swscale/arm/yuv2rgb: save a few instructions by processing the luma line interleaved
authorMatthieu Bouron <matthieu.bouron@gmail.com>
Mon, 28 Mar 2016 11:33:38 +0000 (11:33 +0000)
committerMatthieu Bouron <matthieu.bouron@stupeflix.com>
Fri, 1 Apr 2016 15:23:45 +0000 (17:23 +0200)
libswscale/arm/yuv2rgb_neon.S

index ee11e08..350e360 100644 (file)
 #include "libavutil/arm/asm.S"
 
 
-.macro compute_premult half_u1, half_u2, half_v1, half_v2
-    vmov                d2, \half_u1                                   @ copy left q14 to left q1
-    vmov                d3, \half_u1                                   @ copy left q14 to right q1
-    vmov                d4, \half_u2                                   @ copy right q14 to left q2
-    vmov                d5, \half_u2                                   @ copy right q14 to right q2
-
-    vmov                d6, \half_v1                                   @ copy left q15 to left q3
-    vmov                d7, \half_v1                                   @ copy left q15 to right q3
-    vmov                d8, \half_v2                                   @ copy right q15 to left q4
-    vmov                d9, \half_v2                                   @ copy right q15 to right q4
-
-    vzip.16             d2, d3                                         @ U1U1U2U2U3U3U4U4
-    vzip.16             d4, d5                                         @ U5U5U6U6U7U7U8U8
-
-    vzip.16             d6, d7                                         @ V1V1V2V2V3V3V4V4
-    vzip.16             d8, d9                                         @ V5V5V6V6V7V7V8V8
-
-    vmul.s16            q8,  q3, d1[0]                                 @  V * v2r             (left,  red)
-    vmul.s16            q9,  q4, d1[0]                                 @  V * v2r             (right, red)
-    vmul.s16            q10, q1, d1[1]                                 @  U * u2g
-    vmul.s16            q11, q2, d1[1]                                 @  U * u2g
-    vmla.s16            q10, q3, d1[2]                                 @  U * u2g + V * v2g   (left,  green)
-    vmla.s16            q11, q4, d1[2]                                 @  U * u2g + V * v2g   (right, green)
-    vmul.s16            q12, q1, d1[3]                                 @  U * u2b             (left,  blue)
-    vmul.s16            q13, q2, d1[3]                                 @  U * u2b             (right, blue)
+.macro compute_premult
+    vmul.s16            q8, q15, d1[0]                                 @ q8  = V * v2r
+    vmul.s16            q9, q14, d1[1]                                 @ q9  = U * u2g
+    vmla.s16            q9, q15, d1[2]                                 @ q9  = U * u2g + V * v2g
+    vmul.s16            q10,q14, d1[3]                                 @ q10 = U * u2b
 .endm
 
-.macro compute_color dst_comp1 dst_comp2 pre1 pre2
-    vadd.s16            q1, q14, \pre1
-    vadd.s16            q2, q15, \pre2
+.macro compute_color dst_comp1 dst_comp2 pre
+    vadd.s16            q1, q14, \pre
+    vadd.s16            q2, q15, \pre
     vqrshrun.s16        \dst_comp1, q1, #6
     vqrshrun.s16        \dst_comp2, q2, #6
 .endm
 
 .macro compute_rgba r1 g1 b1 a1 r2 g2 b2 a2
-    compute_color       \r1, \r2, q8, q9
-    compute_color       \g1, \g2, q10, q11
-    compute_color       \b1, \b2, q12, q13
+    compute_color       \r1, \r2, q8
+    compute_color       \g1, \g2, q9
+    compute_color       \b1, \b2, q10
     vmov.u8             \a1, #255
     vmov.u8             \a2, #255
 .endm
 
-.macro compute dst y0 y1 ofmt
-    vmovl.u8            q14, \y0                                       @ 8px of y
-    vmovl.u8            q15, \y1                                       @ 8px of y
-
-    vdup.16             q5, r9                                         @ q5  = y_offset
-    vmov                d14, d0                                        @ q7  = y_coeff
-    vmov                d15, d0                                        @ q7  = y_coeff
-
-    vsub.s16            q14, q5
-    vsub.s16            q15, q5
-
-    vmul.s16            q14, q7                                        @ q14 = (srcY - y_offset) * y_coeff (left)
-    vmul.s16            q15, q7                                        @ q15 = (srcY - y_offset) * y_coeff (right)
-
+.macro compute dst ofmt
+    vmovl.u8            q14, d14                                       @ q14 = Y
+    vmovl.u8            q15, d15                                       @ q15 = Y
+    vsub.s16            q14, q12                                       @ q14 = (Y - y_offset)
+    vsub.s16            q15, q12                                       @ q15 = (Y - y_offset)
+    vmul.s16            q14, q13                                       @ q14 = (Y - y_offset) * y_coeff
+    vmul.s16            q15, q13                                       @ q15 = (Y - y_offset) * y_coeff
 
 .ifc \ofmt,argb
     compute_rgba        d7, d8, d9, d6, d11, d12, d13, d10
 .ifc \ofmt,bgra
     compute_rgba        d8, d7, d6, d9, d12, d11, d10, d13
 .endif
+
+    vzip.8              d6, d10                                        @ d6 = R1R2R3R4R5R6R7R8 d10 = R9R10R11R12R13R14R15R16
+    vzip.8              d7, d11                                        @ d7 = G1G2G3G4G5G6G7G8 d11 = G9G10G11G12G13G14G15G16
+    vzip.8              d8, d12                                        @ d8 = B1B2B3B4B5B6B7B8 d12 = B9B10B11B12B13B14B15B16
+    vzip.8              d9, d13                                        @ d9 = A1A2A3A4A5A6A7A8 d13 = A9A10A11A12A13A14A15A16
     vst4.8              {q3, q4}, [\dst,:128]!
     vst4.8              {q5, q6}, [\dst,:128]!
-
 .endm
 
 .macro process_1l_internal dst src ofmt
-    vld1.8              {q7}, [\src]!
-    compute             \dst, d14, d15, \ofmt
+    vld2.8              {d14, d15}, [\src]!                            @ q7 = Y (interleaved)
+    compute             \dst, \ofmt
 .endm
 
 .macro process_1l ofmt
-    compute_premult     d28, d29, d30, d31
+    compute_premult
     process_1l_internal r2, r4, \ofmt
 .endm
 
 .macro process_2l ofmt
-    compute_premult     d28, d29, d30, d31
+    compute_premult
     process_1l_internal r2, r4, \ofmt
     process_1l_internal r11,r12,\ofmt
 .endm
 .macro declare_func ifmt ofmt
 function ff_\ifmt\()_to_\ofmt\()_neon, export=1
     load_args_\ifmt
+    vdup.16             q12, r9                                        @ q12 = y_offset
+    vmov                d26, d0                                        @ q13 = y_coeff
+    vmov                d27, d0                                        @ q13 = y_coeff
 1:
     mov                 r8, r0                                         @ r8 = width
 2: