upsampling_neon.c revision 2a99a7e74a7f215066514fe81d2bfa6639d9eddd
1// Copyright 2011 Google Inc. All Rights Reserved. 2// 3// This code is licensed under the same terms as WebM: 4// Software License Agreement: http://www.webmproject.org/license/software/ 5// Additional IP Rights Grant: http://www.webmproject.org/license/additional/ 6// ----------------------------------------------------------------------------- 7// 8// NEON version of YUV to RGB upsampling functions. 9// 10// Author: mans@mansr.com (Mans Rullgard) 11// Based on SSE code by: somnath@google.com (Somnath Banerjee) 12 13#include "./dsp.h" 14 15#if defined(__cplusplus) || defined(c_plusplus) 16extern "C" { 17#endif 18 19#if defined(WEBP_USE_NEON) 20 21#include <assert.h> 22#include <arm_neon.h> 23#include <string.h> 24#include "./yuv.h" 25 26#ifdef FANCY_UPSAMPLING 27 28// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels. 29#define UPSAMPLE_16PIXELS(r1, r2, out) { \ 30 uint8x8_t a = vld1_u8(r1); \ 31 uint8x8_t b = vld1_u8(r1 + 1); \ 32 uint8x8_t c = vld1_u8(r2); \ 33 uint8x8_t d = vld1_u8(r2 + 1); \ 34 \ 35 uint16x8_t al = vshll_n_u8(a, 1); \ 36 uint16x8_t bl = vshll_n_u8(b, 1); \ 37 uint16x8_t cl = vshll_n_u8(c, 1); \ 38 uint16x8_t dl = vshll_n_u8(d, 1); \ 39 \ 40 uint8x8_t diag1, diag2; \ 41 uint16x8_t sl; \ 42 \ 43 /* a + b + c + d */ \ 44 sl = vaddl_u8(a, b); \ 45 sl = vaddw_u8(sl, c); \ 46 sl = vaddw_u8(sl, d); \ 47 \ 48 al = vaddq_u16(sl, al); /* 3a + b + c + d */ \ 49 bl = vaddq_u16(sl, bl); /* a + 3b + c + d */ \ 50 \ 51 al = vaddq_u16(al, dl); /* 3a + b + c + 3d */ \ 52 bl = vaddq_u16(bl, cl); /* a + 3b + 3c + d */ \ 53 \ 54 diag2 = vshrn_n_u16(al, 3); \ 55 diag1 = vshrn_n_u16(bl, 3); \ 56 \ 57 a = vrhadd_u8(a, diag1); \ 58 b = vrhadd_u8(b, diag2); \ 59 c = vrhadd_u8(c, diag2); \ 60 d = vrhadd_u8(d, diag1); \ 61 \ 62 { \ 63 const uint8x8x2_t a_b = {{ a, b }}; \ 64 const uint8x8x2_t c_d = {{ c, d }}; \ 65 vst2_u8(out, a_b); \ 66 vst2_u8(out + 32, c_d); \ 67 } \ 68} 69 70// Turn the macro into a function for reducing code-size when non-critical 71static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2, 72 uint8_t *out) { 73 UPSAMPLE_16PIXELS(r1, r2, out); 74} 75 76#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) { \ 77 uint8_t r1[9], r2[9]; \ 78 memcpy(r1, (tb), (num_pixels)); \ 79 memcpy(r2, (bb), (num_pixels)); \ 80 /* replicate last byte */ \ 81 memset(r1 + (num_pixels), r1[(num_pixels) - 1], 9 - (num_pixels)); \ 82 memset(r2 + (num_pixels), r2[(num_pixels) - 1], 9 - (num_pixels)); \ 83 Upsample16Pixels(r1, r2, out); \ 84} 85 86#define CY 76283 87#define CVR 89858 88#define CUG 22014 89#define CVG 45773 90#define CUB 113618 91 92static const int16_t coef[4] = { CVR / 4, CUG, CVG / 2, CUB / 4 }; 93 94#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) { \ 95 int i; \ 96 for (i = 0; i < N; i += 8) { \ 97 int off = ((cur_x) + i) * XSTEP; \ 98 uint8x8_t y = vld1_u8(src_y + (cur_x) + i); \ 99 uint8x8_t u = vld1_u8((src_uv) + i); \ 100 uint8x8_t v = vld1_u8((src_uv) + i + 16); \ 101 int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16)); \ 102 int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128)); \ 103 int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128)); \ 104 \ 105 int16x8_t ud = vshlq_n_s16(uu, 1); \ 106 int16x8_t vd = vshlq_n_s16(vv, 1); \ 107 \ 108 int32x4_t vrl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(vv), 1), \ 109 