Tfloat float and double coexistance -- working towards that goal

Makefile.am

@@ -147,41 +147,53 @@ libtesseract_native_la_CXXFLAGS = -O3 -ffast-math
 if OPENMP_SIMD
 libtesseract_native_la_CXXFLAGS += -fopenmp-simd -DOPENMP_SIMD
 endif
+if HAVE_AVX
+libtesseract_native_la_CXXFLAGS += -mavx
+endif
 if MARCH_NATIVE_OPT
 libtesseract_native_la_CXXFLAGS += -march=native -mtune=native
 endif
+libtesseract_native_la_CXXFLAGS += -I$(top_srcdir)/src/ccutil
 libtesseract_native_la_SOURCES = src/arch/dotproduct.cpp
 
 if HAVE_AVX
 libtesseract_avx_la_CXXFLAGS = -mavx
+libtesseract_avx_la_CXXFLAGS += -I$(top_srcdir)/src/ccutil
 libtesseract_avx_la_SOURCES = src/arch/dotproductavx.cpp
 libtesseract_la_LIBADD += libtesseract_avx.la
 noinst_LTLIBRARIES += libtesseract_avx.la
 endif
 
 if HAVE_AVX2
 libtesseract_avx2_la_CXXFLAGS = -mavx2
+libtesseract_avx2_la_CXXFLAGS += -I$(top_srcdir)/src/ccutil
 libtesseract_avx2_la_SOURCES = src/arch/intsimdmatrixavx2.cpp
 libtesseract_la_LIBADD += libtesseract_avx2.la
 noinst_LTLIBRARIES += libtesseract_avx2.la
 endif
 
 if HAVE_FMA
 libtesseract_fma_la_CXXFLAGS = -mfma
+libtesseract_fma_la_CXXFLAGS += -I$(top_srcdir)/src/ccutil
 libtesseract_fma_la_SOURCES = src/arch/dotproductfma.cpp
 libtesseract_la_LIBADD += libtesseract_fma.la
 noinst_LTLIBRARIES += libtesseract_fma.la
 endif
 
 if HAVE_SSE4_1
 libtesseract_sse_la_CXXFLAGS = -msse4.1
+libtesseract_sse_la_CXXFLAGS += -I$(top_srcdir)/src/ccutil
+if OPENMP_SIMD
+libtesseract_sse_la_CXXFLAGS += -fopenmp-simd -DOPENMP_SIMD
+endif
 libtesseract_sse_la_SOURCES = src/arch/dotproductsse.cpp src/arch/intsimdmatrixsse.cpp
 libtesseract_la_LIBADD += libtesseract_sse.la
 noinst_LTLIBRARIES += libtesseract_sse.la
 endif
 
 if HAVE_NEON
 libtesseract_neon_la_CXXFLAGS = $(NEON_CXXFLAGS)
+libtesseract_neon_la_CXXFLAGS += -I$(top_srcdir)/src/ccutil
 libtesseract_neon_la_SOURCES = src/arch/intsimdmatrixneon.cpp
 libtesseract_la_LIBADD += libtesseract_neon.la
 noinst_LTLIBRARIES += libtesseract_neon.la
@@ -1233,6 +1245,7 @@ check_PROGRAMS += commandlineflags_test
 check_PROGRAMS += dawg_test
 endif # ENABLE_TRAINING
 check_PROGRAMS += denorm_test
+check_PROGRAMS += dotproduct_test
 if !DISABLED_LEGACY_ENGINE
 check_PROGRAMS += equationdetect_test
 endif # !DISABLED_LEGACY_ENGINE
@@ -1359,6 +1372,16 @@ denorm_test_SOURCES = unittest/denorm_test.cc
 denorm_test_CPPFLAGS = $(unittest_CPPFLAGS)
 denorm_test_LDADD = $(TESS_LIBS)
 
