sycl : Support DSv4 OPs: LIGHTNING_INDEXER,DSV4_HC_COMB,DSV4_HC_POST,DSV4_HC_PRE (#26568)

* support DSv4 OPs: LIGHTNING_INDEXER,DSV4_HC_COMB,DSV4_HC_POST,DSV4_HC_PREwq

* update ops.md

* fix format issue
This commit is contained in:
Neo Zhang
2026-08-07 08:22:23 +03:00
committed by GitHub
parent c1f4109898
commit 31558dbb76
7 changed files with 23406 additions and 677 deletions
+35
View File
@@ -62,6 +62,8 @@
#include "ggml-sycl/repeat_back.hpp"
#include "ggml-sycl/set_rows.hpp"
#include "ggml-sycl/set.hpp"
#include "ggml-sycl/dsv4-hc.hpp"
#include "ggml-sycl/lightning-indexer.hpp"
#include "ggml-sycl/conv2d.hpp"
#include "ggml-sycl/conv2d-dw.hpp"
#include "ggml-sycl/conv2d-transpose.hpp"
@@ -4942,6 +4944,18 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
case GGML_OP_SET_ROWS:
ggml_sycl_op_set_rows(ctx, dst);
break;
case GGML_OP_DSV4_HC_PRE:
ggml_sycl_op_dsv4_hc_pre(ctx, dst);
break;
case GGML_OP_DSV4_HC_COMB:
ggml_sycl_op_dsv4_hc_comb(ctx, dst);
break;
case GGML_OP_DSV4_HC_POST:
ggml_sycl_op_dsv4_hc_post(ctx, dst);
break;
case GGML_OP_LIGHTNING_INDEXER:
ggml_sycl_op_lightning_indexer(ctx, dst);
break;
case GGML_OP_DUP:
ggml_sycl_dup(ctx, dst);
break;
@@ -5801,6 +5815,27 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
return res;
}
break;
case GGML_OP_DSV4_HC_PRE:
return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 &&
op->type == GGML_TYPE_F32;
case GGML_OP_DSV4_HC_COMB:
return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 &&
op->src[2]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
case GGML_OP_DSV4_HC_POST:
return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 &&
op->src[2]->type == GGML_TYPE_F32 && op->src[3]->type == GGML_TYPE_F32 &&
op->type == GGML_TYPE_F32;
case GGML_OP_LIGHTNING_INDEXER:
return op->src[0]->type == GGML_TYPE_F32 &&
(op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32 ||
op->src[1]->type == GGML_TYPE_BF16 || op->src[1]->type == GGML_TYPE_Q8_0 ||
op->src[1]->type == GGML_TYPE_Q5_1 || op->src[1]->type == GGML_TYPE_Q5_0 ||
op->src[1]->type == GGML_TYPE_Q4_1 || op->src[1]->type == GGML_TYPE_Q4_0 ||
op->src[1]->type == GGML_TYPE_IQ4_NL) &&
op->src[2]->type == GGML_TYPE_F32 &&
op->src[3]->type == GGML_TYPE_F16 &&
op->type == GGML_TYPE_F32 &&
op->src[0]->ne[0] == WARP_SIZE * 8;
case GGML_OP_CPY:
{
ggml_type src0_type = op->src[0]->type;