From 9ac74f922eefd93f1073a3a2fc4976683ade2c12 Mon Sep 17 00:00:00 2001 From: Parag Salasakar Date: Tue, 19 Jul 2016 20:17:54 +0530 Subject: [PATCH] Add MSA optimized rescaling functions We add the following MSA optimized rescaling functions: - RescalerExportRowExpand - RescalerExportRowShrink Change-Id: Ic1c76065423b02617db94cf0c22bb564219b36e6 --- Android.mk | 1 + Makefile.vc | 1 + build.gradle | 1 + makefile.unix | 1 + src/dsp/Makefile.am | 1 + src/dsp/msa_macro.h | 134 ++++++++++++- src/dsp/rescaler.c | 6 + src/dsp/rescaler_msa.c | 444 +++++++++++++++++++++++++++++++++++++++++ 8 files changed, 585 insertions(+), 4 deletions(-) create mode 100644 src/dsp/rescaler_msa.c diff --git a/Android.mk b/Android.mk index 9b420745..0237abc0 100644 --- a/Android.mk +++ b/Android.mk @@ -65,6 +65,7 @@ dsp_dec_srcs := \ src/dsp/rescaler.c \ src/dsp/rescaler_mips32.c \ src/dsp/rescaler_mips_dsp_r2.c \ + src/dsp/rescaler_msa.c \ src/dsp/rescaler_neon.$(NEON) \ src/dsp/rescaler_sse2.c \ src/dsp/upsampling.c \ diff --git a/Makefile.vc b/Makefile.vc index 67d6201f..ab4200a6 100644 --- a/Makefile.vc +++ b/Makefile.vc @@ -210,6 +210,7 @@ DSP_DEC_OBJS = \ $(DIROBJ)\dsp\rescaler.obj \ $(DIROBJ)\dsp\rescaler_mips32.obj \ $(DIROBJ)\dsp\rescaler_mips_dsp_r2.obj \ + $(DIROBJ)\dsp\rescaler_msa.obj \ $(DIROBJ)\dsp\rescaler_neon.obj \ $(DIROBJ)\dsp\rescaler_sse2.obj \ $(DIROBJ)\dsp\upsampling.obj \ diff --git a/build.gradle b/build.gradle index 31cd140b..e81561aa 100644 --- a/build.gradle +++ b/build.gradle @@ -135,6 +135,7 @@ model { include "rescaler.c" include "rescaler_mips32.c" include "rescaler_mips_dsp_r2.c" + include "rescaler_msa.c" include "rescaler_neon.$NEON" include "rescaler_sse2.c" include "upsampling.c" diff --git a/makefile.unix b/makefile.unix index 8aec72eb..4975f132 100644 --- a/makefile.unix +++ b/makefile.unix @@ -154,6 +154,7 @@ DSP_DEC_OBJS = \ src/dsp/rescaler.o \ src/dsp/rescaler_mips32.o \ src/dsp/rescaler_mips_dsp_r2.o \ + src/dsp/rescaler_msa.o \ src/dsp/rescaler_neon.o \ src/dsp/rescaler_sse2.o \ src/dsp/upsampling.o \ diff --git a/src/dsp/Makefile.am b/src/dsp/Makefile.am index efa19358..212ad38b 100644 --- a/src/dsp/Makefile.am +++ b/src/dsp/Makefile.am @@ -85,6 +85,7 @@ libwebpdspdecode_msa_la_SOURCES = libwebpdspdecode_msa_la_SOURCES += dec_msa.c libwebpdspdecode_msa_la_SOURCES += filters_msa.c libwebpdspdecode_msa_la_SOURCES += lossless_msa.c +libwebpdspdecode_msa_la_SOURCES += rescaler_msa.c libwebpdspdecode_msa_la_SOURCES += msa_macro.h libwebpdspdecode_msa_la_CPPFLAGS = $(libwebpdsp_msa_la_CPPFLAGS) libwebpdspdecode_msa_la_CFLAGS = $(libwebpdsp_msa_la_CFLAGS) diff --git a/src/dsp/msa_macro.h b/src/dsp/msa_macro.h index 8c3b5bce..9b7a2bf7 100644 --- a/src/dsp/msa_macro.h +++ b/src/dsp/msa_macro.h @@ -150,7 +150,7 @@ out3 = LW(ptmp); \ } while (0) -/* Description : Store 4 words with stride +/* Description : Store words with stride * Arguments : Inputs - in0, in1, in2, in3, pdst, stride * Details : Store word from 'in0' to (pdst) * Store word from 'in1' to (pdst + stride) @@ -168,6 +168,22 @@ SW(in3, ptmp); \ } while (0) +#define SW3(in0, in1, in2, pdst, stride) do { \ + uint8_t* ptmp = (uint8_t*)pdst; \ + SW(in0, ptmp); \ + ptmp += stride; \ + SW(in1, ptmp); \ + ptmp += stride; \ + SW(in2, ptmp); \ +} while (0) + +#define SW2(in0, in1, pdst, stride) do { \ + uint8_t* ptmp = (uint8_t*)pdst; \ + SW(in0, ptmp); \ + ptmp += stride; \ + SW(in1, ptmp); \ +} while (0) + /* Description : Store 4 double words with stride * Arguments : Inputs - in0, in1, in2, in3, pdst, stride * Details : Store double word from 'in0' to (pdst) @@ -237,9 +253,11 @@ /* Description : Load vectors with 4 word elements with stride * Arguments : Inputs - psrc, stride - * Outputs - out0, out1 - * Details : Load 4 word elements in 'out0' from (psrc) - * Load 4 word elements in 'out1' from (psrc + stride) + * Outputs - out0, out1, out2, out3 + * Details : Load 4 word elements in 'out0' from (psrc + 0 * stride) + * Load 4 word elements in 'out1' from (psrc + 1 * stride) + * Load 4 word elements in 'out2' from (psrc + 2 * stride) + * Load 4 word elements in 'out3' from (psrc + 3 * stride) */ #define LD_W2(RTYPE, psrc, stride, out0, out1) do { \ out0 = LD_W(RTYPE, psrc); \ @@ -248,6 +266,13 @@ #define LD_UW2(...) LD_W2(v4u32, __VA_ARGS__) #define LD_SW2(...) LD_W2(v4i32, __VA_ARGS__) +#define LD_W3(RTYPE, psrc, stride, out0, out1, out2) do { \ + LD_W2(RTYPE, psrc, stride, out0, out1); \ + out2 = LD_W(RTYPE, psrc + 2 * stride); \ +} while (0) +#define LD_UW3(...) LD_W3(v4u32, __VA_ARGS__) +#define LD_SW3(...) LD_W3(v4i32, __VA_ARGS__) + #define LD_W4(RTYPE, psrc, stride, out0, out1, out2, out3) do { \ LD_W2(RTYPE, psrc, stride, out0, out1); \ LD_W2(RTYPE, psrc + 2 * stride, stride, out2, out3); \ @@ -281,6 +306,34 @@ } while (0) #define ST_UB8(...) ST_B8(v16u8, __VA_ARGS__) +/* Description : Store vectors of 4 word elements with stride + * Arguments : Inputs - in0, in1, in2, in3, pdst, stride + * Details : Store 4 word elements from 'in0' to (pdst + 0 * stride) + * Store 4 word elements from 'in1' to (pdst + 1 * stride) + * Store 4 word elements from 'in2' to (pdst + 2 * stride) + * Store 4 word elements from 'in3' to (pdst + 3 * stride) + */ +#define ST_W2(RTYPE, in0, in1, pdst, stride) do { \ + ST_W(RTYPE, in0, pdst); \ + ST_W(RTYPE, in1, pdst + stride); \ +} while (0) +#define ST_UW2(...) ST_W2(v4u32, __VA_ARGS__) +#define ST_SW2(...) ST_W2(v4i32, __VA_ARGS__) + +#define ST_W3(RTYPE, in0, in1, in2, pdst, stride) do { \ + ST_W2(RTYPE, in0, in1, pdst, stride); \ + ST_W(RTYPE, in2, pdst + 2 * stride); \ +} while (0) +#define ST_UW3(...) ST_W3(v4u32, __VA_ARGS__) +#define ST_SW3(...) ST_W3(v4i32, __VA_ARGS__) + +#define ST_W4(RTYPE, in0, in1, in2, in3, pdst, stride) do { \ + ST_W2(RTYPE, in0, in1, pdst, stride); \ + ST_W2(RTYPE, in2, in3, pdst + 2 * stride, stride); \ +} while (0) +#define ST_UW4(...) ST_W4(v4u32, __VA_ARGS__) +#define ST_SW4(...) ST_W4(v4i32, __VA_ARGS__) + /* Description : Store vectors of 8 halfword elements with stride * Arguments : Inputs - in0, in1, pdst, stride * Details : Store 8 halfword elements from 