Skip to content

Commit f11bda7

Browse files
banach-spacebanach-space
authored
[mlir][linalg] Use vector.shuffle to flatten conv filter (#75038)
Updates the vectorisation of 1D depthwise convolution when flattening the channel dimension (introduced in #71918). In particular - how the convolution filter is "flattened". ATM, the vectoriser will use `vector.shape_cast`: ```mlir %b_filter = vector.broadcast %filter : vector<4xf32> to vector<3x2x4xf32> %sc_filter = vector.shape_cast %b_filter : vector<3x2x4xf32> to vector<3x8xf32> ``` This lowering is not ideal - `vector.shape_cast` can be convenient when it's folded away, but that's not happening in this case. Instead, this patch updates the vectoriser to use `vector.shuffle` (the overall result is identical): ```mlir %sh_filter = vector.shuffle %filter, %filter [0, 1, 2, 3, 0, 1, 2, 3] : vector<4xf32>, vector<4xf32> %b_filter = vector.broadcast %sh_filter : vector<8xf32> to vector<3x8xf32> ```
1 parent 0d94882 commit f11bda7

File tree

2 files changed

+61
-54
lines changed

2 files changed

+61
-54
lines changed

‎mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp

Lines changed: 21 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -2910,17 +2910,16 @@ struct Conv1DGenerator
29102910
for (int64_t w = 0; w < wSize; w += wSizeStep) {
29112911
Value lhsVal = lhsVals[linearIndex(kw, w)];
29122912
Value resVal = resVals[w];
2913-
ShapedType filterBCastTy = cast<ShapedType>(resVal.getType());
29142913
if (flatten) {
2915-
// Flatten the input and filter vectors (collapse the channel
2914+
// Flatten the input and output vectors (collapse the channel
29162915
// dimension)
29172916
lhsVal = rewriter.create<vector::ShapeCastOp>(
29182917
loc, lhsCastType, lhsVals[linearIndex(kw, w)]);
29192918
resVal = rewriter.create<vector::ShapeCastOp>(loc, resCastType,
29202919
resVals[w]);
29212920
}
2922-
resVals[w] = depthwiseConv1dSliceAsMulAcc(
2923-
rewriter, loc, lhsVal, rhsVals[kw], resVal, filterBCastTy, flatten);
2921+
resVals[w] = depthwiseConv1dSliceAsMulAcc(rewriter, loc, lhsVal,
2922+
rhsVals[kw], resVal, flatten);
29242923
if (flatten) {
29252924
// Un-flatten the output vector (restore the channel dimension)
29262925
resVals[w] = rewriter.create<vector::ShapeCastOp>(
@@ -2964,20 +2963,32 @@ struct Conv1DGenerator
29642963
/// to MulAcc.
29652964
Value depthwiseConv1dSliceAsMulAcc(RewriterBase &rewriter, Location loc,
29662965
Value lhs, Value rhs, Value res,
2967-
ShapedType bcastTy, bool flatten) {
2966+
bool flatten) {
29682967
auto rhsTy = cast<ShapedType>(rhs.getType());
29692968
auto resTy = cast<ShapedType>(res.getType());
29702969

29712970
// TODO(suderman): Change this to use a vector.ima intrinsic.
29722971
lhs = promote(rewriter, loc, lhs, resTy);
29732972

2974-
rhs = rewriter.create<vector::BroadcastOp>(
2975-
loc, bcastTy.clone(rhsTy.getElementType()), rhs);
29762973
if (flatten) {
2977-
// Flatten the channel dimension
2978-
rhs = rewriter.create<vector::ShapeCastOp>(
2979-
loc, resTy.clone(rhsTy.getElementType()), rhs);
2974+
// There are two options for handling the filter:
2975+
// * shape_cast(broadcast(filter))
2976+
// * broadcast(shuffle(filter))
2977+
// Opt for the option without shape_cast to simplify the codegen.
2978+
auto rhsSize = rhs.getType().cast<VectorType>().getShape()[0];
2979+
auto resSize = res.getType().cast<VectorType>().getShape()[1];
2980+
2981+
SmallVector<int64_t, 16> indicies;
2982+
for (int i = 0; i < resSize / rhsSize; ++i) {
2983+
for (int j = 0; j < rhsSize; ++j)
2984+
indicies.push_back(j);
2985+
}
2986+
2987+
rhs = rewriter.create<vector::ShuffleOp>(loc, rhs, rhs, indicies);
29802988
}
2989+
// Broadcast the filter to match the output vector
2990+
rhs = rewriter.create<vector::BroadcastOp>(
2991+
loc, resTy.clone(rhsTy.getElementType()), rhs);
29812992

