int write_image(dt_imageio_module_data_t *data, const char *filename, const void *ivoid, void *exif,
int exif_len, int imgid, int num, int total)
{
const dt_imageio_module_data_t *const pfm = data;
int status = 0;
FILE *f = fopen(filename, "wb");
if(f)
{
// INFO: per-line fwrite call seems to perform best. LebedevRI, 18.04.2014
(void)fprintf(f, "PF\n%d %d\n-1.0\n", pfm->width, pfm->height);
void *buf_line = dt_alloc_align(16, 3 * sizeof(float) * pfm->width);
for(int j = 0; j < pfm->height; j++)
{
// NOTE: pfm has rows in reverse order
const int row_in = pfm->height - 1 - j;
const float *in = (const float *)ivoid + 4 * (size_t)pfm->width * row_in;
float *out = (float *)buf_line;
for(int i = 0; i < pfm->width; i++, in += 4, out += 3)
{
memcpy(out, in, 3 * sizeof(float));
}
int cnt = fwrite(buf_line, 3 * sizeof(float), pfm->width, f);
if(cnt != pfm->width)
status = 1;
else
status = 0;
}
dt_free_align(buf_line);
buf_line = NULL;
fclose(f);
}
return status;
}
int dt_dev_pixelpipe_cache_init(dt_dev_pixelpipe_cache_t *cache, int entries, size_t size)
{
cache->entries = entries;
cache->data = (void **)calloc(entries, sizeof(void *));
cache->size = (size_t *)calloc(entries, sizeof(size_t));
cache->hash = (uint64_t *)calloc(entries, sizeof(uint64_t));
cache->used = (int32_t *)calloc(entries, sizeof(int32_t));
for(int k = 0; k < entries; k++)
{
cache->data[k] = (void *)dt_alloc_align(16, size);
if(!cache->data[k]) goto alloc_memory_fail;
cache->size[k] = size;
#ifdef _DEBUG
memset(cache->data[k], 0x5d, size);
#endif
cache->hash[k] = -1;
cache->used[k] = 0;
}
cache->queries = cache->misses = 0;
return 1;
alloc_memory_fail:
for(int k = 0; k < entries; k++)
{
if(cache->data[k]) dt_free_align(cache->data[k]);
}
free(cache->data);
free(cache->size);
free(cache->hash);
free(cache->used);
return 0;
}
void
dt_image_cache_init(dt_image_cache_t *cache)
{
// the image cache does no serialization.
// (unsafe. data should be in db/xmp, not in any other additional cache,
// also, it should be relatively fast to get the image_t structs from sql.)
// TODO: actually an independent conf var?
// too large: dangerous and wasteful?
// can we get away with a fixed size?
const uint32_t max_mem = 50*1024*1024;
uint32_t num = (uint32_t)(1.5f*max_mem/sizeof(dt_image_t));
dt_cache_init(&cache->cache, num, 16, 64, max_mem);
dt_cache_set_allocate_callback(&cache->cache, &dt_image_cache_allocate, cache);
dt_cache_set_cleanup_callback (&cache->cache, &dt_image_cache_deallocate, cache);
// might have been rounded to power of two:
num = dt_cache_capacity(&cache->cache);
cache->images = dt_alloc_align(64, sizeof(dt_image_t)*num);
dt_print(DT_DEBUG_CACHE, "[image_cache] has %d entries\n", num);
// initialize first image as empty data:
dt_image_init(cache->images);
for(uint32_t k=1; k<num; k++)
{
// optimized initialization (avoid accessing conf):
memcpy(cache->images + k, cache->images, sizeof(dt_image_t));
}
}
// callback for the cache backend to initialize payload pointers
int32_t
dt_mipmap_cache_allocate_dynamic(void *data, const uint32_t key, int32_t *cost, void **buf)
{
// for full image buffers
struct dt_mipmap_buffer_dsc* dsc = *buf;
// alloc mere minimum for the header + broken image buffer:
if(!dsc)
{
*buf = dt_alloc_align(16, sizeof(*dsc)+sizeof(float)*4*64);
// fprintf(stderr, "[mipmap cache] alloc dynamic for key %u %lX\n", key, (uint64_t)*buf);
if(!(*buf))
{
fprintf(stderr, "[mipmap cache] memory allocation failed!\n");
exit(1);
}
dsc = *buf;
dsc->width = 0;
dsc->height = 0;
dsc->size = sizeof(*dsc)+sizeof(float)*4*64;
}
assert(dsc->size >= sizeof(*dsc));
dsc->flags = DT_MIPMAP_BUFFER_DSC_FLAG_GENERATE;
// cost is just flat one for the buffer, as the buffers might have different sizes,
// to make sure quota is meaningful.
*cost = 1;
// fprintf(stderr, "dummy allocing %lX\n", (uint64_t)*buf);
return 1; // request write lock
}
int dt_dev_pixelpipe_process(dt_dev_pixelpipe_t *pipe, dt_develop_t *dev, int x, int y, int width, int height, float scale)
{
pipe->processing = 1;
printf("pixelpipe process start\n");
// have backbuf in right size:
if(pipe->backbuf_size < width*height*4*sizeof(uint8_t))
{
pthread_mutex_lock(&pipe->backbuf_mutex);
pipe->backbuf_size = width*height*4*sizeof(uint8_t);
free(pipe->backbuf);
pipe->backbuf = (uint8_t *)dt_alloc_align(16, pipe->backbuf_size);
pthread_mutex_unlock(&pipe->backbuf_mutex);
}
// scale node (is slow):
// scale *= 2;
// FIXME: this seems to be a bug in gegl. need to manually adjust updated roi here.
GeglRectangle roi = (GeglRectangle)
{
x, y, width, height
};
GeglRectangle roio = (GeglRectangle)
{
roi.x/scale, roi.y/scale, roi.width/scale, roi.height/scale
};
roio.x = MAX(0, roio.x);
roio.y = MAX(0, roio.y);
roio.width = MIN(pipe->iwidth -roio.x-1, roio.width);
roio.height = MIN(pipe->iheight-roio.y-1, roio.height);
GeglProcessor *processor = gegl_node_new_processor (pipe->output, &roio);
// gegl_node_set(pipe->scale, "x", scale, "y", scale, NULL);
// GeglProcessor *processor = gegl_node_new_processor (pipe->output, roi);
double progress;
// TODO: insert constant scale node at beginning, maintain lo-res branch of pipeline (shadowed).
// TODO: decide on scale param, which one to use.
while (gegl_processor_work (processor, &progress))
{
// if history changed, abort processing?
if(pipe->changed != DT_DEV_PIPE_UNCHANGED || dev->gui_leaving) return 1;
}
gegl_processor_destroy (processor);
// gegl scale node turned out to be even slower :(
gegl_node_blit (pipe->output, scale, &roi, babl_format("RGBA u8"), pipe->backbuf, GEGL_AUTO_ROWSTRIDE, GEGL_BLIT_CACHE);
// gegl_node_blit (pipe->output, 1.0, roi, babl_format("RGBA u8"), output, GEGL_AUTO_ROWSTRIDE, GEGL_BLIT_CACHE);
// TODO: update histograms here with this data?
printf("pixelpipe process end\n");
pipe->processing = 0;
return 0;
}
// compression stuff: alloc a buffer if needed
uint8_t*
dt_mipmap_cache_alloc_scratchmem(
const dt_mipmap_cache_t *cache)
{
const size_t size = cache->mip[DT_MIPMAP_3].max_width *
cache->mip[DT_MIPMAP_3].max_height;
if(cache->compression_type)
{
return dt_alloc_align(64, size * 4 * sizeof(uint8_t));
}
else // no compression, no buffer:
return NULL;
}
int dt_dev_pixelpipe_cache_get_weighted(dt_dev_pixelpipe_cache_t *cache, const uint64_t hash,
const size_t size, void **data, int weight)
{
cache->queries++;
*data = NULL;
int max_used = -1, max = 0;
size_t sz = 0;
for(int k = 0; k < cache->entries; k++)
{
// search for hash in cache
if(cache->used[k] > max_used)
{
max_used = cache->used[k];
max = k;
}
cache->used[k]++; // age all entries
if(cache->hash[k] == hash)
{
*data = cache->data[k];
sz = cache->size[k];
cache->used[k] = weight; // this is the MRU entry
}
}
if(!*data || sz < size)
{
// kill LRU entry
// printf("[pixelpipe_cache_get] hash not found, returning slot %d/%d age %d\n", max, cache->entries,
// weight);
if(cache->size[max] < size)
{
dt_free_align(cache->data[max]);
cache->data[max] = (void *)dt_alloc_align(16, size);
cache->size[max] = size;
}
*data = cache->data[max];
cache->hash[max] = hash;
cache->used[max] = weight;
cache->misses++;
return 1;
}
else
return 0;
}
dt_gaussian_t *
dt_gaussian_init(
const int width, // width of input image
const int height, // height of input image
const int channels, // channels per pixel
const float *max, // maximum allowed values per channel for clamping
const float *min, // minimum allowed values per channel for clamping
const float sigma, // gaussian sigma
const int order) // order of gaussian blur
{
dt_gaussian_t *g = (dt_gaussian_t *)malloc(sizeof(dt_gaussian_t));
if(!g) return NULL;
g->width = width;
g->height = height;
g->channels = channels;
g->sigma = sigma;
g->order = order;
g->buf = NULL;
g->max = (float *)malloc(channels * sizeof(float));
g->min = (float *)malloc(channels * sizeof(float));
if(!g->min || !g->max) goto error;
for(int k=0; k < channels; k++)
{
g->max[k] = max[k];
g->min[k] = min[k];
}
g->buf = dt_alloc_align(64, width*height*channels*sizeof(float));
if(!g->buf) goto error;
return g;
error:
free(g->buf);
free(g->max);
free(g->min);
free(g);
return NULL;
}
int write_image(dt_imageio_module_data_t *data, const char *filename, const void *ivoid, void *exif,
int exif_len, int imgid, int num, int total)
{
const dt_imageio_module_data_t *const pfm = data;
int status = 0;
FILE *f = fopen(filename, "wb");
if(f)
{
// align pfm header to sse, assuming the file will
// be mmapped to page boundaries.
char header[1024];
snprintf(header, 1024, "PF\n%d %d\n-1.0", pfm->width, pfm->height);
size_t len = strlen(header);
fprintf(f, "PF\n%d %d\n-1.0", pfm->width, pfm->height);
ssize_t off = 0;
while((len + 1 + off) & 0xf) off++;
while(off-- > 0) fprintf(f, "0");
fprintf(f, "\n");
void *buf_line = dt_alloc_align(16, 3 * sizeof(float) * pfm->width);
for(int j = 0; j < pfm->height; j++)
{
// NOTE: pfm has rows in reverse order
const int row_in = pfm->height - 1 - j;
const float *in = (const float *)ivoid + 4 * (size_t)pfm->width * row_in;
float *out = (float *)buf_line;
for(int i = 0; i < pfm->width; i++, in += 4, out += 3)
{
memcpy(out, in, 3 * sizeof(float));
}
// INFO: per-line fwrite call seems to perform best. LebedevRI, 18.04.2014
int cnt = fwrite(buf_line, 3 * sizeof(float), pfm->width, f);
if(cnt != pfm->width)
status = 1;
else
status = 0;
}
dt_free_align(buf_line);
buf_line = NULL;
fclose(f);
}
return status;
}
// callback for the imageio core to allocate memory.
// only needed for _F and _FULL buffers, as they change size
// with the input image. will allocate img->width*img->height*img->bpp bytes.
void*
dt_mipmap_cache_alloc(dt_image_t *img, dt_mipmap_size_t size, dt_mipmap_cache_allocator_t a)
{
assert(size == DT_MIPMAP_FULL);
struct dt_mipmap_buffer_dsc** dsc = (struct dt_mipmap_buffer_dsc**)a;
int32_t wd = img->width;
int32_t ht = img->height;
int32_t bpp = img->bpp;
const uint32_t buffer_size =
((wd*ht*bpp) + sizeof(**dsc));
// buf might have been alloc'ed before,
// so only check size and re-alloc if necessary:
if(!(*dsc) || ((*dsc)->size < buffer_size) || ((void *)*dsc == (void *)dt_mipmap_cache_static_dead_image))
{
if((void *)*dsc != (void *)dt_mipmap_cache_static_dead_image)
dt_free_align(*dsc);
*dsc = dt_alloc_align(64, buffer_size);
// fprintf(stderr, "[mipmap cache] alloc for key %u %p\n", get_key(img->id, size), *buf);
if(!(*dsc))
{
// return fallback: at least alloc size for a dead image:
*dsc = (struct dt_mipmap_buffer_dsc *)dt_mipmap_cache_static_dead_image;
// allocator holds the pointer. but imageio client is tricked to believe allocation failed:
return NULL;
}
// set buffer size only if we're making it larger.
(*dsc)->size = buffer_size;
}
(*dsc)->width = wd;
(*dsc)->height = ht;
(*dsc)->flags = DT_MIPMAP_BUFFER_DSC_FLAG_GENERATE;
// fprintf(stderr, "full buffer allocating img %u %d x %d = %u bytes (%p)\n", img->id, img->width, img->height, buffer_size, *buf);
// trick the user into using a pointer without the header:
return (*dsc)+1;
}
// allocate output buffer with monochrome brightness channel from input, padded
// up by max_supp on all four sides, dimensions written to wd2 ht2
static inline float *ll_pad_input(
const float *const input,
const int wd,
const int ht,
const int max_supp,
int *wd2,
int *ht2)
{
const int stride = 4;
*wd2 = 2*max_supp + wd;
*ht2 = 2*max_supp + ht;
float *const out = dt_alloc_align(16, *wd2**ht2*sizeof(*out));
#ifdef _OPENMP
#pragma omp parallel for schedule(dynamic) default(none) shared(wd2, ht2)
#endif
for(int j=0;j<ht;j++)
{
for(int i=0;i<max_supp;i++)
out[(j+max_supp)**wd2+i] = input[stride*wd*j]* 0.01f; // L -> [0,1]
for(int i=0;i<wd;i++)
out[(j+max_supp)**wd2+i+max_supp] = input[stride*(wd*j+i)] * 0.01f; // L -> [0,1]
for(int i=wd+max_supp;i<*wd2;i++)
out[(j+max_supp)**wd2+i] = input[stride*(j*wd+wd-1)] * 0.01f; // L -> [0,1]
}
#ifdef _OPENMP
#pragma omp parallel for schedule(dynamic) default(none) shared(wd2, ht2)
#endif
for(int j=0;j<max_supp;j++)
memcpy(out + *wd2*j, out+max_supp**wd2, sizeof(float)**wd2);
#ifdef _OPENMP
#pragma omp parallel for schedule(dynamic) default(none) shared(wd2, ht2)
#endif
for(int j=max_supp+ht;j<*ht2;j++)
memcpy(out + *wd2*j, out + *wd2*(max_supp+ht-1), sizeof(float)**wd2);
return out;
}
// callback for the imageio core to allocate memory.
