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

Formatting

Browse source
This commit is contained in:
nscipione 2024-11-29 14:40:35 +00:00
parent a7e15b0355
commit ffd0a998c7
3 changed files with 21 additions and 32 deletions
Download patch file Download diff file Expand all files Collapse all files

28
ggml/src/ggml-sycl/dpct/helper.hpp
View file

@ -1691,12 +1691,11 @@ namespace dpct
auto data_c = get_memory<Tc>(c); auto data_c = get_memory<Tc>(c);
oneapi::mkl::blas::column_major::gemm( oneapi::mkl::blas::column_major::gemm(
#ifdef GGML_SYCL_NVIDIA #ifdef GGML_SYCL_NVIDIA
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{q}, oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
#else #else
q, q,
#endif #endif
a_trans, b_trans, m, n, k, alpha_value, data_a, lda, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb, beta_value, data_c, ldc);
data_b, ldb, beta_value, data_c, ldc);
} }
template <typename VecT, class BinaryOperation, class = void> template <typename VecT, class BinaryOperation, class = void>
@ -1761,16 +1760,14 @@ namespace dpct
sycl::event e = oneapi::mkl::blas::column_major::gemm_batch( sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
#ifdef GGML_SYCL_NVIDIA #ifdef GGML_SYCL_NVIDIA
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{q}, oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
#else #else
q, q,
#endif #endif
matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info,
matrix_info->size_info, matrix_info->size_info + 1, matrix_info->size_info + 1, matrix_info->size_info + 2, matrix_info->value_info,
matrix_info->size_info + 2, matrix_info->value_info, reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
reinterpret_cast<const Ta **>(a), matrix_info->ld_info, matrix_info->ld_info + 1, matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info)); matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
q.submit([&](sycl::handler &cgh) q.submit([&](sycl::handler &cgh)
@ -1795,12 +1792,11 @@ namespace dpct
auto data_c = get_memory<Tc>(c); auto data_c = get_memory<Tc>(c);
oneapi::mkl::blas::column_major::gemm_batch( oneapi::mkl::blas::column_major::gemm_batch(
#ifdef GGML_SYCL_NVIDIA #ifdef GGML_SYCL_NVIDIA
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{q}, oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
#else #else
q, q,
#endif #endif
a_trans, b_trans, m, n, k, alpha_value, data_a, lda, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, stride_a, data_b, ldb, stride_b, beta_value,
stride_a, data_b, ldb, stride_b, beta_value,
data_c, ldc, stride_c, batch_size); data_c, ldc, stride_c, batch_size);
} }

14
ggml/src/ggml-sycl/ggml-sycl.cpp
View file

@ -2562,15 +2562,13 @@ inline void ggml_sycl_op_mul_mat_sycl(
const float beta = 0.0f; const float beta = 0.0f;
#if !GGML_SYCL_DNNL #if !GGML_SYCL_DNNL
SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm( SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
#ifdef GGML_SYCL_NVIDIA # ifdef GGML_SYCL_NVIDIA
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{*stream}, oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream },
#else # else
*stream, *stream,
#endif # endif
oneapi::mkl::transpose::trans, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
dpct::get_value(&alpha, *stream), src0_ddf_i, ne00,
src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
dst_dd_i, ldc))); dst_dd_i, ldc)));
#else #else
auto dnnl_stream = ctx.stream_dnnl(stream); auto dnnl_stream = ctx.stream_dnnl(stream);

11
ggml/src/ggml-sycl/outprod.cpp
View file

@ -42,17 +42,12 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, const ggml_tensor* sr
// Perform matrix multiplication using oneMKL GEMM // Perform matrix multiplication using oneMKL GEMM
oneapi::mkl::blas::column_major::gemm( oneapi::mkl::blas::column_major::gemm(
#ifdef GGML_SYCL_NVIDIA #ifdef GGML_SYCL_NVIDIA
oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{*stream}, oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream },
#else #else
*stream, *stream,
#endif #endif
oneapi::mkl::transpose::nontrans, src1_op, oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha, src0_d, ne00, src1_d, ldb, beta, dst_d,
ne0, ne1, ne01, ne0);
alpha,
src0_d, ne00,
src1_d, ldb,
beta,
dst_d, ne0);
} }
catch (sycl::exception const& exc) { catch (sycl::exception const& exc) {
std::cerr << exc.what() << std::endl; std::cerr << exc.what() << std::endl;