allow AVX512 instructions to be used with MSVC
cover allow AVX512 instructions to be used with MSVC
Up to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
This patchset allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name. When no name is provided 'native' is assumed meaning that the code should be optimized for the machine compiling the code.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool was also added.
Andre Muezerie (2):
config: allow AVX512 instructions to be used with MSVC devtools/dump-cpu-flags: add tool to update CPU flags table
config/x86/meson.build | 87 ++++--- config/x86/msvc/meson.build | 287 +++++++++++++++++++++ devtools/dump-cpu-flags/README.md | 25 ++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 +++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++ lib/acl/meson.build | 8 +- lib/member/meson.build | 11 +- 8 files changed, 648 insertions(+), 50 deletions(-) create mode 100644 config/x86/msvc/meson.build create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.py
--
2.48.1.vfs.0.0
PATCH 2/2 devtools/dump-cpu-flags: add tool to update CPU flags table
This patchset allows users to specify the CPU for which the generated code should be optimized for by passing the CPU name.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool is introduced with this patch. The new tool generates the table entries for all CPUs listed in an input file using a recent version of the compiler, which has all the information needed. This reduces enormously the amount of work needed to update the table in msvc/meson.build and makes the process much less error prone.
--- devtools/dump-cpu-flags/README.md | 25 +++++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 ++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++++++ 4 files changed, 305 insertions(+) create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.py
devtools/dump-cpu-flags/README.md new file +25 -0
@@ -0,0 +1,25 @@ +# Generating updated CPU flags + +File `config\x86\msvc\meson.build` has a table with flags indicating instruction set support for a variety of CPU types. + +Script `dump-cpu-flags.py` can be used to generate updated entries for this table. + +The CPU names are stored in file `cpu-names.txt`, which is consumed by `dump-cpu-flags.py`. The formatting used in that file is described at the top of the file itself. + +The script relies on the information embedded in the g++ compiler. This means that an updated table can automatically be generated by switching to a newer version of the compiler. This avoids the need to manually edit the entries, which is error prone. With the script the table entries can just copied and pasted into `meson.build`. The only thing that might need to be done is adding new CPU names to cpu-names.txt, when new CPUs are released. + +**NOTE**: CPUs not known to the compiler will result in errors, which can be ignored (`dump-cpu-flags.py` will ignore these errors and continue). For best results use the latest g++ compiler available. + +Below is a sample output, where an error was logged because the compiler did not know about a CPU named ‘raptorlake’. + +```sh +$ ./dump-cpu-flags.py + 'x86-64-v2': [], + 'x86-64-v3': ['AVX', 'AVX2'], + 'x86-64-v4': ['AVX', 'AVX2', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'alderlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], +cc1plus: error: bad value (‘raptorlake’) for ‘-march=’ switch +cc1plus: note: valid arguments to ‘-march=’ switch are: nocona core2 nehalem corei7 westmere sandybridge... + 'silvermont': ['PCLMUL', 'RDRND'], + 'slm': ['PCLMUL', 'RDRND'], +```
devtools/dump-cpu-flags/cpu-names.txt new file +120 -0
120 lines not on this page · raw message
devtools/dump-cpu-flags/dump-cpu-flags.cpp new file +119 -0
119 lines not on this page · raw message
devtools/dump-cpu-flags/dump-cpu-flags.py new file +41 -0
41 lines not on this page · raw message
Signature · 1 line
2.48.1.vfs.0.0
PATCH 1/2 config: allow AVX512 instructions to be used with MSVC
In reply to Andre MuezerieUp to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
This patch allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name.
When no explicit CPU name is passed, 'native' is assumed (like it happens with GCC) and the code will be optimized for the same CPU type used to compile the code.
MSVC does not provide this functionality natively, so logic was added to a new meson.build file under config/x86/msvc to handle these differences, detecting which
instruction sets are supported by the CPU(s), passing the best options to MSVC and setting the correct macros (like __AVX512F__) so that the DPDK code can rely on them like it is done with GCC.
--- config/x86/meson.build | 87 +++++------ config/x86/msvc/meson.build | 287 ++++++++++++++++++++++++++++++++++++ lib/acl/meson.build | 8 +- lib/member/meson.build | 11 +- 4 files changed, 343 insertions(+), 50 deletions(-) create mode 100644 config/x86/msvc/meson.build
config/x86/meson.build +46 -41
@@ -1,6 +1,50 @@ # SPDX-License-Identifier: BSD-3-Clause # Copyright(c) 2017-2020 Intel Corporation +dpdk_conf.set('RTE_ARCH_X86', 1) +if dpdk_conf.get('RTE_ARCH_64') + dpdk_conf.set('RTE_ARCH_X86_64', 1) + dpdk_conf.set('RTE_ARCH', 'x86_64') +else + dpdk_conf.set('RTE_ARCH_I686', 1) + dpdk_conf.set('RTE_ARCH', 'i686') +endif + +dpdk_conf.set('RTE_CACHE_LINE_SIZE', 64) +dpdk_conf.set('RTE_MAX_LCORE', 128) + +epyc_zen_cores = { + '__znver5__':768, + '__znver4__':512, + '__znver3__':256, + '__znver2__':256, + '__znver1__':128 + } + +cpu_instruction_set = get_option('cpu_instruction_set') +if cpu_instruction_set == 'native' + foreach m:epyc_zen_cores.keys() + if cc.get_define(m, args: machine_args) != '' + dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) + break + endif + endforeach +else + foreach m:epyc_zen_cores.keys() + if m.contains(cpu_instruction_set) + dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) + break + endif + endforeach +endif + +dpdk_conf.set('RTE_MAX_NUMA_NODES', 32) + +if is_ms_compiler + subdir('msvc') + subdir_done() +endif + # get binutils version for the workaround of Bug 97 binutils_ok = true if is_linux or cc.get_id() == 'gcc'
@@ -14,7 +58,8 @@ if is_linux or cc.get_id() == 'gcc' endif endif -cc_avx512_flags = ['-mavx512f', '-mavx512vl', '-mavx512dq', '-mavx512bw'] +cc_avx2_flags = ['-mavx2'] +cc_avx512_flags = ['-mavx512f', '-mavx512vl', '-mavx512dq', '-mavx512bw', '-mavx512cd'] cc_has_avx512 = false target_has_avx512 = false if (binutils_ok and cc.has_multi_arguments(cc_avx512_flags)
