* * *

mem: support for gpu-style RMWs in ruby

This patch adds support for GPU-style read-modify-write (RMW) operations in
ruby. Such atomic operations are traditionally executed at the memory controller
(instead of through an L1 cache using cache-line locking).

Currently, this patch works by propogating operation functors through the memory
system.

src/base/types.hh

@@ -200,6 +200,19 @@ typedef std::shared_ptr<FaultBase> Fault;
 constexpr decltype(nullptr) NoFault = nullptr;
 #endif
 
+struct AtomicOpFunctor
+{
+    virtual void operator()(uint8_t *p) = 0;
+    virtual ~AtomicOpFunctor() {}
+};
+
+template <class T>
+struct TypedAtomicOpFunctor : public AtomicOpFunctor
+{
+    void operator()(uint8_t *p) { execute((T *)p); }
+    virtual void execute(T * p) = 0;
+};
+
 enum ByteOrder {
     BigEndianByteOrder,
     LittleEndianByteOrder

src/mem/abstract_mem.cc

@@ -341,39 +341,46 @@ AbstractMemory::access(PacketPtr pkt)
     uint8_t *hostAddr = pmemAddr + pkt->getAddr() - range.start();
 
     if (pkt->cmd == MemCmd::SwapReq) {
-        std::vector<uint8_t> overwrite_val(pkt->getSize());
-        uint64_t condition_val64;
-        uint32_t condition_val32;
+        if (pkt->isAtomicOp()) {
+            if (pmemAddr) {
+                memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
+                (*(pkt->getAtomicOp()))(hostAddr);
+            }
+        } else {
+            std::vector<uint8_t> overwrite_val(pkt->getSize());
+            uint64_t condition_val64;
+            uint32_t condition_val32;
 
-        if (!pmemAddr)
-            panic("Swap only works if there is real memory (i.e. null=False)");
+            if (!pmemAddr)
+                panic("Swap only works if there is real memory (i.e. null=False)");
 
-        bool overwrite_mem = true;
-        // keep a copy of our possible write value, and copy what is at the
-        // memory address into the packet
-        std::memcpy(&overwrite_val[0], pkt->getConstPtr<uint8_t>(),
-                    pkt->getSize());
-        std::memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
+            bool overwrite_mem = true;
+            // keep a copy of our possible write value, and copy what is at the
+            // memory address into the packet
+            std::memcpy(&overwrite_val[0], pkt->getConstPtr<uint8_t>(),
+                        pkt->getSize());
+            std::memcpy(pkt->getPtr<uint8_t>(), hostAddr, pkt->getSize());
 
-        if (pkt->req->isCondSwap()) {
-            if (pkt->getSize() == sizeof(uint64_t)) {
-                condition_val64 = pkt->req->getExtraData();
-                overwrite_mem = !std::memcmp(&condition_val64, hostAddr,
-                                             sizeof(uint64_t));
-            } else if (pkt->getSize() == sizeof(uint32_t)) {
-                condition_val32 = (uint32_t)pkt->req->getExtraData();
-                overwrite_mem = !std::memcmp(&condition_val32, hostAddr,
-                                             sizeof(uint32_t));
-            } else
-                panic("Invalid size for conditional read/write\n");
+            if (pkt->req->isCondSwap()) {
+                if (pkt->getSize() == sizeof(uint64_t)) {
+                    condition_val64 = pkt->req->getExtraData();
+                    overwrite_mem = !std::memcmp(&condition_val64, hostAddr,
+                                                 sizeof(uint64_t));
+                } else if (pkt->getSize() == sizeof(uint32_t)) {
+                    condition_val32 = (uint32_t)pkt->req->getExtraData();
+                    overwrite_mem = !std::memcmp(&condition_val32, hostAddr,
+                                                 sizeof(uint32_t));
+                } else
+                    panic("Invalid size for conditional read/write\n");
+            }
+
+            if (overwrite_mem)
+                std::memcpy(hostAddr, &overwrite_val[0], pkt->getSize());
+
+            assert(!pkt->req->isInstFetch());
+            TRACE_PACKET("Read/Write");
+            numOther[pkt->req->masterId()]++;
         }
-
-        if (overwrite_mem)
-            std::memcpy(hostAddr, &overwrite_val[0], pkt->getSize());
-
-        assert(!pkt->req->isInstFetch());
-        TRACE_PACKET("Read/Write");
-        numOther[pkt->req->masterId()]++;
     } else if (pkt->isRead()) {
         assert(!pkt->isWrite());
         if (pkt->isLLSC()) {

src/mem/packet.hh

@@ -660,6 +660,12 @@ class Packet : public Printable
         return _isSecure;
     }
 
