diff options
Diffstat (limited to 'drivers/vulkan')
| -rw-r--r-- | drivers/vulkan/rendering_device_vulkan.cpp | 317 | ||||
| -rw-r--r-- | drivers/vulkan/rendering_device_vulkan.h | 31 | ||||
| -rw-r--r-- | drivers/vulkan/vulkan_context.cpp | 295 | ||||
| -rw-r--r-- | drivers/vulkan/vulkan_context.h | 10 |
4 files changed, 433 insertions, 220 deletions
diff --git a/drivers/vulkan/rendering_device_vulkan.cpp b/drivers/vulkan/rendering_device_vulkan.cpp index bcb6092d87..01d1583ca4 100644 --- a/drivers/vulkan/rendering_device_vulkan.cpp +++ b/drivers/vulkan/rendering_device_vulkan.cpp @@ -47,7 +47,7 @@ static const uint32_t SMALL_ALLOCATION_MAX_SIZE = 4096; // Get the Vulkan object information and possible stage access types (bitwise OR'd with incoming values). -RenderingDeviceVulkan::Buffer *RenderingDeviceVulkan::_get_buffer_from_owner(RID p_buffer, VkPipelineStageFlags &r_stage_mask, VkAccessFlags &r_access_mask, uint32_t p_post_barrier) { +RenderingDeviceVulkan::Buffer *RenderingDeviceVulkan::_get_buffer_from_owner(RID p_buffer, VkPipelineStageFlags &r_stage_mask, VkAccessFlags &r_access_mask, BitField<BarrierMask> p_post_barrier) { Buffer *buffer = nullptr; if (vertex_buffer_owner.owns(p_buffer)) { buffer = vertex_buffer_owner.get_or_null(p_buffer); @@ -55,11 +55,11 @@ RenderingDeviceVulkan::Buffer *RenderingDeviceVulkan::_get_buffer_from_owner(RID r_stage_mask |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT; r_access_mask |= VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT; if (buffer->usage & VK_BUFFER_USAGE_STORAGE_BUFFER_BIT) { - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { r_access_mask |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; r_stage_mask |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; } - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { r_access_mask |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; r_stage_mask |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; } @@ -69,20 +69,20 @@ RenderingDeviceVulkan::Buffer *RenderingDeviceVulkan::_get_buffer_from_owner(RID r_access_mask |= VK_ACCESS_INDEX_READ_BIT; buffer = index_buffer_owner.get_or_null(p_buffer); } else if (uniform_buffer_owner.owns(p_buffer)) { - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { r_stage_mask |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; } - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { r_stage_mask |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; } r_access_mask |= VK_ACCESS_UNIFORM_READ_BIT; buffer = uniform_buffer_owner.get_or_null(p_buffer); } else if (texture_buffer_owner.owns(p_buffer)) { - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { r_stage_mask |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; r_access_mask |= VK_ACCESS_SHADER_READ_BIT; } - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { r_stage_mask |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; r_access_mask |= VK_ACCESS_SHADER_READ_BIT; } @@ -90,11 +90,11 @@ RenderingDeviceVulkan::Buffer *RenderingDeviceVulkan::_get_buffer_from_owner(RID buffer = &texture_buffer_owner.get_or_null(p_buffer)->buffer; } else if (storage_buffer_owner.owns(p_buffer)) { buffer = storage_buffer_owner.get_or_null(p_buffer); - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { r_stage_mask |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; r_access_mask |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { r_stage_mask |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; r_access_mask |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } @@ -1655,9 +1655,7 @@ RID RenderingDeviceVulkan::texture_create(const TextureFormat &p_format, const T image_create_info.pNext = nullptr; image_create_info.flags = 0; -#ifndef _MSC_VER -#warning TODO check for support via RenderingDevice to enable on mobile when possible -#endif + // TODO: Check for support via RenderingDevice to enable on mobile when possible. #ifndef ANDROID_ENABLED @@ -1733,7 +1731,7 @@ RID RenderingDeviceVulkan::texture_create(const TextureFormat &p_format, const T ERR_FAIL_INDEX_V(p_format.samples, TEXTURE_SAMPLES_MAX, RID()); - image_create_info.samples = rasterization_sample_count[p_format.samples]; + image_create_info.samples = _ensure_supported_sample_count(p_format.samples); image_create_info.tiling = (p_format.usage_bits & TEXTURE_USAGE_CPU_READ_BIT) ? VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL; // Usage. @@ -1884,6 +1882,7 @@ RID RenderingDeviceVulkan::texture_create(const TextureFormat &p_format, const T texture.mipmaps = image_create_info.mipLevels; texture.base_mipmap = 0; texture.base_layer = 0; + texture.is_resolve_buffer = p_format.is_resolve_buffer; texture.usage_flags = p_format.usage_bits; texture.samples = p_format.samples; texture.allowed_shared_formats = p_format.shareable_formats; @@ -2010,7 +2009,7 @@ RID RenderingDeviceVulkan::texture_create(const TextureFormat &p_format, const T if (p_data.size()) { for (uint32_t i = 0; i < image_create_info.arrayLayers; i++) { - _texture_update(id, i, p_data[i], RD::BARRIER_MASK_ALL, true); + _texture_update(id, i, p_data[i], RD::BARRIER_MASK_ALL_BARRIERS, true); } } return id; @@ -2161,14 +2160,35 @@ RID RenderingDeviceVulkan::texture_create_from_extension(TextureType p_type, Dat texture.height = p_height; texture.depth = p_depth; texture.layers = p_layers; - texture.mipmaps = 0; // Maybe make this settable too? + texture.mipmaps = 1; texture.usage_flags = p_flags; texture.base_mipmap = 0; texture.base_layer = 0; texture.allowed_shared_formats.push_back(RD::DATA_FORMAT_R8G8B8A8_UNORM); texture.allowed_shared_formats.push_back(RD::DATA_FORMAT_R8G8B8A8_SRGB); - // Do we need to do something with texture.layout? + // Set base layout based on usage priority. + + if (texture.usage_flags & TEXTURE_USAGE_SAMPLING_BIT) { + // First priority, readable. + texture.layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + + } else if (texture.usage_flags & TEXTURE_USAGE_STORAGE_BIT) { + // Second priority, storage. + + texture.layout = VK_IMAGE_LAYOUT_GENERAL; + + } else if (texture.usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { + // Third priority, color or depth. + + texture.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; + + } else if (texture.usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { + texture.layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; + + } else { + texture.layout = VK_IMAGE_LAYOUT_GENERAL; + } if (texture.usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) { texture.read_aspect_mask = VK_IMAGE_ASPECT_DEPTH_BIT; @@ -2394,7 +2414,7 @@ RID RenderingDeviceVulkan::texture_create_shared_from_slice(const TextureView &p return id; } -Error RenderingDeviceVulkan::texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, uint32_t p_post_barrier) { +Error RenderingDeviceVulkan::texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, BitField<BarrierMask> p_post_barrier) { return _texture_update(p_texture, p_layer, p_data, p_post_barrier, false); } @@ -2414,7 +2434,7 @@ static _ALWAYS_INLINE_ void _copy_region(uint8_t const *__restrict p_src, uint8_ } } -Error RenderingDeviceVulkan::_texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, uint32_t p_post_barrier, bool p_use_setup_queue) { +Error RenderingDeviceVulkan::_texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, BitField<BarrierMask> p_post_barrier, bool p_use_setup_queue) { _THREAD_SAFE_METHOD_ ERR_FAIL_COND_V_MSG((draw_list || compute_list) && !p_use_setup_queue, ERR_INVALID_PARAMETER, @@ -2588,15 +2608,15 @@ Error RenderingDeviceVulkan::_texture_update(RID p_texture, uint32_t p_layer, co { uint32_t barrier_flags = 0; uint32_t access_flags = 0; - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { barrier_flags |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } @@ -2849,7 +2869,7 @@ Size2i RenderingDeviceVulkan::texture_size(RID p_texture) { return Size2i(tex->width, tex->height); } -Error RenderingDeviceVulkan::texture_copy(RID p_from_texture, RID p_to_texture, const Vector3 &p_from, const Vector3 &p_to, const Vector3 &p_size, uint32_t p_src_mipmap, uint32_t p_dst_mipmap, uint32_t p_src_layer, uint32_t p_dst_layer, uint32_t p_post_barrier) { +Error RenderingDeviceVulkan::texture_copy(RID p_from_texture, RID p_to_texture, const Vector3 &p_from, const Vector3 &p_to, const