render/vulkan: implement render pass API

3 files changed, 646 insertions, 38 deletions

diff --git a/render/vulkan/meson.build b/render/vulkan/meson.build
index 0c45a713..b08e4f3c 100644
--- a/render/vulkan/meson.build
+++ b/render/vulkan/meson.build
@@ -37,6 +37,7 @@ glslang_version_info = run_command(glslang, '--version', check: true).stdout()
 glslang_version = glslang_version_info.split('\n')[0].split(':')[-1]
 
 wlr_files += files(
+	'pass.c',
 	'renderer.c',
 	'texture.c',
 	'vulkan.c',
diff --git a/render/vulkan/pass.c b/render/vulkan/pass.c
new file mode 100644
index 00000000..17c92277
--- /dev/null
+++ b/render/vulkan/pass.c
@@ -0,0 +1,603 @@
+#include <assert.h>
+#include <drm_fourcc.h>
+#include <stdlib.h>
+#include <wlr/util/log.h>
+
+#include "render/vulkan.h"
+#include "types/wlr_matrix.h"
+
+static const struct wlr_render_pass_impl render_pass_impl;
+
+static struct wlr_vk_render_pass *get_render_pass(struct wlr_render_pass *wlr_pass) {
+	assert(wlr_pass->impl == &render_pass_impl);
+	struct wlr_vk_render_pass *pass = wl_container_of(wlr_pass, pass, base);
+	return pass;
+}
+
+static void bind_pipeline(struct wlr_vk_render_pass *pass, VkPipeline pipeline) {
+	if (pipeline == pass->bound_pipeline) {
+		return;
+	}
+
+	vkCmdBindPipeline(pass->command_buffer->vk, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+	pass->bound_pipeline = pipeline;
+}
+
+static void get_clip_region(struct wlr_vk_render_pass *pass,
+		const pixman_region32_t *in, pixman_region32_t *out) {
+	if (in != NULL) {
+		pixman_region32_init(out);
+		pixman_region32_copy(out, in);
+	} else {
+		struct wlr_buffer *buffer = pass->render_buffer->wlr_buffer;
+		pixman_region32_init_rect(out, 0, 0, buffer->width, buffer->height);
+	}
+}
+
+static void convert_pixman_box_to_vk_rect(const pixman_box32_t *box, VkRect2D *rect) {
+	*rect = (VkRect2D){
+		.offset = { .x = box->x1, .y = box->y1 },
+		.extent = { .width = box->x2 - box->x1, .height = box->y2 - box->y1 },
+	};
+}
+
+static float color_to_linear(float non_linear) {
+	// See https://www.w3.org/Graphics/Color/srgb
+	return (non_linear > 0.04045) ?
+		pow((non_linear + 0.055) / 1.055, 2.4) :
+		non_linear / 12.92;
+}
+
+static void mat3_to_mat4(const float mat3[9], float mat4[4][4]) {
+	memset(mat4, 0, sizeof(float) * 16);
+	mat4[0][0] = mat3[0];
+	mat4[0][1] = mat3[1];
+	mat4[0][3] = mat3[2];
+
+	mat4[1][0] = mat3[3];
+	mat4[1][1] = mat3[4];
+	mat4[1][3] = mat3[5];
+
+	mat4[2][2] = 1.f;
+	mat4[3][3] = 1.f;
+}
+
+static bool render_pass_submit(struct wlr_render_pass *wlr_pass) {
+	struct wlr_vk_render_pass *pass = get_render_pass(wlr_pass);
+	struct wlr_vk_renderer *renderer = pass->renderer;
+	struct wlr_vk_command_buffer *render_cb = pass->command_buffer;
+	struct wlr_vk_render_buffer *render_buffer = pass->render_buffer;
+
+	if (vulkan_record_stage_cb(renderer) == VK_NULL_HANDLE) {
+		return false;
+	}
+
+	struct wlr_vk_command_buffer *stage_cb = renderer->stage.cb;
+	assert(stage_cb != NULL);
+	renderer->stage.cb = NULL;
+
+	if (render_buffer->blend_image) {
+		// Apply output shader to map blend image to actual output image
+		vkCmdNextSubpass(render_cb->vk, VK_SUBPASS_CONTENTS_INLINE);
+
+		VkPipeline pipe = render_buffer->render_setup->output_pipe;
+		if (pipe != renderer->bound_pipe) {
+			vkCmdBindPipeline(render_cb->vk, VK_PIPELINE_BIND_POINT_GRAPHICS, pipe);
+			renderer->bound_pipe = pipe;
+		}
+
+		float final_matrix[9] = {
+			renderer->render_width, 0, -1,
+			0, renderer->render_height, -1,
+			0, 0, 0,
+		};
+		struct wlr_vk_vert_pcr_data vert_pcr_data = {
+			.uv_off = { 0, 0 },
+			.uv_size = { 1, 1 },
+		};
+		mat3_to_mat4(final_matrix, vert_pcr_data.mat4);
+
+		vkCmdPushConstants(render_cb->vk, renderer->output_pipe_layout,
+			VK_SHADER_STAGE_VERTEX_BIT, 0, sizeof(vert_pcr_data), &vert_pcr_data);
+		vkCmdBindDescriptorSets(render_cb->vk,
+			VK_PIPELINE_BIND_POINT_GRAPHICS, renderer->output_pipe_layout,
+			0, 1, &render_buffer->blend_descriptor_set, 0, NULL);
+
+		vkCmdDraw(render_cb->vk, 4, 1, 0, 0);
+	}
+
+	vkCmdEndRenderPass(render_cb->vk);
+
+	// insert acquire and release barriers for dmabuf-images
+	uint32_t barrier_count = wl_list_length(&renderer->foreign_textures) + 1;
+	VkImageMemoryBarrier *acquire_barriers = calloc(barrier_count, sizeof(VkImageMemoryBarrier));
+	VkImageMemoryBarrier *release_barriers = calloc(barrier_count, sizeof(VkImageMemoryBarrier));
+	VkSemaphoreSubmitInfoKHR *render_wait = calloc(barrier_count * WLR_DMABUF_MAX_PLANES, sizeof(VkSemaphoreSubmitInfoKHR));
