42944da5ba
Group all patterns that re-order vector.transpose and vector.broadcast
Ops (*) under `populateSinkVectorOpsPatterns`. These patterns are
normally used to "sink" redundant Vector Ops, hence grouping together.
Example:
```mlir
%at = vector.transpose %a, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
%bt = vector.transpose %b, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
%r = arith.addf %at, %bt : vector<2x4xf32>
```
would get converted to:
```mlir
%0 = arith.addf %a, %b : vector<4x2xf32>
%r = vector.transpose %0, [1, 0] : vector<2x4xf32>
```
This patch also moves all tests for these patterns so that all of them
are:
* run under one test-flag: `test-vector-sink-patterns`,
* located in one file: "vector-sink.mlir".
To facilitate this change:
* `-test-sink-vector-broadcast` is renamed as
`test-vector-sink-patterns`,
* "sink-vector-broadcast.mlir" is renamed as "vector-sink.mlir",
* tests for `ReorderCastOpsOnBroadcast` and
`ReorderElementwiseOpsOnTranspose` patterns are moved from
"vector-reduce-to-contract.mlir" to "vector-sink.mlir",
* `ReorderElementwiseOpsOnTranspose` patterns are removed from
`populateVectorReductionToContractPatterns` and added to (newly
created) `populateSinkVectorOpsPatterns`,
* `ReorderCastOpsOnBroadcast` patterns are removed from
`populateVectorReductionToContractPatterns` - these are already
present in `populateSinkVectorOpsPatterns`.
This should allow us better layering and more straightforward testing.
For the latter, the goal is to be able to easily identify which pattern
a particular test is exercising (especially when it's a specific
pattern).
NOTES FOR DOWNSTREAM USERS
In order to preserve the current functionality, please make sure to add
* `populateSinkVectorOpsPatterns`,
wherever you are using `populateVectorReductionToContractPatterns`.
Also, rename `populateSinkVectorBroadcastPatterns` as
`populateSinkVectorOpsPatterns`.
(*) I didn't notice any other re-order patterns.
323 lines
15 KiB
MLIR
323 lines
15 KiB
MLIR
// RUN: mlir-opt %s -test-vector-sink-patterns -split-input-file | FileCheck %s
|
|
|
|
//-----------------------------------------------------------------------------
|
|
// [Pattern: ReorderElementwiseOpsOnBroadcast]
|
|
//-----------------------------------------------------------------------------
|
|
|
|
// CHECK-LABEL: func.func @broadcast_scalar_with_bcast(
|
|
// CHECK-SAME: %[[ARG_0:.*]]: index, %[[ARG_1:.*]]: index) -> vector<1x4xindex> {
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[ARG_0]], %[[ARG_1]] : index
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x4xindex>
|
|
// CHECK: return %[[BCAST]] : vector<1x4xindex>
|
|
|
|
func.func @broadcast_scalar_with_bcast(%arg1: index, %arg2: index) -> vector<1x4xindex> {
|
|
%0 = vector.broadcast %arg1 : index to vector<1x4xindex>
|
|
%1 = vector.broadcast %arg2 : index to vector<1x4xindex>
|
|
%2 = arith.addi %0, %1 : vector<1x4xindex>
