h264: convert loop filter strength dsp function to yasm.

This completes the conversion of h264dsp to yasm; note that h264 also
uses some dsputil functions, most notably qpel. Performance-wise, the
yasm-version is ~10 cycles faster (182->172) on x86-64, and ~8 cycles
faster (201->193) on x86-32.

2 files changed, 175 insertions, 155 deletions

diff --git a/libavcodec/x86/h264_deblock.asm b/libavcodec/x86/h264_deblock.asm
index 0891ef33da..940a8f77e1 100644
--- a/libavcodec/x86/h264_deblock.asm
+++ b/libavcodec/x86/h264_deblock.asm
@@ -27,6 +27,10 @@
 %include "x86inc.asm"
 %include "x86util.asm"
 
+SECTION_RODATA
+
+pb_3_1: times 4 db 3, 1
+
 SECTION .text
 
 cextern pb_0
@@ -911,3 +915,167 @@ ff_chroma_intra_body_mmx2:
     paddb  m1, m5
     paddb  m2, m6
     ret
+
+;-----------------------------------------------------------------------------
+; void h264_loop_filter_strength(int16_t bs[2][4][4], uint8_t nnz[40],
+;                                int8_t ref[2][40], int16_t mv[2][40][2],
+;                                int bidir,    int edges,    int step,
+;                                int mask_mv0, int mask_mv1, int field);
+;
+; bidir    is 0 or 1
+; edges    is 1 or 4
+; step     is 1 or 2
+; mask_mv0 is 0 or 3
+; mask_mv1 is 0 or 1
+; field    is 0 or 1
+;-----------------------------------------------------------------------------
+%macro loop_filter_strength_iteration 7 ; edges, step, mask_mv,
+                                        ; dir, d_idx, mask_dir, bidir
+%define edgesd    %1
+%define stepd     %2
+%define mask_mvd  %3
+%define dir       %4
+%define d_idx     %5
+%define mask_dir  %6
+%define bidir     %7
+    xor          b_idxd, b_idxd ; for (b_idx = 0; b_idx < edges; b_idx += step)
+%%.b_idx_loop:
+%if mask_dir == 0
+    pxor             m0, m0
+%endif
+    test         b_idxd, dword mask_mvd
+    jnz %%.skip_loop_iter                       ; if (!(b_idx & mask_mv))
+%if bidir == 1
+    movd             m2, [refq+b_idxq+d_idx+12] ; { ref0[bn] }
+    punpckldq        m2, [refq+b_idxq+d_idx+52] ; { ref0[bn], ref1[bn] }
+    pshufw           m0, [refq+b_idxq+12], 0x44 ; { ref0[b],  ref0[b]  }
+    pshufw           m1, [refq+b_idxq+52], 0x44 ; { ref1[b],  ref1[b]  }
+    pshufw           m3, m2, 0x4E               ; { ref1[bn], ref0[bn] }
+    psubb            m0, m2                     ; { ref0[b] != ref0[bn],
+                                                ;   ref0[b] != ref1[bn] }
+    psubb            m1, m3                     ; { ref1[b] != ref1[bn],
+                                                ;   ref1[b] != ref0[bn] }
+
+    por              m0, m1
+    mova             m1, [mvq+b_idxq*4+(d_idx+12)*4]
+    mova             m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize]
+    mova             m3, m1
+    mova             m4, m2
+    psubw            m1, [mvq+b_idxq*4+12*4]
+    psubw            m2, [mvq+b_idxq*4+12*4+mmsize]
+    psubw            m3, [mvq+b_idxq*4+52*4]
+    psubw            m4, [mvq+b_idxq*4+52*4+mmsize]
+    packsswb         m1, m2
+    packsswb         m3, m4
+    paddb            m1, m6
+    paddb            m3, m6
+    psubusb          m1, m5 ; abs(mv[b] - mv[bn]) >= limit
+    psubusb          m3, m5
+    packsswb         m1, m3
+
+    por              m0, m1
+    mova             m1, [mvq+b_idxq*4+(d_idx+52)*4]
+    mova             m2, [mvq+b_idxq*4+(d_idx+52)*4+mmsize]
+    mova             m3, m1
+    mova             m4, m2
+    psubw            m1, [mvq+b_idxq*4+12*4]
