API tools faq
paste
Advertisement
Guest User

Untitled

a guest
Jul 7th, 2017
542
0
Never
Add comment
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
Diff 23.92 KB | None | 0 0
raw download clone embed print report
  1. From 1b9409b5264ee17651a4467529a841b8de7545f7 Mon Sep 17 00:00:00 2001
  2. From: Daniel Kang <daniel.d.kang@gmail.com>
  3. Date: Mon, 27 Dec 2010 17:11:09 -0500
  4. Subject: [PATCH 1/5] pred8x8_top_dc_mmxext
  5.  
  6. ---
  7. libavcodec/x86/h264_intrapred.asm    |   33 +++++++++++++++++++++++++++++++++
  8.  libavcodec/x86/h264_intrapred_init.c |    4 ++++
  9.  2 files changed, 37 insertions(+), 0 deletions(-)
  10.  
  11. diff --git a/libavcodec/x86/h264_intrapred.asm b/libavcodec/x86/h264_intrapred.asm
  12. index 14a6038..1cb22f5 100644
  13. --- a/libavcodec/x86/h264_intrapred.asm
  14. +++ b/libavcodec/x86/h264_intrapred.asm
  15. @@ -828,6 +828,39 @@ PRED8x8_H mmxext
  16.  PRED8x8_H ssse3
  17.  
  18.  ;-----------------------------------------------------------------------------
  19. +; void pred8x8_top_dc_mmxext(uint8_t *src, int stride)
  20. +;-----------------------------------------------------------------------------
  21. +cglobal pred8x8_top_dc_mmxext, 2,5
  22. +    sub         r0, r1
  23. +    movq       mm0, [r0]
  24. +    pxor       mm1, mm1
  25. +    pxor       mm2, mm2
  26. +    lea         r2, [r0+r1*2]
  27. +    punpckhbw  mm1, mm0
  28. +    punpcklbw  mm0, mm2
  29. +    psadbw     mm1, mm2        ; s1
  30. +    lea         r3, [r2+r1*2]
  31. +    psadbw     mm0, mm2        ; s0
  32. +    psrlw      mm1, 1
  33. +    psrlw      mm0, 1
  34. +    pavgw      mm1, mm2
  35. +    lea         r4, [r3+r1*2]
  36. +    pavgw      mm0, mm2
  37. +    pshufw     mm1, mm1, 0
  38. +    pshufw     mm0, mm0, 0     ; dc0 (w)
  39. +    packuswb   mm0, mm1        ; dc0,dc1 (b)
  40. +    movq [r0+r1*1], mm0
  41. +    movq [r0+r1*2], mm0
  42. +    lea         r0, [r3+r1*2]
  43. +    movq [r2+r1*1], mm0
  44. +    movq [r2+r1*2], mm0
  45. +    movq [r3+r1*1], mm0
  46. +    movq [r3+r1*2], mm0
  47. +    movq [r0+r1*1], mm0
  48. +    movq [r0+r1*2], mm0
  49. +    RET
  50. +
  51. +;-----------------------------------------------------------------------------
  52.  ; void pred8x8_dc_rv40(uint8_t *src, int stride)
  53.  ;-----------------------------------------------------------------------------
  54.  
  55. diff --git a/libavcodec/x86/h264_intrapred_init.c b/libavcodec/x86/h264_intrapred_init.c
  56. index 10a6dd6..57a93f7 100644
  57. --- a/libavcodec/x86/h264_intrapred_init.c
  58. +++ b/libavcodec/x86/h264_intrapred_init.c
  59. @@ -44,6 +44,7 @@ void ff_pred16x16_plane_svq3_ssse3 (uint8_t *src, int stride);
  60.  void ff_pred16x16_tm_vp8_mmx       (uint8_t *src, int stride);
  61.  void ff_pred16x16_tm_vp8_mmxext    (uint8_t *src, int stride);
  62.  void ff_pred16x16_tm_vp8_sse2      (uint8_t *src, int stride);
  63. +void ff_pred8x8_top_dc_mmxext      (uint8_t *src, int stride);
  64.  void ff_pred8x8_dc_rv40_mmxext     (uint8_t *src, int stride);
  65.  void ff_pred8x8_vertical_mmx       (uint8_t *src, int stride);
  66.  void ff_pred8x8_horizontal_mmx     (uint8_t *src, int stride);
  67. @@ -97,6 +98,9 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  68.          h->pred4x4  [DC_PRED     ] = ff_pred4x4_dc_mmxext;
  69.          if (codec_id == CODEC_ID_VP8 || codec_id == CODEC_ID_H264)
  70.              h->pred4x4  [DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_mmxext;
  71. +        if (codec_id == CODEC_ID_SVQ3 || codec_id == CODEC_ID_H264) {
  72. +            h->pred8x8[TOP_DC_PRED8x8 ] = ff_pred8x8_top_dc_mmxext;
  73. +        }
  74.          if (codec_id == CODEC_ID_VP8) {
  75.              h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_tm_vp8_mmxext;
  76.              h->pred8x8  [DC_PRED8x8   ] = ff_pred8x8_dc_rv40_mmxext;
  77. --
  78. 1.7.2.2
  79.  
  80.  
  81. From fbafcfba182db3d9539a62cefb028a087e0c473e Mon Sep 17 00:00:00 2001
  82. From: Daniel Kang <daniel.d.kang@gmail.com>
  83. Date: Mon, 27 Dec 2010 17:42:14 -0500
  84. Subject: [PATCH 2/5] pred8x8_dc_mmxext
  85.  
  86. ---
  87. libavcodec/x86/h264_intrapred.asm    |   61 ++++++++++++++++++++++++++++++++++
  88.  libavcodec/x86/h264_intrapred_init.c |    2 +
  89.  2 files changed, 63 insertions(+), 0 deletions(-)
  90.  
  91. diff --git a/libavcodec/x86/h264_intrapred.asm b/libavcodec/x86/h264_intrapred.asm
  92. index 1cb22f5..61a587d 100644
  93. --- a/libavcodec/x86/h264_intrapred.asm
  94. +++ b/libavcodec/x86/h264_intrapred.asm
  95. @@ -861,6 +861,67 @@ cglobal pred8x8_top_dc_mmxext, 2,5
  96.      RET
  97.  
  98.  ;-----------------------------------------------------------------------------
  99. +; void pred8x8_dc_mmxext(uint8_t *src, int stride)
  100. +;-----------------------------------------------------------------------------
  101. +
  102. +INIT_MMX
  103. +cglobal pred8x8_dc_mmxext, 2,5
  104. +    sub       r0, r1
  105. +    pxor      m7, m7
  106. +    movd      m0, [r0+0]
  107. +    movd      m1, [r0+4]
  108. +    psadbw    m0, m7            ; s0
  109. +    mov       r4, r0
  110. +    psadbw    m1, m7            ; s1
  111. +
  112. +    movzx    r2d, byte [r0+r1*1-1]
  113. +    movzx    r3d, byte [r0+r1*2-1]
  114. +    lea       r0, [r0+r1*2]
  115. +    add      r2d, r3d
  116. +    movzx    r3d, byte [r0+r1*1-1]
  117. +    add      r2d, r3d
  118. +    movzx    r3d, byte [r0+r1*2-1]
  119. +    add      r2d, r3d
  120. +    lea       r0, [r0+r1*2]
  121. +    movd      m2, r2d            ; s2
  122. +    movzx    r2d, byte [r0+r1*1-1]
  123. +    movzx    r3d, byte [r0+r1*2-1]
  124. +    lea       r0, [r0+r1*2]
  125. +    add      r2d, r3d
  126. +    movzx    r3d, byte [r0+r1*1-1]
  127. +    add      r2d, r3d
  128. +    movzx    r3d, byte [r0+r1*2-1]
  129. +    add      r2d, r3d
  130. +    movd      m3, r2d            ; s3
  131. +
  132. +    punpcklwd m0, m1
  133. +    mov       r0, r4
  134. +    punpcklwd m2, m3
  135. +    punpckldq m0, m2            ; s0, s1, s2, s3
  136. +    pshufw    m3, m0, 11110110b ; s2, s1, s3, s3
  137. +    lea       r2, [r0+r1*2]
  138. +    pshufw    m0, m0, 01110100b ; s0, s1, s3, s1
  139. +    paddw     m0, m3
  140. +    lea       r3, [r2+r1*2]
  141. +    psrlw     m0, 2
  142. +    pavgw     m0, m7            ; s0+s2, s1, s3, s1+s3
  143. +    lea       r4, [r3+r1*2]
  144. +    packuswb  m0, m0
  145. +    punpcklbw m0, m0
  146. +    movq      m1, m0
  147. +    punpcklbw m0, m0
  148. +    punpckhbw m1, m1
  149. +    movq [r0+r1*1], m0
  150. +    movq [r0+r1*2], m0
  151. +    movq [r2+r1*1], m0
  152. +    movq [r2+r1*2], m0
  153. +    movq [r3+r1*1], m1
  154. +    movq [r3+r1*2], m1
  155. +    movq [r4+r1*1], m1
  156. +    movq [r4+r1*2], m1
  157. +    RET
  158. +
  159. +;-----------------------------------------------------------------------------
  160.  ; void pred8x8_dc_rv40(uint8_t *src, int stride)
  161.  ;-----------------------------------------------------------------------------
  162.  
