x86: dsputilenc: convert ff_sse{8, 16}_mmx() to yasm

Signed-off-by: Timothy Gu <timothygu99@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

2 files changed, 48 insertions, 145 deletions

diff --git a/libavcodec/x86/dsputilenc.asm b/libavcodec/x86/dsputilenc.asm
index de36b95a30..0628550145 100644
--- a/libavcodec/x86/dsputilenc.asm
+++ b/libavcodec/x86/dsputilenc.asm
@@ -278,19 +278,27 @@ INIT_XMM ssse3
 %define ABS_SUM_8x8 ABS_SUM_8x8_64
 HADAMARD8_DIFF 9
 
-INIT_XMM sse2
-; int ff_sse16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
-;                   int line_size, int h);
-cglobal sse16, 5, 5, 8
-    shr      r4d, 1
+; int ff_sse*_*(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+;               int line_size, int h)
+
+%macro SUM_SQUARED_ERRORS 1
+cglobal sse%1, 5,5,8, v, pix1, pix2, lsize, h
+%if %1 == mmsize
+    shr       hd, 1
+%endif
     pxor      m0, m0         ; mm0 = 0
     pxor      m7, m7         ; mm7 holds the sum
 
 .next2lines: ; FIXME why are these unaligned movs? pix1[] is aligned
-    movu      m1, [r1   ]    ; mm1 = pix1[0][0-15]
-    movu      m2, [r2   ]    ; mm2 = pix2[0][0-15]
-    movu      m3, [r1+r3]    ; mm3 = pix1[1][0-15]
-    movu      m4, [r2+r3]    ; mm4 = pix2[1][0-15]
+    movu      m1, [pix1q]    ; m1 = pix1[0][0-15], [0-7] for mmx
+    movu      m2, [pix2q]    ; m2 = pix2[0][0-15], [0-7] for mmx
+%if %1 == mmsize
+    movu      m3, [pix1q+lsizeq] ; m3 = pix1[1][0-15], [0-7] for mmx
+    movu      m4, [pix2q+lsizeq] ; m4 = pix2[1][0-15], [0-7] for mmx
+%else  ; %1 / 2 == mmsize; mmx only
+    mova      m3, [pix1q+8]  ; m3 = pix1[0][8-15]
+    mova      m4, [pix2q+8]  ; m4 = pix2[0][8-15]
+%endif
 
     ; todo: mm1-mm2, mm3-mm4
     ; algo: subtract mm1 from mm2 with saturation and vice versa
@@ -319,20 +327,34 @@ cglobal sse16, 5, 5, 8
     pmaddwd   m1, m1
     pmaddwd   m3, m3
 
-    lea       r1, [r1+r3*2]  ; pix1 += 2*line_size
-    lea       r2, [r2+r3*2]  ; pix2 += 2*line_size
-
     paddd     m1, m2
     paddd     m3, m4
     paddd     m7, m1
     paddd     m7, m3
 
-    dec       r4
+%if %1 == mmsize
+    lea    pix1q, [pix1q + 2*lsizeq]
+    lea    pix2q, [pix2q + 2*lsizeq]
+%else
+    add    pix1q, lsizeq
+    add    pix2q, lsizeq
+%endif
+    dec       hd
     jnz .next2lines
 
     HADDD     m7, m1
     movd     eax, m7         ; return value
     RET
+%endmacro
+
+INIT_MMX mmx
+SUM_SQUARED_ERRORS 8
+
+INIT_MMX mmx
+SUM_SQUARED_ERRORS 16
+
+INIT_XMM sse2
+SUM_SQUARED_ERRORS 16
 
 INIT_MMX mmx
 ; void ff_get_pixels_mmx(int16_t *block, const uint8_t *pixels, int line_size)
diff --git a/libavcodec/x86/dsputilenc_mmx.c b/libavcodec/x86/dsputilenc_mmx.c
index e12ebedfa6..13128d29ad 100644
--- a/libavcodec/x86/dsputilenc_mmx.c
+++ b/libavcodec/x86/dsputilenc_mmx.c
@@ -46,6 +46,10 @@ int ff_sum_abs_dctelem_mmx(int16_t *block);
 int ff_sum_abs_dctelem_mmxext(int16_t *block);
 int ff_sum_abs_dctelem_sse2(int16_t *block);
 int ff_sum_abs_dctelem_ssse3(int16_t *block);
+int ff_sse8_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                int line_size, int h);
+int ff_sse16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
+                 int line_size, int h);
 int ff_sse16_sse2(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
                   int line_size, int h);
 
