[mlir][tosa] Quantized Conv2DOp lowering to linalg added.

[rsuderman]

Summary

Includes a version of a quantized conv2D operations with a lowering from TOSA
to linalg with corresponding test. We keep the quantized and quantized variants
as separate named ops to avoid the additional operations for non-quantized
convolutions.

[hanchung]

We should add a roundtrip test when adding a new op. You can add it to mlir/test/Dialect/Linalg/named-ops.mlir

mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml

1550

Delete the blank line?

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

944

I think the order is strideAttr and then dilationAttr. This is an issue in TC ops. The order attributes of TC ops is alphabetical, but we really should make it be the order of declaration. Hopefully, it is fixed in yaml ops.

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

1187–1188

We can use CHECK-DAG to prevent such change in the future.

[rsuderman]

Updated review comments.

[hanchung]

LGTM if adding a roundtrip test to mlir/test/Dialect/Linalg/named-ops.mlir.

[rsuderman]

Added roundtrip test for conv_2d quantized

mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml

value: !ScalarExpression
             - !ScalarExpression
               symbolic_cast:
                 type_var: U
                 operands:
                 - !ScalarExpression
                   scalar_arg: K
 --- !LinalgOpConfig
 metadata: !LinalgOpMetadata
+  name: conv_2d_input_nhwc_filter_ohwi_poly_q
+  cpp_class_name: Conv2DInputNhwcFilterOhwiPolyQOp
+  doc: |-
+    Performs a 2-D quantized convolution.
+
+    Numeric casting is performed on the operands to the inner multiply, promoting
+    them to the same data type as the accumulator/output. Includes zero point
+    adjustment for quantization.
+structured_op: !LinalgStructuredOpConfig
+  args:
+  - !LinalgOperandDefConfig
+    name: I
+    usage: InputOperand
+    type_var: T1
+    shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12]
+      -> (s0, s1, s2, s3)>
+  - !LinalgOperandDefConfig
+    name: K
+    usage: InputOperand
+    type_var: T2
+    shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12]
+      -> (s4, s5, s6, s3)>
+  - !LinalgOperandDefConfig
+    name: IZp
+    usage: InputOperand
+    type_var: I32
+  - !LinalgOperandDefConfig
+    name: KZp
+    usage: InputOperand
+    type_var: I32
+  - !LinalgOperandDefConfig
+    name: O
+    usage: OutputOperand
+    type_var: U
+    shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12]
+      -> (s0, s7, s8, s4)>
+  - !LinalgOperandDefConfig
+    name: strides
+    usage: IndexAttribute
+    type_var: I64
+    attribute_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
+      s12] -> (s9, s10)>
+  - !LinalgOperandDefConfig
+    name: dilations
+    usage: IndexAttribute
+    type_var: I64
+    attribute_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
+      s12] -> (s11, s12)>
+  indexing_maps: !LinalgIndexingMapsConfig
+    static_indexing_maps:
+    - affine_map<(d0, d1, d2, d3, d4, d5, d6)[s0, s1, s2, s3, s4, s5, s6, s7, s8,
+      s9, s10, s11, s12] -> (d0, d1 * s9 + d3 * s11, d2 * s10 + d4 * s12, d6)>
+    - affine_map<(d0, d1, d2, d3, d4, d5, d6)[s0, s1, s2, s3, s4, s5, s6, s7, s8,
+      s9, s10, s11, s12] -> (d5, d3, d4, d6)>
+    - affine_map<(d0, d1, d2, d3, d4, d5, d6)[s0, s1, s2, s3, s4, s5, s6, s7, s8,
+      s9, s10, s11, s12] -> ()>
+    - affine_map<(d0, d1, d2, d3, d4, d5, d6)[s0, s1, s2, s3, s4, s5, s6, s7, s8,
+      s9, s10, s11, s12] -> ()>
+    - affine_map<(d0, d1, d2, d3, d4, d5, d6)[s0, s1, s2, s3, s4, s5, s6, s7, s8,
+      s9, s10, s11, s12] -> (d0, d1, d2, d5)>
+  iterator_types:
+  - parallel
+  - parallel
+  - parallel
+  - reduction
+  - reduction
+  - parallel
+  - reduction
+  assignments:
+  - !ScalarAssign
+    arg: O
+    value: !ScalarExpression
+      scalar_apply:
+        fn_name: add
+        operands:
+        - !ScalarExpression
+          scalar_arg: O
+        - !ScalarExpression
+          scalar_apply:
+            fn_name: mul
+            operands:
+            - !ScalarExpression
+              scalar_apply:
+                fn_name: sub
+                operands:
+                - !ScalarExpression
+                  symbolic_cast:
+                    type_var: U
+                    operands:
+                    - !ScalarExpression
+                      scalar_arg: I
+                - !ScalarExpression
+                  symbolic_cast:
+                    type_var: U
+                    operands:
+                    - !ScalarExpression
+                      scalar_arg: IZp
+            - !ScalarExpression
+              scalar_apply:
+                fn_name: sub
+                operands:
+                - !ScalarExpression
+                  symbolic_cast:
+                    type_var: U
+                    operands:
+                    - !ScalarExpression
+                      scalar_arg: K
+                - !ScalarExpression
+                  symbolic_cast:
+                    type_var: U
+                    operands:
+                    - !ScalarExpression
+                      scalar_arg: KZp
+--- !LinalgOpConfig
+metadata: !LinalgOpMetadata
   name: depthwise_conv_2d_input_nhwc_filter_hwc_poly
   cpp_class_name: DepthwiseConv2DInputNhwcFilterHwcPolyOp
   doc: |-
     Performs depth-wise 2-D convolution.
 