+    psubw            m2, [mvq+b_idxq*4+12*4+mmsize]
+    psubw            m3, [mvq+b_idxq*4+52*4]
+    psubw            m4, [mvq+b_idxq*4+52*4+mmsize]
+    packsswb         m1, m2
+    packsswb         m3, m4
+    paddb            m1, m6
+    paddb            m3, m6
+    psubusb          m1, m5 ; abs(mv[b] - mv[bn]) >= limit
+    psubusb          m3, m5
+    packsswb         m1, m3
+
+    pshufw           m1, m1, 0x4E
+    por              m0, m1
+    pshufw           m1, m0, 0x4E
+    pminub           m0, m1
+%else ; bidir == 0
+    movd             m0, [refq+b_idxq+12]
+    psubb            m0, [refq+b_idxq+d_idx+12] ; ref[b] != ref[bn]
+
+    mova             m1, [mvq+b_idxq*4+12*4]
+    mova             m2, [mvq+b_idxq*4+12*4+mmsize]
+    psubw            m1, [mvq+b_idxq*4+(d_idx+12)*4]
+    psubw            m2, [mvq+b_idxq*4+(d_idx+12)*4+mmsize]
+    packsswb         m1, m2
+    paddb            m1, m6
+    psubusb          m1, m5 ; abs(mv[b] - mv[bn]) >= limit
+    packsswb         m1, m1
+    por              m0, m1
+%endif ; bidir == 1/0
+
+%%.skip_loop_iter:
+    movd             m1, [nnzq+b_idxq+12]
+    por              m1, [nnzq+b_idxq+d_idx+12] ; nnz[b] || nnz[bn]
+
+    pminub           m1, m7
+    pminub           m0, m7
+    psllw            m1, 1
+    pxor             m2, m2
+    pmaxub           m1, m0
+    punpcklbw        m1, m2
+    movq [bsq+b_idxq+32*dir], m1
+
+    add          b_idxd, dword stepd
+    cmp          b_idxd, dword edgesd
+    jl %%.b_idx_loop
+%endmacro
+
+INIT_MMX mmx2
+cglobal h264_loop_filter_strength, 9, 9, 0, bs, nnz, ref, mv, bidir, edges, \
+                                            step, mask_mv0, mask_mv1, field
+%define b_idxq bidirq
+%define b_idxd bidird
+    cmp    dword fieldm, 0
+    mova             m7, [pb_1]
+    mova             m5, [pb_3]
+    je .nofield
+    mova             m5, [pb_3_1]
+.nofield:
+    mova             m6, m5
+    paddb            m5, m5
+
+    shl     dword stepd, 3
+    shl    dword edgesd, 3
+%if ARCH_X86_32
+%define mask_mv0d mask_mv0m
+%define mask_mv1d mask_mv1m
+%endif
+    shl dword mask_mv1d, 3
+    shl dword mask_mv0d, 3
+
+    cmp    dword bidird, 0
+    jne .bidir
+    loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8,  0, 0
+    loop_filter_strength_iteration     32,     8, mask_mv0d, 0, -1, -1, 0
+
+    mova             m0, [bsq+mmsize*0]
+    mova             m1, [bsq+mmsize*1]
+    mova             m2, [bsq+mmsize*2]
+    mova             m3, [bsq+mmsize*3]
+    TRANSPOSE4x4W 0, 1, 2, 3, 4
+    mova  [bsq+mmsize*0], m0
+    mova  [bsq+mmsize*1], m1
+    mova  [bsq+mmsize*2], m2
+    mova  [bsq+mmsize*3], m3
+    RET
+
+.bidir:
+    loop_filter_strength_iteration edgesd, stepd, mask_mv1d, 1, -8,  0, 1
+    loop_filter_strength_iteration     32,     8, mask_mv0d, 0, -1, -1, 1
+
+    mova             m0, [bsq+mmsize*0]
+    mova             m1, [bsq+mmsize*1]
+    mova             m2, [bsq+mmsize*2]
+    mova             m3, [bsq+mmsize*3]
+    TRANSPOSE4x4W 0, 1, 2, 3, 4
+    mova  [bsq+mmsize*0], m0
+    mova  [bsq+mmsize*1], m1
+    mova  [bsq+mmsize*2], m2
+    mova  [bsq+mmsize*3], m3
+    RET
diff --git a/libavcodec/x86/h264dsp_mmx.c b/libavcodec/x86/h264dsp_mmx.c
index 3f18f64f4b..5d9da993a6 100644
--- a/libavcodec/x86/h264dsp_mmx.c
+++ b/libavcodec/x86/h264dsp_mmx.c
@@ -88,158 +88,10 @@ void ff_h264_luma_dc_dequant_idct_sse2(DCTELEM *output, DCTELEM *input, int qmul
 /***********************************/
 /* deblocking */
 
