youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Improve Swiss rehash and NUMA topology handling

Browse Source
This commit is contained in:
Youmu Liu
2026-10-15 18:15:16 +08:00
parent c64f7ae9f6
commit 9867e1f262
10 changed files with 189 additions and 121 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/Nitload.Native/CMakeLists.txt
+8 -31
View File
@@ -17,39 +17,16 @@ nfn_resolve_target_identifiers(
NFN_RESOLVED_ARCH_NAME
NFN_RESOLVED_ARCH_DEFINE)
set(NFN_NATIVE_SOURCES
${CMAKE_CURRENT_LIST_DIR}/nfn-src/dllmain.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/load_glfw.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/load_intrinsic.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/m.bufop.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/m.solve.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/m.transpose.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/mk.fma.pack.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/mk.fma.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/mk.p_scan.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/s.factorial.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/s.hash.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/simdid_compare.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/simdid.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/sk.fma.x86.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/thread.linux.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/thread.windows.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/vector.c
${CMAKE_CURRENT_LIST_DIR}/nfn-src/vector_dot.c)
# CONFIGURE_DEPENDS makes CMake reconfigure when files are added or removed,
# so native sources, headers, and templates need no per-file CMake edits.
file(GLOB_RECURSE NFN_NATIVE_SOURCES CONFIGURE_DEPENDS
${CMAKE_CURRENT_LIST_DIR}/nfn-src/*.c)
set(NFN_NATIVE_HEADERS
${CMAKE_CURRENT_LIST_DIR}/nfn-head/glfw_function.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_internal.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.context.base.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.context.x86.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.matkernel.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_intrinsic.std.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/nfn_target.h
${CMAKE_CURRENT_LIST_DIR}/nfn-head/NitloadFrameWork.h)
file(GLOB_RECURSE NFN_NATIVE_HEADERS CONFIGURE_DEPENDS
${CMAKE_CURRENT_LIST_DIR}/nfn-head/*.h)
set(NFN_NATIVE_TEMPLATES
${CMAKE_CURRENT_LIST_DIR}/nfn-template/nfn_template.fma.x86.h)
file(GLOB_RECURSE NFN_NATIVE_TEMPLATES CONFIGURE_DEPENDS
${CMAKE_CURRENT_LIST_DIR}/nfn-template/*.h)
# Nitload.Native now lives under <root>/src/Nitload.Native, so walk up two
# levels (project dir -> src -> solution root) to reach the solution root.
src/Nitload.Native/nfn-head/NitloadFrameWork.h
+10 -8
View File
@@ -19,17 +19,19 @@
#ifndef NFN_
#endif
NFN_API void NFN_DECL nfn_load_glfw_functions(GLFWFunctionTable *table);
NFN_API void NFN_DECL nfn_load_intrinsic_function(void *intrinsic_context, uint64_t id);
NFN_API void NFN_DECL nfn_load_glfw_functions(GLFWFunctionTable *table);
NFN_API void NFN_DECL nfn_load_intrinsic_function(void *intrinsic_context, uint64_t id);
NFN_API SIMDID NFN_DECL nfn_get_simd_level_global();
/* Per-core variant: always re-executes CPUID, never caches.
* Must be called on the target core (after affinity is set). */
NFN_API SIMDID NFN_DECL nfn_get_simd_level_local();
NFN_API SIMDID NFN_DECL nfn_get_unite_simd(SIMDID left, SIMDID right);
NFN_API int NFN_DECL nfn_compare_simd(SIMDID left, SIMDID right);
NFN_API void NFN_DECL nfn_set_thread_affinity(int core_id, int numa_id);
NFN_API void* NFN_DECL nfn_numa_alloc(size_t size, int numa_id);
NFN_API void NFN_DECL nfn_numa_free(void *ptr, size_t size, int numa_id);
NFN_API int NFN_DECL nfn_get_numa_each(int numa_id, uint64_t *internal, uint32_t *global);
NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_each_numa);
NFN_API int NFN_DECL nfn_compare_simd(SIMDID left, SIMDID right);
NFN_API void NFN_DECL nfn_set_thread_affinity(int cpu_index, int processor_group_id);
NFN_API void *NFN_DECL nfn_numa_alloc(size_t size, int numa_id);
NFN_API void NFN_DECL nfn_numa_free(void *ptr, size_t size, int numa_id);
NFN_API int NFN_DECL nfn_get_numa_each(int numa_id, uint32_t *processor_group_ids,
uint32_t *socket_ids, uint32_t *cpu_indices,
uint32_t *numa_node_ids);
NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_each_numa);
#endif
src/Nitload.Native/nfn-src/thread.linux.c
+14 -42
View File
@@ -26,18 +26,12 @@ typedef struct nfn_linux_max_context {
int max_value;
} nfn_linux_max_context;
<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
typedef struct nfn_linux_fill_cpu_context {