// only needed for _F and _FULL buffers, as they change size
// with the input image. will allocate img->width*img->height*img->bpp bytes.
void *dt_mipmap_cache_alloc(dt_mipmap_buffer_t *buf, const dt_image_t *img)
{
assert(buf->size == DT_MIPMAP_FULL);
const int wd = img->width;
const int ht = img->height;
struct dt_mipmap_buffer_dsc *dsc = (struct dt_mipmap_buffer_dsc *)buf->cache_entry->data;
const size_t buffer_size = (size_t)wd*ht*img->bpp + sizeof(*dsc);
// buf might have been alloc'ed before,
// so only check size and re-alloc if necessary:
if(!buf->buf || (dsc->size < buffer_size) || ((void *)dsc == (void *)dt_mipmap_cache_static_dead_image))
{
if((void *)dsc != (void *)dt_mipmap_cache_static_dead_image) dt_free_align(buf->cache_entry->data);
buf->cache_entry->data = dt_alloc_align(64, buffer_size);
if(!buf->cache_entry->data)
{
// return fallback: at least alloc size for a dead image:
buf->cache_entry->data = (void*)dt_mipmap_cache_static_dead_image;
// allocator holds the pointer. but let imageio client know that allocation failed:
return NULL;
}
// set buffer size only if we're making it larger.
dsc = (struct dt_mipmap_buffer_dsc *)buf->cache_entry->data;
dsc->size = buffer_size;
}
dsc->width = wd;
dsc->height = ht;
dsc->color_space = DT_COLORSPACE_NONE;
dsc->flags = DT_MIPMAP_BUFFER_DSC_FLAG_GENERATE;
buf->buf = (uint8_t *)(dsc + 1);
// fprintf(stderr, "full buffer allocating img %u %d x %d = %u bytes (%p)\n", img->id, img->width,
// img->height, buffer_size, *buf);
// return pointer to start of payload
return dsc + 1;
}
void local_laplacian_internal(
const float *const input, // input buffer in some Labx or yuvx format
float *const out, // output buffer with colour
const int wd, // width and
const int ht, // height of the input buffer
const float sigma, // user param: separate shadows/midtones/highlights
const float shadows, // user param: lift shadows
const float highlights, // user param: compress highlights
const float clarity, // user param: increase clarity/local contrast
const int use_sse2) // flag whether to use SSE version
{
#define max_levels 30
#define num_gamma 6
// don't divide by 2 more often than we can:
const int num_levels = MIN(max_levels, 31-__builtin_clz(MIN(wd,ht)));
const int max_supp = 1<<(num_levels-1);
int w, h;
float *padded[max_levels] = {0};
padded[0] = ll_pad_input(input, wd, ht, max_supp, &w, &h);
// allocate pyramid pointers for padded input
for(int l=1;l<num_levels;l++)
padded[l] = dt_alloc_align(16, sizeof(float)*dl(w,l)*dl(h,l));
// allocate pyramid pointers for output
float *output[max_levels] = {0};
for(int l=0;l<num_levels;l++)
output[l] = dt_alloc_align(16, sizeof(float)*dl(w,l)*dl(h,l));
// create gauss pyramid of padded input, write coarse directly to output
#if defined(__SSE2__)
if(use_sse2)
{
for(int l=1;l<num_levels-1;l++)
gauss_reduce_sse2(padded[l-1], padded[l], dl(w,l-1), dl(h,l-1));
gauss_reduce_sse2(padded[num_levels-2], output[num_levels-1], dl(w,num_levels-2), dl(h,num_levels-2));
}
else
#endif
{
for(int l=1;l<num_levels-1;l++)
gauss_reduce(padded[l-1], padded[l], dl(w,l-1), dl(h,l-1));
gauss_reduce(padded[num_levels-2], output[num_levels-1], dl(w,num_levels-2), dl(h,num_levels-2));
}
// evenly sample brightness [0,1]:
float gamma[num_gamma] = {0.0f};
for(int k=0;k<num_gamma;k++) gamma[k] = (k+.5f)/(float)num_gamma;
// for(int k=0;k<num_gamma;k++) gamma[k] = k/(num_gamma-1.0f);
// allocate memory for intermediate laplacian pyramids
float *buf[num_gamma][max_levels] = {{0}};
for(int k=0;k<num_gamma;k++) for(int l=0;l<num_levels;l++)
buf[k][l] = dt_alloc_align(16, sizeof(float)*dl(w,l)*dl(h,l));
// the paper says remapping only level 3 not 0 does the trick, too
// (but i really like the additional octave of sharpness we get,
// willing to pay the cost).
for(int k=0;k<num_gamma;k++)
{ // process images
#if defined(__SSE2__)
if(use_sse2)
apply_curve_sse2(buf[k][0], padded[0], w, h, max_supp, gamma[k], sigma, shadows, highlights, clarity);
else // brackets in next line needed for silly gcc warning:
#endif
{apply_curve(buf[k][0], padded[0], w, h, max_supp, gamma[k], sigma, shadows, highlights, clarity);}
// create gaussian pyramids
for(int l=1;l<num_levels;l++)
#if defined(__SSE2__)
if(use_sse2)
gauss_reduce_sse2(buf[k][l-1], buf[k][l], dl(w,l-1), dl(h,l-1));
else
#endif
gauss_reduce(buf[k][l-1], buf[k][l], dl(w,l-1), dl(h,l-1));
}
// assemble output pyramid coarse to fine
for(int l=num_levels-2;l >= 0; l--)
{
const int pw = dl(w,l), ph = dl(h,l);
gauss_expand(output[l+1], output[l], pw, ph);
// go through all coefficients in the upsampled gauss buffer:
#ifdef _OPENMP
#pragma omp parallel for default(none) schedule(static) collapse(2) shared(w,h,buf,output,l,gamma,padded)
#endif
for(int j=0;j<ph;j++) for(int i=0;i<pw;i++)
{
const float v = padded[l][j*pw+i];
int hi = 1;
for(;hi<num_gamma-1 && gamma[hi] <= v;hi++);
int lo = hi-1;
const float a = CLAMPS((v - gamma[lo])/(gamma[hi]-gamma[lo]), 0.0f, 1.0f);
const float l0 = ll_laplacian(buf[lo][l+1], buf[lo][l], i, j, pw, ph);
const float l1 = ll_laplacian(buf[hi][l+1], buf[hi][l], i, j, pw, ph);
output[l][j*pw+i] += l0 * (1.0f-a) + l1 * a;
// we could do this to save on memory (no need for finest buf[][]).
// unfortunately it results in a quite noticable loss of sharpness, i think
// the extra level is worth it.
// else if(l == 0) // use finest scale from input to not amplify noise (and use less memory)
// output[l][j*pw+i] += ll_laplacian(padded[l+1], padded[l], i, j, pw, ph);
}
}
#ifdef _OPENMP
#pragma omp parallel for default(none) schedule(dynamic) collapse(2) shared(w,output,buf)
#endif
for(int j=0;j<ht;j++) for(int i=0;i<wd;i++)
{
out[4*(j*wd+i)+0] = 100.0f * output[0][(j+max_supp)*w+max_supp+i]; // [0,1] -> L
out[4*(j*wd+i)+1] = input[4*(j*wd+i)+1]; // copy original colour channels
out[4*(j*wd+i)+2] = input[4*(j*wd+i)+2];
}
// free all buffers!
for(int l=0;l<max_levels;l++)
{
dt_free_align(padded[l]);
dt_free_align(output[l]);
for(int k = 0; k < num_gamma; k++) dt_free_align(buf[k][l]);
}
#undef num_levels
#undef num_gamma
}
static inline void gauss_reduce_sse2(
const float *const input, // fine input buffer
float *const coarse, // coarse scale, blurred input buf
const int wd, // fine res
const int ht)
{
// blur, store only coarse res
const int cw = (wd-1)/2+1, ch = (ht-1)/2+1;
// this version is inspired by opencv's pyrDown_ :
// - allocate 5 rows of ring buffer (aligned)
// - for coarse res y
// - fill 5 coarse-res row buffers with 1 4 6 4 1 weights (reuse some from last time)
// - do vertical convolution via sse and write to coarse output buf
const int stride = ((cw+8)&~7); // assure sse alignment of rows
float *ringbuf = dt_alloc_align(16, sizeof(*ringbuf)*stride*5);
float *rows[5] = {0};
int rowj = 0; // we initialised this many rows so far
for(int j=1;j<ch-1;j++)
{
// horizontal pass, convolve with 1 4 6 4 1 kernel and decimate
for(;rowj<=2*j+2;rowj++)
{
float *const row = ringbuf + (rowj % 5)*stride;
const float *const in = input + rowj*wd;
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none)
#endif
for(int i=1;i<cw-1;i++)
row[i] = 6*in[2*i] + 4*(in[2*i-1]+in[2*i+1]) + in[2*i-2] + in[2*i+2];
}
// init row pointers
for(int k=0;k<5;k++)
rows[k] = ringbuf + ((2*j-2+k)%5)*stride;
// vertical pass, convolve and decimate using SIMD:
// note that we're ignoring the (1..cw-1) buffer limit, we'll pull in
// garbage and fix it later by border filling.
float *const out = coarse + j*cw;
const float *const row0 = rows[0], *const row1 = rows[1],
*const row2 = rows[2], *const row3 = rows[3], *const row4 = rows[4];
const __m128 four = _mm_set1_ps(4.f), scale = _mm_set1_ps(1.f/256.f);
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none)
#endif
for(int i=0;i<=cw-8;i+=8)
{
__m128 r0, r1, r2, r3, r4, t0, t1;
r0 = _mm_load_ps(row0 + i);
r1 = _mm_load_ps(row1 + i);
r2 = _mm_load_ps(row2 + i);
r3 = _mm_load_ps(row3 + i);
r4 = _mm_load_ps(row4 + i);
r0 = _mm_add_ps(r0, r4);
r1 = _mm_add_ps(_mm_add_ps(r1, r3), r2);
r0 = _mm_add_ps(r0, _mm_add_ps(r2, r2));
t0 = _mm_add_ps(r0, _mm_mul_ps(r1, four));
r0 = _mm_load_ps(row0 + i + 4);
r1 = _mm_load_ps(row1 + i + 4);
r2 = _mm_load_ps(row2 + i + 4);
r3 = _mm_load_ps(row3 + i + 4);
r4 = _mm_load_ps(row4 + i + 4);
r0 = _mm_add_ps(r0, r4);
r1 = _mm_add_ps(_mm_add_ps(r1, r3), r2);
r0 = _mm_add_ps(r0, _mm_add_ps(r2, r2));
t1 = _mm_add_ps(r0, _mm_mul_ps(r1, four));
t0 = _mm_mul_ps(t0, scale);
t1 = _mm_mul_ps(t1, scale);
_mm_storeu_ps(out + i, t0);
_mm_storeu_ps(out + i + 4, t1);
}
// process the rest
for(int i=cw&~7;i<cw-1;i++)
out[i] = (6*row2[i] + 4*(row1[i] + row3[i]) + row0[i] + row4[i])*(1.0f/256.0f);
}
dt_free_align(ringbuf);
ll_fill_boundary1(coarse, cw, ch);
}
// internal function: to avoid exif blob reading + 8-bit byteorder flag + high-quality override
int dt_imageio_export_with_flags(
const uint32_t imgid,
const char *filename,
dt_imageio_module_format_t *format,
dt_imageio_module_data_t *format_params,
const int32_t ignore_exif,
const int32_t display_byteorder,
const gboolean high_quality,
const int32_t thumbnail_export,
const char *filter,
const gboolean copy_metadata,
dt_imageio_module_storage_t *storage,
dt_imageio_module_data_t *storage_params)
{
dt_develop_t dev;
dt_dev_init(&dev, 0);
dt_mipmap_buffer_t buf;
if(thumbnail_export && dt_conf_get_bool("plugins/lighttable/low_quality_thumbnails"))
dt_mipmap_cache_read_get(darktable.mipmap_cache, &buf, imgid, DT_MIPMAP_F, DT_MIPMAP_BLOCKING);
else
dt_mipmap_cache_read_get(darktable.mipmap_cache, &buf, imgid, DT_MIPMAP_FULL, DT_MIPMAP_BLOCKING);
dt_dev_load_image(&dev, imgid);
const dt_image_t *img = &dev.image_storage;
const int wd = img->width;
const int ht = img->height;
int res = 0;
dt_times_t start;
dt_get_times(&start);
dt_dev_pixelpipe_t pipe;
res = thumbnail_export ? dt_dev_pixelpipe_init_thumbnail(&pipe, wd, ht) : dt_dev_pixelpipe_init_export(&pipe, wd, ht, format->levels(format_params));
if(!res)
{
dt_control_log(_("failed to allocate memory for %s, please lower the threads used for export or buy more memory."), thumbnail_export ? C_("noun", "thumbnail export") : C_("noun", "export"));
dt_dev_cleanup(&dev);
dt_mipmap_cache_read_release(darktable.mipmap_cache, &buf);
return 1;
}
if(!buf.buf)
{
dt_control_log(_("image `%s' is not available!"), img->filename);
dt_mipmap_cache_read_release(darktable.mipmap_cache, &buf);
dt_dev_cleanup(&dev);
return 1;
}
