1/* 2 * Copyright (c) 2014 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 <arm_neon.h> 12#include "./vpx_dsp_rtcd.h" 13#include "./vpx_config.h" 14 15#include "vpx_ports/mem.h" 16#include "vpx/vpx_integer.h" 17 18#include "vpx_dsp/variance.h" 19 20static const uint8_t bilinear_filters[8][2] = { 21 { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 }, 22 { 64, 64 }, { 48, 80 }, { 32, 96 }, { 16, 112 }, 23}; 24 25// Process a block exactly 8 wide and any height. 26static void var_filter_block2d_bil_w8(const uint8_t *src_ptr, 27 uint8_t *output_ptr, 28 unsigned int src_pixels_per_line, 29 int pixel_step, 30 unsigned int output_height, 31 const uint8_t *filter) { 32 const uint8x8_t f0 = vmov_n_u8(filter[0]); 33 const uint8x8_t f1 = vmov_n_u8(filter[1]); 34 unsigned int i; 35 for (i = 0; i < output_height; ++i) { 36 const uint8x8_t src_0 = vld1_u8(&src_ptr[0]); 37 const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]); 38 const uint16x8_t a = vmull_u8(src_0, f0); 39 const uint16x8_t b = vmlal_u8(a, src_1, f1); 40 const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS); 41 vst1_u8(&output_ptr[0], out); 42 // Next row... 43 src_ptr += src_pixels_per_line; 44 output_ptr += 8; 45 } 46} 47 48// Process a block which is a mutiple of 16 wide and any height. 49static void var_filter_block2d_bil_w16(const uint8_t *src_ptr, 50 uint8_t *output_ptr, 51 unsigned int src_pixels_per_line, 52 int pixel_step, 53 unsigned int output_height, 54 unsigned int output_width, 55 const uint8_t *filter) { 56 const uint8x8_t f0 = vmov_n_u8(filter[0]); 57 const uint8x8_t f1 = vmov_n_u8(filter[1]); 58 unsigned int i, j; 59 for (i = 0; i < output_height; ++i) { 60 for (j = 0; j < output_width; j += 16) { 61 const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]); 62 const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]); 63 const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0); 64 const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1); 65 const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS); 66 const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0); 67 const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1); 68 const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS); 69 vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi)); 70 } 71 // Next row... 72 src_ptr += src_pixels_per_line; 73 output_ptr += output_width; 74 } 75} 76 77// TODO(johannkoenig): support 4xM block sizes. 78#define sub_pixel_varianceNxM(n, m) \ 79 unsigned int vpx_sub_pixel_variance##n##x##m##_neon( \ 80 const uint8_t *src, int src_stride, int xoffset, int yoffset, \ 81 const uint8_t *dst, int dst_stride, unsigned int *sse) { \ 82 DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + 1)]); \ 83 DECLARE_ALIGNED(16, uint8_t, temp2[n * m]); \ 84 \ 85 if (n == 8) { \ 86 var_filter_block2d_bil_w8(src, fdata3, src_stride, 1, (m + 1), \ 87 bilinear_filters[xoffset]); \ 88 var_filter_block2d_bil_w8(fdata3, temp2, n, n, m, \ 89 bilinear_filters[yoffset]); \ 90 } else { \ 91 var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, (m + 1), n, \ 92 bilinear_filters[xoffset]); \ 93 var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n, \ 94 bilinear_filters[yoffset]); \ 95 } \ 96 return vpx_variance##n##x##m(temp2, n, dst, dst_stride, sse); \ 97 } 98 99sub_pixel_varianceNxM(8, 4); 100sub_pixel_varianceNxM(8, 8); 101sub_pixel_varianceNxM(8, 16); 102sub_pixel_varianceNxM(16, 8); 103sub_pixel_varianceNxM(16, 16); 104sub_pixel_varianceNxM(16, 32); 105sub_pixel_varianceNxM(32, 16); 106sub_pixel_varianceNxM(32, 32); 107sub_pixel_varianceNxM(32, 64); 108sub_pixel_varianceNxM(64, 32); 109sub_pixel_varianceNxM(64, 64); 110