|
|
return %2 : vector<1x4xindex>
|
|
}
|
|
|
|
// CHECK-LABEL: func.func @broadcast_scalar_with_bcast_scalable(
|
|
// CHECK-SAME: %[[ARG_0:.*]]: index, %[[ARG_1:.*]]: index) -> vector<1x[4]xindex> {
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[ARG_0]], %[[ARG_1]] : index
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x[4]xindex>
|
|
// CHECK: return %[[BCAST]] : vector<1x[4]xindex>
|
|
|
|
func.func @broadcast_scalar_with_bcast_scalable(%arg1: index, %arg2: index) -> vector<1x[4]xindex> {
|
|
%0 = vector.broadcast %arg1 : index to vector<1x[4]xindex>
|
|
%1 = vector.broadcast %arg2 : index to vector<1x[4]xindex>
|
|
%2 = arith.addi %0, %1 : vector<1x[4]xindex>
|
|
return %2 : vector<1x[4]xindex>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func.func @broadcast_scalar_with_bcast_and_splat(
|
|
// CHECK-SAME: %[[ARG1:.*]]: index,
|
|
// CHECK-SAME: %[[ARG2:.*]]: index) -> vector<1x4xindex> {
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[ARG1]], %[[ARG2]] : index
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x4xindex>
|
|
// CHECK: return %[[BCAST]] : vector<1x4xindex>
|
|
func.func @broadcast_scalar_with_bcast_and_splat(%arg1: index, %arg2: index) -> vector<1x4xindex> {
|
|
%0 = vector.splat %arg1 : vector<1x4xindex>
|
|
%1 = vector.broadcast %arg2 : index to vector<1x4xindex>
|
|
%2 = arith.addi %0, %1 : vector<1x4xindex>
|
|
return %2 : vector<1x4xindex>
|
|
}
|
|
|
|
// CHECK-LABEL: func.func @broadcast_scalar_with_bcast_and_splat_scalable(
|
|
// CHECK-SAME: %[[ARG1:.*]]: index,
|
|
// CHECK-SAME: %[[ARG2:.*]]: index) -> vector<1x[4]xindex> {
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[ARG1]], %[[ARG2]] : index
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x[4]xindex>
|
|
// CHECK: return %[[BCAST]] : vector<1x[4]xindex>
|
|
func.func @broadcast_scalar_with_bcast_and_splat_scalable(%arg1: index, %arg2: index) -> vector<1x[4]xindex> {
|
|
%0 = vector.splat %arg1 : vector<1x[4]xindex>
|
|
%1 = vector.broadcast %arg2 : index to vector<1x[4]xindex>
|
|
%2 = arith.addi %0, %1 : vector<1x[4]xindex>
|
|
return %2 : vector<1x[4]xindex>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func.func @broadcast_vector(
|
|
// CHECK-SAME: %[[ARG_0:.*]]: vector<4xf32>,
|
|
// CHECK-SAME: %[[ARG_1:.*]]: vector<4xf32>) -> vector<3x4xf32> {
|
|
// CHECK: %[[ADDF:.*]] = arith.addf %[[ARG_0]], %[[ARG_1]] : vector<4xf32>
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ADDF]] : vector<4xf32> to vector<3x4xf32>
|
|
// CHECK: return %[[BCAST]] : vector<3x4xf32>
|
|
|
|
func.func @broadcast_vector(%arg1: vector<4xf32>, %arg2: vector<4xf32>) -> vector<3x4xf32> {
|
|
%arg1_bcast = vector.broadcast %arg1 : vector<4xf32> to vector<3x4xf32>
|
|
%arg2_bcast = vector.broadcast %arg2 : vector<4xf32> to vector<3x4xf32>
|
|
%2 = arith.addf %arg1_bcast, %arg2_bcast : vector<3x4xf32>
|
|
return %2 : vector<3x4xf32>
|
|
}
|
|
|
|
