diff options
Diffstat (limited to '0006-simdmath-Enable-simdmath-on-kunpeng.patch')
| -rw-r--r-- | 0006-simdmath-Enable-simdmath-on-kunpeng.patch | 289 |
1 files changed, 289 insertions, 0 deletions
diff --git a/0006-simdmath-Enable-simdmath-on-kunpeng.patch b/0006-simdmath-Enable-simdmath-on-kunpeng.patch new file mode 100644 index 0000000..9d7bb57 --- /dev/null +++ b/0006-simdmath-Enable-simdmath-on-kunpeng.patch @@ -0,0 +1,289 @@ +From cfd6920125f7968f0c1f5cb225f9fbd5bc8988b9 Mon Sep 17 00:00:00 2001 +From: bule <bule1@huawei.com> +Date: Tue, 13 Jul 2021 15:26:54 +0800 +Subject: [PATCH 06/13] [simdmath] Enable simdmath on kunpeng + +This enable simd math function supported by libmathlib on fortran/c/c++. +Use -fsimdmath to turn on the generation of simdmath function. The +supported functions can be found in simdmath.h. Add more simd declaration +if you need more kinds of math functions. -msimdmath-64 is used to turn +on 64-bit simd math functions which is not supported by libmathlib. +Therefore, this option is default to off. + +diff --git a/gcc/c-family/c-opts.c b/gcc/c-family/c-opts.c +index c51d6d34726..dc1a8984871 100644 +--- a/gcc/c-family/c-opts.c ++++ b/gcc/c-family/c-opts.c +@@ -780,6 +780,10 @@ c_common_post_options (const char **pfilename) + if (cpp_opts->deps.style == DEPS_NONE) + check_deps_environment_vars (); + ++ if (flag_simdmath) ++ { ++ defer_opt (OPT_include, "simdmath.h"); ++ } + handle_deferred_opts (); + + sanitize_cpp_opts (); +diff --git a/gcc/common.opt b/gcc/common.opt +index ec5235c3a41..8eb05570418 100644 +--- a/gcc/common.opt ++++ b/gcc/common.opt +@@ -1977,6 +1977,10 @@ fmath-errno + Common Report Var(flag_errno_math) Init(1) Optimization SetByCombined + Set errno after built-in math functions. + ++fsimdmath ++Common Report Var(flag_simdmath) Init(0) Optimization ++Enable auto-vectorize math functions for mathlib. This option will turn on -fno-math-errno and -fopenmp-simd. ++ + fmax-errors= + Common Joined RejectNegative UInteger Var(flag_max_errors) + -fmax-errors=<number> Maximum number of errors to report. +diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c +index 9b400c49ac6..79dc8f186f4 100644 +--- a/gcc/config/aarch64/aarch64.c ++++ b/gcc/config/aarch64/aarch64.c +@@ -23077,8 +23077,12 @@ aarch64_simd_clone_compute_vecsize_and_simdlen (struct cgraph_node *node, + elt_bits = GET_MODE_BITSIZE (SCALAR_TYPE_MODE (base_type)); + if (clonei->simdlen == 0) + { +- count = 2; +- vec_bits = (num == 0 ? 64 : 128); ++ /* Currently mathlib or sleef hasn't provide function for V2SF mode ++ simdclone of single precision functions. (e.g._ZCVnN2v_expf) ++ Therefore this mode is disabled by default to avoid link error. ++ Use -msimdmath-64 option to enable this mode. */ ++ count = flag_simdmath_64 ? 2 : 1; ++ vec_bits = ((num == 0 && flag_simdmath_64) ? 