From 701c772eed71e6e668aa2a6548ed53669cbc1540 Mon Sep 17 00:00:00 2001 From: Parag Salasakar Date: Mon, 27 Jun 2016 14:36:28 +0530 Subject: [PATCH] Add MSA optimized colorspace conversion functions We add the following MSA optimized colorspace conversion functions: - ConvertBGRAToRGBA - ConvertBGRAToBGR - ConvertBGRAToRGB Change-Id: I76db1c829d593a06d4975d54dbafa385c82b84fb --- Android.mk | 1 + Makefile.vc | 1 + build.gradle | 1 + makefile.unix | 1 + src/dsp/Makefile.am | 1 + src/dsp/lossless.c | 6 ++ src/dsp/lossless_msa.c | 231 +++++++++++++++++++++++++++++++++++++++++ src/dsp/msa_macro.h | 80 ++++++++++---- 8 files changed, 300 insertions(+), 22 deletions(-) create mode 100644 src/dsp/lossless_msa.c diff --git a/Android.mk b/Android.mk index ef35795b..21042b2a 100644 --- a/Android.mk +++ b/Android.mk @@ -50,6 +50,7 @@ dsp_dec_srcs := \ src/dsp/dec_mips32.c \ src/dsp/dec_mips_dsp_r2.c \ src/dsp/dec_msa.c \ + src/dsp/lossless_msa.c \ src/dsp/dec_neon.$(NEON) \ src/dsp/dec_sse2.c \ src/dsp/dec_sse41.c \ diff --git a/Makefile.vc b/Makefile.vc index 5f145e24..c37f945a 100644 --- a/Makefile.vc +++ b/Makefile.vc @@ -195,6 +195,7 @@ DSP_DEC_OBJS = \ $(DIROBJ)\dsp\dec_mips32.obj \ $(DIROBJ)\dsp\dec_mips_dsp_r2.obj \ $(DIROBJ)\dsp\dec_msa.obj \ + $(DIROBJ)\dsp\lossless_msa.obj \ $(DIROBJ)\dsp\dec_neon.obj \ $(DIROBJ)\dsp\dec_sse2.obj \ $(DIROBJ)\dsp\dec_sse41.obj \ diff --git a/build.gradle b/build.gradle index b81c6a82..450e0895 100644 --- a/build.gradle +++ b/build.gradle @@ -128,6 +128,7 @@ model { include "filters_sse2.c" include "lossless.c" include "lossless_mips_dsp_r2.c" + include "lossless_msa.c" include "lossless_neon.$NEON" include "lossless_sse2.c" include "rescaler.c" diff --git a/makefile.unix b/makefile.unix index fe4571e2..9ce2969c 100644 --- a/makefile.unix +++ b/makefile.unix @@ -147,6 +147,7 @@ DSP_DEC_OBJS = \ src/dsp/filters_sse2.o \ src/dsp/lossless.o \ src/dsp/lossless_mips_dsp_r2.o \ + src/dsp/lossless_msa.o \ src/dsp/lossless_neon.o \ src/dsp/lossless_sse2.o \ src/dsp/rescaler.o \ diff --git a/src/dsp/Makefile.am b/src/dsp/Makefile.am index 2ef5c380..fc772243 100644 --- a/src/dsp/Makefile.am +++ b/src/dsp/Makefile.am @@ -83,6 +83,7 @@ libwebpdspdecode_neon_la_CFLAGS = $(libwebpdsp_neon_la_CFLAGS) libwebpdspdecode_msa_la_SOURCES = libwebpdspdecode_msa_la_SOURCES += dec_msa.c +libwebpdspdecode_msa_la_SOURCES += lossless_msa.c libwebpdspdecode_msa_la_SOURCES += msa_macro.h libwebpdspdecode_msa_la_CPPFLAGS = $(libwebpdsp_la_CPPFLAGS) libwebpdspdecode_msa_la_CFLAGS = $(AM_CFLAGS) diff --git a/src/dsp/lossless.c b/src/dsp/lossless.c index af913efc..633ca9fe 100644 --- a/src/dsp/lossless.c +++ b/src/dsp/lossless.c @@ -572,6 +572,7 @@ VP8LMapAlphaFunc VP8LMapColor8b; extern void VP8LDspInitSSE2(void); extern void VP8LDspInitNEON(void); extern void