dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2025-07-16 13:59:51 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

5-7% faster SSE2 versions of YUV->RGB conversion functions

Browse Source 
The C-version gets ~7-8% slower in order to match the SSE2
output exactly. The old (now off-by-1) code is kept under
the WEBP_YUV_USE_TABLE flag for reference.

(note that calc rounding precision is slightly better ~= +0.02dB)

on ARM-neon, we somehow recover the ~4% speed that was lost by mimicking
the initial C-version (see https://gerrit.chromium.org/gerrit/#/c/41610)

Change-Id: Ia4363c5ed9b4c9edff5d932b002e57bb7814bf6f
This commit is contained in:
skal
2013-08-19 12:43:53 -07:00
parent ad6ac32d7c
commit df6cebfa9e
4 changed files with 365 additions and 218 deletions
Download Patch File Download Diff File Expand all files Collapse all files

138
src/dsp/upsampling_neon.c
View File

@ -27,6 +27,9 @@ extern "C" {
#ifdef FANCY_UPSAMPLING
//-----------------------------------------------------------------------------
// U/V upsampling
// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
#define UPSAMPLE_16PIXELS(r1, r2, out) { \
uint8x8_t a = vld1_u8(r1); \
@ -85,105 +88,74 @@ static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
Upsample16Pixels(r1, r2, out); \
}
#define CY 76283
#define CVR 89858
#define CUG 22014
#define CVG 45773
#define CUB 113618
//-----------------------------------------------------------------------------
// YUV->RGB conversion
static const int16_t coef[4] = { CVR / 4, CUG, CVG / 2, CUB / 4 };
#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) { \
int i; \
for (i = 0; i < N; i += 8) { \
int off = ((cur_x) + i) * XSTEP; \
uint8x8_t y = vld1_u8(src_y + (cur_x) + i); \
uint8x8_t u = vld1_u8((src_uv) + i); \
uint8x8_t v = vld1_u8((src_uv) + i + 16); \
int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16)); \
int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128)); \
int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128)); \
\
int16x8_t ud = vshlq_n_s16(uu, 1); \
int16x8_t vd = vshlq_n_s16(vv, 1); \
\
int32x4_t vrl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(vv), 1), \
vget_low_s16(vd), cf16, 0); \
int32x4_t vrh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(vv), 1), \
vget_high_s16(vd), cf16, 0); \
int16x8_t vr = vcombine_s16(vrshrn_n_s32(vrl, 16), \
vrshrn_n_s32(vrh, 16)); \
\
int32x4_t vl = vmovl_s16(vget_low_s16(vv)); \
int32x4_t vh = vmovl_s16(vget_high_s16(vv)); \
int32x4_t ugl = vmlal_lane_s16(vl, vget_low_s16(uu), cf16, 1); \
int32x4_t ugh = vmlal_lane_s16(vh, vget_high_s16(uu), cf16, 1); \
int32x4_t gcl = vqdmlal_lane_s16(ugl, vget_low_s16(vv), cf16, 2); \
int32x4_t gch = vqdmlal_lane_s16(ugh, vget_high_s16(vv), cf16, 2); \
int16x8_t gc = vcombine_s16(vrshrn_n_s32(gcl, 16), \
vrshrn_n_s32(gch, 16)); \
\
int32x4_t ubl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(uu), 1), \
vget_low_s16(ud), cf16, 3); \
int32x4_t ubh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(uu), 1), \
vget_high_s16(ud), cf16, 3); \
int16x8_t ub = vcombine_s16(vrshrn_n_s32(ubl, 16), \
vrshrn_n_s32(ubh, 16)); \
\
int32x4_t rl = vaddl_s16(vget_low_s16(yy), vget_low_s16(vr)); \
int32x4_t rh = vaddl_s16(vget_high_s16(yy), vget_high_s16(vr)); \
int32x4_t gl = vsubl_s16(vget_low_s16(yy), vget_low_s16(gc)); \
int32x4_t gh = vsubl_s16(vget_high_s16(yy), vget_high_s16(gc)); \
int32x4_t bl = vaddl_s16(vget_low_s16(yy), vget_low_s16(ub)); \
int32x4_t bh = vaddl_s16(vget_high_s16(yy), vget_high_s16(ub)); \
\
rl = vmulq_lane_s32(rl, cf32, 0); \
rh = vmulq_lane_s32(rh, cf32, 0); \
gl = vmulq_lane_s32(gl, cf32, 0); \
gh = vmulq_lane_s32(gh, cf32, 0); \
bl = vmulq_lane_s32(bl, cf32, 0); \
bh = vmulq_lane_s32(bh, cf32, 0); \
\
y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, 16), \
vrshrn_n_s32(rh, 16))); \
u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, 16), \
vrshrn_n_s32(gh, 16))); \
v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(bl, 16), \
vrshrn_n_s32(bh, 16))); \
STR_ ## FMT(out + off, y, u, v); \
} \
}
static const int16_t kCoeffs[4] = { kYScale, kVToR, kUToG, kVToG };
