[mlir][linalg] Relax convolution vectorization to support mixed types

[ThomasRaoux]

Summary

Support the case where convolution does float extension of the inputs.

[nicolasvasilache]

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

1377–1389

whose operands ?

1393

also ExtSI/ExtUI Ops ?

[ThomasRaoux]

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

1393

yes I was thinking about that although I'm unclear if contraction with mixed type works for integer type as I don't see a way to distinguish the extSI and extUI when the cast is merged in the contraction op?

[ThomasRaoux]

clarify comment

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

1377–1389

I rephrased the comment to make it more explicit that this is about mul op operands

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

Conv1DNwcGenerator(OpBuilder &builder, LinalgOp linalgOp, int strideW,
     // Check for reduction `add` preceded by `mul`.
     Operation *reduceOp = matchLinalgReduction(linalgOp.getOutputOperand(0));
     if (!reduceOp)
       return;
     llvm::Optional<vector::CombiningKind> maybeKind;
     maybeKind = getCombinerOpKind(reduceOp);
     if (!maybeKind || *maybeKind != vector::CombiningKind::ADD)
       return;
-    maybeKind = getCombinerOpKind(&(linalgOp->getRegion(0).front().front()));
-    if (!maybeKind || *maybeKind != vector::CombiningKind::MUL)
+    // Check for single `mul` predecessor. The `mul` operands must be block
+    // arguments or extension of block arguments.
+    Operation *mulOp = nullptr;
+    for (Value operand : reduceOp->getOperands()) {
+      if (operand.isa<BlockArgument>())
+        continue;
+      if (mulOp)
         return;
-
+      mulOp = operand.getDefiningOp();
+      if (!mulOp || !isa<arith::MulIOp, arith::MulFOp>(mulOp))
+        return;
+    }
+    if (!mulOp)
+      return;
+    for (Value operand : mulOp->getOperands()) {
+      if (Operation *def = operand.getDefiningOp()) {
+        if (!isa<arith::ExtFOp>(def))
+          return;
+        operand = def->getOperand(0);
+      }
+      if (!operand.isa<BlockArgument>())
+        return;
+    }
     // The op is now known to be valid.
     valid = true;
   }
 
   /// Generate a vector implementation for:
   /// ```
   ///   Op def: (     n,     w,     c,    kw,    f  )
   ///    Iters: ({Par(), Par(), Par(), Red(), Red()})

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

 //      CHECK:  %[[FMA_0:.*]] = vector.fma %[[V_INPUT_0]], %[[B_FILTER_0]], %[[V_OUTPUT_R]] : vector<3x2x4xf32>
 
 /// w == 0, kw = 1
 //      CHECK:  %[[B_FILTER_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xf32> to vector<3x2x4xf32>
 //      CHECK:  %[[FMA_1:.*]] = vector.fma %[[V_INPUT_1]], %[[B_FILTER_1]], %[[FMA_0]] : vector<3x2x4xf32>
 
 // Write the result back in one shot.
 //      CHECK:   vector.transfer_write %[[FMA_1]], %[[OUTPUT]][%[[C0]], %[[C0]], %[[C0]]]
+
+
+// -----
+
+func.func @conv_1d_nwc_wcf_mixed_type_memref(%input: memref<1x2x3xf16>, %filter: memref<1x3x2xf16>, %output: memref<1x2x2xf32>) {
+  linalg.conv_1d_nwc_wcf
+  {dilations = dense<1> : vector<1xi64>, strides = dense<1> : vector<1xi64>}
+   ins(%input, %filter : memref<1x2x3xf16>, memref<1x3x2xf16>)
+   outs(%output : memref<1x2x2xf32>)
+  return
+}
+
+//       CHECK: func @conv_1d_nwc_wcf_mixed_type_memref
+//  CHECK-SAME:   (%[[INPUT:[0-9a-z]+]]: memref<1x2x3xf16>, %[[FILTER:[0-9a-z]+]]: memref<1x3x2xf16>, %[[OUTPUT:[0-9a-z]+]]: memref<1x2x2xf32>)
+
+//   CHECK-DAG:   %[[C0:.+]] = arith.constant 0 : index
+//   CHECK-DAG:   %[[F0:.+]] = arith.constant 0.000000e+00 : f32
+
+/// Read the whole data in one shot.
+//      CHECK:   %[[V_INPUT_R:.+]] = vector.transfer_read %[[INPUT]][%[[C0]], %[[C0]], %[[C0]]]
+//      CHECK:   %[[V_FILTER_R:.+]] = vector.transfer_read %[[FILTER]][%[[C0]], %[[C0]], %[[C0]]]
+//      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]]]