VP8LDspInitMIPSdspR2(void); +extern void VP8LDspInitMSA(void); static volatile VP8CPUInfo lossless_last_cpuinfo_used = (VP8CPUInfo)&lossless_last_cpuinfo_used; @@ -625,6 +626,11 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInit(void) { if (VP8GetCPUInfo(kMIPSdspR2)) { VP8LDspInitMIPSdspR2(); } +#endif +#if defined(WEBP_USE_MSA) + if (VP8GetCPUInfo(kMSA)) { + VP8LDspInitMSA(); + } #endif } lossless_last_cpuinfo_used = VP8GetCPUInfo; diff --git a/src/dsp/lossless_msa.c b/src/dsp/lossless_msa.c new file mode 100644 index 00000000..b9516fa4 --- /dev/null +++ b/src/dsp/lossless_msa.c @@ -0,0 +1,231 @@ +// Copyright 2016 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. +// ----------------------------------------------------------------------------- +// +// MSA variant of methods for lossless decoder +// +// Author: Prashant Patil (prashant.patil@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MSA) + +#include "./lossless.h" +#include "./msa_macro.h" + +//------------------------------------------------------------------------------ +// Colorspace conversion functions + +#define CONVERT16_BGRA_XXX(psrc, pdst, m0, m1, m2) do { \ + v16u8 src0, src1, src2, src3, dst0, dst1, dst2; \ + LD_UB4(psrc, 16, src0, src1, src2, src3); \ + VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \ + dst2 = VSHF_UB(src2, src3, m2); \ + ST_UB2(dst0, dst1, pdst, 16); \ + ST_UB(dst2, pdst + 32); \ +} while (0) + +#define CONVERT12_BGRA_XXX(psrc, pdst, m0, m1, m2) do { \ + uint32_t pix_w; \ + v16u8 src0, src1, src2, dst0, dst1, dst2; \ + LD_UB3(psrc, 16, src0, src1, src2); \ + VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \ + dst2 = VSHF_UB(src2, src2, m2); \ + ST_UB2(dst0, dst1, pdst, 16); \ + pix_w = __msa_copy_s_w((v4i32)dst2, 0); \ + SW(pix_w, pdst + 32); \ +} while (0) + +#define CONVERT8_BGRA_XXX(psrc, pdst, m0, m1) do { \ + uint64_t pix_d; \ + v16u8 src0, src1, src2, dst0, dst1; \ + LD_UB2(psrc, 16, src0, src1); \ + VSHF_B2_UB(src0, src1, src1, src2, m0, m1, dst0, dst1); \ + ST_UB(dst0, pdst); \ + pix_d = __msa_copy_s_d((v2i64)dst1, 0); \ + SD(pix_d, pdst + 16); \ +} while (0) + +#define CONVERT4_BGRA_XXX(psrc, pdst, m) do { \ + const v16u8 src0 = LD_UB(psrc); \ + const v16u8 dst0 = VSHF_UB(src0, src0, m); \ + uint64_t pix_d = __msa_copy_s_d((v2i64)dst0, 0); \ + uint32_t pix_w = __msa_copy_s_w((v4i32)dst0, 2); \ + SD(pix_d, pdst + 0); \ + SW(pix_w, pdst + 8); \ +} while (0) + +#define CONVERT1_BGRA_BGR(psrc, pdst) do { \ + const int32_t b = (psrc)[0]; \ + const int32_t g = (psrc)[1]; \ + const int32_t r = (psrc)[2]; \ + (pdst)[0] = b; \ + (pdst)[1] = g; \ + (pdst)[2] = r; \ +} while (0) + +#define CONVERT1_BGRA_RGB(psrc, pdst) do { \ + const int32_t b = (psrc)[0]; \ + const int32_t g = (psrc)[1]; \ + const int32_t r = (psrc)[2]; \ + (pdst)[0] = r; \ + (pdst)[1] = g; \ + (pdst)[2] = b; \ +} while (0) + +static void ConvertBGRAToRGBA(const uint32_t* src, + int num_pixels, uint8_t* dst) { + int i; + const uint8_t* ptemp_src = (const uint8_t*)src; + uint8_t* ptemp_dst = (uint8_t*)dst; + v16u8 src0, dst0; + const v16u8 mask = { 2, 1, 0, 3, 6, 5, 4, 7, 10, 9, 8, 11, 14, 13, 12, 15 }; + + while (num_pixels >= 8) { + v16u8 src1, dst1; + LD_UB2(ptemp_src, 16, src0, src1); + VSHF_B2_UB(src0, src0, src1, src1, mask, mask, dst0, dst1); + ST_UB2(dst0, dst1, ptemp_dst, 16); + ptemp_src += 32; + ptemp_dst += 32; + num_pixels -= 8; + } + if (num_pixels > 0) { + if (num_pixels >= 4) { + src0 = LD_UB(ptemp_src); + dst0 = VSHF_UB(src0, src0, mask); + ST_UB(dst0, ptemp_dst); + ptemp_src += 16; + ptemp_dst += 16; + num_pixels -= 4; + } + for (i = 0; i < num_pixels; i++) { + const uint8_t b = ptemp_src[2]; + const uint8_t g = ptemp_src[1]; + const uint8_t r = ptemp_src[0]; + const uint8_t a = ptemp_src[3]; + ptemp_dst[0] = b; + ptemp_dst[1] = g; + ptemp_dst[2] = r; + ptemp_dst[3] = a; + ptemp_src += 4; + ptemp_dst += 4; + } + } +} + +static void ConvertBGRAToBGR(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint8_t* ptemp_src = (const uint8_t*)src; + uint8_t* ptemp_dst = (uint8_t*)dst; + const v16u8 mask0 = { 0, 1, 2, 4, 5, 6, 8, 9, 10, 12, 13, 14, + 16, 17, 18, 20 }; + const v16u8 mask1 = { 5, 6, 8, 9, 10, 12, 13, 14, 16, 17, 18, 20, + 21, 22, 24, 25 }; + const v16u8 mask2 = { 10, 12, 13, 14, 16, 17, 18, 20, 21, 22, 24, 25, + 26, 28, 29, 30 }; + + while (num_pixels >= 16) { + CONVERT16_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 64; + ptemp_dst += 48; + num_pixels -= 16; + } + if (num_pixels > 0) { + if (num_pixels >= 12) { + CONVERT12_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 48; + ptemp_dst += 36; + num_pixels -= 12; + } else if (num_pixels >= 8) { + CONVERT8_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1); + ptemp_src += 32; + ptemp_dst += 24; + num_pixels -= 8; + } else if (num_pixels >= 4) { + CONVERT4_BGRA_XXX(ptemp_src, ptemp_dst, mask0); + ptemp_src += 16; + ptemp_dst += 12; + num_pixels -= 4; + } + if (num_pixels == 3) { + CONVERT1_BGRA_BGR(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_BGR(ptemp_src + 4, ptemp_dst + 3); + CONVERT1_BGRA_BGR(ptemp_src + 8, ptemp_dst + 6); + } else if (num_pixels == 2) { + CONVERT1_BGRA_BGR(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_BGR(ptemp_src + 4, ptemp_dst + 3); + } else if (num_pixels == 1) { + CONVERT1_BGRA_BGR(ptemp_src, ptemp_dst); + } + } +} + +static void ConvertBGRAToRGB(const uint32_t* src, + int num_pixels, uint8_t* dst) { + const uint8_t* ptemp_src = (const uint8_t*)src; + uint8_t* ptemp_dst = (uint8_t*)dst; + const v16u8 mask0 = { 2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, + 18, 17, 16, 22 }; + const v16u8 mask1 = { 5, 4, 10, 9, 8, 14, 13, 12, 