'in0' to (pdst) @@ -429,6 +482,22 @@ } while (0) #define DOTP_SH2_SW(...) DOTP_SH2(v4i32, __VA_ARGS__) +/* Description : Dot product of unsigned word vector elements + * Arguments : Inputs - mult0, mult1, cnst0, cnst1 + * Outputs - out0, out1 + * Return Type - as per RTYPE + * Details : Unsigned word elements from 'mult0' are multiplied with + * unsigned word elements from 'cnst0' producing a result + * twice the size of input i.e. unsigned double word. + * The multiplication result of adjacent odd-even elements + * are added together and written to the 'out0' vector + */ +#define DOTP_UW2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) do { \ + out0 = (RTYPE)__msa_dotp_u_d((v4u32)mult0, (v4u32)cnst0); \ + out1 = (RTYPE)__msa_dotp_u_d((v4u32)mult1, (v4u32)cnst1); \ +} while (0) +#define DOTP_UW2_UD(...) DOTP_UW2(v2u64, __VA_ARGS__) + /* Description : Dot product & addition of halfword vector elements * Arguments : Inputs - mult0, mult1, cnst0, cnst1 * Outputs - out0, out1 @@ -868,6 +937,7 @@ static WEBP_INLINE uint32_t func_hadd_uh_u32(v8u16 in) { #define ILVRL_W2_UB(...) ILVRL_W2(v16u8, __VA_ARGS__) #define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__) #define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__) +#define ILVRL_W2_UW(...) ILVRL_W2(v4u32, __VA_ARGS__) /* Description : Pack even byte elements of vector pairs * Arguments : Inputs - in0, in1, in2, in3 @@ -913,6 +983,23 @@ static WEBP_INLINE uint32_t func_hadd_uh_u32(v8u16 in) { #define PCKEV_H2_SW(...) PCKEV_H2(v4i32, __VA_ARGS__) #define PCKEV_H2_UW(...) PCKEV_H2(v4u32, __VA_ARGS__) +/* Description : Pack even word elements of vector pairs + * Arguments : Inputs - in0, in1, in2, in3 + * Outputs - out0, out1 + * Return Type - as per RTYPE + * Details : Even word elements of 'in0' are copied to the left half of + * 'out0' & even word elements of 'in1' are copied to the + * right half of 'out0'. + */ +#define PCKEV_W2(RTYPE, in0, in1, in2, in3, out0, out1) do { \ + out0 = (RTYPE)__msa_pckev_w((v4i32)in0, (v4i32)in1); \ + out1 = (RTYPE)__msa_pckev_w((v4i32)in2, (v4i32)in3); \ +} while (0) +#define PCKEV_W2_UH(...) PCKEV_W2(v8u16, __VA_ARGS__) +#define PCKEV_W2_SH(...) PCKEV_W2(v8i16, __VA_ARGS__) +#define PCKEV_W2_SW(...) PCKEV_W2(v4i32, __VA_ARGS__) +#define PCKEV_W2_UW(...) PCKEV_W2(v4u32, __VA_ARGS__) + /* Description : Arithmetic immediate shift right all elements of word vector * Arguments : Inputs - in0, in1, shift * Outputs - in place operation @@ -969,6 +1056,31 @@ static WEBP_INLINE uint32_t func_hadd_uh_u32(v8u16 in) { #define SRARI_W4_UW(...) SRARI_W4(v4u32, __VA_ARGS__) #define SRARI_W4_SW(...) SRARI_W4(v4i32, __VA_ARGS__) +/* Description : Shift right arithmetic rounded double words + * Arguments : Inputs - in0, in1, shift + * Outputs - in place operation + * Return Type - as per RTYPE + * Details : Each element of vector 'in0' is shifted right arithmetically by + * the number of bits in the corresponding element in the vector + * 'shift'. The last discarded bit is added to shifted value for + * rounding and the