1 /* 2 * Copyright 2009 The Android Open Source Project 3 * 4 * Use of this source code is governed by a BSD-style license that can be 5 * found in the LICENSE file. 6 */ 7 8 9 #include "SkBitmapProcState.h" 10 #include "SkColorPriv.h" 11 #include "SkPaint.h" 12 #include "SkTypes.h" 13 #include "SkUtils.h" 14 #include "SkUtilsArm.h" 15 16 #include "SkConvolver.h" 17 18 #if !defined(SK_CPU_ARM64) && SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN) 19 void SI8_D16_nofilter_DX_arm( 20 const SkBitmapProcState& s, 21 const uint32_t* SK_RESTRICT xy, 22 int count, 23 uint16_t* SK_RESTRICT colors) SK_ATTRIBUTE_OPTIMIZE_O1; 24 25 void SI8_D16_nofilter_DX_arm(const SkBitmapProcState& s, 26 const uint32_t* SK_RESTRICT xy, 27 int count, uint16_t* SK_RESTRICT colors) { 28 SkASSERT(count > 0 && colors != NULL); 29 SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)); 30 SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel); 31 32 const uint16_t* SK_RESTRICT table = s.fBitmap->getColorTable()->lock16BitCache(); 33 const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels(); 34 35 // buffer is y32, x16, x16, x16, x16, x16 36 // bump srcAddr to the proper row, since we're told Y never changes 37 SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height()); 38 srcAddr = (const uint8_t*)((const char*)srcAddr + 39 xy[0] * s.fBitmap->rowBytes()); 40 41 uint8_t src; 42 43 if (1 == s.fBitmap->width()) { 44 src = srcAddr[0]; 45 uint16_t dstValue = table[src]; 46 sk_memset16(colors, dstValue, count); 47 } else { 48 int i; 49 int count8 = count >> 3; 50 const uint16_t* SK_RESTRICT xx = (const uint16_t*)(xy + 1); 51 52 asm volatile ( 53 "cmp %[count8], #0 \n\t" // compare loop counter with 0 54 "beq 2f \n\t" // if loop counter == 0, exit 55 "1: \n\t" 56 "ldmia %[xx]!, {r5, r7, r9, r11} \n\t" // load ptrs to pixels 0-7 57 "subs %[count8], %[count8], #1 \n\t" // decrement loop counter 58 "uxth r4, r5 \n\t" // extract ptr 0 59 "mov r5, r5, lsr #16 \n\t" // extract ptr 1 60 "uxth r6, r7 \n\t" // extract ptr 2 61 "mov r7, r7, lsr #16 \n\t" // extract ptr 3 62 "ldrb r4, [%[srcAddr], r4] \n\t" // load pixel 0 from image 63 "uxth r8, r9 \n\t" // extract ptr 4 64 "ldrb r5, [%[srcAddr], r5] \n\t" // load pixel 1 from image 65 "mov r9, r9, lsr #16 \n\t" // extract ptr 5 66 "ldrb r6, [%[srcAddr], r6] \n\t" // load pixel 2 from image 67 "uxth r10, r11 \n\t" // extract ptr 6 68 "ldrb r7, [%[srcAddr], r7] \n\t" // load pixel 3 from image 69 "mov r11, r11, lsr #16 \n\t" // extract ptr 7 70 "ldrb r8, [%[srcAddr], r8] \n\t" // load pixel 4 from image 71 "add r4, r4, r4 \n\t" // double pixel 0 for RGB565 lookup 72 "ldrb r9, [%[srcAddr], r9] \n\t" // load pixel 5 from image 73 "add r5, r5, r5 \n\t" // double pixel 1 for RGB565 lookup 74 "ldrb r10, [%[srcAddr], r10] \n\t" // load pixel 6 from image 75 "add r6, r6, r6 \n\t" // double pixel 2 for RGB565 lookup 76 "ldrb r11, [%[srcAddr], r11] \n\t" // load pixel 7 from image 77 "add r7, r7, r7 \n\t" // double pixel 3 for RGB565 lookup 78 "ldrh r4, [%[table], r4] \n\t" // load pixel 0 RGB565 from colmap 79 "add r8, r8, r8 \n\t" // double pixel 4 for RGB565 lookup 80 "ldrh r5, [%[table], r5] \n\t" // load pixel 1 RGB565 from colmap 81 "add r9, r9, r9 \n\t" // double pixel 5 for RGB565 lookup 82 "ldrh r6, [%[table], r6] \n\t" // load