diff options
Diffstat (limited to 'drivers')
| -rw-r--r-- | drivers/dummy/rasterizer_dummy.h | 1 | ||||
| -rw-r--r-- | drivers/vulkan/rendering_device_vulkan.cpp | 296 | ||||
| -rw-r--r-- | drivers/vulkan/rendering_device_vulkan.h | 17 | ||||
| -rw-r--r-- | drivers/vulkan/vulkan_context.cpp | 62 | ||||
| -rw-r--r-- | drivers/vulkan/vulkan_context.h | 9 |
5 files changed, 342 insertions, 43 deletions
diff --git a/drivers/dummy/rasterizer_dummy.h b/drivers/dummy/rasterizer_dummy.h index 72ab18d115..e69f36e16f 100644 --- a/drivers/dummy/rasterizer_dummy.h +++ b/drivers/dummy/rasterizer_dummy.h @@ -566,6 +566,7 @@ public: AABB mesh_get_custom_aabb(RID p_mesh) const override { return AABB(); } AABB mesh_get_aabb(RID p_mesh, RID p_skeleton = RID()) override { return AABB(); } + void mesh_set_shadow_mesh(RID p_mesh, RID p_shadow_mesh) override {} void mesh_clear(RID p_mesh) override {} /* MULTIMESH API */ diff --git a/drivers/vulkan/rendering_device_vulkan.cpp b/drivers/vulkan/rendering_device_vulkan.cpp index ef331ec4b6..ba80ca3514 100644 --- a/drivers/vulkan/rendering_device_vulkan.cpp +++ b/drivers/vulkan/rendering_device_vulkan.cpp @@ -1627,6 +1627,9 @@ void RenderingDeviceVulkan::_memory_barrier(VkPipelineStageFlags p_src_stage_mas mem_barrier.srcAccessMask = p_src_access; mem_barrier.dstAccessMask = p_dst_sccess; + if (p_src_stage_mask == 0 || p_dst_stage_mask == 0) { + return; //no barrier, since this is invalid + } vkCmdPipelineBarrier(p_sync_with_draw ? frames[frame].draw_command_buffer : frames[frame].setup_command_buffer, p_src_stage_mask, p_dst_stage_mask, 0, 1, &mem_barrier, 0, nullptr, 0, nullptr); } @@ -2477,6 +2480,10 @@ Error RenderingDeviceVulkan::texture_update(RID p_texture, uint32_t p_layer, con access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } + if (barrier_flags == 0) { + barrier_flags = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + VkImageMemoryBarrier image_memory_barrier; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; image_memory_barrier.pNext = nullptr; @@ -2496,6 +2503,13 @@ Error RenderingDeviceVulkan::texture_update(RID p_texture, uint32_t p_layer, con vkCmdPipelineBarrier(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT, barrier_flags, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier); } + if (texture->used_in_frame != frames_drawn) { + texture->used_in_raster = false; + texture->used_in_compute = false; + texture->used_in_frame = frames_drawn; + } + texture->used_in_transfer = true; + return OK; } @@ -2844,6 +2858,10 @@ Error RenderingDeviceVulkan::texture_copy(RID p_from_texture, RID p_to_texture, access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } + if (barrier_flags == 0) { + barrier_flags = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + { //restore src VkImageMemoryBarrier image_memory_barrier; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; @@ -3011,6 +3029,10 @@ Error RenderingDeviceVulkan::texture_resolve_multisample(RID p_from_texture, RID access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } + if (barrier_flags == 0) { + barrier_flags = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + { //restore src VkImageMemoryBarrier image_memory_barrier; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; @@ -3143,6 +3165,10 @@ Error RenderingDeviceVulkan::texture_clear(RID p_texture, const Color &p_color, access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT; } + if (barrier_flags == 0) { + barrier_flags = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + VkImageMemoryBarrier image_memory_barrier; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; image_memory_barrier.pNext = nullptr; @@ -3163,6 +3189,13 @@ Error RenderingDeviceVulkan::texture_clear(RID p_texture, const Color &p_color, vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_TRANSFER_BIT, barrier_flags, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier); } + if (src_tex->used_in_frame != frames_drawn) { + src_tex->used_in_raster = false; + src_tex->used_in_compute = false; + src_tex->used_in_frame = frames_drawn; + } + src_tex->used_in_transfer = true; + return OK; } @@ -3289,6 +3322,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF dependency_from_external.srcStageMask |= reading_stages; } } break; + case INITIAL_ACTION_CLEAR_REGION_CONTINUE: case INITIAL_ACTION_CONTINUE: { if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) { description.