//////////////////////////////////
// Main entry function of the GPU cache model
//////////////////////////////////
int main(int argc, char** argv) {
srand(time(0));
std::cout << SPLIT_STRING << std::endl;
message("");
// Flush messages as soon as possible
std::cout.setf(std::ios_base::unitbuf);
// Read the hardware settings from file
Settings hardware = get_settings();
// Print cache statistics
message("Cache configuration:");
std::cout << "### \t Cache size: ~" << hardware.cache_bytes/1024 << "KB" << std::endl;
std::cout << "### \t Line size: " << hardware.line_size << " bytes" << std::endl;
std::cout << "### \t Layout: " << hardware.cache_ways << " ways, " << hardware.cache_sets << " sets" << std::endl;
message("");
// Parse the input argument and make sure that there is only one
if (argc != 3) {
message("Error: provide one argument only (a folder containing input trace files)");
message("");
std::cout << SPLIT_STRING << std::endl;
exit(1);
}
std::string benchname = argv[1];
std::string suitename = argv[2];
// Loop over all found traces in the folder (one trace per kernel)
for (unsigned kernel_id = 0; kernel_id < 20; kernel_id++) {
std::vector<Thread> threads(MAX_THREADS);
for (unsigned t=0; t<MAX_THREADS; t++) {
threads[t] = Thread();
}
// Set the kernelname and include a counter
std::string kernelname;
if (kernel_id < 10) {
kernelname = benchname+"_0"+std::to_string(kernel_id);
}
else {
kernelname = benchname+"_" +std::to_string(kernel_id);
}
// Load a memory access trace from a file
Dim3 blockdim = read_file(threads, kernelname, benchname, suitename);
unsigned blocksize = blockdim.x*blockdim.y*blockdim.z;
// There was not a single trace that could be found - exit with an error
if (blocksize == 0 && kernel_id == 0) {
std::cout << "### Error: could not read file 'output/" << benchname << "/" << kernelname << ".trc'" << std::endl;
message("");
std::cout << SPLIT_STRING << std::endl;
exit(1);
}
// The final tracefile is already processed, exit the loop
if (blocksize == 0) {
break;
}
// Assign threads to warps, threadblocks and GPU cores
message("");
std::cout << "### Assigning threads to warps/blocks/cores...";
unsigned num_blocks = ceil(threads.size()/(float)(blocksize));
unsigned num_warps_per_block = ceil(blocksize/(float)(hardware.warp_size));
std::vector<std::vector<unsigned>> warps(num_warps_per_block*num_blocks);
std::vector<std::vector<unsigned>> blocks(num_blocks);
std::vector<std::vector<unsigned>> cores(hardware.num_cores);
schedule_threads(threads, warps, blocks, cores, hardware, blocksize);
std::cout << "done" << std::endl;
// Model only a single core, modelling multiple cores requires a loop over 'cid'
unsigned cid = 0;
// Compute the number of active blocks on this core
unsigned hardware_max_active_blocks = std::min(hardware.max_active_threads/blocksize, hardware.max_active_blocks);
unsigned active_blocks = std::min((unsigned)cores[cid].size(), hardware_max_active_blocks);
// Start the computation of the reuse distance profile
message("");
std::cout << "### [core " << cid << "]:" << std::endl;
std::cout << "### Running " << active_blocks << " block(s) at a time" << std::endl;
std::cout << "### Calculating the reuse distances";
// Create a Gaussian distribution to model memory latencies
std::random_device random;
std::mt19937 gen(random());
// Compute the reuse distance for 4 different cases
std::vector<map_type<unsigned,unsigned>> distances(NUM_CASES);
for (unsigned runs = 0; runs < NUM_CASES; runs++) {
std::cout << "...";
unsigned sets, ways;
unsigned ml, ms, nml;
unsigned mshr;
// CASE 0 | Normal - full model
sets = hardware.cache_sets;
ways = hardware.cache_ways;
ml = hardware.mem_latency;
ms = hardware.mem_latency_stddev;
nml = NON_MEM_LATENCY;
mshr = hardware.num_mshr;
// CASE 1 | Only 1 set: don't model associativity
if (runs == 1) {
sets = 1;
ways = hardware.cache_ways*hardware.cache_sets;
}
// CASE 2 | Memory latency to 0: don't model latencies
if (runs == 2) {
ml = 0;
ms = 0;
nml = 0;
}
// CASE 3 | MSHR count to infinite: don't model MSHRs
if (runs == 3) {
mshr = INF;
}
// Calculate the reuse distance profile
std::normal_distribution<> distribution(0,ms);
reuse_distance(cores[cid], blocks, warps, threads, distances[runs], active_blocks, hardware,
sets, ways, ml, nml, mshr, gen, distribution);
}
std::cout << "done" << std::endl;
// Process the reuse distance profile to obtain the cache hit/miss rate
message("");
output_miss_rate(distances, kernelname, benchname,suitename, hardware);
// Display the cache hit/miss rate from the output of the verifier (if available)
message("");
verify_miss_rate(kernelname, benchname);
message("");
}
// End of the program
std::cout << SPLIT_STRING << std::endl;
return 0;
}
typename EMSubpixelCorrelatorView<ImagePixelT>::prerasterize_type
EMSubpixelCorrelatorView<ImagePixelT>::prerasterize(BBox2i const& bbox) const {
vw_out(InfoMessage, "stereo") << "EMSubpixelCorrelatorView: rasterizing image block " << bbox << ".\n";
// Find the range of disparity values for this patch.