+	if (acquire_barriers == NULL || release_barriers == NULL || render_wait == NULL) {
+		wlr_log_errno(WLR_ERROR, "Allocation failed");
+		free(acquire_barriers);
+		free(release_barriers);
+		free(render_wait);
+		return false;
+	}
+
+	struct wlr_vk_texture *texture, *tmp_tex;
+	size_t idx = 0;
+	uint32_t render_wait_len = 0;
+	wl_list_for_each_safe(texture, tmp_tex, &renderer->foreign_textures, foreign_link) {
+		VkImageLayout src_layout = VK_IMAGE_LAYOUT_GENERAL;
+		if (!texture->transitioned) {
+			src_layout = VK_IMAGE_LAYOUT_UNDEFINED;
+			texture->transitioned = true;
+		}
+
+		// acquire
+		acquire_barriers[idx] = (VkImageMemoryBarrier){
+			.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+			.srcQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+			.dstQueueFamilyIndex = renderer->dev->queue_family,
+			.image = texture->image,
+			.oldLayout = src_layout,
+			.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
+			.srcAccessMask = 0, // ignored anyways
+			.dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
+			.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+			.subresourceRange.layerCount = 1,
+			.subresourceRange.levelCount = 1,
+		};
+
+		// release
+		release_barriers[idx] = (VkImageMemoryBarrier){
+			.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+			.srcQueueFamilyIndex = renderer->dev->queue_family,
+			.dstQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+			.image = texture->image,
+			.oldLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL,
+			.newLayout = VK_IMAGE_LAYOUT_GENERAL,
+			.srcAccessMask = VK_ACCESS_SHADER_READ_BIT,
+			.dstAccessMask = 0, // ignored anyways
+			.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+			.subresourceRange.layerCount = 1,
+			.subresourceRange.levelCount = 1,
+		};
+
+		++idx;
+
+		if (!vulkan_sync_foreign_texture(texture)) {
+			wlr_log(WLR_ERROR, "Failed to wait for foreign texture DMA-BUF fence");
+		} else {
+			for (size_t i = 0; i < WLR_DMABUF_MAX_PLANES; i++) {
+				if (texture->foreign_semaphores[i] != VK_NULL_HANDLE) {
+					assert(render_wait_len < barrier_count * WLR_DMABUF_MAX_PLANES);
+					render_wait[render_wait_len++] = (VkSemaphoreSubmitInfoKHR){
+						.sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO_KHR,
+						.semaphore = texture->foreign_semaphores[i],
+						.stageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT_KHR,
+					};
+				}
+			}
+		}
+
+		wl_list_remove(&texture->foreign_link);
+		texture->owned = false;
+	}
+
+	// also add acquire/release barriers for the current render buffer
+	VkImageLayout src_layout = VK_IMAGE_LAYOUT_GENERAL;
+	if (!render_buffer->transitioned) {
+		src_layout = VK_IMAGE_LAYOUT_PREINITIALIZED;
+		render_buffer->transitioned = true;
+	}
+
+	if (render_buffer->blend_image) {
+		// The render pass changes the blend image layout from
+		// color attachment to read only, so on each frame, before
+		// the render pass starts, we change it back
+		VkImageLayout blend_src_layout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
+		if (!render_buffer->blend_transitioned) {
+			blend_src_layout = VK_IMAGE_LAYOUT_UNDEFINED;
+			render_buffer->blend_transitioned = true;
+		}
+
+		VkImageMemoryBarrier blend_acq_barrier = {
+			.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+			.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+			.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+			.image = render_buffer->blend_image,
+			.oldLayout = blend_src_layout,
+			.newLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
+			.srcAccessMask = VK_ACCESS_SHADER_READ_BIT,
+			.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+			.subresourceRange = {
+				.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+				.layerCount = 1,
+				.levelCount = 1,
+			},
+		};
+		vkCmdPipelineBarrier(stage_cb->vk, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
+			VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+			0, 0, NULL, 0, NULL, 1, &blend_acq_barrier);
+	}
+
+	// acquire render buffer before rendering
+	acquire_barriers[idx] = (VkImageMemoryBarrier){
+		.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+		.srcQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+		.dstQueueFamilyIndex = renderer->dev->queue_family,
+		.image = render_buffer->image,
+		.oldLayout = src_layout,
+		.newLayout = VK_IMAGE_LAYOUT_GENERAL,
+		.srcAccessMask = 0, // ignored anyways
