fix rescaler vertical interpolation

* vertical expansion now uses bilinear interpolation
  * heavily assumes that the alpha plane is decoded in full, not row-by-row
  * split the RescalerExportRow and RescalerImportRow methods into Shrink
    and Expand variants.
  * MIPS implementation of ExportRowExpand is missing.

There's room for extra speed optim and code re-org, but let's keep that for later patches.

addresses https://code.google.com/p/webp/issues/detail?id=254

Change-Id: I8f12b855342bf07dd467fe85e4fde5fd814effdb
This commit is contained in:
Pascal Massimino
2015-09-18 10:45:03 +02:00
committed by James Zern
parent cd82440ec7
commit 5ff0079ece
10 changed files with 553 additions and 388 deletions

View File

@ -119,14 +119,16 @@ static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
//------------------------------------------------------------------------------
static int EmitAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
static int EmitAlphaYUV(const VP8Io* const io, WebPDecParams* const p,
int expected_num_lines_out) {
const uint8_t* alpha = io->a;
const WebPYUVABuffer* const buf = &p->output->u.YUVA;
const int mb_w = io->mb_w;
const int mb_h = io->mb_h;
uint8_t* dst = buf->a + io->mb_y * buf->a_stride;
int j;
(void)expected_num_lines_out;
assert(expected_num_lines_out == mb_h);
if (alpha != NULL) {
for (j = 0; j < mb_h; ++j) {
memcpy(dst, alpha, mb_w * sizeof(*dst));
@ -169,7 +171,8 @@ static int GetAlphaSourceRow(const VP8Io* const io,
return start_y;
}
static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p,
int expected_num_lines_out) {
const uint8_t* alpha = io->a;
if (alpha != NULL) {
const int mb_w = io->mb_w;
@ -183,7 +186,8 @@ static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
uint8_t* const dst = base_rgba + (alpha_first ? 0 : 3);
const int has_alpha = WebPDispatchAlpha(alpha, io->width, mb_w,
num_rows, dst, buf->stride);
(void)expected_num_lines_out;
assert(expected_num_lines_out == num_rows);
// has_alpha is true if there's non-trivial alpha to premultiply with.
if (has_alpha && WebPIsPremultipliedMode(colorspace)) {
WebPApplyAlphaMultiply(base_rgba, alpha_first,
@ -193,7 +197,8 @@ static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
return 0;
}
static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p,
int expected_num_lines_out) {
const uint8_t* alpha = io->a;
if (alpha != NULL) {
const int mb_w = io->mb_w;
@ -209,7 +214,6 @@ static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
#endif
uint32_t alpha_mask = 0x0f;
int i, j;
for (j = 0; j < num_rows; ++j) {
for (i = 0; i < mb_w; ++i) {
// Fill in the alpha value (converted to 4 bits).
@ -220,6 +224,8 @@ static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
alpha += io->width;
alpha_dst += buf->stride;
}
(void)expected_num_lines_out;
assert(expected_num_lines_out == num_rows);
if (alpha_mask != 0x0f && WebPIsPremultipliedMode(colorspace)) {
WebPApplyAlphaMultiply4444(base_rgba, mb_w, num_rows, buf->stride);
}
@ -261,12 +267,15 @@ static int EmitRescaledYUV(const VP8Io* const io, WebPDecParams* const p) {
return num_lines_out;
}
static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p) {
static int EmitRescaledAlphaYUV(const VP8Io* const io, WebPDecParams* const p,
int expected_num_lines_out) {
if (io->a != NULL) {
const WebPYUVABuffer* const buf = &p->output->u.YUVA;
uint8_t* dst_y = buf->y + p->last_y * buf->y_stride;
const uint8_t* src_a = buf->a + p->last_y * buf->a_stride;
const int num_lines_out = Rescale(io->a, io->width, io->mb_h, &p->scaler_a);