result is written in-place. + * 'shift' is a vector. + */ +#define SRAR_D2(RTYPE, in0, in1, shift) do { \ + in0 = (RTYPE)__msa_srar_d((v2i64)in0, (v2i64)shift); \ + in1 = (RTYPE)__msa_srar_d((v2i64)in1, (v2i64)shift); \ +} while (0) +#define SRAR_D2_SW(...) SRAR_D2(v4i32, __VA_ARGS__) +#define SRAR_D2_SD(...) SRAR_D2(v2i64, __VA_ARGS__) +#define SRAR_D2_UD(...) SRAR_D2(v2u64, __VA_ARGS__) + +#define SRAR_D4(RTYPE, in0, in1, in2, in3, shift) do { \ + SRAR_D2(RTYPE, in0, in1, shift); \ + SRAR_D2(RTYPE, in2, in3, shift); \ +} while (0) +#define SRAR_D4_SD(...) SRAR_D4(v2i64, __VA_ARGS__) +#define SRAR_D4_UD(...) SRAR_D4(v2u64, __VA_ARGS__) + /* Description : Addition of 2 pairs of half-word vectors * Arguments : Inputs - in0, in1, in2, in3 * Outputs - out0, out1 @@ -1034,6 +1146,20 @@ static WEBP_INLINE uint32_t func_hadd_uh_u32(v8u16 in) { out1 = in2 - in3; \ } while (0) +#define SUB3(in0, in1, in2, in3, in4, in5, out0, out1, out2) do { \ + out0 = in0 - in1; \ + out1 = in2 - in3; \ + out2 = in4 - in5; \ +} while (0) + +#define SUB4(in0, in1, in2, in3, in4, in5, in6, in7, \ + out0, out1, out2, out3) do { \ + out0 = in0 - in1; \ + out1 = in2 - in3; \ + out2 = in4 - in5; \ + out3 = in6 - in7; \ +} while (0) + /* Description : Addition - Subtraction of input vectors * Arguments : Inputs - in0, in1 * Outputs - out0, out1 diff --git a/src/dsp/rescaler.c b/src/dsp/rescaler.c index bc743d5d..b940e6dc 100644 --- a/src/dsp/rescaler.c +++ b/src/dsp/rescaler.c @@ -199,6 +199,7 @@ WebPRescalerExportRowFunc WebPRescalerExportRowShrink; extern void WebPRescalerDspInitSSE2(void); extern void WebPRescalerDspInitMIPS32(void); extern void WebPRescalerDspInitMIPSdspR2(void); +extern void WebPRescalerDspInitMSA(void); extern void WebPRescalerDspInitNEON(void); static volatile VP8CPUInfo rescaler_last_cpuinfo_used = @@ -232,6 +233,11 @@ WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInit(void) { if (VP8GetCPUInfo(kMIPSdspR2)) { WebPRescalerDspInitMIPSdspR2(); } +#endif +#if defined(WEBP_USE_MSA) + if (VP8GetCPUInfo(kMSA)) { + WebPRescalerDspInitMSA(); + } #endif } rescaler_last_cpuinfo_used = VP8GetCPUInfo; diff --git a/src/dsp/rescaler_msa.c b/src/dsp/rescaler_msa.c new file mode 100644 index 00000000..749b6c2c --- /dev/null +++ b/src/dsp/rescaler_msa.c @@ -0,0 +1,444 @@ +// 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 version of rescaling functions +// +// Author: Prashant Patil (prashant.patil@imgtec.com) + +#include "./dsp.h" + +#if defined(WEBP_USE_MSA) + +#include + +#include "../utils/rescaler.h" +#include "./msa_macro.h" + +#define ROUNDER (WEBP_RESCALER_ONE >> 1) +#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX) + +#define CALC_MULT_FIX_16(in0, in1, in2, in3, scale, shift, dst) do { \ + v4u32 tmp0, tmp1, tmp2, tmp3; \ + v16u8 t0, t1, t2, t3, t4, t5; \ + v2u64 out0, out1, out2, out3; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in1, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, scale, scale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_B2_UB(out1, out0, out3, out2, t0, t1); \ + ILVRL_W2_UW(zero, in2, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in3, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, scale, scale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_B2_UB(out1, out0, out3, out2, t2, t3); \ + PCKEV_B2_UB(t1, t0, t3, t2, t4, t5); \ + dst = (v16u8)__msa_pckev_b((v16i8)t5, (v16i8)t4); \ +} while (0) + +#define CALC_MULT_FIX_4(in0, scale, shift, dst) do { \ + v4u32 tmp0, tmp1; \ + v16i8 t0, t1; \ + v2u64 out0, out1; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + t0 = __msa_pckev_b((v16i8)out1, (v16i8)out0); \ + t1 = __msa_pckev_b(t0, t0); \ + t0 = __msa_pckev_b(t1, t1); \ + dst = __msa_copy_s_w((v4i32)t0, 0); \ +} while (0) + +#define CALC_MULT_FIX1_16(in0, in1, in2, in3, fyscale, shift, \ + dst0, dst1, dst2, dst3) do { \ + v4u32 tmp0, tmp1, tmp2, tmp3; \ + v2u64 out0, out1, out2, out3; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in1, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, fyscale, fyscale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, fyscale, fyscale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_W2_UW(out1, out0, out3, out2, dst0, dst1); \ + ILVRL_W2_UW(zero, in2, tmp0, tmp1); \ + ILVRL_W2_UW(zero, in3, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, fyscale, fyscale, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, fyscale, fyscale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_W2_UW(out1, out0, out3, out2, dst2, dst3); \ +} while (0) + +#define CALC_MULT_FIX1_4(in0, scale, shift, dst) do { \ + v4u32 tmp0, tmp1; \ + v2u64 out0, out1; \ + ILVRL_W2_UW(zero, in0, tmp0, tmp1); \ + DOTP_UW2_UD(tmp0, tmp1, scale, scale, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + dst = (v4u32)__msa_pckev_w((v4i32)out1, (v4i32)out0); \ +} while (0) + +#define CALC_MULT_FIX2_16(in0, in1, in2, in3, mult, scale, shift, \ + dst0, dst1) do { \ + v4u32 tmp0, tmp1, tmp2, tmp3; \ + v2u64 out0, out1, out2, out3; \ + ILVRL_W2_UW(in0, in2, tmp0, tmp1); \ + ILVRL_W2_UW(in1, in3, tmp2, tmp3); \ + DOTP_UW2_UD(tmp0, tmp1, mult, mult, out0, out1); \ + DOTP_UW2_UD(tmp2, tmp3, mult, mult, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + DOTP_UW2_UD(out0, out1, scale, scale, out0, out1); \ + DOTP_UW2_UD(out2, out3, scale, scale, out2, out3); \ + SRAR_D4_UD(out0, out1, out2, out3, shift); \ + PCKEV_B2_UB(out1, out0, out3, out2, dst0, dst1); \ +} while (0) + +#define CALC_MULT_FIX2_4(in0, in1, mult, scale, shift, dst) do { \ + v4u32 tmp0, tmp1; \ + v2u64 out0, out1; \ + v16i8 t0, t1; \ + ILVRL_W2_UW(in0, in1, tmp0, tmp1); \ + DOTP_UW2_UD(tmp0, tmp1, mult, mult, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + DOTP_UW2_UD(out0, out1, scale, scale, out0, out1); \ + SRAR_D2_UD(out0, out1, shift); \ + t0 = __msa_pckev_b((v16i8)out1, (v16i8)out0); \ + t1 = __msa_pckev_b(t0, t0); \ + t0 = __msa_pckev_b(t1, t1); \ + dst = __msa_copy_s_w((v4i32)t0, 0); \ +} while (0) + +static WEBP_INLINE void ExportRowExpand_0(const uint32_t* frow, uint8_t* dst, + int length, + WebPRescaler* const wrk) { + const v4u32 scale = (v4u32)__msa_fill_w(wrk->fy_scale); + const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + const v4i32 zero = { 0 }; + + while (length >= 16) { + v4u32 src0, src1, src2, src3; + v16u8 out; + LD_UW4(frow, 4, src0, src1, src2, src3); + CALC_MULT_FIX_16(src0, src1, src2, src3, scale, shift, out); + ST_UB(out, dst); + length -= 16; + frow += 16; + dst += 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 src0, src1, src2; + LD_UW3(frow, 4, src0, src1, src2); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + CALC_MULT_FIX_4(src2, scale, shift, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + length -= 12; + frow += 12; + dst += 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 src0, src1; + LD_UW2(frow, 4, src0, src1); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + SW2(val0_m, val1_m, dst, 4); + length -= 8; + frow += 8; + dst += 8; + } else if (length >= 4) { + uint32_t val0_m; + const v4u32 src0 = LD_UW(frow); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + SW(val0_m, dst); + length -= 4; + frow += 4; + dst += 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const uint32_t J = frow[x_out]; + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } +} + +static WEBP_INLINE void ExportRowExpand_1(const uint32_t* frow, uint32_t* irow, + uint8_t* dst, int length, + WebPRescaler* const wrk) { + const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub); + const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B); + const v4i32 B1 = __msa_fill_w(B); + const v4i32 A1 = __msa_fill_w(A); + const v4i32 AB = __msa_ilvr_w(A1, B1); + const v4u32 scale = (v4u32)__msa_fill_w(wrk->fy_scale); + const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + + while (length >= 16) { + v4u32 frow0, frow1, frow2, frow3, irow0, irow1, irow2, irow3; + v16u8 t0, t1, t2, t3, t4, t5; + LD_UW4(frow, 4, frow0, frow1, frow2, frow3); + LD_UW4(irow, 4, irow0, irow1, irow2, irow3); + CALC_MULT_FIX2_16(frow0, frow1, irow0, irow1, AB, scale, shift, t0, t1); + CALC_MULT_FIX2_16(frow2, frow3, irow2, irow3, AB, scale, shift, t2, t3); + PCKEV_B2_UB(t1, t0, t3, t2, t4, t5); + t0 = (v16u8)__msa_pckev_b((v16i8)t5, (v16i8)t4); + ST_UB(t0, dst); + frow += 16; + irow += 16; + dst += 16; + length -= 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 frow0, frow1, frow2, irow0, irow1, irow2; + LD_UW3(frow, 4, frow0, frow1, frow2); + LD_UW3(irow, 4, irow0, irow1, irow2); + CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m); + CALC_MULT_FIX2_4(frow1, irow1, AB, scale, shift, val1_m); + CALC_MULT_FIX2_4(frow2, irow2, AB, scale, shift, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + frow += 12; + irow += 12; + dst += 12; + length -= 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 frow0, frow1, irow0, irow1; + LD_UW2(frow, 4, frow0, frow1); + LD_UW2(irow, 4, irow0, irow1); + CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m); + CALC_MULT_FIX2_4(frow1, irow1, AB, scale, shift, val1_m); + SW2(val0_m, val1_m, dst, 4); + frow += 4; + irow += 4; + dst += 4; + length -= 4; + } else if (length >= 