uint64_t *internal;
uint32_t *global;
=======
typedef struct nif_linux_fill_cpu_context {
uint32_t *processor_group_ids;
uint32_t *socket_ids;
uint32_t *cpu_indices;
uint32_t *numa_node_ids;
int numa_id;
>>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
int index;
} nfn_linux_fill_cpu_context;
@@ -127,12 +121,7 @@ static int nfn_linux_track_max_id(int value, void *context)
return 1;
}
<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
static int nfn_linux_fill_cpu_id(int value, void *context)
{
nfn_linux_fill_cpu_context *fill_context = (nfn_linux_fill_cpu_context *)context;
=======
static int nif_linux_read_socket_id(int cpu_id, uint32_t *socket_id)
static int nfn_linux_read_socket_id(int cpu_id, uint32_t *socket_id)
{
char path[160];
char buffer[64];
@@ -140,25 +129,24 @@ static int nif_linux_read_socket_id(int cpu_id, uint32_t *socket_id)
long parsed_value;
snprintf(path, sizeof(path),
"/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu_id);
if (!nif_linux_read_first_line(path, buffer, sizeof(buffer)))
if (!nfn_linux_read_first_line(path, buffer, sizeof(buffer)))
return 0;
errno = 0;
parsed_value = strtol(buffer, &end, 10);
if (end == buffer || errno == ERANGE || parsed_value < 0 || parsed_value > UINT32_MAX)
if (end == buffer || errno == ERANGE || parsed_value < 0 || (uintmax_t)parsed_value > UINT32_MAX)
return 0;
*socket_id = (uint32_t)parsed_value;
return 1;
}
static int nif_linux_fill_cpu_id(int value, void *context)
static int nfn_linux_fill_cpu_id(int value, void *context)
{
nif_linux_fill_cpu_context *fill_context = (nif_linux_fill_cpu_context *)context;
nfn_linux_fill_cpu_context *fill_context = (nfn_linux_fill_cpu_context *)context;
uint32_t socket_id = 0;
if (!nif_linux_read_socket_id(value, &socket_id))
if (!nfn_linux_read_socket_id(value, &socket_id))
return 0;
>>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
fill_context->processor_group_ids[fill_context->index] = 0u;
fill_context->socket_ids[fill_context->index] = socket_id;
@@ -249,15 +237,8 @@ static int nfn_linux_populate_writable_mapping(void *address, size_t size)
#endif
}
<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
/* Pin the calling thread to a single logical processor.
* core_id >= 0 : bind to that logical processor (global index, across groups).
* core_id < 0 : clear affinity constraint within the current group. */
NFN_API void NFN_DECL nfn_set_thread_affinity(int core_id, int numa_id)
=======
/* Linux has no Windows-style processor groups; processor_group_id is ignored. */
NIF_API void NIF_DECL nif_set_thread_affinity(int cpu_index, int processor_group_id)
>>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
NFN_API void NFN_DECL nfn_set_thread_affinity(int cpu_index, int processor_group_id)
{
(void)processor_group_id;
@@ -311,15 +292,11 @@ NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_eac
return 1;
}
<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
NFN_API int NFN_DECL nfn_get_numa_each(int numa_id, uint64_t *internal, uint32_t *global)
=======
NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
uint32_t *processor_group_ids,
uint32_t *socket_ids,
uint32_t *cpu_indices,
uint32_t *numa_node_ids)
>>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
NFN_API int NFN_DECL nfn_get_numa_each(int numa_id,
uint32_t *processor_group_ids,
uint32_t *socket_ids,
uint32_t *cpu_indices,
uint32_t *numa_node_ids)
{
if (numa_id < 0 || processor_group_ids == NULL || socket_ids == NULL ||
cpu_indices == NULL || numa_node_ids == NULL)
@@ -331,11 +308,7 @@ NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
if (!nfn_linux_read_first_line(path, buffer, sizeof(buffer)))
return 0;
<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.linux.c
nfn_linux_fill_cpu_context fill_context = { internal, global, 0 };
return nfn_linux_visit_id_list(buffer, nfn_linux_fill_cpu_id, &fill_context);
=======
nif_linux_fill_cpu_context fill_context = {
nfn_linux_fill_cpu_context fill_context = {
processor_group_ids,
socket_ids,
cpu_indices,
@@ -343,8 +316,7 @@ NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
numa_id,
0
};
return nif_linux_visit_id_list(buffer, nif_linux_fill_cpu_id, &fill_context);
>>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.linux.c
return nfn_linux_visit_id_list(buffer, nfn_linux_fill_cpu_id, &fill_context);
}
NFN_API void *NFN_DECL nfn_numa_alloc(size_t size, int numa_id)
src/Nitload.Native/nfn-src/thread.windows.c
+16 -23
View File
@@ -2,15 +2,12 @@
#include "NitloadFrameWork.h"
#if NFN_IS_OS(NFN_OS_WINDOWS)
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include <windows.h>
<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.windows.c
/* Pin the calling thread to a single logical processor.
