swscale/aarch64/yuv2rgb_neon: add BE 16bpp output formats

BE counterparts to the LE paths in 2e142e52ae; pack adds rev16 before
store. nv12/nv21 paths are added but bench-only (no C ref, same as
2e142e52ae).

Test Name                              A55-gcc           M1-clang             A76-gcc
-------------------------------------------------------------------------------------
yuv420p_rgb565be_1920_neon    15086.1 ( 3.91x)    5507.0 ( 4.34x)    19229.1 ( 2.02x)
yuv420p_bgr565be_1920_neon    15291.7 ( 3.84x)    5476.9 ( 4.37x)    19229.4 ( 2.02x)
yuv420p_rgb555be_1920_neon    15091.5 ( 3.67x)    5569.0 ( 3.97x)    19229.3 ( 1.90x)
yuv420p_bgr555be_1920_neon    15298.6 ( 3.62x)    5600.6 ( 3.98x)    19228.8 ( 1.90x)
yuv422p_rgb565be_1920_neon    16862.3 ( 4.00x)    6378.8 ( 4.64x)    22110.3 ( 2.07x)
yuv422p_bgr565be_1920_neon    17139.3 ( 3.93x)    6448.1 ( 4.50x)    22104.1 ( 2.07x)
yuv422p_rgb555be_1920_neon    16853.3 ( 3.98x)    6468.8 ( 4.12x)    22106.4 ( 1.98x)
yuv422p_bgr555be_1920_neon    17202.2 ( 3.89x)    6467.0 ( 4.12x)    22110.2 ( 1.98x)
yuva420p_rgb565be_1920_neon   15050.2 ( 3.92x)    5452.5 ( 4.39x)    19229.5 ( 2.02x)
yuva420p_bgr565be_1920_neon   15346.6 ( 3.84x)    5462.4 ( 4.36x)    19228.9 ( 2.02x)
yuva420p_rgb555be_1920_neon   15050.8 ( 3.69x)    5463.3 ( 3.95x)    19228.6 ( 1.90x)
yuva420p_bgr555be_1920_neon   15352.8 ( 3.61x)    5543.6 ( 3.89x)    19228.6 ( 1.90x)

Co-authored-by: Ramiro Polla <ramiro.polla@gmail.com>
Signed-off-by: DROOdotFOO <drew@axol.io>

libswscale/aarch64/swscale_unscaled.c

@@ -95,6 +95,10 @@ DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, rgb565le)
 DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, bgr565le)                                         \
 DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, rgb555le)                                         \
 DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, bgr555le)                                         \
+DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, rgb565be)                                         \
+DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, bgr565be)                                         \
+DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, rgb555be)                                         \
+DECLARE_FF_YUVX_TO_PACKED_RGB_FUNCS(yuvx, bgr555be)                                         \
 
 DECLARE_FF_YUVX_TO_ALL_RGB16_FUNCS(nv12)
 DECLARE_FF_YUVX_TO_ALL_RGB16_FUNCS(nv21)
@@ -161,6 +165,10 @@ static int nv24_to_yuv420p_neon_wrapper(SwsInternal *c, const uint8_t *const src
     SET_FF_YUVX_TO_RGBX_FUNC(yuvx, YUVX, bgr565le, BGR565LE, accurate_rnd);                 \
     SET_FF_YUVX_TO_RGBX_FUNC(yuvx, YUVX, rgb555le, RGB555LE, accurate_rnd);                 \
     SET_FF_YUVX_TO_RGBX_FUNC(yuvx, YUVX, bgr555le, BGR555LE, accurate_rnd);                 \
+    SET_FF_YUVX_TO_RGBX_FUNC(yuvx, YUVX, rgb565be, RGB565BE, accurate_rnd);                 \
+    SET_FF_YUVX_TO_RGBX_FUNC(yuvx, YUVX, bgr565be, BGR565BE, accurate_rnd);                 \
+    SET_FF_YUVX_TO_RGBX_FUNC(yuvx, YUVX, rgb555be, RGB555BE, accurate_rnd);                 \
+    SET_FF_YUVX_TO_RGBX_FUNC(yuvx, YUVX, bgr555be, BGR555BE, accurate_rnd);                 \
 } while (0)
 
