SWDEV-520384 - Improve postLoad latency during program load (#758)

Change-Id: I3c5a9e148efbd845c8fa3e4b51d9653664e9c499

rocclr/device/device.cpp

@@ -81,6 +81,9 @@ bool VirtualDevice::ActiveWait() const {
   return device_().ActiveWait();
 }
 
+#if defined(USE_COMGR_LIBRARY)
+extern amd_comgr_status_t getMetaBuf(const amd_comgr_metadata_node_t meta, std::string* str);
+#endif
 }
 
 static_assert(static_cast<uint32_t>(device::Memory::MemAccess::kMemAccessNone)
@@ -328,6 +331,61 @@ const Isa* Isa::end() {
   return supportedIsas().second;
 }
 
+#if defined(USE_COMGR_LIBRARY)
+void Isa::setAvailableSgprVgprCached() const {
+  std::call_once(setSgprVgprFlag, [this]() {
+    std::string buf;
+    amd_comgr_metadata_node_t isaMeta;
+    amd_comgr_metadata_node_t sgprMeta;
+    amd_comgr_metadata_node_t vgprMeta;
+    bool hasIsaMeta = false;
+    bool hasSgprMeta = false;
+    bool hasVgprMeta = false;
+
+    amd_comgr_status_t status = amd::Comgr::get_isa_metadata(isaName().c_str(), &isaMeta);
+
+    if (status == AMD_COMGR_STATUS_SUCCESS) {
+      hasIsaMeta = true;
+      status = amd::Comgr::metadata_lookup(isaMeta, "AddressableNumSGPRs", &sgprMeta);
+    }
+
+    if (status == AMD_COMGR_STATUS_SUCCESS) {
+      hasSgprMeta = true;
+      status = amd::device::getMetaBuf(sgprMeta, &buf);
+    }
+
+    sgprPerWavefront_ = (status == AMD_COMGR_STATUS_SUCCESS) ? atoi(buf.c_str()) : 0;
+
+    if (status == AMD_COMGR_STATUS_SUCCESS) {
+      status = amd::Comgr::metadata_lookup(isaMeta, "AddressableNumVGPRs", &vgprMeta);
+    }
+
+    if (status == AMD_COMGR_STATUS_SUCCESS) {
+      hasVgprMeta = true;
+      status = amd::device::getMetaBuf(vgprMeta, &buf);
+    }
+
+    vgprPerWavefront_ = (status == AMD_COMGR_STATUS_SUCCESS) ? atoi(buf.c_str()) : 0;
+
+    if (hasVgprMeta) {
+      amd::Comgr::destroy_metadata(vgprMeta);
+    }
+
+    if (hasSgprMeta) {
+      amd::Comgr::destroy_metadata(sgprMeta);
+    }
+
+    if (hasIsaMeta) {
+      amd::Comgr::destroy_metadata(isaMeta);
+    }
+
+    if (status != AMD_COMGR_STATUS_SUCCESS) {
+      DevLogPrintfError("Failed to set SGPR/VGPR for ISA: %s", isaName().c_str());
+    }
+  });
+}
+#endif
+
 std::vector<Device*>* Device::devices_ = nullptr;
 AppProfile Device::appProfile_;
 

rocclr/device/device.hpp

@@ -1565,6 +1565,20 @@ class Isa {
     return localMemBanks_;
   }
 
+#if defined(USE_COMGR_LIBRARY)
+  /// @returns This Isa's available sgprs per wavefront
+  size_t sgprPerWavefront() const {
+    setAvailableSgprVgprCached();
+    return sgprPerWavefront_;
+  }
+
+  /// @returns This Isa's available vgprs per wavefront
+  size_t vgprPerWavefront() const {
+    setAvailableSgprVgprCached();
+    return vgprPerWavefront_;
+  }
+#endif
+
   /// @returns True if @p codeObjectIsa and @p agentIsa are compatible,
   /// false otherwise.
   static bool isCompatible(const Isa &codeObjectIsa, const Isa &agentIsa);
@@ -1604,11 +1618,19 @@ class Isa {
         simdInstructionWidth_(simdInstructionWidth),
         memChannelBankWidth_(memChannelBankWidth),
         localMemSizePerCU_(localMemSizePerCU),
-        localMemBanks_(localMemBanks) {}
+        localMemBanks_(localMemBanks),
+        sgprPerWavefront_(0),
+        vgprPerWavefront_(0) {}
 
   // @brief Returns the begin and end iterators for the suppported ISAs.
   static std::pair<const Isa*, const Isa*> supportedIsas();
 