// If a style is to be applied during export, add the iop params into the history
if (!thumbnail_export && format_params->style[0] != '\0')
{
GList *stls;
GList *modules = dev.iop;
dt_iop_module_t *m = NULL;
if ((stls=dt_styles_get_item_list(format_params->style, TRUE, -1)) == 0)
{
dt_control_log(_("cannot find the style '%s' to apply during export."), format_params->style);
dt_dev_cleanup(&dev);
dt_mipmap_cache_read_release(darktable.mipmap_cache, &buf);
return 1;
}
// Add each params
while (stls)
{
dt_style_item_t *s = (dt_style_item_t *) stls->data;
modules = dev.iop;
while (modules)
{
m = (dt_iop_module_t *)modules->data;
// since the name in the style is returned with a possible multi-name, just check the start of the name
if (strncmp(m->op, s->name, strlen(m->op)) == 0)
{
dt_dev_history_item_t *h = malloc(sizeof(dt_dev_history_item_t));
h->params = s->params;
h->blend_params = s->blendop_params;
h->enabled = s->enabled;
h->module = m;
h->multi_priority = 1;
g_strlcpy(h->multi_name, "", sizeof(h->multi_name));
if(m->legacy_params && (s->module_version != m->version()))
{
void *new_params = malloc(m->params_size);
m->legacy_params (m, h->params, s->module_version, new_params, labs(m->version()));
free (h->params);
h->params = new_params;
}
dev.history_end++;
dev.history = g_list_append(dev.history, h);
break;
}
modules = g_list_next(modules);
}
stls = g_list_next(stls);
}
}
dt_dev_pixelpipe_set_input(&pipe, &dev, (float *)buf.buf, buf.width, buf.height, 1.0);
dt_dev_pixelpipe_create_nodes(&pipe, &dev);
dt_dev_pixelpipe_synch_all(&pipe, &dev);
dt_dev_pixelpipe_get_dimensions(&pipe, &dev, pipe.iwidth, pipe.iheight, &pipe.processed_width, &pipe.processed_height);
if(filter)
{
if(!strncmp(filter, "pre:", 4))
dt_dev_pixelpipe_disable_after(&pipe, filter+4);
if(!strncmp(filter, "post:", 5))
dt_dev_pixelpipe_disable_before(&pipe, filter+5);
}
dt_show_times(&start, "[export] creating pixelpipe", NULL);
// find output color profile for this image:
int sRGB = 1;
gchar *overprofile = dt_conf_get_string("plugins/lighttable/export/iccprofile");
if(overprofile && !strcmp(overprofile, "sRGB"))
{
sRGB = 1;
}
else if(!overprofile || !strcmp(overprofile, "image"))
{
GList *modules = dev.iop;
dt_iop_module_t *colorout = NULL;
while (modules)
{
colorout = (dt_iop_module_t *)modules->data;
if(colorout->get_p && strcmp(colorout->op, "colorout") == 0)
{
const char *iccprofile = colorout->get_p(colorout->params, "iccprofile");
if(!strcmp(iccprofile, "sRGB")) sRGB = 1;
else sRGB = 0;
}
modules = g_list_next(modules);
}
}
else
{
sRGB = 0;
}
g_free(overprofile);
// get only once at the beginning, in case the user changes it on the way:
const gboolean high_quality_processing = ((format_params->max_width == 0 || format_params->max_width >= pipe.processed_width ) &&
(format_params->max_height == 0 || format_params->max_height >= pipe.processed_height)) ? FALSE :
high_quality;
const int width = high_quality_processing ? 0 : format_params->max_width;
const int height = high_quality_processing ? 0 : format_params->max_height;
const double scalex = width > 0 ? fminf(width /(double)pipe.processed_width, 1.0) : 1.0;
const double scaley = height > 0 ? fminf(height/(double)pipe.processed_height, 1.0) : 1.0;
const double scale = fminf(scalex, scaley);
int processed_width = scale*pipe.processed_width + .5f;
int processed_height = scale*pipe.processed_height + .5f;
const int bpp = format->bpp(format_params);
// downsampling done last, if high quality processing was requested:
uint8_t *outbuf = pipe.backbuf;
uint8_t *moutbuf = NULL; // keep track of alloc'ed memory
dt_get_times(&start);
if(high_quality_processing)
{
dt_dev_pixelpipe_process_no_gamma(&pipe, &dev, 0, 0, processed_width, processed_height, scale);
const double scalex = format_params->max_width > 0 ? fminf(format_params->max_width /(double)pipe.processed_width, 1.0) : 1.0;
const double scaley = format_params->max_height > 0 ? fminf(format_params->max_height/(double)pipe.processed_height, 1.0) : 1.0;
const double scale = fminf(scalex, scaley);
processed_width = scale*pipe.processed_width + .5f;
processed_height = scale*pipe.processed_height + .5f;
moutbuf = (uint8_t *)dt_alloc_align(64, (size_t)sizeof(float)*processed_width*processed_height*4);
outbuf = moutbuf;
// now downscale into the new buffer:
dt_iop_roi_t roi_in, roi_out;
roi_in.x = roi_in.y = roi_out.x = roi_out.y = 0;
roi_in.scale = 1.0;
roi_out.scale = scale;
roi_in.width = pipe.processed_width;
roi_in.height = pipe.processed_height;
roi_out.width = processed_width;
roi_out.height = processed_height;
dt_iop_clip_and_zoom((float *)outbuf, (float *)pipe.backbuf, &roi_out, &roi_in, processed_width, pipe.processed_width);
}
else
{
// do the processing (8-bit with special treatment, to make sure we can use openmp further down):
if(bpp == 8)
dt_dev_pixelpipe_process(&pipe, &dev, 0, 0, processed_width, processed_height, scale);
else
dt_dev_pixelpipe_process_no_gamma(&pipe, &dev, 0, 0, processed_width, processed_height, scale);
outbuf = pipe.backbuf;
}
dt_show_times(&start, thumbnail_export ? "[dev_process_thumbnail] pixel pipeline processing" : "[dev_process_export] pixel pipeline processing", NULL);
// downconversion to low-precision formats:
if(bpp == 8)
{
if(display_byteorder)
{
if(high_quality_processing)
{
const float *const inbuf = (float *)outbuf;
for(size_t k=0; k<(size_t)processed_width*processed_height; k++)
{
// convert in place, this is unfortunately very serial..
const uint8_t r = CLAMP(inbuf[4*k+2]*0xff, 0, 0xff);
const uint8_t g = CLAMP(inbuf[4*k+1]*0xff, 0, 0xff);
const uint8_t b = CLAMP(inbuf[4*k+0]*0xff, 0, 0xff);
outbuf[4*k+0] = r;
outbuf[4*k+1] = g;
outbuf[4*k+2] = b;
}
}
// else processing output was 8-bit already, and no need to swap order
}
else // need to flip
{
// ldr output: char
if(high_quality_processing)
{
const float *const inbuf = (float *)outbuf;
for(size_t k=0; k<(size_t)processed_width*processed_height; k++)
{
// convert in place, this is unfortunately very serial..
const uint8_t r = CLAMP(inbuf[4*k+0]*0xff, 0, 0xff);
const uint8_t g = CLAMP(inbuf[4*k+1]*0xff, 0, 0xff);
const uint8_t b = CLAMP(inbuf[4*k+2]*0xff, 0, 0xff);
outbuf[4*k+0] = r;
outbuf[4*k+1] = g;
outbuf[4*k+2] = b;
}
}
else
{ // !display_byteorder, need to swap:
uint8_t *const buf8 = pipe.backbuf;
#ifdef _OPENMP
#pragma omp parallel for default(none) shared(processed_width, processed_height) schedule(static)
#endif
// just flip byte order
for(size_t k=0; k<(size_t)processed_width*processed_height; k++)
{
uint8_t tmp = buf8[4*k+0];
buf8[4*k+0] = buf8[4*k+2];
buf8[4*k+2] = tmp;
}
}
}
}
else if(bpp == 16)
{
// uint16_t per color channel
float *buff = (float *) outbuf;
uint16_t *buf16 = (uint16_t *)outbuf;
for(int y=0; y<processed_height; y++) for(int x=0; x<processed_width ; x++)
{
// convert in place
const size_t k = (size_t)processed_width*y + x;
for(int i=0; i<3; i++) buf16[4*k+i] = CLAMP(buff[4*k+i]*0x10000, 0, 0xffff);
}
}
// else output float, no further harm done to the pixels :)
format_params->width = processed_width;
format_params->height = processed_height;
if(!ignore_exif)
{
int length;
uint8_t exif_profile[65535]; // C++ alloc'ed buffer is uncool, so we waste some bits here.
char pathname[PATH_MAX];
gboolean from_cache = TRUE;
dt_image_full_path(imgid, pathname, sizeof(pathname), &from_cache);
// last param is dng mode, it's false here
length = dt_exif_read_blob(exif_profile, pathname, imgid, sRGB, processed_width, processed_height, 0);
res = format->write_image (format_params, filename, outbuf, exif_profile, length, imgid);
}
else
{
res = format->write_image (format_params, filename, outbuf, NULL, 0, imgid);
}
dt_dev_pixelpipe_cleanup(&pipe);
dt_dev_cleanup(&dev);
dt_mipmap_cache_read_release(darktable.mipmap_cache, &buf);
dt_free_align(moutbuf);
/* now write xmp into that container, if possible */
if(copy_metadata && (format->flags(format_params) & FORMAT_FLAGS_SUPPORT_XMP)) {
dt_exif_xmp_attach(imgid, filename);
// no need to cancel the export if this fail
}
if(!thumbnail_export && strcmp(format->mime(format_params), "memory"))
{
dt_control_signal_raise(darktable.signals,DT_SIGNAL_IMAGE_EXPORT_TMPFILE,imgid,filename,format,format_params,storage,storage_params);
}
return res;
}
static int dt_group_get_mask_roi(dt_iop_module_t *module, dt_dev_pixelpipe_iop_t *piece,
dt_masks_form_t *form, const dt_iop_roi_t *roi, float *buffer)
{
double start2 = dt_get_wtime();
const guint nb = g_list_length(form->points);
if(nb == 0) return 0;
int nb_ok = 0;
const int width = roi->width;
const int height = roi->height;
// we need to allocate a temporary buffer for intermediate creation of individual shapes
float *bufs = dt_alloc_align(64, (size_t)width * height * sizeof(float));
if(bufs == NULL) return 0;
// empty the output buffer
memset(buffer, 0, (size_t)width * height * sizeof(float));
// and we get all masks
GList *fpts = g_list_first(form->points);
while(fpts)
{
dt_masks_point_group_t *fpt = (dt_masks_point_group_t *)fpts->data;
dt_masks_form_t *sel = dt_masks_get_from_id(module->dev, fpt->formid);
if(sel)
{
const int ok = dt_masks_get_mask_roi(module, piece, sel, roi, bufs);
const float op = fpt->opacity;
const int state = fpt->state;
if(ok)
{
// first see if we need to invert this shape
if(state & DT_MASKS_STATE_INVERSE)
{
#ifdef _OPENMP
#if !defined(__SUNOS__) && !defined(__NetBSD__)
#pragma omp parallel for default(none) shared(bufs)
#else
#pragma omp parallel for shared(bufs)
#endif
#endif
for(int y = 0; y < height; y++)
for(int x = 0; x < width; x++)
{
size_t index = (size_t)y * width + x;
bufs[index] = 1.0f - bufs[index];
}
}
if(state & DT_MASKS_STATE_UNION)
{
#ifdef _OPENMP
#if !defined(__SUNOS__) && !defined(__NetBSD__)
#pragma omp parallel for default(none) shared(bufs, buffer)
#else
#pragma omp parallel for shared(bufs, buffer)
#endif
#endif
for(int y = 0; y < height; y++)
for(int x = 0; x < width; x++)
{
size_t index = (size_t)y * width + x;
buffer[index] = fmaxf(buffer[index], bufs[index] * op);
}
}
else if(state & DT_MASKS_STATE_INTERSECTION)
{
#ifdef _OPENMP
#if !defined(__SUNOS__) && !defined(__NetBSD__)
#pragma omp parallel for default(none) shared(bufs, buffer)
#else
#pragma omp parallel for shared(bufs, buffer)
#endif
#endif
for(int y = 0; y < height; y++)
for(int x = 0; x < width; x++)
{
size_t index = (size_t)y * width + x;
float b1 = buffer[index];
float b2 = b2 = bufs[index]; // FIXME: is this line correct? what it supposed to be doing?
