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

merge main branch

Browse Source
This commit is contained in:
Nitload
2025-12-16 23:16:57 +08:00
parent 3486375470 7bcb460a39
commit 6344242b3c
276 changed files with 53498 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/StgSharpC.c
+75
View File
@@ -0,0 +1,75 @@
#include "StgSharpC.h"
#include "gl.h"
#include "glfw3.h"
#include <stdio.h>
#include <string.h>
extern GladGLContext *currentContext;
extern char infolog[512];
SSCAPI int SSCDECL glCheckShaderStat(GladGLContext *context, uint64_t shaderHandle, int key,
char **logRef)
{
int stat = 0;
context->GetShaderiv(shaderHandle, key, &stat);
if (!stat) {
context->GetShaderInfoLog(shaderHandle, 512, NULL, infolog);
}
*logRef = infolog;
return stat;
}
SSCAPI void SSCDECL initGL(int majorVersion, int minorVersion)
{
glfwWindowHint(GLFW_CONTEXT_VERSION_MAJOR, majorVersion);
glfwWindowHint(GLFW_CONTEXT_VERSION_MINOR, minorVersion);
glfwWindowHint(GLFW_OPENGL_PROFILE, GLFW_OPENGL_CORE_PROFILE);
glfwWindowHint(GLFW_RESIZABLE, GLFW_TRUE);
}
SSCAPI GLFWglproc SSCDECL loadGlfuncDefault(char const *procName)
{
void *ret = glfwGetProcAddress(procName);
// printf("%llu\n", (uint64_t)ret);
if (ret == NULL) {
strcpy_s(infolog,30, "Failed to load OpenGL api:");
for (size_t i = 0; i < sizeof(procName); i++) {
infolog[30 + i] = procName[i];
}
//printf("%s\n", infolog);
}
return (GLFWglproc)ret;
}
SSCAPI unsigned int linkShaderProgram(GladGLContext *context, GLuint shaderProgram)
{
context->LinkProgram(shaderProgram);
int success;
context->GetProgramiv(shaderProgram, GL_LINK_STATUS, &success);
if (!success) {
context->GetProgramInfoLog(shaderProgram, 512, NULL, infolog);
return 0;
} else {
return shaderProgram;
}
}
SSCAPI void SSCDECL loadImageData(char const *location, Image *out, imageLoader loader)
{
if (loader == NULL) {
loader = stbi_load;
}
out->pixelPtr = loader(location, &(out->width), &(out->height), &(out->channel), 0);
}
SSCAPI void SSCDECL unloadImageData(Image *restrict out)
{
free(out->pixelPtr);
}
SSCAPI char *SSCDECL readLog()
{
infolog[511] = '\0';
return infolog;
}
StgSharp.Native/sn-src/dllmain.c
+47
View File
@@ -0,0 +1,47 @@
#include "StgSharpNative.h"
#include "sn_matkernel.h"
#include <stdint.h>
#ifdef _WIN32
/* Define and initialize a global zero kernel aligned to 64 bytes for universal ISA usage */
#ifdef __cplusplus
extern "C" {
#endif
__declspec(align(64)) sn_zero_kernel_u64 SN_ZERO_KERNEL;
#ifdef __cplusplus
}
#endif
static void init_zero_kernel(void)
{
for (int c = 0; c < 4; ++c) {
for (int r = 0; r < 4; ++r) {
SN_ZERO_KERNEL.m[r][c] = 0;
}
}
}
BOOL APIENTRY DllMain(HMODULE hModule, DWORD ul_reason_for_call, LPVOID lpReserved)
{
switch (ul_reason_for_call) {
case DLL_PROCESS_ATTACH:
init_zero_kernel();
return 1;
case DLL_THREAD_ATTACH:
return 1;
case DLL_THREAD_DETACH:
return 1;
case DLL_PROCESS_DETACH:
return 1;
break;
}
return TRUE;
}
#else
attribute((constructor)) void DllMain()
{
init_zero_kernel();
}
#endif // _WIN32
StgSharp.Native/sn-src/k_basicop.c
+3
View File
@@ -1,5 +1,6 @@
#include "sn_intrinsic.h"
<<<<<<< HEAD
DECLARE_PNL_PROC_ANS(float, sse, clear)
{
register __m128 zero = _mm_setzero_ps();
@@ -27,6 +28,8 @@ DECLARE_PNL_PROC_ANS(float, 512, clear)
PANEL_COL_VEC((*ans), 3, 0) = zero;
}
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
// Matrix transpose implementations
DECLARE_KER_PROC_RIGHT_ANS(float, sse, transpose)
{
StgSharp.Native/sn-src/k_fma.c
+3
View File
@@ -54,6 +54,7 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, sse, fma)
// ans = left * right + ans
// right is not transposed
<<<<<<< HEAD
DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
{
static const __m256i idx0 = { 1, 1, 1, 1, 1, 1, 1, 1 };
@@ -97,6 +98,8 @@ DECLARE_PNL_PROC_LEFT_RIGHT_ANS(float, avx_fma, fma)
// ans = left * right + ans
// right is not transposed
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
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
+6
View File
@@ -41,6 +41,7 @@ DECLARE_BUILD_PANEL(avx, float)
}
}
<<<<<<< HEAD
DECLARE_BUILD_PANEL(sse, float)
{
/* Single kernel -> 4 columns, each is one __m128 column vector. */
@@ -86,6 +87,8 @@ DECLARE_BUILD_PANEL(512, float)
}
}
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
DECLARE_STORE_PANEL(avx, float)
{
const int Q = 8; /* fixed for AVX float panel */