+		.dstAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+			VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+		.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+		.subresourceRange.layerCount = 1,
+		.subresourceRange.levelCount = 1,
+	};
+
+	// release render buffer after rendering
+	release_barriers[idx] = (VkImageMemoryBarrier){
+		.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+		.srcQueueFamilyIndex = renderer->dev->queue_family,
+		.dstQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+		.image = render_buffer->image,
+		.oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+		.newLayout = VK_IMAGE_LAYOUT_GENERAL,
+		.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+			VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
+		.dstAccessMask = 0, // ignored anyways
+		.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+		.subresourceRange.layerCount = 1,
+		.subresourceRange.levelCount = 1,
+	};
+
+	++idx;
+
+	vkCmdPipelineBarrier(stage_cb->vk, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+		VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+		0, 0, NULL, 0, NULL, barrier_count, acquire_barriers);
+
+	vkCmdPipelineBarrier(render_cb->vk, VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
+		VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, 0, NULL, 0, NULL,
+		barrier_count, release_barriers);
+
+	free(acquire_barriers);
+	free(release_barriers);
+
+	// No semaphores needed here.
+	// We don't need a semaphore from the stage/transfer submission
+	// to the render submissions since they are on the same queue
+	// and we have a renderpass dependency for that.
+	uint64_t stage_timeline_point = vulkan_end_command_buffer(stage_cb, renderer);
+	if (stage_timeline_point == 0) {
+		return false;
+	}
+
+	VkCommandBufferSubmitInfoKHR stage_cb_info = {
+		.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_SUBMIT_INFO_KHR,
+		.commandBuffer = stage_cb->vk,
+	};
+	VkSemaphoreSubmitInfoKHR stage_signal = {
+		.sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO_KHR,
+		.semaphore = renderer->timeline_semaphore,
+		.value = stage_timeline_point,
+	};
+	VkSubmitInfo2KHR stage_submit = {
+		.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2_KHR,
+		.commandBufferInfoCount = 1,
+		.pCommandBufferInfos = &stage_cb_info,
+		.signalSemaphoreInfoCount = 1,
+		.pSignalSemaphoreInfos = &stage_signal,
+	};
+
+	VkSemaphoreSubmitInfoKHR stage_wait;
+	if (renderer->stage.last_timeline_point > 0) {
+		stage_wait = (VkSemaphoreSubmitInfoKHR){
+			.sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO_KHR,
+			.semaphore = renderer->timeline_semaphore,
+			.value = renderer->stage.last_timeline_point,
+			.stageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT_KHR,
+		};
+
+		stage_submit.waitSemaphoreInfoCount = 1;
+		stage_submit.pWaitSemaphoreInfos = &stage_wait;
+	}
+
+	renderer->stage.last_timeline_point = stage_timeline_point;
+
+	uint64_t render_timeline_point = vulkan_end_command_buffer(render_cb, renderer);
+	if (render_timeline_point == 0) {
+		return false;
+	}
+
+	uint32_t render_signal_len = 1;
+	VkSemaphoreSubmitInfoKHR render_signal[2] = {0};
+	render_signal[0] = (VkSemaphoreSubmitInfoKHR){
+		.sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO_KHR,
+		.semaphore = renderer->timeline_semaphore,
+		.value = render_timeline_point,
+	};
+	if (renderer->dev->implicit_sync_interop) {
+		if (render_cb->binary_semaphore == VK_NULL_HANDLE) {
+			VkExportSemaphoreCreateInfo export_info = {
+				.sType = VK_STRUCTURE_TYPE_EXPORT_SEMAPHORE_CREATE_INFO,
+				.handleTypes = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT,
+			};
+			VkSemaphoreCreateInfo semaphore_info = {
+				.sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
+				.pNext = &export_info,
+			};
+			VkResult res = vkCreateSemaphore(renderer->dev->dev, &semaphore_info,
+				NULL, &render_cb->binary_semaphore);
+			if (res != VK_SUCCESS) {
+				wlr_vk_error("vkCreateSemaphore", res);
+				return false;
+			}
+		}
+
+		render_signal[render_signal_len++] = (VkSemaphoreSubmitInfoKHR){
+			.sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO_KHR,
+			.semaphore = render_cb->binary_semaphore,
+		};
+	}
+
+	VkCommandBufferSubmitInfoKHR render_cb_info = {
+		.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_SUBMIT_INFO_KHR,
+		.commandBuffer = render_cb->vk,
+	};
+	VkSubmitInfo2KHR render_submit = {
+		.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2_KHR,
+		.waitSemaphoreInfoCount = render_wait_len,
+		.pWaitSemaphoreInfos = render_wait,
