Add vertical and horizontal luma deblocking accelerated with Altivec,
authorGuillaume Poirier <gpoirier@mplayerhq.hu>
Fri, 22 Jun 2007 21:42:41 +0000 (21:42 +0000)
committerGuillaume Poirier <gpoirier@mplayerhq.hu>
Fri, 22 Jun 2007 21:42:41 +0000 (21:42 +0000)
based on Graham Booker's code written for FFmpeg with slight modifications
to re-use x264's macros

git-svn-id: svn://svn.videolan.org/x264/trunk@662 df754926-b1dd-0310-bc7b-ec298dee348c

Makefile
common/frame.c
common/ppc/deblock.c [new file with mode: 0644]

index 4400cd5..787d96d 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -48,7 +48,7 @@ endif
 # AltiVec optims
 ifeq ($(ARCH),PPC)
 ALTIVECSRC += common/ppc/mc.c common/ppc/pixel.c common/ppc/dct.c \
-              common/ppc/quant.c
+              common/ppc/quant.c common/ppc/deblock.c
 SRCS += $(ALTIVECSRC)
 $(ALTIVECSRC:%.c=%.o): CFLAGS += $(ALTIVECFLAGS)
 endif
index a852c91..99529d0 100644 (file)
@@ -763,6 +763,11 @@ void x264_deblock_v_luma_mmxext( uint8_t *pix, int stride, int alpha, int beta,
 #endif
 #endif
 
+#ifdef ARCH_PPC
+void x264_deblock_v_luma_altivec( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 );
+void x264_deblock_h_luma_altivec( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 );
+#endif // ARCH_PPC
+
 void x264_deblock_init( int cpu, x264_deblock_function_t *pf )
 {
     pf->deblock_v_luma = deblock_v_luma_c;
@@ -794,6 +799,14 @@ void x264_deblock_init( int cpu, x264_deblock_function_t *pf )
 #endif
     }
 #endif
+
+#ifdef ARCH_PPC
+    if( cpu&X264_CPU_ALTIVEC )
+    {
+        pf->deblock_v_luma = x264_deblock_v_luma_altivec;
+        pf->deblock_h_luma = x264_deblock_h_luma_altivec;
+   }
+#endif // ARCH_PPC
 }
 
 
diff --git a/common/ppc/deblock.c b/common/ppc/deblock.c
new file mode 100644 (file)
index 0000000..127a0ee
--- /dev/null
@@ -0,0 +1,304 @@
+/*****************************************************************************
+* deblock.c: Altivec-accelerated deblocking for h264 encoder
+*****************************************************************************
+* Copyright (C) 2007 x264 project
+*
+* Authors: Guillaume Poirier <gpoirier CHEZ mplayerhq POINT hu>
+*          (based on code written by Graham Booker for FFmpeg)
+*
+* This program is free software; you can redistribute it and/or modify
+* it under the terms of the GNU General Public License as published by
+* the Free Software Foundation; either version 2 of the License, or
+* (at your option) any later version.
+*
+* This program is distributed in the hope that it will be useful,
+* but WITHOUT ANY WARRANTY; without even the implied warranty of
+* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+* GNU General Public License for more details.
+*
+* You should have received a copy of the GNU General Public License
+* along with this program; if not, write to the Free Software
+* Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
+*****************************************************************************/
+
+#if defined SYS_LINUX
+#include <altivec.h>
+#endif
+
+#include "common/common.h"
+#include "ppccommon.h"
+
+#define transpose4x16(r0, r1, r2, r3) {      \
+    register vec_u8_t r4;                    \
+    register vec_u8_t r5;                    \
+    register vec_u8_t r6;                    \
+    register vec_u8_t r7;                    \
+                                             \
+    r4 = vec_mergeh(r0, r2);  /*0, 2 set 0*/ \
+    r5 = vec_mergel(r0, r2);  /*0, 2 set 1*/ \
+    r6 = vec_mergeh(r1, r3);  /*1, 3 set 0*/ \
+    r7 = vec_mergel(r1, r3);  /*1, 3 set 1*/ \
+                                             \
+    r0 = vec_mergeh(r4, r6);  /*all set 0*/  \
+    r1 = vec_mergel(r4, r6);  /*all set 1*/  \
+    r2 = vec_mergeh(r5, r7);  /*all set 2*/  \
+    r3 = vec_mergel(r5, r7);  /*all set 3*/  \
+}
+
+static inline void write16x4(uint8_t *dst, int dst_stride,
+                             register vec_u8_t r0, register vec_u8_t r1,
+                             register vec_u8_t r2, register vec_u8_t r3) {
+    DECLARE_ALIGNED(unsigned char, result[64], 16);
+    uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
+    int int_dst_stride = dst_stride/4;
+
+    vec_st(r0, 0, result);
+    vec_st(r1, 16, result);
+    vec_st(r2, 32, result);
+    vec_st(r3, 48, result);
+    /* FIXME: there has to be a better way!!!! */
+    *dst_int = *src_int;
+    *(dst_int+   int_dst_stride) = *(src_int + 1);
+    *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
+    *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
+    *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
+    *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
+    *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
+    *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
+    *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
+    *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
