Add fast paths in BlitNtoNKey

All following conversions are faster (with colorkey, but no blending).
(ratio isn't very accurate)

ABGR8888 -> BGR888 :  faster x9   (2699035 -> 297425)

ARGB8888 -> RGB888 :  faster x8   (2659266 -> 296137)

BGR24 -> BGR24 :  faster x5   (2232482 -> 445897)
BGR24 -> RGB24 :  faster x4   (2150023 -> 448576)

BGR888 -> ABGR8888 :  faster x8   (2649957 -> 307595)

BGRA8888 -> BGRX8888 :  faster x9   (2696041 -> 297596)

BGRX8888 -> BGRA8888 :  faster x8   (2662011 -> 299463)
BGRX8888 -> BGRX8888 :  faster x9   (2733346 -> 295045)

RGB24 -> BGR24 :  faster x4   (2154551 -> 485262)
RGB24 -> RGB24 :  faster x4   (2149878 -> 484870)

RGB888 -> ARGB8888 :  faster x8   (2762877 -> 324946)

RGBA8888 -> RGBX8888 :  faster x8   (2657855 -> 297753)

RGBX8888 -> RGBA8888 :  faster x8   (2661360 -> 296655)
RGBX8888 -> RGBX8888 :  faster x8   (2649287 -> 308268)
This commit is contained in:
Sylvain Becker 2019-01-30 22:50:20 +01:00
parent cd25c83aad
commit 7b8bac5958
1 changed files with 129 additions and 46 deletions

