|
Packit |
9c6abc |
// Copyright 2016 Google Inc. All Rights Reserved.
|
|
Packit |
9c6abc |
//
|
|
Packit |
9c6abc |
// Use of this source code is governed by a BSD-style license
|
|
Packit |
9c6abc |
// that can be found in the COPYING file in the root of the source
|
|
Packit |
9c6abc |
// tree. An additional intellectual property rights grant can be found
|
|
Packit |
9c6abc |
// in the file PATENTS. All contributing project authors may
|
|
Packit |
9c6abc |
// be found in the AUTHORS file in the root of the source tree.
|
|
Packit |
9c6abc |
// -----------------------------------------------------------------------------
|
|
Packit |
9c6abc |
//
|
|
Packit |
9c6abc |
// MSA version of YUV to RGB upsampling functions.
|
|
Packit |
9c6abc |
//
|
|
Packit |
9c6abc |
// Author: Prashant Patil (prashant.patil@imgtec.com)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#include <string.h>
|
|
Packit |
9c6abc |
#include "src/dsp/dsp.h"
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#if defined(WEBP_USE_MSA)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#include "src/dsp/msa_macro.h"
|
|
Packit |
9c6abc |
#include "src/dsp/yuv.h"
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#ifdef FANCY_UPSAMPLING
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define ILVR_UW2(in, out0, out1) do { \
|
|
Packit |
9c6abc |
const v8i16 t0 = (v8i16)__msa_ilvr_b((v16i8)zero, (v16i8)in); \
|
|
Packit |
9c6abc |
out0 = (v4u32)__msa_ilvr_h((v8i16)zero, t0); \
|
|
Packit |
9c6abc |
out1 = (v4u32)__msa_ilvl_h((v8i16)zero, t0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define ILVRL_UW4(in, out0, out1, out2, out3) do { \
|
|
Packit |
9c6abc |
v16u8 t0, t1; \
|
|
Packit |
9c6abc |
ILVRL_B2_UB(zero, in, t0, t1); \
|
|
Packit |
9c6abc |
ILVRL_H2_UW(zero, t0, out0, out1); \
|
|
Packit |
9c6abc |
ILVRL_H2_UW(zero, t1, out2, out3); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define MULTHI_16(in0, in1, in2, in3, cnst, out0, out1) do { \
|
|
Packit |
9c6abc |
const v4i32 const0 = (v4i32)__msa_fill_w(cnst * 256); \
|
|
Packit |
9c6abc |
v4u32 temp0, temp1, temp2, temp3; \
|
|
Packit |
9c6abc |
MUL4(in0, const0, in1, const0, in2, const0, in3, const0, \
|
|
Packit |
9c6abc |
temp0, temp1, temp2, temp3); \
|
|
Packit |
9c6abc |
PCKOD_H2_UH(temp1, temp0, temp3, temp2, out0, out1); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define MULTHI_8(in0, in1, cnst, out0) do { \
|
|
Packit |
9c6abc |
const v4i32 const0 = (v4i32)__msa_fill_w(cnst * 256); \
|
|
Packit |
9c6abc |
v4u32 temp0, temp1; \
|
|
Packit |
9c6abc |
MUL2(in0, const0, in1, const0, temp0, temp1); \
|
|
Packit |
9c6abc |
out0 = (v8u16)__msa_pckod_h((v8i16)temp1, (v8i16)temp0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_R16(y0, y1, v0, v1, dst) do { \
|
|
Packit |
9c6abc |
const v8i16 const_a = (v8i16)__msa_fill_h(14234); \
|
|
Packit |
9c6abc |
const v8i16 a0 = __msa_adds_s_h((v8i16)y0, (v8i16)v0); \
|
|
Packit |
9c6abc |
const v8i16 a1 = __msa_adds_s_h((v8i16)y1, (v8i16)v1); \
|
|
Packit |
9c6abc |
v8i16 b0 = __msa_subs_s_h(a0, const_a); \
|
|
Packit |
9c6abc |
v8i16 b1 = __msa_subs_s_h(a1, const_a); \
|
|
Packit |
9c6abc |
SRAI_H2_SH(b0, b1, 6); \
|
|
Packit |
9c6abc |
CLIP_SH2_0_255(b0, b1); \
|
|
Packit |
9c6abc |
dst = (v16u8)__msa_pckev_b((v16i8)b1, (v16i8)b0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_R8(y0, v0, dst) do { \
|
|
Packit |
9c6abc |
const v8i16 const_a = (v8i16)__msa_fill_h(14234); \
|
|
Packit |
9c6abc |
const v8i16 a0 = __msa_adds_s_h((v8i16)y0, (v8i16)v0); \
|
|
Packit |
9c6abc |
v8i16 b0 = __msa_subs_s_h(a0, const_a); \
|
|
Packit |
9c6abc |
b0 = SRAI_H(b0, 6); \
|
|
Packit |
9c6abc |
CLIP_SH_0_255(b0); \
|
|
Packit |
9c6abc |
dst = (v16u8)__msa_pckev_b((v16i8)b0, (v16i8)b0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_G16(y0, y1, u0, u1, v0, v1, dst) do { \
|
|
Packit |
9c6abc |
const v8i16 const_a = (v8i16)__msa_fill_h(8708); \
|
|
Packit |
9c6abc |
v8i16 a0 = __msa_subs_s_h((v8i16)y0, (v8i16)u0); \
|
|
Packit |
9c6abc |
v8i16 a1 = __msa_subs_s_h((v8i16)y1, (v8i16)u1); \
|
|
Packit |
9c6abc |
const v8i16 b0 = __msa_subs_s_h(a0, (v8i16)v0); \
|
|
Packit |
9c6abc |
const v8i16 b1 = __msa_subs_s_h(a1, (v8i16)v1); \
|
|
Packit |
9c6abc |
a0 = __msa_adds_s_h(b0, const_a); \
|
|
Packit |
9c6abc |
a1 = __msa_adds_s_h(b1, const_a); \
|
|
Packit |
9c6abc |
SRAI_H2_SH(a0, a1, 6); \
|
|
Packit |
9c6abc |
CLIP_SH2_0_255(a0, a1); \
|
|
Packit |
9c6abc |
dst = (v16u8)__msa_pckev_b((v16i8)a1, (v16i8)a0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_G8(y0, u0, v0, dst) do { \
|
|
Packit |
9c6abc |
const v8i16 const_a = (v8i16)__msa_fill_h(8708); \
|
|
Packit |
9c6abc |
v8i16 a0 = __msa_subs_s_h((v8i16)y0, (v8i16)u0); \
|
|
Packit |
9c6abc |
const v8i16 b0 = __msa_subs_s_h(a0, (v8i16)v0); \
|
|
Packit |
9c6abc |
a0 = __msa_adds_s_h(b0, const_a); \
|
|
Packit |
9c6abc |
a0 = SRAI_H(a0, 6); \
|
|
Packit |
9c6abc |
CLIP_SH_0_255(a0); \
|
|
Packit |
9c6abc |
dst = (v16u8)__msa_pckev_b((v16i8)a0, (v16i8)a0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_B16(y0, y1, u0, u1, dst) do { \
|
|
Packit |
9c6abc |
const v8u16 const_a = (v8u16)__msa_fill_h(17685); \
|
|
Packit |
9c6abc |
const v8u16 a0 = __msa_adds_u_h((v8u16)y0, u0); \
|
|
Packit |
9c6abc |
const v8u16 a1 = __msa_adds_u_h((v8u16)y1, u1); \
|
|
Packit |
9c6abc |
v8u16 b0 = __msa_subs_u_h(a0, const_a); \
|
|
Packit |
9c6abc |
v8u16 b1 = __msa_subs_u_h(a1, const_a); \
|
|
Packit |
9c6abc |
SRAI_H2_UH(b0, b1, 6); \
|
|
Packit |
9c6abc |
CLIP_UH2_0_255(b0, b1); \
|
|
Packit |
9c6abc |
dst = (v16u8)__msa_pckev_b((v16i8)b1, (v16i8)b0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_B8(y0, u0, dst) do { \
|
|
Packit |
9c6abc |
const v8u16 const_a = (v8u16)__msa_fill_h(17685); \
|
|
Packit |
9c6abc |
const v8u16 a0 = __msa_adds_u_h((v8u16)y0, u0); \
|
|
Packit |
9c6abc |
v8u16 b0 = __msa_subs_u_h(a0, const_a); \
|
|
Packit |
9c6abc |
b0 = SRAI_H(b0, 6); \
|
|
Packit |
9c6abc |
CLIP_UH_0_255(b0); \
|
|
Packit |
9c6abc |
dst = (v16u8)__msa_pckev_b((v16i8)b0, (v16i8)b0); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_RGB16(y, u, v, R, G, B) do { \
|
|
Packit |
9c6abc |
const v16u8 zero = { 0 }; \
|
|
Packit |
9c6abc |
v8u16 y0, y1, u0, u1, v0, v1; \
|
|
Packit |
9c6abc |
v4u32 p0, p1, p2, p3; \
|
|
Packit |
9c6abc |
const v16u8 in_y = LD_UB(y); \
|
|
Packit |
9c6abc |
const v16u8 in_u = LD_UB(u); \
|
|
Packit |
9c6abc |
const v16u8 in_v = LD_UB(v); \
|
|
Packit |
9c6abc |
ILVRL_UW4(in_y, p0, p1, p2, p3); \
|
|
Packit |
9c6abc |
MULTHI_16(p0, p1, p2, p3, 19077, y0, y1); \
|
|
Packit |
9c6abc |
ILVRL_UW4(in_v, p0, p1, p2, p3); \
|
|
Packit |
9c6abc |
MULTHI_16(p0, p1, p2, p3, 26149, v0, v1); \
|
|
Packit |
9c6abc |
CALC_R16(y0, y1, v0, v1, R); \
|
|
Packit |
9c6abc |
MULTHI_16(p0, p1, p2, p3, 13320, v0, v1); \
|
|
Packit |
9c6abc |
ILVRL_UW4(in_u, p0, p1, p2, p3); \
|
|
Packit |
9c6abc |
MULTHI_16(p0, p1, p2, p3, 6419, u0, u1); \
|
|
Packit |
9c6abc |
CALC_G16(y0, y1, u0, u1, v0, v1, G); \
|
|
Packit |
9c6abc |
MULTHI_16(p0, p1, p2, p3, 33050, u0, u1); \
|
|
Packit |
9c6abc |
CALC_B16(y0, y1, u0, u1, B); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_RGB8(y, u, v, R, G, B) do { \
|
|
Packit |
9c6abc |
const v16u8 zero = { 0 }; \
|
|
Packit |
9c6abc |
v8u16 y0, u0, v0; \
|
|
Packit |
9c6abc |
v4u32 p0, p1; \
|
|
Packit |
9c6abc |
const v16u8 in_y = LD_UB(y); \
|
|
Packit |
9c6abc |
const v16u8 in_u = LD_UB(u); \
|
|
Packit |
9c6abc |
const v16u8 in_v = LD_UB(v); \
|
|
Packit |
9c6abc |
ILVR_UW2(in_y, p0, p1); \
|
|
Packit |
9c6abc |
MULTHI_8(p0, p1, 19077, y0); \
|
|
Packit |
9c6abc |
ILVR_UW2(in_v, p0, p1); \
|
|
Packit |
9c6abc |
MULTHI_8(p0, p1, 26149, v0); \
|
|
Packit |
9c6abc |
CALC_R8(y0, v0, R); \
|
|
Packit |
9c6abc |
MULTHI_8(p0, p1, 13320, v0); \
|
|
Packit |
9c6abc |
ILVR_UW2(in_u, p0, p1); \
|
|
Packit |
9c6abc |
MULTHI_8(p0, p1, 6419, u0); \
|
|
Packit |
9c6abc |
CALC_G8(y0, u0, v0, G); \
|
|
Packit |
9c6abc |
MULTHI_8(p0, p1, 33050, u0); \
|
|
Packit |
9c6abc |
CALC_B8(y0, u0, B); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define STORE16_3(a0, a1, a2, dst) do { \
|
|
Packit |
9c6abc |