// int num_good; // not used
BBox2i search_range;
try {
search_range = get_disparity_range(crop(m_course_disparity, bbox));
}
catch (const std::exception& e) {
search_range = BBox2i();
}
#ifdef USE_GRAPHICS
ImageWindow window;
if(debug_level >= 0) {
window = vw_create_window("disparity");
}
#endif
// The area in the right image that we'll be searching is
// determined by the bbox of the left image plus the search
// range.
BBox2i left_crop_bbox(bbox);
BBox2i right_crop_bbox(bbox.min() + search_range.min(),
bbox.max() + search_range.max());
// The correlator requires the images to be the same size. The
// search bbox will always be larger than the given left image
// bbox, so we just make the left bbox the same size as the
// right bbox.
left_crop_bbox.max() = left_crop_bbox.min() + Vector2i(right_crop_bbox.width(), right_crop_bbox.height());
// Finally, we must adjust both bounding boxes to account for
// the size of the kernel itself.
right_crop_bbox.min() -= Vector2i(m_kernel_size[0], m_kernel_size[1]);
right_crop_bbox.max() += Vector2i(m_kernel_size[0], m_kernel_size[1]);
left_crop_bbox.min() -= Vector2i(m_kernel_size[0], m_kernel_size[1]);
left_crop_bbox.max() += Vector2i(m_kernel_size[0], m_kernel_size[1]);
// We crop the images to the expanded bounding box and edge
// extend in case the new bbox extends past the image bounds.
ImageView<ImagePixelT> left_image_patch, right_image_patch;
ImageView<disparity_pixel> disparity_map_patch_in;
ImageView<result_type> disparity_map_patch_out;
left_image_patch = crop(edge_extend(m_left_image, ZeroEdgeExtension()),
left_crop_bbox);
right_image_patch = crop(edge_extend(m_right_image, ZeroEdgeExtension()),
right_crop_bbox);
disparity_map_patch_in = crop(edge_extend(m_course_disparity, ZeroEdgeExtension()),
left_crop_bbox);
disparity_map_patch_out.set_size(disparity_map_patch_in.cols(), disparity_map_patch_in.rows());
// Adjust the disparities to be relative to the cropped
// image pixel locations
for (int v = 0; v < disparity_map_patch_in.rows(); ++v) {
for (int u = 0; u < disparity_map_patch_in.cols(); ++u) {
if (disparity_map_patch_in(u,v).valid()) {
disparity_map_patch_in(u,v).child().x() -= search_range.min().x();
disparity_map_patch_in(u,v).child().y() -= search_range.min().y();
}
}
}
double blur_sigma_progressive = .5; // 3*sigma = 1.5 pixels
// create the pyramid first
std::vector<ImageView<ImagePixelT> > left_pyramid(pyramid_levels), right_pyramid(pyramid_levels);
std::vector<BBox2i> regions_of_interest(pyramid_levels);
std::vector<ImageView<Matrix2x2> > warps(pyramid_levels);
std::vector<ImageView<disparity_pixel> > disparity_map_pyramid(pyramid_levels);
// initialize the pyramid at level 0
left_pyramid[0] = channels_to_planes(left_image_patch);
right_pyramid[0] = channels_to_planes(right_image_patch);
disparity_map_pyramid[0] = disparity_map_patch_in;
regions_of_interest[0] = BBox2i(m_kernel_size[0], m_kernel_size[1],
bbox.width(),bbox.height());
// downsample the disparity map and the image pair to initialize the intermediate levels
for(int i = 1; i < pyramid_levels; i++) {
left_pyramid[i] = subsample(gaussian_filter(left_pyramid[i-1], blur_sigma_progressive), 2);
