[mlir][tosa] Update value to values for ConstOp and ConstShapeOp #129943

Jerry-Ge · 2025-03-05T22:01:42Z

Updated the dialect to match TOSA v1.0 specification for ConstOp and ConstShapeOp (https://www.mlplatform.org/tosa/tosa_spec.html#_const).

Also updated lit tests

llvmbot · 2025-03-05T22:02:20Z

@llvm/pr-subscribers-mlir-tosa

@llvm/pr-subscribers-mlir

Author: Jerry-Ge (Jerry-Ge)

Changes

Updated the dialect to match TOSA v1.0 specification (https://www.mlplatform.org/tosa/tosa_spec.html#_const) and the lit tests.

Patch is 264.83 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/129943.diff

25 Files Affected:

(modified) mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td (+3-3)
(modified) mlir/lib/Conversion/TosaToArith/TosaToArith.cpp (+1-1)
(modified) mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp (+1-1)
(modified) mlir/lib/Dialect/Tosa/IR/TosaOps.cpp (+2-2)
(modified) mlir/lib/Dialect/Tosa/Transforms/TosaFolders.cpp (+1-1)
(modified) mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp (+1-1)
(modified) mlir/test/Conversion/TosaToArith/tosa-to-arith.mlir (+1-1)
(modified) mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir (+44-44)
(modified) mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir (+4-4)
(modified) mlir/test/Conversion/TosaToSCF/tosa-to-scf.mlir (+2-2)
(modified) mlir/test/Conversion/TosaToTensor/tosa-to-tensor.mlir (+5-5)
(modified) mlir/test/Dialect/Tosa/availability.mlir (+15-15)
(modified) mlir/test/Dialect/Tosa/canonicalize.mlir (+70-70)
(modified) mlir/test/Dialect/Tosa/constant-op-fold.mlir (+211-211)
(modified) mlir/test/Dialect/Tosa/constant-reciprocal-fold.mlir (+11-11)
(modified) mlir/test/Dialect/Tosa/constant_folding.mlir (+2-2)
(modified) mlir/test/Dialect/Tosa/invalid.mlir (+43-43)
(modified) mlir/test/Dialect/Tosa/invalid_extension.mlir (+2-2)
(modified) mlir/test/Dialect/Tosa/level_check.mlir (+23-23)
(modified) mlir/test/Dialect/Tosa/ops.mlir (+27-27)
(modified) mlir/test/Dialect/Tosa/quant-test.mlir (+2-2)
(modified) mlir/test/Dialect/Tosa/tosa-decompose-depthwise.mlir (+11-11)
(modified) mlir/test/Dialect/Tosa/tosa-decompose-transpose-conv.mlir (+23-23)
(modified) mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir (+22-22)
(modified) mlir/test/Dialect/Tosa/tosa-reduce-transposes.mlir (+36-36)

diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index ff83e4072b6c0..d46e992dc36cb 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -2338,7 +2338,7 @@ def Tosa_RescaleOp: Tosa_Op<"rescale", [Pure,
 // Operator: const
 //===----------------------------------------------------------------------===//
 def Tosa_ConstOp : Tosa_Op<"const", [ConstantLike, Pure,
-                                     AllShapesMatch<["value", "output"]>,
+                                     AllShapesMatch<["values", "output"]>,
                                      FirstAttrDerivedResultType]> {
   let summary = "Constant op.";
 
@@ -2350,12 +2350,12 @@ def Tosa_ConstOp : Tosa_Op<"const", [ConstantLike, Pure,
 
     ```mlir
     // Generic form
-    %out = "tosa.const"() {value = dense<0> : tensor<2x3xi32>} : () -> tensor<2x3xi32>
+    %out = "tosa.const"() {values = dense<0> : tensor<2x3xi32>} : () -> tensor<2x3xi32>
     ```
   }];
 
   let arguments = (ins
-    ElementsAttr:$value
+    ElementsAttr:$values
   );
 
   let results = (outs
diff --git a/mlir/lib/Conversion/TosaToArith/TosaToArith.cpp b/mlir/lib/Conversion/TosaToArith/TosaToArith.cpp
index 593dbaa6c6545..5c84b4063da2e 100644
--- a/mlir/lib/Conversion/TosaToArith/TosaToArith.cpp
+++ b/mlir/lib/Conversion/TosaToArith/TosaToArith.cpp
@@ -28,7 +28,7 @@ class ConstOpConverter : public OpRewritePattern<tosa::ConstOp> {
 