(void)expected_num_lines_out;
assert(expected_num_lines_out == num_lines_out);
if (num_lines_out > 0) { // unmultiply the Y
WebPMultRows(dst_y, buf->y_stride, src_a, buf->a_stride,
p->scaler_a.dst_width, num_lines_out, 1);
@ -287,7 +296,7 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
const size_t work_size = 2 * out_width; // scratch memory for luma rescaler
const size_t uv_work_size = 2 * uv_out_width; // and for each u/v ones
size_t tmp_size;
int32_t* work;
rescaler_t* work;
tmp_size = (work_size + 2 * uv_work_size) * sizeof(*work);
if (has_alpha) {
@ -297,7 +306,7 @@ static int InitYUVRescaler(const VP8Io* const io, WebPDecParams* const p) {
if (p->memory == NULL) {
return 0; // memory error
}
work = (int32_t*)p->memory;
work = (rescaler_t*)p->memory;
WebPRescalerInit(&p->scaler_y, io->mb_w, io->mb_h,
buf->y, out_width, out_height, buf->y_stride, 1,
work);
@ -326,17 +335,17 @@ static int ExportRGB(WebPDecParams* const p, int y_pos) {
const WebPYUV444Converter convert =
WebPYUV444Converters[p->output->colorspace];
const WebPRGBABuffer* const buf = &p->output->u.RGBA;
uint8_t* dst = buf->rgba + (p->last_y + y_pos) * buf->stride;
uint8_t* dst = buf->rgba + y_pos * buf->stride;
int num_lines_out = 0;
// For RGB rescaling, because of the YUV420, current scan position
// U/V can be +1/-1 line from the Y one. Hence the double test.
while (WebPRescalerHasPendingOutput(&p->scaler_y) &&
WebPRescalerHasPendingOutput(&p->scaler_u)) {
assert(p->last_y + y_pos + num_lines_out < p->output->height);
assert(y_pos + num_lines_out < p->output->height);
assert(p->scaler_u.y_accum == p->scaler_v.y_accum);
WebPRescalerExportRow(&p->scaler_y, 0);
WebPRescalerExportRow(&p->scaler_u, 0);
WebPRescalerExportRow(&p->scaler_v, 0);
WebPRescalerExportRow(&p->scaler_y);
WebPRescalerExportRow(&p->scaler_u);
WebPRescalerExportRow(&p->scaler_v);
convert(p->scaler_y.dst, p->scaler_u.dst, p->scaler_v.dst,
dst, p->scaler_y.dst_width);
dst += buf->stride;
@ -354,24 +363,26 @@ static int EmitRescaledRGB(const VP8Io* const io, WebPDecParams* const p) {
const int y_lines_in =
WebPRescalerImport(&p->scaler_y, mb_h - j,
io->y + j * io->y_stride, io->y_stride);
const int u_lines_in =
WebPRescalerImport(&p->scaler_u, uv_mb_h - uv_j,
io->u + uv_j * io->uv_stride, io->uv_stride);
const int v_lines_in =
WebPRescalerImport(&p->scaler_v, uv_mb_h - uv_j,
io->v + uv_j * io->uv_stride, io->uv_stride);
(void)v_lines_in; // remove a gcc warning
assert(u_lines_in == v_lines_in);
j += y_lines_in;
uv_j += u_lines_in;
num_lines_out += ExportRGB(p, num_lines_out);
if (WebPRescaleNeededLines(&p->scaler_u, uv_mb_h - uv_j)) {
const int u_lines_in =
WebPRescalerImport(&p->scaler_u, uv_mb_h - uv_j,
io->u + uv_j * io->uv_stride, io->uv_stride);
const int v_lines_in =
WebPRescalerImport(&p->scaler_v, uv_mb_h - uv_j,
io->v + uv_j * io->uv_stride, io->uv_stride);
(void)v_lines_in; // remove a gcc warning
assert(u_lines_in == v_lines_in);
uv_j += u_lines_in;
}
num_lines_out += ExportRGB(p, p->last_y + num_lines_out);
}
return num_lines_out;
}
static int ExportAlpha(WebPDecParams* const p, int y_pos) {
static int ExportAlpha(WebPDecParams* const p, int y_pos, int max_lines_out) {
const WebPRGBABuffer* const buf = &p->output->u.RGBA;
uint8_t* const base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
uint8_t* const base_rgba = buf->rgba + y_pos * buf->stride;
const WEBP_CSP_MODE colorspace = p->output->colorspace;
const int alpha_first =
(colorspace == MODE_ARGB || colorspace == MODE_Argb);
@ -381,9 +392,10 @@ static int ExportAlpha(WebPDecParams* const p, int y_pos) {
uint32_t non_opaque = 0;
const int width = p->scaler_a.dst_width;
while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
assert(p->last_y + y_pos + num_lines_out < p->output->height);
WebPRescalerExportRow(&p->scaler_a, 0);
while (WebPRescalerHasPendingOutput(&p->scaler_a) &&
num_lines_out < max_lines_out) {
assert(y_pos + num_lines_out < p->output->height);
WebPRescalerExportRow(&p->scaler_a);
non_opaque |= WebPDispatchAlpha(p->scaler_a.dst, 0, width, 1, dst, 0);
dst += buf->stride;
++num_lines_out;
@ -395,9 +407,10 @@ static int ExportAlpha(WebPDecParams* const p, int y_pos) {
return num_lines_out;
}
static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos,
int max_lines_out) {
const WebPRGBABuffer* const buf = &p->output->u.RGBA;
uint8_t* const base_rgba = buf->rgba + (p->last_y + y_pos) * buf->stride;
uint8_t* const base_rgba = buf->rgba + y_pos * buf->stride;
#ifdef WEBP_SWAP_16BIT_CSP
uint8_t* alpha_dst = base_rgba;
#else
@ -409,10 +422,11 @@ static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
const int is_premult_alpha = WebPIsPremultipliedMode(colorspace);
uint32_t alpha_mask = 0x0f;
while (WebPRescalerHasPendingOutput(&p->scaler_a)) {
while (WebPRescalerHasPendingOutput(&p->scaler_a) &&
num_lines_out < max_lines_out) {
int i;
assert(p->last_y + y_pos + num_lines_out < p->output->height);
WebPRescalerExportRow(&p->scaler_a, 0);
assert(y_pos + num_lines_out < p->output->height);
WebPRescalerExportRow(&p->scaler_a);
for (i = 0; i < width; ++i) {
// Fill in the alpha value (converted to 4 bits).
const uint32_t alpha_value = p->scaler_a.dst[i] >> 4;
@ -428,15 +442,17 @@ static int ExportAlphaRGBA4444(WebPDecParams* const p, int y_pos) {
return num_lines_out;
}
static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
static int EmitRescaledAlphaRGB(const VP8Io* const io, WebPDecParams* const p,
int expected_num_out_lines) {
if (io->a != NULL) {
WebPRescaler* const scaler = &p->scaler_a;
int j = 0;
int pos = 0;
while (j < io->mb_h) {
j += WebPRescalerImport(scaler, io->mb_h - j,
io->a + j * io->width, io->width);
pos += p->emit_alpha_row(p, pos);
int lines_left = expected_num_out_lines;
const int y_end = p->last_y + lines_left;
while (lines_left > 0) {
const int row_offset = scaler->src_y - io->mb_y;
WebPRescalerImport(scaler, io->mb_h + io->mb_y - scaler->src_y,
io->a + row_offset * io->width, io->width);
lines_left -= p->emit_alpha_row(p, y_end - lines_left, lines_left);
}
}
return 0;
@ -449,7 +465,7 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
const int uv_in_width = (io->mb_w + 1) >> 1;
const int uv_in_height = (io->mb_h + 1) >> 1;
const size_t work_size = 2 * out_width; // scratch memory for one rescaler
int32_t* work; // rescalers work area
rescaler_t* work; // rescalers work area
uint8_t* tmp; // tmp storage for scaled YUV444 samples before RGB conversion
size_t tmp_size1, tmp_size2, total_size;
@ -464,7 +480,7 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
if (p->memory == NULL) {
return 0; // memory error
}
work = (int32_t*)p->memory;
work = (rescaler_t*)p->memory;
tmp = (uint8_t*)(work + tmp_size1);
WebPRescalerInit(&p->scaler_y, io->mb_w, io->mb_h,
tmp + 0 * out_width, out_width, out_height, 0, 1,
@ -572,7 +588,7 @@ static int CustomPut(const VP8Io* io) {
}
num_lines_out = p->emit(io, p);
if (p->emit_alpha != NULL) {
p->emit_alpha(io, p);
p->emit_alpha(io, p, num_lines_out);
}
p->last_y += num_lines_out;
return 1;