@@ -82,43 +127,3 @@ foreach f:optional_flags compile_time_cpuflags += ['RTE_CPUFLAG_' + f] endif endforeach - - -dpdk_conf.set('RTE_ARCH_X86', 1) -if dpdk_conf.get('RTE_ARCH_64') - dpdk_conf.set('RTE_ARCH_X86_64', 1) - dpdk_conf.set('RTE_ARCH', 'x86_64') -else - dpdk_conf.set('RTE_ARCH_I686', 1) - dpdk_conf.set('RTE_ARCH', 'i686') -endif - -dpdk_conf.set('RTE_CACHE_LINE_SIZE', 64) -dpdk_conf.set('RTE_MAX_LCORE', 128) - -epyc_zen_cores = { - '__znver5__':768, - '__znver4__':512, - '__znver3__':256, - '__znver2__':256, - '__znver1__':128 - } - -cpu_instruction_set = get_option('cpu_instruction_set') -if cpu_instruction_set == 'native' - foreach m:epyc_zen_cores.keys() - if cc.get_define(m, args: machine_args) != '' - dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) - break - endif - endforeach -else - foreach m:epyc_zen_cores.keys() - if m.contains(cpu_instruction_set) - dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) - break - endif - endforeach -endif - -dpdk_conf.set('RTE_MAX_NUMA_NODES', 32)
config/x86/msvc/meson.build new file +287 -0
287 lines not on this page · raw message
lib/acl/meson.build +2 -6
@@ -55,15 +55,11 @@ if dpdk_conf.has('RTE_ARCH_X86') sources += files('acl_run_avx512.c') cflags += '-DCC_AVX512_SUPPORT' - elif cc.has_multi_arguments('-mavx512f', '-mavx512vl', - '-mavx512cd', '-mavx512bw') - + elif cc.has_multi_arguments(cc_avx512_flags) avx512_tmplib = static_library('avx512_tmp', 'acl_run_avx512.c', dependencies: static_rte_eal, - c_args: cflags + - ['-mavx512f', '-mavx512vl', - '-mavx512cd', '-mavx512bw']) + c_args: cflags + cc_avx512_flags) objs += avx512_tmplib.extract_objects( 'acl_run_avx512.c') cflags += '-DCC_AVX512_SUPPORT'
lib/member/meson.build +8 -3
@@ -33,6 +33,12 @@ if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok # compiler flags, and then have the .o file from static lib # linked into main lib. + if is_ms_compiler + member_avx512_args = cc_avx512_flags + else + member_avx512_args = ['-mavx512f', '-mavx512dq', '-mavx512ifma'] + endif + # check if all required flags already enabled sketch_avx512_flags = ['__AVX512F__', '__AVX512DQ__', '__AVX512IFMA__']
@@ -46,13 +52,12 @@ if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok if sketch_avx512_on == true cflags += ['-DCC_AVX512_SUPPORT'] sources += files('rte_member_sketch_avx512.c') - elif cc.has_multi_arguments('-mavx512f', '-mavx512dq', '-mavx512ifma') + elif cc.has_multi_arguments(member_avx512_args) sketch_avx512_tmp = static_library('sketch_avx512_tmp', 'rte_member_sketch_avx512.c', include_directories: includes, dependencies: [static_rte_eal, static_rte_hash], - c_args: cflags + - ['-mavx512f', '-mavx512dq', '-mavx512ifma']) + c_args: cflags + member_avx512_args) objs += sketch_avx512_tmp.extract_objects('rte_member_sketch_avx512.c') cflags += ['-DCC_AVX512_SUPPORT'] endif
Signature · 1 line
2.48.1.vfs.0.0
Re: [PATCH 1/2] config: allow AVX512 instructions to be used with MSVC
Quoted 11 lines · Andre Muezerie
Up to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170This patch allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name.
When no explicit CPU name is passed, 'native' is assumed (like it happens with GCC) and the code will be optimized for the same CPU type used to compile the code.
Quoted 8 lines
MSVC does not provide this functionality natively, so logic was added to a new meson.build file under config/x86/msvc to handle these differences, detecting which
instruction sets are supported by the CPU(s), passing the best options to MSVC and setting the correct macros (like__AVX512F__) so that the DPDK code can rely on them like it is done with GCC.Signed-off-by Andre Muezerie <[email protected]>
Thanks for splitting out this change from the earlier one, it allows more focused review.
However, I think within this, we can split things a bit further two, and I think this patch could do with being split into separate changes. Specifically:
* one patch for reordering the x86/meson.build file, i.e. just move the code about and put a subdir_done() for MSVC once common part is completed. That patch is a quick review since all you are doing is moving things about.
* separate patch for adding the new msvc file + one-line change to add it to the x86/meson.build file.
* library changes in separate patch - or perhaps two patches so individual maintainers can ack.
Some other comments inline below too.
Thanks,
/Bruce
Quoted 107 lines
--- config/x86/meson.build | 87 +++++------ config/x86/msvc/meson.build | 287 ++++++++++++++++++++++++++++++++++++ lib/acl/meson.build | 8 +- lib/member/meson.build | 11 +- 4 files changed, 343 insertions(+), 50 deletions(-) create mode 100644 config/x86/msvc/meson.buildconfig/x86/meson.build +46
-31@@ -1,6 +1,50 @@ # SPDX-License-Identifier: BSD-3-Clause # Copyright(c) 2017-2020 Intel Corporation +dpdk_conf.set('RTE_ARCH_X86', 1) +if dpdk_conf.get('RTE_ARCH_64') + dpdk_conf.set('RTE_ARCH_X86_64', 1) + dpdk_conf.set('RTE_ARCH', 'x86_64') +else + dpdk_conf.set('RTE_ARCH_I686', 1) + dpdk_conf.set('RTE_ARCH', 'i686') +endif + +dpdk_conf.set('RTE_CACHE_LINE_SIZE', 64) +dpdk_conf.set('RTE_MAX_LCORE', 128) + +epyc_zen_cores = { + '__znver5__':768, + '__znver4__':512, + '__znver3__':256, + '__znver2__':256, + '__znver1__':128 + } + +cpu_instruction_set = get_option('cpu_instruction_set') +if cpu_instruction_set == 'native' + foreach m:epyc_zen_cores.keys() + if cc.get_define(m, args: machine_args) != '' + dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) + break + endif + endforeach +else + foreach m:epyc_zen_cores.keys() + if m.contains(cpu_instruction_set) + dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) + break + endif + endforeach +endif + +dpdk_conf.set('RTE_MAX_NUMA_NODES', 32) + +if is_ms_compiler + subdir('msvc') + subdir_done() +endif + # get binutils version for the workaround of Bug 97 binutils_ok = true if is_linux or cc.get_id() == 'gcc'@@ -14,7 +58,8 @@ if is_linux or cc.get_id() == 'gcc' endif endif -cc_avx512_flags = ['-mavx512f', '-mavx512vl', '-mavx512dq', '-mavx512bw'] +cc_avx2_flags = ['-mavx2'] +cc_avx512_flags = ['-mavx512f', '-mavx512vl', '-mavx512dq', '-mavx512bw', '-mavx512cd'] cc_has_avx512 = false target_has_avx512 = false if (binutils_ok and cc.has_multi_arguments(cc_avx512_flags)@@ -82,43 +127,3 @@ foreach f:optional_flags compile_time_cpuflags += ['RTE_CPUFLAG_' + f] endif endforeach - - -dpdk_conf.set('RTE_ARCH_X86', 1) -if dpdk_conf.get('RTE_ARCH_64') - dpdk_conf.set('RTE_ARCH_X86_64', 1) - dpdk_conf.set('RTE_ARCH', 'x86_64') -else - dpdk_conf.set('RTE_ARCH_I686', 1) - dpdk_conf.set('RTE_ARCH', 'i686') -endif - -dpdk_conf.set('RTE_CACHE_LINE_SIZE', 64) -dpdk_conf.set('RTE_MAX_LCORE', 128) - -epyc_zen_cores = { - '__znver5__':768, - '__znver4__':512, - '__znver3__':256, - '__znver2__':256, - '__znver1__':128 - } - -cpu_instruction_set = get_option('cpu_instruction_set') -if cpu_instruction_set == 'native' - foreach m:epyc_zen_cores.keys() - if cc.get_define(m, args: machine_args) != '' - dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) - break - endif - endforeach
config/x86/meson.build foreach f:optional_flags · line 124
Quoted 10 lines
config/x86/meson.build +0
-10-else - foreach m:epyc_zen_cores.keys() - if m.contains(cpu_instruction_set) - dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) - break - endif - endforeach -endif - -dpdk_conf.set('RTE_MAX_NUMA_NODES', 32)
For changes to config/x86/meson.build you can add my Ack if it's a separate patch.
