1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/skia/trunk/src/opts/SkXfermode_opts_arm_neon.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,937 @@
1.4 +#include "SkXfermode.h"
1.5 +#include "SkXfermode_proccoeff.h"
1.6 +#include "SkColorPriv.h"
1.7 +
1.8 +#include <arm_neon.h>
1.9 +#include "SkColor_opts_neon.h"
1.10 +#include "SkXfermode_opts_arm_neon.h"
1.11 +
1.12 +#define SkAlphaMulAlpha(a, b) SkMulDiv255Round(a, b)
1.13 +
1.14 +
1.15 +////////////////////////////////////////////////////////////////////////////////
1.16 +// NEONized skia functions
1.17 +////////////////////////////////////////////////////////////////////////////////
1.18 +
1.19 +static inline uint8x8_t SkAlphaMulAlpha_neon8(uint8x8_t color, uint8x8_t alpha) {
1.20 + uint16x8_t tmp;
1.21 + uint8x8_t ret;
1.22 +
1.23 + tmp = vmull_u8(color, alpha);
1.24 + tmp = vaddq_u16(tmp, vdupq_n_u16(128));
1.25 + tmp = vaddq_u16(tmp, vshrq_n_u16(tmp, 8));
1.26 +
1.27 + ret = vshrn_n_u16(tmp, 8);
1.28 +
1.29 + return ret;
1.30 +}
1.31 +
1.32 +static inline uint16x8_t SkAlphaMulAlpha_neon8_16(uint8x8_t color, uint8x8_t alpha) {
1.33 + uint16x8_t ret;
1.34 +
1.35 + ret = vmull_u8(color, alpha);
1.36 + ret = vaddq_u16(ret, vdupq_n_u16(128));
1.37 + ret = vaddq_u16(ret, vshrq_n_u16(ret, 8));
1.38 +
1.39 + ret = vshrq_n_u16(ret, 8);
1.40 +
1.41 + return ret;
1.42 +}
1.43 +
1.44 +static inline uint8x8_t SkDiv255Round_neon8_32_8(int32x4_t p1, int32x4_t p2) {
1.45 + uint16x8_t tmp;
1.46 +
1.47 + tmp = vcombine_u16(vmovn_u32(vreinterpretq_u32_s32(p1)),
1.48 + vmovn_u32(vreinterpretq_u32_s32(p2)));
1.49 +
1.50 + tmp += vdupq_n_u16(128);
1.51 + tmp += vshrq_n_u16(tmp, 8);
1.52 +
1.53 + return vshrn_n_u16(tmp, 8);
1.54 +}
1.55 +
1.56 +static inline uint16x8_t SkDiv255Round_neon8_16_16(uint16x8_t prod) {
1.57 + prod += vdupq_n_u16(128);
1.58 + prod += vshrq_n_u16(prod, 8);
1.59 +
1.60 + return vshrq_n_u16(prod, 8);
1.61 +}
1.62 +
1.63 +static inline uint8x8_t clamp_div255round_simd8_32(int32x4_t val1, int32x4_t val2) {
1.64 + uint8x8_t ret;
1.65 + uint32x4_t cmp1, cmp2;
1.66 + uint16x8_t cmp16;
1.67 + uint8x8_t cmp8, cmp8_1;
1.68 +
1.69 + // Test if <= 0
1.70 + cmp1 = vcleq_s32(val1, vdupq_n_s32(0));
1.71 + cmp2 = vcleq_s32(val2, vdupq_n_s32(0));
1.72 + cmp16 = vcombine_u16(vmovn_u32(cmp1), vmovn_u32(cmp2));
1.73 + cmp8_1 = vmovn_u16(cmp16);
1.74 +
1.75 + // Init to zero
1.76 + ret = vdup_n_u8(0);
1.77 +
1.78 + // Test if >= 255*255
1.79 + cmp1 = vcgeq_s32(val1, vdupq_n_s32(255*255));
1.80 + cmp2 = vcgeq_s32(val2, vdupq_n_s32(255*255));
1.81 + cmp16 = vcombine_u16(vmovn_u32(cmp1), vmovn_u32(cmp2));
1.82 + cmp8 = vmovn_u16(cmp16);
1.83 +
1.84 + // Insert 255 where true
1.85 + ret = vbsl_u8(cmp8, vdup_n_u8(255), ret);
1.86 +
1.87 + // Calc SkDiv255Round
1.88 + uint8x8_t div = SkDiv255Round_neon8_32_8(val1, val2);
1.89 +
1.90 + // Insert where false and previous test false
1.91 + cmp8 = cmp8 | cmp8_1;
1.92 + ret = vbsl_u8(cmp8, ret, div);
1.93 +
1.94 + // Return the final combination
1.95 + return ret;
1.96 +}
1.97 +
1.98 +////////////////////////////////////////////////////////////////////////////////
1.99 +// 1 pixel modeprocs
1.100 +////////////////////////////////////////////////////////////////////////////////
1.101 +
1.102 +// kSrcATop_Mode, //!< [Da, Sc * Da + (1 - Sa) * Dc]
1.103 +SkPMColor srcatop_modeproc_neon(SkPMColor src, SkPMColor dst) {
1.104 + unsigned sa = SkGetPackedA32(src);
1.105 + unsigned da = SkGetPackedA32(dst);
1.106 + unsigned isa = 255 - sa;
1.107 +
1.108 + uint8x8_t vda, visa, vsrc, vdst;
1.109 +
1.110 + vda = vdup_n_u8(da);
1.111 + visa = vdup_n_u8(isa);
1.112 +
1.113 + uint16x8_t vsrc_wide, vdst_wide;
1.114 + vsrc_wide = vmull_u8(vda, vreinterpret_u8_u32(vdup_n_u32(src)));
1.115 + vdst_wide = vmull_u8(visa, vreinterpret_u8_u32(vdup_n_u32(dst)));
1.116 +
1.117 + vsrc_wide += vdupq_n_u16(128);
1.118 + vsrc_wide += vshrq_n_u16(vsrc_wide, 8);
1.119 +
1.120 + vdst_wide += vdupq_n_u16(128);
1.121 + vdst_wide += vshrq_n_u16(vdst_wide, 8);
1.122 +
1.123 + vsrc = vshrn_n_u16(vsrc_wide, 8);
1.124 + vdst = vshrn_n_u16(vdst_wide, 8);
