Update libtheora to GIT (2020.10)

1 files changed, 187 insertions, 84 deletions

diff --git a/thirdparty/libtheora/x86/mmxloop.h b/thirdparty/libtheora/x86/mmxloop.h
index 2e870c795d..1f6090b567 100644
--- a/thirdparty/libtheora/x86/mmxloop.h
+++ b/thirdparty/libtheora/x86/mmxloop.h
@@ -9,88 +9,191 @@
   On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
    mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}; mm0 and mm3 are clobbered.*/
 #define OC_LOOP_FILTER8_MMX \
- "#OC_LOOP_FILTER8_MMX\n\t" \
- /*mm7=0*/ \
- "pxor %%mm7,%%mm7\n\t" \
- /*mm6:mm0={a0,...,a7}*/ \
- "movq %%mm0,%%mm6\n\t" \
- "punpcklbw %%mm7,%%mm0\n\t" \
- "punpckhbw %%mm7,%%mm6\n\t" \
- /*mm3:mm5={d0,...,d7}*/ \
- "movq %%mm3,%%mm5\n\t" \
- "punpcklbw %%mm7,%%mm3\n\t" \
- "punpckhbw %%mm7,%%mm5\n\t" \
- /*mm6:mm0={a0-d0,...,a7-d7}*/ \
- "psubw %%mm3,%%mm0\n\t" \
- "psubw %%mm5,%%mm6\n\t" \
- /*mm3:mm1={b0,...,b7}*/ \
- "movq %%mm1,%%mm3\n\t" \
- "punpcklbw %%mm7,%%mm1\n\t" \
- "movq %%mm2,%%mm4\n\t" \
- "punpckhbw %%mm7,%%mm3\n\t" \
- /*mm5:mm4={c0,...,c7}*/ \
- "movq %%mm2,%%mm5\n\t" \
- "punpcklbw %%mm7,%%mm4\n\t" \
- "punpckhbw %%mm7,%%mm5\n\t" \
- /*mm7={3}x4 \
-   mm5:mm4={c0-b0,...,c7-b7}*/ \
- "pcmpeqw %%mm7,%%mm7\n\t" \
- "psubw %%mm1,%%mm4\n\t" \
- "psrlw $14,%%mm7\n\t" \
- "psubw %%mm3,%%mm5\n\t" \
- /*Scale by 3.*/ \
- "pmullw %%mm7,%%mm4\n\t" \
- "pmullw %%mm7,%%mm5\n\t" \
- /*mm7={4}x4 \
-   mm5:mm4=f={a0-d0+3*(c0-b0),...,a7-d7+3*(c7-b7)}*/ \
- "psrlw $1,%%mm7\n\t" \
- "paddw %%mm0,%%mm4\n\t" \
- "psllw $2,%%mm7\n\t" \
- "movq (%[ll]),%%mm0\n\t" \
- "paddw %%mm6,%%mm5\n\t" \
- /*R_i has the range [-127,128], so we compute -R_i instead. \
-   mm4=-R_i=-(f+4>>3)=0xFF^(f-4>>3)*/ \
- "psubw %%mm7,%%mm4\n\t" \
- "psubw %%mm7,%%mm5\n\t" \
- "psraw $3,%%mm4\n\t" \
- "psraw $3,%%mm5\n\t" \
- "pcmpeqb %%mm7,%%mm7\n\t" \
- "packsswb %%mm5,%%mm4\n\t" \
- "pxor %%mm6,%%mm6\n\t" \
- "pxor %%mm7,%%mm4\n\t" \
- "packuswb %%mm3,%%mm1\n\t" \
- /*Now compute lflim of -mm4 cf. Section 7.10 of the sepc.*/ \
- /*There's no unsigned byte+signed byte with unsigned saturation op code, so \
-    we have to split things by sign (the other option is to work in 16 bits, \
-    but working in 8 bits gives much better parallelism). \
-   We compute abs(R_i), but save a mask of which terms were negative in mm6. \
-   Then we compute mm4=abs(lflim(R_i,L))=min(abs(R_i),max(2*L-abs(R_i),0)). \
-   Finally, we split mm4 into positive and negative pieces using the mask in \
-    mm6, and add and subtract them as appropriate.*/ \
- /*mm4=abs(-R_i)*/ \
- /*mm7=255-2*L*/ \
- "pcmpgtb %%mm4,%%mm6\n\t" \
- "psubb %%mm0,%%mm7\n\t" \
- "pxor %%mm6,%%mm4\n\t" \
- "psubb %%mm0,%%mm7\n\t" \
- "psubb %%mm6,%%mm4\n\t" \
- /*mm7=255-max(2*L-abs(R_i),0)*/ \
- "paddusb %%mm4,%%mm7\n\t" \
- /*mm4=min(abs(R_i),max(2*L-abs(R_i),0))*/ \
- "paddusb %%mm7,%%mm4\n\t" \
- "psubusb %%mm7,%%mm4\n\t" \
- /*Now split mm4 by the original sign of -R_i.*/ \
- "movq %%mm4,%%mm5\n\t" \
- "pand %%mm6,%%mm4\n\t" \
- "pandn %%mm5,%%mm6\n\t" \
- /*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
- /*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
- "paddusb %%mm4,%%mm1\n\t" \
- "psubusb %%mm4,%%mm2\n\t" \
- "psubusb %%mm6,%%mm1\n\t" \
- "paddusb %%mm6,%%mm2\n\t" \
+  "#OC_LOOP_FILTER8_MMX\n\t" \
+  /*mm7=0*/ \
+  "pxor %%mm7,%%mm7\n\t" \
+  /*mm6:mm0={a0,...,a7}*/ \
+  "movq %%mm0,%%mm6\n\t" \
+  "punpcklbw %%mm7,%%mm0\n\t" \
+  "punpckhbw %%mm7,%%mm6\n\t" \
+  /*mm3:mm5={d0,...,d7}*/ \
+  "movq %%mm3,%%mm5\n\t" \
+  "punpcklbw %%mm7,%%mm3\n\t" \
+  "punpckhbw %%mm7,%%mm5\n\t" \
+  /*mm6:mm0={a0-d0,...,a7-d7}*/ \
+  "psubw %%mm3,%%mm0\n\t" \
+  "psubw %%mm5,%%mm6\n\t" \
+  /*mm3:mm1={b0,...,b7}*/ \
+  "movq %%mm1,%%mm3\n\t" \
+  "punpcklbw %%mm7,%%mm1\n\t" \
+  "movq %%mm2,%%mm4\n\t" \
+  "punpckhbw %%mm7,%%mm3\n\t" \
+  /*mm5:mm4={c0,...,c7}*/ \
+  "movq %%mm2,%%mm5\n\t" \
+  "punpcklbw %%mm7,%%mm4\n\t" \
+  "punpckhbw %%mm7,%%mm5\n\t" \
+  /*mm7={3}x4 \
+    mm5:mm4={c0-b0,...,c7-b7}*/ \
+  "pcmpeqw %%mm7,%%mm7\n\t" \
+  "psubw %%mm1,%%mm4\n\t" \
+  "psrlw $14,%%mm7\n\t" \
+  "psubw %%mm3,%%mm5\n\t" \
+  /*Scale by 3.*/ \
+  "pmullw %%mm7,%%mm4\n\t" \
+  "pmullw %%mm7,%%mm5\n\t" \
+  /*mm7={4}x4 \
+    mm5:mm4=f={a0-d0+3*(c0-b0),...,a7-d7+3*(c7-b7)}*/ \
+  "psrlw $1,%%mm7\n\t" \
+  "paddw %%mm0,%%mm4\n\t" \
+  "psllw $2,%%mm7\n\t" \
+  "movq (%[ll]),%%mm0\n\t" \
+  "paddw %%mm6,%%mm5\n\t" \
+  /*R_i has the range [-127,128], so we compute -R_i instead. \
+    mm4=-R_i=-(f+4>>3)=0xFF^(f-4>>3)*/ \
+  "psubw %%mm7,%%mm4\n\t" \
+  "psubw %%mm7,%%mm5\n\t" \
+  "psraw $3,%%mm4\n\t" \
+  "psraw $3,%%mm5\n\t" \
+  "pcmpeqb %%mm7,%%mm7\n\t" \
+  "packsswb %%mm5,%%mm4\n\t" \
+  "pxor %%mm6,%%mm6\n\t" \
+  "pxor %%mm7,%%mm4\n\t" \
+  "packuswb %%mm3,%%mm1\n\t" \
+  /*Now compute lflim of -mm4 cf. Section 7.10 of the sepc.*/ \
+  /*There's no unsigned byte+signed byte with unsigned saturation op code, so \
+     we have to split things by sign (the other option is to work in 16 bits, \
+     but working in 8 bits gives much better parallelism). \
+    We compute abs(R_i), but save a mask of which terms were negative in mm6. \
+    Then we compute mm4=abs(lflim(R_i,L))=min(abs(R_i),max(2*L-abs(R_i),0)). \
+    Finally, we split mm4 into positive and negative pieces using the mask in \
+     mm6, and add and subtract them as appropriate.*/ \
+  /*mm4=abs(-R_i)*/ \
+  /*mm7=255-2*L*/ \
+  "pcmpgtb %%mm4,%%mm6\n\t" \
+  "psubb %%mm0,%%mm7\n\t" \
+  "pxor %%mm6,%%mm4\n\t" \
+  "psubb %%mm0,%%mm7\n\t" \
+  "psubb %%mm6,%%mm4\n\t" \
+  /*mm7=255-max(2*L-abs(R_i),0)*/ \
