Cache vulkan buffers, remove staging buffers

At this point, runtime is almost the same as the OpenCL implementation.
This commit is contained in:
Stephen Seo 2024-03-26 13:18:24 +09:00
parent 3a97f14199
commit 6aaef83ca9
2 changed files with 62 additions and 108 deletions

View file

@ -186,11 +186,11 @@ void dither::internal::vulkan_invalidate_buffer(VkDevice device,
} }
std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl( std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
VkDevice device, VkPhysicalDevice phys_device, VkDevice device, VkCommandBuffer command_buffer, VkQueue queue,
VkCommandBuffer command_buffer, VkCommandPool command_pool, VkQueue queue, VkPipeline pipeline, VkPipelineLayout pipeline_layout,
VkBuffer pbp_buf, VkPipeline pipeline, VkPipelineLayout pipeline_layout, VkDescriptorSet descriptor_set, float *filter_out_mapped, int *pbp_mapped,
VkDescriptorSet descriptor_set, VkBuffer filter_out_buf, const int width, VkDeviceMemory filter_out_buf_mem, VkDeviceMemory pbp_buf_mem,
const int height) { const int width, const int height) {
const int size = width * height; const int size = width * height;
const int pixel_count = size * 4 / 10; const int pixel_count = size * 4 / 10;
const int local_size = 256; const int local_size = 256;
@ -200,85 +200,21 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
std::vector<bool> pbp = random_noise(size, pixel_count); std::vector<bool> pbp = random_noise(size, pixel_count);
bool reversed_pbp = false; bool reversed_pbp = false;
VkBuffer staging_pbp_buffer; const auto get_filter = [device, command_buffer, queue, pipeline,
VkDeviceMemory staging_pbp_buffer_mem; pipeline_layout, descriptor_set, &pbp, &reversed_pbp,
void *pbp_mapped; global_size, filter_out_buf_mem, pbp_buf_mem,
if (!internal::vulkan_create_buffer(device, phys_device, size * sizeof(int), pbp_mapped]() -> bool {
VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
staging_pbp_buffer,
staging_pbp_buffer_mem)) {
std::clog << "get_filter ERROR: Failed to create staging pbp buffer!\n";
return {};
}
utility::Cleanup cleanup_staging_pbp_buf(
[device](void *ptr) {
vkDestroyBuffer(device, *((VkBuffer *)ptr), nullptr);
},
&staging_pbp_buffer);
utility::Cleanup cleanup_staging_pbp_buf_mem(
[device](void *ptr) {
vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
},
&staging_pbp_buffer_mem);
vkMapMemory(device, staging_pbp_buffer_mem, 0, size * sizeof(int), 0,
&pbp_mapped);
utility::Cleanup cleanup_pbp_mapped(
[device](void *ptr) { vkUnmapMemory(device, *((VkDeviceMemory *)ptr)); },
&staging_pbp_buffer_mem);
int *pbp_mapped_int = (int *)pbp_mapped;
VkBuffer staging_filter_buffer;
VkDeviceMemory staging_filter_buffer_mem;
void *filter_mapped;
if (!internal::vulkan_create_buffer(device, phys_device, size * sizeof(int),
VK_BUFFER_USAGE_TRANSFER_DST_BIT,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
staging_filter_buffer,
staging_filter_buffer_mem)) {
std::clog << "get_filter ERROR: Failed to create staging pbp buffer!\n";
return {};
}
utility::Cleanup cleanup_staging_filter_buf(
[device](void *ptr) {
vkDestroyBuffer(device, *((VkBuffer *)ptr), nullptr);
},
&staging_filter_buffer);
utility::Cleanup cleanup_staging_filter_buf_mem(
[device](void *ptr) {
vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
},
&staging_filter_buffer_mem);
vkMapMemory(device, staging_filter_buffer_mem, 0, size * sizeof(float), 0,
&filter_mapped);
utility::Cleanup cleanup_filter_mapped(
