Merge commit '90bc423212396e96a02edc1118982ab7f7766a63'

[ffmpeg] / libavcodec / x86 / huffyuvdsp_init.c

diff --git a/libavcodec/x86/huffyuvdsp_init.c b/libavcodec/x86/huffyuvdsp_init.c
index 3ced3c0a1cc76354efa5ad128d89c163b3f4bb91..f72d759ef2e8176366230cf97002b03c585f72b6 100644 (file)
--- a/libavcodec/x86/huffyuvdsp_init.c
+++ b/libavcodec/x86/huffyuvdsp_init.c
@@ -21,97 +21,35 @@
 #include "config.h"
 #include "libavutil/attributes.h"
 #include "libavutil/cpu.h"
+#include "libavutil/pixdesc.h"
 #include "libavutil/x86/asm.h"
 #include "libavutil/x86/cpu.h"
 #include "libavcodec/huffyuvdsp.h"
 
-void ff_add_bytes_mmx(uint8_t *dst, uint8_t *src, intptr_t w);
-void ff_add_bytes_sse2(uint8_t *dst, uint8_t *src, intptr_t w);
-
-void ff_add_hfyu_median_pred_mmxext(uint8_t *dst, const uint8_t *top,
-                                    const uint8_t *diff, intptr_t w,
-                                    int *left, int *left_top);
-void ff_add_hfyu_median_pred_sse2(uint8_t *dst, const uint8_t *top,
-                                  const uint8_t *diff, intptr_t w,
-                                  int *left, int *left_top);
-
-int  ff_add_hfyu_left_pred_ssse3(uint8_t *dst, const uint8_t *src,
-                                 intptr_t w, int left);
-int  ff_add_hfyu_left_pred_sse4(uint8_t *dst, const uint8_t *src,
-                                intptr_t w, int left);
-
+void ff_add_int16_mmx(uint16_t *dst, const uint16_t *src, unsigned mask, int w);
+void ff_add_int16_sse2(uint16_t *dst, const uint16_t *src, unsigned mask, int w);
 void ff_add_hfyu_left_pred_bgr32_mmx(uint8_t *dst, const uint8_t *src,
                                      intptr_t w, uint8_t *left);
 void ff_add_hfyu_left_pred_bgr32_sse2(uint8_t *dst, const uint8_t *src,
                                       intptr_t w, uint8_t *left);
+void ff_add_hfyu_median_pred_int16_mmxext(uint16_t *dst, const uint16_t *top, const uint16_t *diff, unsigned mask, int w, int *left, int *left_top);
 
-#if HAVE_INLINE_ASM && HAVE_7REGS && ARCH_X86_32
-static void add_hfyu_median_pred_cmov(uint8_t *dst, const uint8_t *top,
-                                      const uint8_t *diff, intptr_t w,
-                                      int *left, int *left_top)
-{
-    x86_reg w2 = -w;
-    x86_reg x;
-    int l  = *left     & 0xff;
-    int tl = *left_top & 0xff;
-    int t;
-    __asm__ volatile (
-        "mov          %7, %3            \n"
-        "1:                             \n"
-        "movzbl (%3, %4), %2            \n"
-        "mov          %2, %k3           \n"
-        "sub         %b1, %b3           \n"
-        "add         %b0, %b3           \n"
-        "mov          %2, %1            \n"
-        "cmp          %0, %2            \n"
-        "cmovg        %0, %2            \n"
-        "cmovg        %1, %0            \n"
-        "cmp         %k3, %0            \n"
-        "cmovg       %k3, %0            \n"
-        "mov          %7, %3            \n"
-        "cmp          %2, %0            \n"
-        "cmovl        %2, %0            \n"
-        "add    (%6, %4), %b0           \n"
-        "mov         %b0, (%5, %4)      \n"
-        "inc          %4                \n"
-        "jl           1b                \n"
-        : "+&q"(l), "+&q"(tl), "=&r"(t), "=&q"(x), "+&r"(w2)
-        : "r"(dst + w), "r"(diff + w), "rm"(top + w)
-    );
-    *left     = l;
-    *left_top = tl;
-}
-#endif
-
-av_cold void ff_huffyuvdsp_init_x86(HuffYUVDSPContext *c)
+av_cold void ff_huffyuvdsp_init_x86(HuffYUVDSPContext *c, AVCodecContext *avctx)
 {
     int cpu_flags = av_get_cpu_flags();
-
-#if HAVE_INLINE_ASM && HAVE_7REGS && ARCH_X86_32
-    if (cpu_flags & AV_CPU_FLAG_CMOV)
-        c->add_hfyu_median_pred = add_hfyu_median_pred_cmov;
-#endif
+    const AVPixFmtDescriptor *pix_desc = av_pix_fmt_desc_get(avctx->pix_fmt);
 
     if (ARCH_X86_32 && EXTERNAL_MMX(cpu_flags)) {
-        c->add_bytes = ff_add_bytes_mmx;
         c->add_hfyu_left_pred_bgr32 = ff_add_hfyu_left_pred_bgr32_mmx;
+        c->add_int16 = ff_add_int16_mmx;
     }
 
-    if (ARCH_X86_32 && EXTERNAL_MMXEXT(cpu_flags)) {
-        /* slower than cmov version on AMD */
-        if (!(cpu_flags & AV_CPU_FLAG_3DNOW))
-            c->add_hfyu_median_pred = ff_add_hfyu_median_pred_mmxext;
+    if (EXTERNAL_MMXEXT(cpu_flags) && pix_desc && pix_desc->comp[0].depth<16) {
+        c->add_hfyu_median_pred_int16 = ff_add_hfyu_median_pred_int16_mmxext;
     }
 
     if (EXTERNAL_SSE2(cpu_flags)) {
-        c->add_bytes            = ff_add_bytes_sse2;
-        c->add_hfyu_median_pred = ff_add_hfyu_median_pred_sse2;
+        c->add_int16 = ff_add_int16_sse2;
         c->add_hfyu_left_pred_bgr32 = ff_add_hfyu_left_pred_bgr32_sse2;
     }
-
-    if (EXTERNAL_SSSE3(cpu_flags)) {
-        c->add_hfyu_left_pred = ff_add_hfyu_left_pred_ssse3;
-        if (cpu_flags & AV_CPU_FLAG_SSE4) // not really SSE4, just slow on Conroe
-            c->add_hfyu_left_pred = ff_add_hfyu_left_pred_sse4;
-    }
 }