@@ -141,6 +144,7 @@ DECLARE_STORE_PANEL(avx, float)
k_lb->xmm[kc_col] = bot128;
}
}
<<<<<<< HEAD
DECLARE_STORE_PANEL(sse, float)
{
@@ -187,3 +191,5 @@ DECLARE_STORE_PANEL(512, float)
k[blk_col][3]->xmm[kc_col] = r3;
}
}
=======
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
StgSharp.Native/sn-src/load_intrinsic.c
+15
View File
@@ -65,10 +65,15 @@ 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
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);
=======
context->f32_buffer_add = BUF_PROC(float, sse, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
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);
@@ -82,10 +87,15 @@ 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
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);
=======
context->f32_buffer_add = BUF_PROC(float, avx, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
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);
@@ -99,10 +109,15 @@ 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
context->f32_clear_panel = PNL_PROC(float, sse, 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);
=======
context->f32_buffer_add = BUF_PROC(float, 512, add);
context->f32_panel_fma = KER_PROC(float, sse, fma);
>>>>>>> 7bcb460a3994dda40f24cae0044b5a36f4f16515
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/vector.c
+32
View File
@@ -0,0 +1,32 @@
#include "sn_intrinsic.h"
#include "sn_internal.h"
INTERNAL void SN_DECL f32_normalize(VEC(float) *restrict source, VEC(float) *restrict target)
{
register __m128 s = _mm_load_ps(&source->xmm);
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)));
a_length_sq = _mm_add_ss(a_length_sq,
_mm_shuffle_ps(a_length_sq, a_length_sq, _MM_SHUFFLE(0, 0, 3, 2)));
__m128 a_inv_length = _mm_rsqrt_ps(a_length_sq);
s = _mm_mul_ps(s, a_inv_length);
__m128 dot_product = _mm_mul_ps(target->xmm[0], s);
dot_product = _mm_add_ps(dot_product,
_mm_shuffle_ps(dot_product, dot_product, _MM_SHUFFLE(2, 3, 0, 1)));
dot_product = _mm_add_ss(dot_product,
_mm_shuffle_ps(dot_product, dot_product, _MM_SHUFFLE(0, 0, 3, 2)));
__m128 b_orthogonal = _mm_sub_ps(target->xmm[0], _mm_mul_ps(s, dot_product));
__m128 b_length_sq = _mm_mul_ps(b_orthogonal, b_orthogonal);
b_length_sq = _mm_add_ps(b_length_sq,
_mm_shuffle_ps(b_length_sq, b_length_sq, _MM_SHUFFLE(2, 3, 0, 1)));
b_length_sq = _mm_add_ss(b_length_sq,
_mm_shuffle_ps(b_length_sq, b_length_sq, _MM_SHUFFLE(0, 0, 3, 2)));
__m128 b_inv_length = _mm_rsqrt_ps(b_length_sq);
target->xmm[0] = _mm_mul_ps(b_orthogonal, b_inv_length);
}
StgSharp.Native/sn-src/vector_dot.c
+67
View File
@@ -0,0 +1,67 @@
#include "sn_intrinsic.h"
#include "sn_internal.h"
static const int avxdot_indices[] = { 0, 4, 1, 5, 2, 3, 6, 7 };
INTERNAL void FORCEINLINE SN_DECL dot_41_sse(MAT_KERNEL(float) const *transpose,
__m128 const *vector, __m128 *ans)
{
*ans = _mm_dp_ps(transpose->xmm[0], *vector, 0B11110001);
*ans = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[1], *vector, 0B11110001),
SSE_INSERT_SIMPLE(1, 1));
*ans = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[2], *vector, 0B11110001),
SSE_INSERT_SIMPLE(1, 3));
*ans = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[3], *vector, 0B11110001),
SSE_INSERT_SIMPLE(1, 4));
}
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);
tmp = _mm256_blend_ps(tmp, _mm256_dp_ps(transpose->ymm[0], _double, 0B11110010),
0B00100010);
*ans = _mm256_castps256_ps128(
_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)
{
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),
SSE_INSERT_SIMPLE(1, 2));
ans[0] = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[2], vector[0], 0B11110001),
SSE_INSERT_SIMPLE(1, 3));
ans[0] = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[3], vector[0], 0B11110001),
SSE_INSERT_SIMPLE(1, 4));
ans[1] = _mm_dp_ps(transpose->xmm[0], vector[1], 0B11110001);
ans[1] = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[1], vector[1], 0B11110001),
SSE_INSERT_SIMPLE(1, 2));
ans[1] = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[2], vector[1], 0B11110001),
SSE_INSERT_SIMPLE(1, 3));
ans[1] = _mm_insert_ps(*ans, _mm_dp_ps(transpose->xmm[3], vector[1], 0B11110001),
SSE_INSERT_SIMPLE(1, 4));
}
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);
temp = transpose->ymm[1];
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);
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));
}