diff options
Diffstat (limited to 'drivers')
| -rw-r--r-- | drivers/png/image_loader_png.cpp | 6 | ||||
| -rw-r--r-- | drivers/png/resource_saver_png.cpp | 2 | ||||
| -rw-r--r-- | drivers/register_driver_types.cpp | 2 | ||||
| -rw-r--r-- | drivers/unix/dir_access_unix.h | 2 | ||||
| -rw-r--r-- | drivers/unix/file_access_unix.cpp | 2 | ||||
| -rw-r--r-- | drivers/unix/file_access_unix.h | 2 | ||||
| -rw-r--r-- | drivers/unix/net_socket_posix.cpp | 2 | ||||
| -rw-r--r-- | drivers/unix/os_unix.cpp | 4 | ||||
| -rw-r--r-- | drivers/vulkan/rendering_device_vulkan.cpp | 824 | ||||
| -rw-r--r-- | drivers/vulkan/rendering_device_vulkan.h | 127 | ||||
| -rw-r--r-- | drivers/vulkan/vulkan_context.cpp | 129 | ||||
| -rw-r--r-- | drivers/vulkan/vulkan_context.h | 12 | ||||
| -rw-r--r-- | drivers/windows/dir_access_windows.cpp | 2 | ||||
| -rw-r--r-- | drivers/windows/dir_access_windows.h | 2 | ||||
| -rw-r--r-- | drivers/windows/file_access_windows.cpp | 1 | ||||
| -rw-r--r-- | drivers/windows/file_access_windows.h | 2 |
16 files changed, 849 insertions, 272 deletions
diff --git a/drivers/png/image_loader_png.cpp b/drivers/png/image_loader_png.cpp index 8b2e786146..f7f7f2dd22 100644 --- a/drivers/png/image_loader_png.cpp +++ b/drivers/png/image_loader_png.cpp @@ -59,7 +59,7 @@ void ImageLoaderPNG::get_recognized_extensions(List<String> *p_extensions) const Ref<Image> ImageLoaderPNG::load_mem_png(const uint8_t *p_png, int p_size) { Ref<Image> img; - img.instance(); + img.instantiate(); // the value of p_force_linear does not matter since it only applies to 16 bit Error err = PNGDriverCommon::png_to_image(p_png, p_size, false, img); @@ -101,6 +101,6 @@ Vector<uint8_t> ImageLoaderPNG::lossless_pack_png(const Ref<Image> &p_image) { ImageLoaderPNG::ImageLoaderPNG() { Image::_png_mem_loader_func = load_mem_png; - Image::lossless_unpacker = lossless_unpack_png; - Image::lossless_packer = lossless_pack_png; + Image::png_unpacker = lossless_unpack_png; + Image::png_packer = lossless_pack_png; } diff --git a/drivers/png/resource_saver_png.cpp b/drivers/png/resource_saver_png.cpp index b737a287d1..2971880985 100644 --- a/drivers/png/resource_saver_png.cpp +++ b/drivers/png/resource_saver_png.cpp @@ -30,8 +30,8 @@ #include "resource_saver_png.h" +#include "core/io/file_access.h" #include "core/io/image.h" -#include "core/os/file_access.h" #include "drivers/png/png_driver_common.h" #include "scene/resources/texture.h" diff --git a/drivers/register_driver_types.cpp b/drivers/register_driver_types.cpp index 18262c74c4..83702ea2cc 100644 --- a/drivers/register_driver_types.cpp +++ b/drivers/register_driver_types.cpp @@ -40,7 +40,7 @@ void register_core_driver_types() { image_loader_png = memnew(ImageLoaderPNG); ImageLoader::add_image_format_loader(image_loader_png); - resource_saver_png.instance(); + resource_saver_png.instantiate(); ResourceSaver::add_resource_format_saver(resource_saver_png); } diff --git a/drivers/unix/dir_access_unix.h b/drivers/unix/dir_access_unix.h index 28a7053d25..8b19308967 100644 --- a/drivers/unix/dir_access_unix.h +++ b/drivers/unix/dir_access_unix.h @@ -33,7 +33,7 @@ #if defined(UNIX_ENABLED) || defined(LIBC_FILEIO_ENABLED) -#include "core/os/dir_access.h" +#include "core/io/dir_access.h" #include <dirent.h> #include <sys/stat.h> diff --git a/drivers/unix/file_access_unix.cpp b/drivers/unix/file_access_unix.cpp index ec23df62d0..6ea55219bb 100644 --- a/drivers/unix/file_access_unix.cpp +++ b/drivers/unix/file_access_unix.cpp @@ -264,7 +264,7 @@ void FileAccessUnix::store_8(uint8_t p_dest) { void FileAccessUnix::store_buffer(const uint8_t *p_src, uint64_t p_length) { ERR_FAIL_COND_MSG(!f, "File must be opened before use."); - ERR_FAIL_COND(!p_src); + ERR_FAIL_COND(!p_src && p_length > 0); ERR_FAIL_COND(fwrite(p_src, 1, p_length, f) != p_length); } diff --git a/drivers/unix/file_access_unix.h b/drivers/unix/file_access_unix.h index 5b1599c67f..8b27c308e1 100644 --- a/drivers/unix/file_access_unix.h +++ b/drivers/unix/file_access_unix.h @@ -31,7 +31,7 @@ #ifndef FILE_ACCESS_UNIX_H #define FILE_ACCESS_UNIX_H -#include "core/os/file_access.h" +#include "core/io/file_access.h" #include "core/os/memory.h" #include <stdio.h> diff --git a/drivers/unix/net_socket_posix.cpp b/drivers/unix/net_socket_posix.cpp index 222aef998c..efeaf32ff7 100644 --- a/drivers/unix/net_socket_posix.cpp +++ b/drivers/unix/net_socket_posix.cpp @@ -466,7 +466,7 @@ Error NetSocketPosix::poll(PollType p_type, int p_timeout) const { FD_ZERO(&ex); FD_SET(_sock, &ex); struct timeval timeout = { p_timeout / 1000, (p_timeout % 1000) * 1000 }; - // For blocking operation, pass nullptr timeout pointer to select. + // For blocking operation, pass nullptr timeout pointer to select. struct timeval *tp = nullptr; if (p_timeout >= 0) { // If timeout is non-negative, we want to specify the timeout instead. diff --git a/drivers/unix/os_unix.cpp b/drivers/unix/os_unix.cpp index 15cd7bee92..a5c61bbea5 100644 --- a/drivers/unix/os_unix.cpp +++ b/drivers/unix/os_unix.cpp @@ -195,8 +195,8 @@ OS::Time OS_Unix::get_time(bool utc) const { } Time ret; ret.hour = lt.tm_hour; - ret.min = lt.tm_min; - ret.sec = lt.tm_sec; + ret.minute = lt.tm_min; + ret.second = lt.tm_sec; get_time_zone_info(); return ret; } diff --git a/drivers/vulkan/rendering_device_vulkan.cpp b/drivers/vulkan/rendering_device_vulkan.cpp index 43b2a24172..6c4e590586 100644 --- a/drivers/vulkan/rendering_device_vulkan.cpp +++ b/drivers/vulkan/rendering_device_vulkan.cpp @@ -31,7 +31,7 @@ #include "rendering_device_vulkan.h" #include "core/config/project_settings.h" -#include "core/os/file_access.h" +#include "core/io/file_access.h" #include "core/os/os.h" #include "core/templates/hashfuncs.h" #include "drivers/vulkan/vulkan_context.h" @@ -3244,12 +3244,7 @@ bool RenderingDeviceVulkan::texture_is_format_supported_for_usage(DataFormat p_f /**** ATTACHMENT ****/ /********************/ -VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentFormat> &p_format, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, int *r_color_attachment_count) { - Vector<VkAttachmentDescription> attachments; - Vector<VkAttachmentReference> color_references; - Vector<VkAttachmentReference> depth_stencil_references; - Vector<VkAttachmentReference> resolve_references; - +VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentFormat> &p_attachments, const Vector<FramebufferPass> &p_passes, InitialAction p_initial_action, FinalAction p_final_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, uint32_t p_view_count, Vector<TextureSamples> *r_samples) { // Set up dependencies from/to external equivalent to the default (implicit) one, and then amend them const VkPipelineStageFlags default_access_mask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | @@ -3262,27 +3257,31 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF { 0, VK_SUBPASS_EXTERNAL, VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, default_access_mask, 0, 0 } }; VkSubpassDependency &dependency_from_external = dependencies[0]; VkSubpassDependency &dependency_to_external = dependencies[1]; + LocalVector<int32_t> attachment_last_pass; + attachment_last_pass.resize(p_attachments.size()); - for (int i = 0; i < p_format.size(); i++) { - ERR_FAIL_INDEX_V(p_format[i].format, DATA_FORMAT_MAX, VK_NULL_HANDLE); - ERR_FAIL_INDEX_V(p_format[i].samples, TEXTURE_SAMPLES_MAX, VK_NULL_HANDLE); - ERR_FAIL_COND_V_MSG(!(p_format[i].usage_flags & (TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT | TEXTURE_USAGE_RESOLVE_ATTACHMENT_BIT)), + Vector<VkAttachmentDescription> attachments; + + for (int i = 0; i < p_attachments.size(); i++) { + ERR_FAIL_INDEX_V(p_attachments[i].format, DATA_FORMAT_MAX, VK_NULL_HANDLE); + ERR_FAIL_INDEX_V(p_attachments[i].samples, TEXTURE_SAMPLES_MAX, VK_NULL_HANDLE); + ERR_FAIL_COND_V_MSG(!(p_attachments[i].usage_flags & (TEXTURE_USAGE_COLOR_ATTACHMENT_BIT | TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT | TEXTURE_USAGE_RESOLVE_ATTACHMENT_BIT)), VK_NULL_HANDLE, "Texture format for index (" + itos(i) + ") requires an attachment (depth, stencil or resolve) bit set."); VkAttachmentDescription description = {}; description.flags = 0; - description.format = vulkan_formats[p_format[i].format]; - description.samples = rasterization_sample_count[p_format[i].samples]; + description.format = vulkan_formats[p_attachments[i].format]; + description.samples = rasterization_sample_count[p_attachments[i].samples]; - bool is_depth_stencil = p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT; - bool is_sampled = p_format[i].usage_flags & TEXTURE_USAGE_SAMPLING_BIT; - bool is_storage = p_format[i].usage_flags & TEXTURE_USAGE_STORAGE_BIT; + bool is_sampled = p_attachments[i].usage_flags & TEXTURE_USAGE_SAMPLING_BIT; + bool is_storage = p_attachments[i].usage_flags & TEXTURE_USAGE_STORAGE_BIT; + bool is_depth = p_attachments[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT; // For each UNDEFINED, assume the prior use was a *read*, as we'd be discarding the output of a write // Also, each UNDEFINED will do an immediate layout transition (write), s.t. we must ensure execution synchronization vs. - // the read. If this is a performance issue, one could track the actual last accessor of each resource, adding only that + // the read. If this is a performance issue, one could track the actual last accessor of each resource, adding only that // stage - switch (is_depth_stencil ? p_initial_depth_action : p_initial_color_action) { + switch (is_depth ? p_initial_depth_action : p_initial_action) { case INITIAL_ACTION_CLEAR_REGION: case INITIAL_ACTION_CLEAR: { description.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR; @@ -3291,11 +3290,11 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF dependency_from_external.srcStageMask |= reading_stages; } break; case INITIAL_ACTION_KEEP: { - if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + if (p_attachments[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { description.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD; description.initialLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL); description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; - } else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + } else if (p_attachments[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { description.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD; description.initialLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL); description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_LOAD; @@ -3308,11 +3307,11 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF } } break; case INITIAL_ACTION_DROP: { - if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + if (p_attachments[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; description.initialLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL); description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; - } else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + } else if (p_attachments[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; @@ -3326,11 +3325,11 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF } break; case INITIAL_ACTION_CLEAR_REGION_CONTINUE: case INITIAL_ACTION_CONTINUE: { - if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + if (p_attachments[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { description.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD; description.initialLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; - } else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + } else if (p_attachments[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { description.