Skip to content

Commit

Permalink
Add spatiotemporal SSE2 code for 8 bit
Browse files Browse the repository at this point in the history
  • Loading branch information
cantabile committed Aug 20, 2016
1 parent 40d6ebe commit a971d68
Show file tree
Hide file tree
Showing 2 changed files with 126 additions and 13 deletions.
6 changes: 0 additions & 6 deletions readme.rst
View file
Original file line number Diff line number Diff line change
Expand Up @@ -61,9 +61,3 @@ License
The original FluxSmooth author said:
"There is no copyright on this code, and there are no conditions on its distribution or use. Do with it what you will."
Thank you.


To do
=====

Convert the asm for the spatiotemporal version from this 64 bit edition: http://forum.doom9.org/showthread.php?t=158245
133 changes: 126 additions & 7 deletions src/fluxsmooth.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,14 @@ typedef struct FluxSmoothData {


#if defined(FLUXSMOOTH_X86)

#ifdef _WIN32
#define FORCE_INLINE __forceinline
#else
#define FORCE_INLINE inline __attribute__((always_inline))
#endif


static void fluxsmooth_temporal_uint8_sse2(const uint8_t *srcpp, const uint8_t *srccp, const uint8_t *srcnp, uint8_t *dstp, int width, int height, int stride, int temporal_threshold, int spatial_threshold) {
(void)width;
(void)spatial_threshold;
Expand Down Expand Up @@ -216,6 +224,113 @@ static void fluxsmooth_temporal_uint16_sse2(const uint8_t *srcpp, const uint8_t
srcnp += stride;
dstp += stride;
}
}


static FORCE_INLINE void fluxsmooth_spatiotemporal_mmword_uint8_sse2(const uint8_t *srcpp, const uint8_t *srccp, const uint8_t *srcnp, uint8_t *dstp, int x, int stride, const __m128i words_temporal_threshold, const __m128i words_spatial_threshold) {
__m128i prev = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srcpp[x]), zeroes);
__m128i curr = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x]), zeroes);
__m128i next = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srcnp[x]), zeroes);
// "n" for "neighbour"
__m128i n1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x - stride - 1]), zeroes);
__m128i n2 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x - stride]), zeroes);
__m128i n3 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x - stride + 1]), zeroes);
__m128i n4 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x - 1]), zeroes);
__m128i n5 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x + 1]), zeroes);
__m128i n6 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x + stride - 1]), zeroes);
__m128i n7 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x + stride]), zeroes);
__m128i n8 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&srccp[x + stride + 1]), zeroes);

__m128i count = _mm_set1_epi16(11);
__m128i sum = curr;

#define CHECK_NEIGHBOUR(n, threshold) \
__m128i n##_diff = _mm_or_si128(_mm_subs_epu16(n, curr), \
_mm_subs_epu16(curr, n)); \
__m128i n##_mask = _mm_cmpgt_epi16(n##_diff, threshold); \
count = _mm_add_epi16(count, n##_mask); \
sum = _mm_add_epi16(sum, _mm_andnot_si128(n##_mask, n));

CHECK_NEIGHBOUR(prev, words_temporal_threshold)
CHECK_NEIGHBOUR(next, words_temporal_threshold)
CHECK_NEIGHBOUR(n1, words_spatial_threshold)
CHECK_NEIGHBOUR(n2, words_spatial_threshold)
CHECK_NEIGHBOUR(n3, words_spatial_threshold)
CHECK_NEIGHBOUR(n4, words_spatial_threshold)
CHECK_NEIGHBOUR(n5, words_spatial_threshold)
CHECK_NEIGHBOUR(n6, words_spatial_threshold)
CHECK_NEIGHBOUR(n7, words_spatial_threshold)
CHECK_NEIGHBOUR(n8, words_spatial_threshold)

#undef CHECK_NEIGHBOUR

__m128 fcountlo = _mm_cvtepi32_ps(_mm_unpacklo_epi16(count, zeroes));
__m128 fcounthi = _mm_cvtepi32_ps(_mm_unpackhi_epi16(count, zeroes));

__m128 favglo = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpacklo_epi16(sum, zeroes)),
_mm_rcp_ps(fcountlo));
__m128 favghi = _mm_mul_ps(_mm_cvtepi32_ps(_mm_unpackhi_epi16(sum, zeroes)),
_mm_rcp_ps(fcounthi));

const __m128 floats_0_5 = _mm_set1_ps(0.5f);

favglo = _mm_add_ps(favglo, floats_0_5);
favghi = _mm_add_ps(favghi, floats_0_5);

