Skip to content

Commit

Permalink
fix upConv > 0 with RFF flag.
Browse files Browse the repository at this point in the history
  • Loading branch information
@chikuzen
chikuzen committed Aug 31, 2016
1 parent 5ed3558 commit df0dde8
Show file tree
Hide file tree
Showing 4 changed files with 57 additions and 82 deletions.
113 changes: 45 additions & 68 deletions src/dgdecode/AVISynthAPI.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -40,7 +40,12 @@
#define VERSION "MPEG2DecPlus 0.1.0"


MPEG2Source::MPEG2Source(const char* d2v, int cpu, int idct, int iPP, int moderate_h, int moderate_v, bool showQ, bool fastMC, const char* _cpu2, int _info, int _upConv, bool _i420, int iCC, IScriptEnvironment* env)
MPEG2Source::MPEG2Source(const char* d2v, int cpu, int idct, int iPP,
int moderate_h, int moderate_v, bool showQ,
bool fastMC, const char* _cpu2, int _info,
int _upConv, bool _i420, int iCC,
IScriptEnvironment* env) :
bufY(nullptr), bufU(nullptr), bufV(nullptr)
{
int status;

Expand Down Expand Up @@ -150,46 +155,32 @@ MPEG2Source::MPEG2Source(const char* d2v, int cpu, int idct, int iPP, int modera
if (cpu2[5]=='x' || cpu2[5] == 'X') { _PP_MODE |= PP_DERING_C; }
}
#endif
if (ovr_idct > 0 )
if (ovr_idct > 0)
{
dprintf("Overiding iDCT With: %d", ovr_idct);
override(ovr_idct);
}

out = (YV12PICT*)malloc(sizeof(YV12PICT));
if (out == NULL) env->ThrowError("MPEG2Source: malloc failure (yv12 pic out)!");

//m_decoder.pp_mode = _PP_MODE;

bufY = bufU = bufV = NULL;
if (m_decoder.chroma_format != 1 || (m_decoder.chroma_format == 1 && _upConv > 0))
{
bufY = (uint8_t*)_aligned_malloc(vi.width*vi.height+2048,32);
bufU = (uint8_t*)_aligned_malloc(m_decoder.Chroma_Width*vi.height+2048,32);
bufV = (uint8_t*)_aligned_malloc(m_decoder.Chroma_Width*vi.height+2048,32);
if (_upConv == 2) {
size_t ypitch = (vi.width + 31) & ~31;
size_t uvpitch = (m_decoder.Chroma_Width + 15) & ~15;
bufY = (uint8_t*)_aligned_malloc(ypitch * (vi.height + 1), 32);
bufU = (uint8_t*)_aligned_malloc(uvpitch * (vi.height + 1), 32);
bufV = (uint8_t*)_aligned_malloc(uvpitch * (vi.height + 1), 32);
if (bufY == NULL || bufU == NULL || bufV == NULL)
env->ThrowError("MPEG2Source: malloc failure (bufY, bufU, bufV)!");
}

u444 = v444 = NULL;
if (_upConv == 2)
{
u444 = (uint8_t*)_aligned_malloc(vi.width*vi.height+2048,32);
v444 = (uint8_t*)_aligned_malloc(vi.width*vi.height+2048,32);
if (u444 == NULL || v444 == NULL)
env->ThrowError("MPEG2Source: malloc failure (u444, v444)!");
}
}

MPEG2Source::~MPEG2Source()
{
m_decoder.Close();
if (out != NULL) { free(out); out = NULL; }
if (bufY != NULL) { _aligned_free(bufY); bufY = NULL; }
if (bufU != NULL) { _aligned_free(bufU); bufU = NULL; }
if (bufV != NULL) { _aligned_free(bufV); bufV = NULL; }
if (u444 != NULL) { _aligned_free(u444); u444 = NULL; }
if (v444 != NULL) { _aligned_free(v444); v444 = NULL; }
_aligned_free(bufY);
_aligned_free(bufU);
_aligned_free(bufV);
bufY = bufU = bufV = nullptr;
}


Expand Down Expand Up @@ -289,55 +280,41 @@ PVideoFrame __stdcall MPEG2Source::GetFrame(int n, IScriptEnvironment* env)
}

PVideoFrame frame = env->NewVideoFrame(vi);

if (m_decoder.chroma_format == 1 && m_decoder.upConv == 0) // YV12
{
out->y = frame->GetWritePtr(PLANAR_Y);
out->u = frame->GetWritePtr(PLANAR_U);
out->v = frame->GetWritePtr(PLANAR_V);
out->ypitch = frame->GetPitch(PLANAR_Y);
out->uvpitch = frame->GetPitch(PLANAR_U);
out->ywidth = frame->GetRowSize(PLANAR_Y);
out->uvwidth = frame->GetRowSize(PLANAR_U);
out->yheight = frame->GetHeight(PLANAR_Y);
out->uvheight = frame->GetHeight(PLANAR_V);
}
else // its 4:2:2, pass our own buffers
{
out->y = bufY;
out->u = bufU;
out->v = bufV;
out->ypitch = vi.width;
out->uvpitch = m_decoder.Chroma_Width;
out->ywidth = vi.width;
out->uvwidth = m_decoder.Chroma_Width;
out->yheight = vi.height;
out->uvheight = vi.height;
YV12PICT out = {};

if (m_decoder.upConv != 2) { // YV12 || YV16
out.y = frame->GetWritePtr(PLANAR_Y);
out.u = frame->GetWritePtr(PLANAR_U);
out.v = frame->GetWritePtr(PLANAR_V);
out.ypitch = frame->GetPitch(PLANAR_Y);
out.uvpitch = frame->GetPitch(PLANAR_U);
out.ywidth = frame->GetRowSize(PLANAR_Y);
out.uvwidth = frame->GetRowSize(PLANAR_U);
out.yheight = frame->GetHeight(PLANAR_Y);
out.uvheight = frame->GetHeight(PLANAR_V);
} else {
out.y = bufY;
out.u = bufU;
out.v = bufV;
out.ypitch = (vi.width + 31) & ~31;
out.uvpitch = (m_decoder.Chroma_Width + 15) & ~15;
out.ywidth = vi.width;
out.uvwidth = m_decoder.Chroma_Width;
out.yheight = vi.height;
out.uvheight = vi.height;
}

m_decoder.Decode(n, out);
m_decoder.Decode(n, &out);

if ( m_decoder.Luminance_Flag )
m_decoder.LuminanceFilter(out->y, out->ywidth, out->yheight, out->ypitch);
m_decoder.LuminanceFilter(out.y, out.ywidth, out.yheight, out.ypitch);

if ((m_decoder.chroma_format == 2 && m_decoder.upConv != 2) ||
(m_decoder.chroma_format == 1 && m_decoder.upConv == 1)) // convert 4:2:2 (planar) to YUY2 (packed)
{
env->BitBlt(frame->GetWritePtr(PLANAR_Y), frame->GetPitch(PLANAR_Y),
bufY, vi.width, vi.width, vi.height);
env->BitBlt(frame->GetWritePtr(PLANAR_U), frame->GetPitch(PLANAR_U),
bufU, m_decoder.Chroma_Width, m_decoder.Chroma_Width, vi.height);
env->BitBlt(frame->GetWritePtr(PLANAR_V), frame->GetPitch(PLANAR_V),
bufV, m_decoder.Chroma_Width, m_decoder.Chroma_Width, vi.height);
}

if (m_decoder.upConv == 2) // convert 4:2:2 (planar) to 4:4:4 (planar) and then to RGB24
{
if (m_decoder.upConv == 2) { // convert 4:2:2 (planar) to 4:4:4 (planar)
env->BitBlt(frame->GetWritePtr(PLANAR_Y), frame->GetPitch(PLANAR_Y),
bufY, vi.width, vi.width, vi.height);
conv422to444(out->u, frame->GetWritePtr(PLANAR_U), out->uvpitch,
bufY, out.ypitch, vi.width, vi.height);
conv422to444(out.u, frame->GetWritePtr(PLANAR_U), out.uvpitch,
frame->GetPitch(PLANAR_U), vi.width, vi.height);
conv422to444(out->v, frame->GetWritePtr(PLANAR_V), out->uvpitch,
conv422to444(out.v, frame->GetWritePtr(PLANAR_V), out.uvpitch,
frame->GetPitch(PLANAR_V), vi.width, vi.height);
}

