Release of version 2.100

shibatch · Dec 3, 2016 · f9d377c · f9d377c
1 parent 20b4818
commit f9d377c
Show file tree

Hide file tree

Showing 22 changed files with 2,567 additions and 1,562 deletions.
diff --git a/README b/README
@@ -29,6 +29,8 @@ Single precision trigonometric functions : |arg| <= 10000
 
 History
 
+2.100 Added support for AVX-512F and Clang Extended Vectors.
+
 2.90 Added ilogbf. All the reported bugs(listed below) are fixed.
 * Log function returned incorrect values when the argument is very small.
 * Signs of returned values were incorrect when the argument is signed zero.

diff --git a/purec/sleefdp.c b/purec/sleefdp.c
@@ -838,23 +838,23 @@ double xlog(double d) {
   double x, x2, t, m;
   int e;
 
-  e = ilogbk(d * 1.4142);
+  e = ilogbk(d * (1.0/0.75));
   m = ldexpk(d, -e);
 
   x = (m-1) / (m+1);
   x2 = x * x;
 
-  t = 0.148197055177935105296783;
-  t = mla(t, x2, 0.153108178020442575739679);
-  t = mla(t, x2, 0.181837339521549679055568);
-  t = mla(t, x2, 0.22222194152736701733275);
-  t = mla(t, x2, 0.285714288030134544449368);
-  t = mla(t, x2, 0.399999999989941956712869);
-  t = mla(t, x2, 0.666666666666685503450651);
+  t = 0.153487338491425068243146;
+  t = mla(t, x2, 0.152519917006351951593857);
+  t = mla(t, x2, 0.181863266251982985677316);
+  t = mla(t, x2, 0.222221366518767365905163);
+  t = mla(t, x2, 0.285714294746548025383248);
+  t = mla(t, x2, 0.399999999950799600689777);
+  t = mla(t, x2, 0.6666666666667778740063);
   t = mla(t, x2, 2);
 
   x = x * t + 0.693147180559945286226764 * e;
-
+  
   if (xisinf(d)) x = INFINITY;
   if (d < 0) x = NAN;
   if (d == 0) x = -INFINITY;
@@ -894,20 +894,20 @@ static inline double2 logk(double d) {
   double m, t;
   int e;
 
-  e = ilogbk(d * 1.4142);
+  e = ilogbk(d * (1.0/0.75));
   m = ldexpk(d, -e);
 
   x = dddiv_d2_d2_d2(ddadd2_d2_d_d(-1, m), ddadd2_d2_d_d(1, m));
   x2 = ddsqu_d2_d2(x);
 
-  t = 0.134601987501262130076155;
-  t = mla(t, x2.x, 0.132248509032032670243288);
-  t = mla(t, x2.x, 0.153883458318096079652524);
-  t = mla(t, x2.x, 0.181817427573705403298686);
-  t = mla(t, x2.x, 0.222222231326187414840781);
-  t = mla(t, x2.x, 0.285714285651261412873718);
-  t = mla(t, x2.x, 0.400000000000222439910458);
-  t = mla(t, x2.x, 0.666666666666666371239645);
+  t = 0.13860436390467167910856;
+  t = mla(t, x2.x, 0.131699838841615374240845);
+  t = mla(t, x2.x, 0.153914168346271945653214);
+  t = mla(t, x2.x, 0.181816523941564611721589);
+  t = mla(t, x2.x, 0.22222224632662035403996);
+  t = mla(t, x2.x, 0.285714285511134091777308);
+  t = mla(t, x2.x, 0.400000000000914013309483);
+  t = mla(t, x2.x, 0.666666666666664853302393);
 
   return ddadd2_d2_d2_d2(ddmul_d2_d2_d(dd(0.693147180559945286226764, 2.319046813846299558417771e-17), e),
 			 ddadd2_d2_d2_d2(ddscale_d2_d2_d(x, 2), ddmul_d2_d2_d(ddmul_d2_d2_d2(x2, x), t)));
@@ -1041,20 +1041,20 @@ static inline double2 logk2(double2 d) {
   double t;
   int e;
 