pixel 2 RGB565 from colmap 83 "add r10, r10, r10 \n\t" // double pixel 6 for RGB565 lookup 84 "ldrh r7, [%[table], r7] \n\t" // load pixel 3 RGB565 from colmap 85 "add r11, r11, r11 \n\t" // double pixel 7 for RGB565 lookup 86 "ldrh r8, [%[table], r8] \n\t" // load pixel 4 RGB565 from colmap 87 "ldrh r9, [%[table], r9] \n\t" // load pixel 5 RGB565 from colmap 88 "ldrh r10, [%[table], r10] \n\t" // load pixel 6 RGB565 from colmap 89 "ldrh r11, [%[table], r11] \n\t" // load pixel 7 RGB565 from colmap 90 "pkhbt r5, r4, r5, lsl #16 \n\t" // pack pixels 0 and 1 91 "pkhbt r6, r6, r7, lsl #16 \n\t" // pack pixels 2 and 3 92 "pkhbt r8, r8, r9, lsl #16 \n\t" // pack pixels 4 and 5 93 "pkhbt r10, r10, r11, lsl #16 \n\t" // pack pixels 6 and 7 94 "stmia %[colors]!, {r5, r6, r8, r10} \n\t" // store last 8 pixels 95 "bgt 1b \n\t" // loop if counter > 0 96 "2: \n\t" 97 : [xx] "+r" (xx), [count8] "+r" (count8), [colors] "+r" (colors) 98 : [table] "r" (table), [srcAddr] "r" (srcAddr) 99 : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11" 100 ); 101 102 for (i = (count & 7); i > 0; --i) { 103 src = srcAddr[*xx++]; *colors++ = table[src]; 104 } 105 } 106 107 s.fBitmap->getColorTable()->unlock16BitCache(); 108 } 109 110 void SI8_opaque_D32_nofilter_DX_arm( 111 const SkBitmapProcState& s, 112 const uint32_t* SK_RESTRICT xy, 113 int count, 114 SkPMColor* SK_RESTRICT colors) SK_ATTRIBUTE_OPTIMIZE_O1; 115 116 void SI8_opaque_D32_nofilter_DX_arm(const SkBitmapProcState& s, 117 const uint32_t* SK_RESTRICT xy, 118 int count, SkPMColor* SK_RESTRICT colors) { 119 SkASSERT(count > 0 && colors != NULL); 120 SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)); 121 SkASSERT(SkPaint::kNone_FilterLevel == s.fFilterLevel); 122 123 const SkPMColor* SK_RESTRICT table = s.fBitmap->getColorTable()->lockColors(); 124 const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels(); 125 126 // buffer is y32, x16, x16, x16, x16, x16 127 // bump srcAddr to the proper row, since we're told Y never changes 128 SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height()); 129 srcAddr = (const uint8_t*)((const char*)srcAddr + xy[0] * s.fBitmap->rowBytes()); 130 131 if (1 == s.fBitmap->width()) { 132 uint8_t src = srcAddr[0]; 133 SkPMColor dstValue = table[src]; 134 sk_memset32(colors, dstValue, count); 135 } else { 136 const uint16_t* xx = (const uint16_t*)(xy + 1); 137 138 asm volatile ( 139 "subs %[count], %[count], #8 \n\t" // decrement count by 8, set flags 140 "blt 2f \n\t" // if count < 0, branch to singles 141 "1: \n\t" // eights loop 142 "ldmia %[xx]!, {r5, r7, r9, r11} \n\t" // load ptrs to pixels 0-7 143 "uxth r4, r5 \n\t" // extract ptr 0 144 "mov r5, r5, lsr #16 \n\t" // extract ptr 1 145 "uxth r6, r7 \n\t" // extract ptr 2 146 "mov r7, r7, lsr #16 \n\t" // extract ptr 3 147 "ldrb r4, [%[srcAddr], r4] \n\t" // load pixel 0 from image 148 "uxth r8, r9 \n\t" // extract ptr 4 149 "ldrb r5, [%[srcAddr], r5] \n\t" // load pixel 1 from image 150 "mov r9, r9, lsr #16 \n\t" // extract ptr 5 151 "ldrb r6, [%[srcAddr], r6] \n\t" // load pixel 2 from image 152 "uxth r10, r11 \n\t" // extract ptr 6 153 "ldrb r7, [%[srcAddr], r7] \n\t" // load pixel 3 from image 154 "mov r11, r11, lsr #16 \n\t" // extract ptr 7 155 "ldrb r8, [%[srcAddr], r8] \n\t" // load pixel 4 from image 