event/dlb2: add support for single 512B write of 4 QEs
PATCH v2 1/1 event/dlb2: add support for single 512B write of 4 QEs
On Xeon, as 512b accesses are available, movdir64 instruction is able to perform 512b read and write to DLB producer port. In order for movdir64 to be able to pull its data from store buffers (store-buffer-forwarding) (before actual write), data should be in single 512b write format. This commit add change when code is built for Xeon with 512b AVX support to make single 512b write of all 4 QEs instead of 4x64b writes.
===
Changes since V1:
1) Split out dlb2_event_build_hcws into two implementations, one that uses AVX512 instructions, and one that does not. Each implementation is in its own source file in order to avoid build errors if the compiler does not support the newer AVX512 instructions.
2) Update meson.build to and pull in appropriate source file based on whether the compiler supports AVX512VL
3) Check if target supports AVX512VL, and use appropriate implementation based on this runtime check.
--- drivers/event/dlb2/dlb2.c | 206 +--------------------- drivers/event/dlb2/dlb2_avx512.c | 267 +++++++++++++++++++++++++++++ drivers/event/dlb2/dlb2_noavx512.c | 219 +++++++++++++++++++++++ drivers/event/dlb2/dlb2_priv.h | 8 + drivers/event/dlb2/meson.build | 14 ++ 5 files changed, 513 insertions(+), 201 deletions(-) create mode 100644 drivers/event/dlb2/dlb2_avx512.c create mode 100644 drivers/event/dlb2/dlb2_noavx512.c
drivers/event/dlb2/dlb2.c +5 -201
@@ -1834,6 +1834,11 @@ dlb2_eventdev_port_setup(struct rte_eventdev *dev, dev->data->ports[ev_port_id] = &dlb2->ev_ports[ev_port_id]; +if (rte_cpu_get_flag_enabled (RTE_CPUFLAG_AVX512VL )) + ev_port->qm_port.use_avx512 = true; +else + ev_port->qm_port.use_avx512 = false; +return 0 ; }
@@ -2430,21 +2435,6 @@ dlb2_eventdev_start(struct rte_eventdev *dev)return 0 ; } -static uint8_t cmd_byte_map[DLB2_NUM_PORT_TYPES ][DLB2_NUM_HW_SCHED_TYPES ] = { - { -/* Load-balanced cmd bytes */ - [RTE_EVENT_OP_NEW ] =DLB2_NEW_CMD_BYTE , - [RTE_EVENT_OP_FORWARD ] =DLB2_FWD_CMD_BYTE , - [RTE_EVENT_OP_RELEASE ] =DLB2_COMP_CMD_BYTE , - }, - { -/* Directed cmd bytes */ - [RTE_EVENT_OP_NEW ] =DLB2_NEW_CMD_BYTE , - [RTE_EVENT_OP_FORWARD ] =DLB2_NEW_CMD_BYTE , - [RTE_EVENT_OP_RELEASE ] =DLB2_NOOP_CMD_BYTE , - }, -}; -static inline uint32_t dlb2_port_credits_get (struct dlb2_port *qm_port,enum dlb2_hw_queue_types type)
