Adding loads from different value type.

Author: lialan

mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp

@@ -534,7 +534,7 @@ LogicalResult TransposeLoadOp::verify() {
   if (!transferType)
     return emitOpError("destination type must be a vector type");
   size_t numElements = transferType.getNumElements();
-  size_t elementTypeSize = srcType.getElementType().getIntOrFloatBitWidth();
+  size_t elementTypeSize = transferType.getElementType().getIntOrFloatBitWidth();
 
   // ElementSize -> NumElements
   const std::map<size_t, size_t> KValidLoadSizeMap = {

mlir/test/Conversion/AMDGPUToROCDL/transpose_load.mlir

@@ -34,6 +34,27 @@ func.func @transpose_load_to_rocdl_16xi4(%idx1 : index, %idx2 : index, %wgmem :
 // CHECK-LABEL: func @transpose_load_to_rocdl_3xi32
 func.func @transpose_load_to_rocdl_3xi32(%idx1 : index, %idx2 : index, %wgmem : memref<128x32xi32, 3>) -> vector<3xi32> {
   // CHECK: rocdl.ds.read.tr6.b96
+  // CHECK-OLD: error: 'amdgpu.transpose_load' op Non-gfx950 chipset not supported
   %0 = amdgpu.transpose_load %wgmem[%idx1, %idx2] : memref<128x32xi32, 3> -> vector<3xi32>
   return %0 : vector<3xi32>
 }
+
+// -----
+
+// CHECK-LABEL: func @transpose_load_to_rocdl_i4_memrefxi1
+func.func @transpose_load_to_rocdl_i4_memrefxi1(%idx1 : index, %idx2 : index, %wgmem : memref<128x32xi8, 3>) -> vector<16xi4> {
+  // CHECK: rocdl.ds.read.tr4.b64
+  // CHECK-OLD: error: 'amdgpu.transpose_load' op Non-gfx950 chipset not supported
+  %0 = amdgpu.transpose_load %wgmem[%idx1, %idx2] : memref<128x32xi8, 3> -> vector<16xi4>
+  return %0 : vector<16xi4>
+}
+
+// -----
+
+// CHECK-LABEL: func @transpose_load_to_rocdl_i6_memrefxi1
+func.func @transpose_load_to_rocdl_i6_memrefxi1(%idx1 : index, %idx2 : index, %wgmem : memref<128x32xi8, 3>) -> vector<3xi32> {
+  // CHECK: rocdl.ds.read.tr6.b96
+  // CHECK-OLD: error: 'amdgpu.transpose_load' op Non-gfx950 chipset not supported
+  %0 = amdgpu.transpose_load %wgmem[%idx1, %idx2] : memref<128x32xi8, 3> -> vector<3xi32>
+  return %0 : vector<3xi32>
+}