[Mlir-commits] [mlir] [mlir][tosa] Add constant block scaled support (PR #205506)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Mon Jul 6 07:54:47 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir-tosa
Author: Luke Hutton (lhutton1)
<details>
<summary>Changes</summary>
This commit adds support for block scaled tensors.
In particular, the block scaled type has been extended with the `DenseElementTypeInterface` to allow block scaled data values to be specified in dense element attributes.
The block scaled type has also been extended to allow optional scale values to be specified by the type. For now, scale values are not expected to be propagated beyond their use in the attribute input of a constant operation. In the future, we may want to propagate these values to allow certain optimizations.
The `tosa.const` operation has also been updated in the validation pass.
Note: this PR relies on #<!-- -->203583 so also contains the contents of #<!-- -->203583
---
Patch is 32.11 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/205506.diff
9 Files Affected:
- (modified) mlir/include/mlir/Dialect/Tosa/IR/TosaComplianceData.h.inc (+19-1)
- (modified) mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td (+7-3)
- (modified) mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td (+17-3)
- (modified) mlir/lib/Dialect/Tosa/IR/TosaOps.cpp (+136-13)
- (modified) mlir/test/Dialect/Tosa/availability.mlir (+1-1)
- (modified) mlir/test/Dialect/Tosa/ops.mlir (+72)
- (modified) mlir/test/Dialect/Tosa/tosa-validation-version-1p0-invalid.mlir (+18)
- (modified) mlir/test/Dialect/Tosa/tosa-validation-version-1p1-valid.mlir (+13)
- (modified) mlir/test/Dialect/Tosa/verifier.mlir (+69)
``````````diff
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaComplianceData.h.inc b/mlir/include/mlir/Dialect/Tosa/IR/TosaComplianceData.h.inc
index ef644845ed937..291097a7f7514 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaComplianceData.h.inc
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaComplianceData.h.inc
@@ -1305,7 +1305,25 @@ extensionComplianceMap = {
{{fp6e3m2T}, SpecificationVersion::V_1_1_DRAFT},
{{fp6e2m3T}, SpecificationVersion::V_1_1_DRAFT},
{{fp4e2m1T}, SpecificationVersion::V_1_1_DRAFT},
- {{mxint8T}, SpecificationVersion::V_1_1_DRAFT}}}}},
+ {{mxint8T}, SpecificationVersion::V_1_1_DRAFT}}},
+ {{Extension::mx_common, Extension::mx_fp8e4m3},
+ {{{bs32_fp8ue8m0_fp8e4m3T}, SpecificationVersion::V_1_1_DRAFT}},
+ allOf},
+ {{Extension::mx_common, Extension::mx_fp8e5m2},
+ {{{bs32_fp8ue8m0_fp8e5m2T}, SpecificationVersion::V_1_1_DRAFT}},
+ allOf},
+ {{Extension::mx_common, Extension::mx_fp6e3m2},
+ {{{bs32_fp8ue8m0_fp6e3m2T}, SpecificationVersion::V_1_1_DRAFT}},
+ allOf},
+ {{Extension::mx_common, Extension::mx_fp6e2m3},
+ {{{bs32_fp8ue8m0_fp6e2m3T}, SpecificationVersion::V_1_1_DRAFT}},
+ allOf},
+ {{Extension::mx_common, Extension::mx_fp4e2m1},
+ {{{bs32_fp8ue8m0_fp4e2m1T}, SpecificationVersion::V_1_1_DRAFT}},
+ allOf},
+ {{Extension::mx_common, Extension::mx_int8},
+ {{{bs32_fp8ue8m0_mxint8T}, SpecificationVersion::V_1_1_DRAFT}},
+ allOf}}},
{"tosa.identity",
