Add more avx2, sse4.1 flags. Add MSVC's AVX2 ICE workaround.

Author: egorpugin

CMakeLists.txt

@@ -204,21 +204,33 @@ set(tesseract_src ${tesseract_src}
 )
 
 if (WIN32)
-    set_source_files_properties(
-        ${CMAKE_CURRENT_SOURCE_DIR}/arch/dotproductsse.cpp
-        PROPERTIES COMPILE_DEFINITIONS __SSE4_1__)
     if (MSVC)
+        set_source_files_properties(
+            ${CMAKE_CURRENT_SOURCE_DIR}/arch/dotproductsse.cpp
+            PROPERTIES COMPILE_DEFINITIONS __SSE4_1__)
+        set_source_files_properties(
+            ${CMAKE_CURRENT_SOURCE_DIR}/arch/intsimdmatrixsse.cpp
+            PROPERTIES COMPILE_DEFINITIONS __SSE4_1__)
         set_source_files_properties(
             ${CMAKE_CURRENT_SOURCE_DIR}/arch/dotproductavx.cpp
             PROPERTIES COMPILE_FLAGS "/arch:AVX")
+        set_source_files_properties(
+            ${CMAKE_CURRENT_SOURCE_DIR}/arch/intsimdmatrixavx2.cpp
+            PROPERTIES COMPILE_FLAGS "/arch:AVX2")
     endif()
 else()
     set_source_files_properties(
             ${CMAKE_CURRENT_SOURCE_DIR}/arch/dotproductsse.cpp
             PROPERTIES COMPILE_FLAGS "-msse4.1")
+    set_source_files_properties(
+            ${CMAKE_CURRENT_SOURCE_DIR}/arch/intsimdmatrixsse.cpp
+            PROPERTIES COMPILE_FLAGS "-msse4.1")
     set_source_files_properties(
             ${CMAKE_CURRENT_SOURCE_DIR}/arch/dotproductavx.cpp
             PROPERTIES COMPILE_FLAGS "-mavx")
+    set_source_files_properties(
+            ${CMAKE_CURRENT_SOURCE_DIR}/arch/intsimdmatrixavx2.cpp
+            PROPERTIES COMPILE_FLAGS "-mavx2")
 endif()
 
 add_library                     (libtesseract ${LIBRARY_TYPE} ${tesseract_src} ${tesseract_hdr})
@@ -288,69 +300,69 @@ install(FILES
 install(FILES
     # from api/makefile.am
     api/apitypes.h
-    api/baseapi.h 
-    api/capi.h 
+    api/baseapi.h
+    api/capi.h
     api/renderer.h
 
     #from arch/makefile.am
     arch/dotproductavx.h
-    arch/dotproductsse.h 
+    arch/dotproductsse.h
     arch/simddetect.h
 
     #from ccmain/makefile.am
     ccmain/thresholder.h
-    ccmain/ltrresultiterator.h 
-    ccmain/pageiterator.h 
-    ccmain/resultiterator.h 
+    ccmain/ltrresultiterator.h
+    ccmain/pageiterator.h
+    ccmain/resultiterator.h
     ccmain/osdetect.h
 
     #from ccstruct/makefile.am
-    ccstruct/publictypes.h 
+    ccstruct/publictypes.h
 
     #from ccutil/makefile.am
-    ccutil/basedir.h 
-    ccutil/errcode.h 
-    ccutil/fileerr.h 
-    ccutil/genericvector.h 
-    ccutil/helpers.h 
-    ccutil/host.h 
-    ccutil/memry.h 
+    ccutil/basedir.h
+    ccutil/errcode.h
+    ccutil/fileerr.h
+    ccutil/genericvector.h
+    ccutil/helpers.h
+    ccutil/host.h
+    ccutil/memry.h
     ccutil/ndminx.h
     ccutil/params.h
-    ccutil/ocrclass.h 
-    ccutil/platform.h 
-    ccutil/serialis.h 
-    ccutil/strngs.h 
+    ccutil/ocrclass.h
+    ccutil/platform.h
+    ccutil/serialis.h
+    ccutil/strngs.h
     ccutil/tesscallback.h
-    ccutil/unichar.h 
-    ccutil/unicharcompress.h 
-    ccutil/unicharmap.h 
+    ccutil/unichar.h
+    ccutil/unicharcompress.h
+    ccutil/unicharmap.h
     ccutil/unicharset.h
     ccutil/version.h
 
