Changeset 5861

Timestamp:

Feb 3, 2018, 12:02:14 PM (15 months ago)

Author:

cameron

Message:

Using DirectCC builder updates; speedup wc -l

Location:

icGREP/icgrep-devel/icgrep

Files:

• IR_Gen/idisa_builder.h (modified) (2 diffs)
• grep_engine.cpp (modified) (3 diffs)
• kernels/charclasses.cpp (modified) (3 diffs)
• kernels/linebreak_kernel.cpp (modified) (3 diffs)
• kernels/linebreak_kernel.h (modified) (1 diff)
• toolchain/grep_pipeline.cpp (modified) (3 diffs)
• wc.cpp (modified) (9 diffs)

icGREP/icgrep-devel/icgrep/IR_Gen/idisa_builder.h

@@ -16 +16 @@
 namespace IDISA {
 
+    
+inline bool isStreamTy(llvm::Type * t) {
+    return t->isVectorTy() && (t->getVectorNumElements() == 0);
+}
+
+inline bool isStreamSetTy(llvm::Type * t) {
+    return t->isArrayTy() && (isStreamTy(t->getArrayElementType()));
+}
+
+inline unsigned getNumOfStreams (llvm::Type * t) {
+    if (isStreamTy(t)) return 1;
+    assert(isStreamSetTy(t));
+    return t->getArrayNumElements();
+}
+
+inline unsigned getStreamFieldWidth (llvm::Type * t) {
+    if (isStreamTy(t)) return t->getScalarSizeInBits();
+    assert(isStreamSetTy(t));
+    return t->getArrayElementType()->getScalarSizeInBits();
+}
+
+    
 class IDISA_Builder : public CBuilder {
 
@@ -168 +190 @@
         return getStreamSetTy(getContext(), NumElements, FieldWidth);
     }
-
+    
     void CallPrintRegister(const std::string & regName, llvm::Value * const value);
 

icGREP/icgrep-devel/icgrep/grep_engine.cpp

@@ -49 +49 @@
 using namespace llvm;
 using namespace cc;
+using namespace kernel;
 
 static cl::opt<int> Threads("t", cl::desc("Total number of threads."), cl::init(2));
@@ -116 +117 @@
 
     #ifdef USE_DIRECT_LF_BUILDER
-    kernel::Kernel * linefeedK = mGrepDriver->addKernelInstance<kernel::LineFeedKernelBuilder>(idb, encodingBits);
+    kernel::Kernel * linefeedK = mGrepDriver->addKernelInstance<kernel::LineFeedKernelBuilder>(idb, Binding{idb->getStreamSetTy(1, 8), "byteStream", FixedRate(), Principal()});
     mGrepDriver->makeKernelCall(linefeedK, {ByteStream}, {LineFeedStream});
     #endif
@@ -131 +132 @@
 
     #ifndef USE_DIRECT_LF_BUILDER
-    kernel::Kernel * linefeedK = mGrepDriver->addKernelInstance<kernel::LineFeedKernelBuilder>(idb, encodingBits);
+    kernel::Kernel * linefeedK = mGrepDriver->addKernelInstance<kernel::LineFeedKernelBuilder>(idb, Binding{idb->getStreamSetTy(8), "basis", FixedRate(), Principal()});
     mGrepDriver->makeKernelCall(linefeedK, {BasisBits}, {LineFeedStream});
     #endif

icGREP/icgrep-devel/icgrep/kernels/charclasses.cpp

@@ -78 +78 @@
 void CharClassesKernel::generatePabloMethod() {
     PabloBuilder pb(getEntryScope());
-    CC_Compiler * ccc;
+    std::unique_ptr<CC_Compiler> ccc;
     if (mUseDirectCC) {
-        ccc = new Direct_CC_Compiler(this, pb.createExtract(getInputStreamVar("byteData"), pb.getInteger(0)));
+        ccc = make_unique<cc::Direct_CC_Compiler>(this, pb.createExtract(getInput(0), pb.getInteger(0)));
+    } else {
+        ccc = make_unique<cc::Parabix_CC_Compiler>(this, getInputStreamSet("basis"));
     }
-    else {
-        ccc = new Parabix_CC_Compiler(this, getInputStreamSet("basis"));
-    }
-
     unsigned n = mCCs.size();
 