Expand Down
2 changes: 0 additions & 2 deletions src/dgdecode/AvisynthAPI.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -36,9 +36,7 @@ class MPEG2Source: public IClip {
VideoInfo vi;
int ovr_idct;
int _PP_MODE;
YV12PICT *out;
uint8_t *bufY, *bufU, *bufV; // for 4:2:2 input support
uint8_t *u444, *v444; // for RGB24 output
CMPEG2Decoder m_decoder;

void override(int ovr_idct);
Expand Down
14 changes: 7 additions & 7 deletions src/dgdecode/color_convert.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -195,8 +195,8 @@ void conv420to422P(const uint8_t *src, uint8_t *dst, int src_pitch, int dst_pitc
const __m128i one = _mm_set1_epi8(0x01);

for (int x = 0; x < width; x += 16) {
const __m128i sx0 = _mm_load_si128(reinterpret_cast<const __m128i*>(s0 + x));
__m128i sx1 = _mm_load_si128(reinterpret_cast<const __m128i*>(s1 + x));
const __m128i sx0 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(s0 + x));
__m128i sx1 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(s1 + x));

sx1 = _mm_subs_epu8(sx1, one);
sx1 = _mm_avg_epu8(_mm_avg_epu8(sx1, sx0), sx0);
Expand All @@ -212,9 +212,9 @@ void conv420to422P(const uint8_t *src, uint8_t *dst, int src_pitch, int dst_pitc
const uint8_t* s2 = s1 + src_pitch;

for (int x = 0; x < width; x += 16) {
__m128i sx0 = _mm_load_si128(reinterpret_cast<const __m128i*>(s0 + x));
const __m128i sx1 = _mm_load_si128(reinterpret_cast<const __m128i*>(s1 + x));
__m128i sx2 = _mm_load_si128(reinterpret_cast<const __m128i*>(s2 + x));
__m128i sx0 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(s0 + x));
const __m128i sx1 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(s1 + x));
__m128i sx2 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(s2 + x));

sx0 = _mm_subs_epu8(sx0, one);
sx2 = _mm_subs_epu8(sx2, one);
Expand All @@ -231,8 +231,8 @@ void conv420to422P(const uint8_t *src, uint8_t *dst, int src_pitch, int dst_pitc
}

for (int x = 0; x < width; x += 16) {
__m128i sx0 = _mm_load_si128(reinterpret_cast<const __m128i*>(s0 + x));
const __m128i sx1 = _mm_load_si128(reinterpret_cast<const __m128i*>(s1 + x));
__m128i sx0 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(s0 + x));
const __m128i sx1 = _mm_loadu_si128(reinterpret_cast<const __m128i*>(s1 + x));

sx0 = _mm_subs_epu8(sx0, one);
sx0 = _mm_avg_epu8(_mm_avg_epu8(sx0, sx1), sx1);
Expand Down
10 changes: 5 additions & 5 deletions src/dgdecode/misc.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,10 +34,10 @@ void fast_copy(const uint8_t *src, const int src_stride, uint8_t *dst,

size_t __cdecl dprintf(char* fmt, ...);

bool has_sse2() noexcept;

bool has_sse3() noexcept;

bool has_avx2() noexcept;
bool has_sse2() noexcept;

bool has_sse3() noexcept;

bool has_avx2() noexcept;

#endif

0 comments on commit df0dde8

Please sign in to comment.