[device](void *ptr) { vkUnmapMemory(device, *((VkDeviceMemory *)ptr)); },
&staging_filter_buffer_mem);
float *filter_mapped_float = (float *)filter_mapped;
const auto get_filter = [device, command_buffer, command_pool, queue, pbp_buf,
pipeline, pipeline_layout, descriptor_set,
filter_out_buf, size, &pbp, &reversed_pbp,
global_size, pbp_mapped_int, staging_pbp_buffer,
staging_pbp_buffer_mem, staging_filter_buffer_mem,
staging_filter_buffer]() -> bool {
vkResetCommandBuffer(command_buffer, 0); vkResetCommandBuffer(command_buffer, 0);
for (unsigned int i = 0; i < pbp.size(); ++i) { for (unsigned int i = 0; i < pbp.size(); ++i) {
if (reversed_pbp) { if (reversed_pbp) {
pbp_mapped_int[i] = pbp[i] ? 0 : 1; pbp_mapped[i] = pbp[i] ? 0 : 1;
} else { } else {
pbp_mapped_int[i] = pbp[i] ? 1 : 0; pbp_mapped[i] = pbp[i] ? 1 : 0;
} }
} }
vulkan_flush_buffer(device, staging_pbp_buffer_mem); vulkan_flush_buffer(device, pbp_buf_mem);
// Copy pbp buffer.
vulkan_copy_buffer(device, command_pool, queue, staging_pbp_buffer, pbp_buf,
size * sizeof(int));
VkCommandBufferBeginInfo begin_info{}; VkCommandBufferBeginInfo begin_info{};
begin_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO; begin_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
@ -319,11 +255,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
return false; return false;
} }
// Copy back filter_out buffer. vulkan_flush_buffer(device, filter_out_buf_mem);
vulkan_copy_buffer(device, command_pool, queue, filter_out_buf,
staging_filter_buffer, size * sizeof(float));
vulkan_flush_buffer(device, staging_filter_buffer_mem);
return true; return true;
}; };
@ -350,7 +282,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
std::cerr << "Vulkan: Failed to execute get_filter at start!\n"; std::cerr << "Vulkan: Failed to execute get_filter at start!\n";
} else { } else {
#ifndef NDEBUG #ifndef NDEBUG
internal::write_filter(vulkan_buf_to_vec(filter_mapped_float, size), width, internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
"filter_out_start.pgm"); "filter_out_start.pgm");
#endif #endif
} }
@ -370,7 +302,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
int min, max; int min, max;
std::tie(min, max) = std::tie(min, max) =
internal::filter_minmax_raw_array(filter_mapped_float, size, pbp); internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
pbp[max] = false; pbp[max] = false;
@ -382,7 +314,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
// get second buffer's min // get second buffer's min
int second_min; int second_min;
std::tie(second_min, std::ignore) = std::tie(second_min, std::ignore) =
internal::filter_minmax_raw_array(filter_mapped_float, size, pbp); internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
if (second_min == max) { if (second_min == max) {
pbp[max] = true; pbp[max] = true;
@ -414,7 +346,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
std::cerr << "Vulkan: Failed to execute do_filter (at end)\n"; std::cerr << "Vulkan: Failed to execute do_filter (at end)\n";
} else { } else {
#ifndef NDEBUG #ifndef NDEBUG
internal::write_filter(vulkan_buf_to_vec(filter_mapped_float, size), width, internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
"filter_out_final.pgm"); "filter_out_final.pgm");
FILE *blue_noise_image = fopen("blue_noise.pbm", "w"); FILE *blue_noise_image = fopen("blue_noise.pbm", "w");
fprintf(blue_noise_image, "P1\n%d %d\n", width, height); fprintf(blue_noise_image, "P1\n%d %d\n", width, height);
@ -451,7 +383,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
