highperfocused/ffmpeg
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

avcodec/x86/h26x/h2656_inter: add dststride to put

Browse Source 
Signed-off-by: Wu Jianhua <toqsxw@outlook.com>
This commit is contained in:
Wu Jianhua
2024-01-24 02:17:07 +08:00
committed by Nuo Mi
parent 7d9f1f5485
commit fc5ff6b0b8
4 changed files with 19 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
libavcodec/x86/h26x/h2656_inter.asm
View File

@@ -22,8 +22,6 @@
; */ ; */
%include "libavutil/x86/x86util.asm" %include "libavutil/x86/x86util.asm"
%define MAX_PB_SIZE 64
SECTION_RODATA 32 SECTION_RODATA 32
cextern pw_255 cextern pw_255
cextern pw_512 cextern pw_512
@@ -342,7 +340,7 @@ SECTION .text
%endmacro %endmacro
%macro LOOP_END 3 %macro LOOP_END 3
add %1q, 2*MAX_PB_SIZE ; dst += dststride add %1q, dststrideq ; dst += dststride
add %2q, %3q ; src += srcstride add %2q, %3q ; src += srcstride
dec heightd ; cmp height dec heightd ; cmp height
jnz .loop ; height loop jnz .loop ; height loop
@@ -539,7 +537,7 @@ SECTION .text
; ****************************** ; ******************************
; void %1_put_pixels(int16_t *dst, const uint8_t *_src, ptrdiff_t srcstride, ; void %1_put_pixels(int16_t *dst, ptrdiff_t dststride, const uint8_t *_src, ptrdiff_t srcstride,
; int height, const int8_t *hf, const int8_t *vf, int width) ; int height, const int8_t *hf, const int8_t *vf, int width)
; ****************************** ; ******************************
@@ -549,7 +547,7 @@ SECTION .text
%endmacro %endmacro
%macro MC_PIXELS 3 %macro MC_PIXELS 3
cglobal %1_put_pixels%2_%3, 4, 4, 3, dst, src, srcstride, height cglobal %1_put_pixels%2_%3, 5, 5, 3, dst, dststride, src, srcstride, height
pxor m2, m2 pxor m2, m2
.loop: .loop:
SIMPLE_LOAD %2, %3, srcq, m0 SIMPLE_LOAD %2, %3, srcq, m0
@@ -579,10 +577,10 @@ cglobal %1_put_uni_pixels%2_%3, 5, 5, 2, dst, dststride, src, srcstride, height
%endif %endif
; ****************************** ; ******************************
; void %1_put_4tap_hX(int16_t *dst, ; void %1_put_4tap_hX(int16_t *dst, ptrdiff_t dststride,
; const uint8_t *_src, ptrdiff_t _srcstride, int height, int8_t *hf, int8_t *vf, int width); ; const uint8_t *_src, ptrdiff_t _srcstride, int height, int8_t *hf, int8_t *vf, int width);
; ****************************** ; ******************************
cglobal %1_put_4tap_h%2_%3, 5, 5, XMM_REGS, dst, src, srcstride, height, hf cglobal %1_put_4tap_h%2_%3, 6, 6, XMM_REGS, dst, dststride, src, srcstride, height, hf
%assign %%stride ((%3 + 7)/8) %assign %%stride ((%3 + 7)/8)
MC_4TAP_FILTER %3, hf, m4, m5 MC_4TAP_FILTER %3, hf, m4, m5
.loop: .loop:
@@ -612,10 +610,10 @@ cglobal %1_put_uni_4tap_h%2_%3, 6, 7, XMM_REGS, dst, dststride, src, srcstride,
RET RET
; ****************************** ; ******************************
; void %1_put_4tap_v(int16_t *dst, ; void %1_put_4tap_v(int16_t *dst, ptrdiff_t dststride,
; const uint8_t *_src, ptrdiff_t _srcstride, int height, int8_t *hf, int8_t *vf, int width) ; const uint8_t *_src, ptrdiff_t _srcstride, int height, int8_t *hf, int8_t *vf, int width)
; ****************************** ; ******************************
cglobal %1_put_4tap_v%2_%3, 6, 6, XMM_REGS, dst, src, srcstride, height, r3src, vf cglobal %1_put_4tap_v%2_%3, 7, 7, XMM_REGS, dst, dststride, src, srcstride, height, r3src, vf
sub srcq, srcstrideq sub srcq, srcstrideq
MC_4TAP_FILTER %3, vf, m4, m5 MC_4TAP_FILTER %3, vf, m4, m5
lea r3srcq, [srcstrideq*3] lea r3srcq, [srcstrideq*3]
