1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/skia/trunk/src/opts/SkBitmapProcState_matrixProcs_neon.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,234 @@
1.4 +/* NEON optimized code (C) COPYRIGHT 2009 Motorola
1.5 + *
1.6 + * Use of this source code is governed by a BSD-style license that can be
1.7 + * found in the LICENSE file.
1.8 + */
1.9 +
1.10 +#include "SkBitmapProcState.h"
1.11 +#include "SkPerspIter.h"
1.12 +#include "SkShader.h"
1.13 +#include "SkUtilsArm.h"
1.14 +#include "SkBitmapProcState_utils.h"
1.15 +
1.16 +#include <arm_neon.h>
1.17 +
1.18 +extern const SkBitmapProcState::MatrixProc ClampX_ClampY_Procs_neon[];
1.19 +extern const SkBitmapProcState::MatrixProc RepeatX_RepeatY_Procs_neon[];
1.20 +
1.21 +static void decal_nofilter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count);
1.22 +static void decal_filter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count);
1.23 +
1.24 +// TILEX_PROCF(fx, max) SkClampMax((fx) >> 16, max)
1.25 +static inline int16x8_t sbpsm_clamp_tile8(int32x4_t low, int32x4_t high, unsigned max) {
1.26 + int16x8_t res;
1.27 +
1.28 + // get the hi 16s of all those 32s
1.29 + res = vuzpq_s16(vreinterpretq_s16_s32(low), vreinterpretq_s16_s32(high)).val[1];
1.30 +
1.31 + // clamp
1.32 + res = vmaxq_s16(res, vdupq_n_s16(0));
1.33 + res = vminq_s16(res, vdupq_n_s16(max));
1.34 +
1.35 + return res;
1.36 +}
1.37 +
1.38 +// TILEX_PROCF(fx, max) SkClampMax((fx) >> 16, max)
1.39 +static inline int32x4_t sbpsm_clamp_tile4(int32x4_t f, unsigned max) {
1.40 + int32x4_t res;
1.41 +
1.42 + // get the hi 16s of all those 32s
1.43 + res = vshrq_n_s32(f, 16);
1.44 +
1.45 + // clamp
1.46 + res = vmaxq_s32(res, vdupq_n_s32(0));
1.47 + res = vminq_s32(res, vdupq_n_s32(max));
1.48 +
1.49 + return res;
1.50 +}
1.51 +
1.52 +// TILEY_LOW_BITS(fy, max) (((fy) >> 12) & 0xF)
1.53 +static inline int32x4_t sbpsm_clamp_tile4_low_bits(int32x4_t fx) {
1.54 + int32x4_t ret;
1.55 +
1.56 + ret = vshrq_n_s32(fx, 12);
1.57 +
1.58 + /* We don't need the mask below because the caller will
1.59 + * overwrite the non-masked bits
1.60 + */
1.61 + //ret = vandq_s32(ret, vdupq_n_s32(0xF));
1.62 +
1.63 + return ret;
1.64 +}
1.65 +
1.66 +// TILEX_PROCF(fx, max) (((fx)&0xFFFF)*((max)+1)>> 16)
1.67 +static inline int16x8_t sbpsm_repeat_tile8(int32x4_t low, int32x4_t high, unsigned max) {
1.68 + uint16x8_t res;
1.69 + uint32x4_t tmpl, tmph;
1.70 +
1.71 + // get the lower 16 bits
1.72 + res = vuzpq_u16(vreinterpretq_u16_s32(low), vreinterpretq_u16_s32(high)).val[0];
1.73 +
1.74 + // bare multiplication, not SkFixedMul
1.75 + tmpl = vmull_u16(vget_low_u16(res), vdup_n_u16(max+1));
1.76 + tmph = vmull_u16(vget_high_u16(res), vdup_n_u16(max+1));
1.77 +
1.78 + // extraction of the 16 upper bits
1.79 + res = vuzpq_u16(vreinterpretq_u16_u32(tmpl), vreinterpretq_u16_u32(tmph)).val[1];
1.80 +
1.81 + return vreinterpretq_s16_u16(res);
1.82 +}
1.83 +
1.84 +// TILEX_PROCF(fx, max) (((fx)&0xFFFF)*((max)+1)>> 16)
1.85 +static inline int32x4_t sbpsm_repeat_tile4(int32x4_t f, unsigned max) {
1.86 + uint16x4_t res;
1.87 + uint32x4_t tmp;
1.88 +
1.89 + // get the lower 16 bits
1.90 + res = vmovn_u32(vreinterpretq_u32_s32(f));
1.91 +
1.92 + // bare multiplication, not SkFixedMul
1.93 + tmp = vmull_u16(res, vdup_n_u16(max+1));
1.94 +
1.95 + // extraction of the 16 upper bits
1.96 + tmp = vshrq_n_u32(tmp, 16);
1.97 +
1.98 + return vreinterpretq_s32_u32(tmp);
1.99 +}
1.100 +
1.101 +// TILEX_LOW_BITS(fx, max) ((((fx) & 0xFFFF) * ((max) + 1) >> 12) & 0xF)
1.102 +static inline int32x4_t sbpsm_repeat_tile4_low_bits(int32x4_t fx, unsigned max) {
1.103 + uint16x4_t res;
1.104 + uint32x4_t tmp;
1.105 + int32x4_t ret;
1.106 +
1.107 + // get the lower 16 bits
1.108 + res = vmovn_u32(vreinterpretq_u32_s32(fx));
1.109 +
1.110 + // bare multiplication, not SkFixedMul
1.111 + tmp = vmull_u16(res, vdup_n_u16(max + 1));
1.112 +
1.113 + // shift and mask
1.114 + ret = vshrq_n_s32(vreinterpretq_s32_u32(tmp), 12);
1.115 +
1.116 + /* We don't need the mask below because the caller will
1.117 + * overwrite the non-masked bits
1.118 + */
1.119 + //ret = vandq_s32(ret, vdupq_n_s32(0xF));
1.120 +
1.121 + return ret;
1.122 +}
1.123 +
1.124 +#define MAKENAME(suffix) ClampX_ClampY ## suffix ## _neon
1.125 +#define TILEX_PROCF(fx, max) SkClampMax((fx) >> 16, max)
1.126 +#define TILEY_PROCF(fy, max) SkClampMax((fy) >> 16, max)
1.127 +#define TILEX_PROCF_NEON8(l, h, max) sbpsm_clamp_tile8(l, h, max)
1.128 +#define TILEY_PROCF_NEON8(l, h, max) sbpsm_clamp_tile8(l, h, max)
1.129 +#define TILEX_PROCF_NEON4(fx, max) sbpsm_clamp_tile4(fx, max)
1.130 +#define TILEY_PROCF_NEON4(fy, max) sbpsm_clamp_tile4(fy, max)
1.131 +#define TILEX_LOW_BITS(fx, max) (((fx) >> 12) & 0xF)
1.132 +#define TILEY_LOW_BITS(fy, max) (((fy) >> 12) & 0xF)
1.133 +#define TILEX_LOW_BITS_NEON4(fx, max) sbpsm_clamp_tile4_low_bits(fx)
1.134 +#define TILEY_LOW_BITS_NEON4(fy, max) sbpsm_clamp_tile4_low_bits(fy)
1.135 +#define CHECK_FOR_DECAL
1.136 +#include "SkBitmapProcState_matrix_neon.h"
1.137 +
1.138 +#define MAKENAME(suffix) RepeatX_RepeatY ## suffix ## _neon
1.139 +#define TILEX_PROCF(fx, max) SK_USHIFT16(((fx) & 0xFFFF) * ((max) + 1))
1.140 +#define TILEY_PROCF(fy, max) SK_USHIFT16(((fy) & 0xFFFF) * ((max) + 1))
1.141 +#define TILEX_PROCF_NEON8(l, h, max) sbpsm_repeat_tile8(l, h, max)
1.142 +#define TILEY_PROCF_NEON8(l, h, max) sbpsm_repeat_tile8(l, h, max)
1.143 +#define TILEX_PROCF_NEON4(fx, max) sbpsm_repeat_tile4(fx, max)
1.144 +#define TILEY_PROCF_NEON4(fy, max) sbpsm_repeat_tile4(fy, max)
1.145 +#define TILEX_LOW_BITS(fx, max) ((((fx) & 0xFFFF) * ((max) + 1) >> 12) & 0xF)
1.146 +#define TILEY_LOW_BITS(fy, max) ((((fy) & 0xFFFF) * ((max) + 1) >> 12) & 0xF)
1.147 +#define TILEX_LOW_BITS_NEON4(fx, max) sbpsm_repeat_tile4_low_bits(fx, max)
1.148 +#define TILEY_LOW_BITS_NEON4(fy, max) sbpsm_repeat_tile4_low_bits(fy, max)
1.149 +#include "SkBitmapProcState_matrix_neon.h"
1.150 +
1.151 +
1.152 +
1.153 +void decal_nofilter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count) {
1.154 + if (count >= 8) {
1.155 + // SkFixed is 16.16 fixed point
1.156 + SkFixed dx8 = dx * 8;
1.157 + int32x4_t vdx8 = vdupq_n_s32(dx8);
1.158 +
1.159 + // setup lbase and hbase
1.160 + int32x4_t lbase, hbase;
1.161 + lbase = vdupq_n_s32(fx);
1.162 + lbase = vsetq_lane_s32(fx + dx, lbase, 1);
1.163 + lbase = vsetq_lane_s32(fx + dx + dx, lbase, 2);
1.164 + lbase = vsetq_lane_s32(fx + dx + dx + dx, lbase, 3);
1.165 + hbase = lbase + vdupq_n_s32(4 * dx);
1.166 +
1.167 + do {
1.168 + // store the upper 16 bits
1.169 + vst1q_u32(dst, vreinterpretq_u32_s16(
1.170 + vuzpq_s16(vreinterpretq_s16_s32(lbase), vreinterpretq_s16_s32(hbase)).val[1]
1.171 + ));
1.172 +
1.173 + // on to the next group of 8
1.174 + lbase += vdx8;
1.175 + hbase += vdx8;
1.176 + dst += 4; // we did 8 elements but the result is twice smaller
1.177 + count -= 8;
1.178 + fx += dx8;
1.179 + } while (count >= 8);
1.180 + }
1.181 +
1.182 + uint16_t* xx = (uint16_t*)dst;
1.183 + for (int i = count; i > 0; --i) {
1.184 + *xx++ = SkToU16(fx >> 16); fx += dx;
1.185 + }
1.186 +}
1.187 +
1.188 +void decal_filter_scale_neon(uint32_t dst[], SkFixed fx, SkFixed dx, int count) {
1.189 + if (count >= 8) {
1.190 + SkFixed dx8 = dx * 8;
1.191 + int32x4_t vdx8 = vdupq_n_s32(dx8);
1.192 +
1.193 + int32x4_t wide_fx, wide_fx2;
1.194 + wide_fx = vdupq_n_s32(fx);
1.195 + wide_fx = vsetq_lane_s32(fx + dx, wide_fx, 1);
1.196 + wide_fx = vsetq_lane_s32(fx + dx + dx, wide_fx, 2);
1.197 + wide_fx = vsetq_lane_s32(fx + dx + dx + dx, wide_fx, 3);
1.198 +
1.199 + wide_fx2 = vaddq_s32(wide_fx, vdupq_n_s32(4 * dx));
1.200 +
1.201 + while (count >= 8) {
1.202 + int32x4_t wide_out;
1.203 + int32x4_t wide_out2;
1.204 +
1.205 + wide_out = vshlq_n_s32(vshrq_n_s32(wide_fx, 12), 14);
1.206 + wide_out = wide_out | (vshrq_n_s32(wide_fx,16) + vdupq_n_s32(1));
1.207 +
1.208 + wide_out2 = vshlq_n_s32(vshrq_n_s32(wide_fx2, 12), 14);
1.209 + wide_out2 = wide_out2 | (vshrq_n_s32(wide_fx2,16) + vdupq_n_s32(1));
1.210 +
1.211 + vst1q_u32(dst, vreinterpretq_u32_s32(wide_out));
1.212 + vst1q_u32(dst+4, vreinterpretq_u32_s32(wide_out2));
1.213 +
1.214 + dst += 8;
1.215 + fx += dx8;
1.216 + wide_fx += vdx8;
1.217 + wide_fx2 += vdx8;
1.218 + count -= 8;
1.219 + }
1.220 + }
1.221 +
1.222 + if (count & 1)
1.223 + {
1.224 + SkASSERT((fx >> (16 + 14)) == 0);
1.225 + *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
1.226 + fx += dx;
1.227 + }
1.228 + while ((count -= 2) >= 0)
1.229 + {
1.230 + SkASSERT((fx >> (16 + 14)) == 0);
1.231 + *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
1.232 + fx += dx;
1.233 +
1.234 + *dst++ = (fx >> 12 << 14) | ((fx >> 16) + 1);
1.235 + fx += dx;
1.236 + }
1.237 +}
