@@ -83,7 +83,7 @@ struct matmul_kernel_fwd_t {
83
83
}
84
84
85
85
static void store_vec_helper (
86
- out_memory_tensor_t &output, Vec data, int offset) {
86
+ inout_memory_tensor_t &output, Vec data, int offset) {
87
87
data_type_t type = output.md ().data_type ();
88
88
char *offset_ptr = static_cast <char *>(output.ptr ())
89
89
+ data_type_size (type) * offset;
@@ -189,7 +189,7 @@ struct matmul_kernel_fwd_t {
189
189
}
190
190
}
191
191
192
- void store (out_memory_tensor_t &output, int offset, int row_stride) {
192
+ void store (inout_memory_tensor_t &output, int offset, int row_stride) {
193
193
for (int row = 0 ; row < Rows; row++) {
194
194
for (int col = 0 ; col < Cols / vec_len; col++) {
195
195
store_vec_helper (output, data[row][col],
@@ -198,7 +198,7 @@ struct matmul_kernel_fwd_t {
198
198
}
199
199
}
200
200
201
- void store_edge (out_memory_tensor_t &output, int offset, int row_stride,
201
+ void store_edge (inout_memory_tensor_t &output, int offset, int row_stride,
202
202
int rows, int cols) {
203
203
for (int row = 0 ; row < rows; row++) {
204
204
int col;
@@ -215,7 +215,7 @@ struct matmul_kernel_fwd_t {
215
215
}
216
216
}
217
217
218
- void store_generic (out_memory_tensor_t &output, int offset,
218
+ void store_generic (inout_memory_tensor_t &output, int offset,
219
219
int row_stride, bool transpose, bool is_edge_block, int rows,
220
220
int cols) {
221
221
if (is_edge_block) {
@@ -361,7 +361,7 @@ struct matmul_kernel_fwd_t {
361
361
, data_(CTX_IN_SYCL_KERNEL_MEMORY(DNNL_ARG_SRC_0))
362
362
, weights_(CTX_IN_SYCL_KERNEL_MEMORY(DNNL_ARG_WEIGHTS))
363
363
, bias_(CTX_IN_SYCL_KERNEL_MEMORY(DNNL_ARG_BIAS))
364
- , dst_(CTX_OUT_SYCL_KERNEL_MEMORY (DNNL_ARG_DST))
364
+ , dst_(CTX_INOUT_SYCL_KERNEL_MEMORY (DNNL_ARG_DST))
365
365
, data_scale_(CTX_IN_SYCL_KERNEL_MEMORY(
366
366
DNNL_ARG_ATTR_SCALES | DNNL_ARG_SRC_0))
367
367
, data_scales_dt_((conf_.do_scale_data)
@@ -653,7 +653,7 @@ struct matmul_kernel_fwd_t {
653
653
xpu::sycl::in_memory_arg_t data_;
654
654
xpu::sycl::in_memory_arg_t weights_;
655
655
xpu::sycl::in_memory_arg_t bias_;
656
- xpu::sycl::out_memory_arg_t dst_;
656
+ xpu::sycl::inout_memory_arg_t dst_;
657
657
xpu::sycl::in_memory_arg_t data_scale_;
658
658
data_type_t data_scales_dt_;
659
659
xpu::sycl::in_memory_arg_t weights_scale_;
0 commit comments