[FFmpeg-devel] [PATCH] sws/aarch64: add {nv12, nv21, yuv420p, yuv422p}_to_{argb, rgba, abgr, rgba}_neon

Clément Bœsch u at pkh.me
Wed Feb 17 16:18:26 CET 2016


From: Clément Bœsch <clement at stupeflix.com>

---
This code is mirroring the ARM version.

It's not tested on real hardware (qemu only), so I have no idea about
performance. This also means I didn't put any prefetch mechanism (which had a
noticeable impact on performance on ARM).

Last but not least, the 32-bit precision functions are added (and tested) but
are not enabled at all, just like the ARM counterpart.
---
 libswscale/aarch64/Makefile           |   3 +
 libswscale/aarch64/swscale_unscaled.c | 142 +++++++++++++++
 libswscale/aarch64/yuv2rgb_neon.S     | 332 ++++++++++++++++++++++++++++++++++
 libswscale/swscale_internal.h         |   1 +
 libswscale/swscale_unscaled.c         |   2 +
 5 files changed, 480 insertions(+)
 create mode 100644 libswscale/aarch64/Makefile
 create mode 100644 libswscale/aarch64/swscale_unscaled.c
 create mode 100644 libswscale/aarch64/yuv2rgb_neon.S

diff --git a/libswscale/aarch64/Makefile b/libswscale/aarch64/Makefile
new file mode 100644
index 0000000..823806e
--- /dev/null
+++ b/libswscale/aarch64/Makefile
@@ -0,0 +1,3 @@
+OBJS        += aarch64/swscale_unscaled.o
+
+NEON-OBJS   += aarch64/yuv2rgb_neon.o
diff --git a/libswscale/aarch64/swscale_unscaled.c b/libswscale/aarch64/swscale_unscaled.c
new file mode 100644
index 0000000..8b04c2b
--- /dev/null
+++ b/libswscale/aarch64/swscale_unscaled.c
@@ -0,0 +1,142 @@
+/*
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "config.h"
+#include "libswscale/swscale.h"
+#include "libswscale/swscale_internal.h"
+#include "libavutil/aarch64/cpu.h"
+
+#define YUV_TO_RGB_TABLE(precision)                                                         \
+        c->yuv2rgb_v2r_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
+        c->yuv2rgb_u2g_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
+        c->yuv2rgb_v2g_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
+        c->yuv2rgb_u2b_coeff / ((precision) == 16 ? 1 << 7 : 1),                            \
+
+#define DECLARE_FF_YUVX_TO_RGBX_FUNCS(ifmt, ofmt, precision)                                \
+int ff_##ifmt##_to_##ofmt##_neon_##precision(int w, int h,                                  \
+                                 uint8_t *dst, int linesize,                                \
+                                 const uint8_t *srcY, int linesizeY,                        \
+                                 const uint8_t *srcU, int linesizeU,                        \
+                                 const uint8_t *srcV, int linesizeV,                        \
+                                 const int16_t *table,                                      \
+                                 int y_offset,                                              \
+                                 int y_coeff);                                              \
+                                                                                            \
+static int ifmt##_to_##ofmt##_neon_wrapper_##precision(SwsContext *c, const uint8_t *src[], \
+                                           int srcStride[], int srcSliceY, int srcSliceH,   \
+                                           uint8_t *dst[], int dstStride[]) {               \
+    const int16_t yuv2rgb_table[] = { YUV_TO_RGB_TABLE(precision) };                        \
+                                                                                            \
+    ff_##ifmt##_to_##ofmt##_neon_##precision(c->srcW, srcSliceH,                            \
+                                 dst[0] + srcSliceY * dstStride[0], dstStride[0],           \
+                                 src[0], srcStride[0],                                      \
+                                 src[1], srcStride[1],                                      \
+                                 src[2], srcStride[2],                                      \
+                                 yuv2rgb_table,                                             \
+                                 c->yuv2rgb_y_offset >> 9,                                  \
+                                 c->yuv2rgb_y_coeff / ((precision) == 16 ? 1 << 7 : 1));    \
+                                                                                            \
+    return 0;                                                                               \
+}                                                                                           \
+
+#define DECLARE_FF_YUVX_TO_ALL_RGBX_FUNCS(yuvx, precision)                                  \
+DECLARE_FF_YUVX_TO_RGBX_FUNCS(yuvx, argb, precision)                                        \
+DECLARE_FF_YUVX_TO_RGBX_FUNCS(yuvx, rgba, precision)                                        \
+DECLARE_FF_YUVX_TO_RGBX_FUNCS(yuvx, abgr, precision)                                        \
+DECLARE_FF_YUVX_TO_RGBX_FUNCS(yuvx, bgra, precision)                                        \
+
+#define DECLARE_FF_YUVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(yuvx)                               \
+DECLARE_FF_YUVX_TO_ALL_RGBX_FUNCS(yuvx, 16)                                                 \
+//DECLARE_FF_YUVX_TO_ALL_RGBX_FUNCS(yuvx, 32)                                                 \
+
+DECLARE_FF_YUVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(yuv420p)
+DECLARE_FF_YUVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(yuv422p)
+
+#define DECLARE_FF_NVX_TO_RGBX_FUNCS(ifmt, ofmt, precision)                                 \
+int ff_##ifmt##_to_##ofmt##_neon_##precision(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_##precision(SwsContext *c, const uint8_t *src[], \
+                                           int srcStride[], int srcSliceY, int srcSliceH,   \
+                                           uint8_t *dst[], int dstStride[]) {               \
+    const int16_t yuv2rgb_table[] = { YUV_TO_RGB_TABLE(precision) };                        \
+                                                                                            \
+    ff_##ifmt##_to_##ofmt##_neon_##precision(c->srcW, srcSliceH,                            \
+                                 dst[0] + srcSliceY * dstStride[0], dstStride[0],           \
+                                 src[0], srcStride[0], src[1], srcStride[1],                \
+                                 yuv2rgb_table,                                             \
+                                 c->yuv2rgb_y_offset >> 9,                                  \
+                                 c->yuv2rgb_y_coeff / ((precision) == 16 ? 1 << 7 : 1));    \
+                                                                                            \
+    return 0;                                                                               \
+}                                                                                           \
+
+#define DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx, precision)                                    \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, argb, precision)                                          \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, rgba, precision)                                          \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, abgr, precision)                                          \
+DECLARE_FF_NVX_TO_RGBX_FUNCS(nvx, bgra, precision)                                          \
+
+#define DECLARE_FF_NVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(nvx)                                 \
+DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx, 16)                                                   \
+//DECLARE_FF_NVX_TO_ALL_RGBX_FUNCS(nvx, 32)                                                   \
+
+DECLARE_FF_NVX_TO_ALL_RGBX_ALL_PRECISION_FUNCS(nv12)
+DECLARE_FF_NVX_TO_ALL_RGBX_ALL_PRECISION_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, accurate_rnd) do {                  \
+    if (c->srcFormat == AV_PIX_FMT_##IFMT                                                   \
+        && c->dstFormat == AV_PIX_FMT_##OFMT                                                \