@@ -96 +94 @@
     }
 
-    UCD::UCDCompiler ucdCompiler(*ccc);
+    UCD::UCDCompiler ucdCompiler(*ccc.get());
     if (LLVM_UNLIKELY(AlgorithmOptionIsSet(DisableIfHierarchy))) {
         ucdCompiler.generateWithoutIfHierarchy(nameMap, pb);
@@ -114 +112 @@
         }
     }
-    delete ccc;
 }
 

icGREP/icgrep-devel/icgrep/kernels/linebreak_kernel.cpp

@@ -1 +1 @@
 /*
- *  Copyright (c) 2016 International Characters.
+ *  Copyright (c) 2018 International Characters.
  *  This software is licensed to the public under the Open Software License 3.0.
  */
@@ -12 +12 @@
 #include <cc/cc_compiler.h>
 #include <pablo/builder.hpp>
+#include <IR_Gen/idisa_builder.h>
 #include <kernels/kernel_builder.h>
 
@@ -21 +22 @@
 using namespace re;
 using namespace llvm;
+using namespace IDISA;
 
-LineFeedKernelBuilder::LineFeedKernelBuilder(const std::unique_ptr<kernel::KernelBuilder> & b, const unsigned codeUnitWidth)
-: PabloKernel(b, "lf" + std::to_string(codeUnitWidth),
+LineFeedKernelBuilder::LineFeedKernelBuilder(const std::unique_ptr<kernel::KernelBuilder> & b, Binding && inputStreamSet)
+: PabloKernel(b, "lf" + std::to_string(getNumOfStreams(inputStreamSet.getType())) + "x" + std::to_string(getStreamFieldWidth(inputStreamSet.getType())),
 // input
-#ifdef USE_DIRECT_LF_BUILDER
-{Binding{b->getStreamSetTy(1, codeUnitWidth), "byteStream", FixedRate(), Principal()}},
-#else
-{Binding{b->getStreamSetTy(codeUnitWidth), "basis", FixedRate(), Principal()}},
-#endif
-// output
-{Binding{b->getStreamSetTy(1), "lf"}}) {
-
+{inputStreamSet},
+{Binding{b->getStreamSetTy(1), "lf"}}),
+    mNumOfStreams(getNumOfStreams(inputStreamSet.getType())),
+    mStreamFieldWidth(getStreamFieldWidth(inputStreamSet.getType()))
+{
 }
 
 void LineFeedKernelBuilder::generatePabloMethod() {
     PabloBuilder pb(getEntryScope());
-    #ifdef USE_DIRECT_LF_BUILDER
-    cc::Direct_CC_Compiler ccc(this, pb.createExtract(getInputStreamVar("byteStream"), pb.getInteger(0)));
-    #else
-    cc::Parabix_CC_Compiler ccc(this, getInputStreamSet("basis"));
-    #endif
-    PabloAST * LF = ccc.compileCC("LF", makeByte(0x0A), pb);
+    std::unique_ptr<CC_Compiler> ccc;
+    if (mNumOfStreams == 1) {
+        ccc = make_unique<cc::Direct_CC_Compiler>(this, pb.createExtract(getInput(0), pb.getInteger(0)));
+    } else {
+        ccc = make_unique<cc::Parabix_CC_Compiler>(this, getInputStreamSet("basis"));
+    }
+    PabloAST * LF = ccc->compileCC("LF", makeByte(0x0A), pb);
     pb.createAssign(pb.createExtract(getOutput(0), 0), LF);
 }

icGREP/icgrep-devel/icgrep/kernels/linebreak_kernel.h

@@ -11 +11 @@
 
 namespace kernel {
-//#define USE_DIRECT_LF_BUILDER
+#define USE_DIRECT_LF_BUILDER
 
 class LineFeedKernelBuilder final : public pablo::PabloKernel {
 public:
-    LineFeedKernelBuilder(const std::unique_ptr<KernelBuilder> & b, unsigned codeUnitWidth = 8);
+    LineFeedKernelBuilder(const std::unique_ptr<KernelBuilder> & b, kernel::Binding && inputStreamSet);
     bool isCachable() const override { return true; }
     bool hasSignature() const override { return false; }
 protected:
     void generatePabloMethod() override;
+    unsigned mNumOfStreams;
+    unsigned mStreamFieldWidth;
 };
 

icGREP/icgrep-devel/icgrep/toolchain/grep_pipeline.cpp

@@ -26 +26 @@
 using namespace parabix;
 using namespace llvm;
+using namespace kernel;
 
 namespace grep {
@@ -64 +65 @@
     StreamSetBuffer * LineFeedStream = pxDriver.addBuffer<CircularBuffer>(idb, idb->getStreamSetTy(1, 1), segmentSize);
     #ifdef USE_DIRECT_LF_BUILDER
-    kernel::Kernel * linefeedK = pxDriver.addKernelInstance<kernel::LineFeedKernelBuilder>(idb, 8);
+    kernel::Kernel * linefeedK = pxDriver.addKernelInstance<kernel::LineFeedKernelBuilder>(idb, Binding{idb->getStreamSetTy(1, 8), "byteStream", FixedRate(), Principal()});
     pxDriver.makeKernelCall(linefeedK, {ByteStream}, {LineFeedStream});
     #endif
@@ -73 +74 @@
 
     #ifndef USE_DIRECT_LF_BUILDER
-    kernel::Kernel * linefeedK = pxDriver.addKernelInstance<kernel::LineFeedKernelBuilder>(idb, 8);
+    kernel::Kernel * linefeedK = pxDriver.addKernelInstance<kernel::LineFeedKernelBuilder>(idb, Binding{idb->getStreamSetTy(8), "basis", FixedRate(), Principal()});
     pxDriver.makeKernelCall(linefeedK, {BasisBits}, {LineFeedStream});
     #endif

icGREP/icgrep-devel/icgrep/wc.cpp

@@ -42 +42 @@
              CL_ENUM_VAL_SENTINEL), cl::cat(wcFlags), cl::Grouping);
                                                  
-
+static std::string wc_modes = "";
 
 static int defaultFieldWidth = 7;  // default field width
@@ -83 +83 @@
 class WordCountKernel final: public pablo::PabloKernel {
 public:
-    WordCountKernel(const std::unique_ptr<kernel::KernelBuilder> & b);
+    WordCountKernel(const std::unique_ptr<kernel::KernelBuilder> & b, Binding && inputStreamSet);
     bool isCachable() const override { return true; }
     bool hasSignature() const override { return false; }
@@ -90 +90 @@
 };
 
-WordCountKernel::WordCountKernel (const std::unique_ptr<kernel::KernelBuilder> & b)
-: PabloKernel(b, "wc",
-    {Binding{b->getStreamSetTy(8, 1), "u8bit"}},
+WordCountKernel::WordCountKernel (const std::unique_ptr<kernel::KernelBuilder> & b, Binding && inputStreamSet)
+: PabloKernel(b, "wc_" + wc_modes,
+    {inputStreamSet},
     {},
     {},
@@ -101 +101 @@
 void WordCountKernel::generatePabloMethod() {
     PabloBuilder pb(getEntryScope());
-    //  input: 8 basis bit streams
-    std::vector<PabloAST *> u8_bits = getInputStreamSet("u8bit");
+    std::unique_ptr<cc::CC_Compiler> ccc;
+    if (CountWords || CountChars) {
+        ccc = make_unique<cc::Parabix_CC_Compiler>(this, getInputStreamSet("u8bit"));
+    } else {
+        ccc = make_unique<cc::Direct_CC_Compiler>(this, pb.createExtract(getInput(0), pb.getInteger(0)));
+    }
+
     //  output: 3 counters
-
-    cc::Parabix_CC_Compiler ccc(this, u8_bits);
-
     Var * lc = getOutputScalarVar("lineCount");
     Var * wc = getOutputScalarVar("wordCount");
@@ -112 +114 @@
 