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD; description.initialLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_LOAD; @@ -3346,14 +3345,14 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF } } - switch (is_depth_stencil ? p_final_depth_action : p_final_color_action) { + switch (is_depth ? p_final_depth_action : p_final_action) { case FINAL_ACTION_READ: { - if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + if (p_attachments[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { description.storeOp = VK_ATTACHMENT_STORE_OP_STORE; description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE; description.finalLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL); update_external_dependency_for_store(dependency_to_external, is_sampled, is_storage, false); - } else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + } else if (p_attachments[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { description.storeOp = VK_ATTACHMENT_STORE_OP_STORE; description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE; description.finalLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL); @@ -3366,11 +3365,11 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF } } break; case FINAL_ACTION_DISCARD: { - if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + if (p_attachments[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { description.storeOp = VK_ATTACHMENT_STORE_OP_DONT_CARE; description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE; description.finalLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL); - } else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + } else if (p_attachments[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { description.storeOp = VK_ATTACHMENT_STORE_OP_DONT_CARE; description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE; description.finalLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL); @@ -3381,11 +3380,11 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF } } break; case FINAL_ACTION_CONTINUE: { - if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + if (p_attachments[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { description.storeOp = VK_ATTACHMENT_STORE_OP_STORE; description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE; description.finalLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; - } else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + } else if (p_attachments[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { description.storeOp = VK_ATTACHMENT_STORE_OP_STORE; description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE; description.finalLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; @@ -3401,27 +3400,194 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF } } + attachment_last_pass[i] = -1; + attachments.push_back(description); + } + + LocalVector<VkSubpassDescription> subpasses; + LocalVector<LocalVector<VkAttachmentReference>> color_reference_array; + LocalVector<LocalVector<VkAttachmentReference>> input_reference_array; + LocalVector<LocalVector<VkAttachmentReference>> resolve_reference_array; + LocalVector<LocalVector<uint32_t>> preserve_reference_array; + LocalVector<VkAttachmentReference> depth_reference_array; + + subpasses.resize(p_passes.size()); + color_reference_array.resize(p_passes.size()); + input_reference_array.resize(p_passes.size()); + resolve_reference_array.resize(p_passes.size()); + preserve_reference_array.resize(p_passes.size()); + depth_reference_array.resize(p_passes.size()); - VkAttachmentReference reference; - reference.attachment = i; + LocalVector<VkSubpassDependency> subpass_dependencies; - if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { - reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; + for (int i = 0; i < p_passes.size(); i++) { + const FramebufferPass *pass = &p_passes[i]; + + LocalVector<VkAttachmentReference> &color_references = color_reference_array[i]; + + TextureSamples texture_samples = TEXTURE_SAMPLES_1; + bool is_multisample_first = true; + + for (int j = 0; j < pass->color_attachments.size(); j++) { + int32_t attachment = pass->color_attachments[j]; + VkAttachmentReference reference; + if (attachment == FramebufferPass::ATTACHMENT_UNUSED) { + reference.attachment = VK_ATTACHMENT_UNUSED; + reference.layout = VK_IMAGE_LAYOUT_UNDEFINED; + } else { + ERR_FAIL_INDEX_V_MSG(attachment, p_attachments.size(), VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), color attachment (" + itos(j) + ")."); + ERR_FAIL_COND_V_MSG(!(p_attachments[attachment].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT), VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it's marked as depth, but it's not usable as color attachment."); + ERR_FAIL_COND_V_MSG(attachment_last_pass[attachment] == i, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it already was used for something else before in this pass."); + + if (is_multisample_first) { + texture_samples = p_attachments[attachment].samples; + is_multisample_first = false; + } else { + ERR_FAIL_COND_V_MSG(texture_samples != p_attachments[attachment].samples, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), if an attachment is marked as multisample, all of them should be multisample and use the same number of samples."); + } + reference.attachment = attachment; + reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; + attachment_last_pass[attachment] = i; + } color_references.push_back(reference); - } else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { - reference.layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; - depth_stencil_references.push_back(reference); - } else if (p_format[i].usage_flags & TEXTURE_USAGE_RESOLVE_ATTACHMENT_BIT) { - reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; + } + + LocalVector<VkAttachmentReference> &input_references = input_reference_array[i]; + + for (int j = 0; j < pass->input_attachments.size(); j++) { + int32_t attachment = pass->input_attachments[j]; + VkAttachmentReference reference; + if (attachment == FramebufferPass::ATTACHMENT_UNUSED) { + reference.attachment = VK_ATTACHMENT_UNUSED; + reference.layout = VK_IMAGE_LAYOUT_UNDEFINED; + } else { + ERR_FAIL_INDEX_V_MSG(attachment, p_attachments.size(), VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), input attachment (" + itos(j) + ")."); + ERR_FAIL_COND_V_MSG(!(p_attachments[attachment].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT), VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it's marked as depth, but it's not usable as input attachment."); + ERR_FAIL_COND_V_MSG(attachment_last_pass[attachment] == i, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it already was used for something else before in this pass."); + reference.attachment = attachment; + reference.layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + attachment_last_pass[attachment] = i; + } + input_references.push_back(reference); + } + + LocalVector<VkAttachmentReference> &resolve_references = resolve_reference_array[i]; + + if (pass->resolve_attachments.size() > 0) { + ERR_FAIL_COND_V_MSG(pass->resolve_attachments.size() != pass->color_attachments.size(), VK_NULL_HANDLE, "The amount of resolve attachments (" + itos(pass->resolve_attachments.size()) + ") must match the number of color attachments (" + itos(pass->color_attachments.size()) + ")."); + ERR_FAIL_COND_V_MSG(texture_samples == TEXTURE_SAMPLES_1, VK_NULL_HANDLE, "Resolve attachments specified, but color attachments are not multisample."); + } + for (int j = 0; j < pass->resolve_attachments.size(); j++) { + int32_t attachment = pass->resolve_attachments[j]; + VkAttachmentReference reference; + if (attachment == FramebufferPass::ATTACHMENT_UNUSED) { + reference.attachment = VK_ATTACHMENT_UNUSED; + reference.layout = VK_IMAGE_LAYOUT_UNDEFINED; + } else { + ERR_FAIL_INDEX_V_MSG(attachment, p_attachments.size(), VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), resolve attachment (" + itos(j) + ")."); + ERR_FAIL_COND_V_MSG(pass->color_attachments[j] == FramebufferPass::ATTACHMENT_UNUSED, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), resolve attachment (" + itos(j) + "), the respective color attachment is marked as unused."); + ERR_FAIL_COND_V_MSG(!(p_attachments[attachment].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT), VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it's marked as depth, but it's not usable as resolve attachment."); + ERR_FAIL_COND_V_MSG(attachment_last_pass[attachment] == i, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it already was used for something else before in this pass."); + bool multisample = p_attachments[attachment].samples > TEXTURE_SAMPLES_1; + ERR_FAIL_COND_V_MSG(multisample, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), resolve attachments can't be multisample."); + reference.attachment = attachment; + reference.layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + attachment_last_pass[attachment] = i; + } resolve_references.push_back(reference); - // if resolves are done, we need to ensure the copy is safe - dependency_to_external.dstStageMask |= VK_PIPELINE_STAGE_TRANSFER_BIT; - dependency_to_external.dstAccessMask |= VK_ACCESS_TRANSFER_READ_BIT; + } + + LocalVector<uint32_t> &preserve_references = preserve_reference_array[i]; + + for (int j = 0; j < pass->preserve_attachments.size(); j++) { + int32_t attachment = pass->preserve_attachments[j]; + + ERR_FAIL_COND_V_MSG(attachment == FramebufferPass::ATTACHMENT_UNUSED, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), preserve attachment (" + itos(j) + "). Preserve attachments can't be unused."); + + ERR_FAIL_INDEX_V_MSG(attachment, p_attachments.size(), VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), preserve attachment (" + itos(j) + ")."); + ERR_FAIL_COND_V_MSG(attachment_last_pass[attachment] == i, VK_NULL_HANDLE, "Invalid framebuffer format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it already was used for something else before in this pass."); + + attachment_last_pass[attachment] = i; + + preserve_references.push_back(attachment); + } + + VkAttachmentReference &depth_stencil_reference = depth_reference_array[i]; + + if (pass->depth_attachment != FramebufferPass::ATTACHMENT_UNUSED) { + int32_t attachment = pass->depth_attachment; + ERR_FAIL_INDEX_V_MSG(attachment, p_attachments.size(), VK_NULL_HANDLE, "Invalid framebuffer depth format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), depth attachment."); + ERR_FAIL_COND_V_MSG(!(p_attachments[attachment].