+dotproduct_test_SOURCES = unittest/dotproduct_test.cc
+dotproduct_test_CPPFLAGS = $(unittest_CPPFLAGS)
+if HAVE_AVX2
+dotproduct_test_CPPFLAGS += -DHAVE_AVX2
+endif
+if HAVE_SSE4_1
+dotproduct_test_CPPFLAGS += -DHAVE_SSE4_1
+endif
+dotproduct_test_LDADD = $(TESS_LIBS)
+
 if !DISABLED_LEGACY_ENGINE
 equationdetect_test_SOURCES = unittest/equationdetect_test.cc
 equationdetect_test_CPPFLAGS = $(unittest_CPPFLAGS)

configure.ac

@@ -297,13 +297,14 @@ OPENCL_CPPFLAGS=''
 OPENCL_LDFLAGS=''
 case "${host_os}" in
   *darwin* | *-macos10*)
-    echo "checking for OpenCL framework"
-    MY_CHECK_FRAMEWORK([OpenCL])
-    if test $my_cv_framework_OpenCL = yes; then
-       have_opencl_lib=true
+    MY_CHECK_FRAMEWORK([Accelerate])
+    if test $my_cv_framework_Accelerate = yes; then
+      AM_CPPFLAGS="-DHAVE_FRAMEWORK_ACCELERATE $AM_CPPFLAGS"
+      LDFLAGS="$LDFLAGS -framework Accelerate"
     fi
+    MY_CHECK_FRAMEWORK([OpenCL])
     if test "$enable_opencl" = "yes"; then
-      if !($have_opencl_lib); then
+      if test $my_cv_framework_OpenCL = no; then
         AC_MSG_ERROR([Required OpenCL library not found!])
       fi
       AM_CPPFLAGS="-DUSE_OPENCL $AM_CPPFLAGS"

src/api/hocrrenderer.cpp

@@ -319,7 +319,6 @@ char *TessBaseAPI::GetHOCRText(ETEXT_DESC *monitor, int page_number) {
             hocr_str << "\n        </span>";
             ++scnt;
           } else if (lstm_choice_mode == 2) {
-            tesseract::ChoiceIterator ci(*res_it);
             hocr_str << "\n        <span class='ocrx_cinfo'"
                      << " id='"
                      << "lstm_choices_" << page_id << "_" << wcnt << "_" << tcnt << "'>";

src/arch/dotproduct.cpp

@@ -19,15 +19,20 @@
 namespace tesseract {
 
 // Computes and returns the dot product of the two n-vectors u and v.
-double DotProductNative(const double *u, const double *v, int n) {
-  double total = 0.0;
+template <class TFloat>
+TFloat DotProductNative(const TFloat *u, const TFloat *v, int n) {
+  TFloat total = 0;
 #if defined(OPENMP_SIMD) || defined(_OPENMP)
 #pragma omp simd reduction(+:total)
 #endif
-  for (int k = 0; k < n; ++k) {
+  for (int k = 0; k < n; k++) {
     total += u[k] * v[k];
   }
   return total;
 }
 
+// two instantiations: float & double.
+template float DotProductNative<float>(const float *u, const float *v, int n);
+template double DotProductNative<double>(const double *u, const double *v, int n);
+
 } // namespace tesseract

src/arch/dotproduct.h

@@ -17,20 +17,48 @@
 #ifndef TESSERACT_ARCH_DOTPRODUCT_H_
 #define TESSERACT_ARCH_DOTPRODUCT_H_
 
+#include "tfloat.h"
+
 namespace tesseract {
 
 // Computes and returns the dot product of the n-vectors u and v.
-double DotProductNative(const double *u, const double *v, int n);
+template <class TFloat>
+TFloat DotProductNative(const TFloat *u, const TFloat *v, int n);
+
+// ------------ FAST FLOAT specializations -----------------
+
+// Uses Intel AVX intrinsics to access the SIMD instruction set.
+float DotProductAVX(const float *u, const float *v, int n);
+float DotProductAVX1(const float *u, const float *v, int n);
+float DotProductAVX2(const float *u, const float *v, int n);
+float DotProductAVX3(const float *u, const float *v, int n);
+float DotProductAVX4(const float *u, const float *v, int n);
+
+// Use Intel FMA.
+float DotProductFMA(const float *u, const float *v, int n);
+
+// Uses Intel SSE intrinsics to access the SIMD instruction set.
+float DotProductSSE(const float *u, const float *v, int n);
+
+float DotProductAccelerate(const float *u, const float *v, int n);
+
+// ------------ HIGH PRECISION DOUBLE specializations -----------------
 
 // Uses Intel AVX intrinsics to access the SIMD instruction set.
 double DotProductAVX(const double *u, const double *v, int n);
+double DotProductAVX1(const double *u, const double *v, int n);
+double DotProductAVX2(const double *u, const double *v, int n);
+double DotProductAVX3(const double *u, const double *v, int n);
+double DotProductAVX4(const double *u, const double *v, int n);
 
 // Use Intel FMA.
 double DotProductFMA(const double *u, const double *v, int n);
 
 // Uses Intel SSE intrinsics to access the SIMD instruction set.
 double DotProductSSE(const double *u, const double *v, int n);
 
+double DotProductAccelerate(const double *u, const double *v, int n);
+
 } // namespace tesseract.
 
