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

working on matrix mul matrix, rebuilding task model defines in C#

Browse Source
This commit is contained in:
Youmu Liu
2025-12-12 18:00:49 +08:00
parent 9ff438d3b9
commit 2c7277f3f9
12 changed files with 223 additions and 444 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/dllmain.c
+1 -1
View File
@@ -1,4 +1,4 @@
#include "StgSharpNative.h"
#include "StgSharpNative.h"
#include "sn_matkernel.h"
#include <stdint.h>
StgSharp.Native/sn-src/k_panel.c
+6 -6
View File
@@ -54,7 +54,7 @@ DECLARE_BUILD_PANEL(sse, float)
DECLARE_BUILD_PANEL(512, float)
{
const MAT_KERNEL(float) *const zero = (const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
const MAT_KERNEL(float) *k[4][4];
const MAT_KERNEL(float) * k[4][4];
for (int kc = 0; kc < 4; ++kc) {
for (int kr = 0; kr < 4; ++kr) {
@@ -68,8 +68,8 @@ DECLARE_BUILD_PANEL(512, float)
/* Each panel column corresponds to a kernel column (0..3) within a block col (0..3). */
for (int pc = 0; pc < 16; ++pc) {
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
const __m128 r0 = k[blk_col][0]->xmm[kc_col];
const __m128 r1 = k[blk_col][1]->xmm[kc_col];
@@ -155,7 +155,7 @@ DECLARE_STORE_PANEL(sse, float)
DECLARE_STORE_PANEL(512, float)
{
MAT_KERNEL(float) zero;
MAT_KERNEL(float) *k[4][4];
MAT_KERNEL(float) * k[4][4];
for (int kc = 0; kc < 4; ++kc) {
for (int kr = 0; kr < 4; ++kr) {
@@ -168,8 +168,8 @@ DECLARE_STORE_PANEL(512, float)
}
for (int pc = 0; pc < 16; ++pc) {
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
const int blk_col = pc >> 2; /* which kernel column block (0..3) */
const int kc_col = pc & 3; /* column inside a 4x4 kernel */
const __m512 col = PANEL_COL_VEC((*panel), pc, 0);
StgSharp.Native/sn-src/load_glfw.c
+116 -116
View File
@@ -4,120 +4,120 @@
SN_API void SN_DECL load_glfw_functions(GLFWFunctionTable *table)
{
table->glfwCreateCursor = glfwCreateCursor;
table->glfwCreateStandardCursor = glfwCreateStandardCursor;
table->glfwCreateWindow = glfwCreateWindow;
table->glfwDefaultWindowHints = glfwDefaultWindowHints;
table->glfwDestroyCursor = glfwDestroyCursor;
table->glfwDestroyWindow = glfwDestroyWindow;
table->glfwExtensionSupported = glfwExtensionSupported;
table->glfwFocusWindow = glfwFocusWindow;
table->glfwGetClipboardString = glfwGetClipboardString;
table->glfwGetCurrentContext = glfwGetCurrentContext;
table->glfwGetCursorPos = glfwGetCursorPos;
table->glfwGetError = glfwGetError;
table->glfwGetFrameBufferSize = glfwGetFramebufferSize;
table->glfwGetGamepadName = glfwGetGamepadName;
table->glfwGetGamepadState = glfwGetGamepadState;
table->glfwGetGammaRamp = glfwGetGammaRamp;
table->glfwGetInputMode = glfwGetInputMode;
table->glfwGetJoystickAxes = glfwGetJoystickAxes;
table->glfwGetJoystickButtons = glfwGetJoystickButtons;
table->glfwGetJoystickGUID = glfwGetJoystickGUID;
table->glfwGetJoystickHats = glfwGetJoystickHats;
table->glfwGetJoystickName = glfwGetJoystickName;
table->glfwGetJoystickUserPointer = glfwGetJoystickUserPointer;
table->glfwGetKey = glfwGetKey;
table->glfwGetKeyName = glfwGetKeyName;
table->glfwGetKeyScancode = glfwGetKeyScancode;
table->glfwGetMonitorContentScale = glfwGetMonitorContentScale;
table->glfwGetMonitorName = glfwGetMonitorName;
table->glfwGetMonitorPhysicalSize = glfwGetMonitorPhysicalSize;