__m128i avg = _mm_packs_epi32(_mm_cvttps_epi32(favglo),
_mm_cvttps_epi32(favghi));

__m128i m0 = _mm_and_si128(_mm_cmpgt_epi16(curr, prev),
_mm_cmpgt_epi16(curr, next));
__m128i m1 = _mm_and_si128(_mm_cmpgt_epi16(prev, curr),
_mm_cmpgt_epi16(next, curr));

m0 = _mm_or_si128(m0, m1);

m1 = _mm_or_si128(_mm_and_si128(m0, avg),
_mm_andnot_si128(m0, curr));

_mm_storel_epi64((__m128i *)&dstp[x],
_mm_packus_epi16(m1, zeroes));
}


static void fluxsmooth_spatiotemporal_uint8_sse2(const uint8_t *srcpp, const uint8_t *srccp, const uint8_t *srcnp, uint8_t *dstp, int width, int height, int stride, int temporal_threshold, int spatial_threshold) {
const __m128i words_temporal_threshold = _mm_set1_epi16(temporal_threshold);
const __m128i words_spatial_threshold = _mm_set1_epi16(spatial_threshold);

memcpy(dstp, srccp, width);
srcpp += stride;
srccp += stride;
srcnp += stride;
dstp += stride;

const int pixels_in_mm = 8 / sizeof(uint8_t);
const int skip_left = 1;
const int skip_right = 1;

int width_sse2 = (width & ~(pixels_in_mm - 1)) + skip_left + skip_right;
if (width_sse2 > stride)
width_sse2 -= pixels_in_mm;

for (int y = 1; y < height - 1; y++) {
dstp[0] = srccp[0];

for (int x = skip_left; x < width_sse2 - skip_right; x += pixels_in_mm)
fluxsmooth_spatiotemporal_mmword_uint8_sse2(srcpp, srccp, srcnp, dstp, x, stride, words_temporal_threshold, words_spatial_threshold);

if (width + skip_left + skip_right > width_sse2)
fluxsmooth_spatiotemporal_mmword_uint8_sse2(srcpp, srccp, srcnp, dstp, width - pixels_in_mm - skip_right, stride, words_temporal_threshold, words_spatial_threshold);

dstp[width - 1] = srccp[width - 1];

srcpp += stride;
srccp += stride;
srcnp += stride;
dstp += stride;
}

memcpy(dstp, srccp, width);

#undef zeroes
}
Expand Down Expand Up @@ -315,13 +430,6 @@ static void fluxsmooth_temporal_uint16_c(const uint8_t *srcpp, const uint8_t *sr
dstp += stride;
}
}
#endif // FLUXSMOOTH_X86


static void VS_CC fluxSmoothInit(VSMap *in, VSMap *out, void **instanceData, VSNode *node, VSCore *core, const VSAPI *vsapi) {
FluxSmoothData *d = (FluxSmoothData *)*instanceData;
vsapi->setVideoInfo(d->vi, 1, node);
}


static void fluxsmooth_spatiotemporal_uint8_c(const uint8_t *srcpp, const uint8_t *srccp, const uint8_t *srcnp, uint8_t *dstp, int width, int height, int stride, int temporal_threshold, int spatial_threshold) {
Expand Down Expand Up @@ -430,6 +538,13 @@ static void fluxsmooth_spatiotemporal_uint8_c(const uint8_t *srcpp, const uint8_

memcpy(dstp, srccp, stride);
}
#endif // FLUXSMOOTH_X86


static void VS_CC fluxSmoothInit(VSMap *in, VSMap *out, void **instanceData, VSNode *node, VSCore *core, const VSAPI *vsapi) {
FluxSmoothData *d = (FluxSmoothData *)*instanceData;
vsapi->setVideoInfo(d->vi, 1, node);
}


static void fluxsmooth_spatiotemporal_uint16_c(const uint8_t *srcpp, const uint8_t *srccp, const uint8_t *srcnp, uint8_t *dstp, int width, int height, int stride, int temporal_threshold, int spatial_threshold) {
Expand Down Expand Up @@ -696,7 +811,11 @@ static void VS_CC fluxSmoothCreate(const VSMap *in, VSMap *out, void *userData,
// Select the functions.
if (function == SpatioTemporalFlux) {
if (d.vi->format->bytesPerSample == 1) {
#if defined(FLUXSMOOTH_X86)
d.flux_function = fluxsmooth_spatiotemporal_uint8_sse2;
#else
d.flux_function = fluxsmooth_spatiotemporal_uint8_c;
#endif
} else {
d.flux_function = fluxsmooth_spatiotemporal_uint16_c;
}
Expand Down

0 comments on commit a971d68

Please sign in to comment.