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD; @@ -3296,7 +3330,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; } else if (p_format[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; //don't care what is there + description.initialLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL; description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_LOAD; } else { description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE; @@ -3425,8 +3459,13 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF render_pass_create_info.pAttachments = attachments.ptr(); render_pass_create_info.subpassCount = 1; render_pass_create_info.pSubpasses = &subpass; - render_pass_create_info.dependencyCount = 2; - render_pass_create_info.pDependencies = dependencies; + // 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; VkRenderPass render_pass; VkResult res = vkCreateRenderPass(device, &render_pass_create_info, nullptr, &render_pass); @@ -4108,6 +4147,8 @@ RID RenderingDeviceVulkan::shader_create(const Vector<ShaderStageData> &p_stages bool is_compute = false; + uint32_t compute_local_size[3] = { 0, 0, 0 }; + for (int i = 0; i < p_stages.size(); i++) { if (p_stages[i].shader_stage == SHADER_STAGE_COMPUTE) { is_compute = true; @@ -4124,6 +4165,11 @@ RID RenderingDeviceVulkan::shader_create(const Vector<ShaderStageData> &p_stages ERR_FAIL_COND_V_MSG(result != SPV_REFLECT_RESULT_SUCCESS, RID(), "Reflection of SPIR-V shader stage '" + String(shader_stage_names[p_stages[i].shader_stage]) + "' failed parsing shader."); + if (is_compute) { + compute_local_size[0] = module.entry_points->local_size.x; + compute_local_size[1] = module.entry_points->local_size.y; + compute_local_size[2] = module.entry_points->local_size.z; + } uint32_t binding_count = 0; result = spvReflectEnumerateDescriptorBindings(&module, &binding_count, nullptr); ERR_FAIL_COND_V_MSG(result != SPV_REFLECT_RESULT_SUCCESS, RID(), @@ -4328,6 +4374,7 @@ RID RenderingDeviceVulkan::shader_create(const Vector<ShaderStageData> &p_stages } } } + uint32_t pc_count = 0; result = spvReflectEnumeratePushConstantBlocks(&module, &pc_count, nullptr); ERR_FAIL_COND_V_MSG(result != SPV_REFLECT_RESULT_SUCCESS, RID(), @@ -4376,6 +4423,9 @@ RID RenderingDeviceVulkan::shader_create(const Vector<ShaderStageData> &p_stages shader.fragment_outputs = fragment_outputs; shader.push_constant = push_constant; shader.is_compute = is_compute; + shader.compute_local_size[0] = compute_local_size[0]; + shader.compute_local_size[1] = compute_local_size[1]; + shader.compute_local_size[2] = compute_local_size[2]; String error_text; @@ -5216,7 +5266,14 @@ Error RenderingDeviceVulkan::buffer_update(RID p_buffer, uint32_t p_offset, uint #ifdef FORCE_FULL_BARRIER _full_barrier(true); #else - _buffer_memory_barrier(buffer->buffer, p_offset, p_size, VK_PIPELINE_STAGE_TRANSFER_BIT, dst_stage_mask, VK_ACCESS_TRANSFER_WRITE_BIT, dst_access, true); + if (dst_stage_mask == 0) { + dst_stage_mask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + + if (p_post_barrier != RD::BARRIER_MASK_NO_BARRIER) { + _buffer_memory_barrier(buffer->buffer, p_offset, p_size, VK_PIPELINE_STAGE_TRANSFER_BIT, dst_stage_mask, VK_ACCESS_TRANSFER_WRITE_BIT, dst_access, dst_stage_mask); + } + #endif return err; } @@ -5255,7 +5312,12 @@ Error RenderingDeviceVulkan::buffer_clear(RID p_buffer, uint32_t p_offset, uint3 #ifdef FORCE_FULL_BARRIER _full_barrier(true); #else - _buffer_memory_barrier(buffer->buffer, p_offset, p_size, VK_PIPELINE_STAGE_TRANSFER_BIT, dst_stage_mask, VK_ACCESS_TRANSFER_WRITE_BIT, dst_access, p_post_barrier); + if (dst_stage_mask == 0) { + dst_stage_mask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + + _buffer_memory_barrier(buffer->buffer, p_offset, p_size, VK_PIPELINE_STAGE_TRANSFER_BIT, dst_stage_mask, VK_ACCESS_TRANSFER_WRITE_BIT, dst_access, dst_stage_mask); + #endif return OK; } @@ -5710,6 +5772,9 @@ RID RenderingDeviceVulkan::compute_pipeline_create(RID p_shader) { pipeline.pipeline_layout = shader->pipeline_layout; pipeline.shader = p_shader; pipeline.push_constant_size = shader->push_constant.push_constant_size; + pipeline.local_group_size[0] = shader->compute_local_size[0]; + pipeline.local_group_size[1] = shader->compute_local_size[1]; + pipeline.local_group_size[2] = shader->compute_local_size[2]; //create ID to associate with this pipeline RID id = compute_pipeline_owner.make_rid(pipeline); @@ -6019,7 +6084,7 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_framebu _THREAD_SAFE_METHOD_ ERR_FAIL_COND_V_MSG(draw_list != nullptr, INVALID_ID, "Only one draw list can be active at the same time."); - ERR_FAIL_COND_V_MSG(compute_list != nullptr, INVALID_ID, "Only one draw/compute list can be active at the same time."); + ERR_FAIL_COND_V_MSG(compute_list != nullptr && !compute_list->state.allow_draw_overlap, INVALID_ID, "Only one draw/compute list can be active at the same time."); Framebuffer *framebuffer = framebuffer_owner.getornull(p_framebuffer); ERR_FAIL_COND_V(!framebuffer, INVALID_ID); @@ -6040,7 +6105,14 @@ RenderingDevice::DrawListID RenderingDeviceVulkan::draw_list_begin(RID p_framebu viewport_offset = regioni.position; viewport_size = regioni.size; - + if (p_initial_color_action == INITIAL_ACTION_CLEAR_REGION_CONTINUE) { + needs_clear_color = true; + p_initial_color_action = INITIAL_ACTION_CONTINUE; + } + if (p_initial_depth_action == INITIAL_ACTION_CLEAR_REGION_CONTINUE) { + needs_clear_depth = true; + p_initial_depth_action = INITIAL_ACTION_CONTINUE; + } if (p_initial_color_action == INITIAL_ACTION_CLEAR_REGION) { needs_clear_color = true; p_initial_color_action = INITIAL_ACTION_KEEP; @@ -6388,6 +6460,19 @@ void RenderingDeviceVulkan::draw_list_bind_uniform_set(DrawListID p_list, RID p_ dl->state.sets[p_index].uniform_set_format = uniform_set->format; dl->state.sets[p_index].uniform_set = p_uniform_set; + uint32_t mst_count = uniform_set->mutable_storage_textures.size(); + if (mst_count) { + Texture **mst_textures = const_cast<UniformSet *>(uniform_set)->mutable_storage_textures.ptrw(); + for (uint32_t i = 0; i < mst_count; i++) { + if (mst_textures[i]->used_in_frame != frames_drawn) { + mst_textures[i]->used_in_frame = frames_drawn; + mst_textures[i]->used_in_transfer = false; + mst_textures[i]->used_in_compute = false; + } + mst_textures[i]->used_in_raster = true; + } + } + #ifdef DEBUG_ENABLED { //validate that textures bound are not attached as framebuffer bindings uint32_t attachable_count = uniform_set->attachable_textures.size(); @@ -6673,23 +6758,43 @@ void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } if (p_post_barrier & 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; + 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) { barrier_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_TRANSFER_READ_BIT; } + if (barrier_flags == 0) { + barrier_flags = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + draw_list_bound_textures.clear(); - for (int i = 0; i < draw_list_storage_textures.size(); i++) { + VkImageMemoryBarrier *image_barriers = nullptr; + + uint32_t image_barrier_count = draw_list_storage_textures.size(); + + if (image_barrier_count) { + image_barriers = (VkImageMemoryBarrier *)alloca(sizeof(VkImageMemoryBarrier) * draw_list_storage_textures.size()); + } + + uint32_t src_stage = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT; + uint32_t src_access = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT; + + if (image_barrier_count) { + src_stage |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT; + src_access |= VK_ACCESS_SHADER_WRITE_BIT; + } + + for (uint32_t i = 0; i < image_barrier_count; i++) { Texture *texture = texture_owner.getornull(draw_list_storage_textures[i]); - VkImageMemoryBarrier image_memory_barrier; + VkImageMemoryBarrier &image_memory_barrier = image_barriers[i]; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; image_memory_barrier.pNext = nullptr; - image_memory_barrier.srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; + image_memory_barrier.srcAccessMask = src_access; image_memory_barrier.dstAccessMask = access_flags; image_memory_barrier.oldLayout = texture->layout; image_memory_barrier.