cpu,stdlib: Fix Access Trace for Accessing Indices in SpatterGen (#1258)

This change fixes the way indices are generated in a multi generator
setup.
It changes it from all cores generating the same trace of indices for
accessing the index array to each core generating an interleaved subset
of indices.
For an example look below for traces (indices to index array) in a 2
core setup.

Before:
core_0: 0, 1, 2, 3, 4, 5, 6, 7, ...
core_1: 0, 1, 2, 3, 4, 5, 6, 7, ...
After:
core_0: 0, 1, 2, 3, 8, 9, 10, 11, ...
core_1: 4, 5, 6, 7, 12, 13, 14, 15, ...

Additionally, this change fixes the SpatterKernel class in the standard
library to comply with the change in the SpatterGen source code.

src/cpu/testers/spatter_gen/spatter_gen.cc

@@ -205,6 +205,7 @@ void
 SpatterGen::addKernel(
     uint32_t id, uint32_t delta, uint32_t count,
     SpatterKernelType type,
+    uint32_t base_index, uint32_t indices_per_stride, uint32_t stride,
     size_t index_size, Addr base_index_addr,
     size_t value_size, Addr base_value_addr,
     const std::vector<uint32_t>& indices
@@ -218,6 +219,7 @@ SpatterGen::addKernel(
     SpatterKernel new_kernel(
                             requestorId,
                             id, delta, count, type,
+                            base_index, indices_per_stride, stride,
                             index_size, base_index_addr,
                             value_size, base_value_addr
                             );

src/cpu/testers/spatter_gen/spatter_gen.hh

@@ -234,11 +234,11 @@ class SpatterGen: public ClockedObject
 
     void recvReqRetry();
     bool recvTimingResp(PacketPtr pkt);
-
     // PyBindMethod to interface adding a kernel with python JSON frontend.
     void addKernel(
         uint32_t id, uint32_t delta, uint32_t count,
         SpatterKernelType type,
+        uint32_t base_index, uint32_t indices_per_stride, uint32_t stride,
         size_t index_size, Addr base_index_addr,
         size_t value_size, Addr base_value_addr,
         const std::vector<uint32_t>& indices

src/cpu/testers/spatter_gen/utility_structs.hh

@@ -161,7 +161,38 @@ class SpatterKernel
     typedef enums::SpatterKernelType SpatterKernelType;
     typedef SpatterAccess::AccessPair AccessPair;
 
+    class IndexGen
+    {
+      private:
+        uint32_t indicesPerStride;
+        uint32_t stride;
+
+        uint32_t next;
+      public:
+        IndexGen(): indicesPerStride(0), stride(0), next(0)
+        {}
+
+        IndexGen(uint32_t base_index,
+                uint32_t indices_per_stride,
+                uint32_t stride_size):
+            indicesPerStride(indices_per_stride),
+            stride(stride_size), next(base_index)
+        {}
+
+        uint32_t nextIndex() {
+            uint32_t ret = next;
+            // update next index
+            next++;
+            if (next % indicesPerStride == 0) {
+                next += (stride - indicesPerStride);
+            }
+            return ret;
+        }
+    };
+
     RequestorID requestorId;
+    IndexGen indexGen;
+
     uint32_t _id;
     uint32_t delta;
     uint32_t count;
@@ -174,8 +205,6 @@ class SpatterKernel
     size_t valueSize;
     Addr baseValueAddr;
 
-    // needed to iterate over indices multiple times.
-    uint32_t index;
     // current iteration over indices
     uint32_t iteration;
 
@@ -189,15 +218,17 @@ class SpatterKernel
         RequestorID requestor_id,
         uint32_t id, uint32_t delta, uint32_t count,
         SpatterKernelType type,
+        uint32_t base_index, uint32_t indices_per_stride, uint32_t stride,
         size_t index_size, Addr base_index_addr,
         size_t value_size, Addr base_value_addr
     ):
         requestorId(requestor_id),
+        indexGen(base_index, indices_per_stride, stride),
         _id(id), delta(delta), count(count),
         _type(type),
         indexSize(index_size), baseIndexAddr(base_index_addr),
         valueSize(value_size), baseValueAddr(base_value_addr),
-        index(0), iteration(0), remRolls(0)
+        iteration(0), remRolls(0)
     {}
 
     uint32_t id() const { return _id; }
@@ -215,10 +246,10 @@ class SpatterKernel
     SpatterAccess* nextSpatterAccess()
     {
         std::queue<AccessPair> access_pairs;
+        // get the next index for the index array
+        uint32_t index = indexGen.nextIndex();
         Addr index_addr = baseIndexAddr + (index * indexSize);
         access_pairs.emplace(index_addr, indexSize);
-        // update index in the index array
-        index++;
 
         uint32_t front = indices.front();
         uint32_t value_index = (delta * iteration) + front;

src/python/gem5/components/processors/spatter_gen/spatter_kernel.py

@@ -150,22 +150,28 @@ class SpatterKernel:
         kernel_delta: int,
         kernel_count: int,
         kernel_type: SpatterKernelType,
-        kernel_trace: List[int],
+        base_index: int,
+        indices_per_stride: int,
+        stride_size: int,
         index_size: int,
         base_index_addr: Addr,
         value_size: int,
         base_value_addr: Addr,
+        kernel_trace: List[int],
         fix_empty_trace: bool = False,
     ):
         self._id = kernel_id
         self._delta = kernel_delta
         self._count = kernel_count
-        self._trace = kernel_trace
         self._type = kernel_type
+        self._base_index = base_index
+        self._indices_per_stride = indices_per_stride
+        self._stride_size = stride_size
         self._index_size = index_size
         self._base_index_addr = base_index_addr
         self._value_size = value_size
         self._base_value_addr = base_value_addr
+        self._trace = kernel_trace
 
         if fix_empty_trace and len(kernel_trace) == 0:
             inform(
@@ -185,6 +191,9 @@ class SpatterKernel:
             self._delta,
             self._count,
             self._type.getValue(),
+            self._base_index,
+            self._indices_per_stride,
+            self._stride_size,
             self._index_size,
             self._base_index_addr,
             self._value_size,