ryan/mozjpeg
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Simplify the code somewhat. It actually wasn't necessary to have a "fast path" and a "medium path"-- they perform the same.

Browse Source 
git-svn-id: svn+ssh://svn.code.sf.net/p/libjpeg-turbo/code/trunk@1486 632fc199-4ca6-4c93-a231-07263d6284db
This commit is contained in:
DRC
2015-01-10 12:09:11 +00:00
parent 25347882ed
commit a5005751a3
2 changed files with 92 additions and 162 deletions
Download Patch File Download Diff File Expand all files Collapse all files

65
simd/jccolext-altivec.c
View File

@@ -70,27 +70,13 @@ void jsimd_rgb_ycc_convert_altivec (JDIMENSION img_width, JSAMPARRAY input_buf,
__vector unsigned char unaligned_shift_index; __vector unsigned char unaligned_shift_index;
int bytes = num_cols + offset; int bytes = num_cols + offset;
if (bytes >= (RGB_PIXELSIZE + 1) * 16) { if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
/* Fast path -- we have enough buffer space to load all vectors.
* Even if we don't need them all, this is faster than narrowing
* down which ones we need.
*/
rgb0 = vec_ld(0, inptr);
rgb1 = vec_ld(16, inptr);
rgb2 = vec_ld(32, inptr);
rgb3 = vec_ld(48, inptr);
#if RGB_PIXELSIZE == 4
rgb4 = vec_ld(64, inptr);
#endif
} else {
if (bytes & 15) {
/* Slow path to prevent buffer overread. Since there is no way to /* Slow path to prevent buffer overread. Since there is no way to
* read a partial AltiVec register, overread would occur on the * read a partial AltiVec register, overread would occur on the last
* last chunk of the last image row if the right edge is not on a * chunk of the last image row if the right edge is not on a 16-byte
* 16-byte boundary. It could also occur on other rows if the * 16-byte boundary. It could also occur on other rows if the bytes
* bytes per row is low enough. Since we can't determine whether * per row is low enough. Since we can't determine whether we're on
* we're on the last image row, we have to assume every row is the * the last image row, we have to assume every row is the last.
* last.
*/ */
memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
rgb0 = vec_ld(0, tmpbuf); rgb0 = vec_ld(0, tmpbuf);
@@ -99,28 +85,19 @@ void jsimd_rgb_ycc_convert_altivec (JDIMENSION img_width, JSAMPARRAY input_buf,
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
rgb3 = vec_ld(48, tmpbuf); rgb3 = vec_ld(48, tmpbuf);
#endif #endif
goto start; /* Skip permutation */
} else { } else {
/* Medium path -- if the right edge is vector-aligned, then we can /* Fast path */
* read full vectors (but with a lot of branches.)
*/
rgb0 = vec_ld(0, inptr); rgb0 = vec_ld(0, inptr);
if (bytes > 16) { if (bytes > 16)
rgb1 = vec_ld(16, inptr); rgb1 = vec_ld(16, inptr);
if (bytes > 32) { if (bytes > 32)
rgb2 = vec_ld(32, inptr); rgb2 = vec_ld(32, inptr);
if (bytes > 48) { if (bytes > 48)
rgb3 = vec_ld(48, inptr); rgb3 = vec_ld(48, inptr);
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
if (bytes > 64) if (bytes > 64)
rgb4 = vec_ld(64, inptr); rgb4 = vec_ld(64, inptr);
#endif #endif
}
}
}
}
}
unaligned_shift_index = vec_lvsl(0, inptr); unaligned_shift_index = vec_lvsl(0, inptr);
rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index); rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index);
rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index); rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index);
@@ -128,17 +105,9 @@ void jsimd_rgb_ycc_convert_altivec (JDIMENSION img_width, JSAMPARRAY input_buf,
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index); rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index);
#endif #endif
}
} else { } else {
if (num_cols >= RGB_PIXELSIZE * 16) { if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
/* Fast path */
rgb0 = vec_ld(0, inptr);
rgb1 = vec_ld(16, inptr);
rgb2 = vec_ld(32, inptr);
#if RGB_PIXELSIZE == 4
rgb3 = vec_ld(48, inptr);
#endif
} else {
if (num_cols & 15) {
/* Slow path */ /* Slow path */
memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
rgb0 = vec_ld(0, tmpbuf); rgb0 = vec_ld(0, tmpbuf);
@@ -148,11 +117,11 @@ void jsimd_rgb_ycc_convert_altivec (JDIMENSION img_width, JSAMPARRAY input_buf,
rgb3 = vec_ld(48, tmpbuf); rgb3 = vec_ld(48, tmpbuf);
#endif #endif
} else { } else {
/* Medium path */ /* Fast path */
rgb0 = vec_ld(0, inptr); rgb0 = vec_ld(0, inptr);
if (num_cols > 16) { if (num_cols > 16)
rgb1 = vec_ld(16, inptr); rgb1 = vec_ld(16, inptr);
if (num_cols > 32) { if (num_cols > 32)
rgb2 = vec_ld(32, inptr); rgb2 = vec_ld(32, inptr);
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
if (num_cols > 48) if (num_cols > 48)
@@ -160,11 +129,7 @@ void jsimd_rgb_ycc_convert_altivec (JDIMENSION img_width, JSAMPARRAY input_buf,
#endif #endif
} }
} }
}
}
}
start:
#if RGB_PIXELSIZE == 3 #if RGB_PIXELSIZE == 3
/* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5 /* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
* rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga

65
simd/jcgryext-altivec.c
View File

@@ -65,27 +65,13 @@ void jsimd_rgb_gray_convert_altivec (JDIMENSION img_width,
__vector unsigned char unaligned_shift_index; __vector unsigned char unaligned_shift_index;
int bytes = num_cols + offset; int bytes = num_cols + offset;
if (bytes >= (RGB_PIXELSIZE + 1) * 16) { if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
/* Fast path -- we have enough buffer space to load all vectors.
* Even if we don't need them all, this is faster than narrowing
* down which ones we need.
*/
rgb0 = vec_ld(0, inptr);
rgb1 = vec_ld(16, inptr);
rgb2 = vec_ld(32, inptr);
rgb3 = vec_ld(48, inptr);
#if RGB_PIXELSIZE == 4
rgb4 = vec_ld(64, inptr);
#endif
} else {
if (bytes & 15) {
/* Slow path to prevent buffer overread. Since there is no way to /* Slow path to prevent buffer overread. Since there is no way to
* read a partial AltiVec register, overread would occur on the * read a partial AltiVec register, overread would occur on the last
* last chunk of the last image row if the right edge is not on a * chunk of the last image row if the right edge is not on a 16-byte
* 16-byte boundary. It could also occur on other rows if the * 16-byte boundary. It could also occur on other rows if the bytes
* bytes per row is low enough. Since we can't determine whether * per row is low enough. Since we can't determine whether we're on
* we're on the last image row, we have to assume every row is the * the last image row, we have to assume every row is the last.
* last.
*/ */
memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
rgb0 = vec_ld(0, tmpbuf); rgb0 = vec_ld(0, tmpbuf);
@@ -94,28 +80,19 @@ void jsimd_rgb_gray_convert_altivec (JDIMENSION img_width,
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
rgb3 = vec_ld(48, tmpbuf); rgb3 = vec_ld(48, tmpbuf);
#endif #endif
goto start; /* Skip permutation */
} else { } else {
/* Medium path -- if the right edge is vector-aligned, then we can /* Fast path */
* read full vectors (but with a lot of branches.)
*/
rgb0 = vec_ld(0, inptr); rgb0 = vec_ld(0, inptr);
if (bytes > 16) { if (bytes > 16)
rgb1 = vec_ld(16, inptr); rgb1 = vec_ld(16, inptr);
if (bytes > 32) { if (bytes > 32)
rgb2 = vec_ld(32, inptr); rgb2 = vec_ld(32, inptr);
if (bytes > 48) { if (bytes > 48)
rgb3 = vec_ld(48, inptr); rgb3 = vec_ld(48, inptr);
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
if (bytes > 64) if (bytes > 64)
rgb4 = vec_ld(64, inptr); rgb4 = vec_ld(64, inptr);
#endif #endif
}
}
}
}
}
unaligned_shift_index = vec_lvsl(0, inptr); unaligned_shift_index = vec_lvsl(0, inptr);
rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index); rgb0 = vec_perm(rgb0, rgb1, unaligned_shift_index);
rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index); rgb1 = vec_perm(rgb1, rgb2, unaligned_shift_index);
@@ -123,17 +100,9 @@ void jsimd_rgb_gray_convert_altivec (JDIMENSION img_width,
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index); rgb3 = vec_perm(rgb3, rgb4, unaligned_shift_index);
#endif #endif
}
} else { } else {
if (num_cols >= RGB_PIXELSIZE * 16) { if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
/* Fast path */
rgb0 = vec_ld(0, inptr);
rgb1 = vec_ld(16, inptr);
rgb2 = vec_ld(32, inptr);
#if RGB_PIXELSIZE == 4
rgb3 = vec_ld(48, inptr);
#endif
} else {
if (num_cols & 15) {
/* Slow path */ /* Slow path */
memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16)); memcpy(tmpbuf, inptr, min(num_cols, RGB_PIXELSIZE * 16));
rgb0 = vec_ld(0, tmpbuf); rgb0 = vec_ld(0, tmpbuf);
@@ -143,11 +112,11 @@ void jsimd_rgb_gray_convert_altivec (JDIMENSION img_width,
rgb3 = vec_ld(48, tmpbuf); rgb3 = vec_ld(48, tmpbuf);
#endif #endif
} else { } else {
/* Medium path */ /* Fast path */
rgb0 = vec_ld(0, inptr); rgb0 = vec_ld(0, inptr);
if (num_cols > 16) { if (num_cols > 16)
rgb1 = vec_ld(16, inptr); rgb1 = vec_ld(16, inptr);
if (num_cols > 32) { if (num_cols > 32)
rgb2 = vec_ld(32, inptr); rgb2 = vec_ld(32, inptr);
#if RGB_PIXELSIZE == 4 #if RGB_PIXELSIZE == 4
if (num_cols > 48) if (num_cols > 48)
@@ -155,11 +124,7 @@ void jsimd_rgb_gray_convert_altivec (JDIMENSION img_width,
#endif #endif
} }
} }
}
}
}
start:
#if RGB_PIXELSIZE == 3 #if RGB_PIXELSIZE == 3
/* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5 /* rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
* rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga