rocr: Remove KMT usage from AMD ext

Use the core Driver in AMD's HSA extension API to make it
agnostic to the underlying OS and kernel-mode driver.


[ROCm/ROCR-Runtime commit: d3a4dc9687]

projects/rocr-runtime/runtime/hsa-runtime/core/driver/kfd/amd_kfd_driver.cpp

@@ -430,6 +430,28 @@ hsa_status_t KfdDriver::ReleaseShareableHandle(core::ShareableHandle &handle) {
   return HSA_STATUS_SUCCESS;
 }
 
+hsa_status_t KfdDriver::SPMAcquire(uint32_t preferred_node_id) const {
+  if (hsaKmtSPMAcquire(preferred_node_id) != HSAKMT_STATUS_SUCCESS) return HSA_STATUS_ERROR;
+
+  return HSA_STATUS_SUCCESS;
+}
+
+hsa_status_t KfdDriver::SPMRelease(uint32_t preferred_node_id) const {
+  if (hsaKmtSPMRelease(preferred_node_id) != HSAKMT_STATUS_SUCCESS) return HSA_STATUS_ERROR;
+
+  return HSA_STATUS_SUCCESS;
+}
+
+hsa_status_t KfdDriver::SPMSetDestBuffer(uint32_t preferred_node_id, uint32_t size_bytes,
+                                         uint32_t* timeout, uint32_t* size_copied,
+                                         void* dest_mem_addr, bool* is_spm_data_loss) const {
+  if (hsaKmtSPMSetDestBuffer(preferred_node_id, size_bytes, timeout, size_copied, dest_mem_addr,
+                             is_spm_data_loss) != HSAKMT_STATUS_SUCCESS)
+    return HSA_STATUS_ERROR;
+
+  return HSA_STATUS_SUCCESS;
+}
+
 void *KfdDriver::AllocateKfdMemory(const HsaMemFlags &flags, uint32_t node_id,
                                    size_t size) {
   void *mem = nullptr;

projects/rocr-runtime/runtime/hsa-runtime/core/driver/xdna/amd_xdna_driver.cpp

@@ -671,5 +671,22 @@ hsa_status_t XdnaDriver::SubmitCmdChain(hsa_amd_aie_ert_packet_t* first_pkt, uin
   return HSA_STATUS_SUCCESS;
 }
 
+hsa_status_t XdnaDriver::SPMAcquire(uint32_t preferred_node_id) const {
+  // AIE does not support streaming performance monitor.
+  return HSA_STATUS_ERROR_INVALID_AGENT;
+}
+
+hsa_status_t XdnaDriver::SPMRelease(uint32_t preferred_node_id) const {
+  // AIE does not support streaming performance monitor.
+  return HSA_STATUS_ERROR_INVALID_AGENT;
+};
+
+hsa_status_t XdnaDriver::SPMSetDestBuffer(uint32_t preferred_node_id, uint32_t size_bytes,
+                                          uint32_t* timeout, uint32_t* size_copied,
+                                          void* dest_mem_addr, bool* is_spm_data_loss) const {
+  // AIE does not support streaming performance monitor.
+  return HSA_STATUS_ERROR_INVALID_AGENT;
+}
+
 } // namespace AMD
 } // namespace rocr

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

@@ -301,7 +301,7 @@ class Agent : public Checked<0xF6BC25EB17E6F917> {
   __forceinline uint32_t node_id() const { return node_id_; }
 
   // @brief Returns the driver associated with this agent.
-  __forceinline Driver &driver() { return *driver_; }
+  __forceinline Driver& driver() const { return *driver_; }
 
   // @brief Getter for profiling_enabled_.
   __forceinline bool profiling_enabled() const { return profiling_enabled_; }

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

@@ -109,7 +109,13 @@ public:
                      size_t size) override;
   hsa_status_t ReleaseShareableHandle(core::ShareableHandle &handle) override;
 
-private:
+  hsa_status_t SPMAcquire(uint32_t preferred_node_id) const override;
+  hsa_status_t SPMRelease(uint32_t preferred_node_id) const override;
+  hsa_status_t SPMSetDestBuffer(uint32_t preferred_node_id, uint32_t size_bytes, uint32_t* timeout,
+                                uint32_t* size_copied, void* dest_mem_addr,
+                                bool* is_spm_data_loss) const override;
+
+ private:
   /// @brief Allocate agent accessible memory (system / local memory).
   static void *AllocateKfdMemory(const HsaMemFlags &flags, uint32_t node_id,
                                  size_t size);

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

@@ -174,10 +174,16 @@ public:
                      size_t size) override;
   hsa_status_t ReleaseShareableHandle(core::ShareableHandle &handle) override;
 
-  // @brief Submits num_pkts packets in a command chain to the XDNA driver
+  /// @brief Submits num_pkts packets in a command chain to the XDNA driver
   hsa_status_t SubmitCmdChain(hsa_amd_aie_ert_packet_t* first_pkt, uint32_t num_pkts,
                               uint32_t num_operands, uint32_t hw_ctx_handle);
 
+  hsa_status_t SPMAcquire(uint32_t preferred_node_id) const override;
+  hsa_status_t SPMRelease(uint32_t preferred_node_id) const override;
+  hsa_status_t SPMSetDestBuffer(uint32_t preferred_node_id, uint32_t size_bytes, uint32_t* timeout,
+                                uint32_t* size_copied, void* dest_mem_addr,
+                                bool* is_spm_data_loss) const override;
+
  private:
   hsa_status_t QueryDriverVersion();
   /// @brief Allocate device accesible heap space.