64 : 128); + clonei->simdlen = vec_bits / elt_bits; + } + else +diff --git a/gcc/config/aarch64/aarch64.opt b/gcc/config/aarch64/aarch64.opt +index 1b3d942e0f5..4539156d6f4 100644 +--- a/gcc/config/aarch64/aarch64.opt ++++ b/gcc/config/aarch64/aarch64.opt +@@ -190,6 +190,12 @@ precision of square root results to about 16 bits for + single precision and to 32 bits for double precision. + If enabled, it implies -mlow-precision-recip-sqrt. + ++msimdmath-64 ++Target Var(flag_simdmath_64) Optimization ++Allow compiler to generate V2SF 64 bits simdclone of math functions, ++which is not currently supported in mathlib or sleef. ++Therefore this option is disabled by default. ++ + mlow-precision-div + Target Var(flag_mlow_precision_div) Optimization + Enable the division approximation. Enabling this reduces +diff --git a/gcc/fortran/scanner.c b/gcc/fortran/scanner.c +index 6f93508f934..42fd5a8be1e 100644 +--- a/gcc/fortran/scanner.c ++++ b/gcc/fortran/scanner.c +@@ -2737,6 +2737,10 @@ gfc_new_file (void) + && !load_file (flag_pre_include, NULL, false)) + exit (FATAL_EXIT_CODE); + ++ if (flag_simdmath ++ && !load_file ("simdmath_f.h", NULL, false)) ++ exit (FATAL_EXIT_CODE); ++ + if (gfc_cpp_enabled ()) + { + result = gfc_cpp_preprocess (gfc_source_file); +diff --git a/gcc/opts.c b/gcc/opts.c +index 73162528938..e31aa560564 100644 +--- a/gcc/opts.c ++++ b/gcc/opts.c +@@ -189,6 +189,7 @@ static const char use_diagnosed_msg[] = N_("Uses of this option are diagnosed.") + + typedef char *char_p; /* For DEF_VEC_P. */ + ++static void set_simdmath_flags (struct gcc_options *opts, int set); + static void set_debug_level (enum debug_info_type type, int extended, + const char *arg, struct gcc_options *opts, + struct gcc_options *opts_set, +@@ -2469,6 +2470,10 @@ common_handle_option (struct gcc_options *opts, + dc->min_margin_width = value; + break; + ++ case OPT_fsimdmath: ++ set_simdmath_flags (opts, value); ++ break; ++ + case OPT_fdump_: + /* Deferred. */ + break; +@@ -2847,6 +2852,18 @@ common_handle_option (struct gcc_options *opts, + return true; + } + ++/* The following routines are used to set -fno-math-errno and -fopenmp-simd ++ to enable vector mathlib. */ ++static void ++set_simdmath_flags (struct gcc_options *opts, int set) ++{ ++ if (set) ++ { ++ opts->x_flag_errno_math = 0; ++ opts->x_flag_openmp_simd = 1; ++ } ++} ++ + /* Used to set the level of strict aliasing warnings in OPTS, + when no level is specified (i.e., when -Wstrict-aliasing, and not + -Wstrict-aliasing=level was given). +diff --git a/libgomp/Makefile.am b/libgomp/Makefile.am +index 669b9e4defd..0d9cc96481c 100644 +--- a/libgomp/Makefile.am ++++ b/libgomp/Makefile.am +@@ -74,10 +74,10 @@ libgomp_la_SOURCES += openacc.f90 + endif + + nodist_noinst_HEADERS = libgomp_f.h +-nodist_libsubinclude_HEADERS = omp.h openacc.h acc_prof.h ++nodist_libsubinclude_HEADERS = omp.h openacc.h acc_prof.h simdmath.h + if USE_FORTRAN + nodist_finclude_HEADERS = omp_lib.h omp_lib.f90 omp_lib.mod omp_lib_kinds.mod \ +- openacc_lib.h