Quoted 225 lines
config/x86/msvc/meson.build new file +219
-0@@ -0,0 +1,287 @@ +# SPDX-License-Identifier: BSD-3-Clause +# Copyright(c) 2025 Microsoft Corporation + +cc_avx2_flags = ['/arch:AVX2'] +cc_avx512_flags = ['/arch:AVX512'] +cc_has_avx512 = true + +cpuid_code = ''' + #include <stdio.h> + #include <stdint.h> + #include <intrin.h> + + uint32_t f1_ECX = 0; + uint32_t f1_EDX = 0; + uint32_t f7_EBX = 0; + uint32_t f7_ECX = 0; + + void get_support_flags() + { + int ids_max; + int data[4]; + + /* + * Calling __cpuid with 0x0 as the function_id argument + * gets the number of the highest valid function ID. + */ + __cpuid(data, 0); + ids_max = data[0]; + + if (1 <= ids_max) { + __cpuidex(data, 1, 0); + f1_ECX = data[2]; + f1_EDX = data[3]; + + if (7 <= ids_max) { + __cpuidex(data, 7, 0); + f7_EBX = data[1]; + f7_ECX = data[2]; + } + } + } + + int get_instruction_support() + { + get_support_flags(); + + #ifdef SSE3 + return (f1_ECX & (1UL << 0)) ? 1 : 0; + #endif + #ifdef PCLMUL + return (f1_ECX & (1UL << 1)) ? 1 : 0; + #endif + #ifdef SSSE3 + return (f1_ECX & (1UL << 9)) ? 1 : 0; + #endif + #ifdef SSE4_1 + return (f1_ECX & (1UL << 19)) ? 1 : 0; + #endif + #ifdef SSE4_2 + return (f1_ECX & (1UL << 20)) ? 1 : 0; + #endif + #ifdef AES + return (f1_ECX & (1UL << 25)) ? 1 : 0; + #endif + #ifdef AVX + return (f1_ECX & (1UL << 28)) ? 1 : 0; + #endif + #ifdef RDRND + return (f1_ECX & (1UL << 30)) ? 1 : 0; + #endif + #ifdef SSE + return (f1_EDX & (1UL << 25)) ? 1 : 0; + #endif + #ifdef SSE2 + return (f1_EDX & (1UL << 26)) ? 1 : 0; + #endif + #ifdef AVX2 + return (f7_EBX & (1UL << 5)) ? 1 : 0; + #endif + #ifdef AVX512F + return (f7_EBX & (1UL << 16)) ? 1 : 0; + #endif + #ifdef AVX512DQ + return (f7_EBX & (1UL << 17)) ? 1 : 0; + #endif + #ifdef RDSEED + return (f7_EBX & (1UL << 18)) ? 1 : 0; + #endif + #ifdef AVX512IFMA + return (f7_EBX & (1UL << 21)) ? 1 : 0; + #endif + #ifdef AVX512CD + return (f7_EBX & (1UL << 28)) ? 1 : 0; + #endif + #ifdef AVX512BW + return (f7_EBX & (1UL << 30)) ? 1 : 0; + #endif + #ifdef AVX512VL + return (f7_EBX & (1UL << 31)) ? 1 : 0; + #endif + #ifdef GFNI + return (f7_ECX & (1UL << 8)) ? 1 : 0; + #endif + #ifdef VPCLMULQDQ + return (f7_ECX & (1UL << 10)) ? 1 : 0; + #endif + + return -1; + } + + int main(int argc, char *argv[]) + { + int res = get_instruction_support(); + if (res == -1) { + printf("Unknown instruction set"); + return -1; + } + printf("%d", res); + + return 0; + } +''' + +# The data in the table below can be found here: +# https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html +# A tool to easily update this table can be found under devtools/dump-cpu-flags. +# The table only contains CPUs that have SSE4.2, as this instruction set is required by DPDK. +# That means that in addition to the instruction sets mentioned in the table, all these CPUs +# also have ['SSE', 'SSE2', 'SSE3', 'SSEE3', 'SSE4_1', 'SSE4_2'] +cpu_type_to_flags = { + 'x86-64-v2': [], + 'x86-64-v3': ['AVX', 'AVX2'], + 'x86-64-v4': ['AVX', 'AVX2', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'nehalem': [], + 'corei7': [], + 'westmere': ['PCLMUL'], + 'sandybridge': ['AVX', 'PCLMUL'], + 'corei7-avx': ['AVX', 'PCLMUL'], + 'ivybridge': ['AVX', 'PCLMUL', 'RDRND'], + 'core-avx-i': ['AVX', 'PCLMUL', 'RDRND'], + 'haswell': ['AVX', 'PCLMUL', 'RDRND', 'AVX2'], + 'core-avx2': ['AVX', 'PCLMUL', 'RDRND', 'AVX2'], + 'broadwell': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED'], + 'skylake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'skylake-avx512': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'cascadelake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'cannonlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA'], + 'cooperlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'icelake-client': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'icelake-server': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'tigerlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'rocketlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'alderlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'raptorlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'meteorlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'gracemont': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'arrowlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'arrowlake-s': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'lunarlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'pantherlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'sapphirerapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'emeraldrapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'graniterapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'graniterapids-d': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'diamondrapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'silvermont': ['PCLMUL', 'RDRND'], + 'slm': ['PCLMUL', 'RDRND'], + 'goldmont': ['PCLMUL', 'RDRND', 'RDSEED', 'AES'], + 'goldmont-plus': ['PCLMUL', 'RDRND', 'RDSEED', 'AES'], + 'tremont': ['PCLMUL', 'RDRND', 'RDSEED', 'AES', 'GFNI'], + 'sierraforest': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'grandridge': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], +'clearwaterforest': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'bdver1': ['AVX', 'PCLMUL', 'AES'], + 'bdver2': ['AVX', 'PCLMUL', 'AES'], + 'bdver3': ['AVX', 'PCLMUL', 'AES'], + 'bdver4': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'AES'], + 'znver1': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'znver2': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'znver3': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ'], + 'znver4': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'znver5': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'btver2': ['AVX', 'PCLMUL', 'AES'], + 'lujiazui': ['PCLMUL', 'RDRND', 'RDSEED', 'AES'], + 'yongfeng': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'shijidadao': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], +} + +# Determine cpu_flags for a given configuration. +# SSE instructions up to 4.2 are required for DPDK. +cpu_flags = ['SSE', 'SSE2', 'SSE3', 'SSEE3', 'SSE4_1', 'SSE4_2'] + +message('cpu_instruction_set: @0@'.format(cpu_instruction_set)) + +if cpu_instruction_set == '' + # Nothing to do as cpu_flags already holds all the required flags. +elif cpu_instruction_set == 'native' + # MSVC behaves differently than GCC regarding supported instruction sets. + # While GCC will create macros like __AVX512F__ when such instruction set is + # supported by the current CPU, MSVC does not do that. MSVC will create that + # macro when parameter /arch:AVX512 is passed to the compiler, even when the + # CPU does not have that instruction set (by design). So there's a need to + # look at CPUID flags to figure out what is really supported by the CPU, so + # that the correct /arch value can be passed to the compiler. + # The macros also need to be explicitly defined, as /arch will not create all + # macros GCC creates under the same conditions. + # As an example, /arch:AVX512 creates __AVX512BW__, but does not create __SSE2__. + # More details available here: + # https://learn.microsoft.com/en-us/cpp/preprocessor/predefined-macros + + optional_flags = [ + 'PCLMUL', + 'AES', + 'AVX', + 'RDRND', + 'AVX2', + 'AVX512F', + 'AVX512BW', + 'AVX512DQ',
config/x86/msvc/meson.build line 229
Quoted 10 lines
config/x86/msvc/meson.build new file +10
-0+ 'AVX512VL', + 'AVX512CD', + 'AVX512IFMA', + 'GFNI', + 'RDSEED', + 'VPCLMULQDQ', + ] + foreach f:optional_flags + result = cc.run(cpuid_code, args: '-D@0@'.format(f), + name: 'instruction set @0@'.format(f))
Is building a new binary for each instruction set the best way to do this? Would it not be better to have a single binary that outputs all the instruction sets in one go?
