stephenseo/blue_noise_generation
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Make usage of OpenCL conditional (configuring)

Browse source
This commit is contained in:
Stephen Seo 2023-03-31 18:46:10 +09:00
parent 2f768f43bd
commit 62d15771ad
2 changed files with 269 additions and 187 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
CMakeLists.txt
View file

@ -18,11 +18,28 @@ if(NOT DEFINED CMAKE_BUILD_TYPE OR NOT CMAKE_BUILD_TYPE)
endif()
find_package(Threads REQUIRED)
find_package(OpenCL REQUIRED)
if(NOT DEFINED DISABLE_OPENCL)
find_package(OpenCL QUIET)
if(NOT DEFINED OpenCL_FOUND)
set(DISABLE_OPENCL True)
message(WARNING "OpenCL not found, OpenCL usage is disabled.")
endif()
endif()
find_package(PNG REQUIRED)
add_executable(Dithering ${Dithering_SOURCES})
target_compile_features(Dithering PUBLIC cxx_std_17)
if(DEFINED DISABLE_OPENCL AND DISABLE_OPENCL)
message(STATUS "OpenCL usage is disabled.")
target_include_directories(Dithering PUBLIC
Threads::Threads
${PNG_INCLUDE_DIRS})
target_link_libraries(Dithering PUBLIC
Threads::Threads
${PNG_LIBRARIES})
target_compile_definitions(Dithering PRIVATE DITHERING_OPENCL_ENABLED=0)
else()
message(STATUS "OpenCL usage is enabled.")
target_include_directories(Dithering PUBLIC
Threads::Threads
${OpenCL_INCLUDE_DIRS}
@ -31,3 +48,5 @@ target_link_libraries(Dithering PUBLIC
Threads::Threads
${OpenCL_LIBRARIES}
${PNG_LIBRARIES})
target_compile_definitions(Dithering PRIVATE DITHERING_OPENCL_ENABLED=1)
endif()

185
src/blue_noise.cpp
View file

@ -1,23 +1,25 @@
#include "blue_noise.hpp"
#include <random>
#include <cassert>
#include <iostream>
#include <cstdio>
#include <fstream>
#include <iostream>
#include <memory>
#include <random>
#include <string>
#include <unordered_set>
#include <cstdio>
#if DITHERING_OPENCL_ENABLED == 1
#include <CL/opencl.h>
#endif
#include "image.hpp"
image::Bl dither::blue_noise(int width, int height, int threads, bool use_opencl) {
image::Bl dither::blue_noise(int width, int height, int threads,
bool use_opencl) {
bool using_opencl = false;
#if DITHERING_OPENCL_ENABLED == 1
if (use_opencl) {
// try to use OpenCL
do {
@ -36,47 +38,57 @@ image::Bl dither::blue_noise(int width, int height, int threads, bool use_opencl
break;
}
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &device, nullptr);
err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &device,
nullptr);
if (err != CL_SUCCESS) {
std::cerr << "OpenCL: Failed to get a device\n";
break;
}
context = clCreateContext(nullptr, 1, &device, nullptr, nullptr, &err);
context = clCreateContext(nullptr, 1, &device, nullptr, nullptr,
&err);
{
char buf[1024];
std::ifstream program_file("src/blue_noise.cl");
if (!program_file.good()) {
std::cerr << "ERROR: Failed to read \"src/blue_noise.cl\" (not found?)\n";
std::cerr << "ERROR: Failed to read \"src/blue_noise.cl\" "
"(not found?)\n";
break;
}
std::string program_string;
while (program_file.good()) {
program_file.read(buf, 1024);
if(int read_count = program_file.gcount(); read_count > 0) {
if (int read_count = program_file.gcount();
read_count > 0) {
program_string.append(buf, read_count);
}
}
const char *string_ptr = program_string.c_str();
std::size_t program_size = program_string.size();
program = clCreateProgramWithSource(context, 1, (const char**)&string_ptr, &program_size, &err);
program = clCreateProgramWithSource(context, 1,
(const char **)&string_ptr,
&program_size, &err);
if (err != CL_SUCCESS) {
std::cerr << "OpenCL: Failed to create the program\n";
clReleaseContext(context);
break;
}
err = clBuildProgram(program, 1, &device, nullptr, nullptr, nullptr);
err = clBuildProgram(program, 1, &device, nullptr, nullptr,