{{{Extension::int4}, {{{i4T, i4T}, SpecificationVersion::V_1_0}}},
{{Extension::int16}, {{{i48T, i48T}, SpecificationVersion::V_1_0}}},
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index 3518a62800963..5e5760e33b5f6 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -2970,9 +2970,13 @@ def Tosa_ConstOp : Tosa_Op<"const", [ConstantLike, Pure,
Tosa_Tensor:$output
);
- list<Availability> availability = [
- Profile<[Tosa_PRO_INT, Tosa_PRO_FP]>,
- Extension<[Tosa_EXT_INT4, Tosa_EXT_INT16, Tosa_EXT_FP8E4M3, Tosa_EXT_FP8E5M2, Tosa_EXT_BF16, Tosa_EXT_MXFP, Tosa_EXT_INT64]>,
+ list<Availability> availability =
+ [Profile<[Tosa_PRO_INT, Tosa_PRO_FP]>,
+ Extension<[Tosa_EXT_INT4, Tosa_EXT_INT16, Tosa_EXT_FP8E4M3,
+ Tosa_EXT_FP8E5M2, Tosa_EXT_BF16, Tosa_EXT_MXFP,
+ Tosa_EXT_INT64, Tosa_EXT_MX_COMMON, Tosa_EXT_MX_FP4E2M1,
+ Tosa_EXT_MX_FP6E2M3, Tosa_EXT_MX_FP6E3M2, Tosa_EXT_MX_FP8E4M3,
+ Tosa_EXT_MX_FP8E5M2, Tosa_EXT_MX_INT8]>,
];
let hasFolder = 1;
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
index 69f836cf1b9d3..e6dace5a8d3df 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td
@@ -109,7 +109,12 @@ def Tosa_MXFPValue
def Tosa_MXFPScale
: AnyTypeOf<[F8E8M0FNU], "micro-scaling format scale number">;
-def Tosa_BlockScaled : Tosa_Type<"BlockScaled", "block_scaled"> {
+def Tosa_BlockScaled
+ : Tosa_Type<"BlockScaled", "block_scaled",
+ [DeclareTypeInterfaceMethods<
+ DenseElementTypeInterface, ["getDenseElementBitSize",
+ "convertToAttribute",
+ "convertFromAttribute"]>]> {
let summary = "Block scaled tensor element type.";
let description = [{
@@ -121,17 +126,26 @@ def Tosa_BlockScaled : Tosa_Type<"BlockScaled", "block_scaled"> {
specifying the block size along the innermost dimension.
`scale_type` - The type of the scale value associated with each block.
`value_type` - The type of the data values in each block.
+ `scale_values` - Optional array of per-block scale values. If not provided,
+ the type is assumed to be dynamically quantized at runtime.
}];
let parameters = (ins
EnumParameter<Tosa_BlockShape>:$block_shape,
Tosa_MXFPScale:$scale_type,
- Tosa_MXFPValue:$value_type
+ Tosa_MXFPValue:$value_type,
+ OptionalArrayRefParameter<"Attribute">:$scale_values
);
let assemblyFormat =
- "`<` $block_shape```:```$scale_type```:```$value_type `>`";
+ "`<` $block_shape```:```$scale_type```:```$value_type "
+ "(`,` `{` custom<ScaleValues>($scale_values, ref($scale_type))^ `}`)? "
+ "`>`";
+
+ let extraClassDeclaration = [{
+ bool hasScaleValues() const { return !getScaleValues().empty(); }
+ }];
}
//===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
index e25774856f59d..ece1b0ffa29ba 100644
--- a/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
+++ b/mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
@@ -20,6 +20,7 @@
#include "mlir/Dialect/Tosa/Utils/ShapeUtils.h"
#include "mlir/Dialect/Utils/IndexingUtils.h"
#include "mlir/Dialect/Utils/VerificationUtils.h"
+#include "mlir/IR/BuiltinTypeInterfaces.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/DialectImplementation.h"
#include "mlir/IR/Matchers.h"
@@ -740,14 +741,55 @@ LogicalResult mlir::tosa::mxint8Type::convertFromAttribute(
// TOSA block scaling utilities.
