Additional synchronization fixes w/o FORCE_FULL_BARRIER

Additional synchronization fixes from hazards arising from disabling
FORCE_FULL_BARRIER.

1 files changed, 76 insertions, 18 deletions

diff --git a/drivers/vulkan/rendering_device_vulkan.cpp b/drivers/vulkan/rendering_device_vulkan.cpp
index 2947376032..54ec2479f0 100644
--- a/drivers/vulkan/rendering_device_vulkan.cpp
+++ b/drivers/vulkan/rendering_device_vulkan.cpp
@@ -38,7 +38,7 @@
 
 #include "thirdparty/spirv-reflect/spirv_reflect.h"
 
-#define FORCE_FULL_BARRIER
+//#define FORCE_FULL_BARRIER
 
 // 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 &stage_mask, VkAccessFlags &access_mask) {
@@ -67,6 +67,30 @@ RenderingDeviceVulkan::Buffer *RenderingDeviceVulkan::_get_buffer_from_owner(RID
 	return buffer;
 }
 
+static void update_external_dependency_for_store(VkSubpassDependency &dependency, bool is_sampled, bool is_storage, bool is_depth) {
+	// Transitioning from write to read, protect the shaders that may use this next
+	// Allow for copies/image layout transitions
+	dependency.dstStageMask |= VK_PIPELINE_STAGE_TRANSFER_BIT;
+	dependency.dstAccessMask |= VK_ACCESS_TRANSFER_READ_BIT;
+
+	if (is_sampled) {
+		dependency.dstStageMask |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
+		dependency.dstAccessMask |= VK_ACCESS_SHADER_READ_BIT;
+	} else if (is_storage) {
+		dependency.dstStageMask |= VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
+		dependency.dstAccessMask |= VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
+	} else {
+		dependency.dstStageMask |= VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
+		dependency.dstAccessMask |= VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+	}
+
+	if (is_depth) {
+		// Depth resources have addtional stages that may be interested in them
+		dependency.dstStageMask |= VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
+		dependency.dstAccessMask |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
+	}
+}
+
 void RenderingDeviceVulkan::_add_dependency(RID p_id, RID p_depends_on) {
 	if (!dependency_map.has(p_depends_on)) {
 		dependency_map[p_depends_on] = Set<RID>();
@@ -1959,7 +1983,7 @@ RID RenderingDeviceVulkan::texture_create(const TextureFormat &p_format, const T
 		image_memory_barrier.subresourceRange.baseArrayLayer = 0;
 		image_memory_barrier.subresourceRange.layerCount = image_create_info.arrayLayers;
 
-		vkCmdPipelineBarrier(frames[frame].setup_command_buffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+		vkCmdPipelineBarrier(frames[frame].setup_command_buffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 	}
 
 	RID id = texture_owner.make_rid(texture);
@@ -2216,7 +2240,7 @@ Error RenderingDeviceVulkan::texture_update(RID p_texture, uint32_t p_layer, con
 		image_memory_barrier.subresourceRange.baseArrayLayer = p_layer;
 		image_memory_barrier.subresourceRange.layerCount = 1;
 
-		vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+		vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 	}
 
 	uint32_t mipmap_offset = 0;
@@ -2349,7 +2373,7 @@ Error RenderingDeviceVulkan::texture_update(RID p_texture, uint32_t p_layer, con
 		image_memory_barrier.subresourceRange.baseArrayLayer = p_layer;
 		image_memory_barrier.subresourceRange.layerCount = 1;
 
-		vkCmdPipelineBarrier(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+		vkCmdPipelineBarrier(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 	}
 
 	return OK;
@@ -2512,6 +2536,9 @@ Vector<uint8_t> RenderingDeviceVulkan::texture_get_data(RID p_texture, uint32_t
 			image_memory_barrier.pNext = nullptr;
 			image_memory_barrier.srcAccessMask = VK_ACCESS_TRANSFER_READ_BIT;
 			image_memory_barrier.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
+			if (tex->usage_flags & TEXTURE_USAGE_STORAGE_BIT) {
+				image_memory_barrier.dstAccessMask |= VK_ACCESS_SHADER_WRITE_BIT;
+			}
 			image_memory_barrier.oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL;
 			image_memory_barrier.newLayout = tex->layout;
 			image_memory_barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
@@ -2523,7 +2550,7 @@ Vector<uint8_t> RenderingDeviceVulkan::texture_get_data(RID p_texture, uint32_t
 			image_memory_barrier.subresourceRange.baseArrayLayer = p_layer;
 			image_memory_barrier.subresourceRange.layerCount = 1;
 
-			vkCmdPipelineBarrier(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+			vkCmdPipelineBarrier(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 		}
 
 		_flush(true);
@@ -2782,7 +2809,7 @@ Error RenderingDeviceVulkan::texture_resolve_multisample(RID p_from_texture, RID
 			image_memory_barrier.subresourceRange.baseArrayLayer = src_tex->base_layer;
 			image_memory_barrier.subresourceRange.layerCount = 1;
 