nullptr);
if (err != CL_SUCCESS) {
std::cerr << "OpenCL: Failed to build the program\n";
std::size_t log_size;
clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, 0, nullptr, &log_size);
std::unique_ptr<char[]> log = std::make_unique<char[]>(log_size + 1);
clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG,
0, nullptr, &log_size);
std::unique_ptr<char[]> log =
std::make_unique<char[]>(log_size + 1);
log[log_size] = 0;
clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG, log_size, log.get(), nullptr);
clGetProgramBuildInfo(program, device, CL_PROGRAM_BUILD_LOG,
log_size, log.get(), nullptr);
std::cerr << log.get() << std::endl;
clReleaseProgram(program);
@ -98,18 +110,25 @@ image::Bl dither::blue_noise(int width, int height, int threads, bool use_opencl
std::cout << "ERROR: Empty result\n";
} while (false);
}
#else
std::clog << "WARNING: Not compiled with OpenCL support!\n";
#endif
if (!using_opencl) {
std::cout << "OpenCL: Failed to setup/use or is not enabled, using regular impl..."
std::cout << "OpenCL: Failed to setup/use or is not enabled, using "
"regular impl..."
<< std::endl;
return internal::rangeToBl(internal::blue_noise_impl(width, height, threads), width);
return internal::rangeToBl(
internal::blue_noise_impl(width, height, threads), width);
}
std::cout << "ERROR: Invalid state (end of blue_noise fn)\n";
return {};
}
std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int height, int threads) {
std::vector<unsigned int> dither::internal::blue_noise_impl(int width,
int height,
int threads) {
int count = width * height;
std::vector<float> filter_out;
filter_out.resize(count);
@ -119,13 +138,15 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
pbp.resize(count);
#ifndef NDEBUG
printf("Inserting %d pixels into image of max count %d\n", pixel_count, count);
printf("Inserting %d pixels into image of max count %d\n", pixel_count,
count);
// generate image from randomized pbp
FILE *random_noise_image = fopen("random_noise.pbm", "w");
fprintf(random_noise_image, "P1\n%d %d\n", width, height);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
fprintf(random_noise_image, "%d ", pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
fprintf(random_noise_image, "%d ",
pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
}
fputc('\n', random_noise_image);
}
@ -138,10 +159,12 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
int filter_size = (width + height) / 2;
std::unique_ptr<std::vector<float>> precomputed = std::make_unique<std::vector<float>>(internal::precompute_gaussian(filter_size));
std::unique_ptr<std::vector<float>> precomputed =
std::make_unique<std::vector<float>>(
internal::precompute_gaussian(filter_size));
internal::compute_filter(pbp, width, height, count, filter_size,
filter_out, precomputed.get(), threads);
internal::compute_filter(pbp, width, height, count, filter_size, filter_out,
precomputed.get(), threads);
#ifndef NDEBUG
internal::write_filter(filter_out, width, "filter_out_start.pgm");
#endif
@ -176,7 +199,8 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
// get second buffer's min
int second_min;
std::tie(second_min, std::ignore) = internal::filter_minmax(filter_out, pbp);
std::tie(second_min, std::ignore) =
internal::filter_minmax(filter_out, pbp);
if (second_min == max) {
pbp[max] = true;
@ -192,7 +216,9 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
fprintf(blue_noise_image, "P1\n%d %d\n", width, height);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
fprintf(blue_noise_image, "%d ", pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
fprintf(blue_noise_image, "%d ",
