simplify building x86 code with AVX2 support
Subsystems (2)
cover simplify building x86 code with AVX2 support
Inside our optimized vector drivers (and libs), there were always build time checks for various levels of instruction set support, most notably AVX2 and AVX-512 on x86 systems. One of the checks done in each case was verifying that the compiler used was able to generate AVX code appropriately. However, since this was originally done, our minimum compiler support has been updated and so we no longer need to check this - all supported compilers can generate AVX2. This allows us to simplify the build logic for the x86 vector code.
In future, we can do similarly for AVX-512.
Bruce Richardson (2):
build/x86: remove conditional checks for AVX2 support
build: remove unnecessary AVX2 compiler flag
drivers/net/bnxt/bnxt_ethdev.c | 8 ++++---- drivers/net/bnxt/bnxt_rxr.h | 2 +- drivers/net/bnxt/bnxt_txr.h | 2 +- drivers/net/bnxt/meson.build | 26 +++++++++---------------- drivers/net/enic/meson.build | 10 +++------- drivers/net/i40e/i40e_rxtx.c | 14 ++++--------- drivers/net/i40e/meson.build | 22 ++++++--------------- drivers/net/iavf/iavf_rxtx_vec_common.h | 2 +- drivers/net/iavf/meson.build | 22 ++++++--------------- drivers/net/ice/meson.build | 21 ++++++-------------- lib/acl/meson.build | 23 +++++----------------- lib/acl/rte_acl.c | 10 +++------- 12 files changed, 49 insertions(+), 113 deletions(-)
--
2.39.2
PATCH 1/2 build/x86: remove conditional checks for AVX2 support
In a number of libraries and drivers we have multiple levels of checks for enabling AVX2 support. In these cases:
* we first check for AVX2 support in the build-time ISA, i.e. through the instruction-set/-march flag. If present, we add source file to the list of sources.
* if not enabled at the minimum instruction-set level, i.e. a default/generic build, we then check for compiler support for AVX2 and, if available, did a separate build of the AVX2 file using an additional flag to enable the instruction set.
While this works, and was necessary in older releases, we no longer need this level of complexity, as all supported DPDK compilers have support for AVX2. This makes the second check unnecessary.
However, when we look at the effect of the second option above vs the first, the only real difference is that in the second case, we add an additional "-mavx2" to the build flags. This flag simply makes the instruction set available for use, so should be harmless in the case where the "march" flag already has that instruction set available. Therefore, we can remove the check for the first case also, and always use the fallback case with the extra flag.
--- drivers/net/bnxt/meson.build | 27 ++++++++++----------------- drivers/net/enic/meson.build | 10 +++------- drivers/net/i40e/meson.build | 23 +++++++---------------- drivers/net/iavf/meson.build | 23 +++++++---------------- drivers/net/ice/meson.build | 22 +++++++--------------- lib/acl/meson.build | 24 ++++++------------------ 6 files changed, 40 insertions(+), 89 deletions(-)
drivers/net/bnxt/meson.build +10 -17
@@ -53,23 +53,16 @@ subdir('hcapi/cfa') if arch_subdir == 'x86' sources += files('bnxt_rxtx_vec_sse.c') - # compile AVX2 version if either: - # a. we have AVX supported in minimum instruction set baseline - # b. it's not minimum instruction set, but supported by compiler - if cc.get_define('__AVX2__', args: machine_args) != '' - cflags += ['-DCC_AVX2_SUPPORT'] - sources += files('bnxt_rxtx_vec_avx2.c') - elif cc.has_argument('-mavx2') - cflags += ['-DCC_AVX2_SUPPORT'] - bnxt_avx2_lib = static_library('bnxt_avx2_lib', - 'bnxt_rxtx_vec_avx2.c', - dependencies: [static_rte_ethdev, - static_rte_bus_pci, - static_rte_kvargs, static_rte_hash], - include_directories: includes, - c_args: [cflags, '-mavx2']) - objs += bnxt_avx2_lib.extract_objects('bnxt_rxtx_vec_avx2.c') - endif + cflags += ['-DCC_AVX2_SUPPORT'] + # build AVX2 code with instruction set explicitly enabled for runtime selection + bnxt_avx2_lib = static_library('bnxt_avx2_lib', + 'bnxt_rxtx_vec_avx2.c', + dependencies: [static_rte_ethdev, + static_rte_bus_pci, + static_rte_kvargs, static_rte_hash], + include_directories: includes, + c_args: [cflags, '-mavx2']) + objs += bnxt_avx2_lib.extract_objects('bnxt_rxtx_vec_avx2.c') elif arch_subdir == 'arm' and dpdk_conf.get('RTE_ARCH_64') sources += files('bnxt_rxtx_vec_neon.c') endif
drivers/net/enic/meson.build +3 -7
@@ -28,14 +28,10 @@ sources = files( deps += ['hash'] includes += include_directories('base') -# The current implementation assumes 64-bit pointers -if cc.get_define('__AVX2__', args: machine_args) != '' and dpdk_conf.get('RTE_ARCH_64') - sources += files('enic_rxtx_vec_avx2.c') -# Build the avx2 handler if the compiler supports it, even though 'machine' -# does not. This is to support users who build for the min supported machine +# Build the avx2 handler for 64-bit X86 targets, even though 'machine' +# may not. This is to support users who build for the min supported machine # and need to run the binary on newer CPUs too. -# This part is from i40e meson.build -elif cc.has_argument('-mavx2') and dpdk_conf.get('RTE_ARCH_64') +if dpdk_conf.has('RTE_ARCH_X86_64') enic_avx2_lib = static_library('enic_avx2_lib', 'enic_rxtx_vec_avx2.c', dependencies: [static_rte_ethdev, static_rte_bus_pci],
drivers/net/i40e/meson.build +7 -16
@@ -49,22 +49,13 @@ if arch_subdir == 'x86' cflags += ['-fno-asynchronous-unwind-tables'] endif - # compile AVX2 version if either: - # a. we have AVX supported in minimum instruction set baseline - # b. it's not minimum instruction set, but supported by compiler - if cc.get_define('__AVX2__', args: machine_args) != '' - cflags += ['-DCC_AVX2_SUPPORT'] - sources += files('i40e_rxtx_vec_avx2.c') - elif cc.has_argument('-mavx2') - cflags += ['-DCC_AVX2_SUPPORT'] - i40e_avx2_lib = static_library('i40e_avx2_lib', - 'i40e_rxtx_vec_avx2.c', - dependencies: [static_rte_ethdev, - static_rte_kvargs, static_rte_hash], - include_directories: includes, - c_args: [cflags, '-mavx2']) - objs += i40e_avx2_lib.extract_objects('i40e_rxtx_vec_avx2.c') - endif + cflags += ['-DCC_AVX2_SUPPORT'] + i40e_avx2_lib = static_library('i40e_avx2_lib', + 'i40e_rxtx_vec_avx2.c', + dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash], + include_directories: includes, + c_args: [cflags, '-mavx2']) + objs += i40e_avx2_lib.extract_objects('i40e_rxtx_vec_avx2.c') i40e_avx512_cpu_support = ( cc.get_define('__AVX512F__', args: machine_args) != '' and
drivers/net/iavf/meson.build +7 -16
@@ -29,22 +29,13 @@ if arch_subdir == 'x86' cflags += ['-fno-asynchronous-unwind-tables'] endif - # compile AVX2 version if either: - # a. we have AVX supported in minimum instruction set baseline - # b. it's not minimum instruction set, but supported by compiler - if cc.get_define('__AVX2__', args: machine_args) != '' - cflags += ['-DCC_AVX2_SUPPORT'] - sources += files('iavf_rxtx_vec_avx2.c') - elif cc.has_argument('-mavx2') - cflags += ['-DCC_AVX2_SUPPORT'] - iavf_avx2_lib = static_library('iavf_avx2_lib', - 'iavf_rxtx_vec_avx2.c', - dependencies: [static_rte_ethdev, - static_rte_kvargs, static_rte_hash], - include_directories: includes, - c_args: [cflags, '-mavx2']) - objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c') - endif + cflags += ['-DCC_AVX2_SUPPORT'] + iavf_avx2_lib = static_library('iavf_avx2_lib', + 'iavf_rxtx_vec_avx2.c', + dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash], + include_directories: includes, + c_args: [cflags, '-mavx2']) + objs += iavf_avx2_lib.extract_objects('iavf_rxtx_vec_avx2.c') iavf_avx512_cpu_support = ( cc.get_define('__AVX512F__', args: machine_args) != '' and
