libavcodec/lossless_videodsp : add add_bytes avx2 version

libavcodec/lossless_videodsp.h

@@ -29,7 +29,7 @@
 #include "libavutil/cpu.h"
 
 typedef struct LLVidDSPContext {
-    void (*add_bytes)(uint8_t *dst /* align 16 */, uint8_t *src /* align 16 */,
+    void (*add_bytes)(uint8_t *dst /* align 32 */, uint8_t *src /* align 32 */,
                       ptrdiff_t w);
     void (*add_median_pred)(uint8_t *dst, const uint8_t *top,
                             const uint8_t *diff, ptrdiff_t w,

libavcodec/x86/lossless_videodsp.asm

@@ -217,6 +217,11 @@ ADD_BYTES
 INIT_XMM sse2
 ADD_BYTES
 
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+ADD_BYTES
+%endif
+
 %macro ADD_HFYU_LEFT_LOOP_INT16 2 ; %1 = dst alignment (a/u), %2 = src alignment (a/u)
     add     wd, wd
     add     srcq, wq

libavcodec/x86/lossless_videodsp_init.c

@@ -25,6 +25,7 @@
 
 void ff_add_bytes_mmx(uint8_t *dst, uint8_t *src, ptrdiff_t w);
 void ff_add_bytes_sse2(uint8_t *dst, uint8_t *src, ptrdiff_t w);
+void ff_add_bytes_avx2(uint8_t *dst, uint8_t *src, ptrdiff_t w);
 
 void ff_add_median_pred_mmxext(uint8_t *dst, const uint8_t *top,
                                const uint8_t *diff, ptrdiff_t w,
@@ -115,4 +116,7 @@ void ff_llviddsp_init_x86(LLVidDSPContext *c)
     if (EXTERNAL_SSE4(cpu_flags)) {
         c->add_left_pred_int16 = ff_add_left_pred_int16_sse4;
     }
+    if (EXTERNAL_AVX2_FAST(cpu_flags)) {
+        c->add_bytes       = ff_add_bytes_avx2;
+    }
 }