-  e = ilogbk(d.x * 1.4142);
+  e = ilogbk(d.x * (1.0/0.75));
   m = ddscale_d2_d2_d(d, pow2i(-e));
 
   x = dddiv_d2_d2_d2(ddadd2_d2_d2_d(m, -1), ddadd2_d2_d2_d(m, 1));
   x2 = ddsqu_d2_d2(x);
 
-  t = 0.134601987501262130076155;
-  t = mla(t, x2.x, 0.132248509032032670243288);
-  t = mla(t, x2.x, 0.153883458318096079652524);
-  t = mla(t, x2.x, 0.181817427573705403298686);
-  t = mla(t, x2.x, 0.222222231326187414840781);
-  t = mla(t, x2.x, 0.285714285651261412873718);
-  t = mla(t, x2.x, 0.400000000000222439910458);
-  t = mla(t, x2.x, 0.666666666666666371239645);
+  t = 0.13860436390467167910856;
+  t = mla(t, x2.x, 0.131699838841615374240845);
+  t = mla(t, x2.x, 0.153914168346271945653214);
+  t = mla(t, x2.x, 0.181816523941564611721589);
+  t = mla(t, x2.x, 0.22222224632662035403996);
+  t = mla(t, x2.x, 0.285714285511134091777308);
+  t = mla(t, x2.x, 0.400000000000914013309483);
+  t = mla(t, x2.x, 0.666666666666664853302393);
 
   return ddadd2_d2_d2_d2(ddmul_d2_d2_d(dd(0.693147180559945286226764, 2.319046813846299558417771e-17), e),
 			 ddadd2_d2_d2_d2(ddscale_d2_d2_d(x, 2), ddmul_d2_d2_d(ddmul_d2_d2_d2(x2, x), t)));

diff --git a/purec/sleefsp.c b/purec/sleefsp.c
@@ -757,20 +757,20 @@ float xlogf(float d) {
   float x, x2, t, m;
   int e;
 
-  e = ilogbkf(d * 1.4142f);
+  e = ilogbkf(d * (1.0f/0.75f));
   m = ldexpkf(d, -e);
-
+  
   x = (m-1.0f) / (m+1.0f);
   x2 = x * x;
 
-  t = 0.2371599674224853515625f;
-  t = mlaf(t, x2, 0.285279005765914916992188f);
-  t = mlaf(t, x2, 0.400005519390106201171875f);
-  t = mlaf(t, x2, 0.666666567325592041015625f);
+  t = 0.2392828464508056640625f;
+  t = mlaf(t, x2, 0.28518211841583251953125f);
+  t = mlaf(t, x2, 0.400005877017974853515625f);
+  t = mlaf(t, x2, 0.666666686534881591796875f);
   t = mlaf(t, x2, 2.0f);
 
   x = x * t + 0.693147180559945286226764f * e;
-
+  
   if (xisinff(d)) x = INFINITYf;
   if (d < 0) x = NANf;
   if (d == 0) x = -INFINITYf;
@@ -834,16 +834,16 @@ static inline float2 logkf(float d) {
   float m, t;
   int e;
 
-  e = ilogbkf(d * 1.4142f);
+  e = ilogbkf(d * (1.0f/0.75f));
   m = ldexpkf(d, -e);
 
   x = dfdiv_f2_f2_f2(dfadd2_f2_f_f(-1, m), dfadd2_f2_f_f(1, m));
   x2 = dfsqu_f2_f2(x);
-
-  t = 0.2371599674224853515625f;
-  t = mlaf(t, x2.x, 0.285279005765914916992188f);
-  t = mlaf(t, x2.x, 0.400005519390106201171875f);
-  t = mlaf(t, x2.x, 0.666666567325592041015625f);
+  
+  t = 0.2392828464508056640625f;
+  t = mlaf(t, x2.x, 0.28518211841583251953125f);
+  t = mlaf(t, x2.x, 0.400005877017974853515625f);
+  t = mlaf(t, x2.x, 0.666666686534881591796875f);
 
   return dfadd2_f2_f2_f2(dfmul_f2_f2_f(df(0.69314718246459960938f, -1.904654323148236017e-09f), e),
 			 dfadd2_f2_f2_f2(dfscale_f2_f2_f(x, 2), dfmul_f2_f2_f(dfmul_f2_f2_f2(x2, x), t)));
@@ -945,16 +945,16 @@ static inline float2 logk2f(float2 d) {
   float t;
   int e;
 
-  e = ilogbkf(d.x * 1.4142f);
+  e = ilogbkf(d.x * (1.0f/0.75f));
   m = dfscale_f2_f2_f(d, pow2if(-e));
 
   x = dfdiv_f2_f2_f2(dfadd2_f2_f2_f(m, -1), dfadd2_f2_f2_f(m, 1));
   x2 = dfsqu_f2_f2(x);
 
-  t = 0.2371599674224853515625f;
-  t = mlaf(t, x2.x, 0.285279005765914916992188f);
-  t = mlaf(t, x2.x, 0.400005519390106201171875f);
-  t = mlaf(t, x2.x, 0.666666567325592041015625f);
+  t = 0.2392828464508056640625f;
+  t = mlaf(t, x2.x, 0.28518211841583251953125f);
+  t = mlaf(t, x2.x, 0.400005877017974853515625f);
+  t = mlaf(t, x2.x, 0.666666686534881591796875f);
 
   return dfadd2_f2_f2_f2(dfmul_f2_f2_f(df(0.69314718246459960938f, -1.904654323148236017e-09f), e),
 			 dfadd2_f2_f2_f2(dfscale_f2_f2_f(x, 2), dfmul_f2_f2_f(dfmul_f2_f2_f2(x2, x), t)));

diff --git a/simd/Makefile b/simd/Makefile
@@ -1,6 +1,6 @@
-CC=gcc
-OPT=-O -Wall -Wno-unused -Wno-attributes -ffp-contract=off
-SDE=/opt/sde-bdw-external-5.38.0-2013-01-03-lin/sde
+CC=gcc-5
+OPT=-O -Wall -Wno-unused -Wno-attributes -ffp-contract=off -fmax-errors=3
+SDE=sde
 
 all : testsse2 testavx testavx2
 
@@ -13,6 +13,9 @@ iutavx : sleefsimddp.c sleefsimdsp.c helperavx.h iut.c
 iutavx2 : sleefsimddp.c sleefsimdsp.c helperavx2.h iut.c
 	$(CC) $(OPT) -DENABLE_AVX2 -mavx2 -mfma iut.c sleefsimddp.c sleefsimdsp.c -o iutavx2 -lm
 
+iutavx512f : sleefsimddp.c sleefsimdsp.c helperavx512f.h iut.c
+	$(CC) $(OPT) -DENABLE_AVX512F -mavx512f iut.c sleefsimddp.c sleefsimdsp.c -o iutavx512f -lm
+
 iutfma4 : sleefsimddp.c sleefsimdsp.c helperfma4.h iut.c
 	$(CC) $(OPT) -DENABLE_FMA4 -mavx -mfma4 iut.c sleefsimddp.c sleefsimdsp.c -o iutfma4 -lm
 
@@ -30,32 +33,33 @@ iutfma4 : sleefsimddp.c sleefsimdsp.c helperfma4.h iut.c
 
 testsse2 : iutsse2 ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
 	../tester/tester ./iutsse2
-	../tester/testeru1 ./iutsse2
 	../tester/testersp ./iutsse2
+	../tester/testeru1 ./iutsse2
 	../tester/testerspu1 ./iutsse2
 
 testavx : iutavx ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
 	../tester/tester ./iutavx
-	../tester/testeru1 ./iutavx
 	../tester/testersp ./iutavx
+	../tester/testeru1 ./iutavx
 	../tester/testerspu1 ./iutavx
 
 testavx2 : iutavx2 ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
 	../tester/tester ./iutavx2
-	../tester/testeru1 ./iutavx2
 	../tester/testersp ./iutavx2
+	../tester/testeru1 ./iutavx2
 	../tester/testerspu1 ./iutavx2
 
