sysdeps/x86_64/fpu/Makefile

   1 ifeq ($(subdir),mathvec)
   2 libmvec-double-func-list = \
   3   2_core \
   4   4_core \
   5   4_core_avx \
   6   8_core
   7 libmvec-float-func-list = \
   8   f4_core \
   9   f8_core \
  10   f8_core_avx \
  11   f16_core
  12 libmvec-support += \
  13   svml_d_exp_data \
  14   svml_d_log_data \
  15   svml_d_pow_data \
  16   svml_d_trig_data \
  17   svml_s_expf_data \
  18   svml_s_logf_data \
  19   svml_s_powf_data \
  20   svml_s_trig_data \
  21   $(foreach l,$(libmvec-double-func-list), \
  22     $(addprefix svml_d_,$(addsuffix $(l),$(libmvec-funcs)))) \
  23   $(foreach l,$(libmvec-float-func-list), \
  24     $(addprefix svml_s_,$(addsuffix $(l),$(libmvec-funcs))))
  25 endif
  26
  27 # Do not run libmvec tests if multiarch not enabled.
  28 ifneq ($(multi-arch),no)
  29 # Variables for libmvec tests.
  30 ifeq ($(subdir)$(build-mathvec),mathyes)
  31 libmvec-tests += double-vlen2 double-vlen4 double-vlen4-avx2 \
  32                  float-vlen4 float-vlen8 float-vlen8-avx2 \
  33                  double-vlen8 float-vlen16
  34 tests += \
  35   $(libmvec-abi-func-tests) \
  36   $(libmvec-abi-func-avx-tests) \
  37   $(libmvec-abi-func-avx2-tests) \
  38   $(libmvec-abi-func-avx512f-tests)
  39
  40 double-vlen2-funcs = $(libmvec-funcs)
  41 double-vlen4-funcs = $(libmvec-funcs)
  42 double-vlen4-avx2-funcs = $(libmvec-funcs)
  43 double-vlen8-funcs = $(libmvec-funcs)
  44 float-vlen4-funcs = $(libmvec-funcs)
  45 float-vlen8-funcs = $(libmvec-funcs)
  46 float-vlen8-avx2-funcs = $(libmvec-funcs)
  47 float-vlen16-funcs = $(libmvec-funcs)
  48
  49 double-vlen4-arch-ext-cflags = -mavx
  50 double-vlen4-arch-ext2-cflags = -mavx2
  51 double-vlen8-arch-ext-cflags = -mavx512f
  52
  53 float-vlen8-arch-ext-cflags = -mavx
  54 float-vlen8-arch-ext2-cflags = -mavx2
  55 float-vlen16-arch-ext-cflags = -mavx512f
  56
  57 libmvec-abi-test-cflags = \
  58   $(libm-test-fast-math-cflags) \
  59   -fno-inline -fopenmp -Wno-unknown-pragmas
  60
  61 CFLAGS-test-double-vlen4-avx2-wrappers.c = $(double-vlen4-arch-ext2-cflags)
  62
  63 CFLAGS-test-float-vlen8-avx2-wrappers.c = $(float-vlen8-arch-ext2-cflags)
  64 endif
  65
  66 ifeq ($(subdir)$(config-cflags-mprefer-vector-width),mathyes)
  67 # When compiled with -O3 -march=skylake, GCC 8 and 9 optimize some loops
  68 # in branred.c with 256-bit vector instructions, which leads to store
  69 # forward stall:
  70 #
  71 # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90579
  72 #
  73 # Limit vector width to 128 bits to work around this issue.  It improves
  74 # performance of sin and cos by more than 40% on Skylake.
  75 CFLAGS-branred.c = -mprefer-vector-width=128
  76 endif
  77
  78 ifeq ($(subdir)$(build-mathvec),benchtestsyes)
  79 double-vlen4-arch-ext-cflags = -mavx
  80 double-vlen4-arch-ext2-cflags = -mavx2
  81 double-vlen8-arch-ext-cflags = -mavx512f
  82
  83 float-vlen8-arch-ext-cflags = -mavx
  84 float-vlen8-arch-ext2-cflags = -mavx2
  85 float-vlen16-arch-ext-cflags = -mavx512f
  86
  87 bench-libmvec := $(bench-libmvec-double) $(bench-libmvec-float)
  88
  89 ifeq (${STATIC-BENCHTESTS},yes)
  90 libmvec-benchtests = $(common-objpfx)mathvec/libmvec.a $(common-objpfx)math/libm.a
  91 else
  92 libmvec-benchtests = $(libmvec) $(libm)
  93 endif
  94
  95 $(addprefix $(objpfx)bench-,$(bench-libmvec-double)): $(libmvec-benchtests)
  96 $(addprefix $(objpfx)bench-,$(bench-libmvec-float)): $(libmvec-benchtests)
  97 bench-libmvec-deps = $(..)sysdeps/x86_64/fpu/bench-libmvec-skeleton.c bench-timing.h Makefile
  98
  99 $(objpfx)bench-float-%.c: $(bench-libmvec-deps)
 100         { if [ -n "$($*-INCLUDE)" ]; then \
 101           cat $($*-INCLUDE); \
 102         fi; \
 103         $(PYTHON) $(..)sysdeps/x86_64/fpu/scripts/bench_libmvec.py $(basename $(@F)); } > $@-tmp
 104         mv -f $@-tmp $@
 105
 106 $(objpfx)bench-double-%.c: $(bench-libmvec-deps)
 107         { if [ -n "$($*-INCLUDE)" ]; then \
 108           cat $($*-INCLUDE); \
 109         fi; \
 110         $(PYTHON) $(..)sysdeps/x86_64/fpu/scripts/bench_libmvec.py $(basename $(@F)); } > $@-tmp
 111         mv -f $@-tmp $@
 112 endif
 113 endif