     if (CountLines) {
-        PabloAST * LF = ccc.compileCC(re::makeCC(0x0A));
+        PabloAST * LF = ccc->compileCC(re::makeByte(0x0A));
         pb.createAssign(lc, pb.createCount(LF));
     }
     if (CountWords) {
-        PabloAST * WS = ccc.compileCC(re::makeCC(re::makeCC(0x09, 0x0D), re::makeCC(0x20)));
+        PabloAST * WS = ccc->compileCC(re::makeCC(re::makeByte(0x09, 0x0D), re::makeByte(0x20)));
         PabloAST * wordChar = pb.createNot(WS);
         // WS_follow_or_start = 1 past WS or at start of file
@@ -128 +130 @@
         // not UTF-8, or is not valid?
         //
-        PabloAST * u8Begin = ccc.compileCC(re::makeCC(re::makeCC(0, 0x7F), re::makeCC(0xC2, 0xF4)));
+        PabloAST * u8Begin = ccc->compileCC(re::makeCC(re::makeByte(0, 0x7F), re::makeByte(0xC2, 0xF4)));
         pb.createAssign(cc, pb.createCount(u8Begin));
     }
@@ -163 +165 @@
     StreamSetBuffer * const ByteStream = pxDriver.addBuffer<SourceBuffer>(iBuilder, iBuilder->getStreamSetTy(1, 8));
 
-    StreamSetBuffer * const BasisBits = pxDriver.addBuffer<CircularBuffer>(iBuilder, iBuilder->getStreamSetTy(8, 1), segmentSize * bufferSegments);
 
     Kernel * mmapK = pxDriver.addKernelInstance<MMapSourceKernel>(iBuilder);
     mmapK->setInitialArguments({fileDecriptor});
     pxDriver.makeKernelCall(mmapK, {}, {ByteStream});
-
-    Kernel * s2pk = pxDriver.addKernelInstance<S2PKernel>(iBuilder);
-    pxDriver.makeKernelCall(s2pk, {ByteStream}, {BasisBits});
-    
-    Kernel * wck = pxDriver.addKernelInstance<WordCountKernel>(iBuilder);
-    pxDriver.makeKernelCall(wck, {BasisBits}, {});
+    
+    Kernel * wck  = nullptr;
+    if (CountWords || CountChars) {
+        StreamSetBuffer * const BasisBits = pxDriver.addBuffer<CircularBuffer>(iBuilder, iBuilder->getStreamSetTy(8, 1), segmentSize * bufferSegments);
+        Kernel * s2pk = pxDriver.addKernelInstance<S2PKernel>(iBuilder);
+        pxDriver.makeKernelCall(s2pk, {ByteStream}, {BasisBits});
+        
+        wck = pxDriver.addKernelInstance<WordCountKernel>(iBuilder, Binding{iBuilder->getStreamSetTy(8, 1), "u8bit"});
+        pxDriver.makeKernelCall(wck, {BasisBits}, {});
+
+
+    } else {
+        wck = pxDriver.addKernelInstance<WordCountKernel>(iBuilder, Binding{iBuilder->getStreamSetTy(1, 8), "u8byte"});
+        pxDriver.makeKernelCall(wck, {ByteStream}, {});
+    }
 
     pxDriver.generatePipelineIR();
@@ -217 +227 @@
                 case WordOption: CountWords = true; break;
                 case LineOption: CountLines = true; break;
-                case CharOption: CountBytes = true; CountChars = false; break;
-                case ByteOption: CountChars = true; CountBytes = false; break;
+                case CharOption: CountChars = true; CountBytes = false; break;
+                case ByteOption: CountBytes = true; CountChars = false; break;
             }
         }
     }
-    
+    if (CountLines) wc_modes += "l";
+    if (CountWords) wc_modes += "w";
+    if (CountChars) wc_modes += "m";
+    if (CountBytes) wc_modes += "c";
+
     ParabixDriver pxDriver("wc");
     wcPipelineGen(pxDriver);
@@ -243 +257 @@
     if (CountBytes) maxCount = TotalBytes;
     
+    
+    
     int fieldWidth = std::to_string(maxCount).size() + 1;
     if (fieldWidth < defaultFieldWidth) fieldWidth = defaultFieldWidth;