4) { + uint32_t val0_m; + const v4u32 frow0 = LD_UW(frow + 0); + const v4u32 irow0 = LD_UW(irow + 0); + CALC_MULT_FIX2_4(frow0, irow0, AB, scale, shift, val0_m); + SW(val0_m, dst); + frow += 4; + irow += 4; + dst += 4; + length -= 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const uint64_t I = (uint64_t)A * frow[x_out] + + (uint64_t)B * irow[x_out]; + const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX); + const int v = (int)MULT_FIX(J, wrk->fy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + } + } +} + +static void RescalerExportRowExpand(WebPRescaler* const wrk) { + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(wrk->y_expand); + assert(wrk->y_sub != 0); + if (wrk->y_accum == 0) { + ExportRowExpand_0(frow, dst, x_out_max, wrk); + } else { + ExportRowExpand_1(frow, irow, dst, x_out_max, wrk); + } +} + +static WEBP_INLINE void ExportRowShrink_0(const uint32_t* frow, uint32_t* irow, + uint8_t* dst, int length, + const uint32_t yscale, + WebPRescaler* const wrk) { + const v4u32 y_scale = (v4u32)__msa_fill_w(yscale); + const v4u32 fxyscale = (v4u32)__msa_fill_w(wrk->fxy_scale); + const v4u32 shiftval = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + const v4i32 zero = { 0 }; + + while (length >= 16) { + v4u32 src0, src1, src2, src3, frac0, frac1, frac2, frac3; + v16u8 out; + LD_UW4(frow, 4, src0, src1, src2, src3); + CALC_MULT_FIX1_16(src0, src1, src2, src3, y_scale, shiftval, + frac0, frac1, frac2, frac3); + LD_UW4(irow, 4, src0, src1, src2, src3); + SUB4(src0, frac0, src1, frac1, src2, frac2, src3, frac3, + src0, src1, src2, src3); + CALC_MULT_FIX_16(src0, src1, src2, src3, fxyscale, shiftval, out); + ST_UB(out, dst); + ST_UW4(frac0, frac1, frac2, frac3, irow, 4); + frow += 16; + irow += 16; + dst += 16; + length -= 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 src0, src1, src2, frac0, frac1, frac2; + LD_UW3(frow, 4, src0, src1, src2); + CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0); + CALC_MULT_FIX1_4(src1, y_scale, shiftval, frac1); + CALC_MULT_FIX1_4(src2, y_scale, shiftval, frac2); + LD_UW3(irow, 4, src0, src1, src2); + SUB3(src0, frac0, src1, frac1, src2, frac2, src0, src1, src2); + CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m); + CALC_MULT_FIX_4(src1, fxyscale, shiftval, val1_m); + CALC_MULT_FIX_4(src2, fxyscale, shiftval, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + ST_UW3(frac0, frac1, frac2, irow, 4); + frow += 12; + irow += 12; + dst += 12; + length -= 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 src0, src1, frac0, frac1; + LD_UW2(frow, 4, src0, src1); + CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0); + CALC_MULT_FIX1_4(src1, y_scale, shiftval, frac1); + LD_UW2(irow, 4, src0, src1); + SUB2(src0, frac0, src1, frac1, src0, src1); + CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m); + CALC_MULT_FIX_4(src1, fxyscale, shiftval, val1_m); + SW2(val0_m, val1_m, dst, 4); + ST_UW2(frac0, frac1, irow, 4); + frow += 8; + irow += 8; + dst += 8; + length -= 8; + } else if (length >= 4) { + uint32_t val0_m; + v4u32 