Changeset 17071 for stable/HeuristicLab.ExtLibs/HeuristicLab.NativeInterpreter/0.1/NativeInterpreter-0.1/src/interpreter.cpp

Timestamp:

07/04/19 16:03:48 (5 years ago)

Author:

mkommend

Message:

#2958: Merged 16266, 16269, 16274, 16276, 16277, 16285, 16286, 16287, 16289, 16293, 16296, 16297, 16298, 16333, 16334 into stable.

Location:

stable

Files:

• . (modified) (1 prop)
• HeuristicLab.ExtLibs (modified) (1 prop)
• HeuristicLab.ExtLibs/HeuristicLab.NativeInterpreter (copied) (copied from trunk/HeuristicLab.ExtLibs/HeuristicLab.NativeInterpreter)
• HeuristicLab.ExtLibs/HeuristicLab.NativeInterpreter/0.1/NativeInterpreter-0.1 (copied) (copied from trunk/HeuristicLab.ExtLibs/HeuristicLab.NativeInterpreter/0.1/NativeInterpreter-0.1)
• HeuristicLab.ExtLibs/HeuristicLab.NativeInterpreter/0.1/NativeInterpreter-0.1/src/interpreter.cpp (modified) (3 diffs)

stable/HeuristicLab.ExtLibs/HeuristicLab.NativeInterpreter/0.1/NativeInterpreter-0.1/src/interpreter.cpp

@@ -1 @@
-#include <memory>
-
 #include "interpreter.h" 
 
+#ifdef __cplusplus
 extern "C" {
+#endif
+
+constexpr size_t BUFSIZE = BATCHSIZE * sizeof(double);
 
 // slow (ish?)
@@ -22 +24 @@
 
 __declspec(dllexport)
-void __cdecl GetValuesVectorized(instruction* code, int codeLength, int* rows, int totalRows, double* result) noexcept
+void __cdecl GetValuesVectorized(instruction* code, int codeLength, int* rows, int totalRows, double* __restrict result) noexcept
 {
-    std::vector<double[BUFSIZE]> buffers(codeLength);
-    // initialize instruction buffers
+    double* buffer = static_cast<double*>(_aligned_malloc(codeLength * BUFSIZE, 16));
     for (int i = 0; i < codeLength; ++i)
     {
         instruction& in = code[i];
-        in.buf = buffers[i];
+        in.buf = buffer + (i * BATCHSIZE);
 
         if (in.opcode == OpCodes::Const)
@@ -37 +38 @@
     }
 
-    int remainingRows = totalRows % BUFSIZE;
+    int remainingRows = totalRows % BATCHSIZE;
     int total = totalRows - remainingRows;
 
-    for (int rowIndex = 0; rowIndex < total; rowIndex += BUFSIZE)
+    for (int rowIndex = 0; rowIndex < total; rowIndex += BATCHSIZE)
     {
-        evaluate(code, codeLength, rows, rowIndex, BUFSIZE);
-        std::memcpy(result + rowIndex, code[0].buf, BUFSIZE * sizeof(double));
+        evaluate(code, codeLength, rows, rowIndex, BATCHSIZE);
+        std::memcpy(result + rowIndex, code[0].buf, BUFSIZE);
     }
 
     // are there any rows left?
-    if (remainingRows > 0) {
-        for (int rowIndex = total; rowIndex < totalRows; rowIndex += remainingRows)
-        {
-            evaluate(code, codeLength, rows, rowIndex, remainingRows);
-            std::memcpy(result + rowIndex, code[0].buf, remainingRows * sizeof(double));
-        }
+    if (remainingRows > 0) 
+    {
+        evaluate(code, codeLength, rows, total, remainingRows);
+        std::memcpy(result + total, code[0].buf, remainingRows * sizeof(double));
     }
+    _aligned_free(buffer);
 }