+		.commandBufferInfoCount = 1,
+		.pCommandBufferInfos = &render_cb_info,
+		.signalSemaphoreInfoCount = render_signal_len,
+		.pSignalSemaphoreInfos = render_signal,
+	};
+
+	VkSubmitInfo2KHR submit_info[] = { stage_submit, render_submit };
+	VkResult res = renderer->dev->api.vkQueueSubmit2KHR(renderer->dev->queue, 2, submit_info, VK_NULL_HANDLE);
+	if (res == VK_ERROR_DEVICE_LOST) {
+		wlr_log(WLR_ERROR, "vkQueueSubmit failed with VK_ERROR_DEVICE_LOST");
+		wl_signal_emit_mutable(&renderer->wlr_renderer.events.lost, NULL);
+		return false;
+	} else if (res != VK_SUCCESS) {
+		wlr_vk_error("vkQueueSubmit", res);
+		return false;
+	}
+
+	free(render_wait);
+
+	struct wlr_vk_shared_buffer *stage_buf, *stage_buf_tmp;
+	wl_list_for_each_safe(stage_buf, stage_buf_tmp, &renderer->stage.buffers, link) {
+		if (stage_buf->allocs.size == 0) {
+			continue;
+		}
+		wl_list_remove(&stage_buf->link);
+		wl_list_insert(&stage_cb->stage_buffers, &stage_buf->link);
+	}
+
+	if (!vulkan_sync_render_buffer(renderer, render_buffer, render_cb)) {
+		return false;
+	}
+
+	wlr_buffer_unlock(render_buffer->wlr_buffer);
+	free(pass);
+	return true;
+}
+
+static void render_pass_add_rect(struct wlr_render_pass *wlr_pass,
+		const struct wlr_render_rect_options *options) {
+	struct wlr_vk_render_pass *pass = get_render_pass(wlr_pass);
+	struct wlr_vk_renderer *renderer = pass->renderer;
+	VkCommandBuffer cb = pass->command_buffer->vk;
+
+	// Input color values are given in sRGB space, shader expects
+	// them in linear space. The shader does all computation in linear
+	// space and expects in inputs in linear space since it outputs
+	// colors in linear space as well (and vulkan then automatically
+	// does the conversion for out sRGB render targets).
+	float linear_color[] = {
+		color_to_linear(options->color.r),
+		color_to_linear(options->color.g),
+		color_to_linear(options->color.b),
+		options->color.a, // no conversion for alpha
+	};
+
+	pixman_region32_t clip;
+	get_clip_region(pass, options->clip, &clip);
+
+	int clip_rects_len;
+	const pixman_box32_t *clip_rects = pixman_region32_rectangles(&clip, &clip_rects_len);
+
+	switch (options->blend_mode) {
+	case WLR_RENDER_BLEND_MODE_PREMULTIPLIED:;
+		float proj[9], matrix[9];
+		wlr_matrix_identity(proj);
+		wlr_matrix_project_box(matrix, &options->box, WL_OUTPUT_TRANSFORM_NORMAL, 0, proj);
+		wlr_matrix_multiply(matrix, pass->projection, matrix);
+
+		struct wlr_vk_vert_pcr_data vert_pcr_data = {
+			.uv_off = { 0, 0 },
+			.uv_size = { 1, 1 },
+		};
+		mat3_to_mat4(matrix, vert_pcr_data.mat4);
+
+		bind_pipeline(pass, pass->render_buffer->render_setup->quad_pipe);
+		vkCmdPushConstants(cb, renderer->default_pipeline_layout.vk,
+			VK_SHADER_STAGE_VERTEX_BIT, 0, sizeof(vert_pcr_data), &vert_pcr_data);
+		vkCmdPushConstants(cb, renderer->default_pipeline_layout.vk,
+			VK_SHADER_STAGE_FRAGMENT_BIT, sizeof(vert_pcr_data), sizeof(float) * 4,
+			linear_color);
+
+		for (int i = 0; i < clip_rects_len; i++) {
+			VkRect2D rect;
+			convert_pixman_box_to_vk_rect(&clip_rects[i], &rect);
+			vkCmdSetScissor(cb, 0, 1, &rect);
+			vkCmdDraw(cb, 4, 1, 0, 0);
+		}
+		break;
+	case WLR_RENDER_BLEND_MODE_NONE:;
+		VkClearAttachment clear_att = {
+			.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
+			.colorAttachment = 0,
+			.clearValue.color.float32 = {
+				linear_color[0],
+				linear_color[1],
+				linear_color[2],
+				linear_color[3],
+			},
+		};
+		VkClearRect clear_rect = {
+			.rect = {
+				.offset = { options->box.x, options->box.y },
+				.extent = { options->box.width, options->box.height },
+			},
+			.layerCount = 1,
+		};
+		for (int i = 0; i < clip_rects_len; i++) {
+			VkRect2D rect;
+			convert_pixman_box_to_vk_rect(&clip_rects[i], &rect);
+			vkCmdSetScissor(cb, 0, 1, &rect);
+			vkCmdClearAttachments(cb, 1, &clear_att, 1, &clear_rect);
+		}
+		break;
+	}
+
+	pixman_region32_fini(&clip);
+}
+
+static void render_pass_add_texture(struct wlr_render_pass *wlr_pass,
+		const struct wlr_render_texture_options *options) {
+	struct wlr_vk_render_pass *pass = get_render_pass(wlr_pass);
+	struct wlr_vk_renderer *renderer = pass->renderer;
+	struct wlr_vk_render_buffer *render_buffer = pass->render_buffer;
+	VkCommandBuffer cb = pass->command_buffer->vk;
+
+	struct wlr_vk_texture *texture = vulkan_get_texture(options->texture);
+	assert(texture->renderer == renderer);
+
+	if (texture->dmabuf_imported && !texture->owned) {
+		// Store this texture in the list of textures that need to be
