yuv_neon.c (7396B)
1 // Copyright 2017 Google Inc. All Rights Reserved. 2 // 3 // Use of this source code is governed by a BSD-style license 4 // that can be found in the COPYING file in the root of the source 5 // tree. An additional intellectual property rights grant can be found 6 // in the file PATENTS. All contributing project authors may 7 // be found in the AUTHORS file in the root of the source tree. 8 // ----------------------------------------------------------------------------- 9 // 10 // YUV->RGB conversion functions 11 // 12 // Author: Skal (pascal.massimino@gmail.com) 13 14 #include "src/dsp/yuv.h" 15 16 #if defined(WEBP_USE_NEON) 17 18 #include <assert.h> 19 #include <stdlib.h> 20 21 #include "src/dsp/dsp.h" 22 #include "src/dsp/neon.h" 23 24 //----------------------------------------------------------------------------- 25 26 static uint8x8_t ConvertRGBToY_NEON(const uint8x8_t R, 27 const uint8x8_t G, 28 const uint8x8_t B) { 29 const uint16x8_t r = vmovl_u8(R); 30 const uint16x8_t g = vmovl_u8(G); 31 const uint16x8_t b = vmovl_u8(B); 32 const uint16x4_t r_lo = vget_low_u16(r); 33 const uint16x4_t r_hi = vget_high_u16(r); 34 const uint16x4_t g_lo = vget_low_u16(g); 35 const uint16x4_t g_hi = vget_high_u16(g); 36 const uint16x4_t b_lo = vget_low_u16(b); 37 const uint16x4_t b_hi = vget_high_u16(b); 38 const uint32x4_t tmp0_lo = vmull_n_u16( r_lo, 16839u); 39 const uint32x4_t tmp0_hi = vmull_n_u16( r_hi, 16839u); 40 const uint32x4_t tmp1_lo = vmlal_n_u16(tmp0_lo, g_lo, 33059u); 41 const uint32x4_t tmp1_hi = vmlal_n_u16(tmp0_hi, g_hi, 33059u); 42 const uint32x4_t tmp2_lo = vmlal_n_u16(tmp1_lo, b_lo, 6420u); 43 const uint32x4_t tmp2_hi = vmlal_n_u16(tmp1_hi, b_hi, 6420u); 44 const uint16x8_t Y1 = vcombine_u16(vrshrn_n_u32(tmp2_lo, 16), 45 vrshrn_n_u32(tmp2_hi, 16)); 46 const uint16x8_t Y2 = vaddq_u16(Y1, vdupq_n_u16(16)); 47 return vqmovn_u16(Y2); 48 } 49 50 static void ConvertRGB24ToY_NEON(const uint8_t* WEBP_RESTRICT rgb, 51 uint8_t* WEBP_RESTRICT y, int width) { 52 int i; 53 for (i = 0; i + 8 <= width; i += 8, rgb += 3 * 8) { 54 const uint8x8x3_t RGB = vld3_u8(rgb); 55 const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[0], RGB.val[1], RGB.val[2]); 56 vst1_u8(y + i, Y); 57 } 58 for (; i < width; ++i, rgb += 3) { // left-over 59 y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF); 60 } 61 } 62 63 static void ConvertBGR24ToY_NEON(const uint8_t* WEBP_RESTRICT bgr, 64 uint8_t* WEBP_RESTRICT y, int width) { 65 int i; 66 for (i = 0; i + 8 <= width; i += 8, bgr += 3 * 8) { 67 const uint8x8x3_t BGR = vld3_u8(bgr); 68 const uint8x8_t Y = ConvertRGBToY_NEON(BGR.val[2], BGR.val[1], BGR.val[0]); 69 vst1_u8(y + i, Y); 70 } 71 for (; i < width; ++i, bgr += 3) { // left-over 72 y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF); 73 } 74 } 75 76 static void ConvertARGBToY_NEON(const uint32_t* WEBP_RESTRICT argb, 77 uint8_t* WEBP_RESTRICT y, int width) { 78 int i; 79 for (i = 0; i + 8 <= width; i += 8) { 80 const uint8x8x4_t RGB = vld4_u8((const uint8_t*)&argb[i]); 81 const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[2], RGB.val[1], RGB.val[0]); 82 vst1_u8(y + i, Y); 83 } 84 for (; i < width; ++i) { // left-over 85 const uint32_t p = argb[i]; 86 y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff, 87 YUV_HALF); 88 } 89 } 90 91 //----------------------------------------------------------------------------- 92 93 // computes: DST_s16 = [(C0 * r + C1 * g + C2 * b) >> 16] + CST 94 #define MULTIPLY_16b_PREAMBLE(r, g, b) \ 95 const int16x4_t r_lo = vreinterpret_s16_u16(vget_low_u16(r)); \ 