Skip to content

[mlir][tosa] Make RESCALE op input_unsigned and output_unsigned attributes required #129339

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
Mar 3, 2025
Merged

[mlir][tosa] Make RESCALE op input_unsigned and output_unsigned attributes required #129339

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
4 changes: 2 additions & 2 deletions mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2289,8 +2289,8 @@ def Tosa_RescaleOp: Tosa_Op<"rescale", [Pure,
BoolAttr:$scale32,
BoolAttr:$double_round,
BoolAttr:$per_channel,
DefaultValuedOptionalAttr<BoolAttr, "false">:$input_unsigned,
DefaultValuedOptionalAttr<BoolAttr, "false">:$output_unsigned
BoolAttr: $input_unsigned,
BoolAttr: $output_unsigned
);

let results = (outs
Expand Down
2 changes: 1 addition & 1 deletion mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-invalid.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,7 +43,7 @@ func.func @rfft2d_with_non_float_type(%arg0 : tensor<1x1x1xi32>) -> (tensor<1x1x
// CHECK-LABEL: @rescale_unsupported_type
func.func @rescale_unsupported_type(%arg0: tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>> {
// expected-error@+1 {{failed to legalize operation 'tosa.rescale'}}
%0 = tosa.rescale %arg0 {double_round = false, input_zp = 127 : i32, multiplier = array<i32: 1073741824>, output_zp = -1 : i32, per_channel = false, scale32 = true, shift = array<i8: 30>} : (tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
%0 = tosa.rescale %arg0 {double_round = false, input_zp = 127 : i32, multiplier = array<i32: 1073741824>, output_zp = -1 : i32, per_channel = false, scale32 = true, shift = array<i8: 30>, input_unsigned = true, output_unsigned = false} : (tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
return %0 : tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
}

Expand Down
18 changes: 9 additions & 9 deletions mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1149,7 +1149,7 @@ func.func @rescale_i8(%arg0 : tensor<2xi8>) -> () {
// CHECK-DAG: [[BOUNDED:%.+]] = arith.minsi [[CMAX]], [[LOWER]]
// CHECK-DAG: [[TRUNC:%.+]] = arith.trunci [[BOUNDED]]
// CHECK-DAG: linalg.yield [[TRUNC]]
%0 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false} : (tensor<2xi8>) -> tensor<2xi8>
%0 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, input_unsigned = false, output_unsigned = false} : (tensor<2xi8>) -> tensor<2xi8>

// CHECK: return
return
Expand Down Expand Up @@ -1178,7 +1178,7 @@ func.func @rescale_i8_unsigned_output(%arg0 : tensor<2xi8>) -> () {
// CHECK-DAG: [[BOUNDED:%.+]] = arith.minsi [[CMAX]], [[LOWER]]
// CHECK-DAG: [[TRUNC:%.+]] = arith.trunci [[BOUNDED]]
// CHECK: linalg.yield [[TRUNC]]
%1 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, output_unsigned = true} : (tensor<2xi8>) -> tensor<2xi8>
%1 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, input_unsigned = false, output_unsigned = true} : (tensor<2xi8>) -> tensor<2xi8>

// CHECK: return
return
Expand All @@ -1195,13 +1195,13 @@ func.func @rescale_i8_dyn_batch(%arg0 : tensor<?x2xi8>) -> () {
// CHECK: %[[BATCH:.+]] = tensor.dim %[[ARG0]], %[[C0]]
// CHECK: %[[INIT:.+]] = tensor.empty(%[[BATCH]]) : tensor<?x2xi8>
// CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel", "parallel"]} ins(%[[ARG0]] : tensor<?x2xi8>) outs(%[[INIT]] : tensor<?x2xi8>)
%0 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false} : (tensor<?x2xi8>) -> tensor<?x2xi8>
%0 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, input_unsigned = false, output_unsigned = false} : (tensor<?x2xi8>) -> tensor<?x2xi8>

