mirror of
https://github.com/FunKey-Project/FunKey-ProdScreens.git
synced 2025-12-12 08:48:51 +01:00
many tries with __builtin_prefetch to optimize neon
This commit is contained in:
Vincent Buso
parent
78150ef41d
commit
c78ba94025
@ -158,6 +158,12 @@ void fbtft_transpose_neon(uint16_t* src, uint16_t* dst, int w, int h){
|
||||
for (y=0; y<h; y+=4){
|
||||
for (x=0; x<w; x+=4){
|
||||
|
||||
/* 1% CPU gain */
|
||||
__builtin_prefetch(src + (y+0)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+1)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+2)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+3)*w + x + 4);
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp.val[0] = vld1_u16(src + (y+0)*w + x );
|
||||
v_tmp.val[1] = vld1_u16(src + (y+1)*w + x );
|
||||
@ -187,6 +193,12 @@ void fbtft_transpose_inv_neon(uint16_t* src, uint16_t* dst, int w, int h){
|
||||
for (y=0; y<h; y+=4){
|
||||
for (x=0; x<w; x+=4){
|
||||
|
||||
/* 1% CPU gain */
|
||||
__builtin_prefetch(src + (y+0)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+1)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+2)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+3)*w + x + 4);
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp.val[0] = vld1_u16(src + (y+3)*w + x );
|
||||
v_tmp.val[1] = vld1_u16(src + (y+2)*w + x );
|
||||
@ -216,6 +228,12 @@ void fbtft_rotate_90cw_neon(uint16_t* src, uint16_t* dst, int w, int h){
|
||||
for (y=0; y<h; y+=4){
|
||||
for (x=0; x<w; x+=4){
|
||||
|
||||
/* 1% CPU gain */
|
||||
__builtin_prefetch(src + (y+0)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+1)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+2)*w + x + 4);
|
||||
__builtin_prefetch(src + (y+3)*w + x + 4);
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp.val[0] = vld1_u16(src + (y+3)*w + x );
|
||||
v_tmp.val[1] = vld1_u16(src + (y+2)*w + x );
|
||||
@ -259,6 +277,231 @@ void fbtft_rotate_270cw_neon(uint16_t* src, uint16_t* dst, int w, int h){
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/*
|
||||
NEON optimized matrix rotate 270° CW
|
||||
(dimensions multiple of 4, 16bits pixels)
|
||||
*/
|
||||
#define PREFETCH_ORDER_X 32
|
||||
#define PREFETCH_ORDER_Y 4
|
||||
void fbtft_rotate_270cw_neon_prefetch(uint16_t* src, uint16_t* dst, int w, int h){
|
||||
|
||||
/* Vars */
|
||||
uint16x4x4_t v_tmp;
|
||||
int y, x;
|
||||
|
||||
/* Main loop */
|
||||
for (y=0; y<h; y+=4){
|
||||
|
||||
/* Prefetch src */
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+0)*w + x);
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+1)*w + x);
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+2)*w + x);
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+3)*w + x);
|
||||
|
||||
/* Prefetch dst */
|
||||
/*__builtin_prefetch(dst + ( (w-1) - x - 3 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 2 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 1 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 0 + PREFETCH_ORDER_Y )*h + y + 4, 1);*/
|
||||
|
||||
for (x=0; x<w; x+=4){
|
||||
|
||||
/* Prefetch src */
|
||||
__builtin_prefetch(src + (y+0)*w + x + PREFETCH_ORDER_X);
|
||||
__builtin_prefetch(src + (y+1)*w + x + PREFETCH_ORDER_X);
|
||||
__builtin_prefetch(src + (y+2)*w + x + PREFETCH_ORDER_X);
|
||||
__builtin_prefetch(src + (y+3)*w + x + PREFETCH_ORDER_X);
|
||||
|
||||
/* Prefetch dst */
|
||||