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT), VK_NULL_HANDLE, "Invalid framebuffer depth format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it's marked as depth, but it's not a depth attachment."); + ERR_FAIL_COND_V_MSG(attachment_last_pass[attachment] == i, VK_NULL_HANDLE, "Invalid framebuffer depth format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), it already was used for something else before in this pass."); + depth_stencil_reference.attachment = attachment; + depth_stencil_reference.layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; + attachment_last_pass[attachment] = i; + + if (is_multisample_first) { + texture_samples = p_attachments[attachment].samples; + is_multisample_first = false; + } else { + ERR_FAIL_COND_V_MSG(texture_samples != p_attachments[attachment].samples, VK_NULL_HANDLE, "Invalid framebuffer depth format attachment(" + itos(attachment) + "), in pass (" + itos(i) + "), if an attachment is marked as multisample, all of them should be multisample and use the same number of samples including the depth."); + } + + } else { + depth_stencil_reference.attachment = VK_ATTACHMENT_UNUSED; + depth_stencil_reference.layout = VK_IMAGE_LAYOUT_UNDEFINED; + } + + VkSubpassDescription &subpass = subpasses[i]; + subpass.flags = 0; + subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS; + subpass.inputAttachmentCount = input_references.size(); + if (input_references.size()) { + subpass.pInputAttachments = input_references.ptr(); + } else { + subpass.pInputAttachments = nullptr; + } + subpass.colorAttachmentCount = color_references.size(); + if (color_references.size()) { + subpass.pColorAttachments = color_references.ptr(); + } else { + subpass.pColorAttachments = nullptr; + } + if (depth_stencil_reference.attachment != VK_ATTACHMENT_UNUSED) { + subpass.pDepthStencilAttachment = &depth_stencil_reference; } else { - ERR_FAIL_V_MSG(VK_NULL_HANDLE, "Texture index " + itos(i) + " is neither color, depth stencil or resolve so it can't be used as attachment."); + subpass.pDepthStencilAttachment = nullptr; } + if (resolve_references.size()) { + subpass.pResolveAttachments = resolve_references.ptr(); + } else { + subpass.pResolveAttachments = nullptr; + } + + subpass.preserveAttachmentCount = preserve_references.size(); + if (preserve_references.size()) { + subpass.pPreserveAttachments = preserve_references.ptr(); + } else { + subpass.pPreserveAttachments = nullptr; + } + + if (r_samples) { + r_samples->push_back(texture_samples); + } + + if (i > 0) { + VkSubpassDependency dependency; + dependency.srcSubpass = i - 1; + dependency.dstSubpass = i; + dependency.srcStageMask = 0; + dependency.srcStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT; + dependency.dstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; + + dependency.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT; + dependency.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_INPUT_ATTACHMENT_READ_BIT; + dependency.dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT; + subpass_dependencies.push_back(dependency); + } + /* // NOTE: Big Mallet Approach -- any layout transition causes a full barrier if (reference.layout != description.initialLayout) { // NOTE: this should be smarter based on the texture's knowledge of its previous role @@ -3433,57 +3599,82 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF dependency_to_external.dstStageMask |= VK_PIPELINE_STAGE_ALL_COMMANDS_BIT; dependency_to_external.dstAccessMask |= VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT; } + */ } - ERR_FAIL_COND_V_MSG(depth_stencil_references.size() > 1, VK_NULL_HANDLE, - "Formats can only have one depth/stencil attachment, supplied (" + itos(depth_stencil_references.size()) + ")."); - - ERR_FAIL_COND_V_MSG(resolve_references.size() > 1, VK_NULL_HANDLE, - "Formats can only have one resolve attachment, supplied (" + itos(resolve_references.size()) + ")."); - - VkSubpassDescription subpass; - subpass.flags = 0; - subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS; - subpass.inputAttachmentCount = 0; //unsupported for now - subpass.pInputAttachments = nullptr; - subpass.colorAttachmentCount = color_references.size(); - subpass.pColorAttachments = color_references.ptr(); - subpass.pDepthStencilAttachment = depth_stencil_references.ptr(); - subpass.pResolveAttachments = resolve_references.ptr(); - subpass.preserveAttachmentCount = 0; - subpass.pPreserveAttachments = nullptr; - VkRenderPassCreateInfo render_pass_create_info; render_pass_create_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO; render_pass_create_info.pNext = nullptr; render_pass_create_info.flags = 0; render_pass_create_info.attachmentCount = attachments.size(); render_pass_create_info.pAttachments = attachments.ptr(); - render_pass_create_info.subpassCount = 1; - render_pass_create_info.pSubpasses = &subpass; + render_pass_create_info.subpassCount = subpasses.size(); + render_pass_create_info.pSubpasses = subpasses.ptr(); // Commenting this because it seems it just avoids raster and compute to work at the same time. // Other barriers seem to be protecting the render pass fine. // render_pass_create_info.dependencyCount = 2; // render_pass_create_info.pDependencies = dependencies; - render_pass_create_info.dependencyCount = 0; - render_pass_create_info.pDependencies = nullptr; + render_pass_create_info.dependencyCount = subpass_dependencies.size(); + if (subpass_dependencies.size()) { + render_pass_create_info.pDependencies = subpass_dependencies.ptr(); + } else { + render_pass_create_info.pDependencies = nullptr; + } + + const uint32_t view_mask = (1 << p_view_count) - 1; + const uint32_t correlation_mask = (1 << p_view_count) - 1; + VkRenderPassMultiviewCreateInfo render_pass_multiview_create_info; + + if (p_view_count > 1) { + const VulkanContext::MultiviewCapabilities capabilities = context->get_multiview_capabilities(); + + // For now this only works with multiview! + ERR_FAIL_COND_V_MSG(!capabilities.is_supported, VK_NULL_HANDLE, "Multiview not supported"); + + // Make sure we limit this to the number of views we support. + ERR_FAIL_COND_V_MSG(p_view_count > capabilities.max_view_count, VK_NULL_HANDLE, "Hardware does not support requested number of views for Multiview render pass"); + + render_pass_multiview_create_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_MULTIVIEW_CREATE_INFO; + render_pass_multiview_create_info.pNext = nullptr; + render_pass_multiview_create_info.subpassCount = 1; + render_pass_multiview_create_info.pViewMasks = &view_mask; + render_pass_multiview_create_info.dependencyCount = 0; + render_pass_multiview_create_info.pViewOffsets = nullptr; + render_pass_multiview_create_info.correlationMaskCount = 1; + render_pass_multiview_create_info.pCorrelationMasks = &correlation_mask; + + render_pass_create_info.pNext = &render_pass_multiview_create_info; + } VkRenderPass render_pass; VkResult res = vkCreateRenderPass(device, &render_pass_create_info, nullptr, &render_pass); ERR_FAIL_COND_V_MSG(res, VK_NULL_HANDLE, "vkCreateRenderPass failed with error " + itos(res) + "."); - if (r_color_attachment_count) { - *r_color_attachment_count = color_references.size(); - } return render_pass; } -RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_create(const Vector<AttachmentFormat> &p_format) { +RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_create(const Vector<AttachmentFormat> &p_format, uint32_t p_view_count) { + FramebufferPass pass; + for (int i = 0; i < p_format.size(); i++) { + if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + pass.depth_attachment = i; + } else { + pass.color_attachments.push_back(i); + } + } + + Vector<FramebufferPass> passes; + passes.push_back(pass); + return framebuffer_format_create_multipass(p_format, passes, p_view_count); +} +RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_create_multipass(const Vector<AttachmentFormat> &p_attachments, Vector<FramebufferPass> &p_passes, uint32_t p_view_count) { _THREAD_SAFE_METHOD_ FramebufferFormatKey key; - key.attachments = p_format; + key.attachments = p_attachments; + key.passes = p_passes; + key.view_count = p_view_count; const Map<FramebufferFormatKey, FramebufferFormatID>::Element *E = framebuffer_format_cache.find(key); if (E) { @@ -3491,8 +3682,8 @@ RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_c return E->get(); } - int color_references; - VkRenderPass render_pass = _render_pass_create(p_format, INITIAL_ACTION_CLEAR, FINAL_ACTION_READ, INITIAL_ACTION_CLEAR, FINAL_ACTION_READ, &color_references); //actions don't matter for this use case + Vector<TextureSamples> samples; + VkRenderPass render_pass = _render_pass_create(p_attachments, p_passes, INITIAL_ACTION_CLEAR, FINAL_ACTION_READ, INITIAL_ACTION_CLEAR, FINAL_ACTION_READ, p_view_count, &samples); //actions don't matter for this use case if (render_pass == VK_NULL_HANDLE) { //was likely invalid return INVALID_ID; @@ -3502,15 +3693,16 @@ RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_c E = framebuffer_format_cache.insert(key, id); FramebufferFormat fb_format; fb_format.E = E; - fb_format.color_attachments = color_references; fb_format.render_pass = render_pass; - fb_format.samples = p_format[0].samples; + fb_format.pass_samples = samples; + fb_format.view_count = p_view_count; framebuffer_formats[id] = fb_format; return id; } RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_create_empty(TextureSamples p_samples) { FramebufferFormatKey key; + key.passes.push_back(FramebufferPass()); const Map<FramebufferFormatKey, FramebufferFormatID>::Element *E = framebuffer_format_cache.find(key); if (E) { @@ -3556,18 +3748,18 @@ RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_c FramebufferFormat fb_format; fb_format.E = E; - fb_format.color_attachments = 0; fb_format.render_pass = render_pass; - fb_format.samples = p_samples; + fb_format.pass_samples.push_back(p_samples); framebuffer_formats[id] = fb_format; return id; } -RenderingDevice::TextureSamples RenderingDeviceVulkan::framebuffer_format_get_texture_samples(FramebufferFormatID p_format) { +RenderingDevice::TextureSamples RenderingDeviceVulkan::framebuffer_format_get_texture_samples(FramebufferFormatID p_format, uint32_t p_pass) { Map<FramebufferFormatID, FramebufferFormat>::Element *E = framebuffer_formats.find(p_format); ERR_FAIL_COND_V(!E, TEXTURE_SAMPLES_1); + ERR_FAIL_COND_V(p_pass >= uint32_t(E->get().pass_samples.size()), TEXTURE_SAMPLES_1); - return E->get().samples; + return E->get().pass_samples[p_pass]; } /***********************/ @@ -3580,11 +3772,36 @@ RID RenderingDeviceVulkan::framebuffer_create_empty(const Size2i &p_size, Textur framebuffer.format_id = framebuffer_format_create_empty(p_samples); ERR_FAIL_COND_V(p_format_check != INVALID_FORMAT_ID && framebuffer.format_id != p_format_check, RID()); framebuffer.size = p_size; + framebuffer.view_count = 1; return framebuffer_owner.make_rid(framebuffer); } -RID RenderingDeviceVulkan::framebuffer_create(const Vector<RID> &p_texture_attachments, FramebufferFormatID p_format_check) { +RID RenderingDeviceVulkan::framebuffer_create(const Vector<RID> &p_texture_attachments, FramebufferFormatID p_format_check, uint32_t p_view_count) { + _THREAD_SAFE_METHOD_ + + FramebufferPass pass; + + for (int i = 0; i < p_texture_attachments.size(); i++) { + Texture *texture = texture_owner.getornull(p_texture_attachments[i]); + ERR_FAIL_COND_V_MSG(!texture, RID(), "Texture index supplied for framebuffer (" + itos(i) + ") is not a valid texture."); + + ERR_FAIL_COND_V_MSG(texture->layers != p_view_count, RID(), "Layers of our texture doesn't match view count for this framebuffer"); + + if (texture->usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + pass.depth_attachment = i; + } else { + pass.color_attachments.push_back(i); + } + } + + Vector<FramebufferPass> passes; + passes.push_back(pass); + + return framebuffer_create_multipass(p_texture_attachments, passes, p_format_check, p_view_count); +} + +RID RenderingDeviceVulkan::framebuffer_create_multipass(const Vector<RID> &p_texture_attachments, Vector<FramebufferPass> &p_passes, FramebufferFormatID p_format_check, uint32_t p_view_count) { _THREAD_SAFE_METHOD_ Vector<AttachmentFormat> attachments; @@ -3594,6 +3811,8 @@ RID RenderingDeviceVulkan::framebuffer_create(const Vector<RID> &p_texture_attac Texture *texture = texture_owner.getornull(p_texture_attachments[i]); ERR_FAIL_COND_V_MSG(!texture, RID(), "Texture index supplied for framebuffer (" + itos(i) + ") is not a valid texture."); + ERR_FAIL_COND_V_MSG(texture->layers != p_view_count, RID(), "Layers of our texture doesn't match view count for this framebuffer"); + if (i == 0) { size.width = texture->width; size.height = texture->height; @@ -3609,7 +3828,7 @@ RID RenderingDeviceVulkan::framebuffer_create(const Vector<RID> &p_texture_attac attachments.push_back(af); } - FramebufferFormatID format_id = framebuffer_format_create(attachments); + FramebufferFormatID format_id = framebuffer_format_create_multipass(attachments, p_passes, p_view_count); if (format_id == INVALID_ID) { return RID(); } @@ -3621,6 +3840,7 @@ RID RenderingDeviceVulkan::framebuffer_create(const Vector<RID> &p_texture_attac framebuffer.format_id = format_id; framebuffer.texture_ids = p_texture_attachments; framebuffer.size = size; + framebuffer.view_count = p_view_count; RID id = framebuffer_owner.make_rid(framebuffer); @@ -4370,8 +4590,9 @@ RID RenderingDeviceVulkan::shader_create(const Vector<ShaderStageData> &p_stages "Reflection of SPIR-V shader stage '" + String(shader_stage_names[p_stages[i].shader_stage]) + "' failed obtaining output variables."); for (uint32_t j = 0; j < ov_count; j++) { - if (output_vars[j]) { - fragment_outputs = MAX(fragment_outputs, output_vars[j]->location + 1); + const SpvReflectInterfaceVariable *refvar = output_vars[j]; + if (refvar != nullptr && refvar->built_in != SpvBuiltInFragDepth) { + fragment_outputs |= 1 << refvar->location; } } } @@ -4422,7 +4643,7 @@ RID RenderingDeviceVulkan::shader_create(const Vector<ShaderStageData> &p_stages Shader shader; shader.vertex_input_mask = vertex_input_mask; - shader.fragment_outputs = fragment_outputs; + shader.fragment_output_mask = fragment_outputs; shader.push_constant = push_constant; shader.is_compute = is_compute; shader.compute_local_size[0] = compute_local_size[0]; @@ -5163,6 +5384,49 @@ RID RenderingDeviceVulkan::uniform_set_create(const Vector<Uniform> &p_uniforms, write.pTexelBufferView = nullptr; } break; case UNIFORM_TYPE_INPUT_ATTACHMENT: { + ERR_FAIL_COND_V_MSG(shader->is_compute, RID(), "InputAttachment (binding: " + itos(uniform.binding) + ") supplied for compute shader (this is not allowed)."); + + if (uniform.ids.size() != set_uniform.length) { + if (set_uniform.length > 1) { + ERR_FAIL_V_MSG(RID(), "InputAttachment (binding: " + itos(uniform.binding) + ") is an array of (" + itos(set_uniform.length) + ") textures, so it should be provided equal number of texture IDs to satisfy it (IDs provided: " + itos(uniform.ids.size()) + ")."); + } else { + ERR_FAIL_V_MSG(RID(), "InputAttachment (binding: " + itos(uniform.binding) + ") should provide one ID referencing a texture (IDs provided: " + itos(uniform.ids.size()) + ")."); + } + } + + Vector<VkDescriptorImageInfo> image_info; + + for (int j = 0; j < uniform.ids.size(); j++) { + Texture *texture = texture_owner.getornull(uniform.ids[j]); + + ERR_FAIL_COND_V_MSG(!texture, RID(), + "InputAttachment (binding: " + itos(uniform.binding) + ", index " + itos(j) + ") is not a valid texture."); + + ERR_FAIL_COND_V_MSG(!(texture->usage_flags & TEXTURE_USAGE_SAMPLING_BIT), RID(), + "InputAttachment (binding: " + itos(uniform.binding) + ", index " + itos(j) + ") needs the TEXTURE_USAGE_SAMPLING_BIT usage flag set in order to be used as uniform."); + + VkDescriptorImageInfo img_info; + img_info.sampler = VK_NULL_HANDLE; + img_info.imageView = texture->view; + + if (texture->owner.is_valid()) { + texture = texture_owner.getornull(texture->owner); + ERR_FAIL_COND_V(!texture, RID()); //bug, should never happen + } + + img_info.imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + + image_info.push_back(img_info); + } + + write.dstArrayElement = 0; + write.descriptorCount = uniform.ids.size(); + write.descriptorType = VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT; + write.pImageInfo = image_infos.push_back(image_info)->get().ptr(); + write.pBufferInfo = nullptr; + write.pTexelBufferView = nullptr; + + type_size = uniform.ids.size(); } break; default: { } @@ -5373,7 +5637,7 @@ Vector<uint8_t> RenderingDeviceVulkan::buffer_get_data(RID p_buffer) { /**** RENDER PIPELINE ****/ /*************************/ -RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferFormatID p_framebuffer_format, VertexFormatID p_vertex_format, RenderPrimitive p_render_primitive, const PipelineRasterizationState &p_rasterization_state, const PipelineMultisampleState &p_multisample_state, const PipelineDepthStencilState &p_depth_stencil_state, const PipelineColorBlendState &p_blend_state, int p_dynamic_state_flags) { +RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferFormatID p_framebuffer_format, VertexFormatID p_vertex_format, RenderPrimitive p_render_primitive, const PipelineRasterizationState &p_rasterization_state, const PipelineMultisampleState &p_multisample_state, const PipelineDepthStencilState &p_depth_stencil_state, const PipelineColorBlendState &p_blend_state, int p_dynamic_state_flags, uint32_t p_for_render_pass) { _THREAD_SAFE_METHOD_ //needs a shader @@ -5392,8 +5656,16 @@ RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferForma { //validate shader vs framebuffer - ERR_FAIL_COND_V_MSG(shader->fragment_outputs != fb_format.color_attachments, RID(), - "Mismatch fragment output bindings (" + itos(shader->fragment_outputs) + ") and framebuffer color buffers (" + itos(fb_format.color_attachments) + ") when binding both in render pipeline."); + ERR_FAIL_COND_V_MSG(p_for_render_pass >= uint32_t(fb_format.E->key().passes.size()), RID(), "Render pass requested for pipeline creation (" + itos(p_for_render_pass) + ") is out of bounds"); + const FramebufferPass &pass = fb_format.E->key().passes[p_for_render_pass]; + uint32_t output_mask = 0; + for (int i = 0; i < pass.color_attachments.size(); i++) { + if (pass.color_attachments[i] != FramebufferPass::ATTACHMENT_UNUSED) { + output_mask |= 1 << i; + } + } + ERR_FAIL_COND_V_MSG(shader->fragment_output_mask != output_mask, RID(), + "Mismatch fragment shader output mask (" + itos(shader->fragment_output_mask) + ") and framebuffer color output mask (" + itos(output_mask) + ") when binding both in render pipeline."); } //vertex VkPipelineVertexInputStateCreateInfo pipeline_vertex_input_state_create_info; @@ -5578,44 +5850,53 @@ RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferForma ERR_FAIL_INDEX_V(p_blend_state.logic_op, LOGIC_OP_MAX, RID()); color_blend_state_create_info.logicOp = logic_operations[p_blend_state.logic_op]; - ERR_FAIL_COND_V(fb_format.color_attachments != p_blend_state.attachments.size(), RID()); - Vector<VkPipelineColorBlendAttachmentState> attachment_states; + { + const FramebufferPass &pass = fb_format.E->key().passes[p_for_render_pass]; + + for (int i = 0; i < pass.color_attachments.size(); i++) { + if (pass.color_attachments[i] != FramebufferPass::ATTACHMENT_UNUSED) { + int idx = attachment_states.size(); + + ERR_FAIL_INDEX_V(idx, p_blend_state.attachments.size(), RID()); + VkPipelineColorBlendAttachmentState state; + state.blendEnable = p_blend_state.attachments[idx].enable_blend; + + ERR_FAIL_INDEX_V(p_blend_state.attachments[idx].src_color_blend_factor, BLEND_FACTOR_MAX, RID()); + state.srcColorBlendFactor = blend_factors[p_blend_state.attachments[idx].src_color_blend_factor]; + ERR_FAIL_INDEX_V(p_blend_state.attachments[idx].dst_color_blend_factor, BLEND_FACTOR_MAX, RID()); + state.dstColorBlendFactor = blend_factors[p_blend_state.attachments[idx].dst_color_blend_factor]; + ERR_FAIL_INDEX_V(p_blend_state.attachments[idx].color_blend_op, BLEND_OP_MAX, RID()); + state.colorBlendOp = blend_operations[p_blend_state.attachments[idx].color_blend_op]; + + ERR_FAIL_INDEX_V(p_blend_state.attachments[idx].src_alpha_blend_factor, BLEND_FACTOR_MAX, RID()); + state.srcAlphaBlendFactor = blend_factors[p_blend_state.attachments[idx].src_alpha_blend_factor]; + ERR_FAIL_INDEX_V(p_blend_state.attachments[idx].dst_alpha_blend_factor, BLEND_FACTOR_MAX, RID()); + state.dstAlphaBlendFactor = blend_factors[p_blend_state.attachments[idx].dst_alpha_blend_factor]; + ERR_FAIL_INDEX_V(p_blend_state.attachments[idx].alpha_blend_op, BLEND_OP_MAX, RID()); + state.alphaBlendOp = blend_operations[p_blend_state.attachments[idx].alpha_blend_op]; + + state.colorWriteMask = 0; + if (p_blend_state.attachments[idx].write_r) { + state.colorWriteMask |= VK_COLOR_COMPONENT_R_BIT; + } + if (p_blend_state.attachments[idx].write_g) { + state.colorWriteMask |= VK_COLOR_COMPONENT_G_BIT; + } + if (p_blend_state.attachments[idx].write_b) { + state.colorWriteMask |= VK_COLOR_COMPONENT_B_BIT; + } + if (p_blend_state.attachments[idx].write_a) { + state.colorWriteMask |= VK_COLOR_COMPONENT_A_BIT; + } - for (int i = 0; i < p_blend_state.attachments.size(); i++) { - VkPipelineColorBlendAttachmentState state; - state.blendEnable = p_blend_state.attachments[i].enable_blend; - - ERR_FAIL_INDEX_V(p_blend_state.attachments[i].src_color_blend_factor, BLEND_FACTOR_MAX, RID()); - state.srcColorBlendFactor = blend_factors[p_blend_state.attachments[i].src_color_blend_factor]; - ERR_FAIL_INDEX_V(p_blend_state.attachments[i].dst_color_blend_factor, BLEND_FACTOR_MAX, RID()); - state.dstColorBlendFactor = blend_factors[p_blend_state.attachments[i].dst_color_blend_factor]; - ERR_FAIL_INDEX_V(p_blend_state.attachments[i].color_blend_op, BLEND_OP_MAX, RID()); - state.colorBlendOp = blend_operations[p_blend_state.attachments[i].color_blend_op]; - - ERR_FAIL_INDEX_V(p_blend_state.attachments[i].src_alpha_blend_factor, BLEND_FACTOR_MAX, RID()); - state.srcAlphaBlendFactor = blend_factors[p_blend_state.attachments[i].src_alpha_blend_factor]; - ERR_FAIL_INDEX_V(p_blend_state.attachments[i].dst_alpha_blend_factor, BLEND_FACTOR_MAX, RID()); - state.dstAlphaBlendFactor = blend_factors[p_blend_state.attachments[i].dst_alpha_blend_factor]; - ERR_FAIL_INDEX_V(p_blend_state.attachments[i].alpha_blend_op, BLEND_OP_MAX, RID()); - state.alphaBlendOp = blend_operations[p_blend_state.attachments[i].alpha_blend_op]; - - state.colorWriteMask = 0; - if (p_blend_state.attachments[i].write_r) { - state.colorWriteMask |= VK_COLOR_COMPONENT_R_BIT; - } - if (p_blend_state.attachments[i].write_g) { - state.colorWriteMask |= VK_COLOR_COMPONENT_G_BIT; - } - if (p_blend_state.attachments[i].write_b) { - state.colorWriteMask |= VK_COLOR_COMPONENT_B_BIT; - } - if (p_blend_state.attachments[i].write_a) { - state.colorWriteMask |= VK_COLOR_COMPONENT_A_BIT; + attachment_states.push_back(state); + idx++; + }; } - attachment_states.push_back(state); - }; + ERR_FAIL_COND_V(attachment_states.size() != p_blend_state.attachments.size(), RID()); + } color_blend_state_create_info.attachmentCount = attachment_states.size(); color_blend_state_create_info.pAttachments = attachment_states.ptr(); @@ -5688,7 +5969,7 @@ RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferForma graphics_pipeline_create_info.layout = shader->pipeline_layout; graphics_pipeline_create_info.renderPass = fb_format.render_pass; - graphics_pipeline_create_info.subpass = 0; + graphics_pipeline_create_info.subpass = p_for_render_pass; graphics_pipeline_create_info.basePipelineHandle = VK_NULL_HANDLE; graphics_pipeline_create_info.basePipelineIndex = 0; @@ -5705,6 +5986,7 @@ RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferForma #ifdef DEBUG_ENABLED pipeline.validation.dynamic_state = p_dynamic_state_flags; pipeline.validation.framebuffer_format = p_framebuffer_format; + pipeline.validation.render_pass = p_for_render_pass; pipeline.validation.vertex_format = p_vertex_format; pipeline.validation.uses_restart_indices = input_assembly_create_info.primitiveRestartEnable; @@ -5843,13 +6125,14 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin_for_screen(Di ERR_FAIL_COND_V_MSG(compute_list != nullptr, INVALID_ID, "Only one draw/compute list can be active at the same time."); VkCommandBuffer command_buffer = frames[frame].draw_command_buffer; - draw_list = memnew(DrawList); - draw_list->command_buffer = command_buffer; + + Size2i size = Size2i(context->window_get_width(p_screen), context->window_get_height(p_screen)); + + _draw_list_allocate(Rect2i(Vector2i(), size), 0, 0); #ifdef DEBUG_ENABLED - draw_list->validation.framebuffer_format = screen_get_framebuffer_format(); + draw_list_framebuffer_format = screen_get_framebuffer_format(); #endif - draw_list_count = 0; - draw_list_split = false; + draw_list_subpass_count = 1; VkRenderPassBeginInfo render_pass_begin; render_pass_begin.