openacc.f90 openacc.mod openacc_kinds.mod ++ openacc_lib.h openacc.f90 openacc.mod openacc_kinds.mod simdmath_f.h + endif + + LTLDFLAGS = $(shell $(SHELL) $(top_srcdir)/../libtool-ldflags $(LDFLAGS)) +diff --git a/libgomp/Makefile.in b/libgomp/Makefile.in +index ae5d9d54705..dd4b334895e 100644 +--- a/libgomp/Makefile.in ++++ b/libgomp/Makefile.in +@@ -148,7 +148,7 @@ am__CONFIG_DISTCLEAN_FILES = config.status config.cache config.log \ + configure.lineno config.status.lineno + mkinstalldirs = $(SHELL) $(top_srcdir)/../mkinstalldirs + CONFIG_HEADER = config.h +-CONFIG_CLEAN_FILES = omp.h omp_lib.h omp_lib.f90 libgomp_f.h \ ++CONFIG_CLEAN_FILES = omp.h omp_lib.h simdmath.h simdmath_f.h omp_lib.f90 libgomp_f.h \ + libgomp.spec + CONFIG_CLEAN_VPATH_FILES = + am__vpath_adj_setup = srcdirstrip=`echo "$(srcdir)" | sed 's|.|.|g'`; +@@ -609,9 +609,9 @@ libgomp_la_SOURCES = alloc.c atomic.c barrier.c critical.c env.c \ + @PLUGIN_GCN_TRUE@libgomp_plugin_gcn_la_LIBADD = libgomp.la $(PLUGIN_GCN_LIBS) + @PLUGIN_GCN_TRUE@libgomp_plugin_gcn_la_LIBTOOLFLAGS = --tag=disable-static + nodist_noinst_HEADERS = libgomp_f.h +-nodist_libsubinclude_HEADERS = omp.h openacc.h acc_prof.h ++nodist_libsubinclude_HEADERS = omp.h openacc.h acc_prof.h simdmath.h + @USE_FORTRAN_TRUE@nodist_finclude_HEADERS = omp_lib.h omp_lib.f90 omp_lib.mod omp_lib_kinds.mod \ +-@USE_FORTRAN_TRUE@ openacc_lib.h openacc.f90 openacc.mod openacc_kinds.mod ++@USE_FORTRAN_TRUE@ openacc_lib.h openacc.f90 openacc.mod openacc_kinds.mod simdmath_f.h + + LTLDFLAGS = $(shell $(SHELL) $(top_srcdir)/../libtool-ldflags $(LDFLAGS)) + LINK = $(LIBTOOL) --tag CC $(AM_LIBTOOLFLAGS) $(LIBTOOLFLAGS) --mode=link \ +@@ -702,6 +702,10 @@ omp.h: $(top_builddir)/config.status $(srcdir)/omp.h.in + cd $(top_builddir) && $(SHELL) ./config.status $@ + omp_lib.h: $(top_builddir)/config.status $(srcdir)/omp_lib.h.in + cd $(top_builddir) && $(SHELL) ./config.status $@ ++simdmath_f.h: $(top_builddir)/config.status $(srcdir)/simdmath_f.h.in ++ cd $(top_builddir) && $(SHELL) ./config.status $@ ++simdmath.h: $(top_builddir)/config.status $(srcdir)/simdmath.h.in ++ cd $(top_builddir) && $(SHELL) ./config.status $@ + omp_lib.f90: $(top_builddir)/config.status $(srcdir)/omp_lib.f90.in + cd $(top_builddir) && $(SHELL) ./config.status $@ + libgomp_f.h: $(top_builddir)/config.status $(srcdir)/libgomp_f.h.in +diff --git a/libgomp/configure b/libgomp/configure +index 5240f7e9d39..b03036c2738 100644 +--- a/libgomp/configure ++++ b/libgomp/configure +@@ -17050,7 +17050,7 @@ fi + + + +-ac_config_files="$ac_config_files omp.h omp_lib.h omp_lib.f90 libgomp_f.h" ++ac_config_files="$ac_config_files omp.h omp_lib.h simdmath.h simdmath_f.h omp_lib.f90 libgomp_f.h" + + ac_config_files="$ac_config_files Makefile testsuite/Makefile libgomp.spec" + +@@ -18205,6 +18205,8 @@ do + "libtool") CONFIG_COMMANDS="$CONFIG_COMMANDS libtool" ;; + "omp.h") CONFIG_FILES="$CONFIG_FILES