[Runtime][Vulkan] Add RGP support to TVM for vulkan device

RGP(Raedon GPU Profiler) is a tool used to analyze the applications
run on AMD GPU. RGP captures the data based on VKPresent and provides
the hardware specific information. Allowing the developer to optimize
the application. To add RGP support to TVM, debug labels "AmdFrameBegin"
and "AmdFrameEnd" need to be inserted into the vulkan queue.These Labels
helps the RGP tool to understand the start|end of frame when no present
is available. Thus enabling the RGP tool to capture and analyze the data.

At runtime, set the envirnoment variable "TVM_USE_AMD_RGP=1" to start
inserting the Debug Labels into the vulkan queue.

Signed-off-by: Wilkin Chau <Wing-Ki.ChauWilkin@amd.com>
Signed-off-by: Anurag Kumar Vulisha <AnuragKumar.Vulisha@amd.com>

Author: avulisha

src/runtime/vulkan/vulkan_amdrgp.cc

@@ -0,0 +1,53 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+#include "vulkan_device.h"
+
+namespace tvm {
+namespace runtime {
+namespace vulkan {
+
+VulkanStreamProfiler::VulkanStreamProfiler(const VulkanDevice* device)
+    : device_(device), curr_state_(READY), available_(device->UseDebugUtilsLabel()) {}
+
+void AmdRgpProfiler::capture() {
+  if (!available_) {
+    return;
+  }
+
+  // Trigger RGP capture by using dummy present and switch state from READY to RUNNING
+  if (curr_state_ == READY) {
+    VkDebugUtilsLabelEXT frame_end_label = {
+        VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT, NULL, "AmdFrameEnd", {0.0f, 0.0f, 0.0f, 0.0f}};
+    device_->queue_insert_debug_utils_label_functions->vkQueueInsertDebugUtilsLabelEXT(
+        device_->Queue(), &frame_end_label);
+
+    VkDebugUtilsLabelEXT frame_begin_label = {
+        VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT, NULL, "AmdFrameBegin", {0.0f, 0.0f, 0.0f, 0.0f}};
+    device_->queue_insert_debug_utils_label_functions->vkQueueInsertDebugUtilsLabelEXT(
+        device_->Queue(), &frame_begin_label);
+
+    // Set state as RUNNING
+    curr_state_ = RUNNING;
+  }
+}
+
+}  // namespace vulkan
+}  // namespace runtime
+}  // namespace tvm

src/runtime/vulkan/vulkan_amdrgp.h

@@ -0,0 +1,63 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+#ifndef TVM_RUNTIME_VULKAN_VULKAN_AMDRGP_H_
+#define TVM_RUNTIME_VULKAN_VULKAN_AMDRGP_H_
+
+namespace tvm {
+namespace runtime {
+namespace vulkan {
+
+class VulkanDevice;
+
+class VulkanStreamProfiler {
+ public:
+  enum state { READY = 0, RUNNING, RESET };
+
+  explicit VulkanStreamProfiler(const VulkanDevice* device);
+
+  virtual ~VulkanStreamProfiler() {}
+
+  virtual void reset() { curr_state_ = RESET; }
+
+  virtual void ready() {
+    if (curr_state_ == RESET) {
+      curr_state_ = READY;
+    }
+  }
+
+  virtual void capture() = 0;
+
+ protected:
+  const VulkanDevice* device_;
+  state curr_state_;
+  bool available_;
+};
+
+class AmdRgpProfiler : public VulkanStreamProfiler {
+ public:
+  explicit AmdRgpProfiler(const VulkanDevice* device) : VulkanStreamProfiler(device) {}
+
+  void capture();
+};
+
+}  // namespace vulkan
+}  // namespace runtime
+}  // namespace tvm
+#endif  // TVM_RUNTIME_VULKAN_VULKAN_AMDRGP_H_

src/runtime/vulkan/vulkan_device.cc

@@ -228,6 +228,12 @@ VulkanGetBufferMemoryRequirements2Functions::VulkanGetBufferMemoryRequirements2F
       vkGetDeviceProcAddr(device, "vkGetBufferMemoryRequirements2KHR"));
 }
 