pbp[utility::twoToOne(x, y, width, height)] ? 1
: 0);
}
fputc('\n', blue_noise_image);
}
@ -200,8 +226,8 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
#endif
}
}
internal::compute_filter(pbp, width, height, count, filter_size,
filter_out, precomputed.get(), threads);
internal::compute_filter(pbp, width, height, count, filter_size, filter_out,
precomputed.get(), threads);
#ifndef NDEBUG
internal::write_filter(filter_out, width, "filter_out_final.pgm");
#endif
@ -212,7 +238,8 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
fprintf(blue_noise_image, "P1\n%d %d\n", width, height);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
fprintf(blue_noise_image, "%d ", pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
fprintf(blue_noise_image, "%d ",
pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
}
fputc('\n', blue_noise_image);
}
@ -231,14 +258,16 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
#endif
internal::compute_filter(pbp, width, height, count, filter_size,
filter_out, precomputed.get(), threads);
std::tie(std::ignore, max) = internal::filter_minmax(filter_out, pbp);
std::tie(std::ignore, max) =
internal::filter_minmax(filter_out, pbp);
pbp[max] = false;
dither_array[max] = i;
}
pbp = pbp_copy;
}
std::cout << "\nRanking remainder of first half of pixels...\n";
for (unsigned int i = pixel_count; i < (unsigned int)((count + 1) / 2); ++i) {
for (unsigned int i = pixel_count; i < (unsigned int)((count + 1) / 2);
++i) {
#ifndef NDEBUG
std::cout << i << ' ';
#endif
@ -257,8 +286,9 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
for (unsigned int i = 0; i < pbp.size(); ++i) {
reversed_pbp[i] = !pbp[i];
}
internal::compute_filter(reversed_pbp, width, height, count, filter_size,
filter_out, precomputed.get(), threads);
internal::compute_filter(reversed_pbp, width, height, count,
filter_size, filter_out, precomputed.get(),
threads);
std::tie(std::ignore, max) = internal::filter_minmax(filter_out, pbp);
pbp[max] = true;
dither_array[max] = i;
@ -267,8 +297,10 @@ std::vector<unsigned int> dither::internal::blue_noise_impl(int width, int heigh
return dither_array;
}
#if DITHERING_OPENCL_ENABLED == 1
std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
const int width, const int height, const int filter_size, cl_context context, cl_device_id device, cl_program program) {
const int width, const int height, const int filter_size,
cl_context context, cl_device_id device, cl_program program) {
cl_int err;
cl_kernel kernel;
cl_command_queue queue;
@ -284,11 +316,17 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
queue = clCreateCommandQueueWithProperties(context, device, nullptr, &err);
d_filter_out = clCreateBuffer(context, CL_MEM_WRITE_ONLY, count * sizeof(float), nullptr, nullptr);
d_precomputed = clCreateBuffer(context, CL_MEM_READ_ONLY, precomputed.size() * sizeof(float), nullptr, nullptr);
d_pbp = clCreateBuffer(context, CL_MEM_READ_ONLY, count * sizeof(int), nullptr, nullptr);
d_filter_out = clCreateBuffer(context, CL_MEM_WRITE_ONLY,
count * sizeof(float), nullptr, nullptr);
d_precomputed = clCreateBuffer(context, CL_MEM_READ_ONLY,
precomputed.size() * sizeof(float), nullptr,
nullptr);
d_pbp = clCreateBuffer(context, CL_MEM_READ_ONLY, count * sizeof(int),
nullptr, nullptr);
err = clEnqueueWriteBuffer(queue, d_precomputed, CL_TRUE, 0, precomputed.size() * sizeof(float), &precomputed[0], 0, nullptr, nullptr);
err = clEnqueueWriteBuffer(queue, d_precomputed, CL_TRUE, 0,
precomputed.size() * sizeof(float),
&precomputed[0], 0, nullptr, nullptr);
if (err != CL_SUCCESS) {
std::cerr << "OpenCL: Failed to write to d_precomputed buffer\n";
clReleaseMemObject(d_pbp);