@@ -649,10 +647,10 @@ cglobal %1_put_uni_4tap_v%2_%3, 7, 7, XMM_REGS, dst, dststride, src, srcstride,
%macro PUT_4TAP_HV 3 %macro PUT_4TAP_HV 3
; ****************************** ; ******************************
; void put_4tap_hv(int16_t *dst, ; void put_4tap_hv(int16_t *dst, ptrdiff_t dststride,
; const uint8_t *_src, ptrdiff_t _srcstride, int height, int8_t *hf, int8_t *vf, int width) ; const uint8_t *_src, ptrdiff_t _srcstride, int height, int8_t *hf, int8_t *vf, int width)
; ****************************** ; ******************************
cglobal %1_put_4tap_hv%2_%3, 6, 7, 16 , dst, src, srcstride, height, hf, vf, r3src cglobal %1_put_4tap_hv%2_%3, 7, 8, 16 , dst, dststride, src, srcstride, height, hf, vf, r3src
%assign %%stride ((%3 + 7)/8) %assign %%stride ((%3 + 7)/8)
sub srcq, srcstrideq sub srcq, srcstrideq
MC_4TAP_HV_FILTER %3 MC_4TAP_HV_FILTER %3
@@ -784,12 +782,12 @@ cglobal %1_put_uni_4tap_hv%2_%3, 7, 8, 16 , dst, dststride, src, srcstride, heig
%endmacro %endmacro
; ****************************** ; ******************************
; void put_8tap_hX_X_X(int16_t *dst, const uint8_t *_src, ptrdiff_t srcstride, ; void put_8tap_hX_X_X(int16_t *dst, ptrdiff_t dststride, const uint8_t *_src, ptrdiff_t srcstride,
; int height, const int8_t *hf, const int8_t *vf, int width) ; int height, const int8_t *hf, const int8_t *vf, int width)
; ****************************** ; ******************************
%macro PUT_8TAP 3 %macro PUT_8TAP 3
cglobal %1_put_8tap_h%2_%3, 5, 5, 16, dst, src, srcstride, height, hf cglobal %1_put_8tap_h%2_%3, 6, 6, 16, dst, dststride, src, srcstride, height, hf
MC_8TAP_FILTER %3, hf MC_8TAP_FILTER %3, hf
.loop: .loop:
MC_8TAP_H_LOAD %3, srcq, %2, 10 MC_8TAP_H_LOAD %3, srcq, %2, 10
@@ -824,10 +822,10 @@ cglobal %1_put_uni_8tap_h%2_%3, 6, 7, 16 , dst, dststride, src, srcstride, heigh
; ****************************** ; ******************************
; void put_8tap_vX_X_X(int16_t *dst, const uint8_t *_src, ptrdiff_t srcstride, ; void put_8tap_vX_X_X(int16_t *dst, ptrdiff_t dststride, const uint8_t *_src, ptrdiff_t srcstride,
; int height, const int8_t *hf, const int8_t *vf, int width) ; int height, const int8_t *hf, const int8_t *vf, int width)
; ****************************** ; ******************************
cglobal %1_put_8tap_v%2_%3, 6, 8, 16, dst, src, srcstride, height, r3src, vf cglobal %1_put_8tap_v%2_%3, 7, 8, 16, dst, dststride, src, srcstride, height, r3src, vf
MC_8TAP_FILTER %3, vf MC_8TAP_FILTER %3, vf
lea r3srcq, [srcstrideq*3] lea r3srcq, [srcstrideq*3]
.loop: .loop:
@@ -866,11 +864,11 @@ cglobal %1_put_uni_8tap_v%2_%3, 7, 9, 16, dst, dststride, src, srcstride, height
; ****************************** ; ******************************
; void put_8tap_hvX_X(int16_t *dst, const uint8_t *_src, ptrdiff_t srcstride, ; void put_8tap_hvX_X(int16_t *dst, ptrdiff_t dststride, const uint8_t *_src, ptrdiff_t srcstride,
; int height, const int8_t *hf, const int8_t *vf, int width) ; int height, const int8_t *hf, const int8_t *vf, int width)
; ****************************** ; ******************************
%macro PUT_8TAP_HV 3 %macro PUT_8TAP_HV 3
cglobal %1_put_8tap_hv%2_%3, 6, 7, 16, 0 - mmsize*16, dst, src, srcstride, height, hf, vf, r3src cglobal %1_put_8tap_hv%2_%3, 7, 8, 16, 0 - mmsize*16, dst, dststride, src, srcstride, height, hf, vf, r3src
MC_8TAP_FILTER %3, hf, 0 MC_8TAP_FILTER %3, hf, 0
lea hfq, [rsp] lea hfq, [rsp]
MC_8TAP_FILTER %3, vf, 8*mmsize MC_8TAP_FILTER %3, vf, 8*mmsize

