working on panel building, preparing for panel compute

StgSharp.Native/StgSharp.Native.vcxproj

@@ -148,6 +148,7 @@
     </Link>
   </ItemDefinitionGroup>
   <ItemGroup>
+    <ClCompile Include="sn-src\k_build_panel.c" />
     <ClCompile Include="sn-src\k_basicop.c" />
     <ClCompile Include="sn-src\dllmain.c" />
     <ClCompile Include="sn-src\k_bufop.c" />

StgSharp.Native/StgSharp.Native.vcxproj.filters

@@ -51,6 +51,9 @@
     <ClCompile Include="sn-src\k_bufop.c">
       <Filter>源文件</Filter>
     </ClCompile>
+    <ClCompile Include="sn-src\k_build_panel.c">
+      <Filter>源文件</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="sn-head\gl.h">

StgSharp.Native/sn-head/sn_intrinsic.h

@@ -250,6 +250,9 @@ INTERNAL int SN_DECL index_pair_sse(short const *str, uint32_t target, int lengt
 
 #pragma endregion
 
+typedef void(SN_DECL *BUILD_PANEL_PROC)(mat_panel *panel, mat_kernel *kernel, int col_length,
+                                        int row_length, int col_index, int row_index);
+
 // Reordered to match C# IntrinsicContext field order
 // C# order: city_hash_simplify, f32_add, f32_fill, f32_fma, f32_scalar_mul, f32_sub, f32_transpose, factorial_simd, index_pair, normalize_3
 typedef struct sn_intrinsic {

StgSharp.Native/sn-head/sn_matkernel.h

@@ -7,6 +7,7 @@
 #define ZMMCOUNT(T) sizeof(T) / 4
 
 #include <immintrin.h>
+#include <stdint.h>
 
 #define MAT_KERNEL(T)                        \
         union {                              \
@@ -77,7 +78,29 @@
    - Column c occupies vec[c][0..CH-1].
    - Row r inside column c is located at vec[c][r / R] with lane (r % R).
 */
+
+/* Global zero kernel declaration (defined in dllmain.c), aligned to 64 bytes */
+#ifdef __cplusplus
+extern "C" {
+#endif
+__declspec(align(64)) extern MAT_KERNEL(uint64_t) SN_ZERO_KERNEL;
+#ifdef __cplusplus
+}
+#endif
+
 #define PANEL_COL_VEC(panel_obj, c, chunk) ((panel_obj).vec[(int)(c)][(int)(chunk)])
 /* Scalar flat index helper for column-major addressing */
 #define PANEL_ELEM(panel_obj, T, T_VEC, MINQ, r, c) \
         ((panel_obj).flat[(int)(c) * PANEL_Q(T, T_VEC, MINQ) + (int)(r)])
+
+#define BUILD_PANEL(arch, element) build_panel_##arch##_##element
+
+#include "sn_internal.h"
+
+#define DECLARE_BUILD_PANEL(arch, element)                                               \
+        INTERNAL void SN_DECL BUILD_PANEL(arch, element)(                                \
+                MAT_PANEL_##arch(element) *restrict panel,                               \
+                MAT_KERNEL(element) const *const matrix, int col_length, int row_length, \
+                int col_index, int row_index)
+
+DECLARE_BUILD_PANEL(AVX, float);

StgSharp.Native/sn-src/dllmain.c

@@ -1,12 +1,25 @@
-
-#include "StgSharpNative.h"
+#include "StgSharpNative.h"
+#include "sn_matkernel.h"
 
 #ifdef _WIN32
 
+/* Define and initialize a global zero kernel aligned to 64 bytes for universal ISA usage */
+__declspec(align(64)) MAT_KERNEL(uint64_t) SN_ZERO_KERNEL;
+
+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] = 0.0;
+                }
+        }
+}
+
 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;
@@ -21,6 +34,7 @@ BOOL APIENTRY DllMain(HMODULE hModule, DWORD ul_reason_for_call, LPVOID lpReserv
 #else
 attribute((constructor)) void DllMain()
 {
+        init_zero_kernel();
 }
 
 #endif // _WIN32

StgSharp.Native/sn-src/k_build_panel.c

@@ -0,0 +1,42 @@
+#include "sn_matkernel.h"
+
+DECLARE_BUILD_PANEL(AVX, float)
+{
+        const int Q = 8; /* fixed for AVX float panel */
+
+        /* Neighbor kernels; substitute global zero kernel where out-of-bounds */
+        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_rt =
+                has_right ? (matrix + ((col_index + 1) * row_length + row_index)) :
+                            (const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
+        const MAT_KERNEL(float) *const k_lb =
+                has_bottom ? (matrix + (col_index * row_length + (row_index + 1))) :
+                             (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))) :
+                        (const MAT_KERNEL(float) *)&SN_ZERO_KERNEL;
+
+        register __m128 top128, bot128;
+        register __m256 col256;
+
+        for (int pc = 0; pc < Q; ++pc) {
+                const int is_right_half = (pc >= 4);
+                const int kc_col = pc & 3; /* column index inside 4x4 kernel */
+
+                const MAT_KERNEL(float) *src_top = is_right_half ? k_rt : k_lt;
+                const MAT_KERNEL(float) *src_bot = is_right_half ? k_rb : k_lb;
+
+                /* Vector loads: each kernel column is a compact __m128 */
+                top128 = src_top->xmm[kc_col];
+                bot128 = src_bot->xmm[kc_col];
+
+                /* Combine into an AVX column: low 128 = top, high 128 = bottom */
+                col256 = _mm256_set_m128(bot128, top128);
+
+                PANEL_COL_VEC((*panel), pc, 0) = col256;
+        }
+}