const v16u8 mask0 = { 0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, \
|
|
Packit |
9c6abc |
8, 9, 20, 10 }; \
|
|
Packit |
9c6abc |
const v16u8 mask1 = { 0, 21, 1, 2, 22, 3, 4, 23, 5, 6, 24, 7, \
|
|
Packit |
9c6abc |
8, 25, 9, 10 }; \
|
|
Packit |
9c6abc |
const v16u8 mask2 = { 26, 0, 1, 27, 2, 3, 28, 4, 5, 29, 6, 7, \
|
|
Packit |
9c6abc |
30, 8, 9, 31 }; \
|
|
Packit |
9c6abc |
v16u8 out0, out1, out2, tmp0, tmp1, tmp2; \
|
|
Packit |
9c6abc |
ILVRL_B2_UB(a1, a0, tmp0, tmp1); \
|
|
Packit |
9c6abc |
out0 = VSHF_UB(tmp0, a2, mask0); \
|
|
Packit |
9c6abc |
tmp2 = SLDI_UB(tmp1, tmp0, 11); \
|
|
Packit |
9c6abc |
out1 = VSHF_UB(tmp2, a2, mask1); \
|
|
Packit |
9c6abc |
tmp2 = SLDI_UB(tmp1, tmp1, 6); \
|
|
Packit |
9c6abc |
out2 = VSHF_UB(tmp2, a2, mask2); \
|
|
Packit |
9c6abc |
ST_UB(out0, dst + 0); \
|
|
Packit |
9c6abc |
ST_UB(out1, dst + 16); \
|
|
Packit |
9c6abc |
ST_UB(out2, dst + 32); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define STORE8_3(a0, a1, a2, dst) do { \
|
|
Packit |
9c6abc |
int64_t out_m; \
|
|
Packit |
9c6abc |
const v16u8 mask0 = { 0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, \
|
|
Packit |
9c6abc |
8, 9, 20, 10 }; \
|
|
Packit |
9c6abc |
const v16u8 mask1 = { 11, 21, 12, 13, 22, 14, 15, 23, \
|
|
Packit |
9c6abc |
255, 255, 255, 255, 255, 255, 255, 255 }; \
|
|
Packit |
9c6abc |
const v16u8 tmp0 = (v16u8)__msa_ilvr_b((v16i8)a1, (v16i8)a0); \
|
|
Packit |
9c6abc |
v16u8 out0, out1; \
|
|
Packit |
9c6abc |
VSHF_B2_UB(tmp0, a2, tmp0, a2, mask0, mask1, out0, out1); \
|
|
Packit |
9c6abc |
ST_UB(out0, dst); \
|
|
Packit |
9c6abc |
out_m = __msa_copy_s_d((v2i64)out1, 0); \
|
|
Packit |
9c6abc |
SD(out_m, dst + 16); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define STORE16_4(a0, a1, a2, a3, dst) do { \
|
|
Packit |
9c6abc |
v16u8 tmp0, tmp1, tmp2, tmp3; \
|
|
Packit |
9c6abc |
v16u8 out0, out1, out2, out3; \
|
|
Packit |
9c6abc |
ILVRL_B2_UB(a1, a0, tmp0, tmp1); \
|
|
Packit |
9c6abc |
ILVRL_B2_UB(a3, a2, tmp2, tmp3); \
|
|
Packit |
9c6abc |
ILVRL_H2_UB(tmp2, tmp0, out0, out1); \
|
|
Packit |
9c6abc |
ILVRL_H2_UB(tmp3, tmp1, out2, out3); \
|
|
Packit |
9c6abc |
ST_UB(out0, dst + 0); \
|
|
Packit |
9c6abc |
ST_UB(out1, dst + 16); \
|
|
Packit |
9c6abc |
ST_UB(out2, dst + 32); \
|
|
Packit |
9c6abc |
ST_UB(out3, dst + 48); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define STORE8_4(a0, a1, a2, a3, dst) do { \
|
|
Packit |
9c6abc |
v16u8 tmp0, tmp1, tmp2, tmp3; \
|
|
Packit |
9c6abc |
ILVR_B2_UB(a1, a0, a3, a2, tmp0, tmp1); \
|
|
Packit |
9c6abc |
ILVRL_H2_UB(tmp1, tmp0, tmp2, tmp3); \
|
|
Packit |
9c6abc |
ST_UB(tmp2, dst + 0); \
|
|
Packit |
9c6abc |
ST_UB(tmp3, dst + 16); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define STORE2_16(a0, a1, dst) do { \
|
|
Packit |
9c6abc |
v16u8 out0, out1; \
|
|
Packit |
9c6abc |
ILVRL_B2_UB(a1, a0, out0, out1); \
|
|
Packit |
9c6abc |
ST_UB(out0, dst + 0); \
|
|
Packit |
9c6abc |
ST_UB(out1, dst + 16); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define STORE2_8(a0, a1, dst) do { \
|
|
Packit |
9c6abc |
const v16u8 out0 = (v16u8)__msa_ilvr_b((v16i8)a1, (v16i8)a0); \
|
|
Packit |
9c6abc |
ST_UB(out0, dst); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_RGBA4444(y, u, v, out0, out1, N, dst) do { \
|
|
Packit |
9c6abc |
CALC_RGB##N(y, u, v, R, G, B); \
|
|
Packit |
9c6abc |
tmp0 = ANDI_B(R, 0xf0); \
|
|
Packit |
9c6abc |
tmp1 = SRAI_B(G, 4); \
|
|
Packit |
9c6abc |
RG = tmp0 | tmp1; \
|
|
Packit |
9c6abc |
tmp0 = ANDI_B(B, 0xf0); \
|
|
Packit |
9c6abc |
BA = ORI_B(tmp0, 0x0f); \
|
|
Packit |
9c6abc |
STORE2_##N(out0, out1, dst); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define CALC_RGB565(y, u, v, out0, out1, N, dst) do { \
|
|
Packit |
9c6abc |
CALC_RGB##N(y, u, v, R, G, B); \
|
|
Packit |
9c6abc |
tmp0 = ANDI_B(R, 0xf8); \
|
|
Packit |
9c6abc |
tmp1 = SRAI_B(G, 5); \
|
|
Packit |
9c6abc |
RG = tmp0 | tmp1; \
|
|
Packit |
9c6abc |
tmp0 = SLLI_B(G, 3); \
|
|
Packit |
9c6abc |
tmp1 = ANDI_B(tmp0, 0xe0); \
|
|
Packit |
9c6abc |
tmp0 = SRAI_B(B, 3); \
|
|
Packit |
9c6abc |
GB = tmp0 | tmp1; \
|
|
Packit |
9c6abc |