+		// acquired before rendering and released after rendering.
+		// We don't do it here immediately since barriers inside
+		// a renderpass are suboptimal (would require additional renderpass
+		// dependency and potentially multiple barriers) and it's
+		// better to issue one barrier for all used textures anyways.
+		texture->owned = true;
+		assert(texture->foreign_link.prev == NULL);
+		assert(texture->foreign_link.next == NULL);
+		wl_list_insert(&renderer->foreign_textures, &texture->foreign_link);
+	}
+
+	struct wlr_fbox src_box;
+	wlr_render_texture_options_get_src_box(options, &src_box);
+	struct wlr_box dst_box;
+	wlr_render_texture_options_get_dst_box(options, &dst_box);
+	float alpha = wlr_render_texture_options_get_alpha(options);
+
+	pixman_region32_t clip;
+	get_clip_region(pass, options->clip, &clip);
+
+	float proj[9], matrix[9];
+	wlr_matrix_identity(proj);
+	wlr_matrix_project_box(matrix, &dst_box, options->transform, 0, proj);
+	wlr_matrix_multiply(matrix, pass->projection, matrix);
+
+	struct wlr_vk_vert_pcr_data vert_pcr_data = {
+		.uv_off = {
+			src_box.x / options->texture->width,
+			src_box.y / options->texture->height,
+		},
+		.uv_size = {
+			src_box.width / options->texture->width,
+			src_box.height / options->texture->height,
+		},
+	};
+	mat3_to_mat4(matrix, vert_pcr_data.mat4);
+
+	VkPipelineLayout pipeline_layout = texture->pipeline_layout->vk;
+	VkPipeline pipeline = vulkan_get_texture_pipeline(texture, render_buffer->render_setup);
+
+	bind_pipeline(pass, pipeline);
+
+	vkCmdBindDescriptorSets(cb, VK_PIPELINE_BIND_POINT_GRAPHICS,
+		pipeline_layout, 0, 1, &texture->ds, 0, NULL);
+
+	vkCmdPushConstants(cb, pipeline_layout,
+		VK_SHADER_STAGE_VERTEX_BIT, 0, sizeof(vert_pcr_data), &vert_pcr_data);
+	vkCmdPushConstants(cb, pipeline_layout,
+		VK_SHADER_STAGE_FRAGMENT_BIT, sizeof(vert_pcr_data), sizeof(float),
+		&alpha);
+
+	int clip_rects_len;
+	const pixman_box32_t *clip_rects = pixman_region32_rectangles(&clip, &clip_rects_len);
+	for (int i = 0; i < clip_rects_len; i++) {
+		VkRect2D rect;
+		convert_pixman_box_to_vk_rect(&clip_rects[i], &rect);
+		vkCmdSetScissor(cb, 0, 1, &rect);
+		vkCmdDraw(cb, 4, 1, 0, 0);
+	}
+
+	texture->last_used_cb = pass->command_buffer;
+}
+
+static const struct wlr_render_pass_impl render_pass_impl = {
+	.submit = render_pass_submit,
+	.add_rect = render_pass_add_rect,
+	.add_texture = render_pass_add_texture,
+};
+
+struct wlr_vk_render_pass *vulkan_begin_render_pass(struct wlr_vk_renderer *renderer,
+		struct wlr_vk_render_buffer *buffer) {
+	struct wlr_vk_render_pass *pass = calloc(1, sizeof(*pass));
+	if (pass == NULL) {
+		return NULL;
+	}
+	wlr_render_pass_init(&pass->base, &render_pass_impl);
+	pass->renderer = renderer;
+
+	struct wlr_vk_command_buffer *cb = vulkan_acquire_command_buffer(renderer);
+	if (cb == NULL) {
+		free(pass);
+		return NULL;
+	}
+
+	VkCommandBufferBeginInfo begin_info = {
+		.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
+	};
+	VkResult res = vkBeginCommandBuffer(cb->vk, &begin_info);
+	if (res != VK_SUCCESS) {
+		wlr_vk_error("vkBeginCommandBuffer", res);
+		free(pass);
+		return NULL;
+	}
+
+	int width = buffer->wlr_buffer->width;
+	int height = buffer->wlr_buffer->height;
+	VkRect2D rect = { .extent = { width, height } };
+
+	VkRenderPassBeginInfo rp_info = {
+		.sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
+		.renderArea = rect,
+		.renderPass = buffer->render_setup->render_pass,
+		.framebuffer = buffer->framebuffer,
+		.clearValueCount = 0,
+	};
+	vkCmdBeginRenderPass(cb->vk, &rp_info, VK_SUBPASS_CONTENTS_INLINE);
+
+	vkCmdSetViewport(cb->vk, 0, 1, &(VkViewport){
+		.width = width,
+		.height = height,
+		.maxDepth = 1,
+	});
+
+	// matrix_projection() assumes a GL coordinate system so we need
+	// to pass WL_OUTPUT_TRANSFORM_FLIPPED_180 to adjust it for vulkan.
+	matrix_projection(pass->projection, width, height, WL_OUTPUT_TRANSFORM_FLIPPED_180);
+
+	wlr_buffer_lock(buffer->wlr_buffer);
+	pass->render_buffer = buffer;
+	pass->command_buffer = cb;
+	return pass;
+}
diff --git a/render/vulkan/renderer.c b/render/vulkan/renderer.c
index 4a7f8907..f4af9caf 100644
--- a/render/vulkan/renderer.c
+++ b/render/vulkan/renderer.c
@@ -58,13 +58,6 @@ struct wlr_vk_renderer *vulkan_get_renderer(struct wlr_renderer *wlr_renderer) {
 static struct wlr_vk_render_format_setup *find_or_create_render_setup(
 		struct wlr_vk_renderer *renderer, VkFormat format, bool has_blending_buffer);
 
