DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [PATCH] event/dlb2: Optimize Dequeue Operations
@ 2021-03-17 17:02 Timothy McDaniel
  2021-03-21 11:10 ` Jerin Jacob
                   ` (2 more replies)
  0 siblings, 3 replies; 9+ messages in thread
From: Timothy McDaniel @ 2021-03-17 17:02 UTC (permalink / raw)
  To: Bruce Richardson
  Cc: dev, jerinj, harry.van.haaren, mdr, nhorman, nikhil.rao,
	erik.g.carrillo, abhinandan.gujjar, pbhagavatula, hemant.agrawal,
	mattias.ronnblom, peter.mccarthy

Convert code to use x86 vector instructions, thereby significantly
improving dequeue performance.

Signed-off-by: Timothy McDaniel <timothy.mcdaniel@intel.com>
---
 config/rte_config.h            |    1 +
 drivers/event/dlb2/dlb2.c      |  607 ++++++++++++++++++++++++++++++++++++----
 drivers/event/dlb2/dlb2_priv.h |   19 +-
 3 files changed, 574 insertions(+), 53 deletions(-)

diff --git a/config/rte_config.h b/config/rte_config.h
index aedb68c..133ca35 100644
--- a/config/rte_config.h
+++ b/config/rte_config.h
@@ -144,5 +144,6 @@
 #undef RTE_LIBRTE_PMD_DLB2_QUELL_STATS
 #define RTE_LIBRTE_PMD_DLB2_SW_CREDIT_QUANTA 32
 #define RTE_PMD_DLB2_DEFAULT_DEPTH_THRESH 256
+#define RTE_LIBRTE_PMD_DLB2_VECTOR_CODE 1
 
 #endif /* _RTE_CONFIG_H_ */
diff --git a/drivers/event/dlb2/dlb2.c b/drivers/event/dlb2/dlb2.c
index a4a7db4..adcd3dd 100644
--- a/drivers/event/dlb2/dlb2.c
+++ b/drivers/event/dlb2/dlb2.c
@@ -1186,6 +1186,37 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 					 const struct rte_event events[],
 					 uint16_t num);
 
+/* Generate the required bitmask for rotate-style expected QE gen bits.
+ * This requires a pattern of 1's and zeros, starting with expected as
+ * 1 bits, so when hardware writes 0's they're "new". This requires the
+ * ring size to be powers of 2 to wrap correctly.
+ */
+static void
+dlb2_hw_cq_bitmask_init(struct dlb2_port *qm_port, uint32_t cq_depth)
+{
+	uint64_t cq_build_mask = 0;
+	uint32_t i;
+
+	if (cq_depth > 64)
+		return; /* need to fall back to scalar code */
+
+	/*
+	 * all 1's in first u64, all zeros in 2nd is correct bit pattern to
+	 * start. Special casing == 64 easier than adapting complex loop logic.
+	 */
+	if (cq_depth == 64) {
+		qm_port->cq_rolling_mask = 0;
+		qm_port->cq_rolling_mask_2 = -1;
+		return;
+	}
+
+	for (i = 0; i < 64; i += (cq_depth * 2))
+		cq_build_mask |= ((1ULL << cq_depth) - 1) << (i + cq_depth);
+
+	qm_port->cq_rolling_mask = cq_build_mask;
+	qm_port->cq_rolling_mask_2 = cq_build_mask;
+}
+
 static int
 dlb2_hw_create_ldb_port(struct dlb2_eventdev *dlb2,
 			struct dlb2_eventdev_port *ev_port,
@@ -1303,6 +1334,8 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 	/* starting value of gen bit - it toggles at wrap time */
 	qm_port->gen_bit = 1;
 
+	dlb2_hw_cq_bitmask_init(qm_port, qm_port->cq_depth);
+
 	qm_port->int_armed = false;
 
 	/* Save off for later use in info and lookup APIs. */
@@ -1354,6 +1387,17 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 			     dequeue_depth,
 			     qm_port->credits);
 	}
+
+	qm_port->use_scalar = false;
+
+#if (!defined RTE_ARCH_X86_64) || (!defined RTE_LIBRTE_PMD_DLB2_VECTOR_CODE)
+	qm_port->use_scalar = true;
+#else
+	if ((qm_port->cq_depth > 64) ||
+	    (!rte_is_power_of_2(qm_port->cq_depth)))
+		qm_port->use_scalar = true;
+#endif
+
 	rte_spinlock_unlock(&handle->resource_lock);
 
 	return 0;
@@ -1499,6 +1543,7 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 	qm_port->gen_bit_shift = __builtin_popcount(qm_port->cq_depth_mask);
 	/* starting value of gen bit - it toggles at wrap time */
 	qm_port->gen_bit = 1;
+	dlb2_hw_cq_bitmask_init(qm_port, qm_port->cq_depth);
 
 	qm_port->int_armed = false;
 