if(b1 > 0.0f && b2 > 0.0f)
buffer[index] = fminf(b1, b2 * op);
else
buffer[index] = 0.0f;
}
}
else if(state & DT_MASKS_STATE_DIFFERENCE)
{
#ifdef _OPENMP
#if !defined(__SUNOS__) && !defined(__NetBSD__)
#pragma omp parallel for default(none) shared(bufs, buffer)
#else
#pragma omp parallel for shared(bufs, buffer)
#endif
#endif
for(int y = 0; y < height; y++)
for(int x = 0; x < width; x++)
{
size_t index = (size_t)y * width + x;
float b1 = buffer[index];
float b2 = bufs[index] * op;
if(b1 > 0.0f && b2 > 0.0f) buffer[index] = b1 * (1.0f - b2);
}
}
else if(state & DT_MASKS_STATE_EXCLUSION)
{
#ifdef _OPENMP
#if !defined(__SUNOS__) && !defined(__NetBSD__)
#pragma omp parallel for default(none) shared(bufs, buffer)
#else
#pragma omp parallel for shared(bufs, buffer)
#endif
#endif
for(int y = 0; y < height; y++)
for(int x = 0; x < width; x++)
{
size_t index = (size_t)y * width + x;
float b1 = buffer[index];
float b2 = bufs[index] * op;
if(b1 > 0.0f && b2 > 0.0f)
buffer[index] = fmaxf((1.0f - b1) * b2, b1 * (1.0f - b2));
else
buffer[index] = fmaxf(b1, b2);
}
}
else // if we are here, this mean that we just have to copy the shape and null other parts
{
#ifdef _OPENMP
#if !defined(__SUNOS__) && !defined(__NetBSD__)
#pragma omp parallel for default(none) shared(bufs, buffer)
#else
#pragma omp parallel for shared(bufs, buffer)
#endif
#endif
for(int y = 0; y < height; y++)
for(int x = 0; x < width; x++)
{
size_t index = (size_t)y * width + x;
buffer[index] = bufs[index] * op;
}
}
if(darktable.unmuted & DT_DEBUG_PERF)
dt_print(DT_DEBUG_MASKS, "[masks %d] combine took %0.04f sec\n", nb_ok, dt_get_wtime() - start2);
start2 = dt_get_wtime();
nb_ok++;
}
}
fpts = g_list_next(fpts);
}
// and we free the intermediate buffer
dt_free_align(bufs);
return (nb_ok != 0);
}
dt_imageio_retval_t
dt_imageio_open_gm(
dt_image_t *img,
const char *filename,
dt_mipmap_cache_allocator_t a)
{
int err = DT_IMAGEIO_FILE_CORRUPTED;
float *buf = NULL;
ExceptionInfo exception;
Image *image = NULL;
ImageInfo *image_info = NULL;
uint32_t width, height, orientation;
if(!_supported_image(filename)) return DT_IMAGEIO_FILE_CORRUPTED;
if(!img->exif_inited)
(void) dt_exif_read(img, filename);
GetExceptionInfo(&exception);
image_info=CloneImageInfo((ImageInfo *) NULL);
g_strlcpy(image_info->filename,filename,sizeof(image_info->filename));
image=ReadImage(image_info,&exception);
if (exception.severity != UndefinedException)
CatchException(&exception);
if (!image)
{
fprintf(stderr, "[GraphicsMagick_open] image `%s' not found\n", img->filename);
err = DT_IMAGEIO_FILE_NOT_FOUND;
goto error;
}
width = image->columns;
height = image->rows;
orientation = image->orientation;
if(orientation & 4)
{
img->width = height;
img->height = width;
}
else
{
img->width = width;
img->height = height;
}
img->bpp = 4*sizeof(float);
float *mipbuf = (float *)dt_mipmap_cache_alloc(img, DT_MIPMAP_FULL, a);
if(!mipbuf)
{
fprintf(stderr, "[GraphicsMagick_open] could not alloc full buffer for image `%s'\n", img->filename);
err = DT_IMAGEIO_CACHE_FULL;
goto error;
}
buf = (float *)dt_alloc_align(16, width*img->bpp);
if(!buf) goto error;
const int ht2 = orientation & 4 ? img->width : img->height; // pretend unrotated, rotate in write_pos
const int wd2 = orientation & 4 ? img->height : img->width;
for (uint32_t row = 0; row < height; row++)
{
int ret = DispatchImage(image, 0, row, width, 1, "RGBP", FloatPixel, (void *)buf, &exception);
if (exception.severity != UndefinedException)
CatchException(&exception);
if(ret != MagickPass)
{
fprintf(stderr, "[GraphicsMagick_open] error reading image `%s'\n", img->filename);
err = DT_IMAGEIO_FILE_CORRUPTED;
goto error;
}
for(uint32_t i=0; i<width; i++)
for(int k=0; k<4; k++) mipbuf[4*dt_imageio_write_pos(i, row, wd2, ht2, wd2, ht2, orientation) + k] = buf[4*i + k];
}
if(buf) dt_free_align(buf);
if(image) DestroyImage(image);
if(image_info) DestroyImageInfo(image_info);
DestroyExceptionInfo(&exception);
img->filters = 0;
img->flags &= ~DT_IMAGE_RAW;
img->flags &= ~DT_IMAGE_HDR;
img->flags |= DT_IMAGE_LDR;
return DT_IMAGEIO_OK;
error:
if(buf) dt_free_align(buf);
if(image) DestroyImage(image);
if(image_info) DestroyImageInfo(image_info);
DestroyExceptionInfo(&exception);
return err;
}
// callback for the cache backend to initialize payload pointers
void dt_mipmap_cache_allocate_dynamic(void *data, dt_cache_entry_t *entry)
{
dt_mipmap_cache_t *cache = (dt_mipmap_cache_t *)data;
// for full image buffers
struct dt_mipmap_buffer_dsc *dsc = entry->data;
const dt_mipmap_size_t mip = get_size(entry->key);
// alloc mere minimum for the header + broken image buffer:
if(!dsc)
{
if(mip <= DT_MIPMAP_F)
{
// these are fixed-size:
entry->data = dt_alloc_align(16, cache->buffer_size[mip]);
}
else
{
entry->data = dt_alloc_align(16, sizeof(*dsc) + sizeof(float) * 4 * 64);
}
// fprintf(stderr, "[mipmap cache] alloc dynamic for key %u %p\n", key, *buf);
if(!(entry->data))
{
fprintf(stderr, "[mipmap cache] memory allocation failed!\n");
exit(1);
}
dsc = entry->data;
if(mip <= DT_MIPMAP_F)
{
dsc->width = cache->max_width[mip];
dsc->height = cache->max_height[mip];
dsc->size = cache->buffer_size[mip];
dsc->color_space = DT_COLORSPACE_NONE;
}
else
{
dsc->width = 0;
dsc->height = 0;
dsc->color_space = DT_COLORSPACE_NONE;
dsc->size = sizeof(*dsc) + sizeof(float) * 4 * 64;
}
}
assert(dsc->size >= sizeof(*dsc));
int loaded_from_disk = 0;
if(mip < DT_MIPMAP_F)
{
if(cache->cachedir[0] && dt_conf_get_bool("cache_disk_backend"))
{
// try and load from disk, if successful set flag
char filename[PATH_MAX] = {0};
snprintf(filename, sizeof(filename), "%s.d/%d/%d.jpg", cache->cachedir, mip, get_imgid(entry->key));
FILE *f = fopen(filename, "rb");
if(f)
{
long len = 0;
uint8_t *blob = 0;
fseek(f, 0, SEEK_END);
len = ftell(f);
if(len <= 0) goto read_error; // coverity madness
blob = (uint8_t *)malloc(len);
if(!blob) goto read_error;
fseek(f, 0, SEEK_SET);
int rd = fread(blob, sizeof(uint8_t), len, f);
if(rd != len) goto read_error;
dt_colorspaces_color_profile_type_t color_space;
dt_imageio_jpeg_t jpg;
if(dt_imageio_jpeg_decompress_header(blob, len, &jpg)
|| (jpg.width > cache->max_width[mip] || jpg.height > cache->max_height[mip])
|| ((color_space = dt_imageio_jpeg_read_color_space(&jpg)) == DT_COLORSPACE_NONE) // pointless test to keep it in the if clause
|| dt_imageio_jpeg_decompress(&jpg, entry->data + sizeof(*dsc)))
{
fprintf(stderr, "[mipmap_cache] failed to decompress thumbnail for image %d from `%s'!\n", get_imgid(entry->key), filename);
goto read_error;
}
dsc->width = jpg.width;
dsc->height = jpg.height;
dsc->color_space = color_space;
loaded_from_disk = 1;
if(0)
{
read_error:
g_unlink(filename);
}
free(blob);
fclose(f);
}
}
}
if(!loaded_from_disk)
dsc->flags = DT_MIPMAP_BUFFER_DSC_FLAG_GENERATE;
else dsc->flags = 0;
// cost is just flat one for the buffer, as the buffers might have different sizes,
// to make sure quota is meaningful.
if(mip >= DT_MIPMAP_F) entry->cost = 1;
else entry->cost = cache->buffer_size[mip];
}
// internal function: to avoid exif blob reading + 8-bit byteorder flag + high-quality override
int dt_imageio_export_with_flags(
const uint32_t imgid,
const char *filename,
dt_imageio_module_format_t *format,
dt_imageio_module_data_t *format_params,
const int32_t ignore_exif,
const int32_t display_byteorder,
const int32_t high_quality,
const int32_t thumbnail_export)
{
dt_develop_t dev;
dt_dev_init(&dev, 0);
dt_mipmap_buffer_t buf;
dt_mipmap_cache_read_get(darktable.mipmap_cache, &buf, imgid, DT_MIPMAP_FULL, DT_MIPMAP_BLOCKING);
dt_dev_load_image(&dev, imgid);
const dt_image_t *img = &dev.image_storage;
const int wd = img->width;
const int ht = img->height;
int res = 0;
dt_times_t start;
dt_get_times(&start);
dt_dev_pixelpipe_t pipe;
res = thumbnail_export ? dt_dev_pixelpipe_init_thumbnail(&pipe, wd, ht) : dt_dev_pixelpipe_init_export(&pipe, wd, ht);
if(!res)
{
dt_control_log(_("failed to allocate memory for export, please lower the threads used for export or buy more memory."));
dt_dev_cleanup(&dev);
if(buf.buf)
dt_mipmap_cache_read_release(darktable.mipmap_cache, &buf);
return 1;
}
if(!buf.buf)
{
dt_control_log(_("image `%s' is not available!"), img->filename);
dt_dev_cleanup(&dev);
return 1;
}
dt_dev_pixelpipe_set_input(&pipe, &dev, (float *)buf.buf, buf.width, buf.height, 1.0);
dt_dev_pixelpipe_create_nodes(&pipe, &dev);
dt_dev_pixelpipe_synch_all(&pipe, &dev);
dt_dev_pixelpipe_get_dimensions(&pipe, &dev, pipe.iwidth, pipe.iheight, &pipe.processed_width, &pipe.processed_height);
dt_show_times(&start, "[export] creating pixelpipe", NULL);
// find output color profile for this image:
int sRGB = 1;
gchar *overprofile = dt_conf_get_string("plugins/lighttable/export/iccprofile");
if(overprofile && !strcmp(overprofile, "sRGB"))
{
sRGB = 1;
}
else if(!overprofile || !strcmp(overprofile, "image"))
{
GList *modules = dev.iop;
dt_iop_module_t *colorout = NULL;
while (modules)
{
colorout = (dt_iop_module_t *)modules->data;
if (strcmp(colorout->op, "colorout") == 0)
{
dt_iop_colorout_params_t *p = (dt_iop_colorout_params_t *)colorout->params;
if(!strcmp(p->iccprofile, "sRGB")) sRGB = 1;
else sRGB = 0;
}
modules = g_list_next(modules);
}
}
else
{
sRGB = 0;
}
g_free(overprofile);
// get only once at the beginning, in case the user changes it on the way:
const int high_quality_processing = ((format_params->max_width == 0 || format_params->max_width >= pipe.processed_width ) &&
(format_params->max_height == 0 || format_params->max_height >= pipe.processed_height)) ? 0 :
high_quality;
const int width = high_quality_processing ? 0 : format_params->max_width;
const int height = high_quality_processing ? 0 : format_params->max_height;
const float scalex = width > 0 ? fminf(width /(float)pipe.processed_width, 1.0) : 1.0;
const float scaley = height > 0 ? fminf(height/(float)pipe.processed_height, 1.0) : 1.0;
const float scale = fminf(scalex, scaley);
int processed_width = scale*pipe.processed_width;
int processed_height = scale*pipe.processed_height;
const int bpp = format->bpp(format_params);
// downsampling done last, if high quality processing was requested:
uint8_t *outbuf = pipe.backbuf;
uint8_t *moutbuf = NULL; // keep track of alloc'ed memory
if(high_quality_processing)
{
dt_dev_pixelpipe_process_no_gamma(&pipe, &dev, 0, 0, processed_width, processed_height, scale);
const float scalex = format_params->max_width > 0 ? fminf(format_params->max_width /(float)pipe.processed_width, 1.0) : 1.0;
const float scaley = format_params->max_height > 0 ? fminf(format_params->max_height/(float)pipe.processed_height, 1.0) : 1.0;
const float scale = fminf(scalex, scaley);
processed_width = scale*pipe.processed_width + .5f;
processed_height = scale*pipe.processed_height + .5f;
moutbuf = (uint8_t *)dt_alloc_align(64, sizeof(float)*processed_width*processed_height*4);
outbuf = moutbuf;
// now downscale into the new buffer:
dt_iop_roi_t roi_in, roi_out;
roi_in.x = roi_in.y = roi_out.x = roi_out.y = 0;
roi_in.scale = 1.0;
roi_out.scale = scale;
roi_in.width = pipe.processed_width;
roi_in.height = pipe.processed_height;
roi_out.width = processed_width;
roi_out.height = processed_height;
dt_iop_clip_and_zoom((float *)outbuf, (float *)pipe.backbuf, &roi_out, &roi_in, processed_width, pipe.processed_width);
}
else
{
// do the processing (8-bit with special treatment, to make sure we can use openmp further down):
if(bpp == 8)
dt_dev_pixelpipe_process(&pipe, &dev, 0, 0, processed_width, processed_height, scale);
else
dt_dev_pixelpipe_process_no_gamma(&pipe, &dev, 0, 0, processed_width, processed_height, scale);
outbuf = pipe.backbuf;
}
// downconversion to low-precision formats:
if(bpp == 8 && !display_byteorder)
{
// ldr output: char
if(high_quality_processing)
{
const float *const inbuf = (float *)outbuf;
for(int k=0; k<processed_width*processed_height; k++)
{
// convert in place, this is unfortunately very serial..
const uint8_t r = CLAMP(inbuf[4*k+0]*0xff, 0, 0xff);
const uint8_t g = CLAMP(inbuf[4*k+1]*0xff, 0, 0xff);
const uint8_t b = CLAMP(inbuf[4*k+2]*0xff, 0, 0xff);
outbuf[4*k+0] = r;
outbuf[4*k+1] = g;
outbuf[4*k+2] = b;
}
}
else
{
uint8_t *const buf8 = pipe.backbuf;
#ifdef _OPENMP
#pragma omp parallel for default(none) shared(processed_width, processed_height) schedule(static)
#endif
// just flip byte order
for(int k=0; k<processed_width*processed_height; k++)
{
uint8_t tmp = buf8[4*k+0];
buf8[4*k+0] = buf8[4*k+2];
buf8[4*k+2] = tmp;
}
}
}
else if(bpp == 16)
{
// uint16_t per color channel
float *buff = (float *) outbuf;
uint16_t *buf16 = (uint16_t *)outbuf;
for(int y=0; y<processed_height; y++) for(int x=0; x<processed_width ; x++)
{
// convert in place
const int k = x + processed_width*y;
for(int i=0; i<3; i++) buf16[4*k+i] = CLAMP(buff[4*k+i]*0x10000, 0, 0xffff);
}
}
// else output float, no further harm done to the pixels :)
format_params->width = processed_width;
format_params->height = processed_height;
if(!ignore_exif)
{
int length;
uint8_t exif_profile[65535]; // C++ alloc'ed buffer is uncool, so we waste some bits here.
char pathname[1024];
dt_image_full_path(imgid, pathname, 1024);
length = dt_exif_read_blob(exif_profile, pathname, sRGB, imgid);
res = format->write_image (format_params, filename, outbuf, exif_profile, length, imgid);
}
else
{
res = format->write_image (format_params, filename, outbuf, NULL, 0, imgid);
}
dt_dev_pixelpipe_cleanup(&pipe);
dt_dev_cleanup(&dev);
dt_mipmap_cache_read_release(darktable.mipmap_cache, &buf);
free(moutbuf);
return res;
}
static void generate_thumbnail_cache()
{
const int max_mip = DT_MIPMAP_2;
fprintf(stderr, _("creating cache directories\n"));
char filename[PATH_MAX] = {0};
for(int k=DT_MIPMAP_0;k<=max_mip;k++)
{
snprintf(filename, sizeof(filename), "%s.d/%d", darktable.mipmap_cache->cachedir, k);
fprintf(stderr, _("creating cache directory '%s'\n"), filename);
int mkd = g_mkdir_with_parents(filename, 0750);
if(mkd)
{
fprintf(stderr, _("could not create directory '%s'!\n"), filename);
return;
}
}
// some progress counter
sqlite3_stmt *stmt;
uint64_t image_count = 0, counter = 0;
DT_DEBUG_SQLITE3_PREPARE_V2(dt_database_get(darktable.db), "select count(id) from images", -1, &stmt, 0);
if(sqlite3_step(stmt) == SQLITE_ROW)
image_count = sqlite3_column_int(stmt, 0);
sqlite3_finalize(stmt);
// go through all images:
DT_DEBUG_SQLITE3_PREPARE_V2(dt_database_get(darktable.db), "select id from images", -1, &stmt, 0);
// could only alloc max_mip-1, but would need to detect the special case that max==0.