+  "paddusb %%mm4,%%mm7\n\t" \
+  /*mm4=min(abs(R_i),max(2*L-abs(R_i),0))*/ \
+  "paddusb %%mm7,%%mm4\n\t" \
+  "psubusb %%mm7,%%mm4\n\t" \
+  /*Now split mm4 by the original sign of -R_i.*/ \
+  "movq %%mm4,%%mm5\n\t" \
+  "pand %%mm6,%%mm4\n\t" \
+  "pandn %%mm5,%%mm6\n\t" \
+  /*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
+  /*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
+  "paddusb %%mm4,%%mm1\n\t" \
+  "psubusb %%mm4,%%mm2\n\t" \
+  "psubusb %%mm6,%%mm1\n\t" \
+  "paddusb %%mm6,%%mm2\n\t" \
 
-#define OC_LOOP_FILTER_V_MMX(_pix,_ystride,_ll) \
+/*On entry, mm0={a0,...,a7}, mm1={b0,...,b7}, mm2={c0,...,c7}, mm3={d0,...d7}.
+  On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
+   mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}.
+  All other MMX registers are clobbered.*/
+#define OC_LOOP_FILTER8_MMXEXT \
+  "#OC_LOOP_FILTER8_MMXEXT\n\t" \
+  /*R_i=(a_i-3*b_i+3*c_i-d_i+4>>3) has the range [-127,128], so we compute \
+     -R_i=(-a_i+3*b_i-3*c_i+d_i+3>>3) instead.*/ \
+  /*This first part is based on the transformation \
+      f = -(3*(c-b)+a-d+4>>3) \
+        = -(3*(c+255-b)+(a+255-d)+4-1020>>3) \
+        = -(3*(c+~b)+(a+~d)-1016>>3) \
+        = 127-(3*(c+~b)+(a+~d)>>3) \
+        = 128+~(3*(c+~b)+(a+~d)>>3) (mod 256). \
+    Although pavgb(a,b) = (a+b+1>>1) (biased up), we rely heavily on the \
+     fact that ~pavgb(~a,~b) = (a+b>>1) (biased down). \
+    Using this, the last expression above can be computed in 8 bits of working \
+     precision via: \
+      u = ~pavgb(~b,c); \
+      v = pavgb(b,~c); \
+      This mask is 0 or 0xFF, and controls whether t is biased up or down: \
+      m = u-v; \
+      t = m^pavgb(m^~a,m^d); \
+      f = 128+pavgb(pavgb(t,u),v); \
+    This required some careful analysis to ensure that carries are propagated \
+     correctly in all cases, but has been checked exhaustively.*/ \
+  /*input (a, b, c, d, ., ., ., .)*/ \
+  /*ff=0xFF; \
+    u=b; \
+    v=c; \
+    ll=255-2*L;*/ \
+  "pcmpeqb %%mm7,%%mm7\n\t" \
+  "movq %%mm1,%%mm4\n\t" \
+  "movq %%mm2,%%mm5\n\t" \
+  "movq (%[ll]),%%mm6\n\t" \
+  /*allocated u, v, ll, ff: (a, b, c, d, u, v, ll, ff)*/ \
+  /*u^=ff; \
+    v^=ff;*/ \
+  "pxor %%mm7,%%mm4\n\t" \
+  "pxor %%mm7,%%mm5\n\t" \
+  /*allocated ll: (a, b, c, d, u, v, ll, ff)*/ \
+  /*u=pavgb(u,c); \
+    v=pavgb(v,b);*/ \
+  "pavgb %%mm2,%%mm4\n\t" \
+  "pavgb %%mm1,%%mm5\n\t" \
+  /*u^=ff; \
+    a^=ff;*/ \
+  "pxor %%mm7,%%mm4\n\t" \
+  "pxor %%mm7,%%mm0\n\t" \
+  /*m=u-v;*/ \
+  "psubb %%mm5,%%mm4\n\t" \
+  /*freed u, allocated m: (a, b, c, d, m, v, ll, ff)*/ \
+  /*a^=m; \
+    d^=m;*/ \
+  "pxor %%mm4,%%mm0\n\t" \
+  "pxor %%mm4,%%mm3\n\t" \
+  /*t=pavgb(a,d);*/ \
+  "pavgb %%mm3,%%mm0\n\t" \
+  "psllw $7,%%mm7\n\t" \