29822993
rhs = promote(rewriter, loc, rhs, resTy);
29832994

‎mlir/test/Dialect/Linalg/vectorize-convolution-flatten.mlir

Lines changed: 40 additions & 44 deletions
Original file line numberDiff line numberDiff line change
@@ -36,9 +36,10 @@ module attributes {transform.with_named_sequence} {
3636
/// w == 0, kw = 0
3737
// CHECK: %[[SC_INPUT:.*]] = vector.shape_cast %[[V_INPUT_R]] : vector<1x8x3xi8> to vector<1x24xi8>
3838
// CHECK: %[[SC_OUTPUT:.*]] = vector.shape_cast %[[V_OUTPUT_R]] : vector<1x8x3xi8> to vector<1x24xi8>
39-
// CHECK: %[[B_FILTER:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<3xi8> to vector<1x8x3xi8>
40-
// CHECK: %[[SC_FILTER:.*]] = vector.shape_cast %[[B_FILTER]] : vector<1x8x3xi8> to vector<1x24xi8>
41-
// CHECK: %[[MULI:.*]] = arith.muli %[[SC_INPUT]], %[[SC_FILTER]] : vector<1x24xi8>
39+
// CHECK: %[[SH_FILTER_0:.*]] = vector.shuffle %[[V_FILTER_0]], %[[V_FILTER_0]]
40+
// CHECK-SAME: [0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2, 0, 1, 2] : vector<3xi8>, vector<3xi8>
41+
// CHECK: %[[B_FILTER:.*]] = vector.broadcast %[[SH_FILTER_0]] : vector<24xi8> to vector<1x24xi8>
42+
// CHECK: %[[MULI:.*]] = arith.muli %[[SC_INPUT]], %[[B_FILTER]] : vector<1x24xi8>
4243
// CHECK: %[[ADDI:.*]] = arith.addi %[[MULI]], %[[SC_OUTPUT]] : vector<1x24xi8>
4344

4445
// Write the result back in one shot.
@@ -80,15 +81,17 @@ func.func @depthwise_conv1d_nwc_wc_3x5x4xf32_memref_dillation_2(%input: memref<3
8081
/// w == 0, kw = 0
8182
// CHECK: %[[SC_V_INPUT_0:.*]] = vector.shape_cast %[[V_INPUT_0]] : vector<3x2x4xf32> to vector<3x8xf32>
8283
// CHECK: %[[SC_V_OUTPUT_R:.*]] = vector.shape_cast %[[V_OUTPUT_R]] : vector<3x2x4xf32> to vector<3x8xf32>
83-
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xf32> to vector<3x2x4xf32>
84-
// CHECK: %[[SC_B_FILTER_0:.*]] = vector.shape_cast %[[B_FILTER_0]] : vector<3x2x4xf32> to vector<3x8xf32>
85-
// CHECK: %[[FMA_0:.*]] = vector.fma %[[SC_V_INPUT_0]], %[[SC_B_FILTER_0]], %[[SC_V_OUTPUT_R]] : vector<3x8xf32>
84+
// CHECK: %[[SH_FILTER_0:.*]] = vector.shuffle %[[V_FILTER_0]], %[[V_FILTER_0]]
85+
// CHECK-SAME: [0, 1, 2, 3, 0, 1, 2, 3] : vector<4xf32>, vector<4xf32>
86+
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[SH_FILTER_0]] : vector<8xf32> to vector<3x8xf32>
87+
// CHECK: %[[FMA_0:.*]] = vector.fma %[[SC_V_INPUT_0]], %[[B_FILTER_0]], %[[SC_V_OUTPUT_R]] : vector<3x8xf32>
8688