vget_low_s16(vd), cf16, 0); \ 110 int32x4_t vrh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(vv), 1), \ 111 vget_high_s16(vd), cf16, 0); \ 112 int16x8_t vr = vcombine_s16(vrshrn_n_s32(vrl, 16), \ 113 vrshrn_n_s32(vrh, 16)); \ 114 \ 115 int32x4_t vl = vmovl_s16(vget_low_s16(vv)); \ 116 int32x4_t vh = vmovl_s16(vget_high_s16(vv)); \ 117 int32x4_t ugl = vmlal_lane_s16(vl, vget_low_s16(uu), cf16, 1); \ 118 int32x4_t ugh = vmlal_lane_s16(vh, vget_high_s16(uu), cf16, 1); \ 119 int32x4_t gcl = vqdmlal_lane_s16(ugl, vget_low_s16(vv), cf16, 2); \ 120 int32x4_t gch = vqdmlal_lane_s16(ugh, vget_high_s16(vv), cf16, 2); \ 121 int16x8_t gc = vcombine_s16(vrshrn_n_s32(gcl, 16), \ 122 vrshrn_n_s32(gch, 16)); \ 123 \ 124 int32x4_t ubl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(uu), 1), \ 125 vget_low_s16(ud), cf16, 3); \ 126 int32x4_t ubh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(uu), 1), \ 127 vget_high_s16(ud), cf16, 3); \ 128 int16x8_t ub = vcombine_s16(vrshrn_n_s32(ubl, 16), \ 129 vrshrn_n_s32(ubh, 16)); \ 130 \ 131 int32x4_t rl = vaddl_s16(vget_low_s16(yy), vget_low_s16(vr)); \ 132 int32x4_t rh = vaddl_s16(vget_high_s16(yy), vget_high_s16(vr)); \ 133 int32x4_t gl = vsubl_s16(vget_low_s16(yy), vget_low_s16(gc)); \ 134 int32x4_t gh = vsubl_s16(vget_high_s16(yy), vget_high_s16(gc)); \ 135 int32x4_t bl = vaddl_s16(vget_low_s16(yy), vget_low_s16(ub)); \ 136 int32x4_t bh = vaddl_s16(vget_high_s16(yy), vget_high_s16(ub)); \ 137 \ 138 rl = vmulq_lane_s32(rl, cf32, 0); \ 139 rh = vmulq_lane_s32(rh, cf32, 0); \ 140 gl = vmulq_lane_s32(gl, cf32, 0); \ 141 gh = vmulq_lane_s32(gh, cf32, 0); \ 142 bl = vmulq_lane_s32(bl, cf32, 0); \ 143 bh = vmulq_lane_s32(bh, cf32, 0); \ 144 \ 145 y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, 16), \ 146 vrshrn_n_s32(rh, 16))); \ 147 u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, 16), \ 148 vrshrn_n_s32(gh, 16))); \ 149 v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(bl, 16), \ 150 vrshrn_n_s32(bh, 16))); \ 151 STR_ ## FMT(out + off, y, u, v); \ 152 } \ 153} 154 155#define v255 vmov_n_u8(255) 156 157#define STR_Rgb(out, r, g, b) do { \ 158 const uint8x8x3_t r_g_b = {{ r, g, b }}; \ 159 vst3_u8(out, r_g_b); \ 160} while (0) 161 162#define STR_Bgr(out, r, g, b) do { \ 163 const uint8x8x3_t b_g_r = {{ b, g, r }}; \ 164 vst3_u8(out, b_g_r); \ 165} while (0) 166 167#define STR_Rgba(out, r, g, b) do { \ 168 const uint8x8x4_t r_g_b_v255 = {{ r, g, b, v255 }}; \ 169 vst4_u8(out, r_g_b_v255); \ 170} while (0) 171 172#define STR_Bgra(out, r, g, b) do { \ 173 const uint8x8x4_t b_g_r_v255 = {{ b, g, r, v255 }}; \ 174 vst4_u8(out, b_g_r_v255); \ 175} while (0) 176 177#define CONVERT1(FMT, XSTEP, N, src_y, src_uv, rgb, cur_x) { \ 178 int i; \ 179 for (i = 0; i < N; i++) { \ 180 int off = ((cur_x) + i) * XSTEP; \ 181 int y = src_y[(cur_x) + i]; \ 182 int u = (src_uv)[i]; \ 183 int v = (src_uv)[i + 16]; \ 184 VP8YuvTo ## FMT(y, u, v, rgb + off); \ 185 } \ 186} 187 188#define CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, uv, \ 189 top_dst, bottom_dst, cur_x, len) { \ 190 if (top_y) { \ 191 CONVERT8(FMT, XSTEP, len, top_y, uv, top_dst, cur_x) \ 192 } \ 193 if (bottom_y) { \ 194 CONVERT8(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x) \ 195 } \ 196} 197 198#define CONVERT2RGB_1(FMT, XSTEP, top_y, bottom_y, uv, \ 199 top_dst, bottom_dst, cur_x, len) { \ 200 if (top_y) { \ 201 