     Numeric casting is performed on the operands to the inner multiply, promoting
     them to the same data type as the accumulator/output.
 structured_op: !LinalgStructuredOpConfig
value: !ScalarExpression
                 fn_name: exp
                 operands:
                 - !ScalarExpression
                   symbolic_cast:
                     type_var: U
                     operands:
                     - !ScalarExpression
                       scalar_arg: I
+

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

convolutionMatchAndRewriterHelper(Operation *op,
   Value bias = op->getOperand(2);
 
   ShapedType inputTy = input.getType().cast<ShapedType>();
   ShapedType weightTy = weight.getType().cast<ShapedType>();
   ShapedType biasTy = bias.getType().cast<ShapedType>();
   ShapedType resultTy = op->getResult(0).getType().cast<ShapedType>();
 
   Type inputETy = inputTy.getElementType();
-  Type weightETy = weightTy.getElementType();
-  Type biasETy = biasTy.getElementType();
   Type resultETy = resultTy.getElementType();
 
   auto padAttr = op->getAttr("pad").cast<ArrayAttr>();
   auto strideTosaAttr = op->getAttr("stride").cast<ArrayAttr>();
   auto dilationTosaAttr = op->getAttr("dilation").cast<ArrayAttr>();
 
+  bool isQuantized = op->hasAttr("quantization_info");
+  IntegerAttr iZp;
+  IntegerAttr kZp;
+  if (isQuantized) {
+    auto quantizationInfo =
+        op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
+    iZp = rewriter.getI32IntegerAttr(
+        quantizationInfo.input_zp().getValue().getSExtValue());
+    kZp = rewriter.getI32IntegerAttr(
+        quantizationInfo.weight_zp().getValue().getSExtValue());
+  }
+
   if (!inputTy.hasStaticShape() || !weightTy.hasStaticShape() ||
       !biasTy.hasStaticShape() || !resultTy.hasStaticShape())
     return rewriter.notifyMatchFailure(op,
                                        "tosa.conv ops require static shapes");
 
   auto weightShape = weightTy.getShape();
   auto resultShape = resultTy.getShape();
 
-  // TODO(suderman): Support other types.
-  if (!inputETy.isF32() || !weightETy.isF32() || !biasETy.isF32() ||
-      !resultETy.isF32())
-    return failure();
-
   // Apply padding as necessary.
   Attribute zeroAttr = rewriter.getZeroAttr(inputETy);
   llvm::SmallVector<int64_t> pad;
   pad.resize(2, 0);
   getValuesFromIntArrayAttribute(padAttr, pad);
   pad.resize(pad.size() + 2, 0);
 
   input = applyPad(loc, input, pad, zeroAttr, rewriter);
convolutionMatchAndRewriterHelper(Operation *op,
   getValuesFromIntArrayAttribute(dilationTosaAttr, dilation);
 
