swscale/arm: add yuv2planeX_8_neon
authorMatthieu Bouron <matthieu.bouron@stupeflix.com>
Fri, 8 Apr 2016 15:32:24 +0000 (15:32 +0000)
committerMatthieu Bouron <matthieu.bouron@stupeflix.com>
Mon, 11 Apr 2016 16:13:36 +0000 (18:13 +0200)
libswscale/arm/Makefile
libswscale/arm/output.S [new file with mode: 0644]
libswscale/arm/swscale.c

index b8b0134..792da6b 100644 (file)
@@ -4,4 +4,5 @@ OBJS        += arm/swscale.o                    \
 NEON-OBJS   += arm/rgb2yuv_neon_32.o
 NEON-OBJS   += arm/rgb2yuv_neon_16.o
 NEON-OBJS   += arm/hscale.o                     \
+               arm/output.o                     \
                arm/yuv2rgb_neon.o               \
diff --git a/libswscale/arm/output.S b/libswscale/arm/output.S
new file mode 100644 (file)
index 0000000..70846de
--- /dev/null
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 2016 Clément Bœsch <clement stupeflix.com>
+ * Copyright (c) 2016 Matthieu Bouron <matthieu.bouron 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/arm/asm.S"
+
+function ff_yuv2planeX_8_neon, export=1
+    push {r4-r12, lr}
+    vpush {q4-q7}
+    ldr                 r4, [sp, #104]                                 @ dstW
+    ldr                 r5, [sp, #108]                                 @ dither
+    ldr                 r6, [sp, #112]                                 @ offset
+    vld1.8              {d0}, [r5]                                     @ load 8x8-bit dither values
+    cmp                 r6, #0                                         @ check offsetting which can be 0 or 3 only
+    beq                 1f
+    vext.u8             d0, d0, d0, #3                                 @ honor offseting which can be 3 only
+1:  vmovl.u8            q0, d0                                         @ extend dither to 16-bit
+    vshll.u16           q1, d0, #12                                    @ extend dither to 32-bit with left shift by 12 (part 1)
+    vshll.u16           q2, d1, #12                                    @ extend dither to 32-bit with left shift by 12 (part 2)
+    mov                 r7, #0                                         @ i = 0
+2:  vmov.u8             q3, q1                                         @ initialize accumulator with dithering values (part 1)
+    vmov.u8             q4, q2                                         @ initialize accumulator with dithering values (part 2)
+    mov                 r8, r1                                         @ tmpFilterSize = filterSize
+    mov                 r9, r2                                         @ srcp
+    mov                 r10, r0                                        @ filterp
+3:  ldr                 r11, [r9], #4                                  @ get pointer @ src[j]
+    ldr                 r12, [r9], #4                                  @ get pointer @ src[j+1]
+    add                 r11, r11, r7, lsl #1                           @ &src[j][i]
+    add                 r12, r12, r7, lsl #1                           @ &src[j+1][i]
+    vld1.16             {q5}, [r11]                                    @ read 8x16-bit @ src[j  ][i + {0..7}]: A,B,C,D,E,F,G,H
+    vld1.16             {q6}, [r12]                                    @ read 8x16-bit @ src[j+1][i + {0..7}]: I,J,K,L,M,N,O,P
+    ldr                 r11, [r10], #4                                 @ read 2x16-bit coeffs (X, Y) at (filter[j], filter[j+1])
+    vmov.16             q7, q5                                         @ copy 8x16-bit @ src[j  ][i + {0..7}] for following inplace zip instruction
+    vmov.16             q8, q6                                         @ copy 8x16-bit @ src[j+1][i + {0..7}] for following inplace zip instruction
+    vzip.16             q7, q8                                         @ A,I,B,J,C,K,D,L,E,M,F,N,G,O,H,P
+    vdup.32             q15, r11                                       @ X,Y,X,Y,X,Y,X,Y
+    vmull.s16           q9, d14, d30                                   @ A*X,I*Y,B*X,J*Y
+    vmull.s16           q10, d15, d31                                  @ C*X,K*Y,D*X,L*Y
+    vmull.s16           q11, d16, d30                                  @ E*X,M*Y,F*X,N*Y
+    vmull.s16           q12, d17, d31                                  @ G*X,O*Y,H*X,P*Y
+    vpadd.s32           d10, d18, d19                                  @ A*X+I*Y,B*X+J*Y
+    vpadd.s32           d11, d20, d21                                  @ C*X+K*Y,D*X+L*Y
+    vpadd.s32           d12, d22, d23                                  @ E*X+M*Y,F*X+N*Y
+    vpadd.s32           d13, d24, d25                                  @ G*X+O*Y,H*X+P*Y
+    vadd.s32            q3, q5                                         @ update val accumulator (part 1)
+    vadd.s32            q4, q6                                         @ update val accumulator (part 2)
+    subs                r8, #2                                         @ tmpFilterSize -= 2
+    bgt                 3b                                             @ loop until filterSize is consumed
+    vshr.s32            q3, q3, #19                                    @ val>>19 (part 1)
+    vshr.s32            q4, q4, #19                                    @ val>>19 (part 2)
+    vqmovun.s32         d6, q3                                         @ clip16(val>>19) (part 1)
+    vqmovun.s32         d7, q4                                         @ clip16(val>>19) (part 2)
+    vqmovn.u16          d6, q3                                         @ merge part 1 and part 2
+    vst1.8              {d6}, [r3]!                                    @ write destination
+    add                 r7, #8                                         @ i += 8
+    subs                r4, r4, #8                                     @ dstW -= 8
+    bgt                 2b                                             @ loop until width is consumed
+    vpop                {q4-q7}
+    pop                 {r4-r12, lr}
+    mov                 pc, lr
+endfunc
index 6301ac5..1ec360f 100644 (file)
@@ -25,6 +25,10 @@ void ff_hscale_8_to_15_neon(SwsContext *c, int16_t *dst, int dstW,
                             const uint8_t *src, const int16_t *filter,
                             const int32_t *filterPos, int filterSize);
 
+void ff_yuv2planeX_8_neon(const int16_t *filter, int filterSize,
+                          const int16_t **src, uint8_t *dest, int dstW,
+                          const uint8_t *dither, int offset);
+
 av_cold void ff_sws_init_swscale_arm(SwsContext *c)
 {
     int cpu_flags = av_get_cpu_flags();
@@ -33,5 +37,8 @@ av_cold void ff_sws_init_swscale_arm(SwsContext *c)
         if (c->srcBpc == 8 && c->dstBpc <= 14) {
             c->hyScale = c->hcScale = ff_hscale_8_to_15_neon;
         }
+        if (c->dstBpc == 8) {
+            c->yuv2planeX = ff_yuv2planeX_8_neon;
+        }
     }
 }