epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 1 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 2 | /* |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 3 | * Copyright 2009 The Android Open Source Project |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 4 | * |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 5 | * Use of this source code is governed by a BSD-style license that can be |
| 6 | * found in the LICENSE file. |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 7 | */ |
| 8 | |
epoger@google.com | ec3ed6a | 2011-07-28 14:26:00 +0000 | [diff] [blame] | 9 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 10 | #include "SkBitmapProcState.h" |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 11 | #include "SkColorPriv.h" |
| 12 | #include "SkUtils.h" |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 13 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 14 | #if __ARM_ARCH__ >= 6 && !defined(SK_CPU_BENDIAN) |
vandebo@chromium.org | db8b51c | 2011-05-03 02:23:11 +0000 | [diff] [blame] | 15 | void SI8_D16_nofilter_DX_arm( |
| 16 | const SkBitmapProcState& s, |
| 17 | const uint32_t* SK_RESTRICT xy, |
| 18 | int count, |
| 19 | uint16_t* SK_RESTRICT colors) __attribute__((optimize("O1"))); |
| 20 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 21 | void SI8_D16_nofilter_DX_arm(const SkBitmapProcState& s, |
| 22 | const uint32_t* SK_RESTRICT xy, |
| 23 | int count, uint16_t* SK_RESTRICT colors) { |
| 24 | SkASSERT(count > 0 && colors != NULL); |
| 25 | SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)); |
| 26 | SkASSERT(s.fDoFilter == false); |
| 27 | |
| 28 | const uint16_t* SK_RESTRICT table = s.fBitmap->getColorTable()->lock16BitCache(); |
| 29 | const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels(); |
| 30 | |
| 31 | // buffer is y32, x16, x16, x16, x16, x16 |
| 32 | // bump srcAddr to the proper row, since we're told Y never changes |
| 33 | SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height()); |
| 34 | srcAddr = (const uint8_t*)((const char*)srcAddr + |
| 35 | xy[0] * s.fBitmap->rowBytes()); |
| 36 | |
| 37 | uint8_t src; |
| 38 | |
| 39 | if (1 == s.fBitmap->width()) { |
| 40 | src = srcAddr[0]; |
| 41 | uint16_t dstValue = table[src]; |
| 42 | sk_memset16(colors, dstValue, count); |
| 43 | } else { |
| 44 | int i; |
| 45 | int count8 = count >> 3; |
| 46 | const uint16_t* SK_RESTRICT xx = (const uint16_t*)(xy + 1); |
| 47 | |
| 48 | asm volatile ( |
| 49 | "cmp %[count8], #0 \n\t" // compare loop counter with 0 |
| 50 | "beq 2f \n\t" // if loop counter == 0, exit |
| 51 | "1: \n\t" |
| 52 | "ldmia %[xx]!, {r5, r7, r9, r11} \n\t" // load ptrs to pixels 0-7 |
| 53 | "subs %[count8], %[count8], #1 \n\t" // decrement loop counter |
| 54 | "uxth r4, r5 \n\t" // extract ptr 0 |
| 55 | "mov r5, r5, lsr #16 \n\t" // extract ptr 1 |
| 56 | "uxth r6, r7 \n\t" // extract ptr 2 |
| 57 | "mov r7, r7, lsr #16 \n\t" // extract ptr 3 |
| 58 | "ldrb r4, [%[srcAddr], r4] \n\t" // load pixel 0 from image |
| 59 | "uxth r8, r9 \n\t" // extract ptr 4 |
| 60 | "ldrb r5, [%[srcAddr], r5] \n\t" // load pixel 1 from image |
| 61 | "mov r9, r9, lsr #16 \n\t" // extract ptr 5 |