#endif #endif
get_filter(); get_filter();
std::tie(std::ignore, max) = std::tie(std::ignore, max) =
internal::filter_minmax_raw_array(filter_mapped_float, size, pbp); internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
pbp.at(max) = false; pbp.at(max) = false;
dither_array.at(max) = i; dither_array.at(max) = i;
#ifndef NDEBUG #ifndef NDEBUG
@ -475,7 +407,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
#endif #endif
get_filter(); get_filter();
std::tie(min, std::ignore) = std::tie(min, std::ignore) =
internal::filter_minmax_raw_array(filter_mapped_float, size, pbp); internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
pbp.at(min) = true; pbp.at(min) = true;
dither_array.at(min) = i; dither_array.at(min) = i;
#ifndef NDEBUG #ifndef NDEBUG
@ -491,7 +423,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
image::Bl min_pixels = internal::rangeToBl(dither_array, width); 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(); get_filter();
internal::write_filter(vulkan_buf_to_vec(filter_mapped_float, size), width, internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
"filter_mid.pgm"); "filter_mid.pgm");
image::Bl pbp_image = toBl(pbp, width); image::Bl pbp_image = toBl(pbp, width);
pbp_image.writeToFile(image::file_type::PNG, true, "debug_pbp_mid.png"); pbp_image.writeToFile(image::file_type::PNG, true, "debug_pbp_mid.png");
@ -505,7 +437,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
#endif #endif
get_filter(); get_filter();
std::tie(std::ignore, max) = std::tie(std::ignore, max) =
internal::filter_minmax_raw_array(filter_mapped_float, size, pbp); internal::filter_minmax_raw_array(filter_out_mapped, size, pbp);
pbp.at(max) = true; pbp.at(max) = true;
dither_array.at(max) = i; dither_array.at(max) = i;
#ifndef NDEBUG #ifndef NDEBUG
@ -521,7 +453,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
#ifndef NDEBUG #ifndef NDEBUG
{ {
get_filter(); get_filter();
internal::write_filter(vulkan_buf_to_vec(filter_mapped_float, size), width, internal::write_filter(vulkan_buf_to_vec(filter_out_mapped, size), width,
"filter_after.pgm"); "filter_after.pgm");
image::Bl pbp_image = toBl(pbp, width); 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");
@ -530,6 +462,7 @@ std::vector<unsigned int> dither::internal::blue_noise_vulkan_impl(
return dither_array; return dither_array;
} }
std::vector<float> dither::internal::vulkan_buf_to_vec(float *mapped, std::vector<float> dither::internal::vulkan_buf_to_vec(float *mapped,
unsigned int size) { unsigned int size) {
std::vector<float> v(size); std::vector<float> v(size);
@ -1009,11 +942,13 @@ image::Bl dither::blue_noise(int width, int height, int threads,
VkBuffer filter_out_buf; VkBuffer filter_out_buf;
VkDeviceMemory filter_out_buf_mem; VkDeviceMemory filter_out_buf_mem;
if (!internal::vulkan_create_buffer(device, phys_device, filter_out_size, if (!internal::vulkan_create_buffer(
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | device, phys_device, filter_out_size,
VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
filter_out_buf, filter_out_buf_mem)) { VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
filter_out_buf, filter_out_buf_mem)) {
std::clog << "WARNING: Failed to create filter_out buffer!\n"; std::clog << "WARNING: Failed to create filter_out buffer!\n";
goto ENDOF_VULKAN; goto ENDOF_VULKAN;
} }
@ -1027,14 +962,24 @@ image::Bl dither::blue_noise(int width, int height, int threads,
vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr); vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
}, },
&filter_out_buf_mem); &filter_out_buf_mem);
void *filter_out_mapped;
vkMapMemory(device, filter_out_buf_mem, 0, filter_out_size, 0,
&filter_out_mapped);
utility::Cleanup cleanup_filter_out_mapped(
[device](void *ptr) {
vkUnmapMemory(device, *((VkDeviceMemory *)ptr));
},
&filter_out_buf_mem);
float *filter_out_mapped_float = (float *)filter_out_mapped;
VkBuffer pbp_buf; VkBuffer pbp_buf;
VkDeviceMemory pbp_buf_mem; VkDeviceMemory pbp_buf_mem;
if (!internal::vulkan_create_buffer(device, phys_device, pbp_size, if (!internal::vulkan_create_buffer(
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | device, phys_device, pbp_size, VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
VK_BUFFER_USAGE_TRANSFER_DST_BIT, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
pbp_buf, pbp_buf_mem)) { VK_MEMORY_PROPERTY_HOST_CACHED_BIT,
pbp_buf, pbp_buf_mem)) {
std::clog << "WARNING: Failed to create pbp buffer!\n"; std::clog << "WARNING: Failed to create pbp buffer!\n";
goto ENDOF_VULKAN; goto ENDOF_VULKAN;
} }
@ -1048,6 +993,14 @@ image::Bl dither::blue_noise(int width, int height, int threads,
vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr); vkFreeMemory(device, *((VkDeviceMemory *)ptr), nullptr);
}, },
&pbp_buf_mem); &pbp_buf_mem);
void *pbp_mapped;
vkMapMemory(device, pbp_buf_mem, 0, pbp_size, 0, &pbp_mapped);
utility::Cleanup cleanup_pbp_mapped(
[device](void *ptr) {
vkUnmapMemory(device, *((VkDeviceMemory *)ptr));
},
&pbp_buf_mem);
int *pbp_mapped_int = (int *)pbp_mapped;
VkBuffer other_buf; VkBuffer other_buf;
VkDeviceMemory other_buf_mem; VkDeviceMemory other_buf_mem;
@ -1224,9 +1177,10 @@ image::Bl dither::blue_noise(int width, int height, int threads,
} }
auto result = dither::internal::blue_noise_vulkan_impl( auto result = dither::internal::blue_noise_vulkan_impl(
device, phys_device, command_buffer, command_pool, compute_queue, device, command_buffer, compute_queue, compute_pipeline,
pbp_buf, compute_pipeline, compute_pipeline_layout, compute_pipeline_layout, compute_descriptor_set,
compute_descriptor_set, filter_out_buf, width, height); filter_out_mapped_float, pbp_mapped_int, filter_out_buf_mem,
pbp_buf_mem, width, height);
if (!result.empty()) { if (!result.empty()) {
return internal::rangeToBl(result, width); return internal::rangeToBl(result, width);
} }

View file

@ -64,11 +64,11 @@ void vulkan_flush_buffer(VkDevice device, VkDeviceMemory memory);
void vulkan_invalidate_buffer(VkDevice device, VkDeviceMemory memory); void vulkan_invalidate_buffer(VkDevice device, VkDeviceMemory memory);
std::vector<unsigned int> blue_noise_vulkan_impl( std::vector<unsigned int> blue_noise_vulkan_impl(
VkDevice device, VkPhysicalDevice phys_device, VkDevice device, VkCommandBuffer command_buffer, VkQueue queue,
VkCommandBuffer command_buffer, VkCommandPool command_pool, VkQueue queue, VkPipeline pipeline, VkPipelineLayout pipeline_layout,
VkBuffer pbp_buf, VkPipeline pipeline, VkPipelineLayout pipeline_layout, VkDescriptorSet descriptor_set, float *filter_out_mapped, int *pbp_mapped,
VkDescriptorSet descriptor_set, VkBuffer filter_out_buf, const int width, VkDeviceMemory filter_out_buf_mem, VkDeviceMemory pbp_buf_mem,
const int height); const int width, const int height);
std::vector<float> vulkan_buf_to_vec(float *mapped, unsigned int size); std::vector<float> vulkan_buf_to_vec(float *mapped, unsigned int size);
#endif #endif