  163. diff --git a/libavcodec/x86/h264_intrapred_init.c b/libavcodec/x86/h264_intrapred_init.c
  164. index 57a93f7..8f6fd2f 100644
  165. --- a/libavcodec/x86/h264_intrapred_init.c
  166. +++ b/libavcodec/x86/h264_intrapred_init.c
  167. @@ -46,6 +46,7 @@ void ff_pred16x16_tm_vp8_mmxext    (uint8_t *src, int stride);
  168.  void ff_pred16x16_tm_vp8_sse2      (uint8_t *src, int stride);
  169.  void ff_pred8x8_top_dc_mmxext      (uint8_t *src, int stride);
  170.  void ff_pred8x8_dc_rv40_mmxext     (uint8_t *src, int stride);
  171. +void ff_pred8x8_dc_mmxext          (uint8_t *src, int stride);
  172.  void ff_pred8x8_vertical_mmx       (uint8_t *src, int stride);
  173.  void ff_pred8x8_horizontal_mmx     (uint8_t *src, int stride);
  174.  void ff_pred8x8_horizontal_mmxext  (uint8_t *src, int stride);
  175. @@ -100,6 +101,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  176.              h->pred4x4  [DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_mmxext;
  177.          if (codec_id == CODEC_ID_SVQ3 || codec_id == CODEC_ID_H264) {
  178.              h->pred8x8[TOP_DC_PRED8x8 ] = ff_pred8x8_top_dc_mmxext;
  179. +            h->pred8x8[DC_PRED8x8     ] = ff_pred8x8_dc_mmxext;
  180.          }
  181.          if (codec_id == CODEC_ID_VP8) {
  182.              h->pred16x16[PLANE_PRED8x8] = ff_pred16x16_tm_vp8_mmxext;
  183. --
  184. 1.7.2.2
  185.  
  186.  
  187. From 8196e167180f448d637a19e32fb79c9a05c2b71d Mon Sep 17 00:00:00 2001
  188. From: Daniel Kang <daniel.d.kang@gmail.com>
  189. Date: Mon, 27 Dec 2010 17:56:00 -0500
  190. Subject: [PATCH 3/5] pred8x8l_top_dc_(mmx|ssse3)
  191.  
  192. ---
  193. libavcodec/x86/h264_intrapred.asm    |   83 +++++++++++++++++++++++++++++-----
  194.  libavcodec/x86/h264_intrapred_init.c |    4 ++
  195.  2 files changed, 75 insertions(+), 12 deletions(-)
  196.  
  197. diff --git a/libavcodec/x86/h264_intrapred.asm b/libavcodec/x86/h264_intrapred.asm
  198. index 61a587d..c1b7886 100644
  199. --- a/libavcodec/x86/h264_intrapred.asm
  200. +++ b/libavcodec/x86/h264_intrapred.asm
  201. @@ -20,6 +20,7 @@
  202.  ;******************************************************************************
  203.  
  204.  %include "x86inc.asm"
  205. +%include "x86util.asm"
  206.  
  207.  SECTION_RODATA
  208.  
  209. @@ -37,6 +38,7 @@ SECTION .text
  210.  
  211.  cextern pb_1
  212.  cextern pb_3
  213. +cextern pw_4
  214.  cextern pw_5
  215.  cextern pw_16
  216.  cextern pw_17
  217. @@ -1060,6 +1062,75 @@ cglobal pred8x8_tm_vp8_ssse3, 2,3,6
  218.      jg .loop
  219.      REP_RET
  220.  
