Merge pull request #2 from sirzooro/optimizations

Optimizations

boinc_app/Makefile

@@ -1,12 +1,11 @@
 CAMBALA = ../cambala
-ALGLIB = ../../alglib
+ALGLIB = ../alglib-3.14.0/src
 BOINC_DIR = ../../boinc
 BOINC_API_DIR = $(BOINC_DIR)/api
 BOINC_LIB_DIR = $(BOINC_DIR)/lib
 BOINC_ZIP_DIR = $(BOINC_DIR)/zip
 FREETYPE_DIR = /usr/include/freetype2
 
-CPP = g++
 CPPFLAGS = -O3 \
     -std=c++0x -static -Wall -W -Wshadow -Wpointer-arith -Wcast-qual -Wcast-align -Wwrite-strings -fno-common \
     -DAPP_GRAPHICS -D __STDC_LIMIT_MACROS -D __STDC_FORMAT_MACROS -D NDEBUG \
@@ -20,42 +19,65 @@ CPPFLAGS = -O3 \
     -L /usr/X11R6/lib \
     -L.
 
+ifeq ($(MinGW32),1)
+$(info ===== Compiling MinGW 32-bit app version =====)
+CPP = i686-w64-mingw32-g++
+else ifeq ($(MinGW64),1)
+$(info ===== Compiling MinGW 64-bit app version =====)
+CPP = x86_64-w64-mingw32-g++
+else ifeq ($(M32),1)
+$(info ===== Compiling 32-bit app version =====)
+CPP = g++ -m32
+else
+CPP = g++
+endif
+
+ifeq ($(SSE2),1)
+$(info ===== Compiling SSE2 app version =====)
+CPPFLAGS += -msse2
+else ifeq ($(SSE41),1)
+$(info ===== Compiling SSE4.1 app version =====)
+CPPFLAGS += -msse4.1
+else ifeq ($(AVX),1)
+$(info ===== Compiling AVX app version =====)
+CPPFLAGS += -mavx -mtune=sandybridge
+else ifeq ($(AVX2),1)
+$(info ===== Compiling AVX2 app version =====)
+CPPFLAGS += -mavx2 -mfma -mtune=haswell
+else ifeq ($(AVX512),1)
+$(info ===== Compiling AVX512 app version =====)
+CPPFLAGS += -march=skylake-avx512
+ifeq ($(MinGW64),1)
+# MinGW needs workaround for "invalid register for .seh_savexmm" bug
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65782
+CPPFLAGS += \
+	-ffixed-xmm16 -ffixed-xmm17 -ffixed-xmm18 -ffixed-xmm19 -ffixed-xmm20 -ffixed-xmm21 -ffixed-xmm22 -ffixed-xmm23 \
+	-ffixed-xmm24 -ffixed-xmm25 -ffixed-xmm26 -ffixed-xmm27 -ffixed-xmm28 -ffixed-xmm29 -ffixed-xmm30 -ffixed-xmm31
+endif
+endif
+
+ifeq ($(NOWARN),1)
+CPPFLAGS += -w
+endif
+
 release: cambala_boinc_app
 
 libstdc++.a:
-	ln -s `g++ -print-file-name=libstdc++.a`
-
-ttfont.cpp:
-	ln -s ../../api/ttfont.cpp .
+	ln -s `${CPP} -print-file-name=libstdc++.a`
 
 clean: distclean
 
 distclean:
-	/bin/rm -f $(PROGS) *.o libstdc++.a cambala_boinc_app
+	/bin/rm -f $(PROGS) *.o libstdc++.a cambala_boinc_app cambala_boinc_app.exe
 
-cambala_boinc_app: cambala_boinc_app.o alglibinternal.o alglibmisc.o ap.o linalg.o specialfunctions.o sequential.o \
+cambala_boinc_app: cambala_boinc_app.o all.o \
 	libstdc++.a $(BOINC_API_DIR)/libboinc_api.a $(BOINC_LIB_DIR)/libboinc.a
-	$(CPP) $(CPPFLAGS) cambala_boinc_app.o alglibinternal.o alglibmisc.o ap.o linalg.o specialfunctions.o sequential.o -o cambala_boinc_app \
+	$(CPP) $(CPPFLAGS) cambala_boinc_app.o all.o -o cambala_boinc_app \
 	libstdc++.a -pthread $(BOINC_API_DIR)/libboinc_api.a \
 	$(BOINC_LIB_DIR)/libboinc.a
 
-specialfunctions.o: $(ALGLIB)/specialfunctions.cpp 
-	${CPP} ${CPPFLAGS} $(ALGLIB)/specialfunctions.cpp -c
-
-linalg.o: $(ALGLIB)/linalg.cpp 
-	${CPP} ${CPPFLAGS} $(ALGLIB)/linalg.cpp -c
-
-ap.o: $(ALGLIB)/ap.cpp 
-	${CPP} ${CPPFLAGS} $(ALGLIB)/ap.cpp -c
-
-alglibmisc.o: $(ALGLIB)/alglibmisc.cpp 
-	${CPP} ${CPPFLAGS} $(ALGLIB)/alglibmisc.cpp -c
-
-alglibinternal.o: $(ALGLIB)/alglibinternal.cpp 
-	${CPP} ${CPPFLAGS} $(ALGLIB)/alglibinternal.cpp -c
-
-sequential.o: $(CAMBALA)/sequential.cpp
-	${CPP} ${CPPFLAGS} $(CAMBALA)/sequential.cpp -c
+all.o: all.cpp $(ALGLIB)/specialfunctions.cpp $(ALGLIB)/linalg.cpp $(ALGLIB)/ap.cpp $(ALGLIB)/alglibmisc.cpp $(ALGLIB)/alglibinternal.cpp $(CAMBALA)/sequential.cpp
+	${CPP} ${CPPFLAGS} all.cpp -c
 
 cambala_boinc_app.o: cambala_boinc_app.cpp
 	${CPP} ${CPPFLAGS} cambala_boinc_app.cpp -c

boinc_app/all.cpp

@@ -0,0 +1,13 @@
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wunused-parameter"
+#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
+
+#include <ap.cpp>
+#include <alglibinternal.cpp>
+#include <specialfunctions.cpp>
+#include <alglibmisc.cpp>
+#include <linalg.cpp>
+
+#pragma GCC diagnostic pop
+
+#include <sequential.cpp>

boinc_app/cambala_boinc_app.cpp

@@ -32,10 +32,125 @@
 #include "utils.h"
 #include "point.h"
 
+#if defined(__i386__) || defined (__x86_64__)
+#include <cpuid.h>
+#endif
+
 #define CHECKPOINT_FILE "chpt"
 #define INPUT_FILENAME "in"
 #define OUTPUT_FILENAME "out"
 
