LLM inference in C/C++
C++ ggml
sycl: fix multiple warnings in compiling sycl backend (#26713)
Ian Faust6602dd338941
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txtindex a6cc49586b..32e1e7aa19 100644--- a/ggml/src/ggml-cpu/CMakeLists.txt+++ b/ggml/src/ggml-cpu/CMakeLists.txt@@ -737,8 +737,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")endif()- if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")- # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"- target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")- endif()+ if (CMAKE_C_COMPILER_ID STREQUAL "IntelLLVM" OR CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")+ # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"+ target_compile_options(${GGML_CPU_NAME} PRIVATE "$<$<OR:$<COMPILE_LANG_AND_ID:C,IntelLLVM>,$<COMPILE_LANG_AND_ID:CXX,IntelLLVM>>:$<$<BOOL:${WIN32}>:/clang:>-fno-associative-math>")+ endif()+endfunction()diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hppindex 664b8e9697..85af4cab68 100644--- a/ggml/src/ggml-sycl/dpct/helper.hpp+++ b/ggml/src/ggml-sycl/dpct/helper.hpp@@ -62,7 +62,7 @@#define DPCT_UNUSED(x) (void)(x)-inline void _abort(const char * str) {+[[noreturn]] inline void _abort(const char * str) {std::cerr << str << std::endl;std::abort();}diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cppindex 8619ed6f4b..95914873e5 100644--- a/ggml/src/ggml-sycl/element_wise.cpp+++ b/ggml/src/ggml-sycl/element_wise.cpp@@ -10,7 +10,7 @@(ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))static void acc_f32(const char * x, const char * y, float * dst, const int64_t ne,- const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,+ const int64_t ne0, const int64_t ne1, const int64_t ne2,const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,@@ -455,7 +455,7 @@ static void unary_mul_sycl(const T * x, const T * g, T * dst, const int64_t k, cnamespace ggml_sycl_detail {static void acc_f32_sycl(const char *x, const char *y, float *dst,const int64_t n_elements,- const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,+ const int64_t ne0, const int64_t ne1, const int64_t ne2,const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,@@ -466,7 +466,7 @@ static void acc_f32_sycl(const char *x, const char *y, float *dst,sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),[=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {acc_f32(x, y, dst, n_elements,- ne0, ne1, ne2, ne3,+ ne0, ne1, ne2,nb00, nb01, nb02, nb03,ne10, ne11, ne12, ne13,nb10, nb11, nb12, nb13,@@ -970,7 +970,7 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensorconst int64_t offset = (int64_t) ((const int32_t *) dst->op_params)[3] / (int64_t) sizeof(float);ggml_sycl_detail::acc_f32_sycl(src0_d, src1_d, dst_d, ggml_nelements(dst),- dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],+ dst->ne[0], dst->ne[1], dst->ne[2],src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],diff --git a/ggml/src/ggml-sycl/fattn-mkl.cpp b/ggml/src/ggml-sycl/fattn-mkl.cppindex fc22b7bdb8..2d164a0840 100644--- a/ggml/src/ggml-sycl/fattn-mkl.cpp+++ b/ggml/src/ggml-sycl/fattn-mkl.cpp@@ -43,7 +43,7 @@ static void mkl_fa_pack_q_fp16(dpct::queue_ptr stream,sycl::half * __restrict dst,const float * __restrict q_src,- int n_queries, int n_query_rows, int DKQ,+ int n_queries, int DKQ,int gqa_ratio, int kvh_base_head,float q_scale, int64_t q_row_stride, int64_t q_head_stride,int64_t wg_size) {@@ -121,7 +121,7 @@ static void mkl_fa_online_softmax_chunk(float * __restrict VKQ_accum,int q0, int q_rows, int n_queries, int DV,int chunk_size, int chunk_start,- int kvh_head, int gqa_ratio,+ int kvh_head,const sycl::half * mask_data, int64_t mask_head_stride,int64_t mask_row_stride, int mask_n_heads,float logit_softcap, int64_t wg_size) {@@ -473,7 +473,6 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *MKL_ACCUM(dequant_time_us, t_deq);// --- Resolve mask pointers ---- const sycl::half * mask_data = nullptr;int64_t mask_head_stride = 0;int64_t mask_row_stride = 0;int mask_n_heads = 0;@@ -547,7 +546,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *// 1. Pack all GQA Q heads into fp16 (full n_query_rows)mkl_fa_pack_q_fp16(stream,Q_head_f16_ptr, Q_batch,- n_queries, n_query_rows, DKQ,+ n_queries, DKQ,gqa_ratio, kvh_base_head,q_scale, q_row_stride, q_head_stride, wg_size);@@ -605,7 +604,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *KQ_max_ptr, KQ_sum_ptr, VKQ_accum_ptr,q0, q_rows, n_queries, DV,this_chunk, chunk_start,- kvh_base_head, gqa_ratio,+ kvh_base_head,mask_batch, mask_head_stride,mask_row_stride, mask_n_heads,logit_softcap, wg_size);diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cppindex 8372428795..7ebdce7fb9 100644--- a/ggml/src/ggml-sycl/ggml-sycl.cpp+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp@@ -921,16 +921,16 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,void * dev_ptr;if (use_usm_system) {- GGML_SYCL_DEBUG("[SYCL] allocating %lu Bytes with USM system\n", size);+ GGML_SYCL_DEBUG("[SYCL] allocating %zu Bytes with USM system\n", size);dev_ptr = (void *)aligned_malloc_host(alignment, aligned_size);if (!dev_ptr) {- GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);+ GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);return nullptr;}} else {SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)ggml_sycl_malloc_device(size, *stream)));if (!dev_ptr) {- GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size);+ GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device\n", __func__, size);return nullptr;}}@@ -1177,7 +1177,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)ggml_sycl_malloc_device(size, *stream)));if (!buf) {char err_buf[1024];- snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size);+ snprintf(err_buf, 1023, "%s: can't allocate %zu Bytes of memory on device\n", __func__, size);throw std::runtime_error(err_buf);}// set padding to 0 to avoid possible NaN values@@ -1651,7 +1651,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *)ggml_sycl_malloc_device(look_ahead_size, *qptr)));if (!ptr) {- GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device/GPU\n", __func__, look_ahead_size);+ GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device/GPU\n", __func__, look_ahead_size);return nullptr;}@@ -1663,7 +1663,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {(uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));#endif- // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr);+ // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%zu, return %p\n", look_ahead_size, ptr);return ptr;}@@ -1843,7 +1843,7 @@ struct ggml_sycl_pool_host : public ggml_sycl_pool {SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *) sycl::malloc_host(size, *qptr)));if (!ptr) {- GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);+ GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);return nullptr;}pool_size += size;@@ -2779,9 +2779,9 @@ inline void ggml_sycl_op_mul_mat_sycl(const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();{+#if GGML_SYCL_DNNLconst int64_t gemm_flops = (int64_t)row_diff * src1_ncols * ne10;const bool use_mkl_direct = gemm_flops < 256 * 256 * 256;-#if GGML_SYCL_DNNLif (g_ggml_sycl_enable_dnn && !use_mkl_direct) {DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),@@ -3518,7 +3518,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, consfloat * dst_ddf = static_cast<float *>(dst->data);const sycl::half * src1_f16 = static_cast<const