frac0; + v4u32 src0 = LD_UW(frow); + CALC_MULT_FIX1_4(src0, y_scale, shiftval, frac0); + src0 = LD_UW(irow); + src0 = src0 - frac0; + CALC_MULT_FIX_4(src0, fxyscale, shiftval, val0_m); + SW(val0_m, dst); + ST_UW(frac0, irow); + frow += 4; + irow += 4; + dst += 4; + length -= 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const uint32_t frac = (uint32_t)MULT_FIX(frow[x_out], yscale); + const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = frac; + } + } +} + +static WEBP_INLINE void ExportRowShrink_1(uint32_t* irow, uint8_t* dst, + int length, + WebPRescaler* const wrk) { + const v4u32 scale = (v4u32)__msa_fill_w(wrk->fxy_scale); + const v4u32 shift = (v4u32)__msa_fill_w(WEBP_RESCALER_RFIX); + const v4i32 zero = { 0 }; + + while (length >= 16) { + v4u32 src0, src1, src2, src3; + v16u8 dst0; + LD_UW4(irow, 4, src0, src1, src2, src3); + CALC_MULT_FIX_16(src0, src1, src2, src3, scale, shift, dst0); + ST_UB(dst0, dst); + ST_SW4(zero, zero, zero, zero, irow, 4); + length -= 16; + irow += 16; + dst += 16; + } + if (length > 0) { + int x_out; + if (length >= 12) { + uint32_t val0_m, val1_m, val2_m; + v4u32 src0, src1, src2; + LD_UW3(irow, 4, src0, src1, src2); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + CALC_MULT_FIX_4(src2, scale, shift, val2_m); + SW3(val0_m, val1_m, val2_m, dst, 4); + ST_SW3(zero, zero, zero, irow, 4); + length -= 12; + irow += 12; + dst += 12; + } else if (length >= 8) { + uint32_t val0_m, val1_m; + v4u32 src0, src1; + LD_UW2(irow, 4, src0, src1); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + CALC_MULT_FIX_4(src1, scale, shift, val1_m); + SW2(val0_m, val1_m, dst, 4); + ST_SW2(zero, zero, irow, 4); + length -= 8; + irow += 8; + dst += 8; + } else if (length >= 4) { + uint32_t val0_m; + const v4u32 src0 = LD_UW(irow + 0); + CALC_MULT_FIX_4(src0, scale, shift, val0_m); + SW(val0_m, dst); + ST_SW(zero, irow); + length -= 4; + irow += 4; + dst += 4; + } + for (x_out = 0; x_out < length; ++x_out) { + const int v = (int)MULT_FIX(irow[x_out], wrk->fxy_scale); + assert(v >= 0 && v <= 255); + dst[x_out] = v; + irow[x_out] = 0; + } + } +} + +static void RescalerExportRowShrink(WebPRescaler* const wrk) { + uint8_t* dst = wrk->dst; + rescaler_t* irow = wrk->irow; + const int x_out_max = wrk->dst_width * wrk->num_channels; + const rescaler_t* frow = wrk->frow; + const uint32_t yscale = wrk->fy_scale * (-wrk->y_accum); + assert(!WebPRescalerOutputDone(wrk)); + assert(wrk->y_accum <= 0); + assert(!wrk->y_expand); + if (yscale) { + ExportRowShrink_0(frow, irow, dst, x_out_max, yscale, wrk); + } else { + ExportRowShrink_1(irow, dst, x_out_max, wrk); + } +} + +//------------------------------------------------------------------------------ +// Entry point + +extern void WebPRescalerDspInitMSA(void); + +WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitMSA(void) { + WebPRescalerExportRowExpand = RescalerExportRowExpand; + WebPRescalerExportRowShrink = RescalerExportRowShrink; +} + +#else // !WEBP_USE_MSA + +WEBP_DSP_INIT_STUB(WebPRescalerDspInitMSA) + +#endif // WEBP_USE_MSA