   LogicalResult matchAndRewrite(tosa::ConstOp op,
                                 PatternRewriter &rewriter) const final {
-    rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, op.getValue());
+    rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, op.getValues());
     return success();
   }
 };
diff --git a/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp b/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
index 2c0376134b599..ad8dbe312d6cb 100644
--- a/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
+++ b/mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
@@ -882,7 +882,7 @@ OpFoldResult CastOp::fold(FoldAdaptor adaptor) {
   return {};
 }
 
-OpFoldResult ConstOp::fold(FoldAdaptor adaptor) { return getValueAttr(); }
+OpFoldResult ConstOp::fold(FoldAdaptor adaptor) { return getValuesAttr(); }
 
 OpFoldResult ConstShapeOp::fold(FoldAdaptor adaptor) { return getValueAttr(); }
 
diff --git a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
index 11d25e657f904..4ad47173348e1 100644
--- a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
+++ b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
@@ -352,7 +352,7 @@ static LogicalResult verifyConvOp(T op) {
 
 LogicalResult tosa::ConstOp::verify() {
 
-  auto attrType = llvm::dyn_cast<TensorType>(getValueAttr().getType());
+  auto attrType = llvm::dyn_cast<TensorType>(getValuesAttr().getType());
   auto outputType = llvm::dyn_cast<TensorType>(getOutput().getType());
 
   if (!attrType || !outputType) {
@@ -3037,7 +3037,7 @@ LogicalResult tosa::ConstShapeOp::verify() {
   auto valuesRank = getValue().getType().getRank();
   if (valuesRank != 1)
     return emitOpError("expect elements in attribute value with rank 1");
-  // check that number of elements in value attr equal to rank of result shape
+  // check that number of elements in values attr equal to rank of result shape
   auto count = getValue().getNumElements();
   auto rank = (cast<tosa::shapeType>(getResult().getType())).getRank();
   if (!(count == rank || (count == 1 && rank == 0))) {
diff --git a/mlir/lib/Dialect/Tosa/Transforms/TosaFolders.cpp b/mlir/lib/Dialect/Tosa/Transforms/TosaFolders.cpp
index 1f5a906153cd9..9a45d83c61737 100644
--- a/mlir/lib/Dialect/Tosa/Transforms/TosaFolders.cpp
+++ b/mlir/lib/Dialect/Tosa/Transforms/TosaFolders.cpp
@@ -363,7 +363,7 @@ struct ReduceConstantOptimization : public OpRewritePattern<OperationType> {
       return rewriter.notifyMatchFailure(op, "result type shape is not static");
 
     auto reductionAxis = op.getAxis();
-    const auto denseElementsAttr = constOp.getValue();
+    const auto denseElementsAttr = constOp.getValues();
     const auto shapedOldElementsValues =
         cast<ShapedType>(denseElementsAttr.getType());
 
diff --git a/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp b/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
index 36f511fa616d8..585e8b1d7f2cd 100644
--- a/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
+++ b/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
@@ -418,7 +418,7 @@ std::optional<Value> TosaReduceTransposes::buildMappedToValue(
 std::optional<Value> TosaReduceTransposes::buildMappedToValue(
     ConstOp constOp, const DenseMap<Value, Value> &valuesMap,
     IRRewriter &rewriter, ArrayRef<int32_t> hoistedPerms) {
-  auto denseAttr = llvm::dyn_cast<DenseElementsAttr>(constOp.getValue());
+  auto denseAttr = llvm::dyn_cast<DenseElementsAttr>(constOp.getValues());
   if (!denseAttr)
     return std::nullopt;
   auto maybeNewDenseAttr = transposeDenseAttribute(denseAttr, hoistedPerms);
diff --git a/mlir/test/Conversion/TosaToArith/tosa-to-arith.mlir b/mlir/test/Conversion/TosaToArith/tosa-to-arith.mlir
index c4f82d53af982..14f811727c456 100644
--- a/mlir/test/Conversion/TosaToArith/tosa-to-arith.mlir
+++ b/mlir/test/Conversion/TosaToArith/tosa-to-arith.mlir
@@ -4,7 +4,7 @@
 // CHECK-LABEL: func @const_test
 func.func @const_test() -> (tensor<i32>) {
   // CHECK: [[C3:%.+]] = arith.constant dense<3> : tensor<i32>
-  %result = "tosa.const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
+  %result = "tosa.const"() {values = dense<3> : tensor<i32>} : () -> tensor<i32>
 