 static void get_unscaled_swscale_neon(SwsInternal *c) {
@@ -186,6 +194,10 @@ static void get_unscaled_swscale_neon(SwsInternal *c) {
     SET_FF_YUVX_TO_RGBX_FUNC(yuv420p, YUVA420P, bgr565le, BGR565LE, accurate_rnd);
     SET_FF_YUVX_TO_RGBX_FUNC(yuv420p, YUVA420P, rgb555le, RGB555LE, accurate_rnd);
     SET_FF_YUVX_TO_RGBX_FUNC(yuv420p, YUVA420P, bgr555le, BGR555LE, accurate_rnd);
+    SET_FF_YUVX_TO_RGBX_FUNC(yuv420p, YUVA420P, rgb565be, RGB565BE, accurate_rnd);
+    SET_FF_YUVX_TO_RGBX_FUNC(yuv420p, YUVA420P, bgr565be, BGR565BE, accurate_rnd);
+    SET_FF_YUVX_TO_RGBX_FUNC(yuv420p, YUVA420P, rgb555be, RGB555BE, accurate_rnd);
+    SET_FF_YUVX_TO_RGBX_FUNC(yuv420p, YUVA420P, bgr555be, BGR555BE, accurate_rnd);
 
     if (c->opts.dst_format == AV_PIX_FMT_YUV420P &&
         (c->opts.src_format == AV_PIX_FMT_NV24 || c->opts.src_format == AV_PIX_FMT_NV42) &&
@@ -221,6 +233,10 @@ av_cold SwsFunc ff_yuv2rgb_init_aarch64(SwsInternal *c)
         case AV_PIX_FMT_BGR565LE: return yuv420p_to_bgr565le_neon_wrapper;
         case AV_PIX_FMT_RGB555LE: return yuv420p_to_rgb555le_neon_wrapper;
         case AV_PIX_FMT_BGR555LE: return yuv420p_to_bgr555le_neon_wrapper;
+        case AV_PIX_FMT_RGB565BE: return yuv420p_to_rgb565be_neon_wrapper;
+        case AV_PIX_FMT_BGR565BE: return yuv420p_to_bgr565be_neon_wrapper;
+        case AV_PIX_FMT_RGB555BE: return yuv420p_to_rgb555be_neon_wrapper;
+        case AV_PIX_FMT_BGR555BE: return yuv420p_to_bgr555be_neon_wrapper;
         }
     } else if (c->opts.src_format == AV_PIX_FMT_YUVA420P) {
         switch (c->opts.dst_format) {
@@ -238,6 +254,10 @@ av_cold SwsFunc ff_yuv2rgb_init_aarch64(SwsInternal *c)
         case AV_PIX_FMT_BGR565LE: return yuv420p_to_bgr565le_neon_wrapper;
         case AV_PIX_FMT_RGB555LE: return yuv420p_to_rgb555le_neon_wrapper;
         case AV_PIX_FMT_BGR555LE: return yuv420p_to_bgr555le_neon_wrapper;
+        case AV_PIX_FMT_RGB565BE: return yuv420p_to_rgb565be_neon_wrapper;
+        case AV_PIX_FMT_BGR565BE: return yuv420p_to_bgr565be_neon_wrapper;
+        case AV_PIX_FMT_RGB555BE: return yuv420p_to_rgb555be_neon_wrapper;
+        case AV_PIX_FMT_BGR555BE: return yuv420p_to_bgr555be_neon_wrapper;
         }
     } else if (c->opts.src_format == AV_PIX_FMT_YUV422P) {
         switch (c->opts.dst_format) {
@@ -252,6 +272,10 @@ av_cold SwsFunc ff_yuv2rgb_init_aarch64(SwsInternal *c)
         case AV_PIX_FMT_BGR565LE: return yuv422p_to_bgr565le_neon_wrapper;
         case AV_PIX_FMT_RGB555LE: return yuv422p_to_rgb555le_neon_wrapper;
         case AV_PIX_FMT_BGR555LE: return yuv422p_to_bgr555le_neon_wrapper;
+        case AV_PIX_FMT_RGB565BE: return yuv422p_to_rgb565be_neon_wrapper;
+        case AV_PIX_FMT_BGR565BE: return yuv422p_to_bgr565be_neon_wrapper;
+        case AV_PIX_FMT_RGB555BE: return yuv422p_to_rgb555be_neon_wrapper;
+        case AV_PIX_FMT_BGR555BE: return yuv422p_to_bgr555be_neon_wrapper;
         }
     }
     return NULL;

