x86: hpeldsp: implement SSSE3 version of _xy2

Loading pb_1 rather than pw_8192 was benchmarked to be more efficient.
Loading of the 2 yields no advantage. Loading of one saves ~11 cycles.

decicycles count:
put8:  3223(mmx)    -> 2387
avg8:  2863(mmxext) -> 2125
put16: 4356(sse2)   -> 3553
avg16: 4481(sse2)   -> 3513
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/x86/hpeldsp.asm

@@ -30,6 +30,9 @@
 SECTION_RODATA
 cextern pb_1
 cextern pw_2
+pw_8192: times 8 dw (1<<13)
+pb_interleave16: db 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15
+pb_interleave8:  db 0, 4, 1, 5, 2, 6, 3, 7
 
 SECTION_TEXT
 
@@ -635,3 +638,70 @@ SET_PIXELS_XY2 avg
 INIT_XMM sse2
 SET_PIXELS_XY2 put
 SET_PIXELS_XY2 avg
+
+%macro SSSE3_PIXELS_XY2 1-2
+%if %0 == 2 ; sse2
+cglobal %1_pixels16_xy2, 4,5,%2
+    mova        m4, [pb_interleave16]
+%else
+cglobal %1_pixels8_xy2, 4,5
+    mova        m4, [pb_interleave8]
+%endif
+    mova        m5, [pb_1]
+    movu        m0, [r1]
+    movu        m1, [r1+1]
+    pmaddubsw   m0, m5
+    pmaddubsw   m1, m5
+    xor         r4, r4
+    add         r1, r2
+.loop:
+    movu        m2, [r1+r4]
+    movu        m3, [r1+r4+1]
+    pmaddubsw   m2, m5
+    pmaddubsw   m3, m5
+    paddusw     m0, m2
+    paddusw     m1, m3
+    pmulhrsw    m0, [pw_8192]
+    pmulhrsw    m1, [pw_8192]
+%ifidn %1, avg
+    mova        m6, [r0+r4]
+    packuswb    m0, m1
+    pshufb      m0, m4
+    pavgb       m0, m6
+%else
+    packuswb    m0, m1
+    pshufb      m0, m4
+%endif
+    mova   [r0+r4], m0
+    add         r4, r2
+
+    movu        m0, [r1+r4]
+    movu        m1, [r1+r4+1]
+    pmaddubsw   m0, m5
+    pmaddubsw   m1, m5
+    paddusw     m2, m0
+    paddusw     m3, m1
+    pmulhrsw    m2, [pw_8192]
+    pmulhrsw    m3, [pw_8192]
+%ifidn %1, avg
+    mova        m6, [r0+r4]
+    packuswb    m2, m3
+    pshufb      m2, m4
+    pavgb       m2, m6
+%else
+    packuswb    m2, m3
+    pshufb      m2, m4
+%endif
+    mova   [r0+r4], m2
+    add         r4, r2
+    sub        r3d, 2
+    jnz .loop
+    REP_RET
+%endmacro
+
+INIT_MMX ssse3
+SSSE3_PIXELS_XY2 put
+SSSE3_PIXELS_XY2 avg
+INIT_XMM ssse3
+SSSE3_PIXELS_XY2 put, 6
+SSSE3_PIXELS_XY2 avg, 7

libavcodec/x86/hpeldsp_init.c

@@ -95,6 +95,15 @@ void ff_avg_approx_pixels8_xy2_mmxext(uint8_t *block, const uint8_t *pixels,
 void ff_avg_approx_pixels8_xy2_3dnow(uint8_t *block, const uint8_t *pixels,
                                      ptrdiff_t line_size, int h);
 
+void ff_put_pixels8_xy2_ssse3(uint8_t *block, const uint8_t *pixels,
+                              ptrdiff_t line_size, int h);
+void ff_avg_pixels8_xy2_ssse3(uint8_t *block, const uint8_t *pixels,
+                              ptrdiff_t line_size, int h);
+void ff_put_pixels16_xy2_ssse3(uint8_t *block, const uint8_t *pixels,
+                               ptrdiff_t line_size, int h);
+void ff_avg_pixels16_xy2_ssse3(uint8_t *block, const uint8_t *pixels,
+                               ptrdiff_t line_size, int h);
+
 #define avg_pixels8_mmx         ff_avg_pixels8_mmx
 #define avg_pixels8_x2_mmx      ff_avg_pixels8_x2_mmx
 #define avg_pixels16_mmx        ff_avg_pixels16_mmx
@@ -307,6 +316,16 @@ static void hpeldsp_init_sse2(HpelDSPContext *c, int flags, int cpu_flags)
 #endif /* HAVE_SSE2_EXTERNAL */
 }
 
+static void hpeldsp_init_ssse3(HpelDSPContext *c, int flags, int cpu_flags)
+{
+#if HAVE_SSSE3_EXTERNAL
+    c->put_pixels_tab[0][3]            = ff_put_pixels16_xy2_ssse3;
+    c->avg_pixels_tab[0][3]            = ff_avg_pixels16_xy2_ssse3;
+    c->put_pixels_tab[1][3]            = ff_put_pixels8_xy2_ssse3;
+    c->avg_pixels_tab[1][3]            = ff_avg_pixels8_xy2_ssse3;
+#endif
+}
+
 av_cold void ff_hpeldsp_init_x86(HpelDSPContext *c, int flags)
 {
     int cpu_flags = av_get_cpu_flags();
@@ -322,4 +341,7 @@ av_cold void ff_hpeldsp_init_x86(HpelDSPContext *c, int flags)
 
     if (EXTERNAL_SSE2(cpu_flags))
         hpeldsp_init_sse2(c, flags, cpu_flags);
+
+    if (EXTERNAL_SSSE3(cpu_flags))
+        hpeldsp_init_ssse3(c, flags, cpu_flags);
 }