+  /*freed a, d, ff, allocated t, of: (t, b, c, ., m, v, ll, of)*/ \
+  /*t^=m; \
+    u=m+v;*/ \
+  "pxor %%mm4,%%mm0\n\t" \
+  "paddb %%mm5,%%mm4\n\t" \
+  /*freed t, m, allocated f, u: (f, b, c, ., u, v, ll, of)*/ \
+  /*f=pavgb(f,u); \
+    of=128;*/ \
+  "pavgb %%mm4,%%mm0\n\t" \
+  "packsswb %%mm7,%%mm7\n\t" \
+  /*freed u, ff, allocated ll: (f, b, c, ., ll, v, ll, of)*/ \
+  /*f=pavgb(f,v);*/ \
+  "pavgb %%mm5,%%mm0\n\t" \
+  "movq %%mm7,%%mm3\n\t" \
+  "movq %%mm6,%%mm4\n\t" \
+  /*freed v, allocated of: (f, b, c, of, ll, ., ll, of)*/ \
+  /*Now compute lflim of R_i=-(128+mm0) cf. Section 7.10 of the sepc.*/ \
+  /*There's no unsigned byte+signed byte with unsigned saturation op code, so \
+     we have to split things by sign (the other option is to work in 16 bits, \
+     but staying in 8 bits gives much better parallelism).*/ \
+  /*Instead of adding the offset of 128 in mm3, we use it to split mm0. \
+    This is the same number of instructions as computing a mask and splitting \
+     after the lflim computation, but has shorter dependency chains.*/ \
+  /*mm0=R_i<0?-R_i:0 (denoted abs(R_i<0))\
+    mm3=R_i>0?R_i:0* (denoted abs(R_i>0))*/ \
+  "psubusb %%mm0,%%mm3\n\t" \
+  "psubusb %%mm7,%%mm0\n\t" \
+  /*mm6=255-max(2*L-abs(R_i<0),0) \
+    mm4=255-max(2*L-abs(R_i>0),0)*/ \
+  "paddusb %%mm3,%%mm4\n\t" \
+  "paddusb %%mm0,%%mm6\n\t" \
+  /*mm0=min(abs(R_i<0),max(2*L-abs(R_i<0),0)) \
+    mm3=min(abs(R_i>0),max(2*L-abs(R_i>0),0))*/ \
+  "paddusb %%mm4,%%mm3\n\t" \
+  "paddusb %%mm6,%%mm0\n\t" \
+  "psubusb %%mm4,%%mm3\n\t" \
+  "psubusb %%mm6,%%mm0\n\t" \
+  /*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
+  /*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
+  "paddusb %%mm3,%%mm1\n\t" \
+  "psubusb %%mm3,%%mm2\n\t" \
+  "psubusb %%mm0,%%mm1\n\t" \
+  "paddusb %%mm0,%%mm2\n\t" \
+
+#define OC_LOOP_FILTER_V(_filter,_pix,_ystride,_ll) \
   do{ \
     ptrdiff_t ystride3__; \
     __asm__ __volatile__( \
@@ -104,7 +207,7 @@
       "movq (%[pix],%[ystride]),%%mm1\n\t" \
       /*mm2={c0,...,c7}*/ \
       "movq (%[pix],%[ystride],2),%%mm2\n\t" \
-      OC_LOOP_FILTER8_MMX \
+      _filter \
       /*Write it back out.*/ \
       "movq %%mm1,(%[pix],%[ystride])\n\t" \
       "movq %%mm2,(%[pix],%[ystride],2)\n\t" \
@@ -116,7 +219,7 @@
   } \
   while(0)
 
-#define OC_LOOP_FILTER_H_MMX(_pix,_ystride,_ll) \
+#define OC_LOOP_FILTER_H(_filter,_pix,_ystride,_ll) \
   do{ \
     unsigned char *pix__; \
     ptrdiff_t      ystride3__; \
@@ -174,7 +277,7 @@
       "punpckldq %%mm5,%%mm2\n\t" \
       /*mm3=d7 d6 d5 d4 d3 d2 d1 d0*/ \
       "punpckhdq %%mm5,%%mm3\n\t" \
-      OC_LOOP_FILTER8_MMX \
+      _filter \
       /*mm2={b0+R_0'',...,b7+R_7''}*/ \
       "movq %%mm1,%%mm0\n\t" \
       /*mm1={b0+R_0'',c0-R_0'',...,b3+R_3'',c3-R_3''}*/ \