+        && !(c->srcH & 1)                                                                   \
+        && !(c->srcW & 15)                                                                  \
+        && !accurate_rnd) {                                                                 \
+        c->swscale = ifmt##_to_##ofmt##_neon_wrapper_16;                                    \
+    }                                                                                       \
+} while (0)
+
+#define SET_FF_NVX_TO_ALL_RGBX_FUNC(nvx, NVX, accurate_rnd) do {                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, argb, ARGB, accurate_rnd);                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, rgba, RGBA, accurate_rnd);                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, abgr, ABGR, accurate_rnd);                            \
+    SET_FF_NVX_TO_RGBX_FUNC(nvx, NVX, bgra, BGRA, accurate_rnd);                            \
+} while (0)
+
+static void get_unscaled_swscale_neon(SwsContext *c) {
+    int accurate_rnd = c->flags & SWS_ACCURATE_RND;
+
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv12, NV12, accurate_rnd);
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(nv21, NV21, accurate_rnd);
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(yuv420p, YUV420P, accurate_rnd);
+    SET_FF_NVX_TO_ALL_RGBX_FUNC(yuv422p, YUV422P, accurate_rnd);
+}
+
+void ff_get_unscaled_swscale_aarch64(SwsContext *c)
+{
+    int cpu_flags = av_get_cpu_flags();
+    if (have_neon(cpu_flags))
+        get_unscaled_swscale_neon(c);
+}
diff --git a/libswscale/aarch64/yuv2rgb_neon.S b/libswscale/aarch64/yuv2rgb_neon.S
new file mode 100644
index 0000000..360168a
--- /dev/null
+++ b/libswscale/aarch64/yuv2rgb_neon.S
@@ -0,0 +1,332 @@
+/*
+ * Copyright (c) 2015 Matthieu Bouron <matthieu.bouron stupeflix.com>
+ * Copyright (c) 2015 Clément Bœsch <clement stupeflix.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/aarch64/asm.S"
+
+.macro compute_premult_16 u v
+    zip1                v4.8H, \u, \u
+    zip2                v5.8H, \u, \u
+    zip1                v6.8H, \v, \v
+    zip2                v7.8H, \v, \v
+
+    mul                 v20.8H, v6.8H, v1.H[0]                          //  V * v2r             (left,  red)
+    mul                 v21.8H, v7.8H, v1.H[0]                          //  V * v2r             (right, red)
+    mul                 v22.8H, v4.8H, v1.H[1]                          //  U * u2g
+    mul                 v23.8H, v5.8H, v1.H[1]                          //  U * u2g
+    mla                 v22.8H, v6.8H, v1.H[2]                          //  U * u2g + V * v2g   (left,  green)
+    mla                 v23.8H, v7.8H, v1.H[2]                          //  U * u2g + V * v2g   (right, green)
+    mul                 v24.8H, v4.8H, v1.H[3]                          //  U * u2b             (left,  blue)
+    mul                 v25.8H, v5.8H, v1.H[3]                          //  U * u2b             (right, blue)
+.endm
+
+.macro compute_premult_32 part u v
+    zip\part            v4.8H, \u, \u
+    zip\part            v5.8H, \v, \v
+
+    smull               v20.4S, v5.4H, v1.H[0]                          //  V * v2r             (left,  red)
+    smull2              v21.4S, v5.8H, v1.H[0]                          //  V * v2r             (right, red)
+    smull               v22.4S, v4.4H, v1.H[1]                          //  U * u2g
+    smull2              v23.4S, v4.8H, v1.H[1]                          //  U * u2g
+    smlal               v22.4S, v5.4H, v1.H[2]                          //  U * u2g + V * v2g   (left,  green)
+    smlal2              v23.4S, v5.8H, v1.H[2]                          //  U * u2g + V * v2g   (right, green)
+    smull               v24.4S, v4.4H, v1.H[3]                          //  U * u2b             (left,  blue)
+    smull2              v25.4S, v4.8H, v1.H[3]                          //  U * u2b             (right, blue)
+.endm
+
+.macro compute_color_16 dst_comp1 dst_comp2 pre1 pre2
+    add                 v28.8H, v26.8H, \pre1
+    add                 v29.8H, v27.8H, \pre2
+    sqrshrun            \dst_comp1, v28.8H, #6
