|
|
@ -30,13 +30,8 @@ |
|
|
|
|
|
|
|
|
|
|
|
#include "pixels.h" |
|
|
|
#include "pixels.h" |
|
|
|
|
|
|
|
|
|
|
|
#include "bit_depth_template.c" |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#include "hpel_template.c" |
|
|
|
|
|
|
|
#include "tpel_template.c" |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#define PIXOP2(OPNAME, OP) \ |
|
|
|
#define PIXOP2(OPNAME, OP) \ |
|
|
|
static inline void FUNC(OPNAME ## _no_rnd_pixels8_l2)(uint8_t *dst, \
|
|
|
|
static inline void OPNAME ## _no_rnd_pixels8_l2_8(uint8_t *dst, \
|
|
|
|
const uint8_t *src1, \
|
|
|
|
const uint8_t *src1, \
|
|
|
|
const uint8_t *src2, \
|
|
|
|
const uint8_t *src2, \
|
|
|
|
int dst_stride, \
|
|
|
|
int dst_stride, \
|
|
|
@ -47,104 +42,101 @@ static inline void FUNC(OPNAME ## _no_rnd_pixels8_l2)(uint8_t *dst, \ |
|
|
|
int i; \
|
|
|
|
int i; \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
for (i = 0; i < h; i++) { \
|
|
|
|
for (i = 0; i < h; i++) { \
|
|
|
|
pixel4 a, b; \
|
|
|
|
uint32_t a, b; \
|
|
|
|
a = AV_RN4P(&src1[i * src_stride1]); \
|
|
|
|
a = AV_RN32(&src1[i * src_stride1]); \
|
|
|
|
b = AV_RN4P(&src2[i * src_stride2]); \
|
|
|
|
b = AV_RN32(&src2[i * src_stride2]); \
|
|
|
|
OP(*((pixel4 *) &dst[i * dst_stride]), \
|
|
|
|
OP(*((uint32_t *) &dst[i * dst_stride]), \
|
|
|
|
no_rnd_avg_pixel4(a, b)); \
|
|
|
|
no_rnd_avg32(a, b)); \
|
|
|
|
a = AV_RN4P(&src1[i * src_stride1 + 4 * sizeof(pixel)]); \
|
|
|
|
a = AV_RN32(&src1[i * src_stride1 + 4]); \
|
|
|
|
b = AV_RN4P(&src2[i * src_stride2 + 4 * sizeof(pixel)]); \
|
|
|
|
b = AV_RN32(&src2[i * src_stride2 + 4]); \
|
|
|
|
OP(*((pixel4 *) &dst[i * dst_stride + 4 * sizeof(pixel)]), \
|
|
|
|
OP(*((uint32_t *) &dst[i * dst_stride + 4]), \
|
|
|
|
no_rnd_avg_pixel4(a, b)); \
|
|
|
|
no_rnd_avg32(a, b)); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _no_rnd_pixels8_x2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _no_rnd_pixels8_x2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _no_rnd_pixels8_l2)(block, pixels, pixels + sizeof(pixel), \
|
|
|
|
OPNAME ## _no_rnd_pixels8_l2_8(block, pixels, pixels + 1, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels8_x2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels8_x2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _pixels8_l2)(block, pixels, pixels + sizeof(pixel), \
|
|
|
|
OPNAME ## _pixels8_l2_8(block, pixels, pixels + 1, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _no_rnd_pixels8_y2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _no_rnd_pixels8_y2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _no_rnd_pixels8_l2)(block, pixels, pixels + line_size, \
|
|
|
|
OPNAME ## _no_rnd_pixels8_l2_8(block, pixels, pixels + line_size, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels8_y2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels8_y2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _pixels8_l2)(block, pixels, pixels + line_size, \
|
|
|
|
OPNAME ## _pixels8_l2_8(block, pixels, pixels + line_size, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels4_x2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels4_x2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _pixels4_l2)(block, pixels, pixels + sizeof(pixel), \
|
|
|
|
OPNAME ## _pixels4_l2_8(block, pixels, pixels + 1, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels4_y2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels4_y2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _pixels4_l2)(block, pixels, pixels + line_size, \
|
|
|
|
OPNAME ## _pixels4_l2_8(block, pixels, pixels + line_size, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels2_x2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels2_x2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _pixels2_l2)(block, pixels, pixels + sizeof(pixel), \
|
|
|
|
OPNAME ## _pixels2_l2_8(block, pixels, pixels + 1, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels2_y2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels2_y2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
FUNC(OPNAME ## _pixels2_l2)(block, pixels, pixels + line_size, \
|
|
|
|
OPNAME ## _pixels2_l2_8(block, pixels, pixels + line_size, \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
line_size, line_size, line_size, h); \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels2_xy2)(uint8_t *_block, \
|
|
|
|
static inline void OPNAME ## _pixels2_xy2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *_pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
|
{ \
|
|
|
|
{ \
|
|
|
|
pixel *block = (pixel *) _block; \
|
|
|
|
|
|
|
|
const pixel *pixels = (const pixel *) _pixels; \
|
|
|
|
|
|
|
|
int i, a1, b1; \
|
|
|
|
int i, a1, b1; \
|
|
|
|
int a0 = pixels[0]; \
|
|
|
|
int a0 = pixels[0]; \
|
|
|
|
int b0 = pixels[1] + 2; \
|
|
|
|
int b0 = pixels[1] + 2; \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
a0 += b0; \
|
|
|
|
a0 += b0; \
|
|
|
|
b0 += pixels[2]; \
|
|
|
|
b0 += pixels[2]; \
|
|
|
|
line_size /= sizeof(pixel); \
|
|
|
|
|
|
|
|
pixels += line_size; \
|
|
|
|
pixels += line_size; \
|
|
|
|
for (i = 0; i < h; i += 2) { \
|
|
|
|
for (i = 0; i < h; i += 2) { \
|
|
|
|
a1 = pixels[0]; \
|
|
|
|
a1 = pixels[0]; \
|
|
|
@ -170,7 +162,7 @@ static inline void FUNCC(OPNAME ## _pixels2_xy2)(uint8_t *_block, \ |
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels4_xy2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels4_xy2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
@ -212,7 +204,7 @@ static inline void FUNCC(OPNAME ## _pixels4_xy2)(uint8_t *block, \ |
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _pixels8_xy2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _pixels8_xy2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
@ -260,7 +252,7 @@ static inline void FUNCC(OPNAME ## _pixels8_xy2)(uint8_t *block, \ |
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
static inline void FUNCC(OPNAME ## _no_rnd_pixels8_xy2)(uint8_t *block, \
|
|
|
|
static inline void OPNAME ## _no_rnd_pixels8_xy2_8_c(uint8_t *block, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
const uint8_t *pixels, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
ptrdiff_t line_size, \
|
|
|
|
int h) \
|
|
|
|
int h) \
|
|
|
@ -308,34 +300,32 @@ static inline void FUNCC(OPNAME ## _no_rnd_pixels8_xy2)(uint8_t *block, \ |
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
} \
|
|
|
|
\
|
|
|
|
\
|
|
|
|
CALL_2X_PIXELS(FUNCC(OPNAME ## _pixels16_x2), \
|
|
|
|
CALL_2X_PIXELS(OPNAME ## _pixels16_x2_8_c, \
|
|
|
|
FUNCC(OPNAME ## _pixels8_x2), \
|
|
|
|
OPNAME ## _pixels8_x2_8_c, \
|
|
|
|
8 * sizeof(pixel)) \
|
|
|
|
8) \
|
|
|
|
CALL_2X_PIXELS(FUNCC(OPNAME ## _pixels16_y2), \
|
|
|
|
CALL_2X_PIXELS(OPNAME ## _pixels16_y2_8_c, \
|
|
|
|
FUNCC(OPNAME ## _pixels8_y2), \
|
|
|
|
OPNAME ## _pixels8_y2_8_c, \
|
|
|
|
8 * sizeof(pixel)) \
|
|
|
|
8) \
|
|
|
|
CALL_2X_PIXELS(FUNCC(OPNAME ## _pixels16_xy2), \
|
|
|
|
CALL_2X_PIXELS(OPNAME ## _pixels16_xy2_8_c, \
|
|
|
|
FUNCC(OPNAME ## _pixels8_xy2), \
|
|
|
|
OPNAME ## _pixels8_xy2_8_c, \
|
|
|
|
8 * sizeof(pixel)) \
|
|
|
|
8) \
|
|
|
|
av_unused CALL_2X_PIXELS(FUNCC(OPNAME ## _no_rnd_pixels16), \
|
|
|
|
av_unused CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_8_c, \
|
|
|
|
FUNCC(OPNAME ## _pixels8), \
|
|
|
|
OPNAME ## _pixels8_8_c, \
|
|
|
|
8 * sizeof(pixel)) \
|
|
|
|
8) \
|
|
|
|
CALL_2X_PIXELS(FUNCC(OPNAME ## _no_rnd_pixels16_x2), \
|
|
|
|
CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_x2_8_c, \
|
|
|
|
FUNCC(OPNAME ## _no_rnd_pixels8_x2), \
|
|
|
|
OPNAME ## _no_rnd_pixels8_x2_8_c, \
|
|
|
|
8 * sizeof(pixel)) \
|
|
|
|
8) \
|
|
|
|
CALL_2X_PIXELS(FUNCC(OPNAME ## _no_rnd_pixels16_y2), \
|
|
|
|
CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_y2_8_c, \
|
|
|
|
FUNCC(OPNAME ## _no_rnd_pixels8_y2), \
|
|
|
|
OPNAME ## _no_rnd_pixels8_y2_8_c, \
|
|
|
|
8 * sizeof(pixel)) \
|
|
|
|
8) \
|
|
|
|
CALL_2X_PIXELS(FUNCC(OPNAME ## _no_rnd_pixels16_xy2), \
|
|
|
|
CALL_2X_PIXELS(OPNAME ## _no_rnd_pixels16_xy2_8_c, \
|
|
|
|
FUNCC(OPNAME ## _no_rnd_pixels8_xy2), \
|
|
|
|
OPNAME ## _no_rnd_pixels8_xy2_8_c, \
|
|
|
|
8 * sizeof(pixel)) \
|
|
|
|
8) \
|
|
|
|
|
|
|
|
|
|
|
|
#define op_avg(a, b) a = rnd_avg_pixel4(a, b) |
|
|
|
#define op_avg(a, b) a = rnd_avg32(a, b) |
|
|
|
#define op_put(a, b) a = b |
|
|
|
#define op_put(a, b) a = b |
|
|
|
#if BIT_DEPTH == 8 |
|
|
|
|
|
|
|
#define put_no_rnd_pixels8_8_c put_pixels8_8_c |
|
|
|
#define put_no_rnd_pixels8_8_c put_pixels8_8_c |
|
|
|
PIXOP2(avg, op_avg) |
|
|
|
PIXOP2(avg, op_avg) |
|
|
|
PIXOP2(put, op_put) |
|
|
|
PIXOP2(put, op_put) |
|
|
|
#endif |
|
|
|
|
|
|
|
#undef op_avg |
|
|
|
#undef op_avg |
|
|
|
#undef op_put |
|
|
|
#undef op_put |
|
|
|