View File

@ -2329,35 +2329,120 @@ BlitNtoNKey(SDL_BlitInfo * info)
int dstbpp = dstfmt->BytesPerPixel;
unsigned alpha = dstfmt->Amask ? info->a : 0;
Uint32 rgbmask = ~srcfmt->Amask;
int sfmt = srcfmt->format;
int dfmt = dstfmt->format;
/* Set up some basic variables */
ckey &= rgbmask;
/* Fastpath: same source/destination format, no Amask, bpp 32, loop is vectorized. ~10x faster */
if (srcfmt->format == dstfmt->format &&
(srcfmt->format == SDL_PIXELFORMAT_RGB888 || srcfmt->format == SDL_PIXELFORMAT_BGR888)) {
/* BPP 4, same rgb */
if (srcbpp == 4 && dstbpp == 4 && srcfmt->Rmask == dstfmt->Rmask && srcfmt->Gmask == dstfmt->Gmask && srcfmt->Bmask == dstfmt->Bmask) {
Uint32 *src32 = (Uint32*)src;
Uint32 *dst32 = (Uint32*)dst;
srcskip /= sizeof(Uint32);
dstskip /= sizeof(Uint32);
if (dstfmt->Amask) {
/* RGB->RGBA, SET_ALPHA */
Uint32 mask = info->a << dstfmt->Ashift;
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
{
if ((*src32 & rgbmask) != ckey) {
*dst32 = *src32 | mask;
}
++dst32;
++src32;
}, width);
/* *INDENT-ON* */
src32 = (Uint32 *) ((Uint8 *) src32 + srcskip);
dst32 = (Uint32 *) ((Uint8 *) dst32 + dstskip);
}
return;
} else {
/* RGBA->RGB, NO_ALPHA */
Uint32 mask = srcfmt->Rmask | srcfmt->Gmask | srcfmt->Bmask;
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
{
if ((*src32 & rgbmask) != ckey) {
*dst32 = *src32 & mask;
}
++dst32;
++src32;
}, width);
/* *INDENT-ON* */
src32 = (Uint32 *) ((Uint8 *) src32 + srcskip);
dst32 = (Uint32 *) ((Uint8 *) dst32 + dstskip);
}
return;
}
}
/* BPP 3, same rgb triplet */
if ((sfmt == SDL_PIXELFORMAT_RGB24 && dfmt == SDL_PIXELFORMAT_RGB24) ||
(sfmt == SDL_PIXELFORMAT_BGR24 && dfmt == SDL_PIXELFORMAT_BGR24)) {
Uint8 k0 = ckey & 0x000000FF;
Uint8 k1 = (ckey & 0x0000FF00) >> 8;
Uint8 k2 = (ckey & 0x00FF0000) >> 16;
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
{
if (*src32 != ckey) {
*dst32 = *src32;
Uint8 s0 = src[0];
Uint8 s1 = src[1];
Uint8 s2 = src[2];
if (k0 != s0 || k1 != s1 || k2 != s2) {
dst[0] = s0;
dst[1] = s1;
dst[2] = s2;
}
++src32;
++dst32;
src += 3;
dst += 3;
},
width);
/* *INDENT-ON* */
src32 += srcskip;
dst32 += dstskip;
src += srcskip;
dst += dstskip;
}
return;
}
/* BPP 3, inversed rgb triplet */
if ((sfmt == SDL_PIXELFORMAT_RGB24 && dfmt == SDL_PIXELFORMAT_BGR24) ||
(sfmt == SDL_PIXELFORMAT_BGR24 && dfmt == SDL_PIXELFORMAT_RGB24)) {
Uint8 k0 = ckey & 0xFF;
Uint8 k1 = (ckey >> 8) & 0xFF;
Uint8 k2 = (ckey >> 16) & 0xFF;
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
{
Uint8 s0 = src[0];
Uint8 s1 = src[1];
Uint8 s2 = src[2];
if (k0 != s0 || k1 != s1 || k2 != s2) {
/* Inversed RGB */
dst[0] = s2;
dst[1] = s1;
dst[2] = s0;
}
src += 3;
dst += 3;
},
width);
/* *INDENT-ON* */
src += srcskip;
dst += dstskip;
}
return;
}
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
@ -2406,29 +2491,30 @@ BlitNtoNKeyCopyAlpha(SDL_BlitInfo * info)
ckey &= rgbmask;
/* Fastpath: same source/destination format, with Amask, bpp 32, loop is vectorized. ~10x faster */
if (srcfmt->format == dstfmt->format &&
(srcfmt->format == SDL_PIXELFORMAT_ARGB8888 ||
srcfmt->format == SDL_PIXELFORMAT_ABGR8888 ||
srcfmt->format == SDL_PIXELFORMAT_BGRA8888 ||
srcfmt->format == SDL_PIXELFORMAT_RGBA8888)) {
Uint32 *src32 = (Uint32*)src;
Uint32 *dst32 = (Uint32*)dst;
