Replace dynamically allocated IntSimdMatrix instances by constants

Two header files are no longer needed and could be removed.

Signed-off-by: Stefan Weil <[email protected]>

Author: stweil

CMakeLists.txt

@@ -356,8 +356,6 @@ install(FILES
     src/arch/dotproductavx.h
     src/arch/dotproductsse.h
     src/arch/intsimdmatrix.h
-    src/arch/intsimdmatrixavx2.h
-    src/arch/intsimdmatrixsse.h
     src/arch/simddetect.h
 
     #from ccmain/makefile.am

src/arch/Makefile.am

@@ -11,7 +11,7 @@ endif
 pkginclude_HEADERS =
 
 noinst_HEADERS = dotproduct.h dotproductavx.h dotproductsse.h
-noinst_HEADERS += intsimdmatrix.h intsimdmatrixavx2.h intsimdmatrixsse.h
+noinst_HEADERS += intsimdmatrix.h
 noinst_HEADERS += simddetect.h
 
 noinst_LTLIBRARIES = libtesseract_native.la

src/arch/intsimdmatrix.cpp

@@ -18,25 +18,26 @@
 
 #include "intsimdmatrix.h"
 #include "genericvector.h"      // for GenericVector
-#include "intsimdmatrixavx2.h"  // for IntSimdMatrixAVX2
-#include "intsimdmatrixsse.h"   // for IntSimdMatrixSSE
 #include "matrix.h"             // for GENERIC_2D_ARRAY
 #include "simddetect.h"         // for SIMDDetect
 
 namespace tesseract {
 
+const IntSimdMatrix IntSimdMatrix::IntSimdMatrixNative =
+  IntSimdMatrix(1, 1, 1, 1, 1, {});
+
 // Factory makes and returns an IntSimdMatrix (sub)class of the best
 // available type for the current architecture.
 /* static */
 const IntSimdMatrix* IntSimdMatrix::GetFastestMultiplier() {
   const IntSimdMatrix* multiplier;
   if (SIMDDetect::IsAVX2Available()) {
-    multiplier = new IntSimdMatrixAVX2();
+    multiplier = &IntSimdMatrixAVX2;
   } else if (SIMDDetect::IsSSEAvailable()) {
-    multiplier = new IntSimdMatrixSSE();
+    multiplier = &IntSimdMatrixSSE;
   } else {
     // Default c++ implementation.
-    multiplier = new IntSimdMatrix(1, 1, 1, 1, 1, {});
+    multiplier = &IntSimdMatrixNative;
   }
   return multiplier;
 }

src/arch/intsimdmatrix.h

@@ -115,6 +115,10 @@ class IntSimdMatrix {
                        const GenericVector<double>& scales, const int8_t* u,
                        double* v) const;
 
+  static const IntSimdMatrix IntSimdMatrixAVX2;
+  static const IntSimdMatrix IntSimdMatrixSSE;
+  static const IntSimdMatrix IntSimdMatrixNative;
+
  protected:
   // Rounds the input up to a multiple of the given factor.
   static int Roundup(int input, int factor) {

src/arch/intsimdmatrixavx2.cpp

@@ -16,7 +16,7 @@
 // limitations under the License.
 ///////////////////////////////////////////////////////////////////////
 
-#include "intsimdmatrixavx2.h"
+#include "intsimdmatrix.h"
 
 #ifdef __AVX2__
 #include <immintrin.h>
@@ -269,14 +269,12 @@ static void PartialMatrixDotVector8(const int8_t* wi, const double* scales,
 namespace tesseract {
 #endif  // __AVX2__
 
-IntSimdMatrixAVX2::IntSimdMatrixAVX2()
 #ifdef __AVX2__
-: IntSimdMatrix(kNumOutputsPerRegister, kMaxOutputRegisters, kNumInputsPerRegister, kNumInputsPerGroup, kNumInputGroups, {PartialMatrixDotVector64, PartialMatrixDotVector32,
-   PartialMatrixDotVector16, PartialMatrixDotVector8})
+const IntSimdMatrix IntSimdMatrix::IntSimdMatrixAVX2 =
+  IntSimdMatrix(kNumOutputsPerRegister, kMaxOutputRegisters, kNumInputsPerRegister, kNumInputsPerGroup, kNumInputGroups, {PartialMatrixDotVector64, PartialMatrixDotVector32,
+   PartialMatrixDotVector16, PartialMatrixDotVector8});
 #else
-: IntSimdMatrix(1, 1, 1, 1, 1, {})
+const IntSimdMatrix IntSimdMatrix::IntSimdMatrixAVX2 = IntSimdMatrix(1, 1, 1, 1, 1, {});
 #endif  // __AVX2__
-{
-}
 
 }  // namespace tesseract.

src/arch/intsimdmatrixavx2.h

@@ -15,7 +15,7 @@
 // limitations under the License.
 ///////////////////////////////////////////////////////////////////////
 
-#include "intsimdmatrixsse.h"
+#include "intsimdmatrix.h"
 
 #include <cstdint>
 #include "dotproductsse.h"
@@ -33,13 +33,12 @@ static void PartialMatrixDotVector1(const int8_t* wi, const double* scales,
 }
 #endif  // __SSE4_1__
 
-IntSimdMatrixSSE::IntSimdMatrixSSE()
 #ifdef __SSE4_1__
-: IntSimdMatrix(1, 1, 1, 1, 1, {PartialMatrixDotVector1})
+const IntSimdMatrix IntSimdMatrix::IntSimdMatrixSSE =
+  IntSimdMatrix(1, 1, 1, 1, 1, {PartialMatrixDotVector1});
 #else
-: IntSimdMatrix(1, 1, 1, 1, 1, {})
+const IntSimdMatrix IntSimdMatrix::IntSimdMatrixSSE =
+  IntSimdMatrix(1, 1, 1, 1, 1, {});
 #endif  // __SSE4_1__
-{
-}
 
 }  // namespace tesseract.

src/arch/intsimdmatrixsse.cpp

@@ -143,7 +143,7 @@ void WeightMatrix::ConvertToInt() {
   }
   wf_.Resize(1, 1, 0.0);
   int_mode_ = true;
-  multiplier_.reset(IntSimdMatrix::GetFastestMultiplier());
+  multiplier_ = IntSimdMatrix::GetFastestMultiplier();
   multiplier_->Init(wi_, shaped_w_);
 }
 
@@ -196,7 +196,7 @@ bool WeightMatrix::DeSerialize(bool training, TFile* fp) {
   if (int_mode_) {
     if (!wi_.DeSerialize(fp)) return false;
     if (!scales_.DeSerialize(fp)) return false;
-    multiplier_.reset(IntSimdMatrix::GetFastestMultiplier());
+    multiplier_ = IntSimdMatrix::GetFastestMultiplier();
     multiplier_->Init(wi_, shaped_w_);
   } else {
     if (!wf_.DeSerialize(fp)) return false;

src/arch/intsimdmatrixsse.h

@@ -64,7 +64,7 @@ class TransposedArray : public GENERIC_2D_ARRAY<double> {
 // backward steps with the matrix and updates to the weights.
 class WeightMatrix {
  public:
-  WeightMatrix() : int_mode_(false), use_adam_(false) {}
+  WeightMatrix() : int_mode_(false), use_adam_(false), multiplier_(nullptr) {}
   // Sets up the network for training. Initializes weights using weights of
   // scale `range` picked according to the random number generator `randomizer`.
   // Note the order is outputs, inputs, as this is the order of indices to
@@ -179,7 +179,7 @@ class WeightMatrix {
   // The weights matrix reorganized in whatever way suits this instance.
   std::vector<int8_t> shaped_w_;
   // Holds the optimal integer multiplier for this machine.
-  std::unique_ptr<const IntSimdMatrix> multiplier_;
+  const IntSimdMatrix* multiplier_;
 };
 