// CHECK: %[[C0:.+]] = arith.constant 0
// CHECK: %[[BATCH:.+]] = tensor.dim %[[ARG0]], %[[C0]]
// CHECK: %[[INIT:.+]] = tensor.empty(%[[BATCH]]) : tensor<?x2xi8>
// CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel", "parallel"]} ins(%[[ARG0]] : tensor<?x2xi8>) outs(%[[INIT]] : tensor<?x2xi8>)
%1 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, output_unsigned = true} : (tensor<?x2xi8>) -> tensor<?x2xi8>
%1 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, input_unsigned = false, output_unsigned = true} : (tensor<?x2xi8>) -> tensor<?x2xi8>

return
}
Expand All @@ -1219,7 +1219,7 @@ func.func @rescale_dyn(%arg0 : tensor<1x?x?x32xi32>) -> () {
// CHECK: %[[DIM2:.+]] = tensor.dim %[[ARG0]], %[[C2]]
// CHECK: %[[INIT:.+]] = tensor.empty(%[[DIM1]], %[[DIM2]])
// CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%[[ARG0]] : tensor<1x?x?x32xi32>) outs(%[[INIT]] : tensor<1x?x?x32xi8>)
%0 = tosa.rescale %arg0 {double_round = true, input_zp = 0 : i32, multiplier = array<i32: 1376784203>, output_zp = 0 : i32, per_channel = false, scale32 = true, shift = array<i8: 38>} : (tensor<1x?x?x32xi32>) -> tensor<1x?x?x32xi8>
%0 = tosa.rescale %arg0 {double_round = true, input_zp = 0 : i32, multiplier = array<i32: 1376784203>, output_zp = 0 : i32, per_channel = false, scale32 = true, shift = array<i8: 38>, input_unsigned = false, output_unsigned = false} : (tensor<1x?x?x32xi32>) -> tensor<1x?x?x32xi8>
return
}

Expand Down Expand Up @@ -1247,7 +1247,7 @@ func.func @rescale_i8_unsigned_input(%arg0 : tensor<2xi8>) -> () {
// CHECK-DAG: [[BOUNDED:%.+]] = arith.minsi [[CMAX]], [[LOWER]]
// CHECK-DAG: [[TRUNC:%.+]] = arith.trunci [[BOUNDED]]
// CHECK: linalg.yield [[TRUNC]]
%0 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, input_unsigned = true} : (tensor<2xi8>) -> tensor<2xi8>
%0 = tosa.rescale %arg0 {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = false, double_round = false, per_channel = false, input_unsigned = true, output_unsigned = false} : (tensor<2xi8>) -> tensor<2xi8>

return
}
Expand Down Expand Up @@ -1277,7 +1277,7 @@ func.func @rescale_per_channel(%arg0 : tensor<3xi8>) -> (tensor<3xi8>) {
// CHECK-DAG: [[BOUNDED:%.+]] = arith.minsi [[CMAX]], [[LOWER]]
// CHECK-DAG: [[TRUNC:%.+]] = arith.trunci [[BOUNDED]]
// CHECK-DAG: linalg.yield [[TRUNC]]
%0 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 42, 43, 44>, shift = array<i8: 14, 15, 64>, scale32 = false, double_round = false, per_channel = false} : (tensor<3xi8>) -> tensor<3xi8>
%0 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 42, 43, 44>, shift = array<i8: 14, 15, 64>, scale32 = false, double_round = false, per_channel = false, input_unsigned = false, output_unsigned = false} : (tensor<3xi8>) -> tensor<3xi8>

// CHECK: return [[GENERIC]]
return %0 : tensor<3xi8>
Expand All @@ -1290,7 +1290,7 @@ func.func @rescaleDoubleRound(%arg0 : tensor<2xi8>) -> (tensor<2xi8>) {
// CHECK: linalg.generic
// CHECK: tosa.apply_scale
// CHECK-SAME: {double_round = true}
%0 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 19689>, shift = array<i8: 33>, scale32 = true, double_round = true, per_channel = false} : (tensor<2xi8>) -> tensor<2xi8>
%0 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 19689>, shift = array<i8: 33>, scale32 = true, double_round = true, per_channel = false, input_unsigned = false, output_unsigned = false} : (tensor<2xi8>) -> tensor<2xi8>
return %0 : tensor<2xi8>
}