+    /**
+     * Accessor function to atomic op.
+     */
+    AtomicOpFunctor *getAtomicOp() const { return req->getAtomicOpFunctor(); }
+    bool isAtomicOp() const { return req->isAtomic(); }
+
     /**
      * It has been determined that the SC packet should successfully update
      * memory. Therefore, convert this SC packet to a normal write.

src/mem/protocol/RubySlicc_Exports.sm

@@ -56,6 +56,7 @@ bool testAndWrite(Addr addr, DataBlock datablk, Packet *pkt);
 enumeration(AccessPermission, desc="...", default="AccessPermission_NotPresent") {
   // Valid data
   Read_Only,  desc="block is Read Only (modulo functional writes)";
+  Write_Only, desc="block is Write Only";
   Read_Write, desc="block is Read/Write";
 
   // Possibly Invalid data
@@ -144,7 +145,9 @@ enumeration(TransitionResult, desc="...") {
 enumeration(RubyRequestType, desc="...", default="RubyRequestType_NULL") {
   LD,                desc="Load";
   ST,                desc="Store";
-  ATOMIC,            desc="Atomic Load/Store";
+  ATOMIC,            desc="Atomic Load/Store -- depricated. use ATOMIC_RETURN or ATOMIC_NO_RETURN";
+  ATOMIC_RETURN,     desc="Atomic Load/Store, return data";
+  ATOMIC_NO_RETURN,  desc="Atomic Load/Store, do not return data";
   IFETCH,            desc="Instruction fetch";
   IO,                desc="I/O";
   REPLACEMENT,       desc="Replacement";
@@ -166,6 +169,8 @@ enumeration(SequencerRequestType, desc="...", default="SequencerRequestType_NULL
   Default,     desc="Replace this with access_types passed to the DMA Ruby object";
   LD,          desc="Load";
   ST,          desc="Store";
+  ATOMIC,      desc="Atomic Load/Store";
+  REPLACEMENT, desc="Replacement";
   FLUSH,       desc="Flush request type";
   NULL,        desc="Invalid request type";
 }

src/mem/protocol/RubySlicc_Types.sm

@@ -126,6 +126,7 @@ structure(RubyRequest, desc="...", interface="Message", external="yes") {
   int Size,                  desc="size in bytes of access";
   PrefetchBit Prefetch,      desc="Is this a prefetch request";
   int contextId,             desc="this goes away but must be replace with Nilay";
+  int wfid,                  desc="Writethrough wavefront";
   HSAScope scope,            desc="HSA scope";
   HSASegment segment,        desc="HSA segment";
 }

src/mem/request.hh

@@ -160,6 +160,11 @@ class Request
         /** The request should be marked with RELEASE. */
         RELEASE                     = 0x00040000,
 
+        /** The request is an atomic that returns data. */
+        ATOMIC_RETURN_OP            = 0x40000000,
+        /** The request is an atomic that does not return data. */
+        ATOMIC_NO_RETURN_OP         = 0x80000000,
+
         /** The request should be marked with KERNEL.
           * Used to indicate the synchronization associated with a GPU kernel
           * launch or completion.
@@ -345,6 +350,9 @@ class Request
     /** Sequence number of the instruction that creates the request */
     InstSeqNum _reqInstSeqNum;
 
+    /** A pointer to an atomic operation */
+    AtomicOpFunctor *atomicOpFunctor;
+
   public:
 
     /**
@@ -356,7 +364,8 @@ class Request
         : _paddr(0), _size(0), _masterId(invldMasterId), _time(0),
           _taskId(ContextSwitchTaskId::Unknown), _asid(0), _vaddr(0),
           _extraData(0), _contextId(0), _threadId(0), _pc(0),
-          _reqInstSeqNum(0), translateDelta(0), accessDelta(0), depth(0)
+          _reqInstSeqNum(0), atomicOpFunctor(nullptr), translateDelta(0),
+          accessDelta(0), depth(0)
     {}
 
     Request(Addr paddr, unsigned size, Flags flags, MasterID mid,
@@ -364,7 +373,8 @@ class Request
         : _paddr(0), _size(0), _masterId(invldMasterId), _time(0),
           _taskId(ContextSwitchTaskId::Unknown), _asid(0), _vaddr(0),
           _extraData(0), _contextId(0), _threadId(0), _pc(0),
-          _reqInstSeqNum(seq_num), translateDelta(0), accessDelta(0), depth(0)
+          _reqInstSeqNum(seq_num), atomicOpFunctor(nullptr), translateDelta(0),
+          accessDelta(0), depth(0)
     {
         setPhys(paddr, size, flags, mid, curTick());
         setThreadContext(cid, tid);
@@ -380,7 +390,8 @@ class Request
         : _paddr(0), _size(0), _masterId(invldMasterId), _time(0),
           _taskId(ContextSwitchTaskId::Unknown), _asid(0), _vaddr(0),
           _extraData(0), _contextId(0), _threadId(0), _pc(0),
-          _reqInstSeqNum(0), translateDelta(0), accessDelta(0), depth(0)
+          _reqInstSeqNum(0), atomicOpFunctor(nullptr), translateDelta(0),
+          accessDelta(0), depth(0)
     {
         setPhys(paddr, size, flags, mid, curTick());
     }
@@ -389,7 +400,8 @@ class Request
         : _paddr(0), _size(0), _masterId(invldMasterId), _time(0),
           _taskId(ContextSwitchTaskId::Unknown), _asid(0), _vaddr(0),
           _extraData(0), _contextId(0), _threadId(0), _pc(0),
-          _reqInstSeqNum(0), translateDelta(0), accessDelta(0), depth(0)
+          _reqInstSeqNum(0), atomicOpFunctor(nullptr), translateDelta(0),
+          accessDelta(0), depth(0)
     {
         setPhys(paddr, size, flags, mid, time);
     }
@@ -398,12 +410,12 @@ class Request
             Addr pc)
         : _paddr(0), _size(0), _masterId(invldMasterId), _time(0),
           _taskId(ContextSwitchTaskId::Unknown), _asid(0), _vaddr(0),
-          _extraData(0), _contextId(0), _threadId(0), _pc(0),
-          _reqInstSeqNum(0), translateDelta(0), accessDelta(0), depth(0)
+          _extraData(0), _contextId(0), _threadId(0), _pc(pc),
+          _reqInstSeqNum(0), atomicOpFunctor(nullptr), translateDelta(0),
+          accessDelta(0), depth(0)
     {
         setPhys(paddr, size, flags, mid, time);
         privateFlags.set(VALID_PC);
-        _pc = pc;
     }
 
     Request(int asid, Addr vaddr, unsigned size, Flags flags, MasterID mid,
@@ -411,13 +423,27 @@ class Request
         : _paddr(0), _size(0), _masterId(invldMasterId), _time(0),
           _taskId(ContextSwitchTaskId::Unknown), _asid(0), _vaddr(0),
           _extraData(0), _contextId(0), _threadId(0), _pc(0),
-          _reqInstSeqNum(0), translateDelta(0), accessDelta(0), depth(0)
+          _reqInstSeqNum(0), atomicOpFunctor(nullptr), translateDelta(0),
+          accessDelta(0), depth(0)
     {
         setVirt(asid, vaddr, size, flags, mid, pc);
         setThreadContext(cid, tid);
     }
 
-    ~Request() {}
+    Request(int asid, Addr vaddr, int size, Flags flags, MasterID mid, Addr pc,
+            int cid, ThreadID tid, AtomicOpFunctor *atomic_op)
+        : atomicOpFunctor(atomic_op)
+    {
+        setVirt(asid, vaddr, size, flags, mid, pc);
+        setThreadContext(cid, tid);
+    }
+
+    ~Request()
+    {
+        if (hasAtomicOpFunctor()) {
+            delete atomicOpFunctor;
+        }
+    }
 
     /**
      * Set up CPU and thread numbers.
@@ -541,6 +567,22 @@ class Request
         return _time;
     }
 
+    /**
+     * Accessor for atomic-op functor.
+     */
+    bool
+    hasAtomicOpFunctor()
+    {
+        return atomicOpFunctor != NULL;
+    }
+
+    AtomicOpFunctor *
+    getAtomicOpFunctor()
+    {
+        assert(atomicOpFunctor != NULL);
+        return atomicOpFunctor;
+    }
+
     /** Accessor for flags. */
     Flags
     getFlags()
@@ -749,6 +791,15 @@ class Request
     bool isAcquire() const { return _flags.isSet(ACQUIRE); }
     bool isRelease() const { return _flags.isSet(RELEASE); }
     bool isKernel() const { return _flags.isSet(KERNEL); }
+    bool isAtomicReturn() const { return _flags.isSet(ATOMIC_RETURN_OP); }
+    bool isAtomicNoReturn() const { return _flags.isSet(ATOMIC_NO_RETURN_OP); }
+
+    bool
+    isAtomic() const
+    {
+        return _flags.isSet(ATOMIC_RETURN_OP) ||
+               _flags.isSet(ATOMIC_NO_RETURN_OP);
+    }
 
     /**
      * Accessor functions for the memory space configuration flags and used by