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

Daily update on 20260111-005714 by Nitload Youmu. Additional Message : Use optimized tile operation to accelerate matrix mul, still has problem

Browse Source 
Signed-off-by: Nitload <YoumuLiu_NSI@whv0t.onmicrosoft.com>
This commit is contained in:
Nitload
2026-01-11 00:57:14 +08:00
parent 6827259891
commit 4be5c1c7c9
9 changed files with 141 additions and 219 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-src/m.fma_tile.x86.c
+10 -52
View File
@@ -1,4 +1,4 @@
#include "sn_target.h"
#include "sn_target.h"
#if SN_IS_ARCH(SN_ARCH_X86_64)
@@ -293,7 +293,7 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(float, 512, , fma)
a_hi = _mm_load_pd(left_cur + (3 * 2 + 1)); \
b = _mm_shuffle_pd(b_hi, b_hi, 0x3); \
c##i##_lo = _mm_add_pd(c##i##_lo, _mm_mul_pd(a_lo, b)); \
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b));
c##i##_hi = _mm_add_pd(c##i##_hi, _mm_mul_pd(a_hi, b));\
DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, sse, , fma)
{
@@ -316,10 +316,12 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, sse, , fma)
const *left_cur = left + GET_INDEX(ansHeight, k, j),
const *right_cur = right + GET_INDEX(commonK, j, k); // GetKernelAddressUnsafe
DOUBLE_KERNEL_FMA_CYCLE(0)
DOUBLE_KERNEL_FMA_CYCLE(1)
DOUBLE_KERNEL_FMA_CYCLE(2)
DOUBLE_KERNEL_FMA_CYCLE(3)
{DOUBLE_KERNEL_FMA_CYCLE(0)}
{DOUBLE_KERNEL_FMA_CYCLE(1)}
{DOUBLE_KERNEL_FMA_CYCLE(2)}
{
DOUBLE_KERNEL_FMA_CYCLE(3)
}
}
_mm_store_pd(ans->f64_x + (0 * 2 + 0), c0_lo);
@@ -519,8 +521,8 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, 512, , fma)
c23 = _mm512_fmadd_pd(a01, b23_perm, c23);
//group 1
register __m512d a12 = _mm512_permute2f128_pd(a01, a23, 0x21);
register __m512d a30 = _mm512_permute2f128_pd(a01, a23, 0x30);
register __m512d a12 = _mm512_shuffle_f64x2(a01, a23, 0x21);
register __m512d a30 = _mm512_shuffle_f64x2(a01, a23, 0x30);
b01_perm = _mm512_shuffle_pd(b01, b01, _FMA_SELECT(1));
b23_perm = _mm512_shuffle_pd(b23, b23, _FMA_SELECT(1));
c01 = _mm512_fmadd_pd(a12, b01_perm, c01);
@@ -534,50 +536,6 @@ DECLARE_KERTILE_PROC_LEFT_RIGHT_ANS(double, 512, , fma)
}
_mm512_store_pd((double *)&ans_cur->f64_z[0], c01);
_mm512_store_pd((double *)&ans_cur->f64_z[1], c23);
/*
static const __m512i b01_shift = { .m512i_i64 = { 0, 1, 2, 3, 5, 6, 7, 4 } };
static const __m512i b23_shift = { .m512i_i64 = { 2, 3, 0, 1, 7, 4, 5, 6 } };
register __m512d a01 = _mm512_load_pd((double *)&left->f64_z[0]);
register __m512d a23 = _mm512_load_pd((double *)&left->f64_z[1]);
register __m512d b01 = _mm512_load_pd((double *)&right->f64_z[0]);
register __m512d b23 = _mm512_load_pd((double *)&right->f64_z[1]);
b01 = _mm512_permutexvar_pd(b01_shift, b01);
b23 = _mm512_permutexvar_pd(b23_shift, b23);
register __m512d c01 = _mm512_load_pd((float *)&ans->f64_z[0]);
register __m512d c23 = _mm512_load_pd((float *)&ans->f64_z[1]);
//group 0
register __m512d b01_perm = _mm512_shuffle_pd(b01, b01, _FMA_SELECT(0));
register __m512d b23_perm = _mm512_shuffle_pd(b23, b23, _FMA_SELECT(0));
c01 = _mm512_fmadd_pd(a01, b01_perm, c01);
c23 = _mm512_fmadd_pd(a23, b23_perm, c23);
//group 2
b01_perm = _mm512_shuffle_pd(b01, b01, _FMA_SELECT(2));
b23_perm = _mm512_shuffle_pd(b23, b23, _FMA_SELECT(2));
c01 = _mm512_fmadd_pd(a23, b01_perm, c01);
c23 = _mm512_fmadd_pd(a01, b23_perm, c23);
//group 1
register __m512d a12 = _mm512_shuffle_f64x2(a01, a23, _MM_SHUFFLE(1, 0, 3, 2));
register __m512d a30 = _mm512_shuffle_f64x2(a23, a01, _MM_SHUFFLE(1, 0, 3, 2));
b01_perm = _mm512_shuffle_pd(b01, b01, _FMA_SELECT(1));
b23_perm = _mm512_shuffle_pd(b23, b23, _FMA_SELECT(1));
c01 = _mm512_fmadd_pd(a12, b01_perm, c01);
c23 = _mm512_fmadd_pd(a30, b23_perm, c23);
//group 3
b01_perm = _mm512_shuffle_pd(b01, b01, _FMA_SELECT(3));
b23_perm = _mm512_shuffle_pd(b23, b23, _FMA_SELECT(3));
c01 = _mm512_fmadd_pd(a30, b01_perm, c01);
c23 = _mm512_fmadd_pd(a12, b23_perm, c23);
ans->f64_z[0] = c01;
ans->f64_z[1] = c23;
/**/
}
#endif