+VulkanQueueInsertDebugUtilsLabelFunctions::VulkanQueueInsertDebugUtilsLabelFunctions(
+    VkInstance instance) {
+  vkQueueInsertDebugUtilsLabelEXT = (PFN_vkQueueInsertDebugUtilsLabelEXT)ICHECK_NOTNULL(
+      vkGetInstanceProcAddr(instance, "vkQueueInsertDebugUtilsLabelEXT"));
+}
+
 VulkanDevice::VulkanDevice(const VulkanInstance& instance, VkPhysicalDevice phy_device)
     : physical_device_(phy_device) {
   queue_family_index = SelectComputeQueueFamily();
@@ -325,6 +331,11 @@ VulkanDevice::VulkanDevice(const VulkanInstance& instance, VkPhysicalDevice phy_
     get_buffer_memory_requirements_2_functions =
         std::make_unique<VulkanGetBufferMemoryRequirements2Functions>(device_);
   }
+
+  if (instance.HasExtension("VK_EXT_debug_utils")) {
+    queue_insert_debug_utils_label_functions =
+        std::make_unique<VulkanQueueInsertDebugUtilsLabelFunctions>(instance);
+  }
 }
 
 VulkanDevice::~VulkanDevice() {
@@ -363,6 +374,8 @@ void VulkanDevice::do_swap(VulkanDevice&& other) {
   std::swap(descriptor_template_khr_functions, other.descriptor_template_khr_functions);
   std::swap(get_buffer_memory_requirements_2_functions,
             other.get_buffer_memory_requirements_2_functions);
+  std::swap(queue_insert_debug_utils_label_functions,
+            other.queue_insert_debug_utils_label_functions);
   std::swap(compute_mtype_index, other.compute_mtype_index);
   std::swap(queue, other.queue);
   std::swap(queue_family_index, other.queue_family_index);

src/runtime/vulkan/vulkan_device.h

@@ -57,6 +57,12 @@ struct VulkanGetBufferMemoryRequirements2Functions {
   PFN_vkGetBufferMemoryRequirements2KHR vkGetBufferMemoryRequirements2KHR{nullptr};
 };
 
+struct VulkanQueueInsertDebugUtilsLabelFunctions {
+  explicit VulkanQueueInsertDebugUtilsLabelFunctions(VkInstance instance);
+
+  PFN_vkQueueInsertDebugUtilsLabelEXT vkQueueInsertDebugUtilsLabelEXT{nullptr};
+};
+
 /*!
  * \brief Stores the capabilities/limits queried from the physical device.
  *
@@ -212,6 +218,8 @@ class VulkanDevice {
   std::unique_ptr<VulkanDescriptorTemplateKHRFunctions> descriptor_template_khr_functions{nullptr};
   std::unique_ptr<VulkanGetBufferMemoryRequirements2Functions>
       get_buffer_memory_requirements_2_functions{nullptr};
+  std::unique_ptr<VulkanQueueInsertDebugUtilsLabelFunctions>
+      queue_insert_debug_utils_label_functions{nullptr};
   // Memory type index for compute
   uint32_t compute_mtype_index{0};
 
@@ -220,6 +228,10 @@ class VulkanDevice {
 
   bool UseImmediate() const { return descriptor_template_khr_functions != nullptr; }
 
+  bool UseDebugUtilsLabel() const { return queue_insert_debug_utils_label_functions != nullptr; }
+
+  VkQueue Queue() const { return queue; }
+
  private:
   /*! \brief Helper function for move assignment/construction
    *

src/runtime/vulkan/vulkan_device_api.cc

@@ -367,6 +367,7 @@ void VulkanDeviceAPI::CopyDataFromTo(const void* from, size_t from_offset, void*
                       &copy_info);
     });
     stream.Synchronize();
+    stream.ProfilerReset();
     if (!device.coherent_staging) {
       VkMappedMemoryRange mrange;
       mrange.sType = VK_STRUCTURE_TYPE_MAPPED_MEMORY_RANGE;
@@ -413,6 +414,8 @@ void VulkanDeviceAPI::CopyDataFromTo(const void* from, size_t from_offset, void*
       vkCmdCopyBuffer(state->cmd_buffer_, staging_buffer.vk_buf.buffer, to_buf->buffer, 1,
                       &copy_info);
     });
+
+    stream.ProfilerReady();
     // TODO(tulloch): should we instead make the staging buffer a property of the
     // Stream? This would allow us to elide synchronizations here.
     stream.Synchronize();

src/runtime/vulkan/vulkan_instance.cc

@@ -59,6 +59,14 @@ VulkanInstance::VulkanInstance() {
     std::vector<const char*> required_extensions{};
     std::vector<const char*> optional_extensions{"VK_KHR_get_physical_device_properties2"};
 