+testavx512f : iutavx512f ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
+	../tester/tester $(SDE) -- ./iutavx512f
+	../tester/testersp $(SDE) -- ./iutavx512f
+	../tester/testeru1 $(SDE) -- ./iutavx512f
+	../tester/testerspu1 $(SDE) -- ./iutavx512f
+
 testfma4 : iutfma4 ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
 	../tester/tester ./iutfma4
-	../tester/testeru1 ./iutfma4
 	../tester/testersp ./iutfma4
+	../tester/testeru1 ./iutfma4
 	../tester/testerspu1 ./iutfma4
 
 clean :
-	rm -f *~ *.o *.s iutsse2 iutavx iutavx2 iutfma4 iutneon
-
-# ../tester/tester $(SDE) -- ./iutavx2
-# ../tester/testeru1 $(SDE) -- ./iutavx2
-# ../tester/testersp $(SDE) -- ./iutavx2
-# ../tester/testerspu1 $(SDE) -- ./iutavx2
+	rm -f *~ *.o *.s iutsse2 iutavx iutavx2 iutavx512f iutfma4 iutneon iutclangvec
diff --git a/simd/Makefile.clang b/simd/Makefile.clang
@@ -0,0 +1,74 @@
+CC=clang
+OPT=-O -Wall -Wno-unused -Wno-attributes -ffp-contract=off -ferror-limit=3
+SDE=sde
+
+all : testclangvec
+
+iutsse2 : sleefsimddp.c sleefsimdsp.c helpersse2.h iut.c
+	$(CC) $(OPT) -DENABLE_SSE2 -msse2 iut.c sleefsimddp.c sleefsimdsp.c -o iutsse2 -lm
+
+iutavx : sleefsimddp.c sleefsimdsp.c helperavx.h iut.c
+	$(CC) $(OPT) -DENABLE_AVX -mavx iut.c sleefsimddp.c sleefsimdsp.c -o iutavx -lm
+
+iutavx2 : sleefsimddp.c sleefsimdsp.c helperavx2.h iut.c
+	$(CC) $(OPT) -DENABLE_AVX2 -mavx2 -mfma iut.c sleefsimddp.c sleefsimdsp.c -o iutavx2 -lm
+
+iutavx512f : sleefsimddp.c sleefsimdsp.c helperavx512f.h iut.c
+	$(CC) $(OPT) -DENABLE_AVX512F -mavx512f iut.c sleefsimddp.c sleefsimdsp.c -o iutavx512f -lm
+
+iutfma4 : sleefsimddp.c sleefsimdsp.c helperfma4.h iut.c
+	$(CC) $(OPT) -DENABLE_FMA4 -mavx -mfma4 iut.c sleefsimddp.c sleefsimdsp.c -o iutfma4 -lm
+
+iutclangvec : sleefsimddp.c sleefsimdsp.c helperclangvec.h iut.c
+	$(CC) $(OPT) -DENABLE_CLANGVEC iut.c sleefsimddp.c sleefsimdsp.c -o iutclangvec -lm
+
+../tester/tester :
+	cd ../tester; make tester
+
+../tester/testeru1 :
+	cd ../tester; make testeru1
+
+../tester/testersp :
+	cd ../tester; make testersp
+
+../tester/testerspu1 :
+	cd ../tester; make testerspu1
+
+testsse2 : iutsse2 ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
+	../tester/tester ./iutsse2
+	../tester/testersp ./iutsse2
+	../tester/testeru1 ./iutsse2
+	../tester/testerspu1 ./iutsse2
+
+testavx : iutavx ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
+	../tester/tester ./iutavx
+	../tester/testersp ./iutavx
+	../tester/testeru1 ./iutavx
+	../tester/testerspu1 ./iutavx
+
+testavx2 : iutavx2 ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
+	../tester/tester ./iutavx2
+	../tester/testersp ./iutavx2
+	../tester/testeru1 ./iutavx2
+	../tester/testerspu1 ./iutavx2
+
+testavx512f : iutavx512f ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
+	../tester/tester $(SDE) -- ./iutavx512f
+	../tester/testersp $(SDE) -- ./iutavx512f
+	../tester/testeru1 $(SDE) -- ./iutavx512f
+	../tester/testerspu1 $(SDE) -- ./iutavx512f
+
+testfma4 : iutfma4 ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
+	../tester/tester ./iutfma4
+	../tester/testersp ./iutfma4
+	../tester/testeru1 ./iutfma4
+	../tester/testerspu1 ./iutfma4
+
+testclangvec : iutclangvec ../tester/tester ../tester/testeru1 ../tester/testersp ../tester/testerspu1
+	../tester/tester ./iutclangvec
+	../tester/testersp ./iutclangvec
+	../tester/testeru1 ./iutclangvec
+	../tester/testerspu1 ./iutclangvec
+
+clean :
+	rm -f *~ *.o *.s iutsse2 iutavx iutavx2 iutavx512f iutfma4 iutneon iutclangvec
diff --git a/simd/dd.h b/simd/dd.h
@@ -2,6 +2,10 @@ typedef struct {
   vdouble x, y;
 } vdouble2;
 
+static INLINE vdouble vupper_vd_vd(vdouble d) {
+  return (vdouble)vand_vm_vm_vm((vmask)d, vcast_vm_i_i(0xffffffff, 0xf8000000));
+}
+
 static INLINE vdouble2 vcast_vd2_vd_vd(vdouble h, vdouble l) {
   vdouble2 ret = {h, l};
   return ret;
@@ -12,10 +16,10 @@ static INLINE vdouble2 vcast_vd2_d_d(double h, double l) {
   return ret;
 }
 
-static INLINE vdouble2 vsel_vd2_vm_vd2_vd2(vmask m, vdouble2 x, vdouble2 y) {
+static INLINE vdouble2 vsel_vd2_vo_vd2_vd2(vopmask m, vdouble2 x, vdouble2 y) {
   vdouble2 r;
-  r.x = vsel_vd_vm_vd_vd(m, x.x, y.x);
-  r.y = vsel_vd_vm_vd_vd(m, x.y, y.y);
+  r.x = vsel_vd_vo_vd_vd(m, x.x, y.x);
+  r.y = vsel_vd_vo_vd_vd(m, x.y, y.y);
   return r;
 }
 
@@ -145,7 +149,7 @@ static INLINE vdouble2 ddadd2_vd2_vd2_vd2(vdouble2 x, vdouble2 y) {
   return r;
 }
 
-static inline vdouble2 ddsub_vd2_vd_vd(vdouble x, vdouble y) {
+static INLINE vdouble2 ddsub_vd2_vd_vd(vdouble x, vdouble y) {
   // |x| >= |y|
 
   vdouble2 r;
@@ -170,19 +174,6 @@ static INLINE vdouble2 ddsub_vd2_vd2_vd2(vdouble2 x, vdouble2 y) {
   return r;
 }
 
-#if 0
-static inline vdouble2 ddsub_vd2_vd2_vd2(vdouble2 x, vdouble2 y) {
-  // |x| >= |y|
-
-  vdouble2 r;
-
-  r.x = vsub_vd_vd_vd(x.x, y.x);
-  r.y = vsub_vd_vd_vd(vadd_vd_vd_vd(vsub_vd_vd_vd(vsub_vd_vd_vd(x.x, r.x), y.x), x.y), y.y);
-
-  return r;
-}
-#endif
-
 #ifdef ENABLE_FMA_DP
 static INLINE vdouble2 dddiv_vd2_vd2_vd2(vdouble2 n, vdouble2 d) {
   vdouble2 q;
@@ -232,7 +223,7 @@ static INLINE vdouble2 ddmul_vd2_vd2_vd(vdouble2 x, vdouble y) {
   return r;
 }
 
-static inline vdouble2 ddrec_vd2_vd(vdouble d) {
+static INLINE vdouble2 ddrec_vd2_vd(vdouble d) {
   vdouble2 q;
 
   q.x = vrec_vd_vd(d);