const size_t bufsize = (size_t)4 * darktable.mipmap_cache->max_width[max_mip]
* darktable.mipmap_cache->max_height[max_mip];
uint8_t *tmp = (uint8_t *)dt_alloc_align(16, bufsize);
if(!tmp)
{
fprintf(stderr, "couldn't allocate temporary memory!\n");
sqlite3_finalize(stmt);
return;
}
const int cache_quality = MIN(100, MAX(10, dt_conf_get_int("database_cache_quality")));
while(sqlite3_step(stmt) == SQLITE_ROW)
{
const int32_t imgid = sqlite3_column_int(stmt, 0);
// check whether all of these files are already there
int all_exist = 1;
for(int k=max_mip;k>=DT_MIPMAP_0;k--)
{
snprintf(filename, sizeof(filename), "%s.d/%d/%d.jpg", darktable.mipmap_cache->cachedir, k, imgid);
all_exist &= !access(filename, R_OK);
}
if(all_exist) goto next;
dt_mipmap_buffer_t buf;
// get largest thumbnail for this image
// this one will take care of itself, we'll just write out the lower thumbs manually:
dt_mipmap_cache_get(darktable.mipmap_cache, &buf, imgid, max_mip, DT_MIPMAP_BLOCKING, 'r');
if(buf.width > 8 && buf.height > 8) // don't create for skulls
for(int k=max_mip-1;k>=DT_MIPMAP_0;k--)
{
uint32_t width, height;
const int wd = darktable.mipmap_cache->max_width[k];
const int ht = darktable.mipmap_cache->max_height[k];
// use exactly the same mechanism as the cache internally to rescale the thumbnail:
dt_iop_flip_and_zoom_8(buf.buf, buf.width, buf.height, tmp, wd, ht, 0, &width, &height);
snprintf(filename, sizeof(filename), "%s.d/%d/%d.jpg", darktable.mipmap_cache->cachedir, k, imgid);
FILE *f = fopen(filename, "wb");
if(f)
{
// allocate temp memory:
uint8_t *blob = (uint8_t *)malloc(bufsize);
if(!blob) goto write_error;
const int32_t length
= dt_imageio_jpeg_compress(tmp, blob, width, height, cache_quality);
assert(length <= bufsize);
int written = fwrite(blob, sizeof(uint8_t), length, f);
if(written != length)
{
write_error:
unlink(filename);
}
free(blob);
fclose(f);
}
}
dt_mipmap_cache_release(darktable.mipmap_cache, &buf);
next:
counter ++;
fprintf(stderr, "\rimage %lu/%lu (%.02f%%) ", counter, image_count, 100.0*counter/(float)image_count);
}
dt_free_align(tmp);
sqlite3_finalize(stmt);
fprintf(stderr, "done \n");
}
void process(struct dt_iop_module_t *self, dt_dev_pixelpipe_iop_t *piece, void *ivoid, void *ovoid,
const dt_iop_roi_t *roi_in, const dt_iop_roi_t *roi_out)
{
const dt_iop_colorout_data_t *const d = (dt_iop_colorout_data_t *)piece->data;
const int ch = piece->colors;
const int gamutcheck = (d->mode == DT_PROFILE_GAMUTCHECK);
if(!isnan(d->cmatrix[0]))
{
// fprintf(stderr,"Using cmatrix codepath\n");
// convert to rgb using matrix
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none) shared(roi_in, roi_out, ivoid, ovoid)
#endif
for(int j = 0; j < roi_out->height; j++)
{
float *in = (float *)ivoid + (size_t)ch * roi_in->width * j;
float *out = (float *)ovoid + (size_t)ch * roi_out->width * j;
const __m128 m0 = _mm_set_ps(0.0f, d->cmatrix[6], d->cmatrix[3], d->cmatrix[0]);
const __m128 m1 = _mm_set_ps(0.0f, d->cmatrix[7], d->cmatrix[4], d->cmatrix[1]);
const __m128 m2 = _mm_set_ps(0.0f, d->cmatrix[8], d->cmatrix[5], d->cmatrix[2]);
for(int i = 0; i < roi_out->width; i++, in += ch, out += ch)
{
const __m128 xyz = dt_Lab_to_XYZ_SSE(_mm_load_ps(in));
const __m128 t
= _mm_add_ps(_mm_mul_ps(m0, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(0, 0, 0, 0))),
_mm_add_ps(_mm_mul_ps(m1, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(1, 1, 1, 1))),
_mm_mul_ps(m2, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(2, 2, 2, 2)))));
_mm_stream_ps(out, t);
}
}
_mm_sfence();
// apply profile
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none) shared(roi_in, roi_out, ivoid, ovoid)
#endif
for(int j = 0; j < roi_out->height; j++)
{
float *in = (float *)ivoid + (size_t)ch * roi_in->width * j;
float *out = (float *)ovoid + (size_t)ch * roi_out->width * j;
for(int i = 0; i < roi_out->width; i++, in += ch, out += ch)
{
for(int i = 0; i < 3; i++)
if(d->lut[i][0] >= 0.0f)
{
out[i] = (out[i] < 1.0f) ? lerp_lut(d->lut[i], out[i])
: dt_iop_eval_exp(d->unbounded_coeffs[i], out[i]);
}
}
}
}
else
{
// fprintf(stderr,"Using xform codepath\n");
const __m128 outofgamutpixel = _mm_set_ps(0.0f, 1.0f, 1.0f, 0.0f);
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none) shared(ivoid, ovoid, roi_out)
#endif
for(int k = 0; k < roi_out->height; k++)
{
const float *in = ((float *)ivoid) + (size_t)ch * k * roi_out->width;
float *out = ((float *)ovoid) + (size_t)ch * k * roi_out->width;
if(!gamutcheck)
{
cmsDoTransform(d->xform, in, out, roi_out->width);
}
else
{
void *rgb = dt_alloc_align(16, 4 * sizeof(float) * roi_out->width);
cmsDoTransform(d->xform, in, rgb, roi_out->width);
float *rgbptr = (float *)rgb;
for(int j = 0; j < roi_out->width; j++, rgbptr += 4, out += 4)
{
const __m128 pixel = _mm_load_ps(rgbptr);
__m128 ingamut = _mm_cmplt_ps(pixel, _mm_set_ps(-FLT_MAX, 0.0f, 0.0f, 0.0f));
ingamut = _mm_or_ps(_mm_unpacklo_ps(ingamut, ingamut), _mm_unpackhi_ps(ingamut, ingamut));
ingamut = _mm_or_ps(_mm_unpacklo_ps(ingamut, ingamut), _mm_unpackhi_ps(ingamut, ingamut));
const __m128 result
= _mm_or_ps(_mm_and_ps(ingamut, outofgamutpixel), _mm_andnot_ps(ingamut, pixel));
_mm_stream_ps(out, result);
}
dt_free_align(rgb);
}
}
_mm_sfence();
}
if(piece->pipe->mask_display) dt_iop_alpha_copy(ivoid, ovoid, roi_out->width, roi_out->height);
}
/** Prepares a 1D resampling plan
*
* This consists of the following informations
* <ul>
* <li>A list of lengths that tell how many pixels are relevant for the
* next output</li>
* <li>A list of required filter kernels</li>
* <li>A list of sample indexes</li>
* </ul>
*
* How to apply the resampling plan:
* <ol>
* <li>Pick a length from the length array</li>
* <li>until length is reached
* <ol>
* <li>pick a kernel tap></li>
* <li>pick the relevant sample according to the picked index</li>
* <li>multiply them and accumulate</li>
* </ol>
* </li>
* <li>here goes a single output sample</li>
* </ol>
*
* This until you reach the number of output pixels
*
* @param itor interpolator used to resample
* @param in [in] Number of input samples
* @param out [in] Number of output samples
* @param plength [out] Array of lengths for each pixel filtering (number
* of taps/indexes to use). This array mus be freed with fre() when you're
* done with the plan.
* @param pkernel [out] Array of filter kernel taps
* @param pindex [out] Array of sample indexes to be used for applying each kernel tap
* arrays of informations
* @param pmeta [out] Array of int triplets (length, kernel, index) telling where to start for an arbitrary out position meta[3*out]
* @return 0 for success, !0 for failure
*/
static int
prepare_resampling_plan(
const struct dt_interpolation* itor,
int in,
const int in_x0,
int out,
const int out_x0,
float scale,
int** plength,
float** pkernel,
int** pindex,
int** pmeta)
{
// Safe return values
*plength = NULL;
*pkernel = NULL;
*pindex = NULL;
if (pmeta) {
*pmeta = NULL;
}
if (scale == 1.f) {
// No resampling required
return 0;
}
// Compute common upsampling/downsampling memory requirements
int maxtapsapixel;
if (scale > 1.f) {
// Upscale... the easy one. The values are exact
maxtapsapixel = 2*itor->width;
} else {
// Downscale... going for worst case values memory wise
maxtapsapixel = ceil_fast((float)2*(float)itor->width/scale);
}
int nlengths = out;
int nindex = maxtapsapixel*out;
int nkernel = maxtapsapixel*out;
size_t lengthreq = increase_for_alignment(nlengths*sizeof(int), SSE_ALIGNMENT);
size_t indexreq = increase_for_alignment(nindex*sizeof(int), SSE_ALIGNMENT);
size_t kernelreq = increase_for_alignment(nkernel*sizeof(float), SSE_ALIGNMENT);
size_t scratchreq = maxtapsapixel*sizeof(float) + 4*sizeof(float);
// NB: because sse versions compute four taps a time
size_t metareq = pmeta ? 3*sizeof(int)*out : 0;
void *blob = NULL;
size_t totalreq = kernelreq + lengthreq + indexreq + scratchreq + metareq;
blob = dt_alloc_align(SSE_ALIGNMENT, totalreq);
if (!blob) {
return 1;
}
int* lengths = (int*)blob;
blob = (char*)blob + lengthreq;
int* index = (int*)blob;
blob = (char*)blob + indexreq;
float* kernel = (float*)blob;
blob = (char*)blob + kernelreq;
float* scratchpad = scratchreq ? (float*)blob : NULL;
blob = (char*)blob + scratchreq;
int* meta = metareq ? (int*)blob : NULL;
blob = (char*)blob + metareq;
/* setting this as a const should help the compilers trim all unecessary
* codepaths */
const enum border_mode bordermode = RESAMPLING_BORDER_MODE;
/* Upscale and downscale differ in subtle points, getting rid of code
* duplication might have been tricky and i prefer keeping the code
* as straight as possible */
if (scale > 1.f) {
int kidx = 0;
int iidx = 0;
int lidx = 0;
int midx = 0;
for (int x=0; x<out; x++) {
if (meta) {
meta[midx++] = lidx;
meta[midx++] = kidx;
meta[midx++] = iidx;
}
// Projected position in input samples
float fx = (float)(out_x0 + x)/scale;
// Compute the filter kernel at that position
int first;
compute_upsampling_kernel_sse(itor, scratchpad, NULL, &first, fx);
/* Check lower and higher bound pixel index and skip as many pixels as
* necessary to fall into range */
int tap_first;
int tap_last;
prepare_tap_boundaries(&tap_first, &tap_last, bordermode, 2*itor->width, first, in);
// Track number of taps that will be used
lengths[lidx++] = tap_last - tap_first;
// Precompute the inverse of the norm
float norm = 0.f;
for (int tap=tap_first; tap<tap_last; tap++) {
norm += scratchpad[tap];
}
norm = 1.f/norm;
/* Unlike single pixel or single sample code, here it's interesting to
* precompute the normalized filter kernel as this will avoid dividing
* by the norm for all processed samples/pixels
* NB: use the same loop to put in place the index list */
first += tap_first;
for (int tap=tap_first; tap<tap_last; tap++) {
kernel[kidx++] = scratchpad[tap]*norm;
index[iidx++] = clip(first++, 0, in-1, bordermode);
}
}
} else {
int kidx = 0;
int iidx = 0;
int lidx = 0;
int midx = 0;
for (int x=0; x<out; x++) {
if (meta) {
meta[midx++] = lidx;
meta[midx++] = kidx;
meta[midx++] = iidx;
}
// Compute downsampling kernel centered on output position
int taps;
int first;
compute_downsampling_kernel_sse(itor, &taps, &first, scratchpad, NULL, scale, out_x0 + x);
/* Check lower and higher bound pixel index and skip as many pixels as
* necessary to fall into range */
int tap_first;
int tap_last;
prepare_tap_boundaries(&tap_first, &tap_last, bordermode, taps, first, in);
// Track number of taps that will be used
lengths[lidx++] = tap_last - tap_first;
// Precompute the inverse of the norm
float norm = 0.f;
for (int tap=tap_first; tap<tap_last; tap++) {
norm += scratchpad[tap];
}
norm = 1.f/norm;
/* Unlike single pixel or single sample code, here it's interesting to
* precompute the normalized filter kernel as this will avoid dividing
* by the norm for all processed samples/pixels
* NB: use the same loop to put in place the index list */
first += tap_first;
for (int tap=tap_first; tap<tap_last; tap++) {
kernel[kidx++] = scratchpad[tap]*norm;
index[iidx++] = clip(first++, 0, in-1, bordermode);
}
}
}
// Validate plan wrt caller
*plength = lengths;
*pindex = index;
*pkernel = kernel;
if (pmeta) {
*pmeta = meta;
}
return 0;
}
/* if a module does not implement process_tiling() by itself, this function is called instead.
