| Message ID | 20230928210848.95565-1-ffmpeg@haasn.xyz |
|---|---|
| State | Accepted |
| Commit | 22530ad1cee9002903a5a4302e6b20bd7823c445 |
| Headers | show |
| Series | [FFmpeg-devel] lavc/h274: transpose IDCT | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/make_loongarch64 | success | Make finished |
| yinshiyou/make_fate_loongarch64 | success | Make fate finished |
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
On Thu, 28 Sep 2023 23:08:48 +0200 Niklas Haas <ffmpeg@haasn.xyz> wrote: > From: Niklas Haas <git@haasn.dev> > > This is mathematically equivalent to what we were doing before, but > gives subtly different results due to rounding (rows first vs columns > first). Doing it this way makes our film grain database generation match > reference implementation and now produces bit-exact outputs in my > testing. > > Rename the transposed variables to be a bit less confusing. > --- > libavcodec/h274.c | 20 ++++++++++---------- > 1 file changed, 10 insertions(+), 10 deletions(-) > > diff --git a/libavcodec/h274.c b/libavcodec/h274.c > index a5caf09564d..5709200322e 100644 > --- a/libavcodec/h274.c > +++ b/libavcodec/h274.c > @@ -59,13 +59,13 @@ static void init_slice_c(int8_t out[64][64], uint8_t h, uint8_t v, > // > // Note: To make the subsequent matrix multiplication cache friendlier, we > // store each *column* of the starting image in a *row* of `out` > - for (int y = 0; y <= freq_v; y++) { > - for (int x = 0; x <= freq_h; x += 4) { > + for (int l = 0; l <= freq_v; l++) { > + for (int k = 0; k <= freq_h; k += 4) { > uint16_t offset = seed % 2048; > - out[x + 0][y] = Gaussian_LUT[offset + 0]; > - out[x + 1][y] = Gaussian_LUT[offset + 1]; > - out[x + 2][y] = Gaussian_LUT[offset + 2]; > - out[x + 3][y] = Gaussian_LUT[offset + 3]; > + out[l][k + 0] = Gaussian_LUT[offset + 0]; > + out[l][k + 1] = Gaussian_LUT[offset + 1]; > + out[l][k + 2] = Gaussian_LUT[offset + 2]; > + out[l][k + 3] = Gaussian_LUT[offset + 3]; > prng_shift(&seed); > } > } > @@ -74,9 +74,9 @@ static void init_slice_c(int8_t out[64][64], uint8_t h, uint8_t v, > > // 64x64 inverse integer transform > for (int y = 0; y < 64; y++) { > - for (int x = 0; x <= freq_h; x++) { > + for (int x = 0; x <= freq_v; x++) { > int32_t sum = 0; > - for (int p = 0; p <= freq_v; p++) > + for (int p = 0; p <= freq_h; p++) > sum += R64T[y][p] * out[x][p]; > tmp[y][x] = (sum + 128) >> 8; > } > @@ -85,8 +85,8 @@ static void init_slice_c(int8_t out[64][64], uint8_t h, uint8_t v, > for (int y = 0; y < 64; y++) { > for (int x = 0; x < 64; x++) { > int32_t sum = 0; > - for (int p = 0; p <= freq_h; p++) > - sum += tmp[y][p] * R64T[x][p]; // R64T^T = R64 > + for (int p = 0; p <= freq_v; p++) > + sum += tmp[x][p] * R64T[y][p]; // R64T^T = R64 > // Renormalize and clip to [-127, 127] > out[y][x] = av_clip((sum + 128) >> 8, -127, 127); > } > -- > 2.42.0 > Merged as 22530ad1ce
diff --git a/libavcodec/h274.c b/libavcodec/h274.c index a5caf09564d..5709200322e 100644 --- a/libavcodec/h274.c +++ b/libavcodec/h274.c @@ -59,13 +59,13 @@ static void init_slice_c(int8_t out[64][64], uint8_t h, uint8_t v, // // Note: To make the subsequent matrix multiplication cache friendlier, we // store each *column* of the starting image in a *row* of `out` - for (int y = 0; y <= freq_v; y++) { - for (int x = 0; x <= freq_h; x += 4) { + for (int l = 0; l <= freq_v; l++) { + for (int k = 0; k <= freq_h; k += 4) { uint16_t offset = seed % 2048; - out[x + 0][y] = Gaussian_LUT[offset + 0]; - out[x + 1][y] = Gaussian_LUT[offset + 1]; - out[x + 2][y] = Gaussian_LUT[offset + 2]; - out[x + 3][y] = Gaussian_LUT[offset + 3]; + out[l][k + 0] = Gaussian_LUT[offset + 0]; + out[l][k + 1] = Gaussian_LUT[offset + 1]; + out[l][k + 2] = Gaussian_LUT[offset + 2]; + out[l][k + 3] = Gaussian_LUT[offset + 3]; prng_shift(&seed); } } @@ -74,9 +74,9 @@ static void init_slice_c(int8_t out[64][64], uint8_t h, uint8_t v, // 64x64 inverse integer transform for (int y = 0; y < 64; y++) { - for (int x = 0; x <= freq_h; x++) { + for (int x = 0; x <= freq_v; x++) { int32_t sum = 0; - for (int p = 0; p <= freq_v; p++) + for (int p = 0; p <= freq_h; p++) sum += R64T[y][p] * out[x][p]; tmp[y][x] = (sum + 128) >> 8; } @@ -85,8 +85,8 @@ static void init_slice_c(int8_t out[64][64], uint8_t h, uint8_t v, for (int y = 0; y < 64; y++) { for (int x = 0; x < 64; x++) { int32_t sum = 0; - for (int p = 0; p <= freq_h; p++) - sum += tmp[y][p] * R64T[x][p]; // R64T^T = R64 + for (int p = 0; p <= freq_v; p++) + sum += tmp[x][p] * R64T[y][p]; // R64T^T = R64 // Renormalize and clip to [-127, 127] out[y][x] = av_clip((sum + 128) >> 8, -127, 127); }
From: Niklas Haas <git@haasn.dev> This is mathematically equivalent to what we were doing before, but gives subtly different results due to rounding (rows first vs columns first). Doing it this way makes our film grain database generation match reference implementation and now produces bit-exact outputs in my testing. Rename the transposed variables to be a bit less confusing. --- libavcodec/h274.c | 20 ++++++++++---------- 1 file changed, 10 insertions(+), 10 deletions(-)