STORE2_##N(out0, out1, dst); \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static WEBP_INLINE int Clip8(int v) {
|
|
Packit |
9c6abc |
return v < 0 ? 0 : v > 255 ? 255 : v;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToRgb(int y, int u, int v, uint8_t* const rgb) {
|
|
Packit |
9c6abc |
const int y1 = MultHi(y, 19077);
|
|
Packit |
9c6abc |
const int r1 = y1 + MultHi(v, 26149) - 14234;
|
|
Packit |
9c6abc |
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
|
|
Packit |
9c6abc |
const int b1 = y1 + MultHi(u, 33050) - 17685;
|
|
Packit |
9c6abc |
rgb[0] = Clip8(r1 >> 6);
|
|
Packit |
9c6abc |
rgb[1] = Clip8(g1 >> 6);
|
|
Packit |
9c6abc |
rgb[2] = Clip8(b1 >> 6);
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToBgr(int y, int u, int v, uint8_t* const bgr) {
|
|
Packit |
9c6abc |
const int y1 = MultHi(y, 19077);
|
|
Packit |
9c6abc |
const int r1 = y1 + MultHi(v, 26149) - 14234;
|
|
Packit |
9c6abc |
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
|
|
Packit |
9c6abc |
const int b1 = y1 + MultHi(u, 33050) - 17685;
|
|
Packit |
9c6abc |
bgr[0] = Clip8(b1 >> 6);
|
|
Packit |
9c6abc |
bgr[1] = Clip8(g1 >> 6);
|
|
Packit |
9c6abc |
bgr[2] = Clip8(r1 >> 6);
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#if !defined(WEBP_REDUCE_CSP)
|
|
Packit |
9c6abc |
static void YuvToRgb565(int y, int u, int v, uint8_t* const rgb) {
|
|
Packit |
9c6abc |
const int y1 = MultHi(y, 19077);
|
|
Packit |
9c6abc |
const int r1 = y1 + MultHi(v, 26149) - 14234;
|
|
Packit |
9c6abc |
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
|
|
Packit |
9c6abc |
const int b1 = y1 + MultHi(u, 33050) - 17685;
|
|
Packit |
9c6abc |
const int r = Clip8(r1 >> 6);
|
|
Packit |
9c6abc |
const int g = Clip8(g1 >> 6);
|
|
Packit |
9c6abc |
const int b = Clip8(b1 >> 6);
|
|
Packit |
9c6abc |
const int rg = (r & 0xf8) | (g >> 5);
|
|
Packit |
9c6abc |
const int gb = ((g << 3) & 0xe0) | (b >> 3);
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
rgb[0] = gb;
|
|
Packit |
9c6abc |
rgb[1] = rg;
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
rgb[0] = rg;
|
|
Packit |
9c6abc |
rgb[1] = gb;
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToRgba4444(int y, int u, int v, uint8_t* const argb) {
|
|
Packit |
9c6abc |
const int y1 = MultHi(y, 19077);
|
|
Packit |
9c6abc |
const int r1 = y1 + MultHi(v, 26149) - 14234;
|
|
Packit |
9c6abc |
const int g1 = y1 - MultHi(u, 6419) - MultHi(v, 13320) + 8708;
|
|
Packit |
9c6abc |
const int b1 = y1 + MultHi(u, 33050) - 17685;
|
|
Packit |
9c6abc |
const int r = Clip8(r1 >> 6);
|
|
Packit |
9c6abc |
const int g = Clip8(g1 >> 6);
|
|
Packit |
9c6abc |
const int b = Clip8(b1 >> 6);
|
|
Packit |
9c6abc |
const int rg = (r & 0xf0) | (g >> 4);
|
|
Packit |
9c6abc |
const int ba = (b & 0xf0) | 0x0f; // overwrite the lower 4 bits
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
argb[0] = ba;
|
|
Packit |
9c6abc |
argb[1] = rg;
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
argb[0] = rg;
|
|
Packit |
9c6abc |
argb[1] = ba;
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToArgb(uint8_t y, uint8_t u, uint8_t v, uint8_t* const argb) {
|
|
Packit |
9c6abc |
argb[0] = 0xff;
|
|
Packit |
9c6abc |
YuvToRgb(y, u, v, argb + 1);
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
#endif // WEBP_REDUCE_CSP
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToBgra(uint8_t y, uint8_t u, uint8_t v, uint8_t* const bgra) {
|
|
Packit |
9c6abc |
YuvToBgr(y, u, v, bgra);
|
|
Packit |
9c6abc |
bgra[3] = 0xff;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToRgba(uint8_t y, uint8_t u, uint8_t v, uint8_t* const rgba) {
|
|
Packit |
9c6abc |
YuvToRgb(y, u, v, rgba);
|
|
Packit |
9c6abc |
rgba[3] = 0xff;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#if !defined(WEBP_REDUCE_CSP)
|
|
Packit |
9c6abc |
static void YuvToRgbLine(const uint8_t* y, const uint8_t* u,
|
|
Packit |
9c6abc |
const uint8_t* v, uint8_t* dst, int length) {
|
|
Packit |
9c6abc |
v16u8 R, G, B;
|
|
Packit |
9c6abc |
while (length >= 16) {
|
|
Packit |
9c6abc |
CALC_RGB16(y, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_3(R, G, B, dst);