   // Create the convolution op.
   auto strideAttr = DenseIntElementsAttr::get(
       RankedTensorType::get({2}, rewriter.getI64Type()), stride);
   auto dilationAttr = DenseIntElementsAttr::get(
       RankedTensorType::get({2}, rewriter.getI64Type()), dilation);
 
-  if (isa<tosa::Conv2DOp>(op)) {
+  if (isa<tosa::Conv2DOp>(op) && !isQuantized) {
     rewriter.replaceOpWithNewOp<linalg::Conv2DInputNhwcFilterOhwiPolyOp>(
         op, resultTy, ValueRange{input, weight}, ValueRange{biasBroadcast},
         strideAttr, dilationAttr);
     return success();
   }
 
-  if (isa<tosa::DepthwiseConv2DOp>(op)) {
+  if (isa<tosa::Conv2DOp>(op) && isQuantized) {
+    auto iZpVal = rewriter.create<ConstantOp>(loc, iZp);
+    auto kZpVal = rewriter.create<ConstantOp>(loc, kZp);
+    rewriter.replaceOpWithNewOp<linalg::Conv2DInputNhwcFilterOhwiPolyQOp>(
+        op, resultTy, ValueRange{input, weight, iZpVal, kZpVal},
+        ValueRange{biasBroadcast}, dilationAttr, strideAttr);
+    return success();
+  }
+
+  if (isa<tosa::DepthwiseConv2DOp>(op) && !isQuantized) {
     ShapedType linalgConvTy =
         RankedTensorType::get({resultShape[0], resultShape[1], resultShape[2],
                                weightShape[2], weightShape[3]},
                               resultETy);
 
     Value biasReshape =
         rewriter.create<tosa::ReshapeOp>(loc, linalgConvTy, biasBroadcast);
     Value conv = rewriter

mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py

def conv_2d_input_nhwc_filter_ohwi_poly(
   domain(D.n, D.oh, D.ow, D.kh, D.kw, D.oc, D.ic)
   O[D.n, D.oh, D.ow, D.oc] += cast(
       U, I[D.n,
            D.oh * S.SH + D.kh * S.DH,
            D.ow * S.SW + D.kw * S.DW,
            D.ic]) * cast(U, K[D.oc, D.kh, D.kw, D.ic])
 
 @linalg_structured_op
+def conv_2d_input_nhwc_filter_ohwi_poly_q(
+    I=TensorDef(T1, S.N, S.IH, S.IW, S.IC),
+    K=TensorDef(T2, S.OC, S.KH, S.KW, S.IC),
+    IZp=ScalarDef(I32),
+    KZp=ScalarDef(I32),
+    O=TensorDef(U, S.N, S.OH, S.OW, S.OC, output=True),
+    strides=AttributeDef(S.SH, S.SW),
+    dilations=AttributeDef(S.DH, S.DW)):
+  """Performs a 2-D quantized convolution.
+
+  Numeric casting is performed on the operands to the inner multiply, promoting
+  them to the same data type as the accumulator/output. Includes zero point
+  adjustment for quantization.
+  """
+  domain(D.n, D.oh, D.ow, D.kh, D.kw, D.oc, D.ic)
+  O[D.n, D.oh, D.ow, D.oc] += ((cast(
+      U, I[D.n,
+           D.oh * S.SH + D.kh * S.DH,
+           D.ow * S.SW + D.kw * S.DW,
+           D.ic]) - cast(U, IZp)) *
+           (cast(U, K[D.oc, D.kh, D.kw, D.ic]) - cast(U, KZp)))
+
+
+@linalg_structured_op
 def depthwise_conv_2d_input_nhwc_filter_hwc_poly(
     I=TensorDef(T1, S.N, S.IH, S.IW, S.C),
     K=TensorDef(T2, S.KH, S.KW, S.C),
     O=TensorDef(U, S.N, S.OH, S.OW, S.C, output=True),
     strides=AttributeDef(S.SH, S.SW),
     dilations=AttributeDef(S.DH, S.DW)):
   """Performs depth-wise 2-D convolution.
 