-			vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+			vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 		}
 		{ //Dest
 			VkImageMemoryBarrier image_memory_barrier;
@@ -2963,7 +2990,7 @@ Error RenderingDeviceVulkan::texture_clear(RID p_texture, const Color &p_color,
 		image_memory_barrier.subresourceRange.baseArrayLayer = src_tex->base_layer + p_base_layer;
 		image_memory_barrier.subresourceRange.layerCount = p_layers;
 
-		vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+		vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 	}
 
 	return OK;
@@ -3018,6 +3045,19 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF
 	Vector<VkAttachmentReference> depth_stencil_references;
 	Vector<VkAttachmentReference> resolve_references;
 
+	// Set up a dependencies from/to external equivalent to the default (implicit) one, and then amend them
+	const VkPipelineStageFlags default_access_mask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT |
+													 VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+													 VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
+													 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
+													 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT; // From Section 7.1 of Vulkan API Spec v1.1.148
+
+	VkPipelineStageFlags reading_stages = VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT;
+	VkSubpassDependency dependencies[2] = { { VK_SUBPASS_EXTERNAL, 0, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT, 0, default_access_mask, 0 },
+		{ 0, VK_SUBPASS_EXTERNAL, VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, default_access_mask, 0, 0 } };
+	VkSubpassDependency &dependency_from_external = dependencies[0];
+	VkSubpassDependency &dependency_to_external = dependencies[1];
+
 	for (int i = 0; i < p_format.size(); i++) {
 		ERR_FAIL_INDEX_V(p_format[i].format, DATA_FORMAT_MAX, VK_NULL_HANDLE);
 		ERR_FAIL_INDEX_V(p_format[i].samples, TEXTURE_SAMPLES_MAX, VK_NULL_HANDLE);
@@ -3033,11 +3073,16 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF
 		bool is_sampled = p_format[i].usage_flags & TEXTURE_USAGE_SAMPLING_BIT;
 		bool is_storage = p_format[i].usage_flags & TEXTURE_USAGE_STORAGE_BIT;
 
+		// For each UNDEFINED, assume the prior use was a *read*, as we'd be discarding the output of a write
+		// Also, each UNDEFINED will do an immediate layout transition (write), s.t. we must ensure execution syncronization vs.
+		// the read.  If this is a performance issue, one could track the actual last accessor of each resource, adding only that
+		// stage
 		switch (is_depth_stencil ? p_initial_depth_action : p_initial_color_action) {
 			case INITIAL_ACTION_CLEAR: {
 				description.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR;
 				description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_CLEAR;
 				description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there
+				dependency_from_external.srcStageMask |= reading_stages;
 			} break;
 			case INITIAL_ACTION_KEEP: {
 				if (p_format[i].usage_flags & TEXTURE_USAGE_COLOR_ATTACHMENT_BIT) {
@@ -3048,10 +3093,12 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF
 					description.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR;
 					description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there
 					description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_CLEAR;
+					dependency_from_external.srcStageMask |= reading_stages;
 				} else {
 					description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there
+					dependency_from_external.srcStageMask |= reading_stages;
 				}
 			} break;
 			case INITIAL_ACTION_DROP: {
@@ -3063,10 +3110,12 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF
 					description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there
 					description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
+					dependency_from_external.srcStageMask |= reading_stages;
 				} else {
 					description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there
+					dependency_from_external.srcStageMask |= reading_stages;
 				}
 			} break;
 			case INITIAL_ACTION_CONTINUE: {
@@ -3082,6 +3131,7 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF
 					description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there
+					dependency_from_external.srcStageMask |= reading_stages;
 				}
 			} break;
 			default: {
@@ -3095,14 +3145,17 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF
 					description.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
 					description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
 					description.finalLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL);
+					update_external_dependency_for_store(dependency_to_external, is_sampled, is_storage, false);
 				} else if (p_format[i].usage_flags & TEXTURE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
 					description.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
 					description.stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE;
 					description.finalLayout = is_sampled ? VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL : (is_storage ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL);
+					update_external_dependency_for_store(dependency_to_external, is_sampled, is_storage, true);
 				} else {
 					description.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
 					description.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED; //don't care what is there
+					// TODO: What does this mean about the next usage (and thus appropriate dependency masks
 				}
 			} break;
 			case FINAL_ACTION_DISCARD: {
@@ -3155,6 +3208,9 @@ VkRenderPass RenderingDeviceVulkan::_render_pass_create(const Vector<AttachmentF
 		} else if (p_format[i].usage_flags & TEXTURE_USAGE_RESOLVE_ATTACHMENT_BIT) {
 			reference.layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
 			resolve_references.push_back(reference);
+			// if resolves are done, we need to ensure the copy is safe
+			dependency_to_external.dstStageMask |= VK_PIPELINE_STAGE_TRANSFER_BIT;
+			dependency_to_external.dstAccessMask |= VK_ACCESS_TRANSFER_READ_BIT;
 		} else {
 			ERR_FAIL_V_MSG(VK_NULL_HANDLE, "Texture index " + itos(i) + " is neither color, depth stencil or resolve so it can't be used as attachment.");
 		}
@@ -3186,8 +3242,8 @@ 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 = 0;
-	render_pass_create_info.pDependencies = nullptr;
+	render_pass_create_info.dependencyCount = 2;
+	render_pass_create_info.pDependencies = dependencies;
 