-// vertex shader push constant range data
-struct vert_pcr_data {
-	float mat4[4][4];
-	float uv_off[2];
-	float uv_size[2];
-};
-
 // https://www.w3.org/Graphics/Color/srgb
 static float color_to_linear(float non_linear) {
 	return (non_linear > 0.04045) ?
@@ -360,16 +353,9 @@ error_alloc:
 	};
 }
 
-static struct wlr_vk_command_buffer *acquire_command_buffer(
-	struct wlr_vk_renderer *renderer);
-static uint64_t end_command_buffer(struct wlr_vk_command_buffer *cb,
-	struct wlr_vk_renderer *renderer);
-static bool wait_command_buffer(struct wlr_vk_command_buffer *cb,
-	struct wlr_vk_renderer *renderer);
-
 VkCommandBuffer vulkan_record_stage_cb(struct wlr_vk_renderer *renderer) {
 	if (renderer->stage.cb == NULL) {
-		renderer->stage.cb = acquire_command_buffer(renderer);
+		renderer->stage.cb = vulkan_acquire_command_buffer(renderer);
 		if (renderer->stage.cb == NULL) {
 			return VK_NULL_HANDLE;
 		}
@@ -391,7 +377,7 @@ bool vulkan_submit_stage_wait(struct wlr_vk_renderer *renderer) {
 	struct wlr_vk_command_buffer *cb = renderer->stage.cb;
 	renderer->stage.cb = NULL;
 
