From d702b3224cd6272de99473c5725d48181338f680 Mon Sep 17 00:00:00 2001 From: Youmu Liu <105042599+Nitload-NSI@users.noreply.github.com> Date: Fri, 3 Apr 2026 18:40:11 +0800 Subject: [PATCH] adjust for clang-gcc --- StgSharp.Native/CMakeLists.txt | 53 ++++++++++--- StgSharp.Native/StgSharp.Native.vcxproj | 2 +- .../StgSharp.Native.vcxproj.filters | 4 + StgSharp.Native/sn-head/StgSharpNative.h | 6 +- StgSharp.Native/sn-head/sn_internal.h | 7 +- .../sn-head/sn_intrinsic.context.x86.h | 4 +- StgSharp.Native/sn-head/sn_intrinsic.h | 12 ++- .../sn-head/sn_intrinsic.matkernel.h | 6 +- StgSharp.Native/sn-head/sn_intrinsic.std.h | 29 ++++--- StgSharp.Native/sn-src/load_intrinsic.x86.c | 11 ++- StgSharp.Native/sn-src/m.bufop.x86.c | 4 +- StgSharp.Native/sn-src/m.transpose.x86.c | 16 ++-- StgSharp.Native/sn-src/mk.fma.x86.c | 79 +++++++++++-------- StgSharp.Native/sn-src/s.hash.x86.c | 14 ++-- StgSharp.Native/sn-src/simdid.x86.c | 10 +++ StgSharp.Native/sn-src/sk.fma.x86.c | 71 ++++++++++------- StgSharp.Native/sn-src/sn_native.c | 2 +- StgSharp.Native/sn-src/thread.linux.c | 11 ++- StgSharp.Native/sn-src/vector.c | 2 +- 19 files changed, 217 insertions(+), 126 deletions(-) diff --git a/StgSharp.Native/CMakeLists.txt b/StgSharp.Native/CMakeLists.txt index 96403b67..9bbc2912 100644 --- a/StgSharp.Native/CMakeLists.txt +++ b/StgSharp.Native/CMakeLists.txt @@ -20,9 +20,14 @@ sn_resolve_target_identifiers( file(GLOB STGSHARP_NATIVE_SOURCES CONFIGURE_DEPENDS ${CMAKE_CURRENT_LIST_DIR}/sn-src/*.c) -list(REMOVE_ITEM STGSHARP_NATIVE_SOURCES - ${CMAKE_CURRENT_LIST_DIR}/sn-src/formatted.c - ${CMAKE_CURRENT_LIST_DIR}/sn-src/StgSharpC.c) +file(GLOB STGSHARP_NATIVE_HEADERS CONFIGURE_DEPENDS + ${CMAKE_CURRENT_LIST_DIR}/sn-head/*.h) + +file(GLOB STGSHARP_NATIVE_TEMPLATES CONFIGURE_DEPENDS + ${CMAKE_CURRENT_LIST_DIR}/sn-template/*.h + ${CMAKE_CURRENT_LIST_DIR}/sn-template/*.inl + ${CMAKE_CURRENT_LIST_DIR}/sn-template/*.inc + ${CMAKE_CURRENT_LIST_DIR}/sn-template/*.ipp) if(NOT STGSHARP_NATIVE_SOURCES) message(FATAL_ERROR "No source files found under sn-src/.") @@ -52,8 +57,14 @@ endif() set(STGSHARP_BUILD_OUTPUT_DIR ${STGSHARP_SOLUTION_ROOT}/build/${STGSHARP_OS_TAG}-${STGSHARP_ARCH_TAG}) set(STGSHARP_OBJECT_ROOT_DIR ${STGSHARP_SOLUTION_ROOT}/obj/${PROJECT_NAME}) +set(STGSHARP_NATIVE_PROJECT_FILES + ${STGSHARP_NATIVE_SOURCES} + ${STGSHARP_NATIVE_HEADERS} + ${STGSHARP_NATIVE_TEMPLATES}) -add_library(${PROJECT_NAME} SHARED ${STGSHARP_NATIVE_SOURCES}) +add_library(${PROJECT_NAME} SHARED ${STGSHARP_NATIVE_PROJECT_FILES}) + +source_group(TREE ${CMAKE_CURRENT_LIST_DIR} FILES ${STGSHARP_NATIVE_PROJECT_FILES}) set_target_properties(${PROJECT_NAME} PROPERTIES C_STANDARD 11 @@ -75,23 +86,43 @@ endforeach() set_target_properties(${PROJECT_NAME} PROPERTIES COMPILE_PDB_OUTPUT_DIRECTORY ${STGSHARP_OBJECT_ROOT_DIR}/$/${STGSHARP_ARCH_TAG}) -if(MSVC OR CMAKE_C_COMPILER_ID STREQUAL "Clang") +if(MSVC OR (CMAKE_C_COMPILER_ID STREQUAL "Clang" AND CMAKE_C_COMPILER_FRONTEND_VARIANT STREQUAL "MSVC")) target_compile_options(${PROJECT_NAME} PRIVATE /Fd${STGSHARP_OBJECT_ROOT_DIR}/$/${STGSHARP_ARCH_TAG}/) endif() if(CMAKE_C_COMPILER_ID STREQUAL "Clang" AND SN_RESOLVED_ARCH_NAME STREQUAL "x86_64") - target_compile_options(${PROJECT_NAME} PRIVATE - /clang:-mavx - /clang:-mavx2 - /clang:-mfma - /clang:-mavx512f - /clang:-mavx512vl) + # Keep the intrinsic type system and builtin declarations visible to all TUs. + # The runtime dispatch contract still follows T_arch/T_feature at the function level. + if(CMAKE_C_COMPILER_FRONTEND_VARIANT STREQUAL "MSVC") + target_compile_options(${PROJECT_NAME} PRIVATE + /clang:-mavx + /clang:-mavx2 + /clang:-mfma + /clang:-mf16c + /clang:-mavx512f + /clang:-mavx512cd + /clang:-mavx512dq + /clang:-mavx512bw + /clang:-mavx512vl) + else() + target_compile_options(${PROJECT_NAME} PRIVATE + -mavx + -mavx2 + -mfma + -mf16c + -mavx512f + -mavx512cd + -mavx512dq + -mavx512bw + -mavx512vl) + endif() endif() target_include_directories(${PROJECT_NAME} PRIVATE ${CMAKE_CURRENT_LIST_DIR}/sn-head + ${CMAKE_CURRENT_LIST_DIR}/sn-template ${CMAKE_CURRENT_LIST_DIR}/lib) target_compile_definitions(${PROJECT_NAME} diff --git a/StgSharp.Native/StgSharp.Native.vcxproj b/StgSharp.Native/StgSharp.Native.vcxproj index 3abfa5da..4505f211 100644 --- a/StgSharp.Native/StgSharp.Native.vcxproj +++ b/StgSharp.Native/StgSharp.Native.vcxproj @@ -110,7 +110,7 @@ stdc11 true Speed - $(ProjectDir)lib;$(ProjectDir)sn-head;%(AdditionalIncludeDirectories) + $(ProjectDir)lib;$(ProjectDir)sn-head;$(ProjectDir)sn-template;%(AdditionalIncludeDirectories) diff --git a/StgSharp.Native/StgSharp.Native.vcxproj.filters b/StgSharp.Native/StgSharp.Native.vcxproj.filters index 11bae1bb..ec136964 100644 --- a/StgSharp.Native/StgSharp.Native.vcxproj.filters +++ b/StgSharp.Native/StgSharp.Native.vcxproj.filters @@ -16,6 +16,10 @@ {fbd74b53-5314-4d9d-a788-07bd1b413f74} + + {C8E0D8C1-42A1-4F7E-9C35-7E6D32D49A7B} + h;hh;hpp;hxx;h++;hm;inl;inc;ipp;xsd + diff --git a/StgSharp.Native/sn-head/StgSharpNative.h b/StgSharp.Native/sn-head/StgSharpNative.h index 2ebb8ccf..6a262e93 100644 --- a/StgSharp.Native/sn-head/StgSharpNative.h +++ b/StgSharp.Native/sn-head/StgSharpNative.h @@ -13,7 +13,6 @@ #include "gl.h" #include "wgl.h" -#define GLFWAPI INTERNAL #include "glfw\glfw3.h" #include #include @@ -71,11 +70,12 @@ SN_API void SN_DECL load_intrinsic_function(void *intrinsic_context, uint64_t id SN_API SIMDID SN_DECL sn_get_unite_simd(SIMDID left, SIMDID right); SN_API int SN_DECL sn_compare_simd(SIMDID left, SIMDID right); SN_API void SN_DECL sn_set_thread_affinity(int core_id); -SN_API void SN_DECL sn_get_numa_overall(int *out_numa_count, int *count_of_each_numa); -SN_API void SN_DECL sn_get_numa_each(int numa_id, uint64_t *in_numa_cpuid, uint64_t *global_cpuid); +SN_API int SN_DECL sn_get_numa_overall(int *out_numa_count, int *count_of_each_numa); +SN_API int SN_DECL sn_get_numa_each(int numa_id, uint64_t *in_numa_cpuid, uint64_t *global_cpuid); /* Per-core variant: always re-executes CPUID, never caches. * Must be called on the target core (after affinity is set). */ SN_API SIMDID SN_DECL sn_get_simd_level_local(); +SN_API SIMDID SN_DECL sn_get_simd_level_global(); #pragma endregion #ifdef __cplusplus diff --git a/StgSharp.Native/sn-head/sn_internal.h b/StgSharp.Native/sn-head/sn_internal.h index 1ee66c90..377364f1 100644 --- a/StgSharp.Native/sn-head/sn_internal.h +++ b/StgSharp.Native/sn-head/sn_internal.h @@ -2,7 +2,7 @@ #ifndef SN_INTERNAL #define SN_INTERNAL -#include +#include #define UNSAFE_AS(type, expr) *(type *)&(expr) @@ -37,4 +37,9 @@ #define PRIVATE static #define INTERNAL extern +#define SN_PTR_CAST(type, expr) ((type *)(void *)(expr)) +#define SN_CONST_PTR_CAST(type, expr) ((type const *)(void const *)(expr)) +#define SN_REF_CAST(type, expr) ((type *)(void *)&(expr)) +#define SN_CONST_REF_CAST(type, expr) ((type const *)(void const *)&(expr)) + #endif // diff --git a/StgSharp.Native/sn-head/sn_intrinsic.context.x86.h b/StgSharp.Native/sn-head/sn_intrinsic.context.x86.h index e55f893e..9c446e05 100644 --- a/StgSharp.Native/sn-head/sn_intrinsic.context.x86.h +++ b/StgSharp.Native/sn-head/sn_intrinsic.context.x86.h @@ -239,8 +239,8 @@ INTERNAL uint64_t SN_DECL factorial_simd_sse(int n); #pragma region dot_product -INTERNAL void FORCEINLINE SN_DECL dot_41_sse(MAT_KERNEL(float) const *transpose, - __m128 const *vector, __m128 *ans); +INTERNAL void FORCEINLINE SN_DECL dot_41_sse( + MAT_KERNEL(float) const *transpose, __m128 const *vector, __m128 *ans); INTERNAL void SN_DECL dot_41_avx(MAT_KERNEL(float) const *transpose, __m128 const *vector, __m128 *ans); INTERNAL void SN_DECL dot_42_sse(MAT_KERNEL(float) const *transpose, __m128 const *vector, diff --git a/StgSharp.Native/sn-head/sn_intrinsic.h b/StgSharp.Native/sn-head/sn_intrinsic.h index fc9980d5..15dea136 100644 --- a/StgSharp.Native/sn-head/sn_intrinsic.h +++ b/StgSharp.Native/sn-head/sn_intrinsic.h @@ -39,6 +39,11 @@ typedef union uint64_u { uint32_t u32[2]; } uint64_u; +/* Opaque placeholder used by generic single-kernel dispatch. + * It preserves pointer semantics without pretending every branch shares the + * same concrete MAT_KERNEL(T) type. */ +typedef struct sn_mat_kernel_opaque sn_mat_kernel_opaque; + typedef struct mat_kernel mat_kernel; typedef struct mat_panel mat_panel; @@ -71,9 +76,10 @@ static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mis static_assert(sizeof(matrix_task) == 128, "matrix_parallel_task_package size mismatch"); #endif -typedef void(SN_DECL *UNI_MK_PROC)(matrix_task *task); -typedef void(SN_DECL *UNI_SK_PROC)(mat_kernel const *left, mat_kernel const *right, - mat_kernel *restrict ans, uint64_t scalar); +typedef void(SN_DECL *UNI_MK_PROC)(matrix_task const *task); +typedef void(SN_DECL *UNI_SK_PROC)(sn_mat_kernel_opaque const *left, + sn_mat_kernel_opaque const *right, + sn_mat_kernel_opaque *restrict ans, uint64_t scalar); typedef void(SN_DECL *VECTORNORMALIZEPROC)(sn_vec_f32 *source, sn_vec_f32 *target); typedef void(SN_DECL *DOTPROC)(void *transpose, __m128 *vector, __m128 *ans); diff --git a/StgSharp.Native/sn-head/sn_intrinsic.matkernel.h b/StgSharp.Native/sn-head/sn_intrinsic.matkernel.h index dfd35a5a..14cd1dc5 100644 --- a/StgSharp.Native/sn-head/sn_intrinsic.matkernel.h +++ b/StgSharp.Native/sn-head/sn_intrinsic.matkernel.h @@ -138,7 +138,7 @@ extern __declspec(align(64)) sn_zero_kernel_u64 SN_ZERO_KERNEL; int col_length, int row_length, \ int col_index, int row_index) -/* Declarations for AVX float variant +#if 0 DECLARE_BUILD_PANEL(avx, float); DECLARE_BUILD_PANEL(512, float); @@ -149,6 +149,6 @@ DECLARE_STORE_PANEL(512, float); DECLARE_BUILD_PANEL(512, double); DECLARE_STORE_PANEL(512, double); -/**/ +#endif -#endif /* SN_MATKERNEL */ +#endif /* SN_MATKERNEL */ \ No newline at end of file diff --git a/StgSharp.Native/sn-head/sn_intrinsic.std.h b/StgSharp.Native/sn-head/sn_intrinsic.std.h index 913452df..6f63025d 100644 --- a/StgSharp.Native/sn-head/sn_intrinsic.std.h +++ b/StgSharp.Native/sn-head/sn_intrinsic.std.h @@ -11,19 +11,21 @@ #define MK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op) \ T_type##_multi_kernel_proc_##T_op##_##T_arch##_##T_feature #define GET_MK_PROC_STD(T_type, T_arch, T_feature, T_op) \ - ((UNI_MK_PROC)MK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op)) + MK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op) #define SN_MK_PROC_DECL(T_type, P_name) INTERNAL void SN_DECL P_name(MAT_TASK(T_type) const *task) -#define SN_MK_PROC_DECL_STD(T_type, T_arch, T_feature, T_op) \ - INTERNAL FORCEINLINE void SN_DECL MK_PROC_NAME_STD(T_type, T_arch, T_feature, \ - T_op)(MAT_TASK(T_type) const *task) +#define SN_MK_PROC_DECL_STD(T_type, T_arch, T_feature, T_op) \ + INTERNAL void SN_DECL MK_PROC_NAME_STD(T_type, T_arch, T_feature, \ + T_op)(MAT_TASK(T_type) const *task) #define SK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op) \ T_type##_single_kernel_proc_##T_op##_##T_arch##_##T_feature #define GET_SK_PROC_STD(T_type, T_arch, T_feature, T_op) \ - ((UNI_SK_PROC)SK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op)) -#define SN_SK_PROC_DECL_STD(T_type, T_arch, T_feature, T_op) \ - INTERNAL FORCEINLINE void SN_DECL SK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op)( \ - MAT_KERNEL(T_type) const *left, MAT_KERNEL(T_type) const *right, \ - MAT_KERNEL(T_type) *restrict ans, uint64_t scalar) + SK_PROC_NAME_STD(T_type, T_arch, T_feature, T_op) +#define SN_SK_PROC_DECL_STD(T_type, T_arch, T_feature, T_op) \ + INTERNAL void SN_DECL SK_PROC_NAME_STD(T_type, T_arch, T_feature, \ + T_op)(sn_mat_kernel_opaque const *left_ptr, \ + sn_mat_kernel_opaque const *right_ptr, \ + sn_mat_kernel_opaque *restrict ans_ptr, \ + uint64_t scalar) #define _SN_STD__PROBE_0 0 #define _SN_STD__PROBE_1 1 @@ -100,4 +102,13 @@ matrix_task const *_ = (_SN_PARAM_STYLE(param), (_SN_MEM_LAYOUT(mem)), task); \ _SN_MK_EXPAND_PARAM(param); \ _SN_MK_EXPAND_MEM(mem) + +#define __sk_param_std(T) \ + typedef MAT_KERNEL(T) SN_MAT_KERNEL_LOCAL; \ + SN_MAT_KERNEL_LOCAL const *left = \ + (SN_MAT_KERNEL_LOCAL const *)(left_ptr); \ + SN_MAT_KERNEL_LOCAL const *right = \ + (SN_MAT_KERNEL_LOCAL const *)(right_ptr); \ + SN_MAT_KERNEL_LOCAL *restrict ans = \ + (SN_MAT_KERNEL_LOCAL *)(ans_ptr) #endif // !SN_INTRINSIC_PARAM_EXPAND diff --git a/StgSharp.Native/sn-src/load_intrinsic.x86.c b/StgSharp.Native/sn-src/load_intrinsic.x86.c index 7c4ccd51..3364f4cd 100644 --- a/StgSharp.Native/sn-src/load_intrinsic.x86.c +++ b/StgSharp.Native/sn-src/load_intrinsic.x86.c @@ -18,7 +18,7 @@ PRIVATE void load_intrinsic_sse(sn_intrinsic *context, SIMDID id); PRIVATE void load_intrinsic_avx(sn_intrinsic *context, SIMDID id); PRIVATE void load_intrinsic_512(sn_intrinsic *context, SIMDID id); -SN_API void load_intrinsic_function(void *intrinsic_context, uint64_t id) +SN_API void SN_DECL load_intrinsic_function(void *intrinsic_context, uint64_t id) { sn_null_assert(intrinsic_context); @@ -74,7 +74,6 @@ void load_intrinsic_avx(sn_intrinsic *context, SIMDID id) { char avx_seg = id.make_byte[5]; char is_fma = (avx_seg & SIMDID_AVX_FMA) != 0; - char