//===----------------------------------------------------------------------===//
-LogicalResult OpTrait::tosa::verifyBlockScaledTensorType(Operation &op,
- mlir::Type type) {
+static ParseResult parseScaleValues(AsmParser &parser,
+ SmallVector<Attribute> &scaleValues,
+ Type scaleType) {
+ const auto parseScaleValue = [&]() -> ParseResult {
+ const SMLoc loc = parser.getCurrentLocation();
+
+ double floatValue;
+ if (parser.parseFloat(floatValue))
+ return failure();
+
+ if (floatValue < 0.0)
+ return parser.emitError(loc, "scale value must be non-negative, got ")
+ << floatValue;
+
+ Type attrType = scaleType;
+ if (succeeded(parser.parseOptionalColon()) && parser.parseType(attrType))
+ return failure();
+
+ if (attrType != scaleType)
+ return parser.emitError(loc, "parsed attribute type ")
+ << attrType << " does not match expected scale type " << scaleType;
+
+ scaleValues.push_back(FloatAttr::get(attrType, floatValue));
+ return success();
+ };
+
+ return parser.parseCommaSeparatedList(parseScaleValue);
+}
+
+static void printScaleValues(AsmPrinter &printer,
+ ArrayRef<Attribute> scaleValues, Type) {
+ llvm::interleaveComma(scaleValues, printer, [&](Attribute scaleValue) {
+ printer.printAttributeWithoutType(scaleValue);
+ });
+}
+
+static LogicalResult verifyBlockScaledTensorType(Operation &op, mlir::Type type,
+ bool allowScaleValues) {
const auto tensorType = llvm::cast<ShapedType>(type);
const BlockScaledType elemType =
llvm::dyn_cast<BlockScaledType>(tensorType.getElementType());
if (!elemType)
return success();
+ if (!allowScaleValues && elemType.hasScaleValues())
+ return op.emitError()
+ << "tensor type " << type
+ << " does not support scale values for this operation";
+
if (!tensorType.hasRank())
return success();
@@ -756,12 +798,24 @@ LogicalResult OpTrait::tosa::verifyBlockScaledTensorType(Operation &op,
<< "tensor type " << type
<< " does not support block scaling on scalar tensors";
- const int64_t blockedDimension = tensorType.getShape().back();
- if (ShapedType::isDynamic(blockedDimension))
- return success();
-
+ const ArrayRef<int64_t> tensorShape = tensorType.getShape();
const uint32_t blockSize =
BlockShapeAttr::getBlockShapeValue(elemType.getBlockShape());
+
+ if (allowScaleValues && elemType.hasScaleValues() &&
+ tensorType.hasStaticShape()) {
+ const size_t numBlocks =
+ llvm::accumulate(tensorShape, int64_t(1), std::multiplies<int64_t>());
+ if (elemType.getScaleValues().size() != numBlocks / blockSize)
+ return op.emitError()
+ << "tensor type " << type << " has " << numBlocks / blockSize
+ << " blocks but got " << elemType.getScaleValues().size()
+ << " scale values";
+ }
+
+ const int64_t blockedDimension = tensorShape.back();
+ if (ShapedType::isDynamic(blockedDimension))
+ return success();
if (blockedDimension % blockSize != 0)
return op.emitError()
<< "tensor type " << type
@@ -771,6 +825,43 @@ LogicalResult OpTrait::tosa::verifyBlockScaledTensorType(Operation &op,
return success();
}
+LogicalResult OpTrait::tosa::verifyBlockScaledTensorType(Operation &op,
+ mlir::Type type) {
+ return ::verifyBlockScaledTensorType(op, type, /*allowScaleValues=*/false);
+}
+
+size_t mlir::tosa::BlockScaledType::getDenseElementBitSize() const {
+ const Type valueType = getValueType();
+ if (isa<tosa::mxint8Type>(valueType))
+ return 8;
+ return valueType.getIntOrFloatBitWidth();
+}
+
+Attribute
+mlir::tosa::BlockScaledType::convertToAttribute(ArrayRef<char> rawData) const {
+ assert(rawData.size() == 1 && "expected 1 byte for block_scaled element");
+ const Type valueType = getValueType();
+ if (const auto mxint8Value = dyn_cast<tosa::mxint8Type>(valueType))
+ return mxint8Value.convertToAttribute(rawData);
+ if (!isa<FloatType>(valueType))
+ return {};
+ return mlir::detail::convertFloatTypeToAttribute(valueType, rawData);
+}
+
+LogicalResult mlir::tosa::BlockScaledType::convertFromAttribute(
+ Attribute attr, SmallVectorImpl<char> &result) const {
+ const Type valueType = getValueType();
+ if (const auto mxint8Value = dyn_cast<tosa::mxint8Type>(valueType))
+ return mxint8Value.convertFromAttribute(attr, result);
+
+ const auto floatAttr = dyn_cast<FloatAttr>(attr);
+ if (!floatAttr || floatAttr.getType() != valueType)
+ return failure();
+ const APFloat value = floatAttr.getValue();
+ return mlir::detail::convertFloatTypeFromAttribute(
+ valueType, FloatAttr::get(valueType, value), result);
+}
+
//===----------------------------------------------------------------------===//
// TOSA Operator Verifiers.