  221. +; dest, left, right, src, tmp
  222. +; output: %1 = (t[n-1] + t[n]*2 + t[n+1] + 2) >> 2
  223. +%macro PRED4x4_LOWPASS 5
  224. +    mova    %5, %2
  225. +    pavgb   %2, %3
  226. +    pxor    %3, %5
  227. +    mova    %1, %4
  228. +    pand    %3, [pb_1]
  229. +    psubusb %2, %3
  230. +    pavgb   %1, %2
  231. +%endmacro
  232. +
  233. +;-----------------------------------------------------------------------------
  234. +; void pred8x8l_top_dc(uint8_t *src, int has_topleft, int has_topright, int stride)
  235. +;-----------------------------------------------------------------------------
  236. +
  237. +%macro PRED8x8L_TOP_DC 1
  238. +cglobal pred8x8l_top_dc_%1, 4,4
  239. +    sub          r0, r3
  240. +    pxor        mm7, mm7
  241. +    movq        mm0, [r0-8]
  242. +    movq        mm3, [r0]
  243. +    movq        mm1, [r0+8]
  244. +    movq        mm2, mm3
  245. +    movq        mm4, mm3
  246. +    PALIGNR     mm2, mm0, 7, mm0
  247. +    PALIGNR     mm1, mm4, 1, mm4
  248. +    test         r1, r1 ; top_left
  249. +    jz .fix_lt_2
  250. +    test         r2, r2 ; top_right
  251. +    jz .fix_tr_1
  252. +    jmp .body
  253. +.fix_lt_2:
  254. +    movq        mm5, mm3
  255. +    pxor        mm5, mm2
  256. +    psllq       mm5, 56
  257. +    psrlq       mm5, 56
  258. +    pxor        mm2, mm5
  259. +    test         r2, r2 ; top_right
  260. +    jnz .body
  261. +.fix_tr_1:
  262. +    movq        mm5, mm3
  263. +    pxor        mm5, mm1
  264. +    psrlq       mm5, 56
  265. +    psllq       mm5, 56
  266. +    pxor        mm1, mm5
  267. +.body
  268. +    PRED4x4_LOWPASS mm0, mm2, mm1, mm3, mm5
  269. +    psadbw   mm7, mm0
  270. +    paddw    mm7, [pw_4]
  271. +    psrlw    mm7, 3
  272. +    pshufw   mm7, mm7, 0
  273. +    packuswb mm7, mm7
  274. +%rep 3
  275. +    movq [r0+r3*1], mm7
  276. +    movq [r0+r3*2], mm7
  277. +    lea    r0, [r0+r3*2]
  278. +%endrep
  279. +    movq [r0+r3*1], mm7
  280. +    movq [r0+r3*2], mm7
  281. +    RET
  282. +%endmacro
  283. +
  284. +INIT_MMX
  285. +%define PALIGNR PALIGNR_MMX
  286. +PRED8x8L_TOP_DC mmxext
  287. +%define PALIGNR PALIGNR_SSSE3
  288. +PRED8x8L_TOP_DC ssse3
  289. +
  290.  ;-----------------------------------------------------------------------------
  291.  ; void pred4x4_dc_mmxext(uint8_t *src, const uint8_t *topright, int stride)
  292.  ;-----------------------------------------------------------------------------
  293. @@ -1167,18 +1238,6 @@ cglobal pred4x4_tm_vp8_ssse3, 3,3
  294.      movd [r1+r2*2], mm5
  295.      RET
  296.  
  297. -; dest, left, right, src, tmp
  298. -; output: %1 = (t[n-1] + t[n]*2 + t[n+1] + 2) >> 2
  299. -%macro PRED4x4_LOWPASS 5
  300. -    mova    %5, %2
  301. -    pavgb   %2, %3
  302. -    pxor    %3, %5
  303. -    mova    %1, %4
  304. -    pand    %3, [pb_1]
  305. -    psubusb %2, %3
  306. -    pavgb   %1, %2
  307. -%endmacro
  308. -
  309.  ;-----------------------------------------------------------------------------
  310.  ; void pred4x4_vertical_vp8_mmxext(uint8_t *src, const uint8_t *topright, int stride)
  311.  ;-----------------------------------------------------------------------------
  312. diff --git a/libavcodec/x86/h264_intrapred_init.c b/libavcodec/x86/h264_intrapred_init.c
  313. index 8f6fd2f..1a5c6de 100644
  314. --- a/libavcodec/x86/h264_intrapred_init.c
  315. +++ b/libavcodec/x86/h264_intrapred_init.c
  316. @@ -59,6 +59,8 @@ void ff_pred8x8_tm_vp8_mmx         (uint8_t *src, int stride);
  317.  void ff_pred8x8_tm_vp8_mmxext      (uint8_t *src, int stride);
  318.  void ff_pred8x8_tm_vp8_sse2        (uint8_t *src, int stride);
  319.  void ff_pred8x8_tm_vp8_ssse3       (uint8_t *src, int stride);
  320. +void ff_pred8x8l_top_dc_mmxext     (uint8_t *src, int has_topleft, int has_topright, int stride);
  321. +void ff_pred8x8l_top_dc_ssse3      (uint8_t *src, int has_topleft, int has_topright, int stride);
  322.  void ff_pred4x4_dc_mmxext          (uint8_t *src, const uint8_t *topright, int stride);
  323.  void ff_pred4x4_down_left_mmxext   (uint8_t *src, const uint8_t *topright, int stride);
  324.  void ff_pred4x4_tm_vp8_mmx         (uint8_t *src, const uint8_t *topright, int stride);
  325. @@ -96,6 +98,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  326.          h->pred16x16[HOR_PRED8x8 ] = ff_pred16x16_horizontal_mmxext;
  327.          h->pred16x16[DC_PRED8x8  ] = ff_pred16x16_dc_mmxext;
  328.          h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_mmxext;
  329. +        h->pred8x8l[TOP_DC_PRED  ] = ff_pred8x8l_top_dc_mmxext;
  330.          h->pred4x4  [DC_PRED     ] = ff_pred4x4_dc_mmxext;
  331.          if (codec_id == CODEC_ID_VP8 || codec_id == CODEC_ID_H264)
  332.              h->pred4x4  [DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_mmxext;
  333. @@ -146,6 +149,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  334.          h->pred16x16[HOR_PRED8x8 ] = ff_pred16x16_horizontal_ssse3;
  335.          h->pred16x16[DC_PRED8x8  ] = ff_pred16x16_dc_ssse3;
  336.          h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_ssse3;
  337. +        h->pred8x8l[TOP_DC_PRED  ] = ff_pred8x8l_top_dc_ssse3;
  338.          if (codec_id == CODEC_ID_VP8) {
  339.              h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_tm_vp8_ssse3;
  340.              h->pred4x4  [TM_VP8_PRED  ] = ff_pred4x4_tm_vp8_ssse3;
  341. --
  342. 1.7.2.2
  343.  