|
|
Packit |
9c6abc |
y += 16;
|
|
Packit |
9c6abc |
u += 16;
|
|
Packit |
9c6abc |
v += 16;
|
|
Packit |
9c6abc |
dst += 16 * 3;
|
|
Packit |
9c6abc |
length -= 16;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
if (length > 8) {
|
|
Packit |
9c6abc |
uint8_t temp[3 * 16] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB16(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_3(R, G, B, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 3 * sizeof(*dst));
|
|
Packit |
9c6abc |
} else if (length > 0) {
|
|
Packit |
9c6abc |
uint8_t temp[3 * 8] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB8(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE8_3(R, G, B, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 3 * sizeof(*dst));
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToBgrLine(const uint8_t* y, const uint8_t* u,
|
|
Packit |
9c6abc |
const uint8_t* v, uint8_t* dst, int length) {
|
|
Packit |
9c6abc |
v16u8 R, G, B;
|
|
Packit |
9c6abc |
while (length >= 16) {
|
|
Packit |
9c6abc |
CALC_RGB16(y, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_3(B, G, R, dst);
|
|
Packit |
9c6abc |
y += 16;
|
|
Packit |
9c6abc |
u += 16;
|
|
Packit |
9c6abc |
v += 16;
|
|
Packit |
9c6abc |
dst += 16 * 3;
|
|
Packit |
9c6abc |
length -= 16;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
if (length > 8) {
|
|
Packit |
9c6abc |
uint8_t temp[3 * 16] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB16(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_3(B, G, R, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 3 * sizeof(*dst));
|
|
Packit |
9c6abc |
} else if (length > 0) {
|
|
Packit |
9c6abc |
uint8_t temp[3 * 8] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB8(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE8_3(B, G, R, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 3 * sizeof(*dst));
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
#endif // WEBP_REDUCE_CSP
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToRgbaLine(const uint8_t* y, const uint8_t* u,
|
|
Packit |
9c6abc |
const uint8_t* v, uint8_t* dst, int length) {
|
|
Packit |
9c6abc |
v16u8 R, G, B;
|
|
Packit |
9c6abc |
const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL);
|
|
Packit |
9c6abc |
while (length >= 16) {
|
|
Packit |
9c6abc |
CALC_RGB16(y, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_4(R, G, B, A, dst);
|
|
Packit |
9c6abc |
y += 16;
|
|
Packit |
9c6abc |
u += 16;
|
|
Packit |
9c6abc |
v += 16;
|
|
Packit |
9c6abc |
dst += 16 * 4;
|
|
Packit |
9c6abc |
length -= 16;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
if (length > 8) {
|
|
Packit |
9c6abc |
uint8_t temp[4 * 16] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB16(&temp[0], u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_4(R, G, B, A, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 4 * sizeof(*dst));
|
|
Packit |
9c6abc |
} else if (length > 0) {
|
|
Packit |
9c6abc |
uint8_t temp[4 * 8] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB8(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE8_4(R, G, B, A, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 4 * sizeof(*dst));
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToBgraLine(const uint8_t* y, const uint8_t* u,
|
|
Packit |
9c6abc |
const uint8_t* v, uint8_t* dst, int length) {
|
|
Packit |
9c6abc |
v16u8 R, G, B;
|
|
Packit |
9c6abc |
const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL);
|
|
Packit |
9c6abc |
while (length >= 16) {
|
|
Packit |
9c6abc |
CALC_RGB16(y, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_4(B, G, R, A, dst);
|
|
Packit |
9c6abc |
y += 16;
|
|
Packit |
9c6abc |
u += 16;
|
|
Packit |
9c6abc |
v += 16;
|
|
Packit |
9c6abc |
dst += 16 * 4;
|
|
Packit |
9c6abc |
length -= 16;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
if (length > 8) {
|
|
Packit |
9c6abc |
uint8_t temp[4 * 16] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB16(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_4(B, G, R, A, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 4 * sizeof(*dst));
