x86/videodsp: add ff_emu_edge_{hfix,hvar}_avx2

~15% faster than sse2.
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>

libavcodec/x86/videodsp.asm

@@ -97,7 +97,10 @@ cglobal emu_edge_hvar, 5, 6, 1, dst, dst_stride, start_x, n_words, h, w
     neg        n_wordsq
     lea        start_xq, [start_xq+n_wordsq*2]
 .y_loop:                                        ; do {
-    ; FIXME also write a ssse3 version using pshufb
+%if cpuflag(avx2)
+    vpbroadcastb     m0, [dstq+start_xq]
+    mov              wq, n_wordsq               ;   initialize w
+%else
     movzx            wd, byte [dstq+start_xq]   ;   w = read(1)
     imul             wd, 0x01010101             ;   w *= 0x01010101
     movd             m0, wd
@@ -107,6 +110,7 @@ cglobal emu_edge_hvar, 5, 6, 1, dst, dst_stride, start_x, n_words, h, w
 %else ; mmx
     punpckldq        m0, m0                     ;   splat
 %endif ; mmx/sse
+%endif ; avx2
 .x_loop:                                        ;   do {
     movu    [dstq+wq*2], m0                     ;     write($reg, $mmsize)
     add              wq, mmsize/2               ;     w -= $mmsize/2
@@ -127,6 +131,11 @@ hvar_fn
 INIT_XMM sse2
 hvar_fn
 
+%if HAVE_AVX2_EXTERNAL
+INIT_XMM avx2
+hvar_fn
+%endif
+
 ; macro to read/write a horizontal number of pixels (%2) to/from registers
 ; on sse, - fills xmm0-15 for consecutive sets of 16 pixels
 ;         - if (%2 & 8)  fills 8 bytes into xmm$next
@@ -344,6 +353,9 @@ VERTICAL_EXTEND 16, 22
 ; obviously not the same on both sides.
 
 %macro READ_V_PIXEL 2
+%if cpuflag(avx2)
+    vpbroadcastb     m0, %2
+%else
     movzx          vald, byte %2
     imul           vald, 0x01010101
 %if %1 >= 8
@@ -354,6 +366,7 @@ VERTICAL_EXTEND 16, 22
     punpckldq        m0, m0
 %endif ; mmsize == 16
 %endif ; %1 > 16
+%endif ; avx2
 %endmacro ; READ_V_PIXEL
 
 %macro WRITE_V_PIXEL 2
@@ -398,14 +411,22 @@ VERTICAL_EXTEND 16, 22
 %endif ; %1 >=/< 8
 
 %if %1-%%off == 2
+%if cpuflag(avx2)
+    movd     [%2+%%off-2], m0
+%else
     mov      [%2+%%off], valw
+%endif ; avx2
 %endif ; (%1-%%off)/2
 %endmacro ; WRITE_V_PIXEL
 
 %macro H_EXTEND 2
 %assign %%n %1
 %rep 1+(%2-%1)/2
+%if cpuflag(avx2)
+cglobal emu_edge_hfix %+ %%n, 4, 4, 1, dst, dst_stride, start_x, bh
+%else
 cglobal emu_edge_hfix %+ %%n, 4, 5, 1, dst, dst_stride, start_x, bh, val
+%endif
 .loop_y:                                        ; do {
     READ_V_PIXEL    %%n, [dstq+start_xq]        ;   $variable_regs = read($n)
     WRITE_V_PIXEL   %%n, dstq                   ;   write($variable_regs, $n)
@@ -426,6 +447,11 @@ H_EXTEND 16, 22
 INIT_XMM sse2
 H_EXTEND 16, 22
 
+%if HAVE_AVX2_EXTERNAL
+INIT_XMM avx2
+H_EXTEND 8, 22
+%endif
+
 %macro PREFETCH_FN 1
 cglobal prefetch, 3, 3, 0, buf, stride, h
 .loop:

libavcodec/x86/videodsp_init.c

@@ -128,6 +128,23 @@ static emu_edge_hfix_func * const hfixtbl_sse2[11] = {
     ff_emu_edge_hfix20_sse2, ff_emu_edge_hfix22_sse2
 };
 extern emu_edge_hvar_func ff_emu_edge_hvar_sse2;
+#if HAVE_AVX2_EXTERNAL
+extern emu_edge_hfix_func ff_emu_edge_hfix8_avx2;
+extern emu_edge_hfix_func ff_emu_edge_hfix10_avx2;
+extern emu_edge_hfix_func ff_emu_edge_hfix12_avx2;
+extern emu_edge_hfix_func ff_emu_edge_hfix14_avx2;
+extern emu_edge_hfix_func ff_emu_edge_hfix16_avx2;
+extern emu_edge_hfix_func ff_emu_edge_hfix18_avx2;
+extern emu_edge_hfix_func ff_emu_edge_hfix20_avx2;
+extern emu_edge_hfix_func ff_emu_edge_hfix22_avx2;
+static emu_edge_hfix_func * const hfixtbl_avx2[11] = {
+    ff_emu_edge_hfix2_mmx,  ff_emu_edge_hfix4_mmx,  ff_emu_edge_hfix6_mmx,
+    ff_emu_edge_hfix8_avx2,  ff_emu_edge_hfix10_avx2, ff_emu_edge_hfix12_avx2,
+    ff_emu_edge_hfix14_avx2, ff_emu_edge_hfix16_avx2, ff_emu_edge_hfix18_avx2,
+    ff_emu_edge_hfix20_avx2, ff_emu_edge_hfix22_avx2
+};
+extern emu_edge_hvar_func ff_emu_edge_hvar_avx2;
+#endif
 
 static av_always_inline void emulated_edge_mc(uint8_t *dst, const uint8_t *src,
                                               ptrdiff_t dst_stride,
@@ -238,6 +255,20 @@ static av_noinline void emulated_edge_mc_sse2(uint8_t *buf, const uint8_t *src,
                      src_x, src_y, w, h, vfixtbl_sse, &ff_emu_edge_vvar_sse,
                      hfixtbl_sse2, &ff_emu_edge_hvar_sse2);
 }
+
+#if HAVE_AVX2_EXTERNAL
+static av_noinline void emulated_edge_mc_avx2(uint8_t *buf, const uint8_t *src,
+                                              ptrdiff_t buf_stride,
+                                              ptrdiff_t src_stride,
+                                              int block_w, int block_h,
+                                              int src_x, int src_y, int w,
+                                              int h)
+{
+    emulated_edge_mc(buf, src, buf_stride, src_stride, block_w, block_h,
+                     src_x, src_y, w, h, vfixtbl_sse, &ff_emu_edge_vvar_sse,
+                     hfixtbl_avx2, &ff_emu_edge_hvar_avx2);
+}
+#endif /* HAVE_AVX2_EXTERNAL */
 #endif /* HAVE_YASM */
 
 void ff_prefetch_mmxext(uint8_t *buf, ptrdiff_t stride, int h);
@@ -267,5 +298,10 @@ av_cold void ff_videodsp_init_x86(VideoDSPContext *ctx, int bpc)
     if (EXTERNAL_SSE2(cpu_flags) && bpc <= 8) {
         ctx->emulated_edge_mc = emulated_edge_mc_sse2;
     }
+#if HAVE_AVX2_EXTERNAL
+    if (EXTERNAL_AVX2(cpu_flags) && bpc <= 8) {
+        ctx->emulated_edge_mc = emulated_edge_mc_avx2;
+    }
+#endif
 #endif /* HAVE_YASM */
 }