1.125 +
1.126 + vsrc += vdst;
1.127 + vsrc = vset_lane_u8(da, vsrc, 3);
1.128 +
1.129 + return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
1.130 +}
1.131 +
1.132 +// kDstATop_Mode, //!< [Sa, Sa * Dc + Sc * (1 - Da)]
1.133 +SkPMColor dstatop_modeproc_neon(SkPMColor src, SkPMColor dst) {
1.134 + unsigned sa = SkGetPackedA32(src);
1.135 + unsigned da = SkGetPackedA32(dst);
1.136 + unsigned ida = 255 - da;
1.137 +
1.138 + uint8x8_t vsa, vida, vsrc, vdst;
1.139 +
1.140 + vsa = vdup_n_u8(sa);
1.141 + vida = vdup_n_u8(ida);
1.142 +
1.143 + uint16x8_t vsrc_wide, vdst_wide;
1.144 + vsrc_wide = vmull_u8(vida, vreinterpret_u8_u32(vdup_n_u32(src)));
1.145 + vdst_wide = vmull_u8(vsa, vreinterpret_u8_u32(vdup_n_u32(dst)));
1.146 +
1.147 + vsrc_wide += vdupq_n_u16(128);
1.148 + vsrc_wide += vshrq_n_u16(vsrc_wide, 8);
1.149 +
1.150 + vdst_wide += vdupq_n_u16(128);
1.151 + vdst_wide += vshrq_n_u16(vdst_wide, 8);
1.152 +
1.153 + vsrc = vshrn_n_u16(vsrc_wide, 8);
1.154 + vdst = vshrn_n_u16(vdst_wide, 8);
1.155 +
1.156 + vsrc += vdst;
1.157 + vsrc = vset_lane_u8(sa, vsrc, 3);
1.158 +
1.159 + return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
1.160 +}
1.161 +
1.162 +// kXor_Mode [Sa + Da - 2 * Sa * Da, Sc * (1 - Da) + (1 - Sa) * Dc]
1.163 +SkPMColor xor_modeproc_neon(SkPMColor src, SkPMColor dst) {
1.164 + unsigned sa = SkGetPackedA32(src);
1.165 + unsigned da = SkGetPackedA32(dst);
1.166 + unsigned ret_alpha = sa + da - (SkAlphaMulAlpha(sa, da) << 1);
1.167 + unsigned isa = 255 - sa;
1.168 + unsigned ida = 255 - da;
1.169 +
1.170 + uint8x8_t vsrc, vdst, visa, vida;
1.171 + uint16x8_t vsrc_wide, vdst_wide;
1.172 +
1.173 + visa = vdup_n_u8(isa);
1.174 + vida = vdup_n_u8(ida);
1.175 + vsrc = vreinterpret_u8_u32(vdup_n_u32(src));
1.176 + vdst = vreinterpret_u8_u32(vdup_n_u32(dst));
1.177 +
1.178 + vsrc_wide = vmull_u8(vsrc, vida);
1.179 + vdst_wide = vmull_u8(vdst, visa);
1.180 +
1.181 + vsrc_wide += vdupq_n_u16(128);
1.182 + vsrc_wide += vshrq_n_u16(vsrc_wide, 8);
1.183 +
1.184 + vdst_wide += vdupq_n_u16(128);
1.185 + vdst_wide += vshrq_n_u16(vdst_wide, 8);
1.186 +
1.187 + vsrc = vshrn_n_u16(vsrc_wide, 8);
1.188 + vdst = vshrn_n_u16(vdst_wide, 8);
1.189 +
1.190 + vsrc += vdst;
1.191 +
1.192 + vsrc = vset_lane_u8(ret_alpha, vsrc, 3);
1.193 +
1.194 + return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
1.195 +}
1.196 +
1.197 +// kPlus_Mode
1.198 +SkPMColor plus_modeproc_neon(SkPMColor src, SkPMColor dst) {
1.199 + uint8x8_t vsrc, vdst;
1.200 + vsrc = vreinterpret_u8_u32(vdup_n_u32(src));
1.201 + vdst = vreinterpret_u8_u32(vdup_n_u32(dst));
1.202 + vsrc = vqadd_u8(vsrc, vdst);
1.203 +
1.204 + return vget_lane_u32(vreinterpret_u32_u8(vsrc), 0);
1.205 +}
1.206 +
1.207 +// kModulate_Mode
1.208 +SkPMColor modulate_modeproc_neon(SkPMColor src, SkPMColor dst) {
1.209 + uint8x8_t vsrc, vdst, vres;
1.210 + uint16x8_t vres_wide;
1.211 +
1.212 + vsrc = vreinterpret_u8_u32(vdup_n_u32(src));
1.213 + vdst = vreinterpret_u8_u32(vdup_n_u32(dst));
1.214 +
1.215 + vres_wide = vmull_u8(vsrc, vdst);
1.216 +
1.217 + vres_wide += vdupq_n_u16(128);
1.218 + vres_wide += vshrq_n_u16(vres_wide, 8);
1.219 +
1.220 + vres = vshrn_n_u16(vres_wide, 8);
1.221 +
1.222 + return vget_lane_u32(vreinterpret_u32_u8(vres), 0);
1.223 +}
1.224 +
1.225 +////////////////////////////////////////////////////////////////////////////////
1.226 +// 8 pixels modeprocs
1.227 +////////////////////////////////////////////////////////////////////////////////
1.228 +
1.229 +uint8x8x4_t dstover_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.230 + uint8x8x4_t ret;
1.231 + uint16x8_t src_scale;
1.232 +
1.233 + src_scale = vsubw_u8(vdupq_n_u16(256), dst.val[NEON_A]);
1.234 +
1.235 + ret.val[NEON_A] = dst.val[NEON_A] + SkAlphaMul_neon8(src.val[NEON_A], src_scale);
1.236 + ret.val[NEON_R] = dst.val[NEON_R] + SkAlphaMul_neon8(src.val[NEON_R], src_scale);
1.237 + ret.val[NEON_G] = dst.val[NEON_G] + SkAlphaMul_neon8(src.val[NEON_G], src_scale);
1.238 + ret.val[NEON_B] = dst.val[NEON_B] + SkAlphaMul_neon8(src.val[NEON_B], src_scale);
1.239 +
1.240 + return ret;
1.241 +}
1.242 +
1.243 +uint8x8x4_t srcin_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.244 + uint8x8x4_t ret;
1.245 + uint16x8_t scale;
1.246 +
1.247 + scale = SkAlpha255To256_neon8(dst.val[NEON_A]);
1.248 +
1.249 + ret.val[NEON_A] = SkAlphaMul_neon8(src.val[NEON_A], scale);
1.250 + ret.val[NEON_R] = SkAlphaMul_neon8(src.val[NEON_R], scale);
1.251 + ret.val[NEON_G] = SkAlphaMul_neon8(src.val[NEON_G], scale);
1.252 + ret.val[NEON_B] = SkAlphaMul_neon8(src.val[NEON_B], scale);
1.253 +
1.254 + return ret;
1.255 +}
1.256 +
1.257 +uint8x8x4_t dstin_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.258 + uint8x8x4_t ret;
1.259 + uint16x8_t scale;
1.260 +
1.261 + scale = SkAlpha255To256_neon8(src.val[NEON_A]);
1.262 +
1.263 + ret = SkAlphaMulQ_neon8(dst, scale);
1.264 +
1.265 + return ret;
1.266 +}
1.267 +
1.268 +uint8x8x4_t srcout_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.269 + uint8x8x4_t ret;
1.270 + uint16x8_t scale = vsubw_u8(vdupq_n_u16(256), dst.val[NEON_A]);
1.271 +
1.272 + ret = SkAlphaMulQ_neon8(src, scale);
1.273 +
1.274 + return ret;
1.275 +}
1.276 +
1.277 +uint8x8x4_t dstout_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.278 + uint8x8x4_t ret;
1.279 + uint16x8_t scale = vsubw_u8(vdupq_n_u16(256), src.val[NEON_A]);
1.280 +
1.281 + ret = SkAlphaMulQ_neon8(dst, scale);
1.282 +
1.283 + return ret;
1.284 +}
1.285 +
1.286 +uint8x8x4_t srcatop_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.287 + uint8x8x4_t ret;
1.288 + uint8x8_t isa;
1.289 +
1.290 + isa = vsub_u8(vdup_n_u8(255), src.val[NEON_A]);
1.291 +
1.292 + ret.val[NEON_A] = dst.val[NEON_A];
1.293 + ret.val[NEON_R] = SkAlphaMulAlpha_neon8(src.val[NEON_R], dst.val[NEON_A])
1.294 + + SkAlphaMulAlpha_neon8(dst.val[NEON_R], isa);
1.295 + ret.val[NEON_G] = SkAlphaMulAlpha_neon8(src.val[NEON_G], dst.val[NEON_A])
1.296 + + SkAlphaMulAlpha_neon8(dst.val[NEON_G], isa);
1.297 + ret.val[NEON_B] = SkAlphaMulAlpha_neon8(src.val[NEON_B], dst.val[NEON_A])
1.298 + + SkAlphaMulAlpha_neon8(dst.val[NEON_B], isa);
1.299 +
1.300 + return ret;
1.301 +}
1.302 +
1.303 +uint8x8x4_t dstatop_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.304 + uint8x8x4_t ret;
1.305 + uint8x8_t ida;
1.306 +
1.307 + ida = vsub_u8(vdup_n_u8(255), dst.val[NEON_A]);
1.308 +
1.309 + ret.val[NEON_A] = src.val[NEON_A];
1.310 + ret.val[NEON_R] = SkAlphaMulAlpha_neon8(src.val[NEON_R], ida)
1.311 + + SkAlphaMulAlpha_neon8(dst.val[NEON_R], src.val[NEON_A]);
1.312 + ret.val[NEON_G] = SkAlphaMulAlpha_neon8(src.val[NEON_G], ida)
1.313 + + SkAlphaMulAlpha_neon8(dst.val[NEON_G], src.val[NEON_A]);
1.314 + ret.val[NEON_B] = SkAlphaMulAlpha_neon8(src.val[NEON_B], ida)
1.315 + + SkAlphaMulAlpha_neon8(dst.val[NEON_B], src.val[NEON_A]);
1.316 +
1.317 + return ret;
1.318 +}
1.319 +
1.320 +uint8x8x4_t xor_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.321 + uint8x8x4_t ret;
1.322 + uint8x8_t isa, ida;
1.323 + uint16x8_t tmp_wide, tmp_wide2;
1.324 +
1.325 + isa = vsub_u8(vdup_n_u8(255), src.val[NEON_A]);
1.326 + ida = vsub_u8(vdup_n_u8(255), dst.val[NEON_A]);
1.327 +
1.328 + // First calc alpha
1.329 + tmp_wide = vmovl_u8(src.val[NEON_A]);
1.330 + tmp_wide = vaddw_u8(tmp_wide, dst.val[NEON_A]);
1.331 + tmp_wide2 = vshll_n_u8(SkAlphaMulAlpha_neon8(src.val[NEON_A], dst.val[NEON_A]), 1);
1.332 + tmp_wide = vsubq_u16(tmp_wide, tmp_wide2);
1.333 + ret.val[NEON_A] = vmovn_u16(tmp_wide);
1.334 +
1.335 + // Then colors
1.336 + ret.val[NEON_R] = SkAlphaMulAlpha_neon8(src.val[NEON_R], ida)
1.337 + + SkAlphaMulAlpha_neon8(dst.val[NEON_R], isa);
1.338 + ret.val[NEON_G] = SkAlphaMulAlpha_neon8(src.val[NEON_G], ida)
1.339 + + SkAlphaMulAlpha_neon8(dst.val[NEON_G], isa);
1.340 + ret.val[NEON_B] = SkAlphaMulAlpha_neon8(src.val[NEON_B], ida)
1.341 + + SkAlphaMulAlpha_neon8(dst.val[NEON_B], isa);
1.342 +
1.343 + return ret;
1.344 +}
1.345 +
1.346 +uint8x8x4_t plus_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.347 + uint8x8x4_t ret;
1.348 +
1.349 + ret.val[NEON_A] = vqadd_u8(src.val[NEON_A], dst.val[NEON_A]);
1.350 + ret.val[NEON_R] = vqadd_u8(src.val[NEON_R], dst.val[NEON_R]);
1.351 + ret.val[NEON_G] = vqadd_u8(src.val[NEON_G], dst.val[NEON_G]);
1.352 + ret.val[NEON_B] = vqadd_u8(src.val[NEON_B], dst.val[NEON_B]);
1.353 +
1.354 + return ret;
1.355 +}
1.356 +
1.357 +uint8x8x4_t modulate_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.358 + uint8x8x4_t ret;
1.359 +
1.360 + ret.val[NEON_A] = SkAlphaMulAlpha_neon8(src.val[NEON_A], dst.val[NEON_A]);
1.361 + ret.val[NEON_R] = SkAlphaMulAlpha_neon8(src.val[NEON_R], dst.val[NEON_R]);
1.362 + ret.val[NEON_G] = SkAlphaMulAlpha_neon8(src.val[NEON_G], dst.val[NEON_G]);
1.363 + ret.val[NEON_B] = SkAlphaMulAlpha_neon8(src.val[NEON_B], dst.val[NEON_B]);
1.364 +
1.365 + return ret;
1.366 +}
1.367 +
1.368 +static inline uint8x8_t srcover_color(uint8x8_t a, uint8x8_t b) {
1.369 + uint16x8_t tmp;
1.370 +
1.371 + tmp = vaddl_u8(a, b);
1.372 + tmp -= SkAlphaMulAlpha_neon8_16(a, b);
1.373 +
1.374 + return vmovn_u16(tmp);
1.375 +}
1.376 +
1.377 +uint8x8x4_t screen_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.378 + uint8x8x4_t ret;
1.379 +
1.380 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.381 + ret.val[NEON_R] = srcover_color(src.val[NEON_R], dst.val[NEON_R]);
1.382 + ret.val[NEON_G] = srcover_color(src.val[NEON_G], dst.val[NEON_G]);
1.383 + ret.val[NEON_B] = srcover_color(src.val[NEON_B], dst.val[NEON_B]);
1.384 +
1.385 + return ret;
1.386 +}
1.387 +
1.388 +template <bool overlay>
1.389 +static inline uint8x8_t overlay_hardlight_color(uint8x8_t sc, uint8x8_t dc,
1.390 + uint8x8_t sa, uint8x8_t da) {
1.391 + /*
1.392 + * In the end we're gonna use (rc + tmp) with a different rc
1.393 + * coming from an alternative.
1.394 + * The whole value (rc + tmp) can always be expressed as
1.395 + * VAL = COM - SUB in the if case
1.396 + * VAL = COM + SUB - sa*da in the else case
1.397 + *
1.398 + * with COM = 255 * (sc + dc)
1.399 + * and SUB = sc*da + dc*sa - 2*dc*sc
1.400 + */
1.401 +
1.402 + // Prepare common subexpressions
1.403 + uint16x8_t const255 = vdupq_n_u16(255);
1.404 + uint16x8_t sc_plus_dc = vaddl_u8(sc, dc);
1.405 + uint16x8_t scda = vmull_u8(sc, da);
1.406 + uint16x8_t dcsa = vmull_u8(dc, sa);
1.407 + uint16x8_t sada = vmull_u8(sa, da);
1.408 +
1.409 + // Prepare non common subexpressions
1.410 + uint16x8_t dc2, sc2;
1.411 + uint32x4_t scdc2_1, scdc2_2;
1.412 + if (overlay) {
1.413 + dc2 = vshll_n_u8(dc, 1);
1.414 + scdc2_1 = vmull_u16(vget_low_u16(dc2), vget_low_u16(vmovl_u8(sc)));
1.415 + scdc2_2 = vmull_u16(vget_high_u16(dc2), vget_high_u16(vmovl_u8(sc)));
1.416 + } else {
1.417 + sc2 = vshll_n_u8(sc, 1);
1.418 + scdc2_1 = vmull_u16(vget_low_u16(sc2), vget_low_u16(vmovl_u8(dc)));
1.419 + scdc2_2 = vmull_u16(vget_high_u16(sc2), vget_high_u16(vmovl_u8(dc)));
1.420 + }
1.421 +
1.422 + // Calc COM
1.423 + int32x4_t com1, com2;
1.424 + com1 = vreinterpretq_s32_u32(
1.425 + vmull_u16(vget_low_u16(const255), vget_low_u16(sc_plus_dc)));
1.426 + com2 = vreinterpretq_s32_u32(
1.427 + vmull_u16(vget_high_u16(const255), vget_high_u16(sc_plus_dc)));
1.428 +
1.429 + // Calc SUB
1.430 + int32x4_t sub1, sub2;
1.431 + sub1 = vreinterpretq_s32_u32(vaddl_u16(vget_low_u16(scda), vget_low_u16(dcsa)));
1.432 + sub2 = vreinterpretq_s32_u32(vaddl_u16(vget_high_u16(scda), vget_high_u16(dcsa)));
1.433 + sub1 = vsubq_s32(sub1, vreinterpretq_s32_u32(scdc2_1));
1.434 + sub2 = vsubq_s32(sub2, vreinterpretq_s32_u32(scdc2_2));
1.435 +
1.436 + // Compare 2*dc <= da
1.437 + uint16x8_t cmp;
1.438 +
1.439 + if (overlay) {
1.440 + cmp = vcleq_u16(dc2, vmovl_u8(da));
1.441 + } else {
1.442 + cmp = vcleq_u16(sc2, vmovl_u8(sa));
1.443 + }
1.444 +
1.445 + // Prepare variables
1.446 + int32x4_t val1_1, val1_2;
1.447 + int32x4_t val2_1, val2_2;
1.448 + uint32x4_t cmp1, cmp2;
1.449 +
1.450 + cmp1 = vmovl_u16(vget_low_u16(cmp));
1.451 + cmp1 |= vshlq_n_u32(cmp1, 16);
1.452 + cmp2 = vmovl_u16(vget_high_u16(cmp));
1.453 + cmp2 |= vshlq_n_u32(cmp2, 16);
1.454 +
1.455 + // Calc COM - SUB
1.456 + val1_1 = com1 - sub1;
1.457 + val1_2 = com2 - sub2;
1.458 +
1.459 + // Calc COM + SUB - sa*da
1.460 + val2_1 = com1 + sub1;
1.461 + val2_2 = com2 + sub2;
1.462 +
1.463 + val2_1 = vsubq_s32(val2_1, vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(sada))));
1.464 + val2_2 = vsubq_s32(val2_2, vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(sada))));
1.465 +
1.466 + // Insert where needed
1.467 + val1_1 = vbslq_s32(cmp1, val1_1, val2_1);
1.468 + val1_2 = vbslq_s32(cmp2, val1_2, val2_2);
1.469 +
1.470 + // Call the clamp_div255round function
1.471 + return clamp_div255round_simd8_32(val1_1, val1_2);
1.472 +}
1.473 +
1.474 +static inline uint8x8_t overlay_color(uint8x8_t sc, uint8x8_t dc,
1.475 + uint8x8_t sa, uint8x8_t da) {
1.476 + return overlay_hardlight_color<true>(sc, dc, sa, da);
1.477 +}
1.478 +
1.479 +uint8x8x4_t overlay_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.480 + uint8x8x4_t ret;
1.481 +
1.482 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.483 + ret.val[NEON_R] = overlay_color(src.val[NEON_R], dst.val[NEON_R],
1.484 + src.val[NEON_A], dst.val[NEON_A]);
1.485 + ret.val[NEON_G] = overlay_color(src.val[NEON_G], dst.val[NEON_G],
1.486 + src.val[NEON_A], dst.val[NEON_A]);
1.487 + ret.val[NEON_B] = overlay_color(src.val[NEON_B], dst.val[NEON_B],
1.488 + src.val[NEON_A], dst.val[NEON_A]);
1.489 +
1.490 + return ret;
1.491 +}
1.492 +
1.493 +template <bool lighten>
1.494 +static inline uint8x8_t lighten_darken_color(uint8x8_t sc, uint8x8_t dc,
1.495 + uint8x8_t sa, uint8x8_t da) {
1.496 + uint16x8_t sd, ds, cmp, tmp, tmp2;
1.497 +
1.498 + // Prepare
1.499 + sd = vmull_u8(sc, da);
1.500 + ds = vmull_u8(dc, sa);
1.501 +
1.502 + // Do test
1.503 + if (lighten) {
1.504 + cmp = vcgtq_u16(sd, ds);
1.505 + } else {
1.506 + cmp = vcltq_u16(sd, ds);
1.507 + }
1.508 +
1.509 + // Assign if
1.510 + tmp = vaddl_u8(sc, dc);
1.511 + tmp2 = tmp;
1.512 + tmp -= SkDiv255Round_neon8_16_16(ds);
1.513 +
1.514 + // Calc else
1.515 + tmp2 -= SkDiv255Round_neon8_16_16(sd);
1.516 +
1.517 + // Insert where needed
1.518 + tmp = vbslq_u16(cmp, tmp, tmp2);
1.519 +
1.520 + return vmovn_u16(tmp);
1.521 +}
1.522 +
1.523 +static inline uint8x8_t darken_color(uint8x8_t sc, uint8x8_t dc,
1.524 + uint8x8_t sa, uint8x8_t da) {
1.525 + return lighten_darken_color<false>(sc, dc, sa, da);
1.526 +}
1.527 +
1.528 +uint8x8x4_t darken_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.529 + uint8x8x4_t ret;
1.530 +
1.531 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.532 + ret.val[NEON_R] = darken_color(src.val[NEON_R], dst.val[NEON_R],
1.533 + src.val[NEON_A], dst.val[NEON_A]);
1.534 + ret.val[NEON_G] = darken_color(src.val[NEON_G], dst.val[NEON_G],
1.535 + src.val[NEON_A], dst.val[NEON_A]);
1.536 + ret.val[NEON_B] = darken_color(src.val[NEON_B], dst.val[NEON_B],
1.537 + src.val[NEON_A], dst.val[NEON_A]);
1.538 +
1.539 + return ret;
1.540 +}
1.541 +
1.542 +static inline uint8x8_t lighten_color(uint8x8_t sc, uint8x8_t dc,
1.543 + uint8x8_t sa, uint8x8_t da) {
1.544 + return lighten_darken_color<true>(sc, dc, sa, da);
1.545 +}
1.546 +
1.547 +uint8x8x4_t lighten_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.548 + uint8x8x4_t ret;
1.549 +
1.550 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.551 + ret.val[NEON_R] = lighten_color(src.val[NEON_R], dst.val[NEON_R],
1.552 + src.val[NEON_A], dst.val[NEON_A]);
1.553 + ret.val[NEON_G] = lighten_color(src.val[NEON_G], dst.val[NEON_G],
1.554 + src.val[NEON_A], dst.val[NEON_A]);
1.555 + ret.val[NEON_B] = lighten_color(src.val[NEON_B], dst.val[NEON_B],
1.556 + src.val[NEON_A], dst.val[NEON_A]);
1.557 +
1.558 + return ret;
1.559 +}
1.560 +
1.561 +static inline uint8x8_t hardlight_color(uint8x8_t sc, uint8x8_t dc,
1.562 + uint8x8_t sa, uint8x8_t da) {
1.563 + return overlay_hardlight_color<false>(sc, dc, sa, da);
1.564 +}
1.565 +
1.566 +uint8x8x4_t hardlight_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.567 + uint8x8x4_t ret;
1.568 +
1.569 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.570 + ret.val[NEON_R] = hardlight_color(src.val[NEON_R], dst.val[NEON_R],
1.571 + src.val[NEON_A], dst.val[NEON_A]);
1.572 + ret.val[NEON_G] = hardlight_color(src.val[NEON_G], dst.val[NEON_G],
1.573 + src.val[NEON_A], dst.val[NEON_A]);
1.574 + ret.val[NEON_B] = hardlight_color(src.val[NEON_B], dst.val[NEON_B],
1.575 + src.val[NEON_A], dst.val[NEON_A]);
1.576 +
1.577 + return ret;
1.578 +}
1.579 +
1.580 +static inline uint8x8_t difference_color(uint8x8_t sc, uint8x8_t dc,
1.581 + uint8x8_t sa, uint8x8_t da) {
1.582 + uint16x8_t sd, ds, tmp;
1.583 + int16x8_t val;
1.584 +
1.585 + sd = vmull_u8(sc, da);
1.586 + ds = vmull_u8(dc, sa);
1.587 +
1.588 + tmp = vminq_u16(sd, ds);
1.589 + tmp = SkDiv255Round_neon8_16_16(tmp);
1.590 + tmp = vshlq_n_u16(tmp, 1);
1.591 +
1.592 + val = vreinterpretq_s16_u16(vaddl_u8(sc, dc));
1.593 +
1.594 + val -= vreinterpretq_s16_u16(tmp);
1.595 +
1.596 + val = vmaxq_s16(val, vdupq_n_s16(0));
1.597 + val = vminq_s16(val, vdupq_n_s16(255));
1.598 +
1.599 + return vmovn_u16(vreinterpretq_u16_s16(val));
1.600 +}
1.601 +
1.602 +uint8x8x4_t difference_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.603 + uint8x8x4_t ret;
1.604 +
1.605 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.606 + ret.val[NEON_R] = difference_color(src.val[NEON_R], dst.val[NEON_R],
1.607 + src.val[NEON_A], dst.val[NEON_A]);
1.608 + ret.val[NEON_G] = difference_color(src.val[NEON_G], dst.val[NEON_G],
1.609 + src.val[NEON_A], dst.val[NEON_A]);
1.610 + ret.val[NEON_B] = difference_color(src.val[NEON_B], dst.val[NEON_B],
1.611 + src.val[NEON_A], dst.val[NEON_A]);
1.612 +
1.613 + return ret;
1.614 +}
1.615 +
1.616 +static inline uint8x8_t exclusion_color(uint8x8_t sc, uint8x8_t dc,
1.617 + uint8x8_t sa, uint8x8_t da) {
1.618 + /* The equation can be simplified to 255(sc + dc) - 2 * sc * dc */
1.619 +
1.620 + uint16x8_t sc_plus_dc, scdc, const255;
1.621 + int32x4_t term1_1, term1_2, term2_1, term2_2;
1.622 +
1.623 + /* Calc (sc + dc) and (sc * dc) */
1.624 + sc_plus_dc = vaddl_u8(sc, dc);
1.625 + scdc = vmull_u8(sc, dc);
1.626 +
1.627 + /* Prepare constants */
1.628 + const255 = vdupq_n_u16(255);
1.629 +
1.630 + /* Calc the first term */
1.631 + term1_1 = vreinterpretq_s32_u32(
1.632 + vmull_u16(vget_low_u16(const255), vget_low_u16(sc_plus_dc)));
1.633 + term1_2 = vreinterpretq_s32_u32(
1.634 + vmull_u16(vget_high_u16(const255), vget_high_u16(sc_plus_dc)));
1.635 +
1.636 + /* Calc the second term */
1.637 + term2_1 = vreinterpretq_s32_u32(vshll_n_u16(vget_low_u16(scdc), 1));
1.638 + term2_2 = vreinterpretq_s32_u32(vshll_n_u16(vget_high_u16(scdc), 1));
1.639 +
1.640 + return clamp_div255round_simd8_32(term1_1 - term2_1, term1_2 - term2_2);
1.641 +}
1.642 +
1.643 +uint8x8x4_t exclusion_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.644 + uint8x8x4_t ret;
1.645 +
1.646 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.647 + ret.val[NEON_R] = exclusion_color(src.val[NEON_R], dst.val[NEON_R],
1.648 + src.val[NEON_A], dst.val[NEON_A]);
1.649 + ret.val[NEON_G] = exclusion_color(src.val[NEON_G], dst.val[NEON_G],
1.650 + src.val[NEON_A], dst.val[NEON_A]);
1.651 + ret.val[NEON_B] = exclusion_color(src.val[NEON_B], dst.val[NEON_B],
1.652 + src.val[NEON_A], dst.val[NEON_A]);
1.653 +
1.654 + return ret;
1.655 +}
1.656 +
1.657 +static inline uint8x8_t blendfunc_multiply_color(uint8x8_t sc, uint8x8_t dc,
1.658 + uint8x8_t sa, uint8x8_t da) {
1.659 + uint32x4_t val1, val2;
1.660 + uint16x8_t scdc, t1, t2;
1.661 +
1.662 + t1 = vmull_u8(sc, vdup_n_u8(255) - da);
1.663 + t2 = vmull_u8(dc, vdup_n_u8(255) - sa);
1.664 + scdc = vmull_u8(sc, dc);
1.665 +
1.666 + val1 = vaddl_u16(vget_low_u16(t1), vget_low_u16(t2));
1.667 + val2 = vaddl_u16(vget_high_u16(t1), vget_high_u16(t2));
1.668 +
1.669 + val1 = vaddw_u16(val1, vget_low_u16(scdc));
1.670 + val2 = vaddw_u16(val2, vget_high_u16(scdc));
1.671 +
1.672 + return clamp_div255round_simd8_32(
1.673 + vreinterpretq_s32_u32(val1), vreinterpretq_s32_u32(val2));
1.674 +}
1.675 +
1.676 +uint8x8x4_t multiply_modeproc_neon8(uint8x8x4_t src, uint8x8x4_t dst) {
1.677 + uint8x8x4_t ret;
1.678 +
1.679 + ret.val[NEON_A] = srcover_color(src.val[NEON_A], dst.val[NEON_A]);
1.680 + ret.val[NEON_R] = blendfunc_multiply_color(src.val[NEON_R], dst.val[NEON_R],
1.681 + src.val[NEON_A], dst.val[NEON_A]);
1.682 + ret.val[NEON_G] = blendfunc_multiply_color(src.val[NEON_G], dst.val[NEON_G],
1.683 + src.val[NEON_A], dst.val[NEON_A]);
1.684 + ret.val[NEON_B] = blendfunc_multiply_color(src.val[NEON_B], dst.val[NEON_B],
1.685 + src.val[NEON_A], dst.val[NEON_A]);
1.686 +
1.687 + return ret;
1.688 +}
1.689 +
1.690 +////////////////////////////////////////////////////////////////////////////////
1.691 +
1.692 +typedef uint8x8x4_t (*SkXfermodeProcSIMD)(uint8x8x4_t src, uint8x8x4_t dst);
1.693 +
1.694 +extern SkXfermodeProcSIMD gNEONXfermodeProcs[];
1.695 +
1.696 +SkNEONProcCoeffXfermode::SkNEONProcCoeffXfermode(SkReadBuffer& buffer)
1.697 + : INHERITED(buffer) {
1.698 + fProcSIMD = reinterpret_cast<void*>(gNEONXfermodeProcs[this->getMode()]);
1.699 +}
1.700 +
1.701 +void SkNEONProcCoeffXfermode::xfer32(SkPMColor dst[], const SkPMColor src[],
1.702 + int count, const SkAlpha aa[]) const {
1.703 + SkASSERT(dst && src && count >= 0);
1.704 +
1.705 + SkXfermodeProc proc = this->getProc();
1.706 + SkXfermodeProcSIMD procSIMD = reinterpret_cast<SkXfermodeProcSIMD>(fProcSIMD);
1.707 + SkASSERT(procSIMD != NULL);
1.708 +
1.709 + if (NULL == aa) {
1.710 + // Unrolled NEON code
1.711 + while (count >= 8) {
1.712 + uint8x8x4_t vsrc, vdst, vres;
1.713 +
1.714 +#if (__GNUC__ > 4) || ((__GNUC__ == 4) && (__GNUC_MINOR__ > 6))
1.715 + asm volatile (
1.716 + "vld4.u8 %h[vsrc], [%[src]]! \t\n"
1.717 + "vld4.u8 %h[vdst], [%[dst]] \t\n"
1.718 + : [vsrc] "=w" (vsrc), [vdst] "=w" (vdst), [src] "+&r" (src)
1.719 + : [dst] "r" (dst)
1.720 + :
1.721 + );
1.722 +#else
1.723 + register uint8x8_t d0 asm("d0");
1.724 + register uint8x8_t d1 asm("d1");
1.725 + register uint8x8_t d2 asm("d2");
1.726 + register uint8x8_t d3 asm("d3");
1.727 + register uint8x8_t d4 asm("d4");
1.728 + register uint8x8_t d5 asm("d5");
1.729 + register uint8x8_t d6 asm("d6");
1.730 + register uint8x8_t d7 asm("d7");
1.731 +
1.732 + asm volatile (
1.733 + "vld4.u8 {d0-d3},[%[src]]!;"
1.734 + "vld4.u8 {d4-d7},[%[dst]];"
1.735 + : "=w" (d0), "=w" (d1), "=w" (d2), "=w" (d3),
1.736 + "=w" (d4), "=w" (d5), "=w" (d6), "=w" (d7),
1.737 + [src] "+&r" (src)
1.738 + : [dst] "r" (dst)
1.739 + :
1.740 + );
1.741 + vsrc.val[0] = d0; vdst.val[0] = d4;
1.742 + vsrc.val[1] = d1; vdst.val[1] = d5;
1.743 + vsrc.val[2] = d2; vdst.val[2] = d6;
1.744 + vsrc.val[3] = d3; vdst.val[3] = d7;
1.745 +#endif
1.746 +
1.747 + vres = procSIMD(vsrc, vdst);
1.748 +
1.749 + vst4_u8((uint8_t*)dst, vres);
1.750 +
1.751 + count -= 8;
1.752 + dst += 8;
1.753 + }
1.754 + // Leftovers
1.755 + for (int i = 0; i < count; i++) {
1.756 + dst[i] = proc(src[i], dst[i]);
1.757 + }
1.758 + } else {
1.759 + for (int i = count - 1; i >= 0; --i) {
1.760 + unsigned a = aa[i];
1.761 + if (0 != a) {
1.762 + SkPMColor dstC = dst[i];
1.763 + SkPMColor C = proc(src[i], dstC);
1.764 + if (a != 0xFF) {
1.765 + C = SkFourByteInterp_neon(C, dstC, a);
1.766 + }
1.767 + dst[i] = C;
1.768 + }
1.769 + }
1.770 + }
1.771 +}
1.772 +
1.773 +void SkNEONProcCoeffXfermode::xfer16(uint16_t* SK_RESTRICT dst,
1.774 + const SkPMColor* SK_RESTRICT src, int count,
1.775 + const SkAlpha* SK_RESTRICT aa) const {
1.776 + SkASSERT(dst && src && count >= 0);
1.777 +
1.778 + SkXfermodeProc proc = this->getProc();
1.779 + SkXfermodeProcSIMD procSIMD = reinterpret_cast<SkXfermodeProcSIMD>(fProcSIMD);
1.780 + SkASSERT(procSIMD != NULL);
1.781 +
1.782 + if (NULL == aa) {
1.783 + while(count >= 8) {
1.784 + uint16x8_t vdst, vres16;
1.785 + uint8x8x4_t vdst32, vsrc, vres;
1.786 +
1.787 + vdst = vld1q_u16(dst);
1.788 +
1.789 +#if (__GNUC__ > 4) || ((__GNUC__ == 4) && (__GNUC_MINOR__ > 6))
1.790 + asm volatile (
1.791 + "vld4.u8 %h[vsrc], [%[src]]! \t\n"
1.792 + : [vsrc] "=w" (vsrc), [src] "+&r" (src)
1.793 + : :
1.794 + );
1.795 +#else
1.796 + register uint8x8_t d0 asm("d0");
1.797 + register uint8x8_t d1 asm("d1");
1.798 + register uint8x8_t d2 asm("d2");
1.799 + register uint8x8_t d3 asm("d3");
1.800 +
1.801 + asm volatile (
1.802 + "vld4.u8 {d0-d3},[%[src]]!;"
1.803 + : "=w" (d0), "=w" (d1), "=w" (d2), "=w" (d3),
1.804 + [src] "+&r" (src)
1.805 + : :
1.806 + );
1.807 + vsrc.val[0] = d0;
1.808 + vsrc.val[1] = d1;
1.809 + vsrc.val[2] = d2;
1.810 + vsrc.val[3] = d3;
1.811 +#endif
1.812 +
1.813 + vdst32 = SkPixel16ToPixel32_neon8(vdst);
1.814 + vres = procSIMD(vsrc, vdst32);
1.815 + vres16 = SkPixel32ToPixel16_neon8(vres);
1.816 +
1.817 + vst1q_u16(dst, vres16);
1.818 +
1.819 + count -= 8;
1.820 + dst += 8;
1.821 + }
1.822 + for (int i = 0; i < count; i++) {
1.823 + SkPMColor dstC = SkPixel16ToPixel32(dst[i]);
1.824 + dst[i] = SkPixel32ToPixel16_ToU16(proc(src[i], dstC));
1.825 + }
1.826 + } else {
1.827 + for (int i = count - 1; i >= 0; --i) {
1.828 + unsigned a = aa[i];
1.829 + if (0 != a) {
1.830 + SkPMColor dstC = SkPixel16ToPixel32(dst[i]);
1.831 + SkPMColor C = proc(src[i], dstC);
1.832 + if (0xFF != a) {
1.833 + C = SkFourByteInterp_neon(C, dstC, a);
1.834 + }
1.835 + dst[i] = SkPixel32ToPixel16_ToU16(C);
1.836 + }
1.837 + }
1.838 + }
1.839 +}
1.840 +
1.841 +#ifndef SK_IGNORE_TO_STRING
1.842 +void SkNEONProcCoeffXfermode::toString(SkString* str) const {
1.843 + this->INHERITED::toString(str);
1.844 +}
1.845 +#endif
1.846 +
1.847 +////////////////////////////////////////////////////////////////////////////////
1.848 +
1.849 +SkXfermodeProcSIMD gNEONXfermodeProcs[] = {
1.850 + NULL, // kClear_Mode
1.851 + NULL, // kSrc_Mode
1.852 + NULL, // kDst_Mode
1.853 + NULL, // kSrcOver_Mode
1.854 + dstover_modeproc_neon8,
1.855 + srcin_modeproc_neon8,
1.856 + dstin_modeproc_neon8,
1.857 + srcout_modeproc_neon8,
1.858 + dstout_modeproc_neon8,
1.859 + srcatop_modeproc_neon8,
1.860 + dstatop_modeproc_neon8,
1.861 + xor_modeproc_neon8,
1.862 + plus_modeproc_neon8,
1.863 + modulate_modeproc_neon8,
1.864 + screen_modeproc_neon8,
1.865 +
1.866 + overlay_modeproc_neon8,
1.867 + darken_modeproc_neon8,
1.868 + lighten_modeproc_neon8,
1.869 + NULL, // kColorDodge_Mode
1.870 + NULL, // kColorBurn_Mode
1.871 + hardlight_modeproc_neon8,
1.872 + NULL, // kSoftLight_Mode
1.873 + difference_modeproc_neon8,
1.874 + exclusion_modeproc_neon8,
1.875 + multiply_modeproc_neon8,
1.876 +
1.877 + NULL, // kHue_Mode
1.878 + NULL, // kSaturation_Mode
1.879 + NULL, // kColor_Mode
1.880 + NULL, // kLuminosity_Mode
1.881 +};
1.882 +
1.883 +SK_COMPILE_ASSERT(
1.884 + SK_ARRAY_COUNT(gNEONXfermodeProcs) == SkXfermode::kLastMode + 1,
1.885 + mode_count_arm
1.886 +);
1.887 +
1.888 +SkXfermodeProc gNEONXfermodeProcs1[] = {
1.889 + NULL, // kClear_Mode
1.890 + NULL, // kSrc_Mode
1.891 + NULL, // kDst_Mode
1.892 + NULL, // kSrcOver_Mode
1.893 + NULL, // kDstOver_Mode
1.894 + NULL, // kSrcIn_Mode
1.895 + NULL, // kDstIn_Mode
1.896 + NULL, // kSrcOut_Mode
1.897 + NULL, // kDstOut_Mode
1.898 + srcatop_modeproc_neon,
1.899 + dstatop_modeproc_neon,
1.900 + xor_modeproc_neon,
1.901 + plus_modeproc_neon,
1.902 + modulate_modeproc_neon,
1.903 + NULL, // kScreen_Mode
1.904 +
1.905 + NULL, // kOverlay_Mode
1.906 + NULL, // kDarken_Mode
1.907 + NULL, // kLighten_Mode
1.908 + NULL, // kColorDodge_Mode
1.909 + NULL, // kColorBurn_Mode
1.910 + NULL, // kHardLight_Mode
1.911 + NULL, // kSoftLight_Mode
1.912 + NULL, // kDifference_Mode
1.913 + NULL, // kExclusion_Mode
1.914 + NULL, // kMultiply_Mode
1.915 +
1.916 + NULL, // kHue_Mode
1.917 + NULL, // kSaturation_Mode
1.918 + NULL, // kColor_Mode
1.919 + NULL, // kLuminosity_Mode
1.920 +};
1.921 +
1.922 +SK_COMPILE_ASSERT(
1.923 + SK_ARRAY_COUNT(gNEONXfermodeProcs1) == SkXfermode::kLastMode + 1,
1.924 + mode1_count_arm
1.925 +);
1.926 +
1.927 +SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_neon(const ProcCoeff& rec,
1.928 + SkXfermode::Mode mode) {
1.929 +
1.930 + void* procSIMD = reinterpret_cast<void*>(gNEONXfermodeProcs[mode]);
1.931 +
1.932 + if (procSIMD != NULL) {
1.933 + return SkNEW_ARGS(SkNEONProcCoeffXfermode, (rec, mode, procSIMD));
1.934 + }
1.935 + return NULL;
1.936 +}
1.937 +
1.938 +SkXfermodeProc SkPlatformXfermodeProcFactory_impl_neon(SkXfermode::Mode mode) {
1.939 + return gNEONXfermodeProcs1[mode];
1.940 +}