is_f16c = (avx_seg & SIMDID_AVX_F16C) != 0; context->city_hash_simplify = city_hash_simplify_sse; context->factorial_simd = factorial_simd_sse; @@ -84,8 +83,8 @@ void load_intrinsic_avx(sn_intrinsic *context, SIMDID id) context->mat_mk[F32].scalar_mul = GET_MK_PROC_STD(float, avx, , scalar_mul); context->mat_mk[F32].sub = GET_MK_PROC_STD(float, avx, , sub); context->mat_mk[F32].transpose = GET_MK_PROC_STD(float, avx, , transpose); - context->mat_mk[F32].fma = is_fma ? GET_MK_PROC_STD(float, avx, , fma) : - GET_MK_PROC_STD(float, avx, _fma, fma); + context->mat_mk[F32].fma = is_fma ? GET_MK_PROC_STD(float, avx, _fma, fma) : + GET_MK_PROC_STD(float, avx, , fma); context->mat_mk[F32].pivot = NULL; context->mat_mk[F64].add = GET_MK_PROC_STD(double, avx, , add); @@ -93,8 +92,8 @@ void load_intrinsic_avx(sn_intrinsic *context, SIMDID id) context->mat_mk[F64].scalar_mul = GET_MK_PROC_STD(double, avx, , scalar_mul); context->mat_mk[F64].sub = GET_MK_PROC_STD(double, avx, , sub); context->mat_mk[F64].transpose = NULL; - context->mat_mk[F64].fma = is_fma ? GET_MK_PROC_STD(double, avx, , fma) : - GET_MK_PROC_STD(double, avx, _fma, fma); + context->mat_mk[F64].fma = is_fma ? GET_MK_PROC_STD(double, avx, _fma, fma) : + GET_MK_PROC_STD(double, avx, , fma); context->mat_mk[F64].pivot = NULL; context->vec_f32_normalize_3 = f32_normalize; diff --git a/StgSharp.Native/sn-src/m.bufop.x86.c b/StgSharp.Native/sn-src/m.bufop.x86.c index 34b8f3c9..6b76e984 100644 --- a/StgSharp.Native/sn-src/m.bufop.x86.c +++ b/StgSharp.Native/sn-src/m.bufop.x86.c @@ -68,7 +68,7 @@ SN_MK_PROC_DECL_STD(double, avx, , fill) SN_MK_PROC_DECL_STD(double, 512, , fill) { __mk_param_std(BUFFER, ANS_SCALAR, double); - register __m512d vec = _mm512_broadcastsd_pd(_mm_loadu_pd(scalar)); + register __m512d vec = _mm512_set1_pd(SN_CONST_PTR_CAST(double, scalar->data)[0]); for (size_t i = 0; i < count; i++) { ans[i].f64_z[0] = vec; ans[i].f64_z[1] = vec; @@ -354,7 +354,7 @@ SN_MK_PROC_DECL_STD(double, 512, , scalar_mul) { __mk_param_std(BUFFER, RIGHT_ANS_SCALAR, double); - register __m512d vec = _mm512_broadcastsd_pd(_mm_load_pd(scalar)); + register __m512d vec = _mm512_set1_pd(SN_CONST_PTR_CAST(double, scalar->data)[0]); for (size_t i = 0; i < count; i++) { register __m512d c0 = _mm512_mul_pd(right[i].f64_z[0], vec); register __m512d c1 = _mm512_mul_pd(right[i].f64_z[1], vec); diff --git a/StgSharp.Native/sn-src/m.transpose.x86.c b/StgSharp.Native/sn-src/m.transpose.x86.c index 2f6e2272..21cf6c28 100644 --- a/StgSharp.Native/sn-src/m.transpose.x86.c +++ b/StgSharp.Native/sn-src/m.transpose.x86.c @@ -20,10 +20,10 @@ SN_MK_PROC_DECL_STD(float, sse, , transpose) register __m128 tmp2 = _mm_unpacklo_ps(col2, col3); register __m128 tmp3 = _mm_unpackhi_ps(col2, col3); - _mm_store_ps(&(ans->f32_x[0]), _mm_movelh_ps(tmp0, tmp2)); - _mm_store_ps(&(ans->f32_x[1]), _mm_movehl_ps(tmp2, tmp0)); - _mm_store_ps(&(ans->f32_x[2]), _mm_movelh_ps(tmp1, tmp3)); - _mm_store_ps(&(ans->f32_x[3]), _mm_movehl_ps(tmp3, tmp1)); + ans->f32_x[0] = _mm_movelh_ps(tmp0, tmp2); + ans->f32_x[1] = _mm_movehl_ps(tmp2, tmp0); + ans->f32_x[2] = _mm_movelh_ps(tmp1, tmp3); + ans->f32_x[3] = _mm_movehl_ps(tmp3, tmp1); } SN_MK_PROC_DECL_STD(float, avx, , transpose) @@ -38,10 +38,8 @@ SN_MK_PROC_DECL_STD(float, avx, , transpose) col_low = _mm256_permute2f128_ps(unpack_low, unpack_high, 0x20); col_high = _mm256_permute2f128_ps(unpack_low, unpack_high, 0x31); - _mm256_store_ps(&(ans->f32_y[0]), - _mm256_shuffle_ps(col_high, col_low, _MM_SHUFFLE(2, 0, 2, 0))); - _mm256_store_ps(&(ans->f32_y[1]), - _mm256_shuffle_ps(col_high, col_low, _MM_SHUFFLE(3, 1, 3, 1))); + ans->f32_y[0] = _mm256_shuffle_ps(col_high, col_low, _MM_SHUFFLE(2, 0, 2, 0)); + ans->f32_y[1] = _mm256_shuffle_ps(col_high, col_low, _MM_SHUFFLE(3, 1, 3, 1)); } SN_MK_PROC_DECL_STD(float, 512, , transpose) @@ -52,7 +50,7 @@ SN_MK_PROC_DECL_STD(float, 512, , transpose) const __m512i transpose_indices = _mm512_setr_epi32(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15); - _mm512_store_ps(&(ans->f32_z[0]), _mm512_permutexvar_ps(transpose_indices, matrix)); + ans->f32_z[0] = _mm512_permutexvar_ps(transpose_indices, matrix); } #endif diff --git a/StgSharp.Native/sn-src/mk.fma.x86.c b/StgSharp.Native/sn-src/mk.fma.x86.c index 0a5f82c3..4cf78551 100644 --- a/StgSharp.Native/sn-src/mk.fma.x86.c +++ b/StgSharp.Native/sn-src/mk.fma.x86.c @@ -1,5 +1,5 @@ #include "sn_target.h" - +/* clang-format off */ #if SN_IS_ARCH(SN_ARCH_X86_64) #include "sn_intrinsic.h" @@ -16,7 +16,8 @@ SN_MK_PROC_DECL_STD(float, sse, , fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, float); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe + SN_MAT_KERNEL_LOCAL *restrict ans_cur = + ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe register __m128 c0 = _mm_load_ps((float *)&ans_cur->f32_x[0]); register __m128 c1 = _mm_load_ps((float *)&ans_cur->f32_x[1]); @@ -72,7 +73,8 @@ SN_MK_PROC_DECL_STD(float, avx, , fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, float); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe + SN_MAT_KERNEL_LOCAL *restrict ans_cur = + ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe register __m256 c01 = _mm256_load_ps((float *)&ans_cur->f32_y[0]); register __m256 c23 = _mm256_load_ps((float *)&ans_cur->f32_y[1]); @@ -105,7 +107,8 @@ SN_MK_PROC_DECL_STD(float, avx, _fma, fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, float); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe + SN_MAT_KERNEL_LOCAL *restrict ans_cur = + ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe register __m256 c01 = _mm256_load_ps((float *)&ans_cur->f32_y[0]); register __m256 c23 = _mm256_load_ps((float *)&ans_cur->f32_y[1]); @@ -136,8 +139,11 @@ SN_MK_PROC_DECL_STD(float, avx, _fma, fma) SN_MK_PROC_DECL_STD(float, 512, , fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, float); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); + SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y); register __m512 c = _mm512_load_ps((float *)ans_cur); + register __m512 c_1 = c; + register __m512 c_2 = c; + register __m512 c_3 = c; for (size_t z = 0; z < common_z; z++) { SN_MAT_KERNEL_LOCAL const *right_cur = right + GET_INDEX(common_z, x, z); @@ -170,31 +176,31 @@ SN_MK_PROC_DECL_STD(float, 512, , fma) #define DOUBLE_KERNEL_FMA_CYCLE(i) \ do { \ /* k = 0 */ \ - register __m128d b_lo = _mm_load_pd(right_cur + (2 * i + 0)); \ - register __m128d b_hi = _mm_load_pd(right_cur + (2 * i + 1)); \ - register __m128d a_lo = _mm_load_pd(left_cur + (0 * 2 + 0)); \ - register __m128d a_hi = _mm_load_pd(left_cur + (0 * 2 + 1)); \ + register __m128d b_lo = _mm_load_pd(right_values + (2 * i + 0)); \ + register __m128d b_hi = _mm_load_pd(right_values + (2 * i + 1)); \ + register __m128d a_lo = _mm_load_pd(left_values + (0 * 2 + 0)); \ + register __m128d a_hi = _mm_load_pd(left_values + (0 * 2 + 1)); \ register __m128d b = _mm_shuffle_pd(b_lo, b_lo, 0x0); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ \ /* group = 1 */ \ - a_lo = _mm_load_pd(left_cur + (1 * 2 + 0)); \ - a_hi = _mm_load_pd(left_cur + (1 * 2 + 1)); \ + a_lo = _mm_load_pd(left_values + (1 * 2 + 0)); \ + a_hi = _mm_load_pd(left_values + (1 * 2 + 1)); \ b = _mm_shuffle_pd(b_lo, b_lo, 0x3); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ \ /* groupp = 2 */ \ - a_lo = _mm_load_pd(left_cur + (2 * 2 + 0)); \ - a_hi = _mm_load_pd(left_cur + (2 * 2 + 1)); \ + a_lo = _mm_load_pd(left_values + (2 * 2 + 0)); \ + a_hi = _mm_load_pd(left_values + (2 * 2 + 1)); \ b = _mm_shuffle_pd(b_hi, b_hi, 0x0); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ \ /* froup = 3 */ \ - a_lo = _mm_load_pd(left_cur + (3 * 2 + 0)); \ - a_hi = _mm_load_pd(left_cur + (3 * 2 + 1)); \ + a_lo = _mm_load_pd(left_values + (3 * 2 + 0)); \ + a_hi = _mm_load_pd(left_values + (3 * 2 + 1)); \ b = _mm_shuffle_pd(b_hi, b_hi, 0x3); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ @@ -204,20 +210,23 @@ SN_MK_PROC_DECL_STD(double, sse, , fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, double); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe + SN_MAT_KERNEL_LOCAL *restrict ans_cur = + ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe - register __m128d c0_lo = _mm_load_pd(ans_cur->f64_x + (0 * 2 + 0)); - register __m128d c0_hi = _mm_load_pd(ans_cur->f64_x + (0 * 2 + 1)); - register __m128d c1_lo = _mm_load_pd(ans_cur->f64_x + (1 * 2 + 0)); - register __m128d c1_hi = _mm_load_pd(ans_cur->f64_x + (1 * 2 + 1)); - register __m128d c2_lo = _mm_load_pd(ans_cur->f64_x + (2 * 2 + 0)); - register __m128d c2_hi = _mm_load_pd(ans_cur->f64_x + (2 * 2 + 1)); - register __m128d c3_lo = _mm_load_pd(ans_cur->f64_x + (3 * 2 + 0)); - register __m128d c3_hi = _mm_load_pd(ans_cur->f64_x + (3 * 2 + 1)); + register __m128d c0_lo = ans_cur->f64_x[0]; + register __m128d c0_hi = ans_cur->f64_x[1]; + register __m128d c1_lo = ans_cur->f64_x[2]; + register __m128d c1_hi = ans_cur->f64_x[3]; + register __m128d c2_lo = ans_cur->f64_x[4]; + register __m128d c2_hi = ans_cur->f64_x[5]; + register __m128d c3_lo = ans_cur->f64_x[6]; + register __m128d c3_hi = ans_cur->f64_x[7]; for (size_t z = 0; z < common_z; z++) { SN_MAT_KERNEL_LOCAL const *right_cur = right + GET_INDEX(common_z, x, z); SN_MAT_KERNEL_LOCAL const *left_cur = left + GET_INDEX(common_y, z, y); + const double *right_values = &right_cur->m[0][0]; + const double *left_values = &left_cur->m[0][0]; DOUBLE_KERNEL_FMA_CYCLE(0); @@ -228,21 +237,21 @@ SN_MK_PROC_DECL_STD(double, sse, , fma) DOUBLE_KERNEL_FMA_CYCLE(3); } - _mm_store_pd(ans_cur->f64_x + (0 * 2 + 0), c0_lo); - _mm_store_pd(ans_cur->f64_x + (0 * 2 + 1), c0_hi); - _mm_store_pd(ans_cur->f64_x + (1 * 2 + 0), c1_lo); - _mm_store_pd(ans_cur->f64_x + (1 * 2 + 1), c1_hi); - _mm_store_pd(ans_cur->f64_x + (2 * 2 + 0), c2_lo); - _mm_store_pd(ans_cur->f64_x + (2 * 2 + 1), c2_hi); - _mm_store_pd(ans_cur->f64_x + (3 * 2 + 0), c3_lo); - _mm_store_pd(ans_cur->f64_x + (3 * 2 + 1), c3_hi); + ans_cur->f64_x[0] = c0_lo; + ans_cur->f64_x[1] = c0_hi; + ans_cur->f64_x[2] = c1_lo; + ans_cur->f64_x[3] = c1_hi; + ans_cur->f64_x[4] = c2_lo; + ans_cur->f64_x[5] = c2_hi; + ans_cur->f64_x[6] = c3_lo; + ans_cur->f64_x[7] = c3_hi; } SN_MK_PROC_DECL_STD(double, avx, , fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, double); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe + SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y); // tKernelAddressUnsafe register __m256d c0 = _mm256_load_pd((double *)&ans_cur->f64_y[0]); register __m256d c1 = _mm256_load_pd((double *)&ans_cur->f64_y[1]); @@ -314,7 +323,7 @@ SN_MK_PROC_DECL_STD(double, avx, _fma, fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, double); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe + SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y); // GetKernelAddressUnsafe register __m256d c0 = _mm256_load_pd((double *)&ans_cur->f64_y[0]); register __m256d c1 = _mm256_load_pd((double *)&ans_cur->f64_y[1]); @@ -370,7 +379,7 @@ SN_MK_PROC_DECL_STD(double, 512, , fma) { __mk_param_std(TILE, LEFT_RIGHT_ANS, double); - SN_MAT_KERNEL_LOCAL const *ans_cur = ans + GET_INDEX(common_y, x, y); + SN_MAT_KERNEL_LOCAL *restrict ans_cur = ans + GET_INDEX(common_y, x, y); register __m512d c01 = _mm512_load_pd((double *)&ans_cur->f64_z[0]); register __m512d c23 = _mm512_load_pd((double *)&ans_cur->f64_z[1]); diff --git a/StgSharp.Native/sn-src/s.hash.x86.c b/StgSharp.Native/sn-src/s.hash.x86.c index 295b192d..88a36947 100644 --- a/StgSharp.Native/sn-src/s.hash.x86.c +++ b/StgSharp.Native/sn-src/s.hash.x86.c @@ -8,33 +8,33 @@ INTERNAL int SN_DECL city_hash_simplify_sse(byte const *str, int const length) { uint32_t hash = length, size = length << 1; - byte *end = str + size, *ptr = str; + byte const *end = str + size; + byte const *ptr = str; if (size < 4) { for (; ptr < end; ptr += 2) { hash = _mm_crc32_u16(hash, *ptr); } } else if (size < 16) { - byte *last = end - 4; for (; ptr < end; ptr += 4) { - hash = _mm_crc32_u32(hash, *(uint32_t *)ptr); + hash = _mm_crc32_u32(hash, SN_CONST_PTR_CAST(uint32_t, ptr)[0]); } for (; ptr < end; ptr++) { hash = _mm_crc32_u16(hash, *ptr); } } else { - byte *last_128 = end - 16; - byte *last_32 = end - 4; + byte const *last_128 = end - 16; + byte const *last_32 = end - 4; uint64_t hash64 = size; for (; ptr < last_128; ptr += 16) { - __m128i vec = _mm_loadu_si32((int *)ptr); + __m128i vec = _mm_loadu_si128((__m128i const *)ptr); hash64 = _mm_crc32_u64(hash64, _mm_extract_epi64(vec, 0)); hash64 = _mm_crc32_u64(hash64, _mm_extract_epi64(vec, 1)); } hash = _mm_crc32_u8(hash, hash64); hash = _mm_crc32_u8(hash, hash64 >> 32); for (; ptr < last_32; ptr += 4) { - hash = _mm_crc32_u32(hash, *(uint32_t *)ptr); + hash = _mm_crc32_u32(hash, SN_CONST_PTR_CAST(uint32_t, ptr)[0]); } for (; ptr < end; ptr += 2) { hash = _mm_crc32_u8(hash, *ptr); diff --git a/StgSharp.Native/sn-src/simdid.x86.c b/StgSharp.Native/sn-src/simdid.x86.c index 8cf545e8..4717149f 100644 --- a/StgSharp.Native/sn-src/simdid.x86.c +++ b/StgSharp.Native/sn-src/simdid.x86.c @@ -340,6 +340,16 @@ SN_API SIMDID SN_DECL