+    sqrshrun            \dst_comp2, v29.8H, #6
+.endm
+
+.macro compute_color_32 dst_comp pre1 pre2
+    add                 v28.4S, v26.4S, \pre1
+    add                 v29.4S, v27.4S, \pre2
+    sqrshrun            v16.4H, v28.4S, #13
+    sqrshrun2           v16.8H, v29.4S, #13
+    uqxtn               \dst_comp, v16.8H
+.endm
+
+.macro compute_rgba_16 r1 g1 b1 a1 r2 g2 b2 a2
+    compute_color_16    \r1, \r2, v20.8H, v21.8H
+    compute_color_16    \g1, \g2, v22.8H, v23.8H
+    compute_color_16    \b1, \b2, v24.8H, v25.8H
+    movi                \a1, #255
+    movi                \a2, #255
+.endm
+
+.macro compute_rgba_32 r g b a
+    compute_color_32    \r, v20.4S, v21.4S
+    compute_color_32    \g, v22.4S, v23.4S
+    compute_color_32    \b, v24.4S, v25.4S
+    movi                \a, #255
+.endm
+
+.macro compute_16px_16 dst y ofmt
+    uxtl                v26.8H, \y\().8B
+    uxtl2               v27.8H, \y\().16B
+
+    sub                 v26.8H, v26.8H, v3.8H
+    sub                 v27.8H, v27.8H, v3.8H
+
+    mul                 v26.8H, v26.8H, v0.8H
+    mul                 v27.8H, v27.8H, v0.8H
+
+.ifc \ofmt,argb // 1 2 3 0
+    compute_rgba_16     v5.8B,v6.8B,v7.8B,v4.8B, v17.8B,v18.8B,v19.8B,v16.8B
+.endif
+
+.ifc \ofmt,rgba // 0 1 2 3
+    compute_rgba_16     v4.8B,v5.8B,v6.8B,v7.8B, v16.8B,v17.8B,v18.8B,v19.8B
+.endif
+
+.ifc \ofmt,abgr // 3 2 1 0
+    compute_rgba_16     v7.8B,v6.8B,v5.8B,v4.8B, v19.8B,v18.8B,v17.8B,v16.8B
+.endif
+
+.ifc \ofmt,bgra // 2 1 0 3
+    compute_rgba_16     v6.8B,v5.8B,v4.8B,v7.8B, v18.8B,v17.8B,v16.8B,v19.8B
+.endif
+
+    st4                 { v4.8B, v5.8B, v6.8B, v7.8B}, [\dst], #32
+    st4                 {v16.8B,v17.8B,v18.8B,v19.8B}, [\dst], #32
+.endm
+
+.macro compute_8px_32 dst half_y ofmt
+    sub                 \half_y\().8H, \half_y\().8H, v3.8H
+    smull               v26.4S, \half_y\().4H, v0.4H                    // q1 = (srcY - y_offset) * y_coeff (left)
+    smull2              v27.4S, \half_y\().8H, v0.8H                    // q2 = (srcY - y_offset) * y_coeff (right)
+
+.ifc \ofmt,argb // 1 2 3 0
+    compute_rgba_32     v5.8B,v6.8B,v7.8B,v4.8B
+.endif
+
+.ifc \ofmt,rgba // 0 1 2 3
+    compute_rgba_32     v4.8B,v5.8B,v6.8B,v7.8B
+.endif
+
+.ifc \ofmt,abgr // 3 2 1 0
+    compute_rgba_32     v7.8B,v6.8B,v5.8B,v4.8B
+.endif
+
+.ifc \ofmt,bgra // 2 1 0 3
+    compute_rgba_32     v6.8B,v5.8B,v4.8B,v7.8B
+.endif
+
+    st4                 {v4.8B, v5.8B, v6.8B, v7.8B}, [\dst], #32
+.endm
+
+.macro process_1l_16px_16 ofmt
+    compute_premult_16  v18.8H, v19.8H
+    ld1                 {v2.16B}, [x4], #16
+    compute_16px_16     x2, v2, \ofmt
+.endm
+
+.macro process_1l_16px_32 ofmt
+    compute_premult_32  1, v18.8H, v19.8H
+    ld1                 {v30.16B}, [x4], #16
+    uxtl                v2.8H, v30.8B
+    compute_8px_32      x2, v2, \ofmt
+
+    compute_premult_32  2, v18.8H, v19.8H
+    uxtl2               v2.8H, v30.16B
+    compute_8px_32      x2,  v2, \ofmt
+.endm
+
+.macro process_2l_16px_16 ofmt
+    compute_premult_16  v18.8H, v19.8H
+
+    ld1                 {v2.16B}, [x4], #16                             // first line of luma
+    compute_16px_16     x2, v2, \ofmt
+
+    ld1                 {v2.16B}, [x12], #16                            // second line of luma
+    compute_16px_16     x11, v2, \ofmt
+.endm
+
+.macro process_2l_16px_32 ofmt
+    compute_premult_32  1, v18.8H, v19.8H
+
+    ld1                 {v30.16B}, [x4], #16                            // first line of luma
+
+    uxtl                v2.8H, v30.8B
+    compute_8px_32      x2, v2, \ofmt
+
+    ld1                 {v31.16B}, [x12], #16                           // second line of luma
+    uxtl                v2.8H, v31.8B
+    compute_8px_32      x11, v2, \ofmt
+
+    compute_premult_32  2, v18.8H, v19.8H
+    uxtl2               v2.8H, v30.16B
+    compute_8px_32      x2,  v2, \ofmt
+    uxtl2               v2.8H, v31.16B
+    compute_8px_32      x11, v2, \ofmt
+.endm
+
+.macro load_args_nv12
+    ldr                 x8,  [sp]                                       // table