Vector3 &p_size, uint32_t p_src_mipmap, uint32_t p_dst_mipmap, uint32_t p_src_layer, uint32_t p_dst_layer, BitField<BarrierMask> p_post_barrier) { _THREAD_SAFE_METHOD_ Texture *src_tex = texture_owner.get_or_null(p_from_texture); @@ -2974,15 +2994,15 @@ Error RenderingDeviceVulkan::texture_copy(RID p_from_texture, RID p_to_texture, uint32_t barrier_flags = 0; uint32_t access_flags = 0; - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { barrier_flags |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } @@ -3044,7 +3064,7 @@ Error RenderingDeviceVulkan::texture_copy(RID p_from_texture, RID p_to_texture, return OK; } -Error RenderingDeviceVulkan::texture_resolve_multisample(RID p_from_texture, RID p_to_texture, uint32_t p_post_barrier) { +Error RenderingDeviceVulkan::texture_resolve_multisample(RID p_from_texture, RID p_to_texture, BitField<BarrierMask> p_post_barrier) { _THREAD_SAFE_METHOD_ Texture *src_tex = texture_owner.get_or_null(p_from_texture); @@ -3152,15 +3172,15 @@ Error RenderingDeviceVulkan::texture_resolve_multisample(RID p_from_texture, RID uint32_t barrier_flags = 0; uint32_t access_flags = 0; - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { barrier_flags |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } @@ -3215,7 +3235,7 @@ Error RenderingDeviceVulkan::texture_resolve_multisample(RID p_from_texture, RID return OK; } -Error RenderingDeviceVulkan::texture_clear(RID p_texture, const Color &p_color, uint32_t p_base_mipmap, uint32_t p_mipmaps, uint32_t p_base_layer, uint32_t p_layers, uint32_t p_post_barrier) { +Error RenderingDeviceVulkan::texture_clear(RID p_texture, const Color &p_color, uint32_t p_base_mipmap, uint32_t p_mipmaps, uint32_t p_base_layer, uint32_t p_layers, BitField<BarrierMask> p_post_barrier) { _THREAD_SAFE_METHOD_ Texture *src_tex = texture_owner.get_or_null(p_texture); @@ -3288,15 +3308,15 @@ Error RenderingDeviceVulkan::texture_clear(RID p_texture, const Color &p_color, uint32_t barrier_flags = 0; uint32_t access_flags = 0; - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { barrier_flags |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } @@ -3402,6 +3422,16 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF LocalVector<int32_t> attachment_last_pass; attachment_last_pass.resize(p_attachments.size()); + if (p_view_count > 1) { + const VulkanContext::MultiviewCapabilities capabilities = context->get_multiview_capabilities(); + + // 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"); + } + // These are only used if we use multiview but we need to define them in scope. const uint32_t view_mask = (1 << p_view_count) - 1; const uint32_t correlation_mask = (1 << p_view_count) - 1; @@ -3425,7 +3455,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF description.pNext = nullptr; description.flags = 0; description.format = vulkan_formats[p_attachments[i].format]; - description.samples = rasterization_sample_count[p_attachments[i].samples]; + description.samples = _ensure_supported_sample_count(p_attachments[i].samples); bool is_sampled = p_attachments[i].usage_flags & TEXTURE_USAGE_SAMPLING_BIT; bool is_storage = p_attachments[i].usage_flags & TEXTURE_USAGE_STORAGE_BIT; @@ -3546,7 +3576,8 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF break; } } - } else { + } + if (!used_last) { for (int j = 0; j < p_passes[last_pass].color_attachments.size(); j++) { if (p_passes[last_pass].color_attachments[j] == i) { used_last = true; @@ -3701,7 +3732,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; attachment_last_pass[attachment] = i; } - reference.aspectMask = 0; + reference.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; color_references.push_back(reference); } @@ -3723,7 +3754,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF reference.layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; attachment_last_pass[attachment] = i; } - reference.aspectMask = 0; // TODO: We need to set this here, possibly VK_IMAGE_ASPECT_COLOR_BIT? + reference.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; input_references.push_back(reference); } @@ -3752,7 +3783,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL; // VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL attachment_last_pass[attachment] = i; } - reference.aspectMask = 0; + reference.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; resolve_references.push_back(reference); } @@ -3767,7 +3798,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF 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_remap[attachment]; depth_stencil_reference.layout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; - depth_stencil_reference.aspectMask = 0; + depth_stencil_reference.aspectMask = VK_IMAGE_ASPECT_NONE; attachment_last_pass[attachment] = i; if (is_multisample_first) { @@ -3793,9 +3824,9 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF vrs_reference.pNext = nullptr; vrs_reference.attachment = attachment_remap[attachment]; vrs_reference.layout = VK_IMAGE_LAYOUT_FRAGMENT_SHADING_RATE_ATTACHMENT_OPTIMAL_KHR; - vrs_reference.aspectMask = 0; + vrs_reference.aspectMask = VK_IMAGE_ASPECT_NONE; - Size2i texel_size = context->get_vrs_capabilities().max_texel_size; + Size2i texel_size = context->get_vrs_capabilities().texel_size; VkFragmentShadingRateAttachmentInfoKHR &vrs_attachment_info = vrs_attachment_info_array[i]; vrs_attachment_info.sType = VK_STRUCTURE_TYPE_FRAGMENT_SHADING_RATE_ATTACHMENT_INFO_KHR; @@ -3934,16 +3965,9 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF Vector<uint32_t> view_masks; VkRenderPassMultiviewCreateInfo render_pass_multiview_create_info; - if (p_view_count > 1) { - // This may no longer be needed with the new settings already including this. - - 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"); + if ((p_view_count > 1) && !context->supports_renderpass2()) { + // This is only required when using vkCreateRenderPass, we add it if vkCreateRenderPass2KHR is not supported + // resulting this in being passed to our vkCreateRenderPass fallback. // Set view masks for each subpass. for (uint32_t i = 0; i < subpasses.size(); i++) { @@ -4116,7 +4140,11 @@ RID RenderingDeviceVulkan::framebuffer_create(const Vector<RID> &p_texture_attac } else if (texture && texture->usage_flags & TEXTURE_USAGE_VRS_ATTACHMENT_BIT) { pass.vrs_attachment = i; } else { - pass.color_attachments.push_back(texture ? i : FramebufferPass::ATTACHMENT_UNUSED); + if (texture && texture->is_resolve_buffer) { + pass.resolve_attachments.push_back(i); + } else { + pass.color_attachments.push_back(texture ? i : FramebufferPass::ATTACHMENT_UNUSED); + } } } @@ -4146,8 +4174,8 @@ RID RenderingDeviceVulkan::framebuffer_create_multipass(const Vector<RID> &p_tex size.height = texture->height; size_set = true; } else if (texture->usage_flags & TEXTURE_USAGE_VRS_ATTACHMENT_BIT) { - // If this is not the first attachement we assume this is used as the VRS attachment. - // In this case this texture will be 1/16th the size of the color attachement. + // If this is not the first attachment we assume this is used as the VRS attachment. + // In this case this texture will be 1/16th the size of the color attachment. // So we skip the size check. } else { ERR_FAIL_COND_V_MSG((uint32_t)size.width != texture->width || (uint32_t)size.height != texture->height, RID(), @@ -4350,7 +4378,7 @@ RenderingDevice::VertexFormatID RenderingDeviceVulkan::vertex_format_create(cons return id; } -RID RenderingDeviceVulkan::vertex_array_create(uint32_t p_vertex_count, VertexFormatID p_vertex_format, const Vector<RID> &p_src_buffers) { +RID RenderingDeviceVulkan::vertex_array_create(uint32_t p_vertex_count, VertexFormatID p_vertex_format, const Vector<RID> &p_src_buffers, const Vector<uint64_t> &p_offsets) { _THREAD_SAFE_METHOD_ ERR_FAIL_COND_V(!vertex_formats.has(p_vertex_format), RID()); @@ -4364,6 +4392,13 @@ RID RenderingDeviceVulkan::vertex_array_create(uint32_t p_vertex_count, VertexFo VertexArray vertex_array; + if (p_offsets.is_empty()) { + vertex_array.offsets.resize_zeroed(p_src_buffers.size()); + } else { + ERR_FAIL_COND_V(p_offsets.size() != p_src_buffers.size(), RID()); + vertex_array.offsets = p_offsets; + } + vertex_array.vertex_count = p_vertex_count; vertex_array.description = p_vertex_format; vertex_array.max_instances_allowed = 0xFFFFFFFF; // By