CONVERT1(FMT, XSTEP, len, top_y, uv, top_dst, cur_x); \ 202 } \ 203 if (bottom_y) { \ 204 CONVERT1(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \ 205 } \ 206} 207 208#define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP) \ 209static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y, \ 210 const uint8_t *top_u, const uint8_t *top_v, \ 211 const uint8_t *cur_u, const uint8_t *cur_v, \ 212 uint8_t *top_dst, uint8_t *bottom_dst, int len) { \ 213 int block; \ 214 /* 16 byte aligned array to cache reconstructed u and v */ \ 215 uint8_t uv_buf[2 * 32 + 15]; \ 216 uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15); \ 217 const int uv_len = (len + 1) >> 1; \ 218 /* 9 pixels must be read-able for each block */ \ 219 const int num_blocks = (uv_len - 1) >> 3; \ 220 const int leftover = uv_len - num_blocks * 8; \ 221 const int last_pos = 1 + 16 * num_blocks; \ 222 \ 223 const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \ 224 const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \ 225 \ 226 const int16x4_t cf16 = vld1_s16(coef); \ 227 const int32x2_t cf32 = vmov_n_s32(CY); \ 228 const uint8x8_t u16 = vmov_n_u8(16); \ 229 const uint8x8_t u128 = vmov_n_u8(128); \ 230 \ 231 /* Treat the first pixel in regular way */ \ 232 if (top_y) { \ 233 const int u0 = (top_u[0] + u_diag) >> 1; \ 234 const int v0 = (top_v[0] + v_diag) >> 1; \ 235 VP8YuvTo ## FMT(top_y[0], u0, v0, top_dst); \ 236 } \ 237 if (bottom_y) { \ 238 const int u0 = (cur_u[0] + u_diag) >> 1; \ 239 const int v0 = (cur_v[0] + v_diag) >> 1; \ 240 VP8YuvTo ## FMT(bottom_y[0], u0, v0, bottom_dst); \ 241 } \ 242 \ 243 for (block = 0; block < num_blocks; ++block) { \ 244 UPSAMPLE_16PIXELS(top_u, cur_u, r_uv); \ 245 UPSAMPLE_16PIXELS(top_v, cur_v, r_uv + 16); \ 246 CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, r_uv, \ 247 top_dst, bottom_dst, 16 * block + 1, 16); \ 248 top_u += 8; \ 249 cur_u += 8; \ 250 top_v += 8; \ 251 cur_v += 8; \ 252 } \ 253 \ 254 UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv); \ 255 UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16); \ 256 CONVERT2RGB_1(FMT, XSTEP, top_y, bottom_y, r_uv, \ 257 top_dst, bottom_dst, last_pos, len - last_pos); \ 258} 259 260// NEON variants of the fancy upsampler. 261NEON_UPSAMPLE_FUNC(UpsampleRgbLinePairNEON, Rgb, 3) 262NEON_UPSAMPLE_FUNC(UpsampleBgrLinePairNEON, Bgr, 3) 263NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePairNEON, Rgba, 4) 264NEON_UPSAMPLE_FUNC(UpsampleBgraLinePairNEON, Bgra, 4) 265 266#endif // FANCY_UPSAMPLING 267 268#endif // WEBP_USE_NEON 269 270//------------------------------------------------------------------------------ 271 272extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */]; 273 274void WebPInitUpsamplersNEON(void) { 275#if defined(WEBP_USE_NEON) 276 WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePairNEON; 277 WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairNEON; 278 WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePairNEON; 279 WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairNEON; 280#endif // WEBP_USE_NEON 281} 282 283void WebPInitPremultiplyNEON(void) { 284#if defined(WEBP_USE_NEON) 285 WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairNEON; 286 WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairNEON; 287#endif // WEBP_USE_NEON 288} 289 290#if defined(__cplusplus) || defined(c_plusplus) 291} // extern "C" 292#endif 293