8789
/// w == 0, kw = 1
8890
// CHECK: %[[SC_V_INPUT_1:.*]] = vector.shape_cast %[[V_INPUT_1]] : vector<3x2x4xf32> to vector<3x8xf32>
89-
// CHECK: %[[B_V_FILTER_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xf32> to vector<3x2x4xf32>
90-
// CHECK: %[[SC_B_FILTER_1:.*]] = vector.shape_cast %[[B_V_FILTER_1]] : vector<3x2x4xf32> to vector<3x8xf32>
91-
// CHECK: %[[FMA_1:.*]] = vector.fma %[[SC_V_INPUT_1]], %[[SC_B_FILTER_1]], %[[FMA_0]] : vector<3x8xf32>
91+
// CHECK: %[[SH_FILTER_1:.*]] = vector.shuffle %[[V_FILTER_1]], %[[V_FILTER_1]]
92+
// CHECK-SAME: [0, 1, 2, 3, 0, 1, 2, 3] : vector<4xf32>, vector<4xf32>
93+
// CHECK: %[[B_FILTER_1:.*]] = vector.broadcast %[[SH_FILTER_1]] : vector<8xf32> to vector<3x8xf32>
94+
// CHECK: %[[FMA_1:.*]] = vector.fma %[[SC_V_INPUT_1]], %[[B_FILTER_1]], %[[FMA_0]] : vector<3x8xf32>
9295

9396
// Write the result back in one shot.
9497
// CHECK: %[[SC_FMA_1:.*]] = vector.shape_cast %[[FMA_1]] : vector<3x8xf32> to vector<3x2x4xf32>
@@ -138,19 +141,21 @@ func.func @depthwise_conv1d_nwc_wc_3x5x4xi8_memref_dilation_2(%input: memref<3x5
138141
// CHECK: %[[SC_V_INPUT_0:.*]] = vector.shape_cast %[[V_INPUT_0]] : vector<3x2x4xi8> to vector<3x8xi8>
139142
// CHECK: %[[SC_V_OUTPUT_R:.*]] = vector.shape_cast %[[V_OUTPUT_R]] : vector<3x2x4xi32> to vector<3x8xi32>
140143
// CHECK: %[[EXT_INPUT_0:.*]] = arith.extsi %[[SC_V_INPUT_0]] : vector<3x8xi8> to vector<3x8xi32>
141-
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xi8> to vector<3x2x4xi8>
142-
// CHECK: %[[SC_B_FILTER_0:.*]] = vector.shape_cast %[[B_FILTER_0]] : vector<3x2x4xi8> to vector<3x8xi8>
143-
// CHECK: %[[EXT_FILTER_0:.*]] = arith.extsi %[[SC_B_FILTER_0]] : vector<3x8xi8> to vector<3x8xi32>
144-
// CHECK: %[[MUL_0:.*]] = arith.muli %[[EXT_INPUT_0]], %[[EXT_FILTER_0]] : vector<3x8xi32>
144+
// CHECK: %[[SH_FILTER_0:.*]] = vector.shuffle %[[V_FILTER_0]], %[[V_FILTER_0]]
145+
// CHECK-SAME: [0, 1, 2, 3, 0, 1, 2, 3] : vector<4xi8>, vector<4xi8>
146+
// CHECK: %[[EXT_FILTER_0:.*]] = arith.extsi %[[SH_FILTER_0]] : vector<8xi8> to vector<8xi32>
147+
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[EXT_FILTER_0]] : vector<8xi32> to vector<3x8xi32>
148+
// CHECK: %[[MUL_0:.*]] = arith.muli %[[EXT_INPUT_0]], %[[B_FILTER_0]] : vector<3x8xi32>
145149
// CHECK: %[[ADD_0:.*]] = arith.addi %[[MUL_0]], %[[SC_V_OUTPUT_R]] : vector<3x8xi32>
146150