-#define h264_loop_filter_strength_iteration_mmx2(bS, nz, ref, mv, bidir, edges, step, mask_mv, dir, d_idx, mask_dir) \
-    do { \
-        x86_reg b_idx; \
-        mask_mv <<= 3; \
-        for( b_idx=0; b_idx<edges; b_idx+=step ) { \
-            if (!mask_dir) \
-            __asm__ volatile( \
-                    "pxor %%mm0, %%mm0 \n\t" \
-                    :: \
-            ); \
-            if(!(mask_mv & b_idx)) { \
-                if(bidir) { \
-                    __asm__ volatile( \
-                        "movd         %a3(%0,%2), %%mm2 \n" \
-                        "punpckldq    %a4(%0,%2), %%mm2 \n" /* { ref0[bn], ref1[bn] } */ \
-                        "pshufw $0x44, 12(%0,%2), %%mm0 \n" /* { ref0[b], ref0[b] } */ \
-                        "pshufw $0x44, 52(%0,%2), %%mm1 \n" /* { ref1[b], ref1[b] } */ \
-                        "pshufw $0x4E, %%mm2, %%mm3 \n" \
-                        "psubb         %%mm2, %%mm0 \n" /* { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] } */ \
-                        "psubb         %%mm3, %%mm1 \n" /* { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] } */ \
- \
-                        "por           %%mm1, %%mm0 \n" \
-                        "movq   %a5(%1,%2,4), %%mm1 \n" \
-                        "movq   %a6(%1,%2,4), %%mm2 \n" \
-                        "movq          %%mm1, %%mm3 \n" \
-                        "movq          %%mm2, %%mm4 \n" \
-                        "psubw   48(%1,%2,4), %%mm1 \n" \
-                        "psubw   56(%1,%2,4), %%mm2 \n" \
-                        "psubw  208(%1,%2,4), %%mm3 \n" \
-                        "psubw  216(%1,%2,4), %%mm4 \n" \
-                        "packsswb      %%mm2, %%mm1 \n" \
-                        "packsswb      %%mm4, %%mm3 \n" \
-                        "paddb         %%mm6, %%mm1 \n" \
-                        "paddb         %%mm6, %%mm3 \n" \
-                        "psubusb       %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
-                        "psubusb       %%mm5, %%mm3 \n" \
-                        "packsswb      %%mm3, %%mm1 \n" \
- \
-                        "por           %%mm1, %%mm0 \n" \
-                        "movq   %a7(%1,%2,4), %%mm1 \n" \
-                        "movq   %a8(%1,%2,4), %%mm2 \n" \
-                        "movq          %%mm1, %%mm3 \n" \
-                        "movq          %%mm2, %%mm4 \n" \
-                        "psubw   48(%1,%2,4), %%mm1 \n" \
-                        "psubw   56(%1,%2,4), %%mm2 \n" \
-                        "psubw  208(%1,%2,4), %%mm3 \n" \
-                        "psubw  216(%1,%2,4), %%mm4 \n" \
-                        "packsswb      %%mm2, %%mm1 \n" \
-                        "packsswb      %%mm4, %%mm3 \n" \
-                        "paddb         %%mm6, %%mm1 \n" \
-                        "paddb         %%mm6, %%mm3 \n" \
-                        "psubusb       %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
-                        "psubusb       %%mm5, %%mm3 \n" \
-                        "packsswb      %%mm3, %%mm1 \n" \
- \
-                        "pshufw $0x4E, %%mm1, %%mm1 \n" \
-                        "por           %%mm1, %%mm0 \n" \
-                        "pshufw $0x4E, %%mm0, %%mm1 \n" \
-                        "pminub        %%mm1, %%mm0 \n" \
-                        ::"r"(ref), \
-                          "r"(mv), \
-                          "r"(b_idx), \
-                          "i"(d_idx+12), \
-                          "i"(d_idx+52), \
-                          "i"(d_idx*4+48), \
-                          "i"(d_idx*4+56), \
-                          "i"(d_idx*4+208), \
-                          "i"(d_idx*4+216) \
-                    ); \
-                } else { \
-                    __asm__ volatile( \
-                        "movd   12(%0,%2), %%mm0 \n" \
-                        "psubb %a3(%0,%2), %%mm0 \n" /* ref[b] != ref[bn] */ \
-                        "movq   48(%1,%2,4), %%mm1 \n" \