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO; @@ -5857,8 +6140,8 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin_for_screen(Di render_pass_begin.renderPass = context->window_get_render_pass(p_screen); render_pass_begin.framebuffer = context->window_get_framebuffer(p_screen); - render_pass_begin.renderArea.extent.width = context->window_get_width(p_screen); - render_pass_begin.renderArea.extent.height = context->window_get_height(p_screen); + render_pass_begin.renderArea.extent.width = size.width; + render_pass_begin.renderArea.extent.height = size.height; render_pass_begin.renderArea.offset.x = 0; render_pass_begin.renderArea.offset.y = 0; @@ -5898,18 +6181,19 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin_for_screen(Di return int64_t(ID_TYPE_DRAW_LIST) << ID_BASE_SHIFT; } -Error RenderingDeviceVulkan::_draw_list_setup_framebuffer(Framebuffer *p_framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, VkFramebuffer *r_framebuffer, VkRenderPass *r_render_pass) { +Error RenderingDeviceVulkan::_draw_list_setup_framebuffer(Framebuffer *p_framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, VkFramebuffer *r_framebuffer, VkRenderPass *r_render_pass, uint32_t *r_subpass_count) { Framebuffer::VersionKey vk; vk.initial_color_action = p_initial_color_action; vk.final_color_action = p_final_color_action; vk.initial_depth_action = p_initial_depth_action; vk.final_depth_action = p_final_depth_action; + vk.view_count = p_framebuffer->view_count; if (!p_framebuffer->framebuffers.has(vk)) { //need to create this version Framebuffer::Version version; - version.render_pass = _render_pass_create(framebuffer_formats[p_framebuffer->format_id].E->key().attachments, p_initial_color_action, p_final_color_action, p_initial_depth_action, p_final_depth_action); + version.render_pass = _render_pass_create(framebuffer_formats[p_framebuffer->format_id].E->key().attachments, framebuffer_formats[p_framebuffer->format_id].E->key().passes, p_initial_color_action, p_final_color_action, p_initial_depth_action, p_final_depth_action, p_framebuffer->view_count); VkFramebufferCreateInfo framebuffer_create_info; framebuffer_create_info.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO; @@ -5933,11 +6217,14 @@ Error RenderingDeviceVulkan::_draw_list_setup_framebuffer(Framebuffer *p_framebu VkResult err = vkCreateFramebuffer(device, &framebuffer_create_info, nullptr, &version.framebuffer); ERR_FAIL_COND_V_MSG(err, ERR_CANT_CREATE, "vkCreateFramebuffer failed with error " + itos(err) + "."); + version.subpass_count = framebuffer_formats[p_framebuffer->format_id].E->key().passes.size(); + p_framebuffer->framebuffers.insert(vk, version); } const Framebuffer::Version &version = p_framebuffer->framebuffers[vk]; *r_framebuffer = version.framebuffer; *r_render_pass = version.render_pass; + *r_subpass_count = version.subpass_count; return OK; } @@ -6127,15 +6414,23 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_framebu if (p_initial_color_action == INITIAL_ACTION_CLEAR) { //check clear values - int color_attachments = framebuffer_formats[framebuffer->format_id].color_attachments; - ERR_FAIL_COND_V_MSG(p_clear_color_values.size() != color_attachments, INVALID_ID, - "Clear color values supplied (" + itos(p_clear_color_values.size()) + ") differ from the amount required for framebuffer (" + itos(color_attachments) + ")."); + int color_count = 0; + for (int i = 0; i < framebuffer->texture_ids.size(); i++) { + Texture *texture = texture_owner.getornull(framebuffer->texture_ids[i]); + + if (texture->usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + color_count++; + } + } + + ERR_FAIL_COND_V_MSG(p_clear_color_values.size() != color_count, INVALID_ID, + "Clear color values supplied (" + itos(p_clear_color_values.size()) + ") differ from the amount required for framebuffer color attachments (" + itos(color_count) + ")."); } VkFramebuffer vkframebuffer; VkRenderPass render_pass; - Error err = _draw_list_setup_framebuffer(framebuffer, p_initial_color_action, p_final_color_action, p_initial_depth_action, p_final_depth_action, &vkframebuffer, &render_pass); + Error err = _draw_list_setup_framebuffer(framebuffer, p_initial_color_action, p_final_color_action, p_initial_depth_action, p_final_depth_action, &vkframebuffer, &render_pass, &draw_list_subpass_count); ERR_FAIL_COND_V(err != OK, INVALID_ID); VkCommandBuffer command_buffer = frames[frame].draw_command_buffer; @@ -6145,13 +6440,14 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_framebu return INVALID_ID; } - draw_list = memnew(DrawList); - draw_list->command_buffer = command_buffer; + draw_list_render_pass = render_pass; + draw_list_vkframebuffer = vkframebuffer; + + _draw_list_allocate(Rect2i(viewport_offset, viewport_size), 0, 0); #ifdef DEBUG_ENABLED - draw_list->validation.framebuffer_format = framebuffer->format_id; + draw_list_framebuffer_format = framebuffer->format_id; #endif - draw_list_count = 0; - draw_list_split = false; + draw_list_current_subpass = 0; if (needs_clear_color || needs_clear_depth) { _draw_list_insert_clear_region(draw_list, framebuffer, viewport_offset, viewport_size, needs_clear_color, p_clear_color_values, needs_clear_depth, p_clear_depth, p_clear_stencil); @@ -6175,7 +6471,6 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_framebu vkCmdSetScissor(command_buffer, 0, 1, &scissor); - draw_list->viewport = Rect2i(viewport_offset, viewport_size); return int64_t(ID_TYPE_DRAW_LIST) << ID_BASE_SHIFT; } @@ -6217,47 +6512,23 @@ Error RenderingDeviceVulkan::draw_list_begin_split(RID p_framebuffer, uint32_t p if (p_initial_color_action == INITIAL_ACTION_CLEAR) { //check clear values - int color_attachments = framebuffer_formats[framebuffer->format_id].color_attachments; - ERR_FAIL_COND_V_MSG(p_clear_color_values.size() != color_attachments, ERR_INVALID_PARAMETER, - "Clear color values supplied (" + itos(p_clear_color_values.size()) + ") differ from the amount required for framebuffer (" + itos(color_attachments) + ")."); - } - - if (p_splits > (uint32_t)split_draw_list_allocators.size()) { - uint32_t from = split_draw_list_allocators.size(); - split_draw_list_allocators.resize(p_splits); - for (uint32_t i = from; i < p_splits; i++) { - VkCommandPoolCreateInfo cmd_pool_info; - cmd_pool_info.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO; - cmd_pool_info.pNext = nullptr; - cmd_pool_info.queueFamilyIndex = context->get_graphics_queue(); - cmd_pool_info.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT; - - VkResult res = vkCreateCommandPool(device, &cmd_pool_info, nullptr, &split_draw_list_allocators.write[i].command_pool); - ERR_FAIL_COND_V_MSG(res, ERR_CANT_CREATE, "vkCreateCommandPool failed with error " + itos(res) + "."); - - for (int j = 0; j < frame_count; j++) { - VkCommandBuffer command_buffer; - - VkCommandBufferAllocateInfo cmdbuf; - //no command buffer exists, create it. - cmdbuf.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO; - cmdbuf.pNext = nullptr; - cmdbuf.commandPool = split_draw_list_allocators[i].command_pool; - cmdbuf.level = VK_COMMAND_BUFFER_LEVEL_SECONDARY; - cmdbuf.commandBufferCount = 1; - - VkResult err = vkAllocateCommandBuffers(device, &cmdbuf, &command_buffer); - ERR_FAIL_COND_V_MSG(err, ERR_CANT_CREATE, "vkAllocateCommandBuffers failed with error " + itos(err) + "."); + int color_count = 0; + for (int i = 0; i < framebuffer->texture_ids.size(); i++) { + Texture *texture = texture_owner.getornull(framebuffer->texture_ids[i]); - split_draw_list_allocators.write[i].command_buffers.push_back(command_buffer); + if (texture->usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + color_count++; } } + + ERR_FAIL_COND_V_MSG(p_clear_color_values.size() != color_count, ERR_INVALID_PARAMETER, + "Clear color values supplied (" + itos(p_clear_color_values.size()) + ") differ from the amount required for framebuffer (" + itos(color_count) + ")."); } VkFramebuffer vkframebuffer; VkRenderPass render_pass; - Error err = _draw_list_setup_framebuffer(framebuffer, p_initial_color_action, p_final_color_action, p_initial_depth_action, p_final_depth_action, &vkframebuffer, &render_pass); + Error err = _draw_list_setup_framebuffer(framebuffer, p_initial_color_action, p_final_color_action, p_initial_depth_action, p_final_depth_action, &vkframebuffer, &render_pass, &draw_list_subpass_count); ERR_FAIL_COND_V(err != OK, ERR_CANT_CREATE); VkCommandBuffer frame_command_buffer = frames[frame].draw_command_buffer; @@ -6267,53 +6538,24 @@ Error RenderingDeviceVulkan::draw_list_begin_split(RID p_framebuffer, uint32_t p return ERR_CANT_CREATE; } - draw_list = memnew_arr(DrawList, p_splits); - draw_list_count = p_splits; - draw_list_split = true; - - for (uint32_t i = 0; i < p_splits; i++) { - //take a command buffer and initialize it - VkCommandBuffer command_buffer = split_draw_list_allocators[i].command_buffers[frame]; - - VkCommandBufferInheritanceInfo inheritance_info; - inheritance_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_INHERITANCE_INFO; - inheritance_info.pNext = nullptr; - inheritance_info.renderPass = render_pass; - inheritance_info.subpass = 0; - inheritance_info.framebuffer = vkframebuffer; - inheritance_info.occlusionQueryEnable = false; - inheritance_info.queryFlags = 0; //? - inheritance_info.pipelineStatistics = 0; - - VkCommandBufferBeginInfo cmdbuf_begin; - cmdbuf_begin.