From a73e7b97a414d3a87833157e78d3b1aa93128572 Mon Sep 17 00:00:00 2001
From: Stefan Weil <stefan.weil@bib.uni-mannheim.de>
Date: Mon, 2 Aug 2021 13:23:30 +0200
Subject: [PATCH] Add float dotproduct implementation for NEON

Signed-off-by: Stefan Weil <stefan.weil@bib.uni-mannheim.de>
---
 Makefile.am                 |  5 +++
 src/arch/dotproduct.h       |  3 ++
 src/arch/dotproductneon.cpp | 67 +++++++++++++++++++++++++++++++++++++
 src/arch/simddetect.cpp     |  4 +--
 4 files changed, 77 insertions(+), 2 deletions(-)
 create mode 100644 src/arch/dotproductneon.cpp

diff --git a/Makefile.am b/Makefile.am
index 741cd02b..97f0d434 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -187,8 +187,13 @@ endif
 
 if HAVE_NEON
 libtesseract_neon_la_CXXFLAGS = $(NEON_CXXFLAGS)
+libtesseract_neon_la_CXXFLAGS += -O3
+if OPENMP_SIMD
+libtesseract_neon_la_CXXFLAGS += -fopenmp-simd -DOPENMP_SIMD
+endif
 libtesseract_neon_la_CXXFLAGS += -I$(top_srcdir)/src/ccutil
 libtesseract_neon_la_SOURCES = src/arch/intsimdmatrixneon.cpp
+libtesseract_neon_la_SOURCES += src/arch/dotproductneon.cpp
 libtesseract_la_LIBADD += libtesseract_neon.la
 noinst_LTLIBRARIES += libtesseract_neon.la
 endif
diff --git a/src/arch/dotproduct.h b/src/arch/dotproduct.h
index 94ba4310..4ee8ddd4 100644
--- a/src/arch/dotproduct.h
+++ b/src/arch/dotproduct.h
@@ -33,6 +33,9 @@ TFloat DotProductFMA(const TFloat *u, const TFloat *v, int n);
 // Uses Intel SSE intrinsics to access the SIMD instruction set.
 TFloat DotProductSSE(const TFloat *u, const TFloat *v, int n);
 
+// Use NEON intrinsics.
+TFloat DotProductNEON(const TFloat *u, const TFloat *v, int n);
+
 } // namespace tesseract.
 
 #endif // TESSERACT_ARCH_DOTPRODUCT_H_
diff --git a/src/arch/dotproductneon.cpp b/src/arch/dotproductneon.cpp
new file mode 100644
index 00000000..de6ea2f8
--- /dev/null
+++ b/src/arch/dotproductneon.cpp
@@ -0,0 +1,67 @@
+///////////////////////////////////////////////////////////////////////
+// File:        dotproductneon.cpp
+// Description: Dot product function for ARM NEON.
+// Author:      Stefan Weil
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+// http://www.apache.org/licenses/LICENSE-2.0
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+///////////////////////////////////////////////////////////////////////
+
+#include <arm_neon.h>
+#include "dotproduct.h"
+
+namespace tesseract {
+
+// Documentation:
+// https://developer.arm.com/architectures/instruction-sets/intrinsics/
+
+#if defined(FAST_FLOAT) && defined(__ARM_ARCH_ISA_A64)
+
+float DotProductNEON(const float *u, const float *v, int n) {
+  float32x4_t result0123 = vdupq_n_f32(0.0f);
+  float32x4_t result4567 = vdupq_n_f32(0.0f);
+  while (n > 7) {
+    // Calculate 8 dot products per iteration.
+    float32x4_t u0 = vld1q_f32(u);
+    float32x4_t v0 = vld1q_f32(v);
+    float32x4_t u4 = vld1q_f32(u + 4);
+    float32x4_t v4 = vld1q_f32(v + 4);
+    result0123 = vfmaq_f32(result0123, u0, v0);
+    result4567 = vfmaq_f32(result4567, u4, v4);
+    u += 8;
+    v += 8;
+    n -= 8;
+  }
+  float total = vaddvq_f32(result0123);
+  total += vaddvq_f32(result4567);
+  while (n > 0) {
+    total += *u++ * *v++;
+    n--;
+  }
+  return total;
+}
+
+#else
+
+// Computes and returns the dot product of the two n-vectors u and v.
+TFloat DotProductNEON(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++) {
+    total += u[k] * v[k];
+  }
+  return total;
+}
+
+#endif
+
+} // namespace tesseract
diff --git a/src/arch/simddetect.cpp b/src/arch/simddetect.cpp
index a8244c82..4bc0ba03 100644
--- a/src/arch/simddetect.cpp
+++ b/src/arch/simddetect.cpp
@@ -237,7 +237,7 @@ SIMDDetect::SIMDDetect() {
 #if defined(HAVE_NEON) || defined(__aarch64__)
   } else if (neon_available_) {
     // NEON detected.
-    SetDotProduct(DotProductNative, &IntSimdMatrix::intSimdMatrixNEON);
+    SetDotProduct(DotProductNEON, &IntSimdMatrix::intSimdMatrixNEON);
 #endif
   }
 
@@ -294,7 +294,7 @@ void SIMDDetect::Update() {
 #if defined(HAVE_NEON) || defined(__aarch64__)
   } else if (dotproduct == "neon" && neon_available_) {
     // NEON selected by config variable.
-    SetDotProduct(DotProductNative, &IntSimdMatrix::intSimdMatrixNEON);
+    SetDotProduct(DotProductNEON, &IntSimdMatrix::intSimdMatrixNEON);
     dotproduct_method = "neon";
 #endif
   } else if (dotproduct == "std::inner_product") {