dump1090-fa/starch/templates/dispatcher.c.template

207 lines
6.4 KiB
Plaintext
Raw Normal View History

Move all converters to starch-based implementations (#97) * Switch all conversion routines to use starch. main user-visible changes: * ensure you check out submodules ('git clone --recurse-submodules") * --version shows the CPU features and DSP implementations in use * --wisdom allows overriding of the built-in architecture wisdom * --dcfilter no longer supported * "starch-benchmark" binary will benchmark all options on the current machine and can produce a wisdom file to feed to the --wisdom option If you have a usecase for --dcfilter, please get in touch and let me know - it's an edge case and for now there's no starch/DSP support for it, but support can be written if needed. In almost all cases the new conversion routines are slightly or substantially faster than the old conversion routines. The only case that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower due to changing from heavily approximated lookup tables to higher quality results (but SC16 is probably already out of reach of a Pi 0) * No need to build with SC16Q11_TABLE_BITS any more * Add oneoff/uc8_capture_stats (reads a UC8 capture; measures min/max/mean I and Q) * Switch UC8 conversion to 127.4 center, 128 range. Looking at actual UC8 captures from a RTL2832, the mean I and Q are actually at 127.4, so use that as the zero point. This means that the resulting I/Q maximum values could be as large as 127.6. Switch to 128 for simplicity. * Switch to the new UC8 zero offset in benchmarks, fix some bugs * Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements * Ditch UC8 approximation path, add a NEON VRQSQRTE path. * Tweak the SC16 exact path, add a new impl that uses a mix of u32 & floats. * SC16Q11 impl tweaks: * add a u32->float exact path * ditch the approximation path * add a NEON VRSQRTE path * add a 12-bit table path (using the full signed I/Q value, not absolute value) * Ditch SC16 approximation path, add NEON vrsqrte path * Add oneoff/dsp_error_measurement This runs sample input through the DSP functions that are allowed to be inexact and dumps the results as a TSV suitable for feeding to gnuplot to look at the actual errors. * Update make clean, make wisdom targets * Update wisdom based on benchmarking * Preserve the raw wisdom benchmark data * Update to latest starch * Update .gitignore for new wisdom files * Update starch generated code * Build starch-benchmark as part of the 'all' target * Use wisdom from /etc/dump1090-fa/wisdom.local if present * Package starch-benchmark and a helper script to generate local wisdom data * Remove submodules in preparation for importing them directly * Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0 * Import starch at commit a725c8491dc33a321565d451b385131e589d8490 from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
### Copyright (c) 2020, FlightAware LLC.
### All rights reserved.
### See the LICENSE file for licensing terms.
/* starch generated code. Do not edit. */
#include <stdlib.h>
#include <stdio.h>
#include <ctype.h>
#include <string.h>
#include "${os.path.relpath(gen.generated_include_path, current_dir)}"
/* helper for re-sorting registries */
struct starch_regentry_prefix {
int rank;
};
static int starch_regentry_rank_compare (const void *l, const void *r)
{
const struct starch_regentry_prefix *left = l, *right = r;
return left->rank - right->rank;
}
% for function in sorted(gen.functions.values()):
Move all converters to starch-based implementations (#97) * Switch all conversion routines to use starch. main user-visible changes: * ensure you check out submodules ('git clone --recurse-submodules") * --version shows the CPU features and DSP implementations in use * --wisdom allows overriding of the built-in architecture wisdom * --dcfilter no longer supported * "starch-benchmark" binary will benchmark all options on the current machine and can produce a wisdom file to feed to the --wisdom option If you have a usecase for --dcfilter, please get in touch and let me know - it's an edge case and for now there's no starch/DSP support for it, but support can be written if needed. In almost all cases the new conversion routines are slightly or substantially faster than the old conversion routines. The only case that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower due to changing from heavily approximated lookup tables to higher quality results (but SC16 is probably already out of reach of a Pi 0) * No need to build with SC16Q11_TABLE_BITS any more * Add oneoff/uc8_capture_stats (reads a UC8 capture; measures min/max/mean I and Q) * Switch UC8 conversion to 127.4 center, 128 range. Looking at actual UC8 captures from a RTL2832, the mean I and Q are actually at 127.4, so use that as the zero point. This means that the resulting I/Q maximum values could be as large as 127.6. Switch to 128 for simplicity. * Switch to the new UC8 zero offset in benchmarks, fix some bugs * Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements * Ditch UC8 approximation path, add a NEON VRQSQRTE path. * Tweak the SC16 exact path, add a new impl that uses a mix of u32 & floats. * SC16Q11 impl tweaks: * add a u32->float exact path * ditch the approximation path * add a NEON VRSQRTE path * add a 12-bit table path (using the full signed I/Q value, not absolute value) * Ditch SC16 approximation path, add NEON vrsqrte path * Add oneoff/dsp_error_measurement This runs sample input through the DSP functions that are allowed to be inexact and dumps the results as a TSV suitable for feeding to gnuplot to look at the actual errors. * Update make clean, make wisdom targets * Update wisdom based on benchmarking * Preserve the raw wisdom benchmark data * Update to latest starch * Update .gitignore for new wisdom files * Update starch generated code * Build starch-benchmark as part of the 'all' target * Use wisdom from /etc/dump1090-fa/wisdom.local if present * Package starch-benchmark and a helper script to generate local wisdom data * Remove submodules in preparation for importing them directly * Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0 * Import starch at commit a725c8491dc33a321565d451b385131e589d8490 from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
/* dispatcher / registry for ${function.name} */
${function.regentry_type} * ${function.select_symbol}() {
for (${function.regentry_type} *entry = ${function.registry_symbol};
entry->name;
++entry)
{
if (entry->flavor_supported && !(entry->flavor_supported()))
continue;
return entry;
}
return NULL;
}
static ${function.returntype} ${function.dispatcher_symbol} ( ${function.declaration_arglist} ) {
${function.regentry_type} *entry = ${function.select_symbol}();
if (!entry)
abort();
${function.callable_symbol} = entry->callable;
% if function.returntype == 'void':
${function.callable_symbol} ( ${function.named_arglist} );
% else:
return ${function.callable_symbol} ( ${function.named_arglist} );
% endif
}
${function.pointer_type} ${function.callable_symbol} = ${function.dispatcher_symbol};
void ${function.set_wisdom_symbol} (const char * const * received_wisdom)
{
/* re-rank the registry based on received wisdom */
${function.regentry_type} *entry;
for (entry = ${function.registry_symbol}; entry->name; ++entry) {
const char * const *search;
for (search = received_wisdom; *search; ++search) {
if (!strcmp(*search, entry->name)) {
break;
}
}
if (*search) {
/* matches an entry in the wisdom list, order by position in the list */
entry->rank = search - received_wisdom;
} else {
/* no match, rank after all possible matches, retaining existing order */
entry->rank = (search - received_wisdom) + (entry - ${function.registry_symbol});
}
}
/* re-sort based on the new ranking */
qsort(${function.registry_symbol}, entry - ${function.registry_symbol}, sizeof(${function.regentry_type}), starch_regentry_rank_compare);
/* reset the implementation pointer so the next call will re-select */
${function.callable_symbol} = ${function.dispatcher_symbol};
}
${function.regentry_type} ${function.registry_symbol}[] = {
% for mix in sorted(gen.mixes.values()):
Move all converters to starch-based implementations (#97) * Switch all conversion routines to use starch. main user-visible changes: * ensure you check out submodules ('git clone --recurse-submodules") * --version shows the CPU features and DSP implementations in use * --wisdom allows overriding of the built-in architecture wisdom * --dcfilter no longer supported * "starch-benchmark" binary will benchmark all options on the current machine and can produce a wisdom file to feed to the --wisdom option If you have a usecase for --dcfilter, please get in touch and let me know - it's an edge case and for now there's no starch/DSP support for it, but support can be written if needed. In almost all cases the new conversion routines are slightly or substantially