//===----------------------------------------------------------------------===//
@@ -857,7 +948,7 @@ static LogicalResult verifyConvOp(T op) {
}
LogicalResult tosa::ConstOp::verify() {
-
+ Operation &op = *getOperation();
auto attrType = llvm::dyn_cast<TensorType>(getValuesAttr().getType());
auto outputType = llvm::dyn_cast<TensorType>(getOutput().getType());
@@ -866,17 +957,49 @@ LogicalResult tosa::ConstOp::verify() {
return failure();
}
- if (auto result = llvm::dyn_cast<mlir::quant::QuantizedType>(
- outputType.getElementType())) {
- if (getStorageElementTypeFromQuantized(result) == attrType.getElementType())
+ const Type attrElemType = attrType.getElementType();
+ const Type resultElemType = outputType.getElementType();
+
+ if (auto result =
+ llvm::dyn_cast<mlir::quant::QuantizedType>(resultElemType)) {
+ if (getStorageElementTypeFromQuantized(result) == attrElemType)
return success();
}
- if (attrType.getElementType() != outputType.getElementType()) {
- emitOpError("expected same attr/result element types");
- return failure();
+ if (auto attrBlockScaledType =
+ llvm::dyn_cast<mlir::tosa::BlockScaledType>(attrElemType)) {
+ if (failed(verifyBlockScaledTensorType(op, attrType, true)) ||
+ failed(verifyBlockScaledTensorType(op, outputType, false)))
+ return failure();
+
+ if (!attrBlockScaledType.hasScaleValues())
+ return op.emitOpError(
+ "attribute block scaled type must have scale values");
+
+ const BlockScaledType resultBlockScaledType =
+ llvm::dyn_cast<mlir::tosa::BlockScaledType>(resultElemType);
+ if (!resultBlockScaledType)
+ return op.emitOpError(
+ "result type must be block scaled type if attribute is block "
+ "scaled type");
+
+ if (attrBlockScaledType.getValueType() !=
+ resultBlockScaledType.getValueType() ||
+ attrBlockScaledType.getScaleType() !=
+ resultBlockScaledType.getScaleType() ||
+ attrBlockScaledType.getBlockShape() !=
+ resultBlockScaledType.getBlockShape())
+ return op.emitOpError(
+ "expected block scaled element type to be compatible "
+ "between attr and result, got ")
+ << attrBlockScaledType << " vs. " << resultBlockScaledType;
+
+ return success();
}
+ if (attrElemType != resultElemType)
+ return emitOpError("expected same attr/result element types");
+
return success();
}
diff --git a/mlir/test/Dialect/Tosa/availability.mlir b/mlir/test/Dialect/Tosa/availability.mlir
index c358460bbc662..bc014fb9b928e 100644
--- a/mlir/test/Dialect/Tosa/availability.mlir
+++ b/mlir/test/Dialect/Tosa/availability.mlir
@@ -660,7 +660,7 @@ func.func @test_rescale(%arg0: tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439
// CHECK-LABEL: test_const
func.func @test_const(%arg0 : index) -> tensor<4xi32> {
// CHECK: profiles: [ [pro_int, pro_fp] ]
- // CHECK: extensions: [ [int4, int16, fp8e4m3, fp8e5m2, bf16, mxfp, int64] ]
+ // CHECK: extensions: [ [int4, int16, fp8e4m3, fp8e5m2, bf16, mxfp, int64, mx_common, mx_fp4e2m1, mx_fp6e2m3, mx_fp6e3m2, mx_fp8e4m3, mx_fp8e5m2, mx_int8] ]
%0 = "tosa.const"() {values = dense<[3, 0, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
return %0 : tensor<4xi32>
}
diff --git a/mlir/test/Dialect/Tosa/ops.mlir b/mlir/test/Dialect/Tosa/ops.mlir
index ec3d0c881d857..4e71d512def2d 100644
--- a/mlir/test/Dialect/Tosa/ops.mlir
+++ b/mlir/test/Dialect/Tosa/ops.mlir
@@ -1938,3 +1938,75 @@ func.func @test_assert_equal_shape() {
tosa.assert_equal_shape %0, %1 {allow_broadcast = true} : (!tosa.shape<2>, !tosa.shape<2>) -> ()
return
}
+
+// -----
+
+// CHECK-LABEL: test_block_scaled_const
