From 4d454a23c87e6003389fba4a59a161e940f099af Mon Sep 17 00:00:00 2001 From: Dmitry Zolotukhin Date: Sun, 28 Jan 2024 20:46:01 +0100 Subject: [PATCH] Improve device scoring. Improve searching for the right buffer. Keep trying all heaps when attempting to allocate memory. Fixed incorrect type in shader. Added a few more tests for sending the matrix and vectors. --- .../shaders/cross_check_filter.comp.glsl | 1 + .../shaders/init_out_data.comp.glsl | 3 +- src/correlation/shaders/init_out_data.spv | Bin 2060 -> 2076 bytes src/correlation/vk.rs | 158 ++++++++---------- 4 files changed, 77 insertions(+), 85 deletions(-) diff --git a/src/correlation/shaders/cross_check_filter.comp.glsl b/src/correlation/shaders/cross_check_filter.comp.glsl index 8f8fca5..4eb2803 100644 --- a/src/correlation/shaders/cross_check_filter.comp.glsl +++ b/src/correlation/shaders/cross_check_filter.comp.glsl @@ -38,4 +38,5 @@ void main() { // TODO: remove this debug code img1[0] = ivec2(img1_width, img1_height); + img1[1] = ivec2(img2_width, img2_height); } diff --git a/src/correlation/shaders/init_out_data.comp.glsl b/src/correlation/shaders/init_out_data.comp.glsl index 138da8d..91d53b0 100644 --- a/src/correlation/shaders/init_out_data.comp.glsl +++ b/src/correlation/shaders/init_out_data.comp.glsl @@ -45,7 +45,7 @@ layout(std430, set = 0, binding = 3) buffer Internals_Int { // Layout: // Contains [min, max, neighbor_count] for the corridor range - int internals_int[]; + ivec3 internals_int[]; }; layout(std430, set = 0, binding = 4) buffer Result_Matches { @@ -69,4 +69,5 @@ void main() { // TODO: remove this debug code result_corr[0] = threshold; + result_corr[1] = fundamental_matrix[2][0]; } diff --git a/src/correlation/shaders/init_out_data.spv b/src/correlation/shaders/init_out_data.spv index 0ed0416a34ffe52ee78998cba345f6126abb6b9d..d0282371cab1d06a32c46ffefab97cb417e7cf0f 100644 GIT binary patch delta 373 zcmYk2%?*M;6oj9Ef-xlAj0cUvQBU|U;1#>D0E?gjZ(eP}SegwuAE05AVRmNU4g>e- z8HKKs_+~vDSzzZ^cy+6_nGJ+UxLDsdVph_SmFCu-D_xprJ8}F!YxF|;cYe^Gyg(rD zC}wr@@Y?y=nm|wBg~dDcZca7UHkCCIxI5L8?szlp>5ZOiZYy>KysXvq#v^Ah rCI?Lpo$<(_8=9Pxm>GF#9fUzY%%ukS&|Ki>t)zzHqV@8}#Q%H&QfU-$ delta 357 zcmYk1y$!-J6oemfAVDZnIz$12q(V^f4+(YHf&myJ3((QB1Y$H6;67pymQLq8KcCO{ z^qM9v$Ro3s^(?mIBRRWzVP+j66i(K*rI?jBWR Result<(vk::PhysicalDevice, &'static str, u32), Box> { + ) -> Result<(vk::PhysicalDevice, String, u32), Box> { let devices = instance.enumerate_physical_devices()?; let device = devices .iter() @@ -923,16 +934,7 @@ impl Device { let queue_index = Device::find_compute_queue(instance, device)?; let device_name = CStr::from_ptr(props.device_name.as_ptr()); - let device_name = device_name.to_str().unwrap(); - println!( - "Device {} type {} {}-{}-{}-{}", - device_name, - props.device_type.as_raw(), - props.limits.max_push_constants_size, - props.limits.max_bound_descriptor_sets, - props.limits.max_storage_buffer_range, - max_buffer_size - ); + let device_name = String::from_utf8_lossy(device_name.to_bytes()).to_string(); // TODO: allow to specify a device name filter/regex? let score = match props.device_type { vk::PhysicalDeviceType::DISCRETE_GPU => 3, @@ -941,24 +943,25 @@ impl Device { _ => 0, }; // Prefer real devices instead of dzn emulation. - let dzn_multiplier = if device_name + let is_dzn = device_name .to_lowercase() - .starts_with("microsoft direct3d12") - { - 1 - } else { - 10 - }; - Some((device, device_name, queue_index, score * dzn_multiplier)) + .starts_with("microsoft direct3d12"); + let score = (score, is_dzn); + Some((device, device_name, queue_index, score)) }) - .max_by_key(|(_device, _name, _queue_index, score)| *score); - let (device, name, queue_index) = if let Some((device, name, queue_index, _score)) = device - { + .max_by(|(_, _, _, a), (_, _, _, b)| { + if a.1 && !b.1 { + return Ordering::Less; + } else if !a.1 && b.1 { + return Ordering::Greater; + } + return a.0.cmp(&b.0); + }); + let (device, name, queue_index) = if let Some((device, name, queue_index, score)) = device { (device, name, queue_index) } else { return Err(GpuError::new("Device not found").into()); }; - println!("selected device {}", name); Ok((device, name, queue_index)) } @@ -1009,6 +1012,7 @@ impl Device { let max_pixels = img1_pixels.max(img2_pixels); let mut buffers: Vec = vec![]; let cleanup_err = |buffers: &[Buffer], err| { + println!("buffers count is {}", buffers.len()); buffers.iter().for_each(|buffer| { device.free_memory(buffer.buffer_memory, None); device.destroy_buffer(buffer.buffer, None) @@ -1095,13 +1099,13 @@ impl Device { buffer_type: BufferType, ) -> Result> { let size = size as u64; - let gpu_local = match buffer_type { - BufferType::GpuOnly | BufferType::GpuDestination | BufferType::GpuSource => true, - BufferType::HostSource | BufferType::HostDestination => false, - }; - let host_visible = match buffer_type { - BufferType::HostSource | BufferType::HostDestination => true, - BufferType::GpuOnly | BufferType::GpuDestination | BufferType::GpuSource => false, + let required_memory_properties = match buffer_type { + BufferType::GpuOnly | BufferType::GpuDestination | BufferType::GpuSource => { + vk::MemoryPropertyFlags::DEVICE_LOCAL + } + BufferType::HostSource | BufferType::HostDestination => { + vk::MemoryPropertyFlags::HOST_VISIBLE + } }; let extra_usage_flags = match buffer_type { BufferType::HostSource => vk::BufferUsageFlags::TRANSFER_SRC, @@ -1122,58 +1126,41 @@ impl Device { }; let buffer = device.create_buffer(&buffer_create_info, None)?; let memory_requirements = device.get_buffer_memory_requirements(buffer); - let memory_type_index = memory_properties.memory_types - [..memory_properties.memory_type_count as usize] - .iter() - .enumerate() - .find(|(memory_type_index, memory_type)| { + let buffer_memory = (0..memory_properties.memory_type_count as usize) + .flat_map(|i| { + let memory_type = memory_properties.memory_types[i]; if memory_properties.memory_heaps[memory_type.heap_index as usize].size < memory_requirements.size { - return false; - }; - if (1 << memory_type_index) & memory_requirements.memory_type_bits == 0 { - return false; + return None; } - - if gpu_local - && memory_type - .property_flags - .contains(vk::MemoryPropertyFlags::DEVICE_LOCAL) - { - return true; + if ((1 << i) & memory_requirements.memory_type_bits) == 0 { + return None; } - if host_visible - && memory_type - .property_flags - .contains(vk::MemoryPropertyFlags::HOST_VISIBLE) - { - return true; + let property_flags = memory_type.property_flags; + if !property_flags.contains(required_memory_properties) { + return None; } - false - }); - let memory_type_index = if let