 #endif // TESSERACT_ARCH_DOTPRODUCT_H_

src/arch/dotproductavx.cpp

@@ -15,20 +15,95 @@
 // limitations under the License.
 ///////////////////////////////////////////////////////////////////////
 
-#if !defined(__AVX__)
-#  if defined(__i686__) || defined(__x86_64__)
-#    error Implementation only for AVX capable architectures
-#  endif
-#else
+#include "intsimdmatrix.h"
+
+#if defined(__AVX__)
 
 #  include <immintrin.h>
 #  include <cstdint>
 #  include "dotproduct.h"
 
 namespace tesseract {
 
+// ---------------------------- FAST FLOAT section ------------------------
+
 // Computes and returns the dot product of the n-vectors u and v.
 // Uses Intel AVX intrinsics to access the SIMD instruction set.
+float DotProductAVX(const float *u, const float *v, int n) {
+  const unsigned quot = n / 8;
+  const unsigned rem = n % 8;
+  __m256 t0 = _mm256_setzero_ps();
+  for (unsigned k = 0; k < quot; k++) {
+    __m256 f0 = _mm256_loadu_ps(u);
+    __m256 f1 = _mm256_loadu_ps(v);
+    f0 = _mm256_mul_ps(f0, f1);
+    t0 = _mm256_add_ps(t0, f0);
+    u += 8;
+    v += 8;
+  }
+  alignas(32) float tmp[8];
+  _mm256_store_ps(tmp, t0);
+  float result = tmp[0] + tmp[1] + tmp[2] + tmp[3] + tmp[4] + tmp[5] + tmp[6] + tmp[7];
+  for (unsigned k = 0; k < rem; k++) {
+    result += *u++ * *v++;
+  }
+  return result;
+}
+
+float DotProductAVX1(const float *u, const float *v, int n) {
+  const unsigned quot = n / 16;
+  const unsigned rem = n % 16;
+  __m256 t0 = _mm256_setzero_ps();
+  __m256 t1 = _mm256_setzero_ps();
+  for (unsigned k = 0; k < quot; k++) {
+    __m256 f0 = _mm256_loadu_ps(u);
+    __m256 f1 = _mm256_loadu_ps(v);
+    __m256 f2 = _mm256_loadu_ps(u + 8);
+    __m256 f3 = _mm256_loadu_ps(v + 8);
+    f0 = _mm256_mul_ps(f0, f1);
+    f2 = _mm256_mul_ps(f2, f3);
+    t0 = _mm256_add_ps(t0, f0);
+    t1 = _mm256_add_ps(t1, f2);
+    u += 16;
+    v += 16;
+  }
+  t0 = _mm256_hadd_ps(t0, t1);
+  alignas(32) float tmp[8];
+  _mm256_store_ps(tmp, t0);
+  float result = tmp[0] + tmp[1] + tmp[2] + tmp[3] + tmp[4] + tmp[5] + tmp[6] + tmp[7];
+  for (unsigned k = 0; k < rem; k++) {
+    result += *u++ * *v++;
+  }
+  return result;
+}
+
+// ---------------------------- HIGH-PRECISION DOUBLE section ------------------------
+
+double DotProductAVX1(const double *u, const double *v, int n) {
+  __m256d t0 = _mm256_setzero_pd();
+  __m256d t1 = _mm256_setzero_pd();
+  for (unsigned quot = n / 8; quot > 0; quot--) {
+    __m256d f0 = _mm256_loadu_pd(u);
+    __m256d f1 = _mm256_loadu_pd(v);
+    __m256d f2 = _mm256_loadu_pd(u + 4);
+    __m256d f3 = _mm256_loadu_pd(v + 4);
+    f0 = _mm256_mul_pd(f0, f1);
+    f2 = _mm256_mul_pd(f2, f3);
+    t0 = _mm256_add_pd(t0, f0);
+    t1 = _mm256_add_pd(t1, f2);
+    u += 8;
+    v += 8;
+  }
+  t0 = _mm256_hadd_pd(t0, t1);
+  alignas(32) double tmp[4];
+  _mm256_store_pd(tmp, t0);
+  double result = tmp[0] + tmp[1] + tmp[2] + tmp[3];
+  for (unsigned rem = n % 8; rem > 0; rem--) {
+    result += *u++ * *v++;
+  }
+  return result;
+}
+
 double DotProductAVX(const double *u, const double *v, int n) {
   const unsigned quot = n / 8;
   const unsigned rem = n % 8;
@@ -58,6 +133,20 @@ double DotProductAVX(const double *u, const double *v, int n) {
   return result;
 }
 