Quoted 70 lines
config/x86/msvc/meson.build new file +58
-0+ has_instr_set = result.returncode() == 0 and result.stdout() == '1' + if has_instr_set + cpu_flags += f + endif + message('Target has @0@: @1@'.format(f, has_instr_set)) + endforeach +else + # An explicit cpu_instruction_set was provided. Get cpu_flags + # from cpu_type_to_flags table. + if cpu_instruction_set not in cpu_type_to_flags + error('CPU not known or not supported. Please update the table with known CPUs if needed.') + endif + cpu_flags += cpu_type_to_flags[cpu_instruction_set] +endif + +# Now that all cpu_flags are known, set compile_time_cpuflags and also +# machine_args to ensure that the instruction set #defines (like __SSE2__) +# are always present in the preprocessor. +message('cpu_flags: @0@'.format(cpu_flags)) + +foreach flag:cpu_flags + machine_args += '/D__@0@__'.format(flag) + if flag == 'PCLMUL' + flag = 'PCLMULQDQ' + elif flag == 'RDRND' + flag = 'RDRAND' + endif + compile_time_cpuflags += ['RTE_CPUFLAG_' + flag] +endforeach + +# Per https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170 +# option '/arch:AVX512' enables all five flags used in the expression below. +target_has_avx512 = ('AVX512F' in cpu_flags and + 'AVX512BW' in cpu_flags and + 'AVX512DQ' in cpu_flags and + 'AVX512CD' in cpu_flags and + 'AVX512VL' in cpu_flags) + +# Decide which instruction sets should be used by the compiler. +# With MSVC, intrinsic functions are always enabled. However, for the +# compiler to use an extended instruction set for automatically +# generated code "/arch" needs to be passed. So we instruct the compiler +# to use the largest set that is supported by the CPU. It is implied that +# smaller sets than the largest selected are included, as described here: +# https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170 +if 'RTE_CPUFLAG_AVX512F' in compile_time_cpuflags + machine_args += ['/arch:AVX512'] +elif 'RTE_CPUFLAG_AVX2' in compile_time_cpuflags + machine_args += ['/arch:AVX2'] +elif 'RTE_CPUFLAG_AVX' in compile_time_cpuflags + machine_args += ['/arch:AVX'] +else + # SSE4.2 is expected to always be available + machine_args += ['/arch:SSE4.2'] +endif + +message('machine_args: @0@'.format(machine_args)) +message('compile_time_cpuflags: @0@'.format(compile_time_cpuflags))lib/acl/meson.build +1
-3@@ -55,15 +55,11 @@ if dpdk_conf.has('RTE_ARCH_X86') sources += files('acl_run_avx512.c') cflags += '-DCC_AVX512_SUPPORT' - elif cc.has_multi_arguments('-mavx512f', '-mavx512vl', - '-mavx512cd', '-mavx512bw') - + elif cc.has_multi_arguments(cc_avx512_flags)
lib/acl/meson.build if dpdk_conf.has('RTE_ARCH_X86') · line 65
Quoted 10 lines
lib/acl/meson.build +1
-3avx512_tmplib = static_library('avx512_tmp', 'acl_run_avx512.c', dependencies: static_rte_eal, - c_args: cflags + - ['-mavx512f', '-mavx512vl', - '-mavx512cd', '-mavx512bw']) + c_args: cflags + cc_avx512_flags) objs += avx512_tmplib.extract_objects( 'acl_run_avx512.c') cflags += '-DCC_AVX512_SUPPORT'
Ack from me for this change too.
Quoted 5 lines
lib/member/meson.build +0
-0
lib/member/meson.build if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok · line 41
Quoted 9 lines
lib/member/meson.build +6
-0@@ -33,6 +33,12 @@ if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok # compiler flags, and then have the .o file from static lib # linked into main lib. + if is_ms_compiler + member_avx512_args = cc_avx512_flags + else + member_avx512_args = ['-mavx512f', '-mavx512dq', '-mavx512ifma'] + endif +
is ifma included as msvc as part of the AVX512 flags? How does this work for the windows build?
Quoted 22 lines
lib/member/meson.build +2
-3# check if all required flags already enabled sketch_avx512_flags = ['__AVX512F__', '__AVX512DQ__', '__AVX512IFMA__']@@ -46,13 +52,12 @@ if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok if sketch_avx512_on == true cflags += ['-DCC_AVX512_SUPPORT'] sources += files('rte_member_sketch_avx512.c') - elif cc.has_multi_arguments('-mavx512f', '-mavx512dq', '-mavx512ifma') + elif cc.has_multi_arguments(member_avx512_args) sketch_avx512_tmp = static_library('sketch_avx512_tmp', 'rte_member_sketch_avx512.c', include_directories: includes, dependencies: [static_rte_eal, static_rte_hash], - c_args: cflags + - ['-mavx512f', '-mavx512dq', '-mavx512ifma']) + c_args: cflags + member_avx512_args) objs += sketch_avx512_tmp.extract_objects('rte_member_sketch_avx512.c') cflags += ['-DCC_AVX512_SUPPORT'] endifSignature · 1 line
2.48.1.vfs.0.0
Re: [PATCH 2/2] devtools/dump-cpu-flags: add tool to update CPU flags table
In reply to Andre MuezerieQuoted 47 lines · Andre Muezerie
This patchset allows users to specify the CPU for which the generated code should be optimized for by passing the CPU name.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool is introduced with this patch. The new tool generates the table entries for all CPUs listed in an input file using a recent version of the compiler, which has all the information needed. This reduces enormously the amount of work needed to update the table in msvc/meson.build and makes the process much less error prone.Signed-off-by Andre Muezerie <[email protected]>--- devtools/dump-cpu-flags/README.md | 25 +++++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 ++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++++++ 4 files changed, 305 insertions(+) create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.pydevtools/dump-cpu-flags/README.md new file +16
-0@@ -0,0 +1,25 @@ +# Generating updated CPU flags + +File `config\x86\msvc\meson.build` has a table with flags indicating instruction set support for a variety of CPU types. + +Script `dump-cpu-flags.py` can be used to generate updated entries for this table. + +The CPU names are stored in file `cpu-names.txt`, which is consumed by `dump-cpu-flags.py`. The formatting used in that file is described at the top of the file itself. + +The script relies on the information embedded in the g++ compiler. This means that an updated table can automatically be generated by switching to a newer version of the compiler. This avoids the need to manually edit the entries, which is error prone. With the script the table entries can just copied and pasted into `meson.build`. The only thing that might need to be done is adding new CPU names to cpu-names.txt, when new CPUs are released. + +**NOTE**: CPUs not known to the compiler will result in errors, which can be ignored (`dump-cpu-flags.py` will ignore these errors and continue). For best results use the latest g++ compiler available. + +Below is a sample output, where an error was logged because the compiler did not know about a CPU named ‘raptorlake’. + +```sh +$ ./dump-cpu-flags.py
devtools/dump-cpu-flags/README.md line 25
Quoted 10 lines
devtools/dump-cpu-flags/README.md new file +9
-0+ 'x86-64-v2': [], + 'x86-64-v3': ['AVX', 'AVX2'], + 'x86-64-v4': ['AVX', 'AVX2', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'alderlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], +cc1plus: error: bad value (‘raptorlake’) for ‘-march=’ switch +cc1plus: note: valid arguments to ‘-march=’ switch are: nocona core2 nehalem corei7 westmere sandybridge... + 'silvermont': ['PCLMUL', 'RDRND'], + 'slm': ['PCLMUL', 'RDRND'], +```
How about having the tool output a valid meson.build file, that can then be used directly without copy-paste. While I know such a thing would end up with us having deep subdir structures, it could be just loaded from e.g. config/x86/msvc/cpu-flags/meson.build, for example.