+    // Check if RGP support is needed. If needed, enable VK_EXT_debug_utils extension for
+    // inserting debug labels into the queue.
+    const char* val = getenv("TVM_USE_AMD_RGP");
+    if (val != nullptr && atoi(val) == 1) {
+      LOG(INFO) << "Push VK_EXT_debug_utils";
+      required_extensions.push_back("VK_EXT_debug_utils");
+    }
+
     uint32_t inst_extension_prop_count;
     VULKAN_CALL(
         vkEnumerateInstanceExtensionProperties(nullptr, &inst_extension_prop_count, nullptr));

src/runtime/vulkan/vulkan_stream.cc

@@ -55,11 +55,15 @@ VulkanStream::VulkanStream(const VulkanDevice* device)
   cb_begin.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
   cb_begin.pInheritanceInfo = 0;
   VULKAN_CALL(vkBeginCommandBuffer(state_->cmd_buffer_, &cb_begin));
+
+  profiler_ = new AmdRgpProfiler(device_);
 }
 
 VulkanStream::~VulkanStream() {
   vkDestroyFence(*device_, state_->fence_, nullptr);
   vkDestroyCommandPool(*device_, cmd_pool_, nullptr);
+
+  delete (profiler_);
 }
 
 void VulkanStream::Launch(const std::function<void(VulkanStreamState*)>& kernel) {
@@ -132,6 +136,8 @@ void VulkanStream::Synchronize() {
   cb_submit.signalSemaphoreCount = 0;
   cb_submit.pSignalSemaphores = nullptr;
 
+  profiler_->capture();
+
   device_->QueueSubmit(cb_submit, state_->fence_);
 
   uint64_t timeout = 1UL << 30UL;

src/runtime/vulkan/vulkan_stream.h

@@ -25,6 +25,7 @@
 #include <unordered_map>
 #include <vector>
 
+#include "vulkan_amdrgp.h"
 #include "vulkan_common.h"
 
 namespace tvm {
@@ -99,6 +100,20 @@ class VulkanStream {
                       const std::function<void(VulkanStreamState*)>& deferred_kernel,
                       const VulkanStreamToken& deferred_token);
 
+  // reset profiler state
+  void ProfilerReset() {
+    if (profiler_) {
+      profiler_->reset();
+    }
+  }
+
+  // set profiler to READY state after reset
+  void ProfilerReady() {
+    if (profiler_) {
+      profiler_->ready();
+    }
+  }
+
   // Synchronize the current stream `state_` with respect to the host.
   void Synchronize();
 
@@ -110,6 +125,7 @@ class VulkanStream {
   std::unordered_map<VkDescriptorSet, std::vector<VulkanStreamToken>> deferred_tokens_;
   std::vector<std::function<void(VulkanStreamState*)>> deferred_kernels_;
   VkCommandPool cmd_pool_;
+  VulkanStreamProfiler* profiler_;
 };
 
 }  // namespace vulkan

src/runtime/vulkan/vulkan_wrapped_func.cc

@@ -164,6 +164,15 @@ void VulkanWrappedFunc::operator()(TVMArgs args, TVMRetValue* rv,
     deferred_token.buffers_[i] = descriptor_buffers[i].buffer;
   }
   device.ThreadLocalStream().LaunchDeferred(deferred_initializer, deferred_kernel, deferred_token);
+
+  if (device.UseDebugUtilsLabel()) {
+    VkDebugUtilsLabelEXT dispatch_label = {VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT,
+                                           NULL,
+                                           func_name_.c_str(),
+                                           {0.0f, 0.0f, 0.0f, 0.0f}};
+    device.queue_insert_debug_utils_label_functions->vkQueueInsertDebugUtilsLabelEXT(
+        device.Queue(), &dispatch_label);
+  }
 }
 
 VulkanModuleNode::~VulkanModuleNode() {