Align umath native bench with both Numba and Numpy versions (#14)

* Use high accuracy VML/SVML functions in native code

* Drop -fp-model precise flag

* Minor cleanup in umath_bench.c.src: remove outer layer of numpy templating and replace with preprocessor directives

* Fix python benchmark's logic to look up erf, invsqrt from numpy.core.umath.

* Use 1/sqrt for SVML "invsqrt" to match Numba

* Parse command-line options in umath native benchmarks using getopt_long - use ./<executable> -h for help.

Author: bibikar

numpy/linalg/Makefile

@@ -4,7 +4,7 @@
 
 CXX = icc
 CXXFLAGS = -O3 -g -xCORE-AVX2 -axCOMMON-AVX512 -qopenmp \
-		   -qopt-report=5 -qopt-report-phase=openmp,par,vec
+	   -qopt-report=5 -qopt-report-phase=openmp,par,vec
 LDFLAGS = -lmkl_rt -qopenmp
 
 TARGET = linalg

numpy/random/Makefile

@@ -7,8 +7,8 @@ SOURCES = $(addsuffix .c,$(BENCHMARKS))
 CC = icc
 CLANG_FORMAT = clang-format
 CFLAGS += -m64 -fPIC -fomit-frame-pointer -xSSE4.2 -axCORE-AVX2,CORE-AVX512 \
-		  -O3 -fp-model fast=2 -fimf-precision=high -prec-sqrt -prec-div \
-		  -fprotect-parens
+	  -O3 -fp-model fast=2 -fimf-precision=high -prec-sqrt -prec-div \
+	  -fprotect-parens
 LDFLAGS += -lmkl_rt
 
 run: $(BENCHMARKS)

numpy/umath/Makefile

@@ -4,14 +4,17 @@
 
 CC = icc
 CFLAGS = -qopenmp -xCORE-AVX2 -axCOMMON-AVX512 -O3 \
-		 -g -lmkl_rt
+	 -g -lmkl_rt -Wall -pedantic
+
+ifneq ($(CONDA_PREFIX),)
+		CFLAGS += -I$(CONDA_PREFIX)/include
+endif
 
 PYTHON ?= python
 
-ACC ?= la
+ACC ?= ha
 ifeq ($(ACC), ha)
 	CFLAGS += -fimf-precision=high -D_VML_ACCURACY_HA_
-	CFLAGS += -fp-model precise
 endif
 ifeq ($(ACC), la)
 	CFLAGS += -fimf-precision=medium -D_VML_ACCURACY_LA_