   // CHECK: return [[C3]]
   return %result : tensor<i32>
diff --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
index d4afc468eeea4..5bb4a3bddb51b 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
@@ -290,8 +290,8 @@ func.func @avg_pool_f32(%arg0: tensor<1x6x34x62xf32>) -> (tensor<1x5x33x62xf32>)
   // CHECK:   %[[FLT:.+]] = arith.sitofp %[[CAST]]
   // CHECK:   %[[DIV:.+]] = arith.divf %[[IN]], %[[FLT]]
   // CHECK:   linalg.yield %[[DIV]]
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %output_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %output_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = f32, pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<1x6x34x62xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x5x33x62xf32>
   return %0 : tensor<1x5x33x62xf32>
 }
@@ -377,8 +377,8 @@ func.func @avg_pool_f16_f32acc(%arg0: tensor<1x6x34x62xf16>) -> (tensor<1x5x33x6
   // CHECK:   %[[DIV:.+]] = arith.divf %[[IN]], %[[FLT]]
   // CHECK:   %[[TRUNC:.+]] = arith.truncf %[[DIV]]
   // CHECK:   linalg.yield %[[TRUNC]]
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf16>}> : () -> tensor<1xf16>
-  %output_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf16>}> : () -> tensor<1xf16>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf16>}> : () -> tensor<1xf16>
+  %output_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf16>}> : () -> tensor<1xf16>
   %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = f32, pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<1x6x34x62xf16>, tensor<1xf16>, tensor<1xf16>) -> tensor<1x5x33x62xf16>
   return %0 : tensor<1x5x33x62xf16>
 }
@@ -420,8 +420,8 @@ func.func @avg_pool_i8(%arg0: tensor<1x6x34x62xi8>) -> (tensor<1x5x33x62xi8>) {
   // CHECK: %[[CLAMP:.+]] = arith.minsi %[[CMAX]], %[[LOW]]
   // CHECK: %[[TRUNC:.+]] = arith.trunci %[[CLAMP]]
   // CHECK: linalg.yield %[[TRUNC]]
-  %input_zp = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
-  %output_zp = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %input_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %output_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
   %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = i32, pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<1x6x34x62xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x5x33x62xi8>
   return %0 : tensor<1x5x33x62xi8>
 }
@@ -445,8 +445,8 @@ func.func @avg_pool_dyn(%arg0: tensor<?x6x34x62xf32>) -> (tensor<?x5x33x62xf32>)
   // CHECK-SAME: outs(%[[FILL]] : tensor<?x5x33x62xf32>) -> tensor<?x5x33x62xf32>
   // CHECK: %[[EMPTY:.+]] = tensor.empty(%[[BATCH]]) : tensor<?x5x33x62xf32>
   // CHECK: %[[GENERIC:.+]] = linalg.generic
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %output_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %output_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.avg_pool2d %arg0, %input_zp, %output_zp {acc_type = f32, pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<?x6x34x62xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x5x33x62xf32>
   return %0 : tensor<?x5x33x62xf32>
 }
@@ -460,8 +460,8 @@ func.func @avg_pool_dyn(%arg0: tensor<?x6x34x62xf32>) -> (tensor<?x5x33x62xf32>)
 func.func @conv2d_scalar_bias_f32(%input: tensor<1x49x42x27xf32>, %weights: tensor<28x3x3x27xf32>, %bias: tensor<1xf32>) -> () {
   // CHECK: %[[INIT:.+]] = tensor.empty() : tensor<1x45x40x28xf32>
   // CHECK: %[[BROADCAST:.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2 : tensor<1xf32>) outs(%[[INIT]] : tensor<1x45x40x28xf32>) {
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 2, 1>} : (tensor<1x49x42x27xf32>, tensor<28x3x3x27xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x45x40x28xf32>
   return
 }
@@ -482,8 +482,8 @@ func.func @conv2d_i8(%input: tensor<1x49x42x27xi8>, %weights: tensor<28x1x1x27xi
   // CHECK: linalg.conv_2d_nhwc_fhwc {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x49x42x27xi8>, tensor<28x1x1x27xi8>) outs(%[[BROADCAST]] : tensor<1x49x42x28xi32>) -> tensor<1x49x42x28xi32>
   // HWCF: linalg.conv_2d_nhwc_hwcf {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %[[TRANSPOSE]] : tensor<1x49x42x27xi8>, tensor<1x1x27x28xi8>) outs(%{{[a-zA-Z0-9_]*}} : tensor<1x49x42x28xi32>) -> tensor<1x49x42x28xi32>
 
-  %input_zp = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
-  %weight_zp = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %input_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %weight_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
   %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = i32, dilation = array<i64: 2, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<1x49x42x27xi8>, tensor<28x1x1x27xi8>, tensor<28xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x49x42x28xi32>
   return
 }
@@ -504,8 +504,8 @@ func.func @conv2d_f32(%input: tensor<1x49x42x27xf32>, %weights: tensor<28x3x3x27
   // CHECK: linalg.conv_2d_nhwc_fhwc {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x49x42x27xf32>, tensor<28x3x3x27xf32>) outs(%[[BROADCAST]] : tensor<1x45x40x28xf32>) -> tensor<1x45x40x28xf32>
 
