[mlir][tosa]Fix Rescale shift attr data type

[Tai78641]

Summary

Change Rescale shift attribute to be DenseI8ArrayAttr to match spec
(instead of DenseI32ArrayAttr)

Signed-off-by: Tai Ly <[e-mail]>
Change-Id: Ib9f492845eb7e71cb6bf32cbcd785011cae6eb3d

[rsuderman]

mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td

1760

For these kind of changes it is better to relax the requirement (support i32s and i8s) for a fixed period of time, then constrain back to i8. This allows downstream users to incrementally transition vs breaking all users immediately.

[Tai78641]

mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td

1760

not clear how to allow both i32 and i8 attr types for same attribute.
have considered adding a second attribute, eg, shift_i8, for a while and then remove shift later, then rename shift_i8 to shift.
seems there is no way to make this change without breaking users one way or another.
would appreciate any suggestions to phase this in.

[jpienaar]

If this one is still active, could we move it to PR?

[Tai78641]

In D157439#4655999, @jpienaar wrote:

If this one is still active, could we move it to PR?

done. https://github.com/llvm/llvm-project/pull/71084

[Tai78641]

this is replaced by https://github.com/llvm/llvm-project/pull/71084

mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td

let description = [{
     signed 8 to unsigned 8  int8    uint8
   }];
 
   let arguments = (ins
     Tosa_Tensor:$input,
     I32Attr:$input_zp,
     I32Attr:$output_zp,
     DenseI32ArrayAttr:$multiplier,
-    DenseI32ArrayAttr:$shift,
+    DenseI8ArrayAttr:$shift,
     BoolAttr:$scale32,
     BoolAttr:$double_round,
     BoolAttr:$per_channel
   );
 
   let results = (outs
     Tosa_Tensor:$output
   );

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

func.func @rescale_i8(%arg0 : tensor<2xi8>) -> () {
   // CHECK-DAG: [[CMIN:%.+]] = arith.constant -128
   // CHECK-DAG: [[CMAX:%.+]] = arith.constant 127
   // CHECK-DAG: [[MINLT:%.+]] = arith.cmpi slt, [[SCALED_ZEROED]], [[CMIN]]
   // CHECK-DAG: [[MAXLT:%.+]] = arith.cmpi slt, [[CMAX]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[LOWER:%.+]] = arith.select [[MINLT]], [[CMIN]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[BOUNDED:%.+]] = arith.select [[MAXLT]], [[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<i32: 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} : (tensor<2xi8>)  -> (tensor<2xi8>)
 
   // CHECK: [[C0:%.+]] = arith.constant 19689
   // CHECK: [[C1:%.+]] = arith.constant 15
   // CHECK: [[INIT:%.+]] = tensor.empty()
   // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel"]} ins(%[[ARG0]] : tensor<2xi8>) outs([[INIT]] : tensor<2xui8>)
   // CHECK: ^bb0([[IN:%.+]]: i8, [[UNUSED:%.+]]: ui8):
   // CHECK: [[C17:%.+]] = arith.constant 17
   // CHECK: [[C22:%.+]] = arith.constant 22
   // CHECK-DAG: [[IN32:%.+]] = arith.extsi [[IN]]
   // CHECK-DAG: [[IN_ZEROED:%.+]] = arith.subi [[IN32]], [[C17]]
   // CHECK-DAG: [[SCALED:%.+]] = "tosa.apply_scale"([[IN_ZEROED]], [[C0]], [[C1]]) <{double_round = false}
   // CHECK-DAG: [[SCALED_ZEROED:%.+]] = arith.addi [[SCALED]], [[C22]]
   // CHECK-DAG: [[CMIN:%.+]] = arith.constant 0
   // CHECK-DAG: [[CMAX:%.+]] = arith.constant 255
   // CHECK-DAG: [[MINLT:%.+]] = arith.cmpi slt, [[SCALED_ZEROED]], [[CMIN]]
   // CHECK-DAG: [[LOWER:%.+]] = arith.select [[MINLT]], [[CMIN]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[MAXLT:%.+]] = arith.cmpi slt, [[CMAX]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[BOUNDED:%.+]] = arith.select [[MAXLT]], [[CMAX]], [[LOWER]]
   // CHECK-DAG: [[TRUNC:%.+]] = arith.trunci [[BOUNDED]]
   // CHECK-DAG: [[CAST:%.+]] = builtin.unrealized_conversion_cast [[TRUNC]] : i8 to ui8
   // CHECK: linalg.yield [[CAST]]
-  %1 = "tosa.rescale"(%arg0) {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i32: 15>, scale32 = false, double_round = false, per_channel = false} : (tensor<2xi8>)  -> (tensor<2xui8>)
+  %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} : (tensor<2xi8>)  -> (tensor<2xui8>)
 