  344.  
  345. From 26d0c78dc0149a60d128765e48660f99c5a978f0 Mon Sep 17 00:00:00 2001
  346. From: Daniel Kang <daniel.d.kang@gmail.com>
  347. Date: Mon, 27 Dec 2010 18:35:22 -0500
  348. Subject: [PATCH 4/5] pred8x8l_dc_(mmx|ssse3)
  349.  
  350. ---
  351. libavcodec/x86/h264_intrapred.asm    |  104 ++++++++++++++++++++++++++++++++++
  352.  libavcodec/x86/h264_intrapred_init.c |    4 +
  353.  2 files changed, 108 insertions(+), 0 deletions(-)
  354.  
  355. diff --git a/libavcodec/x86/h264_intrapred.asm b/libavcodec/x86/h264_intrapred.asm
  356. index c1b7886..dcc8d67 100644
  357. --- a/libavcodec/x86/h264_intrapred.asm
  358. +++ b/libavcodec/x86/h264_intrapred.asm
  359. @@ -40,6 +40,7 @@ cextern pb_1
  360.  cextern pb_3
  361.  cextern pw_4
  362.  cextern pw_5
  363. +cextern pw_8
  364.  cextern pw_16
  365.  cextern pw_17
  366.  cextern pw_32
  367. @@ -1132,6 +1133,109 @@ PRED8x8L_TOP_DC mmxext
  368.  PRED8x8L_TOP_DC ssse3
  369.  
  370.  ;-----------------------------------------------------------------------------
  371. +;void pred8x8l_dc(uint8_t *src, int has_topleft, int has_topright, int stride)
  372. +;-----------------------------------------------------------------------------
  373. +
  374. +%macro PRED8x8L_DC 1
  375. +cglobal pred8x8l_dc_%1, 4,5
  376. +    sub          r0, r3
  377. +    lea          r4, [r0+r3*2]
  378. +    movq        mm0, [r0+r3*1-8]
  379. +    punpckhbw   mm0, [r0+r3*0-8]
  380. +    movq        mm1, [r4+r3*1-8]
  381. +    punpckhbw   mm1, [r0+r3*2-8]
  382. +    mov          r4, r0
  383. +    punpckhwd   mm1, mm0
  384. +    lea          r0, [r0+r3*4]
  385. +    movq        mm2, [r0+r3*1-8]
  386. +    punpckhbw   mm2, [r0+r3*0-8]
  387. +    lea          r0, [r0+r3*2]
  388. +    movq        mm3, [r0+r3*1-8]
  389. +    punpckhbw   mm3, [r0+r3*0-8]
  390. +    punpckhwd   mm3, mm2
  391. +    punpckhdq   mm3, mm1
  392. +    lea          r0, [r0+r3*2]
  393. +    movq        mm0, [r0+r3*0-8]
  394. +    movq        mm1, [r4]
  395. +    mov          r0, r4
  396. +    movq        mm4, mm3
  397. +    movq        mm2, mm3
  398. +    PALIGNR     mm4, mm0, 7, mm0
  399. +    PALIGNR     mm1, mm2, 1, mm2
  400. +    test        r1, r1
  401. +    jnz .do_left
  402. +.fix_lt_1:
  403. +    movq        mm5, mm3
  404. +    pxor        mm5, mm4
  405. +    psrlq       mm5, 56
  406. +    psllq       mm5, 48
  407. +    pxor        mm1, mm5
  408. +    jmp .do_left
  409. +.fix_lt_2:
  410. +    movq        mm5, mm3
  411. +    pxor        mm5, mm2
  412. +    psllq       mm5, 56
  413. +    psrlq       mm5, 56
  414. +    pxor        mm2, mm5
  415. +    test         r2, r2
  416. +    jnz .body
  417. +.fix_tr_1:
  418. +    movq        mm5, mm3
  419. +    pxor        mm5, mm1
  420. +    psrlq       mm5, 56
  421. +    psllq       mm5, 56
  422. +    pxor        mm1, mm5
  423. +    jmp .body
  424. +.do_left:
  425. +    movq        mm0, mm4
  426. +    PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
  427. +    movq        mm4, mm0
  428. +    movq        mm7, mm2
  429. +    PRED4x4_LOWPASS mm1, mm3, mm0, mm4, mm5
  430. +    psllq       mm1, 56
  431. +    PALIGNR     mm7, mm1, 7, mm3
  432. +    movq        mm0, [r0-8]
  433. +    movq        mm3, [r0]
  434. +    movq        mm1, [r0+8]
  435. +    movq        mm2, mm3
  436. +    movq        mm4, mm3
  437. +    PALIGNR     mm2, mm0, 7, mm0
  438. +    PALIGNR     mm1, mm4, 1, mm4
  439. +    test         r1, r1
  440. +    jz .fix_lt_2
  441. +    test         r2, r2
