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

use individual panel define as compute unit, kernel is now storage and loading unit

Browse Source
This commit is contained in:
Youmu Liu
2025-11-12 21:13:00 +08:00
parent bbe69f0384
commit 2dfb4ae86a
33 changed files with 522 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Native/sn-head/sn_intrinsic.h
+1
View File
@@ -39,6 +39,7 @@ typedef union scalar_pack {
typedef void(SN_DECL *VECTORNORMALIZEPROC)(__m128 *source, __m128 *target);
typedef void(SN_DECL *DOTPROC)(void *transpose, __m128 *vector, __m128 *ans);
typedef void(SN_DECL *PANEL_PROC_AS)(MAT_PANEL_SSE(int));
typedef void(SN_DECL *MAT_PROC_AS)(mat_kernel *ans, scalar_pack *scalar);
typedef void(SN_DECL *MAT_PROC_LA)(mat_kernel *left, mat_kernel *ans);
typedef void(SN_DECL *MAT_PROC_LRA)(mat_kernel *left, mat_kernel *right, mat_kernel *ans);
StgSharp.Native/sn-head/sn_matkernel.h
+102 -6
View File
@@ -8,12 +8,12 @@
#include <immintrin.h>
#define MAT_KERNEL(T) \
union { \
__m128 xmm[XMMCOUNT(T)]; \
__m256 ymm[YMMCOUNT(T)]; \
__m512 zmm[ZMMCOUNT(T)]; \
T m[4][4]; \
#define MAT_KERNEL(T) \
union { \
__m128 xmm[XMMCOUNT(T) / 4]; \
__m256 ymm[YMMCOUNT(T) / 4]; \
__m512 zmm[ZMMCOUNT(T) / 4]; \
T m[4][4]; \
}
#define VEC(T) \
@@ -25,3 +25,99 @@
}
#endif
#include <stddef.h>
#include <stdint.h>
/* align decoration */
#if __cplusplus && defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
#define ML_ALIGNAS(n) _Alignas(n)
#elif defined(_MSC_VER)
#define ML_ALIGNAS(n) __declspec(align(n))
#else
#define ML_ALIGNAS(n) __attribute__((aligned(n)))
#endif
/* Round up to the alignment boundary in bytes */
#ifndef ML_ALIGN_UP
#define ML_ALIGN_UP(x, a) (((size_t)(x) + ((size_t)(a) - 1)) & ~((size_t)(a) - 1))
#endif
/* Vector register width in bytes */
#define ML_BYTES_XMM 16u
#define ML_BYTES_YMM 32u
#define ML_BYTES_ZMM 64u
/* Compute the aligned leading-direction element stride (LEp) */
#define ML_PANEL_STRIDE_ELEMS(T, LE, ALIGN) \
(ML_ALIGN_UP((size_t)(LE) * sizeof(T), (size_t)(ALIGN)) / sizeof(T))
/* Total element count and total byte count */
#define ML_PANEL_TOT_ELEMS(T, LE, MINOR, ALIGN) \
((size_t)ML_PANEL_STRIDE_ELEMS(T, (LE), (ALIGN)) * (size_t)(MINOR))
#define ML_TOT_BYTES(T, ELEMS) ((size_t)(ELEMS) * sizeof(T))
/* Number of SIMD view slots (ceil by total bytes) */
#define ML_XMM_COUNT(TOT_BYTES) (((size_t)(TOT_BYTES) + ML_BYTES_XMM - 1) / ML_BYTES_XMM)
#define ML_YMM_COUNT(TOT_BYTES) (((size_t)(TOT_BYTES) + ML_BYTES_YMM - 1) / ML_BYTES_YMM)
#define ML_ZMM_COUNT(TOT_BYTES) (((size_t)(TOT_BYTES) + ML_BYTES_ZMM - 1) / ML_BYTES_ZMM)
/* ========== Unified "panel" union (no enum/descriptor fields) ==========
Usage:
MAT_PANEL(float, 8, 32, 32) P; // LE=8, MINOR=32, aligned to 32 bytes
Views:
P.xmm[] / P.ymm[] / P.zmm[] // raw SIMD blocks
P.blk[MINOR][LEp] // 2D scalar view (includes padding for alignment)
P.flat[MINOR*LEp] // flat array
Notes:
- LE/MINOR/ALIGN must be compile-time constants
- LEp = ML_PANEL_STRIDE_ELEMS(T, LE, ALIGN)
- Layout/semantics defined by the calling convention (e.g. blk[k][r] is element r of block k)
*/
#define MAT_PANEL(T, LE, MINOR, ALIGN) \
ML_ALIGNAS(ALIGN) union { \
/* Raw SIMD view: cast to __m256/__m512d inside the micro-kernel on demand */ \
__m128i xmm[ML_XMM_COUNT( \
ML_TOT_BYTES(T, ML_PANEL_TOT_ELEMS(T, (LE), (MINOR), (ALIGN))))]; \
__m256i ymm[ML_YMM_COUNT( \
ML_TOT_BYTES(T, ML_PANEL_TOT_ELEMS(T, (LE), (MINOR), (ALIGN))))]; \
__m512i zmm[ML_ZMM_COUNT( \
ML_TOT_BYTES(T, ML_PANEL_TOT_ELEMS(T, (LE), (MINOR), (ALIGN))))]; \
/* Scalar 2D view: minor �� LEp (LE dimension contains alignment padding) */ \
T blk[(MINOR)][ML_PANEL_STRIDE_ELEMS(T, (LE), (ALIGN))]; \
/* Flat view */ \
T flat[ML_PANEL_TOT_ELEMS(T, (LE), (MINOR), (ALIGN))]; \
}
/* ---------- SSE preset ---------- */
#undef PANEL_LE
#undef PANEL_MINOR
#undef PANEL_ALIGN
#define PANEL_LE 4
#define PANEL_MINOR 32
#define PANEL_ALIGN 16
/* ======== SSE preset ========
4 x 32, 16B aligned
*/
#define MAT_PANEL_SSE(T) MAT_PANEL(T, PANEL_LE, PANEL_MINOR, PANEL_ALIGN)
/* ---------- AVX2 preset ---------- */
#undef PANEL_LE
#undef PANEL_MINOR
#undef PANEL_ALIGN
#define PANEL_LE 8
#define PANEL_MINOR 64
#define PANEL_ALIGN 32
#define MAT_PANEL_AVX(T) MAT_PANEL(T, PANEL_LE, PANEL_MINOR, PANEL_ALIGN) /* 8 x 64, 32B aligned */
/* ---------- AVX-512 preset ---------- */
#undef PANEL_LE
#undef PANEL_MINOR
#undef PANEL_ALIGN
#define PANEL_LE 16
#define PANEL_MINOR 64
#define PANEL_ALIGN 64
#define MAT_PANEL_512(T) \
MAT_PANEL(T, PANEL_LE, PANEL_MINOR, PANEL_ALIGN); /* 16 x 64, 64B aligned */
StgSharp.Native/sn-src/k_fma.c
+30 -3
View File
@@ -1,5 +1,9 @@
#include "sn_intrinsic.h"
INTERNAL void SN_DECL f32_fma_sse(MAT_PANEL_SSE(int) * left)
{
}
// ans = left * right + ans
// right is not transposed
INTERNAL void SN_DECL f32_fma_sse(MAT_KERNEL(float) const *restrict left,
@@ -41,9 +45,32 @@ INTERNAL void SN_DECL f32_fma_avx(MAT_KERNEL(float) const *restrict left,
MAT_KERNEL(float) const *restrict right,
MAT_KERNEL(float) *restrict ans)
{
/*
* to be implemented
*/
register __m256 source0 = _mm256_load_ps(&left->ymm[0]);
register __m256 source1 = _mm256_load_ps(&left->ymm[1]);
register __m256 ans0 = _mm256_load_ps(&ans->ymm[0]);
register __m256 ans1 = _mm256_load_ps(&ans->ymm[1]);
register __m128 bs0, bs1;
register __m256 broadcast;
for (size_t i = 0; i < 4; i++) {
bs0 = _mm_load_ps(&right->xmm[i]);
__m128 lo = _mm_shuffle_ps(bs0, bs0, 0x00);
__m128 hi = _mm_shuffle_ps(bs0, bs0, 0x55);
broadcast = _mm256_castps128_ps256(lo);
broadcast = _mm256_insertf128_ps(broadcast, bs0, 1);
ans0 = _mm256_add_ps(ans0, _mm256_mul_ps(source0, broadcast));
lo = _mm_shuffle_ps(bs0, bs0, 0xAA);
hi = _mm_shuffle_ps(bs0, bs0, 0xFF);
broadcast = _mm256_castps128_ps256(lo);
broadcast = _mm256_insertf128_ps(broadcast, bs0, 1);
ans1 = _mm256_add_ps(ans1, _mm256_mul_ps(source1, broadcast));
}
_mm256_store_ps(&ans->ymm[0], ans0);
_mm256_store_ps(&ans->ymm[1], ans1);
}
// ans = left * right + ans