| 62 | "ldrb r6, [%[srcAddr], r6] \n\t" // load pixel 2 from image |
| 63 | "uxth r10, r11 \n\t" // extract ptr 6 |
| 64 | "ldrb r7, [%[srcAddr], r7] \n\t" // load pixel 3 from image |
| 65 | "mov r11, r11, lsr #16 \n\t" // extract ptr 7 |
| 66 | "ldrb r8, [%[srcAddr], r8] \n\t" // load pixel 4 from image |
| 67 | "add r4, r4, r4 \n\t" // double pixel 0 for RGB565 lookup |
| 68 | "ldrb r9, [%[srcAddr], r9] \n\t" // load pixel 5 from image |
| 69 | "add r5, r5, r5 \n\t" // double pixel 1 for RGB565 lookup |
| 70 | "ldrb r10, [%[srcAddr], r10] \n\t" // load pixel 6 from image |
| 71 | "add r6, r6, r6 \n\t" // double pixel 2 for RGB565 lookup |
| 72 | "ldrb r11, [%[srcAddr], r11] \n\t" // load pixel 7 from image |
| 73 | "add r7, r7, r7 \n\t" // double pixel 3 for RGB565 lookup |
| 74 | "ldrh r4, [%[table], r4] \n\t" // load pixel 0 RGB565 from colmap |
| 75 | "add r8, r8, r8 \n\t" // double pixel 4 for RGB565 lookup |
| 76 | "ldrh r5, [%[table], r5] \n\t" // load pixel 1 RGB565 from colmap |
| 77 | "add r9, r9, r9 \n\t" // double pixel 5 for RGB565 lookup |
| 78 | "ldrh r6, [%[table], r6] \n\t" // load pixel 2 RGB565 from colmap |
| 79 | "add r10, r10, r10 \n\t" // double pixel 6 for RGB565 lookup |
| 80 | "ldrh r7, [%[table], r7] \n\t" // load pixel 3 RGB565 from colmap |
| 81 | "add r11, r11, r11 \n\t" // double pixel 7 for RGB565 lookup |
| 82 | "ldrh r8, [%[table], r8] \n\t" // load pixel 4 RGB565 from colmap |
| 83 | "ldrh r9, [%[table], r9] \n\t" // load pixel 5 RGB565 from colmap |
| 84 | "ldrh r10, [%[table], r10] \n\t" // load pixel 6 RGB565 from colmap |
| 85 | "ldrh r11, [%[table], r11] \n\t" // load pixel 7 RGB565 from colmap |
| 86 | "pkhbt r5, r4, r5, lsl #16 \n\t" // pack pixels 0 and 1 |
| 87 | "pkhbt r6, r6, r7, lsl #16 \n\t" // pack pixels 2 and 3 |
| 88 | "pkhbt r8, r8, r9, lsl #16 \n\t" // pack pixels 4 and 5 |
| 89 | "pkhbt r10, r10, r11, lsl #16 \n\t" // pack pixels 6 and 7 |
| 90 | "stmia %[colors]!, {r5, r6, r8, r10} \n\t" // store last 8 pixels |
| 91 | "bgt 1b \n\t" // loop if counter > 0 |
| 92 | "2: \n\t" |
| 93 | : [xx] "+r" (xx), [count8] "+r" (count8), [colors] "+r" (colors) |
| 94 | : [table] "r" (table), [srcAddr] "r" (srcAddr) |
| 95 | : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11" |
| 96 | ); |
| 97 | |
| 98 | for (i = (count & 7); i > 0; --i) { |
| 99 | src = srcAddr[*xx++]; *colors++ = table[src]; |
| 100 | } |
| 101 | } |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 102 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 103 | s.fBitmap->getColorTable()->unlock16BitCache(); |
| 104 | } |
| 105 | |
vandebo@chromium.org | db8b51c | 2011-05-03 02:23:11 +0000 | [diff] [blame] | 106 | void SI8_opaque_D32_nofilter_DX_arm( |
| 107 | const SkBitmapProcState& s, |
| 108 | const uint32_t* SK_RESTRICT xy, |
| 109 | int count, |
| 110 | SkPMColor* SK_RESTRICT colors) __attribute__((optimize("O1"))); |
| 111 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 112 | void SI8_opaque_D32_nofilter_DX_arm(const SkBitmapProcState& s, |
| 113 | const uint32_t* SK_RESTRICT xy, |
| 114 | int count, SkPMColor* SK_RESTRICT colors) { |
| 115 | SkASSERT(count > 0 && colors != NULL); |
| 116 | SkASSERT(s.fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)); |