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO; - cmdbuf_begin.pNext = nullptr; - cmdbuf_begin.flags = VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT | VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT; - cmdbuf_begin.pInheritanceInfo = &inheritance_info; - - VkResult res = vkResetCommandBuffer(command_buffer, 0); - if (res) { - memdelete_arr(draw_list); - draw_list = nullptr; - ERR_FAIL_V_MSG(ERR_CANT_CREATE, "vkResetCommandBuffer failed with error " + itos(res) + "."); - } - - res = vkBeginCommandBuffer(command_buffer, &cmdbuf_begin); - if (res) { - memdelete_arr(draw_list); - draw_list = nullptr; - ERR_FAIL_V_MSG(ERR_CANT_CREATE, "vkBeginCommandBuffer failed with error " + itos(res) + "."); - } + draw_list_current_subpass = 0; - draw_list[i].command_buffer = command_buffer; #ifdef DEBUG_ENABLED - draw_list[i].validation.framebuffer_format = framebuffer->format_id; + draw_list_framebuffer_format = framebuffer->format_id; #endif + draw_list_render_pass = render_pass; + draw_list_vkframebuffer = vkframebuffer; - if (i == 0 && (needs_clear_color || needs_clear_depth)) { - _draw_list_insert_clear_region(draw_list, framebuffer, viewport_offset, viewport_size, needs_clear_color, p_clear_color_values, needs_clear_depth, p_clear_depth, p_clear_stencil); - } + err = _draw_list_allocate(Rect2i(viewport_offset, viewport_size), p_splits, 0); + if (err != OK) { + return err; + } + + if (needs_clear_color || needs_clear_depth) { + _draw_list_insert_clear_region(&draw_list[0], framebuffer, viewport_offset, viewport_size, needs_clear_color, p_clear_color_values, needs_clear_depth, p_clear_depth, p_clear_stencil); + } + for (uint32_t i = 0; i < p_splits; i++) { VkViewport viewport; viewport.x = viewport_offset.x; viewport.y = viewport_offset.y; @@ -6322,7 +6564,7 @@ Error RenderingDeviceVulkan::draw_list_begin_split(RID p_framebuffer, uint32_t p viewport.minDepth = 0; viewport.maxDepth = 1.0; - vkCmdSetViewport(command_buffer, 0, 1, &viewport); + vkCmdSetViewport(draw_list[i].command_buffer, 0, 1, &viewport); VkRect2D scissor; scissor.offset.x = viewport_offset.x; @@ -6330,10 +6572,8 @@ Error RenderingDeviceVulkan::draw_list_begin_split(RID p_framebuffer, uint32_t p scissor.extent.width = viewport_size.width; scissor.extent.height = viewport_size.height; - vkCmdSetScissor(command_buffer, 0, 1, &scissor); + vkCmdSetScissor(draw_list[i].command_buffer, 0, 1, &scissor); r_split_ids[i] = (int64_t(ID_TYPE_SPLIT_DRAW_LIST) << ID_BASE_SHIFT) + i; - - draw_list[i].viewport = Rect2i(viewport_offset, viewport_size); } return OK; @@ -6378,7 +6618,7 @@ void RenderingDeviceVulkan::draw_list_bind_render_pipeline(DrawListID p_list, RI const RenderPipeline *pipeline = render_pipeline_owner.getornull(p_render_pipeline); ERR_FAIL_COND(!pipeline); #ifdef DEBUG_ENABLED - ERR_FAIL_COND(pipeline->validation.framebuffer_format != dl->validation.framebuffer_format); + ERR_FAIL_COND(pipeline->validation.framebuffer_format != draw_list_framebuffer_format && pipeline->validation.render_pass != draw_list_current_subpass); #endif if (p_render_pipeline == dl->state.pipeline) { @@ -6717,30 +6957,162 @@ void RenderingDeviceVulkan::draw_list_disable_scissor(DrawListID p_list) { vkCmdSetScissor(dl->command_buffer, 0, 1, &scissor); } -void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { - _THREAD_SAFE_METHOD_ +RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_switch_to_next_pass() { + ERR_FAIL_COND_V(draw_list == nullptr, INVALID_ID); + ERR_FAIL_COND_V(draw_list_current_subpass >= draw_list_subpass_count - 1, INVALID_FORMAT_ID); - ERR_FAIL_COND_MSG(!draw_list, "Immediate draw list is already inactive."); + draw_list_current_subpass++; + + Rect2i viewport; + _draw_list_free(&viewport); + + vkCmdNextSubpass(frames[frame].draw_command_buffer, VK_SUBPASS_CONTENTS_INLINE); + + _draw_list_allocate(viewport, 0, draw_list_current_subpass); + return int64_t(ID_TYPE_DRAW_LIST) << ID_BASE_SHIFT; +} +Error RenderingDeviceVulkan::draw_list_switch_to_next_pass_split(uint32_t p_splits, DrawListID *r_split_ids) { + ERR_FAIL_COND_V(draw_list == nullptr, ERR_INVALID_PARAMETER); + ERR_FAIL_COND_V(draw_list_current_subpass >= draw_list_subpass_count - 1, ERR_INVALID_PARAMETER); + + draw_list_current_subpass++; + + Rect2i viewport; + _draw_list_free(&viewport); + + vkCmdNextSubpass(frames[frame].draw_command_buffer, VK_SUBPASS_CONTENTS_INLINE); + + _draw_list_allocate(viewport, p_splits, draw_list_current_subpass); + + for (uint32_t i = 0; i < p_splits; i++) { + r_split_ids[i] = (int64_t(ID_TYPE_SPLIT_DRAW_LIST) << ID_BASE_SHIFT) + i; + } + + return OK; +} + +Error RenderingDeviceVulkan::_draw_list_allocate(const Rect2i &p_viewport, uint32_t p_splits, uint32_t p_subpass) { + if (p_splits == 0) { + draw_list = memnew(DrawList); + draw_list->command_buffer = frames[frame].draw_command_buffer; + draw_list->viewport = p_viewport; + draw_list_count = 0; + draw_list_split = false; + } else { + if (p_splits > (uint32_t)split_draw_list_allocators.size()) { + uint32_t from = split_draw_list_allocators.size(); + split_draw_list_allocators.resize(p_splits); + for (uint32_t i = from; i < p_splits; i++) { + VkCommandPoolCreateInfo cmd_pool_info; + cmd_pool_info.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO; + cmd_pool_info.pNext = nullptr; + cmd_pool_info.queueFamilyIndex = context->get_graphics_queue(); + cmd_pool_info.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT; + + VkResult res = vkCreateCommandPool(device, &cmd_pool_info, nullptr, &split_draw_list_allocators.write[i].command_pool); + ERR_FAIL_COND_V_MSG(res, ERR_CANT_CREATE, "vkCreateCommandPool failed with error " + itos(res) + "."); + + for (int j = 0; j < frame_count; j++) { + VkCommandBuffer command_buffer; + + VkCommandBufferAllocateInfo cmdbuf; + //no command buffer exists, create it. + cmdbuf.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO; + cmdbuf.pNext = nullptr; + cmdbuf.commandPool = split_draw_list_allocators[i].command_pool; + cmdbuf.level = VK_COMMAND_BUFFER_LEVEL_SECONDARY; + cmdbuf.commandBufferCount = 1; + + VkResult err = vkAllocateCommandBuffers(device, &cmdbuf, &command_buffer); + ERR_FAIL_COND_V_MSG(err, ERR_CANT_CREATE, "vkAllocateCommandBuffers failed with error " + itos(err) + "."); + + split_draw_list_allocators.write[i].command_buffers.push_back(command_buffer); + } + } + } + draw_list = memnew_arr(DrawList, p_splits); + draw_list_count = p_splits; + draw_list_split = true; + + for (uint32_t i = 0; i < p_splits; i++) { + //take a command buffer and initialize it + VkCommandBuffer command_buffer = split_draw_list_allocators[i].command_buffers[frame]; + + VkCommandBufferInheritanceInfo inheritance_info; + inheritance_info.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_INHERITANCE_INFO; + inheritance_info.pNext = nullptr; + inheritance_info.renderPass = draw_list_render_pass; + inheritance_info.subpass = p_subpass; + inheritance_info.framebuffer = draw_list_vkframebuffer; + inheritance_info.occlusionQueryEnable = false; + inheritance_info.queryFlags = 0; //? + inheritance_info.pipelineStatistics = 0; + + VkCommandBufferBeginInfo cmdbuf_begin; + cmdbuf_begin.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO; + cmdbuf_begin.pNext = nullptr; + cmdbuf_begin.flags = VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT | VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT; + cmdbuf_begin.pInheritanceInfo = &inheritance_info; + + VkResult res = vkResetCommandBuffer(command_buffer, 0); + if (res) { + memdelete_arr(draw_list); + draw_list = nullptr; + ERR_FAIL_V_MSG(ERR_CANT_CREATE, "vkResetCommandBuffer failed with error " + itos(res) + "."); + } + + res = vkBeginCommandBuffer(command_buffer, &cmdbuf_begin); + if (res) { + memdelete_arr(draw_list); + draw_list = nullptr; + ERR_FAIL_V_MSG(ERR_CANT_CREATE, "vkBeginCommandBuffer failed with error " + itos(res) + "."); + } + + draw_list[i].command_buffer = command_buffer; + draw_list[i].viewport = p_viewport; + } + } + + return OK; +} + +void RenderingDeviceVulkan::_draw_list_free(Rect2i *r_last_viewport) { if (draw_list_split) { //send all command buffers VkCommandBuffer *command_buffers = (VkCommandBuffer *)alloca(sizeof(VkCommandBuffer) * draw_list_count); for (uint32_t i = 0; i < draw_list_count; i++) { vkEndCommandBuffer(draw_list[i].command_buffer); command_buffers[i] = draw_list[i].command_buffer; + if (r_last_viewport) { + if (i == 0 || draw_list[i].viewport_set) { + *r_last_viewport = draw_list[i].viewport; + } + } } vkCmdExecuteCommands(frames[frame].draw_command_buffer, draw_list_count, command_buffers); - vkCmdEndRenderPass(frames[frame].draw_command_buffer); memdelete_arr(draw_list); draw_list = nullptr; } else { + if (r_last_viewport) { + *r_last_viewport = draw_list->viewport; + } //just end the list - vkCmdEndRenderPass(draw_list->command_buffer); memdelete(draw_list); draw_list = nullptr; } +} + +void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { + _THREAD_SAFE_METHOD_ + + ERR_FAIL_COND_MSG(!draw_list, "Immediate draw list is already inactive."); + + _draw_list_free(); + + vkCmdEndRenderPass(frames[frame].draw_command_buffer); for (int i = 0; i < draw_list_bound_textures.size(); i++) { Texture *texture = texture_owner.getornull(draw_list_bound_textures[i]); diff --git a/drivers/vulkan/rendering_device_vulkan.h b/drivers/vulkan/rendering_device_vulkan.h index a2527d5c33..9f5830103d 100644 --- a/drivers/vulkan/rendering_device_vulkan.h +++ b/drivers/vulkan/rendering_device_vulkan.h @@ -234,7 +234,87 @@ class RenderingDeviceVulkan : public RenderingDevice { struct FramebufferFormatKey { Vector<AttachmentFormat> attachments; + Vector<FramebufferPass> passes; + uint32_t view_count = 1; bool operator<(const FramebufferFormatKey &p_key) const { + if (view_count != p_key.view_count) { + return view_count < p_key.view_count; + } + + uint32_t pass_size = passes.size(); + uint32_t key_pass_size = p_key.passes.size(); + if (pass_size != key_pass_size) { + return pass_size < key_pass_size; + } + const FramebufferPass *pass_ptr = passes.ptr(); + const FramebufferPass *key_pass_ptr = p_key.passes.ptr(); + + for (uint32_t i = 0; i < pass_size; i++) { + { //compare color attachments + uint32_t attachment_size = pass_ptr[i].color_attachments.size(); + uint32_t key_attachment_size = key_pass_ptr[i].color_attachments.size(); + if (attachment_size != key_attachment_size) { + return attachment_size < key_attachment_size; + } + const int32_t *pass_attachment_ptr = pass_ptr[i].color_attachments.ptr(); + const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].color_attachments.ptr(); + + for (uint32_t j = 0; j < attachment_size; j++) { + if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { + return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; + } + } + } + { //compare input attachments + uint32_t attachment_size = pass_ptr[i].input_attachments.size(); + uint32_t key_attachment_size = key_pass_ptr[i].input_attachments.size(); + if (attachment_size != key_attachment_size) { + return attachment_size < key_attachment_size; + } + const int32_t *pass_attachment_ptr = pass_ptr[i].input_attachments.ptr(); + const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].input_attachments.ptr(); + + for (uint32_t j = 0; j < attachment_size; j++) { + if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { + return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; + } + } + } + { //compare resolve attachments + uint32_t attachment_size = pass_ptr[i].resolve_attachments.size(); + uint32_t key_attachment_size = key_pass_ptr[i].resolve_attachments.size(); + if (attachment_size != key_attachment_size) { + return attachment_size < key_attachment_size; + } + const int32_t *pass_attachment_ptr = pass_ptr[i].resolve_attachments.ptr(); + const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].resolve_attachments.ptr(); + + for (uint32_t j = 0; j < attachment_size; j++) { + if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { + return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; + } + } + } + { //compare preserve attachments + uint32_t attachment_size = pass_ptr[i].preserve_attachments.size(); + uint32_t key_attachment_size = key_pass_ptr[i].preserve_attachments.size(); + if (attachment_size != key_attachment_size) { + return attachment_size < key_attachment_size; + } + const int32_t *pass_attachment_ptr = pass_ptr[i].preserve_attachments.ptr(); + const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].preserve_attachments.ptr(); + + for (uint32_t j = 0; j < attachment_size; j++) { + if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { + return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; + } + } + } + if (pass_ptr[i].depth_attachment != key_pass_ptr[i].depth_attachment) { + return pass_ptr[i].depth_attachment < key_pass_ptr[i].depth_attachment; + } + } + int as = attachments.size(); int bs = p_key.attachments.size(); if (as != bs) { @@ -261,16 +341,15 @@ class RenderingDeviceVulkan : public RenderingDevice { } }; - VkRenderPass _render_pass_create(const