dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2025-07-16 22:09:57 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

Rescaler: fix rounding error

Browse Source 
We saturate the result to [0..255]
It's the easiest and safest, given the wide variety of scaling
range we cover: we're not using floats, so precision is always
an issue at one end or the other of the scaling spectrum.

we also use:
  round(a - floor(b))
instead of:
  floor(a - round(b))
to handle difficult cases (ratio ~= .99, e.g.)

MIPS code is still disabled (and wrong)

Change-Id: I18d3f5ddc4c524879c257b928329b1c648fa7fb5
This commit is contained in:
Pascal Massimino
2019-03-28 19:44:39 +01:00
parent aa65f89a8f
commit ab2dc8939f
5 changed files with 43 additions and 97 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
src/dsp/rescaler_sse2.c
View File

@ -225,35 +225,6 @@ static WEBP_INLINE void ProcessRow_SSE2(const __m128i* const A0,
_mm_storel_epi64((__m128i*)dst, G);
}
static WEBP_INLINE void ProcessRow_Floor_SSE2(const __m128i* const A0,
const __m128i* const A1,
const __m128i* const A2,
const __m128i* const A3,
const __m128i* const mult,
uint8_t* const dst) {
const __m128i mask = _mm_set_epi32(0xffffffffu, 0, 0xffffffffu, 0);
const __m128i B0 = _mm_mul_epu32(*A0, *mult);
const __m128i B1 = _mm_mul_epu32(*A1, *mult);
const __m128i B2 = _mm_mul_epu32(*A2, *mult);
const __m128i B3 = _mm_mul_epu32(*A3, *mult);
const __m128i D0 = _mm_srli_epi64(B0, WEBP_RESCALER_RFIX);
const __m128i D1 = _mm_srli_epi64(B1, WEBP_RESCALER_RFIX);
#if (WEBP_RESCALER_RFIX < 32)
const __m128i D2 =
_mm_and_si128(_mm_slli_epi64(B2, 32 - WEBP_RESCALER_RFIX), mask);
const __m128i D3 =
_mm_and_si128(_mm_slli_epi64(B3, 32 - WEBP_RESCALER_RFIX), mask);
#else
const __m128i D2 = _mm_and_si128(B2, mask);
const __m128i D3 = _mm_and_si128(B3, mask);
#endif
const __m128i E0 = _mm_or_si128(D0, D2);
const __m128i E1 = _mm_or_si128(D1, D3);
const __m128i F = _mm_packs_epi32(E0, E1);
const __m128i G = _mm_packus_epi16(F, F);
_mm_storel_epi64((__m128i*)dst, G);
}
static void RescalerExportRowExpand_SSE2(WebPRescaler* const wrk) {
int x_out;
uint8_t* const dst = wrk->dst;
@ -274,8 +245,7 @@ static void RescalerExportRowExpand_SSE2(WebPRescaler* const wrk) {
for (; x_out < x_out_max; ++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;
dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
}
} else {
const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub);
@ -308,8 +278,7 @@ static void RescalerExportRowExpand_SSE2(WebPRescaler* const wrk) {
+ (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;
dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
}
}
}
@ -328,20 +297,15 @@ static void RescalerExportRowShrink_SSE2(WebPRescaler* const wrk) {
const int scale_xy = wrk->fxy_scale;
const __m128i mult_xy = _mm_set_epi32(0, scale_xy, 0, scale_xy);
const __m128i mult_y = _mm_set_epi32(0, yscale, 0, yscale);
const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER);
for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) {
__m128i A0, A1, A2, A3, B0, B1, B2, B3;
LoadDispatchAndMult_SSE2(irow + x_out, NULL, &A0, &A1, &A2, &A3);
LoadDispatchAndMult_SSE2(frow + x_out, &mult_y, &B0, &B1, &B2, &B3);
{
const __m128i C0 = _mm_add_epi64(B0, rounder);
const __m128i C1 = _mm_add_epi64(B1, rounder);
const __m128i C2 = _mm_add_epi64(B2, rounder);
const __m128i C3 = _mm_add_epi64(B3, rounder);
const __m128i D0 = _mm_srli_epi64(C0, WEBP_RESCALER_RFIX); // = frac
const __m128i D1 = _mm_srli_epi64(C1, WEBP_RESCALER_RFIX);
const __m128i D2 = _mm_srli_epi64(C2, WEBP_RESCALER_RFIX);
const __m128i D3 = _mm_srli_epi64(C3, WEBP_RESCALER_RFIX);
const __m128i D0 = _mm_srli_epi64(B0, WEBP_RESCALER_RFIX); // = frac
const __m128i D1 = _mm_srli_epi64(B1, WEBP_RESCALER_RFIX);
const __m128i D2 = _mm_srli_epi64(B2, WEBP_RESCALER_RFIX);
const __m128i D3 = _mm_srli_epi64(B3, WEBP_RESCALER_RFIX);
const __m128i E0 = _mm_sub_epi64(A0, D0); // irow[x] - frac
const __m128i E1 = _mm_sub_epi64(A1, D1);
const __m128i E2 = _mm_sub_epi64(A2, D2);
@ -352,14 +316,13 @@ static void RescalerExportRowShrink_SSE2(WebPRescaler* const wrk) {
const __m128i G1 = _mm_or_si128(D1, F3);
_mm_storeu_si128((__m128i*)(irow + x_out + 0), G0);
_mm_storeu_si128((__m128i*)(irow + x_out + 4), G1);
ProcessRow_Floor_SSE2(&E0, &E1, &E2, &E3, &mult_xy, dst + x_out);
ProcessRow_SSE2(&E0, &E1, &E2, &E3, &mult_xy, dst + x_out);
}
}
for (; x_out < x_out_max; ++x_out) {
const uint32_t frac = (int)MULT_FIX(frow[x_out], yscale);
const int v = (int)MULT_FIX_FLOOR(irow[x_out] - frac, wrk->fxy_scale);
assert(v >= 0 && v <= 255);
dst[x_out] = v;
const uint32_t frac = (int)MULT_FIX_FLOOR(frow[x_out], yscale);
const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale);
dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
irow[x_out] = frac; // new fractional start
}
} else {
@ -375,8 +338,7 @@ static void RescalerExportRowShrink_SSE2(WebPRescaler* const wrk) {
}
for (; x_out < x_out_max; ++x_out) {
const int v = (int)MULT_FIX(irow[x_out], scale);
assert(v >= 0 && v <= 255);
dst[x_out] = v;
dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
irow[x_out] = 0;
}
}