radv: use less AMDGPU contexts by creating only one per queue priority

It should be more efficient. Suggested by Bas.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8878>

diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index fe4cba2..ff4e6b6 100644 (file)
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -2376,21 +2376,15 @@ radv_queue_init(struct radv_device *device, struct radv_queue *queue,
        queue->queue_idx = idx;
        queue->priority = radv_get_queue_global_priority(global_priority);
        queue->flags = flags;
+       queue->hw_ctx = device->hw_ctx[queue->priority];
 
        vk_object_base_init(&device->vk, &queue->base, VK_OBJECT_TYPE_QUEUE);
 
-       VkResult result = device->ws->ctx_create(device->ws, queue->priority, &queue->hw_ctx);
-       if (result != VK_SUCCESS) {
-               vk_object_base_finish(&queue->base);
-               return vk_error(device->instance, result);
-       }
-
        list_inithead(&queue->pending_submissions);
        mtx_init(&queue->pending_mutex, mtx_plain);
 
        mtx_init(&queue->thread_mutex, mtx_plain);
        if (u_cnd_monotonic_init(&queue->thread_cond)) {
-               device->ws->ctx_destroy(queue->hw_ctx);
                vk_object_base_finish(&queue->base);
                return vk_error(device->instance, VK_ERROR_INITIALIZATION_FAILED);
        }
@@ -2415,8 +2409,6 @@ radv_queue_finish(struct radv_queue *queue)
 
                mtx_destroy(&queue->pending_mutex);
                mtx_destroy(&queue->thread_mutex);
-
-               queue->device->ws->ctx_destroy(queue->hw_ctx);
        }
 
        if (queue->initial_full_flush_preamble_cs)
@@ -2747,6 +2739,22 @@ VkResult radv_CreateDevice(
 
        radv_bo_list_init(&device->bo_list);
 
+       /* Create one context per queue priority. */
+       for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
+               const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
+               const VkDeviceQueueGlobalPriorityCreateInfoEXT *global_priority =
+                       vk_find_struct_const(queue_create->pNext, DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_EXT);
+               enum radeon_ctx_priority priority = radv_get_queue_global_priority(global_priority);
+
+               if (device->hw_ctx[priority])
+                       continue;
+
+               result = device->ws->ctx_create(device->ws, priority,
+                                               &device->hw_ctx[priority]);
+               if (result != VK_SUCCESS)
+                       goto fail;
+       }
+
        for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
                const VkDeviceQueueCreateInfo *queue_create = &pCreateInfo->pQueueCreateInfos[i];
                uint32_t qfi = queue_create->queueFamilyIndex;
@@ -2980,6 +2988,11 @@ fail:
                        vk_free(&device->vk.alloc, device->queues[i]);
        }
 
+       for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++) {
+               if (device->hw_ctx[i])
+                       device->ws->ctx_destroy(device->hw_ctx[i]);
+       }
+
        vk_device_finish(&device->vk);
        vk_free(&device->vk.alloc, device);
        return result;
@@ -3010,6 +3023,12 @@ void radv_DestroyDevice(
                if (device->empty_cs[i])
                        device->ws->cs_destroy(device->empty_cs[i]);
        }
+
+       for (unsigned i = 0; i < RADV_NUM_HW_CTX; i++) {
+               if (device->hw_ctx[i])
+                       device->ws->ctx_destroy(device->hw_ctx[i]);
+       }
+
        radv_device_finish_meta(device);
 
        VkPipelineCache pc = radv_pipeline_cache_to_handle(device->mem_cache);

diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
index 56c281b..34df1c8 100644 (file)
--- a/src/amd/vulkan/radv_private.h
+++ b/src/amd/vulkan/radv_private.h
@@ -663,6 +663,8 @@ struct radv_meta_state {
 
 #define RADV_MAX_QUEUE_FAMILIES 3
 
+#define RADV_NUM_HW_CTX (RADEON_CTX_PRIORITY_REALTIME + 1)
+
 struct radv_deferred_queue_submission;
 
 enum ring_type radv_queue_family_to_ring(int f);
@@ -742,6 +744,7 @@ struct radv_device {
        struct radv_instance *                       instance;
        struct radeon_winsys *ws;
 
+       struct radeon_winsys_ctx *hw_ctx[RADV_NUM_HW_CTX];
        struct radv_meta_state                       meta_state;
 
        struct radv_queue *queues[RADV_MAX_QUEUE_FAMILIES];