     #from lstm/makefile.am
     lstm/convolve.h
-    lstm/ctc.h 
-    lstm/fullyconnected.h 
-    lstm/functions.h 
+    lstm/ctc.h
+    lstm/fullyconnected.h
+    lstm/functions.h
     lstm/input.h
-    lstm/lstm.h 
-    lstm/lstmrecognizer.h 
-    lstm/lstmtrainer.h 
+    lstm/lstm.h
+    lstm/lstmrecognizer.h
+    lstm/lstmtrainer.h
     lstm/maxpool.h
-    lstm/networkbuilder.h 
-    lstm/network.h 
-    lstm/networkio.h 
+    lstm/networkbuilder.h
+    lstm/network.h
+    lstm/networkio.h
     lstm/networkscratch.h
-    lstm/parallel.h 
-    lstm/plumbing.h 
-    lstm/recodebeam.h 
-    lstm/reconfig.h 
+    lstm/parallel.h
+    lstm/plumbing.h
+    lstm/recodebeam.h
+    lstm/reconfig.h
     lstm/reversed.h
-    lstm/series.h 
-    lstm/static_shape.h 
-    lstm/stridemap.h 
-    lstm/tfnetwork.h 
+    lstm/series.h
+    lstm/static_shape.h
+    lstm/stridemap.h
+    lstm/tfnetwork.h
     lstm/weightmatrix.h
 
     #${CMAKE_BINARY_DIR}/src/endianness.h

arch/intsimdmatrixavx2.cpp

@@ -21,6 +21,7 @@
 #ifdef __AVX2__
 #include <immintrin.h>
 #include <stdint.h>
+#include <algorithm>
 #include <vector>
 
 namespace tesseract {
@@ -74,7 +75,15 @@ inline void ExtractResults(__m256i& result, __m256i& shift_id,
                            const int8_t*& wi, const double*& scales,
                            int num_out, double*& v) {
   for (int out = 0; out < num_out; ++out) {
-    int32_t res = _mm256_extract_epi32(result, 0);
+    int32_t res =
+#ifndef _MSC_VER
+        _mm256_extract_epi32(result, 0)
+#else
+        // Workaround MSVC's ICE
+        // _mm256_extract_epi32(X, Y) == ((int32_t*)&X)[Y]
+        ((int32_t*)&result)[0]
+#endif
+        ;
     *v++ = (static_cast<double>(res) / MAX_INT8 + *wi++) * *scales++;
     // Rotate the results in int32_t units, so the next result is ready.
     result = _mm256_permutevar8x32_epi32(result, shift_id);

cppan.yml

@@ -124,13 +124,19 @@ projects:
 
         post_sources: |
             if (WIN32)
-                set_source_files_properties(
-                    ${SDIR}/arch/dotproductsse.cpp
-                    PROPERTIES COMPILE_DEFINITIONS __SSE4_1__)
                 if (MSVC)
                     set_source_files_properties(
-                        ${SDIR}/arch/dotproductavx.cpp
+                        ${CMAKE_CURRENT_SOURCE_DIR}/arch/dotproductsse.cpp
+                        PROPERTIES COMPILE_DEFINITIONS __SSE4_1__)
+                    set_source_files_properties(
+                        ${CMAKE_CURRENT_SOURCE_DIR}/arch/intsimdmatrixsse.cpp
+                        PROPERTIES COMPILE_DEFINITIONS __SSE4_1__)
+                    set_source_files_properties(
+                        ${CMAKE_CURRENT_SOURCE_DIR}/arch/dotproductavx.cpp
                         PROPERTIES COMPILE_FLAGS "/arch:AVX")
+                    set_source_files_properties(
+                        ${CMAKE_CURRENT_SOURCE_DIR}/arch/intsimdmatrixavx2.cpp
+                        PROPERTIES COMPILE_FLAGS "/arch:AVX2")
                 endif()
             else()
                 remove_src_dir(vs2010/port/*)