@@ -1539,6 +1584,15 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 			     dequeue_depth,
 			     credit_high_watermark);
 	}
+
+#if (!defined RTE_ARCH_X86_64) || (!defined RTE_LIBRTE_PMD_DLB2_VECTOR_CODE)
+	qm_port->use_scalar = true;
+#else
+	if ((qm_port->cq_depth > 64) ||
+	    (!rte_is_power_of_2(qm_port->cq_depth)))
+		qm_port->use_scalar = true;
+#endif
+
 	rte_spinlock_unlock(&handle->resource_lock);
 
 	return 0;
@@ -2424,6 +2478,203 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 }
 
 static inline void
+dlb2_event_build_hcws_vec(struct dlb2_port *qm_port,
+			  const struct rte_event ev[],
+			  int num,
+			  uint8_t *sched_type,
+			  uint8_t *queue_id)
+{
+	const __m128i *v_evs = (const __m128i *)ev;
+
+	/* Useful for mask creation in all sections */
+	const __m128i v_zeros = _mm_setzero_si128();
+	const __m128i v_ones = _mm_cmpeq_epi8(v_zeros, v_zeros);
+
+	/* LSB lanes are pkt 0, then pkt 1, pkt 2, MSB lanes are pkt 3 */
+	__m128i v_unpack_02 = _mm_unpacklo_epi32(v_evs[0], v_evs[2]);
+	__m128i v_unpack_13 = _mm_unpacklo_epi32(v_evs[1], v_evs[3]);
+	__m128i v_unpack_b32_63 = _mm_unpackhi_epi32(v_unpack_02, v_unpack_13);
+	__m128i v_unpack_b00_31 = _mm_unpacklo_epi32(v_unpack_02, v_unpack_13);
+
+	/* HCW contents for 4x QEs built up in here at u32 per QE */
+	__m128i v_qe_hcws;
+
+	/* Event OP processing to cmd_byte:
+	 * qe[i].cmd_byte = cmd_byte_map[qm_port->is_directed][ev[i].op];
+	 */
+	{
+		__m128i v_op_bits = _mm_srli_epi32(v_ones, 30);
+		__m128i v_ev_op = _mm_and_si128(v_op_bits, v_unpack_b32_63);
+
+		/*
+		 * Is directed -> shuffle index:
+		 * - OP field is 2 bits, so indexes 0,1,2,3 are used by LB.
+		 * - DIR results stored in 4,5,6,7, so bit OR in a bit to
+		 *   select the higher shuffle-mask range if the enqueues are
+		 *   for DIR ports.
+		 */
+		uint32_t port_is_directed = qm_port->is_directed;
+		__m128i v_is_dir = _mm_insert_epi32(v_zeros,
+						    0xffffff00 |
+						    port_is_directed << 2, 0);
+		__m128i v_dir_inject = _mm_shuffle_epi32(v_is_dir, 0);
+		__m128i v_ev_op_w_dir = _mm_or_si128(v_dir_inject, v_ev_op);
+
+		/* Load shuffle mask, perform shuffle for CMD byte */
+		__m128i v_op_shuffle = _mm_loadu_si128((__m128i *)cmd_byte_map);
+		__m128i v_qe_op_preshift = _mm_shuffle_epi8(v_op_shuffle,
+							    v_ev_op_w_dir);
+
+		/* move to MSB u8 lane where cmp_byte is located */
+		__m128i v_cmd_byte = _mm_slli_epi32(v_qe_op_preshift, 24);
+
+
+		/* Use burst size as provided in "num" to ZERO cmd_bytes for
+		 * the relevant QE. Note that if num == 4, then no zeroing need
+		 * take place, and num > 4 is invalid.
+		 * We take advantage of the num < 4 fact in the below scalar
+		 * mask generation code.
+		 */
+		if (num < 4) {
+			uint64_t burst_mask = (1ULL << (num * 16)) - 1;
+			__m128i v_mask_u64 = _mm_insert_epi64(v_zeros,
+							      burst_mask, 0);
+			__m128i v_burst_mask = _mm_unpacklo_epi8(v_mask_u64,
+								 v_mask_u64);