+__attribute__((noreturn))
+void PrintFatalError(const char* str)
+{
+	// print error to unredirected stderr first
+	fprintf(stderr, "Error: %s", str);
+
+	// now try to send it back to server
+	int retval = boinc_init();
+    if (0 == retval)
+	{
+		fprintf(stderr, "Error: %s", str);
+		boinc_finish(1);
+    }
+
+	exit(1);
+}
+
+void VerifyCpu()
+{
+#if (defined(__i386__) || defined (__x86_64__)) && defined(__SSE2__)
+	unsigned int a, b, c, d;
+
+	if (!__get_cpuid(1, &a, &b, &c, &d))
+	{
+		PrintFatalError("CPUID instruction is not supported by your CPU!\n");
+	}
+
+	if (0 == (d & bit_SSE2))
+	{
+		PrintFatalError("SSE2 instructions are not supported by your CPU!\n");
+	}
+
+#ifdef __SSE4_1__
+	if (0 == (c & bit_SSE4_1))
+	{
+		PrintFatalError("SSE4.1 instructions are not supported by your CPU!\n");
+	}
+#endif
+
+#ifdef __AVX__
+	if (0 == (c & bit_AVX))
+	{
+		PrintFatalError("AVX instructions are not supported by your CPU!\n");
+	}
+
+	// AVX also needs OS support, check for it
+	if (0 == (c & bit_OSXSAVE))
+	{
+		PrintFatalError("OSXSAVE instructions are not supported by your CPU!\n");
+	}
+
+	unsigned int eax, edx;
+	unsigned int ecx = 0; // _XCR_XFEATURE_ENABLED_MASK
+	__asm__ ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (ecx));
+	if (0x6 != (eax & 0x6)) // XSTATE_SSE | XSTATE_YMM
+	{
+		PrintFatalError("AVX instructions are not supported by your OS!\n");
+	}
+#endif
+
+#ifdef __AVX2__
+	if (__get_cpuid_max(0, 0) < 7)
+	{
+		PrintFatalError("CPUID level 7 is not supported by your CPU!\n");
+	}
+
+	unsigned int a2, b2, c2, d2;
+	__cpuid_count(7, 0, a2, b2, c2, d2);
+
+	if (0 == (b2 & bit_AVX2))
+	{
+		PrintFatalError("AVX2 instructions are not supported by your CPU!\n");
+	}
+#endif
+
+#ifdef __FMA__
+	// Some AMD CPUs(s) support FMA but no AVX2. FMA does not provide significat 
+	// boost for this app, so it should be enabled on AVX2 CPUs only.
+#ifndef __AVX2__
+#error AVX2 is not enabled!
+#endif
+	if (0 == (c & bit_FMA))
+	{
+		PrintFatalError("FMA instructions are not supported by your CPU!\n");
+	}
+#endif
+
+#ifdef __AVX512F__
+	// AVX512 consists of few subsets. Skylake-AVX512 target supports F, BW, DQ, VL and CD.
+	// We need data loaded during AVX and AVX2 checks, make sure we have it
+#if !defined(__AVX__) || !defined(__AVX2__)
+#error AVX or AVX2 is not enabled!
+#endif
+
+	const unsigned int avx512bits = bit_AVX512F | bit_AVX512DQ | bit_AVX512CD | bit_AVX512BW | bit_AVX512VL;
+
+	if (avx512bits != (b2 & avx512bits))
+	{
+		PrintFatalError("AVX512 instructions are not supported by your CPU!\n");
+	}
+
+	// AVX512 also needs OS support, check for it
+	if (0xe6 != (eax & 0xe6)) // XSTATE_SSE | XSTATE_YMM | XSTATE_OPMASK | XSTATE_ZMM | XSTATE_HI_ZMM
+	{
+		PrintFatalError("AVX512 instructions are not supported by your OS!\n");
+	}
+#endif
+
+#endif
+}
+
 using namespace std;
 char buf[256];
 
@@ -46,14 +161,16 @@ int do_checkpoint( const long long &total_points,
 
 int main(int argc, char **argv)
 {
+	VerifyCpu();
+
 	int retval = boinc_init();
     if ( retval ) {
         fprintf(stderr, "%s APP: boinc_init() returned %d\n",
             boinc_msg_prefix(buf, sizeof(buf)), retval
         );
         exit( retval );
     }
-	
+
 	search_space_point cur_record_point;
 	cur_record_point.residual = START_HUGE_VALUE;
 

boinc_app/test/cambala_depths_out.ref

@@ -0,0 +1 @@
+23 33 50 300 

boinc_app/test/cambala_out.ref

@@ -0,0 +1,26 @@
+Input parameters :
+launch_type bruteforce
+object_function_type uniform
+ppm 2
+init_iterated_local_search_runs 10
+cw1_init_arr :
+1506 1490 1460 
+cw2_init_arr :
+1506 1490 1480 
+ncpl_init_arr :
+1 1 11 
+nR 1
+R1 7000
+R2 7000
+ntau 1
+tau1 0
+tau2 0
+nrhob 1
+rhob1 1.7
+rhob2 1.7
+ncb 1
+cb1 1700
+cb2 1700
+dtimes_file ac_modes_R7km_dtimes1.txt
+spmag_file no
+launch_type bruteforce

boinc_app/test/out.ref

@@ -0,0 +1 @@
+0.003945 1700 1.7 7000 0 1506 1490 1462 23 33 50 300 

boinc_app/test/test.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/bash
+
+rm -f boinc_finish_called cambala_depths_out cambala_out chpt out stderr.txt
+
+time ../cambala_boinc_app
+
+echo
+
+#diff -qs cambala_depths_out cambala_depths_out.ref
+#diff -qs cambala_out cambala_out.ref
+diff -qs out out.ref