 }  // namespace tesseract.

src/lstm/weightmatrix.cpp

@@ -18,8 +18,6 @@
 #include <memory>
 #include "genericvector.h"
 #include "include_gunit.h"
-#include "intsimdmatrixavx2.h"
-#include "intsimdmatrixsse.h"
 #include "matrix.h"
 #include "simddetect.h"
 #include "tprintf.h"
@@ -56,21 +54,21 @@ class IntSimdMatrixTest : public ::testing::Test {
     }
     return v;
   }
-  // Tests a range of sizes and compares the results against the base_ version.
-  void ExpectEqualResults(const IntSimdMatrix* matrix) {
+  // Tests a range of sizes and compares the results against the generic version.
+  void ExpectEqualResults(const IntSimdMatrix& matrix) {
     double total = 0.0;
     for (int num_out = 1; num_out < 130; ++num_out) {
       for (int num_in = 1; num_in < 130; ++num_in) {
         GENERIC_2D_ARRAY<int8_t> w = InitRandom(num_out, num_in + 1);
-        std::vector<int8_t> u = RandomVector(num_in, *matrix);
+        std::vector<int8_t> u = RandomVector(num_in, matrix);
         GenericVector<double> scales = RandomScales(num_out);
         std::vector<double> base_result(num_out);
         std::vector<int8_t> dummy;
-        base_.MatrixDotVector(w, dummy, scales, u.data(), base_result.data());
+        IntSimdMatrix::IntSimdMatrixNative.MatrixDotVector(w, dummy, scales, u.data(), base_result.data());
         std::vector<double> test_result(num_out);
         std::vector<int8_t> shaped_wi;
-        matrix->Init(w, shaped_wi);
-        matrix->MatrixDotVector(w, shaped_wi, scales, u.data(), test_result.data());
+        matrix.Init(w, shaped_wi);
+        matrix.MatrixDotVector(w, shaped_wi, scales, u.data(), test_result.data());
         for (int i = 0; i < num_out; ++i) {
           EXPECT_FLOAT_EQ(base_result[i], test_result[i]) << "i=" << i;
           total += base_result[i];
@@ -82,13 +80,12 @@ class IntSimdMatrixTest : public ::testing::Test {
   }
 
   TRand random_;
-  IntSimdMatrix base_ = IntSimdMatrix(1, 1, 1, 1, 1, {});
 };
 
 // Test the C++ implementation without SIMD.
 TEST_F(IntSimdMatrixTest, C) {
-  std::unique_ptr<IntSimdMatrix> matrix(new IntSimdMatrix());
-  ExpectEqualResults(matrix.get());
+  static const IntSimdMatrix matrix(1, 1, 1, 1, 1, {});
+  ExpectEqualResults(matrix);
 }
 
 // Tests that the SSE implementation gets the same result as the vanilla.
@@ -99,8 +96,7 @@ TEST_F(IntSimdMatrixTest, SSE) {
     tprintf("No SSE found! Not Tested!");
     return;
   }
-  std::unique_ptr<IntSimdMatrix> matrix(new IntSimdMatrixSSE());
-  ExpectEqualResults(matrix.get());
+  ExpectEqualResults(IntSimdMatrix::IntSimdMatrixSSE);
 }
 
 // Tests that the AVX2 implementation gets the same result as the vanilla.
@@ -111,8 +107,7 @@ TEST_F(IntSimdMatrixTest, AVX2) {
     tprintf("No AVX2 found! Not Tested!");
     return;
   }
-  std::unique_ptr<IntSimdMatrix> matrix(new IntSimdMatrixAVX2());
-  ExpectEqualResults(matrix.get());
+  ExpectEqualResults(IntSimdMatrix::IntSimdMatrixAVX2);
 }
 
 }  // namespace