mirror of
https://github.com/Karaka-Management/cOMS.git
synced 2026-01-10 19:08:39 +00:00
132 lines
3.7 KiB
C
Executable File
132 lines
3.7 KiB
C
Executable File
/**
|
|
* Jingga
|
|
*
|
|
* @copyright Jingga
|
|
* @license OMS License 2.0
|
|
* @version 1.0.0
|
|
* @link https://jingga.app
|
|
*/
|
|
#ifndef COMS_STDLIB_SIMD_I16_H
|
|
#define COMS_STDLIB_SIMD_I16_H
|
|
|
|
#include <immintrin.h>
|
|
#include <xmmintrin.h>
|
|
|
|
#include "../../../stdlib/Types.h"
|
|
|
|
#ifdef __SSE4_2__
|
|
#include "SIMD_I16_SSE.h"
|
|
#endif
|
|
|
|
#ifdef __AVX2__
|
|
#include "SIMD_I16_AVX2.h"
|
|
#endif
|
|
|
|
#ifdef __AVX512F__
|
|
#include "SIMD_I16_AVX512.h"
|
|
#endif
|
|
|
|
// @todo from down here we can optimize some of the code by NOT using the wrappers
|
|
// the code is self contained and we could use te intrinsic functions directly
|
|
|
|
inline
|
|
void simd_mult(const int16* a, f32 b, int16* result, int32 size, int32 steps = 16)
|
|
{
|
|
int32 i = 0;
|
|
steps = intrin_validate_steps((const byte*) a, steps);
|
|
steps = intrin_validate_steps((const byte*) result, steps);
|
|
|
|
#ifdef __AVX512F__
|
|
if (steps >= 16) {
|
|
steps = 16;
|
|
__m512i a_16;
|
|
__m512 af_lo, af_hi;
|
|
__m512 b_16 = _mm512_set1_ps(b);
|
|
__m512 result_lo, result_hi;
|
|
__m512i result_16;
|
|
|
|
for (; i <= size - steps; i += steps) {
|
|
a_16 = _mm512_load_si512((__m512i*) a);
|
|
|
|
af_lo = _mm512_cvtepi32_ps(_mm512_cvtepi16_epi32(_mm512_extracti64x4_epi64(a_16, 0)));
|
|
af_hi = _mm512_cvtepi32_ps(_mm512_cvtepi16_epi32(_mm512_extracti64x4_epi64(a_16, 1)));
|
|
|
|
result_lo = _mm512_mul_ps(af_lo, b_16);
|
|
result_hi = _mm512_mul_ps(af_hi, b_16);
|
|
|
|
result_16 = _mm512_packs_epi32(_mm512_cvtps_epi32(result_lo), _mm512_cvtps_epi32(result_hi));
|
|
_mm512_store_si512((__m512i*) result, result_16);
|
|
|
|
a += steps;
|
|
result += steps;
|
|
}
|
|
|
|
steps = 1;
|
|
}
|
|
#endif
|
|
|
|
#ifdef __AVX2__
|
|
if (steps >= 8) {
|
|
steps = 8;
|
|
__m256i a_8;
|
|
__m256 af_lo, af_hi;
|
|
__m256 b_8 = _mm256_set1_ps(b);
|
|
__m256 result_lo, result_hi;
|
|
__m256i result_8;
|
|
|
|
for (; i <= size - steps; i += steps) {
|
|
a_8 = _mm256_load_si256((__m256i*) a);
|
|
|
|
af_lo = _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(a_8, 0)));
|
|
af_hi = _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(a_8, 1)));
|
|
|
|
result_lo = _mm256_mul_ps(af_lo, b_8);
|
|
result_hi = _mm256_mul_ps(af_hi, b_8);
|
|
|
|
result_8 = _mm256_packs_epi32(_mm256_cvtps_epi32(result_lo), _mm256_cvtps_epi32(result_hi));
|
|
_mm256_store_si256((__m256i*) result, result_8);
|
|
|
|
a += steps;
|
|
result += steps;
|
|
}
|
|
|
|
steps = 1;
|
|
}
|
|
#endif
|
|
|
|
#ifdef __SSE4_2__
|
|
if (steps >= 4) {
|
|
steps = 4;
|
|
__m128i a_4;
|
|
__m128 af_lo, af_hi;
|
|
__m128 b_4 = _mm_set1_ps(b);
|
|
__m128 result_lo, result_hi;
|
|
__m128i result_4;
|
|
|
|
for (; i <= size - steps; i += steps) {
|
|
a_4 = _mm_load_si128((__m128i*) a);
|
|
|
|
af_lo = _mm_cvtepi32_ps(_mm_cvtepi16_epi32(a_4));
|
|
af_hi = _mm_cvtepi32_ps(_mm_cvtepi16_epi32(_mm_srli_si128(a_4, 8)));
|
|
|
|
result_lo = _mm_mul_ps(af_lo, b_4);
|
|
result_hi = _mm_mul_ps(af_hi, b_4);
|
|
|
|
result_4 = _mm_packs_epi32(_mm_cvtps_epi32(result_lo), _mm_cvtps_epi32(result_hi));
|
|
_mm_store_si128((__m128i*) result, result_4);
|
|
|
|
a += steps;
|
|
result += steps;
|
|
}
|
|
}
|
|
#endif
|
|
|
|
// Handle any remaining elements
|
|
for (; i < size; ++i) {
|
|
*result = (int16) ((f32) (*a) * b);
|
|
++a;
|
|
++result;
|
|
}
|
|
}
|
|
|
|
#endif |