table->glfwGetMonitorPos = glfwGetMonitorPos;
table->glfwGetMonitors = glfwGetMonitors;
table->glfwGetMonitorUserPointer = glfwGetMonitorUserPointer;
table->glfwGetMonitorWorkarea = glfwGetMonitorWorkarea;
table->glfwGetMouseButton = glfwGetMouseButton;
table->glfwGetPrimaryMonitor = glfwGetPrimaryMonitor;
table->glfwGetProcAddress = glfwGetProcAddress;
table->glfwGetRequiredInstanceExtensions = glfwGetRequiredInstanceExtensions;
table->glfwGetTime = glfwGetTime;
table->glfwGetTimerFrequency = glfwGetTimerFrequency;
table->glfwGetTimerValue = glfwGetTimerValue;
table->glfwGetVersion = glfwGetVersion;
table->glfwGetVersionString = glfwGetVersionString;
table->glfwGetVideoMode = glfwGetVideoMode;
table->glfwGetVideoModes = glfwGetVideoModes;
table->glfwGetWindowAttrib = glfwGetWindowAttrib;
table->glfwGetWindowContentScale = glfwGetWindowContentScale;
table->glfwGetWindowFrameSize = glfwGetWindowFrameSize;
table->glfwGetWindowMonitor = glfwGetWindowMonitor;
table->glfwGetWindowOpacity = glfwGetWindowOpacity;
table->glfwGetWindowPos = glfwGetWindowPos;
table->glfwGetWindowSize = glfwGetWindowSize;
table->glfwGetWindowUserPointer = glfwGetWindowUserPointer;
table->glfwHideWindow = glfwHideWindow;
table->glfwIconifyWindow = glfwIconifyWindow;
table->glfwInit = glfwInit;
table->glfwInitHint = glfwInitHint;
table->glfwJoystickIsGamePad = glfwJoystickIsGamepad;
table->glfwJoystickPresent = glfwJoystickPresent;
table->glfwMakeContextCurrent = glfwMakeContextCurrent;
table->glfwMaximizeWindow = glfwMaximizeWindow;
table->glfwPollEvents = glfwPollEvents;
table->glfwPostEmptyEvent = glfwPostEmptyEvent;
table->glfwRawMouseMotionSupported = glfwRawMouseMotionSupported;
table->glfwRequestWindowAttention = glfwRequestWindowAttention;
table->glfwRestoreWindow = glfwRestoreWindow;
table->glfwSetCharCallback = glfwSetCharCallback;
table->glfwSetCharModsCallback = glfwSetCharModsCallback;
table->glfwSetClipboardString = glfwSetClipboardString;
table->glfwSetCursor = glfwSetCursor;
table->glfwSetCursorEnterCallback = glfwSetCursorEnterCallback;
table->glfwSetCursorPos = glfwSetCursorPos;
table->glfwSetCursorPosCallback = glfwSetCursorPosCallback;
table->glfwSetDropCallback = glfwSetDropCallback;
table->glfwSetErrorCallback = glfwSetErrorCallback;
table->glfwSetFramebufferSizeCallback = glfwSetFramebufferSizeCallback;
table->glfwSetGamma = glfwSetGamma;
table->glfwSetGammaRamp = glfwSetGammaRamp;
table->glfwSetInputMode = glfwSetInputMode;
table->glfwSetJoystickCallback = glfwSetJoystickCallback;
table->glfwSetJoystickUserPointer = glfwSetJoystickUserPointer;
table->glfwSetKeyCallback = glfwSetKeyCallback;
table->glfwSetMonitorCallback = glfwSetMonitorCallback;
table->glfwSetMonitorUserPointer = glfwSetMonitorUserPointer;
table->glfwSetMouseButtonCallback = glfwSetMouseButtonCallback;
table->glfwSetScrollCallback = glfwSetScrollCallback;
table->glfwSetTime = glfwSetTime;
table->glfwSetWindowAspectRatio = glfwSetWindowAspectRatio;
table->glfwSetWindowAttrib = glfwSetWindowAttrib;
table->glfwSetWindowCloseCallback = glfwSetWindowCloseCallback;
table->glfwSetWindowContentScaleCallback = glfwSetWindowContentScaleCallback;
table->glfwSetWindowFocusCallback = glfwSetWindowFocusCallback;
table->glfwSetWindowIcon = glfwSetWindowIcon;
table->glfwSetWindowIconifyCallback = glfwSetWindowIconifyCallback;