| 117 | SkASSERT(s.fDoFilter == false); |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 118 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 119 | const SkPMColor* SK_RESTRICT table = s.fBitmap->getColorTable()->lockColors(); |
| 120 | const uint8_t* SK_RESTRICT srcAddr = (const uint8_t*)s.fBitmap->getPixels(); |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 121 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 122 | // buffer is y32, x16, x16, x16, x16, x16 |
| 123 | // bump srcAddr to the proper row, since we're told Y never changes |
| 124 | SkASSERT((unsigned)xy[0] < (unsigned)s.fBitmap->height()); |
| 125 | srcAddr = (const uint8_t*)((const char*)srcAddr + xy[0] * s.fBitmap->rowBytes()); |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 126 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 127 | if (1 == s.fBitmap->width()) { |
| 128 | uint8_t src = srcAddr[0]; |
| 129 | SkPMColor dstValue = table[src]; |
| 130 | sk_memset32(colors, dstValue, count); |
| 131 | } else { |
| 132 | const uint16_t* xx = (const uint16_t*)(xy + 1); |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 133 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 134 | asm volatile ( |
| 135 | "subs %[count], %[count], #8 \n\t" // decrement count by 8, set flags |
| 136 | "blt 2f \n\t" // if count < 0, branch to singles |
| 137 | "1: \n\t" // eights loop |
| 138 | "ldmia %[xx]!, {r5, r7, r9, r11} \n\t" // load ptrs to pixels 0-7 |
| 139 | "uxth r4, r5 \n\t" // extract ptr 0 |
| 140 | "mov r5, r5, lsr #16 \n\t" // extract ptr 1 |
| 141 | "uxth r6, r7 \n\t" // extract ptr 2 |
| 142 | "mov r7, r7, lsr #16 \n\t" // extract ptr 3 |
| 143 | "ldrb r4, [%[srcAddr], r4] \n\t" // load pixel 0 from image |
| 144 | "uxth r8, r9 \n\t" // extract ptr 4 |
| 145 | "ldrb r5, [%[srcAddr], r5] \n\t" // load pixel 1 from image |
| 146 | "mov r9, r9, lsr #16 \n\t" // extract ptr 5 |
| 147 | "ldrb r6, [%[srcAddr], r6] \n\t" // load pixel 2 from image |
| 148 | "uxth r10, r11 \n\t" // extract ptr 6 |
| 149 | "ldrb r7, [%[srcAddr], r7] \n\t" // load pixel 3 from image |
| 150 | "mov r11, r11, lsr #16 \n\t" // extract ptr 7 |
| 151 | "ldrb r8, [%[srcAddr], r8] \n\t" // load pixel 4 from image |
| 152 | "ldrb r9, [%[srcAddr], r9] \n\t" // load pixel 5 from image |
| 153 | "ldrb r10, [%[srcAddr], r10] \n\t" // load pixel 6 from image |
| 154 | "ldrb r11, [%[srcAddr], r11] \n\t" // load pixel 7 from image |
| 155 | "ldr r4, [%[table], r4, lsl #2] \n\t" // load pixel 0 SkPMColor from colmap |
| 156 | "ldr r5, [%[table], r5, lsl #2] \n\t" // load pixel 1 SkPMColor from colmap |
| 157 | "ldr r6, [%[table], r6, lsl #2] \n\t" // load pixel 2 SkPMColor from colmap |
| 158 | "ldr r7, [%[table], r7, lsl #2] \n\t" // load pixel 3 SkPMColor from colmap |
| 159 | "ldr r8, [%[table], r8, lsl #2] \n\t" // load pixel 4 SkPMColor from colmap |
| 160 | "ldr r9, [%[table], r9, lsl #2] \n\t" // load pixel 5 SkPMColor from colmap |
| 161 | "ldr r10, [%[table], r10, lsl #2] \n\t" // load pixel 6 SkPMColor from colmap |
| 162 | "ldr r11, [%[table], r11, lsl #2] \n\t" // load pixel 7 SkPMColor from colmap |
| 163 | "subs %[count], %[count], #8 \n\t" // decrement loop counter |