-	uint64_t timeline_point = end_command_buffer(cb, renderer);
+	uint64_t timeline_point = vulkan_end_command_buffer(cb, renderer);
 	if (timeline_point == 0) {
 		return false;
 	}
@@ -418,7 +404,7 @@ bool vulkan_submit_stage_wait(struct wlr_vk_renderer *renderer) {
 	// NOTE: don't release stage allocations here since they may still be
 	// used for reading. Will be done next frame.
 
-	return wait_command_buffer(cb, renderer);
+	return vulkan_wait_command_buffer(cb, renderer);
 }
 
 struct wlr_vk_format_props *vulkan_format_props_from_drm(
@@ -456,7 +442,7 @@ static bool init_command_buffer(struct wlr_vk_command_buffer *cb,
 	return true;
 }
 
-static bool wait_command_buffer(struct wlr_vk_command_buffer *cb,
+bool vulkan_wait_command_buffer(struct wlr_vk_command_buffer *cb,
 		struct wlr_vk_renderer *renderer) {
 	VkResult res;
 
@@ -546,13 +532,13 @@ static struct wlr_vk_command_buffer *get_command_buffer(
 	}
 
 	// Block until a busy command buffer becomes available
-	if (!wait_command_buffer(wait, renderer)) {
+	if (!vulkan_wait_command_buffer(wait, renderer)) {
 		return NULL;
 	}
 	return wait;
 }
 
-static struct wlr_vk_command_buffer *acquire_command_buffer(
+struct wlr_vk_command_buffer *vulkan_acquire_command_buffer(
 		struct wlr_vk_renderer *renderer) {
 	struct wlr_vk_command_buffer *cb = get_command_buffer(renderer);
 	if (cb == NULL) {
@@ -565,7 +551,7 @@ static struct wlr_vk_command_buffer *acquire_command_buffer(
 	return cb;
 }
 
-static uint64_t end_command_buffer(struct wlr_vk_command_buffer *cb,
+uint64_t vulkan_end_command_buffer(struct wlr_vk_command_buffer *cb,
 		struct wlr_vk_renderer *renderer) {
 	assert(cb->recording);
 	cb->recording = false;
@@ -900,7 +886,7 @@ static bool vulkan_begin(struct wlr_renderer *wlr_renderer,
 	struct wlr_vk_renderer *renderer = vulkan_get_renderer(wlr_renderer);
 	assert(renderer->current_render_buffer);
 
-	struct wlr_vk_command_buffer *cb = acquire_command_buffer(renderer);
+	struct wlr_vk_command_buffer *cb = vulkan_acquire_command_buffer(renderer);
 	if (cb == NULL) {
 		return false;
 	}
@@ -949,7 +935,7 @@ static bool vulkan_begin(struct wlr_renderer *wlr_renderer,
 	return true;
 }
 
-static bool vulkan_sync_foreign_texture(struct wlr_vk_texture *texture) {
+bool vulkan_sync_foreign_texture(struct wlr_vk_texture *texture) {
 	struct wlr_vk_renderer *renderer = texture->renderer;
 	VkResult res;
 
@@ -1019,18 +1005,17 @@ static bool vulkan_sync_foreign_texture(struct wlr_vk_texture *texture) {
 	return true;
 }
 
-static bool vulkan_sync_render_buffer(struct wlr_vk_renderer *renderer,
-		struct wlr_vk_command_buffer *cb) {
+bool vulkan_sync_render_buffer(struct wlr_vk_renderer *renderer,
+		struct wlr_vk_render_buffer *render_buffer, struct wlr_vk_command_buffer *cb) {
 	VkResult res;
 
 	if (!renderer->dev->implicit_sync_interop) {
 		// We have no choice but to block here sadly
-		return wait_command_buffer(cb, renderer);
+		return vulkan_wait_command_buffer(cb, renderer);
 	}
 
 	struct wlr_dmabuf_attributes dmabuf = {0};
-	if (!wlr_buffer_get_dmabuf(renderer->current_render_buffer->wlr_buffer,
-			&dmabuf)) {
+	if (!wlr_buffer_get_dmabuf(render_buffer->wlr_buffer, &dmabuf)) {
 		wlr_log(WLR_ERROR, "wlr_buffer_get_dmabuf failed");
 		return false;
 	}
@@ -1095,7 +1080,7 @@ static void vulkan_end(struct wlr_renderer *wlr_renderer) {
 			0.f, renderer->render_height, -1.f,
 			0.f, 0.f, 0.f,
 		};
-		struct vert_pcr_data vert_pcr_data;
+		struct wlr_vk_vert_pcr_data vert_pcr_data;
 		mat3_to_mat4(final_matrix, vert_pcr_data.mat4);
 		vert_pcr_data.uv_off[0] = 0.f;
 		vert_pcr_data.uv_off[1] = 0.f;
@@ -1275,7 +1260,7 @@ static void vulkan_end(struct wlr_renderer *wlr_renderer) {
 	// We don't need a semaphore from the stage/transfer submission
 	// to the render submissions since they are on the same queue
 	// and we have a renderpass dependency for that.
-	uint64_t stage_timeline_point = end_command_buffer(stage_cb, renderer);
+	uint64_t stage_timeline_point = vulkan_end_command_buffer(stage_cb, renderer);
 	if (stage_timeline_point == 0) {
 		return;
 	}
@@ -1312,7 +1297,7 @@ static void vulkan_end(struct wlr_renderer *wlr_renderer) {
 
 	renderer->stage.last_timeline_point = stage_timeline_point;
 
-	uint64_t render_timeline_point = end_command_buffer(render_cb, renderer);
+	uint64_t render_timeline_point = vulkan_end_command_buffer(render_cb, renderer);
 	if (render_timeline_point == 0) {
 		return;
 	}
@@ -1384,12 +1369,12 @@ static void vulkan_end(struct wlr_renderer *wlr_renderer) {
 		wl_list_insert(&stage_cb->stage_buffers, &stage_buf->link);
 	}
 
