Merge pull request #731 from ROCm/rvs-rel-6.1

Merge pull request #730 from jkottiku/master
ROCm · Apr 10, 2024 · f590ccb · f590ccb
2 parents 8ac8ffc + ddc928b
commit f590ccb
Show file tree

Hide file tree

Showing 7 changed files with 538 additions and 145 deletions.
diff --git a/babel.so/CMakeLists.txt b/babel.so/CMakeLists.txt
@@ -61,11 +61,13 @@ set(HIP_HCC_BUILD_FLAGS "${HIP_HCC_BUILD_FLAGS} -DHIP_VERSION_MAJOR=${HIP_VERSIO
 set(HIP_HCC_BUILD_FLAGS)
 set(HIP_HCC_BUILD_FLAGS "${HIP_HCC_BUILD_FLAGS} -fPIC ${HCC_CXX_FLAGS} -I${HSA_PATH}/include ${ASAN_CXX_FLAGS}")
 
+set(HIP_STREAM_BUILD_FLAGS "-DNONTEMPORAL=1 -DDWORDS_PER_LANE=4 -DTBSIZE=1024 -DCHUNKS_PER_BLOCK=2 -O3 -std=c++17")
+
 # Set compiler and compiler flags
 set(CMAKE_CXX_COMPILER "${HIPCC_PATH}/bin/hipcc")
 set(CMAKE_C_COMPILER   "${HIPCC_PATH}/bin/hipcc")
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${HIP_HCC_BUILD_FLAGS}")
-set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${HIP_HCC_BUILD_FLAGS}")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${HIP_HCC_BUILD_FLAGS} ${HIP_STREAM_BUILD_FLAGS}")
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${HIP_HCC_BUILD_FLAGS} ${HIP_STREAM_BUILD_FLAGS}")
 set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${ASAN_LD_FLAGS}")
 set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} ${ASAN_LD_FLAGS}")
 

diff --git a/babel.so/include/HIPStream.h b/babel.so/include/HIPStream.h
@@ -5,23 +5,62 @@
 // For full license terms please see the LICENSE file distributed with this
 // source code
 
-#ifndef MEM_SO_INCLUDE_HIP_STREAM_H_
-#define MEM_SO_INCLUDE_HIP_STREAM_H_
+#pragma once
 
+#include <algorithm>
 #include <iostream>
 #include <stdexcept>
 #include <sstream>
 
 #include "Stream.h"
+#include "hip/hip_runtime.h"
+#ifndef __HIP_PLATFORM_NVCC__
+#include "hip/hip_ext.h"
+#endif
 
 #define IMPLEMENTATION_STRING "HIP"
 
 template <class T>
 class HIPStream : public Stream<T>
 {
+#ifdef __HIP_PLATFORM_NVCC__
+  #ifndef DWORDS_PER_LANE
+  #define DWORDS_PER_LANE 1
+  #endif
+  #ifndef CHUNKS_PER_BLOCK
+  #define CHUNKS_PER_BLOCK 8
+  #endif
+#else
+  #ifndef DWORDS_PER_LANE
+  #define DWORDS_PER_LANE 4
+  #endif
+  #ifndef CHUNKS_PER_BLOCK
+  #define CHUNKS_PER_BLOCK 1
+  #endif
+#endif
+  // make sure that either:
+  //    DWORDS_PER_LANE is less than sizeof(T), in which case we default to 1 element
+  //    or
+  //    DWORDS_PER_LANE is divisible by sizeof(T)
+  static_assert((DWORDS_PER_LANE * sizeof(unsigned int) < sizeof(T)) ||
+                (DWORDS_PER_LANE * sizeof(unsigned int) % sizeof(T) == 0),
+                "DWORDS_PER_LANE not divisible by sizeof(element_type)");
+
+  static constexpr unsigned int chunks_per_block{CHUNKS_PER_BLOCK};
+  // take into account the datatype size
+  // that is, if we specify 4 DWORDS_PER_LANE, this is 2 FP64 elements
+  // and 4 FP32 elements
+  static constexpr unsigned int elements_per_lane{
+    (DWORDS_PER_LANE * sizeof(unsigned int)) < sizeof(T) ? 1 : (
+     DWORDS_PER_LANE * sizeof(unsigned int) / sizeof(T))};
   protected:
     // Size of arrays
-    unsigned int array_size;
+    const unsigned int array_size;
+    const unsigned int block_cnt;
+    const bool evt_timing;
+    hipEvent_t start_ev;
+    hipEvent_t stop_ev;
+    hipEvent_t coherent_ev;
 
     // Host array for partial sums for dot kernel
     T *sums;
@@ -30,22 +69,19 @@ class HIPStream : public Stream<T>
     T *d_a;
     T *d_b;
     T *d_c;
-    T *d_sum;
-
 
   public:
-
-    HIPStream(const unsigned int, const int);
+    HIPStream(const unsigned int, const bool, const int);
     ~HIPStream();
 
-    virtual void copy() override;
-    virtual void add() override;
-    virtual void mul() override;
-    virtual void triad() override;
+    virtual float read() override;
+    virtual float write() override;
+    virtual float copy() override;
+    virtual float add() override;
+    virtual float mul() override;
+    virtual float triad() override;
     virtual T dot() override;
 
     virtual void init_arrays(T initA, T initB, T initC) override;
     virtual void read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
-
 };
-#endif
diff --git a/babel.so/include/Stream.h b/babel.so/include/Stream.h
@@ -5,9 +5,7 @@
 // For full license terms please see the LICENSE file distributed with this
 // source code
 
-
-#ifndef RVS_INCLUDE_STREAM_H_
-#define RVS_INCLUDE_STREAM_H_
+#pragma once
 
 #include <vector>
 #include <string>
@@ -27,10 +25,12 @@ class Stream
 
     // Kernels
     // These must be blocking calls
-    virtual void copy() = 0;
-    virtual void mul() = 0;
-    virtual void add() = 0;
-    virtual void triad() = 0;
+    virtual float read() = 0;
+    virtual float write() = 0;
+    virtual float copy() = 0;
+    virtual float mul() = 0;
+    virtual float add() = 0;
+    virtual float triad() = 0;
     virtual T dot() = 0;
 
     // Copy memory between host and device
@@ -44,6 +44,3 @@ class Stream
 void listDevices(void);
 std::string getDeviceName(const int);
 std::string getDeviceDriver(const int);
-
-#endif
-
diff --git a/babel.so/src/rvs_memworker.cpp b/babel.so/src/rvs_memworker.cpp
@@ -63,8 +63,8 @@ void MemWorker::run() {
     std::pair<int, uint16_t> device;
 
     // log MEM stress test - start message
-    msg = "[" + action_name + "] " + MODULE_NAME + " " +
-            std::to_string(gpu_id) + " "  + " Starting the Memory stress test ";
+    msg = "[" + action_name + "] " + "[GPU:: " +
+            std::to_string(gpu_id) + "] " + "Starting the Babel memory stress test";
     rvs::lp::Log(msg, rvs::logresults);
 
     /* Device Index */