drivers/net/ice/meson.build +7 -15
@@ -28,21 +28,13 @@ if arch_subdir == 'x86' cflags += ['-fno-asynchronous-unwind-tables'] endif - # compile AVX2 version if either: - # a. we have AVX supported in minimum instruction set baseline - # b. it's not minimum instruction set, but supported by compiler - if cc.get_define('__AVX2__', args: machine_args) != '' - cflags += ['-DCC_AVX2_SUPPORT'] - sources += files('ice_rxtx_vec_avx2.c') - elif cc.has_argument('-mavx2') - cflags += ['-DCC_AVX2_SUPPORT'] - ice_avx2_lib = static_library('ice_avx2_lib', - 'ice_rxtx_vec_avx2.c', - dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash], - include_directories: includes, - c_args: [cflags, '-mavx2']) - objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c') - endif + cflags += ['-DCC_AVX2_SUPPORT'] + ice_avx2_lib = static_library('ice_avx2_lib', + 'ice_rxtx_vec_avx2.c', + dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash], + include_directories: includes, + c_args: [cflags, '-mavx2']) + objs += ice_avx2_lib.extract_objects('ice_rxtx_vec_avx2.c') ice_avx512_cpu_support = ( cc.get_define('__AVX512F__', args: machine_args) != '' and
lib/acl/meson.build +6 -18
@@ -14,24 +14,12 @@ headers = files('rte_acl.h', 'rte_acl_osdep.h') if dpdk_conf.has('RTE_ARCH_X86') sources += files('acl_run_sse.c') - # compile AVX2 version if either: - # a. we have AVX supported in minimum instruction set baseline - # b. it's not minimum instruction set, but supported by compiler - # - # in former case, just add avx2 C file to files list - # in latter case, compile c file to static lib, using correct compiler - # flags, and then have the .o file from static lib linked into main lib. - if cc.get_define('__AVX2__', args: machine_args) != '' - sources += files('acl_run_avx2.c') - cflags += '-DCC_AVX2_SUPPORT' - elif cc.has_argument('-mavx2') - avx2_tmplib = static_library('avx2_tmp', - 'acl_run_avx2.c', - dependencies: static_rte_eal, - c_args: cflags + ['-mavx2']) - objs += avx2_tmplib.extract_objects('acl_run_avx2.c') - cflags += '-DCC_AVX2_SUPPORT' - endif + cflags += '-DCC_AVX2_SUPPORT' + avx2_tmplib = static_library('avx2_tmp', + 'acl_run_avx2.c', + dependencies: static_rte_eal, + c_args: cflags + ['-mavx2']) + objs += avx2_tmplib.extract_objects('acl_run_avx2.c') # compile AVX512 version if: # we are building 64-bit binary AND binutils can generate proper code
Signature · 1 line
2.39.2
PATCH 2/2 build: remove unnecessary AVX2 compiler flag
In reply to Bruce RichardsonSince AVX2 is always available at build time, the CC_AVX2_SUPPORT cflag became equivalent to the RTE_ARCH_X86 one. Therefore, we can just remove all use of the flag, replacing it will the RTE_ARCH_X86 one, and also simplifying the code in a few places where we can merge with other X86/non-X86 blocks e.g. for providing SSE fallbacks.
--- drivers/net/bnxt/bnxt_ethdev.c | 8 ++++---- drivers/net/bnxt/bnxt_rxr.h | 2 +- drivers/net/bnxt/bnxt_txr.h | 2 +- drivers/net/bnxt/meson.build | 1 - drivers/net/i40e/i40e_rxtx.c | 14 ++++---------- drivers/net/i40e/meson.build | 1 - drivers/net/iavf/iavf_rxtx_vec_common.h | 2 +- drivers/net/iavf/meson.build | 1 - drivers/net/ice/meson.build | 1 - lib/acl/meson.build | 1 - lib/acl/rte_acl.c | 10 +++------- 11 files changed, 14 insertions(+), 29 deletions(-)
drivers/net/bnxt/bnxt_ethdev.c +4 -4