4
libavcodec/x86/h26x/h2656dsp.c
View File

@@ -24,7 +24,7 @@
#include "h2656dsp.h" #include "h2656dsp.h"
#define mc_rep_func(name, bitd, step, W, opt) \ #define mc_rep_func(name, bitd, step, W, opt) \
void ff_h2656_put_##name##W##_##bitd##_##opt(int16_t *_dst, \ void ff_h2656_put_##name##W##_##bitd##_##opt(int16_t *_dst, ptrdiff_t dststride, \
const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width) \ const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width) \
{ \ { \
int i; \ int i; \
@@ -32,7 +32,7 @@ void ff_h2656_put_##name##W##_##bitd##_##opt(int16_t *_dst,
for (i = 0; i < W; i += step) { \ for (i = 0; i < W; i += step) { \
const uint8_t *src = _src + (i * ((bitd + 7) / 8)); \ const uint8_t *src = _src + (i * ((bitd + 7) / 8)); \
dst = _dst + i; \ dst = _dst + i; \
ff_h2656_put_##name##step##_##bitd##_##opt(dst, src, _srcstride, height, hf, vf, width); \ ff_h2656_put_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, height, hf, vf, width); \
} \ } \
} }

2
libavcodec/x86/h26x/h2656dsp.h
View File

@@ -30,7 +30,7 @@
#include <stdlib.h> #include <stdlib.h>
#define H2656_PEL_PROTOTYPE(name, D, opt) \ #define H2656_PEL_PROTOTYPE(name, D, opt) \
void ff_h2656_put_ ## name ## _ ## D ## _##opt(int16_t *dst, const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width); \ void ff_h2656_put_ ## name ## _ ## D ## _##opt(int16_t *dst, ptrdiff_t dststride, const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width); \
void ff_h2656_put_uni_ ## name ## _ ## D ## _##opt(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width); \ void ff_h2656_put_uni_ ## name ## _ ## D ## _##opt(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, int height, const int8_t *hf, const int8_t *vf, int width); \
#define H2656_MC_8TAP_PROTOTYPES(fname, bitd, opt) \ #define H2656_MC_8TAP_PROTOTYPES(fname, bitd, opt) \

2
libavcodec/x86/hevcdsp_init.c
View File

@@ -96,7 +96,7 @@ void ff_hevc_put_hevc_ ## a ## _ ## depth ## _##opt(int16_t *dst, const uint8_t
int height, intptr_t mx, intptr_t my,int width) \ int height, intptr_t mx, intptr_t my,int width) \
{ \ { \
DECL_HV_FILTER(p) \ DECL_HV_FILTER(p) \
ff_h2656_put_ ## b ## _ ## depth ## _##opt(dst, src, srcstride, height, hf, vf, width); \ ff_h2656_put_ ## b ## _ ## depth ## _##opt(dst, 2 * MAX_PB_SIZE, src, srcstride, height, hf, vf, width); \
} }
#define FW_PUT_UNI(p, a, b, depth, opt) \ #define FW_PUT_UNI(p, a, b, depth, opt) \