dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-11-20 04:18:26 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

revamp the point-sampling functions by processing a full plane

Browse Source 
-nofancy is slower than fancy upsampler, because the latter has SSE2 optim.

Change-Id: Ibf22e5a8ea1de86a54248d4a4ecc63d514c01b88
This commit is contained in:
Pascal Massimino 2014-05-20 09:43:02 -07:00
parent 2b5cb32612
commit a2f8b28905
4 changed files with 115 additions and 121 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
src/dec/io.c
View File

@ -45,27 +45,12 @@ static int EmitYUV(const VP8Io* const io, WebPDecParams* const p) {
// Point-sampling U/V sampler. // Point-sampling U/V sampler.
static int EmitSampledRGB(const VP8Io* const io, WebPDecParams* const p) { static int EmitSampledRGB(const VP8Io* const io, WebPDecParams* const p) {
WebPDecBuffer* output = p->output; WebPDecBuffer* const output = p->output;
const WebPRGBABuffer* const buf = &output->u.RGBA; WebPRGBABuffer* const buf = &output->u.RGBA;
uint8_t* dst = buf->rgba + io->mb_y * buf->stride; uint8_t* const dst = buf->rgba + io->mb_y * buf->stride;
const uint8_t* y_src = io->y; WebPSamplers[output->colorspace](io->y, io->y_stride,
const uint8_t* u_src = io->u; io->u, io->v, io->uv_stride,
const uint8_t* v_src = io->v; dst, buf->stride, io->mb_w, io->mb_h);
const WebPSampleLinePairFunc sample = WebPSamplers[output->colorspace];
const int mb_w = io->mb_w;
const int last = io->mb_h - 1;
int j;
for (j = 0; j < last; j += 2) {
sample(y_src, y_src + io->y_stride, u_src, v_src,
dst, dst + buf->stride, mb_w);
y_src += 2 * io->y_stride;
u_src += io->uv_stride;
v_src += io->uv_stride;
dst += 2 * buf->stride;
}
if (j == last) { // Just do the last line twice
sample(y_src, y_src, u_src, v_src, dst, dst, mb_w);
}
return io->mb_h; return io->mb_h;
} }
@ -637,4 +622,3 @@ void WebPInitCustomIo(WebPDecParams* const params, VP8Io* const io) {
} }
//------------------------------------------------------------------------------ //------------------------------------------------------------------------------

20
src/dsp/dsp.h
View File

@ -195,13 +195,23 @@ void WebPInitUpsamplersNEON(void);
#endif // FANCY_UPSAMPLING #endif // FANCY_UPSAMPLING
// Point-sampling methods. // Point-sampling methods.
typedef void (*WebPSampleLinePairFunc)( typedef void (*WebPSamplePlaneFunc)(const uint8_t* y, int y_stride,
const uint8_t* top_y, const uint8_t* bottom_y, const uint8_t* u, const uint8_t* v,
const uint8_t* u, const uint8_t* v, int uv_stride,
uint8_t* top_dst, uint8_t* bottom_dst, int len); uint8_t* dst, int dst_stride,
int width, int height);
typedef void (*WebPSamplerRowFunc)(const uint8_t* y,
const uint8_t* u, const uint8_t* v,
uint8_t* dst, int len);
// Generic function to apply 'WebPSamplerRowFunc' to the whole plane:
void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
const uint8_t* u, const uint8_t* v, int uv_stride,
uint8_t* dst, int dst_stride,
int width, int height, WebPSamplerRowFunc func);
// Sampling functions to convert YUV to RGB(A) modes // Sampling functions to convert YUV to RGB(A) modes
extern WebPSampleLinePairFunc WebPSamplers[/* MODE_LAST */]; extern WebPSamplePlaneFunc WebPSamplers[/* MODE_LAST */];
// Initializes MIPS version of the samplers. // Initializes MIPS version of the samplers.
void WebPInitSamplersMIPS32(void); void WebPInitSamplersMIPS32(void);