/Bruce
PATCH v2 1/2 config: allow AVX512 instructions to be used with MSVC
In reply to Andre MuezerieUp to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
This patch allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name.
When no explicit CPU name is passed, 'native' is assumed (like it happens with GCC) and the code will be optimized for the same CPU type used to compile the code.
MSVC does not provide this functionality natively, so logic was added to a new meson.build file under config/x86/msvc to handle these differences, detecting which
instruction sets are supported by the CPU(s), passing the best options to MSVC and setting the correct macros (like __AVX512F__) so that the DPDK code can rely on them like it is done with GCC.
--- config/x86/msvc/meson.build | 287 ++++++++++++++++++++++++++++++++++++ 1 file changed, 287 insertions(+) create mode 100644 config/x86/msvc/meson.build
config/x86/msvc/meson.build new file +287 -0
@@ -0,0 +1,287 @@ +# SPDX-License-Identifier: BSD-3-Clause +# Copyright(c) 2025 Microsoft Corporation + +cc_avx2_flags = ['/arch:AVX2'] +cc_avx512_flags = ['/arch:AVX512'] +cc_has_avx512 = true + +cpuid_code = ''' + #include <stdio.h> + #include <stdint.h> + #include <intrin.h> + + uint32_t f1_ECX = 0; + uint32_t f1_EDX = 0; + uint32_t f7_EBX = 0; + uint32_t f7_ECX = 0; + + void get_support_flags() + { + int ids_max; + int data[4]; + + /* + * Calling __cpuid with 0x0 as the function_id argument + * gets the number of the highest valid function ID. + */ + __cpuid(data, 0); + ids_max = data[0]; + + if (1 <= ids_max) { + __cpuidex(data, 1, 0); + f1_ECX = data[2]; + f1_EDX = data[3]; + + if (7 <= ids_max) { + __cpuidex(data, 7, 0); + f7_EBX = data[1]; + f7_ECX = data[2]; + } + } + } + + int get_instruction_support() + { + get_support_flags(); + + #ifdef SSE3 + return (f1_ECX & (1UL << 0)) ? 1 : 0; + #endif + #ifdef PCLMUL + return (f1_ECX & (1UL << 1)) ? 1 : 0; + #endif + #ifdef SSSE3 + return (f1_ECX & (1UL << 9)) ? 1 : 0; + #endif + #ifdef SSE4_1 + return (f1_ECX & (1UL << 19)) ? 1 : 0; + #endif + #ifdef SSE4_2 + return (f1_ECX & (1UL << 20)) ? 1 : 0; + #endif + #ifdef AES + return (f1_ECX & (1UL << 25)) ? 1 : 0; + #endif + #ifdef AVX + return (f1_ECX & (1UL << 28)) ? 1 : 0; + #endif + #ifdef RDRND + return (f1_ECX & (1UL << 30)) ? 1 : 0; + #endif + #ifdef SSE + return (f1_EDX & (1UL << 25)) ? 1 : 0; + #endif + #ifdef SSE2 + return (f1_EDX & (1UL << 26)) ? 1 : 0; + #endif + #ifdef AVX2 + return (f7_EBX & (1UL << 5)) ? 1 : 0; + #endif + #ifdef AVX512F + return (f7_EBX & (1UL << 16)) ? 1 : 0; + #endif + #ifdef AVX512DQ + return (f7_EBX & (1UL << 17)) ? 1 : 0; + #endif + #ifdef RDSEED + return (f7_EBX & (1UL << 18)) ? 1 : 0; + #endif + #ifdef AVX512IFMA + return (f7_EBX & (1UL << 21)) ? 1 : 0; + #endif + #ifdef AVX512CD + return (f7_EBX & (1UL << 28)) ? 1 : 0; + #endif + #ifdef AVX512BW + return (f7_EBX & (1UL << 30)) ? 1 : 0; + #endif + #ifdef AVX512VL + return (f7_EBX & (1UL << 31)) ? 1 : 0; + #endif + #ifdef GFNI + return (f7_ECX & (1UL << 8)) ? 1 : 0; + #endif + #ifdef VPCLMULQDQ + return (f7_ECX & (1UL << 10)) ? 1 : 0; + #endif + + return -1; + } + + int main(int argc, char *argv[]) + { + int res = get_instruction_support(); + if (res == -1) { + printf("Unknown instruction set"); + return -1; + } + printf("%d", res); + + return 0; + } +''' + +# The data in the table below can be found here: +# https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html +# A tool to easily update this table can be found under devtools/dump-cpu-flags. +# The table only contains CPUs that have SSE4.2, as this instruction set is required by DPDK. +# That means that in addition to the instruction sets mentioned in the table, all these CPUs +# also have ['SSE', 'SSE2', 'SSE3', 'SSEE3', 'SSE4_1', 'SSE4_2'] +cpu_type_to_flags = { + 'x86-64-v2': [], + 'x86-64-v3': ['AVX', 'AVX2'], + 'x86-64-v4': ['AVX', 'AVX2', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'nehalem': [], + 'corei7': [], + 'westmere': ['PCLMUL'], + 'sandybridge': ['AVX', 'PCLMUL'], + 'corei7-avx': ['AVX', 'PCLMUL'], + 'ivybridge': ['AVX', 'PCLMUL', 'RDRND'], + 'core-avx-i': ['AVX', 'PCLMUL', 'RDRND'], + 'haswell': ['AVX', 'PCLMUL', 'RDRND', 'AVX2'], + 'core-avx2': ['AVX', 'PCLMUL', 'RDRND', 'AVX2'], + 'broadwell': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED'], + 'skylake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'skylake-avx512': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'cascadelake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'cannonlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA'], + 'cooperlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'icelake-client': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'icelake-server': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'tigerlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'rocketlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'alderlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'raptorlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'meteorlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'gracemont': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'arrowlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'arrowlake-s': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'lunarlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'pantherlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'sapphirerapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'emeraldrapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'graniterapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'graniterapids-d': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'diamondrapids': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'silvermont': ['PCLMUL', 'RDRND'], + 'slm': ['PCLMUL', 'RDRND'], + 'goldmont': ['PCLMUL', 'RDRND', 'RDSEED', 'AES'], + 'goldmont-plus': ['PCLMUL', 'RDRND', 'RDSEED', 'AES'], + 'tremont': ['PCLMUL', 'RDRND', 'RDSEED', 'AES', 'GFNI'], + 'sierraforest': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'grandridge': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], +'clearwaterforest': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], + 