@ -334,7 +372,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
return {};
}
if(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_filter_out) != CL_SUCCESS) {
if (clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_filter_out) !=
CL_SUCCESS) {
std::cerr << "OpenCL: Failed to set kernel arg 0\n";
clReleaseKernel(kernel);
clReleaseMemObject(d_pbp);
@ -343,7 +382,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
clReleaseCommandQueue(queue);
return {};
}
if(clSetKernelArg(kernel, 1, sizeof(cl_mem), &d_precomputed) != CL_SUCCESS) {
if (clSetKernelArg(kernel, 1, sizeof(cl_mem), &d_precomputed) !=
CL_SUCCESS) {
std::cerr << "OpenCL: Failed to set kernel arg 1\n";
clReleaseKernel(kernel);
clReleaseMemObject(d_pbp);
@ -381,7 +421,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
}
if (filter_size % 2 == 0) {
int filter_size_odd = filter_size + 1;
if(clSetKernelArg(kernel, 5, sizeof(int), &filter_size_odd) != CL_SUCCESS) {
if (clSetKernelArg(kernel, 5, sizeof(int), &filter_size_odd) !=
CL_SUCCESS) {
std::cerr << "OpenCL: Failed to set kernel arg 4\n";
clReleaseKernel(kernel);
clReleaseMemObject(d_pbp);
@ -391,7 +432,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
return {};
}
} else {
if(clSetKernelArg(kernel, 5, sizeof(int), &filter_size) != CL_SUCCESS) {
if (clSetKernelArg(kernel, 5, sizeof(int), &filter_size) !=
CL_SUCCESS) {
std::cerr << "OpenCL: Failed to set kernel arg 4\n";
clReleaseKernel(kernel);
clReleaseMemObject(d_pbp);
@ -402,7 +444,9 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
}
}
if(clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE, sizeof(std::size_t), &local_size, nullptr) != CL_SUCCESS) {
if (clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE,
sizeof(std::size_t), &local_size,
nullptr) != CL_SUCCESS) {
std::cerr << "OpenCL: Failed to get work group size\n";
clReleaseKernel(kernel);
clReleaseMemObject(d_pbp);
@ -411,17 +455,19 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
clReleaseCommandQueue(queue);
return {};
}
global_size = (std::size_t)std::ceil(count / (float)local_size) * local_size;
global_size =
(std::size_t)std::ceil(count / (float)local_size) * local_size;
std::cout << "OpenCL: global = " << global_size << ", local = " << local_size
<< std::endl;
std::cout << "OpenCL: global = " << global_size
<< ", local = " << local_size << std::endl;
std::vector<float> filter(count);
bool reversed_pbp = false;
const auto get_filter = [&queue, &kernel, &global_size, &local_size,
&d_filter_out, &d_pbp, &pbp, &pbp_i, &count, &filter, &err, &reversed_pbp] () -> bool {
&d_filter_out, &d_pbp, &pbp, &pbp_i, &count,
&filter, &err, &reversed_pbp]() -> bool {
for (unsigned int i = 0; i < pbp.size(); ++i) {
if (reversed_pbp) {
pbp_i[i] = pbp[i] ? 0 : 1;
@ -429,14 +475,17 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
pbp_i[i] = pbp[i] ? 1 : 0;
}
}
if(clEnqueueWriteBuffer(queue, d_pbp, CL_TRUE, 0, count * sizeof(int), &pbp_i[0], 0, nullptr, nullptr) != CL_SUCCESS) {
if (clEnqueueWriteBuffer(queue, d_pbp, CL_TRUE, 0, count * sizeof(int),
&pbp_i[0], 0, nullptr,
nullptr) != CL_SUCCESS) {
std::cerr << "OpenCL: Failed to write to d_pbp buffer\n";
return false;
}
if(err = clEnqueueNDRangeKernel(
queue, kernel, 1, nullptr, &global_size, &local_size,
0, nullptr, nullptr); err != CL_SUCCESS) {
if (err = clEnqueueNDRangeKernel(queue, kernel, 1, nullptr,
&global_size, &local_size, 0, nullptr,
nullptr);
err != CL_SUCCESS) {
std::cerr << "OpenCL: Failed to enqueue task: ";
switch (err) {
case CL_INVALID_PROGRAM_EXECUTABLE:
@ -481,20 +530,24 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
clFinish(queue);