   // HWCF: linalg.conv_2d_nhwc_hwcf {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %[[TRANSPOSE]] : tensor<1x49x42x27xf32>, tensor<3x3x27x28xf32>) outs(%{{[a-zA-Z0-9_]*}} : tensor<1x45x40x28xf32>
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 2, 1>} : (tensor<1x49x42x27xf32>, tensor<28x3x3x27xf32>, tensor<28xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x45x40x28xf32>
   return
 }
@@ -525,8 +525,8 @@ func.func @conv2d_dyn(%input: tensor<?x49x42x27xf32>, %weights: tensor<28x3x3x27
   // CHECK:   linalg.yield %[[IN]] : f32
   // CHECK: } -> tensor<?x45x40x28xf32>
   // CHECK: linalg.conv_2d_nhwc_fhwc {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<?x49x42x27xf32>, tensor<28x3x3x27xf32>) outs(%[[BROADCAST]] : tensor<?x45x40x28xf32>) -> tensor<?x45x40x28xf32>
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 2, 1>} : (tensor<?x49x42x27xf32>, tensor<28x3x3x27xf32>, tensor<28xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x45x40x28xf32>
   return
 }
@@ -583,8 +583,8 @@ func.func @conv2d_dyn_w_h(%input: tensor<1x?x?x27xf32>, %weights: tensor<28x3x3x
   // CHECK: } -> tensor<1x?x?x28xf32>
   // CHECK: linalg.conv_2d_nhwc_fhwc {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x?x?x27xf32>, tensor<28x3x3x27xf32>) outs(%[[BROADCAST]] : tensor<1x?x?x28xf32>) -> tensor<1x?x?x28xf32>
 
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 2, 1>} : (tensor<1x?x?x27xf32>, tensor<28x3x3x27xf32>, tensor<28xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x?x?x28xf32>
   return
 }
@@ -608,8 +608,8 @@ func.func @conv2d_dyn_output(%input: tensor<2x6x5x4xf32>, %weights: tensor<4x3x3
   //   linalg.yield %[[ADD]] : f32
   // } -> tensor<?x4x3x4xf32>
 
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x6x5x4xf32>, tensor<4x3x3x4xf32>, tensor<4xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x4x3x4xf32>
   return
 }
@@ -622,8 +622,8 @@ func.func @conv2d_padded_f32(%input: tensor<1x47x40x28xf32>, %weights: tensor<28
   // CHECK: tensor.pad %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
   // CHECK:   tensor.yield %[[C0]]
   // CHECK: linalg.conv_2d_nhwc_fhwc
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %0 = tosa.conv2d %input, %weights, %bias, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 1, 1, 1, 1>, stride = array<i64: 1, 1>, dilation = array<i64: 2, 1>}
     : (tensor<1x47x40x28xf32>, tensor<28x3x3x28xf32>, tensor<28xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x45x40x28xf32>
   return
@@ -637,8 +637,8 @@ func.func @conv2d_quant(%arg0 : tensor<1x12x12x1xi8>, %arg1 : tensor<1024x3x3x1x
   // CHECK: tensor.pad %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
   // CHECK:   tensor.yield %[[C22]]
   // CHECK: linalg.conv_2d_nhwc_fhwc_q
-  %input_zp = "tosa.const"() <{value = dense<-22> : tensor<1xi8>}> : () -> tensor<1xi8>
-  %weight_zp = "tosa.const"() <{value = dense<42> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %input_zp = "tosa.const"() <{values = dense<-22> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %weight_zp = "tosa.const"() <{values = dense<42> : tensor<1xi8>}> : () -> tensor<1xi8>
   %0 = tosa.conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 1, 1, 1, 1>, stride = array<i64: 1, 1>}
     : (tensor<1x12x12x1xi8>, tensor<1024x3x3x1xi8>, tensor<1024xi32>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x12x12x1024xi32>
   return
@@ -662,8 +662,8 @@ func.func @depthwise_conv(%arg0 : tensor<1x7x5x3xf32>, %arg1 : tensor<3x1x3x11xf
   // CHECK:   [[ADD:%.+]] = arith.addf %[[ARG3]], %[[ARG4]] : f32
   // CHECK:   linalg.yield [[ADD]] : f32
   // CHECK: } -> tensor<1x5x5x33xf32>
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %2 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1> } : (tensor<1x7x5x3xf32>, tensor<3x1x3x11xf32>, tensor<33xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x5x5x33xf32>
   return
 }
@@ -680,8 +680,8 @@ func.func @depthwise_conv_scalar_bias(%arg0 : tensor<1x7x5x3xf32>, %arg1 : tenso
   // CHECK:   [[ADD:%.+]] = arith.addf %[[ARG3]], %[[ARG4]] : f32
   // CHECK:   linalg.yield [[ADD]] : f32
   // CHECK: } -> tensor<1x5x5x33xf32>
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %input_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
+  %weight_zp = "tosa.const"() <{values = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
   %2 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {acc_type = f32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1> } : (tensor<1x7x5x3xf32>, tensor<3x1x3x11xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x5x5x33xf32>
   return
 }
@@ -706,8 +706,8 @@ func.func @depthwise_conv_dyn(%arg0 : tensor<?x7x5x3xf32>, %arg1 : tensor<3x1x3x
   // CHECK:   %[[ADD:.+]] = arith.addf %[[ARG3]], %[[ARG4]] : f32
   // CHECK:   linalg.yield %[[ADD]] : f32
   // CHECK: } -> tensor<?x5x5x33xf32>
-  %input_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> : () -> tensor<1xf32>
-  %weight_zp = "tosa.const"() <{value = dense<0.0> : tensor<1xf32>}> ...
[truncated]