   // CHECK: return
   return
 }
 
 // -----
 
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, d1)>
 
 // CHECK-LABEL: @rescale_i8_dyn_batch
 // CHECK-SAME: (%[[ARG0:[0-9a-zA-Z_]*]]:
 func.func @rescale_i8_dyn_batch(%arg0 : 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>)
-  %0 = "tosa.rescale"(%arg0) {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i32: 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} : (tensor<?x2xi8>)  -> (tensor<?x2xi8>)
 
   // CHECK: %[[C0:.+]] = arith.constant 0
   // CHECK: %[[BATCH:.+]] = tensor.dim %[[ARG0]], %[[C0]]
   // CHECK: %[[INIT:.+]] = tensor.empty(%[[BATCH]]) : tensor<?x2xui8>
   // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel", "parallel"]} ins(%[[ARG0]] : tensor<?x2xi8>) outs(%[[INIT]] : tensor<?x2xui8>)
-  %1 = "tosa.rescale"(%arg0) {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = array<i32: 19689>, shift = array<i32: 15>, scale32 = false, double_round = false, per_channel = false} : (tensor<?x2xi8>)  -> (tensor<?x2xui8>)
+  %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} : (tensor<?x2xi8>)  -> (tensor<?x2xui8>)
 
   return
 }
 
 // -----
 
 // CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
 
 // CHECK-LABEL: @rescale_dyn
 // CHECK-SAME: (%[[ARG0:[0-9a-zA-Z_]*]]:
 func.func @rescale_dyn(%arg0 : tensor<1x?x?x32xi32>) -> () {
   // CHECK: %[[C1:.+]] = arith.constant 1
   // CHECK: %[[DIM1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
   // CHECK: %[[C2:.+]] = arith.constant 2
   // 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<i32: 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>} : (tensor<1x?x?x32xi32>) -> tensor<1x?x?x32xi8>
   return
 }
 
 // -----
 
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0) -> (d0)>
 
 // CHECK-LABEL: @rescale_ui8
func.func @rescale_ui8(%arg0 : tensor<2xui8>) -> () {
   // CHECK-DAG: [[CMIN:%.+]] = arith.constant -128
   // CHECK-DAG: [[CMAX:%.+]] = arith.constant 127
   // CHECK-DAG: [[MINLT:%.+]] = arith.cmpi slt, [[SCALED_ZEROED]], [[CMIN]]
   // CHECK-DAG: [[LOWER:%.+]] = arith.select [[MINLT]], [[CMIN]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[MAXLT:%.+]] = arith.cmpi slt, [[CMAX]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[BOUNDED:%.+]] = arith.select [[MAXLT]], [[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<i32: 15>, scale32 = false, double_round = false, per_channel = false} : (tensor<2xui8>)  -> (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} : (tensor<2xui8>)  -> (tensor<2xi8>)
 
   return
 }
 
 // -----
 
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0) -> (d0)>
 
func.func @rescale_per_channel(%arg0 : tensor<3xi8>) -> (tensor<3xi8>) {
   // CHECK-DAG: [[CMIN:%.+]] = arith.constant -128
   // CHECK-DAG: [[CMAX:%.+]] = arith.constant 127
   // CHECK-DAG: [[MINLT:%.+]] = arith.cmpi slt, [[SCALED_ZEROED]], [[CMIN]]
   // CHECK-DAG: [[MAXLT:%.+]] = arith.cmpi slt, [[CMAX]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[LOWER:%.+]] = arith.select [[MINLT]], [[CMIN]], [[SCALED_ZEROED]]
   // CHECK-DAG: [[BOUNDED:%.+]] = arith.select [[MAXLT]], [[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<i32: 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} : (tensor<3xi8>)  -> (tensor<3xi8>)
 
   // CHECK: return [[GENERIC]]
   return %0 : tensor<3xi8>
 }
 
 // -----
 
 // CHECK-LABEL: @rescaleDoubleRound
 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<i32: 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} : (tensor<2xi8>)  -> (tensor<2xi8>)
   return %0 : tensor<2xi8>
 }
 
 // CHECK-LABEL: @rescaleUnnecessaryDoubleRound
 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<i32: 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} : (tensor<2xi8>)  -> (tensor<2xi8>)
   return %0 : tensor<2xi8>
 }
 