default_process_tiling() is able to handle standard cases where pixels change their values
but not their places. */
void
default_process_tiling (struct dt_iop_module_t *self, struct dt_dev_pixelpipe_iop_t *piece, void *ivoid, void *ovoid, const dt_iop_roi_t *roi_in, const dt_iop_roi_t *roi_out, const int in_bpp)
{
void *input = NULL;
void *output = NULL;
/* we only care for the most simple cases ATM. else try to process the standard way, i.e. in one chunk. let's hope for the best... */
if(memcmp(roi_in, roi_out, sizeof(struct dt_iop_roi_t)))
{
dt_print(DT_DEBUG_DEV, "[default_process_tiling] cannot handle requested roi's. fall back to standard method for module '%s'\n", self->op);
goto fallback;
}
const int out_bpp = self->output_bpp(self, piece->pipe, piece);
const int ipitch = roi_in->width * in_bpp;
const int opitch = roi_out->width * out_bpp;
/* get tiling requirements of module */
dt_develop_tiling_t tiling = { 0 };
self->tiling_callback(self, piece, roi_in, roi_out, &tiling);
/* tiling really does not make sense in these cases. standard process() is not better or worse than we are */
if(tiling.factor < 2.2f && tiling.overhead < 0.2f * roi_out->width * roi_out->height * max(in_bpp, out_bpp))
{
dt_print(DT_DEBUG_DEV, "[default_process_tiling] don't use tiling for module '%s'. no real memory saving could be reached\n", self->op);
goto fallback;
}
/* calculate optimal size of tiles */
float available = dt_conf_get_int("host_memory_limit")*1024*1024;
assert(available >= 500*1024*1024);
/* correct for size of ivoid and ovoid which are needed on top of tiling */
available = max(available - roi_out->width * roi_out->height * (in_bpp + out_bpp) - tiling.overhead, 0);
/* we ignore the above value if singlebuffer_limit (is defined and) is higher than available/tiling.factor.
this will mainly allow tiling for modules with high and "unpredictable" memory demand which is
reflected in high values of tiling.factor (take bilateral noise reduction as an example). */
float singlebuffer = dt_conf_get_int("singlebuffer_limit")*1024*1024;
singlebuffer = max(singlebuffer, 1024*1024);
assert(tiling.factor > 1.0f);
singlebuffer = max(available / tiling.factor, singlebuffer);
int width = roi_out->width;
int height = roi_out->height;
/* shrink tile size in case it would exceed singlebuffer size */
if(width*height*max(in_bpp, out_bpp) > singlebuffer)
{
const float scale = singlebuffer/(width*height*max(in_bpp, out_bpp));
/* TODO: can we make this more efficient to minimize total overlap between tiles? */
if(width < height && scale >= 0.333f)
{
height = floorf(height * scale);
}
else if(height <= width && scale >= 0.333f)
{
width = floorf(width * scale);
}
else
{
width = floorf(width * sqrt(scale));
height = floorf(height * sqrt(scale));
}
}
/* make sure we have a reasonably effective tile dimension. if not try square tiles */
if(3*tiling.overlap > width || 3*tiling.overlap > height)
{
width = height = floorf(sqrtf((float)width*height));
}
#if 0
/* we might want to grow dimensions a bit */
width = max(4*tiling.overlap, width);
height = max(4*tiling.overlap, height);
#endif
/* Alignment rules: we need to make sure that alignment requirements of module are fulfilled.
Modules will report alignment requirements via xalign and yalign within tiling_callback().
Typical use case is demosaic where Bayer pattern requires alignment to a multiple of 2 in x and y
direction.
We guarantee alignment by selecting image width/height and overlap accordingly. For a tile width/height
that is identical to image width/height no special alignment is needed. */
const unsigned int xyalign = _lcm(tiling.xalign, tiling.yalign);
assert(xyalign != 0);
/* properly align tile width and height by making them smaller if needed */
if(width < roi_out->width) width = (width / xyalign) * xyalign;
if(height < roi_out->height) height = (height / xyalign) * xyalign;
/* also make sure that overlap follows alignment rules by making it wider when needed */
const int overlap = tiling.overlap % xyalign != 0 ? (tiling.overlap / xyalign + 1) * xyalign : tiling.overlap;
/* calculate effective tile size */
const int tile_wd = width - 2*overlap > 0 ? width - 2*overlap : 1;
const int tile_ht = height - 2*overlap > 0 ? height - 2*overlap : 1;
/* calculate number of tiles */
const int tiles_x = width < roi_out->width ? ceilf(roi_out->width /(float)tile_wd) : 1;
const int tiles_y = height < roi_out->height ? ceilf(roi_out->height/(float)tile_ht) : 1;
/* sanity check: don't run wild on too many tiles */
if(tiles_x * tiles_y > DT_TILING_MAXTILES)
{
dt_print(DT_DEBUG_DEV, "[default_process_tiling] gave up tiling for module '%s'. too many tiles: %d x %d\n", self->op, tiles_x, tiles_y);
goto error;
}
dt_print(DT_DEBUG_DEV, "[default_process_tiling] use tiling on module '%s' for image with full size %d x %d\n", self->op, roi_out->width, roi_out->height);
dt_print(DT_DEBUG_DEV, "[default_process_tiling] (%d x %d) tiles with max dimensions %d x %d and overlap %d\n", tiles_x, tiles_y, width, height, overlap);
/* reserve input and output buffers for tiles */
input = dt_alloc_align(64, width*height*in_bpp);
if(input == NULL)
{
dt_print(DT_DEBUG_DEV, "[default_process_tiling] could not alloc input buffer for module '%s'\n", self->op);
goto error;
}
output = dt_alloc_align(64, width*height*out_bpp);
if(output == NULL)
{
dt_print(DT_DEBUG_DEV, "[default_process_tiling] could not alloc output buffer for module '%s'\n", self->op);
goto error;
}
/* store processed_maximum to be re-used and aggregated */
float processed_maximum_saved[3];
float processed_maximum_new[3] = { 1.0f };
for(int k=0; k<3; k++)
processed_maximum_saved[k] = piece->pipe->processed_maximum[k];
/* iterate over tiles */
for(int tx=0; tx<tiles_x; tx++)
for(int ty=0; ty<tiles_y; ty++)
{
size_t wd = tx * tile_wd + width > roi_out->width ? roi_out->width - tx * tile_wd : width;
size_t ht = ty * tile_ht + height > roi_out->height ? roi_out->height- ty * tile_ht : height;
/* no need to process end-tiles that are smaller than overlap */
if((wd <= overlap && tx > 0) || (ht <= overlap && ty > 0)) continue;
/* origin and region of effective part of tile, which we want to store later */
size_t origin[] = { 0, 0, 0 };
size_t region[] = { wd, ht, 1 };
/* roi_in and roi_out for process_cl on subbuffer */
dt_iop_roi_t iroi = { 0, 0, wd, ht, roi_in->scale };
dt_iop_roi_t oroi = { 0, 0, wd, ht, roi_out->scale };
/* offsets of tile into ivoid and ovoid */
size_t ioffs = (ty * tile_ht)*ipitch + (tx * tile_wd)*in_bpp;
size_t ooffs = (ty * tile_ht)*opitch + (tx * tile_wd)*out_bpp;
dt_print(DT_DEBUG_DEV, "[default_process_tiling] tile (%d, %d) with %d x %d at origin [%d, %d]\n", tx, ty, wd, ht, tx*tile_wd, ty*tile_ht);
/* prepare input tile buffer */
#ifdef _OPENMP
#pragma omp parallel for default(none) shared(input,width,ivoid,ioffs,wd,ht) schedule(static)
#endif
for(int j=0; j<ht; j++)
memcpy((char *)input+j*wd*in_bpp, (char *)ivoid+ioffs+j*ipitch, wd*in_bpp);
/* take original processed_maximum as starting point */
for(int k=0; k<3; k++)
piece->pipe->processed_maximum[k] = processed_maximum_saved[k];
/* call process() of module */
self->process(self, piece, input, output, &iroi, &oroi);
/* aggregate resulting processed_maximum */
/* TODO: check if there really can be differences between tiles and take
appropriate action (calculate minimum, maximum, average, ...?) */
for(int k=0; k<3; k++)
{
if(tx+ty > 0 && fabs(processed_maximum_new[k] - piece->pipe->processed_maximum[k]) > 1.0e-6f)
dt_print(DT_DEBUG_DEV, "[default_process_tiling] processed_maximum[%d] differs between tiles in module '%s'\n", k, self->op);
processed_maximum_new[k] = piece->pipe->processed_maximum[k];
}
/* correct origin and region of tile for overlap.
make sure that we only copy back the "good" part. */
if(tx > 0)
{
origin[0] += overlap;
region[0] -= overlap;
ooffs += overlap*out_bpp;
}
if(ty > 0)
{
origin[1] += overlap;
region[1] -= overlap;
ooffs += overlap*opitch;
}
/* copy "good" part of tile to output buffer */
#ifdef _OPENMP
#pragma omp parallel for default(none) shared(ovoid,ooffs,output,width,origin,region,wd) schedule(static)
#endif
for(int j=0; j<region[1]; j++)
memcpy((char *)ovoid+ooffs+j*opitch, (char *)output+((j+origin[1])*wd+origin[0])*out_bpp, region[0]*out_bpp);
}
/* copy back final processed_maximum */
for(int k=0; k<3; k++)
piece->pipe->processed_maximum[k] = processed_maximum_new[k];
if(input != NULL) free(input);
if(output != NULL) free(output);
return;
error:
if(input != NULL) free(input);
if(output != NULL) free(output);
dt_print(DT_DEBUG_DEV, "[default_process_tiling] tiling failed for module '%s'\n", self->op);
/* TODO: give a warning message to user */
return;
fallback:
if(input != NULL) free(input);
if(output != NULL) free(output);
dt_print(DT_DEBUG_DEV, "[default_process_tiling] fall back to standard processing for module '%s'\n", self->op);
self->process(self, piece, ivoid, ovoid, roi_in, roi_out);
return;
}
dt_imageio_retval_t dt_imageio_open_png(dt_image_t *img, const char *filename, dt_mipmap_buffer_t *mbuf)
{
const char *ext = filename + strlen(filename);
while(*ext != '.' && ext > filename) ext--;
if(strncmp(ext, ".png", 4) && strncmp(ext, ".PNG", 4)) return DT_IMAGEIO_FILE_CORRUPTED;
if(!img->exif_inited) (void)dt_exif_read(img, filename);
dt_imageio_png_t image;
uint8_t *buf = NULL;
uint32_t width, height;
uint16_t bpp;
if(read_header(filename, &image) != 0) return DT_IMAGEIO_FILE_CORRUPTED;
width = img->width = image.width;
height = img->height = image.height;
bpp = image.bit_depth;
img->bpp = 4 * sizeof(float);
float *mipbuf = (float *)dt_mipmap_cache_alloc(mbuf, img);
if(!mipbuf)
{
fclose(image.f);
png_destroy_read_struct(&image.png_ptr, &image.info_ptr, NULL);
fprintf(stderr, "[png_open] could not alloc full buffer for image `%s'\n", img->filename);
return DT_IMAGEIO_CACHE_FULL;
}
buf = dt_alloc_align(16, (size_t)width * height * 3 * (bpp < 16 ? 1 : 2));
if(!buf)
{
fclose(image.f);
png_destroy_read_struct(&image.png_ptr, &image.info_ptr, NULL);
fprintf(stderr, "[png_open] could not alloc intermediate buffer for image `%s'\n", img->filename);
return DT_IMAGEIO_CACHE_FULL;
}
if(read_image(&image, (void *)buf) != 0)
{
dt_free_align(buf);
fprintf(stderr, "[png_open] could not read image `%s'\n", img->filename);
return DT_IMAGEIO_FILE_CORRUPTED;
}
for(size_t j = 0; j < height; j++)
{
if(bpp < 16)
for(size_t i = 0; i < width; i++)
for(int k = 0; k < 3; k++)
mipbuf[4 * (j * width + i) + k] = buf[3 * (j * width + i) + k] * (1.0f / 255.0f);
else
for(size_t i = 0; i < width; i++)
for(int k = 0; k < 3; k++)
mipbuf[4 * (j * width + i) + k] = (256.0f * buf[2 * (3 * (j * width + i) + k)]
+ buf[2 * (3 * (j * width + i) + k) + 1]) * (1.0f / 65535.0f);
}
dt_free_align(buf);
return DT_IMAGEIO_OK;
}
void dt_mipmap_cache_init(dt_mipmap_cache_t *cache)
{
// make sure static memory is initialized
struct dt_mipmap_buffer_dsc *dsc = (struct dt_mipmap_buffer_dsc *)dt_mipmap_cache_static_dead_image;
dead_image_f((dt_mipmap_buffer_t *)(dsc+1));
cache->compression_type = 0;
gchar *compression = dt_conf_get_string("cache_compression");
if(compression)
{
if(!strcmp(compression, "low quality (fast)"))
cache->compression_type = 1;
else if(!strcmp(compression, "high quality (slow)"))
cache->compression_type = 2;
g_free(compression);
}
dt_print(DT_DEBUG_CACHE, "[mipmap_cache_init] using %s\n", cache->compression_type == 0 ? "no compression" :
(cache->compression_type == 1 ? "low quality compression" : "slow high quality compression"));
// adjust numbers to be large enough to hold what mem limit suggests.
// we want at least 100MB, and consider 8G just still reasonable.
size_t max_mem = CLAMPS(dt_conf_get_int64("cache_memory"), 100u<<20, ((uint64_t)8)<<30);
const uint32_t parallel = CLAMP(dt_conf_get_int ("worker_threads")*dt_conf_get_int("parallel_export"), 1, 8);
const int32_t max_size = 2048, min_size = 32;
int32_t wd = darktable.thumbnail_width;
int32_t ht = darktable.thumbnail_height;
wd = CLAMPS(wd, min_size, max_size);
ht = CLAMPS(ht, min_size, max_size);
// round up to a multiple of 8, so we can divide by two 3 times
if(wd & 0xf) wd = (wd & ~0xf) + 0x10;
if(ht & 0xf) ht = (ht & ~0xf) + 0x10;
// cache these, can't change at runtime:
cache->mip[DT_MIPMAP_F].max_width = wd;
cache->mip[DT_MIPMAP_F].max_height = ht;
cache->mip[DT_MIPMAP_F-1].max_width = wd;
cache->mip[DT_MIPMAP_F-1].max_height = ht;
for(int k=DT_MIPMAP_F-2; k>=DT_MIPMAP_0; k--)
{
cache->mip[k].max_width = cache->mip[k+1].max_width / 2;
cache->mip[k].max_height = cache->mip[k+1].max_height / 2;
}
// initialize some per-thread cached scratchmem for uncompressed buffers during thumb creation:
if(cache->compression_type)
{
cache->scratchmem.max_width = wd;
cache->scratchmem.max_height = ht;
cache->scratchmem.buffer_size = wd*ht*sizeof(uint32_t);
cache->scratchmem.size = DT_MIPMAP_3; // at max.
