From 51b02627b1b138e3ecb5dd331c54c8b38642fd0e Mon Sep 17 00:00:00 2001
From: Matthew <matthew@sleepy.day>
Date: Sun, 30 Mar 2025 09:10:43 +1100
Subject: [PATCH] fixed some queue selection cases

---
 src/renderer_vulkan.c | 70 +++++++++++++++++++++++++++++++++----------
 src/renderer_vulkan.h |  1 +
 src/util.h            |  2 +-
 3 files changed, 56 insertions(+), 17 deletions(-)

diff --git a/src/renderer_vulkan.c b/src/renderer_vulkan.c
index 73877bd..83ba3d5 100644
--- a/src/renderer_vulkan.c
+++ b/src/renderer_vulkan.c
@@ -341,21 +341,55 @@ static DeviceQueues CheckDeviceQueueSupport(VkPhysicalDevice device, VkSurfaceKH
 	VkQueueFamilyProperties *families = ArenaAlloc(renderer.arena, sizeof(VkQueueFamilyProperties) * queue_count);
 	vkGetPhysicalDeviceQueueFamilyProperties(device, &queue_count, families);
 
-	for (i32 i = 0; i < queue_count; i++) 
+	if (queue_count == 1 && 
+			BitEq(families[0].queueFlags, VK_QUEUE_TRANSFER_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT) &&
+			families[0].queueCount == 1)
 	{
-		if (i == 0 && CheckQueueSurfaceSupport(i, device, surface)) 
+		queues.graphics = queues.transfer = 0;
+		queues.single_queue = true;
+	}
+	else 
+	{
+		b8 sparse_binding = false;
+		b8 transfer_only = false;
+		for (i32 i = 0; i < queue_count; i++) 
 		{
-			queues.graphics = i;
-			continue;
+			if (queues.graphics < 0 && CheckQueueSurfaceSupport(i, device, surface) && BitEq(families[i].queueFlags, VK_QUEUE_GRAPHICS_BIT)) 
+			{
+				queues.graphics = i;
+				continue;
+			}
+
+			if (BitEq(families[i].queueFlags, VK_QUEUE_TRANSFER_BIT | VK_QUEUE_SPARSE_BINDING_BIT) && !BitEq(families[i].queueFlags, VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT))
+			{
+				sparse_binding = true;
+				transfer_only = true;
+				queues.transfer = i;
+				continue;
+			}
+
+			if (BitEq(families[i].queueFlags, VK_QUEUE_TRANSFER_BIT | VK_QUEUE_SPARSE_BINDING_BIT) && !(sparse_binding && transfer_only))
+			{
+				sparse_binding = true;
+				queues.transfer = i;
+				continue;
+			}
+
+			if (BitEq(families[i].queueFlags, VK_QUEUE_TRANSFER_BIT) && !BitEq(families[i].queueFlags, VK_QUEUE_COMPUTE_BIT) && !sparse_binding)
+			{
+				transfer_only = true;
+				queues.transfer = i;
+				continue;
+			}
+
+			if (BitEq(families[i].queueFlags, VK_QUEUE_TRANSFER_BIT) && !sparse_binding && !transfer_only)
+				queues.transfer = i;
 		}
 
-		if (BitEq(families[i].queueFlags, VK_QUEUE_TRANSFER_BIT)) 
-			queues.transfer = i;
+		if (queues.transfer < 0) 
+			queues.transfer = queues.graphics;
 	}
 
-	if (queues.transfer < 0) 
-		queues.transfer = queues.graphics;
-
 	return queues;
 }
 
@@ -466,6 +500,7 @@ static b32 CreateDevice()
 		VkDeviceQueueCreateInfo queue_info[2] = {0};
 		f32 priority = 1.0f;
 		u32 count = 1;
+		u32 transfer_queue_index = 0;
 
 		queue_info[0].sType = STYPE(DEVICE_QUEUE_CREATE_INFO);
 		queue_info[0].queueFamilyIndex = queues->graphics;
@@ -473,12 +508,16 @@ static b32 CreateDevice()
 		queue_info[0].pQueuePriorities = &priority;
 		queue_info[0].flags = 0;
 
-		if (queues->graphics != queues->transfer) {
+		if (!queues->single_queue) {
 			queue_info[1].sType = STYPE(DEVICE_QUEUE_CREATE_INFO);
 			queue_info[1].queueFamilyIndex = queues->transfer;
 			queue_info[1].queueCount = 1;
 			queue_info[1].pQueuePriorities = &priority;
 			queue_info[1].flags = 0;
+
+			if (queues->transfer == queues->graphics)
+				transfer_queue_index = 1;
+
 			count++;
 		}
 
@@ -492,8 +531,9 @@ static b32 CreateDevice()
 		else 
 		{
 			Assert(InitVkDeviceFunctions(), "Failed to initialize device functions");
+			
 			vkGetDeviceQueue(renderer.vk.device, queues->graphics, 0, &queues->graphics_queue);
-			vkGetDeviceQueue(renderer.vk.device, queues->transfer, 0, &queues->transfer_queue);
+			vkGetDeviceQueue(renderer.vk.device, queues->transfer, transfer_queue_index, &queues->transfer_queue);
 			success = true;
 		}
 	}
@@ -757,12 +797,10 @@ static b32 CreateImmediateStructures()
 	ImmediateStructures *imm          = &renderer.vk.imm;
 	pool_create_info.queueFamilyIndex = renderer.vk.queues.transfer;
 	
-	if (renderer.vk_conf.avail_threads >= 10)
-		renderer.vk_conf.avail_threads = 3;
-	else if (renderer.vk_conf.avail_threads >= 8)
-		renderer.vk_conf.avail_threads = 2;
-	else
+	if (renderer.vk_conf.avail_threads >= 4 && !renderer.vk.queues.single_queue)
 		renderer.vk_conf.avail_threads = 1;
+	else
+		renderer.vk_conf.avail_threads = 0;
 
 	imm->pools          = ArenaAlloc(renderer.perm_arena, sizeof(VkCommandPool) * renderer.vk_conf.avail_threads);
 	imm->cmds           = ArenaAlloc(renderer.perm_arena, sizeof(VkCommandBuffer) * renderer.vk_conf.avail_threads);
diff --git a/src/renderer_vulkan.h b/src/renderer_vulkan.h
index fd8af47..584d618 100644
--- a/src/renderer_vulkan.h
+++ b/src/renderer_vulkan.h
@@ -236,6 +236,7 @@ typedef struct
 typedef struct { 
 	i32     graphics, transfer; 
 	VkQueue graphics_queue, transfer_queue;
+	b8 			single_queue;
 } DeviceQueues;
 
 typedef struct {
diff --git a/src/util.h b/src/util.h
index 5461f2c..b8be589 100644
--- a/src/util.h
+++ b/src/util.h
@@ -15,7 +15,7 @@
 
 #define Len(x) ((sizeof(x)/sizeof(0[x])) / ((size_t)(!(sizeof(x) % sizeof(0[x])))))
 #define MakeString(x) #x
-#define BitEq(var, bits) (((var) & (bits)) == bits)
+#define BitEq(var, bits) (((var) & (bits)) == (bits))
 #define AlignPow2(x, b)	(((x) + (b) - 1) & (~((b) - 1)))	
 #define IsPow2(x)				((x) != 0 && ((x) &((x) - 1)) == 0)
 #define PtrAdd(ptr, add) (((char *)ptr) + add)