 	VkRenderPass render_pass;
 	VkResult res = vkCreateRenderPass(device, &render_pass_create_info, nullptr, &render_pass);
@@ -3212,7 +3268,7 @@ RenderingDevice::FramebufferFormatID RenderingDeviceVulkan::framebuffer_format_c
 	}
 
 	int color_references;
-	VkRenderPass render_pass = _render_pass_create(p_format, INITIAL_ACTION_CLEAR, FINAL_ACTION_DISCARD, INITIAL_ACTION_CLEAR, FINAL_ACTION_DISCARD, &color_references); //actions don't matter for this use case
+	VkRenderPass render_pass = _render_pass_create(p_format, INITIAL_ACTION_CLEAR, FINAL_ACTION_READ, INITIAL_ACTION_CLEAR, FINAL_ACTION_READ, &color_references); //actions don't matter for this use case
 
 	if (render_pass == VK_NULL_HANDLE) { //was likely invalid
 		return INVALID_ID;
@@ -4913,8 +4969,9 @@ Error RenderingDeviceVulkan::buffer_update(RID p_buffer, uint32_t p_offset, uint
 	ERR_FAIL_COND_V_MSG(draw_list && p_sync_with_draw, ERR_INVALID_PARAMETER,
 			"Updating buffers in 'sync to draw' mode is forbidden during creation of a draw list");
 
-	VkPipelineStageFlags dst_stage_mask = 0;
-	VkAccessFlags dst_access = 0;
+	// Protect subsequent updates...
+	VkPipelineStageFlags dst_stage_mask = VK_PIPELINE_STAGE_TRANSFER_BIT;
+	VkAccessFlags dst_access = VK_ACCESS_TRANSFER_WRITE_BIT;
 
 	Buffer *buffer = _get_buffer_from_owner(p_buffer, dst_stage_mask, dst_access);
 	if (!buffer) {
@@ -5621,7 +5678,7 @@ Error RenderingDeviceVulkan::_draw_list_render_pass_begin(Framebuffer *framebuff
 			image_memory_barrier.subresourceRange.baseArrayLayer = texture->base_layer;
 			image_memory_barrier.subresourceRange.layerCount = texture->layers;
 
-			vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, VK_PIPELINE_STAGE_VERTEX_INPUT_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+			vkCmdPipelineBarrier(command_buffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 
 			texture->layout = VK_IMAGE_LAYOUT_GENERAL;
 
@@ -6501,7 +6558,7 @@ 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_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+			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;
 
@@ -6698,13 +6755,13 @@ void RenderingDeviceVulkan::compute_list_add_barrier(ComputeListID p_list) {
 
 void RenderingDeviceVulkan::compute_list_end() {
 	ERR_FAIL_COND(!compute_list);
-
+	const VkPipelineStageFlags dest_stage_mask = VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT | VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT;
 	for (Set<Texture *>::Element *E = compute_list->state.textures_to_sampled_layout.front(); E; E = E->next()) {
 		VkImageMemoryBarrier image_memory_barrier;
 		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.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
+		image_memory_barrier.dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_INDIRECT_COMMAND_READ_BIT;
 		image_memory_barrier.oldLayout = E->get()->layout;
 		image_memory_barrier.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 
@@ -6717,7 +6774,8 @@ void RenderingDeviceVulkan::compute_list_end() {
 		image_memory_barrier.subresourceRange.baseArrayLayer = E->get()->base_layer;
 		image_memory_barrier.subresourceRange.layerCount = E->get()->layers;
 
-		vkCmdPipelineBarrier(compute_list->command_buffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
+		// 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, dest_stage_mask, 0, 0, nullptr, 0, nullptr, 1, &image_memory_barrier);
 
 		E->get()->layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 	}
@@ -6727,7 +6785,7 @@ void RenderingDeviceVulkan::compute_list_end() {
 #ifdef FORCE_FULL_BARRIER
 	_full_barrier(true);
 #else
-	_memory_barrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_SHADER_WRITE_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_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);
+	_memory_barrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, dest_stage_mask, VK_ACCESS_SHADER_WRITE_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | VK_ACCESS_INDEX_READ_BIT | VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_INDIRECT_COMMAND_READ_BIT, true);
 #endif
 }