// TODO: use thread local storage instead (zero performance penalty on linux)
dt_cache_init(&cache->scratchmem.cache, parallel, parallel, 64, 0.9f*parallel*wd*ht*sizeof(uint32_t));
// might have been rounded to power of two:
const int cnt = dt_cache_capacity(&cache->scratchmem.cache);
cache->scratchmem.buf = dt_alloc_align(64, cnt * wd*ht*sizeof(uint32_t));
dt_cache_static_allocation(&cache->scratchmem.cache, (uint8_t *)cache->scratchmem.buf, wd*ht*sizeof(uint32_t));
dt_cache_set_allocate_callback(&cache->scratchmem.cache,
scratchmem_allocate, &cache->scratchmem);
dt_print(DT_DEBUG_CACHE,
"[mipmap_cache_init] cache has % 5d entries for temporary compression buffers (% 4.02f MB).\n",
cnt, cnt* wd*ht*sizeof(uint32_t)/(1024.0*1024.0));
}
for(int k=DT_MIPMAP_3; k>=0; k--)
{
// clear stats:
cache->mip[k].stats_requests = 0;
cache->mip[k].stats_near_match = 0;
cache->mip[k].stats_misses = 0;
cache->mip[k].stats_fetches = 0;
cache->mip[k].stats_standin = 0;
// buffer stores width and height + actual data
const int width = cache->mip[k].max_width;
const int height = cache->mip[k].max_height;
// header + adjusted for dxt compression:
cache->mip[k].buffer_size = 4*sizeof(uint32_t) + compressed_buffer_size(cache->compression_type, width, height);
cache->mip[k].size = k;
// level of parallelism also gives minimum size (which is twice that)
// is rounded to a power of two by the cache anyways, we might as well.
// XXX this needs adjustment for video mode (more full-res thumbs for replay)
// TODO: collect hit/miss stats and auto-adjust to user browsing behaviour
// TODO: can #prefetches be collected this way, too?
const size_t max_mem2 = MAX(0, (k == 0) ? (max_mem) : (max_mem/(k+4)));
uint32_t thumbnails = MAX(2, nearest_power_of_two((uint32_t)((double)max_mem2/cache->mip[k].buffer_size)));
while(thumbnails > parallel && (size_t)thumbnails * cache->mip[k].buffer_size > max_mem2) thumbnails /= 2;
// try to utilize that memory well (use 90% quota), the hopscotch paper claims good scalability up to
// even more than that.
dt_cache_init(&cache->mip[k].cache, thumbnails,
parallel,
64, 0.9f*thumbnails*cache->mip[k].buffer_size);
// might have been rounded to power of two:
thumbnails = dt_cache_capacity(&cache->mip[k].cache);
max_mem -= thumbnails * cache->mip[k].buffer_size;
// dt_print(DT_DEBUG_CACHE, "[mipmap mem] %4.02f left\n", max_mem/(1024.0*1024.0));
cache->mip[k].buf = dt_alloc_align(64, thumbnails * cache->mip[k].buffer_size);
dt_cache_static_allocation(&cache->mip[k].cache, (uint8_t *)cache->mip[k].buf, cache->mip[k].buffer_size);
dt_cache_set_allocate_callback(&cache->mip[k].cache,
dt_mipmap_cache_allocate, &cache->mip[k]);
// dt_cache_set_cleanup_callback(&cache->mip[k].cache,
// &dt_mipmap_cache_deallocate, &cache->mip[k]);
dt_print(DT_DEBUG_CACHE,
"[mipmap_cache_init] cache has % 5d entries for mip %d (% 4.02f MB).\n",
thumbnails, k, thumbnails * cache->mip[k].buffer_size/(1024.0*1024.0));
}
// full buffer needs dynamic alloc:
const int full_entries = MAX(2, parallel); // even with one thread you want two buffers. one for dr one for thumbs.
int32_t max_mem_bufs = nearest_power_of_two(full_entries);
// for this buffer, because it can be very busy during import, we want the minimum
// number of entries in the hashtable to be 16, but leave the quota as is. the dynamic
// alloc/free properties of this cache take care that no more memory is required.
dt_cache_init(&cache->mip[DT_MIPMAP_FULL].cache, max_mem_bufs, parallel, 64, max_mem_bufs);
dt_cache_set_allocate_callback(&cache->mip[DT_MIPMAP_FULL].cache,
dt_mipmap_cache_allocate_dynamic, &cache->mip[DT_MIPMAP_FULL]);
// dt_cache_set_cleanup_callback(&cache->mip[DT_MIPMAP_FULL].cache,
// &dt_mipmap_cache_deallocate_dynamic, &cache->mip[DT_MIPMAP_FULL]);
cache->mip[DT_MIPMAP_FULL].buffer_size = 0;
cache->mip[DT_MIPMAP_FULL].size = DT_MIPMAP_FULL;
cache->mip[DT_MIPMAP_FULL].buf = NULL;
// same for mipf:
dt_cache_init(&cache->mip[DT_MIPMAP_F].cache, max_mem_bufs, parallel, 64, max_mem_bufs);
dt_cache_set_allocate_callback(&cache->mip[DT_MIPMAP_F].cache,
dt_mipmap_cache_allocate_dynamic, &cache->mip[DT_MIPMAP_F]);
dt_cache_set_cleanup_callback(&cache->mip[DT_MIPMAP_F].cache,
dt_mipmap_cache_deallocate_dynamic, &cache->mip[DT_MIPMAP_F]);
cache->mip[DT_MIPMAP_F].buffer_size = 4*sizeof(uint32_t) +
4*sizeof(float) * cache->mip[DT_MIPMAP_F].max_width * cache->mip[DT_MIPMAP_F].max_height;
cache->mip[DT_MIPMAP_F].size = DT_MIPMAP_F;
cache->mip[DT_MIPMAP_F].buf = NULL;
dt_mipmap_cache_deserialize(cache);
}
int process_cl(struct dt_iop_module_t *self, dt_dev_pixelpipe_iop_t *piece, cl_mem dev_in, cl_mem dev_out,
const dt_iop_roi_t *const roi_in, const dt_iop_roi_t *const roi_out)
{
dt_iop_global_tonemap_data_t *d = (dt_iop_global_tonemap_data_t *)piece->data;
dt_iop_global_tonemap_global_data_t *gd = (dt_iop_global_tonemap_global_data_t *)self->data;
dt_iop_global_tonemap_gui_data_t *g = (dt_iop_global_tonemap_gui_data_t *)self->gui_data;
dt_bilateral_cl_t *b = NULL;
cl_int err = -999;
cl_mem dev_m = NULL;
cl_mem dev_r = NULL;
float *maximum = NULL;
const int devid = piece->pipe->devid;
int gtkernel = -1;
const int width = roi_out->width;
const int height = roi_out->height;
float parameters[4] = { 0.0f };
switch(d->operator)
{
case OPERATOR_REINHARD:
gtkernel = gd->kernel_global_tonemap_reinhard;
break;
case OPERATOR_DRAGO:
gtkernel = gd->kernel_global_tonemap_drago;
break;
case OPERATOR_FILMIC:
gtkernel = gd->kernel_global_tonemap_filmic;
break;
}
if(d->operator== OPERATOR_DRAGO)
{
const float eps = 0.0001f;
float tmp_lwmax = NAN;
// see comments in process() about lwmax value
if(self->dev->gui_attached && g && piece->pipe->type == DT_DEV_PIXELPIPE_FULL)
{
dt_pthread_mutex_lock(&g->lock);
const uint64_t hash = g->hash;
dt_pthread_mutex_unlock(&g->lock);
if(hash != 0 && !dt_dev_sync_pixelpipe_hash(self->dev, piece->pipe, 0, self->priority, &g->lock, &g->hash))
dt_control_log(_("inconsistent output"));
dt_pthread_mutex_lock(&g->lock);
tmp_lwmax = g->lwmax;
dt_pthread_mutex_unlock(&g->lock);
}
if(isnan(tmp_lwmax))
{
dt_opencl_local_buffer_t flocopt
= (dt_opencl_local_buffer_t){ .xoffset = 0, .xfactor = 1, .yoffset = 0, .yfactor = 1,
.cellsize = sizeof(float), .overhead = 0,
.sizex = 1 << 4, .sizey = 1 << 4 };
if(!dt_opencl_local_buffer_opt(devid, gd->kernel_pixelmax_first, &flocopt))
goto error;
const size_t bwidth = ROUNDUP(width, flocopt.sizex);
const size_t bheight = ROUNDUP(height, flocopt.sizey);
const int bufsize = (bwidth / flocopt.sizex) * (bheight / flocopt.sizey);
dt_opencl_local_buffer_t slocopt
= (dt_opencl_local_buffer_t){ .xoffset = 0, .xfactor = 1, .yoffset = 0, .yfactor = 1,
.cellsize = sizeof(float), .overhead = 0,
.sizex = 1 << 16, .sizey = 1 };
if(!dt_opencl_local_buffer_opt(devid, gd->kernel_pixelmax_second, &slocopt))
goto error;
const int reducesize = MIN(REDUCESIZE, ROUNDUP(bufsize, slocopt.sizex) / slocopt.sizex);
size_t sizes[3];
size_t local[3];
dev_m = dt_opencl_alloc_device_buffer(devid, (size_t)bufsize * sizeof(float));
if(dev_m == NULL) goto error;
dev_r = dt_opencl_alloc_device_buffer(devid, (size_t)reducesize * sizeof(float));
if(dev_r == NULL) goto error;
sizes[0] = bwidth;
sizes[1] = bheight;
sizes[2] = 1;
local[0] = flocopt.sizex;
local[1] = flocopt.sizey;
local[2] = 1;
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_first, 0, sizeof(cl_mem), &dev_in);
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_first, 1, sizeof(int), &width);
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_first, 2, sizeof(int), &height);
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_first, 3, sizeof(cl_mem), &dev_m);
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_first, 4, flocopt.sizex * flocopt.sizey * sizeof(float), NULL);
err = dt_opencl_enqueue_kernel_2d_with_local(devid, gd->kernel_pixelmax_first, sizes, local);
if(err != CL_SUCCESS) goto error;
sizes[0] = reducesize * slocopt.sizex;
sizes[1] = 1;
sizes[2] = 1;
local[0] = slocopt.sizex;
local[1] = 1;
local[2] = 1;
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_second, 0, sizeof(cl_mem), &dev_m);
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_second, 1, sizeof(cl_mem), &dev_r);
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_second, 2, sizeof(int), &bufsize);
dt_opencl_set_kernel_arg(devid, gd->kernel_pixelmax_second, 3, slocopt.sizex * sizeof(float), NULL);
err = dt_opencl_enqueue_kernel_2d_with_local(devid, gd->kernel_pixelmax_second, sizes, local);
if(err != CL_SUCCESS) goto error;
maximum = dt_alloc_align(16, reducesize * sizeof(float));
err = dt_opencl_read_buffer_from_device(devid, (void *)maximum, dev_r, 0,
(size_t)reducesize * sizeof(float), CL_TRUE);
if(err != CL_SUCCESS) goto error;
dt_opencl_release_mem_object(dev_r);
dt_opencl_release_mem_object(dev_m);
dev_r = dev_m = NULL;
for(int k = 1; k < reducesize; k++)
{
float mine = maximum[0];
float other = maximum[k];
maximum[0] = (other > mine) ? other : mine;
}
tmp_lwmax = MAX(eps, (maximum[0] * 0.01f));
dt_free_align(maximum);
maximum = NULL;
}
const float lwmax = tmp_lwmax;
const float ldc = d->drago.max_light * 0.01f / log10f(lwmax + 1.0f);
const float bl = logf(MAX(eps, d->drago.bias)) / logf(0.5f);
parameters[0] = eps;
parameters[1] = ldc;
parameters[2] = bl;
parameters[3] = lwmax;
if(self->dev->gui_attached && g && piece->pipe->type == DT_DEV_PIXELPIPE_PREVIEW)
{
uint64_t hash = dt_dev_hash_plus(self->dev, piece->pipe, 0, self->priority);
dt_pthread_mutex_lock(&g->lock);
g->lwmax = lwmax;
g->hash = hash;
dt_pthread_mutex_unlock(&g->lock);
}
}
const float scale = piece->iscale / roi_in->scale;
const float sigma_r = 8.0f; // does not depend on scale
const float iw = piece->buf_in.width / scale;
const float ih = piece->buf_in.height / scale;
const float sigma_s = fminf(iw, ih) * 0.03f;
if(d->detail != 0.0f)
{
b = dt_bilateral_init_cl(devid, roi_in->width, roi_in->height, sigma_s, sigma_r);
if(!b) goto error;
// get detail from unchanged input buffer
err = dt_bilateral_splat_cl(b, dev_in);
if(err != CL_SUCCESS) goto error;
}
size_t sizes[2] = { ROUNDUPWD(width), ROUNDUPHT(height) };
dt_opencl_set_kernel_arg(devid, gtkernel, 0, sizeof(cl_mem), &dev_in);
dt_opencl_set_kernel_arg(devid, gtkernel, 1, sizeof(cl_mem), &dev_out);
dt_opencl_set_kernel_arg(devid, gtkernel, 2, sizeof(int), &width);
dt_opencl_set_kernel_arg(devid, gtkernel, 3, sizeof(int), &height);
dt_opencl_set_kernel_arg(devid, gtkernel, 4, 4 * sizeof(float), ¶meters);
err = dt_opencl_enqueue_kernel_2d(devid, gtkernel, sizes);
if(err != CL_SUCCESS) goto error;
if(d->detail != 0.0f)
{
err = dt_bilateral_blur_cl(b);
if(err != CL_SUCCESS) goto error;
// and apply it to output buffer after logscale
err = dt_bilateral_slice_to_output_cl(b, dev_in, dev_out, d->detail);
if(err != CL_SUCCESS) goto error;
dt_bilateral_free_cl(b);
}
return TRUE;
error:
if(b) dt_bilateral_free_cl(b);
dt_opencl_release_mem_object(dev_m);
dt_opencl_release_mem_object(dev_r);
dt_free_align(maximum);
dt_print(DT_DEBUG_OPENCL, "[opencl_global_tonemap] couldn't enqueue kernel! %d\n", err);
return FALSE;
}
#endif
void tiling_callback(struct dt_iop_module_t *self, struct dt_dev_pixelpipe_iop_t *piece,
const dt_iop_roi_t *roi_in, const dt_iop_roi_t *roi_out,
struct dt_develop_tiling_t *tiling)
{
dt_iop_global_tonemap_data_t *d = (dt_iop_global_tonemap_data_t *)piece->data;
const float scale = piece->iscale / roi_in->scale;
const float iw = piece->buf_in.width / scale;
const float ih = piece->buf_in.height / scale;
const float sigma_s = fminf(iw, ih) * 0.03f;
const float sigma_r = 8.0f;
const int detail = (d->detail != 0.0f);
const int width = roi_in->width;
const int height = roi_in->height;
const int channels = piece->colors;
const size_t basebuffer = width * height * channels * sizeof(float);
tiling->factor = 2.0f + (detail ? (float)dt_bilateral_memory_use2(width, height, sigma_s, sigma_r) / basebuffer : 0.0f);
tiling->maxbuf
= (detail ? MAX(1.0f, (float)dt_bilateral_singlebuffer_size2(width, height, sigma_s, sigma_r) / basebuffer) : 1.0f);
tiling->overhead = 0;
tiling->overlap = (detail ? ceilf(4 * sigma_s) : 0);
tiling->xalign = 1;
tiling->yalign = 1;
return;
}
void commit_params(struct dt_iop_module_t *self, dt_iop_params_t *p1, dt_dev_pixelpipe_t *pipe,
dt_dev_pixelpipe_iop_t *piece)
{
dt_iop_global_tonemap_params_t *p = (dt_iop_global_tonemap_params_t *)p1;
dt_iop_global_tonemap_data_t *d = (dt_iop_global_tonemap_data_t *)piece->data;
d->operator= p->operator;
d->drago.bias = p->drago.bias;
d->drago.max_light = p->drago.max_light;
d->detail = p->detail;
// drago needs the maximum L-value of the whole image so it must not use tiling
if(d->operator == OPERATOR_DRAGO) piece->process_tiling_ready = 0;
#ifdef HAVE_OPENCL
if(d->detail != 0.0f)
piece->process_cl_ready = (piece->process_cl_ready && !(darktable.opencl->avoid_atomics));
#endif
}
// if found, the data void* is returned. if not, it is set to be
// the given *data and a new hash table entry is created, which can be
// found using the given key later on.