* core_id >= 0 : bind to that logical processor (global index, across groups).
* core_id < 0 : clear affinity constraint within the current group. */
NFN_API void NFN_DECL nfn_set_thread_affinity(int core_id, int numa_id)
=======
static int nif_windows_get_socket_map(uint32_t **out_socket_map, DWORD *out_group_count)
>>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.windows.c
static int nfn_windows_get_socket_map(uint32_t **out_socket_map, DWORD *out_group_count)
{
DWORD group_count = GetActiveProcessorGroupCount();
DWORD buffer_size = 0;
@@ -71,7 +68,7 @@ static int nif_windows_get_socket_map(uint32_t **out_socket_map, DWORD *out_grou
return 1;
}
static int nif_windows_get_processor_node(WORD group, BYTE cpu, USHORT *node_id)
static int nfn_windows_get_processor_node(WORD group, BYTE cpu, USHORT *node_id)
{
PROCESSOR_NUMBER processor;
memset(&processor, 0, sizeof(processor));
@@ -81,7 +78,7 @@ static int nif_windows_get_processor_node(WORD group, BYTE cpu, USHORT *node_id)
}
/* Pin the calling thread using its processor group and group-local CPU index. */
NIF_API void NIF_DECL nif_set_thread_affinity(int cpu_index, int processor_group_id)
NFN_API void NFN_DECL nfn_set_thread_affinity(int cpu_index, int processor_group_id)
{
if (cpu_index < 0 || processor_group_id < 0) {
GROUP_AFFINITY current;
@@ -98,7 +95,7 @@ NIF_API void NIF_DECL nif_set_thread_affinity(int cpu_index, int processor_group
if ((DWORD)processor_group_id >= GetActiveProcessorGroupCount() || cpu_index >= 64)
return;
USHORT node_id = 0;
if (!nif_windows_get_processor_node((WORD)processor_group_id, (BYTE)cpu_index, &node_id))
if (!nfn_windows_get_processor_node((WORD)processor_group_id, (BYTE)cpu_index, &node_id))
return;
GROUP_AFFINITY affinity;
@@ -124,7 +121,7 @@ NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_eac
for (DWORD group = 0; group < group_count; group++) {
for (DWORD cpu = 0; cpu < 64; cpu++) {
USHORT node_id = 0;
if (nif_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) &&
if (nfn_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) &&
count_of_each_numa != NULL && node_id <= highest)
count_of_each_numa[node_id]++;
}
@@ -132,15 +129,11 @@ NFN_API int NFN_DECL nfn_get_numa_overall(int *out_numa_count, int *count_of_eac
return 1;
}
<<<<<<< HEAD:src/Nitload.Native/nfn-src/thread.windows.c
NFN_API int NFN_DECL nfn_get_numa_each(int numa_id, uint64_t *internal, uint32_t *global)
=======
NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
uint32_t *processor_group_ids,
uint32_t *socket_ids,
uint32_t *cpu_indices,
uint32_t *numa_node_ids)
>>>>>>> 3a2b1234782f893f5b71655289b372ffed8dfd93:src/Nitload.Native/nif-src/thread.windows.c
NFN_API int NFN_DECL nfn_get_numa_each(int numa_id,
uint32_t *processor_group_ids,
uint32_t *socket_ids,
uint32_t *cpu_indices,
uint32_t *numa_node_ids)
{
if (numa_id < 0 || processor_group_ids == NULL || socket_ids == NULL ||
cpu_indices == NULL || numa_node_ids == NULL)
@@ -148,14 +141,14 @@ NIF_API int NIF_DECL nif_get_numa_each(int numa_id,
uint32_t *socket_map = NULL;
DWORD group_count = 0;
if (!nif_windows_get_socket_map(&socket_map, &group_count))
if (!nfn_windows_get_socket_map(&socket_map, &group_count))
return 0;
int index = 0;
for (DWORD group = 0; group < group_count; group++) {
for (DWORD cpu = 0; cpu < 64; cpu++) {
USHORT node_id = 0;
if (!nif_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) ||
if (!nfn_windows_get_processor_node((WORD)group, (BYTE)cpu, &node_id) ||
node_id != (USHORT)numa_id)
continue;