+#if defined(USE_COMGR_LIBRARY)
+  // @brief Populate this Isa's available sgprs/vgprs per wavefront from comgr.
+  // Only called once per Isa.
+  void setAvailableSgprVgprCached() const;
+#endif
+
   // @brief Isa's target ID name. Used for LLVM COde Object Manager
   // compilations.
   const char* targetId_;
@@ -1631,6 +1653,10 @@ class Isa {
   uint32_t memChannelBankWidth_;   //!< Memory channel bank width.
   uint32_t localMemSizePerCU_;     //!< Local memory size per CU.
   uint32_t localMemBanks_;         //!< Number of banks of local memory.
+
+  mutable size_t sgprPerWavefront_;        //!< Number of sgpr per wavefront.
+  mutable size_t vgprPerWavefront_;        //!< Number of vgpr per wavefront.
+  mutable std::once_flag setSgprVgprFlag;  //!< Once flag for sgpr and vgpr retrieval.
 }; // class Isa
 
 /*! \addtogroup Runtime

rocclr/device/devkernel.cpp

@@ -1155,56 +1155,6 @@ bool Kernel::GetAttrCodePropMetadata() {
   return true;
 }
 
-bool Kernel::SetAvailableSgprVgpr() {
-  std::string buf;
-
-  amd_comgr_metadata_node_t isaMeta;
-  amd_comgr_metadata_node_t sgprMeta;
-  amd_comgr_metadata_node_t vgprMeta;
-  bool hasIsaMeta = false;
-  bool hasSgprMeta = false;
-  bool hasVgprMeta = false;
-
-  amd_comgr_status_t status = amd::Comgr::get_isa_metadata(
-                                prog().device().isa().isaName().c_str(), &isaMeta);
-
-  if (status == AMD_COMGR_STATUS_SUCCESS) {
-    hasIsaMeta = true;
-    status = amd::Comgr::metadata_lookup(isaMeta, "AddressableNumSGPRs", &sgprMeta);
-  }
-
-  if (status == AMD_COMGR_STATUS_SUCCESS) {
-    hasSgprMeta = true;
-    status = getMetaBuf(sgprMeta, &buf);
-  }
-
-  workGroupInfo_.availableSGPRs_ = (status == AMD_COMGR_STATUS_SUCCESS) ? atoi(buf.c_str()) : 0;
-
-  if (status == AMD_COMGR_STATUS_SUCCESS) {
-    status = amd::Comgr::metadata_lookup(isaMeta, "AddressableNumVGPRs", &vgprMeta);
-  }
-
-  if (status == AMD_COMGR_STATUS_SUCCESS) {
-    hasVgprMeta = true;
-    status = getMetaBuf(vgprMeta, &buf);
-  }
-  workGroupInfo_.availableVGPRs_ = (status == AMD_COMGR_STATUS_SUCCESS) ? atoi(buf.c_str()) : 0;
-
-  if (hasVgprMeta) {
-    amd::Comgr::destroy_metadata(vgprMeta);
-  }
-
-  if (hasSgprMeta) {
-    amd::Comgr::destroy_metadata(sgprMeta);
-  }
-
-  if (hasIsaMeta) {
-    amd::Comgr::destroy_metadata(isaMeta);
-  }
-
-  return (status == AMD_COMGR_STATUS_SUCCESS);
-}
-
 bool Kernel::GetPrintfStr(std::vector<std::string>* printfStr) {
   const amd_comgr_metadata_node_t programMD = prog().metadata();
   amd_comgr_metadata_node_t printfMeta;

rocclr/device/devkernel.hpp

@@ -367,9 +367,6 @@ class Kernel : public amd::HeapObject {
   //! Retrieve kernel attribute and code properties metadata
   bool GetAttrCodePropMetadata();
 
-  //! Retrieve the available SGPRs and VGPRs
-  bool SetAvailableSgprVgpr();
-
   //! Retrieve the printf string metadata
   bool GetPrintfStr(std::vector<std::string>* printfStr);
 

rocclr/device/rocm/rockernel.cpp

@@ -47,11 +47,6 @@ bool Kernel::postLoad() {
   workGroupInfo_.availableLDSSize_ = device().info().localMemSizePerCU_;
   assert(workGroupInfo_.availableLDSSize_ > 0);
 
-  if (!SetAvailableSgprVgpr()) {
-    DevLogError("Cannot set available SGPR/VGPR\n");
-    return false;
-  }
-
   // Get the kernel code handle
   hsa_status_t hsaStatus;
   hsa_executable_symbol_t symbol;
@@ -145,6 +140,8 @@ bool Kernel::postLoad() {
   }
   assert(wavefront_size > 0);
 
+  workGroupInfo_.availableVGPRs_ = device().isa().vgprPerWavefront();
+  workGroupInfo_.availableSGPRs_ = device().isa().sgprPerWavefront();
   workGroupInfo_.privateMemSize_ = workitemPrivateSegmentByteSize_;
   workGroupInfo_.localMemSize_ = workgroupGroupSegmentByteSize_;
   workGroupInfo_.usedLDSSize_ = workgroupGroupSegmentByteSize_;