[mlir][TOSA] Fix linalg lowering of depthwise conv2d

Current lowering for tosa.depthwise_conv2d assumes if both zero points
are zero then it's a floating-point operation by hardcoding the use of a
arith.addf in the lowered code. Fix code to check for the element type
to decide what add operation to use.

Author: RoboTux

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp

@@ -477,13 +477,13 @@ class DepthwiseConvConverter
       return rewriter.notifyMatchFailure(
           op, "weight zero point must be zero for non-int8 integer types");
 
-    bool hasZp = (inputZpVal != 0) || (weightZpVal != 0);
+    bool hasNullZps = (inputZpVal == 0) && (weightZpVal == 0);
     auto weightShape = weightTy.getShape();
     auto resultShape = resultTy.getShape();
 
     // Apply padding as necessary.
     TypedAttr zeroAttr = rewriter.getZeroAttr(inputETy);
-    if (hasZp) {
+    if (!hasNullZps) {
       int64_t intMin =
           APInt::getSignedMinValue(inputETy.getIntOrFloatBitWidth())
               .getSExtValue();
@@ -536,7 +536,7 @@ class DepthwiseConvConverter
     indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultRank));
     indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultRank));
 
-    if (!hasZp) {
+    if (hasNullZps) {
       Value conv = rewriter
                        .create<linalg::DepthwiseConv2DNhwcHwcmOp>(
                            loc, linalgConvTy, ValueRange{input, weight},
@@ -556,8 +556,13 @@ class DepthwiseConvConverter
                   getNParallelLoopsAttrs(resultRank),
                   [&](OpBuilder &nestedBuilder, Location nestedLoc,
                       ValueRange args) {
-                    Value added = nestedBuilder.create<arith::AddFOp>(
-                        loc, args[0], args[1]);
+                    Value added;
+                    if (llvm::isa<FloatType>(inputETy))
+                      added = nestedBuilder.create<arith::AddFOp>(loc, args[0],
+                                                                  args[1]);
+                    else
+                      added = nestedBuilder.create<arith::AddIOp>(loc, args[0],
+                                                                  args[1]);
                     nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
                   })
               .getResult(0);

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir

@@ -812,6 +812,30 @@ func.func @depthwise_conv2d_dyn_w_h(%arg0: tensor<2x?x?x3xf32>, %arg1: tensor<3x
 
 // -----
 
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @depthwise_int_conv_zero_zp
+func.func @depthwise_int_conv_zero_zp(%arg0 : tensor<1x7x5x3xi8>, %arg1 : tensor<3x1x3x11xi8>, %arg2 : tensor<33xi32>) -> () {
+  // CHECK: [[INIT:%.+]] = tensor.empty()
+  // CHECK: [[CST0:%.+]] = arith.constant 0
+  // CHECK: [[FILL:%.+]] = linalg.fill ins([[CST0]]{{.*}}outs([[INIT]]
+  // CHECK: [[OUT:%.+]] = tensor.empty()
+  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x7x5x3xi8>, tensor<3x1x3x11xi8>) outs([[FILL]] : tensor<1x5x5x3x11xi32>)
+  // CHECK: [[COLLAPSED:%.+]] = tensor.collapse_shape [[DEPTH]] {{\[}}[0], [1], [2], [3, 4]]
+  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<33xi32>, tensor<1x5x5x33xi32>) outs([[OUT]] : tensor<1x5x5x33xi32>) {
+  // CHECK: ^bb0(%[[ARG3:[0-9a-zA-Z_]+]]: i32, %[[ARG4:[0-9a-zA-Z_]+]]: i32, %[[ARG5:[0-9a-zA-Z_]+]]: i32):
+  // CHECK:   [[ADD:%.+]] = arith.addi %[[ARG3]], %[[ARG4]] : i32
+  // CHECK:   linalg.yield [[ADD]] : i32
+  // CHECK: } -> tensor<1x5x5x33xi32>
+  %input_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %weight_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %2 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {acc_type = i32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1> } : (tensor<1x7x5x3xi8>, tensor<3x1x3x11xi8>, tensor<33xi32>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x5x5x33xi32>
+  return
+}
+
+// -----
+
 // CHECK: #[[$MAP1:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d4)>
 // CHECK: #[[$MAP2:.+]] = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>