 // -----
 
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, d1)>
 
 // CHECK-LABEL: @reverse

mlir/test/Dialect/Tosa/ops.mlir

 }
 
 // -----
 // CHECK-LABEL: conv2d_q8xi4
 func.func @test_conv2d_q8xi4(%arg0: tensor<1x11x11x3xi8>) -> tensor<1x1x1x3xi8> {
   %0 = "tosa.const"() {value = dense<0> : tensor<3x11x11x3xi4>} : () -> tensor<3x11x11x3xi4>
   %1 = "tosa.const"() {value = dense<[12, 23, 55]> : tensor<3xi32>} : () -> tensor<3xi32>
   %2 = "tosa.conv2d"(%arg0, %0, %1) {dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, quantization_info = #tosa.conv_quant<input_zp = 0, weight_zp = 0>, stride = array<i64: 1, 1>} : (tensor<1x11x11x3xi8>, tensor<3x11x11x3xi4>, tensor<3xi32>) -> 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<i32: 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>} : (tensor<1x1x1x3xi32>) -> tensor<1x1x1x3xi8>
   return %3 : tensor<1x1x1x3xi8>
 }
 
 // -----
 // CHECK-LABEL: depthwise_conv2d
 func.func @test_depthwise_conv2d(%arg0: tensor<1x4x4x4xf32>, %arg1: tensor<1x1x4x2xf32>, %arg2: tensor<8xf32>) -> tensor<1x4x4x8xf32> {
   %2 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<1x4x4x4xf32>, tensor<1x1x4x2xf32>, tensor<8xf32>) -> tensor<1x4x4x8xf32>
   return %2 : tensor<1x4x4x8xf32>

 func.func @test_cast3(%arg0: tensor<13x21x3xi32>) -> tensor<13x21x3x!quant.uniform<i16:f32, 0.078431375324726104:128>> {
   %0 = "tosa.cast"(%arg0) : (tensor<13x21x3xi32>) -> tensor<13x21x3x!quant.uniform<i16:f32, 0.078431375324726104:128>>
   return %0 : tensor<13x21x3x!quant.uniform<i16:f32, 0.078431375324726104:128>>
 }
 
 // -----
 // 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<i32: 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>} : (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>>
 }
 
 // -----
 // CHECK-LABEL: const
 func.func @test_const(%arg0 : index) -> tensor<4xi32> {
     %0 = "tosa.const"() {value = dense<[3, 0, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
     return %0 : tensor<4xi32>

mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir

func.func @test_unary_i32(%arg0 : tensor<4xi32>) -> () {
 
   // CHECK: "tosa.negate"(%arg0) : (tensor<4xi32>) -> tensor<4xi32>
   %4 = "tosa.negate"(%arg0) : (tensor<4xi32>) -> tensor<*xi32>
 
   // CHECK: "tosa.reverse"(%arg0) <{axis = 0 : i64}> : (tensor<4xi32>) -> tensor<4xi32>
   %5 = "tosa.reverse"(%arg0) { axis = 0 : i64 } : (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<i32: 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} : (tensor<4xi32>)  -> (tensor<*xi16>)
 
   // CHECK: "tosa.identity"(%arg0) : (tensor<4xi32>) -> tensor<4xi32>
   %7 = "tosa.identity"(%arg0) : (tensor<4xi32>) -> tensor<?xi32>
   return
 }
 
 // -----
 

mlir/test/lib/Dialect/Tosa/TosaTestPasses.cpp

ConvertTosaConv2DOp::matchAndRewrite(Operation *op,
 
   // Obtain the quantized scale = multiplier and shift.
   computeMultiplierAndShift(opTensorScale, multiplier, shift, 32);
 
   auto newTosaRescaleOp = rewriter.create<tosa::RescaleOp>(
       op->getLoc(), outputType, newTosaConv2DOp.getResult(),
       rewriter.getI32IntegerAttr(0), rewriter.getI32IntegerAttr(outputZp),
       rewriter.getDenseI32ArrayAttr({multiplier}),
-      rewriter.getDenseI32ArrayAttr({shift}), rewriter.getBoolAttr(true),
+      rewriter.getDenseI8ArrayAttr({shift}), rewriter.getBoolAttr(true),
       rewriter.getBoolAttr(true), rewriter.getBoolAttr(false));
 
   rewriter.replaceOp(op, {newTosaRescaleOp.getResult()});
   return success();
 }
 
 namespace {