|
|
Packit |
9c6abc |
} else if (length > 0) {
|
|
Packit |
9c6abc |
uint8_t temp[4 * 8] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB8(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE8_4(B, G, R, A, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 4 * sizeof(*dst));
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#if !defined(WEBP_REDUCE_CSP)
|
|
Packit |
9c6abc |
static void YuvToArgbLine(const uint8_t* y, const uint8_t* u,
|
|
Packit |
9c6abc |
const uint8_t* v, uint8_t* dst, int length) {
|
|
Packit |
9c6abc |
v16u8 R, G, B;
|
|
Packit |
9c6abc |
const v16u8 A = (v16u8)__msa_ldi_b(ALPHAVAL);
|
|
Packit |
9c6abc |
while (length >= 16) {
|
|
Packit |
9c6abc |
CALC_RGB16(y, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_4(A, R, G, B, dst);
|
|
Packit |
9c6abc |
y += 16;
|
|
Packit |
9c6abc |
u += 16;
|
|
Packit |
9c6abc |
v += 16;
|
|
Packit |
9c6abc |
dst += 16 * 4;
|
|
Packit |
9c6abc |
length -= 16;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
if (length > 8) {
|
|
Packit |
9c6abc |
uint8_t temp[4 * 16] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB16(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE16_4(A, R, G, B, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 4 * sizeof(*dst));
|
|
Packit |
9c6abc |
} else if (length > 0) {
|
|
Packit |
9c6abc |
uint8_t temp[4 * 8] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
CALC_RGB8(temp, u, v, R, G, B);
|
|
Packit |
9c6abc |
STORE8_4(A, R, G, B, temp);
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 4 * sizeof(*dst));
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToRgba4444Line(const uint8_t* y, const uint8_t* u,
|
|
Packit |
9c6abc |
const uint8_t* v, uint8_t* dst, int length) {
|
|
Packit |
9c6abc |
v16u8 R, G, B, RG, BA, tmp0, tmp1;
|
|
Packit |
9c6abc |
while (length >= 16) {
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
CALC_RGBA4444(y, u, v, BA, RG, 16, dst);
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
CALC_RGBA4444(y, u, v, RG, BA, 16, dst);
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
y += 16;
|
|
Packit |
9c6abc |
u += 16;
|
|
Packit |
9c6abc |
v += 16;
|
|
Packit |
9c6abc |
dst += 16 * 2;
|
|
Packit |
9c6abc |
length -= 16;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
if (length > 8) {
|
|
Packit |
9c6abc |
uint8_t temp[2 * 16] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
CALC_RGBA4444(temp, u, v, BA, RG, 16, temp);
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
CALC_RGBA4444(temp, u, v, RG, BA, 16, temp);
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 2 * sizeof(*dst));
|
|
Packit |
9c6abc |
} else if (length > 0) {
|
|
Packit |
9c6abc |
uint8_t temp[2 * 8] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
CALC_RGBA4444(temp, u, v, BA, RG, 8, temp);
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
CALC_RGBA4444(temp, u, v, RG, BA, 8, temp);
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 2 * sizeof(*dst));
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
static void YuvToRgb565Line(const uint8_t* y, const uint8_t* u,
|
|
Packit |
9c6abc |
const uint8_t* v, uint8_t* dst, int length) {
|
|
Packit |
9c6abc |
v16u8 R, G, B, RG, GB, tmp0, tmp1;
|
|
Packit |
9c6abc |
while (length >= 16) {
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
CALC_RGB565(y, u, v, GB, RG, 16, dst);
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
CALC_RGB565(y, u, v, RG, GB, 16, dst);
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
y += 16;
|
|
Packit |
9c6abc |
u += 16;
|
|
Packit |
9c6abc |
v += 16;
|
|
Packit |
9c6abc |
dst += 16 * 2;
|
|
Packit |
9c6abc |
length -= 16;
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
if (length > 8) {
|
|
Packit |
9c6abc |
uint8_t temp[2 * 16] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
CALC_RGB565(temp, u, v, GB, RG, 16, temp);
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
CALC_RGB565(temp, u, v, RG, GB, 16, temp);
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 2 * sizeof(*dst));
|
|
Packit |
9c6abc |
} else if (length > 0) {
|
|
Packit |
9c6abc |
uint8_t temp[2 * 8] = { 0 };
|
|
Packit |
9c6abc |
memcpy(temp, y, length * sizeof(*temp));
|
|
Packit |
9c6abc |
#if (WEBP_SWAP_16BIT_CSP == 1)
|
|
Packit |
9c6abc |
CALC_RGB565(temp, u, v, GB, RG, 8, temp);
|
|
Packit |
9c6abc |
#else
|
|
Packit |
9c6abc |
CALC_RGB565(temp, u, v, RG, GB, 8, temp);
|
|
Packit |
9c6abc |
#endif
|
|
Packit |
9c6abc |
memcpy(dst, temp, length * 2 * sizeof(*dst));
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
#endif // WEBP_REDUCE_CSP
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define UPSAMPLE_32PIXELS(a, b, c, d) do { \
|
|
Packit |
9c6abc |
v16u8 s = __msa_aver_u_b(a, d); \
|
|
Packit |
9c6abc |
v16u8 t = __msa_aver_u_b(b, c); \
|
|
Packit |
9c6abc |
const v16u8 st = s ^ t; \
|
|
Packit |
9c6abc |
v16u8 ad = a ^ d; \
|
|
Packit |
9c6abc |
v16u8 bc = b ^ c; \
|
|
Packit |
9c6abc |
v16u8 t0 = ad | bc; \
|
|
Packit |
9c6abc |
v16u8 t1 = t0 | st; \
|
|
Packit |
9c6abc |
v16u8 t2 = ANDI_B(t1, 1); \
|
|
Packit |
9c6abc |
v16u8 t3 = __msa_aver_u_b(s, t); \
|
|
Packit |
9c6abc |
const v16u8 k = t3 - t2; \
|
|
Packit |
9c6abc |
v16u8 diag1, diag2; \
|
|
Packit |
9c6abc |
AVER_UB2_UB(t, k, s, k, t0, t1); \
|
|
Packit |
9c6abc |
bc = bc & st; \
|
|
Packit |
9c6abc |
ad = ad & st; \
|
|
Packit |
9c6abc |
t = t ^ k; \
|
|
Packit |
9c6abc |
s = s ^ k; \
|
|
Packit |
9c6abc |
t2 = bc | t; \
|
|
Packit |
9c6abc |
t3 = ad | s; \
|
|
Packit |
9c6abc |
t2 = ANDI_B(t2, 1); \
|
|
Packit |
9c6abc |
t3 = ANDI_B(t3, 1); \
|
|
Packit |
9c6abc |
SUB2(t0, t2, t1, t3, diag1, diag2); \
|
|
Packit |
9c6abc |
AVER_UB2_UB(a, diag1, b, diag2, t0, t1); \
|
|
Packit |
9c6abc |
ILVRL_B2_UB(t1, t0, a, b); \
|
|
Packit |
9c6abc |
if (pbot_y != NULL) { \
|
|
Packit |
9c6abc |
AVER_UB2_UB(c, diag2, d, diag1, t0, t1); \
|
|
Packit |
9c6abc |
ILVRL_B2_UB(t1, t0, c, d); \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
} while (0)
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \
|
|
Packit |
9c6abc |
static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y, \
|
|
Packit |
9c6abc |
const uint8_t* top_u, const uint8_t* top_v, \
|
|
Packit |
9c6abc |
const uint8_t* cur_u, const uint8_t* cur_v, \
|
|
Packit |
9c6abc |
uint8_t* top_dst, uint8_t* bot_dst, int len) \
|
|
Packit |
9c6abc |
{ \
|
|
Packit |
9c6abc |
int size = (len - 1) >> 1; \
|
|
Packit |
9c6abc |
uint8_t temp_u[64]; \
|
|
Packit |
9c6abc |
uint8_t temp_v[64]; \
|
|
Packit |
9c6abc |
const uint32_t tl_uv = ((top_u[0]) | ((top_v[0]) << 16)); \
|
|
Packit |
9c6abc |
const uint32_t l_uv = ((cur_u[0]) | ((cur_v[0]) << 16)); \
|
|
Packit |
9c6abc |
const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \
|
|
Packit |
9c6abc |
const uint8_t* ptop_y = &top_y[1]; \
|
|
Packit |
9c6abc |
uint8_t *ptop_dst = top_dst + XSTEP; \
|
|
Packit |
9c6abc |
const uint8_t* pbot_y = &bot_y[1]; \
|
|
Packit |
9c6abc |
uint8_t *pbot_dst = bot_dst + XSTEP; \
|
|
Packit |
9c6abc |
\
|
|
Packit |
9c6abc |
FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \
|
|
Packit |
9c6abc |
if (bot_y != NULL) { \
|
|
Packit |
9c6abc |
const uint32_t uv1 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \
|
|
Packit |
9c6abc |
FUNC(bot_y[0], uv1 & 0xff, (uv1 >> 16), bot_dst); \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
while (size >= 16) { \
|
|
Packit |
9c6abc |
v16u8 tu0, tu1, tv0, tv1, cu0, cu1, cv0, cv1; \
|
|
Packit |
9c6abc |
LD_UB2(top_u, 1, tu0, tu1); \
|
|
Packit |
9c6abc |
LD_UB2(cur_u, 1, cu0, cu1); \
|
|
Packit |
9c6abc |
LD_UB2(top_v, 1, tv0, tv1); \
|
|
Packit |
9c6abc |
LD_UB2(cur_v, 1, cv0, cv1); \
|
|
Packit |
9c6abc |
UPSAMPLE_32PIXELS(tu0, tu1, cu0, cu1); \
|
|
Packit |
9c6abc |
UPSAMPLE_32PIXELS(tv0, tv1, cv0, cv1); \
|
|
Packit |
9c6abc |
ST_UB4(tu0, tu1, cu0, cu1, &temp_u[0], 16); \
|
|
Packit |
9c6abc |
ST_UB4(tv0, tv1, cv0, cv1, &temp_v[0], 16); \
|
|
Packit |
9c6abc |
FUNC##Line(ptop_y, &temp_u[ 0], &temp_v[0], ptop_dst, 32); \
|
|
Packit |
9c6abc |
if (bot_y != NULL) { \
|
|
Packit |
9c6abc |
FUNC##Line(pbot_y, &temp_u[32], &temp_v[32], pbot_dst, 32); \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
