1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/common/arm/neon/vp9_convolve8_neon.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,280 @@
1.4 +;
1.5 +; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
1.6 +;
1.7 +; Use of this source code is governed by a BSD-style license
1.8 +; that can be found in the LICENSE file in the root of the source
1.9 +; tree. An additional intellectual property rights grant can be found
1.10 +; in the file PATENTS. All contributing project authors may
1.11 +; be found in the AUTHORS file in the root of the source tree.
1.12 +;
1.13 +
1.14 +
1.15 + ; These functions are only valid when:
1.16 + ; x_step_q4 == 16
1.17 + ; w%4 == 0
1.18 + ; h%4 == 0
1.19 + ; taps == 8
1.20 + ; VP9_FILTER_WEIGHT == 128
1.21 + ; VP9_FILTER_SHIFT == 7
1.22 +
1.23 + EXPORT |vp9_convolve8_horiz_neon|
1.24 + EXPORT |vp9_convolve8_vert_neon|
1.25 + IMPORT |vp9_convolve8_horiz_c|
1.26 + IMPORT |vp9_convolve8_vert_c|
1.27 + ARM
1.28 + REQUIRE8
1.29 + PRESERVE8
1.30 +
1.31 + AREA ||.text||, CODE, READONLY, ALIGN=2
1.32 +
1.33 + ; Multiply and accumulate by q0
1.34 + MACRO
1.35 + MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
1.36 + vmull.s16 $dst, $src0, d0[0]
1.37 + vmlal.s16 $dst, $src1, d0[1]
1.38 + vmlal.s16 $dst, $src2, d0[2]
1.39 + vmlal.s16 $dst, $src3, d0[3]
1.40 + vmlal.s16 $dst, $src4, d1[0]
1.41 + vmlal.s16 $dst, $src5, d1[1]
1.42 + vmlal.s16 $dst, $src6, d1[2]
1.43 + vmlal.s16 $dst, $src7, d1[3]
1.44 + MEND
1.45 +
1.46 +; r0 const uint8_t *src
1.47 +; r1 int src_stride
1.48 +; r2 uint8_t *dst
1.49 +; r3 int dst_stride
1.50 +; sp[]const int16_t *filter_x
1.51 +; sp[]int x_step_q4
1.52 +; sp[]const int16_t *filter_y ; unused
1.53 +; sp[]int y_step_q4 ; unused
1.54 +; sp[]int w
1.55 +; sp[]int h
1.56 +
1.57 +|vp9_convolve8_horiz_neon| PROC
1.58 + ldr r12, [sp, #4] ; x_step_q4
1.59 + cmp r12, #16
1.60 + bne vp9_convolve8_horiz_c
1.61 +
1.62 + push {r4-r10, lr}
1.63 +
1.64 + sub r0, r0, #3 ; adjust for taps
1.65 +
1.66 + ldr r5, [sp, #32] ; filter_x
1.67 + ldr r6, [sp, #48] ; w
1.68 + ldr r7, [sp, #52] ; h
1.69 +
1.70 + vld1.s16 {q0}, [r5] ; filter_x
1.71 +
1.72 + sub r8, r1, r1, lsl #2 ; -src_stride * 3
1.73 + add r8, r8, #4 ; -src_stride * 3 + 4
1.74 +
1.75 + sub r4, r3, r3, lsl #2 ; -dst_stride * 3
1.76 + add r4, r4, #4 ; -dst_stride * 3 + 4
1.77 +
1.78 + rsb r9, r6, r1, lsl #2 ; reset src for outer loop
1.79 + sub r9, r9, #7
1.80 + rsb r12, r6, r3, lsl #2 ; reset dst for outer loop