| 164 | "stmia %[colors]!, {r4-r11} \n\t" // store 8 pixels |
| 165 | "bge 1b \n\t" // loop if counter >= 0 |
| 166 | "2: \n\t" |
| 167 | "adds %[count], %[count], #8 \n\t" // fix up counter, set flags |
| 168 | "beq 4f \n\t" // if count == 0, branch to exit |
| 169 | "3: \n\t" // singles loop |
| 170 | "ldrh r4, [%[xx]], #2 \n\t" // load pixel ptr |
| 171 | "subs %[count], %[count], #1 \n\t" // decrement loop counter |
| 172 | "ldrb r5, [%[srcAddr], r4] \n\t" // load pixel from image |
| 173 | "ldr r6, [%[table], r5, lsl #2] \n\t" // load SkPMColor from colmap |
| 174 | "str r6, [%[colors]], #4 \n\t" // store pixel, update ptr |
| 175 | "bne 3b \n\t" // loop if counter != 0 |
| 176 | "4: \n\t" // exit |
| 177 | : [xx] "+r" (xx), [count] "+r" (count), [colors] "+r" (colors) |
| 178 | : [table] "r" (table), [srcAddr] "r" (srcAddr) |
| 179 | : "memory", "cc", "r4", "r5", "r6", "r7", "r8", "r9", "r10", "r11" |
| 180 | ); |
| 181 | } |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 182 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 183 | s.fBitmap->getColorTable()->unlockColors(false); |
| 184 | } |
| 185 | #endif //__ARM_ARCH__ >= 6 && !defined(SK_CPU_BENDIAN) |
| 186 | |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 187 | /////////////////////////////////////////////////////////////////////////////// |
| 188 | |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 189 | /* If we replace a sampleproc, then we null-out the associated shaderproc, |
| 190 | otherwise the shader won't even look at the matrix/sampler |
| 191 | */ |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 192 | void SkBitmapProcState::platformProcs() { |
| 193 | bool doFilter = fDoFilter; |
| 194 | bool isOpaque = 256 == fAlphaScale; |
| 195 | bool justDx = false; |
| 196 | |
| 197 | if (fInvType <= (SkMatrix::kTranslate_Mask | SkMatrix::kScale_Mask)) { |
| 198 | justDx = true; |
| 199 | } |
| 200 | |
| 201 | switch (fBitmap->config()) { |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 202 | case SkBitmap::kIndex8_Config: |
| 203 | #if __ARM_ARCH__ >= 6 && !defined(SK_CPU_BENDIAN) |
| 204 | if (justDx && !doFilter) { |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 205 | #if 0 /* crashing on android device */ |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 206 | fSampleProc16 = SI8_D16_nofilter_DX_arm; |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 207 | fShaderProc16 = NULL; |
| 208 | #endif |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 209 | if (isOpaque) { |
reed@android.com | 152f748 | 2009-08-07 19:14:34 +0000 | [diff] [blame] | 210 | // this one is only very slighty faster than the C version |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 211 | fSampleProc32 = SI8_opaque_D32_nofilter_DX_arm; |
reed@android.com | 6123e47 | 2009-08-04 01:52:27 +0000 | [diff] [blame] | 212 | fShaderProc32 = NULL; |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 213 | } |
| 214 | } |
| 215 | #endif |
reed@android.com | 152f748 | 2009-08-07 19:14:34 +0000 | [diff] [blame] | 216 | break; |
reed@android.com | a0bd7f4 | 2009-08-03 17:22:46 +0000 | [diff] [blame] | 217 | default: |
| 218 | break; |
| 219 | } |
| 220 | } |
| 221 | |