Huge progress on Matrix Parallel, a parallel pool has its basic structure.

2025-09-02 21:56:10 +08:00
parent 2475a3b0ed
commit 125040620b
6 changed files with 307 additions and 39 deletions
@@ -1,6 +1,7 @@
 #include "sn_intrinsic.h"

-void kernel_fma_sse(mat_kernel *const left, mat_kernel *const right, mat_kernel *ans)
+INTERNAL void SN_DECL kernel_fma_sse(mat_kernel const *restrict left,
+                                     mat_kernel const *restrict right, mat_kernel *restrict ans)
 {
        register __m128 source0 = _mm_load_ps(&left->xmm[0]);
        register __m128 source1 = _mm_load_ps(&left->xmm[1]);