sn_get_simd_level_local() return simdid; } +SN_API SIMDID SN_DECL sn_get_simd_level_global() +{ + if (global_simdid.mask) { + return global_simdid; + } + + global_simdid = sn_get_simd_level_local(); + return global_simdid; +} + #else SN_API SIMDID SN_DECL sn_get_simd_level_local() diff --git a/StgSharp.Native/sn-src/sk.fma.x86.c b/StgSharp.Native/sn-src/sk.fma.x86.c index 52433927..84e4cef5 100644 --- a/StgSharp.Native/sn-src/sk.fma.x86.c +++ b/StgSharp.Native/sn-src/sk.fma.x86.c @@ -14,6 +14,7 @@ SN_SK_PROC_DECL_STD(float, sse, , fma) { + __sk_param_std(float); register __m128 c0 = _mm_load_ps((float *)&ans->f32_x[0]); register __m128 c1 = _mm_load_ps((float *)&ans->f32_x[1]); register __m128 c2 = _mm_load_ps((float *)&ans->f32_x[2]); @@ -61,6 +62,7 @@ SN_SK_PROC_DECL_STD(float, sse, , fma) SN_SK_PROC_DECL_STD(float, avx, , fma) { + __sk_param_std(float); register __m256 c01 = _mm256_load_ps((float *)&ans->f32_y[0]); register __m256 c23 = _mm256_load_ps((float *)&ans->f32_y[1]); @@ -86,22 +88,23 @@ SN_SK_PROC_DECL_STD(float, avx, , fma) SN_SK_PROC_DECL_STD(float, avx, _fma, fma) { + __sk_param_std(float); register __m256 c01 = _mm256_load_ps((float *)&ans->f32_y[0]); register __m256 c23 = _mm256_load_ps((float *)&ans->f32_y[1]); register __m256 b01 = _mm256_load_ps((float *)&right->f32_y[0]); register __m256 b23 = _mm256_load_ps((float *)&right->f32_y[1]); - register __m256 a0 = _mm256_broadcast_ps((float *)&left->f32_x[0]); + register __m256 a0 = _mm256_broadcast_ps(&left->f32_x[0]); c01 = _mm256_fmadd_ps(a0, b01, c01); c23 = _mm256_fmadd_ps(a0, b23, c23); - register __m256 a1 = _mm256_broadcast_ps((float *)&left->f32_x[1]); + register __m256 a1 = _mm256_broadcast_ps(&left->f32_x[1]); c01 = _mm256_fmadd_ps(a1, b01, c01); c23 = _mm256_fmadd_ps(a1, b23, c23); - register __m256 a2 = _mm256_broadcast_ps((float *)&left->f32_x[2]); + register __m256 a2 = _mm256_broadcast_ps(&left->f32_x[2]); c01 = _mm256_fmadd_ps(a2, b01, c01); c23 = _mm256_fmadd_ps(a2, b23, c23); - register __m256 a3 = _mm256_broadcast_ps((float *)&left->f32_x[3]); + register __m256 a3 = _mm256_broadcast_ps(&left->f32_x[3]); c01 = _mm256_fmadd_ps(a3, b01, c01); c23 = _mm256_fmadd_ps(a3, b23, c23); @@ -111,6 +114,7 @@ SN_SK_PROC_DECL_STD(float, avx, _fma, fma) SN_SK_PROC_DECL_STD(float, 512, , fma) { + __sk_param_std(float); register __m512 c = _mm512_load_ps((float *)ans); // full right load：[kb_col0 | kb_col1 | kb_col2 | kb_col3] register __m512 b = _mm512_load_ps((float *)right); @@ -137,31 +141,31 @@ SN_SK_PROC_DECL_STD(float, 512, , fma) #define DOUBLE_KERNEL_FMA_CYCLE(i) \ do { \ /* k = 0 */ \ - register __m128d b_lo = _mm_load_pd(right + (2 * i + 0)); \ - register __m128d b_hi = _mm_load_pd(right + (2 * i + 1)); \ - register __m128d a_lo = _mm_load_pd(left + (0 * 2 + 0)); \ - register __m128d a_hi = _mm_load_pd(left + (0 * 2 + 1)); \ + register __m128d b_lo = _mm_load_pd(right_values + (2 * i + 0)); \ + register __m128d b_hi = _mm_load_pd(right_values + (2 * i + 1)); \ + register __m128d a_lo = _mm_load_pd(left_values + (0 * 2 + 0)); \ + register __m128d a_hi = _mm_load_pd(left_values + (0 * 2 + 1)); \ register __m128d b = _mm_shuffle_pd(b_lo, b_lo, 0x0); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ \ /* group = 1 */ \ - a_lo = _mm_load_pd(left + (1 * 2 + 0)); \ - a_hi = _mm_load_pd(left + (1 * 2 + 1)); \ + a_lo = _mm_load_pd(left_values + (1 * 2 + 0)); \ + a_hi = _mm_load_pd(left_values + (1 * 2 + 1)); \ b = _mm_shuffle_pd(b_lo, b_lo, 0x3); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ \ /* groupp = 2 */ \ - a_lo = _mm_load_pd(left + (2 * 2 + 0)); \ - a_hi = _mm_load_pd(left + (2 * 2 + 1)); \ + a_lo = _mm_load_pd(left_values + (2 * 2 + 0)); \ + a_hi = _mm_load_pd(left_values + (2 * 2 + 1)); \ b = _mm_shuffle_pd(b_hi, b_hi, 0x0); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ \ /* froup = 3 */ \ - a_lo = _mm_load_pd(left + (3 * 2 + 0)); \ - a_hi = _mm_load_pd(left + (3 * 2 + 1)); \ + a_lo = _mm_load_pd(left_values + (3 * 2 + 0)); \ + a_hi = _mm_load_pd(left_values + (3 * 2 + 1)); \ b = _mm_shuffle_pd(b_hi, b_hi, 0x3); \ c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \ c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b)); \ @@ -171,33 +175,38 @@ SN_SK_PROC_DECL_STD(float, 512, , fma) SN_SK_PROC_DECL_STD(double, sse, , fma) { - register __m128d c0_lo = _mm_load_pd(ans->f64_x + (0 * 2 + 0)); - register __m128d c0_hi = _mm_load_pd(ans->f64_x + (0 * 2 + 1)); - register __m128d c1_lo = _mm_load_pd(ans->f64_x + (1 * 2 + 0)); - register __m128d c1_hi = _mm_load_pd(ans->f64_x + (1 * 2 + 1)); - register __m128d c2_lo = _mm_load_pd(ans->f64_x + (2 * 2 + 0)); - register __m128d c2_hi = _mm_load_pd(ans->f64_x + (2 * 2 + 1)); - register __m128d c3_lo = _mm_load_pd(ans->f64_x + (3 * 2 + 0)); - register __m128d c3_hi = _mm_load_pd(ans->f64_x + (3 * 2 + 1)); + __sk_param_std(double); + const double *left_values = &left->m[0][0]; + const double *right_values = &right->m[0][0]; + + register __m128d c0_lo = ans->f64_x[0]; + register __m128d c0_hi = ans->f64_x[1]; + register __m128d c1_lo = ans->f64_x[2]; + register __m128d c1_hi = ans->f64_x[3]; + register __m128d c2_lo = ans->f64_x[4]; + register __m128d c2_hi = ans->f64_x[5]; + register __m128d c3_lo = ans->f64_x[6]; + register __m128d c3_hi = ans->f64_x[7]; DOUBLE_KERNEL_FMA_CYCLE(0); DOUBLE_KERNEL_FMA_CYCLE(1); DOUBLE_KERNEL_FMA_CYCLE(2); DOUBLE_KERNEL_FMA_CYCLE(3); - _mm_store_pd(ans->f64_x + (0 * 2 + 0), c0_lo); - _mm_store_pd(ans->f64_x + (0 * 2 + 1), c0_hi); - _mm_store_pd(ans->f64_x + (1 * 2 + 0), c1_lo); - _mm_store_pd(ans->f64_x + (1 * 2 + 1), c1_hi); - _mm_store_pd(ans->f64_x + (2 * 2 + 0), c2_lo); - _mm_store_pd(ans->f64_x + (2 * 2 + 1), c2_hi); - _mm_store_pd(ans->f64_x + (3 * 2 + 0), c3_lo); - _mm_store_pd(ans->f64_x + (3 * 2 + 1), c3_hi); + ans->f64_x[0] = c0_lo; + ans->f64_x[1] = c0_hi; + ans->f64_x[2] = c1_lo; + ans->f64_x[3] = c1_hi; + ans->f64_x[4] = c2_lo; + ans->f64_x[5] = c2_hi; + ans->f64_x[6] = c3_lo; + ans->f64_x[7] = c3_hi; } // clang-format off SN_SK_PROC_DECL_STD(double, avx, , fma) { + __sk_param_std(double); register __m256d c0 = _mm256_load_pd((double *)&ans->f64_y[0]); register __m256d c1 = _mm256_load_pd((double *)&ans->f64_y[1]); register __m256d c2 = _mm256_load_pd((double *)&ans->f64_y[2]); @@ -246,6 +255,7 @@ SN_SK_PROC_DECL_STD(double, avx, , fma) SN_SK_PROC_DECL_STD(double, avx, _fma, fma) { + __sk_param_std(double); register __m256d c0 = _mm256_load_pd((double *)&ans->f64_y[0]); register __m256d c1 = _mm256_load_pd((double *)&ans->f64_y[1]); register __m256d c2 = _mm256_load_pd((double *)&ans->f64_y[2]); @@ -293,6 +303,7 @@ SN_SK_PROC_DECL_STD(double, avx, _fma, fma) SN_SK_PROC_DECL_STD(double, 512, , fma) { + __sk_param_std(double); register __m512d c01 = _mm512_load_pd((double *)&ans->f64_z[0]); register __m512d c23 = _mm512_load_pd((double *)&ans->f64_z[1]); diff --git a/StgSharp.Native/sn-src/sn_native.c b/StgSharp.Native/sn-src/sn_native.c index 9e2876bd..2a62c7ce 100644 --- a/StgSharp.Native/sn-src/sn_native.c +++ b/StgSharp.Native/sn-src/sn_native.c @@ -44,7 +44,7 @@ SN_API GLFWglproc SN_DECL loadGlfuncDefault(char *procName) return (GLFWglproc)ret; } -SN_API unsigned int linkShaderProgram(GladGLContext *context, GLuint shaderProgram) +SN_API unsigned int SN_DECL linkShaderProgram(GladGLContext *context, GLuint shaderProgram) { context->LinkProgram(shaderProgram); diff --git a/StgSharp.Native/sn-src/thread.linux.c b/StgSharp.Native/sn-src/thread.linux.c index 566338bd..e0eb9f68 100644 --- a/StgSharp.Native/sn-src/thread.linux.c +++ b/StgSharp.Native/sn-src/thread.linux.c @@ -29,11 +29,18 @@ SN_API void SN_DECL sn_set_thread_affinity(int core_id) CPU_FREE(set); } -SN_API void SN_DECL sn_get_numa_overall(int *out_numa_count, int *count_of_each_numa) +SN_API int SN_DECL sn_get_numa_overall(int *out_numa_count, int *count_of_each_numa) { + (void)out_numa_count; + (void)count_of_each_numa; + return 0; } -SN_API void SN_DECL sn_get_numa_each(int numa_id, int *in_numa_cpuid, int *global_cpuid) +SN_API int SN_DECL sn_get_numa_each(int numa_id, uint64_t *in_numa_cpuid, uint64_t *global_cpuid) { + (void)numa_id; + (void)in_numa_cpuid; + (void)global_cpuid; + return 0; } #endif diff --git a/StgSharp.Native/sn-src/vector.c b/StgSharp.Native/sn-src/vector.c index 7f6267a7..e7012802 100644 --- a/StgSharp.Native/sn-src/vector.c +++ b/StgSharp.Native/sn-src/vector.c @@ -3,7 +3,7 @@ INTERNAL void SN_DECL f32_normalize(sn_vec_f32 *restrict source, sn_vec_f32 *restrict target) { - register __m128 s = _mm_load_ps(&source->f32_x); + register __m128 s = source->f32_x[0]; register __m128 a_length_sq = _mm_mul_ps(s, s); a_length_sq = _mm_add_ps(a_length_sq, _mm_shuffle_ps(a_length_sq, a_length_sq, _MM_SHUFFLE(2, 3, 0, 1)));