faster than the old conversion routines. The only case that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower due to changing from heavily approximated lookup tables to higher quality results (but SC16 is probably already out of reach of a Pi 0) * No need to build with SC16Q11_TABLE_BITS any more * Add oneoff/uc8_capture_stats (reads a UC8 capture; measures min/max/mean I and Q) * Switch UC8 conversion to 127.4 center, 128 range. Looking at actual UC8 captures from a RTL2832, the mean I and Q are actually at 127.4, so use that as the zero point. This means that the resulting I/Q maximum values could be as large as 127.6. Switch to 128 for simplicity. * Switch to the new UC8 zero offset in benchmarks, fix some bugs * Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements * Ditch UC8 approximation path, add a NEON VRQSQRTE path. * Tweak the SC16 exact path, add a new impl that uses a mix of u32 & floats. * SC16Q11 impl tweaks: * add a u32->float exact path * ditch the approximation path * add a NEON VRSQRTE path * add a 12-bit table path (using the full signed I/Q value, not absolute value) * Ditch SC16 approximation path, add NEON vrsqrte path * Add oneoff/dsp_error_measurement This runs sample input through the DSP functions that are allowed to be inexact and dumps the results as a TSV suitable for feeding to gnuplot to look at the actual errors. * Update make clean, make wisdom targets * Update wisdom based on benchmarking * Preserve the raw wisdom benchmark data * Update to latest starch * Update .gitignore for new wisdom files * Update starch generated code * Build starch-benchmark as part of the 'all' target * Use wisdom from /etc/dump1090-fa/wisdom.local if present * Package starch-benchmark and a helper script to generate local wisdom data * Remove submodules in preparation for importing them directly * Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0 * Import starch at commit a725c8491dc33a321565d451b385131e589d8490 from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
<%
# gather all implementations for this mix, sort by wisdom
def rank_key(value, wisdom=mix.function_wisdom(function)):
impl, flavor = value
try:
return wisdom.index(impl.wisdom_name(flavor))
except ValueError:
return len(wisdom)
mix_impls = []
for flavor in mix.flavors:
if function.aligned and function.aligned_pair:
if flavor.alignment > 1:
# add aligned impls
for impl in function.impls:
if impl.feature is None or impl.feature in flavor.features:
mix_impls.append( (impl, flavor) )
# add unaligned impls
for impl in function.aligned_pair.impls:
if impl.feature is None or impl.feature in flavor.features:
mix_impls.append( (impl, flavor) )
else:
# no alignment specialization
for impl in function.impls:
if impl.feature is None or impl.feature in flavor.features:
mix_impls.append( (impl, flavor) )
mix_impls.sort(key=rank_key)
%>
#ifdef ${mix.macro}
% for rank, (impl, flavor) in enumerate(mix_impls):
{ ${rank}, "${impl.wisdom_name(flavor)}", "${flavor.name}", ${impl.impl_symbol(flavor)}, ${flavor.test_function_expr} },
% endfor
#endif /* ${mix.macro} */
% endfor
{ 0, NULL, NULL, NULL, NULL }
};
% endfor
int ${gen.sym("read_wisdom")} (const char * path)
{
FILE *fp = fopen(path, "r");
if (!fp)
return -1;
/* reset all ranks to identify entries not listed in the wisdom file; we'll assign ranks at the end to produce a stable sort */
% for function in sorted(gen.functions.values()):
Move all converters to starch-based implementations (#97) * Switch all conversion routines to use starch. main user-visible changes: * ensure you check out submodules ('git clone --recurse-submodules") * --version shows the CPU features and DSP implementations in use * --wisdom allows overriding of the built-in architecture wisdom * --dcfilter no longer supported * "starch-benchmark" binary will benchmark all options on the current machine and can produce a wisdom file to feed to the --wisdom option If you have a usecase for --dcfilter, please get in touch and let me know - it's an edge case and for now there's no starch/DSP support for it, but support can be written if needed. In almost all cases the new conversion routines are slightly or substantially faster than the old conversion routines. The only case that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower due to changing from heavily approximated lookup tables to higher quality results (but SC16 is probably already out of reach of a Pi 0) * No need to build with