CANN: Support Ascend310P to accelerate F32 and F16 Model (#10216)
* CANN Support Ascend310P to accelerate F32 and F16 Model * Add compile option soc type macro ASCEND_310P to ggml-cann lib * Remove unused code * Remove the ascend soc_type hard code compile option in CMakelist.txt
This commit is contained in:
parent
a5e47592b6
commit
c18610b4ee
7 changed files with 123 additions and 41 deletions
|
@ -2312,6 +2312,14 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
|
|||
|
||||
switch (src0->type) {
|
||||
case GGML_TYPE_F32:
|
||||
{
|
||||
#ifdef ASCEND_310P
|
||||
// Special operation for get_row_f32 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
|
||||
if ((src0->ne[0] % 8) != 0) {
|
||||
size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
|
||||
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
|
||||
}
|
||||
#endif
|
||||
aclrtlaunch_ascendc_get_row_f32(
|
||||
24, ctx.stream(), src0->data, src1->data, dst->data,
|
||||
((ggml_tensor*)src0->extra)->ne,
|
||||
|
@ -2320,7 +2328,16 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
|
|||
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
|
||||
((ggml_tensor*)dst->extra)->nb);
|
||||
break;
|
||||
}
|
||||
case GGML_TYPE_F16:
|
||||
{
|
||||
#ifdef ASCEND_310P
|
||||
// Special operation for get_row_f16 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
|
||||
if ((src0->ne[0] % 16) != 0) {
|
||||
size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32); // out is also f32, even input is f16
|
||||
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
|
||||
}
|
||||
#endif
|
||||
aclrtlaunch_ascendc_get_row_f16(
|
||||
24, ctx.stream(), src0->data, src1->data, dst->data,
|
||||
((ggml_tensor*)src0->extra)->ne,
|
||||
|
@ -2329,6 +2346,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
|
|||
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
|
||||
((ggml_tensor*)dst->extra)->nb);
|
||||
break;
|
||||
}
|
||||
case GGML_TYPE_Q4_0:
|
||||
aclrtlaunch_ascendc_get_row_q4_0(
|
||||
24, ctx.stream(), src0->data, src1->data, dst->data,
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue