[Do-not-merge] PIM API fusion initial support

libpimeval/src/libpimeval.cpp

@@ -531,6 +531,14 @@ pimShiftBitsLeft(PimObjId src, PimObjId dest, unsigned shiftAmount)
   return ok ? PIM_OK : PIM_ERROR;
 }
 
+//! @brief  Execute fused PIM APIs
+PimStatus
+pimFuse(PimProg prog)
+{
+  bool ok = pimSim::get()->pimFuse(prog);
+  return ok ? PIM_OK : PIM_ERROR;
+}
+
 //! @brief  BitSIMD-V: Read a row to SA
 PimStatus
 pimOpReadRowToSa(PimObjId src, unsigned ofst)

libpimeval/src/libpimeval.h

@@ -9,6 +9,8 @@
 
 #include <cstdint>
 #include <cstdarg>
+#include <vector>
+#include <functional>
 
 //! @brief  PIM API return status
 enum PimStatus {
@@ -197,6 +199,18 @@ PimStatus pimShiftBitsRight(PimObjId src, PimObjId dest, unsigned shiftAmount);
 PimStatus pimShiftBitsLeft(PimObjId src, PimObjId dest, unsigned shiftAmount);
 
 
+////////////////////////////////////////////////////////////////////////////////
+// Experimental Feature: PIM API Fusion                                       //
+////////////////////////////////////////////////////////////////////////////////
+struct PimProg {
+  template <typename... Args>
+  void add(PimStatus(*api)(Args...), Args... args) {
+    m_apis.push_back([=]() { return api(args...); });
+  }
+  std::vector<std::function<PimStatus()>> m_apis;
+};
+PimStatus pimFuse(PimProg prog);
+
 ////////////////////////////////////////////////////////////////////////////////
 // Warning: Avoid using below customized APIs for functional simulation       //
 //          Some are PIM architecture dependent, some are in progress         //

libpimeval/src/pimCmdFuse.cpp

@@ -0,0 +1,42 @@
+// File: pimCmdFuse.cpp
+// PIMeval Simulator - PIM API Fusion
+// Copyright (c) 2024 University of Virginia
+// This file is licensed under the MIT License.
+// See the LICENSE file in the root of this repository for more details.
+
+#include "pimCmdFuse.h"
+#include <cstdio>
+
+
+//! @brief  Pim CMD: PIM API Fusion
+bool
+pimCmdFuse::execute()
+{
+  if (m_debugCmds) {
+    std::printf("PIM-Cmd: API Fusion\n");
+  }
+
+  // Functional simulation
+  // TODO: skip original updateStats
+  bool success = true;
+  for (auto& api : m_prog.m_apis) {
+    PimStatus status = api();
+    if (status != PIM_OK) {
+      success = false;
+      break;
+    }
+  }
+
+  // Analyze API fusion opportunities
+  success = success && updateStats();
+  return success;
+}
+
+//! @brief  Pim CMD: PIM API Fusion - update stats
+bool
+pimCmdFuse::updateStats() const
+{
+  // TODO: Parse m_prog and update stats
+  return true;
+}
+

libpimeval/src/pimCmdFuse.h

@@ -0,0 +1,28 @@
+// File: pimCmdFuse.h
+// PIMeval Simulator - PIM API Fusion
+// Copyright (c) 2024 University of Virginia
+// This file is licensed under the MIT License.
+// See the LICENSE file in the root of this repository for more details.
+
+#ifndef LAVA_PIM_CMD_FUSE_H
+#define LAVA_PIM_CMD_FUSE_H
+
+#include "libpimeval.h"
+#include "pimCmd.h"
+
+
+//! @class  pimCmdFuse
+//! @brief  Pim CMD: PIM API Fusion
+class pimCmdFuse : public pimCmd
+{
+public:
+  pimCmdFuse(PimProg prog) : pimCmd(PimCmdEnum::NOOP), m_prog(prog) {}
+  virtual ~pimCmdFuse() {}
+  virtual bool execute() override;
+  virtual bool updateStats() const override;
+private:
+  PimProg m_prog;
+};
+
+#endif
+

libpimeval/src/pimSim.cpp

@@ -6,6 +6,7 @@
 
 #include "pimSim.h"
 #include "pimCmd.h"
+#include "pimCmdFuse.h"
 #include "pimParamsDram.h"
 #include "pimStats.h"
 #include "pimUtils.h"
@@ -874,6 +875,15 @@ pimSim::pimShiftBitsLeft(PimObjId src, PimObjId dest, unsigned shiftAmount)
   return m_device->executeCmd(std::move(cmd));
 }
 
+bool
+pimSim::pimFuse(PimProg prog)
+{
+  pimPerfMon perfMon("pimFuse");
+  if (!isValidDevice()) { return false; }
+  std::unique_ptr<pimCmd> cmd = std::make_unique<pimCmdFuse>(prog);
+  return m_device->executeCmd(std::move(cmd));
+}
+
 bool
 pimSim::pimOpReadRowToSa(PimObjId objId, unsigned ofst)
 {

libpimeval/src/pimSim.h

@@ -125,6 +125,9 @@ class pimSim
   bool pimShiftBitsRight(PimObjId src, PimObjId dest, unsigned shiftAmount);
   bool pimShiftBitsLeft(PimObjId src, PimObjId dest, unsigned shiftAmount);
 