right_pyramid[i] = subsample(gaussian_filter(right_pyramid[i-1], blur_sigma_progressive), 2);
disparity_map_pyramid[i] = detail::subsample_disp_map_by_two(disparity_map_pyramid[i-1]);
regions_of_interest[i] = BBox2i(regions_of_interest[i-1].min()/2, regions_of_interest[i-1].max()/2);
}
// initialize warps at the lowest resolution level
warps[pyramid_levels-1].set_size(left_pyramid[pyramid_levels-1].cols(),
left_pyramid[pyramid_levels-1].rows());
for(int y = 0; y < warps[pyramid_levels-1].rows(); y++) {
for(int x = 0; x < warps[pyramid_levels-1].cols(); x++) {
warps[pyramid_levels-1](x, y).set_identity();
}
}
#ifdef USE_GRAPHICS
vw_initialize_graphics(0, NULL);
if(debug_level >= 0) {
for(int i = 0; i < pyramid_levels; i++) {
vw_show_image(window, left_pyramid[i]);
usleep((int)(.2*1000*1000));
}
}
#endif
// go up the pyramid; first run refinement, then upsample result for the next level
for(int i = pyramid_levels-1; i >=0; i--) {
vw_out() << "processing pyramid level "
<< i << " of " << pyramid_levels-1 << std::endl;
if(debug_level >= 0) {
std::stringstream stream;
stream << "pyramid_level_" << i << ".tif";
write_image(stream.str(), disparity_map_pyramid[i]);
}
ImageView<ImagePixelT> process_left_image = left_pyramid[i];
ImageView<ImagePixelT> process_right_image = right_pyramid[i];
if(i > 0) { // in this case take refine the upsampled disparity map from the previous level,
// and upsample for the next level
m_subpixel_refine(edge_extend(process_left_image, ZeroEdgeExtension()), edge_extend(process_right_image, ZeroEdgeExtension()),
disparity_map_pyramid[i], disparity_map_pyramid[i], warps[i],
regions_of_interest[i], false, debug_level == i);
// upsample the warps and the refined map for the next level of processing
int up_width = left_pyramid[i-1].cols();
int up_height = left_pyramid[i-1].rows();
warps[i-1] = copy(resize(warps[i], up_width , up_height, ConstantEdgeExtension(), NearestPixelInterpolation())); //upsample affine transforms
disparity_map_pyramid[i-1] = copy(detail::upsample_disp_map_by_two(disparity_map_pyramid[i], up_width, up_height));
}
else { // here there is no next level so we refine directly to the output patch
m_subpixel_refine(edge_extend(process_left_image, ZeroEdgeExtension()), edge_extend(process_right_image, ZeroEdgeExtension()),
disparity_map_pyramid[i], disparity_map_patch_out, warps[i],
regions_of_interest[i], true, debug_level == i);
}
}
#ifdef USE_GRAPHICS
if(debug_level >= 0) {
vw_show_image(window, .5 + select_plane(channels_to_planes(disparity_map_patch_out)/6., 0));
usleep(10*1000*1000);
}
#endif
// Undo the above adjustment
for (int v = 0; v < disparity_map_patch_out.rows(); ++v) {
for (int u = 0; u < disparity_map_patch_out.cols(); ++u) {
if (disparity_map_patch_out(u,v).valid()) {
disparity_map_patch_out(u,v).child().x() += search_range.min().x();
disparity_map_patch_out(u,v).child().y() += search_range.min().y();
}
}
}
#ifdef USE_GRAPHICS
if(debug_level >= 0 ) {
vw_destroy_window(window);
}
#endif
return crop(disparity_map_patch_out, BBox2i(m_kernel_size[0]-bbox.min().x(),
m_kernel_size[1]-bbox.min().y(),
m_left_image.cols(),
m_left_image.rows()));
}