94
src/dsp/upsampling.c
View File

@ -109,40 +109,59 @@ UPSAMPLE_FUNC(UpsampleRgb565LinePair, VP8YuvToRgb565, 2)
//------------------------------------------------------------------------------ //------------------------------------------------------------------------------
// simple point-sampling // simple point-sampling
WebPSamplePlaneFunc WebPSamplers[MODE_LAST];
WebPSampleLinePairFunc WebPSamplers[MODE_LAST]; void WebPSamplerProcessPlane(const uint8_t* y, int y_stride,
const uint8_t* u, const uint8_t* v, int uv_stride,
uint8_t* dst, int dst_stride,
int width, int height, WebPSamplerRowFunc func) {
int j;
for (j = 0; j < height; ++j) {
func(y, u, v, dst, width);
y += y_stride;
if (j & 1) {
u += uv_stride;
v += uv_stride;
}
dst += dst_stride;
}
}
// TODO(skal): maybe some of these per-row functions should be in yuv.h?
#define SAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ #define SAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \
static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ static void FUNC_NAME##Row(const uint8_t* y, \
const uint8_t* u, const uint8_t* v, \ const uint8_t* u, const uint8_t* v, \
uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ uint8_t* dst, int len) { \
int i; \ const uint8_t* const end = dst + (len & ~1) * XSTEP; \
for (i = 0; i < len - 1; i += 2) { \ while (dst != end) { \
FUNC(top_y[0], u[0], v[0], top_dst); \ FUNC(y[0], u[0], v[0], dst); \
FUNC(top_y[1], u[0], v[0], top_dst + XSTEP); \ FUNC(y[1], u[0], v[0], dst + XSTEP); \
FUNC(bottom_y[0], u[0], v[0], bottom_dst); \ y += 2; \
FUNC(bottom_y[1], u[0], v[0], bottom_dst + XSTEP); \ ++u; \
top_y += 2; \ ++v; \
bottom_y += 2; \ dst += 2 * XSTEP; \
u++; \
v++; \
top_dst += 2 * XSTEP; \
bottom_dst += 2 * XSTEP; \
} \ } \
if (i == len - 1) { /* last one */ \ if (len & 1) { \
FUNC(top_y[0], u[0], v[0], top_dst); \ FUNC(y[0], u[0], v[0], dst); \
FUNC(bottom_y[0], u[0], v[0], bottom_dst); \
} \ } \
} \
static void FUNC_NAME(const uint8_t* y, int y_stride, \
const uint8_t* u, const uint8_t* v, int uv_stride, \
uint8_t* dst, int dst_stride, \
int width, int height) { \
WebPSamplerProcessPlane(y, y_stride, u, v, uv_stride, \
dst, dst_stride, width, height, \
FUNC_NAME##Row); \
} }
// All variants implemented. // All variants implemented.
SAMPLE_FUNC(SampleRgbLinePair, VP8YuvToRgb, 3) SAMPLE_FUNC(SampleRgbPlane, VP8YuvToRgb, 3)
SAMPLE_FUNC(SampleBgrLinePair, VP8YuvToBgr, 3) SAMPLE_FUNC(SampleBgrPlane, VP8YuvToBgr, 3)
SAMPLE_FUNC(SampleRgbaLinePair, VP8YuvToRgba, 4) SAMPLE_FUNC(SampleRgbaPlane, VP8YuvToRgba, 4)
SAMPLE_FUNC(SampleBgraLinePair, VP8YuvToBgra, 4) SAMPLE_FUNC(SampleBgraPlane, VP8YuvToBgra, 4)
SAMPLE_FUNC(SampleArgbLinePair, VP8YuvToArgb, 4) SAMPLE_FUNC(SampleArgbPlane, VP8YuvToArgb, 4)
SAMPLE_FUNC(SampleRgba4444LinePair, VP8YuvToRgba4444, 2) SAMPLE_FUNC(SampleRgba4444Plane, VP8YuvToRgba4444, 2)
SAMPLE_FUNC(SampleRgb565LinePair, VP8YuvToRgb565, 2) SAMPLE_FUNC(SampleRgb565Plane, VP8YuvToRgb565, 2)
#undef SAMPLE_FUNC #undef SAMPLE_FUNC
@ -341,17 +360,17 @@ void WebPInitUpsamplers(void) {
} }
void WebPInitSamplers(void) { void WebPInitSamplers(void) {
WebPSamplers[MODE_RGB] = SampleRgbLinePair; WebPSamplers[MODE_RGB] = SampleRgbPlane;
WebPSamplers[MODE_RGBA] = SampleRgbaLinePair; WebPSamplers[MODE_RGBA] = SampleRgbaPlane;
WebPSamplers[MODE_BGR] = SampleBgrLinePair; WebPSamplers[MODE_BGR] = SampleBgrPlane;
WebPSamplers[MODE_BGRA] = SampleBgraLinePair; WebPSamplers[MODE_BGRA] = SampleBgraPlane;
WebPSamplers[MODE_ARGB] = SampleArgbLinePair; WebPSamplers[MODE_ARGB] = SampleArgbPlane;
WebPSamplers[MODE_RGBA_4444] = SampleRgba4444LinePair; WebPSamplers[MODE_RGBA_4444] = SampleRgba4444Plane;
WebPSamplers[MODE_RGB_565] = SampleRgb565LinePair; WebPSamplers[MODE_RGB_565] = SampleRgb565Plane;
WebPSamplers[MODE_rgbA] = SampleRgbaLinePair; WebPSamplers[MODE_rgbA] = SampleRgbaPlane;
WebPSamplers[MODE_bgrA] = SampleBgraLinePair; WebPSamplers[MODE_bgrA] = SampleBgraPlane;
WebPSamplers[MODE_Argb] = SampleArgbLinePair; WebPSamplers[MODE_Argb] = SampleArgbPlane;
WebPSamplers[MODE_rgbA_4444] = SampleRgba4444LinePair; WebPSamplers[MODE_rgbA_4444] = SampleRgba4444Plane;
// If defined, use CPUInfo() to overwrite some pointers with faster versions. // If defined, use CPUInfo() to overwrite some pointers with faster versions.
if (VP8GetCPUInfo != NULL) { if (VP8GetCPUInfo != NULL) {
@ -387,4 +406,3 @@ void WebPInitPremultiply(void) {
} }
#endif // FANCY_UPSAMPLING #endif // FANCY_UPSAMPLING
} }