147151
/// w == 0, kw = 1
148152
// CHECK: %[[SC_V_INPUT_1:.*]] = vector.shape_cast %[[V_INPUT_1]] : vector<3x2x4xi8> to vector<3x8xi8>
149153
// CHECK: %[[EXT_INPUT_1:.*]] = arith.extsi %[[SC_V_INPUT_1]] : vector<3x8xi8> to vector<3x8xi32>
150-
// CHECK: %[[B_FILTER_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xi8> to vector<3x2x4xi8>
151-
// CHECK: %[[SC_B_FILTER_1:.*]] = vector.shape_cast %[[B_FILTER_1]] : vector<3x2x4xi8> to vector<3x8xi8>
152-
// CHECK: %[[EXT_FILTER_1:.*]] = arith.extsi %[[SC_B_FILTER_1]] : vector<3x8xi8> to vector<3x8xi32>
153-
// CHECK: %[[MUL_1:.*]] = arith.muli %[[EXT_INPUT_1]], %[[EXT_FILTER_1]] : vector<3x8xi32>
154+
// CHECK: %[[SH_FILTER_1:.*]] = vector.shuffle %[[V_FILTER_1]], %[[V_FILTER_1]]
155+
// CHECK-SAME: [0, 1, 2, 3, 0, 1, 2, 3] : vector<4xi8>, vector<4xi8>
156+
// CHECK: %[[EXT_FILTER_1:.*]] = arith.extsi %[[SH_FILTER_1]] : vector<8xi8> to vector<8xi32>
157+
// CHECK: %[[B_FILTER_1:.*]] = vector.broadcast %[[EXT_FILTER_1]] : vector<8xi32> to vector<3x8xi32>
158+
// CHECK: %[[MUL_1:.*]] = arith.muli %[[EXT_INPUT_1]], %[[B_FILTER_1]] : vector<3x8xi32>
154159
// CHECK: %[[ADD_1:.*]] = arith.addi %[[MUL_1]], %[[ADD_0]] : vector<3x8xi32>
155160

156161
// Write the result back in one shot.
@@ -223,69 +228,60 @@ func.func @depthwise_conv1d_nwc_wc_3x9x4xi8_tensor_stride_2(%input: tensor<3x9x4
223228
/// w == 0, kw == 0
224229
// CHECK: %[[VAL_23:.*]] = vector.shape_cast %[[V_INPUT_0]] : vector<3x1x4xi8> to vector<3x4xi8>
225230
// CHECK: %[[VAL_24:.*]] = vector.shape_cast %[[V_OUTPUT_0]] : vector<3x1x4xi8> to vector<3x4xi8>
226-
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xi8> to vector<3x1x4xi8>
227-
// CHECK: %[[VAL_26:.*]] = vector.shape_cast %[[B_FILTER_0]] : vector<3x1x4xi8> to vector<3x4xi8>
228-
// CHECK: %[[VAL_27:.*]] = arith.muli %[[VAL_23]], %[[VAL_26]] : vector<3x4xi8>
231+
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xi8> to vector<3x4xi8>
232+
// CHECK: %[[VAL_27:.*]] = arith.muli %[[VAL_23]], %[[B_FILTER_0]] : vector<3x4xi8>
229233
// CHECK: %[[VAL_28:.*]] = arith.addi %[[VAL_27]], %[[VAL_24]] : vector<3x4xi8>
230234

231235
/// w == 1, kw == 0
232236
// CHECK: %[[VAL_29:.*]] = vector.shape_cast %[[V_INPUT_1]] : vector<3x1x4xi8> to vector<3x4xi8>
233237
// CHECK: %[[VAL_30:.*]] = vector.shape_cast %[[V_OUTPUT_1]] : vector<3x1x4xi8> to vector<3x4xi8>
234-
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xi8> to vector<3x1x4xi8>
235-
// CHECK: %[[VAL_32:.*]] = vector.shape_cast %[[B_FILTER_0]] : vector<3x1x4xi8> to vector<3x4xi8>
236-
// CHECK: %[[VAL_33:.*]] = arith.muli %[[VAL_29]], %[[VAL_32]] : vector<3x4xi8>
238+
// CHECK: %[[B_FILTER_0_1:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xi8> to vector<3x4xi8>
239+
// CHECK: %[[VAL_33:.*]] = arith.muli %[[VAL_29]], %[[B_FILTER_0_1]] : vector<3x4xi8>
237240
// CHECK: %[[VAL_34:.*]] = arith.addi %[[VAL_33]], %[[VAL_30]] : vector<3x4xi8>
238241

