vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

use fp16fp16fp16

Browse source
This commit is contained in:
luoyu-intel 2024-08-20 16:01:29 +08:00
parent af1b276a34
commit 2ee02b290b
1 changed files with 0 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

5
ggml/src/ggml-sycl.cpp
View file

@ -2495,15 +2495,10 @@ inline void ggml_sycl_op_mul_mat_sycl(
to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream); to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
#else #else
auto dnnl_stream = ctx.stream_dnnl(stream); auto dnnl_stream = ctx.stream_dnnl(stream);
#if 0
DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>()); src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>());
const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream); to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
#else
DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
#endif
#endif #endif
} }
else { else {