// CHECK-LABEL: func.func @broadcast_vector_scalable(
|
|
// CHECK-SAME: %[[ARG_0:.*]]: vector<[4]xf32>,
|
|
// CHECK-SAME: %[[ARG_1:.*]]: vector<[4]xf32>) -> vector<3x[4]xf32> {
|
|
// CHECK: %[[ADDF:.*]] = arith.addf %[[ARG_0]], %[[ARG_1]] : vector<[4]xf32>
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ADDF]] : vector<[4]xf32> to vector<3x[4]xf32>
|
|
// CHECK: return %[[BCAST]] : vector<3x[4]xf32>
|
|
|
|
func.func @broadcast_vector_scalable(%arg1: vector<[4]xf32>, %arg2: vector<[4]xf32>) -> vector<3x[4]xf32> {
|
|
%arg1_bcast = vector.broadcast %arg1 : vector<[4]xf32> to vector<3x[4]xf32>
|
|
%arg2_bcast = vector.broadcast %arg2 : vector<[4]xf32> to vector<3x[4]xf32>
|
|
%2 = arith.addf %arg1_bcast, %arg2_bcast : vector<3x[4]xf32>
|
|
return %2 : vector<3x[4]xf32>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func.func @broadcast_scalar_and_vec(
|
|
// CHECK-SAME: %[[ARG1:.*]]: index,
|
|
// CHECK-SAME: %[[ARG2:.*]]: vector<4xindex>) -> vector<1x4xindex> {
|
|
// CHECK: %[[SPLAT:.*]] = vector.splat %[[ARG1]] : vector<1x4xindex>
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ARG2]] : vector<4xindex> to vector<1x4xindex>
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[SPLAT]], %[[BCAST]] : vector<1x4xindex>
|
|
// CHECK: return %[[ADD]] : vector<1x4xindex>
|
|
func.func @broadcast_scalar_and_vec(%arg1: index, %arg2: vector<4xindex>) -> vector<1x4xindex> {
|
|
%0 = vector.splat %arg1 : vector<1x4xindex>
|
|
%1 = vector.broadcast %arg2 : vector<4xindex> to vector<1x4xindex>
|
|
%2 = arith.addi %0, %1 : vector<1x4xindex>
|
|
return %2 : vector<1x4xindex>
|
|
}
|
|
|
|
// CHECK-LABEL: func.func @broadcast_scalar_and_vec_scalable(
|
|
// CHECK-SAME: %[[ARG1:.*]]: index,
|
|
// CHECK-SAME: %[[ARG2:.*]]: vector<[4]xindex>) -> vector<1x[4]xindex> {
|
|
// CHECK: %[[SPLAT:.*]] = vector.splat %[[ARG1]] : vector<1x[4]xindex>
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ARG2]] : vector<[4]xindex> to vector<1x[4]xindex>
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[SPLAT]], %[[BCAST]] : vector<1x[4]xindex>
|
|
// CHECK: return %[[ADD]] : vector<1x[4]xindex>
|
|
func.func @broadcast_scalar_and_vec_scalable(%arg1: index, %arg2: vector<[4]xindex>) -> vector<1x[4]xindex> {
|
|
%0 = vector.splat %arg1 : vector<1x[4]xindex>
|
|
%1 = vector.broadcast %arg2 : vector<[4]xindex> to vector<1x[4]xindex>
|
|
%2 = arith.addi %0, %1 : vector<1x[4]xindex>
|
|
return %2 : vector<1x[4]xindex>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func.func @broadcast_vector_and_scalar(
|
|
// CHECK-SAME: %[[ARG_0:.*]]: i32,
|
|
// CHECK-SAME: %[[ARG_1:.*]]: vector<4xi32>) -> vector<4xi32> {
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ARG_0]] : i32 to vector<4xi32>
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[BCAST]], %[[ARG_1]] : vector<4xi32>
|
|
// CHECK: return %[[ADD]] : vector<4xi32>
|
|
|
|
func.func @broadcast_vector_and_scalar(%arg1: i32, %arg2: vector<4xi32>) -> vector<4xi32> {
