T1 & DWT multithreading decoding optimizations

.travis.yml

@@ -8,6 +8,9 @@ matrix:
     - os: linux
       compiler: gcc
       env: OPJ_CI_ARCH=x86_64 OPJ_CI_BUILD_CONFIGURATION=Release OPJ_CI_INCLUDE_IF_DEPLOY=1
+    - os: linux
+      compiler: gcc
+      env: OPJ_CI_ARCH=x86_64 OPJ_CI_BUILD_CONFIGURATION=Release OPJ_NUM_THREADS=2
     - os: linux
       compiler: gcc
       env: OPJ_CI_ARCH=i386 OPJ_CI_BUILD_CONFIGURATION=Release

src/bin/jp2/CMakeLists.txt

@@ -57,6 +57,9 @@ foreach(exe opj_decompress opj_compress opj_dump)
   # On unix you need to link to the math library:
   if(UNIX)
     target_link_libraries(${exe} m)
+    IF("${CMAKE_SYSTEM_NAME}" MATCHES "Linux")
+      target_link_libraries(${exe} rt)
+    endif()
   endif()
   # Install exe
   install(TARGETS ${exe}

src/bin/jp2/opj_decompress.c

@@ -43,6 +43,7 @@
 #include <string.h>
 #include <stdlib.h>
 #include <math.h>
+#include <time.h>
 
 #ifdef _WIN32
 #include "windirent.h"
@@ -150,6 +151,8 @@ typedef struct opj_decompress_params
 	int upsample;
 	/* split output components to different files */
 	int split_pnm;
+    /** number of threads */
+    int num_threads;
 }opj_decompress_parameters;
 
 /* -------------------------------------------------------------------------- */
@@ -224,8 +227,11 @@ static void decode_help_display(void) {
 	               "  -upsample\n"
 	               "    Downsampled components will be upsampled to image size\n"
 	               "  -split-pnm\n"
-	               "    Split output components to different files when writing to PNM\n"
-	               "\n");
+	               "    Split output components to different files when writing to PNM\n");
+	if( opj_has_thread_support() ) {
+	  fprintf(stdout,"  -threads <num_threads>\n"
+					"    Number of threads to use for decoding.\n");
+	}
 /* UniPG>> */
 #ifdef USE_JPWL
 	fprintf(stdout,"  -W <options>\n"
@@ -520,7 +526,8 @@ int parse_cmdline_decoder(int argc, char **argv, opj_decompress_parameters *para
 		{"OutFor",    REQ_ARG, NULL,'O'},
 		{"force-rgb", NO_ARG,  NULL, 1},
 		{"upsample",  NO_ARG,  NULL, 1},
-		{"split-pnm", NO_ARG,  NULL, 1}
+		{"split-pnm", NO_ARG,  NULL, 1},
+		{"threads",   REQ_ARG, NULL, 'T'}
 	};
 
 	const char optlist[] = "i:o:r:l:x:d:t:p:"
@@ -808,6 +815,22 @@ int parse_cmdline_decoder(int argc, char **argv, opj_decompress_parameters *para
 			break;	
 #endif /* USE_JPWL */
 /* <<UniPG */            
+
+				/* ----------------------------------------------------- */
+			case 'T':  /* Number of threads */
+				{
+					if( strcmp(opj_optarg, "ALL_CPUS") == 0 )
+					{
+						parameters->num_threads = opj_get_num_cpus();
+						if( parameters->num_threads == 1 )
+							parameters->num_threads = 0;
+					}
+					else
+					{
+					  sscanf(opj_optarg, "%d", &parameters->num_threads);
+					}
+				}
+				break;
 
 				/* ----------------------------------------------------- */
 
@@ -885,17 +908,22 @@ OPJ_FLOAT64 opj_clock(void) {
     /* t is the high resolution performance counter (see MSDN) */
     QueryPerformanceCounter ( & t ) ;
 	return freq.QuadPart ? (t.QuadPart / (OPJ_FLOAT64)freq.QuadPart) : 0;
+#elif defined(__linux)
+	struct timespec ts;
+	clock_gettime(CLOCK_REALTIME, &ts);
+	return( ts.tv_sec + ts.tv_nsec * 1e-9 );
 #else
-	/* Unix or Linux: use resource usage */
-    struct rusage t;
-    OPJ_FLOAT64 procTime;
-    /* (1) Get the rusage data structure at this moment (man getrusage) */
-    getrusage(0,&t);
-    /* (2) What is the elapsed time ? - CPU time = User time + System time */
+	/* Unix : use resource usage */
+	/* FIXME: this counts the total CPU time, instead of the user perceived time */
+	struct rusage t;
+	OPJ_FLOAT64 procTime;
+	/* (1) Get the rusage data structure at this moment (man getrusage) */
+	getrusage(0,&t);
+	/* (2) What is the elapsed time ? - CPU time = User time + System time */
 	/* (2a) Get the seconds */
-    procTime = (OPJ_FLOAT64)(t.ru_utime.tv_sec + t.ru_stime.tv_sec);
-    /* (2b) More precisely! Get the microseconds part ! */
-    return ( procTime + (OPJ_FLOAT64)(t.ru_utime.tv_usec + t.ru_stime.tv_usec) * 1e-6 ) ;
+	procTime = (OPJ_FLOAT64)(t.ru_utime.tv_sec + t.ru_stime.tv_sec);
+	/* (2b) More precisely! Get the microseconds part ! */
+	return ( procTime + (OPJ_FLOAT64)(t.ru_utime.tv_usec + t.ru_stime.tv_usec) * 1e-6 ) ;
 #endif
 }
 