@@ -1280,7 +1280,7 @@ bnxt_receive_function(struct rte_eth_dev *eth_dev)if (bp->ieee_1588) goto use_scalar_rx; -#if defined(RTE_ARCH_X86 ) && defined(CC_AVX2_SUPPORT ) +#if defined(RTE_ARCH_X86 )if (rte_vect_get_max_simd_bitwidth () >=RTE_VECT_SIMD_256 &&rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX2 ) ==1 ) { PMD_DRV_LOG(INFO,
@@ -1332,7 +1332,7 @@ bnxt_transmit_function(struct rte_eth_dev *eth_dev) BNXT_TRUFLOW_EN(bp) || bp->ieee_1588) goto use_scalar_tx; -#if defined(RTE_ARCH_X86 ) && defined(CC_AVX2_SUPPORT ) +#if defined(RTE_ARCH_X86 )if (rte_vect_get_max_simd_bitwidth () >=RTE_VECT_SIMD_256 &&rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX2 ) ==1 ) { PMD_DRV_LOG(INFO,
@@ -3019,7 +3019,7 @@ static const struct { #if defined(RTE_ARCH_X86) {bnxt_recv_pkts_vec, "Vector SSE"}, #endif -#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT) +#if defined(RTE_ARCH_X86) {bnxt_recv_pkts_vec_avx2, "Vector AVX2"}, #endif #if defined(RTE_ARCH_ARM64)
@@ -3053,7 +3053,7 @@ static const struct { #if defined(RTE_ARCH_X86) {bnxt_xmit_pkts_vec, "Vector SSE"}, #endif -#if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT) +#if defined(RTE_ARCH_X86) {bnxt_xmit_pkts_vec_avx2, "Vector AVX2"}, #endif #if defined(RTE_ARCH_ARM64)
drivers/net/bnxt/bnxt_rxr.h +1 -1
@@ -112,7 +112,7 @@ uint16_t bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,int bnxt_rxq_vec_setup (struct bnxt_rx_queue *rxq);#endif -#if defined(RTE_ARCH_X86 ) && defined(CC_AVX2_SUPPORT ) +#if defined(RTE_ARCH_X86 )uint16_t bnxt_recv_pkts_vec_avx2 (void *rx_queue,struct rte_mbuf **rx_pkts,uint16_t nb_pkts);#endif
drivers/net/bnxt/bnxt_txr.h +1 -1
@@ -52,7 +52,7 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,uint16_t bnxt_xmit_pkts_vec (void *tx_queue,struct rte_mbuf **tx_pkts,uint16_t nb_pkts);#endif -#if defined(RTE_ARCH_X86 ) && defined(CC_AVX2_SUPPORT ) +#if defined(RTE_ARCH_X86 )uint16_t bnxt_xmit_pkts_vec_avx2 (void *tx_queue,struct rte_mbuf **tx_pkts,uint16_t nb_pkts);#endif
drivers/net/bnxt/meson.build +0 -1
@@ -53,7 +53,6 @@ subdir('hcapi/cfa') if arch_subdir == 'x86' sources += files('bnxt_rxtx_vec_sse.c') - cflags += ['-DCC_AVX2_SUPPORT'] # build AVX2 code with instruction set explicitly enabled for runtime selection bnxt_avx2_lib = static_library('bnxt_avx2_lib', 'bnxt_rxtx_vec_avx2.c',
drivers/net/i40e/i40e_rxtx.c +4 -10
@@ -3216,15 +3216,9 @@ get_avx_supported(bool request_avx512)#endif }else {if (rte_vect_get_max_simd_bitwidth () >=RTE_VECT_SIMD_256 && -rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX2 ) ==1 && - rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) == 1) -#ifdef CC_AVX2_SUPPORT +rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX2 ) ==1 && +rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX512F ) ==1 )return true; -#else -PMD_DRV_LOG (NOTICE , -"AVX2 is not supported in build env" ); -return false; -#endif }return false;
@@ -3608,7 +3602,7 @@ i40e_set_default_pctype_table(struct rte_eth_dev *dev) } } -#ifndef CC_AVX2_SUPPORT +#ifndef RTE_ARCH_X86 uint16_t i40e_recv_pkts_vec_avx2(void __rte_unused *rx_queue, struct rte_mbuf __rte_unused **rx_pkts,
@@ -3632,4 +3626,4 @@ i40e_xmit_pkts_vec_avx2(void __rte_unused * tx_queue, {return 0 ; } -#endif /* ifndef CC_AVX2_SUPPORT */ +#endif /* ifndef RTE_ARCH_X86 */
drivers/net/i40e/meson.build +0 -1
@@ -49,7 +49,6 @@ if arch_subdir == 'x86' cflags += ['-fno-asynchronous-unwind-tables'] endif - cflags += ['-DCC_AVX2_SUPPORT'] i40e_avx2_lib = static_library('i40e_avx2_lib', 'i40e_rxtx_vec_avx2.c', dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash],
drivers/net/iavf/iavf_rxtx_vec_common.h +1 -1