omp.h" ;; + "omp_lib.h") CONFIG_FILES="$CONFIG_FILES omp_lib.h" ;; ++ "simdmath.h") CONFIG_FILES="$CONFIG_FILES simdmath.h" ;; ++ "simdmath_f.h") CONFIG_FILES="$CONFIG_FILES simdmath_f.h" ;; + "omp_lib.f90") CONFIG_FILES="$CONFIG_FILES omp_lib.f90" ;; + "libgomp_f.h") CONFIG_FILES="$CONFIG_FILES libgomp_f.h" ;; + "Makefile") CONFIG_FILES="$CONFIG_FILES Makefile" ;; +diff --git a/libgomp/configure.ac b/libgomp/configure.ac +index ef5d293c31e..569c2065a66 100644 +--- a/libgomp/configure.ac ++++ b/libgomp/configure.ac +@@ -433,7 +433,7 @@ CFLAGS="$save_CFLAGS" + # Determine what GCC version number to use in filesystem paths. + GCC_BASE_VER + +-AC_CONFIG_FILES(omp.h omp_lib.h omp_lib.f90 libgomp_f.h) ++AC_CONFIG_FILES(omp.h omp_lib.h simdmath.h simdmath_f.h omp_lib.f90 libgomp_f.h) + AC_CONFIG_FILES(Makefile testsuite/Makefile libgomp.spec) + AC_CONFIG_FILES([testsuite/libgomp-test-support.pt.exp:testsuite/libgomp-test-support.exp.in]) + AC_CONFIG_FILES([testsuite/libgomp-site-extra.exp]) +diff --git a/libgomp/simdmath.h.in b/libgomp/simdmath.h.in +new file mode 100644 +index 00000000000..ab91a4ec317 +--- /dev/null ++++ b/libgomp/simdmath.h.in +@@ -0,0 +1,40 @@ ++#ifdef __cplusplus ++extern "C" { ++#endif ++ ++#pragma omp declare simd simdlen(2) notinbranch ++double cos (double x); ++ ++#pragma omp declare simd simdlen(4) notinbranch ++float cosf (float x); ++ ++#pragma omp declare simd simdlen(2) notinbranch ++double sin (double x); ++ ++#pragma omp declare simd simdlen(4) notinbranch ++float sinf (float x); ++ ++#pragma omp declare simd simdlen(2) notinbranch ++double exp (double x); ++ ++#pragma omp declare simd simdlen(4) notinbranch ++float expf (float x); ++ ++#pragma omp declare simd simdlen(2) notinbranch ++double log (double x); ++ ++#pragma omp declare simd simdlen(4) notinbranch ++float logf (float x); ++ ++#pragma omp declare simd simdlen(2) notinbranch ++double pow (double x, double y); ++ ++#pragma omp declare simd simdlen(4) notinbranch ++float powf (float x, float y); ++ ++#pragma omp declare simd simdlen(4) notinbranch ++float exp2f (float x); ++ ++#ifdef __cplusplus ++} // extern "C" ++#endif +diff --git a/libgomp/simdmath_f.h.in b/libgomp/simdmath_f.h.in +new file mode 100644 +index 00000000000..550595015db +--- /dev/null ++++ b/libgomp/simdmath_f.h.in +@@ -0,0 +1,11 @@ ++!GCC$ builtin (cos) attributes simd (notinbranch) ++!GCC$ builtin (cosf) attributes simd (notinbranch) ++!GCC$ builtin (sin) attributes simd (notinbranch) ++!GCC$ builtin (sinf) attributes simd (notinbranch) ++!GCC$ builtin (exp) attributes simd (notinbranch) ++!GCC$ builtin (expf) attributes simd (notinbranch) ++!GCC$ builtin (exp2f) attributes simd (notinbranch) ++!GCC$ builtin (log) attributes simd (notinbranch) ++!GCC$ builtin (logf) attributes simd (notinbranch) ++!GCC$ builtin (pow) attributes simd (notinbranch) ++!GCC$ builtin (powf) attributes simd (notinbranch) +-- +2.21.0.windows.1 + |