From 52245424b0e7a760ea87602772074d8807c9e4b9 Mon Sep 17 00:00:00 2001 From: Pascal Massimino Date: Tue, 18 Apr 2017 18:22:30 +0200 Subject: [PATCH] NEON implementation of some Sharp-YUV420 functions Change-Id: I449ef9c76b06f971f6e2ad7f9db96bf906d8fe1f new-file: dsp/yuv_neon.c --- Android.mk | 1 + Makefile.vc | 1 + build.gradle | 1 + makefile.unix | 1 + src/dsp/Makefile.am | 1 + src/dsp/yuv.c | 7 +++ src/dsp/yuv_neon.c | 136 ++++++++++++++++++++++++++++++++++++++++++++ 7 files changed, 148 insertions(+) create mode 100644 src/dsp/yuv_neon.c diff --git a/Android.mk b/Android.mk index 6abf5f6d..53a285c9 100644 --- a/Android.mk +++ b/Android.mk @@ -79,6 +79,7 @@ dsp_dec_srcs := \ src/dsp/yuv.c \ src/dsp/yuv_mips32.c \ src/dsp/yuv_mips_dsp_r2.c \ + src/dsp/yuv_neon.$(NEON) \ src/dsp/yuv_sse2.c \ dsp_enc_srcs := \ diff --git a/Makefile.vc b/Makefile.vc index 33ed6963..558e9579 100644 --- a/Makefile.vc +++ b/Makefile.vc @@ -229,6 +229,7 @@ DSP_DEC_OBJS = \ $(DIROBJ)\dsp\yuv.obj \ $(DIROBJ)\dsp\yuv_mips32.obj \ $(DIROBJ)\dsp\yuv_mips_dsp_r2.obj \ + $(DIROBJ)\dsp\yuv_neon.obj \ $(DIROBJ)\dsp\yuv_sse2.obj \ DSP_ENC_OBJS = \ diff --git a/build.gradle b/build.gradle index 81fd74de..f52bdf34 100644 --- a/build.gradle +++ b/build.gradle @@ -148,6 +148,7 @@ model { include "yuv.c" include "yuv_mips32.c" include "yuv_mips_dsp_r2.c" + include "yuv_neon.$NEON" include "yuv_sse2.c" srcDir "src/utils" include "bit_reader_utils.c" diff --git a/makefile.unix b/makefile.unix index 929f2e7a..7c8a9968 100644 --- a/makefile.unix +++ b/makefile.unix @@ -169,6 +169,7 @@ DSP_DEC_OBJS = \ src/dsp/yuv.o \ src/dsp/yuv_mips32.o \ src/dsp/yuv_mips_dsp_r2.o \ + src/dsp/yuv_neon.o \ src/dsp/yuv_sse2.o \ DSP_ENC_OBJS = \ diff --git a/src/dsp/Makefile.am b/src/dsp/Makefile.am index 14687798..4605d2fb 100644 --- a/src/dsp/Makefile.am +++ b/src/dsp/Makefile.am @@ -82,6 +82,7 @@ libwebpdspdecode_neon_la_SOURCES += lossless_neon.c libwebpdspdecode_neon_la_SOURCES += neon.h libwebpdspdecode_neon_la_SOURCES += rescaler_neon.c libwebpdspdecode_neon_la_SOURCES += upsampling_neon.c +libwebpdspdecode_neon_la_SOURCES += yuv_neon.c libwebpdspdecode_neon_la_CPPFLAGS = $(libwebpdsp_neon_la_CPPFLAGS) libwebpdspdecode_neon_la_CFLAGS = $(libwebpdsp_neon_la_CFLAGS) diff --git a/src/dsp/yuv.c b/src/dsp/yuv.c index dd7d9ded..716a7278 100644 --- a/src/dsp/yuv.c +++ b/src/dsp/yuv.c @@ -309,6 +309,7 @@ static volatile VP8CPUInfo rgba_to_yuv_last_cpuinfo_used = extern void WebPInitConvertARGBToYUVSSE2(void); extern void WebPInitSharpYUVSSE2(void); +extern void WebPInitSharpYUVNEON(void); WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) { if (rgba_to_yuv_last_cpuinfo_used == VP8GetCPUInfo) return; @@ -332,6 +333,12 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUV(void) { WebPInitSharpYUVSSE2(); } #endif // WEBP_USE_SSE2 +#if defined(WEBP_USE_NEON) + if (VP8GetCPUInfo(kNEON)) { + WebPInitSharpYUVNEON(); + } +#endif // WEBP_USE_NEON + } rgba_to_yuv_last_cpuinfo_used = VP8GetCPUInfo; } diff --git a/src/dsp/yuv_neon.c b/src/dsp/yuv_neon.c new file mode 100644 index 00000000..7a963d4f --- /dev/null +++ b/src/dsp/yuv_neon.c @@ -0,0 +1,136 @@ +// Copyright 2017 Google Inc. All Rights Reserved. +// +// Use of this source code is governed by a BSD-style license +// that can be found in the COPYING file in the root of the source +// tree. An additional intellectual property rights grant can be found +// in the file PATENTS. All contributing project authors may +// be found in the AUTHORS file in the root of the source tree. +// ----------------------------------------------------------------------------- +// +// YUV->RGB conversion functions +// +// Author: Skal (pascal.massimino@gmail.com) + +#include "./yuv.h" + +#if defined(WEBP_USE_NEON) + +#include +#include + +#include "./neon.h" + +//----------------------------------------------------------------------------- + +#define MAX_Y ((1 << 10) - 1) // 10b precision over 16b-arithmetic +static uint16_t clip_y(int v) { + return (v < 0) ? 0 : (v > MAX_Y) ? MAX_Y : (uint16_t)v; +} + +static uint64_t SharpYUVUpdateY_NEON(const uint16_t* ref, const uint16_t* src, + uint16_t* dst, int len) { + int i; + const int16x8_t zero = vdupq_n_s16(0); + const int16x8_t max = vdupq_n_s16(MAX_Y); + uint64x2_t sum = vdupq_n_u64(0); + uint64_t diff; + + for (i = 0; i + 8 <= len; i += 8) { + const int16x8_t A = vreinterpretq_s16_u16(vld1q_u16(ref + i)); + const int16x8_t B = vreinterpretq_s16_u16(vld1q_u16(src + i)); + const int16x8_t C = vreinterpretq_s16_u16(vld1q_u16(dst + i)); + const int16x8_t D = vsubq_s16(A, B); // diff_y + const int16x8_t F = vaddq_s16(C, D); // new_y + const uint16x8_t H = + vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(F, max), zero)); + const int16x8_t I = vabsq_s16(D); // abs(diff_y) + vst1q_u16(dst + i, H); + sum = vpadalq_u32(sum, vpaddlq_u16(vreinterpretq_u16_s16(I))); + } + diff = vgetq_lane_u64(sum, 0) + vgetq_lane_u64(sum, 1); + for (; i < len; ++i) { + const int diff_y = ref[i] - src[i]; + const int new_y = (int)(dst[i]) + diff_y; + dst[i] = clip_y(new_y); + diff += (uint64_t)(abs(diff_y)); + } + return diff; +} + +static void SharpYUVUpdateRGB_NEON(const int16_t* ref, const int16_t* src, + int16_t* dst, int len) { + int i; + for (i = 0; i + 8 <= len; i += 8) { + const int16x8_t A = vld1q_s16(ref + i); + const int16x8_t B = vld1q_s16(src + i); + const int16x8_t C = vld1q_s16(dst + i); + const int16x8_t D = vsubq_s16(A, B); // diff_uv + const int16x8_t E = vaddq_s16(C, D); // new_uv + vst1q_s16(dst + i, E); + } + for (; i < len; ++i) { + const int diff_uv = ref[i] - src[i]; + dst[i] += diff_uv; + } +} + +static void SharpYUVFilterRow_NEON(const int16_t* A, const int16_t* B, int len, + const uint16_t* best_y, uint16_t* out) { + int i; + const int16x8_t max = vdupq_n_s16(MAX_Y); + const int16x8_t zero = vdupq_n_s16(0); + for (i = 0; i + 8 <= len; i += 8) { + const int16x8_t a0 = vld1q_s16(A + i + 0); + const int16x8_t a1 = vld1q_s16(A + i + 1); + const int16x8_t b0 = vld1q_s16(B + i + 0); + const int16x8_t b1 = vld1q_s16(B + i + 1); + const int16x8_t a0b1 = vaddq_s16(a0, b1); + const int16x8_t a1b0 = vaddq_s16(a1, b0); + const int16x8_t a0a1b0b1 = vaddq_s16(a0b1, a1b0); // A0+A1+B0+B1 + const int16x8_t a0b1_2 = vaddq_s16(a0b1, a0b1); // 2*(A0+B1) + const int16x8_t a1b0_2 = vaddq_s16(a1b0, a1b0); // 2*(A1+B0) + const int16x8_t c0 = vshrq_n_s16(vaddq_s16(a0b1_2, a0a1b0b1), 3); + const int16x8_t c1 = vshrq_n_s16(vaddq_s16(a1b0_2, a0a1b0b1), 3); + const int16x8_t d0 = vaddq_s16(c1, a0); + const int16x8_t d1 = vaddq_s16(c0, a1); + const int16x8_t e0 = vrshrq_n_s16(d0, 1); + const int16x8_t e1 = vrshrq_n_s16(d1, 1); + const int16x8x2_t f = vzipq_s16(e0, e1); + const int16x8_t g0 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 0)); + const int16x8_t g1 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 8)); + const int16x8_t h0 = vaddq_s16(g0, f.val[0]); + const int16x8_t h1 = vaddq_s16(g1, f.val[1]); + const int16x8_t i0 = vmaxq_s16(vminq_s16(h0, max), zero); + const int16x8_t i1 = vmaxq_s16(vminq_s16(h1, max), zero); + vst1q_u16(out + 2 * i + 0, vreinterpretq_u16_s16(i0)); + vst1q_u16(out + 2 * i + 8, vreinterpretq_u16_s16(i1)); + } + for (; i < len; ++i) { + const int a0b1 = A[i + 0] + B[i + 1]; + const int a1b0 = A[i + 1] + B[i + 0]; + const int a0a1b0b1 = a0b1 + a1b0 + 8; + const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4; + const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4; + out[2 * i + 0] = clip_y(best_y[2 * i + 0] + v0); + out[2 * i + 1] = clip_y(best_y[2 * i + 1] + v1); + } +} +#undef MAX_Y + +//------------------------------------------------------------------------------ + +extern void WebPInitSharpYUVNEON(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPInitSharpYUVNEON(void) { + WebPSharpYUVUpdateY = SharpYUVUpdateY_NEON; + WebPSharpYUVUpdateRGB = SharpYUVUpdateRGB_NEON; + WebPSharpYUVFilterRow = SharpYUVFilterRow_NEON; +} + +#else // !WEBP_USE_NEON + +WEBP_DSP_INIT_STUB(WebPInitSamplersNEON) +WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVNEON) +WEBP_DSP_INIT_STUB(WebPInitSharpYUVNEON) + +#endif // WEBP_USE_NEON