+    ldr                 w9,  [sp, #8]                                   // y_offset
+    ldr                 w10, [sp, #16]                                  // y_coeff
+    ld1                 {v1.1D}, [x8]
+    dup                 v0.8H, w10
+    dup                 v3.8H, w9
+    add                 x11, x2, w3, UXTW                               // w11 = dst + linesize (dst2)
+    add                 x12, x4, w5, UXTW                               // w12 = srcY + linesizeY (srcY2)
+    lsl                 w3, w3, #1
+    lsl                 w5, w5, #1
+    lsl                 w8, w0, #2
+    sub                 w3, w3, w8                                      // w3 = linesize  * 2 - width * 4 (padding)
+    sub                 w5, w5, w0                                      // w5 = linesizeY * 2 - width     (paddingY)
+    sub                 w7, w7, w0                                      // w7 = linesizeC     - width     (paddingC)
+.endm
+
+.macro load_args_nv21
+    load_args_nv12
+.endm
+
+.macro load_args_yuv420p
+    ldr                 x13, [sp]                                       // srcV
+    ldr                 w14, [sp, #8]                                   // linesizeV
+    ldr                 x8,  [sp, #16]                                  // table
+    ldr                 w9,  [sp, #24]                                  // y_offset
+    ldr                 w10, [sp, #32]                                  // y_coeff
+    ld1                 {v1.1D}, [x8]
+    dup                 v0.8H, w10
+    dup                 v3.8H, w9
+    add                 x11, x2, w3, UXTW                               // w11 = dst + linesize (dst2)
+    add                 x12, x4, w5, UXTW                               // w12 = srcY + linesizeY (srcY2)
+    lsl                 w3, w3, #1
+    lsl                 w5, w5, #1
+    lsl                 w8, w0, #2
+    sub                 w3, w3, w8                                      // w3 = linesize  * 2 - width * 4 (padding)
+    sub                 w5, w5, w0                                      // w5 = linesizeY * 2 - width     (paddingY)
+    sub                 w7,  w7,  w0, lsr #1                            // w6  = linesizeU - width / 2 (paddingU)
+    sub                 w14, w14, w0, lsr #1                            // w14 = linesizeV - width / 2 (paddingV)
+.endm
+
+.macro load_args_yuv422p
+    ldr                 x13, [sp]                                       // srcV
+    ldr                 w14, [sp, #8]                                   // linesizeV
+    ldr                 x8,  [sp, #16]                                  // table
+    ldr                 w9,  [sp, #24]                                  // y_offset
+    ldr                 w10, [sp, #32]                                  // y_coeff
+    ld1                 {v1.1D}, [x8]
+    dup                 v0.8H, w10
+    dup                 v3.8H, w9
+    add                 x12, x4, w5, UXTW                               // w12 = srcY + linesizeY (srcY2)
+    lsl                 w8, w0, #2
+    sub                 w3, w3, w8                                      // w3 = linesize  - width * 4 (padding)
+    sub                 w5, w5, w0                                      // w5 = linesizeY - width     (paddingY)
+    sub                 w7,  w7,  w0, lsr #1                            // w6  = linesizeU - width / 2 (paddingU)
+    sub                 w14, w14, w0, lsr #1                            // w14 = linesizeV - width / 2 (paddingV)
+.endm
+
+.macro process_nv12 ofmt precision
+    ld2                 {v16.8B, v17.8B}, [x6], #16
+    usubl               v18.8H, v16.8B, v5.8B
+    usubl               v19.8H, v17.8B, v5.8B
+    process_2l_16px_\precision \ofmt
+.endm
+
+.macro process_nv21 ofmt precision
+    ld2                 {v16.8B, v17.8B}, [x6], #16
+    usubl               v19.8H, v16.8B, v5.8B
+    usubl               v18.8H, v17.8B, v5.8B
+    process_2l_16px_\precision \ofmt
+.endm
+
+.macro process_yuv420p ofmt precision
+    ld1                 {v16.8B}, [ x6], #8
+    ld1                 {v17.8B}, [x13], #8
+    usubl               v18.8H, v16.8B, v5.8B
+    usubl               v19.8H, v17.8B, v5.8B