default as many as you want. @@ -4395,7 +4430,6 @@ RID RenderingDeviceVulkan::vertex_array_create(uint32_t p_vertex_count, VertexFo } vertex_array.buffers.push_back(buffer->buffer); - vertex_array.offsets.push_back(0); // Offset unused, but passing anyway. } RID id = vertex_array_owner.make_rid(vertex_array); @@ -4822,7 +4856,7 @@ Vector<uint8_t> RenderingDeviceVulkan::shader_compile_binary_from_spirv(const Ve for (uint32_t j = 0; j < binding_count; j++) { const SpvReflectDescriptorBinding &binding = *bindings[j]; - RenderingDeviceVulkanShaderBinaryDataBinding info; + RenderingDeviceVulkanShaderBinaryDataBinding info{}; bool need_array_dimensions = false; bool need_block_size = false; @@ -4973,7 +5007,7 @@ Vector<uint8_t> RenderingDeviceVulkan::shader_compile_binary_from_spirv(const Ve for (uint32_t j = 0; j < sc_count; j++) { int32_t existing = -1; - RenderingDeviceVulkanShaderBinarySpecializationConstant sconst; + RenderingDeviceVulkanShaderBinarySpecializationConstant sconst{}; SpvReflectSpecializationConstant *spc = spec_constants[j]; sconst.constant_id = spc->constant_id; @@ -5166,9 +5200,9 @@ Vector<uint8_t> RenderingDeviceVulkan::shader_compile_binary_from_spirv(const Ve uint32_t offset = 0; uint8_t *binptr = ret.ptrw(); binptr[0] = 'G'; - binptr[1] = 'V'; + binptr[1] = 'S'; binptr[2] = 'B'; - binptr[3] = 'D'; // Godot vulkan binary data. + binptr[3] = 'D'; // Godot Shader Binary Data. offset += 4; encode_uint32(SHADER_BINARY_VERSION, binptr + offset); offset += sizeof(uint32_t); @@ -5229,7 +5263,7 @@ RID RenderingDeviceVulkan::shader_create_from_bytecode(const Vector<uint8_t> &p_ uint32_t read_offset = 0; // Consistency check. ERR_FAIL_COND_V(binsize < sizeof(uint32_t) * 3 + sizeof(RenderingDeviceVulkanShaderBinaryData), RID()); - ERR_FAIL_COND_V(binptr[0] != 'G' || binptr[1] != 'V' || binptr[2] != 'B' || binptr[3] != 'D', RID()); + ERR_FAIL_COND_V(binptr[0] != 'G' || binptr[1] != 'S' || binptr[2] != 'B' || binptr[3] != 'D', RID()); uint32_t bin_version = decode_uint32(binptr + 4); ERR_FAIL_COND_V(bin_version != SHADER_BINARY_VERSION, RID()); @@ -6287,7 +6321,7 @@ void RenderingDeviceVulkan::uniform_set_set_invalidation_callback(RID p_uniform_ us->invalidated_callback_userdata = p_userdata; } -Error RenderingDeviceVulkan::buffer_update(RID p_buffer, uint32_t p_offset, uint32_t p_size, const void *p_data, uint32_t p_post_barrier) { +Error RenderingDeviceVulkan::buffer_update(RID p_buffer, uint32_t p_offset, uint32_t p_size, const void *p_data, BitField<BarrierMask> p_post_barrier) { _THREAD_SAFE_METHOD_ ERR_FAIL_COND_V_MSG(draw_list, ERR_INVALID_PARAMETER, @@ -6297,7 +6331,7 @@ Error RenderingDeviceVulkan::buffer_update(RID p_buffer, uint32_t p_offset, uint VkPipelineStageFlags dst_stage_mask = 0; VkAccessFlags dst_access = 0; - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { // Protect subsequent updates. dst_stage_mask = VK_PIPELINE_STAGE_TRANSFER_BIT; dst_access = VK_ACCESS_TRANSFER_WRITE_BIT; @@ -6333,7 +6367,7 @@ Error RenderingDeviceVulkan::buffer_update(RID p_buffer, uint32_t p_offset, uint return err; } -Error RenderingDeviceVulkan::buffer_clear(RID p_buffer, uint32_t p_offset, uint32_t p_size, uint32_t p_post_barrier) { +Error RenderingDeviceVulkan::buffer_clear(RID p_buffer, uint32_t p_offset, uint32_t p_size, BitField<BarrierMask> p_post_barrier) { _THREAD_SAFE_METHOD_ ERR_FAIL_COND_V_MSG((p_size % 4) != 0, ERR_INVALID_PARAMETER, @@ -6345,7 +6379,7 @@ Error RenderingDeviceVulkan::buffer_clear(RID p_buffer, uint32_t p_offset, uint3 VkPipelineStageFlags dst_stage_mask = 0; VkAccessFlags dst_access = 0; - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { // Protect subsequent updates. dst_stage_mask = VK_PIPELINE_STAGE_TRANSFER_BIT; dst_access = VK_ACCESS_TRANSFER_WRITE_BIT; @@ -6384,7 +6418,7 @@ Vector<uint8_t> RenderingDeviceVulkan::buffer_get_data(RID p_buffer) { VkPipelineShaderStageCreateFlags src_stage_mask = VK_PIPELINE_STAGE_TRANSFER_BIT; VkAccessFlags src_access_mask = VK_ACCESS_TRANSFER_WRITE_BIT; // Get the vulkan buffer and the potential stage/access possible. - Buffer *buffer = _get_buffer_from_owner(p_buffer, src_stage_mask, src_access_mask, BARRIER_MASK_ALL); + Buffer *buffer = _get_buffer_from_owner(p_buffer, src_stage_mask, src_access_mask, BARRIER_MASK_ALL_BARRIERS); if (!buffer) { ERR_FAIL_V_MSG(Vector<uint8_t>(), "Buffer is either invalid or this type of buffer can't be retrieved. Only Index and Vertex buffers allow retrieving."); } @@ -6555,7 +6589,7 @@ RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferForma ERR_FAIL_INDEX_V(p_rasterization_state.cull_mode, 3, RID()); rasterization_state_create_info.cullMode = cull_mode[p_rasterization_state.cull_mode]; rasterization_state_create_info.frontFace = (p_rasterization_state.front_face == POLYGON_FRONT_FACE_CLOCKWISE ? VK_FRONT_FACE_CLOCKWISE : VK_FRONT_FACE_COUNTER_CLOCKWISE); - rasterization_state_create_info.depthBiasEnable = p_rasterization_state.depth_bias_enable; + rasterization_state_create_info.depthBiasEnable = p_rasterization_state.depth_bias_enabled; rasterization_state_create_info.depthBiasConstantFactor = p_rasterization_state.depth_bias_constant_factor; rasterization_state_create_info.depthBiasClamp = p_rasterization_state.depth_bias_clamp; rasterization_state_create_info.depthBiasSlopeFactor = p_rasterization_state.depth_bias_slope_factor; @@ -6567,7 +6601,7 @@ RID RenderingDeviceVulkan::render_pipeline_create(RID p_shader, FramebufferForma multisample_state_create_info.pNext = nullptr; multisample_state_create_info.flags = 0; - multisample_state_create_info.rasterizationSamples = rasterization_sample_count[p_multisample_state.sample_count]; + multisample_state_create_info.rasterizationSamples = _ensure_supported_sample_count(p_multisample_state.sample_count); multisample_state_create_info.sampleShadingEnable = p_multisample_state.enable_sample_shading; multisample_state_create_info.minSampleShading = p_multisample_state.min_sample_shading; Vector<VkSampleMask> sample_mask; @@ -7256,12 +7290,12 @@ Error RenderingDeviceVulkan::_draw_list_render_pass_begin(Framebuffer *framebuff return OK; } -void RenderingDeviceVulkan::_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) { +void RenderingDeviceVulkan::_draw_list_insert_clear_region(DrawList *p_draw_list, Framebuffer *p_framebuffer, Point2i p_viewport_offset, Point2i p_viewport_size, bool p_clear_color, const Vector<Color> &p_clear_colors, bool p_clear_depth, float p_depth, uint32_t p_stencil) { Vector<VkClearAttachment> clear_attachments; int color_index = 0; int texture_index = 0; - for (int i = 0; i < framebuffer->texture_ids.size(); i++) { - Texture *texture = texture_owner.get_or_null(framebuffer->texture_ids[i]); + for (int i = 0; i < p_framebuffer->texture_ids.size(); i++) { + Texture *texture = texture_owner.get_or_null(p_framebuffer->texture_ids[i]); if (!texture) { texture_index++; @@ -7294,12 +7328,12 @@ void RenderingDeviceVulkan::_draw_list_insert_clear_region(DrawList *draw_list, VkClearRect cr; cr.baseArrayLayer = 0; cr.layerCount = 1; - cr.rect.offset.x = viewport_offset.x; - cr.rect.offset.y = viewport_offset.y; - cr.rect.extent.width = viewport_size.width; - cr.rect.extent.height = viewport_size.height; + cr.rect.offset.x = p_viewport_offset.x; + cr.rect.offset.y = p_viewport_offset.y; + cr.rect.extent.width = p_viewport_size.width; + cr.rect.extent.height = p_viewport_size.height; - vkCmdClearAttachments(draw_list->command_buffer, clear_attachments.size(), clear_attachments.ptr(), 1, &cr); + vkCmdClearAttachments(p_draw_list->command_buffer, clear_attachments.size(), clear_attachments.ptr(), 1, &cr); } RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_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_color_values, float p_clear_depth, uint32_t p_clear_stencil, const Rect2 &p_region, const Vector<RID> &p_storage_textures) { @@ -7353,7 +7387,9 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_framebu // If it is the first we're likely populating our VRS texture. // Bit dirty but... if (!texture || (!(texture->usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) && !(i != 0 && texture->usage_flags & TEXTURE_USAGE_VRS_ATTACHMENT_BIT))) { - color_count++; + if (!texture || !texture->is_resolve_buffer) { + 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) + ")."); @@ -7543,6 +7579,16 @@ RenderingDeviceVulkan::DrawList *RenderingDeviceVulkan::_get_draw_list_ptr(DrawL } } +void RenderingDeviceVulkan::draw_list_set_blend_constants(DrawListID p_list, const Color &p_color) { + DrawList *dl = _get_draw_list_ptr(p_list); + ERR_FAIL_COND(!dl); +#ifdef DEBUG_ENABLED + ERR_FAIL_COND_MSG(!dl->validation.active, "Submitted Draw Lists can no longer be modified."); +#endif + + vkCmdSetBlendConstants(dl->command_buffer, p_color.components); +} + void RenderingDeviceVulkan::draw_list_bind_render_pipeline(DrawListID p_list, RID p_render_pipeline) { DrawList *dl = _get_draw_list_ptr(p_list); ERR_FAIL_COND(!dl); @@ -7711,7 +7757,7 @@ void RenderingDeviceVulkan::draw_list_bind_index_array(DrawListID p_list, RID p_ dl->validation.index_array_size = index_array->indices; dl->validation.index_array_offset = index_array->offset; - vkCmdBindIndexBuffer(dl->command_buffer, index_array->buffer, index_array->offset, index_array->index_type); + vkCmdBindIndexBuffer(dl->command_buffer, index_array->buffer, 0, index_array->index_type); } void RenderingDeviceVulkan::draw_list_set_line_width(DrawListID p_list, float p_width) { @@ -7805,12 +7851,6 @@ void RenderingDeviceVulkan::draw_list_draw(DrawListID p_list, bool p_use_indices ERR_FAIL_COND_MSG(!dl->validation.index_array_size, "Draw command requested indices, but no index buffer was set."); - if (dl->validation.pipeline_vertex_format != INVALID_ID) { - // Uses vertices, do some vertex validations. - ERR_FAIL_COND_MSG(dl->validation.vertex_array_size < dl->validation.index_array_max_index, - "Index array references (max index: " + itos(dl->validation.index_array_max_index) + ") indices beyond the vertex array size (" + itos(dl->validation.vertex_array_size) + ")."); - } - ERR_FAIL_COND_MSG(dl->validation.pipeline_uses_restart_indices != dl->validation.index_buffer_uses_restart_indices, "The usage of restart indices in index buffer does not match the render primitive in the pipeline."); #endif @@ -8050,7 +8090,7 @@ void RenderingDeviceVulkan::_draw_list_free(Rect2i *r_last_viewport) { _THREAD_SAFE_UNLOCK_ } -void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { +void RenderingDeviceVulkan::draw_list_end(BitField<BarrierMask> p_post_barrier) { _THREAD_SAFE_METHOD_ ERR_FAIL_COND_MSG(!draw_list, "Immediate draw list is already inactive."); @@ -8072,15 +8112,15 @@ void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { uint32_t barrier_flags = 0; uint32_t access_flags = 0; - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { barrier_flags |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT /*| VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT*/; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_INDEX_READ_BIT | VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT /*| VK_ACCESS_INDIRECT_COMMAND_READ_BIT*/; } - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_TRANSFER_READ_BIT; } @@ -8558,20 +8598,20 @@ void RenderingDeviceVulkan::compute_list_add_barrier(ComputeListID p_list) { #endif } -void RenderingDeviceVulkan::compute_list_end(uint32_t p_post_barrier) { +void RenderingDeviceVulkan::compute_list_end(BitField<BarrierMask> p_post_barrier) { ERR_FAIL_COND(!compute_list); uint32_t barrier_flags = 0; uint32_t access_flags = 0; - if (p_post_barrier & BARRIER_MASK_COMPUTE) { + if (p_post_barrier.has_flag(BARRIER_MASK_COMPUTE)) { barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } - if (p_post_barrier & BARRIER_MASK_RASTER) { + if (p_post_barrier.has_flag(BARRIER_MASK_RASTER)) { barrier_flags |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT; access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_INDEX_READ_BIT | VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | VK_ACCESS_INDIRECT_COMMAND_READ_BIT; } - if (p_post_barrier & BARRIER_MASK_TRANSFER) { + if (p_post_barrier.has_flag(BARRIER_MASK_TRANSFER)) { barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_TRANSFER_READ_BIT; } @@ -8639,43 +8679,45 @@ void RenderingDeviceVulkan::compute_list_end(uint32_t p_post_barrier) { _THREAD_SAFE_UNLOCK_ } -void RenderingDeviceVulkan::barrier(uint32_t p_from, uint32_t p_to) { +void RenderingDeviceVulkan::barrier(BitField<BarrierMask> p_from, BitField<BarrierMask> p_to) { uint32_t src_barrier_flags = 0; uint32_t src_access_flags = 0; - if (p_from & BARRIER_MASK_COMPUTE) { - src_barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; - src_access_flags |= VK_ACCESS_SHADER_WRITE_BIT; - } - if (p_from & BARRIER_MASK_RASTER) { - src_barrier_flags |= VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT; - src_access_flags |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT; - } - if (p_from & BARRIER_MASK_TRANSFER) { - src_barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; - src_access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; - } if (p_from == 0) { src_barrier_flags = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT; + } else { + if (p_from.has_flag(BARRIER_MASK_COMPUTE)) { + src_barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; + src_access_flags |= VK_ACCESS_SHADER_WRITE_BIT; + } + if (p_from.has_flag(BARRIER_MASK_RASTER)) { + src_barrier_flags |= VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT; + src_access_flags |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT; + } + if (p_from.has_flag(BARRIER_MASK_TRANSFER)) { + src_barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; + src_access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; + } } uint32_t dst_barrier_flags = 0; uint32_t dst_access_flags = 0; - if (p_to & BARRIER_MASK_COMPUTE) { - dst_barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; - dst_access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; - } - if (p_to & BARRIER_MASK_RASTER) { - dst_barrier_flags |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT; - dst_access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_INDEX_READ_BIT | VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | VK_ACCESS_INDIRECT_COMMAND_READ_BIT; - } - if (p_to & BARRIER_MASK_TRANSFER) { - dst_barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; - dst_access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_TRANSFER_READ_BIT; - } if (p_to == 0) { dst_barrier_flags = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } else { + if (p_to.has_flag(BARRIER_MASK_COMPUTE)) { + dst_barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; + dst_access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; + } + if (p_to.has_flag(BARRIER_MASK_RASTER)) { + dst_barrier_flags |= VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT; + dst_access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_INDEX_READ_BIT | VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | VK_ACCESS_INDIRECT_COMMAND_READ_BIT; + } + if (p_to.has_flag(BARRIER_MASK_TRANSFER)) { + dst_barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; + dst_access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_TRANSFER_READ_BIT; + } } _memory_barrier(src_barrier_flags, dst_barrier_flags, src_access_flags, dst_access_flags, true); @@ -8989,6 +9031,25 @@ void RenderingDeviceVulkan::_begin_frame() { frames[frame].index = Engine::get_singleton()->get_frames_drawn(); } +VkSampleCountFlagBits RenderingDeviceVulkan::_ensure_supported_sample_count(TextureSamples p_requested_sample_count) const { + VkSampleCountFlags sample_count_flags = limits.framebufferColorSampleCounts & limits.framebufferDepthSampleCounts; + + if (sample_count_flags & rasterization_sample_count[p_requested_sample_count]) { + // The requested sample count is supported. + return rasterization_sample_count[p_requested_sample_count]; + } else { + // Find the closest lower supported sample count. + VkSampleCountFlagBits sample_count = rasterization_sample_count[p_requested_sample_count]; + while (sample_count > VK_SAMPLE_COUNT_1_BIT) { + if (sample_count_flags & sample_count) { + return sample_count; + } + sample_count = (VkSampleCountFlagBits)(sample_count >> 1); + } + } + return VK_SAMPLE_COUNT_1_BIT; +} + void RenderingDeviceVulkan::swap_buffers() { ERR_FAIL_COND_MSG(local_device.is_valid(), "Local devices can't swap buffers."); _THREAD_SAFE_METHOD_ @@ -9133,7 +9194,7 @@ void RenderingDeviceVulkan::_free_pending_resources(int p_frame) { Texture *texture = &frames[p_frame].textures_to_dispose_of.front()->get(); if (texture->bound) { - WARN_PRINT("Deleted a texture while it was bound.."); + WARN_PRINT("Deleted a texture while it was bound."); } vkDestroyImageView(device, texture->view, nullptr); if (texture->owner.is_null()) { @@ -9334,12 +9395,10 @@ void RenderingDeviceVulkan::initialize(VulkanContext *p_context, bool p_local_de } } - // NOTE: If adding new project settings here, also duplicate their definition in - // rendering_server.cpp for headless doctool. - staging_buffer_block_size = GLOBAL_DEF("rendering/vulkan/staging_buffer/block_size_kb", 256); + staging_buffer_block_size = GLOBAL_GET("rendering/rendering_device/staging_buffer/block_size_kb"); staging_buffer_block_size = MAX(4u, staging_buffer_block_size); staging_buffer_block_size *= 1024; // Kb -> bytes. - staging_buffer_max_size = GLOBAL_DEF("rendering/vulkan/staging_buffer/max_size_mb", 128); + staging_buffer_max_size = GLOBAL_GET("rendering/rendering_device/staging_buffer/max_size_mb"); staging_buffer_max_size = MAX(1u, staging_buffer_max_size); staging_buffer_max_size *= 1024 * 1024; @@ -9347,7 +9406,7 @@ void RenderingDeviceVulkan::initialize(VulkanContext *p_context, bool p_local_de // Validate enough blocks. staging_buffer_max_size = staging_buffer_block_size * 4; } - texture_upload_region_size_px = GLOBAL_DEF("rendering/vulkan/staging_buffer/texture_upload_region_size_px", 64); + texture_upload_region_size_px = GLOBAL_GET("rendering/rendering_device/staging_buffer/texture_upload_region_size_px"); texture_upload_region_size_px = nearest_power_of_2_templated(texture_upload_region_size_px); frames_drawn = frame_count; // Start from frame count, so everything else is immediately old. @@ -9362,7 +9421,7 @@ void RenderingDeviceVulkan::initialize(VulkanContext *p_context, bool p_local_de ERR_CONTINUE(err != OK); } - max_descriptors_per_pool = GLOBAL_DEF("rendering/vulkan/descriptor_pools/max_descriptors_per_pool", 64); + max_descriptors_per_pool = GLOBAL_GET("rendering/rendering_device/vulkan/max_descriptors_per_pool"); // Check to make sure DescriptorPoolKey is good. static_assert(sizeof(uint64_t) * 3 >= UNIFORM_TYPE_MAX * sizeof(uint16_t)); @@ -9657,6 +9716,10 @@ uint64_t RenderingDeviceVulkan::limit_get(Limit p_limit) const { return limits.maxComputeWorkGroupSize[1]; case LIMIT_MAX_COMPUTE_WORKGROUP_SIZE_Z: return limits.maxComputeWorkGroupSize[2]; + case LIMIT_MAX_VIEWPORT_DIMENSIONS_X: + return limits.maxViewportDimensions[0]; + case LIMIT_MAX_VIEWPORT_DIMENSIONS_Y: + return limits.maxViewportDimensions[1]; case LIMIT_SUBGROUP_SIZE: { VulkanContext::SubgroupCapabilities subgroup_capabilities = context->get_subgroup_capabilities(); return subgroup_capabilities.size; @@ -9669,6 +9732,12 @@ uint64_t RenderingDeviceVulkan::limit_get(Limit p_limit) const { VulkanContext::SubgroupCapabilities subgroup_capabilities = context->get_subgroup_capabilities(); return subgroup_capabilities.supported_operations_flags_rd(); } + case LIMIT_VRS_TEXEL_WIDTH: { + return context->get_vrs_capabilities().texel_size.x; + } + case LIMIT_VRS_TEXEL_HEIGHT: { + return context->get_vrs_capabilities().texel_size.y; + } default: ERR_FAIL_V(0); } diff --git a/drivers/vulkan/rendering_device_vulkan.h b/drivers/vulkan/rendering_device_vulkan.h index 6572de7c52..537ad88f5a 100644 --- a/drivers/vulkan/rendering_device_vulkan.h +++ b/drivers/vulkan/rendering_device_vulkan.h @@ -150,6 +150,8 @@ class RenderingDeviceVulkan : public RenderingDevice { bool used_in_raster = false; bool used_in_compute = false; + bool is_resolve_buffer = false; + uint32_t read_aspect_mask = 0; uint32_t barrier_aspect_mask = 0; bool bound = false; // Bound to framebffer. @@ -160,7 +162,7 @@ class RenderingDeviceVulkan : public RenderingDevice { uint32_t texture_upload_region_size_px = 0; Vector<uint8_t> _texture_get_data_from_image(Texture *tex, VkImage p_image, VmaAllocation p_allocation, uint32_t p_layer, bool p_2d = false); - Error _texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, uint32_t p_post_barrier, bool p_use_setup_queue); + Error _texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, BitField<BarrierMask> p_post_barrier, bool p_use_setup_queue); /*****************/ /**** SAMPLER ****/ @@ -903,11 +905,11 @@ class RenderingDeviceVulkan : public RenderingDevice { bool draw_list_unbind_color_textures = false; 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); + void _draw_list_insert_clear_region(DrawList *p_draw_list, Framebuffer *p_framebuffer, Point2i p_viewport_offset, Point2i p_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, 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); + Buffer *_get_buffer_from_owner(RID p_buffer, VkPipelineStageFlags &dst_stage_mask, VkAccessFlags &dst_access, BitField<BarrierMask> 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); @@ -1042,13 +1044,15 @@ class RenderingDeviceVulkan : public RenderingDevice { HashMap<RID, String> resource_names; #endif + VkSampleCountFlagBits _ensure_supported_sample_count(TextureSamples p_requested_sample_count) const; + public: virtual RID texture_create(const TextureFormat &p_format, const TextureView &p_view, const Vector<Vector<uint8_t>> &p_data = Vector<Vector<uint8_t>>()); virtual RID texture_create_shared(const TextureView &p_view, RID p_with_texture); virtual RID texture_create_from_extension(TextureType p_type, DataFormat p_format, TextureSamples p_samples, uint64_t p_flags, uint64_t p_image, uint64_t p_width, uint64_t p_height, uint64_t p_depth, uint64_t p_layers); virtual RID texture_create_shared_from_slice(const TextureView &p_view, RID p_with_texture, uint32_t p_layer, uint32_t p_mipmap, uint32_t p_mipmaps = 1, TextureSliceType p_slice_type = TEXTURE_SLICE_2D); - virtual Error texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, uint32_t p_post_barrier = BARRIER_MASK_ALL); + virtual Error texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); virtual Vector<uint8_t> texture_get_data(RID p_texture, uint32_t p_layer); virtual bool texture_is_format_supported_for_usage(DataFormat p_format, uint32_t p_usage) const; @@ -1056,9 +1060,9 @@ public: virtual bool texture_is_valid(RID p_texture); virtual Size2i texture_size(RID p_texture); - virtual Error texture_copy(RID p_from_texture, RID p_to_texture, const Vector3 &p_from, const Vector3 &p_to, const Vector3 &p_size, uint32_t p_src_mipmap, uint32_t p_dst_mipmap, uint32_t p_src_layer, uint32_t p_dst_layer, uint32_t p_post_barrier = BARRIER_MASK_ALL); - virtual Error texture_clear(RID p_texture, const Color &p_color, uint32_t p_base_mipmap, uint32_t p_mipmaps, uint32_t p_base_layer, uint32_t p_layers, uint32_t p_post_barrier = BARRIER_MASK_ALL); - virtual Error texture_resolve_multisample(RID p_from_texture, RID p_to_texture, uint32_t p_post_barrier = BARRIER_MASK_ALL); + virtual Error texture_copy(RID p_from_texture, RID p_to_texture, const Vector3 &p_from, const Vector3 &p_to, const Vector3 &p_size, uint32_t p_src_mipmap, uint32_t p_dst_mipmap, uint32_t p_src_layer, uint32_t p_dst_layer, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); + virtual Error texture_clear(RID p_texture, const Color &p_color, uint32_t p_base_mipmap, uint32_t p_mipmaps, uint32_t p_base_layer, uint32_t p_layers, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); + virtual Error texture_resolve_multisample(RID p_from_texture, RID p_to_texture, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); /*********************/ /**** FRAMEBUFFER ****/ @@ -1091,7 +1095,7 @@ public: // Internally reference counted, this ID is warranted to be unique for the same description, but needs to be freed as many times as it was allocated. virtual VertexFormatID vertex_format_create(const Vector<VertexAttribute> &p_vertex_formats); - virtual RID vertex_array_create(uint32_t p_vertex_count, VertexFormatID p_vertex_format, const Vector<RID> &p_src_buffers); + virtual RID vertex_array_create(uint32_t p_vertex_count, VertexFormatID p_vertex_format, const Vector<RID> &p_src_buffers, const Vector<uint64_t> &p_offsets = Vector<uint64_t>()); virtual RID index_buffer_create(uint32_t p_size_indices, IndexBufferFormat p_format, const Vector<uint8_t> &p_data = Vector<uint8_t>(), bool p_use_restart_indices = false); @@ -1120,8 +1124,8 @@ public: virtual bool uniform_set_is_valid(RID p_uniform_set); virtual void uniform_set_set_invalidation_callback(RID p_uniform_set, InvalidationCallback p_callback, void *p_userdata); - virtual Error buffer_update(RID p_buffer, uint32_t p_offset, uint32_t p_size, const void *p_data, uint32_t p_post_barrier = BARRIER_MASK_ALL); // Works for any buffer. - virtual Error buffer_clear(RID p_buffer, uint32_t p_offset, uint32_t p_size, uint32_t p_post_barrier = BARRIER_MASK_ALL); + virtual Error buffer_update(RID p_buffer, uint32_t p_offset, uint32_t p_size, const void *p_data, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); // Works for any buffer. + virtual Error buffer_clear(RID p_buffer, uint32_t p_offset, uint32_t p_size, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); virtual Vector<uint8_t> buffer_get_data(RID p_buffer); /*************************/ @@ -1155,6 +1159,7 @@ public: virtual DrawListID draw_list_begin(RID p_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_color_values = Vector<Color>(), float p_clear_depth = 1.0, uint32_t p_clear_stencil = 0, const Rect2 &p_region = Rect2(), const Vector<RID> &p_storage_textures = Vector<RID>()); virtual Error draw_list_begin_split(RID p_framebuffer, uint32_t p_splits, DrawListID *r_split_ids, 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_color_values = Vector<Color>(), float p_clear_depth = 1.0, uint32_t p_clear_stencil = 0, const Rect2 &p_region = Rect2(), const Vector<RID> &p_storage_textures = Vector<RID>()); + virtual void draw_list_set_blend_constants(DrawListID p_list, const Color &p_color); virtual void draw_list_bind_render_pipeline(DrawListID p_list, RID p_render_pipeline); virtual void draw_list_bind_uniform_set(DrawListID p_list, RID p_uniform_set, uint32_t p_index); virtual void draw_list_bind_vertex_array(DrawListID p_list, RID p_vertex_array); @@ -1171,7 +1176,7 @@ public: 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); + virtual void draw_list_end(BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); /***********************/ /**** COMPUTE LISTS ****/ @@ -1186,9 +1191,9 @@ public: virtual void compute_list_dispatch(ComputeListID p_list, uint32_t p_x_groups, uint32_t p_y_groups, uint32_t p_z_groups); virtual void compute_list_dispatch_threads(ComputeListID p_list, uint32_t p_x_threads, uint32_t p_y_threads, uint32_t p_z_threads); virtual void compute_list_dispatch_indirect(ComputeListID p_list, RID p_buffer, uint32_t p_offset); - virtual void compute_list_end(uint32_t p_post_barrier = BARRIER_MASK_ALL); + virtual void compute_list_end(BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); - virtual void barrier(uint32_t p_from = BARRIER_MASK_ALL, uint32_t p_to = BARRIER_MASK_ALL); + virtual void barrier(BitField<BarrierMask> p_from = BARRIER_MASK_ALL_BARRIERS, BitField<BarrierMask> p_to = BARRIER_MASK_ALL_BARRIERS); virtual void full_barrier(); /**************/ diff --git a/drivers/vulkan/vulkan_context.cpp b/drivers/vulkan/vulkan_context.cpp index afc3e78372..381df6d65e 100644 --- a/drivers/vulkan/vulkan_context.cpp +++ b/drivers/vulkan/vulkan_context.cpp @@ -48,15 +48,119 @@ VulkanHooks *VulkanContext::vulkan_hooks = nullptr; -VkResult VulkanContext::vkCreateRenderPass2KHR(VkDevice device, const VkRenderPassCreateInfo2 *pCreateInfo, const VkAllocationCallbacks *pAllocator, VkRenderPass *pRenderPass) { - if (fpCreateRenderPass2KHR == nullptr) { - fpCreateRenderPass2KHR = (PFN_vkCreateRenderPass2KHR)vkGetInstanceProcAddr(inst, "vkCreateRenderPass2KHR"); +Vector<VkAttachmentReference> VulkanContext::_convert_VkAttachmentReference2(uint32_t p_count, const VkAttachmentReference2 *p_refs) { + Vector<VkAttachmentReference> att_refs; + + if (p_refs != nullptr) { + for (uint32_t i = 0; i < p_count; i++) { + // We lose aspectMask in this conversion but we don't use it currently. + + VkAttachmentReference ref = { + p_refs[i].attachment, /* attachment */ + p_refs[i].layout /* layout */ + }; + + att_refs.push_back(ref); + } } - if (fpCreateRenderPass2KHR == nullptr) { - return VK_ERROR_EXTENSION_NOT_PRESENT; + return att_refs; +} + +VkResult VulkanContext::vkCreateRenderPass2KHR(VkDevice p_device, const VkRenderPassCreateInfo2 *p_create_info, const VkAllocationCallbacks *p_allocator, VkRenderPass *p_render_pass) { + if (has_renderpass2_ext) { + if (fpCreateRenderPass2KHR == nullptr) { + fpCreateRenderPass2KHR = (PFN_vkCreateRenderPass2KHR)vkGetDeviceProcAddr(p_device, "vkCreateRenderPass2KHR"); + } + + if (fpCreateRenderPass2KHR == nullptr) { + return VK_ERROR_EXTENSION_NOT_PRESENT; + } else { + return (fpCreateRenderPass2KHR)(p_device, p_create_info, p_allocator, p_render_pass); + } } else { - return (fpCreateRenderPass2KHR)(device, pCreateInfo, pAllocator, pRenderPass); + // need to fall back on vkCreateRenderPass + + const void *next = p_create_info->pNext; // ATM we only support multiview which should work if supported. + + Vector<VkAttachmentDescription> attachments; + for (uint32_t i = 0; i < p_create_info->attachmentCount; i++) { + // Basically the old layout just misses type and next. + VkAttachmentDescription att = { + p_create_info->pAttachments[i].flags, /* flags */ + p_create_info->pAttachments[i].format, /* format */ + p_create_info->pAttachments[i].samples, /* samples */ + p_create_info->pAttachments[i].loadOp, /* loadOp */ + p_create_info->pAttachments[i].storeOp, /* storeOp */ + p_create_info->pAttachments[i].stencilLoadOp, /* stencilLoadOp */ + p_create_info->pAttachments[i].stencilStoreOp, /* stencilStoreOp */ + p_create_info->pAttachments[i].initialLayout, /* initialLayout */ + p_create_info->pAttachments[i].finalLayout /* finalLayout */ + }; + + attachments.push_back(att); + } + + Vector<VkSubpassDescription> subpasses; + for (uint32_t i = 0; i < p_create_info->subpassCount; i++) { + // Here we need to do more, again it's just stripping out type and next + // but we have VkAttachmentReference2 to convert to VkAttachmentReference. + // Also viewmask is not supported but we don't use it outside of multiview. + + Vector<VkAttachmentReference> input_attachments = _convert_VkAttachmentReference2(p_create_info->pSubpasses[i].inputAttachmentCount, p_create_info->pSubpasses[i].pInputAttachments); + Vector<VkAttachmentReference> color_attachments = _convert_VkAttachmentReference2(p_create_info->pSubpasses[i].colorAttachmentCount, p_create_info->pSubpasses[i].pColorAttachments); + Vector<VkAttachmentReference> resolve_attachments = _convert_VkAttachmentReference2(p_create_info->pSubpasses[i].colorAttachmentCount, p_create_info->pSubpasses[i].pResolveAttachments); + Vector<VkAttachmentReference> depth_attachments = _convert_VkAttachmentReference2(p_create_info->pSubpasses[i].colorAttachmentCount, p_create_info->pSubpasses[i].pDepthStencilAttachment); + + VkSubpassDescription subpass = { + p_create_info->pSubpasses[i].flags, /* flags */ + p_create_info->pSubpasses[i].pipelineBindPoint, /* pipelineBindPoint */ + p_create_info->pSubpasses[i].inputAttachmentCount, /* inputAttachmentCount */ + input_attachments.size() == 0 ? nullptr : input_attachments.ptr(), /* pInputAttachments */ + p_create_info->pSubpasses[i].colorAttachmentCount, /* colorAttachmentCount */ + color_attachments.size() == 0 ? nullptr : color_attachments.ptr(), /* pColorAttachments */ + resolve_attachments.size() == 0 ? nullptr : resolve_attachments.ptr(), /* pResolveAttachments */ + depth_attachments.size() == 0 ? nullptr : depth_attachments.ptr(), /* pDepthStencilAttachment */ + p_create_info->pSubpasses[i].preserveAttachmentCount, /* preserveAttachmentCount */ + p_create_info->pSubpasses[i].pPreserveAttachments /* pPreserveAttachments */ + }; + + subpasses.push_back(subpass); + } + + Vector<VkSubpassDependency> dependencies; + for (uint32_t i = 0; i < p_create_info->dependencyCount; i++) { + // We lose viewOffset here but again I don't believe we use this anywhere. + VkSubpassDependency dep = { + p_create_info->pDependencies[i].srcSubpass, /* srcSubpass */ + p_create_info->pDependencies[i].dstSubpass, /* dstSubpass */ + p_create_info->pDependencies[i].srcStageMask, /* srcStageMask */ + p_create_info->pDependencies[i].dstStageMask, /* dstStageMask */ + p_create_info->pDependencies[i].srcAccessMask, /* srcAccessMask */ + p_create_info->pDependencies[i].dstAccessMask, /* dstAccessMask */ + p_create_info->pDependencies[i].dependencyFlags, /* dependencyFlags */ + }; + + dependencies.push_back(dep); + } + + // CorrelatedViewMask is not supported in vkCreateRenderPass but we + // currently only use this for multiview. + // We'll need to look into this. + + VkRenderPassCreateInfo create_info = { + VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO, /* sType */ + next, /* pNext*/ + p_create_info->flags, /* flags */ + (uint32_t)attachments.size(), /* attachmentCount */ + attachments.ptr(), /* pAttachments */ + (uint32_t)subpasses.size(), /* subpassCount */ + subpasses.ptr(), /* pSubpasses */ + (uint32_t)dependencies.size(), /* */ + dependencies.ptr(), /* */ + }; + + return vkCreateRenderPass(device, &create_info, p_allocator, p_render_pass); } } @@ -85,19 +189,6 @@ VKAPI_ATTR VkBool32 VKAPI_CALL VulkanContext::_debug_messenger_callback( return VK_FALSE; } - // Workaround for Vulkan-Loader usability bug: https://github.com/KhronosGroup/Vulkan-Loader/issues/262. - if (strstr(pCallbackData->pMessage, "wrong ELF class: ELFCLASS32") != nullptr) { - return VK_FALSE; - } - -#ifdef WINDOWS_ENABLED - // Some software installs Vulkan overlays in Windows registry and never cleans them up on uninstall. - // So we get spammy error level messages from the loader about those - make them verbose instead. - if (strstr(pCallbackData->pMessage, "loader_get_json: Failed to open JSON file") != nullptr) { - messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT; - } -#endif - if (pCallbackData->pMessageIdName && strstr(pCallbackData->pMessageIdName, "UNASSIGNED-CoreValidation-DrawState-ClearCmdBeforeDraw") != nullptr) { return VK_FALSE; } @@ -320,6 +411,16 @@ Error VulkanContext::_initialize_extensions() { VkBool32 platformSurfaceExtFound = 0; memset(extension_names, 0, sizeof(extension_names)); + // Only enable debug utils in verbose mode or DEV_ENABLED. + // End users would get spammed with messages of varying verbosity due to the + // mess that thirdparty layers/extensions and drivers seem to leave in their + // wake, making the Windows registry a bottomless pit of broken layer JSON. +#ifdef DEV_ENABLED + bool want_debug_utils = true; +#else + bool want_debug_utils = OS::get_singleton()->is_stdout_verbose(); +#endif + VkResult err = vkEnumerateInstanceExtensionProperties(nullptr, &instance_extension_count, nullptr); ERR_FAIL_COND_V(err != VK_SUCCESS && err != VK_INCOMPLETE, ERR_CANT_CREATE); @@ -347,8 +448,10 @@ Error VulkanContext::_initialize_extensions() { } } if (!strcmp(VK_EXT_DEBUG_UTILS_EXTENSION_NAME, instance_extensions[i].extensionName)) { - extension_names[enabled_extension_count++] = VK_EXT_DEBUG_UTILS_EXTENSION_NAME; - enabled_debug_utils = true; + if (want_debug_utils) { + 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; @@ -522,6 +625,9 @@ Error VulkanContext::_check_capabilities() { vrs_capabilities.pipeline_vrs_supported = false; vrs_capabilities.primitive_vrs_supported = false; vrs_capabilities.attachment_vrs_supported = false; + vrs_capabilities.min_texel_size = Size2i(); + vrs_capabilities.max_texel_size = Size2i(); + vrs_capabilities.texel_size = Size2i(); multiview_capabilities.is_supported = false; multiview_capabilities.geometry_shader_is_supported = false; multiview_capabilities.tessellation_shader_is_supported = false; @@ -584,21 +690,43 @@ Error VulkanContext::_check_capabilities() { vkGetPhysicalDeviceFeatures2_func(gpu, &device_features); - vrs_capabilities.pipeline_vrs_supported = vrs_features.pipelineFragmentShadingRate; - vrs_capabilities.primitive_vrs_supported = vrs_features.primitiveFragmentShadingRate; - vrs_capabilities.attachment_vrs_supported = vrs_features.attachmentFragmentShadingRate; + // We must check that the relative extension is present before assuming a + // feature as enabled. Actually, according to the spec we shouldn't add the + // structs in pNext at all, but this works fine. + // See also: https://github.com/godotengine/godot/issues/65409 + for (uint32_t i = 0; i < enabled_extension_count; ++i) { + if (!strcmp(VK_KHR_FRAGMENT_SHADING_RATE_EXTENSION_NAME, extension_names[i])) { + vrs_capabilities.pipeline_vrs_supported = vrs_features.pipelineFragmentShadingRate; + vrs_capabilities.primitive_vrs_supported = vrs_features.primitiveFragmentShadingRate; + vrs_capabilities.attachment_vrs_supported = vrs_features.attachmentFragmentShadingRate; - multiview_capabilities.is_supported = multiview_features.multiview; - multiview_capabilities.geometry_shader_is_supported = multiview_features.multiviewGeometryShader; - multiview_capabilities.tessellation_shader_is_supported = multiview_features.multiviewTessellationShader; + continue; + } + + if (!strcmp(VK_KHR_MULTIVIEW_EXTENSION_NAME, extension_names[i])) { + multiview_capabilities.is_supported = multiview_features.multiview; + multiview_capabilities.geometry_shader_is_supported = multiview_features.multiviewGeometryShader; + multiview_capabilities.tessellation_shader_is_supported = multiview_features.multiviewTessellationShader; - shader_capabilities.shader_float16_is_supported = shader_features.shaderFloat16; - shader_capabilities.shader_int8_is_supported = shader_features.shaderInt8; + continue; + } - storage_buffer_capabilities.storage_buffer_16_bit_access_is_supported = storage_feature.storageBuffer16BitAccess; - storage_buffer_capabilities.uniform_and_storage_buffer_16_bit_access_is_supported = storage_feature.uniformAndStorageBuffer16BitAccess; - storage_buffer_capabilities.storage_push_constant_16_is_supported = storage_feature.storagePushConstant16; - storage_buffer_capabilities.storage_input_output_16 = storage_feature.storageInputOutput16; + if (!strcmp(VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME, extension_names[i])) { + shader_capabilities.shader_float16_is_supported = shader_features.shaderFloat16; + shader_capabilities.shader_int8_is_supported = shader_features.shaderInt8; + + continue; + } + + if (!strcmp(VK_KHR_16BIT_STORAGE_EXTENSION_NAME, extension_names[i])) { + storage_buffer_capabilities.storage_buffer_16_bit_access_is_supported = storage_feature.storageBuffer16BitAccess; + storage_buffer_capabilities.uniform_and_storage_buffer_16_bit_access_is_supported = storage_feature.uniformAndStorageBuffer16BitAccess; + storage_buffer_capabilities.storage_push_constant_16_is_supported = storage_feature.storagePushConstant16; + storage_buffer_capabilities.storage_input_output_16 = storage_feature.storageInputOutput16; + + continue; + } + } } // Check extended properties. @@ -608,15 +736,18 @@ Error VulkanContext::_check_capabilities() { device_properties_func = (PFN_vkGetPhysicalDeviceProperties2)vkGetInstanceProcAddr(inst, "vkGetPhysicalDeviceProperties2KHR"); } if (device_properties_func != nullptr) { - VkPhysicalDeviceFragmentShadingRatePropertiesKHR vrsProperties; - VkPhysicalDeviceMultiviewProperties multiviewProperties; - VkPhysicalDeviceSubgroupProperties subgroupProperties; - VkPhysicalDeviceProperties2 physicalDeviceProperties; + VkPhysicalDeviceFragmentShadingRatePropertiesKHR vrsProperties{}; + VkPhysicalDeviceMultiviewProperties multiviewProperties{}; + VkPhysicalDeviceSubgroupProperties subgroupProperties{}; + VkPhysicalDeviceProperties2 physicalDeviceProperties{}; void *nextptr = nullptr; - subgroupProperties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES; - subgroupProperties.pNext = nextptr; - nextptr = &subgroupProperties; + if (!