'bdver1': ['AVX', 'PCLMUL', 'AES'], + 'bdver2': ['AVX', 'PCLMUL', 'AES'], + 'bdver3': ['AVX', 'PCLMUL', 'AES'], + 'bdver4': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'AES'], + 'znver1': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'znver2': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'znver3': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ'], + 'znver4': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'znver5': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD', 'AVX512IFMA', 'GFNI'], + 'btver2': ['AVX', 'PCLMUL', 'AES'], + 'lujiazui': ['PCLMUL', 'RDRND', 'RDSEED', 'AES'], + 'yongfeng': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], + 'shijidadao': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES'], +} + +# Determine cpu_flags for a given configuration. +# SSE instructions up to 4.2 are required for DPDK. +cpu_flags = ['SSE', 'SSE2', 'SSE3', 'SSEE3', 'SSE4_1', 'SSE4_2'] + +message('cpu_instruction_set: @0@'.format(cpu_instruction_set)) + +if cpu_instruction_set == '' + # Nothing to do as cpu_flags already holds all the required flags. +elif cpu_instruction_set == 'native' + # MSVC behaves differently than GCC regarding supported instruction sets. + # While GCC will create macros like __AVX512F__ when such instruction set is + # supported by the current CPU, MSVC does not do that. MSVC will create that + # macro when parameter /arch:AVX512 is passed to the compiler, even when the + # CPU does not have that instruction set (by design). So there's a need to + # look at CPUID flags to figure out what is really supported by the CPU, so + # that the correct /arch value can be passed to the compiler. + # The macros also need to be explicitly defined, as /arch will not create all + # macros GCC creates under the same conditions. + # As an example, /arch:AVX512 creates __AVX512BW__, but does not create __SSE2__. + # More details available here: + # https://learn.microsoft.com/en-us/cpp/preprocessor/predefined-macros + + optional_flags = [ + 'PCLMUL', + 'AES', + 'AVX', + 'RDRND', + 'AVX2', + 'AVX512F', + 'AVX512BW', + 'AVX512DQ', + 'AVX512VL', + 'AVX512CD', + 'AVX512IFMA', + 'GFNI', + 'RDSEED', + 'VPCLMULQDQ', + ] + foreach f:optional_flags + result = cc.run(cpuid_code, args: '-D@0@'.format(f), + name: 'instruction set @0@'.format(f)) + has_instr_set = result.returncode() == 0 and result.stdout() == '1' + if has_instr_set + cpu_flags += f + endif + message('Target has @0@: @1@'.format(f, has_instr_set)) + endforeach +else + # An explicit cpu_instruction_set was provided. Get cpu_flags + # from cpu_type_to_flags table. + if cpu_instruction_set not in cpu_type_to_flags + error('CPU not known or not supported. Please update the table with known CPUs if needed.') + endif + cpu_flags += cpu_type_to_flags[cpu_instruction_set] +endif + +# Now that all cpu_flags are known, set compile_time_cpuflags and also +# machine_args to ensure that the instruction set #defines (like __SSE2__) +# are always present in the preprocessor. +message('cpu_flags: @0@'.format(cpu_flags)) + +foreach flag:cpu_flags + machine_args += '/D__@0@__'.format(flag) + if flag == 'PCLMUL' + flag = 'PCLMULQDQ' + elif flag == 'RDRND' + flag = 'RDRAND' + endif + compile_time_cpuflags += ['RTE_CPUFLAG_' + flag] +endforeach + +# Per https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170 +# option '/arch:AVX512' enables all five flags used in the expression below. +target_has_avx512 = ('AVX512F' in cpu_flags and + 'AVX512BW' in cpu_flags and + 'AVX512DQ' in cpu_flags and + 'AVX512CD' in cpu_flags and + 'AVX512VL' in cpu_flags) + +# Decide which instruction sets should be used by the compiler. +# With MSVC, intrinsic functions are always enabled. However, for the +# compiler to use an extended instruction set for automatically +# generated code "/arch" needs to be passed. So we instruct the compiler +# to use the largest set that is supported by the CPU. It is implied that +# smaller sets than the largest selected are included, as described here: +# https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170 +if 'RTE_CPUFLAG_AVX512F' in compile_time_cpuflags + machine_args += ['/arch:AVX512'] +elif 'RTE_CPUFLAG_AVX2' in compile_time_cpuflags + machine_args += ['/arch:AVX2'] +elif 'RTE_CPUFLAG_AVX' in compile_time_cpuflags + machine_args += ['/arch:AVX'] +else + # SSE4.2 is expected to always be available + machine_args += ['/arch:SSE4.2'] +endif + +message('machine_args: @0@'.format(machine_args)) +message('compile_time_cpuflags: @0@'.format(compile_time_cpuflags))
Signature · 1 line
2.48.1.vfs.0.0
v2 cover allow AVX512 instructions to be used with MSVC
In reply to Andre MuezerieUp to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
This patchset allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name. When no name is provided 'native' is assumed meaning that the code should be optimized for the machine compiling the code.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool was also added.
TODO: Add subdir('msvc'). We can't do this until the common part in config/meson.build is merged.
Andre Muezerie (2):
config: allow AVX512 instructions to be used with MSVC devtools/dump-cpu-flags: add tool to update CPU flags table
config/x86/msvc/meson.build | 287 +++++++++++++++++++++ devtools/dump-cpu-flags/README.md | 25 ++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 +++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++ 5 files changed, 592 insertions(+) create mode 100644 config/x86/msvc/meson.build create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.py
--
2.48.1.vfs.0.0
PATCH v2 2/2 devtools/dump-cpu-flags: add tool to update CPU flags table
This patchset allows users to specify the CPU for which the generated code should be optimized for by passing the CPU name.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool is introduced with this patch. The new tool generates the table entries for all CPUs listed in an input file using a recent version of the compiler, which has all the information needed. This reduces enormously the amount of work needed to update the table in msvc/meson.build and makes the process much less error prone.
