Add HSA_CU_MASK

New environment variable HSA_CU_MASK allows users to
specify a cu mask to every queue allocated from any
GPU.  hsa_amd_queue_cu_set_mask is restricted from
escaping this mask.

A new API hsa_amd_queue_cu_get_mask is added to query
the current cu mask.

Change-Id: I846c03a5faaca9b95067c31db84b59cc9fce2f03


[ROCm/ROCR-Runtime commit: 4455250be1]

projects/rocr-runtime/runtime/hsa-runtime/core/inc/amd_aql_queue.h

@@ -185,7 +185,16 @@ class AqlQueue : public core::Queue, private core::LocalSignal, public core::Doo
   /// @param cu_mask pointer to cu mask
   ///
   /// @return hsa_status_t
-  hsa_status_t SetCUMasking(const uint32_t num_cu_mask_count, const uint32_t* cu_mask) override;
+  hsa_status_t SetCUMasking(uint32_t num_cu_mask_count, const uint32_t* cu_mask) override;
+
+  /// @brief Get CU Masking
+  ///
+  /// @param num_cu_mask_count size of mask bit array
+  ///
+  /// @param cu_mask pointer to cu mask
+  ///
+  /// @return hsa_status_t
+  hsa_status_t GetCUMasking(uint32_t num_cu_mask_count, uint32_t* cu_mask) override;
 
   // @brief Submits a block of PM4 and waits until it has been executed.
   void ExecutePM4(uint32_t* cmd_data, size_t cmd_size_b) override;
@@ -277,6 +286,12 @@ class AqlQueue : public core::Queue, private core::LocalSignal, public core::Doo
   // Exception notification signal
   Signal* exception_signal_;
 
+  // CU mask lock
+  KernelMutex mask_lock_;
+
+  // Current CU mask
+  std::vector<uint32_t> cu_mask_;
+
   // Shared event used for queue errors
   static HsaEvent* queue_event_;
 

projects/rocr-runtime/runtime/hsa-runtime/core/inc/amd_gpu_agent.h

@@ -174,7 +174,7 @@ class GpuAgent : public GpuAgentInt {
   // id.
   // @param [in] node_props Node property.
   // @param [in] xnack_mode XNACK mode of device.
-  GpuAgent(HSAuint32 node, const HsaNodeProperties& node_props, bool xnack_mode);
+  GpuAgent(HSAuint32 node, const HsaNodeProperties& node_props, bool xnack_mode, uint32_t index);
 
   // @brief GPU agent destructor.
   ~GpuAgent();
@@ -322,6 +322,10 @@ class GpuAgent : public GpuAgentInt {
     return memory_max_frequency_;
   }
 
+  // @brief Order the device is surfaced in hsa_iterate_agents counting only
+  // GPU devices.
+  __forceinline uint32_t enumeration_index() const { return enum_index_; }
+
   void Trim() override;
 
  protected:
@@ -454,6 +458,9 @@ class GpuAgent : public GpuAgentInt {
   // @brief The GPU memory maximum frequency in MHz.
   uint32_t memory_max_frequency_;
 
+  // @brief Enumeration index
+  uint32_t enum_index_;
+
   // @brief HDP flush registers
   hsa_amd_hdp_flush_t HDP_flush_ = {nullptr, nullptr};
 

projects/rocr-runtime/runtime/hsa-runtime/core/inc/host_queue.h

@@ -144,8 +144,12 @@ class HostQueue : public Queue {
                        std::memory_order_release);
   }
 
-  hsa_status_t SetCUMasking(const uint32_t num_cu_mask_count, const uint32_t* cu_mask) override {
-    return HSA_STATUS_ERROR;
+  hsa_status_t SetCUMasking(uint32_t num_cu_mask_count, const uint32_t* cu_mask) override {
+    return HSA_STATUS_ERROR_INVALID_QUEUE;
+  }
+
+  hsa_status_t GetCUMasking(uint32_t num_cu_mask_count, uint32_t* cu_mask) override {
+    return HSA_STATUS_ERROR_INVALID_QUEUE;
   }
 
   void ExecutePM4(uint32_t* cmd_data, size_t cmd_size_b) override {

projects/rocr-runtime/runtime/hsa-runtime/core/inc/hsa_ext_amd_impl.h

@@ -113,6 +113,10 @@ hsa_status_t hsa_amd_queue_cu_set_mask(const hsa_queue_t* queue,
                                                uint32_t num_cu_mask_count,
                                                const uint32_t* cu_mask);
 
+// Mirrors Amd Extension Apis
+hsa_status_t HSA_API hsa_amd_queue_cu_get_mask(const hsa_queue_t* queue, uint32_t num_cu_mask_count,
+                                               uint32_t* cu_mask);
+
 // Mirrors Amd Extension Apis
 hsa_status_t
     hsa_amd_memory_pool_get_info(hsa_amd_memory_pool_t memory_pool,

projects/rocr-runtime/runtime/hsa-runtime/core/inc/intercept_queue.h

@@ -114,9 +114,12 @@ class QueueWrapper : public Queue {
   uint64_t AddWriteIndexRelease(uint64_t value) override {
     return wrapped->AddWriteIndexRelease(value);
   }
-  hsa_status_t SetCUMasking(const uint32_t num_cu_mask_count, const uint32_t* cu_mask) override {
+  hsa_status_t SetCUMasking(uint32_t num_cu_mask_count, const uint32_t* cu_mask) override {
     return wrapped->SetCUMasking(num_cu_mask_count, cu_mask);
   }
+  hsa_status_t GetCUMasking(uint32_t num_cu_mask_count, uint32_t* cu_mask) override {
+    return wrapped->GetCUMasking(num_cu_mask_count, cu_mask);
+  }
   void ExecutePM4(uint32_t* cmd_data, size_t cmd_size_b) override {
     wrapped->ExecutePM4(cmd_data, cmd_size_b);
   }

projects/rocr-runtime/runtime/hsa-runtime/core/inc/queue.h

@@ -295,8 +295,16 @@ class Queue : public Checked<0xFA3906A679F9DB49>, private LocalQueue {
   /// @param cu_mask pointer to cu mask
   ///
   /// @return hsa_status_t
-  virtual hsa_status_t SetCUMasking(const uint32_t num_cu_mask_count,
-                                    const uint32_t* cu_mask) = 0;
+  virtual hsa_status_t SetCUMasking(uint32_t num_cu_mask_count, const uint32_t* cu_mask) = 0;
+
+  /// @brief Get CU Masking
+  ///
+  /// @param num_cu_mask_count size of mask bit array
+  ///
+  /// @param cu_mask pointer to cu mask
+  ///
+  /// @return hsa_status_t
+  virtual hsa_status_t GetCUMasking(uint32_t num_cu_mask_count, uint32_t* cu_mask) = 0;
 
   // @brief Submits a block of PM4 and waits until it has been executed.
   virtual void ExecutePM4(uint32_t* cmd_data, size_t cmd_size_b) = 0;