/*__builtin_prefetch(dst + ( (w-1) - x - 3 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 2 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 1 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 0 + PREFETCH_ORDER_X )*h + y, 1);*/
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp.val[0] = vld1_u16(src + (y+0)*w + x );
|
||||
v_tmp.val[1] = vld1_u16(src + (y+1)*w + x );
|
||||
v_tmp.val[2] = vld1_u16(src + (y+2)*w + x );
|
||||
v_tmp.val[3] = vld1_u16(src + (y+3)*w + x );
|
||||
|
||||
/* Neon store (4 interleaved) */
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 3 )*h + y, v_tmp, 3);
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 2 )*h + y, v_tmp, 2);
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 1 )*h + y, v_tmp, 1);
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 0 )*h + y, v_tmp, 0);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/*
|
||||
NEON optimized matrix rotate 270° CW
|
||||
(dimensions multiple of 4, 16bits pixels)
|
||||
*/
|
||||
#define PREFETCH_ORDER_X 32
|
||||
#define PREFETCH_ORDER_Y 4
|
||||
void fbtft_rotate_270cw_neon_prefetch2(uint16_t* src, uint16_t* dst, int w, int h){
|
||||
|
||||
/* Vars */
|
||||
uint16x4x4_t v_tmp, v_tmp2, v_tmp3, v_tmp4;
|
||||
int y, x;
|
||||
|
||||
/* Main loop */
|
||||
for (y=0; y<h; y+=4){
|
||||
|
||||
/* Prefetch src */
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+0)*w + x);
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+1)*w + x);
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+2)*w + x);
|
||||
__builtin_prefetch(src + (y+PREFETCH_ORDER_Y+3)*w + x);
|
||||
|
||||
/* Prefetch dst */
|
||||
/*__builtin_prefetch(dst + ( (w-1) - x - 3 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 2 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 1 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 0 + PREFETCH_ORDER_Y )*h + y + 4, 1);*/
|
||||
|
||||
for (x=0; x<w; x+=16){
|
||||
|
||||
/* Prefetch src */
|
||||
__builtin_prefetch(src + (y+0)*w + x + PREFETCH_ORDER_X);
|
||||
__builtin_prefetch(src + (y+1)*w + x + PREFETCH_ORDER_X);
|
||||
__builtin_prefetch(src + (y+2)*w + x + PREFETCH_ORDER_X);
|
||||
__builtin_prefetch(src + (y+3)*w + x + PREFETCH_ORDER_X);
|
||||
|
||||
/* Prefetch dst */
|
||||
/*__builtin_prefetch(dst + ( (w-1) - x - 3 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 2 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 1 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 0 + PREFETCH_ORDER_X )*h + y, 1);*/
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp.val[0] = vld1_u16(src + (y+0)*w + x );
|
||||
v_tmp.val[1] = vld1_u16(src + (y+1)*w + x );
|
||||
v_tmp.val[2] = vld1_u16(src + (y+2)*w + x );
|
||||
v_tmp.val[3] = vld1_u16(src + (y+3)*w + x );
|
||||
|
||||
/* Neon store (4 interleaved) */
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 3 )*h + y, v_tmp, 3);
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 2 )*h + y, v_tmp, 2);
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 1 )*h + y, v_tmp, 1);
|
||||
vst4_lane_u16(dst + ( (w-1) - x - 0 )*h + y, v_tmp, 0);
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp2.val[0] = vld1_u16(src + (y+0)*w + x+4 );
|
||||
v_tmp2.val[1] = vld1_u16(src + (y+1)*w + x+4 );
|
||||
v_tmp2.val[2] = vld1_u16(src + (y+2)*w + x+4 );
|
||||
v_tmp2.val[3] = vld1_u16(src + (y+3)*w + x+4 );
|
||||
|
||||
/* Neon store (4 interleaved) */
|
||||
vst4_lane_u16(dst + ( (w-1) - x+4 - 3 )*h + y, v_tmp2, 3);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+4 - 2 )*h + y, v_tmp2, 2);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+4 - 1 )*h + y, v_tmp2, 1);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+4 - 0 )*h + y, v_tmp2, 0);