+// ---------------------------- END FLOAT/DOUBLE sections ------------------------
+
 } // namespace tesseract.
 
+#else
+
+namespace tesseract {
+
+	// Computes and returns the dot product of the n-vectors u and v.
+	// Uses Intel FMA intrinsics to access the SIMD instruction set.
+	inline TFloat DotProductAVX(const TFloat* u, const TFloat* v, int n) {
+		return DotProductFMA(u, v, n);
+	}
+
+}
+
 #endif

src/arch/dotproductfma.cpp

@@ -15,20 +15,47 @@
 // limitations under the License.
 ///////////////////////////////////////////////////////////////////////
 
-#if !defined(__FMA__)
-#  if defined(__i686__) || defined(__x86_64__)
-#    error Implementation only for FMA capable architectures
-#  endif
-#else
+#if defined(__FMA__)
 
 #  include <immintrin.h>
 #  include <cstdint>
 #  include "dotproduct.h"
 
 namespace tesseract {
 
+// ---------------------------- FAST FLOAT section ------------------------
+
 // Computes and returns the dot product of the n-vectors u and v.
 // Uses Intel FMA intrinsics to access the SIMD instruction set.
+float DotProductFMA(const float *u, const float *v, int n) {
+  const unsigned quot = n / 16;
+  const unsigned rem = n % 16;
+  __m256 t0 = _mm256_setzero_ps();
+  __m256 t1 = _mm256_setzero_ps();
+  for (unsigned k = 0; k < quot; k++) {
+    __m256 f0 = _mm256_loadu_ps(u);
+    __m256 f1 = _mm256_loadu_ps(v);
+    t0 = _mm256_fmadd_ps(f0, f1, t0);
+    u += 8;
+    v += 8;
+    __m256 f2 = _mm256_loadu_ps(u);
+    __m256 f3 = _mm256_loadu_ps(v);
+    t1 = _mm256_fmadd_ps(f2, f3, t1);
+    u += 8;
+    v += 8;
+  }
+  t0 = _mm256_hadd_ps(t0, t1);
+  alignas(32) float tmp[8];
+  _mm256_store_ps(tmp, t0);
+  float result = tmp[0] + tmp[1] + tmp[2] + tmp[3] + tmp[4] + tmp[5] + tmp[6] + tmp[7];
+  for (unsigned k = 0; k < rem; k++) {
+    result += *u++ * *v++;
+  }
+  return result;
+}
+
+// ---------------------------- HIGH-PRECISION DOUBLE section ------------------------
+
 double DotProductFMA(const double *u, const double *v, int n) {
   const unsigned quot = n / 8;
   const unsigned rem = n % 8;
@@ -56,6 +83,23 @@ double DotProductFMA(const double *u, const double *v, int n) {
   return result;
 }
 
+// ---------------------------- END section ------------------------
+
 } // namespace tesseract.
 
+#else
+
+namespace tesseract {
+
+// Computes and returns the dot product of the n-vectors u and v.
+// Uses Intel FMA intrinsics to access the SIMD instruction set.
+inline float DotProductFMA(const float *u, const float *v, int n) {
+	return DotProductSSE(u, v, n);
+}
+inline double DotProductFMA(const double *u, const double *v, int n) {
+  return DotProductSSE(u, v, n);
+}
+
+}
+
 #endif