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

@@ -193,6 +193,24 @@ public:
   virtual hsa_status_t
   ReleaseShareableHandle(core::ShareableHandle &handle) = 0;
 
+  /// @brief Acquire a streaming performance monitor on an agent.
+  /// @param[in] preferred_node_id Node ID of the preferred agent.
+  virtual hsa_status_t SPMAcquire(uint32_t preferred_node_id) const = 0;
+  /// @brief Release a streaming performance monitor on an agent.
+  /// @param[in] preferred_node_id Node ID of the preferred agent.
+  virtual hsa_status_t SPMRelease(uint32_t preferred_node_id) const = 0;
+  /// @brief Setup the destination user-mode buffer for streaming performance monitor data.
+  /// @param[in] preferred_node_id Node ID of the preferred agent.
+  /// @param[in] size_bytes Size of the destination buffer in bytes.
+  /// @param[in, out] timeout Timeout in milliseconds.
+  /// @param[out] size_copied Size of data copied in bytes.
+  /// @param[in] dest_mem_addr Destination address for streaming performance data. Set to NULL to
+  /// stop copy on previous buffer.
+  /// @param[out] is_spm_data_loss Data was lost if true.
+  virtual hsa_status_t SPMSetDestBuffer(uint32_t preferred_node_id, uint32_t size_bytes,
+                                        uint32_t* timeout, uint32_t* size_copied,
+                                        void* dest_mem_addr, bool* is_spm_data_loss) const = 0;
+
   /// Unique identifier for supported kernel-mode drivers.
   const DriverType kernel_driver_type_;
 

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/hsa_ext_amd.cpp

@@ -1218,11 +1218,11 @@ hsa_status_t hsa_amd_spm_acquire(hsa_agent_t preferred_agent) {
   TRY;
   IS_OPEN();
   const core::Agent* agent = core::Agent::Convert(preferred_agent);
+  // Currently, the SPM API is only supported for GPU agents.
   if (agent == NULL || !agent->IsValid() || agent->device_type() != core::Agent::kAmdGpuDevice)
     return HSA_STATUS_ERROR_INVALID_AGENT;
 
-  if (hsaKmtSPMAcquire(agent->node_id()) != HSAKMT_STATUS_SUCCESS) return HSA_STATUS_ERROR;
+  return agent->driver().SPMAcquire(agent->node_id());
-  return HSA_STATUS_SUCCESS;
 
   CATCH;
 }
@@ -1232,12 +1232,11 @@ hsa_status_t hsa_amd_spm_release(hsa_agent_t preferred_agent) {
   IS_OPEN();
 
   const core::Agent* agent = core::Agent::Convert(preferred_agent);
+  // Currently, the SPM API is only supported for GPU agents.
   if (agent == NULL || !agent->IsValid() || agent->device_type() != core::Agent::kAmdGpuDevice)
     return HSA_STATUS_ERROR_INVALID_AGENT;
 
-  if (hsaKmtSPMRelease(agent->node_id()) != HSAKMT_STATUS_SUCCESS) return HSA_STATUS_ERROR;
+  return agent->driver().SPMRelease(agent->node_id());
-
-  return HSA_STATUS_SUCCESS;
 
   CATCH;
 }
@@ -1249,14 +1248,12 @@ hsa_status_t hsa_amd_spm_set_dest_buffer(hsa_agent_t preferred_agent, size_t siz
   IS_OPEN();
 
   const core::Agent* agent = core::Agent::Convert(preferred_agent);
+  // Currently, the SPM API is only supported for GPU agents.
   if (agent == NULL || !agent->IsValid() || agent->device_type() != core::Agent::kAmdGpuDevice)
     return HSA_STATUS_ERROR_INVALID_AGENT;
 
-  if (hsaKmtSPMSetDestBuffer(agent->node_id(), size_in_bytes, timeout, size_copied, dest,
+  return agent->driver().SPMSetDestBuffer(agent->node_id(), size_in_bytes, timeout, size_copied,
-                             is_data_loss) != HSAKMT_STATUS_SUCCESS)
+                                          dest, is_data_loss);
-    return HSA_STATUS_ERROR;
-
-  return HSA_STATUS_SUCCESS;
   CATCH;
 }