--- devtools/dump-cpu-flags/README.md | 25 +++++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 ++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++++++ 4 files changed, 305 insertions(+) create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.py
devtools/dump-cpu-flags/README.md new file +25 -0
@@ -0,0 +1,25 @@ +# Generating updated CPU flags + +File `config\x86\msvc\meson.build` has a table with flags indicating instruction set support for a variety of CPU types. + +Script `dump-cpu-flags.py` can be used to generate updated entries for this table. + +The CPU names are stored in file `cpu-names.txt`, which is consumed by `dump-cpu-flags.py`. The formatting used in that file is described at the top of the file itself. + +The script relies on the information embedded in the g++ compiler. This means that an updated table can automatically be generated by switching to a newer version of the compiler. This avoids the need to manually edit the entries, which is error prone. With the script the table entries can just copied and pasted into `meson.build`. The only thing that might need to be done is adding new CPU names to cpu-names.txt, when new CPUs are released. + +**NOTE**: CPUs not known to the compiler will result in errors, which can be ignored (`dump-cpu-flags.py` will ignore these errors and continue). For best results use the latest g++ compiler available. + +Below is a sample output, where an error was logged because the compiler did not know about a CPU named ‘raptorlake’. + +```sh +$ ./dump-cpu-flags.py + 'x86-64-v2': [], + 'x86-64-v3': ['AVX', 'AVX2'], + 'x86-64-v4': ['AVX', 'AVX2', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'alderlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], +cc1plus: error: bad value (‘raptorlake’) for ‘-march=’ switch +cc1plus: note: valid arguments to ‘-march=’ switch are: nocona core2 nehalem corei7 westmere sandybridge... + 'silvermont': ['PCLMUL', 'RDRND'], + 'slm': ['PCLMUL', 'RDRND'], +```
devtools/dump-cpu-flags/cpu-names.txt new file +120 -0
120 lines not on this page · raw message
devtools/dump-cpu-flags/dump-cpu-flags.cpp new file +119 -0
119 lines not on this page · raw message
devtools/dump-cpu-flags/dump-cpu-flags.py new file +41 -0
41 lines not on this page · raw message
Signature · 1 line
2.48.1.vfs.0.0
Re: [PATCH 2/2] devtools/dump-cpu-flags: add tool to update CPU flags table
In reply to Bruce RichardsonQuoted 64 lines · Bruce Richardson
Quoted 47 lines · Andre Muezerie
This patchset allows users to specify the CPU for which the generated code should be optimized for by passing the CPU name.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool is introduced with this patch. The new tool generates the table entries for all CPUs listed in an input file using a recent version of the compiler, which has all the information needed. This reduces enormously the amount of work needed to update the table in msvc/meson.build and makes the process much less error prone.Signed-off-by Andre Muezerie <[email protected]>--- devtools/dump-cpu-flags/README.md | 25 +++++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 ++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++++++ 4 files changed, 305 insertions(+) create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.pydevtools/dump-cpu-flags/README.md new file +16
-0@@ -0,0 +1,25 @@ +# Generating updated CPU flags + +File `config\x86\msvc\meson.build` has a table with flags indicating instruction set support for a variety of CPU types. + +Script `dump-cpu-flags.py` can be used to generate updated entries for this table. + +The CPU names are stored in file `cpu-names.txt`, which is consumed by `dump-cpu-flags.py`. The formatting used in that file is described at the top of the file itself. + +The script relies on the information embedded in the g++ compiler. This means that an updated table can automatically be generated by switching to a newer version of the compiler. This avoids the need to manually edit the entries, which is error prone. With the script the table entries can just copied and pasted into `meson.build`. The only thing that might need to be done is adding new CPU names to cpu-names.txt, when new CPUs are released. + +**NOTE**: CPUs not known to the compiler will result in errors, which can be ignored (`dump-cpu-flags.py` will ignore these errors and continue). For best results use the latest g++ compiler available. + +Below is a sample output, where an error was logged because the compiler did not know about a CPU named ‘raptorlake’. + +```sh +$ ./dump-cpu-flags.pydevtools/dump-cpu-flags/README.md line 25
Quoted 10 lines
devtools/dump-cpu-flags/README.md new file +9
-0+ 'x86-64-v2': [], + 'x86-64-v3': ['AVX', 'AVX2'], + 'x86-64-v4': ['AVX', 'AVX2', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'alderlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], +cc1plus: error: bad value (‘raptorlake’) for ‘-march=’ switch +cc1plus: note: valid arguments to ‘-march=’ switch are: nocona core2 nehalem corei7 westmere sandybridge... + 'silvermont': ['PCLMUL', 'RDRND'], + 'slm': ['PCLMUL', 'RDRND'], +```How about having the tool output a valid meson.build file, that can then be used directly without copy-paste. While I know such a thing would end up with us having deep subdir structures, it could be just loaded from e.g. config/x86/msvc/cpu-flags/meson.build, for example.
Quoted 1 line
/Bruce
That's an interesting idea. It could be done, but when I think that this table will probably get updated no more than once per release I don't see much reason to to come up with something more sofisticated.
Sometimes it's better to keep things simple.
Re: [PATCH 1/2] config: allow AVX512 instructions to be used with MSVC
In reply to Bruce RichardsonQuoted 21 lines · Bruce Richardson
Quoted 11 lines · Andre Muezerie
Up to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170This patch allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name.
When no explicit CPU name is passed, 'native' is assumed (like it happens with GCC) and the code will be optimized for the same CPU type used to compile the code.Quoted 8 lines
MSVC does not provide this functionality natively, so logic was added to a new meson.build file under config/x86/msvc to handle these differences, detecting which
instruction sets are supported by the CPU(s), passing the best options to MSVC and setting the correct macros (like__AVX512F__) so that the DPDK code can rely on them like it is done with GCC.Signed-off-by Andre Muezerie <[email protected]>
Quoted 14 lines
Thanks for splitting out this change from the earlier one, it allows more focused review.
However, I think within this, we can split things a bit further two, and I think this patch could do with being split into separate changes. Specifically:
* one patch for reordering the x86/meson.build file, i.e. just move the code about and put asubdir_done()for MSVC once common part is completed. That patch is a quick review since all you are doing is moving things about.
* separate patch for adding the new msvc file + one-line change to add it to the x86/meson.build file.
* library changes in separate patch - or perhaps two patches so individual maintainers can ack.
I split them as suggested. In that process I found an old bug for which I sent a patch as well.
Quoted 6 lines
Some other comments inline below too.
Thanks,
/BruceQuoted 2 lines
--- config/x86/meson.build | 87 +++++------
<snip>
Quoted 14 lines
Quoted 14 lines
config/x86/meson.build +0
-14- -cpu_instruction_set = get_option('cpu_instruction_set') -if cpu_instruction_set == 'native' - foreach m:epyc_zen_cores.keys() - if cc.get_define(m, args: machine_args) != '' - dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) - break - endif - endforeach -else - foreach m:epyc_zen_cores.keys() - if m.contains(cpu_instruction_set) - dpdk_conf.set('RTE_MAX_LCORE', epyc_zen_cores[m]) - break
Quoted 8 lines
config/x86/meson.build foreach f:optional_flags · line 124
Quoted 5 lines
config/x86/meson.build +0
-5- endif - endforeach -endif - -dpdk_conf.set('RTE_MAX_NUMA_NODES', 32)For changes to config/x86/meson.build you can add my Ack if it's a separate patch.
Done, and added the ACK.
Quoted 7 lines
Quoted 7 lines
config/x86/msvc/meson.build new file +0
-0@@ -0,0 +1,287 @@
<snip>
Quoted 9 lines
config/x86/msvc/meson.build line 229
Quoted 5 lines
config/x86/msvc/meson.build new file +5
-0+ 'VPCLMULQDQ', + ] + foreach f:optional_flags + result = cc.run(cpuid_code, args: '-D@0@'.format(f), + name: 'instruction set @0@'.format(f))Is building a new binary for each instruction set the best way to do this? Would it not be better to have a single binary that outputs all the instruction sets in one go?
Building a new binary every time is probably the simplest approach. It also makes teh code easy to read and understand. It is fast, and since it will probably not be run more than once per DPDK release there's not much incentive to come up with something better.
Quoted 14 lines
config/x86/msvc/meson.build
Quoted 13 lines
config/x86/msvc/meson.build new file +6
-3+ has_instr_set = result.returncode() == 0 and result.stdout() == '1' + if has_instr_set + cpu_flags += f + endif + message('Target has @0@: @1@'.format(f, has_instr_set)) - c_args: cflags + - ['-mavx512f', '-mavx512vl', - '-mavx512cd', '-mavx512bw']) + c_args: cflags + cc_avx512_flags) objs += avx512_tmplib.extract_objects( 'acl_run_avx512.c') cflags += '-DCC_AVX512_SUPPORT'
Quoted 1 line
Ack from me for this change too.
I forgot to add the ACK.
Quoted 9 lines
Quoted 9 lines
lib/member/meson.build +0
-0@@ -33,6 +33,12 @@ if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok # compiler flags, and then have the .o file from static lib # linked into main lib.