18, 17, 16, 22, + 21, 20, 26, 25 }; + const v16u8 mask2 = { 8, 14, 13, 12, 18, 17, 16, 22, 21, 20, 26, 25, + 24, 30, 29, 28 }; + + while (num_pixels >= 16) { + CONVERT16_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 64; + ptemp_dst += 48; + num_pixels -= 16; + } + if (num_pixels) { + if (num_pixels >= 12) { + CONVERT12_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1, mask2); + ptemp_src += 48; + ptemp_dst += 36; + num_pixels -= 12; + } else if (num_pixels >= 8) { + CONVERT8_BGRA_XXX(ptemp_src, ptemp_dst, mask0, mask1); + ptemp_src += 32; + ptemp_dst += 24; + num_pixels -= 8; + } else if (num_pixels >= 4) { + CONVERT4_BGRA_XXX(ptemp_src, ptemp_dst, mask0); + ptemp_src += 16; + ptemp_dst += 12; + num_pixels -= 4; + } + if (num_pixels == 3) { + CONVERT1_BGRA_RGB(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_RGB(ptemp_src + 4, ptemp_dst + 3); + CONVERT1_BGRA_RGB(ptemp_src + 8, ptemp_dst + 6); + } else if (num_pixels == 2) { + CONVERT1_BGRA_RGB(ptemp_src + 0, ptemp_dst + 0); + CONVERT1_BGRA_RGB(ptemp_src + 4, ptemp_dst + 3); + } else if (num_pixels == 1) { + CONVERT1_BGRA_RGB(ptemp_src, ptemp_dst); + } + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void VP8LDspInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitMSA(void) { + VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; + VP8LConvertBGRAToBGR = ConvertBGRAToBGR; + VP8LConvertBGRAToRGB = ConvertBGRAToRGB; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(VP8LDspInitMSA) + +#endif // WEBP_USE_MSA diff --git a/src/dsp/msa_macro.h b/src/dsp/msa_macro.h index 82bb598e..b381d6be 100644 --- a/src/dsp/msa_macro.h +++ b/src/dsp/msa_macro.h @@ -196,6 +196,13 @@ #define LD_UB2(...) LD_B2(v16u8, __VA_ARGS__) #define LD_SB2(...) LD_B2(v16i8, __VA_ARGS__) +#define LD_B3(RTYPE, psrc, stride, out0, out1, out2) { \ + LD_B2(RTYPE, psrc, stride, out0, out1); \ + out2 = LD_B(RTYPE, psrc + 2 * stride); \ +} +#define LD_UB3(...) LD_B3(v16u8, __VA_ARGS__) +#define LD_SB3(...) LD_B3(v16i8, __VA_ARGS__) + #define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3) { \ LD_B2(RTYPE, psrc, stride, out0, out1); \ LD_B2(RTYPE, psrc + 2 * stride , stride, out2, out3); \ @@ -316,6 +323,31 @@ #define SLDI_SB(...) SLDI_B(v16i8, __VA_ARGS__) #define SLDI_SH(...) SLDI_B(v8i16, __VA_ARGS__) +/* Description : Shuffle byte vector elements as per mask vector + * Arguments : Inputs - in0, in1, in2, in3, mask0, mask1 + * Outputs - out0, out1 + * Return Type - as per RTYPE + * Details : Byte elements from 'in0' & 'in1' are copied selectively to + * 'out0' as per control vector 'mask0' + */ +#define VSHF_B(RTYPE, in0, in1, mask) \ + (RTYPE)__msa_vshf_b((v16i8)mask, (v16i8)in1, (v16i8)in0) + +#define VSHF_UB(...) VSHF_B(v16u8, __VA_ARGS__) +#define VSHF_SB(...) VSHF_B(v16i8, __VA_ARGS__) +#define