SWDEV-440866 - [hip-roclr] Adds support to batch memory operations APIs

Change-Id: I449ffca44bbb04d13348d112e896d603c70fd485

hipamd/src/CMakeLists.txt

@@ -299,7 +299,7 @@ if(WIN32 OR NOT BUILD_SHARED_LIBS)
     # rocprofiler-register is not support on Windows
     set(HIP_ENABLE_ROCPROFILER_REGISTER OFF)
 else()
-    option(HIP_ENABLE_ROCPROFILER_REGISTER "Enable rocprofiler-register support" ON)
+  option(HIP_ENABLE_ROCPROFILER_REGISTER "Enable rocprofiler-register support" ON)
 endif()
 
 if(HIP_ENABLE_ROCPROFILER_REGISTER)

hipamd/src/amdhip.def

@@ -480,3 +480,4 @@ hipGraphExecNodeSetParams
 hipDrvGraphMemcpyNodeSetParams
 hipDrvGraphMemcpyNodeGetParams
 hipExtHostAlloc
+hipStreamBatchMemOp

hipamd/src/hip_api_trace.cpp

@@ -602,6 +602,8 @@ hipError_t hipStreamWaitValue64(hipStream_t stream, void* ptr, uint64_t value, u
                                 uint64_t mask);
 hipError_t hipStreamWriteValue32(hipStream_t stream, void* ptr, uint32_t value, unsigned int flags);
 hipError_t hipStreamWriteValue64(hipStream_t stream, void* ptr, uint64_t value, unsigned int flags);
+hipError_t hipStreamBatchMemOp(hipStream_t stream, unsigned int count, hipStreamBatchMemOpParams* paramArray,
+                               unsigned int flags);
 hipError_t hipTexObjectCreate(hipTextureObject_t* pTexObject, const HIP_RESOURCE_DESC* pResDesc,
                               const HIP_TEXTURE_DESC* pTexDesc,
                               const HIP_RESOURCE_VIEW_DESC* pResViewDesc);
@@ -1197,6 +1199,7 @@ void UpdateDispatchTable(HipDispatchTable* ptrDispatchTable) {
   ptrDispatchTable->hipStreamWaitValue64_fn = hip::hipStreamWaitValue64;
   ptrDispatchTable->hipStreamWriteValue32_fn = hip::hipStreamWriteValue32;
   ptrDispatchTable->hipStreamWriteValue64_fn = hip::hipStreamWriteValue64;
+  ptrDispatchTable->hipStreamBatchMemOp_fn = hip::hipStreamBatchMemOp;
   ptrDispatchTable->hipTexObjectCreate_fn = hip::hipTexObjectCreate;
   ptrDispatchTable->hipTexObjectDestroy_fn = hip::hipTexObjectDestroy;
   ptrDispatchTable->hipTexObjectGetResourceDesc_fn = hip::hipTexObjectGetResourceDesc;
@@ -1887,6 +1890,9 @@ HIP_ENFORCE_ABI(HipDispatchTable, hipDrvGraphMemcpyNodeSetParams_fn, 460)
 HIP_ENFORCE_ABI(HipDispatchTable, hipExtHostAlloc_fn, 461)
 // HIP_RUNTIME_API_TABLE_STEP_VERSION == 6
 HIP_ENFORCE_ABI(HipDispatchTable, hipDeviceGetTexture1DLinearMaxWidth_fn, 462)
+// HIP_RUNTIME_API_TABLE_STEP_VERSION == 7
+HIP_ENFORCE_ABI(HipDispatchTable, hipStreamBatchMemOp_fn, 463);
+
 
 // if HIP_ENFORCE_ABI entries are added for each new function pointer in the table, the number below
 // will be +1 of the number in the last HIP_ENFORCE_ABI line. E.g.:
@@ -1894,9 +1900,9 @@ HIP_ENFORCE_ABI(HipDispatchTable, hipDeviceGetTexture1DLinearMaxWidth_fn, 462)
 //  HIP_ENFORCE_ABI(<table>, <functor>, 8)
 //
 //  HIP_ENFORCE_ABI_VERSIONING(<table>, 9) <- 8 + 1 = 9
-HIP_ENFORCE_ABI_VERSIONING(HipDispatchTable, 463)
+HIP_ENFORCE_ABI_VERSIONING(HipDispatchTable, 464)
 
-static_assert(HIP_RUNTIME_API_TABLE_MAJOR_VERSION == 0 && HIP_RUNTIME_API_TABLE_STEP_VERSION == 6,
+static_assert(HIP_RUNTIME_API_TABLE_MAJOR_VERSION == 0 && HIP_RUNTIME_API_TABLE_STEP_VERSION == 7,
               "If you get this error, add new HIP_ENFORCE_ABI(...) code for the new function "
               "pointers and then update this check so it is true");
 #endif

hipamd/src/hip_hcc.map.in

@@ -583,6 +583,7 @@ local:
 hip_6.3 {
 global:
     hipExtHostAlloc;
+    hipStreamBatchMemOp;
 local:
     *;
 } hip_6.2;

hipamd/src/hip_stream_ops.cpp

@@ -23,8 +23,36 @@
 #include "platform/command_utils.hpp"
 
 namespace hip {
+hipError_t ihipBatchMemOperation(hipStream_t stream, cl_command_type cmdType, unsigned int count,
+                                 hipStreamBatchMemOpParams* paramArray, unsigned int flags) {
+  if (paramArray == nullptr || flags != 0 || count > 256) {
+    return hipErrorInvalidValue;
+  }
+
+  if (!hip::isValid(stream)) {
+    return hipErrorContextIsDestroyed;
+  }
+
+  hip::Stream* hip_stream = hip::getStream(stream);
+  amd::Command::EventWaitList waitList;
+
+  amd::BatchMemoryOperationCommand* command =
+    new amd::BatchMemoryOperationCommand(*hip_stream, cmdType, count,
+                                         flags, waitList, paramArray,
+                                         sizeof(hipStreamBatchMemOpParams));
+
+  if (command == nullptr) {
+    return hipErrorOutOfMemory;
+  }
+  command->enqueue();
+  command->release();
+  HIP_RETURN(hipSuccess);
+}
+
+
 hipError_t ihipStreamOperation(hipStream_t stream, cl_command_type cmdType, void* ptr,
-                               uint64_t value, uint64_t mask, unsigned int flags, size_t sizeBytes) {
+                               uint64_t value, uint64_t mask, unsigned int flags,
+                               size_t sizeBytes) {
   size_t offset = 0;
   unsigned int outFlags = 0;
 
@@ -136,4 +164,15 @@ hipError_t hipStreamWriteValue64(hipStream_t stream, void* ptr, uint64_t value,
       0,  // flags un-used for now set it to 0
       sizeof(uint64_t)));
 }
+
+hipError_t hipStreamBatchMemOp(hipStream_t stream, unsigned int count,
+                               hipStreamBatchMemOpParams* paramArray, unsigned int flags) {
+  HIP_INIT_API(hipStreamBatchMemOp, count, paramArray, flags);
+  HIP_RETURN_DURATION(ihipBatchMemOperation(
+                      stream,
+                      ROCCLR_COMMAND_BATCH_STREAM,
+                      count,
+                      paramArray,
+                      flags));
+  }
 }  // namespace hip

hipamd/src/hip_table_interface.cpp

@@ -1413,6 +1413,10 @@ hipError_t hipStreamWriteValue64(hipStream_t stream, void* ptr, uint64_t value,
                                  unsigned int flags) {
   return hip::GetHipDispatchTable()->hipStreamWriteValue64_fn(stream, ptr, value, flags);
 }
+hipError_t hipStreamBatchMemOp(hipStream_t stream, unsigned int count,
+                               hipStreamBatchMemOpParams* paramArray, unsigned int flags) {
+  return hip::GetHipDispatchTable()->hipStreamBatchMemOp_fn(stream, count, paramArray, flags);
+}
 hipError_t hipTexObjectCreate(hipTextureObject_t* pTexObject, const HIP_RESOURCE_DESC* pResDesc,
                               const HIP_TEXTURE_DESC* pTexDesc,
                               const HIP_RESOURCE_VIEW_DESC* pResViewDesc) {