dt_cache_entry_t *dt_cache_get_with_caller(dt_cache_t *cache, const uint32_t key, char mode, const char *file, int line)
{
gpointer orig_key, value;
gboolean res;
int result;
double start = dt_get_wtime();
restart:
dt_pthread_mutex_lock(&cache->lock);
res = g_hash_table_lookup_extended(
cache->hashtable, GINT_TO_POINTER(key), &orig_key, &value);
if(res)
{ // yay, found. read lock and pass on.
dt_cache_entry_t *entry = (dt_cache_entry_t *)value;
if(mode == 'w') result = dt_pthread_rwlock_trywrlock_with_caller(&entry->lock, file, line);
else result = dt_pthread_rwlock_tryrdlock_with_caller(&entry->lock, file, line);
if(result)
{ // need to give up mutex so other threads have a chance to get in between and
// free the lock we're trying to acquire:
dt_pthread_mutex_unlock(&cache->lock);
g_usleep(5);
goto restart;
}
// bubble up in lru list:
cache->lru = g_list_remove_link(cache->lru, entry->link);
cache->lru = g_list_concat(cache->lru, entry->link);
dt_pthread_mutex_unlock(&cache->lock);
#ifdef _DEBUG
const pthread_t writer = dt_pthread_rwlock_get_writer(&entry->lock);
if(mode == 'w')
{
assert(pthread_equal(writer, pthread_self()));
}
else
{
assert(!pthread_equal(writer, pthread_self()));
}
#endif
if(mode == 'w')
{
assert(entry->data_size);
ASAN_POISON_MEMORY_REGION(entry->data, entry->data_size);
}
// WARNING: do *NOT* unpoison here. it must be done by the caller!
return entry;
}
// else, not found, need to allocate.
// first try to clean up.
// also wait if we can't free more than the requested fill ratio.
if(cache->cost > 0.8f * cache->cost_quota)
{
// need to roll back all the way to get a consistent lock state:
dt_cache_gc(cache, 0.8f);
}
// here dies your 32-bit system:
dt_cache_entry_t *entry = (dt_cache_entry_t *)g_slice_alloc(sizeof(dt_cache_entry_t));
int ret = dt_pthread_rwlock_init(&entry->lock, 0);
if(ret) fprintf(stderr, "rwlock init: %d\n", ret);
entry->data = 0;
entry->data_size = cache->entry_size;
entry->cost = 1;
entry->link = g_list_append(0, entry);
entry->key = key;
entry->_lock_demoting = 0;
g_hash_table_insert(cache->hashtable, GINT_TO_POINTER(key), entry);
assert(cache->allocate || entry->data_size);
if(cache->allocate)
cache->allocate(cache->allocate_data, entry);
else
entry->data = dt_alloc_align(16, entry->data_size);
assert(entry->data_size);
ASAN_POISON_MEMORY_REGION(entry->data, entry->data_size);
// if allocate callback is given, always return a write lock
const int write = ((mode == 'w') || cache->allocate);
// write lock in case the caller requests it:
if(write) dt_pthread_rwlock_wrlock_with_caller(&entry->lock, file, line);
else dt_pthread_rwlock_rdlock_with_caller(&entry->lock, file, line);
cache->cost += entry->cost;
// put at end of lru list (most recently used):
cache->lru = g_list_concat(cache->lru, entry->link);
dt_pthread_mutex_unlock(&cache->lock);
double end = dt_get_wtime();
if(end - start > 0.1)
fprintf(stderr, "wait time %.06fs\n", end - start);
// WARNING: do *NOT* unpoison here. it must be done by the caller!
return entry;
}
/** process, all real work is done here. */
void process (struct dt_iop_module_t *self, dt_dev_pixelpipe_iop_t *piece, void *ivoid, void *ovoid, const dt_iop_roi_t *roi_in, const dt_iop_roi_t *roi_out)
{
// this is called for preview and full pipe separately, each with its own pixelpipe piece.
// get our data struct:
dt_iop_nlmeans_params_t *d = (dt_iop_nlmeans_params_t *)piece->data;
// adjust to zoom size:
const int P = ceilf(3 * roi_in->scale / piece->iscale); // pixel filter size
const int K = ceilf(7 * roi_in->scale / piece->iscale); // nbhood
if(P <= 1)
{
// nothing to do from this distance:
memcpy (ovoid, ivoid, sizeof(float)*4*roi_out->width*roi_out->height);
return;
}
// adjust to Lab, make L more important
// float max_L = 100.0f, max_C = 256.0f;
// float nL = 1.0f/(d->luma*max_L), nC = 1.0f/(d->chroma*max_C);
float max_L = 120.0f, max_C = 512.0f;
float nL = 1.0f/max_L, nC = 1.0f/max_C;
const float norm2[4] = { nL*nL, nC*nC, nC*nC, 1.0f };
float *Sa = dt_alloc_align(64, sizeof(float)*roi_out->width*dt_get_num_threads());
// we want to sum up weights in col[3], so need to init to 0:
memset(ovoid, 0x0, sizeof(float)*roi_out->width*roi_out->height*4);
// for each shift vector
for(int kj=-K;kj<=K;kj++)
{
for(int ki=-K;ki<=K;ki++)
{
int inited_slide = 0;
// don't construct summed area tables but use sliding window! (applies to cpu version res < 1k only, or else we will add up errors)
// do this in parallel with a little threading overhead. could parallelize the outer loops with a bit more memory
#ifdef _OPENMP
# pragma omp parallel for schedule(static) default(none) firstprivate(inited_slide) shared(kj, ki, roi_out, roi_in, ivoid, ovoid, Sa)
#endif
for(int j=0; j<roi_out->height; j++)
{
if(j+kj < 0 || j+kj >= roi_out->height) continue;
float *S = Sa + dt_get_thread_num() * roi_out->width;
const float *ins = ((float *)ivoid) + 4*(roi_in->width *(j+kj) + ki);
float *out = ((float *)ovoid) + 4*roi_out->width*j;
const int Pm = MIN(MIN(P, j+kj), j);
const int PM = MIN(MIN(P, roi_out->height-1-j-kj), roi_out->height-1-j);
// first line of every thread
// TODO: also every once in a while to assert numerical precision!
if(!inited_slide)
{
// sum up a line
memset(S, 0x0, sizeof(float)*roi_out->width);
for(int jj=-Pm;jj<=PM;jj++)
{
int i = MAX(0, -ki);
float *s = S + i;
const float *inp = ((float *)ivoid) + 4*i + 4* roi_in->width *(j+jj);
const float *inps = ((float *)ivoid) + 4*i + 4*(roi_in->width *(j+jj+kj) + ki);
const int last = roi_out->width + MIN(0, -ki);
for(; i<last; i++, inp+=4, inps+=4, s++)
{
for(int k=0;k<3;k++)
s[0] += (inp[k] - inps[k])*(inp[k] - inps[k]) * norm2[k];
}
}
// only reuse this if we had a full stripe
if(Pm == P && PM == P) inited_slide = 1;
}
// sliding window for this line:
float *s = S;
float slide = 0.0f;
// sum up the first -P..P
for(int i=0;i<2*P+1;i++) slide += s[i];
for(int i=0; i<roi_out->width; i++)
{
if(i-P > 0 && i+P<roi_out->width)
slide += s[P] - s[-P-1];
if(i+ki >= 0 && i+ki < roi_out->width)
{
const __m128 iv = { ins[0], ins[1], ins[2], 1.0f };
_mm_store_ps(out, _mm_load_ps(out) + iv * _mm_set1_ps(gh(slide)));
}
s ++;
ins += 4;
out += 4;
}
if(inited_slide && j+P+1+MAX(0,kj) < roi_out->height)
{
// sliding window in j direction:
int i = MAX(0, -ki);
float *s = S + i;
const float *inp = ((float *)ivoid) + 4*i + 4* roi_in->width *(j+P+1);
const float *inps = ((float *)ivoid) + 4*i + 4*(roi_in->width *(j+P+1+kj) + ki);
const float *inm = ((float *)ivoid) + 4*i + 4* roi_in->width *(j-P);
const float *inms = ((float *)ivoid) + 4*i + 4*(roi_in->width *(j-P+kj) + ki);
const int last = roi_out->width + MIN(0, -ki);
for(; ((unsigned long)s & 0xf) != 0 && i<last; i++, inp+=4, inps+=4, inm+=4, inms+=4, s++)
{
float stmp = s[0];
for(int k=0;k<3;k++)
stmp += ((inp[k] - inps[k])*(inp[k] - inps[k])
- (inm[k] - inms[k])*(inm[k] - inms[k])) * norm2[k];
s[0] = stmp;
}
/* Process most of the line 4 pixels at a time */
for(; i<last-4; i+=4, inp+=16, inps+=16, inm+=16, inms+=16, s+=4)
{
__m128 sv = _mm_load_ps(s);
const __m128 inp1 = _mm_load_ps(inp) - _mm_load_ps(inps);
const __m128 inp2 = _mm_load_ps(inp+4) - _mm_load_ps(inps+4);
const __m128 inp3 = _mm_load_ps(inp+8) - _mm_load_ps(inps+8);
const __m128 inp4 = _mm_load_ps(inp+12) - _mm_load_ps(inps+12);
const __m128 inp12lo = _mm_unpacklo_ps(inp1,inp2);
const __m128 inp34lo = _mm_unpacklo_ps(inp3,inp4);
const __m128 inp12hi = _mm_unpackhi_ps(inp1,inp2);
const __m128 inp34hi = _mm_unpackhi_ps(inp3,inp4);
const __m128 inpv0 = _mm_movelh_ps(inp12lo,inp34lo);
sv += inpv0*inpv0 * _mm_set1_ps(norm2[0]);
const __m128 inpv1 = _mm_movehl_ps(inp34lo,inp12lo);
sv += inpv1*inpv1 * _mm_set1_ps(norm2[1]);
const __m128 inpv2 = _mm_movelh_ps(inp12hi,inp34hi);
sv += inpv2*inpv2 * _mm_set1_ps(norm2[2]);
const __m128 inm1 = _mm_load_ps(inm) - _mm_load_ps(inms);
const __m128 inm2 = _mm_load_ps(inm+4) - _mm_load_ps(inms+4);
const __m128 inm3 = _mm_load_ps(inm+8) - _mm_load_ps(inms+8);
const __m128 inm4 = _mm_load_ps(inm+12) - _mm_load_ps(inms+12);
const __m128 inm12lo = _mm_unpacklo_ps(inm1,inm2);
const __m128 inm34lo = _mm_unpacklo_ps(inm3,inm4);
const __m128 inm12hi = _mm_unpackhi_ps(inm1,inm2);
const __m128 inm34hi = _mm_unpackhi_ps(inm3,inm4);
const __m128 inmv0 = _mm_movelh_ps(inm12lo,inm34lo);
sv -= inmv0*inmv0 * _mm_set1_ps(norm2[0]);
const __m128 inmv1 = _mm_movehl_ps(inm34lo,inm12lo);
sv -= inmv1*inmv1 * _mm_set1_ps(norm2[1]);
const __m128 inmv2 = _mm_movelh_ps(inm12hi,inm34hi);
sv -= inmv2*inmv2 * _mm_set1_ps(norm2[2]);
_mm_store_ps(s, sv);
}
for(; i<last; i++, inp+=4, inps+=4, inm+=4, inms+=4, s++)
{
float stmp = s[0];
for(int k=0;k<3;k++)
stmp += ((inp[k] - inps[k])*(inp[k] - inps[k])
- (inm[k] - inms[k])*(inm[k] - inms[k])) * norm2[k];
s[0] = stmp;
}
}
else inited_slide = 0;
}
}
}
// normalize and apply chroma/luma blending
// bias a bit towards higher values for low input values:
const __m128 weight = _mm_set_ps(1.0f, powf(d->chroma, 0.6), powf(d->chroma, 0.6), powf(d->luma, 0.6));
const __m128 invert = _mm_sub_ps(_mm_set1_ps(1.0f), weight);
#ifdef _OPENMP
#pragma omp parallel for default(none) schedule(static) shared(ovoid,ivoid,roi_out,d)
#endif
for(int j=0; j<roi_out->height; j++)
{
float *out = ((float *)ovoid) + 4*roi_out->width*j;
float *in = ((float *)ivoid) + 4*roi_out->width*j;
for(int i=0; i<roi_out->width; i++)
{
_mm_store_ps(out, _mm_add_ps(
_mm_mul_ps(_mm_load_ps(in), invert),
_mm_mul_ps(_mm_load_ps(out), _mm_div_ps(weight, _mm_set1_ps(out[3])))));
out += 4;
in += 4;
}
}
// free shared tmp memory:
free(Sa);
}