94
src/dsp/upsampling_mips32.c
View File

@ -22,9 +22,9 @@
// simple point-sampling // simple point-sampling
#define SAMPLE_FUNC_MIPS(FUNC_NAME, XSTEP, R, G, B, A) \ #define SAMPLE_FUNC_MIPS(FUNC_NAME, XSTEP, R, G, B, A) \
static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ static void FUNC_NAME##Row(const uint8_t* y, \
const uint8_t* u, const uint8_t* v, \ const uint8_t* u, const uint8_t* v, \
uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ uint8_t* dst, int len) { \
int i, r, g, b; \ int i, r, g, b; \
int temp0, temp1, temp2, temp3, temp4; \ int temp0, temp1, temp2, temp3, temp4; \
for (i = 0; i < (len >> 1); i++) { \ for (i = 0; i < (len >> 1); i++) { \
@ -32,7 +32,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \
temp3 = kVToG * v[0]; \ temp3 = kVToG * v[0]; \
temp2 = kUToG * u[0]; \ temp2 = kUToG * u[0]; \
temp4 = kUToB * u[0]; \ temp4 = kUToB * u[0]; \
temp0 = kYScale * top_y[0]; \ temp0 = kYScale * y[0]; \
temp1 += kRCst; \ temp1 += kRCst; \
temp3 -= kGCst; \ temp3 -= kGCst; \
temp2 += temp3; \ temp2 += temp3; \
@ -40,47 +40,29 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \
r = VP8Clip8(temp0 + temp1); \ r = VP8Clip8(temp0 + temp1); \
g = VP8Clip8(temp0 - temp2); \ g = VP8Clip8(temp0 - temp2); \
b = VP8Clip8(temp0 + temp4); \ b = VP8Clip8(temp0 + temp4); \
temp0 = kYScale * top_y[1]; \ temp0 = kYScale * y[1]; \
top_dst[R] = r; \ dst[R] = r; \
top_dst[G] = g; \ dst[G] = g; \
top_dst[B] = b; \ dst[B] = b; \
if (A) top_dst[A] = 0xff; \ if (A) dst[A] = 0xff; \
r = VP8Clip8(temp0 + temp1); \ r = VP8Clip8(temp0 + temp1); \
g = VP8Clip8(temp0 - temp2); \ g = VP8Clip8(temp0 - temp2); \
b = VP8Clip8(temp0 + temp4); \ b = VP8Clip8(temp0 + temp4); \
temp0 = kYScale * bottom_y[0]; \ dst[R + XSTEP] = r; \
top_dst[R + XSTEP] = r; \ dst[G + XSTEP] = g; \
top_dst[G + XSTEP] = g; \ dst[B + XSTEP] = b; \
top_dst[B + XSTEP] = b; \ if (A) dst[A + XSTEP] = 0xff; \
if (A) top_dst[A + XSTEP] = 0xff; \ y += 2; \
r = VP8Clip8(temp0 + temp1); \ ++u; \
g = VP8Clip8(temp0 - temp2); \ ++v; \
b = VP8Clip8(temp0 + temp4); \ dst += 2 * XSTEP; \
temp0 = kYScale * bottom_y[1]; \
bottom_dst[R] = r; \
bottom_dst[G] = g; \
bottom_dst[B] = b; \
if (A) bottom_dst[A] = 0xff; \
r = VP8Clip8(temp0 + temp1); \
g = VP8Clip8(temp0 - temp2); \
b = VP8Clip8(temp0 + temp4); \
bottom_dst[R + XSTEP] = r; \
bottom_dst[G + XSTEP] = g; \
bottom_dst[B + XSTEP] = b; \
if (A) bottom_dst[A + XSTEP] = 0xff; \
top_y += 2; \