  442. +    jz .fix_tr_1
  443. +.body
  444. +    lea          r1, [r0+r3*2]
  445. +    PRED4x4_LOWPASS mm6, mm2, mm1, mm3, mm5
  446. +    pxor        mm0, mm0
  447. +    pxor        mm1, mm1
  448. +    lea          r2, [r1+r3*2]
  449. +    psadbw      mm0, mm7
  450. +    psadbw      mm1, mm6
  451. +    paddw       mm0, [pw_8]
  452. +    paddw       mm0, mm1
  453. +    lea          r4, [r2+r3*2]
  454. +    psrlw       mm0, 4
  455. +    pshufw      mm0, mm0, 0
  456. +    packuswb    mm0, mm0
  457. +    movq [r0+r3*1], mm0
  458. +    movq [r0+r3*2], mm0
  459. +    movq [r1+r3*1], mm0
  460. +    movq [r1+r3*2], mm0
  461. +    movq [r2+r3*1], mm0
  462. +    movq [r2+r3*2], mm0
  463. +    movq [r4+r3*1], mm0
  464. +    movq [r4+r3*2], mm0
  465. +    RET
  466. +%endmacro
  467. +INIT_MMX
  468. +%define PALIGNR PALIGNR_MMX
  469. +PRED8x8L_DC mmxext
  470. +%define PALIGNR PALIGNR_SSSE3
  471. +PRED8x8L_DC ssse3
  472. +
  473. +;-----------------------------------------------------------------------------
  474.  ; void pred4x4_dc_mmxext(uint8_t *src, const uint8_t *topright, int stride)
  475.  ;-----------------------------------------------------------------------------
  476.  
  477. diff --git a/libavcodec/x86/h264_intrapred_init.c b/libavcodec/x86/h264_intrapred_init.c
  478. index 1a5c6de..b8c7ff1 100644
  479. --- a/libavcodec/x86/h264_intrapred_init.c
  480. +++ b/libavcodec/x86/h264_intrapred_init.c
  481. @@ -61,6 +61,8 @@ void ff_pred8x8_tm_vp8_sse2        (uint8_t *src, int stride);
  482.  void ff_pred8x8_tm_vp8_ssse3       (uint8_t *src, int stride);
  483.  void ff_pred8x8l_top_dc_mmxext     (uint8_t *src, int has_topleft, int has_topright, int stride);
  484.  void ff_pred8x8l_top_dc_ssse3      (uint8_t *src, int has_topleft, int has_topright, int stride);
  485. +void ff_pred8x8l_dc_mmxext         (uint8_t *src, int has_topleft, int has_topright, int stride);
  486. +void ff_pred8x8l_dc_ssse3          (uint8_t *src, int has_topleft, int has_topright, int stride);
  487.  void ff_pred4x4_dc_mmxext          (uint8_t *src, const uint8_t *topright, int stride);
  488.  void ff_pred4x4_down_left_mmxext   (uint8_t *src, const uint8_t *topright, int stride);
  489.  void ff_pred4x4_tm_vp8_mmx         (uint8_t *src, const uint8_t *topright, int stride);
  490. @@ -99,6 +101,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  491.          h->pred16x16[DC_PRED8x8  ] = ff_pred16x16_dc_mmxext;
  492.          h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_mmxext;
  493.          h->pred8x8l[TOP_DC_PRED  ] = ff_pred8x8l_top_dc_mmxext;
  494. +        h->pred8x8l[DC_PRED      ] = ff_pred8x8l_dc_mmxext;
  495.          h->pred4x4  [DC_PRED     ] = ff_pred4x4_dc_mmxext;
  496.          if (codec_id == CODEC_ID_VP8 || codec_id == CODEC_ID_H264)
  497.              h->pred4x4  [DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_mmxext;
  498. @@ -150,6 +153,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  499.          h->pred16x16[DC_PRED8x8  ] = ff_pred16x16_dc_ssse3;
  500.          h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_ssse3;
  501.          h->pred8x8l[TOP_DC_PRED  ] = ff_pred8x8l_top_dc_ssse3;
  502. +        h->pred8x8l[DC_PRED      ] = ff_pred8x8l_dc_ssse3;
  503.          if (codec_id == CODEC_ID_VP8) {
  504.              h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_tm_vp8_ssse3;
  505.              h->pred4x4  [TM_VP8_PRED  ] = ff_pred4x4_tm_vp8_ssse3;
  506. --
  507. 1.7.2.2
  508.  