1.81 +
1.82 + mov r10, r6 ; w loop counter
1.83 +
1.84 +loop_horiz_v
1.85 + vld1.8 {d24}, [r0], r1
1.86 + vld1.8 {d25}, [r0], r1
1.87 + vld1.8 {d26}, [r0], r1
1.88 + vld1.8 {d27}, [r0], r8
1.89 +
1.90 + vtrn.16 q12, q13
1.91 + vtrn.8 d24, d25
1.92 + vtrn.8 d26, d27
1.93 +
1.94 + pld [r0, r1, lsl #2]
1.95 +
1.96 + vmovl.u8 q8, d24
1.97 + vmovl.u8 q9, d25
1.98 + vmovl.u8 q10, d26
1.99 + vmovl.u8 q11, d27
1.100 +
1.101 + ; save a few instructions in the inner loop
1.102 + vswp d17, d18
1.103 + vmov d23, d21
1.104 +
1.105 + add r0, r0, #3
1.106 +
1.107 +loop_horiz
1.108 + add r5, r0, #64
1.109 +
1.110 + vld1.32 {d28[]}, [r0], r1
1.111 + vld1.32 {d29[]}, [r0], r1
1.112 + vld1.32 {d31[]}, [r0], r1
1.113 + vld1.32 {d30[]}, [r0], r8
1.114 +
1.115 + pld [r5]
1.116 +
1.117 + vtrn.16 d28, d31
1.118 + vtrn.16 d29, d30
1.119 + vtrn.8 d28, d29
1.120 + vtrn.8 d31, d30
1.121 +
1.122 + pld [r5, r1]
1.123 +
1.124 + ; extract to s16
1.125 + vtrn.32 q14, q15
1.126 + vmovl.u8 q12, d28
1.127 + vmovl.u8 q13, d29
1.128 +
1.129 + pld [r5, r1, lsl #1]
1.130 +
1.131 + ; src[] * filter_x
1.132 + MULTIPLY_BY_Q0 q1, d16, d17, d20, d22, d18, d19, d23, d24
1.133 + MULTIPLY_BY_Q0 q2, d17, d20, d22, d18, d19, d23, d24, d26
1.134 + MULTIPLY_BY_Q0 q14, d20, d22, d18, d19, d23, d24, d26, d27
1.135 + MULTIPLY_BY_Q0 q15, d22, d18, d19, d23, d24, d26, d27, d25
1.136 +
1.137 + pld [r5, -r8]
1.138 +
1.139 + ; += 64 >> 7
1.140 + vqrshrun.s32 d2, q1, #7
1.141 + vqrshrun.s32 d3, q2, #7
1.142 + vqrshrun.s32 d4, q14, #7
1.143 + vqrshrun.s32 d5, q15, #7
1.144 +
1.145 + ; saturate
1.146 + vqmovn.u16 d2, q1
1.147 + vqmovn.u16 d3, q2
1.148 +
1.149 + ; transpose
1.150 + vtrn.16 d2, d3
1.151 + vtrn.32 d2, d3
1.152 + vtrn.8 d2, d3
1.153 +
1.154 + vst1.u32 {d2[0]}, [r2@32], r3
1.155 + vst1.u32 {d3[0]}, [r2@32], r3
1.156 + vst1.u32 {d2[1]}, [r2@32], r3
1.157 + vst1.u32 {d3[1]}, [r2@32], r4
1.158 +
1.159 + vmov q8, q9
1.160 + vmov d20, d23
1.161 + vmov q11, q12
1.162 + vmov q9, q13
1.163 +
1.164 + subs r6, r6, #4 ; w -= 4
1.165 + bgt loop_horiz
1.166 +
1.167 + ; outer loop
1.168 + mov r6, r10 ; restore w counter
1.169 + add r0, r0, r9 ; src += src_stride * 4 - w
1.170 + add r2, r2, r12 ; dst += dst_stride * 4 - w
1.171 + subs r7, r7, #4 ; h -= 4
1.172 + bgt loop_horiz_v
1.173 +
1.174 + pop {r4-r10, pc}
1.175 +
1.176 + ENDP
1.177 +
1.178 +|vp9_convolve8_vert_neon| PROC
1.179 + ldr r12, [sp, #12]
1.180 + cmp r12, #16
1.181 + bne vp9_convolve8_vert_c
