Skip to content

Commit 43fe07c

Browse files
abhilash1910abhilash1910
authored
ggml-sycl: Replace 3d ops with macro (#5458)
* use macro * use macro * fix format
1 parent 4a46d2b commit 43fe07c

File tree

1 file changed

+14
-61
lines changed

1 file changed

+14
-61
lines changed

‎ggml-sycl.cpp

Lines changed: 14 additions & 61 deletions
Original file line numberDiff line numberDiff line change
@@ -11578,11 +11578,8 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
1157811578
}
1157911579
char * dst_ptr = (char *) dst;
1158011580

11581-
const int64_t ne0 = src->ne[0];
11582-
const int64_t nb0 = src->nb[0];
11583-
const int64_t nb1 = src->nb[1];
11584-
const int64_t nb2 = src->nb[2];
11585-
const int64_t nb3 = src->nb[3];
11581+
GGML_TENSOR_LOCALS_1(int64_t, ne, src, ne);
11582+
GGML_TENSOR_LOCALS(int64_t, nb, src, nb);
1158611583
const enum ggml_type type = src->type;
1158711584
const int64_t ts = ggml_type_size(type);
1158811585
const int64_t bs = ggml_blck_size(type);
@@ -12426,9 +12423,7 @@ inline void ggml_sycl_op_alibi(const ggml_tensor *src0, const ggml_tensor *src1,
1242612423
GGML_ASSERT(src0->type == GGML_TYPE_F32);
1242712424
GGML_ASSERT( dst->type == GGML_TYPE_F32);
1242812425

12429-
const int64_t ne00 = src0->ne[0];
12430-
const int64_t ne01 = src0->ne[1];
12431-
const int64_t ne02 = src0->ne[2];
12426+
GGML_TENSOR_LOCALS_3(int64_t, ne0, src0, ne);
1243212427
const int64_t nrows = ggml_nrows(src0);
1243312428

1243412429
//const int n_past = ((int32_t *) dst->op_params)[0];
@@ -12758,15 +12753,9 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
1275812753
ggml_sycl_op_mul_mat_t op,
1275912754
const bool convert_src1_to_q8_1) try {
1276012755

12761-
const int64_t ne00 = src0->ne[0];
12762-
const int64_t ne01 = src0->ne[1];
12763-
const int64_t ne02 = src0->ne[2];
12764-
const int64_t ne03 = src0->ne[3];
12756+
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
1276512757

12766-
const int64_t ne10 = src1->ne[0];
12767-
const int64_t ne11 = src1->ne[1];
12768-
const int64_t ne12 = src1->ne[2];
12769-
const int64_t ne13 = src1->ne[3];
12758+
GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
1277012759
const int64_t nrows1 = ggml_nrows(src1);
1277112760

1277212761
GGML_ASSERT(ne03 == ne13);
@@ -13337,23 +13326,13 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0,
1333713326
GGML_ASSERT(src0->type == GGML_TYPE_F16);
1333813327
GGML_ASSERT(src1->type == GGML_TYPE_F32);
1333913328

13340-
const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00);
13341-
const int64_t ne01 = src0->ne[1];
13342-
const int64_t ne02 = src0->ne[2];
13343-
const int64_t ne03 = src0->ne[3];
13329+
GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
1334413330

13345-
const int64_t nb01 = src0->nb[1];
13346-
const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02);
13347-
const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03);
13331+
GGML_TENSOR_LOCALS(int64_t, nb0, src0, nb);
1334813332

13349-
const int64_t ne10 = src1->ne[0];
13350-
const int64_t ne11 = src1->ne[1];
13351-
const int64_t ne12 = src1->ne[2];
13352-
const int64_t ne13 = src1->ne[3];
13333+
GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
1335313334

13354-
const int64_t nb11 = src1->nb[1];
13355-
const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
13356-
const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
13335+
GGML_TENSOR_LOCALS(int64_t, nb1, src1, nb);
1335713336

1335813337
const int64_t ne1 = ggml_nelements(src1);
1335913338
const int64_t ne = ggml_nelements(dst);
@@ -13655,23 +13634,15 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
1365513634
GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
1365613635
GGML_ASSERT(src1->type == GGML_TYPE_F32);
1365713636

13658-
const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
13659-
const int64_t ne01 = src00->ne[1];
13660-
const int64_t ne02 = src00->ne[2];
13661-
const int64_t ne03 = src00->ne[3];
13637+
GGML_TENSOR_LOCALS(int64_t, ne0, src00, ne);
1366213638

1366313639
//const int64_t nb01 = src00->nb[1];
13664-
const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02);
13665-
const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03);
13640+
GGML_TENSOR_LOCALS(int64_t, nb0, src00, nb);
1366613641

13667-
const int64_t ne10 = src1->ne[0];
13668-
const int64_t ne11 = src1->ne[1];
13669-
const int64_t ne12 = src1->ne[2];
13670-
const int64_t ne13 = src1->ne[3];
13642+
GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
1367113643

13644+
GGML_TENSOR_LOCALS(int64_t, nb1, src1, nb);
1367213645
//const int64_t nb11 = src1->nb[1];
13673-
const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
13674-
const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
1367513646

1367613647
const int64_t ne1 = ggml_nelements(src1);
1367713648
const int64_t ne = ggml_nelements(dst);
@@ -13940,25 +13911,7 @@ static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1,
1394013911
GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
1394113912
GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
1394213913

13943-
const int64_t ne00 = src0->ne[0];
13944-
const int64_t ne01 = src0->ne[1];
13945-
const int64_t ne02 = src0->ne[2];
13946-
13947-
13948-
const int64_t nb00 = src0->nb[0];
13949-
const int64_t nb01 = src0->nb[1];
13950-
const int64_t nb02 = src0->nb[2];
13951-
const int64_t nb03 = src0->nb[3];
13952-
13953-
const int64_t ne10 = src1->ne[0];
13954-
const int64_t ne11 = src1->ne[1];
13955-
const int64_t ne12 = src1->ne[2];
13956-
13957-
13958-
const int64_t nb10 = src1->nb[0];
13959-
const int64_t nb11 = src1->nb[1];
13960-
const int64_t nb12 = src1->nb[2];
13961-
const int64_t nb13 = src1->nb[3];
13914+
GGML_TENSOR_BINARY_OP_LOCALS;
1396213915

1396313916
SYCL_CHECK(ggml_sycl_set_device(g_main_device));
1396413917
dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0];

0 commit comments

Comments
 (0)