Expand All @@ -1299,7 +1299,7 @@ func.func @rescaleUnnecessaryDoubleRound(%arg0 : tensor<2xi8>) -> (tensor<2xi8>)
// CHECK: linalg.generic
// CHECK: tosa.apply_scale
// CHECK-SAME: {double_round = false}
%0 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = true, double_round = true, per_channel = false} : (tensor<2xi8>) -> tensor<2xi8>
%0 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 19689>, shift = array<i8: 15>, scale32 = true, double_round = true, per_channel = false, input_unsigned = false, output_unsigned = false} : (tensor<2xi8>) -> tensor<2xi8>
return %0 : tensor<2xi8>
}

Expand Down
2 changes: 1 addition & 1 deletion mlir/test/Dialect/Tosa/availability.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -611,7 +611,7 @@ func.func @test_cast1(%arg0: tensor<13x21x3xi32>) -> tensor<13x21x3xf32> {
func.func @test_rescale(%arg0: tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>, %multiplier : tensor<1xi32>, %shift : tensor<1xi8>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>> {
// CHECK: profiles: [ [pro_int] ]
// CHECK: extensions: [ [int16] ]
%0 = tosa.rescale %arg0 {double_round = false, input_zp = 127 : i32, multiplier = array<i32: 1073741824>, output_zp = -1 : i32, per_channel = false, scale32 = true, shift = array<i8: 30>} : (tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
%0 = tosa.rescale %arg0 {double_round = false, input_zp = 127 : i32, multiplier = array<i32: 1073741824>, output_zp = -1 : i32, per_channel = false, scale32 = true, shift = array<i8: 30>, input_unsigned = true, output_unsigned = false} : (tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
return %0 : tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
}

Expand Down
2 changes: 0 additions & 2 deletions mlir/test/Dialect/Tosa/canonicalize.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -897,8 +897,6 @@ func.func @reshape_quant_nofold() -> tensor<1x1x1x1xi32> {
%0 = "tosa.const"() {value = dense<127> : tensor<i8>} : () -> tensor<!quant.uniform<i8:f32, 3.0757404601899907E-5:-128>>
%cst0 = "tosa.const_shape"() {value = dense<[1, 1, 1, 1]> : tensor<4xindex>} : () -> !tosa.shape<4>
%1 = tosa.reshape %0, %cst0 : (tensor<!quant.uniform<i8:f32, 3.0757404601899907E-5:-128>>, !tosa.shape<4>) -> tensor<1x1x1x1x!quant.uniform<i8:f32, 3.0757404601899907E-5:-128>>
%multiplier = "tosa.const"() {value = dense<1073741824> : tensor<1xi32> } : () -> tensor<1xi32>
%shift = "tosa.const"() {value = dense<30> : tensor<1xi8> } : () -> tensor<1xi8>
%2 = tosa.rescale %1 {double_round = true, input_zp = -128 : i32, multiplier = array<i32: 1073741824>, output_zp = 0 : i32, per_channel = false, scale32 = true, shift = array<i8: 30>, input_unsigned = false, output_unsigned = false} : (tensor<1x1x1x1x!quant.uniform<i8:f32, 3.0757404601899907E-5:-128>>) -> tensor<1x1x1x1xi32>
return %2 : tensor<1x1x1x1xi32>
}
Expand Down
4 changes: 2 additions & 2 deletions mlir/test/Dialect/Tosa/ops.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,7 +82,7 @@ func.func @test_conv2d_q8xi4(%arg0: tensor<1x11x11x3xi8>) -> tensor<1x1x1x3xi8>
%izp = "tosa.const"() {value = dense<0> : tensor<1xi8>} : () -> tensor<1xi8>
%wzp = "tosa.const"() {value = dense<0> : tensor<1xi4>} : () -> tensor<1xi4>
%2 = "tosa.conv2d"(%arg0, %0, %1, %izp, %wzp) {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<1x11x11x3xi8>, tensor<3x11x11x3xi4>, tensor<3xi32>, tensor<1xi8>, tensor<1xi4>) -> tensor<1x1x1x3xi32>
%3 = "tosa.rescale"(%2) {double_round = true, input_zp = 0 : i32, multiplier = array<i32: 2026291432, 1079222024, 1693132724>, output_zp = 27 : i32, per_channel = true, scale32 = true, shift = array<i8: 37, 36, 37>} : (tensor<1x1x1x3xi32>) -> tensor<1x1x1x3xi8>
%3 = "tosa.rescale"(%2) {double_round = true, input_zp = 0 : i32, multiplier = array<i32: 2026291432, 1079222024, 1693132724>, output_zp = 27 : i32, per_channel = true, scale32 = true, shift = array<i8: 37, 36, 37>, input_unsigned = false, output_unsigned = false} : (tensor<1x1x1x3xi32>) -> tensor<1x1x1x3xi8>
return %3 : tensor<1x1x1x3xi8>
}