table->glfwSetWindowMaximizeCallback = glfwSetWindowMaximizeCallback;
table->glfwSetWindowMonitor = glfwSetWindowMonitor;
table->glfwSetWindowOpacity = glfwSetWindowOpacity;
table->glfwSetWindowPos = glfwSetWindowPos;
table->glfwSetWindowPosCallback = glfwSetWindowPosCallback;
table->glfwSetWindowRefreshCallback = glfwSetWindowRefreshCallback;
table->glfwSetWindowShouldClose = glfwSetWindowShouldClose;
table->glfwSetWindowSize = glfwSetWindowSize;
table->glfwSetWindowSizeCallback = glfwSetWindowSizeCallback;
table->glfwSetWindowSizeLimits = glfwSetWindowSizeLimits;
table->glfwSetWindowTitle = glfwSetWindowTitle;
table->glfwSetWindowUserPointer = glfwSetWindowUserPointer;
table->glfwShowWindow = glfwShowWindow;
table->glfwSwapBuffers = glfwSwapBuffers;
table->glfwSwapInterval = glfwSwapInterval;
table->glfwTerminate = glfwTerminate;
table->glfwUpdateGamePadMappings = glfwUpdateGamepadMappings;
table->glfwVulkanSupported = glfwVulkanSupported;
table->glfwWaitEvents = glfwWaitEvents;
table->glfwWaitEventsTimeout = glfwWaitEventsTimeout;
table->glfwWindowHint = glfwWindowHint;
table->glfwWindowHintString = glfwWindowHintString;
table->glfwWindowShouldClose = glfwWindowShouldClose;
table->glfwCreateCursor = glfwCreateCursor;
table->glfwCreateStandardCursor = glfwCreateStandardCursor;
table->glfwCreateWindow = glfwCreateWindow;
table->glfwDefaultWindowHints = glfwDefaultWindowHints;
table->glfwDestroyCursor = glfwDestroyCursor;
table->glfwDestroyWindow = glfwDestroyWindow;
table->glfwExtensionSupported = glfwExtensionSupported;
table->glfwFocusWindow = glfwFocusWindow;
table->glfwGetClipboardString = glfwGetClipboardString;
table->glfwGetCurrentContext = glfwGetCurrentContext;
table->glfwGetCursorPos = glfwGetCursorPos;
table->glfwGetError = glfwGetError;
table->glfwGetFrameBufferSize = glfwGetFramebufferSize;
table->glfwGetGamepadName = glfwGetGamepadName;
table->glfwGetGamepadState = glfwGetGamepadState;
table->glfwGetGammaRamp = glfwGetGammaRamp;
table->glfwGetInputMode = glfwGetInputMode;
table->glfwGetJoystickAxes = glfwGetJoystickAxes;
table->glfwGetJoystickButtons = glfwGetJoystickButtons;
table->glfwGetJoystickGUID = glfwGetJoystickGUID;
table->glfwGetJoystickHats = glfwGetJoystickHats;
table->glfwGetJoystickName = glfwGetJoystickName;
table->glfwGetJoystickUserPointer = glfwGetJoystickUserPointer;
table->glfwGetKey = glfwGetKey;
table->glfwGetKeyName = glfwGetKeyName;
table->glfwGetKeyScancode = glfwGetKeyScancode;
table->glfwGetMonitorContentScale = glfwGetMonitorContentScale;
table->glfwGetMonitorName = glfwGetMonitorName;
table->glfwGetMonitorPhysicalSize = glfwGetMonitorPhysicalSize;
table->glfwGetMonitorPos = glfwGetMonitorPos;
table->glfwGetMonitors = glfwGetMonitors;
table->glfwGetMonitorUserPointer = glfwGetMonitorUserPointer;
table->glfwGetMonitorWorkarea = glfwGetMonitorWorkarea;
table->glfwGetMouseButton = glfwGetMouseButton;
table->glfwGetPrimaryMonitor = glfwGetPrimaryMonitor;
table->glfwGetProcAddress = glfwGetProcAddress;
table->glfwGetRequiredInstanceExtensions = glfwGetRequiredInstanceExtensions;
table->glfwGetTime = glfwGetTime;
table->glfwGetTimerFrequency = glfwGetTimerFrequency;
table->glfwGetTimerValue = glfwGetTimerValue;
table->glfwGetVersion = glfwGetVersion;
table->glfwGetVersionString = glfwGetVersionString;
table->glfwGetVideoMode = glfwGetVideoMode;
table->glfwGetVideoModes = glfwGetVideoModes;
table->glfwGetWindowAttrib = glfwGetWindowAttrib;
table->glfwGetWindowContentScale = glfwGetWindowContentScale;
table->glfwGetWindowFrameSize = glfwGetWindowFrameSize;
table->glfwGetWindowMonitor = glfwGetWindowMonitor;
table->glfwGetWindowOpacity = glfwGetWindowOpacity;
table->glfwGetWindowPos = glfwGetWindowPos;
table->glfwGetWindowSize = glfwGetWindowSize;
table->glfwGetWindowUserPointer = glfwGetWindowUserPointer;
table->glfwHideWindow = glfwHideWindow;
table->glfwIconifyWindow = glfwIconifyWindow;
table->glfwInit = glfwInit;
table->glfwInitHint = glfwInitHint;
table->glfwJoystickIsGamePad = glfwJoystickIsGamepad;
table->glfwJoystickPresent = glfwJoystickPresent;
table->glfwMakeContextCurrent = glfwMakeContextCurrent;
table->glfwMaximizeWindow = glfwMaximizeWindow;
table->glfwPollEvents = glfwPollEvents;
table->glfwPostEmptyEvent = glfwPostEmptyEvent;
table->glfwRawMouseMotionSupported = glfwRawMouseMotionSupported;
table->glfwRequestWindowAttention = glfwRequestWindowAttention;
table->glfwRestoreWindow = glfwRestoreWindow;
table->glfwSetCharCallback = glfwSetCharCallback;
table->glfwSetCharModsCallback = glfwSetCharModsCallback;
table->glfwSetClipboardString = glfwSetClipboardString;
table->glfwSetCursor = glfwSetCursor;
table->glfwSetCursorEnterCallback = glfwSetCursorEnterCallback;
table->glfwSetCursorPos = glfwSetCursorPos;
table->glfwSetCursorPosCallback = glfwSetCursorPosCallback;
table->glfwSetDropCallback = glfwSetDropCallback;
table->glfwSetErrorCallback = glfwSetErrorCallback;
table->glfwSetFramebufferSizeCallback = glfwSetFramebufferSizeCallback;
table->glfwSetGamma = glfwSetGamma;
table->glfwSetGammaRamp = glfwSetGammaRamp;
table->glfwSetInputMode = glfwSetInputMode;
table->glfwSetJoystickCallback = glfwSetJoystickCallback;
table->glfwSetJoystickUserPointer = glfwSetJoystickUserPointer;
table->glfwSetKeyCallback = glfwSetKeyCallback;
table->glfwSetMonitorCallback = glfwSetMonitorCallback;
table->glfwSetMonitorUserPointer = glfwSetMonitorUserPointer;
table->glfwSetMouseButtonCallback = glfwSetMouseButtonCallback;
table->glfwSetScrollCallback = glfwSetScrollCallback;
table->glfwSetTime = glfwSetTime;
table->glfwSetWindowAspectRatio = glfwSetWindowAspectRatio;
table->glfwSetWindowAttrib = glfwSetWindowAttrib;
table->glfwSetWindowCloseCallback = glfwSetWindowCloseCallback;
table->glfwSetWindowContentScaleCallback = glfwSetWindowContentScaleCallback;
table->glfwSetWindowFocusCallback = glfwSetWindowFocusCallback;
table->glfwSetWindowIcon = glfwSetWindowIcon;
table->glfwSetWindowIconifyCallback = glfwSetWindowIconifyCallback;
table->glfwSetWindowMaximizeCallback = glfwSetWindowMaximizeCallback;
table->glfwSetWindowMonitor = glfwSetWindowMonitor;
table->glfwSetWindowOpacity = glfwSetWindowOpacity;
table->glfwSetWindowPos = glfwSetWindowPos;
table->glfwSetWindowPosCallback = glfwSetWindowPosCallback;
table->glfwSetWindowRefreshCallback = glfwSetWindowRefreshCallback;
table->glfwSetWindowShouldClose = glfwSetWindowShouldClose;
table->glfwSetWindowSize = glfwSetWindowSize;
table->glfwSetWindowSizeCallback = glfwSetWindowSizeCallback;
table->glfwSetWindowSizeLimits = glfwSetWindowSizeLimits;
table->glfwSetWindowTitle = glfwSetWindowTitle;
table->glfwSetWindowUserPointer = glfwSetWindowUserPointer;
table->glfwShowWindow = glfwShowWindow;
table->glfwSwapBuffers = glfwSwapBuffers;
table->glfwSwapInterval = glfwSwapInterval;
table->glfwTerminate = glfwTerminate;
table->glfwUpdateGamePadMappings = glfwUpdateGamepadMappings;
table->glfwVulkanSupported = glfwVulkanSupported;
table->glfwWaitEvents = glfwWaitEvents;
table->glfwWaitEventsTimeout = glfwWaitEventsTimeout;
table->glfwWindowHint = glfwWindowHint;
table->glfwWindowHintString = glfwWindowHintString;
table->glfwWindowShouldClose = glfwWindowShouldClose;
}
StgSharp.Native/sn-src/str_hash.c
+3 -3
View File
@@ -15,7 +15,7 @@ INTERNAL int SN_DECL city_hash_simplify_sse(byte const *str, int const length)
for (; ptr < end; ptr += 4) {
hash = _mm_crc32_u32(hash, *(uint32_t *)ptr);
}
for (; ptr < end; ptr++) {
for (; ptr < end; ptr++) {
hash = _mm_crc32_u16(hash, *ptr);
}
} else {
@@ -30,10 +30,10 @@ INTERNAL int SN_DECL city_hash_simplify_sse(byte const *str, int const length)
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, *(uint32_t *)ptr);
}
for (; ptr < end; ptr += 2) {
hash = _mm_crc32_u8(hash,*ptr);
hash = _mm_crc32_u8(hash, *ptr);
}
}
hash ^= hash >> 16;
StgSharp.Native/sn-src/vector_dot.c
+10 -10
View File
@@ -15,7 +15,8 @@ INTERNAL void FORCEINLINE SN_DECL dot_41_sse(MAT_KERNEL(float) const *transpose,
SSE_INSERT_SIMPLE(1, 4));
}
INTERNAL void SN_DECL dot_41_avx(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)
{
__m256 _double = _mm256_set_m128(vector[0], vector[0]);
__m256 tmp = _mm256_dp_ps(transpose->ymm[0], _double, 0B11110001);
@@ -25,7 +26,8 @@ INTERNAL void SN_DECL dot_41_avx(MAT_KERNEL(float) const *transpose, __m128 cons
_mm256_permutevar8x32_ps(tmp, _mm256_loadu_epi32(avxdot_indices)));
}
INTERNAL void SN_DECL dot_42_sse(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,
__m128 *ans)
{
ans[0] = _mm_dp_ps(transpose->xmm[0], vector[0], 0B11110001);
ans[0] = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[1], vector[0], 0B11110001),
@@ -44,24 +46,22 @@ INTERNAL void SN_DECL dot_42_sse(MAT_KERNEL(float) const *transpose, __m128 cons
SSE_INSERT_SIMPLE(1, 4));
}
INTERNAL void SN_DECL dot_42_avx(MAT_KERNEL(float) const *transpose, __m128 const *vector, __m128 *ans)
INTERNAL void SN_DECL dot_42_avx(MAT_KERNEL(float) const *transpose, __m128 const *vector,
__m128 *ans)
{
__m256 temp = transpose->ymm[0];
__m256 vector_stream = _mm256_set_m128(vector[1], vector[0]);
__m256 result = _mm256_dp_ps(temp, vector_stream, 0B11110001);
temp = _mm256_permute2f128_ps(temp, temp, 0B00000001);
result = _mm256_blend_ps(
result, _mm256_dp_ps(temp, vector_stream, 0B11110010), 0B00100010);
result = _mm256_blend_ps(result, _mm256_dp_ps(temp, vector_stream, 0B11110010), 0B00100010);
temp = transpose->ymm[1];
result = _mm256_blend_ps(
result, _mm256_dp_ps(temp, vector_stream, 0B11110100), 0B01000100);
result = _mm256_blend_ps(result, _mm256_dp_ps(temp, vector_stream, 0B11110100), 0B01000100);
temp = _mm256_permute2f128_ps(temp, temp, 0B00000001);
result = _mm256_blend_ps(
result, _mm256_dp_ps(temp, vector_stream, 0B11111000), 0B10001000);
result = _mm256_blend_ps(result, _mm256_dp_ps(temp, vector_stream, 0B11111000), 0B10001000);
ans[0] = _mm256_castps256_ps128(result);
ans[1] = _mm256_extractf128_ps(result, 1);
ans[1] = _mm_shuffle_ps(ans[1], ans[1], _MM_SHUFFLE(2, 3, 0, 1));
}
}