SC16Q11_TABLE_BITS any more * Add oneoff/uc8_capture_stats (reads a UC8 capture; measures min/max/mean I and Q) * Switch UC8 conversion to 127.4 center, 128 range. Looking at actual UC8 captures from a RTL2832, the mean I and Q are actually at 127.4, so use that as the zero point. This means that the resulting I/Q maximum values could be as large as 127.6. Switch to 128 for simplicity. * Switch to the new UC8 zero offset in benchmarks, fix some bugs * Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements * Ditch UC8 approximation path, add a NEON VRQSQRTE path. * Tweak the SC16 exact path, add a new impl that uses a mix of u32 & floats. * SC16Q11 impl tweaks: * add a u32->float exact path * ditch the approximation path * add a NEON VRSQRTE path * add a 12-bit table path (using the full signed I/Q value, not absolute value) * Ditch SC16 approximation path, add NEON vrsqrte path * Add oneoff/dsp_error_measurement This runs sample input through the DSP functions that are allowed to be inexact and dumps the results as a TSV suitable for feeding to gnuplot to look at the actual errors. * Update make clean, make wisdom targets * Update wisdom based on benchmarking * Preserve the raw wisdom benchmark data * Update to latest starch * Update .gitignore for new wisdom files * Update starch generated code * Build starch-benchmark as part of the 'all' target * Use wisdom from /etc/dump1090-fa/wisdom.local if present * Package starch-benchmark and a helper script to generate local wisdom data * Remove submodules in preparation for importing them directly * Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0 * Import starch at commit a725c8491dc33a321565d451b385131e589d8490 from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
int rank_${function.name} = 0;
for (${function.regentry_type} *entry = ${function.registry_symbol}; entry->name; ++entry) {
entry->rank = 0;
}
% endfor
char linebuf[512];
while (fgets(linebuf, sizeof(linebuf), fp)) {
/* split name and impl on whitespace, handle comments etc */
char *name = linebuf;
while (*name && isspace(*name))
++name;
if (!*name || *name == '#')
continue;
char *end = name;
while (*end && !isspace(*end))
++end;
if (!*end)
continue;
*end = 0;
char *impl = end + 1;
while (*impl && isspace(*impl))
++impl;
if (!*impl)
continue;
end = impl;
while (*end && !isspace(*end))
++end;
*end = 0;
/* try to find a matching registry entry */
% for function in sorted(gen.functions.values()):
Move all converters to starch-based implementations (#97) * Switch all conversion routines to use starch. main user-visible changes: * ensure you check out submodules ('git clone --recurse-submodules") * --version shows the CPU features and DSP implementations in use * --wisdom allows overriding of the built-in architecture wisdom * --dcfilter no longer supported * "starch-benchmark" binary will benchmark all options on the current machine and can produce a wisdom file to feed to the --wisdom option If you have a usecase for --dcfilter, please get in touch and let me know - it's an edge case and for now there's no starch/DSP support for it, but support can be written if needed. In almost all cases the new conversion routines are slightly or substantially faster than the old conversion routines. The only case that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower due to changing from heavily approximated lookup tables to higher quality results (but SC16 is probably already out of reach of a Pi 0) * No need to build with SC16Q11_TABLE_BITS any more * Add oneoff/uc8_capture_stats (reads a UC8 capture; measures min/max/mean I and Q) * Switch UC8 conversion to 127.4 center, 128 range. Looking at actual UC8 captures from a RTL2832, the mean I and Q are actually at 127.4, so use that as the zero point. This means that the resulting I/Q maximum values could be as large as 127.6. Switch to 128 for simplicity. * Switch to the new UC8 zero offset in benchmarks, fix some bugs * Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements * Ditch UC8 approximation path, add a NEON VRQSQRTE path. * Tweak the SC16 exact path, add a new impl that uses a mix of u32 & floats. * SC16Q11 impl tweaks: * add a u32->float exact path * ditch the approximation path * add a NEON VRSQRTE path * add a 12-bit table path (using the full signed I/Q value, not absolute value) * Ditch SC16 approximation path, add NEON vrsqrte path * Add oneoff/dsp_error_measurement This runs sample input through the DSP