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp3.val[0] = vld1_u16(src + (y+0)*w + x+8 );
|
||||
v_tmp3.val[1] = vld1_u16(src + (y+1)*w + x+8 );
|
||||
v_tmp3.val[2] = vld1_u16(src + (y+2)*w + x+8 );
|
||||
v_tmp3.val[3] = vld1_u16(src + (y+3)*w + x+8 );
|
||||
|
||||
/* Neon store (4 interleaved) */
|
||||
vst4_lane_u16(dst + ( (w-1) - x+8 - 3 )*h + y, v_tmp3, 3);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+8 - 2 )*h + y, v_tmp3, 2);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+8 - 1 )*h + y, v_tmp3, 1);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+8 - 0 )*h + y, v_tmp3, 0);
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp4.val[0] = vld1_u16(src + (y+0)*w + x+12 );
|
||||
v_tmp4.val[1] = vld1_u16(src + (y+1)*w + x+12 );
|
||||
v_tmp4.val[2] = vld1_u16(src + (y+2)*w + x+12 );
|
||||
v_tmp4.val[3] = vld1_u16(src + (y+3)*w + x+12 );
|
||||
|
||||
/* Neon store (4 interleaved) */
|
||||
vst4_lane_u16(dst + ( (w-1) - x+12 - 3 )*h + y, v_tmp4, 3);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+12 - 2 )*h + y, v_tmp4, 2);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+12 - 1 )*h + y, v_tmp4, 1);
|
||||
vst4_lane_u16(dst + ( (w-1) - x+12 - 0 )*h + y, v_tmp4, 0);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
#define prefetch(x) __builtin_prefetch(x)
|
||||
#ifndef PREFETCH_STRIDE
|
||||
#define PREFETCH_STRIDE (2*32)
|
||||
#endif
|
||||
static inline void prefetch_range(void *addr, size_t len)
|
||||
{
|
||||
char *cp;
|
||||
char *end = addr + len;
|
||||
|
||||
for (cp = addr; cp < end; cp += PREFETCH_STRIDE)
|
||||
prefetch(cp);
|
||||
}
|
||||
|
||||
|
||||
|
||||
/*
|
||||
NEON optimized matrix rotate 270° CW
|
||||
(dimensions multiple of 4, 16bits pixels)
|
||||
*/
|
||||
#define PREFETCH_ORDER_X 20
|
||||
#define PREFETCH_ORDER_Y 4
|
||||
#define PREFETCH_RANGE 1
|
||||
void fbtft_rotate_270cw_neon_prefetch3(uint16_t* src, uint16_t* dst, int w, int h){
|
||||
|
||||
/* Vars */
|
||||
uint16x4x4_t v_tmp;
|
||||
int y, x;
|
||||
|
||||
/* Main loop */
|
||||
for (y=0; y<h; y+=4){
|
||||
|
||||
/* Prefetch src */
|
||||
prefetch_range(src + (y+PREFETCH_ORDER_Y+0)*w + x, PREFETCH_RANGE);
|
||||
prefetch_range(src + (y+PREFETCH_ORDER_Y+1)*w + x, PREFETCH_RANGE);
|
||||
prefetch_range(src + (y+PREFETCH_ORDER_Y+2)*w + x, PREFETCH_RANGE);
|
||||
prefetch_range(src + (y+PREFETCH_ORDER_Y+3)*w + x, PREFETCH_RANGE);
|
||||
|
||||
/* Prefetch dst */
|
||||
/*__builtin_prefetch(dst + ( (w-1) - x - 3 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 2 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 1 + PREFETCH_ORDER_Y )*h + y + 4, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 0 + PREFETCH_ORDER_Y )*h + y + 4, 1);*/
|
||||
|
||||
for (x=0; x<w; x+=4){
|
||||
|
||||
/* Prefetch src */
|
||||
prefetch_range(src + (y+0)*w + x + PREFETCH_ORDER_X, PREFETCH_RANGE);
|
||||
prefetch_range(src + (y+1)*w + x + PREFETCH_ORDER_X, PREFETCH_RANGE);
|
||||
prefetch_range(src + (y+2)*w + x + PREFETCH_ORDER_X, PREFETCH_RANGE);
|
||||
prefetch_range(src + (y+3)*w + x + PREFETCH_ORDER_X, PREFETCH_RANGE);
|
||||
|
||||
/* Prefetch dst */
|
||||
/*__builtin_prefetch(dst + ( (w-1) - x - 3 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 2 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 1 + PREFETCH_ORDER_X )*h + y, 1);
|
||||