239242
/// w == 2, kw == 0
240243
// CHECK: %[[VAL_35:.*]] = vector.shape_cast %[[V_INPUT_2]] : vector<3x1x4xi8> to vector<3x4xi8>
241244
// CHECK: %[[VAL_36:.*]] = vector.shape_cast %[[V_OUTPUT_2]] : vector<3x1x4xi8> to vector<3x4xi8>
242-
// CHECK: %[[B_FILTER_0:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xi8> to vector<3x1x4xi8>
243-
// CHECK: %[[VAL_38:.*]] = vector.shape_cast %[[B_FILTER_0]] : vector<3x1x4xi8> to vector<3x4xi8>
244-
// CHECK: %[[VAL_39:.*]] = arith.muli %[[VAL_35]], %[[VAL_38]] : vector<3x4xi8>
245+
// CHECK: %[[B_FILTER_0_2:.*]] = vector.broadcast %[[V_FILTER_0]] : vector<4xi8> to vector<3x4xi8>
246+
// CHECK: %[[VAL_39:.*]] = arith.muli %[[VAL_35]], %[[B_FILTER_0_2]] : vector<3x4xi8>
245247
// CHECK: %[[VAL_40:.*]] = arith.addi %[[VAL_39]], %[[VAL_36]] : vector<3x4xi8>
246248

247249
/// w == 3, kw == 1
248250
// CHECK: %[[VAL_41:.*]] = vector.shape_cast %[[V_INPUT_3]] : vector<3x1x4xi8> to vector<3x4xi8>
249-
// CHECK: %[[B_FILTER_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xi8> to vector<3x1x4xi8>
250-
// CHECK: %[[VAL_43:.*]] = vector.shape_cast %[[B_FILTER_1]] : vector<3x1x4xi8> to vector<3x4xi8>
251-
// CHECK: %[[VAL_44:.*]] = arith.muli %[[VAL_41]], %[[VAL_43]] : vector<3x4xi8>
251+
// CHECK: %[[B_FILTER_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xi8> to vector<3x4xi8>
252+
// CHECK: %[[VAL_44:.*]] = arith.muli %[[VAL_41]], %[[B_FILTER_1]] : vector<3x4xi8>
252253
// CHECK: %[[VAL_45:.*]] = arith.addi %[[VAL_44]], %[[VAL_28]] : vector<3x4xi8>
253254

254255
/// w == 4, kw == 1
255256
// CHECK: %[[VAL_46:.*]] = vector.shape_cast %[[V_INPUT_4]] : vector<3x1x4xi8> to vector<3x4xi8>
256-
// CHECK: %[[B_FILTER_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xi8> to vector<3x1x4xi8>
257-
// CHECK: %[[VAL_48:.*]] = vector.shape_cast %[[B_FILTER_1]] : vector<3x1x4xi8> to vector<3x4xi8>
258-
// CHECK: %[[VAL_49:.*]] = arith.muli %[[VAL_46]], %[[VAL_48]] : vector<3x4xi8>
257+
// CHECK: %[[B_FILTER_1_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xi8> to vector<3x4xi8>
258+
// CHECK: %[[VAL_49:.*]] = arith.muli %[[VAL_46]], %[[B_FILTER_1_1]] : vector<3x4xi8>
259259
// CHECK: %[[VAL_50:.*]] = arith.addi %[[VAL_49]], %[[VAL_34]] : vector<3x4xi8>
260260

261261
/// w == 5, kw == 1
262262
// CHECK: %[[VAL_51:.*]] = vector.shape_cast %[[V_INPUT_5]] : vector<3x1x4xi8> to vector<3x4xi8>
263-
// CHECK: %[[B_FILTER_1:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xi8> to vector<3x1x4xi8>
264-
// CHECK: %[[VAL_53:.*]] = vector.shape_cast %[[B_FILTER_1]] : vector<3x1x4xi8> to vector<3x4xi8>
265-
// CHECK: %[[VAL_54:.*]] = arith.muli %[[VAL_51]], %[[VAL_53]] : vector<3x4xi8>
263+
// CHECK: %[[B_FILTER_1_2:.*]] = vector.broadcast %[[V_FILTER_1]] : vector<4xi8> to vector<3x4xi8>
264+
// CHECK: %[[VAL_54:.*]] = arith.muli %[[VAL_51]], %[[B_FILTER_1_2]] : vector<3x4xi8>
266265
// CHECK: %[[VAL_55:.*]] = arith.addi %[[VAL_54]], %[[VAL_40]] : vector<3x4xi8>
267266

