msarett | 3a24f45 | 2016-01-13 14:31:59 -0800 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright 2016 Google Inc. |
| 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 | #ifndef SkSwizzler_opts_DEFINED |
| 9 | #define SkSwizzler_opts_DEFINED |
| 10 | |
| 11 | #include "SkColorPriv.h" |
| 12 | |
| 13 | namespace SK_OPTS_NS { |
| 14 | |
| 15 | // These variable names in these functions just pretend the input is BGRA. |
| 16 | // They work fine with both RGBA and BGRA. |
| 17 | |
| 18 | static void premul_xxxa_portable(uint32_t dst[], const uint32_t src[], int count) { |
| 19 | for (int i = 0; i < count; i++) { |
| 20 | uint8_t a = src[i] >> 24, |
| 21 | r = src[i] >> 16, |
| 22 | g = src[i] >> 8, |
| 23 | b = src[i] >> 0; |
| 24 | r = (r*a+127)/255; |
| 25 | g = (g*a+127)/255; |
| 26 | b = (b*a+127)/255; |
| 27 | dst[i] = (uint32_t)a << 24 |
| 28 | | (uint32_t)r << 16 |
| 29 | | (uint32_t)g << 8 |
| 30 | | (uint32_t)b << 0; |
| 31 | } |
| 32 | } |
| 33 | |
| 34 | static void premul_swaprb_xxxa_portable(uint32_t dst[], const uint32_t src[], int count) { |
| 35 | for (int i = 0; i < count; i++) { |
| 36 | uint8_t a = src[i] >> 24, |
| 37 | r = src[i] >> 16, |
| 38 | g = src[i] >> 8, |
| 39 | b = src[i] >> 0; |
| 40 | r = (r*a+127)/255; |
| 41 | g = (g*a+127)/255; |
| 42 | b = (b*a+127)/255; |
| 43 | dst[i] = (uint32_t)a << 24 |
| 44 | | (uint32_t)b << 16 |
| 45 | | (uint32_t)g << 8 |
| 46 | | (uint32_t)r << 0; |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | #if defined(SK_ARM_HAS_NEON) |
| 51 | |
| 52 | // Rounded divide by 255, (x + 127) / 255 |
| 53 | static uint8x8_t div255_round(uint16x8_t x) { |
| 54 | // result = (x + 127) / 255 |
| 55 | // result = (x + 127) / 256 + error1 |
| 56 | // |
| 57 | // error1 = (x + 127) / (255 * 256) |
| 58 | // error1 = (x + 127) / (256 * 256) + error2 |
| 59 | // |
| 60 | // error2 = (x + 127) / (255 * 256 * 256) |
| 61 | // |
| 62 | // The maximum value of error2 is too small to matter. Thus: |
| 63 | // result = (x + 127) / 256 + (x + 127) / (256 * 256) |
| 64 | // result = ((x + 127) / 256 + x + 127) / 256 |
| 65 | // result = ((x + 127) >> 8 + x + 127) >> 8 |
| 66 | // |
| 67 | // Use >>> to represent "rounded right shift" which, conveniently, |
| 68 | // NEON supports in one instruction. |
| 69 | // result = ((x >>> 8) + x) >>> 8 |
| 70 | // |
| 71 | // Note that the second right shift is actually performed as an |
| 72 | // "add, round, and narrow back to 8-bits" instruction. |
| 73 | return vraddhn_u16(x, vrshrq_n_u16(x, 8)); |
| 74 | } |
| 75 | |
| 76 | // Scale a byte by another, (x * y + 127) / 255 |
| 77 | static uint8x8_t scale(uint8x8_t x, uint8x8_t y) { |
| 78 | return div255_round(vmull_u8(x, y)); |
| 79 | } |
| 80 | |
| 81 | template <bool kSwapRB> |
| 82 | static void premul_xxxa_should_swaprb(uint32_t dst[], const uint32_t src[], int count) { |
| 83 | while (count >= 8) { |
| 84 | // Load 8 pixels. |
| 85 | uint8x8x4_t bgra = vld4_u8((const uint8_t*) src); |
| 86 | |
| 87 | uint8x8_t a = bgra.val[3], |
| 88 | r = bgra.val[2], |
| 89 | g = bgra.val[1], |
| 90 | b = bgra.val[0]; |
| 91 | |
| 92 | // Premultiply. |
| 93 | r = scale(r, a); |
| 94 | g = scale(g, a); |
| 95 | b = scale(b, a); |
| 96 | |
| 97 | // Store 8 premultiplied pixels. |
| 98 | if (kSwapRB) { |
| 99 | bgra.val[2] = b; |
| 100 | bgra.val[1] = g; |
| 101 | bgra.val[0] = r; |
| 102 | } else { |
| 103 | bgra.val[2] = r; |
| 104 | bgra.val[1] = g; |
| 105 | bgra.val[0] = b; |
| 106 | } |
| 107 | vst4_u8((uint8_t*) dst, bgra); |
| 108 | src += 8; |
| 109 | dst += 8; |
| 110 | count -= 8; |
| 111 | } |
| 112 | |
| 113 | // Call portable code to finish up the tail of [0,8) pixels. |
| 114 | auto proc = kSwapRB ? premul_swaprb_xxxa_portable : premul_xxxa_portable; |
| 115 | proc(dst, src, count); |
| 116 | } |
| 117 | |
| 118 | static void premul_xxxa(uint32_t dst[], const uint32_t src[], int count) { |
| 119 | premul_xxxa_should_swaprb<false>(dst, src, count); |
| 120 | } |
| 121 | |
| 122 | static void premul_swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) { |
| 123 | premul_xxxa_should_swaprb<true>(dst, src, count); |
| 124 | } |
| 125 | |
| 126 | #else |
| 127 | |
| 128 | static void premul_xxxa(uint32_t dst[], const uint32_t src[], int count) { |
| 129 | premul_xxxa_portable(dst, src, count); |
| 130 | } |
| 131 | |
| 132 | static void premul_swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) { |
| 133 | premul_swaprb_xxxa_portable(dst, src, count); |
| 134 | } |
| 135 | |
| 136 | #endif |
| 137 | |
| 138 | static void swaprb_xxxa(uint32_t dst[], const uint32_t src[], int count) { |
| 139 | for (int i = 0; i < count; i++) { |
| 140 | uint8_t a = src[i] >> 24, |
| 141 | r = src[i] >> 16, |
| 142 | g = src[i] >> 8, |
| 143 | b = src[i] >> 0; |
| 144 | dst[i] = (uint32_t)a << 24 |
| 145 | | (uint32_t)b << 16 |
| 146 | | (uint32_t)g << 8 |
| 147 | | (uint32_t)r << 0; |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | } |
| 152 | |
| 153 | #endif // SkSwizzler_opts_DEFINED |