@@ -396,7 +396,7 @@ iavf_txd_enable_offload(__rte_unused struct rte_mbuf *tx_pkt, *txd_hi |= ((uint64_t )td_cmd) <<IAVF_TXD_QW1_CMD_SHIFT ; } -#ifdef CC_AVX2_SUPPORT +#ifdef RTE_ARCH_X86 static __rte_always_inline void iavf_rxq_rearm_common(struct iavf_rx_queue *rxq, __rte_unused bool avx512) {
drivers/net/iavf/meson.build +0 -1
@@ -29,7 +29,6 @@ if arch_subdir == 'x86' cflags += ['-fno-asynchronous-unwind-tables'] endif - cflags += ['-DCC_AVX2_SUPPORT'] iavf_avx2_lib = static_library('iavf_avx2_lib', 'iavf_rxtx_vec_avx2.c', dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash],
drivers/net/ice/meson.build +0 -1
@@ -28,7 +28,6 @@ if arch_subdir == 'x86' cflags += ['-fno-asynchronous-unwind-tables'] endif - cflags += ['-DCC_AVX2_SUPPORT'] ice_avx2_lib = static_library('ice_avx2_lib', 'ice_rxtx_vec_avx2.c', dependencies: [static_rte_ethdev, static_rte_kvargs, static_rte_hash],
lib/acl/meson.build +0 -1
@@ -14,7 +14,6 @@ headers = files('rte_acl.h', 'rte_acl_osdep.h') if dpdk_conf.has('RTE_ARCH_X86') sources += files('acl_run_sse.c') - cflags += '-DCC_AVX2_SUPPORT' avx2_tmplib = static_library('avx2_tmp', 'acl_run_avx2.c', dependencies: static_rte_eal,
lib/acl/rte_acl.c +3 -7
@@ -42,10 +42,9 @@ rte_acl_classify_avx512x32(__rte_unused const struct rte_acl_ctx *ctx, }#endif -#ifndef CC_AVX2_SUPPORT +#ifndef RTE_ARCH_X86 /* -* If the compiler doesn't support AVX2 instructions, -* then the dummy one would be used instead for AVX2 classify method. +* If ISA doesn't have AVX2 or SSE, provide dummy fallbacks */ int rte_acl_classify_avx2(__rte_unused const struct rte_acl_ctx *ctx,
@@ -56,9 +55,6 @@ rte_acl_classify_avx2(__rte_unused const struct rte_acl_ctx *ctx, {return -ENOTSUP ; } -#endif - -#ifndef RTE_ARCH_X86 int rte_acl_classify_sse(__rte_unused const struct rte_acl_ctx *ctx, __rte_unused const uint8_t **data,
@@ -182,7 +178,7 @@ acl_check_alg_x86(enum rte_acl_classify_alg alg) }if (alg ==RTE_ACL_CLASSIFY_AVX2 ) { -#ifdef CC_AVX2_SUPPORT +#ifdef RTE_ARCH_X86 if (rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX2 ) &&rte_vect_get_max_simd_bitwidth () >=RTE_VECT_SIMD_256 )return 0 ;
Signature · 1 line
2.39.2
Hello Bruce,
Quoted 16 lines · Bruce Richardson
Inside our optimized vector drivers (and libs), there were always build time checks for various levels of instruction set support, most notably AVX2 and AVX-512 on x86 systems. One of the checks done in each case was verifying that the compiler used was able to generate AVX code appropriately. However, since this was originally done, our minimum compiler support has been updated and so we no longer need to check this - all supported compilers can generate AVX2. This allows us to simplify the build logic for the x86 vector code.
In future, we can do similarly for AVX-512.
Bruce Richardson (2):
build/x86: remove conditional checks for AVX2 support
build: remove unnecessary AVX2 compiler flag
Quoted 13 lines
drivers/net/bnxt/bnxt_ethdev.c | 8 ++++---- drivers/net/bnxt/bnxt_rxr.h | 2 +- drivers/net/bnxt/bnxt_txr.h | 2 +- drivers/net/bnxt/meson.build | 26 +++++++++---------------- drivers/net/enic/meson.build | 10 +++------- drivers/net/i40e/i40e_rxtx.c | 14 ++++--------- drivers/net/i40e/meson.build | 22 ++++++--------------- drivers/net/iavf/iavf_rxtx_vec_common.h | 2 +- drivers/net/iavf/meson.build | 22 ++++++--------------- drivers/net/ice/meson.build | 21 ++++++-------------- lib/acl/meson.build | 23 +++++----------------- lib/acl/rte_acl.c | 10 +++------- 12 files changed, 49 insertions(+), 113 deletions(-)
Sorry, this series had fallen through the cracks... though I intended to apply with other build update series.
It still applies fine, I ran my checks and I see nothing wrong with the series. Applied, thanks for the cleanup.
Signature · 1 line
David Marchand