268267
/// w == 6, kw == 2
269268
// CHECK: %[[VAL_56:.*]] = vector.shape_cast %[[V_INPUT_6]] : vector<3x1x4xi8> to vector<3x4xi8>
270-
// CHECK: %[[B_FILTER_2:.*]] = vector.broadcast %[[V_FILTER_2]] : vector<4xi8> to vector<3x1x4xi8>
271-
// CHECK: %[[VAL_58:.*]] = vector.shape_cast %[[B_FILTER_2]] : vector<3x1x4xi8> to vector<3x4xi8>
272-
// CHECK: %[[VAL_59:.*]] = arith.muli %[[VAL_56]], %[[VAL_58]] : vector<3x4xi8>
269+
// CHECK: %[[B_FILTER_2:.*]] = vector.broadcast %[[V_FILTER_2]] : vector<4xi8> to vector<3x4xi8>
270+
// CHECK: %[[VAL_59:.*]] = arith.muli %[[VAL_56]], %[[B_FILTER_2]] : vector<3x4xi8>
273271
// CHECK: %[[VAL_60:.*]] = arith.addi %[[VAL_59]], %[[VAL_45]] : vector<3x4xi8>
274272

275273
/// w == 7, kw == 2
276274
// CHECK: %[[VAL_61:.*]] = vector.shape_cast %[[VAL_60]] : vector<3x4xi8> to vector<3x1x4xi8>
277275
// CHECK: %[[VAL_62:.*]] = vector.shape_cast %[[V_INPUT_7]] : vector<3x1x4xi8> to vector<3x4xi8>
278-
// CHECK: %[[B_FILTER_2:.*]] = vector.broadcast %[[V_FILTER_2]] : vector<4xi8> to vector<3x1x4xi8>
279-
// CHECK: %[[VAL_64:.*]] = vector.shape_cast %[[B_FILTER_2]] : vector<3x1x4xi8> to vector<3x4xi8>
280-
// CHECK: %[[VAL_65:.*]] = arith.muli %[[VAL_62]], %[[VAL_64]] : vector<3x4xi8>
276+
// CHECK: %[[B_FILTER_2_1:.*]] = vector.broadcast %[[V_FILTER_2]] : vector<4xi8> to vector<3x4xi8>
277+
// CHECK: %[[VAL_65:.*]] = arith.muli %[[VAL_62]], %[[B_FILTER_2_1]] : vector<3x4xi8>
281278
// CHECK: %[[VAL_66:.*]] = arith.addi %[[VAL_65]], %[[VAL_50]] : vector<3x4xi8>
282279

283280
/// w == 8, kw == 2
284281
// CHECK: %[[VAL_67:.*]] = vector.shape_cast %[[VAL_66]] : vector<3x4xi8> to vector<3x1x4xi8>
285282
// CHECK: %[[VAL_68:.*]] = vector.shape_cast %[[V_INPUT_8]] : vector<3x1x4xi8> to vector<3x4xi8>
286-
// CHECK: %[[B_FILTER_2:.*]] = vector.broadcast %[[V_FILTER_2]] : vector<4xi8> to vector<3x1x4xi8>
287-
// CHECK: %[[VAL_70:.*]] = vector.shape_cast %[[B_FILTER_2]] : vector<3x1x4xi8> to vector<3x4xi8>
288-
// CHECK: %[[VAL_71:.*]] = arith.muli %[[VAL_68]], %[[VAL_70]] : vector<3x4xi8>
283+
// CHECK: %[[B_FILTER_2_2:.*]] = vector.broadcast %[[V_FILTER_2]] : vector<4xi8> to vector<3x4xi8>
284+
// CHECK: %[[VAL_71:.*]] = arith.muli %[[VAL_68]], %[[B_FILTER_2_2]] : vector<3x4xi8>
289285
// CHECK: %[[VAL_72:.*]] = arith.addi %[[VAL_71]], %[[VAL_55]] : vector<3x4xi8>
290286

291287
// Write the result back.

0 commit comments

Comments
 (0)