VSHF_UH(...) VSHF_B(v8u16, __VA_ARGS__) +#define VSHF_SH(...) VSHF_B(v8i16, __VA_ARGS__) + +#define VSHF_B2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1) do { \ + out0 = VSHF_B(RTYPE, in0, in1, mask0); \ + out1 = VSHF_B(RTYPE, in2, in3, mask1); \ +} while (0) + +#define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__) +#define VSHF_B2_SB(...) VSHF_B2(v16i8, __VA_ARGS__) +#define VSHF_B2_UH(...) VSHF_B2(v8u16, __VA_ARGS__) +#define VSHF_B2_SH(...) VSHF_B2(v8i16, __VA_ARGS__) + /* Description : Shuffle halfword vector elements as per mask vector * Arguments : Inputs - in0, in1, in2, in3, mask0, mask1 * Outputs - out0, out1 @@ -323,10 +355,11 @@ * Details : halfword elements from 'in0' & 'in1' are copied selectively to * 'out0' as per control vector 'mask0' */ -#define VSHF_H2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1) { \ - out0 = (RTYPE)__msa_vshf_h((v8i16)mask0, (v8i16)in1, (v8i16)in0); \ - out1 = (RTYPE)__msa_vshf_h((v8i16)mask1, (v8i16)in3, (v8i16)in2); \ -} +#define VSHF_H2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1) do { \ + out0 = (RTYPE)__msa_vshf_h((v8i16)mask0, (v8i16)in1, (v8i16)in0); \ + out1 = (RTYPE)__msa_vshf_h((v8i16)mask1, (v8i16)in3, (v8i16)in2); \ +} while (0) + #define VSHF_H2_UH(...) VSHF_H2(v8u16, __VA_ARGS__) #define VSHF_H2_SH(...) VSHF_H2(v8i16, __VA_ARGS__) @@ -335,36 +368,39 @@ * Arguments : Input/output - val * Return Type - signed halfword */ -#define CLIP_SH_0_255(val) { \ +#define CLIP_SH_0_255(val) do { \ const v8i16 max_m = __msa_ldi_h(255); \ val = __msa_maxi_s_h((v8i16)val, 0); \ val = __msa_min_s_h(max_m, (v8i16)val); \ -} -#define CLIP_SH2_0_255(in0, in1) { \ - CLIP_SH_0_255(in0); \ - CLIP_SH_0_255(in1); \ -} -#define CLIP_SH4_0_255(in0, in1, in2, in3) { \ - CLIP_SH2_0_255(in0, in1); \ - CLIP_SH2_0_255(in2, in3); \ -} +} while (0) + +#define CLIP_SH2_0_255(in0, in1) do { \ + CLIP_SH_0_255(in0); \ + CLIP_SH_0_255(in1); \ +} while (0) + +#define CLIP_SH4_0_255(in0, in1, in2, in3) do { \ + CLIP_SH2_0_255(in0, in1); \ + CLIP_SH2_0_255(in2, in3); \ +} while (0) /* Description : Clips all signed word elements of input vector * between 0 & 255 * Arguments : Input/output - val * Return Type - signed word */ -#define CLIP_SW_0_255(val) { \ +#define CLIP_SW_0_255(val) do { \ const v4i32 max_m = __msa_ldi_w(255); \ val = __msa_maxi_s_w((v4i32)val, 0); \ val = __msa_min_s_w(max_m, (v4i32)val); \ -} -#define CLIP_SW4_0_255(in0, in1, in2, in3) { \ - CLIP_SW_0_255(in0); \ - CLIP_SW_0_255(in1); \ - CLIP_SW_0_255(in2); \ - CLIP_SW_0_255(in3); \ -} +} while (0) + +#define CLIP_SW4_0_255(in0, in1, in2, in3) do { \ + CLIP_SW_0_255(in0); \ + CLIP_SW_0_255(in1); \ + CLIP_SW_0_255(in2); \ + CLIP_SW_0_255(in3); \ +} while (0) /* Description : Horizontal addition of 8 unsigned halfword elements * Arguments : Input - in (unsigned halfword vector)