  509.  
  510. From 9535a1017f05a108370bfd9a4bf28a0dfaf89bab Mon Sep 17 00:00:00 2001
  511. From: Daniel Kang <daniel.d.kang@gmail.com>
  512. Date: Mon, 27 Dec 2010 21:05:24 -0500
  513. Subject: [PATCH 5/5] pred8x8l_horizontal_(mmx|ssse3)
  514.  
  515. ---
  516. libavcodec/x86/h264_intrapred.asm    |   78 ++++++++++++++++++++++++++++++++++
  517.  libavcodec/x86/h264_intrapred_init.c |    4 ++
  518.  2 files changed, 82 insertions(+), 0 deletions(-)
  519.  
  520. diff --git a/libavcodec/x86/h264_intrapred.asm b/libavcodec/x86/h264_intrapred.asm
  521. index dcc8d67..ba27e0f 100644
  522. --- a/libavcodec/x86/h264_intrapred.asm
  523. +++ b/libavcodec/x86/h264_intrapred.asm
  524. @@ -1236,6 +1236,84 @@ PRED8x8L_DC mmxext
  525.  PRED8x8L_DC ssse3
  526.  
  527.  ;-----------------------------------------------------------------------------
  528. +; void pred8x8l_horizontal(uint8_t *src, int has_topleft, int has_topright, int stride)
  529. +;-----------------------------------------------------------------------------
  530. +
  531. +%macro PRED8x8L_HORIZONTAL 1
  532. +cglobal pred8x8l_horizontal_%1, 4,4
  533. +    sub          r0, r3
  534. +    lea          r2, [r0+r3*2]
  535. +    movq        mm0, [r0+r3*1-8]
  536. +    punpckhbw   mm0, [r0+r3*0-8]
  537. +    movq        mm1, [r2+r3*1-8]
  538. +    punpckhbw   mm1, [r0+r3*2-8]
  539. +    mov          r2, r0
  540. +    punpckhwd   mm1, mm0
  541. +    lea          r0, [r0+r3*4]
  542. +    movq        mm2, [r0+r3*1-8]
  543. +    punpckhbw   mm2, [r0+r3*0-8]
  544. +    lea          r0, [r0+r3*2]
  545. +    movq        mm3, [r0+r3*1-8]
  546. +    punpckhbw   mm3, [r0+r3*0-8]
  547. +    punpckhwd   mm3, mm2
  548. +    punpckhdq   mm3, mm1
  549. +    lea          r0, [r0+r3*2]
  550. +    movq        mm0, [r0+r3*0-8]
  551. +    movq        mm1, [r2]
  552. +    mov          r0, r2
  553. +    movq        mm4, mm3
  554. +    movq        mm2, mm3
  555. +    PALIGNR     mm4, mm0, 7, mm0
  556. +    PALIGNR     mm1, mm2, 1, mm2
  557. +    test        r1, r1 ; top_left
  558. +    jnz .do_left
  559. +.fix_lt_1:
  560. +    movq        mm5, mm3
  561. +    pxor        mm5, mm4
  562. +    psrlq       mm5, 56
  563. +    psllq       mm5, 48
  564. +    pxor        mm1, mm5
  565. +.do_left:
  566. +    movq        mm0, mm4
  567. +    PRED4x4_LOWPASS mm2, mm1, mm4, mm3, mm5
  568. +    movq        mm4, mm0
  569. +    movq        mm7, mm2
  570. +    PRED4x4_LOWPASS mm1, mm3, mm0, mm4, mm5
  571. +    psllq       mm1, 56
  572. +    PALIGNR     mm7, mm1, 7, mm3
  573. +    movq        mm3, mm7
  574. +    lea         r1, [r0+r3*2]
  575. +    movq       mm7, mm3
  576. +    punpckhbw  mm3, mm3
  577. +    punpcklbw  mm7, mm7
  578. +    pshufw     mm0, mm3, 0xff
  579. +    pshufw     mm1, mm3, 0xaa
  580. +    lea         r2, [r1+r3*2]
  581. +    pshufw     mm2, mm3, 0x55
  582. +    pshufw     mm3, mm3, 0x00
  583. +    pshufw     mm4, mm7, 0xff
  584. +    pshufw     mm5, mm7, 0xaa
  585. +    pshufw     mm6, mm7, 0x55
  586. +    pshufw     mm7, mm7, 0x00
  587. +    movq [r0+r3*1], mm0
  588. +    movq [r0+r3*2], mm1
  589. +    movq [r1+r3*1], mm2
  590. +    movq [r1+r3*2], mm3
  591. +    movq [r2+r3*1], mm4
  592. +    movq [r2+r3*2], mm5
  593. +    lea         r0, [r2+r3*2]
  594. +    movq [r0+r3*1], mm6
  595. +    movq [r0+r3*2], mm7
  596. +    RET
  597. +%endmacro
  598. +
  599. +INIT_MMX
  600. +%define PALIGNR PALIGNR_MMX
  601. +PRED8x8L_HORIZONTAL mmxext
  602. +%define PALIGNR PALIGNR_SSSE3
  603. +PRED8x8L_HORIZONTAL ssse3
  604. +
  605. +;-----------------------------------------------------------------------------
  606.  ; void pred4x4_dc_mmxext(uint8_t *src, const uint8_t *topright, int stride)
  607.  ;-----------------------------------------------------------------------------
  608.  
