[mlir][linalg] Fix linalg.conv vectorization for mixed int-fp types

[rsuderman]

Summary

We always assume mixed same type values. Instead of ExtF or ExtSI, we
need SIToFp when the values must be promoted.

[dcaballe]

LGTM! Just one comments. THank!

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

2630–2634

Could you please also add an extra condition to these if statements to make sure that both dst and src have the expected type? IIUC, we were incorrectly going through one of them and not hitting the assert below.

[rsuderman]

Add matching element type

mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

Value promote(RewriterBase &rewriter, Location loc, Value val, Type ty) {
     if (srcElementType == dstElementType)
       return val;
 
     const int64_t srcWidth = srcElementType.getIntOrFloatBitWidth();
     const int64_t dstWidth = dstElementType.getIntOrFloatBitWidth();
     const Type dstType =
         cast<ShapedType>(val.getType()).cloneWith(std::nullopt, dstElementType);
 
+    if (isa<IntegerType>(srcElementType) && isa<FloatType>(dstElementType)) {
+      return rewriter.create<arith::SIToFPOp>(loc, dstType, val);
+    }
+
     if (isa<FloatType>(dstElementType) && srcWidth < dstWidth)
       return rewriter.create<arith::ExtFOp>(loc, dstType, val);
 
     if (isa<IntegerType>(dstElementType) && srcWidth < dstWidth)
       return rewriter.create<arith::ExtSIOp>(loc, dstType, val);
 
     assert(false && "unhandled promotion case");
     return nullptr;
   }
 
   // Create a contraction: lhs{n, w, c} * rhs{c, f} -> res{n, w, f}
   Value conv1dSliceAsContraction(RewriterBase &rewriter, Location loc,
                                  Value lhs, Value rhs, Value res) {

mlir/test/Dialect/Linalg/vectorize-convolution.mlir

 //      CHECK:   %[[V_OUTPUT_R:.+]] = vector.transfer_read %[[OUTPUT]][%[[C0]], %[[C0]], %[[C0]]]
 //      CHECK:   %[[V_FILTER_1:.+]] = vector.extract %[[V_FILTER_R]][0] : vector<1x3x2xf16>
 //      CHECK:   %[[CONT:.*]] = vector.contract
 //        {{.*}} %[[V_INPUT_R]], %[[V_FILTER_1]], %[[V_OUTPUT_R]] : vector<1x2x3xf16>, vector<3x2xf16> into vector<1x2x2xf32>
 //      CHECK:   vector.transfer_write %[[CONT]], %[[OUTPUT]][%[[C0]], %[[C0]], %[[C0]]]
 
 // -----
 
+func.func @conv_1d_nwc_wcf_mixed_int_fp_memref(%input: memref<1x2x3xi8>, %filter: memref<1x3x2xi8>, %output: memref<1x2x2xf32>) {
+  linalg.conv_1d_nwc_wcf
+  {dilations = dense<1> : vector<1xi64>, strides = dense<1> : vector<1xi64>}
+   ins(%input, %filter : memref<1x2x3xi8>, memref<1x3x2xi8>)
+   outs(%output : memref<1x2x2xf32>)
+  return
+}
+
+
+// CHECK-LABEL: func @conv_1d_nwc_wcf_mixed_int_fp_memref
+// CHECK-SAME:   (%[[INPUT:[0-9a-z]+]]: memref<1x2x3xi8>, %[[FILTER:[0-9a-z]+]]: memref<1x3x2xi8>, %[[OUTPUT:[0-9a-z]+]]: memref<1x2x2xf32>)
+// CHECK-DAG: %[[CST:.+]] = arith.constant 0.000000e+00 : f32
+// CHECK-DAG: %[[I0:.+]] = arith.constant 0 : index
+// CHECK-DAG: %[[C0:.+]] = arith.constant 0 : i8
+// CHECK: %[[READ0:.+]] = vector.transfer_read %arg0[%[[I0]], %[[I0]], %[[I0]]], %[[C0]]
+// CHECK: %[[READ1:.+]] = vector.transfer_read %arg1[%[[I0]], %[[I0]], %[[I0]]], %[[C0]]
+// CHECK: %[[READ2:.+]] = vector.transfer_read %arg2[%[[I0]], %[[I0]], %[[I0]]], %[[CST]]
+// CHECK: %[[EXT:.+]] = vector.extract %[[READ1]][0] : vector<1x3x2xi8>
+// CHECK: %[[CAST0:.+]] = arith.sitofp %[[READ0]] : vector<1x2x3xi8> to vector<1x2x3xf32>
+// CHECK: %[[CAST1:.+]] = arith.sitofp %[[EXT]] : vector<3x2xi8> to vector<3x2xf32>
+// CHECK: %[[CONTRACT:.+]] = vector.contract {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "parallel", "reduction"], kind = #vector.kind<add>} %[[CAST0]], %[[CAST1]], %[[READ2]]
+// CHECK: vector.transfer_write %[[CONTRACT]], %arg2[%[[I0]], %[[I0]], %[[I0]]]
+
+// -----
+
 func.func @pooling_nwc_sum_memref_1_2_1_3(%input: memref<4x4x3xf32>, %filter: memref<1xf32>, %output: memref<4x2x3xf32>) {
   linalg.pooling_nwc_sum
     {dilations = dense<1> : tensor<1xi64>, strides = dense<3> : tensor<1xi64>}
     ins(%input, %filter : memref<4x4x3xf32>, memref<1xf32>)
     outs(%output : memref<4x2x3xf32>)
   return
 }