@@ -62,134 +66,7 @@ hadamard_func(ssse3)
 
 #if HAVE_INLINE_ASM
 
-static int sse8_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
-                    int line_size, int h)
-{
-    int tmp;
-
-    __asm__ volatile (
-        "movl         %4, %%ecx          \n"
-        "shr          $1, %%ecx          \n"
-        "pxor      %%mm0, %%mm0          \n" /* mm0 = 0 */
-        "pxor      %%mm7, %%mm7          \n" /* mm7 holds the sum */
-        "1:                              \n"
-        "movq       (%0), %%mm1          \n" /* mm1 = pix1[0][0 - 7] */
-        "movq       (%1), %%mm2          \n" /* mm2 = pix2[0][0 - 7] */
-        "movq   (%0, %3), %%mm3          \n" /* mm3 = pix1[1][0 - 7] */
-        "movq   (%1, %3), %%mm4          \n" /* mm4 = pix2[1][0 - 7] */
-
-        /* todo: mm1-mm2, mm3-mm4 */
-        /* algo: subtract mm1 from mm2 with saturation and vice versa */
-        /*       OR the results to get absolute difference */
-        "movq      %%mm1, %%mm5          \n"
-        "movq      %%mm3, %%mm6          \n"
-        "psubusb   %%mm2, %%mm1          \n"
-        "psubusb   %%mm4, %%mm3          \n"
-        "psubusb   %%mm5, %%mm2          \n"
-        "psubusb   %%mm6, %%mm4          \n"
-
-        "por       %%mm1, %%mm2          \n"
-        "por       %%mm3, %%mm4          \n"
-
-        /* now convert to 16-bit vectors so we can square them */
-        "movq      %%mm2, %%mm1          \n"
-        "movq      %%mm4, %%mm3          \n"
-
-        "punpckhbw %%mm0, %%mm2          \n"
-        "punpckhbw %%mm0, %%mm4          \n"
-        "punpcklbw %%mm0, %%mm1          \n" /* mm1 now spread over (mm1, mm2) */
-        "punpcklbw %%mm0, %%mm3          \n" /* mm4 now spread over (mm3, mm4) */
-
-        "pmaddwd   %%mm2, %%mm2          \n"
-        "pmaddwd   %%mm4, %%mm4          \n"
-        "pmaddwd   %%mm1, %%mm1          \n"
-        "pmaddwd   %%mm3, %%mm3          \n"
-
-        "lea (%0, %3, 2), %0             \n" /* pix1 += 2 * line_size */
-        "lea (%1, %3, 2), %1             \n" /* pix2 += 2 * line_size */
-
-        "paddd     %%mm2, %%mm1          \n"
-        "paddd     %%mm4, %%mm3          \n"
-        "paddd     %%mm1, %%mm7          \n"
-        "paddd     %%mm3, %%mm7          \n"
-
-        "decl      %%ecx                 \n"
-        "jnz       1b                    \n"
-
-        "movq      %%mm7, %%mm1          \n"
-        "psrlq       $32, %%mm7          \n" /* shift hi dword to lo */
-        "paddd     %%mm7, %%mm1          \n"
-        "movd      %%mm1, %2             \n"
-        : "+r" (pix1), "+r" (pix2), "=r" (tmp)
-        : "r" ((x86_reg) line_size), "m" (h)
-        : "%ecx");
-
-    return tmp;
-}
-
-static int sse16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
-                     int line_size, int h)
-{
-    int tmp;
-
-    __asm__ volatile (
-        "movl %4, %%ecx\n"
-        "pxor %%mm0, %%mm0\n"    /* mm0 = 0 */
-        "pxor %%mm7, %%mm7\n"    /* mm7 holds the sum */
-        "1:\n"
-        "movq (%0), %%mm1\n"     /* mm1 = pix1[0 -  7] */
-        "movq (%1), %%mm2\n"     /* mm2 = pix2[0 -  7] */
-        "movq 8(%0), %%mm3\n"    /* mm3 = pix1[8 - 15] */
-        "movq 8(%1), %%mm4\n"    /* mm4 = pix2[8 - 15] */
-
-        /* todo: mm1-mm2, mm3-mm4 */
-        /* algo: subtract mm1 from mm2 with saturation and vice versa */
-        /*       OR the results to get absolute difference */
-        "movq %%mm1, %%mm5\n"
-        "movq %%mm3, %%mm6\n"
-        "psubusb %%mm2, %%mm1\n"
-        "psubusb %%mm4, %%mm3\n"
-        "psubusb %%mm5, %%mm2\n"
-        "psubusb %%mm6, %%mm4\n"
-
-        "por %%mm1, %%mm2\n"
-        "por %%mm3, %%mm4\n"
-
-        /* now convert to 16-bit vectors so we can square them */
-        "movq %%mm2, %%mm1\n"
-        "movq %%mm4, %%mm3\n"
-
-        "punpckhbw %%mm0, %%mm2\n"
-        "punpckhbw %%mm0, %%mm4\n"
-        "punpcklbw %%mm0, %%mm1\n" /* mm1 now spread over (mm1, mm2) */
-        "punpcklbw %%mm0, %%mm3\n" /* mm4 now spread over (mm3, mm4) */
-
-        "pmaddwd %%mm2, %%mm2\n"
-        "pmaddwd %%mm4, %%mm4\n"
-        "pmaddwd %%mm1, %%mm1\n"
-        "pmaddwd %%mm3, %%mm3\n"
-
-        "add %3, %0\n"
-        "add %3, %1\n"
-
-        "paddd %%mm2, %%mm1\n"
-        "paddd %%mm4, %%mm3\n"
-        "paddd %%mm1, %%mm7\n"
-        "paddd %%mm3, %%mm7\n"
-
-        "decl %%ecx\n"
-        "jnz 1b\n"
-
-        "movq %%mm7, %%mm1\n"
-        "psrlq $32, %%mm7\n"    /* shift hi dword to lo */
-        "paddd %%mm7, %%mm1\n"
-        "movd %%mm1, %2\n"
-        : "+r" (pix1), "+r" (pix2), "=r" (tmp)
-        : "r" ((x86_reg) line_size), "m" (h)
-        : "%ecx");
-
-    return tmp;
-}
+#if HAVE_YASM
 