+    process_2l_16px_\precision \ofmt
+.endm
+
+.macro process_yuv422p ofmt precision
+    ld1                 {v16.8B}, [ x6], #8
+    ld1                 {v17.8B}, [x13], #8
+    usubl               v18.8H, v16.8B, v5.8B
+    usubl               v19.8H, v17.8B, v5.8B
+    process_1l_16px_\precision \ofmt
+.endm
+
+.macro increment_nv12
+    add                 x11, x11, w3, UXTW                              // dst2  += padding
+    add                 x12, x12, w5, UXTW                              // srcY2 += paddingY
+    add                 x6, x6, w7, UXTW                                // srcC  += paddingC
+    subs                w1, w1, #2                                      // height -= 2
+.endm
+
+.macro increment_nv21
+    add                 x11, x11, w3, UXTW                              // dst2  += padding
+    add                 x12, x12, w5, UXTW                              // srcY2 += paddingY
+    add                 x6, x6, w7, UXTW                                // srcC  += paddingC
+    subs                w1, w1, #2                                      // height -= 2
+.endm
+
+.macro increment_yuv420p
+    add                 x11, x11, w3, UXTW                              // dst2  += padding
+    add                 x12, x12, w5, UXTW                              // srcY2 += paddingY
+    add                 x6,  x6,  w7, UXTW                              // srcU  += paddingU
+    add                 x13, x13, w14, UXTW                             // srcV  += paddingV
+    subs                w1, w1, #2                                      // height -= 2
+.endm
+
+.macro increment_yuv422p
+    add                 x6,  x6,  w7, UXTW                              // srcU  += paddingU
+    add                 x13, x13, w14, UXTW                             // srcV  += paddingV
+    subs                w1, w1, #1                                      // height -= 1
+.endm
+
+.macro declare_func ifmt ofmt precision
+function ff_\ifmt\()_to_\ofmt\()_neon_\precision\(), export=1
+    load_args_\ifmt
+1:
+    mov                 w8, w0                                          // w8 = width
+2:
+    movi                v5.8B, #128
+    process_\ifmt       \ofmt \precision
+    subs                w8, w8, #16                                     // width -= 16
+    b.gt                2b
+    add                 x2, x2, w3, UXTW                                // dst   += padding
+    add                 x4, x4, w5, UXTW                                // srcY  += paddingY
+    increment_\ifmt
+    b.gt                1b
+    ret
+endfunc
+.endm
+
+.macro declare_rgb_funcs ifmt precision
+    declare_func \ifmt, argb, \precision
+    declare_func \ifmt, rgba, \precision
+    declare_func \ifmt, abgr, \precision
+    declare_func \ifmt, bgra, \precision
+.endm
+
+declare_rgb_funcs nv12, 16
+declare_rgb_funcs nv21, 16
+declare_rgb_funcs nv12, 32
+declare_rgb_funcs nv21, 32
+declare_rgb_funcs yuv420p, 16
+declare_rgb_funcs yuv420p, 32
+declare_rgb_funcs yuv422p, 16
+declare_rgb_funcs yuv422p, 32
diff --git a/libswscale/swscale_internal.h b/libswscale/swscale_internal.h
index 1e29ec3..f0bab78 100644
--- a/libswscale/swscale_internal.h
+++ b/libswscale/swscale_internal.h
@@ -877,6 +877,7 @@ extern const AVClass ff_sws_context_class;
 void ff_get_unscaled_swscale(SwsContext *c);
 void ff_get_unscaled_swscale_ppc(SwsContext *c);
 void ff_get_unscaled_swscale_arm(SwsContext *c);
+void ff_get_unscaled_swscale_aarch64(SwsContext *c);
 
 /**
  * Return function pointer to fastest main scaler path function depending
diff --git a/libswscale/swscale_unscaled.c b/libswscale/swscale_unscaled.c
index 74f3467..4bf2447 100644
--- a/libswscale/swscale_unscaled.c
+++ b/libswscale/swscale_unscaled.c
@@ -1776,6 +1776,8 @@ void ff_get_unscaled_swscale(SwsContext *c)
         ff_get_unscaled_swscale_ppc(c);
      if (ARCH_ARM)
          ff_get_unscaled_swscale_arm(c);
+    if (ARCH_AARCH64)
+        ff_get_unscaled_swscale_aarch64(c);
 }
 
 /* Convert the palette to the same packed 32-bit format as the palette */
-- 
2.7.1



More information about the ffmpeg-devel mailing list