libswscale/aarch64/yuv2rgb_neon.S

@@ -301,6 +301,22 @@
         dst_load_args_packed 2
 .endm
 
+.macro dst_load_args_rgb565be
+        dst_load_args_packed 2
+.endm
+
+.macro dst_load_args_bgr565be
+        dst_load_args_packed 2
+.endm
+
+.macro dst_load_args_rgb555be
+        dst_load_args_packed 2
+.endm
+
+.macro dst_load_args_bgr555be
+        dst_load_args_packed 2
+.endm
+
 .macro dst_load_args_gbrp
         dst_load_args_planar
 .endm
@@ -365,6 +381,22 @@
         dst_load_args_packed_2l 2
 .endm
 
+.macro dst_load_args_rgb565be_2l
+        dst_load_args_packed_2l 2
+.endm
+
+.macro dst_load_args_bgr565be_2l
+        dst_load_args_packed_2l 2
+.endm
+
+.macro dst_load_args_rgb555be_2l
+        dst_load_args_packed_2l 2
+.endm
+
+.macro dst_load_args_bgr555be_2l
+        dst_load_args_packed_2l 2
+.endm
+
 // 2-lines-at-a-time planar dst loader. \sp_off is the byte offset at
 // which the caller's [sp+0] arg now lives (i.e., however many bytes the
 // caller pushed before invoking this macro). declare_2l_gbrp spills
@@ -639,11 +671,11 @@
  .endif
         compute_rgb     v4, v5, v6, v16, v17, v18
  .if r_first
-        // rgb*le: (R << hshift) | (G << 5) | B
+        // rgb*: (R << hshift) | (G << 5) | B
         pack_rgb16_2l   v8,  v6,  v5,  v4,  gshift, hshift
         pack_rgb16_2l   v9,  v18, v17, v16, gshift, hshift
  .else
-        // bgr*le: (B << hshift) | (G << 5) | R
+        // bgr*: (B << hshift) | (G << 5) | R
         pack_rgb16_2l   v8,  v4,  v5,  v6,  gshift, hshift
         pack_rgb16_2l   v9,  v16, v17, v18, gshift, hshift
  .endif
@@ -651,15 +683,16 @@
 .endif
 .endm
 
-// Map ofmt to .set predicates: rgb16=1 for the four 16bpp LE ofmts
+// Map ofmt to .set predicates: rgb16=1 for the eight 16bpp ofmts
 // (r_first=1 for rgb*, 0 for bgr*; gshift/hshift = 2/11 for 565,
-// 3/10 for 555), letting sibling macros branch on .if rgb16 instead of
+// 3/10 for 555; is_be=1 for the BE variants), letting sibling macros
-// repeating a four-way .ifc cascade.
+// branch on .if rgb16 / .if is_be instead of repeating .ifc cascades.
 .macro set_rgb16_predicates ofmt
         .set rgb16,     0
         .set r_first,   0
         .set gshift,    0
         .set hshift,    0
+        .set is_be,     0
 .ifc \ofmt,rgb565le
         .set rgb16,     1
         .set r_first,   1
@@ -682,6 +715,32 @@
         .set gshift,    3
         .set hshift,   10
 .endif
+.ifc \ofmt,rgb565be
+        .set rgb16,     1
+        .set r_first,   1
+        .set gshift,    2
+        .set hshift,   11
+        .set is_be,     1
+.endif
+.ifc \ofmt,bgr565be
+        .set rgb16,     1
+        .set gshift,    2
+        .set hshift,   11
+        .set is_be,     1
+.endif
+.ifc \ofmt,rgb555be
+        .set rgb16,     1
+        .set r_first,   1
+        .set gshift,    3
+        .set hshift,   10
+        .set is_be,     1
+.endif
+.ifc \ofmt,bgr555be
+        .set rgb16,     1
+        .set gshift,    3
+        .set hshift,   10
+        .set is_be,     1
+.endif
 .endm
 