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

func @avg_pool(%arg0: tensor<1x6x34x62xf32>) -> (tensor<1x5x33x62xf32>) {
   // CHECK:   [[RESULT:%.+]] = divf %arg1, [[CF]]
   // CHECK:   linalg.yield [[RESULT]]
   %0 = "tosa.avg_pool2d"(%arg0) {pad = [1, 1, 1, 1], kernel = [4, 4], stride = [1, 1]} : (tensor<1x6x34x62xf32>)  -> (tensor<1x5x33x62xf32>)
   return %0 : tensor<1x5x33x62xf32>
 }
 
 // -----
 
-// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
 
+// CHECK-LABEL: @conv2d_f32
 func @conv2d_f32(%input: tensor<1x49x42x27xf32>, %weights: tensor<28x3x3x27xf32>, %bias: tensor<28xf32>) -> () {
   // CHECK: %[[INIT:.+]] = linalg.init_tensor [1, 45, 40, 28]
-  // CHECK: %[[BROADCAST:.+]] = linalg.generic {indexing_maps = [#[[$MAP2]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2 : tensor<28xf32>) outs(%[[INIT]] : tensor<1x45x40x28xf32>)
+  // CHECK: %[[BROADCAST:.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2 : tensor<28xf32>) outs(%[[INIT]] : tensor<1x45x40x28xf32>)
   // CHECK: linalg.conv_2d_input_nhwc_filter_ohwi_poly {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x49x42x27xf32>, tensor<28x3x3x27xf32>) outs(%[[BROADCAST]] : tensor<1x45x40x28xf32>)
   %0 = "tosa.conv2d"(%input, %weights, %bias) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [2, 1]} : (tensor<1x49x42x27xf32>, tensor<28x3x3x27xf32>, tensor<28xf32>)  -> (tensor<1x45x40x28xf32>)
   return
 }
 
+// -----
+
+// CHECK-LABEL: @conv2d_padded_f32
 func @conv2d_padded_f32(%input: tensor<1x47x40x28xf32>, %weights: tensor<28x3x3x28xf32>, %bias: tensor<28xf32>) -> () {
   // CHECK: linalg.pad_tensor %arg0
   // CHECK: linalg.conv_2d_input_nhwc_filter_ohwi_poly
   %0 = "tosa.conv2d"(%input, %weights, %bias) {pad = [1, 1, 1, 1], stride = [1, 1], dilation = [2, 1]} : (tensor<1x47x40x28xf32>, tensor<28x3x3x28xf32>, tensor<28xf32>)  -> (tensor<1x45x40x28xf32>)
   return
 }
 
 // -----
 
+// CHECK: #map0 = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @conv2d_quant
+func @conv2d_quant(%arg0 : tensor<1x12x12x1xi8>, %arg1 : tensor<1024x3x3x1xi8>, %arg2 : tensor<1024xi32>) -> () {
+  // CHECK:   %[[INIT:.+]] = linalg.init_tensor [1, 10, 10, 1024]
+  // CHECK:   %[[BROADCAST:.+]] = linalg.generic {indexing_maps = [#map0, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2 : tensor<1024xi32>) outs(%[[INIT]] : tensor<1x10x10x1024xi32>)
+  // CHECK:   ^bb0(%arg3: i32, %arg4: i32): 
+  // CHECK:     linalg.yield %arg3 : i32
+  // CHECK:   %[[C128:.+]] = constant -128 
+  // CHECK:   %[[C42:.+]] = constant 42 
+  // CHECK:   linalg.conv_2d_input_nhwc_filter_ohwi_poly_q {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1, %[[C128]], %[[C42]] : tensor<1x12x12x1xi8>, tensor<1024x3x3x1xi8>, i32, i32) outs(%1 : tensor<1x10x10x1024xi32>) 
+  %0 = "tosa.conv2d"(%arg0, %arg1, %arg2) {dilation = [1, 1], pad = [0, 0, 0, 0], quantization_info = {input_zp = -128 : i32, weight_zp = 42 : i32}, stride = [1, 1]} : (tensor<1x12x12x1xi8>, tensor<1024x3x3x1xi8>, tensor<1024xi32>) -> tensor<1x10x10x1024xi32>
+  return
+}
+
+// -----
+
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
 // CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
 
 // CHECK-LABEL: @depthwise_conv
 func @depthwise_conv(%arg0 : tensor<1x7x5x3xf32>, %arg1 : tensor<3x1x3x11xf32>, %arg2 : tensor<33xf32>) -> () {
   // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 5, 33]
   // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2 : tensor<33xf32>) outs([[INIT]] : tensor<1x5x5x33xf32>) {
   // CHECK: ^bb0(%arg3: f32, %arg4: f32):  // no predecessors