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

matrix mul matrix is full function

Browse Source
This commit is contained in:
Nitload
2025-12-16 23:17:30 +08:00
parent 6344242b3c 88ed5a95c5
commit b13adc1a1b
26 changed files with 1049 additions and 861 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+16
View File
@@ -141,12 +141,18 @@ typedef void(SN_DECL *PNL_PROC_RAS)(mat_panel *right, mat_panel *ans, scalar_pac
#define PNL_PROC(T_type, T_arch, T_op_name) T_type##_pnl_##T_op_name##_##T_arch
<<<<<<< HEAD
<<<<<<< HEAD
#define DECLARE_PNL_PROC_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) *restrict ans)
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
#define DECLARE_PNL_PROC_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) *restrict ans)
>>>>>>> stgsharp-dev/giga
#define DECLARE_PNL_PROC_LEFT_RIGHT_ANS(T_type, T_arch, T_op_name) \
INTERNAL void SN_DECL PNL_PROC(T_type, T_arch, \
T_op_name)(MAT_PANEL_##T_arch##(T_type) const *left, \
@@ -230,6 +236,9 @@ DECLARE_BUF_PROC_ANS_SCALAR(float, 512, fill);
#pragma region fma
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_ANS(float, sse, clear);
DECLARE_PNL_PROC_ANS(float, avx, clear);
DECLARE_PNL_PROC_ANS(float, 512, clear);
@@ -238,10 +247,13 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, 512, fma);
<<<<<<< HEAD
=======
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma);
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
#pragma endregion
@@ -283,10 +295,14 @@ typedef struct sn_intrinsic {
BUF_PROC_LRA f32_buffer_sub;
KER_PROC_RA f32_buffer_transpose;
PANEL_PROC f32_build_panel;
<<<<<<< HEAD
<<<<<<< HEAD
PANEL_PROC f32_clear_panel;
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
PANEL_PROC f32_clear_panel;
>>>>>>> stgsharp-dev/giga
VECTORNORMALIZEPROC f32_normalize_3;
PNL_PROC_LRA f32_panel_fma;
PANEL_PROC f32_store_panel;
StgSharp.Native/sn-head/sn_matkernel.h
+12 -2
View File
@@ -110,14 +110,15 @@ extern __declspec(align(64)) sn_zero_kernel_u64 SN_ZERO_KERNEL;
MAT_KERNEL(element) const *const matrix, int col_length, int row_length, \
int col_index, int row_index)
#define DECLARE_STORE_PANEL(arch, element) \
INTERNAL void SN_DECL STORE_PANEL(arch, element)(MAT_KERNEL(element) *restrict matrix, \
MAT_PANEL_##arch(element) \
INTERNAL void SN_DECL STORE_PANEL(arch, element)(MAT_PANEL_##arch(element) \
const *const panel, \
MAT_KERNEL(element) *restrict matrix, \
int col_length, int row_length, \
int col_index, int row_index)
/* Declarations for AVX float variant */
<<<<<<< HEAD
<<<<<<< HEAD
DECLARE_BUILD_PANEL(sse, float);
DECLARE_BUILD_PANEL(avx, float);
DECLARE_BUILD_PANEL(512, float);
@@ -125,8 +126,17 @@ DECLARE_STORE_PANEL(sse, float);
DECLARE_STORE_PANEL(avx, float);
DECLARE_STORE_PANEL(512, float);
=======
=======
DECLARE_BUILD_PANEL(sse, float);
>>>>>>> stgsharp-dev/giga
DECLARE_BUILD_PANEL(avx, float);
DECLARE_BUILD_PANEL(512, float);
DECLARE_STORE_PANEL(sse, float);
DECLARE_STORE_PANEL(avx, float);
<<<<<<< HEAD
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
DECLARE_STORE_PANEL(512, float);
>>>>>>> stgsharp-dev/giga
#endif /* SN_MATKERNEL */
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_basicop.c
+17
View File
@@ -1,6 +1,9 @@
#include "sn_intrinsic.h"
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_ANS(float, sse, clear)
{
register __m128 zero = _mm_setzero_ps();
@@ -13,15 +16,22 @@ DECLARE_PNL_PROC_ANS(float, sse, clear)
DECLARE_PNL_PROC_ANS(float, avx, clear)
{
register __m256 zero = _mm256_setzero_ps();
<<<<<<< HEAD
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
PANEL_COL_VEC((*ans), 3, 0) = zero;
=======
for (size_t i = 0; i < 8; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
>>>>>>> stgsharp-dev/giga
}
DECLARE_PNL_PROC_ANS(float, 512, clear)
{
register __m512 zero = _mm512_setzero_ps();
<<<<<<< HEAD
PANEL_COL_VEC((*ans), 0, 0) = zero;
PANEL_COL_VEC((*ans), 1, 0) = zero;
PANEL_COL_VEC((*ans), 2, 0) = zero;
@@ -30,6 +40,13 @@ DECLARE_PNL_PROC_ANS(float, 512, clear)
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
for (size_t i = 0; i < 16; i++) {
PANEL_COL_VEC((*ans), i, 0) = zero;
}
}
>>>>>>> stgsharp-dev/giga
// Matrix transpose implementations
DECLARE_KER_PROC_RIGHT_ANS(float, sse, transpose)
{
StgSharp.Native/sn-src/k_fma.c
+6
View File
@@ -55,6 +55,9 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma)
// ans = left * right + ans
// right is not transposed
<<<<<<< HEAD
<<<<<<< HEAD
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
@@ -98,8 +101,11 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
// ans = left * right + ans
// right is not transposed
<<<<<<< HEAD
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx, fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
StgSharp.Native/sn-src/k_panel.c
+54 -8
View File
@@ -8,16 +8,16 @@ DECLARE_BUILD_PANEL(avx, float)
const int has_right = (col_index + 1) < col_length;
const int has_bottom = (row_index + 1) < row_length;
const MAT_KERNEL(float) *const k_lt = matrix + (col_index * row_length + row_index);
const MAT_KERNEL(float) *const k_lt = matrix + (row_index * col_length + col_index);
const MAT_KERNEL(float) *const k_rt =
has_right ? (matrix + ((col_index + 1) * row_length + row_index)) :
has_right ? (matrix + (row_index * col_length + (col_index + 1))) :
(const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
const MAT_KERNEL(float) *const k_lb =
has_bottom ? (matrix + (col_index * row_length + (row_index + 1))) :
has_bottom ? (matrix + ((row_index + 1) * col_length + col_index)) :
(const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
const MAT_KERNEL(float) *const k_rb =
(has_right && has_bottom) ?
(matrix + ((col_index + 1) * row_length + (row_index + 1))) :
(matrix + ((row_index + 1) * col_length + (col_index + 1))) :
(const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
register __m128 top128, bot128;
@@ -41,11 +41,18 @@ DECLARE_BUILD_PANEL(avx, float)
}
}
<<<<<<< HEAD
<<<<<<< HEAD
DECLARE_BUILD_PANEL(sse, float)
{
/* Single kernel -> 4 columns, each is one __m128 column vector. */
const MAT_KERNEL(float) *src = matrix + (col_index * row_length + row_index);
=======
DECLARE_BUILD_PANEL(sse, float)
{
/* Single kernel -> 4 columns, each is one __m128 column vector. */
const MAT_KERNEL(float) *src = matrix + (row_index * col_length + col_index);
>>>>>>> stgsharp-dev/giga
for (int pc = 0; pc < 4; ++pc) {
PANEL_COL_VEC((*panel), pc, 0) = src->xmm[pc];
@@ -55,22 +62,35 @@ DECLARE_BUILD_PANEL(sse, float)
DECLARE_BUILD_PANEL(512, float)
{
const MAT_KERNEL(float) *const zero = (const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
<<<<<<< HEAD
const MAT_KERNEL(float) *k[4][4];
=======
const MAT_KERNEL(float) * k[4][4];
>>>>>>> stgsharp-dev/giga
for (int kc = 0; kc < 4; ++kc) {
for (int kr = 0; kr < 4; ++kr) {
const int col = col_index + kc;
const int row = row_index + kr;
k[kc][kr] = (col < col_length && row < row_length) ?
<<<<<<< HEAD
(matrix + (col * row_length + row)) :
=======
(matrix + (row * col_length + col)) :
>>>>>>> stgsharp-dev/giga
zero;
}
}
/* Each panel column corresponds to a kernel column (0..3) within a block col (0..3). */
for (int pc = 0; pc < 16; ++pc) {
<<<<<<< HEAD
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 */
>>>>>>> stgsharp-dev/giga
const __m128 r0 = k[blk_col][0]->xmm[kc_col];
const __m128 r1 = k[blk_col][1]->xmm[kc_col];
@@ -87,8 +107,11 @@ DECLARE_BUILD_PANEL(512, float)
}
}
<<<<<<< HEAD
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
DECLARE_STORE_PANEL(avx, float)
{
const int Q = 8; /* fixed for AVX float panel */
@@ -100,19 +123,19 @@ DECLARE_STORE_PANEL(avx, float)
const int has_bottom = (row_index + 1) < row_length;
MAT_KERNEL(float) *const k_lt = (col_index < col_length && row_index < row_length) ?
(matrix + (col_index * row_length + row_index)) :
(matrix + (row_index * col_length + col_index)) :
&zero;
MAT_KERNEL(float) *const k_rt =
has_right && (row_index < row_length) ?
(matrix + ((col_index + 1) * row_length + row_index)) :
(matrix + (row_index * col_length + (col_index + 1))) :
&zero;
MAT_KERNEL(float) *const k_lb =
has_bottom && (col_index < col_length) ?
(matrix + (col_index * row_length + (row_index + 1))) :
(matrix + ((row_index + 1) * col_length + col_index)) :
&zero;
MAT_KERNEL(float) *const k_rb =
(has_right && has_bottom) ?
(matrix + ((col_index + 1) * row_length + (row_index + 1))) :
(matrix + ((row_index + 1) * col_length + (col_index + 1))) :
&zero;
register __m128 top128, bot128;
@@ -145,10 +168,17 @@ DECLARE_STORE_PANEL(avx, float)
}
}
<<<<<<< HEAD
<<<<<<< HEAD
DECLARE_STORE_PANEL(sse, float)
{
MAT_KERNEL(float) *dst = matrix + (col_index * row_length + row_index);
=======
DECLARE_STORE_PANEL(sse, float)
{
MAT_KERNEL(float) *dst = matrix + (row_index * col_length + col_index);
>>>>>>> stgsharp-dev/giga
for (int pc = 0; pc < 4; ++pc) {
const __m128 col = PANEL_COL_VEC((*panel), pc, 0);
@@ -159,21 +189,34 @@ DECLARE_STORE_PANEL(sse, float)
DECLARE_STORE_PANEL(512, float)
{
MAT_KERNEL(float) zero;
<<<<<<< HEAD
MAT_KERNEL(float) *k[4][4];
=======
MAT_KERNEL(float) * k[4][4];
>>>>>>> stgsharp-dev/giga
for (int kc = 0; kc < 4; ++kc) {
for (int kr = 0; kr < 4; ++kr) {
const int col = col_index + kc;
const int row = row_index + kr;
k[kc][kr] = (col < col_length && row < row_length) ?
<<<<<<< HEAD
(matrix + (col * row_length + row)) :
=======
(matrix + (row * col_length + col)) :
>>>>>>> stgsharp-dev/giga
&zero;
}
}
for (int pc = 0; pc < 16; ++pc) {
<<<<<<< HEAD
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 */
>>>>>>> stgsharp-dev/giga
const __m512 col = PANEL_COL_VEC((*panel), pc, 0);
@@ -191,5 +234,8 @@ DECLARE_STORE_PANEL(512, float)
k[blk_col][3]->xmm[kc_col] = r3;
}
}
<<<<<<< HEAD
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
>>>>>>> stgsharp-dev/giga
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/load_intrinsic.c
+24
View File
@@ -65,6 +65,7 @@ void load_intrinsic_sse(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
<<<<<<< HEAD
<<<<<<< HEAD
context->f32_clear_panel = PNL_PROC(float, sse, clear);
context->f32_build_panel = BUILD_PANEL(sse, float);
@@ -74,6 +75,13 @@ void load_intrinsic_sse(sn_intrinsic *context)
context->f32_buffer_add = BUF_PROC(float, sse, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
context->f32_store_panel = STORE_PANEL(sse, float);
context->f32_clear_panel = PNL_PROC(float, sse, clear);
context->f32_build_panel = BUILD_PANEL(sse, float);
context->f32_buffer_add = BUF_PROC(float, sse, add);
context->f32_panel_fma = PNL_PROC(float, sse, fma);
>>>>>>> stgsharp-dev/giga
context->f32_buffer_fill = BUF_PROC(float, sse, fill);
context->variant = BUF_PROC(float, sse, scalar_mul);
context->f32_buffer_sub = BUF_PROC(float, sse, sub);
@@ -87,6 +95,7 @@ void load_intrinsic_avx(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
<<<<<<< HEAD
<<<<<<< HEAD
context->f32_clear_panel = PNL_PROC(float, avx, clear);
context->f32_build_panel = BUILD_PANEL(avx, float);
@@ -96,6 +105,13 @@ void load_intrinsic_avx(sn_intrinsic *context)
context->f32_buffer_add = BUF_PROC(float, avx, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
context->f32_store_panel = STORE_PANEL(avx, float);
context->f32_clear_panel = PNL_PROC(float, avx, clear);
context->f32_build_panel = BUILD_PANEL(avx, float);
context->f32_buffer_add = BUF_PROC(float, avx, add);
context->f32_panel_fma = PNL_PROC(float, avx, fma);
>>>>>>> stgsharp-dev/giga
context->f32_buffer_fill = BUF_PROC(float, avx, fill);
context->variant = BUF_PROC(float, avx, scalar_mul);
context->f32_buffer_sub = BUF_PROC(float, avx, sub);
@@ -109,6 +125,7 @@ void load_intrinsic_512(sn_intrinsic *context)
context->city_hash_simplify = city_hash_simplify_sse;
context->factorial_simd = factorial_simd_sse;
context->index_pair = index_pair_sse;
<<<<<<< HEAD
<<<<<<< HEAD
context->f32_clear_panel = PNL_PROC(float, sse, clear);
context->f32_build_panel = BUILD_PANEL(512, float);
@@ -118,6 +135,13 @@ void load_intrinsic_512(sn_intrinsic *context)
context->f32_buffer_add = BUF_PROC(float, 512, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
=======
context->f32_store_panel = STORE_PANEL(512, float);
context->f32_clear_panel = PNL_PROC(float, 512, clear);
context->f32_build_panel = BUILD_PANEL(512, float);
context->f32_buffer_add = BUF_PROC(float, 512, add);
context->f32_panel_fma = PNL_PROC(float, 512, fma);
>>>>>>> stgsharp-dev/giga
context->f32_buffer_fill = BUF_PROC(float, 512, fill);
context->variant = BUF_PROC(float, 512, scalar_mul);
context->f32_buffer_sub = BUF_PROC(float, 512, sub);
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));
}
}