@@ -2639,192 +2629,6 @@ dlb2_construct_token_pop_qe(struct dlb2_port *qm_port, int idx) qm_port->owed_tokens =0 ; } -static inline void -dlb2_event_build_hcws (struct dlb2_port *qm_port, -const struct rte_event ev[], -int num, -uint8_t *sched_type, -uint8_t *queue_id) -{ -struct dlb2_enqueue_qe *qe; -uint16_t sched_word[4 ]; -__m128i sse_qe[2 ]; -int i; - - qe = qm_port->qe4; - - sse_qe[0 ] =_mm_setzero_si128 (); - sse_qe[1 ] =_mm_setzero_si128 (); - -switch (num) { -case 4 : -/* Construct the metadata portion of two HCWs in one 128b SSE -* register. HCW metadata is constructed in the SSE registers -* like so: -* sse_qe[0][63:0]: qe[0]'s metadata -* sse_qe[0][127:64]: qe[1]'s metadata -* sse_qe[1][63:0]: qe[2]'s metadata -* sse_qe[1][127:64]: qe[3]'s metadata -*/ - -/* Convert the event operation into a command byte and store it -* in the metadata: -* sse_qe[0][63:56] = cmd_byte_map[is_directed][ev[0].op] -* sse_qe[0][127:120] = cmd_byte_map[is_directed][ev[1].op] -* sse_qe[1][63:56] = cmd_byte_map[is_directed][ev[2].op] -* sse_qe[1][127:120] = cmd_byte_map[is_directed][ev[3].op] -*/ -#define DLB2_QE_CMD_BYTE 7 - sse_qe[0 ] =_mm_insert_epi8 (sse_qe[0 ], - cmd_byte_map[qm_port->is_directed][ev[0 ].op], -DLB2_QE_CMD_BYTE ); - sse_qe[0 ] =_mm_insert_epi8 (sse_qe[0 ], - cmd_byte_map[qm_port->is_directed][ev[1 ].op], -DLB2_QE_CMD_BYTE +8 ); - sse_qe[1 ] =_mm_insert_epi8 (sse_qe[1 ], - cmd_byte_map[qm_port->is_directed][ev[2 ].op], -DLB2_QE_CMD_BYTE ); - sse_qe[1 ] =_mm_insert_epi8 (sse_qe[1 ], - cmd_byte_map[qm_port->is_directed][ev[3 ].op], -DLB2_QE_CMD_BYTE +8 ); - -/* Store priority, scheduling type, and queue ID in the sched -* word array because these values are re-used when the -* destination is a directed queue. -*/ - sched_word[0 ] =EV_TO_DLB2_PRIO (ev[0 ].priority) <<10 | - sched_type[0 ] <<8 | - queue_id[0 ]; - sched_word[1 ] =EV_TO_DLB2_PRIO (ev[1 ].priority) <<10 | - sched_type[1 ] <<8 | - queue_id[1 ]; - sched_word[2 ] =EV_TO_DLB2_PRIO (ev[2 ].priority) <<10 | - sched_type[2 ] <<8 | - queue_id[2 ]; - sched_word[3 ] =EV_TO_DLB2_PRIO (ev[3 ].priority) <<10 | - sched_type[3 ] <<8 | - queue_id[3 ]; - -/* Store the event priority, scheduling type, and queue ID in -* the metadata: -* sse_qe[0][31:16] = sched_word[0] -* sse_qe[0][95:80] = sched_word[1] -* sse_qe[1][31:16] = sched_word[2] -* sse_qe[1][95:80] = sched_word[3] -*/ -#define DLB2_QE_QID_SCHED_WORD 1 - sse_qe[0 ] =_mm_insert_epi16 (sse_qe[0 ], - sched_word[0 ], -DLB2_QE_QID_SCHED_WORD ); - sse_qe[0 ] =_mm_insert_epi16 (sse_qe[0 ], - sched_word[1 ], -DLB2_QE_QID_SCHED_WORD +4 ); - sse_qe[1 ] =_mm_insert_epi16 (sse_qe[1 ], - sched_word[2 ], -DLB2_QE_QID_SCHED_WORD ); - sse_qe[1 ] =_mm_insert_epi16 (sse_qe[1 ], - sched_word[3 ], -DLB2_QE_QID_SCHED_WORD +4 ); - -/* If the destination is a load-balanced queue, store the lock -* ID. If it is a directed queue, DLB places this field in -* bytes 10-11 of the received QE, so we format it accordingly: -* sse_qe[0][47:32] = dir queue ? sched_word[0] : flow_id[0] -* sse_qe[0][111:96] = dir queue ? sched_word[1] : flow_id[1] -* sse_qe[1][47:32] = dir queue ? sched_word[2] : flow_id[2] -* sse_qe[1][111:96] = dir