 static int hf_noise8_mmx(uint8_t *pix1, int line_size, int h)
 {
@@ -444,7 +321,7 @@ static int nsse16_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
     if (c)
         score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h);
     else
-        score1 = sse16_mmx(c, pix1, pix2, line_size, h);
+        score1 = ff_sse16_mmx(c, pix1, pix2, line_size, h);
     score2 = hf_noise16_mmx(pix1, line_size, h) -
              hf_noise16_mmx(pix2, line_size, h);
 
@@ -457,7 +334,7 @@ static int nsse16_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
 static int nsse8_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
                      int line_size, int h)
 {
-    int score1 = sse8_mmx(c, pix1, pix2, line_size, h);
+    int score1 = ff_sse8_mmx(c, pix1, pix2, line_size, h);
     int score2 = hf_noise8_mmx(pix1, line_size, h) -
                  hf_noise8_mmx(pix2, line_size, h);
 
@@ -467,6 +344,8 @@ static int nsse8_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
         return score1 + FFABS(score2) * 8;
 }
 
+#endif /* HAVE_YASM */
+
 static int vsad_intra16_mmx(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
                             int line_size, int h)
 {
@@ -808,12 +687,12 @@ av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx,
             (dct_algo == FF_DCT_AUTO || dct_algo == FF_DCT_MMX))
             c->fdct = ff_fdct_mmx;
 
-        c->sse[0]  = sse16_mmx;
-        c->sse[1]  = sse8_mmx;
         c->vsad[4] = vsad_intra16_mmx;
 
+#if HAVE_YASM
         c->nsse[0] = nsse16_mmx;
         c->nsse[1] = nsse8_mmx;
+#endif /* HAVE_YASM */
         if (!(avctx->flags & CODEC_FLAG_BITEXACT)) {
             c->vsad[0]      = vsad16_mmx;
             c->try_8x8basis = try_8x8basis_mmx;
@@ -860,6 +739,8 @@ av_cold void ff_dsputilenc_init_mmx(DSPContext *c, AVCodecContext *avctx,
         c->hadamard8_diff[0] = ff_hadamard8_diff16_mmx;
         c->hadamard8_diff[1] = ff_hadamard8_diff_mmx;
         c->sum_abs_dctelem   = ff_sum_abs_dctelem_mmx;
+        c->sse[0]            = ff_sse16_mmx;
+        c->sse[1]            = ff_sse8_mmx;
     }
 
     if (EXTERNAL_MMXEXT(cpu_flags)) {