-	if (!vulkan_sync_render_buffer(renderer, render_cb)) {
+	if (!vulkan_sync_render_buffer(renderer, renderer->current_render_buffer, render_cb)) {
 		return;
 	}
 }
 
-static VkPipeline get_texture_pipeline(struct wlr_vk_texture *texture,
+VkPipeline vulkan_get_texture_pipeline(struct wlr_vk_texture *texture,
 		struct wlr_vk_render_format_setup *render_setup) {
 	if (texture->format->is_ycbcr) {
 		size_t pipeline_layout_index = texture->pipeline_layout - texture->renderer->ycbcr_pipeline_layouts;
@@ -1426,7 +1411,7 @@ static bool vulkan_render_subtexture_with_matrix(struct wlr_renderer *wlr_render
 	}
 
 	VkPipelineLayout pipe_layout = texture->pipeline_layout->vk;
-	VkPipeline pipe = get_texture_pipeline(texture, renderer->current_render_buffer->render_setup);
+	VkPipeline pipe = vulkan_get_texture_pipeline(texture, renderer->current_render_buffer->render_setup);
 
 	if (pipe != renderer->bound_pipe) {
 		vkCmdBindPipeline(cb, VK_PIPELINE_BIND_POINT_GRAPHICS, pipe);
@@ -1439,7 +1424,7 @@ static bool vulkan_render_subtexture_with_matrix(struct wlr_renderer *wlr_render
 	float final_matrix[9];
 	wlr_matrix_multiply(final_matrix, renderer->projection, matrix);
 
-	struct vert_pcr_data vert_pcr_data;
+	struct wlr_vk_vert_pcr_data vert_pcr_data;
 	mat3_to_mat4(final_matrix, vert_pcr_data.mat4);
 
 	vert_pcr_data.uv_off[0] = box->x / wlr_texture->width;
@@ -1529,7 +1514,7 @@ static void vulkan_render_quad_with_matrix(struct wlr_renderer *wlr_renderer,
 	float final_matrix[9];
 	wlr_matrix_multiply(final_matrix, renderer->projection, matrix);
 
-	struct vert_pcr_data vert_pcr_data;
+	struct wlr_vk_vert_pcr_data vert_pcr_data;
 	mat3_to_mat4(final_matrix, vert_pcr_data.mat4);
 	vert_pcr_data.uv_off[0] = 0.f;
 	vert_pcr_data.uv_off[1] = 0.f;
@@ -1915,6 +1900,24 @@ static uint32_t vulkan_get_render_buffer_caps(struct wlr_renderer *wlr_renderer)
 	return WLR_BUFFER_CAP_DMABUF;
 }
 
+static struct wlr_render_pass *vulkan_begin_buffer_pass(struct wlr_renderer *wlr_renderer, struct wlr_buffer *buffer) {
+	struct wlr_vk_renderer *renderer = vulkan_get_renderer(wlr_renderer);
+
+	struct wlr_vk_render_buffer *render_buffer = get_render_buffer(renderer, buffer);
+	if (!render_buffer) {
+		render_buffer = create_render_buffer(renderer, buffer);
+		if (!render_buffer) {
+			return NULL;
+		}
+	}
+
+	struct wlr_vk_render_pass *render_pass = vulkan_begin_render_pass(renderer, render_buffer);
+	if (render_pass == NULL) {
+		return NULL;
+	}
+	return &render_pass->base;
+}
+
 static const struct wlr_renderer_impl renderer_impl = {
 	.bind_buffer = vulkan_bind_buffer,
 	.begin = vulkan_begin,
@@ -1932,6 +1935,7 @@ static const struct wlr_renderer_impl renderer_impl = {
 	.get_drm_fd = vulkan_get_drm_fd,
 	.get_render_buffer_caps = vulkan_get_render_buffer_caps,
 	.texture_from_buffer = vulkan_texture_from_buffer,
+	.begin_buffer_pass = vulkan_begin_buffer_pass,
 };
 
 static bool init_sampler(struct wlr_vk_renderer *renderer, VkSampler *sampler,
@@ -1996,7 +2000,7 @@ static bool init_tex_layouts(struct wlr_vk_renderer *renderer,
 
 	VkPushConstantRange pc_ranges[2] = {
 		{
-			.size = sizeof(struct vert_pcr_data),
+			.size = sizeof(struct wlr_vk_vert_pcr_data),
 			.stageFlags = VK_SHADER_STAGE_VERTEX_BIT,
 		},
 		{
@@ -2052,7 +2056,7 @@ static bool init_blend_to_output_layouts(struct wlr_vk_renderer *renderer,
 	// pipeline layout -- standard vertex uniforms, no shader uniforms
 	VkPushConstantRange pc_ranges[1] = {
 		{
-			.size = sizeof(struct vert_pcr_data),
+			.size = sizeof(struct wlr_vk_vert_pcr_data),
 			.stageFlags = VK_SHADER_STAGE_VERTEX_BIT,
 		},
 	};