| 1 | // Copyright 2015 Google Inc. All Rights Reserved. |
| 2 | // |
| 3 | // Use of this source code is governed by a BSD-style license |
| 4 | // that can be found in the COPYING file in the root of the source |
| 5 | // tree. An additional intellectual property rights grant can be found |
| 6 | // in the file PATENTS. All contributing project authors may |
| 7 | // be found in the AUTHORS file in the root of the source tree. |
| 8 | // ----------------------------------------------------------------------------- |
| 9 | // |
| 10 | // NEON variant of methods for lossless encoder |
| 11 | // |
| 12 | // Author: Skal (pascal.massimino@gmail.com) |
| 13 | |
| 14 | #include "src/dsp/dsp.h" |
| 15 | |
| 16 | #if defined(WEBP_USE_NEON) |
| 17 | |
| 18 | #include <arm_neon.h> |
| 19 | |
| 20 | #include "src/dsp/lossless.h" |
| 21 | #include "src/dsp/neon.h" |
| 22 | |
| 23 | //------------------------------------------------------------------------------ |
| 24 | // Subtract-Green Transform |
| 25 | |
| 26 | // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use |
| 27 | // non-standard versions there. |
| 28 | #if defined(__APPLE__) && WEBP_AARCH64 && \ |
| 29 | defined(__apple_build_version__) && (__apple_build_version__< 6020037) |
| 30 | #define USE_VTBLQ |
| 31 | #endif |
| 32 | |
| 33 | #ifdef USE_VTBLQ |
| 34 | // 255 = byte will be zeroed |
| 35 | static const uint8_t kGreenShuffle[16] = { |
| 36 | 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255 |
| 37 | }; |
| 38 | |
| 39 | static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb, |
| 40 | const uint8x16_t shuffle) { |
| 41 | return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)), |
| 42 | vtbl1q_u8(argb, vget_high_u8(shuffle))); |
| 43 | } |
| 44 | #else // !USE_VTBLQ |
| 45 | // 255 = byte will be zeroed |
| 46 | static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 }; |
| 47 | |
| 48 | static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb, |
| 49 | const uint8x8_t shuffle) { |
| 50 | return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle), |
| 51 | vtbl1_u8(vget_high_u8(argb), shuffle)); |
| 52 | } |
| 53 | #endif // USE_VTBLQ |
| 54 | |
| 55 | static void SubtractGreenFromBlueAndRed_NEON(uint32_t* argb_data, |
| 56 | int num_pixels) { |
| 57 | const uint32_t* const end = argb_data + (num_pixels & ~3); |
| 58 | #ifdef USE_VTBLQ |
| 59 | const uint8x16_t shuffle = vld1q_u8(kGreenShuffle); |
| 60 | #else |
| 61 | const uint8x8_t shuffle = vld1_u8(kGreenShuffle); |
| 62 | #endif |
| 63 | for (; argb_data < end; argb_data += 4) { |
| 64 | const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data); |
| 65 | const uint8x16_t greens = DoGreenShuffle_NEON(argb, shuffle); |
| 66 | vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens)); |
| 67 | } |
| 68 | // fallthrough and finish off with plain-C |
| 69 | VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3); |
| 70 | } |
| 71 | |
| 72 | //------------------------------------------------------------------------------ |
| 73 | // Color Transform |
| 74 | |
| 75 | static void TransformColor_NEON(const VP8LMultipliers* const m, |
| 76 | uint32_t* argb_data, int num_pixels) { |
| 77 | // sign-extended multiplying constants, pre-shifted by 6. |
| 78 | #define CST(X) (((int16_t)(m->X << 8)) >> 6) |
| 79 | const int16_t rb[8] = { |
| 80 | CST(green_to_blue_), CST(green_to_red_), |
| 81 | CST(green_to_blue_), CST(green_to_red_), |
| 82 | CST(green_to_blue_), CST(green_to_red_), |
| 83 | CST(green_to_blue_), CST(green_to_red_) |
| 84 | }; |
| 85 | const int16x8_t mults_rb = vld1q_s16(rb); |
| 86 | const int16_t b2[8] = { |
| 87 | 0, CST(red_to_blue_), 0, CST(red_to_blue_), |
| 88 | 0, CST(red_to_blue_), 0, CST(red_to_blue_), |
| 89 | }; |
| 90 | const int16x8_t mults_b2 = vld1q_s16(b2); |
| 91 | #undef CST |
| 92 | #ifdef USE_VTBLQ |
| 93 | static const uint8_t kg0g0[16] = { |
| 94 | 255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13 |
| 95 | }; |
| 96 | const uint8x16_t shuffle = vld1q_u8(kg0g0); |
| 97 | #else |
| 98 | static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 }; |
| 99 | const uint8x8_t shuffle = vld1_u8(k0g0g); |
| 100 | #endif |
| 101 | const uint32x4_t mask_rb = vdupq_n_u32(0x00ff00ffu); // red-blue masks |
| 102 | int i; |
| 103 | for (i = 0; i + 4 <= num_pixels; i += 4) { |
| 104 | const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i)); |
| 105 | // 0 g 0 g |
| 106 | const uint8x16_t greens = DoGreenShuffle_NEON(in, shuffle); |
| 107 | // x dr x db1 |
| 108 | const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb); |
| 109 | // r 0 b 0 |
| 110 | const int16x8_t B = vshlq_n_s16(vreinterpretq_s16_u8(in), 8); |
| 111 | // x db2 0 0 |
| 112 | const int16x8_t C = vqdmulhq_s16(B, mults_b2); |
| 113 | // 0 0 x db2 |
| 114 | const uint32x4_t D = vshrq_n_u32(vreinterpretq_u32_s16(C), 16); |
| 115 | // x dr x db |
| 116 | const int8x16_t E = vaddq_s8(vreinterpretq_s8_u32(D), |
| 117 | vreinterpretq_s8_s16(A)); |
| 118 | // 0 dr 0 db |
| 119 | const uint32x4_t F = vandq_u32(vreinterpretq_u32_s8(E), mask_rb); |
| 120 | const int8x16_t out = vsubq_s8(vreinterpretq_s8_u8(in), |
| 121 | vreinterpretq_s8_u32(F)); |
| 122 | vst1q_s8((int8_t*)(argb_data + i), out); |
| 123 | } |
| 124 | // fallthrough and finish off with plain-C |
| 125 | VP8LTransformColor_C(m, argb_data + i, num_pixels - i); |
| 126 | } |
| 127 | |
| 128 | #undef USE_VTBLQ |
| 129 | |
| 130 | //------------------------------------------------------------------------------ |
| 131 | // Entry point |
| 132 | |
| 133 | extern void VP8LEncDspInitNEON(void); |
| 134 | |
| 135 | WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitNEON(void) { |
| 136 | VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_NEON; |
| 137 | VP8LTransformColor = TransformColor_NEON; |
| 138 | } |
| 139 | |
| 140 | #else // !WEBP_USE_NEON |
| 141 | |
| 142 | WEBP_DSP_INIT_STUB(VP8LEncDspInitNEON) |
| 143 | |
| 144 | #endif // WEBP_USE_NEON |
| 145 | |