|
|
%arg1_bcast = vector.broadcast %arg1 : i32 to vector<4xi32>
|
|
%2 = arith.addi %arg1_bcast, %arg2 : vector<4xi32>
|
|
return %2 : vector<4xi32>
|
|
}
|
|
|
|
// CHECK-LABEL: func.func @broadcast_vector_and_scalar_scalable(
|
|
// CHECK-SAME: %[[ARG_0:.*]]: i32,
|
|
// CHECK-SAME: %[[ARG_1:.*]]: vector<[4]xi32>) -> vector<[4]xi32> {
|
|
// CHECK: %[[BCAST:.*]] = vector.broadcast %[[ARG_0]] : i32 to vector<[4]xi32>
|
|
// CHECK: %[[ADD:.*]] = arith.addi %[[BCAST]], %[[ARG_1]] : vector<[4]xi32>
|
|
// CHECK: return %[[ADD]] : vector<[4]xi32>
|
|
|
|
func.func @broadcast_vector_and_scalar_scalable(%arg1: i32, %arg2: vector<[4]xi32>) -> vector<[4]xi32> {
|
|
%arg1_bcast = vector.broadcast %arg1 : i32 to vector<[4]xi32>
|
|
%2 = arith.addi %arg1_bcast, %arg2 : vector<[4]xi32>
|
|
return %2 : vector<[4]xi32>
|
|
}
|
|
|
|
// -----
|
|
|
|
#matmat_accesses = [
|
|
affine_map<(i, j, k) -> (i, k)>,
|
|
affine_map<(i, j, k) -> (k, j)>,
|
|
affine_map<(i, j, k) -> (i, j)>
|
|
]
|
|
#matmat_trait = {
|
|
indexing_maps = #matmat_accesses,
|
|
iterator_types = ["parallel", "parallel", "reduction"]
|
|
}
|
|
|
|
// CHECK-LABEL: func.func @negative_not_elementwise
|
|
// CHECK-DAG: %[[F1:.*]] = arith.constant dense<1.000000e+00> : vector<2x2xf32>
|
|
// CHECK-DAG: %[[F2:.*]] = arith.constant dense<2.000000e+00> : vector<2x2xf32>
|
|
// CHECK-DAG: %[[F3:.*]] = arith.constant dense<3.000000e+00> : vector<2x2xf32>
|
|
// CHECK: %[[RES:.*]] = vector.contract {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %[[F1]], %[[F2]], %[[F3]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
|
|
func.func @negative_not_elementwise() -> vector<2x2xf32> {
|
|
%f1 = arith.constant 1.0: f32
|
|
%f2 = arith.constant 2.0: f32
|
|
%f3 = arith.constant 3.0: f32
|
|
|
|
%A = vector.broadcast %f1 : f32 to vector<2x2xf32>
|
|
%B = vector.broadcast %f2 : f32 to vector<2x2xf32>
|
|
%C = vector.broadcast %f3 : f32 to vector<2x2xf32>
|
|
%res = vector.contract #matmat_trait %A, %B, %C
|
|
: vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
|
|
|
|
return %res : vector<2x2xf32>
|
|
}
|
|
|
|
// -----
|
|
|
|
// The source and the result for arith.cmp have different types - not supported
|
|
|
|
// CHECK-LABEL: func.func @negative_source_and_result_mismatch
|
|
// CHECK: %[[BROADCAST:.+]] = vector.broadcast
|
|
// CHECK: %[[RETURN:.+]] = arith.cmpf uno, %[[BROADCAST]], %[[BROADCAST]]
|
|
// CHECK: return %[[RETURN]]
|
|
func.func @negative_source_and_result_mismatch(%arg0 : f32, %arg1 : vector<1xf32>) -> vector<1xi1> {
|
|
%0 = vector.broadcast %arg0 : f32 to vector<1xf32>
|
|
%1 = arith.cmpf uno, %0, %0 : vector<1xf32>
|
|
return %1 : vector<1xi1>
|
|
}
|
|
|
|
// -----
|
|
|
|
// vector.fma only supports vectors - currently it's not possible to replace this with e.g.:
|
|
// %scalar_res = vector.fma %scalar_1, %scalar2
|
|
// %vec_res = vector.broadcast %scalar_res
|
|
//
|
|
// TODO: It should be possible to support this case
|
|
|
|
// CHECK-LABEL: func.func @negative_op_only_supports_vectors
|
|
// CHECK: %[[BROADCAST:.+]] = vector.broadcast
|
|
// CHECK: %[[RESULT:.+]] = vector.fma %[[BROADCAST]]
|
|
// CHECK: return %[[RESULT]]
|
|
func.func @negative_op_only_supports_vectors(%arg0 : f32) -> vector<1xf32> {
|
|
%0 = vector.broadcast %arg0 : f32 to vector<1xf32>
|
|
%1 = vector.fma %0, %0, %0 : vector<1xf32>
|
|
return %1 : vector<1xf32>
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// [Pattern: ReorderCastOpsOnBroadcast]
|
|
//
|
|
// Reorder casting ops and vector ops. The casting ops have almost identical
|
|
// pattern, so only arith.extsi op is tested.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// -----
|
|
|
|
func.func @broadcast_vector_extsi(%a : vector<4xi8>) -> vector<2x4xi32> {
|
|
// CHECK: %[[EXT:.+]] = arith.extsi %{{.+}} : vector<4xi8> to vector<4xi32>
|
|
// CHECK: vector.broadcast %[[EXT:.+]] : vector<4xi32> to vector<2x4xi32>
|
|
%b = vector.broadcast %a : vector<4xi8> to vector<2x4xi8>
|
|
%r = arith.extsi %b : vector<2x4xi8> to vector<2x4xi32>
|
|
return %r : vector<2x4xi32>
|
|
}
|
|
|
|
// -----
|
|
|
|
func.func @broadcast_scalar_extsi(%a : i8) -> vector<2x4xi32> {
|
|
// CHECK: %[[EXT:.+]] = arith.extsi %{{.+}} : i8 to i32
|
|
// CHECK: vector.broadcast %[[EXT]] : i32 to vector<2x4xi32>
|
|
%b = vector.broadcast %a : i8 to vector<2x4xi8>
|
|
%r = arith.extsi %b : vector<2x4xi8> to vector<2x4xi32>
|
|
return %r : vector<2x4xi32>
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// [Pattern: ReorderElementwiseOpsOnTranspose]
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
func.func @transpose_extsi(%a : vector<4x2xi8>) -> vector<2x4xi32> {
|
|
// CHECK: %[[EXT:.+]] = arith.extsi %{{.+}} : vector<4x2xi8> to vector<4x2xi32>
|
|
// CHECK: vector.transpose %[[EXT]], [1, 0] : vector<4x2xi32> to vector<2x4xi32>
|
|
%b = vector.transpose %a, [1, 0]: vector<4x2xi8> to vector<2x4xi8>
|
|
%r = arith.extsi %b : vector<2x4xi8> to vector<2x4xi32>
|
|
return %r : vector<2x4xi32>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func @transpose_elementwise_same_type
|
|
// CHECK-SAME: (%[[A:.+]]: vector<4x2xf32>, %[[B:.+]]: vector<4x2xf32>)
|
|
// CHECK: %[[ADD:.+]] = arith.addf %[[A]], %[[B]] : vector<4x2xf32>
|
|
// CHECK: %[[T:.+]] = vector.transpose %[[ADD]], [1, 0]
|
|
// CHECK: return %[[T]]
|
|
|
|
func.func @transpose_elementwise_same_type(%a : vector<4x2xf32>, %b : vector<4x2xf32>) -> vector<2x4xf32> {
|
|
%at = vector.transpose %a, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
|
|
%bt = vector.transpose %b, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
|
|
%r = arith.addf %at, %bt : vector<2x4xf32>
|
|
return %r : vector<2x4xf32>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func @transpose_elementwise_diff_operand_types
|
|
