1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/skia/trunk/src/opts/SkBitmapProcState_opts_arm.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,234 @@
1.4 +/*
1.5 + * Copyright 2009 The Android Open Source Project
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license that can be
1.8 + * found in the LICENSE file.
1.9 + */
1.10 +
1.11 +
1.12 +#include "SkBitmapProcState.h"
1.13 +#include "SkColorPriv.h"
1.14 +#include "SkPaint.h"
1.15 +#include "SkTypes.h"
1.16 +#include "SkUtils.h"
1.17 +#include "SkUtilsArm.h"
1.18 +
1.19 +#include "SkConvolver.h"
1.20 +
1.21 +#if SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN)
1.22 +void SI8_D16_nofilter_DX_arm(
1.23 + const SkBitmapProcState& s,
1.24 + const uint32_t* SK_RESTRICT xy,
1.25 + int count,
1.26 + uint16_t* SK_RESTRICT colors) SK_ATTRIBUTE_OPTIMIZE_O1;
1.27 +
1.28 +void SI8_D16_nofilter_DX_arm(const SkBitmapProcState& s,
1.29 + const uint32_t* SK_RESTRICT xy,
1.30 + int count, uint16_t* SK_RESTRICT colors) {
1.31 + SkASSERT(count > 0 && colors != NULL);
1.32 + SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask));
1.33 + SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
1.34 +
1.35 + const uint16_t* SK_RESTRICT table = s.fBitmap->getColorTable()->lock16BitCache();
1.36 + const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels();
1.37 +
1.38 + // buffer is y32, x16, x16, x16, x16, x16
1.39 + // bump srcAddr to the proper row, since we're told Y never changes
1.40 + SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height());
1.41 + srcAddr = (const uint8_t*)((const char*)srcAddr +
1.42 + xy[0] * s.fBitmap->rowBytes());
1.43 +
1.44 + uint8_t src;
1.45 +
1.46 + if (1 == s.fBitmap->width()) {
1.47 + src = srcAddr[0];
1.48 + uint16_t dstValue = table[src];
1.49 + sk_memset16(colors, dstValue, count);
1.50 + } else {
1.51 + int i;
1.52 + int count8 = count >> 3;
1.53 + const uint16_t* SK_RESTRICT xx = (const uint16_t*)(xy + 1);
1.54 +
1.55 + asm volatile (
1.56 + "cmp %[count8], #0 \n\t" // compare loop counter with 0
1.57 + "beq 2f \n\t" // if loop counter == 0, exit
1.58 + "1: \n\t"
1.59 + "ldmia %[xx]!, {r5, r7, r9, r11} \n\t" // load ptrs to pixels 0-7
1.60 + "subs %[count8], %[count8], #1 \n\t" // decrement loop counter
1.61 + "uxth r4, r5 \n\t" // extract ptr 0
1.62 + "mov r5, r5, lsr #16 \n\t" // extract ptr 1
1.63 + "uxth r6, r7 \n\t" // extract ptr 2
1.64 + "mov r7, r7, lsr #16 \n\t" // extract ptr 3
1.65 + "ldrb r4, [%[srcAddr], r4] \n\t" // load pixel 0 from image
1.66 + "uxth r8, r9 \n\t" // extract ptr 4
1.67 + "ldrb r5, [%[srcAddr], r5] \n\t" // load pixel 1 from image
1.68 + "mov r9, r9, lsr #16 \n\t" // extract ptr 5
1.69 + "ldrb r6, [%[srcAddr], r6] \n\t" // load pixel 2 from image
1.70 + "uxth r10, r11 \n\t" // extract ptr 6
1.71 + "ldrb r7, [%[srcAddr], r7] \n\t" // load pixel 3 from image
1.72 + "mov r11, r11, lsr #16 \n\t" // extract ptr 7
1.73 + "ldrb r8, [%[srcAddr], r8] \n\t" // load pixel 4 from image
1.74 + "add r4, r4, r4 \n\t" // double pixel 0 for RGB565 lookup
1.75 + "ldrb r9, [%[srcAddr], r9] \n\t" // load pixel 5 from image
1.76 + "add r5, r5, r5 \n\t" // double pixel 1 for RGB565 lookup
1.77 + "ldrb r10, [%[srcAddr], r10] \n\t" // load pixel 6 from image
1.78 + "add r6, r6, r6 \n\t" // double pixel 2 for RGB565 lookup
1.79 + "ldrb r11, [%[srcAddr], r11] \n\t" // load pixel 7 from image
1.80 + "add r7, r7, r7 \n\t" // double pixel 3 for RGB565 lookup
1.81 + "ldrh r4, [%[table], r4] \n\t" // load pixel 0 RGB565 from colmap
1.82 + "add r8, r8, r8 \n\t" // double pixel 4 for RGB565 lookup
1.83 + "ldrh r5, [%[table], r5] \n\t" // load pixel 1 RGB565 from colmap
1.84 + "add r9, r9, r9 \n\t" // double pixel 5 for RGB565 lookup
1.85 + "ldrh r6, [%[table], r6] \n\t" // load pixel 2 RGB565 from colmap
1.86 + "add r10, r10, r10 \n\t" // double pixel 6 for RGB565 lookup
1.87 + "ldrh r7, [%[table], r7] \n\t" // load pixel 3 RGB565 from colmap
1.88 + "add r11, r11, r11 \n\t" // double pixel 7 for RGB565 lookup
1.89 + "ldrh r8, [%[table], r8] \n\t" // load pixel 4 RGB565 from colmap
1.90 + "ldrh r9, [%[table], r9] \n\t" // load pixel 5 RGB565 from colmap
1.91 + "ldrh r10, [%[table], r10] \n\t" // load pixel 6 RGB565 from colmap
1.92 + "ldrh r11, [%[table], r11] \n\t" // load pixel 7 RGB565 from colmap
1.93 + "pkhbt r5, r4, r5, lsl #16 \n\t" // pack pixels 0 and 1
1.94 + "pkhbt r6, r6, r7, lsl #16 \n\t" // pack pixels 2 and 3
1.95 + "pkhbt r8, r8, r9, lsl #16 \n\t" // pack pixels 4 and 5
1.96 + "pkhbt r10, r10, r11, lsl #16 \n\t" // pack pixels 6 and 7
1.97 + "stmia %[colors]!, {r5, r6, r8, r10} \n\t" // store last 8 pixels
1.98 + "bgt 1b \n\t" // loop if counter > 0
1.99 + "2: \n\t"
1.100 + : [xx] "+r" (xx), [count8] "+r" (count8), [colors] "+r" (colors)
1.101 + : [table] "r" (table), [srcAddr] "r" (srcAddr)
1.102 + : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
1.103 + );
1.104 +
1.105 + for (i = (count & 7); i > 0; --i) {
1.106 + src = srcAddr[*xx++]; *colors++ = table[src];
1.107 + }
1.108 + }
1.109 +
1.110 + s.fBitmap->getColorTable()->unlock16BitCache();
1.111 +}
1.112 +
1.113 +void SI8_opaque_D32_nofilter_DX_arm(
1.114 + const SkBitmapProcState& s,
1.115 + const uint32_t* SK_RESTRICT xy,
1.116 + int count,
1.117 + SkPMColor* SK_RESTRICT colors) SK_ATTRIBUTE_OPTIMIZE_O1;
1.118 +
1.119 +void SI8_opaque_D32_nofilter_DX_arm(const SkBitmapProcState& s,
1.120 + const uint32_t* SK_RESTRICT xy,
1.121 + int count, SkPMColor* SK_RESTRICT colors) {
1.122 + SkASSERT(count > 0 && colors != NULL);
1.123 + SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask));
1.124 + SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel);
1.125 +
1.126 + const SkPMColor* SK_RESTRICT table = s.fBitmap->getColorTable()->lockColors();
1.127 + const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels();
1.128 +
1.129 + // buffer is y32, x16, x16, x16, x16, x16
1.130 + // bump srcAddr to the proper row, since we're told Y never changes
1.131 + SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height());
1.132 + srcAddr = (const uint8_t*)((const char*)srcAddr + xy[0] * s.fBitmap->rowBytes());
1.133 +
1.134 + if (1 == s.fBitmap->width()) {
1.135 + uint8_t src = srcAddr[0];
1.136 + SkPMColor dstValue = table[src];
1.137 + sk_memset32(colors, dstValue, count);
1.138 + } else {
1.139 + const uint16_t* xx = (const uint16_t*)(xy + 1);
1.140 +
1.141 + asm volatile (
1.142 + "subs %[count], %[count], #8 \n\t" // decrement count by 8, set flags
1.143 + "blt 2f \n\t" // if count < 0, branch to singles
1.144 + "1: \n\t" // eights loop
1.145 + "ldmia %[xx]!, {r5, r7, r9, r11} \n\t" // load ptrs to pixels 0-7
1.146 + "uxth r4, r5 \n\t" // extract ptr 0
1.147 + "mov r5, r5, lsr #16 \n\t" // extract ptr 1
1.148 + "uxth r6, r7 \n\t" // extract ptr 2
1.149 + "mov r7, r7, lsr #16 \n\t" // extract ptr 3
1.150 + "ldrb r4, [%[srcAddr], r4] \n\t" // load pixel 0 from image
1.151 + "uxth r8, r9 \n\t" // extract ptr 4
1.152 + "ldrb r5, [%[srcAddr], r5] \n\t" // load pixel 1 from image
1.153 + "mov r9, r9, lsr #16 \n\t" // extract ptr 5
1.154 + "ldrb r6, [%[srcAddr], r6] \n\t" // load pixel 2 from image
1.155 + "uxth r10, r11 \n\t" // extract ptr 6
1.156 + "ldrb r7, [%[srcAddr], r7] \n\t" // load pixel 3 from image
1.157 + "mov r11, r11, lsr #16 \n\t" // extract ptr 7
