Add queue serialization

ARM-software · Jan 20, 2025 · b883898 · b883898
1 parent 939fc52
commit b883898
Show file tree

Hide file tree

Showing 4 changed files with 219 additions and 4 deletions.
diff --git a/layer_gpu_support/source/device.cpp b/layer_gpu_support/source/device.cpp
@@ -93,4 +93,25 @@ Device::Device(
     device(_device)
 {
     initDriverDeviceDispatchTable(device, nlayerGetProcAddress, driver);
+
+    VkSemaphoreTypeCreateInfo timelineCreateInfo {
+        .sType = VK_STRUCTURE_TYPE_SEMAPHORE_TYPE_CREATE_INFO,
+        .pNext = nullptr,
+        .semaphoreType = VK_SEMAPHORE_TYPE_TIMELINE,
+        .initialValue = queueSerializationTimelineSemCount
+    };
+
+    VkSemaphoreCreateInfo createInfo {
+        .sType = VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO,
+        .pNext = &timelineCreateInfo,
+        .flags = 0
+    };
+
+    auto result = driver.vkCreateSemaphore(
+        device, &createInfo, nullptr, &queueSerializationTimelineSem);
+    if (result != VK_SUCCESS)
+    {
+        LAYER_ERR("Failed vkCreateSemaphore() for queue serialization");
+        queueSerializationTimelineSem = nullptr;
+    }
 }
diff --git a/layer_gpu_support/source/device.hpp b/layer_gpu_support/source/device.hpp
@@ -147,6 +147,16 @@ class Device
      */
     static const std::vector<std::string> extraExtensions;
 
+    /**
+     * @brief The timeline sem use for queue serialization.
+     */
+    VkSemaphore queueSerializationTimelineSem { nullptr };
+
+    /**
+     * @brief The current timeline sem target value the next use waits for.
+     */
+    uint64_t queueSerializationTimelineSemCount { 0 };
+
 private:
     /**
      * @brief The physical device this device is created with.

diff --git a/layer_gpu_support/source/layer_device_functions_queue.cpp b/layer_gpu_support/source/layer_device_functions_queue.cpp
@@ -44,9 +44,65 @@ VKAPI_ATTR VkResult VKAPI_CALL layer_vkQueueSubmit<user_tag>(
     std::unique_lock<std::mutex> lock { g_vulkanLock };
     auto* layer = Device::retrieve(queue);
 
+    // Serialize in the order submits are called
+    // TODO: This assumes a forward progress guarantee which is no longer
+    // guaranteed if the user is using timeline semaphores for syncs
+
+    const uint64_t waitValue = layer->queueSerializationTimelineSemCount;
+    layer->queueSerializationTimelineSemCount++;
+    const uint64_t signalValue = layer->queueSerializationTimelineSemCount;
+
+    VkPipelineStageFlags waitMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
+
+    VkTimelineSemaphoreSubmitInfo timelineInfo {
+        .sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO,
+        .pNext = nullptr,
+        .waitSemaphoreValueCount = 1,
+        .pWaitSemaphoreValues = &waitValue,
+        .signalSemaphoreValueCount = 1,
+        .pSignalSemaphoreValues = &signalValue
+    };
+
+    VkSubmitInfo submitInfoPre {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+        .pNext = &timelineInfo,
+        .waitSemaphoreCount = 1,
+        .pWaitSemaphores = &(layer->queueSerializationTimelineSem),
+        .pWaitDstStageMask = &waitMask,
+        .commandBufferCount = 0,
+        .pCommandBuffers = 0,
+        .signalSemaphoreCount = 0,
+        .pSignalSemaphores = nullptr
+    };
+
+    VkSubmitInfo submitInfoPost {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
+        .pNext = &timelineInfo,
+        .waitSemaphoreCount = 0,
+        .pWaitSemaphores = 0,
+        .pWaitDstStageMask = 0,
+        .commandBufferCount = 0,
+        .pCommandBuffers = 0,
+        .signalSemaphoreCount = 1,
+        .pSignalSemaphores = &(layer->queueSerializationTimelineSem)
+    };
+
     // Release the lock to call into the driver
     lock.unlock();
-    return layer->driver.vkQueueSubmit(queue, submitCount, pSubmits, fence);
+
+    if (layer->instance->config.serialize_queue())
+    {
+        layer->driver.vkQueueSubmit(queue, 1, &submitInfoPre, VK_NULL_HANDLE);
+    }
+
+    auto result = layer->driver.vkQueueSubmit(queue, submitCount, pSubmits, fence);
+
+    if (layer->instance->config.serialize_queue())
+    {
+        layer->driver.vkQueueSubmit(queue, 1, &submitInfoPost, VK_NULL_HANDLE);
+    }
+
+    return result;
 }
 
 /* See Vulkan API for documentation. */
@@ -63,9 +119,73 @@ VKAPI_ATTR VkResult VKAPI_CALL layer_vkQueueSubmit2<user_tag>(
     std::unique_lock<std::mutex> lock { g_vulkanLock };
     auto* layer = Device::retrieve(queue);
 