+    *(dst_int+10*int_dst_stride) = *(src_int + 10);
+    *(dst_int+11*int_dst_stride) = *(src_int + 11);
+    *(dst_int+12*int_dst_stride) = *(src_int + 12);
+    *(dst_int+13*int_dst_stride) = *(src_int + 13);
+    *(dst_int+14*int_dst_stride) = *(src_int + 14);
+    *(dst_int+15*int_dst_stride) = *(src_int + 15);
+}
+
+/** \brief performs a 6x16 transpose of data in src, and stores it to dst
+    \todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
+    out of unaligned_load() */
+#define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
+    register vec_u8_t r0, r1, r2, r3, r4, r5, r6, r7, r14, r15;\
+    VEC_LOAD(src,                  r0, 16, vec_u8_t);          \
+    VEC_LOAD(src +    src_stride,  r1, 16, vec_u8_t);          \
+    VEC_LOAD(src +  2*src_stride,  r2, 16, vec_u8_t);          \
+    VEC_LOAD(src +  3*src_stride,  r3, 16, vec_u8_t);          \
+    VEC_LOAD(src +  4*src_stride,  r4, 16, vec_u8_t);          \
+    VEC_LOAD(src +  5*src_stride,  r5, 16, vec_u8_t);          \
+    VEC_LOAD(src +  6*src_stride,  r6, 16, vec_u8_t);          \
+    VEC_LOAD(src +  7*src_stride,  r7, 16, vec_u8_t);          \
+    VEC_LOAD(src + 14*src_stride, r14, 16, vec_u8_t);          \
+    VEC_LOAD(src + 15*src_stride, r15, 16, vec_u8_t);          \
+                                                               \
+    VEC_LOAD(src + 8*src_stride,   r8, 16, vec_u8_t);          \
+    VEC_LOAD(src + 9*src_stride,   r9, 16, vec_u8_t);          \
+    VEC_LOAD(src + 10*src_stride, r10, 16, vec_u8_t);          \
+    VEC_LOAD(src + 11*src_stride, r11, 16, vec_u8_t);          \
+    VEC_LOAD(src + 12*src_stride, r12, 16, vec_u8_t);          \
+    VEC_LOAD(src + 13*src_stride, r13, 16, vec_u8_t);          \
+                                                               \
+    /*Merge first pairs*/                                      \
+    r0 = vec_mergeh(r0, r8);    /*0, 8*/                       \
+    r1 = vec_mergeh(r1, r9);    /*1, 9*/                       \
+    r2 = vec_mergeh(r2, r10);   /*2,10*/                       \
+    r3 = vec_mergeh(r3, r11);   /*3,11*/                       \
+    r4 = vec_mergeh(r4, r12);   /*4,12*/                       \
+    r5 = vec_mergeh(r5, r13);   /*5,13*/                       \
+    r6 = vec_mergeh(r6, r14);   /*6,14*/                       \
+    r7 = vec_mergeh(r7, r15);   /*7,15*/                       \
+                                                               \
+    /*Merge second pairs*/                                     \
+    r8  = vec_mergeh(r0, r4);   /*0,4, 8,12 set 0*/            \
+    r9  = vec_mergel(r0, r4);   /*0,4, 8,12 set 1*/            \
+    r10 = vec_mergeh(r1, r5);   /*1,5, 9,13 set 0*/            \
+    r11 = vec_mergel(r1, r5);   /*1,5, 9,13 set 1*/            \
+    r12 = vec_mergeh(r2, r6);   /*2,6,10,14 set 0*/            \
+    r13 = vec_mergel(r2, r6);   /*2,6,10,14 set 1*/            \
+    r14 = vec_mergeh(r3, r7);   /*3,7,11,15 set 0*/            \
+    r15 = vec_mergel(r3, r7);   /*3,7,11,15 set 1*/            \
+                                                               \
+    /*Third merge*/                                            \
+    r0 = vec_mergeh(r8, r12);   /*0,2,4,6,8,10,12,14 set 0*/   \
+    r1 = vec_mergel(r8, r12);   /*0,2,4,6,8,10,12,14 set 1*/   \
+    r2 = vec_mergeh(r9, r13);   /*0,2,4,6,8,10,12,14 set 2*/   \
+    r4 = vec_mergeh(r10, r14);  /*1,3,5,7,9,11,13,15 set 0*/   \
+    r5 = vec_mergel(r10, r14);  /*1,3,5,7,9,11,13,15 set 1*/   \
+    r6 = vec_mergeh(r11, r15);  /*1,3,5,7,9,11,13,15 set 2*/   \
+    /* Don't need to compute 3 and 7*/                         \
+                                                               \
+    /*Final merge*/                                            \
+    r8  = vec_mergeh(r0, r4);   /*all set 0*/                  \
+    r9  = vec_mergel(r0, r4);   /*all set 1*/                  \
+    r10 = vec_mergeh(r1, r5);   /*all set 2*/                  \
+    r11 = vec_mergel(r1, r5);   /*all set 3*/                  \
+    r12 = vec_mergeh(r2, r6);   /*all set 4*/                  \
+    r13 = vec_mergel(r2, r6);   /*all set 5*/                  \
+    /* Don't need to compute 14 and 15*/                       \
+                                                               \
+}
+
+// out: o = |x-y| < a
+static inline vec_u8_t diff_lt_altivec ( register vec_u8_t x,
+                                         register vec_u8_t y,
+                                         register vec_u8_t a) {
+
+    register vec_u8_t diff = vec_subs(x, y);
+    register vec_u8_t diffneg = vec_subs(y, x);
+    register vec_u8_t o = vec_or(diff, diffneg); /* |x-y| */
+    o = (vec_u8_t)vec_cmplt(o, a);
+    return o;
+}
+
+static inline vec_u8_t h264_deblock_mask ( register vec_u8_t p0,
+                                           register vec_u8_t p1,