clEnqueueReadBuffer(queue, d_filter_out, CL_TRUE, 0, count * sizeof(float), &filter[0], 0, nullptr, nullptr);
clEnqueueReadBuffer(queue, d_filter_out, CL_TRUE, 0,
count * sizeof(float), &filter[0], 0, nullptr,
nullptr);
return true;
};
{
#ifndef NDEBUG
printf("Inserting %d pixels into image of max count %d\n", pixel_count, count);
printf("Inserting %d pixels into image of max count %d\n", pixel_count,
count);
// generate image from randomized pbp
FILE *random_noise_image = fopen("random_noise.pbm", "w");
fprintf(random_noise_image, "P1\n%d %d\n", width, height);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
fprintf(random_noise_image, "%d ", pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
fprintf(random_noise_image, "%d ",
pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
}
fputc('\n', random_noise_image);
}
@ -541,7 +594,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
// get second buffer's min
int second_min;
std::tie(second_min, std::ignore) = internal::filter_minmax(filter, pbp);
std::tie(second_min, std::ignore) =
internal::filter_minmax(filter, pbp);
if (second_min == max) {
pbp[max] = true;
@ -559,7 +613,9 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
fprintf(blue_noise_image, "P1\n%d %d\n", width, height);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
fprintf(blue_noise_image, "%d ", pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
fprintf(blue_noise_image, "%d ",
pbp[utility::twoToOne(x, y, width, height)] ? 1
: 0);
}
fputc('\n', blue_noise_image);
}
@ -577,7 +633,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
fprintf(blue_noise_image, "P1\n%d %d\n", width, height);
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
fprintf(blue_noise_image, "%d ", pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
fprintf(blue_noise_image, "%d ",
pbp[utility::twoToOne(x, y, width, height)] ? 1 : 0);
}
fputc('\n', blue_noise_image);
}
@ -588,7 +645,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
#ifndef NDEBUG
{
image::Bl pbp_image = toBl(pbp, width);
pbp_image.writeToFile(image::file_type::PNG, true, "debug_pbp_before.png");
pbp_image.writeToFile(image::file_type::PNG, true,
"debug_pbp_before.png");
}
#endif
@ -620,11 +678,13 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
pbp = pbp_copy;
#ifndef NDEBUG
image::Bl min_pixels = internal::rangeToBl(dither_array, width);
min_pixels.writeToFile(image::file_type::PNG, true, "da_min_pixels.png");
min_pixels.writeToFile(image::file_type::PNG, true,
"da_min_pixels.png");
#endif
}
std::cout << "\nRanking remainder of first half of pixels...\n";
for (unsigned int i = pixel_count; i < (unsigned int)((count + 1) / 2); ++i) {
for (unsigned int i = pixel_count; i < (unsigned int)((count + 1) / 2);
++i) {
#ifndef NDEBUG
std::cout << i << ' ';
#endif
@ -643,7 +703,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
#ifndef NDEBUG
{
image::Bl min_pixels = internal::rangeToBl(dither_array, width);
min_pixels.writeToFile(image::file_type::PNG, true, "da_mid_pixels.png");
min_pixels.writeToFile(image::file_type::PNG, true,
"da_mid_pixels.png");
get_filter();
internal::write_filter(filter, width, "filter_mid.pgm");
image::Bl pbp_image = toBl(pbp, width);
@ -675,7 +736,8 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
get_filter();
internal::write_filter(filter, width, "filter_after.pgm");
image::Bl pbp_image = toBl(pbp, width);
pbp_image.writeToFile(image::file_type::PNG, true, "debug_pbp_after.png");
pbp_image.writeToFile(image::file_type::PNG, true,
"debug_pbp_after.png");
}
#endif
@ -686,3 +748,4 @@ std::vector<unsigned int> dither::internal::blue_noise_cl_impl(
clReleaseCommandQueue(queue);
return dither_array;
}
#endif