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
%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 -= $mmsize/2
INIT_XMM sse2
hvar_fn
+%if HAVE_AVX2_EXTERNAL
+INIT_XMM avx2
+hvar_fn
+%endif
+
; macro to read/write a horizontal number of pixels (%2) to/from registers
; on sse, - fills xmm0-15 for consecutive sets of 16 pixels
; - if (%2 & 8) fills 8 bytes into xmm$next
; 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
punpckldq m0, m0
%endif ; mmsize == 16
%endif ; %1 > 16
+%endif ; avx2
%endmacro ; READ_V_PIXEL
%macro WRITE_V_PIXEL 2
%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)
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: