void WebPRescalerInit(WebPRescaler* const wrk, int src_width, int src_height,
uint8_t* const dst,
int dst_width, int dst_height, int dst_stride,
int num_channels, rescaler_t* const work) {
const int x_add = src_width, x_sub = dst_width;
const int y_add = src_height, y_sub = dst_height;
wrk->x_expand = (src_width < dst_width);
wrk->y_expand = (src_height < dst_height);
wrk->src_width = src_width;
wrk->src_height = src_height;
wrk->dst_width = dst_width;
wrk->dst_height = dst_height;
wrk->src_y = 0;
wrk->dst_y = 0;
wrk->dst = dst;
wrk->dst_stride = dst_stride;
wrk->num_channels = num_channels;
// for 'x_expand', we use bilinear interpolation
wrk->x_add = wrk->x_expand ? (x_sub - 1) : x_add;
wrk->x_sub = wrk->x_expand ? (x_add - 1) : x_sub;
if (!wrk->x_expand) { // fx_scale is not used otherwise
wrk->fx_scale = WEBP_RESCALER_FRAC(1, wrk->x_sub);
}
// vertical scaling parameters
wrk->y_add = wrk->y_expand ? y_add - 1 : y_add;
wrk->y_sub = wrk->y_expand ? y_sub - 1 : y_sub;
wrk->y_accum = wrk->y_expand ? wrk->y_sub : wrk->y_add;
if (!wrk->y_expand) {
// this is WEBP_RESCALER_FRAC(dst_height, x_add * y_add) without the cast.
const uint64_t ratio =
(uint64_t)dst_height * WEBP_RESCALER_ONE / (wrk->x_add * wrk->y_add);
if (ratio != (uint32_t)ratio) {
// We can't represent the ratio with the current fixed-point precision.
// => We special-case fxy_scale = 0, in WebPRescalerExportRow().
wrk->fxy_scale = 0;
} else {
wrk->fxy_scale = (uint32_t)ratio;
}
wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->y_sub);
} else {
wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->x_add);
// wrk->fxy_scale is unused here.
}
wrk->irow = work;
wrk->frow = work + num_channels * dst_width;
memset(work, 0, 2 * dst_width * num_channels * sizeof(*work));
WebPRescalerDspInit();
}
void WebPRescalerExportRowExpandC(WebPRescaler* const wrk) {
int x_out;
uint8_t* const dst = wrk->dst;
rescaler_t* const irow = wrk->irow;
const int x_out_max = wrk->dst_width * wrk->num_channels;
const rescaler_t* const 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) {
for (x_out = 0; 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;
}
} else {
const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub);
const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B);
for (x_out = 0; x_out < x_out_max; ++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 RescalerExportRowExpandSSE2(WebPRescaler* const wrk) {
int x_out;
uint8_t* const dst = wrk->dst;
rescaler_t* const irow = wrk->irow;
const int x_out_max = wrk->dst_width * wrk->num_channels;
const rescaler_t* const frow = wrk->frow;
const __m128i mult = _mm_set_epi32(0, wrk->fy_scale, 0, wrk->fy_scale);
assert(!WebPRescalerOutputDone(wrk));
assert(wrk->y_accum <= 0 && wrk->y_sub + wrk->y_accum >= 0);
assert(wrk->y_expand);
if (wrk->y_accum == 0) {
for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) {
__m128i A0, A1, A2, A3;
LoadDispatchAndMult(frow + x_out, NULL, &A0, &A1, &A2, &A3);
ProcessRow(&A0, &A1, &A2, &A3, &mult, dst + x_out);
}
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;
}
} else {
const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub);
const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B);
const __m128i mA = _mm_set_epi32(0, A, 0, A);
const __m128i mB = _mm_set_epi32(0, B, 0, B);
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(frow + x_out, &mA, &A0, &A1, &A2, &A3);
LoadDispatchAndMult(irow + x_out, &mB, &B0, &B1, &B2, &B3);
{
const __m128i C0 = _mm_add_epi64(A0, B0);
const __m128i C1 = _mm_add_epi64(A1, B1);
const __m128i C2 = _mm_add_epi64(A2, B2);
const __m128i C3 = _mm_add_epi64(A3, B3);
const __m128i D0 = _mm_add_epi64(C0, rounder);
const __m128i D1 = _mm_add_epi64(C1, rounder);
const __m128i D2 = _mm_add_epi64(C2, rounder);
const __m128i D3 = _mm_add_epi64(C3, rounder);
const __m128i E0 = _mm_srli_epi64(D0, WEBP_RESCALER_RFIX);
const __m128i E1 = _mm_srli_epi64(D1, WEBP_RESCALER_RFIX);
const __m128i E2 = _mm_srli_epi64(D2, WEBP_RESCALER_RFIX);
const __m128i E3 = _mm_srli_epi64(D3, WEBP_RESCALER_RFIX);
ProcessRow(&E0, &E1, &E2, &E3, &mult, dst + x_out);
}
}
for (; x_out < x_out_max; ++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;
}
}
}
void WebPRescalerInit(WebPRescaler* const wrk, int src_width, int src_height,
uint8_t* const dst,
int dst_width, int dst_height, int dst_stride,
int num_channels, rescaler_t* const work) {
const int x_add = src_width, x_sub = dst_width;
const int y_add = src_height, y_sub = dst_height;
wrk->x_expand = (src_width < dst_width);
wrk->y_expand = (src_height < dst_height);
wrk->src_width = src_width;
wrk->src_height = src_height;
wrk->dst_width = dst_width;
wrk->dst_height = dst_height;
wrk->src_y = 0;
wrk->dst_y = 0;
wrk->dst = dst;
wrk->dst_stride = dst_stride;
wrk->num_channels = num_channels;
// for 'x_expand', we use bilinear interpolation
wrk->x_add = wrk->x_expand ? (x_sub - 1) : x_add;
wrk->x_sub = wrk->x_expand ? (x_add - 1) : x_sub;
if (!wrk->x_expand) { // fx_scale is not used otherwise
wrk->fx_scale = WEBP_RESCALER_FRAC(1, wrk->x_sub);
}
// vertical scaling parameters
wrk->y_add = wrk->y_expand ? y_add - 1 : y_add;
wrk->y_sub = wrk->y_expand ? y_sub - 1 : y_sub;
wrk->y_accum = wrk->y_expand ? wrk->y_sub : wrk->y_add;
if (!wrk->y_expand) {
// note the very special case where x_add = y_add = 1 cannot be represented.
// We special-case fxy_scale = 0 in this case, in WebPRescalerExportRow().
wrk->fxy_scale = WEBP_RESCALER_FRAC(dst_height, wrk->x_add * wrk->y_add);
wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->y_sub);
} else {
wrk->fy_scale = WEBP_RESCALER_FRAC(1, wrk->x_add);
// wrk->fxy_scale is unused here.
}
wrk->irow = work;
wrk->frow = work + num_channels * dst_width;
memset(work, 0, 2 * dst_width * num_channels * sizeof(*work));
WebPRescalerDspInit();
}
