Skip to content

[mlir][tosa] Fix mul folder conformance to the spec #137601

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 8, 2025
Merged

[mlir][tosa] Fix mul folder conformance to the spec #137601

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 25 additions & 6 deletions mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -918,6 +918,27 @@ OpFoldResult IntDivOp::fold(FoldAdaptor adaptor) {
}

namespace {
// calculate lhs * rhs >> shift according to TOSA Spec
// return nullopt if result is not in range of int32_t when shift > 0
std::optional<APInt> mulInt(APInt lhs, APInt rhs, int32_t shift,
unsigned bitwidth) {
APInt result = lhs.sext(64) * rhs.sext(64);

if (shift > 0) {
auto round = APInt(64, 1) << (shift - 1);
result += round;
result.ashrInPlace(shift);
// REQUIRE(product >= minimum_s<i32_t>() && product <= maximum_s<i32_t>())
if (!(result.getSExtValue() >= INT32_MIN &&
result.getSExtValue() <= INT32_MAX)) {
// REQUIRE failed
return std::nullopt;
}
}

return result.trunc(bitwidth);
}

DenseElementsAttr mulBinaryFolder(DenseElementsAttr lhs, DenseElementsAttr rhs,
RankedTensorType ty, int32_t shift) {
if (rhs && lhs && rhs.isSplat() && lhs.isSplat()) {
Expand All @@ -930,12 +951,10 @@ DenseElementsAttr mulBinaryFolder(DenseElementsAttr lhs, DenseElementsAttr rhs,
}

auto bitwidth = ty.getElementType().getIntOrFloatBitWidth();
l = l.sext(bitwidth * 2);
r = r.sext(bitwidth * 2);
auto result = l * r;
result.lshrInPlace(shift);
result = result.trunc(bitwidth);
return DenseElementsAttr::get(ty, result);
const std::optional<APInt> result = mulInt(l, r, shift, bitwidth);
Copy link
Contributor

@udaya-ranga udaya-ranga Apr 28, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM, reviewed

if (!result)
return {};
return DenseElementsAttr::get(ty, result.value());
}

if (llvm::isa<FloatType>(ty.getElementType())) {
Expand Down
41 changes: 40 additions & 1 deletion mlir/test/Dialect/Tosa/canonicalize.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1226,4 +1226,43 @@ func.func @slice_dynamic_size_static_output_canonicalize(%arg0: tensor<2x60x59x?
%1 = tosa.const_shape {values = dense<[-1, 60, 58, -1]> : tensor<4xindex>} : () -> !tosa.shape<4>
%2 = tosa.slice %arg0, %0, %1 : (tensor<2x60x59x?xf32>, !tosa.shape<4>, !tosa.shape<4>) -> tensor<2x60x58x?xf32>
return %2 : tensor<2x60x58x?xf32>
}
}

// -----

// CHECK-LABEL: @fold_mul_shift
// CHECK-DAG: "tosa.const"() <{values = dense<1> : tensor<i32>}> : () -> tensor<i32>
func.func @fold_mul_shift() -> tensor<i32> {
%0 = "tosa.const"() <{values = dense<-23661> : tensor<i32>}> : () -> tensor<i32>
%1 = "tosa.const"() <{values = dense<-33022> : tensor<i32>}> : () -> tensor<i32>
%2 = "tosa.const"() <{values = dense<30> : tensor<1xi8>}> : () -> tensor<1xi8>
%3 = tosa.mul %0, %1, %2 : (tensor<i32>, tensor<i32>, tensor<1xi8>) -> tensor<i32>
return %3 : tensor<i32>
}

// -----

// CHECK-LABEL: @fold_mul_no_shift
// CHECK-DAG: "tosa.const"() <{values = dense<781333542> : tensor<i32>}> : () -> tensor<i32>
func.func @fold_mul_no_shift() -> tensor<i32> {
%0 = "tosa.const"() <{values = dense<-23661> : tensor<i32>}> : () -> tensor<i32>
%1 = "tosa.const"() <{values = dense<-33022> : tensor<i32>}> : () -> tensor<i32>
%2 = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
%3 = tosa.mul %0, %1, %2 : (tensor<i32>, tensor<i32>, tensor<1xi8>) -> tensor<i32>
return %3 : tensor<i32>
}

// -----

// CHECK-LABEL: @no_fold_mul_result_exceeds_i32
// CHECK-DAG: %[[LHS:.*]] = "tosa.const"() <{values = dense<23661> : tensor<i32>}> : () -> tensor<i32>
// CHECK-DAG: %[[RHS:.*]] = "tosa.const"() <{values = dense<330222> : tensor<i32>}> : () -> tensor<i32>
// CHECK-DAG: %[[SHIFT:.*]] = "tosa.const"() <{values = dense<1> : tensor<1xi8>}> : () -> tensor<1xi8>
// CHECK: tosa.mul %[[LHS]], %[[RHS]], %[[SHIFT]] : (tensor<i32>, tensor<i32>, tensor<1xi8>) -> tensor<i32>
func.func @no_fold_mul_result_exceeds_i32() -> tensor<i32> {
%0 = "tosa.const"() <{values = dense<23661> : tensor<i32>}> : () -> tensor<i32>
%1 = "tosa.const"() <{values = dense<330222> : tensor<i32>}> : () -> tensor<i32>
%2 = "tosa.const"() <{values = dense<1> : tensor<1xi8>}> : () -> tensor<1xi8>
%3 = tosa.mul %0, %1, %2 : (tensor<i32>, tensor<i32>, tensor<1xi8>) -> tensor<i32>
return %3 : tensor<i32>
}