+  // PIM API Fusion
+  bool pimFuse(PimProg prog);
+
   // BitSIMD-V micro ops
   bool pimOpReadRowToSa(PimObjId src, unsigned ofst);
   bool pimOpWriteSaToRow(PimObjId src, unsigned ofst);

tests/test-api-fusion/Makefile

@@ -0,0 +1,19 @@
+# Makefile: Test PIM API Fusion
+# Copyright (c) 2024 University of Virginia
+# This file is licensed under the MIT License.
+# See the LICENSE file in the root of this repository for more details.
+
+PROJ_ROOT = ../..
+include ${PROJ_ROOT}/Makefile.common
+
+EXEC := test-api-fusion.out
+SRC := test-api-fusion.cpp
+
+debug perf dramsim3_integ: $(EXEC)
+
+$(EXEC): $(SRC) $(DEPS)
+	$(CXX) $< $(CXXFLAGS) -o $@
+
+clean:
+	rm -rf $(EXEC) *.dSYM
+

tests/test-api-fusion/test-api-fusion.cpp

@@ -0,0 +1,104 @@
+// Test: Test PIM API Fusion
+// Copyright (c) 2024 University of Virginia
+// This file is licensed under the MIT License.
+// See the LICENSE file in the root of this repository for more details.
+
+#include "libpimeval.h"
+#include <iostream>
+#include <vector>
+#include <algorithm>
+#include <bitset>
+#include <cassert>
+#include <cstdlib>
+#include <cstdio>
+
+
+bool testFused(PimDeviceEnum deviceType)
+{
+  // 1GB capacity
+  unsigned numRanks = 1;
+  unsigned numBankPerRank = 1;
+  unsigned numSubarrayPerBank = 8;
+  unsigned numRows = 1024;
+  unsigned numCols = 8192;
+
+  uint64_t numElements = 16 * 1024;
+
+  std::vector<int> src1(numElements);
+  std::vector<int> src2(numElements);
+  std::vector<int> dest1(numElements);
+  std::vector<int> dest2(numElements);
+  const int scalarVal = -123;
+
+  for (uint64_t i = 0; i < numElements; ++i) {
+    src1[i] = static_cast<int>(i);
+    src2[i] = static_cast<int>(i + 1);
+  }
+
+  PimStatus status = pimCreateDevice(deviceType, numRanks, numBankPerRank, numSubarrayPerBank, numRows, numCols);
+  assert(status == PIM_OK);
+
+  PimObjId objSrc1 = pimAlloc(PIM_ALLOC_AUTO, numElements, PIM_INT32);
+  PimObjId objSrc2 = pimAllocAssociated(objSrc1, PIM_INT32);
+  PimObjId objDest1 = pimAllocAssociated(objSrc1, PIM_INT32);
+  PimObjId objDest2 = pimAllocAssociated(objSrc1, PIM_INT32);
+  assert(objSrc1 != -1 && objSrc2 != -1 && objDest1 != -1 && objDest2 != -1);
+
+  // copy host to device
+  status = pimCopyHostToDevice((void*)src1.data(), objSrc1);
+  assert(status == PIM_OK);
+  status = pimCopyHostToDevice((void*)src2.data(), objSrc2);
+  assert(status == PIM_OK);
+
+
+  // Fused PIM APIs
+  PimProg prog;
+  prog.add(pimMulScalar, objSrc1, objDest1, static_cast<uint64_t>(scalarVal));
+  prog.add(pimAdd, objDest1, objSrc2, objDest1);
+  status = pimFuse(prog);
+  assert(status == PIM_OK);
+
+  // Direct APIs
+  status = pimScaledAdd(objSrc1, objSrc2, objDest2, static_cast<uint64_t>(scalarVal));
+  assert(status == PIM_OK);
+
+
+  status = pimCopyDeviceToHost(objDest1, (void*)dest1.data());
+  assert(status == PIM_OK);
+  status = pimCopyDeviceToHost(objDest2, (void*)dest2.data());
+  assert(status == PIM_OK);
+
+  bool ok = true;
+  for (uint64_t i = 0; i < numElements; ++i) {
+    if (dest1[i] != (src1[i] * scalarVal + src2[i]) || dest1[i] != dest2[i]) {
+      ok = false;
+      std::printf("Fused Test Error: src1 %d src2 %d dest1 %d dest2 %d\n", src1[i], src2[i], dest1[i], dest2[i]);
+    }
+  }
+
+  pimFree(objSrc1);
+  pimFree(objSrc2);
+  pimFree(objDest1);
+  pimFree(objDest2);
+
+  pimShowStats();
+  pimResetStats();
+  pimDeleteDevice();
+
+  std::cout << "Fused Test " << (ok ? "PASSED" : "FAILED") << std::endl;
+  return ok;
+}
+
+int main()
+{
+  std::cout << "PIM Regression Test: PIM fused operations" << std::endl;
+
+  bool ok = true;
+  ok &= testFused(PIM_DEVICE_BITSIMD_V);
+  ok &= testFused(PIM_DEVICE_FULCRUM);
+  ok &= testFused(PIM_DEVICE_BANK_LEVEL);
+
+  std::cout << (ok ? "PASSED" : "FAILED") << std::endl;
+  return 0;
+}
+