Quoted 9 lines
lib/member/meson.build if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok · line 41
Quoted 6 lines
lib/member/meson.build +6
-0+ if is_ms_compiler + member_avx512_args = cc_avx512_flags + else + member_avx512_args = ['-mavx512f', '-mavx512dq', '-mavx512ifma'] + endif +is ifma included as msvc as part of the AVX512 flags? How does this work for the windows build?
Option /arch:AVX512 does not set __AVX512IFMA__ so, my educated guess is that msvc does not automatically generate instructions from that set.
There's also no msvc option specific for AVX512IFMA.
That being said, compiler intrinsics would still work as these do not require any options to be passed to msvc. This is other difference when compared to GCC.
Relying on cc_avx512_flags for msvc would still bring the benefit of using other AVX512 (non-IFMA) instructions.
Quoted 23 lines
Quoted 23 lines
lib/member/meson.build +2
-3# check if all required flags already enabled sketch_avx512_flags = ['__AVX512F__', '__AVX512DQ__', '__AVX512IFMA__']@@ -46,13 +52,12 @@ if dpdk_conf.has('RTE_ARCH_X86_64') and binutils_ok if sketch_avx512_on == true cflags += ['-DCC_AVX512_SUPPORT'] sources += files('rte_member_sketch_avx512.c') - elif cc.has_multi_arguments('-mavx512f', '-mavx512dq', '-mavx512ifma') + elif cc.has_multi_arguments(member_avx512_args) sketch_avx512_tmp = static_library('sketch_avx512_tmp', 'rte_member_sketch_avx512.c', include_directories: includes, dependencies: [static_rte_eal, static_rte_hash], - c_args: cflags + - ['-mavx512f', '-mavx512dq', '-mavx512ifma']) + c_args: cflags + member_avx512_args) objs += sketch_avx512_tmp.extract_objects('rte_member_sketch_avx512.c') cflags += ['-DCC_AVX512_SUPPORT'] endifSignature · 1 line
2.48.1.vfs.0.0
v3 cover allow AVX512 instructions to be used with MSVC
In reply to Andre MuezerieUp to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
This patchset allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name. When no name is provided 'native' is assumed meaning that the code should be optimized for the machine compiling the code.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool was also added.
v3:
- Added subdir('msvc') to config/x86/meson.build (now that other required patches are merged).
Andre Muezerie (2):
config: allow AVX512 instructions to be used with MSVC devtools/dump-cpu-flags: add tool to update CPU flags table
config/x86/meson.build | 1 + config/x86/msvc/meson.build | 287 +++++++++++++++++++++ devtools/dump-cpu-flags/README.md | 25 ++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 +++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++ 6 files changed, 593 insertions(+) create mode 100644 config/x86/msvc/meson.build create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.py
--
2.48.1.vfs.0.0
PATCH v3 1/2 config: allow AVX512 instructions to be used with MSVC
Up to now MSVC has being used with the default mode, which uses SSE2 instructions for scalar floating-point and vector calculations.
https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170
This patch allows users to specify the CPU for which the generated code should be optimized for in the same way it's done for GCC: by passing the CPU name.
When no explicit CPU name is passed, 'native' is assumed (like it happens with GCC) and the code will be optimized for the same CPU type used to compile the code.
MSVC does not provide this functionality natively, so logic was added to a new meson.build file under config/x86/msvc to handle these differences, detecting which
instruction sets are supported by the CPU(s), passing the best options to MSVC and setting the correct macros (like __AVX512F__) so that the DPDK code can rely on them like it is done with GCC.
--- config/x86/meson.build | 1 + config/x86/msvc/meson.build | 287 ++++++++++++++++++++++++++++++++++++ 2 files changed, 288 insertions(+) create mode 100644 config/x86/msvc/meson.build
config/x86/meson.build +1 -0
@@ -49,6 +49,7 @@ endif dpdk_conf.set('RTE_MAX_NUMA_NODES', 32) if is_ms_compiler + subdir('msvc') subdir_done() endif
config/x86/msvc/meson.build new file +287 -0
287 lines not on this page · raw message
Signature · 1 line
2.48.1.vfs.0.0
PATCH v3 2/2 devtools/dump-cpu-flags: add tool to update CPU flags table
In reply to Andre MuezerieThis patchset allows users to specify the CPU for which the generated code should be optimized for by passing the CPU name.
MSVC does not provide this functionality natively, so logic was added. This additional logic relies on a table which stores instruction set availability (like AXV512F) for different CPUs.
To make it easier to update this table a new devtool is introduced with this patch. The new tool generates the table entries for all CPUs listed in an input file using a recent version of the compiler, which has all the information needed. This reduces enormously the amount of work needed to update the table in msvc/meson.build and makes the process much less error prone.
--- devtools/dump-cpu-flags/README.md | 25 +++++ devtools/dump-cpu-flags/cpu-names.txt | 120 +++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.cpp | 119 ++++++++++++++++++++ devtools/dump-cpu-flags/dump-cpu-flags.py | 41 +++++++ 4 files changed, 305 insertions(+) create mode 100644 devtools/dump-cpu-flags/README.md create mode 100644 devtools/dump-cpu-flags/cpu-names.txt create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.cpp create mode 100644 devtools/dump-cpu-flags/dump-cpu-flags.py
devtools/dump-cpu-flags/README.md new file +25 -0
@@ -0,0 +1,25 @@ +# Generating updated CPU flags + +File `config\x86\msvc\meson.build` has a table with flags indicating instruction set support for a variety of CPU types. + +Script `dump-cpu-flags.py` can be used to generate updated entries for this table. + +The CPU names are stored in file `cpu-names.txt`, which is consumed by `dump-cpu-flags.py`. The formatting used in that file is described at the top of the file itself. + +The script relies on the information embedded in the g++ compiler. This means that an updated table can automatically be generated by switching to a newer version of the compiler. This avoids the need to manually edit the entries, which is error prone. With the script the table entries can just copied and pasted into `meson.build`. The only thing that might need to be done is adding new CPU names to cpu-names.txt, when new CPUs are released. + +**NOTE**: CPUs not known to the compiler will result in errors, which can be ignored (`dump-cpu-flags.py` will ignore these errors and continue). For best results use the latest g++ compiler available. + +Below is a sample output, where an error was logged because the compiler did not know about a CPU named ‘raptorlake’. + +```sh +$ ./dump-cpu-flags.py + 'x86-64-v2': [], + 'x86-64-v3': ['AVX', 'AVX2'], + 'x86-64-v4': ['AVX', 'AVX2', 'AVX512F', 'AVX512VL', 'AVX512BW', 'AVX512DQ', 'AVX512CD'], + 'alderlake': ['AVX', 'PCLMUL', 'RDRND', 'AVX2', 'RDSEED', 'AES', 'VPCLMULQDQ', 'GFNI'], +cc1plus: error: bad value (‘raptorlake’) for ‘-march=’ switch +cc1plus: note: valid arguments to ‘-march=’ switch are: nocona core2 nehalem corei7 westmere sandybridge... + 'silvermont': ['PCLMUL', 'RDRND'], + 'slm': ['PCLMUL', 'RDRND'], +```
devtools/dump-cpu-flags/cpu-names.txt new file +120 -0
120 lines not on this page · raw message
devtools/dump-cpu-flags/dump-cpu-flags.cpp new file +119 -0
119 lines not on this page · raw message
devtools/dump-cpu-flags/dump-cpu-flags.py new file +41 -0
41 lines not on this page · raw message
Signature · 1 line
2.48.1.vfs.0.0