 // 16bpp packing uses v8/v9 as the accumulator. AAPCS-64 requires d8/d9
@@ -704,10 +763,10 @@
 // Pack 8 pixels of 16bpp output. The three channels are extracted via ushr,
 // widened to u16, then merged via shift-left-insert:
 //   dst = (high << high_shl) | (mid << 5) | low
-// For RGB565LE  pass (B, G, R) as (low, mid, high), g_shr=2, high_shl=11.
+// For RGB565LE/BE  pass (B, G, R) as (low, mid, high), g_shr=2, high_shl=11.
-// For BGR565LE  pass (R, G, B), g_shr=2, high_shl=11.
+// For BGR565LE/BE  pass (R, G, B), g_shr=2, high_shl=11.
-// For RGB555LE  pass (B, G, R), g_shr=3, high_shl=10.
+// For RGB555LE/BE  pass (B, G, R), g_shr=3, high_shl=10.
-// For BGR555LE  pass (R, G, B), g_shr=3, high_shl=10.
+// For BGR555LE/BE  pass (R, G, B), g_shr=3, high_shl=10.
 // Clobbers v20-v23.
 .macro pack_rgb16 dst, low_ch, mid_ch, high_ch, g_shr, high_shl
         ushr            v20.8b, \high_ch\().8b, #3
@@ -718,6 +777,9 @@
         sli             \dst\().8h, v23.8h, #5
         uxtl            v23.8h, v20.8b
         sli             \dst\().8h, v23.8h, #\high_shl
+.if is_be
+        rev16           \dst\().16b, \dst\().16b
+.endif
 .endm
 
 // As pack_rgb16 but uses v26-v29 as scratch (luma temps, dead after
@@ -733,6 +795,9 @@
         sli             \dst\().8h, v29.8h, #5
         uxtl            v29.8h, v26.8b
         sli             \dst\().8h, v29.8h, #\high_shl
+.if is_be
+        rev16           \dst\().16b, \dst\().16b
+.endif
 .endm
 
 .macro declare_func ifmt ofmt
@@ -827,11 +892,11 @@ function ff_\ifmt\()_to_\ofmt\()_neon, export=1
    .if rgb16
         compute_rgb     v4,v5,v6, v16,v17,v18
     .if r_first
-        // rgb*le: (R << hshift) | (G << 5) | B
+        // rgb*: (R << hshift) | (G << 5) | B
         pack_rgb16      v8,  v6,  v5,  v4,  gshift, hshift
         pack_rgb16      v9,  v18, v17, v16, gshift, hshift
     .else
-        // bgr*le: (B << hshift) | (G << 5) | R
+        // bgr*: (B << hshift) | (G << 5) | R
         pack_rgb16      v8,  v4,  v5,  v6,  gshift, hshift
         pack_rgb16      v9,  v16, v17, v18, gshift, hshift
     .endif
@@ -970,25 +1035,43 @@ declare_rgb_funcs_2l_packed yuv420p
 declare_2l_gbrp             yuv420p
 declare_rgb_funcs           yuv422p
 
-.macro declare_rgb16_funcs ifmt
+.macro declare_rgb16le_funcs ifmt
         declare_func    \ifmt, rgb565le
         declare_func    \ifmt, bgr565le
         declare_func    \ifmt, rgb555le
         declare_func    \ifmt, bgr555le
 .endm
 