sycl::half *>(src1->data);+#if GGML_SYCL_DNNLconst size_t type_size_src0 = ggml_type_size(src0->type);+#endifconst size_t type_size_src1 = ggml_type_size(src1->type);bool is_src0_cont_2 = ggml_is_contiguous_2(src0);@@ -3535,6 +3537,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, consscope_op_debug_print scope_dbg_print(__func__, "/to_fp16_nc_sycl", dst, /*num_src=*/2," : converting src1 to fp16");+#if GGML_SYCL_DNNL// iterate tensor dims and find the slowest moving dim and strideint last_dim=0;int last_str=0;@@ -3554,7 +3557,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons}}-#if GGML_SYCL_DNNL// oneDNN handles strided data and does not need overhead of ggml_get_to_fp16_nc_syclconst int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;src1_f16_alloc.alloc(ne_src1);diff --git a/ggml/src/ggml-sycl/im2col.cpp b/ggml/src/ggml-sycl/im2col.cppindex 7bf3584fb9..e666167594 100644--- a/ggml/src/ggml-sycl/im2col.cpp+++ b/ggml/src/ggml-sycl/im2col.cpp@@ -85,7 +85,7 @@ static void im2col_sycl(const float * x,*/stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),- [=](sycl::nd_item<3> item_ct1) {+ [=](sycl::nd_item<3>) {im2col_kernel(x, dst, IC, IW, IH, OH, OW, KW, KH, IC_IH_IW, IH_IW, N_OH, KH_KW, IC_KH_KW,s0, s1, p0, p1, d0, d1);});@@ -271,7 +271,7 @@ static void im2col_3d_sycl(const float * src,*/stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),- [=](sycl::nd_item<3> item_ct1) {+ [=](sycl::nd_item<3>) {im2col_3d_kernel(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, OH_OW, KD_KH_KW,ID_IH_IW, KH_KW, IH_IW, IC_ID_IH_IW, IC_KD_KH_KW, OW_KD_KH_KW,OD_OH_OW_IC_KD_KH_KW, OH_OW_IC_KD_KH_KW, OW_IC_KD_KH_KW, N_OD_OH, OD_OH,diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cppindex 682a9f51ee..f98a7a9542 100644--- a/ggml/src/ggml-sycl/norm.cpp+++ b/ggml/src/ggml-sycl/norm.cpp@@ -7,9 +7,6 @@ static void norm_f32(const float* x, float* dst, const int ncols,const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {- const int nrows = item_ct1.get_group_range(2);- const int nchannels = item_ct1.get_group_range(1);-const int nthreads = item_ct1.get_local_range(2);const int sample = item_ct1.get_group(0);const int channel = item_ct1.get_group(1);@@ -155,9 +152,6 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols,const float* mul = nullptr, const int64_t mul_stride_row = 0, const int64_t mul_stride_channel = 0,const int64_t mul_stride_sample = 0, const int mul_nrows = 0, const int mul_nchannels = 0, const int mul_nsamples = 0) {- const int nrows = item_ct1.get_group_range(2);- const int nchannels = item_ct1.get_group_range(1);-const int sample = item_ct1.get_group(0);const int channel = item_ct1.get_group(1);const int row = item_ct1.get_group(2);@@ -225,8 +219,6 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,const int64_t src_stride_sample, const int64_t dst_stride_col, const int64_t dst_stride_row,const int64_t dst_stride_channel, const int64_t dst_stride_sample, const float eps,const sycl::nd_item<3>& item_ct1, float* s_sum, const int block_size) {- const int nrows = item_ct1.get_group_range(2);- const int nchannels = item_ct1.get_group_range(1);const int row = item_ct1.get_group(2);const int channel = item_ct1.get_group(1);diff --git a/ggml/src/ggml-sycl/set_rows.cpp b/ggml/src/ggml-sycl/set_rows.cppindex 52a0bcb6eb..5f8d881a29 100644--- a/ggml/src/ggml-sycl/set_rows.cpp+++ b/ggml/src/ggml-sycl/set_rows.cpp@@ -291,7 +291,7 @@ static void set_rows_sycl(stream->parallel_for(sycl::nd_range<1>(grid_size * block_size, block_size),- [=](sycl::nd_item<1> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {+ [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {k_set_rows<TIn, TIdx, TOut>(src0_d, src1_d, dst_d,ne00, ne01, ne02,