1
0
mirror of https://github.com/wolfpld/tracy.git synced 2025-03-20 07:40:02 +08:00

Force inline ProcessRGB functions.

This commit is contained in:
Bartosz Taudul 2019-06-30 03:33:14 +02:00
parent ddd89dcce5
commit 52b6bdb55a

View File

@ -1,4 +1,5 @@
#include "TracyDxt1.hpp" #include "TracyDxt1.hpp"
#include "../common/TracyForceInline.hpp"
#include <assert.h> #include <assert.h>
#include <stdint.h> #include <stdint.h>
@ -157,7 +158,7 @@ static const uint16_t DivTableAVX[255*3] = {
}; };
static uint64_t ProcessRGB( const uint8_t* src ) static tracy_force_inline uint64_t ProcessRGB( const uint8_t* src )
{ {
#ifdef __SSE4_1__ #ifdef __SSE4_1__
__m128i px0 = _mm_loadu_si128(((__m128i*)src) + 0); __m128i px0 = _mm_loadu_si128(((__m128i*)src) + 0);
@ -460,7 +461,7 @@ static uint64_t ProcessRGB( const uint8_t* src )
} }
#ifdef __AVX2__ #ifdef __AVX2__
static void ProcessRGB_AVX( const uint8_t* src, char*& dst ) static tracy_force_inline void ProcessRGB_AVX( const uint8_t* src, char*& dst )
{ {
__m256i px0 = _mm256_loadu_si256(((__m256i*)src) + 0); __m256i px0 = _mm256_loadu_si256(((__m256i*)src) + 0);
__m256i px1 = _mm256_loadu_si256(((__m256i*)src) + 1); __m256i px1 = _mm256_loadu_si256(((__m256i*)src) + 1);