+                                           register vec_u8_t q0,
+                                           register vec_u8_t q1,
+                                           register vec_u8_t alpha,
+                                           register vec_u8_t beta) {
+
+    register vec_u8_t mask;
+    register vec_u8_t tempmask;
+
+    mask = diff_lt_altivec(p0, q0, alpha);
+    tempmask = diff_lt_altivec(p1, p0, beta);
+    mask = vec_and(mask, tempmask);
+    tempmask = diff_lt_altivec(q1, q0, beta);
+    mask = vec_and(mask, tempmask);
+
+    return mask;
+}
+
+// out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
+static inline vec_u8_t h264_deblock_q1(register vec_u8_t p0,
+                                       register vec_u8_t p1,
+                                       register vec_u8_t p2,
+                                       register vec_u8_t q0,
+                                       register vec_u8_t tc0) {
+
+    register vec_u8_t average = vec_avg(p0, q0);
+    register vec_u8_t temp;
+    register vec_u8_t uncliped;
+    register vec_u8_t ones;
+    register vec_u8_t max;
+    register vec_u8_t min;
+    register vec_u8_t newp1;
+
+    temp = vec_xor(average, p2);
+    average = vec_avg(average, p2);     /*avg(p2, avg(p0, q0)) */
+    ones = vec_splat_u8(1);
+    temp = vec_and(temp, ones);         /*(p2^avg(p0, q0)) & 1 */
+    uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
+    max = vec_adds(p1, tc0);
+    min = vec_subs(p1, tc0);
+    newp1 = vec_max(min, uncliped);
+    newp1 = vec_min(max, newp1);
+    return newp1;
+}
+
+#define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) {                                         \
+                                                                                                \
+    const vec_u8_t A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4));                             \
+                                                                                                \
+    register vec_u8_t pq0bit = vec_xor(p0,q0);                                                  \
+    register vec_u8_t q1minus;                                                                  \
+    register vec_u8_t p0minus;                                                                  \
+    register vec_u8_t stage1;                                                                   \
+    register vec_u8_t stage2;                                                                   \
+    register vec_u8_t vec160;                                                                   \
+    register vec_u8_t delta;                                                                    \
+    register vec_u8_t deltaneg;                                                                 \
+                                                                                                \
+    q1minus = vec_nor(q1, q1);                /* 255 - q1 */                                    \
+    stage1 = vec_avg(p1, q1minus);            /* (p1 - q1 + 256)>>1 */                          \
+    stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */    \
+    p0minus = vec_nor(p0, p0);                /* 255 - p0 */                                    \
+    stage1 = vec_avg(q0, p0minus);            /* (q0 - p0 + 256)>>1 */                          \
+    pq0bit = vec_and(pq0bit, vec_splat_u8(1));                                                  \
+    stage2 = vec_avg(stage2, pq0bit);         /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */\
+    stage2 = vec_adds(stage2, stage1);        /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \
+    vec160 = vec_ld(0, &A0v);                                                                   \
+    deltaneg = vec_subs(vec160, stage2);      /* -d */                                          \
+    delta = vec_subs(stage2, vec160);         /*  d */                                          \
+    deltaneg = vec_min(tc0masked, deltaneg);                                                    \
+    delta = vec_min(tc0masked, delta);                                                          \
+    p0 = vec_subs(p0, deltaneg);                                                                \
+    q0 = vec_subs(q0, delta);                                                                   \
+    p0 = vec_adds(p0, delta);                                                                   \
+    q0 = vec_adds(q0, deltaneg);                                                                \
+}
+
+#define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) {            \
+    DECLARE_ALIGNED(unsigned char, temp[16], 16);                                            \
+    register vec_u8_t alphavec;                                                              \
+    register vec_u8_t betavec;                                                               \
+    register vec_u8_t mask;                                                                  \
+    register vec_u8_t p1mask;                                                                \
+    register vec_u8_t q1mask;                                                                \
+    register vec_s8_t tc0vec;                                                                \
+    register vec_u8_t finaltc0;                                                              \
+    register vec_u8_t tc0masked;                                                             \
+    register vec_u8_t newp1;                                                                 \
+    register vec_u8_t newq1;                                                                 \
+                                                                                             \
+    temp[0] = alpha;                                                                         \
+    temp[1] = beta;                                                                          \
+    alphavec = vec_ld(0, temp);                                                              \
+    betavec = vec_splat(alphavec, 0x1);                                                      \
+    alphavec = vec_splat(alphavec, 0x0);                                                     \
+    mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */            \
+                                                                                             \
+    *((int *)temp) = *((int *)tc0);                                                          \
+    tc0vec = vec_ld(0, (signed char*)temp);                                                  \
+    tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
+    tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
+    mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1)));  /* if tc0[i] >= 0 */         \
+    finaltc0 = vec_and((vec_u8_t)tc0vec, mask);                 /* tc = tc0 */               \
+                                                                                             \
+    p1mask = diff_lt_altivec(p2, p0, betavec);                                               \
+    p1mask = vec_and(p1mask, mask);                             /* if( |p2 - p0| < beta) */  \
+    tc0masked = vec_and(p1mask, (vec_u8_t)tc0vec);                                           \
+    finaltc0 = vec_sub(finaltc0, p1mask);                       /* tc++ */                   \
+    newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked);                                      \
+    /*end if*/                                                                               \
+                                                                                             \
+    q1mask = diff_lt_altivec(q2, q0, betavec);                                               \
+    q1mask = vec_and(q1mask, mask);                             /* if ( |q2 - q0| < beta ) */\
+    tc0masked = vec_and(q1mask, (vec_u8_t)tc0vec);                                           \
+    finaltc0 = vec_sub(finaltc0, q1mask);                       /* tc++ */                   \
+    newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked);                                      \
+    /*end if*/                                                                               \
+                                                                                             \
+    h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0);                                            \
+    p1 = newp1;                                                                              \
+    q1 = newq1;                                                                              \
+}
+
+void x264_deblock_v_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
+
+    if((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
+        register vec_u8_t p2 = vec_ld(-3*stride, pix);
+        register vec_u8_t p1 = vec_ld(-2*stride, pix);
+        register vec_u8_t p0 = vec_ld(-1*stride, pix);
+        register vec_u8_t q0 = vec_ld(0, pix);
+        register vec_u8_t q1 = vec_ld(stride, pix);
+        register vec_u8_t q2 = vec_ld(2*stride, pix);
+        h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
+        vec_st(p1, -2*stride, pix);
+        vec_st(p0, -1*stride, pix);
+        vec_st(q0, 0, pix);
+        vec_st(q1, stride, pix);
+    }
+}
+
+void x264_deblock_h_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
+
+    register vec_u8_t line0, line1, line2, line3, line4, line5;
+    if((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
+        return;
+    PREP_LOAD;
+    readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
+    h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
+    transpose4x16(line1, line2, line3, line4);
+    write16x4(pix-2, stride, line1, line2, line3, line4);
+}