functions that are allowed to be inexact and dumps the results as a TSV suitable for feeding to gnuplot to look at the actual errors. * Update make clean, make wisdom targets * Update wisdom based on benchmarking * Preserve the raw wisdom benchmark data * Update to latest starch * Update .gitignore for new wisdom files * Update starch generated code * Build starch-benchmark as part of the 'all' target * Use wisdom from /etc/dump1090-fa/wisdom.local if present * Package starch-benchmark and a helper script to generate local wisdom data * Remove submodules in preparation for importing them directly * Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0 * Import starch at commit a725c8491dc33a321565d451b385131e589d8490 from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
if (!strcmp(name, "${function.name}")) {
for (${function.regentry_type} *entry = ${function.registry_symbol}; entry->name; ++entry) {
if (!strcmp(impl, entry->name)) {
entry->rank = ++rank_${function.name};
break;
}
}
continue;
}
% endfor
}
if (ferror(fp)) {
fclose(fp);
return -1;
}
fclose(fp);
/* assign ranks to unmatched items to (stable) sort them last; re-sort everything */
% for function in sorted(gen.functions.values()):
Move all converters to starch-based implementations (#97) * Switch all conversion routines to use starch. main user-visible changes: * ensure you check out submodules ('git clone --recurse-submodules") * --version shows the CPU features and DSP implementations in use * --wisdom allows overriding of the built-in architecture wisdom * --dcfilter no longer supported * "starch-benchmark" binary will benchmark all options on the current machine and can produce a wisdom file to feed to the --wisdom option If you have a usecase for --dcfilter, please get in touch and let me know - it's an edge case and for now there's no starch/DSP support for it, but support can be written if needed. In almost all cases the new conversion routines are slightly or substantially faster than the old conversion routines. The only case that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower due to changing from heavily approximated lookup tables to higher quality results (but SC16 is probably already out of reach of a Pi 0) * No need to build with SC16Q11_TABLE_BITS any more * Add oneoff/uc8_capture_stats (reads a UC8 capture; measures min/max/mean I and Q) * Switch UC8 conversion to 127.4 center, 128 range. Looking at actual UC8 captures from a RTL2832, the mean I and Q are actually at 127.4, so use that as the zero point. This means that the resulting I/Q maximum values could be as large as 127.6. Switch to 128 for simplicity. * Switch to the new UC8 zero offset in benchmarks, fix some bugs * Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements * Ditch UC8 approximation path, add a NEON VRQSQRTE path. * Tweak the SC16 exact path, add a new impl that uses a mix of u32 & floats. * SC16Q11 impl tweaks: * add a u32->float exact path * ditch the approximation path * add a NEON VRSQRTE path * add a 12-bit table path (using the full signed I/Q value, not absolute value) * Ditch SC16 approximation path, add NEON vrsqrte path * Add oneoff/dsp_error_measurement This runs sample input through the DSP functions that are allowed to be inexact and dumps the results as a TSV suitable for feeding to gnuplot to look at the actual errors. * Update make clean, make wisdom targets * Update wisdom based on benchmarking * Preserve the raw wisdom benchmark data * Update to latest starch * Update .gitignore for new wisdom files * Update starch generated code * Build starch-benchmark as part of the 'all' target * Use wisdom from /etc/dump1090-fa/wisdom.local if present * Package starch-benchmark and a helper script to generate local wisdom data * Remove submodules in preparation for importing them directly * Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0 * Import starch at commit a725c8491dc33a321565d451b385131e589d8490 from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
{
${function.regentry_type} *entry;
for (entry = ${function.registry_symbol}; entry->name; ++entry) {
if (!entry->rank)
entry->rank = ++rank_${function.name};
}
qsort(${function.registry_symbol}, entry - ${function.registry_symbol}, sizeof(${function.regentry_type}), starch_regentry_rank_compare);
/* reset the implementation pointer so the next call will re-select */
${function.callable_symbol} = ${function.dispatcher_symbol};
}
% endfor
return 0;
}