x86/videodsp: add emulated_edge_mc_mmxext

This also changes hfix8_mmx and above to use mmx regs instead of
gprs, and makes emulated_edge_mc_sse and emulated_edge_mc_sse2 use
mmxext hfix and hvar functions instead of mmx where possible.

This is mostly in preparation for an ssse3 version.

Signed-off-by: James Almer <jamrial@gmail.com>

code is about 1 cpu cycle faster approximately

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

2 files changed, 74 insertions, 29 deletions

diff --git a/libavcodec/x86/videodsp.asm b/libavcodec/x86/videodsp.asm
index 1ac02574d6..ad15af9a27 100644
--- a/libavcodec/x86/videodsp.asm
+++ b/libavcodec/x86/videodsp.asm
@@ -92,21 +92,13 @@ INIT_XMM sse
 vvar_fn
 
 %macro hvar_fn 0
-cglobal emu_edge_hvar, 5, 6, 1, dst, dst_stride, start_x, n_words, h, w
+cglobal emu_edge_hvar, 5, 6, 2, dst, dst_stride, start_x, n_words, h, w
     lea            dstq, [dstq+n_wordsq*2]
     neg        n_wordsq
     lea        start_xq, [start_xq+n_wordsq*2]
 .y_loop:                                        ; do {
-    ; FIXME also write a ssse3 version using pshufb
-    movzx            wd, byte [dstq+start_xq]   ;   w = read(1)
-    imul             wd, 0x01010101             ;   w *= 0x01010101
-    movd             m0, wd
     mov              wq, n_wordsq               ;   initialize w
-%if cpuflag(sse2)
-    pshufd           m0, m0, q0000              ;   splat
-%else ; mmx
-    punpckldq        m0, m0                     ;   splat
-%endif ; mmx/sse
+    SPLATB_LOAD      m0, dstq+start_xq, m1      ;   read(1); splat
 .x_loop:                                        ;   do {
     movu    [dstq+wq*2], m0                     ;     write($reg, $mmsize)
     add              wq, mmsize/2               ;     w -= $mmsize/2
@@ -122,6 +114,8 @@ cglobal emu_edge_hvar, 5, 6, 1, dst, dst_stride, start_x, n_words, h, w
 %if ARCH_X86_32
 INIT_MMX mmx
 hvar_fn
+INIT_MMX mmxext
+hvar_fn
 %endif
 
 INIT_XMM sse2
@@ -344,16 +338,12 @@ VERTICAL_EXTEND 16, 22
 ; obviously not the same on both sides.
 
 %macro READ_V_PIXEL 2
-    movzx          vald, byte %2
+%if notcpuflag(mmxext) && %1 < 8
+    movzx          vald, byte [%2]
     imul           vald, 0x01010101
-%if %1 >= 8
-    movd             m0, vald
-%if mmsize == 16
-    pshufd           m0, m0, q0000
 %else
-    punpckldq        m0, m0
-%endif ; mmsize == 16
-%endif ; %1 > 16
+    SPLATB_LOAD      m0, %2, m1
+%endif ; %1 < 8
 %endmacro ; READ_V_PIXEL
 
 %macro WRITE_V_PIXEL 2
@@ -388,26 +378,42 @@ VERTICAL_EXTEND 16, 22
 %endif
 %endif ; %1-%%off >= 4
 
+%if %1-%%off == 2
+    movd     [%2+%%off-2], m0
+%endif ; (%1-%%off)/2
+
 %else ; %1 < 8
 
+%if cpuflag(mmxext)
+    movd     [%2+%%off], m0
+%if %1 == 6
+    movd     [%2+%%off+2], m0
+%endif ; (%1-%%off)/2
+
+%else ; notcpuflag(mmxext)
+
 %rep %1/4
     mov      [%2+%%off], vald
 %assign %%off %%off+4
 %endrep ; %1/4
 
-%endif ; %1 >=/< 8
-
 %if %1-%%off == 2
     mov      [%2+%%off], valw
 %endif ; (%1-%%off)/2
+%endif ; cpuflag
+%endif ; %1 >=/< 8
 %endmacro ; WRITE_V_PIXEL
 
 %macro H_EXTEND 2
 %assign %%n %1
 %rep 1+(%2-%1)/2
-cglobal emu_edge_hfix %+ %%n, 4, 5, 1, dst, dst_stride, start_x, bh, val
+%if %%n < 8 && notcpuflag(mmxext)
+cglobal emu_edge_hfix %+ %%n, 4, 5, 2, dst, dst_stride, start_x, bh, val
+%else
+cglobal emu_edge_hfix %+ %%n, 4, 4, 2, dst, dst_stride, start_x, bh
+%endif
 .loop_y:                                        ; do {
-    READ_V_PIXEL    %%n, [dstq+start_xq]        ;   $variable_regs = read($n)
+    READ_V_PIXEL    %%n, dstq+start_xq          ;   $variable_regs = read($n)
     WRITE_V_PIXEL   %%n, dstq                   ;   write($variable_regs, $n)
     add            dstq, dst_strideq            ;   dst += dst_stride
     dec             bhq                         ; } while (--bh)
@@ -418,11 +424,16 @@ cglobal emu_edge_hfix %+ %%n, 4, 5, 1, dst, dst_stride, start_x, bh, val
 %endmacro ; H_EXTEND
 