156 "ldrb r9, [%[srcAddr], r9] \n\t" // load pixel 5 from image 157 "ldrb r10, [%[srcAddr], r10] \n\t" // load pixel 6 from image 158 "ldrb r11, [%[srcAddr], r11] \n\t" // load pixel 7 from image 159 "ldr r4, [%[table], r4, lsl #2] \n\t" // load pixel 0 SkPMColor from colmap 160 "ldr r5, [%[table], r5, lsl #2] \n\t" // load pixel 1 SkPMColor from colmap 161 "ldr r6, [%[table], r6, lsl #2] \n\t" // load pixel 2 SkPMColor from colmap 162 "ldr r7, [%[table], r7, lsl #2] \n\t" // load pixel 3 SkPMColor from colmap 163 "ldr r8, [%[table], r8, lsl #2] \n\t" // load pixel 4 SkPMColor from colmap 164 "ldr r9, [%[table], r9, lsl #2] \n\t" // load pixel 5 SkPMColor from colmap 165 "ldr r10, [%[table], r10, lsl #2] \n\t" // load pixel 6 SkPMColor from colmap 166 "ldr r11, [%[table], r11, lsl #2] \n\t" // load pixel 7 SkPMColor from colmap 167 "subs %[count], %[count], #8 \n\t" // decrement loop counter 168 "stmia %[colors]!, {r4-r11} \n\t" // store 8 pixels 169 "bge 1b \n\t" // loop if counter >= 0 170 "2: \n\t" 171 "adds %[count], %[count], #8 \n\t" // fix up counter, set flags 172 "beq 4f \n\t" // if count == 0, branch to exit 173 "3: \n\t" // singles loop 174 "ldrh r4, [%[xx]], #2 \n\t" // load pixel ptr 175 "subs %[count], %[count], #1 \n\t" // decrement loop counter 176 "ldrb r5, [%[srcAddr], r4] \n\t" // load pixel from image 177 "ldr r6, [%[table], r5, lsl #2] \n\t" // load SkPMColor from colmap 178 "str r6, [%[colors]], #4 \n\t" // store pixel, update ptr 179 "bne 3b \n\t" // loop if counter != 0 180 "4: \n\t" // exit 181 : [xx] "+r" (xx), [count] "+r" (count), [colors] "+r" (colors) 182 : [table] "r" (table), [srcAddr] "r" (srcAddr) 183 : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11" 184 ); 185 } 186 187 s.fBitmap->getColorTable()->unlockColors(); 188 } 189 #endif // !defined(SK_CPU_ARM64) && SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN) 190 191 /////////////////////////////////////////////////////////////////////////////// 192 193 /* If we replace a sampleproc, then we null-out the associated shaderproc, 194 otherwise the shader won't even look at the matrix/sampler 195 */ 196 void SkBitmapProcState::platformProcs() { 197 #if !defined(SK_CPU_ARM64) && SK_ARM_ARCH >= 6 && !defined(SK_CPU_BENDIAN) 198 bool isOpaque = 256 == fAlphaScale; 199 bool justDx = false; 200 201 if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) { 202 justDx = true; 203 } 204 205 switch (fBitmap->colorType()) { 206 case kIndex_8_SkColorType: 207 if (justDx && SkPaint::kNone_FilterLevel == fFilterLevel) { 208 #if 0 /* crashing on android device */ 209 fSampleProc16 = SI8_D16_nofilter_DX_arm; 210 fShaderProc16 = NULL; 211 #endif 212 if (isOpaque) { 213 // this one is only very slighty faster than the C version 214 fSampleProc32 = SI8_opaque_D32_nofilter_DX_arm; 215 fShaderProc32 = NULL; 216 } 217 } 218 break; 219 default: 220 break; 221 } 222 #endif 223 } 224 225 /////////////////////////////////////////////////////////////////////////////// 226 227 extern void platformConvolutionProcs_arm_neon(SkConvolutionProcs* procs); 228 229 void platformConvolutionProcs_arm(SkConvolutionProcs* procs) { 230 } 231 232 void SkBitmapProcState::platformConvolutionProcs(SkConvolutionProcs* procs) { 233 SK_ARM_NEON_WRAP(platformConvolutionProcs_arm)(procs); 234 } 235