  609. diff --git a/libavcodec/x86/h264_intrapred_init.c b/libavcodec/x86/h264_intrapred_init.c
  610. index b8c7ff1..a22c5cd 100644
  611. --- a/libavcodec/x86/h264_intrapred_init.c
  612. +++ b/libavcodec/x86/h264_intrapred_init.c
  613. @@ -63,6 +63,8 @@ void ff_pred8x8l_top_dc_mmxext     (uint8_t *src, int has_topleft, int has_topri
  614.  void ff_pred8x8l_top_dc_ssse3      (uint8_t *src, int has_topleft, int has_topright, int stride);
  615.  void ff_pred8x8l_dc_mmxext         (uint8_t *src, int has_topleft, int has_topright, int stride);
  616.  void ff_pred8x8l_dc_ssse3          (uint8_t *src, int has_topleft, int has_topright, int stride);
  617. +void ff_pred8x8l_horizontal_mmxext (uint8_t *src, int has_topleft, int has_topright, int stride);
  618. +void ff_pred8x8l_horizontal_ssse3  (uint8_t *src, int has_topleft, int has_topright, int stride);
  619.  void ff_pred4x4_dc_mmxext          (uint8_t *src, const uint8_t *topright, int stride);
  620.  void ff_pred4x4_down_left_mmxext   (uint8_t *src, const uint8_t *topright, int stride);
  621.  void ff_pred4x4_tm_vp8_mmx         (uint8_t *src, const uint8_t *topright, int stride);
  622. @@ -102,6 +104,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  623.          h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_mmxext;
  624.          h->pred8x8l[TOP_DC_PRED  ] = ff_pred8x8l_top_dc_mmxext;
  625.          h->pred8x8l[DC_PRED      ] = ff_pred8x8l_dc_mmxext;
  626. +        h->pred8x8l[HOR_PRED     ] = ff_pred8x8l_horizontal_mmxext;
  627.          h->pred4x4  [DC_PRED     ] = ff_pred4x4_dc_mmxext;
  628.          if (codec_id == CODEC_ID_VP8 || codec_id == CODEC_ID_H264)
  629.              h->pred4x4  [DIAG_DOWN_LEFT_PRED ] = ff_pred4x4_down_left_mmxext;
  630. @@ -154,6 +157,7 @@ void ff_h264_pred_init_x86(H264PredContext *h, int codec_id)
  631.          h->pred8x8  [HOR_PRED8x8 ] = ff_pred8x8_horizontal_ssse3;
  632.          h->pred8x8l[TOP_DC_PRED  ] = ff_pred8x8l_top_dc_ssse3;
  633.          h->pred8x8l[DC_PRED      ] = ff_pred8x8l_dc_ssse3;
  634. +        h->pred8x8l[HOR_PRED     ] = ff_pred8x8l_horizontal_ssse3;
  635.          if (codec_id == CODEC_ID_VP8) {
  636.              h->pred8x8  [PLANE_PRED8x8] = ff_pred8x8_tm_vp8_ssse3;
  637.              h->pred4x4  [TM_VP8_PRED  ] = ff_pred4x4_tm_vp8_ssse3;
  638. --
  639. 1.7.2.2
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement
Public Pastes
Advertisement
We use cookies for various purposes including analytics. By continuing to use Pastebin, you agree to our use of cookies as described in the Cookies Policy.  OK, I Understand
Not a member of Pastebin yet?
Sign Up, it unlocks many cool features!