPIRV, CreateDescToVCPattern, DpasToVCPattern,

test/Integration/Dialect/XeGPU/exp_f32.vc.mlir

@@ -35,7 +35,7 @@ module @gemm attributes {gpu.container_module} {
       // do DPAS
       %val4 = xegpu.dpas %val0, %val2 : vector<8x8x2xf16>, vector<8x16x2xf16> -> vector<8x16xf32>
       // take exp
-      %t6 = spirv.CL.exp %val4 : vector<8x16xf32>
+      %t6 = math.exp %val4 : vector<8x16xf32>
       // store
       %out_tile = xegpu.create_nd_tdesc %Out [%c0, %c0] { mode = vc } : memref<8x16xf32> -> !xegpu.tensor_desc<8x16xf32>
       xegpu.store_nd %t6, %out_tile { mode = vc} : vector<8x16xf32>, !xegpu.tensor_desc<8x16xf32>
@@ -65,14 +65,14 @@ module @gemm attributes {gpu.container_module} {
       %v6 = vector.extract %val4[6] : vector<16xf32> from vector<8x16xf32>
       %v7 = vector.extract %val4[7] : vector<16xf32> from vector<8x16xf32>
       // do generic size exp
-      %v0_exp = spirv.CL.exp %v0 : vector<16xf32>
-      %v1_exp = spirv.CL.exp %v1 : vector<16xf32>
-      %v2_exp = spirv.CL.exp %v2 : vector<16xf32>
-      %v3_exp = spirv.CL.exp %v3 : vector<16xf32>
-      %v4_exp = spirv.CL.exp %v4 : vector<16xf32>
-      %v5_exp = spirv.CL.exp %v5 : vector<16xf32>
-      %v6_exp = spirv.CL.exp %v6 : vector<16xf32>
-      %v7_exp = spirv.CL.exp %v7 : vector<16xf32>
+      %v0_exp = math.exp %v0 : vector<16xf32>
+      %v1_exp = math.exp %v1 : vector<16xf32>
+      %v2_exp = math.exp %v2 : vector<16xf32>
+      %v3_exp = math.exp %v3 : vector<16xf32>
+      %v4_exp = math.exp %v4 : vector<16xf32>
+      %v5_exp = math.exp %v5 : vector<16xf32>
+      %v6_exp = math.exp %v6 : vector<16xf32>
+      %v7_exp = math.exp %v7 : vector<16xf32>
       %v0_exp_cast = vector.shape_cast %v0_exp : vector<16xf32> to vector<1x16xf32>
       %v1_exp_cast = vector.shape_cast %v1_exp : vector<16xf32> to vector<1x16xf32>
       %v2_exp_cast = vector.shape_cast %v2_exp : vector<16xf32> to vector<1x16xf32>

test/Integration/Dialect/XeGPU/fmax_f32.vc.mlir

@@ -37,7 +37,7 @@ module @gemm attributes {gpu.container_module} {
       %val4 = xegpu.dpas %val0, %val2 : vector<8x8x2xf16>, vector<8x16x2xf16> -> vector<8x16xf32>
       %val5 = xegpu.dpas %val1, %val3 : vector<8x8x2xf16>, vector<8x16x2xf16> -> vector<8x16xf32>
       // take fmax
-      %val6 = spirv.CL.fmax %val4, %val5 : vector<8x16xf32>
+      %val6 = arith.maximumf %val4, %val5 fastmath<nnan> : vector<8x16xf32>
       // store fmax
       %out_tile = xegpu.create_nd_tdesc %Out [%c0, %c0] { mode = vc } : memref<8x16xf32> -> !xegpu.tensor_desc<8x16xf32>
       xegpu.store_nd %val6, %out_tile { mode = vc} : vector<8x16xf32>, !xegpu.tensor_desc<8x16xf32>

test/Integration/Dialect/XeGPU/xegpu-to-llvm.pp

@@ -1,6 +1,7 @@
 // linalg dialect to gpu dialect lowering pipeline
 // Ready for vulkan runner or narrow scope l0/sycl runner starting from GPU dialect.
 builtin.module(
+    imex-vector-linearize
     imex-convert-gpu-to-spirv{enable-vc-intrinsic=true}
     spirv.module(spirv-lower-abi-attrs
              spirv-update-vce)