 INIT_MMX mmx
-H_EXTEND 2, 14
+H_EXTEND 2, 2
 %if ARCH_X86_32
-H_EXTEND 16, 22
+H_EXTEND 4, 22
 %endif
 
+INIT_MMX mmxext
+H_EXTEND 4, 14
+%if ARCH_X86_32
+H_EXTEND 16, 22
+%endif
 INIT_XMM sse2
 H_EXTEND 16, 22
 
diff --git a/libavcodec/x86/videodsp_init.c b/libavcodec/x86/videodsp_init.c
index 3218abdd88..bd61ab461f 100644
--- a/libavcodec/x86/videodsp_init.c
+++ b/libavcodec/x86/videodsp_init.c
@@ -117,15 +117,34 @@ static emu_edge_hfix_func *hfixtbl_mmx[11] = {
 };
 #endif
 extern emu_edge_hvar_func ff_emu_edge_hvar_mmx;
+extern emu_edge_hfix_func ff_emu_edge_hfix4_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix6_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix8_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix10_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix12_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix14_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix16_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix18_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix20_mmxext;
+extern emu_edge_hfix_func ff_emu_edge_hfix22_mmxext;
+#if ARCH_X86_32
+static emu_edge_hfix_func *hfixtbl_mmxext[11] = {
+    ff_emu_edge_hfix2_mmx,     ff_emu_edge_hfix4_mmxext,  ff_emu_edge_hfix6_mmxext,
+    ff_emu_edge_hfix8_mmxext,  ff_emu_edge_hfix10_mmxext, ff_emu_edge_hfix12_mmxext,
+    ff_emu_edge_hfix14_mmxext, ff_emu_edge_hfix16_mmxext, ff_emu_edge_hfix18_mmxext,
+    ff_emu_edge_hfix20_mmxext, ff_emu_edge_hfix22_mmxext
+};
+#endif
+extern emu_edge_hvar_func ff_emu_edge_hvar_mmxext;
 extern emu_edge_hfix_func ff_emu_edge_hfix16_sse2;
 extern emu_edge_hfix_func ff_emu_edge_hfix18_sse2;
 extern emu_edge_hfix_func ff_emu_edge_hfix20_sse2;
 extern emu_edge_hfix_func ff_emu_edge_hfix22_sse2;
 static emu_edge_hfix_func *hfixtbl_sse2[11] = {
-    ff_emu_edge_hfix2_mmx,  ff_emu_edge_hfix4_mmx,  ff_emu_edge_hfix6_mmx,
-    ff_emu_edge_hfix8_mmx,  ff_emu_edge_hfix10_mmx, ff_emu_edge_hfix12_mmx,
-    ff_emu_edge_hfix14_mmx, ff_emu_edge_hfix16_sse2, ff_emu_edge_hfix18_sse2,
-    ff_emu_edge_hfix20_sse2, ff_emu_edge_hfix22_sse2
+    ff_emu_edge_hfix2_mmx,     ff_emu_edge_hfix4_mmxext,  ff_emu_edge_hfix6_mmxext,
+    ff_emu_edge_hfix8_mmxext,  ff_emu_edge_hfix10_mmxext, ff_emu_edge_hfix12_mmxext,
+    ff_emu_edge_hfix14_mmxext, ff_emu_edge_hfix16_sse2,   ff_emu_edge_hfix18_sse2,
+    ff_emu_edge_hfix20_sse2,   ff_emu_edge_hfix22_sse2
 };
 extern emu_edge_hvar_func ff_emu_edge_hvar_sse2;
 
@@ -215,6 +234,17 @@ static av_noinline void emulated_edge_mc_mmx(uint8_t *buf, const uint8_t *src,
                      hfixtbl_mmx, &ff_emu_edge_hvar_mmx);
 }
 
+static av_noinline void emulated_edge_mc_mmxext(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_mmx, &ff_emu_edge_vvar_mmx,
+                     hfixtbl_mmxext, &ff_emu_edge_hvar_mmxext);
+}
+
 static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
                                              ptrdiff_t buf_stride,
                                              ptrdiff_t src_stride,
@@ -223,7 +253,7 @@ static av_noinline void emulated_edge_mc_sse(uint8_t *buf, const uint8_t *src,
 {
     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_mmx, &ff_emu_edge_hvar_mmx);
+                     hfixtbl_mmxext, &ff_emu_edge_hvar_mmxext);
 }
 #endif
 
@@ -258,6 +288,10 @@ av_cold void ff_videodsp_init_x86(VideoDSPContext *ctx, int bpc)
 #endif /* ARCH_X86_32 */
     if (EXTERNAL_MMXEXT(cpu_flags)) {
         ctx->prefetch = ff_prefetch_mmxext;
+#if ARCH_X86_32
+        if (bpc <= 8)
+            ctx->emulated_edge_mc = emulated_edge_mc_mmxext;
+#endif /* ARCH_X86_32 */
     }
 #if ARCH_X86_32
     if (EXTERNAL_SSE(cpu_flags) && bpc <= 8) {