1.158 + "ldrb r8, [%[srcAddr], r8] \n\t" // load pixel 4 from image
1.159 + "ldrb r9, [%[srcAddr], r9] \n\t" // load pixel 5 from image
1.160 + "ldrb r10, [%[srcAddr], r10] \n\t" // load pixel 6 from image
1.161 + "ldrb r11, [%[srcAddr], r11] \n\t" // load pixel 7 from image
1.162 + "ldr r4, [%[table], r4, lsl #2] \n\t" // load pixel 0 SkPMColor from colmap
1.163 + "ldr r5, [%[table], r5, lsl #2] \n\t" // load pixel 1 SkPMColor from colmap
1.164 + "ldr r6, [%[table], r6, lsl #2] \n\t" // load pixel 2 SkPMColor from colmap
1.165 + "ldr r7, [%[table], r7, lsl #2] \n\t" // load pixel 3 SkPMColor from colmap
1.166 + "ldr r8, [%[table], r8, lsl #2] \n\t" // load pixel 4 SkPMColor from colmap
1.167 + "ldr r9, [%[table], r9, lsl #2] \n\t" // load pixel 5 SkPMColor from colmap
1.168 + "ldr r10, [%[table], r10, lsl #2] \n\t" // load pixel 6 SkPMColor from colmap
1.169 + "ldr r11, [%[table], r11, lsl #2] \n\t" // load pixel 7 SkPMColor from colmap
1.170 + "subs %[count], %[count], #8 \n\t" // decrement loop counter
1.171 + "stmia %[colors]!, {r4-r11} \n\t" // store 8 pixels
1.172 + "bge 1b \n\t" // loop if counter >= 0
1.173 + "2: \n\t"
1.174 + "adds %[count], %[count], #8 \n\t" // fix up counter, set flags
1.175 + "beq 4f \n\t" // if count == 0, branch to exit
1.176 + "3: \n\t" // singles loop
1.177 + "ldrh r4, [%[xx]], #2 \n\t" // load pixel ptr
1.178 + "subs %[count], %[count], #1 \n\t" // decrement loop counter
1.179 + "ldrb r5, [%[srcAddr], r4] \n\t" // load pixel from image
1.180 + "ldr r6, [%[table], r5, lsl #2] \n\t" // load SkPMColor from colmap
1.181 + "str r6, [%[colors]], #4 \n\t" // store pixel, update ptr
1.182 + "bne 3b \n\t" // loop if counter != 0
1.183 + "4: \n\t" // exit
1.184 + : [xx] "+r" (xx), [count] "+r" (count), [colors] "+r" (colors)
1.185 + : [table] "r" (table), [srcAddr] "r" (srcAddr)
1.186 + : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11"
1.187 + );
1.188 + }
1.189 +
1.190 + s.fBitmap->getColorTable()->unlockColors();
1.191 +}
1.192 +#endif // SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN)
1.193 +
1.194 +///////////////////////////////////////////////////////////////////////////////
1.195 +
1.196 +/* If we replace a sampleproc, then we null-out the associated shaderproc,
1.197 + otherwise the shader won't even look at the matrix/sampler
1.198 + */
1.199 +void SkBitmapProcState::platformProcs() {
1.200 + bool isOpaque = 256 == fAlphaScale;
1.201 + bool justDx = false;
1.202 +
1.203 + if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) {
1.204 + justDx = true;
1.205 + }
1.206 +
1.207 + switch (fBitmap->config()) {
1.208 + case SkBitmap::kIndex8_Config:
1.209 +#if SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN)
1.210 + if (justDx && SkPaint::kNone_FilterLevel == fFilterLevel) {
1.211 +#if 0 /* crashing on android device */
1.212 + fSampleProc16 = SI8_D16_nofilter_DX_arm;
1.213 + fShaderProc16 = NULL;
1.214 +#endif
1.215 + if (isOpaque) {
1.216 + // this one is only very slighty faster than the C version
1.217 + fSampleProc32 = SI8_opaque_D32_nofilter_DX_arm;
1.218 + fShaderProc32 = NULL;
1.219 + }
1.220 + }
1.221 +#endif
1.222 + break;
1.223 + default:
1.224 + break;
1.225 + }
1.226 +}
1.227 +
1.228 +///////////////////////////////////////////////////////////////////////////////
1.229 +
1.230 +extern void platformConvolutionProcs_arm_neon(SkConvolutionProcs* procs);
1.231 +
1.232 +void platformConvolutionProcs_arm(SkConvolutionProcs* procs) {
1.233 +}
1.234 +
1.235 +void SkBitmapProcState::platformConvolutionProcs(SkConvolutionProcs* procs) {
1.236 + SK_ARM_NEON_WRAP(platformConvolutionProcs_arm)(procs);
1.237 +}