+func.func @test_block_scaled_const() -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>> {
+ %0 = "tosa.const"() <{values = dense<tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN, {1.0, 1.0}>> : [[0.0 : f8E4M3FN, 1.0 : f8E4M3FN, 0.001953125 : f8E4M3FN, 0.0078125 : f8E4M3FN,
+ 2.0 : f8E4M3FN, 2.25 : f8E4M3FN, 2.5 : f8E4M3FN, 2.75 : f8E4M3FN,
+ 15.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN],
+ [0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN,
+ 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN, 0.0 : f8E4M3FN]]>}> : () -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+ return %0 : tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+}
+
+// -----
+
+// CHECK-LABEL: test_block_scaled_const_splat
+func.func @test_block_scaled_const_splat() -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f4E2M1FN>> {
+ %0 = "tosa.const"() <{values = dense<tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f4E2M1FN, {1.0, 1.0}>> : 0.0 : f4E2M1FN>}> : () -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f4E2M1FN>>
+ return %0 : tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f4E2M1FN>>
+}
+
+// -----
+
+// CHECK-LABEL: test_block_scaled_const_splat_mxint8
+func.func @test_block_scaled_const_splat_mxint8() -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:!tosa.mxint8>> {
+ %0 = "tosa.const"() <{values = dense<tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:!tosa.mxint8, {1.0, 1.0}>> : 0 : i8>}> : () -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:!tosa.mxint8>>
+ return %0 : tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:!tosa.mxint8>>
+}
+
+// -----
+
+// CHECK-LABEL: test_block_scaled_const_scale_values
+func.func @test_block_scaled_const_scale_values() -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>> {
+ %0 = "tosa.const"() <{values = dense<tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN, {1.0, 2.0}>> : 0.0 : f8E4M3FN>}> : () -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+ return %0 : tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+}
+
+// -----
+
+// CHECK-LABEL: test_block_scaled_const_scale_values_explicit_type
+func.func @test_block_scaled_const_scale_values_explicit_type() -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>> {
+ %0 = "tosa.const"() <{values = dense<tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN, {1.0 : f8E8M0FNU, 2.0 : f8E8M0FNU}>> : 0.0 : f8E4M3FN>}> : () -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+ return %0 : tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+}
+
+// -----
+
+// CHECK-LABEL: test_block_scaled_const_scale_values_wide_inner_dim
+func.func @test_block_scaled_const_scale_values_wide_inner_dim() -> tensor<2x64x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>> {
+ %0 = "tosa.const"() <{values = dense<tensor<2x64x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN, {1.0, 2.0, 4.0, 8.0}>> : 0.0 : f8E4M3FN>}> : () -> tensor<2x64x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+ return %0 : tensor<2x64x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN>>
+}
+
+// -----
+
+// CHECK-LABEL: test_block_scaled_const_cast_scale_values_no_propagate
+func.func @test_block_scaled_const_cast_scale_values_no_propagate() -> tensor<2x32xf32> {
+ %0 = "tosa.const"() <{values = dense<tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M3FN, {2.0, 4.0}>> : 0.0 : f8E4M3FN>}> : () -> tensor<2x32x!tosa.block_scaled<BLOCK_SHAPE_32:f8E8M0FNU:f8E4M...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/205506
More information about the Mlir-commits
mailing list