Expand Down Expand Up @@ -707,7 +707,7 @@ func.func @test_cast3(%arg0: tensor<13x21x3xi32>) -> tensor<13x21x3x!quant.unifo
// -----
// CHECK-LABEL: rescale
func.func @test_rescale(%arg0: tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>> {
%0 = tosa.rescale %arg0 {double_round = false, input_zp = 127 : i32, multiplier = array<i32: 1073741824>, output_zp = -1 : i32, per_channel = false, scale32 = true, shift = array<i8: 30>} : (tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
%0 = tosa.rescale %arg0 {double_round = false, input_zp = 127 : i32, multiplier = array<i32: 1073741824>, output_zp = -1 : i32, per_channel = false, scale32 = true, shift = array<i8: 30>, input_unsigned = false, output_unsigned = false} : (tensor<13x21x3x!quant.uniform<u8:f32, 0.015655439347028732:127>>) -> tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
return %0 : tensor<13x21x3x!quant.uniform<i8:f32, 0.015655439347028732:-1>>
}

Expand Down
2 changes: 1 addition & 1 deletion mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ func.func @test_unary_i32(%arg0 : tensor<4xi32>) -> () {
%5 = tosa.reverse %arg0 { axis = 0 : i32 } : (tensor<4xi32>) -> tensor<?xi32>

// CHECK: tosa.rescale %arg0 {{.+}} : (tensor<4xi32>) -> tensor<4xi16>
%6 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 42, 43>, shift = array<i8: 14, 15>, scale32 = false, double_round = false, per_channel = false} : (tensor<4xi32>) -> tensor<*xi16>
%6 = tosa.rescale %arg0 {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = array<i32: 42, 43>, shift = array<i8: 14, 15>, scale32 = false, double_round = false, per_channel = false, input_unsigned = false, output_unsigned = false} : (tensor<4xi32>) -> tensor<*xi16>

// CHECK: tosa.identity %arg0 : (tensor<4xi32>) -> tensor<4xi32>
%7 = tosa.identity %arg0 : (tensor<4xi32>) -> tensor<?xi32>
Expand Down
7 changes: 6 additions & 1 deletion mlir/test/lib/Dialect/Tosa/TosaTestPasses.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,13 +165,18 @@ ConvertTosaConv2DOp::matchAndRewrite(Operation *op,
// Obtain the quantized scale = multiplier and shift.
computeMultiplierAndShift(opTensorScale, multiplier, shift, 32);

bool input_unsigned =
newTosaConv2DOp.getResult().getType().isUnsignedInteger();
bool output_unsigned = outputType.isUnsignedInteger();

auto newTosaRescaleOp = rewriter.create<tosa::RescaleOp>(
op->getLoc(), outputType, newTosaConv2DOp.getResult(),
rewriter.getI32IntegerAttr(0), rewriter.getI32IntegerAttr(outputZp),
rewriter.getDenseI32ArrayAttr({multiplier}),
rewriter.getDenseI8ArrayAttr({static_cast<int8_t>(shift)}),
rewriter.getBoolAttr(true), rewriter.getBoolAttr(true),
rewriter.getBoolAttr(false));
rewriter.getBoolAttr(false), rewriter.getBoolAttr(input_unsigned),
rewriter.getBoolAttr(output_unsigned));

rewriter.replaceOp(op, {newTosaRescaleOp.getResult()});
return success();
Expand Down