-.macro declare_rgb16_funcs_2l ifmt
+.macro declare_rgb16le_funcs_2l ifmt
         declare_2l_packed \ifmt, rgb565le
         declare_2l_packed \ifmt, bgr565le
         declare_2l_packed \ifmt, rgb555le
         declare_2l_packed \ifmt, bgr555le
 .endm
 
+.macro declare_rgb16be_funcs ifmt
+        declare_func    \ifmt, rgb565be
+        declare_func    \ifmt, bgr565be
+        declare_func    \ifmt, rgb555be
+        declare_func    \ifmt, bgr555be
+.endm
+
+.macro declare_rgb16be_funcs_2l ifmt
+        declare_2l_packed \ifmt, rgb565be
+        declare_2l_packed \ifmt, bgr565be
+        declare_2l_packed \ifmt, rgb555be
+        declare_2l_packed \ifmt, bgr555be
+.endm
+
 // Subsampled inputs take the 2-line rgb16 path; yuv422p stays single-row.
-declare_rgb16_funcs_2l nv12
+declare_rgb16le_funcs_2l nv12
-declare_rgb16_funcs_2l nv21
+declare_rgb16be_funcs_2l nv12
-declare_rgb16_funcs_2l yuv420p
+declare_rgb16le_funcs_2l nv21
-declare_rgb16_funcs    yuv422p
+declare_rgb16be_funcs_2l nv21
+declare_rgb16le_funcs_2l yuv420p
+declare_rgb16be_funcs_2l yuv420p
+declare_rgb16le_funcs    yuv422p
+declare_rgb16be_funcs    yuv422p
 
 .macro declare_yuva_funcs ifmt
         declare_func    \ifmt, argb

tests/checkasm/sw_yuv2rgb.c

@@ -46,10 +46,14 @@ static const int dst_fmts[] = {
     AV_PIX_FMT_BGRA,
     AV_PIX_FMT_RGB24,
     AV_PIX_FMT_BGR24,
-    AV_PIX_FMT_RGB565,
+    AV_PIX_FMT_RGB565LE,
-    AV_PIX_FMT_BGR565,
+    AV_PIX_FMT_BGR565LE,
-    AV_PIX_FMT_RGB555,
+    AV_PIX_FMT_RGB555LE,
-    AV_PIX_FMT_BGR555,
+    AV_PIX_FMT_BGR555LE,
+    AV_PIX_FMT_RGB565BE,
+    AV_PIX_FMT_BGR565BE,
+    AV_PIX_FMT_RGB555BE,
+    AV_PIX_FMT_BGR555BE,
 //     AV_PIX_FMT_RGB444,
 //     AV_PIX_FMT_BGR444,
 //     AV_PIX_FMT_RGB8,
@@ -71,31 +75,31 @@ static int cmp_off_by_n(const uint8_t *ref, const uint8_t *test, size_t n, int a
     return 0;
 }
 
-static int cmp_555_by_n(const uint8_t *ref, const uint8_t *test, size_t n, int accuracy)
+static int cmp_555_by_n(const uint8_t *ref, const uint8_t *test, size_t n, int accuracy, int is_be)
 {
-    const uint16_t *ref16  = (const uint16_t *) ref;
-    const uint16_t *test16 = (const uint16_t *) test;
     for (size_t i = 0; i < n; i++) {
-        if (abs(( ref16[i]        & 0x1f) - ( test16[i]        & 0x1f)) > accuracy)
+        uint16_t r = is_be ? AV_RB16(ref  + i * 2) : AV_RL16(ref  + i * 2);
+        uint16_t t = is_be ? AV_RB16(test + i * 2) : AV_RL16(test + i * 2);
+        if (abs(( r        & 0x1f) - ( t        & 0x1f)) > accuracy)
             return 1;
-        if (abs(((ref16[i] >>  5) & 0x1f) - ((test16[i] >>  5) & 0x1f)) > accuracy)
+        if (abs(((r >>  5) & 0x1f) - ((t >>  5) & 0x1f)) > accuracy)
             return 1;
-        if (abs(((ref16[i] >> 10) & 0x1f) - ((test16[i] >> 10) & 0x1f)) > accuracy)
+        if (abs(((r >> 10) & 0x1f) - ((t >> 10) & 0x1f)) > accuracy)
             return 1;
     }
     return 0;
 }
 