queue ? sched_word[3] : flow_id[3] -*/ -#define DLB2_QE_LOCK_ID_WORD 2 - sse_qe[0 ] =_mm_insert_epi16 (sse_qe[0 ], - (sched_type[0 ] ==DLB2_SCHED_DIRECTED ) ? - sched_word[0 ] : ev[0 ].flow_id, -DLB2_QE_LOCK_ID_WORD ); - sse_qe[0 ] =_mm_insert_epi16 (sse_qe[0 ], - (sched_type[1 ] ==DLB2_SCHED_DIRECTED ) ? - sched_word[1 ] : ev[1 ].flow_id, -DLB2_QE_LOCK_ID_WORD +4 ); - sse_qe[1 ] =_mm_insert_epi16 (sse_qe[1 ], - (sched_type[2 ] ==DLB2_SCHED_DIRECTED ) ? - sched_word[2 ] : ev[2 ].flow_id, -DLB2_QE_LOCK_ID_WORD ); - sse_qe[1 ] =_mm_insert_epi16 (sse_qe[1 ], - (sched_type[3 ] ==DLB2_SCHED_DIRECTED ) ? - sched_word[3 ] : ev[3 ].flow_id, -DLB2_QE_LOCK_ID_WORD +4 ); - -/* Store the event type and sub event type in the metadata: -* sse_qe[0][15:0] = flow_id[0] -* sse_qe[0][79:64] = flow_id[1] -* sse_qe[1][15:0] = flow_id[2] -* sse_qe[1][79:64] = flow_id[3] -*/ -#define DLB2_QE_EV_TYPE_WORD 0 - sse_qe[0 ] =_mm_insert_epi16 (sse_qe[0 ], - ev[0 ].sub_event_type <<8 | - ev[0 ].event_type, -DLB2_QE_EV_TYPE_WORD ); - sse_qe[0 ] =_mm_insert_epi16 (sse_qe[0 ], - ev[1 ].sub_event_type <<8 | - ev[1 ].event_type, -DLB2_QE_EV_TYPE_WORD +4 ); - sse_qe[1 ] =_mm_insert_epi16 (sse_qe[1 ], - ev[2 ].sub_event_type <<8 | - ev[2 ].event_type, -DLB2_QE_EV_TYPE_WORD ); - sse_qe[1 ] =_mm_insert_epi16 (sse_qe[1 ], - ev[3 ].sub_event_type <<8 | - ev[3 ].event_type, -DLB2_QE_EV_TYPE_WORD +4 ); - -/* Store the metadata to memory (use the double-precision -* _mm_storeh_pd because there is no integer function for -* storing the upper 64b): -* qe[0] metadata = sse_qe[0][63:0] -* qe[1] metadata = sse_qe[0][127:64] -* qe[2] metadata = sse_qe[1][63:0] -* qe[3] metadata = sse_qe[1][127:64] -*/ -_mm_storel_epi64 ((__m128i *)&qe[0 ].u.opaque_data, sse_qe[0 ]); -_mm_storeh_pd ((double *)&qe[1 ].u.opaque_data, - (__m128d )sse_qe[0 ]); -_mm_storel_epi64 ((__m128i *)&qe[2 ].u.opaque_data, sse_qe[1 ]); -_mm_storeh_pd ((double *)&qe[3 ].u.opaque_data, - (__m128d )sse_qe[1 ]); - - qe[0 ].data = ev[0 ].u64; - qe[1 ].data = ev[1 ].u64; - qe[2 ].data = ev[2 ].u64; - qe[3 ].data = ev[3 ].u64; - -break ; -case 3 : -case 2 : -case 1 : -for (i =0 ; i < num; i++) { - qe[i].cmd_byte = - cmd_byte_map[qm_port->is_directed][ev[i].op]; - qe[i].sched_type = sched_type[i]; - qe[i].data = ev[i].u64; - qe[i].qid = queue_id[i]; - qe[i].priority =EV_TO_DLB2_PRIO (ev[i].priority); - qe[i].lock_id = ev[i].flow_id; -if (sched_type[i] ==DLB2_SCHED_DIRECTED ) { -struct dlb2_msg_info *info = - (struct dlb2_msg_info *)&qe[i].lock_id; - - info->qid = queue_id[i]; - info->sched_type =DLB2_SCHED_DIRECTED ; - info->priority = qe[i].priority; - } - qe[i].u.event_type.major = ev[i].event_type; - qe[i].u.event_type.sub = ev[i].sub_event_type; - } -break ; -case 0 : -break ; - } -} - static inline int dlb2_event_enqueue_prep(struct dlb2_eventdev_port *ev_port, struct dlb2_port *qm_port,
drivers/event/dlb2/dlb2_avx512.c new file +267 -0
267 lines not on this page · raw message
drivers/event/dlb2/dlb2_noavx512.c new file +219 -0
219 lines not on this page · raw message
drivers/event/dlb2/dlb2_priv.h +8 -0
@@ -377,6 +377,7 @@ struct dlb2_port {struct dlb2_eventdev_port *ev_port;/* back ptr */ bool use_scalar;/* force usage of scalar code */ uint16_t hw_credit_quanta; +bool use_avx512; };/* Per-process per-port mmio and memory pointers */
@@ -685,6 +686,13 @@ int dlb2_parse_params(const char *params,struct dlb2_devargs *dlb2_args, uint8_t version); +void dlb2_event_build_hcws (struct dlb2_port *qm_port, +const struct rte_event ev[], +int num, +uint8_t *sched_type, +uint8_t *queue_id); + +/* Extern globals */ extern struct process_local_port_data dlb2_port[][DLB2_NUM_PORT_TYPES ];
drivers/event/dlb2/meson.build +14 -0
@@ -19,6 +19,20 @@ sources = files( 'dlb2_selftest.c', ) +dlb2_avx512_support = false + +if dpdk_conf.has('RTE_ARCH_X86_64') + dlb2_avx512_support = ( + cc.get_define('__AVX512VL__', args: machine_args) != '' + ) +endif + +if dlb2_avx512_support == true + sources += files('dlb2_avx512.c') +else + sources += files('dlb2_noavx512.c') +endif + headers = files('rte_pmd_dlb2.h') deps += ['mbuf', 'mempool', 'ring', 'pci', 'bus_pci']
Signature · 1 line
2.25.1
Quoted 21 lines · Timothy McDaniel
On Xeon, as 512b accesses are available, movdir64 instruction is able to perform 512b read and write to DLB producer port. In order for movdir64 to be able to pull its data from store buffers (store-buffer-forwarding) (before actual write), data should be in single 512b write format. This commit add change when code is built for Xeon with 512b AVX support to make single 512b write of all 4 QEs instead of 4x64b writes.
Signed-off-by Timothy McDaniel <[email protected]>Acked-by Kent Wires <[email protected]>===
Changes since V1:
1) Split outdlb2_event_build_hcwsinto two implementations, one that uses AVX512 instructions, and one that does not. Each implementation is in its own source file in order to avoid build errors if the compiler does not support the newer AVX512 instructions.
2) Update meson.build to and pull in appropriate source file based on whether the compiler supports AVX512VL
3) Check if target supports AVX512VL, and use appropriate implementation based on this runtime check.
Quoted 4 lines
--- drivers/event/dlb2/dlb2.c | 206 +--------------------- drivers/event/dlb2/dlb2_avx512.c | 267 +++++++++++++++++++++++++++++ drivers/event/dlb2/dlb2_noavx512.c | 219 +++++++++++++++++++++++
Could you change the file name to dlb2_sve.c as noavx512 means it can be NEON too.
Rest looks good to me. Will merge the next version.