+    // Serialize in the order submits are called
+    // TODO: This assumes a forward progress guarantee which is no longer
+    // guaranteed if the user is using timeline semaphores for syncs
+
+    const uint64_t waitValue = layer->queueSerializationTimelineSemCount;
+    layer->queueSerializationTimelineSemCount++;
+    const uint64_t signalValue = layer->queueSerializationTimelineSemCount;
+
+    VkSemaphoreSubmitInfo timelineInfoPre {
+        .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO,
+        .pNext = nullptr,
+        .semaphore = layer->queueSerializationTimelineSem,
+        .value = waitValue,
+        .stageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+        .deviceIndex = 0
+    };
+
+    VkSemaphoreSubmitInfo timelineInfoPost {
+        .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO,
+        .pNext = nullptr,
+        .semaphore = layer->queueSerializationTimelineSem,
+        .value = signalValue,
+        .stageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+        .deviceIndex = 0
+    };
+
+    VkSubmitInfo2 submitInfoPre {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2,
+        .pNext = nullptr,
+        .flags = 0,
+        .waitSemaphoreInfoCount = 1,
+        .pWaitSemaphoreInfos = &timelineInfoPre,
+        .commandBufferInfoCount = 0,
+        .pCommandBufferInfos = nullptr,
+        .signalSemaphoreInfoCount = 0,
+        .pSignalSemaphoreInfos = nullptr
+    };
+
+    VkSubmitInfo2 submitInfoPost {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2,
+        .pNext = nullptr,
+        .flags = 0,
+        .waitSemaphoreInfoCount = 0,
+        .pWaitSemaphoreInfos = nullptr,
+        .commandBufferInfoCount = 0,
+        .pCommandBufferInfos = nullptr,
+        .signalSemaphoreInfoCount = 1,
+        .pSignalSemaphoreInfos = &timelineInfoPost
+    };
+
     // Release the lock to call into the driver
     lock.unlock();
-    return layer->driver.vkQueueSubmit2(queue, submitCount, pSubmits, fence);
+
+    if (layer->instance->config.serialize_queue())
+    {
+        layer->driver.vkQueueSubmit2(queue, 1, &submitInfoPre, VK_NULL_HANDLE);
+    }
+
+    auto result = layer->driver.vkQueueSubmit2(queue, submitCount, pSubmits, fence);
+
+    if (layer->instance->config.serialize_queue())
+    {
+        layer->driver.vkQueueSubmit2(queue, 1, &submitInfoPost, VK_NULL_HANDLE);
+    }
+
+    return result;
+
 }
 
 /* See Vulkan API for documentation. */
@@ -82,7 +202,70 @@ VKAPI_ATTR VkResult VKAPI_CALL layer_vkQueueSubmit2KHR<user_tag>(
     std::unique_lock<std::mutex> lock { g_vulkanLock };
     auto* layer = Device::retrieve(queue);
 
+    // Serialize in the order submits are called
+    // TODO: This assumes a forward progress guarantee which is no longer
+    // guaranteed if the user is using timeline semaphores for syncs
+
+    const uint64_t waitValue = layer->queueSerializationTimelineSemCount;
+    layer->queueSerializationTimelineSemCount++;
+    const uint64_t signalValue = layer->queueSerializationTimelineSemCount;
+
+    VkSemaphoreSubmitInfo timelineInfoPre {
+        .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO,
+        .pNext = nullptr,
+        .semaphore = layer->queueSerializationTimelineSem,
+        .value = waitValue,
+        .stageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+        .deviceIndex = 0
+    };
+
+    VkSemaphoreSubmitInfo timelineInfoPost {
+        .sType = VK_STRUCTURE_TYPE_SEMAPHORE_SUBMIT_INFO,
+        .pNext = nullptr,
+        .semaphore = layer->queueSerializationTimelineSem,
+        .value = signalValue,
+        .stageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+        .deviceIndex = 0
+    };
+
+    VkSubmitInfo2 submitInfoPre {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2,
+        .pNext = nullptr,
+        .flags = 0,
+        .waitSemaphoreInfoCount = 1,
+        .pWaitSemaphoreInfos = &timelineInfoPre,
+        .commandBufferInfoCount = 0,
+        .pCommandBufferInfos = nullptr,
+        .signalSemaphoreInfoCount = 0,
+        .pSignalSemaphoreInfos = nullptr
+    };
+
+    VkSubmitInfo2 submitInfoPost {
+        .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO_2,
+        .pNext = nullptr,
+        .flags = 0,
+        .waitSemaphoreInfoCount = 0,
+        .pWaitSemaphoreInfos = nullptr,
+        .commandBufferInfoCount = 0,
+        .pCommandBufferInfos = nullptr,
+        .signalSemaphoreInfoCount = 1,
+        .pSignalSemaphoreInfos = &timelineInfoPost
+    };
+
     // Release the lock to call into the driver
     lock.unlock();
-    return layer->driver.vkQueueSubmit2KHR(queue, submitCount, pSubmits, fence);
+
+    if (layer->instance->config.serialize_queue())
+    {
+        layer->driver.vkQueueSubmit2KHR(queue, 1, &submitInfoPre, VK_NULL_HANDLE);
+    }
+
+    auto result = layer->driver.vkQueueSubmit2KHR(queue, submitCount, pSubmits, fence);
+
+    if (layer->instance->config.serialize_queue())
+    {
+        layer->driver.vkQueueSubmit2KHR(queue, 1, &submitInfoPost, VK_NULL_HANDLE);
+    }
+
+    return result;
 }
diff --git a/source_common/framework/manual_functions.cpp b/source_common/framework/manual_functions.cpp
@@ -909,9 +909,10 @@ VKAPI_ATTR VkResult VKAPI_CALL layer_vkCreateDevice_default(
         return res;
     }
 
+    auto device = std::make_unique<Device>(layer, physicalDevice, *pDevice, fpGetDeviceProcAddr);
+
     // Retake the lock to access layer-wide global store
     lock.lock();
-    auto device = std::make_unique<Device>(layer, physicalDevice, *pDevice, fpGetDeviceProcAddr);
     Device::store(*pDevice, std::move(device));
 
     return VK_SUCCESS;