lhutton1

LGTM! - probably a good idea to retrigger CI before merging

Signed-off-by: Jerry Ge <[email protected]> Change-Id: I59501284d8612038f819002b250ed4d86a0b05cd

Signed-off-by: Jerry Ge <[email protected]> Change-Id: Ia31f233f4e051eb7a565c26f496a02cb17cf9828

…m#129943) Updated the dialect to match TOSA v1.0 specification for ConstOp and ConstShapeOp (https://www.mlplatform.org/tosa/tosa_spec.html#_const). Also updated lit tests --------- Signed-off-by: Jerry Ge <[email protected]>

Jerry-Ge requested review from lhutton1 and GeorgeARM March 5, 2025 22:01

llvmbot added mlir:linalg mlir mlir:tosa labels Mar 5, 2025

Jerry-Ge force-pushed the values branch from 26f1dd8 to 4226de9 Compare March 5, 2025 22:27

Jerry-Ge changed the title ~~[mlir][tosa] Update value to values for ConstOp~~ [mlir][tosa] Update value to values for ConstOp and ConstShapeOp Mar 5, 2025

GeorgeARM approved these changes Mar 6, 2025

View reviewed changes

lhutton1 approved these changes Mar 6, 2025

View reviewed changes

Jerry-Ge added 2 commits March 6, 2025 08:59

[mlir][tosa] Update value to values for ConstOp

043a747

Signed-off-by: Jerry Ge <[email protected]> Change-Id: I59501284d8612038f819002b250ed4d86a0b05cd

[mlir][tosa] Update value to values for ConstShapeOp

e69f6db

Signed-off-by: Jerry Ge <[email protected]> Change-Id: Ia31f233f4e051eb7a565c26f496a02cb17cf9828

Jerry-Ge force-pushed the values branch from 0ad4dd9 to e69f6db Compare March 6, 2025 17:01

Jerry-Ge merged commit 472c2e3 into llvm:main Mar 6, 2025
11 checks passed

RoboTux mentioned this pull request Mar 7, 2025

[mlir][TOSA] Fix linalg lowering of depthwise conv2d (#130282) #130293

Merged

Jerry-Ge deleted the values branch March 20, 2025 21:40

wonjeon mentioned this pull request Mar 24, 2025

[mlir][tosa] Update Tensorflow to match TOSA v1.0 specification (part 4) tensorflow/tensorflow#89907

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[mlir][tosa] Update value to values for ConstOp and ConstShapeOp #129943

[mlir][tosa] Update value to values for ConstOp and ConstShapeOp #129943

Uh oh!

Jerry-Ge commented Mar 5, 2025 •

edited

Loading

Uh oh!

llvmbot commented Mar 5, 2025 •

edited

Loading

Uh oh!

lhutton1 left a comment

Uh oh!

Uh oh!

Uh oh!

[mlir][tosa] Update value to values for ConstOp and ConstShapeOp #129943

[mlir][tosa] Update value to values for ConstOp and ConstShapeOp #129943

Uh oh!

Conversation

Jerry-Ge commented Mar 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Mar 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lhutton1 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Jerry-Ge commented Mar 5, 2025 •

edited

Loading

llvmbot commented Mar 5, 2025 •

edited

Loading