@@ -1306,7 +1334,13 @@ int main(int argc, char **argv)
 			opj_destroy_codec(l_codec);
 			failed = 1; goto fin;
 		}
-
+
+		if( parameters.num_threads >= 1 && !opj_codec_set_threads(l_codec, parameters.num_threads) ) {
+			fprintf(stderr, "ERROR -> opj_decompress: failed to set number of threads\n");
+			opj_stream_destroy(l_stream);
+			opj_destroy_codec(l_codec);
+			failed = 1; goto fin;
+		}
 
 		/* Read the main header of the codestream and if necessary the JP2 boxes*/
 		if(! opj_read_header(l_stream, l_codec, &image)){

src/lib/openjp2/CMakeLists.txt

@@ -9,6 +9,8 @@ include_directories(
 )
 # Defines the source code for the library
 set(OPENJPEG_SRCS
+  ${CMAKE_CURRENT_SOURCE_DIR}/thread.c
+  ${CMAKE_CURRENT_SOURCE_DIR}/thread.h
   ${CMAKE_CURRENT_SOURCE_DIR}/bio.c
   ${CMAKE_CURRENT_SOURCE_DIR}/bio.h
   ${CMAKE_CURRENT_SOURCE_DIR}/cio.c
@@ -29,6 +31,7 @@ set(OPENJPEG_SRCS
   ${CMAKE_CURRENT_SOURCE_DIR}/mct.h
   ${CMAKE_CURRENT_SOURCE_DIR}/mqc.c
   ${CMAKE_CURRENT_SOURCE_DIR}/mqc.h
+  ${CMAKE_CURRENT_SOURCE_DIR}/mqc_inl.h
   ${CMAKE_CURRENT_SOURCE_DIR}/openjpeg.c
   ${CMAKE_CURRENT_SOURCE_DIR}/openjpeg.h
   ${CMAKE_CURRENT_SOURCE_DIR}/opj_clock.c
@@ -73,6 +76,11 @@ if(OPJ_DISABLE_TPSOT_FIX)
   add_definitions(-DOPJ_DISABLE_TPSOT_FIX)
 endif()
 
+# Special case for old i586-mingw32msvc-gcc cross compiler
+if(NOT WIN32 AND CMAKE_COMPILER_IS_GNUCC AND CMAKE_C_COMPILER MATCHES ".*mingw32msvc.*" )
+  set(WIN32 YES)
+endif()
+
 # Build the library
 if(WIN32)
   if(BUILD_SHARED_LIBS)
@@ -142,3 +150,36 @@ if(OPJ_USE_DSYMUTIL)
     DEPENDS ${OPENJPEG_LIBRARY_NAME})
   endif()
 endif()
+
+#################################################################################
+# threading configuration
+#################################################################################
+set(CMAKE_THREAD_PREFER_PTHREAD TRUE)
+
+option(USE_THREAD "Build with thread/mutex support " ON)
+if(NOT USE_THREAD)
+   add_definitions( -DMUTEX_stub)
+endif(NOT USE_THREAD)
+
+find_package(Threads QUIET)
+
+if(USE_THREAD AND WIN32 AND NOT Threads_FOUND )
+    add_definitions( -DMUTEX_win32)
+    set(Threads_FOUND YES)
+endif()
+
+if(USE_THREAD AND Threads_FOUND AND CMAKE_USE_WIN32_THREADS_INIT )
+   add_definitions( -DMUTEX_win32)
+endif(USE_THREAD AND Threads_FOUND AND CMAKE_USE_WIN32_THREADS_INIT )
+
+if(USE_THREAD AND Threads_FOUND AND CMAKE_USE_PTHREADS_INIT )
+   add_definitions( -DMUTEX_pthread)
+endif(USE_THREAD AND Threads_FOUND AND CMAKE_USE_PTHREADS_INIT )
+
+if(USE_THREAD AND NOT Threads_FOUND)
+  message(FATAL_ERROR "No thread library found and thread/mutex support is required by USE_THREAD option")
+endif(USE_THREAD AND NOT Threads_FOUND)
+
+if(USE_THREAD AND Threads_FOUND AND CMAKE_USE_PTHREADS_INIT)
+   TARGET_LINK_LIBRARIES(${OPENJPEG_LIBRARY_NAME} ${CMAKE_THREAD_LIBS_INIT})
+endif(USE_THREAD AND Threads_FOUND AND CMAKE_USE_PTHREADS_INIT)