swscale/arm: add ff_nv{12,21}_to_{argb,rgba,abgr,bgra}_neon
[ffmpeg.git] / libswscale / arm / swscale_unscaled.c
index 04be762..5d0c8e9 100644 (file)
@@ -23,6 +23,7 @@
 #include "libswscale/swscale_internal.h"
 #include "libavutil/arm/cpu.h"
 
+#if 0
 extern void rgbx_to_nv12_neon_32(const uint8_t *src, uint8_t *y, uint8_t *chroma,
                 int width, int height,
                 int y_stride, int c_stride, int src_stride,
@@ -60,8 +61,69 @@ static int rgbx_to_nv12_neon_16_wrapper(SwsContext *context, const uint8_t *src[
 
     return 0;
 }
+#endif
+
+#define DECLARE_FF_NVX_TO_RGBX_FUNCS(ifmt, ofmt)                                            \
+int ff_##ifmt##_to_##ofmt##_neon(int w, int h,                                              \
+                                 uint8_t *dst, int linesize,                                \
+                                 const uint8_t *srcY, int linesizeY,                        \
+                                 const uint8_t *srcC, int linesizeC,                        \
+                                 const int16_t *table,                                      \
+                                 int y_offset,                                              \
+                                 int y_coeff);                                              \
+                                                                                            \
+static int ifmt##_to_##ofmt##_neon_wrapper(SwsContext *c, const uint8_t *src[],             \
+                                           int srcStride[], int srcSliceY, int srcSliceH,   \
+                                           uint8_t *dst[], int dstStride[]) {               \
+    const int16_t yuv2rgb_table[] = {                                                       \
+        c->yuv2rgb_v2r_coeff,                                                               \
+        c->yuv2rgb_u2g_coeff,                                                               \
+        c->yuv2rgb_v2g_coeff,                                                               \
+        c->yuv2rgb_u2b_coeff,                                                               \
+    };                                                                                      \
+                                                                                            \
+    ff_##ifmt##_to_##ofmt##_neon(c->srcW, srcSliceH,                                        \
+                                 dst[0] +  srcSliceY      * dstStride[0], dstStride[0],     \
+                                 src[0] +  srcSliceY      * srcStride[0], srcStride[0],     \
+                                 src[1] + (srcSliceY / 2) * srcStride[1], srcStride[1],     \
+                                 yuv2rgb_table,                                             \
+                                 c->yuv2rgb_y_offset >> 9,                                  \
+                                 c->yuv2rgb_y_coeff);                                       \
+                                                                                            \
+    return 0;                                                                               \
+}
+
+#define DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx)                                               \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, argb)                                                     \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, rgba)                                                     \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, abgr)                                                     \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, bgra)                                                     \
+
+DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nv12)
+DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nv21)
+
+/* We need a 16 pixel width alignment. This constraint can easily be removed
+ * for input reading but for the output which is 4-bytes per pixel (RGBA) the
+ * assembly might be writing as much as 4*15=60 extra bytes at the end of the
+ * line, which won't fit the 32-bytes buffer alignment. */
+#define SET_FF_NVX_TO_RGBX_FUNC(ifmt, IFMT, ofmt, OFMT) do {                                \
+    if (c->srcFormat == AV_PIX_FMT_##IFMT                                                   \
+        && c->dstFormat == AV_PIX_FMT_##OFMT                                                \
+        && !(c->srcH & 1)                                                                   \
+        && !(c->srcW & 15)) {                                                               \
+        c->swscale = ifmt##_to_##ofmt##_neon_wrapper;                                       \
+    }                                                                                       \
+} while (0)
+
+#define SET_FF_NVX_TO_ALL_RGBX_FUNC(nvx, NVX) do {                                          \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, argb, ARGB);                                          \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, rgba, RGBA);                                          \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, abgr, ABGR);                                          \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, bgra, BGRA);                                          \
+} while (0)
 
 static void get_unscaled_swscale_neon(SwsContext *c) {
+#if 0
     int accurate_rnd = c->flags & SWS_ACCURATE_RND;
     if (c->srcFormat == AV_PIX_FMT_RGBA
             && c->dstFormat == AV_PIX_FMT_NV12
@@ -69,6 +131,10 @@ static void get_unscaled_swscale_neon(SwsContext *c) {
         c->swscale = accurate_rnd ? rgbx_to_nv12_neon_32_wrapper
                         : rgbx_to_nv12_neon_16_wrapper;
     }
+#endif
+
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv12, NV12);
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv21, NV21);
 }
 
 void ff_get_unscaled_swscale_arm(SwsContext *c)