1 /* 2 * Copyright (c) 2013 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include <assert.h> 12 13 #include "./vpx_dsp_rtcd.h" 14 #include "vpx_dsp/vpx_dsp_common.h" 15 #include "vpx_ports/mem.h" 16 17 void vpx_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, 18 ptrdiff_t dst_stride, const int16_t *filter_x, 19 int x_step_q4, const int16_t *filter_y, int y_step_q4, 20 int w, int h) { 21 /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the 22 * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4). 23 */ 24 DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]); 25 26 // Account for the vertical phase needing 3 lines prior and 4 lines post 27 const int intermediate_height = h + 7; 28 29 assert(y_step_q4 == 16); 30 assert(x_step_q4 == 16); 31 32 /* Filter starting 3 lines back. The neon implementation will ignore the given 33 * height and filter a multiple of 4 lines. Since this goes in to the temp 34 * buffer which has lots of extra room and is subsequently discarded this is 35 * safe if somewhat less than ideal. */ 36 vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, w, filter_x, 37 x_step_q4, filter_y, y_step_q4, w, 38 intermediate_height); 39 40 /* Step into the temp buffer 3 lines to get the actual frame data */ 41 vpx_convolve8_vert_neon(temp + w * 3, w, dst, dst_stride, filter_x, x_step_q4, 42 filter_y, y_step_q4, w, h); 43 } 44 45 void vpx_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride, 46 uint8_t *dst, ptrdiff_t dst_stride, 47 const int16_t *filter_x, int x_step_q4, 48 const int16_t *filter_y, int y_step_q4, int w, 49 int h) { 50 DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]); 51 const int intermediate_height = h + 7; 52 53 assert(y_step_q4 == 16); 54 assert(x_step_q4 == 16); 55 56 /* This implementation has the same issues as above. In addition, we only want 57 * to average the values after both passes. 58 */ 59 vpx_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, w, filter_x, 60 x_step_q4, filter_y, y_step_q4, w, 61 intermediate_height); 62 vpx_convolve8_avg_vert_neon(temp + w * 3, w, dst, dst_stride, filter_x, 63 x_step_q4, filter_y, y_step_q4, w, h); 64 } 65