-static int cmp_565_by_n(const uint8_t *ref, const uint8_t *test, size_t n, int accuracy)
+static int cmp_565_by_n(const uint8_t *ref, const uint8_t *test, size_t n, int accuracy, int is_be)
 {
-    const uint16_t *ref16  = (const uint16_t *) ref;
-    const uint16_t *test16 = (const uint16_t *) test;
     for (size_t i = 0; i < n; i++) {
-        if (abs(( ref16[i]        & 0x1f) - ( test16[i]        & 0x1f)) > accuracy)
+        uint16_t r = is_be ? AV_RB16(ref  + i * 2) : AV_RL16(ref  + i * 2);
+        uint16_t t = is_be ? AV_RB16(test + i * 2) : AV_RL16(test + i * 2);
+        if (abs(( r        & 0x1f) - ( t        & 0x1f)) > accuracy)
             return 1;
-        if (abs(((ref16[i] >>  5) & 0x3f) - ((test16[i] >>  5) & 0x3f)) > accuracy)
+        if (abs(((r >>  5) & 0x3f) - ((t >>  5) & 0x3f)) > accuracy)
             return 1;
-        if (abs(((ref16[i] >> 11) & 0x1f) - ((test16[i] >> 11) & 0x1f)) > accuracy)
+        if (abs(((r >> 11) & 0x1f) - ((t >> 11) & 0x1f)) > accuracy)
             return 1;
     }
     return 0;
@@ -199,19 +203,27 @@ static void check_yuv2rgb(int src_pix_fmt)
                                          dst1_0 + row * dstStride[0],
                                          width * sample_size, 3))
                             fail();
-                } else if (dst_pix_fmt == AV_PIX_FMT_RGB565 ||
+                } else if (dst_pix_fmt == AV_PIX_FMT_RGB565LE ||
-                           dst_pix_fmt == AV_PIX_FMT_BGR565) {
+                           dst_pix_fmt == AV_PIX_FMT_BGR565LE ||
+                           dst_pix_fmt == AV_PIX_FMT_RGB565BE ||
+                           dst_pix_fmt == AV_PIX_FMT_BGR565BE) {
+                    int is_be = dst_pix_fmt == AV_PIX_FMT_RGB565BE ||
+                                dst_pix_fmt == AV_PIX_FMT_BGR565BE;
                     for (int row = 0; row < srcSliceH; row++)
                         if (cmp_565_by_n(dst0_0 + row * dstStride[0],
                                          dst1_0 + row * dstStride[0],
-                                         width, 2))
+                                         width, 2, is_be))
                             fail();
-                } else if (dst_pix_fmt == AV_PIX_FMT_RGB555 ||
+                } else if (dst_pix_fmt == AV_PIX_FMT_RGB555LE ||
-                           dst_pix_fmt == AV_PIX_FMT_BGR555) {
+                           dst_pix_fmt == AV_PIX_FMT_BGR555LE ||
+                           dst_pix_fmt == AV_PIX_FMT_RGB555BE ||
+                           dst_pix_fmt == AV_PIX_FMT_BGR555BE) {
+                    int is_be = dst_pix_fmt == AV_PIX_FMT_RGB555BE ||
+                                dst_pix_fmt == AV_PIX_FMT_BGR555BE;
                     for (int row = 0; row < srcSliceH; row++)
                         if (cmp_555_by_n(dst0_0 + row * dstStride[0],
                                          dst1_0 + row * dstStride[0],
-                                         width, 2))
+                                         width, 2, is_be))
                             fail();
                 } else if (dst_pix_fmt == AV_PIX_FMT_GBRP) {
                     for (int p = 0; p < 3; p++)