render/vulkan: Recycle memory maps for stage spans

Remapping buffers on every use causes a lot of unwanted pagefaults.
Reuse the mapping to significantly speed up the memcpy.

include/render/vulkan.h

@@ -483,6 +483,7 @@ struct wlr_vk_shared_buffer {
 	VkBuffer buffer;
 	VkDeviceMemory memory;
 	VkDeviceSize buf_size;
+	void *cpu_mapping;
 	struct wl_array allocs; // struct wlr_vk_allocation
 };
 

render/vulkan/renderer.c

@@ -187,6 +187,10 @@ static void shared_buffer_destroy(struct wlr_vk_renderer *r,
 	if (buffer->memory) {
 		vkFreeMemory(r->dev->dev, buffer->memory, NULL);
 	}
+	if (buffer->cpu_mapping) {
+		vkUnmapMemory(r->dev->dev, buffer->memory);
+		buffer->cpu_mapping = NULL;
+	}
 
 	wl_list_remove(&buffer->link);
 	free(buffer);

render/vulkan/texture.c

@@ -80,19 +80,19 @@ static bool write_pixels(struct wlr_vk_texture *texture,
 		return false;
 	}
 
-	void *vmap;
-	res = vkMapMemory(dev, span.buffer->memory, span.alloc.start,
-		bsize, 0, &vmap);
+	if (!span.buffer->cpu_mapping) {
+		res = vkMapMemory(dev, span.buffer->memory, 0, VK_WHOLE_SIZE, 0, &span.buffer->cpu_mapping);
 		if (res != VK_SUCCESS) {
 			wlr_vk_error("vkMapMemory", res);
 			free(copies);
 			return false;
 		}
-	char *map = (char *)vmap;
+	}
+	char *map = (char*)span.buffer->cpu_mapping + span.alloc.start;
 
 	// upload data
 
-	uint32_t buf_off = span.alloc.start + (map - (char *)vmap);
+	uint32_t buf_off = span.alloc.start;
 	for (int i = 0; i < rects_len; i++) {
 		pixman_box32_t rect = rects[i];
 		uint32_t width = rect.x2 - rect.x1;
@@ -137,9 +137,6 @@ static bool write_pixels(struct wlr_vk_texture *texture,
 		buf_off += height * packed_stride;
 	}
 
-	assert((uint32_t)(map - (char *)vmap) == bsize);
-	vkUnmapMemory(dev, span.buffer->memory);
-
 	// record staging cb
 	// will be executed before next frame
 	VkCommandBuffer cb = vulkan_record_stage_cb(renderer);