96 const int16x4_t r_hi = vreinterpret_s16_u16(vget_high_u16(r)); \ 97 const int16x4_t g_lo = vreinterpret_s16_u16(vget_low_u16(g)); \ 98 const int16x4_t g_hi = vreinterpret_s16_u16(vget_high_u16(g)); \ 99 const int16x4_t b_lo = vreinterpret_s16_u16(vget_low_u16(b)); \ 100 const int16x4_t b_hi = vreinterpret_s16_u16(vget_high_u16(b)) 101 102 #define MULTIPLY_16b(C0, C1, C2, CST, DST_s16) do { \ 103 const int32x4_t tmp0_lo = vmull_n_s16( r_lo, C0); \ 104 const int32x4_t tmp0_hi = vmull_n_s16( r_hi, C0); \ 105 const int32x4_t tmp1_lo = vmlal_n_s16(tmp0_lo, g_lo, C1); \ 106 const int32x4_t tmp1_hi = vmlal_n_s16(tmp0_hi, g_hi, C1); \ 107 const int32x4_t tmp2_lo = vmlal_n_s16(tmp1_lo, b_lo, C2); \ 108 const int32x4_t tmp2_hi = vmlal_n_s16(tmp1_hi, b_hi, C2); \ 109 const int16x8_t tmp3 = vcombine_s16(vshrn_n_s32(tmp2_lo, 16), \ 110 vshrn_n_s32(tmp2_hi, 16)); \ 111 DST_s16 = vaddq_s16(tmp3, vdupq_n_s16(CST)); \ 112 } while (0) 113 114 // This needs to be a macro, since (128 << SHIFT) needs to be an immediate. 115 #define CONVERT_RGB_TO_UV(r, g, b, SHIFT, U_DST, V_DST) do { \ 116 MULTIPLY_16b_PREAMBLE(r, g, b); \ 117 MULTIPLY_16b(-9719, -19081, 28800, 128 << SHIFT, U_DST); \ 118 MULTIPLY_16b(28800, -24116, -4684, 128 << SHIFT, V_DST); \ 119 } while (0) 120 121 static void ConvertRGBA32ToUV_NEON(const uint16_t* WEBP_RESTRICT rgb, 122 uint8_t* WEBP_RESTRICT u, 123 uint8_t* WEBP_RESTRICT v, int width) { 124 int i; 125 for (i = 0; i + 8 <= width; i += 8, rgb += 4 * 8) { 126 const uint16x8x4_t RGB = vld4q_u16((const uint16_t*)rgb); 127 int16x8_t U, V; 128 CONVERT_RGB_TO_UV(RGB.val[0], RGB.val[1], RGB.val[2], 2, U, V); 129 vst1_u8(u + i, vqrshrun_n_s16(U, 2)); 130 vst1_u8(v + i, vqrshrun_n_s16(V, 2)); 131 } 132 for (; i < width; i += 1, rgb += 4) { 133 const int r = rgb[0], g = rgb[1], b = rgb[2]; 134 u[i] = VP8RGBToU(r, g, b, YUV_HALF << 2); 135 v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2); 136 } 137 } 138 139 static void ConvertARGBToUV_NEON(const uint32_t* WEBP_RESTRICT argb, 140 uint8_t* WEBP_RESTRICT u, 141 uint8_t* WEBP_RESTRICT v, 142 int src_width, int do_store) { 143 int i; 144 for (i = 0; i + 16 <= src_width; i += 16, u += 8, v += 8) { 145 const uint8x16x4_t RGB = vld4q_u8((const uint8_t*)&argb[i]); 146 const uint16x8_t R = vpaddlq_u8(RGB.val[2]); // pair-wise adds 147 const uint16x8_t G = vpaddlq_u8(RGB.val[1]); 148 const uint16x8_t B = vpaddlq_u8(RGB.val[0]); 149 int16x8_t U_tmp, V_tmp; 150 CONVERT_RGB_TO_UV(R, G, B, 1, U_tmp, V_tmp); 151 { 152 const uint8x8_t U = vqrshrun_n_s16(U_tmp, 1); 153 const uint8x8_t V = vqrshrun_n_s16(V_tmp, 1); 154 if (do_store) { 155 vst1_u8(u, U); 156 vst1_u8(v, V); 157 } else { 158 const uint8x8_t prev_u = vld1_u8(u); 159 const uint8x8_t prev_v = vld1_u8(v); 160 vst1_u8(u, vrhadd_u8(U, prev_u)); 161 vst1_u8(v, vrhadd_u8(V, prev_v)); 162 } 163 } 164 } 165 if (i < src_width) { // left-over 166 WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store); 167 } 168 } 169 170 171 //------------------------------------------------------------------------------ 172 173 extern void WebPInitConvertARGBToYUVNEON(void); 174 175 WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVNEON(void) { 176 WebPConvertRGB24ToY = ConvertRGB24ToY_NEON; 177 WebPConvertBGR24ToY = ConvertBGR24ToY_NEON; 178 WebPConvertARGBToY = ConvertARGBToY_NEON; 179 WebPConvertARGBToUV = ConvertARGBToUV_NEON; 180 WebPConvertRGBA32ToUV = ConvertRGBA32ToUV_NEON; 181 } 182 183 #else // !WEBP_USE_NEON 184 185 WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVNEON) 186 187 #endif // WEBP_USE_NEON