-                        "movq   56(%1,%2,4), %%mm2 \n" \
-                        "psubw %a4(%1,%2,4), %%mm1 \n" \
-                        "psubw %a5(%1,%2,4), %%mm2 \n" \
-                        "packsswb   %%mm2, %%mm1 \n" \
-                        "paddb      %%mm6, %%mm1 \n" \
-                        "psubusb    %%mm5, %%mm1 \n" /* abs(mv[b] - mv[bn]) >= limit */ \
-                        "packsswb   %%mm1, %%mm1 \n" \
-                        "por        %%mm1, %%mm0 \n" \
-                        ::"r"(ref), \
-                          "r"(mv), \
-                          "r"(b_idx), \
-                          "i"(d_idx+12), \
-                          "i"(d_idx*4+48), \
-                          "i"(d_idx*4+56) \
-                    ); \
-                } \
-            } \
-            __asm__ volatile( \
-                "movd 12(%0,%1), %%mm1 \n" \
-                "por %a2(%0,%1), %%mm1 \n" /* nnz[b] || nnz[bn] */ \
-                ::"r"(nnz), \
-                  "r"(b_idx), \
-                  "i"(d_idx+12) \
-            ); \
-            __asm__ volatile( \
-                "pminub    %%mm7, %%mm1 \n" \
-                "pminub    %%mm7, %%mm0 \n" \
-                "psllw        $1, %%mm1 \n" \
-                "pxor      %%mm2, %%mm2 \n" \
-                "pmaxub    %%mm0, %%mm1 \n" \
-                "punpcklbw %%mm2, %%mm1 \n" \
-                "movq      %%mm1, %a1(%0,%2) \n" \
-                ::"r"(bS), \
-                  "i"(32*dir), \
-                  "r"(b_idx) \
-                :"memory" \
-            ); \
-        } \
-    } while (0)
-
-static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
-                                            int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
-    __asm__ volatile(
-        "movq %0, %%mm7 \n"
-        "movq %1, %%mm6 \n"
-        ::"m"(ff_pb_1), "m"(ff_pb_3)
-    );
-    if(field)
-        __asm__ volatile(
-            "movq %0, %%mm6 \n"
-            ::"m"(ff_pb_3_1)
-        );
-    __asm__ volatile(
-        "movq  %%mm6, %%mm5 \n"
-        "paddb %%mm5, %%mm5 \n"
-    :);
-
-    // could do a special case for dir==0 && edges==1, but it only reduces the
-    // average filter time by 1.2%
-    step  <<= 3;
-    edges <<= 3;
-    h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir, edges, step, mask_mv1, 1, -8,  0);
-    h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir,    32,    8, mask_mv0, 0, -1, -1);
-
-    __asm__ volatile(
-        "movq   (%0), %%mm0 \n\t"
-        "movq  8(%0), %%mm1 \n\t"
-        "movq 16(%0), %%mm2 \n\t"
-        "movq 24(%0), %%mm3 \n\t"
-        TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
-        "movq %%mm0,   (%0) \n\t"
-        "movq %%mm3,  8(%0) \n\t"
-        "movq %%mm4, 16(%0) \n\t"
-        "movq %%mm2, 24(%0) \n\t"
-        ::"r"(bS[0])
-        :"memory"
-    );
-}
+void ff_h264_loop_filter_strength_mmx2(int16_t bS[2][4][4], uint8_t nnz[40],
+                                       int8_t ref[2][40], int16_t mv[2][40][2],
+                                       int bidir, int edges, int step,
+                                       int mask_mv0, int mask_mv1, int field);
 
 #define LF_FUNC(DIR, TYPE, DEPTH, OPT) \
 void ff_deblock_ ## DIR ## _ ## TYPE ## _ ## DEPTH ## _ ## OPT (uint8_t *pix, int stride, \
@@ -344,12 +196,12 @@ void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth, const int chrom
 {
     int mm_flags = av_get_cpu_flags();
 
+#if HAVE_YASM
     if (chroma_format_idc == 1 && mm_flags & AV_CPU_FLAG_MMX2) {
-        c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2;
+        c->h264_loop_filter_strength = ff_h264_loop_filter_strength_mmx2;
     }
 
     if (bit_depth == 8) {
-#if HAVE_YASM
     if (mm_flags & AV_CPU_FLAG_MMX) {
         c->h264_idct_dc_add         =
         c->h264_idct_add            = ff_h264_idct_add_8_mmx;
@@ -510,6 +362,6 @@ void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth, const int chrom
 #endif /* HAVE_AVX */
         }
     }
-#endif
     }
+#endif
 }