1.182 +
1.183 + push {r4-r8, lr}
1.184 +
1.185 + ; adjust for taps
1.186 + sub r0, r0, r1
1.187 + sub r0, r0, r1, lsl #1
1.188 +
1.189 + ldr r4, [sp, #32] ; filter_y
1.190 + ldr r6, [sp, #40] ; w
1.191 + ldr lr, [sp, #44] ; h
1.192 +
1.193 + vld1.s16 {q0}, [r4] ; filter_y
1.194 +
1.195 + lsl r1, r1, #1
1.196 + lsl r3, r3, #1
1.197 +
1.198 +loop_vert_h
1.199 + mov r4, r0
1.200 + add r7, r0, r1, asr #1
1.201 + mov r5, r2
1.202 + add r8, r2, r3, asr #1
1.203 + mov r12, lr ; h loop counter
1.204 +
1.205 + vld1.u32 {d16[0]}, [r4], r1
1.206 + vld1.u32 {d16[1]}, [r7], r1
1.207 + vld1.u32 {d18[0]}, [r4], r1
1.208 + vld1.u32 {d18[1]}, [r7], r1
1.209 + vld1.u32 {d20[0]}, [r4], r1
1.210 + vld1.u32 {d20[1]}, [r7], r1
1.211 + vld1.u32 {d22[0]}, [r4], r1
1.212 +
1.213 + vmovl.u8 q8, d16
1.214 + vmovl.u8 q9, d18
1.215 + vmovl.u8 q10, d20
1.216 + vmovl.u8 q11, d22
1.217 +
1.218 +loop_vert
1.219 + ; always process a 4x4 block at a time
1.220 + vld1.u32 {d24[0]}, [r7], r1
1.221 + vld1.u32 {d26[0]}, [r4], r1
1.222 + vld1.u32 {d26[1]}, [r7], r1
1.223 + vld1.u32 {d24[1]}, [r4], r1
1.224 +
1.225 + ; extract to s16
1.226 + vmovl.u8 q12, d24
1.227 + vmovl.u8 q13, d26
1.228 +
1.229 + pld [r5]
1.230 + pld [r8]
1.231 +
1.232 + ; src[] * filter_y
1.233 + MULTIPLY_BY_Q0 q1, d16, d17, d18, d19, d20, d21, d22, d24
1.234 +
1.235 + pld [r5, r3]
1.236 + pld [r8, r3]
1.237 +
1.238 + MULTIPLY_BY_Q0 q2, d17, d18, d19, d20, d21, d22, d24, d26
1.239 +
1.240 + pld [r7]
1.241 + pld [r4]
1.242 +
1.243 + MULTIPLY_BY_Q0 q14, d18, d19, d20, d21, d22, d24, d26, d27
1.244 +
1.245 + pld [r7, r1]
1.246 + pld [r4, r1]
1.247 +
1.248 + MULTIPLY_BY_Q0 q15, d19, d20, d21, d22, d24, d26, d27, d25
1.249 +
1.250 + ; += 64 >> 7
1.251 + vqrshrun.s32 d2, q1, #7
1.252 + vqrshrun.s32 d3, q2, #7
1.253 + vqrshrun.s32 d4, q14, #7
1.254 + vqrshrun.s32 d5, q15, #7
1.255 +
1.256 + ; saturate
1.257 + vqmovn.u16 d2, q1
1.258 + vqmovn.u16 d3, q2
1.259 +
1.260 + vst1.u32 {d2[0]}, [r5@32], r3
1.261 + vst1.u32 {d2[1]}, [r8@32], r3
1.262 + vst1.u32 {d3[0]}, [r5@32], r3
1.263 + vst1.u32 {d3[1]}, [r8@32], r3
1.264 +
1.265 + vmov q8, q10
1.266 + vmov d18, d22
1.267 + vmov d19, d24
1.268 + vmov q10, q13
1.269 + vmov d22, d25
1.270 +
1.271 + subs r12, r12, #4 ; h -= 4
1.272 + bgt loop_vert
1.273 +
1.274 + ; outer loop
1.275 + add r0, r0, #4
1.276 + add r2, r2, #4
1.277 + subs r6, r6, #4 ; w -= 4
1.278 + bgt loop_vert_h
1.279 +
1.280 + pop {r4-r8, pc}
1.281 +
1.282 + ENDP
1.283 + END