(vulkan_major == 1 && vulkan_minor == 0)) { + subgroupProperties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES; + subgroupProperties.pNext = nextptr; + + nextptr = &subgroupProperties; + } if (multiview_capabilities.is_supported) { multiviewProperties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES; @@ -660,6 +791,10 @@ Error VulkanContext::_check_capabilities() { vrs_capabilities.max_texel_size.x = vrsProperties.maxFragmentShadingRateAttachmentTexelSize.width; vrs_capabilities.max_texel_size.y = vrsProperties.maxFragmentShadingRateAttachmentTexelSize.height; + // We'll attempt to default to a texel size of 16x16 + vrs_capabilities.texel_size.x = CLAMP(16, vrs_capabilities.min_texel_size.x, vrs_capabilities.max_texel_size.x); + vrs_capabilities.texel_size.y = CLAMP(16, vrs_capabilities.min_texel_size.y, vrs_capabilities.max_texel_size.y); + print_verbose(String(" Attachment fragment shading rate") + String(", min texel size: (") + itos(vrs_capabilities.min_texel_size.x) + String(", ") + itos(vrs_capabilities.min_texel_size.y) + String(")") + String(", max texel size: (") + itos(vrs_capabilities.max_texel_size.x) + String(", ") + itos(vrs_capabilities.max_texel_size.y) + String(")")); } @@ -704,7 +839,7 @@ Error VulkanContext::_create_instance() { } } - CharString cs = ProjectSettings::get_singleton()->get("application/config/name").operator String().utf8(); + CharString cs = GLOBAL_GET("application/config/name").operator String().utf8(); const VkApplicationInfo app = { /*sType*/ VK_STRUCTURE_TYPE_APPLICATION_INFO, /*pNext*/ nullptr, @@ -1061,6 +1196,7 @@ Error VulkanContext::_create_physical_device(VkSurfaceKHR p_surface) { extension_names[enabled_extension_count++] = VK_KHR_FRAGMENT_SHADING_RATE_EXTENSION_NAME; } if (!strcmp(VK_KHR_CREATE_RENDERPASS_2_EXTENSION_NAME, device_extensions[i].extensionName)) { + has_renderpass2_ext = true; extension_names[enabled_extension_count++] = VK_KHR_CREATE_RENDERPASS_2_EXTENSION_NAME; } if (enabled_extension_count >= MAX_EXTENSIONS) { @@ -1207,7 +1343,7 @@ Error VulkanContext::_create_device() { vulkan11features.shaderDrawParameters = 0; nextptr = &vulkan11features; } else { - // On Vulkan 1.0 and 1.1 we use our older structs to initialise these features. + // On Vulkan 1.0 and 1.1 we use our older structs to initialize these features. storage_feature.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR; storage_feature.pNext = nextptr; storage_feature.storageBuffer16BitAccess = storage_buffer_capabilities.storage_buffer_16_bit_access_is_supported; @@ -1354,7 +1490,7 @@ Error VulkanContext::_initialize_queues(VkSurfaceKHR p_surface) { color_space = surfFormats[0].colorSpace; } else { // These should be ordered with the ones we want to use on top and fallback modes further down - // we want a 32bit RGBA unsigned normalised buffer or similar. + // we want a 32bit RGBA unsigned normalized buffer or similar. const VkFormat allowed_formats[] = { VK_FORMAT_B8G8R8A8_UNORM, VK_FORMAT_R8G8B8A8_UNORM @@ -1436,6 +1572,24 @@ bool VulkanContext::_use_validation_layers() { return Engine::get_singleton()->is_validation_layers_enabled(); } +VkExtent2D VulkanContext::_compute_swapchain_extent(const VkSurfaceCapabilitiesKHR &p_surf_capabilities, int *p_window_width, int *p_window_height) const { + // Width and height are either both 0xFFFFFFFF, or both not 0xFFFFFFFF. + if (p_surf_capabilities.currentExtent.width == 0xFFFFFFFF) { + // If the surface size is undefined, the size is set to the size + // of the images requested, which must fit within the minimum and + // maximum values. + VkExtent2D extent = {}; + extent.width = CLAMP((uint32_t)(*p_window_width), p_surf_capabilities.minImageExtent.width, p_surf_capabilities.maxImageExtent.width); + extent.height = CLAMP((uint32_t)(*p_window_height), p_surf_capabilities.minImageExtent.height, p_surf_capabilities.maxImageExtent.height); + return extent; + } else { + // If the surface size is defined, the swap chain size must match. + *p_window_width = p_surf_capabilities.currentExtent.width; + *p_window_height = p_surf_capabilities.currentExtent.height; + return p_surf_capabilities.currentExtent; + } +} + Error VulkanContext::_window_create(DisplayServer::WindowID p_window_id, DisplayServer::VSyncMode p_vsync_mode, VkSurfaceKHR p_surface, int p_width, int p_height) { ERR_FAIL_COND_V(windows.has(p_window_id), ERR_INVALID_PARAMETER); @@ -1576,32 +1730,7 @@ Error VulkanContext::_update_swap_chain(Window *window) { ERR_FAIL_V(ERR_CANT_CREATE); } - VkExtent2D swapchainExtent; - // Width and height are either both 0xFFFFFFFF, or both not 0xFFFFFFFF. - if (surfCapabilities.currentExtent.width == 0xFFFFFFFF) { - // If the surface size is undefined, the size is set to the size - // of the images requested, which must fit within the minimum and - // maximum values. - swapchainExtent.width = window->width; - swapchainExtent.height = window->height; - - if (swapchainExtent.width < surfCapabilities.minImageExtent.width) { - swapchainExtent.width = surfCapabilities.minImageExtent.width; - } else if (swapchainExtent.width > surfCapabilities.maxImageExtent.width) { - swapchainExtent.width = surfCapabilities.maxImageExtent.width; - } - - if (swapchainExtent.height < surfCapabilities.minImageExtent.height) { - swapchainExtent.height = surfCapabilities.minImageExtent.height; - } else if (swapchainExtent.height > surfCapabilities.maxImageExtent.height) { - swapchainExtent.height = surfCapabilities.maxImageExtent.height; - } - } else { - // If the surface size is defined, the swap chain size must match. - swapchainExtent = surfCapabilities.currentExtent; - window->width = surfCapabilities.currentExtent.width; - window->height = surfCapabilities.currentExtent.height; - } + VkExtent2D swapchainExtent = _compute_swapchain_extent(surfCapabilities, &window->width, &window->height); if (window->width == 0 || window->height == 0) { free(presentModes); @@ -1710,18 +1839,22 @@ Error VulkanContext::_update_swap_chain(Window *window) { preTransform = surfCapabilities.currentTransform; } - // Find a supported composite alpha mode - one of these is guaranteed to be set. VkCompositeAlphaFlagBitsKHR compositeAlpha = VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR; - VkCompositeAlphaFlagBitsKHR compositeAlphaFlags[4] = { - VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR, - VK_COMPOSITE_ALPHA_PRE_MULTIPLIED_BIT_KHR, - VK_COMPOSITE_ALPHA_POST_MULTIPLIED_BIT_KHR, - VK_COMPOSITE_ALPHA_INHERIT_BIT_KHR, - }; - for (uint32_t i = 0; i < ARRAY_SIZE(compositeAlphaFlags); i++) { - if (surfCapabilities.supportedCompositeAlpha & compositeAlphaFlags[i]) { - compositeAlpha = compositeAlphaFlags[i]; - break; + + if (OS::get_singleton()->is_layered_allowed() || !(surfCapabilities.supportedCompositeAlpha & compositeAlpha)) { + // Find a supported composite alpha mode - one of these is guaranteed to be set. + VkCompositeAlphaFlagBitsKHR compositeAlphaFlags[4] = { + VK_COMPOSITE_ALPHA_PRE_MULTIPLIED_BIT_KHR, + VK_COMPOSITE_ALPHA_POST_MULTIPLIED_BIT_KHR, + VK_COMPOSITE_ALPHA_INHERIT_BIT_KHR, + VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR, + }; + + for (uint32_t i = 0; i < ARRAY_SIZE(compositeAlphaFlags); i++) { + if (surfCapabilities.supportedCompositeAlpha & compositeAlphaFlags[i]) { + compositeAlpha = compositeAlphaFlags[i]; + break; + } } } @@ -2268,8 +2401,6 @@ Error VulkanContext::swap_buffers() { } } #endif - static int total_frames = 0; - total_frames++; // print_line("current buffer: " + itos(current_buffer)); err = fpQueuePresentKHR(present_queue, &present); diff --git a/drivers/vulkan/vulkan_context.h b/drivers/vulkan/vulkan_context.h index 5cc3b515d9..8cf33fa463 100644 --- a/drivers/vulkan/vulkan_context.h +++ b/drivers/vulkan/vulkan_context.h @@ -76,6 +76,8 @@ public: Size2i min_texel_size; Size2i max_texel_size; + + Size2i texel_size; // The texel size we'll use }; struct ShaderCapabilities { @@ -188,6 +190,7 @@ private: uint32_t enabled_extension_count = 0; const char *extension_names[MAX_EXTENSIONS]; bool enabled_debug_utils = false; + bool has_renderpass2_ext = false; /** * True if VK_EXT_debug_report extension is used. VK_EXT_debug_report is deprecated but it is @@ -257,6 +260,8 @@ private: Error _create_swap_chain(); Error _create_semaphores(); + Vector<VkAttachmentReference> _convert_VkAttachmentReference2(uint32_t p_count, const VkAttachmentReference2 *p_refs); + protected: virtual const char *_get_platform_surface_extension() const = 0; @@ -266,9 +271,12 @@ protected: Error _get_preferred_validation_layers(uint32_t *count, const char *const **names); + virtual VkExtent2D _compute_swapchain_extent(const VkSurfaceCapabilitiesKHR &p_surf_capabilities, int *p_window_width, int *p_window_height) const; + public: // Extension calls. - VkResult vkCreateRenderPass2KHR(VkDevice device, const VkRenderPassCreateInfo2 *pCreateInfo, const VkAllocationCallbacks *pAllocator, VkRenderPass *pRenderPass); + bool supports_renderpass2() const { return has_renderpass2_ext; } + VkResult vkCreateRenderPass2KHR(VkDevice p_device, const VkRenderPassCreateInfo2 *p_create_info, const VkAllocationCallbacks *p_allocator, VkRenderPass *p_render_pass); uint32_t get_vulkan_major() const { return vulkan_major; }; uint32_t get_vulkan_minor() const { return vulkan_minor; }; |