__builtin_prefetch(dst + ( (w-1) - x - 0 + PREFETCH_ORDER_X )*h + y, 1);*/
|
||||
|
||||
/* Neon Load */
|
||||
v_tmp.val[0] = vld1_u16( &src[(y+0)*w + x] );
|
||||
v_tmp.val[1] = vld1_u16( &src[(y+1)*w + x] );
|
||||
v_tmp.val[2] = vld1_u16( &src[(y+2)*w + x] );
|
||||
v_tmp.val[3] = vld1_u16( &src[(y+3)*w + x] );
|
||||
|
||||
/* Neon store (4 interleaved) */
|
||||
vst4_lane_u16( &dst[ ( (w-1) - x - 3 )*h + y ], v_tmp, 3);
|
||||
vst4_lane_u16( &dst[ ( (w-1) - x - 2 )*h + y ], v_tmp, 2);
|
||||
vst4_lane_u16( &dst[ ( (w-1) - x - 1 )*h + y ], v_tmp, 1);
|
||||
vst4_lane_u16( &dst[ ( (w-1) - x - 0 )*h + y ], v_tmp, 0);
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif //__ARM_FP
|
||||
|
||||
|
||||
@ -891,7 +1134,7 @@ int launch_prod_screen_tests(int argc, char *argv[]){
|
||||
|
||||
|
||||
/****************************************** 3 bis (non squared) ********************************/
|
||||
#if 1
|
||||
#if 0
|
||||
/* Vars */
|
||||
int w = image_rgb_16b->w/2, h = image_rgb_16b->h;
|
||||
SDL_Surface *image_rgb_16b_notsquare = SDL_CreateRGBSurface(SDL_SWSURFACE, w, h, 16, 0,0,0,0);
|
||||
@ -1008,6 +1251,7 @@ int h = image_rgb_16b->h, w = image_rgb_16b->w;
|
||||
uint16_t * p = (uint16_t *)image_rgb_16b->pixels;
|
||||
SDL_Surface *image_rgb_16b_transposed = SDL_CreateRGBSurface(SDL_SWSURFACE, h, w, 16, 0,0,0,0);
|
||||
uint16_t * p2 = (uint16_t *)image_rgb_16b_transposed->pixels;
|
||||
|
||||
int i;
|
||||
uint32_t now = SDL_GetTicks();
|
||||
|
||||
@ -1068,6 +1312,65 @@ now = SDL_GetTicks();
|
||||
|
||||
|
||||
|
||||
/****************************************** 4 bis (optims neon) ********************************/
|
||||
#if 1
|
||||
//* Vars */
|
||||
int h = image_rgb_16b->h, w = image_rgb_16b->w;
|
||||
uint16_t * p = (uint16_t *)image_rgb_16b->pixels;
|
||||
SDL_Surface *image_rgb_16b_transposed = SDL_CreateRGBSurface(SDL_SWSURFACE, h, w, 16, 0,0,0,0);
|
||||
uint16_t * p2 = (uint16_t *)image_rgb_16b_transposed->pixels;
|
||||
|
||||
int i;
|
||||
uint32_t now = SDL_GetTicks();
|
||||
|
||||
|
||||
/* Saved perfs for 10000 iterations: */
|
||||
/* Rotate square optimized with memcpy: 2381ms
|
||||
* Rotate square optimized exported: 9847ms
|
||||
* Translate soft: 8645ms
|
||||
* Translate soft a la mano 4x4: 11110ms
|
||||
* Translate neon: 7800ms
|
||||
*/
|
||||
|
||||
|
||||
|
||||
#define ITERATIONS 3000
|
||||
printf("\n");
|
||||
|
||||
#ifdef __ARM_FP
|
||||
|
||||
/* Rotate 270 with prefetch neon */
|
||||
for (i=0; i<ITERATIONS; i++) fbtft_rotate_270cw_neon_prefetch2(p, p2, w, h);
|
||||
printf("Rotate 270 cw neon with prefetch2: %dms\n", SDL_GetTicks()-now);
|
||||
now = SDL_GetTicks();
|
||||
|
||||
/* Rotate 270 neon */
|
||||
for (i=0; i<ITERATIONS; i++) fbtft_rotate_270cw_neon(p, p2, w, h);
|
||||
printf("Rotate 270 cw neon: %dms\n", SDL_GetTicks()-now);
|
||||
now = SDL_GetTicks();
|
||||
|
||||
/* Rotate 270 with prefetch neon */
|
||||
for (i=0; i<ITERATIONS; i++) fbtft_rotate_270cw_neon_prefetch(p, p2, w, h);
|
||||
printf("Rotate 270 cw neon with prefetch: %dms\n", SDL_GetTicks()-now);
|
||||
now = SDL_GetTicks();
|
||||
|
||||
#endif //__ARM_FP
|
||||
|
||||
#endif
|
||||
/*************************************************************************************/
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
/****************************************** 4 (perfs tests 320x240) ********************************/
|
||||
#if 0
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user