#define v255 vmov_n_u8(255)
#define STR_Rgb(out, r, g, b) do { \
#define STORE_Rgb(out, r, g, b) do { \
const uint8x8x3_t r_g_b = {{ r, g, b }}; \
vst3_u8(out, r_g_b); \
} while (0)
#define STR_Bgr(out, r, g, b) do { \
#define STORE_Bgr(out, r, g, b) do { \
const uint8x8x3_t b_g_r = {{ b, g, r }}; \
vst3_u8(out, b_g_r); \
} while (0)
#define STR_Rgba(out, r, g, b) do { \
#define STORE_Rgba(out, r, g, b) do { \
const uint8x8x4_t r_g_b_v255 = {{ r, g, b, v255 }}; \
vst4_u8(out, r_g_b_v255); \
} while (0)
#define STR_Bgra(out, r, g, b) do { \
#define STORE_Bgra(out, r, g, b) do { \
const uint8x8x4_t b_g_r_v255 = {{ b, g, r, v255 }}; \
vst4_u8(out, b_g_r_v255); \
} while (0)
#define CONVERT1(FMT, XSTEP, N, src_y, src_uv, rgb, cur_x) { \
#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) { \
int i; \
for (i = 0; i < N; i += 8) { \
const int off = ((cur_x) + i) * XSTEP; \
uint8x8_t y = vld1_u8((src_y) + (cur_x) + i); \
uint8x8_t u = vld1_u8((src_uv) + i); \
uint8x8_t v = vld1_u8((src_uv) + i + 16); \
const int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16)); \
const int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128)); \
const int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128)); \
int32x4_t yl = vmull_lane_s16(vget_low_s16(yy), cf16, 0); \
int32x4_t yh = vmull_lane_s16(vget_high_s16(yy), cf16, 0); \
const int32x4_t rl = vmlal_lane_s16(yl, vget_low_s16(vv), cf16, 1);\
const int32x4_t rh = vmlal_lane_s16(yh, vget_high_s16(vv), cf16, 1);\
int32x4_t gl = vmlsl_lane_s16(yl, vget_low_s16(uu), cf16, 2); \
int32x4_t gh = vmlsl_lane_s16(yh, vget_high_s16(uu), cf16, 2); \
const int32x4_t bl = vmovl_s16(vget_low_s16(uu)); \
const int32x4_t bh = vmovl_s16(vget_high_s16(uu)); \
gl = vmlsl_lane_s16(gl, vget_low_s16(vv), cf16, 3); \
gh = vmlsl_lane_s16(gh, vget_high_s16(vv), cf16, 3); \
yl = vmlaq_lane_s32(yl, bl, cf32, 0); \
yh = vmlaq_lane_s32(yh, bh, cf32, 0); \
/* vrshrn_n_s32() already incorporates the rounding constant */ \
y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, YUV_FIX2), \
vrshrn_n_s32(rh, YUV_FIX2))); \
u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, YUV_FIX2), \
vrshrn_n_s32(gh, YUV_FIX2))); \
v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(yl, YUV_FIX2), \
vrshrn_n_s32(yh, YUV_FIX2))); \
STORE_ ## FMT(out + off, y, u, v); \
} \
}
#define CONVERT1(FUNC, XSTEP, N, src_y, src_uv, rgb, cur_x) { \
int i; \
for (i = 0; i < N; i++) { \
int off = ((cur_x) + i) * XSTEP; \
int y = src_y[(cur_x) + i]; \
int u = (src_uv)[i]; \
int v = (src_uv)[i + 16]; \
VP8YuvTo ## FMT(y, u, v, rgb + off); \
const int off = ((cur_x) + i) * XSTEP; \
const int y = src_y[(cur_x) + i]; \
const int u = (src_uv)[i]; \
const int v = (src_uv)[i + 16]; \
FUNC(y, u, v, rgb + off); \
} \
}
@ -195,11 +167,11 @@ static const int16_t coef[4] = { CVR / 4, CUG, CVG / 2, CUB / 4 };
} \
}
#define CONVERT2RGB_1(FMT, XSTEP, top_y, bottom_y, uv, \
#define CONVERT2RGB_1(FUNC, XSTEP, top_y, bottom_y, uv, \
top_dst, bottom_dst, cur_x, len) { \
CONVERT1(FMT, XSTEP, len, top_y, uv, top_dst, cur_x); \
CONVERT1(FUNC, XSTEP, len, top_y, uv, top_dst, cur_x); \
if (bottom_y != NULL) { \
CONVERT1(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \
CONVERT1(FUNC, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \
} \
}
@ -221,8 +193,8 @@ static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y, \
const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1; \
const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1; \
\
const int16x4_t cf16 = vld1_s16(coef); \
const int32x2_t cf32 = vmov_n_s32(CY); \
const int16x4_t cf16 = vld1_s16(kCoeffs); \
const int32x2_t cf32 = vmov_n_s32(kUToB); \
const uint8x8_t u16 = vmov_n_u8(16); \
const uint8x8_t u128 = vmov_n_u8(128); \
\
@ -252,7 +224,7 @@ static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y, \
\
UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv); \
UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16); \
CONVERT2RGB_1(FMT, XSTEP, top_y, bottom_y, r_uv, \
CONVERT2RGB_1(VP8YuvTo ## FMT, XSTEP, top_y, bottom_y, r_uv, \
top_dst, bottom_dst, last_pos, len - last_pos); \
}