[mlir][tosa] Add accumulator type attribute to TOSA dialect

Tosa supports fp16 and fp32 accumulator type for fp16 input, but
no way to tell for computational operators whether accumulator
should be fp16 or fp32 from input type. Add this new attribute
to specify the type.

Set to fp32 by default for now. When the time fp16 is supported,
the accumulator type can be selected based on trade-off between
performance and accuracy.

Reviewed By: jpienaar

Differential Revision: https://reviews.llvm.org/D146317

3 files changed, 11 insertions, 11 deletions

diff --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
index 838e957ae16b..901dcf78380b 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
@@ -286,7 +286,7 @@ 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]]
-  %0 = "tosa.avg_pool2d"(%arg0) {pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<1x6x34x62xf32>)  -> (tensor<1x5x33x62xf32>)
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = f32, pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<1x6x34x62xf32>)  -> (tensor<1x5x33x62xf32>)
   return %0 : tensor<1x5x33x62xf32>
 }
 
@@ -329,7 +329,7 @@ func.func @avg_pool_i8(%arg0: tensor<1x6x34x62xi8>) -> (tensor<1x5x33x62xi8>) {
   // CHECK: %[[CLAMP:.+]] = arith.select %[[CMP]], %[[CMAX]], %[[SEL]]
   // CHECK: %[[TRUNC:.+]] = arith.trunci %[[CLAMP]]
   // CHECK: linalg.yield %[[TRUNC]]
-  %0 = "tosa.avg_pool2d"(%arg0) {pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<1x6x34x62xi8>)  -> (tensor<1x5x33x62xi8>)
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = i32, pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<1x6x34x62xi8>)  -> (tensor<1x5x33x62xi8>)
   return %0 : tensor<1x5x33x62xi8>
 }
 
@@ -352,7 +352,7 @@ 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
-  %0 = "tosa.avg_pool2d"(%arg0) {pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<?x6x34x62xf32>)  -> (tensor<?x5x33x62xf32>)
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = f32, pad = array<i64: 1, 1, 1, 1>, kernel = array<i64: 4, 4>, stride = array<i64: 1, 1>} : (tensor<?x6x34x62xf32>)  -> (tensor<?x5x33x62xf32>)
   return %0 : tensor<?x5x33x62xf32>
 }
 
diff --git a/mlir/test/Dialect/Tosa/ops.mlir b/mlir/test/Dialect/Tosa/ops.mlir
index 68eca320aa46..bf3cf3d7084b 100644
--- a/mlir/test/Dialect/Tosa/ops.mlir
+++ b/mlir/test/Dialect/Tosa/ops.mlir
@@ -12,28 +12,28 @@ func.func @test_argmax(%arg0: tensor<14x19xf32>) -> tensor<14xi32> {
 // -----
 // CHECK-LABEL: avg_pool2d_f32
 func.func @test_avg_pool2d_f32(%arg0: tensor<1x7x7x9xf32>) -> tensor<1x7x7x9xf32> {
-    %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9xf32>) -> tensor<1x7x7x9xf32>
+    %0 = "tosa.avg_pool2d"(%arg0) {acc_type = f32, kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9xf32>) -> tensor<1x7x7x9xf32>
     return %0 : tensor<1x7x7x9xf32>
 }
 
 // -----
 // CHECK-LABEL: avg_pool2d_i8
 func.func @test_avg_pool2d_i8(%arg0: tensor<1x7x7x9xi8>) -> tensor<1x7x7x9xi8> {
-    %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9xi8>) -> tensor<1x7x7x9xi8>
+    %0 = "tosa.avg_pool2d"(%arg0) {acc_type = i32, kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9xi8>) -> tensor<1x7x7x9xi8>
     return %0 : tensor<1x7x7x9xi8>
 }
 
 // -----
 // CHECK-LABEL: avg_pool2d_i16
 func.func @test_avg_pool2d_i16(%arg0: tensor<1x7x7x9xi16>) -> tensor<1x7x7x9xi16> {
-    %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9xi16>) -> tensor<1x7x7x9xi16>
+    %0 = "tosa.avg_pool2d"(%arg0) {acc_type = i32, kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9xi16>) -> tensor<1x7x7x9xi16>
     return %0 : tensor<1x7x7x9xi16>
 }
 
 // -----
 // CHECK-LABEL: avg_pool2d_q8
 func.func @test_avg_pool2d_q8(%arg0: tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>) -> tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>> {
-    %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>) -> tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>
+    %0 = "tosa.avg_pool2d"(%arg0) {acc_type = i32, kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>) -> tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>
     return %0 : tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>
 }
 
diff --git a/mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir b/mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir
index ea2bd42106a2..56b088784cbf 100644
--- a/mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir
+++ b/mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir
@@ -659,7 +659,7 @@ func.func @scatter_minimum_static(%arg0 : tensor<?x4x?xi32>, %arg1 : tensor<3x?x
 // CHECK-LABEL: @test_pool_static
 func.func @test_pool_static(%arg0: tensor<3x5x6x7xf32>) {
   // CHECK: -> tensor<3x2x4x7xf32>
-  %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>
 
   // CHECK: -> tensor<3x2x4x7xf32>
   %1 = "tosa.max_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>
@@ -689,7 +689,7 @@ func.func @conv2d_dynamic_input(%input: tensor<?x?x?x?xf32>, %weights: tensor<5x
 // CHECK-LABEL: @test_pool_dynamic_input
 func.func @test_pool_dynamic_input(%arg0: tensor<?x?x?x?xf32>) {
   // CHECK: -> tensor<?x?x?x?xf32>
-  %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
 
   // CHECK: -> tensor<?x?x?x?xf32>
   %1 = "tosa.max_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
@@ -701,7 +701,7 @@ func.func @test_pool_dynamic_input(%arg0: tensor<?x?x?x?xf32>) {
 // CHECK-LABEL: @test_pool_padded
 func.func @test_pool_padded(%arg0: tensor<3x5x6x7xf32>) {
   // CHECK: -> tensor<3x5x11x7xf32>
-  %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 1, 2, 3, 4>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 1, 2, 3, 4>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>
 
   // CHECK: -> tensor<3x5x11x7xf32>
   %1 = "tosa.max_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 1, 2, 3, 4>, stride = array<i64: 1, 1>} : (tensor<3x5x6x7xf32>) -> tensor<?x?x?x?xf32>
@@ -731,7 +731,7 @@ func.func @conv2d_dynamic_bias(%input: tensor<2x8x9x3xf32>, %weights: tensor<5x3
 // CHECK-LABEL: @test_pool_stride
 func.func @test_pool_stride(%arg0: tensor<3x11x12x7xf32>) {
   // CHECK: -> tensor<3x4x4x7xf32>
-  %0 = "tosa.avg_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 3>} : (tensor<3x11x12x7xf32>) -> tensor<?x?x?x?xf32>
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = f32, kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 3>} : (tensor<3x11x12x7xf32>) -> tensor<?x?x?x?xf32>
 
   // CHECK: -> tensor<3x4x4x7xf32>
   %1 = "tosa.max_pool2d"(%arg0) {kernel = array<i64: 4, 3>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 3>} : (tensor<3x11x12x7xf32>) -> tensor<?x?x?x?xf32>