// CHECK-SAME: (%[[COND:.+]]: vector<4x2xi1>, %[[A:.+]]: vector<4x2xf32>, %[[B:.+]]: vector<4x2xf32>)
|
|
// CHECK: %[[S:.+]] = arith.select %[[COND]], %[[A]], %[[B]] : vector<4x2xi1>, vector<4x2xf32>
|
|
// CHECK: %[[T:.+]] = vector.transpose %[[S]], [1, 0] : vector<4x2xf32> to vector<2x4xf32>
|
|
// CHECK: return %[[T]]
|
|
func.func @transpose_elementwise_diff_operand_types(%cond: vector<4x2xi1>, %a : vector<4x2xf32>, %b : vector<4x2xf32>) -> vector<2x4xf32> {
|
|
%condt = vector.transpose %cond, [1, 0]: vector<4x2xi1> to vector<2x4xi1>
|
|
%at = vector.transpose %a, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
|
|
%bt = vector.transpose %b, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
|
|
%r = arith.select %condt, %at, %bt : vector<2x4xi1>, vector<2x4xf32>
|
|
return %r : vector<2x4xf32>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func @transpose_elementwise_diff_operand_result_type
|
|
// CHECK-SAME: (%[[A:.+]]: vector<4x2xf32>, %[[B:.+]]: vector<4x2xf32>)
|
|
// CHECK: %[[CMP:.+]] = arith.cmpf olt, %[[A]], %[[B]] : vector<4x2xf32>
|
|
// CHECK: %[[T:.+]] = vector.transpose %[[CMP]], [1, 0] : vector<4x2xi1> to vector<2x4xi1>
|
|
// CHECK: return %[[T]]
|
|
func.func @transpose_elementwise_diff_operand_result_type(%a : vector<4x2xf32>, %b : vector<4x2xf32>) -> vector<2x4xi1> {
|
|
%at = vector.transpose %a, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
|
|
%bt = vector.transpose %b, [1, 0]: vector<4x2xf32> to vector<2x4xf32>
|
|
%r = arith.cmpf olt, %at, %bt : vector<2x4xf32>
|
|
return %r : vector<2x4xi1>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func @transpose_elementwise_splat_constant
|
|
// CHECK-SAME: (%[[A:.+]]: vector<4x6x3x2xf32>)
|
|
// CHECK: %[[B:.+]] = arith.constant dense<5.000000e+00> : vector<4x6x3x2xf32>
|
|
// CHECK: %[[ADD:.+]] = arith.addf %[[A]], %[[B]] : vector<4x6x3x2xf32>
|
|
// CHECK: %[[T:.+]] = vector.transpose %[[ADD]], [1, 0, 3, 2] : vector<4x6x3x2xf32> to vector<6x4x2x3xf32>
|
|
// CHECK: return %[[T:.+]] : vector<6x4x2x3xf32>
|
|
|
|
func.func @transpose_elementwise_splat_constant(%a : vector<4x6x3x2xf32>) -> vector<6x4x2x3xf32> {
|
|
%b = arith.constant dense<5.0> : vector<6x4x2x3xf32>
|
|
%at = vector.transpose %a, [1, 0, 3, 2]: vector<4x6x3x2xf32> to vector<6x4x2x3xf32>
|
|
%r = arith.addf %at, %b : vector<6x4x2x3xf32>
|
|
return %r : vector<6x4x2x3xf32>
|
|
}
|
|
|
|
// -----
|
|
|
|
// CHECK-LABEL: func @transpose_elementwise_diff_map
|
|
// CHECK: vector.transpose
|
|
// CHECK: vector.transpose
|
|
// CHECK: arith.addf
|
|
func.func @transpose_elementwise_diff_map(%a : vector<4x6x3x2xf32>, %b: vector<6x2x4x3xf32>) -> vector<6x4x2x3xf32> {
|
|
%at = vector.transpose %a, [1, 0, 3, 2]: vector<4x6x3x2xf32> to vector<6x4x2x3xf32>
|
|
%bt = vector.transpose %b, [0, 2, 1, 3]: vector<6x2x4x3xf32> to vector<6x4x2x3xf32>
|
|
%r = arith.addf %at, %bt : vector<6x4x2x3xf32>
|
|
return %r : vector<6x4x2x3xf32>
|
|
}
|