Some((index, _)) = memory_type_index { - index as u32 + let host_visible = property_flags.contains(vk::MemoryPropertyFlags::HOST_VISIBLE); + let host_coherent = property_flags.contains(vk::MemoryPropertyFlags::HOST_COHERENT); + let allocate_info = vk::MemoryAllocateInfo { + allocation_size: memory_requirements.size, + memory_type_index: i as u32, + ..Default::default() + }; + // Some buffers may fill up, in this case allocating memory can fail. + let mem = device.allocate_memory(&allocate_info, None).ok()?; + + Some((mem, host_visible, host_coherent)) + }) + .next(); + + let (buffer_memory, host_visible, host_coherent) = if let Some(mem) = buffer_memory { + mem } else { + device.destroy_buffer(buffer, None); return Err(GpuError::new("Cannot find suitable memory").into()); }; - let property_flags = - memory_properties.memory_types[memory_type_index as usize].property_flags; - let host_visible = property_flags.contains(vk::MemoryPropertyFlags::HOST_VISIBLE); - let host_coherent = property_flags.contains(vk::MemoryPropertyFlags::HOST_COHERENT); - let allocate_info = vk::MemoryAllocateInfo { - allocation_size: memory_requirements.size, - memory_type_index, - ..Default::default() - }; - let buffer_memory = device.allocate_memory(&allocate_info, None); - let buffer_memory = match buffer_memory { - Ok(mem) => mem, - Err(err) => { - device.destroy_buffer(buffer, None); - return Err(err.into()); - } - }; let result = Buffer { buffer, buffer_memory, @@ -1192,13 +1179,12 @@ impl Device { ) -> Result> { let create_layout_bindings = |count| { let bindings = (0..count) - .map(|i| { - vk::DescriptorSetLayoutBinding::builder() - .binding(i as u32) - .descriptor_type(vk::DescriptorType::STORAGE_BUFFER) - .descriptor_count(1) - .stage_flags(vk::ShaderStageFlags::COMPUTE) - .build() + .map(|i| vk::DescriptorSetLayoutBinding { + binding: i as u32, + descriptor_type: vk::DescriptorType::STORAGE_BUFFER, + descriptor_count: 1, + stage_flags: vk::ShaderStageFlags::COMPUTE, + ..Default::default() }) .collect::>(); let layout_info = @@ -1207,10 +1193,10 @@ impl Device { }; let descriptor_pool_size = [vk::DescriptorPoolSize::builder() .ty(vk::DescriptorType::STORAGE_BUFFER) - .descriptor_count(2) + .descriptor_count(8) .build()]; let descriptor_pool_info = vk::DescriptorPoolCreateInfo::builder() - .max_sets(1) + .max_sets(2) .pool_sizes(&descriptor_pool_size); let descriptor_pool = device.create_descriptor_pool(&descriptor_pool_info, None)?; let cleanup_err = |err| { @@ -1251,12 +1237,15 @@ impl Device { device.destroy_descriptor_pool(descriptor_pool, None); err }; + println!("creating descriptor pools 6"); let descriptor_set_allocate_info = vk::DescriptorSetAllocateInfo::builder() .descriptor_pool(descriptor_pool) .set_layouts(&layouts); + println!("creating descriptor pools 6.1"); let descriptor_sets = device .allocate_descriptor_sets(&descriptor_set_allocate_info) .map_err(cleanup_err)?; + println!("creating descriptor pools 7"); // TODO: extract this to allow switching direction on the fly. let create_buffer_infos = |buffers: &[Buffer]| { @@ -1398,6 +1387,7 @@ impl Device { .create_fence(&fence_create_info, None) .map_err(cleanup_err)?; let cleanup_err = |err| { + println!("Failed to alloc command buffer"); device.destroy_command_pool(command_pool, None); device.destroy_fence(fence, None); err