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; @@ -6703,8 +6808,6 @@ void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { image_memory_barrier.subresourceRange.baseArrayLayer = texture->base_layer; image_memory_barrier.subresourceRange.layerCount = texture->layers; - vkCmdPipelineBarrier(frames[frame].draw_command_buffer, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, barrier_flags, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier); - texture->layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; } @@ -6717,7 +6820,17 @@ void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { #ifdef FORCE_FULL_BARRIER _full_barrier(true); #else - _memory_barrier(VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT, barrier_flags, VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT, access_flags, true); + + VkMemoryBarrier mem_barrier; + mem_barrier.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER; + mem_barrier.pNext = nullptr; + mem_barrier.srcAccessMask = src_access; + mem_barrier.dstAccessMask = access_flags; + + if (image_barrier_count > 0 || p_post_barrier != BARRIER_MASK_NO_BARRIER) { + vkCmdPipelineBarrier(frames[frame].draw_command_buffer, src_stage, barrier_flags, 0, 1, &mem_barrier, 0, nullptr, image_barrier_count, image_barriers); + } + #endif } @@ -6725,12 +6838,13 @@ void RenderingDeviceVulkan::draw_list_end(uint32_t p_post_barrier) { /**** COMPUTE LISTS ****/ /***********************/ -RenderingDevice::ComputeListID RenderingDeviceVulkan::compute_list_begin() { - ERR_FAIL_COND_V_MSG(draw_list != nullptr, INVALID_ID, "Only one draw list can be active at the same time."); +RenderingDevice::ComputeListID RenderingDeviceVulkan::compute_list_begin(bool p_allow_draw_overlap) { + ERR_FAIL_COND_V_MSG(!p_allow_draw_overlap && draw_list != nullptr, INVALID_ID, "Only one draw list can be active at the same time."); ERR_FAIL_COND_V_MSG(compute_list != nullptr, INVALID_ID, "Only one draw/compute list can be active at the same time."); compute_list = memnew(ComputeList); compute_list->command_buffer = frames[frame].draw_command_buffer; + compute_list->state.allow_draw_overlap = p_allow_draw_overlap; return ID_TYPE_COMPUTE_LIST; } @@ -6787,6 +6901,9 @@ void RenderingDeviceVulkan::compute_list_bind_compute_pipeline(ComputeListID p_l } cl->state.pipeline_shader = pipeline->shader; + cl->state.local_group_size[0] = pipeline->local_group_size[0]; + cl->state.local_group_size[1] = pipeline->local_group_size[1]; + cl->state.local_group_size[2] = pipeline->local_group_size[2]; } #ifdef DEBUG_ENABLED @@ -6824,11 +6941,24 @@ void RenderingDeviceVulkan::compute_list_bind_uniform_set(ComputeListID p_list, cl->state.sets[p_index].uniform_set = p_uniform_set; uint32_t textures_to_sampled_count = uniform_set->mutable_sampled_textures.size(); + uint32_t textures_to_storage_count = uniform_set->mutable_storage_textures.size(); + Texture **textures_to_sampled = uniform_set->mutable_sampled_textures.ptrw(); + VkImageMemoryBarrier *texture_barriers = nullptr; + + if (textures_to_sampled_count + textures_to_storage_count) { + texture_barriers = (VkImageMemoryBarrier *)alloca(sizeof(VkImageMemoryBarrier) * (textures_to_sampled_count + textures_to_storage_count)); + } + uint32_t texture_barrier_count = 0; + + uint32_t src_stage_flags = 0; + for (uint32_t i = 0; i < textures_to_sampled_count; i++) { if (textures_to_sampled[i]->layout != VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL) { - VkImageMemoryBarrier image_memory_barrier; + src_stage_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; + + VkImageMemoryBarrier &image_memory_barrier = texture_barriers[texture_barrier_count++]; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; image_memory_barrier.pNext = nullptr; image_memory_barrier.srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; @@ -6845,23 +6975,55 @@ void RenderingDeviceVulkan::compute_list_bind_uniform_set(ComputeListID p_list, image_memory_barrier.subresourceRange.baseArrayLayer = textures_to_sampled[i]->base_layer; image_memory_barrier.subresourceRange.layerCount = textures_to_sampled[i]->layers; - vkCmdPipelineBarrier(cl->command_buffer, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier); - textures_to_sampled[i]->layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; cl->state.textures_to_sampled_layout.erase(textures_to_sampled[i]); } + + if (textures_to_sampled[i]->used_in_frame != frames_drawn) { + textures_to_sampled[i]->used_in_frame = frames_drawn; + textures_to_sampled[i]->used_in_transfer = false; + textures_to_sampled[i]->used_in_raster = false; + } + textures_to_sampled[i]->used_in_compute = true; } - uint32_t textures_to_storage_count = uniform_set->mutable_storage_textures.size(); Texture **textures_to_storage = uniform_set->mutable_storage_textures.ptrw(); for (uint32_t i = 0; i < textures_to_storage_count; i++) { if (textures_to_storage[i]->layout != VK_IMAGE_LAYOUT_GENERAL) { - VkImageMemoryBarrier image_memory_barrier; + uint32_t src_access_flags = 0; + + if (textures_to_storage[i]->used_in_frame == frames_drawn) { + if (textures_to_storage[i]->used_in_compute) { + src_stage_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; + src_access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; + } + if (textures_to_storage[i]->used_in_raster) { + src_stage_flags |= VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT; + src_access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; + } + if (textures_to_storage[i]->used_in_transfer) { + src_stage_flags |= VK_PIPELINE_STAGE_TRANSFER_BIT; + src_access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_TRANSFER_READ_BIT; + } + + textures_to_storage[i]->used_in_compute = false; + textures_to_storage[i]->used_in_raster = false; + textures_to_storage[i]->used_in_compute = false; + + } else { + src_access_flags = 0; + textures_to_storage[i]->used_in_compute = false; + textures_to_storage[i]->used_in_raster = false; + textures_to_storage[i]->used_in_compute = false; + textures_to_storage[i]->used_in_frame = frames_drawn; + } + + VkImageMemoryBarrier &image_memory_barrier = texture_barriers[texture_barrier_count++]; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; image_memory_barrier.pNext = nullptr; - image_memory_barrier.srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; + image_memory_barrier.srcAccessMask = src_access_flags; image_memory_barrier.dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; image_memory_barrier.oldLayout = textures_to_storage[i]->layout; image_memory_barrier.newLayout = VK_IMAGE_LAYOUT_GENERAL; @@ -6875,14 +7037,20 @@ void RenderingDeviceVulkan::compute_list_bind_uniform_set(ComputeListID p_list, image_memory_barrier.subresourceRange.baseArrayLayer = textures_to_storage[i]->base_layer; image_memory_barrier.subresourceRange.layerCount = textures_to_storage[i]->layers; - vkCmdPipelineBarrier(cl->command_buffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier); - textures_to_storage[i]->layout = VK_IMAGE_LAYOUT_GENERAL; cl->state.textures_to_sampled_layout.insert(textures_to_storage[i]); //needs to go back to sampled layout afterwards } } + if (texture_barrier_count) { + if (src_stage_flags == 0) { + src_stage_flags = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT; + } + + vkCmdPipelineBarrier(cl->command_buffer, src_stage_flags, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, texture_barrier_count, texture_barriers); + } + #if 0 { //validate that textures bound are not attached as framebuffer bindings uint32_t attachable_count = uniform_set->attachable_textures.size(); @@ -6976,6 +7144,27 @@ void RenderingDeviceVulkan::compute_list_dispatch(ComputeListID p_list, uint32_t vkCmdDispatch(cl->command_buffer, p_x_groups, p_y_groups, p_z_groups); } +void RenderingDeviceVulkan::compute_list_dispatch_threads(ComputeListID p_list, uint32_t p_x_threads, uint32_t p_y_threads, uint32_t p_z_threads) { + ERR_FAIL_COND(p_list != ID_TYPE_COMPUTE_LIST); + ERR_FAIL_COND(!compute_list); + + ComputeList *cl = compute_list; + +#ifdef DEBUG_ENABLED + + ERR_FAIL_COND_MSG(!cl->validation.pipeline_active, "No compute pipeline was set before attempting to draw."); + + if (cl->validation.pipeline_push_constant_size > 0) { + //using push constants, check that they were supplied + ERR_FAIL_COND_MSG(!cl->validation.pipeline_push_constant_supplied, + "The shader in this pipeline requires a push constant to be set before drawing, but it's not present."); + } + +#endif + + compute_list_dispatch(p_list, (p_x_threads - 1) / cl->state.local_group_size[0] + 1, (p_y_threads - 1) / cl->state.local_group_size[1] + 1, (p_z_threads - 1) / cl->state.local_group_size[2] + 1); +} + void RenderingDeviceVulkan::compute_list_dispatch_indirect(ComputeListID p_list, RID p_buffer, uint32_t p_offset) { ERR_FAIL_COND(p_list != ID_TYPE_COMPUTE_LIST); ERR_FAIL_COND(!compute_list); @@ -7047,7 +7236,7 @@ void RenderingDeviceVulkan::compute_list_end(uint32_t p_post_barrier) { uint32_t access_flags = 0; if (p_post_barrier & BARRIER_MASK_COMPUTE) { barrier_flags |= VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT; - access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_INDIRECT_COMMAND_READ_BIT; + access_flags |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; } if (p_post_barrier & 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; @@ -7058,8 +7247,22 @@ void RenderingDeviceVulkan::compute_list_end(uint32_t p_post_barrier) { access_flags |= VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_TRANSFER_READ_BIT; } + if (barrier_flags == 0) { + barrier_flags = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT; + } + + VkImageMemoryBarrier *image_barriers = nullptr; + + uint32_t image_barrier_count = compute_list->state.textures_to_sampled_layout.size(); + + if (image_barrier_count) { + image_barriers = (VkImageMemoryBarrier *)alloca(sizeof(VkImageMemoryBarrier) * image_barrier_count); + } + + uint32_t barrier_idx = 0; + for (Set<Texture *>::Element *E = compute_list->state.textures_to_sampled_layout.front(); E; E = E->next()) { - VkImageMemoryBarrier image_memory_barrier; + VkImageMemoryBarrier &image_memory_barrier = image_barriers[barrier_idx++]; image_memory_barrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER; image_memory_barrier.pNext = nullptr; image_memory_barrier.srcAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT; @@ -7076,19 +7279,33 @@ void RenderingDeviceVulkan::compute_list_end(uint32_t p_post_barrier) { image_memory_barrier.subresourceRange.baseArrayLayer = E->get()->base_layer; image_memory_barrier.subresourceRange.layerCount = E->get()->layers; - // TODO: Look at the usages in the compute list and determine tighter dst stage and access masks based on some "final" usage equivalent - vkCmdPipelineBarrier(compute_list->command_buffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, barrier_flags, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier); - E->get()->layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + + if (E->get()->used_in_frame != frames_drawn) { + E->get()->used_in_transfer = false; + E->get()->used_in_raster = false; + E->get()->used_in_compute = false; + E->get()->used_in_frame = frames_drawn; + } } - memdelete(compute_list); - compute_list = nullptr; #ifdef FORCE_FULL_BARRIER _full_barrier(true); #else - _memory_barrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, barrier_flags, VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_INDEX_READ_BIT | VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_INDIRECT_COMMAND_READ_BIT, true); + VkMemoryBarrier mem_barrier; + mem_barrier.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER; + mem_barrier.pNext = nullptr; + mem_barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT; + mem_barrier.dstAccessMask = access_flags; + + if (image_barrier_count > 0 || p_post_barrier != BARRIER_MASK_NO_BARRIER) { + vkCmdPipelineBarrier(compute_list->command_buffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, barrier_flags, 0, 1, &mem_barrier, 0, nullptr, image_barrier_count, image_barriers); + } + #endif + + memdelete(compute_list); + compute_list = nullptr; } void RenderingDeviceVulkan::barrier(uint32_t p_from, uint32_t p_to) { @@ -7111,7 +7328,7 @@ void RenderingDeviceVulkan::barrier(uint32_t p_from, uint32_t p_to) { 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 | VK_ACCESS_INDIRECT_COMMAND_READ_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; @@ -7325,6 +7542,16 @@ void RenderingDeviceVulkan::draw_command_end_label() { context->command_end_label(frames[frame].draw_command_buffer); } +String RenderingDeviceVulkan::get_device_vendor_name() const { + return context->get_device_vendor_name(); +} +String