def gen_bench_asm_function(f, simd, typ, category):
bench_args_init, bench_args_decl, \
bench_args_call, bench_call = gen_bench_info_from(f, simd, typ)
## Add function that can easily be parsed to get assembly and plain code
return \
'''
void {bench_name}__asm__({type}* _r, {bench_args_decl}, int sz) {{
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
// code:{{
int n = {step};
#pragma clang loop unroll(disable)
for (int i = 0; i < sz; i += n) {{
{bench_call};
}}
// code:}}
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
__asm__ __volatile__("nop");
}}
'''.format(
bench_name=gen_bench_name(category, f.function_name),
type=typ,
step=f.code_ptr_step(typ),
bench_call=bench_call,
bench_args_decl=common.pprint_commas(bench_args_decl)
)
def gen_bench_from_basic_fun(f, simd, typ, category, unroll=None):
bench_args_init, bench_args_decl, bench_args_call, bench_call = \
gen_bench_info_from(f, simd, typ)
bench_name = gen_bench_name(category, f.function_name, unroll)
return \
'''
{code_before}
extern "C" {{ void __asm_marker__{bench_name}() {{}} }}
void {bench_name}(benchmark::State& state, {type}* _r, {bench_args_decl}, int sz) {{
// Normalize size depending on the step so that we're not going out of boundaies
// (Required when the size is'nt a multiple of `n`, like for unrolling benches)
sz = (sz / {step}) * {step};
try {{
for (auto _ : state) {{
{asm_marker}
// code: {bench_name}
int n = {step};
#pragma clang loop unroll(disable)
for (int i = 0; i < sz; i += n) {{
{bench_call};
}}
// code: {bench_name}
{asm_marker}
}}
}} catch (std::exception const& e) {{
std::string message("ERROR: ");
message += e.what();
state.SkipWithError(message.c_str());
}}
}}
BENCHMARK_CAPTURE({bench_name}, {type}, make_data(sz), {bench_args_init}, sz);
'''.format(
bench_name=bench_name,
type=typ,
step=f.code_ptr_step(typ),
bench_call=bench_call,
bench_args_init=common.pprint_commas(bench_args_init),
bench_args_decl=common.pprint_commas(bench_args_decl),
bench_args_call=common.pprint_commas(bench_args_call),
code_before=f.bench_code_before(typ),
asm_marker=asm_marker(simd, bench_name)
)
def code_call(self, typ, args):
return 'nsimd::{}({}, {}(), nsimd::cpu())'. \
format(self.name, common.pprint_commas(args), typ)
def code_call(self, typ, args):
return '{}({})'.format(self.name, common.pprint_commas(args))