Vector<AttachmentFormat> &p_format, InitialAction p_initial_action, FinalAction p_final_action, InitialAction p_initial_depth_action, FinalAction p_final_depthcolor_action, int *r_color_attachment_count = nullptr); - + VkRenderPass _render_pass_create(const Vector<AttachmentFormat> &p_attachments, const Vector<FramebufferPass> &p_passes, InitialAction p_initial_action, FinalAction p_final_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, uint32_t p_view_count = 1, Vector<TextureSamples> *r_samples = nullptr); // This is a cache and it's never freed, it ensures // IDs for a given format are always unique. Map<FramebufferFormatKey, FramebufferFormatID> framebuffer_format_cache; struct FramebufferFormat { const Map<FramebufferFormatKey, FramebufferFormatID>::Element *E; VkRenderPass render_pass = VK_NULL_HANDLE; //here for constructing shaders, never used, see section (7.2. Render Pass Compatibility from Vulkan spec) - int color_attachments = 0; //used for pipeline validation - TextureSamples samples; + Vector<TextureSamples> pass_samples; + uint32_t view_count = 1; // number of views }; Map<FramebufferFormatID, FramebufferFormat> framebuffer_formats; @@ -282,11 +361,17 @@ class RenderingDeviceVulkan : public RenderingDevice { FinalAction final_color_action; InitialAction initial_depth_action; FinalAction final_depth_action; + uint32_t view_count; + bool operator<(const VersionKey &p_key) const { if (initial_color_action == p_key.initial_color_action) { if (final_color_action == p_key.final_color_action) { if (initial_depth_action == p_key.initial_depth_action) { - return final_depth_action < p_key.final_depth_action; + if (final_depth_action == p_key.final_depth_action) { + return view_count < p_key.view_count; + } else { + return final_depth_action < p_key.final_depth_action; + } } else { return initial_depth_action < p_key.initial_depth_action; } @@ -305,10 +390,12 @@ class RenderingDeviceVulkan : public RenderingDevice { struct Version { VkFramebuffer framebuffer = VK_NULL_HANDLE; VkRenderPass render_pass = VK_NULL_HANDLE; //this one is owned + uint32_t subpass_count = 1; }; Map<VersionKey, Version> framebuffers; Size2 size; + uint32_t view_count; }; RID_Owner<Framebuffer, true> framebuffer_owner; @@ -524,7 +611,7 @@ class RenderingDeviceVulkan : public RenderingDevice { }; uint32_t vertex_input_mask = 0; //inputs used, this is mostly for validation - int fragment_outputs = 0; + uint32_t fragment_output_mask = 0; struct PushConstant { uint32_t push_constant_size = 0; @@ -668,6 +755,7 @@ class RenderingDeviceVulkan : public RenderingDevice { #ifdef DEBUG_ENABLED struct Validation { FramebufferFormatID framebuffer_format = 0; + uint32_t render_pass = 0; uint32_t dynamic_state = 0; VertexFormatID vertex_format = 0; bool uses_restart_indices = false; @@ -723,6 +811,7 @@ class RenderingDeviceVulkan : public RenderingDevice { struct DrawList { VkCommandBuffer command_buffer = VK_NULL_HANDLE; // If persistent, this is owned, otherwise it's shared with the ringbuffer. Rect2i viewport; + bool viewport_set = false; struct SetState { uint32_t pipeline_expected_format = 0; @@ -746,7 +835,6 @@ class RenderingDeviceVulkan : public RenderingDevice { #ifdef DEBUG_ENABLED struct Validation { bool active = true; // Means command buffer was not closed, so you can keep adding things. - FramebufferFormatID framebuffer_format = INVALID_ID; // Actual render pass values. uint32_t dynamic_state = 0; VertexFormatID vertex_format = INVALID_ID; @@ -782,7 +870,15 @@ class RenderingDeviceVulkan : public RenderingDevice { }; DrawList *draw_list = nullptr; // One for regular draw lists, multiple for split. + uint32_t draw_list_subpass_count = 0; uint32_t draw_list_count = 0; + VkRenderPass draw_list_render_pass; + VkFramebuffer draw_list_vkframebuffer; +#ifdef DEBUG_ENABLED + FramebufferFormatID draw_list_framebuffer_format = INVALID_ID; +#endif + uint32_t draw_list_current_subpass = 0; + bool draw_list_split = false; Vector<RID> draw_list_bound_textures; Vector<RID> draw_list_storage_textures; @@ -790,10 +886,12 @@ class RenderingDeviceVulkan : public RenderingDevice { bool draw_list_unbind_depth_textures = false; void _draw_list_insert_clear_region(DrawList *draw_list, Framebuffer *framebuffer, Point2i viewport_offset, Point2i viewport_size, bool p_clear_color, const Vector<Color> &p_clear_colors, bool p_clear_depth, float p_depth, uint32_t p_stencil); - Error _draw_list_setup_framebuffer(Framebuffer *p_framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, VkFramebuffer *r_framebuffer, VkRenderPass *r_render_pass); + Error _draw_list_setup_framebuffer(Framebuffer *p_framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, VkFramebuffer *r_framebuffer, VkRenderPass *r_render_pass, uint32_t *r_subpass_count); Error _draw_list_render_pass_begin(Framebuffer *framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, const Vector<Color> &p_clear_colors, float p_clear_depth, uint32_t p_clear_stencil, Point2i viewport_offset, Point2i viewport_size, VkFramebuffer vkframebuffer, VkRenderPass render_pass, VkCommandBuffer command_buffer, VkSubpassContents subpass_contents, const Vector<RID> &p_storage_textures); _FORCE_INLINE_ DrawList *_get_draw_list_ptr(DrawListID p_id); Buffer *_get_buffer_from_owner(RID p_buffer, VkPipelineStageFlags &dst_stage_mask, VkAccessFlags &dst_access, uint32_t p_post_barrier); + Error _draw_list_allocate(const Rect2i &p_viewport, uint32_t p_splits, uint32_t p_subpass); + void _draw_list_free(Rect2i *r_last_viewport = nullptr); /**********************/ /**** COMPUTE LIST ****/ @@ -938,11 +1036,13 @@ public: /**** FRAMEBUFFER ****/ /*********************/ - virtual FramebufferFormatID framebuffer_format_create(const Vector<AttachmentFormat> &p_format); + virtual FramebufferFormatID framebuffer_format_create(const Vector<AttachmentFormat> &p_format, uint32_t p_view_count = 1); + virtual FramebufferFormatID framebuffer_format_create_multipass(const Vector<AttachmentFormat> &p_attachments, Vector<FramebufferPass> &p_passes, uint32_t p_view_count = 1); virtual FramebufferFormatID framebuffer_format_create_empty(TextureSamples p_samples = TEXTURE_SAMPLES_1); - virtual TextureSamples framebuffer_format_get_texture_samples(FramebufferFormatID p_format); + virtual TextureSamples framebuffer_format_get_texture_samples(FramebufferFormatID p_format, uint32_t p_pass = 0); - virtual RID framebuffer_create(const Vector<RID> &p_texture_attachments, FramebufferFormatID p_format_check = INVALID_ID); + virtual RID framebuffer_create(const Vector<RID> &p_texture_attachments, FramebufferFormatID p_format_check = INVALID_ID, uint32_t p_view_count = 1); + virtual RID framebuffer_create_multipass(const Vector<RID> &p_texture_attachments, Vector<FramebufferPass> &p_passes, FramebufferFormatID p_format_check = INVALID_ID, uint32_t p_view_count = 1); virtual RID framebuffer_create_empty(const Size2i &p_size, TextureSamples p_samples = TEXTURE_SAMPLES_1, FramebufferFormatID p_format_check = INVALID_ID); virtual FramebufferFormatID framebuffer_get_format(RID p_framebuffer); @@ -993,7 +1093,7 @@ public: /**** RENDER PIPELINE ****/ /*************************/ - virtual RID render_pipeline_create(RID p_shader, FramebufferFormatID p_framebuffer_format, VertexFormatID p_vertex_format, RenderPrimitive p_render_primitive, const PipelineRasterizationState &p_rasterization_state, const PipelineMultisampleState &p_multisample_state, const PipelineDepthStencilState &p_depth_stencil_state, const PipelineColorBlendState &p_blend_state, int p_dynamic_state_flags = 0); + virtual RID render_pipeline_create(RID p_shader, FramebufferFormatID p_framebuffer_format, VertexFormatID p_vertex_format, RenderPrimitive p_render_primitive, const PipelineRasterizationState &p_rasterization_state, const PipelineMultisampleState &p_multisample_state, const PipelineDepthStencilState &p_depth_stencil_state, const PipelineColorBlendState &p_blend_state, int p_dynamic_state_flags = 0, uint32_t p_for_render_pass = 0); virtual bool render_pipeline_is_valid(RID p_pipeline); /**************************/ @@ -1032,6 +1132,9 @@ public: virtual void draw_list_enable_scissor(DrawListID p_list, const Rect2 &p_rect); virtual void draw_list_disable_scissor(DrawListID p_list); + virtual DrawListID draw_list_switch_to_next_pass(); + virtual Error draw_list_switch_to_next_pass_split(uint32_t p_splits, DrawListID *r_split_ids); + virtual void draw_list_end(uint32_t p_post_barrier = BARRIER_MASK_ALL); /***********************/ diff --git a/drivers/vulkan/vulkan_context.cpp b/drivers/vulkan/vulkan_context.cpp index 38455bdbed..9939611a83 100644 --- a/drivers/vulkan/vulkan_context.cpp +++ b/drivers/vulkan/vulkan_context.cpp @@ -337,6 +337,9 @@ Error VulkanContext::_initialize_extensions() { extension_names[enabled_extension_count++] = VK_EXT_DEBUG_UTILS_EXTENSION_NAME; enabled_debug_utils = true; } + if (!strcmp(VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME, instance_extensions[i].extensionName)) { + extension_names[enabled_extension_count++] = VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME; + } if (enabled_extension_count >= MAX_EXTENSIONS) { free(instance_extensions); ERR_FAIL_V_MSG(ERR_BUG, "Enabled extension count reaches MAX_EXTENSIONS, BUG"); @@ -504,6 +507,8 @@ Error VulkanContext::_check_capabilities() { // assume not supported until proven otherwise multiview_capabilities.is_supported = false; + multiview_capabilities.geometry_shader_is_supported = false; + multiview_capabilities.tessellation_shader_is_supported = false; multiview_capabilities.max_view_count = 0; multiview_capabilities.max_instance_count = 0; subgroup_capabilities.size = 0; @@ -529,7 +534,8 @@ Error VulkanContext::_check_capabilities() { device_features_func(gpu, &device_features); multiview_capabilities.is_supported = multiview_features.multiview; - // For now we ignore if multiview is available in geometry and tessellation as we do not currently support those + multiview_capabilities.geometry_shader_is_supported = multiview_features.multiviewGeometryShader; + multiview_capabilities.tessellation_shader_is_supported = multiview_features.multiviewTessellationShader; } // check extended properties @@ -573,7 +579,7 @@ Error VulkanContext::_check_capabilities() { #ifdef DEBUG_ENABLED print_line("- Vulkan multiview supported:"); - print_line(" max views: " + itos(multiview_capabilities.max_view_count)); + print_line(" max view count: " + itos(multiview_capabilities.max_view_count)); print_line(" max instances: " + itos(multiview_capabilities.max_instance_count)); } else { print_line("- Vulkan multiview not supported"); @@ -694,8 +700,25 @@ Error VulkanContext::_create_physical_device() { free(physical_devices); ERR_FAIL_V(ERR_CANT_CREATE); } - /* for now, just grab the first physical device */ + + // TODO: At least on Linux Laptops integrated GPUs fail with Vulkan in many instances. + // The device should really be a preference, but for now choosing a discrete GPU over the + // integrated one is better than the default. + + // Default to first device uint32_t device_index = 0; + + for (uint32_t i = 0; i < gpu_count; ++i) { + VkPhysicalDeviceProperties props; + vkGetPhysicalDeviceProperties(physical_devices[i], &props); + + if (props.deviceType == VkPhysicalDeviceType::VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) { + // Prefer discrete GPU. + device_index = i; + break; + } + } + gpu = physical_devices[device_index]; free(physical_devices); @@ -755,6 +778,10 @@ Error VulkanContext::_create_physical_device() { swapchainExtFound = 1; extension_names[enabled_extension_count++] = VK_KHR_SWAPCHAIN_EXTENSION_NAME; } + if (!strcmp(VK_KHR_MULTIVIEW_EXTENSION_NAME, device_extensions[i].extensionName)) { + // if multiview is supported, enable it + extension_names[enabled_extension_count++] = VK_KHR_MULTIVIEW_EXTENSION_NAME; + } if (enabled_extension_count >= MAX_EXTENSIONS) { free(device_extensions); ERR_FAIL_V_MSG(ERR_BUG, "Enabled extension count reaches MAX_EXTENSIONS, BUG"); @@ -947,17 +974,50 @@ Error VulkanContext::_create_device() { queues[1].flags = 0; sdevice.queueCreateInfoCount = 2; } + +#ifdef VK_VERSION_1_2 + VkPhysicalDeviceVulkan11Features vulkan11features; + + vulkan11features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES; + vulkan11features.pNext = nullptr; + // !BAS! Need to figure out which ones of these we want enabled... + vulkan11features.storageBuffer16BitAccess = 0; + vulkan11features.uniformAndStorageBuffer16BitAccess = 0; + vulkan11features.storagePushConstant16 = 0; + vulkan11features.storageInputOutput16 = 0; + vulkan11features.multiview = multiview_capabilities.is_supported; + vulkan11features.multiviewGeometryShader = multiview_capabilities.geometry_shader_is_supported; + vulkan11features.multiviewTessellationShader = multiview_capabilities.tessellation_shader_is_supported; + vulkan11features.variablePointersStorageBuffer = 0; + vulkan11features.variablePointers = 0; + vulkan11features.protectedMemory = 0; + vulkan11features.samplerYcbcrConversion = 0; + vulkan11features.shaderDrawParameters = 0; + + sdevice.pNext = &vulkan11features; +#elif VK_VERSION_1_1 + VkPhysicalDeviceMultiviewFeatures multiview_features; + + multiview_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES; + multiview_features.pNext = nullptr; + multiview_features.multiview = multiview_capabilities.is_supported; + multiview_features.multiviewGeometryShader = multiview_capabilities.geometry_shader_is_supported; + multiview_features.multiviewTessellationShader = multiview_capabilities.tessellation_shader_is_supported; + + sdevice.pNext = &multiview_features; +#endif + err = vkCreateDevice(gpu, &sdevice, nullptr, &device); ERR_FAIL_COND_V(err, ERR_CANT_CREATE); return OK; } -Error VulkanContext::_initialize_queues(VkSurfaceKHR surface) { +Error VulkanContext::_initialize_queues(VkSurfaceKHR p_surface) { // Iterate over each queue to learn whether it supports presenting: VkBool32 *supportsPresent = (VkBool32 *)malloc(queue_family_count * sizeof(VkBool32)); for (uint32_t i = 0; i < queue_family_count; i++) { - fpGetPhysicalDeviceSurfaceSupportKHR(gpu, i, surface, &supportsPresent[i]); + fpGetPhysicalDeviceSurfaceSupportKHR(gpu, i, p_surface, &supportsPresent[i]); } // Search for a graphics and a present queue in the array of queue @@ -1031,10 +1091,10 @@ Error VulkanContext::_initialize_queues(VkSurfaceKHR surface) { // Get the list of VkFormat's that are supported: uint32_t formatCount; - VkResult err = fpGetPhysicalDeviceSurfaceFormatsKHR(gpu, surface, &formatCount, nullptr); + VkResult err = fpGetPhysicalDeviceSurfaceFormatsKHR(gpu, p_surface, &formatCount, nullptr); ERR_FAIL_COND_V(err, ERR_CANT_CREATE); VkSurfaceFormatKHR *surfFormats = (VkSurfaceFormatKHR *)malloc(formatCount * sizeof(VkSurfaceFormatKHR)); - err = fpGetPhysicalDeviceSurfaceFormatsKHR(gpu, surface, &formatCount, surfFormats); + err = fpGetPhysicalDeviceSurfaceFormatsKHR(gpu, p_surface, &formatCount, surfFormats); if (err) { free(surfFormats); ERR_FAIL_V(ERR_CANT_CREATE); @@ -1109,9 +1169,6 @@ Error VulkanContext::_create_semaphores() { err = vkCreateFence(device, &fence_ci, nullptr, &fences[i]); ERR_FAIL_COND_V(err, ERR_CANT_CREATE); - err = vkCreateSemaphore(device, &semaphoreCreateInfo, nullptr, &image_acquired_semaphores[i]); - ERR_FAIL_COND_V(err, ERR_CANT_CREATE); - err = vkCreateSemaphore(device, &semaphoreCreateInfo, nullptr, &draw_complete_semaphores[i]); ERR_FAIL_COND_V(err, ERR_CANT_CREATE); @@ -1141,6 +1198,19 @@ Error VulkanContext::_window_create(DisplayServer::WindowID p_window_id, VkSurfa // is created. Error err = _initialize_queues(p_surface); ERR_FAIL_COND_V(err != OK, ERR_CANT_CREATE); + } else { + // make sure any of the surfaces supports present (validation layer complains if this is not done). + bool any_supports_present = false; + for (uint32_t i = 0; i < queue_family_count; i++) { + VkBool32 supports; + fpGetPhysicalDeviceSurfaceSupportKHR(gpu, i, p_surface, &supports); + if (supports) { + any_supports_present = true; + break; + } + } + + ERR_FAIL_COND_V_MSG(!any_supports_present, ERR_CANT_CREATE, "Surface passed for sub-window creation does not support presenting"); } Window window; @@ -1150,6 +1220,17 @@ Error VulkanContext::_window_create(DisplayServer::WindowID p_window_id, VkSurfa Error err = _update_swap_chain(&window); ERR_FAIL_COND_V(err != OK, ERR_CANT_CREATE); + VkSemaphoreCreateInfo semaphoreCreateInfo = { + /*sType*/ VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO, + /*pNext*/ nullptr, + /*flags*/ 0, + }; + + for (uint32_t i = 0; i < FRAME_LAG; i++) { + VkResult vkerr = vkCreateSemaphore(device, &semaphoreCreateInfo, nullptr, &window.image_acquired_semaphores[i]); + ERR_FAIL_COND_V(vkerr, ERR_CANT_CREATE); + } + windows[p_window_id] = window; return OK; } @@ -1189,6 +1270,10 @@ VkFramebuffer VulkanContext::window_get_framebuffer(DisplayServer::WindowID p_wi void VulkanContext::window_destroy(DisplayServer::WindowID p_window_id) { ERR_FAIL_COND(!windows.has(p_window_id)); _clean_up_swap_chain(&windows[p_window_id]); + for (uint32_t i = 0; i < FRAME_LAG; i++) { + vkDestroySemaphore(device, windows[p_window_id].image_acquired_semaphores[i], nullptr); + } + vkDestroySurfaceKHR(inst, windows[p_window_id].surface, nullptr); windows.erase(p_window_id); } @@ -1643,6 +1728,8 @@ Error VulkanContext::prepare_buffers() { for (Map<int, Window>::Element *E = windows.front(); E; E = E->next()) { Window *w = &E->get(); + w->semaphore_acquired = false; + if (w->swapchain == VK_NULL_HANDLE) { continue; } @@ -1651,7 +1738,7 @@ Error VulkanContext::prepare_buffers() { // Get the index of the next available swapchain image: err = fpAcquireNextImageKHR(device, w->swapchain, UINT64_MAX, - image_acquired_semaphores[frame_index], VK_NULL_HANDLE, &w->current_buffer); + w->image_acquired_semaphores[frame_index], VK_NULL_HANDLE, &w->current_buffer); if (err == VK_ERROR_OUT_OF_DATE_KHR) { // swapchain is out of date (e.g. the window was resized) and @@ -1664,8 +1751,10 @@ Error VulkanContext::prepare_buffers() { // presentation engine will still present the image correctly. print_verbose("Vulkan: Early suboptimal swapchain."); break; + } else if (err != VK_SUCCESS) { + ERR_BREAK(ERR_CANT_CREATE); } else { - ERR_FAIL_COND_V(err, ERR_CANT_CREATE); + w->semaphore_acquired = true; } } while (err != VK_SUCCESS); } @@ -1715,14 +1804,25 @@ Error VulkanContext::swap_buffers() { commands_to_submit = command_buffer_count; } + VkSemaphore *semaphores_to_acquire = (VkSemaphore *)alloca(windows.size() * sizeof(VkSemaphore)); + uint32_t semaphores_to_acquire_count = 0; + + for (Map<int, Window>::Element *E = windows.front(); E; E = E->next()) { + Window *w = &E->get(); + + if (w->semaphore_acquired) { + semaphores_to_acquire[semaphores_to_acquire_count++] = w->image_acquired_semaphores[frame_index]; + } + } + VkPipelineStageFlags pipe_stage_flags; VkSubmitInfo submit_info; submit_info.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO; submit_info.pNext = nullptr; submit_info.pWaitDstStageMask = &pipe_stage_flags; pipe_stage_flags = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT; - submit_info.waitSemaphoreCount = 1; - submit_info.pWaitSemaphores = &image_acquired_semaphores[frame_index]; + submit_info.waitSemaphoreCount = semaphores_to_acquire_count; + submit_info.pWaitSemaphores = semaphores_to_acquire; submit_info.commandBufferCount = commands_to_submit; submit_info.pCommandBuffers = commands_ptr; submit_info.signalSemaphoreCount = 1; @@ -2074,7 +2174,6 @@ VulkanContext::~VulkanContext() { if (device_initialized) { for (uint32_t i = 0; i < FRAME_LAG; i++) { vkDestroyFence(device, fences[i], nullptr); - vkDestroySemaphore(device, image_acquired_semaphores[i], nullptr); vkDestroySemaphore(device, draw_complete_semaphores[i], nullptr); if (separate_present_queue) { vkDestroySemaphore(device, image_ownership_semaphores[i], nullptr); diff --git a/drivers/vulkan/vulkan_context.h b/drivers/vulkan/vulkan_context.h index 3d9b295c5a..8f1005d07f 100644 --- a/drivers/vulkan/vulkan_context.h +++ b/drivers/vulkan/vulkan_context.h @@ -56,8 +56,10 @@ public: struct MultiviewCapabilities { bool is_supported; - int32_t max_view_count; - int32_t max_instance_count; + bool geometry_shader_is_supported; + bool tessellation_shader_is_supported; + uint32_t max_view_count; + uint32_t max_instance_count; }; private: @@ -68,7 +70,6 @@ private: }; VkInstance inst = VK_NULL_HANDLE; - VkSurfaceKHR surface = VK_NULL_HANDLE; VkPhysicalDevice gpu = VK_NULL_HANDLE; VkPhysicalDeviceProperties gpu_props; uint32_t queue_family_count = 0; @@ -99,7 +100,6 @@ private: VkQueue present_queue = VK_NULL_HANDLE; VkColorSpaceKHR color_space; VkFormat format; - VkSemaphore image_acquired_semaphores[FRAME_LAG]; VkSemaphore draw_complete_semaphores[FRAME_LAG]; VkSemaphore image_ownership_semaphores[FRAME_LAG]; int frame_index = 0; @@ -119,6 +119,8 @@ private: VkSwapchainKHR swapchain = VK_NULL_HANDLE; SwapchainImageResources *swapchain_image_resources = VK_NULL_HANDLE; VkPresentModeKHR presentMode = VK_PRESENT_MODE_FIFO_KHR; + VkSemaphore image_acquired_semaphores[FRAME_LAG]; + bool semaphore_acquired = false; uint32_t current_buffer = 0; int width = 0; int height = 0; @@ -206,7 +208,7 @@ private: Error _create_physical_device(); - Error _initialize_queues(VkSurfaceKHR surface); + Error _initialize_queues(VkSurfaceKHR p_surface); Error _create_device(); diff --git a/drivers/windows/dir_access_windows.cpp b/drivers/windows/dir_access_windows.cpp index 55f9e6de17..325bae5b56 100644 --- a/drivers/windows/dir_access_windows.cpp +++ b/drivers/windows/dir_access_windows.cpp @@ -208,7 +208,7 @@ String DirAccessWindows::get_current_dir(bool p_include_drive) { bool DirAccessWindows::file_exists(String p_file) { GLOBAL_LOCK_FUNCTION - if (!p_file.is_abs_path()) { + if (!p_file.is_absolute_path()) { p_file = get_current_dir().plus_file(p_file); } diff --git a/drivers/windows/dir_access_windows.h b/drivers/windows/dir_access_windows.h index b151b631e9..1ba4e70e42 100644 --- a/drivers/windows/dir_access_windows.h +++ b/drivers/windows/dir_access_windows.h @@ -33,7 +33,7 @@ #ifdef WINDOWS_ENABLED -#include "core/os/dir_access.h" +#include "core/io/dir_access.h" /** @author Juan Linietsky <reduz@gmail.com> diff --git a/drivers/windows/file_access_windows.cpp b/drivers/windows/file_access_windows.cpp index 1f46b44f5e..d6deda7b5d 100644 --- a/drivers/windows/file_access_windows.cpp +++ b/drivers/windows/file_access_windows.cpp @@ -294,6 +294,7 @@ void FileAccessWindows::store_8(uint8_t p_dest) { void FileAccessWindows::store_buffer(const uint8_t *p_src, uint64_t p_length) { ERR_FAIL_COND(!f); + ERR_FAIL_COND(!p_src && p_length > 0); if (flags == READ_WRITE || flags == WRITE_READ) { if (prev_op == READ) { if (last_error != ERR_FILE_EOF) { diff --git a/drivers/windows/file_access_windows.h b/drivers/windows/file_access_windows.h index 95f3a75548..7280fc3237 100644 --- a/drivers/windows/file_access_windows.h +++ b/drivers/windows/file_access_windows.h @@ -33,7 +33,7 @@ #ifdef WINDOWS_ENABLED -#include "core/os/file_access.h" +#include "core/io/file_access.h" #include "core/os/memory.h" #include <stdio.h> |