RenderingDeviceVulkan::get_device_name() const { + return context->get_device_name(); +} +String RenderingDeviceVulkan::get_device_pipeline_cache_uuid() const { + return context->get_device_pipeline_cache_uuid(); +} + void RenderingDeviceVulkan::_finalize_command_bufers() { if (draw_list) { ERR_PRINT("Found open draw list at the end of the frame, this should never happen (further drawing will likely not work)."); @@ -7377,6 +7604,7 @@ void RenderingDeviceVulkan::_begin_frame() { if (frames[frame].timestamp_count) { vkGetQueryPoolResults(device, frames[frame].timestamp_pool, 0, frames[frame].timestamp_count, sizeof(uint64_t) * max_timestamp_query_elements, frames[frame].timestamp_result_values, sizeof(uint64_t), VK_QUERY_RESULT_64_BIT); + vkCmdResetQueryPool(frames[frame].setup_command_buffer, frames[frame].timestamp_pool, 0, frames[frame].timestamp_count); SWAP(frames[frame].timestamp_names, frames[frame].timestamp_result_names); SWAP(frames[frame].timestamp_cpu_values, frames[frame].timestamp_cpu_result_values); } diff --git a/drivers/vulkan/rendering_device_vulkan.h b/drivers/vulkan/rendering_device_vulkan.h index 4bea17e4a1..a2527d5c33 100644 --- a/drivers/vulkan/rendering_device_vulkan.h +++ b/drivers/vulkan/rendering_device_vulkan.h @@ -141,6 +141,11 @@ class RenderingDeviceVulkan : public RenderingDevice { VkImageLayout layout; + uint64_t used_in_frame = 0; + bool used_in_transfer = false; + bool used_in_raster = false; + bool used_in_compute = false; + uint32_t read_aspect_mask = 0; uint32_t barrier_aspect_mask = 0; bool bound = false; //bound to framebffer @@ -528,6 +533,8 @@ class RenderingDeviceVulkan : public RenderingDevice { PushConstant push_constant; + uint32_t compute_local_size[3] = { 0, 0, 0 }; + bool is_compute = false; int max_output = 0; Vector<Set> sets; @@ -686,6 +693,7 @@ class RenderingDeviceVulkan : public RenderingDevice { VkPipeline pipeline = VK_NULL_HANDLE; uint32_t push_constant_size = 0; uint32_t push_constant_stages = 0; + uint32_t local_group_size[3] = { 0, 0, 0 }; }; RID_Owner<ComputePipeline, true> compute_pipeline_owner; @@ -808,8 +816,10 @@ class RenderingDeviceVulkan : public RenderingDevice { uint32_t set_count = 0; RID pipeline; RID pipeline_shader; + uint32_t local_group_size[3] = { 0, 0, 0 }; VkPipelineLayout pipeline_layout = VK_NULL_HANDLE; uint32_t pipeline_push_constant_stages = 0; + bool allow_draw_overlap; } state; #ifdef DEBUG_ENABLED @@ -1028,13 +1038,14 @@ public: /**** COMPUTE LISTS ****/ /***********************/ - virtual ComputeListID compute_list_begin(); + virtual ComputeListID compute_list_begin(bool p_allow_draw_overlap = false); virtual void compute_list_bind_compute_pipeline(ComputeListID p_list, RID p_compute_pipeline); virtual void compute_list_bind_uniform_set(ComputeListID p_list, RID p_uniform_set, uint32_t p_index); virtual void compute_list_set_push_constant(ComputeListID p_list, const void *p_data, uint32_t p_data_size); virtual void compute_list_add_barrier(ComputeListID p_list); 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); @@ -1085,6 +1096,10 @@ public: virtual void draw_command_insert_label(String p_label_name, const Color p_color = Color(1, 1, 1, 1)); virtual void draw_command_end_label(); + virtual String get_device_vendor_name() const; + virtual String get_device_name() const; + virtual String get_device_pipeline_cache_uuid() const; + RenderingDeviceVulkan(); ~RenderingDeviceVulkan(); }; diff --git a/drivers/vulkan/vulkan_context.cpp b/drivers/vulkan/vulkan_context.cpp index 98966477a5..c564cee757 100644 --- a/drivers/vulkan/vulkan_context.cpp +++ b/drivers/vulkan/vulkan_context.cpp @@ -380,7 +380,8 @@ Error VulkanContext::_create_physical_device() { ERR_FAIL_V(ERR_CANT_CREATE); } /* for now, just grab the first physical device */ - gpu = physical_devices[0]; + uint32_t device_index = 0; + gpu = physical_devices[device_index]; free(physical_devices); /* Look for device extensions */ @@ -389,6 +390,40 @@ Error VulkanContext::_create_physical_device() { enabled_extension_count = 0; memset(extension_names, 0, sizeof(extension_names)); + /* Get identifier properties */ + vkGetPhysicalDeviceProperties(gpu, &gpu_props); + + static const struct { + uint32_t id; + const char *name; + } vendor_names[] = { + { 0x1002, "AMD" }, + { 0x1010, "ImgTec" }, + { 0x10DE, "NVIDIA" }, + { 0x13B5, "ARM" }, + { 0x5143, "Qualcomm" }, + { 0x8086, "INTEL" }, + { 0, nullptr }, + }; + device_name = gpu_props.deviceName; + pipeline_cache_id = String::hex_encode_buffer(gpu_props.pipelineCacheUUID, VK_UUID_SIZE); + pipeline_cache_id += "-driver-" + itos(gpu_props.driverVersion); + { + device_vendor = "Unknown"; + uint32_t vendor_idx = 0; + while (vendor_names[vendor_idx].name != nullptr) { + if (gpu_props.vendorID == vendor_names[vendor_idx].id) { + device_vendor = vendor_names[vendor_idx].name; + break; + } + vendor_idx++; + } + } +#ifdef DEBUG_ENABLED + print_line("Using Vulkan Device #" + itos(device_index) + ": " + device_vendor + " - " + device_name); +#endif + device_api_version = gpu_props.apiVersion; + err = vkEnumerateDeviceExtensionProperties(gpu, nullptr, &device_extension_count, nullptr); ERR_FAIL_COND_V(err, ERR_CANT_CREATE); @@ -498,7 +533,6 @@ Error VulkanContext::_create_physical_device() { break; } } - vkGetPhysicalDeviceProperties(gpu, &gpu_props); /* Call with NULL data to get count */ vkGetPhysicalDeviceQueueFamilyProperties(gpu, &queue_family_count, nullptr); @@ -565,6 +599,7 @@ Error VulkanContext::_create_device() { } err = vkCreateDevice(gpu, &sdevice, nullptr, &device); ERR_FAIL_COND_V(err, ERR_CANT_CREATE); + return OK; } @@ -1590,11 +1625,12 @@ void VulkanContext::command_begin_label(VkCommandBuffer p_command_buffer, String if (!enabled_debug_utils) { return; } + + CharString cs = p_label_name.utf8().get_data(); VkDebugUtilsLabelEXT label; label.sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT; label.pNext = nullptr; - CharString label_name = p_label_name.utf8(); - label.pLabelName = label_name.get_data(); + label.pLabelName = cs.get_data(); label.color[0] = p_color[0]; label.color[1] = p_color[1]; label.color[2] = p_color[2]; @@ -1606,11 +1642,11 @@ void VulkanContext::command_insert_label(VkCommandBuffer p_command_buffer, Strin if (!enabled_debug_utils) { return; } + CharString cs = p_label_name.utf8().get_data(); VkDebugUtilsLabelEXT label; label.sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT; label.pNext = nullptr; - CharString label_name = p_label_name.utf8(); - label.pLabelName = label_name.get_data(); + label.pLabelName = cs.get_data(); label.color[0] = p_color[0]; label.color[1] = p_color[1]; label.color[2] = p_color[2]; @@ -1629,16 +1665,26 @@ void VulkanContext::set_object_name(VkObjectType p_object_type, uint64_t p_objec if (!enabled_debug_utils) { return; } + CharString obj_data = p_object_name.utf8(); VkDebugUtilsObjectNameInfoEXT name_info; name_info.sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_OBJECT_NAME_INFO_EXT; name_info.pNext = nullptr; name_info.objectType = p_object_type; name_info.objectHandle = p_object_handle; - CharString object_name = p_object_name.utf8(); - name_info.pObjectName = object_name.get_data(); + name_info.pObjectName = obj_data.get_data(); SetDebugUtilsObjectNameEXT(device, &name_info); } +String VulkanContext::get_device_vendor_name() const { + return device_vendor; +} +String VulkanContext::get_device_name() const { + return device_name; +} +String VulkanContext::get_device_pipeline_cache_uuid() const { + return pipeline_cache_id; +} + VulkanContext::VulkanContext() { use_validation_layers = Engine::get_singleton()->is_validation_layers_enabled(); diff --git a/drivers/vulkan/vulkan_context.h b/drivers/vulkan/vulkan_context.h index 5cb762aca8..dc6b0410bc 100644 --- a/drivers/vulkan/vulkan_context.h +++ b/drivers/vulkan/vulkan_context.h @@ -57,6 +57,11 @@ class VulkanContext { bool device_initialized = false; bool inst_initialized = false; + String device_vendor; + String device_name; + String pipeline_cache_id; + uint32_t device_api_version = 0; + bool buffers_prepared = false; // Present queue. @@ -215,6 +220,10 @@ public: void command_end_label(VkCommandBuffer p_command_buffer); void set_object_name(VkObjectType p_object_type, uint64_t p_object_handle, String p_object_name); + String get_device_vendor_name() const; + String get_device_name() const; + String get_device_pipeline_cache_uuid() const; + VulkanContext(); virtual ~VulkanContext(); }; |