srcskip /= sizeof(Uint32);
dstskip /= sizeof(Uint32);
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
{
if ((*src32 & rgbmask) != ckey) {
*dst32 = *src32;
}
++src32;
++dst32;
},
width);
/* *INDENT-ON* */
src32 += srcskip;
dst32 += dstskip;
if (srcfmt->format == dstfmt->format) {
if (srcfmt->format == SDL_PIXELFORMAT_ARGB8888 ||
srcfmt->format == SDL_PIXELFORMAT_ABGR8888 ||
srcfmt->format == SDL_PIXELFORMAT_BGRA8888 ||
srcfmt->format == SDL_PIXELFORMAT_RGBA8888) {
Uint32 *src32 = (Uint32*)src;
Uint32 *dst32 = (Uint32*)dst;
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
{
if ((*src32 & rgbmask) != ckey) {
*dst32 = *src32;
}
++src32;
++dst32;
},
width);
/* *INDENT-ON* */
src32 = (Uint32 *)((Uint8 *)src32 + srcskip);
dst32 = (Uint32 *)((Uint8 *)dst32 + dstskip);
}
}
return;
}
@ -2532,8 +2618,7 @@ Blit_3or4_to_3or4__same_rgb(SDL_BlitInfo * info)
if (dstfmt->Amask) {
/* SET_ALPHA */
Uint32 alpha = info->a;
Uint32 alphashift = alpha << 24;
Uint32 mask = info->a << dstfmt->Ashift;
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
@ -2542,7 +2627,7 @@ Blit_3or4_to_3or4__same_rgb(SDL_BlitInfo * info)
Uint8 s0 = src[0];
Uint8 s1 = src[1];
Uint8 s2 = src[2];
*dst32 = (s0) | (s1 << 8) | (s2 << 16) | alphashift;
*dst32 = (s0) | (s1 << 8) | (s2 << 16) | mask;
dst += dstbpp;
src += srcbpp;
}, width);
@ -2588,7 +2673,6 @@ Blit_3or4_to_3or4__inversed_rgb(SDL_BlitInfo * info)
int dstbpp = dstfmt->BytesPerPixel;
if (dstfmt->Amask) {
if (srcfmt->Amask) {
/* COPY_ALPHA */
/* Only to switch ABGR8888 <-> ARGB8888 */
@ -2596,11 +2680,11 @@ Blit_3or4_to_3or4__inversed_rgb(SDL_BlitInfo * info)
/* *INDENT-OFF* */
DUFFS_LOOP(
{
Uint32 *dst32 = (Uint32*)dst;
Uint32 *dst32 = (Uint32*)dst;
Uint8 s0 = src[0];
Uint8 s1 = src[1];
Uint8 s2 = src[2];
Uint32 alphashift = src[3] << 24;
Uint32 alphashift = src[3] << dstfmt->Ashift;
/* inversed, compared to Blit_3or4_to_3or4__same_rgb */
*dst32 = (s0 << 16) | (s1 << 8) | (s2) | alphashift;
dst += dstbpp;
@ -2612,18 +2696,17 @@ Blit_3or4_to_3or4__inversed_rgb(SDL_BlitInfo * info)
}
} else {
/* SET_ALPHA */
Uint32 alpha = info->a;
Uint32 alphashift = alpha << 24;
Uint32 mask = info->a << dstfmt->Ashift;
while (height--) {
/* *INDENT-OFF* */
DUFFS_LOOP(
{
Uint32 *dst32 = (Uint32*)dst;
Uint32 *dst32 = (Uint32*)dst;
Uint8 s0 = src[0];
Uint8 s1 = src[1];
Uint8 s2 = src[2];
/* inversed, compared to Blit_3or4_to_3or4__same_rgb */
*dst32 = (s0 << 16) | (s1 << 8) | (s2) | alphashift;
*dst32 = (s0 << 16) | (s1 << 8) | (s2) | mask;
dst += dstbpp;
src += srcbpp;
}, width);
@ -2638,7 +2721,7 @@ Blit_3or4_to_3or4__inversed_rgb(SDL_BlitInfo * info)
/* *INDENT-OFF* */
DUFFS_LOOP(
{
Uint32 *dst32 = (Uint32*)dst;
Uint32 *dst32 = (Uint32*)dst;
Uint8 s0 = src[0];
Uint8 s1 = src[1];
Uint8 s2 = src[2];
@ -2745,7 +2828,7 @@ static const struct blit_table normal_blit_4[] = {
0, Blit_3or4_to_3or4__inversed_rgb, NO_ALPHA | SET_ALPHA | COPY_ALPHA},
{0x00FF0000, 0x0000FF00, 0x000000FF, 4, 0x000000FF, 0x0000FF00, 0x00FF0000,
0, Blit_3or4_to_3or4__inversed_rgb, NO_ALPHA | SET_ALPHA | COPY_ALPHA},
/* RBG 888 and RGB 565 */
/* RGB 888 and RGB 565 */
{0x00FF0000, 0x0000FF00, 0x000000FF, 2, 0x0000F800, 0x000007E0, 0x0000001F,
0, Blit_RGB888_RGB565, NO_ALPHA},
{0x00FF0000, 0x0000FF00, 0x000000FF, 2, 0x00007C00, 0x000003E0, 0x0000001F,