ptop_y += 32; \
|
|
Packit |
9c6abc |
pbot_y += 32; \
|
|
Packit |
9c6abc |
ptop_dst += XSTEP * 32; \
|
|
Packit |
9c6abc |
pbot_dst += XSTEP * 32; \
|
|
Packit |
9c6abc |
top_u += 16; \
|
|
Packit |
9c6abc |
top_v += 16; \
|
|
Packit |
9c6abc |
cur_u += 16; \
|
|
Packit |
9c6abc |
cur_v += 16; \
|
|
Packit |
9c6abc |
size -= 16; \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
if (size > 0) { \
|
|
Packit |
9c6abc |
v16u8 tu0, tu1, tv0, tv1, cu0, cu1, cv0, cv1; \
|
|
Packit |
9c6abc |
memcpy(&temp_u[ 0], top_u, 17 * sizeof(uint8_t)); \
|
|
Packit |
9c6abc |
memcpy(&temp_u[32], cur_u, 17 * sizeof(uint8_t)); \
|
|
Packit |
9c6abc |
memcpy(&temp_v[ 0], top_v, 17 * sizeof(uint8_t)); \
|
|
Packit |
9c6abc |
memcpy(&temp_v[32], cur_v, 17 * sizeof(uint8_t)); \
|
|
Packit |
9c6abc |
LD_UB2(&temp_u[ 0], 1, tu0, tu1); \
|
|
Packit |
9c6abc |
LD_UB2(&temp_u[32], 1, cu0, cu1); \
|
|
Packit |
9c6abc |
LD_UB2(&temp_v[ 0], 1, tv0, tv1); \
|
|
Packit |
9c6abc |
LD_UB2(&temp_v[32], 1, cv0, cv1); \
|
|
Packit |
9c6abc |
UPSAMPLE_32PIXELS(tu0, tu1, cu0, cu1); \
|
|
Packit |
9c6abc |
UPSAMPLE_32PIXELS(tv0, tv1, cv0, cv1); \
|
|
Packit |
9c6abc |
ST_UB4(tu0, tu1, cu0, cu1, &temp_u[0], 16); \
|
|
Packit |
9c6abc |
ST_UB4(tv0, tv1, cv0, cv1, &temp_v[0], 16); \
|
|
Packit |
9c6abc |
FUNC##Line(ptop_y, &temp_u[ 0], &temp_v[0], ptop_dst, size * 2); \
|
|
Packit |
9c6abc |
if (bot_y != NULL) { \
|
|
Packit |
9c6abc |
FUNC##Line(pbot_y, &temp_u[32], &temp_v[32], pbot_dst, size * 2); \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
top_u += size; \
|
|
Packit |
9c6abc |
top_v += size; \
|
|
Packit |
9c6abc |
cur_u += size; \
|
|
Packit |
9c6abc |
cur_v += size; \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
if (!(len & 1)) { \
|
|
Packit |
9c6abc |
const uint32_t t0 = ((top_u[0]) | ((top_v[0]) << 16)); \
|
|
Packit |
9c6abc |
const uint32_t c0 = ((cur_u[0]) | ((cur_v[0]) << 16)); \
|
|
Packit |
9c6abc |
const uint32_t tmp0 = (3 * t0 + c0 + 0x00020002u) >> 2; \
|
|
Packit |
9c6abc |
FUNC(top_y[len - 1], tmp0 & 0xff, (tmp0 >> 16), \
|
|
Packit |
9c6abc |
top_dst + (len - 1) * XSTEP); \
|
|
Packit |
9c6abc |
if (bot_y != NULL) { \
|
|
Packit |
9c6abc |
const uint32_t tmp1 = (3 * c0 + t0 + 0x00020002u) >> 2; \
|
|
Packit |
9c6abc |
FUNC(bot_y[len - 1], tmp1 & 0xff, (tmp1 >> 16), \
|
|
Packit |
9c6abc |
bot_dst + (len - 1) * XSTEP); \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
} \
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
UPSAMPLE_FUNC(UpsampleRgbaLinePair, YuvToRgba, 4)
|
|
Packit |
9c6abc |
UPSAMPLE_FUNC(UpsampleBgraLinePair, YuvToBgra, 4)
|
|
Packit |
9c6abc |
#if !defined(WEBP_REDUCE_CSP)
|
|
Packit |
9c6abc |
UPSAMPLE_FUNC(UpsampleRgbLinePair, YuvToRgb, 3)
|
|
Packit |
9c6abc |
UPSAMPLE_FUNC(UpsampleBgrLinePair, YuvToBgr, 3)
|
|
Packit |
9c6abc |
UPSAMPLE_FUNC(UpsampleArgbLinePair, YuvToArgb, 4)
|
|
Packit |
9c6abc |
UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2)
|
|
Packit |
9c6abc |
UPSAMPLE_FUNC(UpsampleRgb565LinePair, YuvToRgb565, 2)
|
|
Packit |
9c6abc |
#endif // WEBP_REDUCE_CSP
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
//------------------------------------------------------------------------------
|
|
Packit |
9c6abc |
// Entry point
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
extern void WebPInitUpsamplersMSA(void);
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMSA(void) {
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
|
|
Packit |
9c6abc |
#if !defined(WEBP_REDUCE_CSP)
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
|
|
Packit |
9c6abc |
WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
|
|
Packit |
9c6abc |
#endif // WEBP_REDUCE_CSP
|
|
Packit |
9c6abc |
}
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#endif // FANCY_UPSAMPLING
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#endif // WEBP_USE_MSA
|
|
Packit |
9c6abc |
|
|
Packit |
9c6abc |
#if !(defined(FANCY_UPSAMPLING) && defined(WEBP_USE_MSA))
|
|
Packit |
9c6abc |
WEBP_DSP_INIT_STUB(WebPInitUpsamplersMSA)
|
|
Packit |
9c6abc |
#endif
|