bottom_y += 2; \
u++; \
v++; \
top_dst += 2 * XSTEP; \
bottom_dst += 2 * XSTEP; \
} \ } \
if (len & 1) { \ if (len & 1) { \
temp1 = kVToR * v[0]; \ temp1 = kVToR * v[0]; \
temp3 = kVToG * v[0]; \ temp3 = kVToG * v[0]; \
temp2 = kUToG * u[0]; \ temp2 = kUToG * u[0]; \
temp4 = kUToB * u[0]; \ temp4 = kUToB * u[0]; \
temp0 = kYScale * top_y[0]; \ temp0 = kYScale * y[0]; \
temp1 += kRCst; \ temp1 += kRCst; \
temp3 -= kGCst; \ temp3 -= kGCst; \
temp2 += temp3; \ temp2 += temp3; \
@ -88,25 +70,25 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \
r = VP8Clip8(temp0 + temp1); \ r = VP8Clip8(temp0 + temp1); \
g = VP8Clip8(temp0 - temp2); \ g = VP8Clip8(temp0 - temp2); \
b = VP8Clip8(temp0 + temp4); \ b = VP8Clip8(temp0 + temp4); \
temp0 = kYScale * bottom_y[0]; \ dst[R] = r; \
top_dst[R] = r; \ dst[G] = g; \
top_dst[G] = g; \ dst[B] = b; \
top_dst[B] = b; \ if (A) dst[A] = 0xff; \
if (A) top_dst[A] = 0xff; \
r = VP8Clip8(temp0 + temp1); \
g = VP8Clip8(temp0 - temp2); \
b = VP8Clip8(temp0 + temp4); \
bottom_dst[R] = r; \
bottom_dst[G] = g; \
bottom_dst[B] = b; \
if (A) bottom_dst[A] = 0xff; \
} \ } \
} \
static void FUNC_NAME(const uint8_t* y, int y_stride, \
const uint8_t* u, const uint8_t* v, int uv_stride, \
uint8_t* dst, int dst_stride, \
int width, int height) { \
WebPSamplerProcessPlane(y, y_stride, u, v, uv_stride, \
dst, dst_stride, width, height, \
FUNC_NAME##Row); \
} }
SAMPLE_FUNC_MIPS(SampleRgbLinePair, 3, 0, 1, 2, 0) SAMPLE_FUNC_MIPS(SampleRgbPlane, 3, 0, 1, 2, 0)
SAMPLE_FUNC_MIPS(SampleRgbaLinePair, 4, 0, 1, 2, 3) SAMPLE_FUNC_MIPS(SampleRgbaPlane, 4, 0, 1, 2, 3)
SAMPLE_FUNC_MIPS(SampleBgrLinePair, 3, 2, 1, 0, 0) SAMPLE_FUNC_MIPS(SampleBgrPlane, 3, 2, 1, 0, 0)
SAMPLE_FUNC_MIPS(SampleBgraLinePair, 4, 2, 1, 0, 3) SAMPLE_FUNC_MIPS(SampleBgraPlane, 4, 2, 1, 0, 3)
#endif // WEBP_USE_MIPS32 #endif // WEBP_USE_MIPS32
@ -114,9 +96,9 @@ SAMPLE_FUNC_MIPS(SampleBgraLinePair, 4, 2, 1, 0, 3)
void WebPInitSamplersMIPS32(void) { void WebPInitSamplersMIPS32(void) {
#if defined(WEBP_USE_MIPS32) #if defined(WEBP_USE_MIPS32)
WebPSamplers[MODE_RGB] = SampleRgbLinePair; WebPSamplers[MODE_RGB] = SampleRgbPlane;
WebPSamplers[MODE_RGBA] = SampleRgbaLinePair; WebPSamplers[MODE_RGBA] = SampleRgbaPlane;
WebPSamplers[MODE_BGR] = SampleBgrLinePair; WebPSamplers[MODE_BGR] = SampleBgrPlane;
WebPSamplers[MODE_BGRA] = SampleBgraLinePair; WebPSamplers[MODE_BGRA] = SampleBgraPlane;
#endif // WEBP_USE_MIPS32 #endif // WEBP_USE_MIPS32
} }