+			v_cmd_byte = _mm_blendv_epi8(v_zeros, v_cmd_byte,
+						     v_burst_mask);
+		}
+		v_qe_hcws = v_cmd_byte;
+	}
+
+	/* Priority:
+	 * Generate 3-bit wide mask in correct place, shift prio value itself
+	 * between lanes with a u32 shift, then AND then OR into HCW result.
+	 */
+	{
+		__m128i v_prio_mask = _mm_srli_epi32(v_ones, 29);
+		v_prio_mask = _mm_slli_epi32(v_prio_mask, 2 + 8);
+
+		__m128i v_ev_prio = _mm_srli_epi32(v_unpack_b32_63, 3 + 8);
+		__m128i v_prio_final = _mm_and_si128(v_ev_prio, v_prio_mask);
+		v_qe_hcws = _mm_or_si128(v_qe_hcws, v_prio_final);
+	}
+
+	/* Scalar Loading & inserts:
+	 * - Re-use scalar provided version until "prep_enq" is SSE)
+	 * SCHED_TYPE:
+	 * - Load 4x u8 as u32, insert
+	 * QUEUE ID:
+	 * - Load 4x u8 as u32, insert
+	 * Finalize:
+	 * - Shuffle u8 values to correct locations, OR into HCWs
+	 */
+	/* Expose for re-use in SCHED==DIRECTED mask creation later */
+	__m128i v_4x_qid_sched;
+	{
+		uint32_t qtmp = queue_id[0] |
+				queue_id[1] << 8 |
+				queue_id[2] << 8*2 |
+				queue_id[3] << 8*3;
+		__m128i v_4x_qid = _mm_insert_epi32(_mm_setzero_si128(),
+						qtmp, 0);
+
+		uint32_t stmp = sched_type[0] |
+				sched_type[1] << 8 |
+				sched_type[2] << 8*2 |
+				sched_type[3] << 8*3;
+		v_4x_qid_sched = _mm_insert_epi32(v_4x_qid, stmp, 1);
+
+		static const uint8_t data_move[] = {
+			 0, 4, 0xFF, 0xFF,
+			 1, 5, 0xFF, 0xFF,
+			 2, 6, 0xFF, 0xFF,
+			 3, 7, 0xFF, 0xFF,
+		};
+		__m128i v_qid_sched_move = _mm_loadu_si128((const void *)
+							   data_move);
+		__m128i v_qid_sched_done = _mm_shuffle_epi8(v_4x_qid_sched,
+							    v_qid_sched_move);
+		v_qe_hcws = _mm_or_si128(v_qe_hcws, v_qid_sched_done);
+	}
+
+	/* FlowID, SubEv Type:
+	 * - Process 4x in single reg
+	 * - Shift by 4 bits, blend to align for shuffle mask
+	 * - Shuffle FID vs sub/ev type order
+	 * - unpack as u16s to interleave with previous work
+	 */
+	__m128i v_sev_fid_done;
+	{
+		__m128i v_subevt_shift = _mm_srli_epi16(v_unpack_b00_31, 4);
+		__m128i v_subevt_fixed = _mm_blend_epi16(v_unpack_b00_31,
+							 v_subevt_shift, 0xAA);
+		static const uint8_t data_move[16] = {
+			 3,  2,  0,  1,
+			 7,  6,  4,  5,
+			11, 10,  8,  9,
+			15, 14, 12, 13,
+		};
+		__m128i v_data_move = _mm_loadu_si128((const void *)data_move);
+		v_sev_fid_done = _mm_shuffle_epi8(v_subevt_fixed, v_data_move);
+	}
+
+	/* If sched type == DIRECTED, copy QID/SCHED/PRIO fields to FLOWID u16:
+	 * v_qe_hcws u16 lane 0 gets written to v_sev_fid_done u16 lane 1
+	 * (x4 for QEs) blend results based on == DIR compare
+	 */
+	{
+		__m128i v_q_sp_shift = _mm_slli_epi32(v_qe_hcws, 16);
+		__m128i v_sched_dir = _mm_insert_epi32(v_zeros,
+						       0x3 | 0x3 << 8 |
+						       0x3 << 16 | 0x3 << 24,
+						       1);
+		__m128i v_dir_sched_mask = _mm_cmpeq_epi8(v_4x_qid_sched,
+							  v_sched_dir);
+
+		/* Duplicate cmp mask to u16 as required for larger blend */
+		static const uint8_t data_move[] = {
+			 0xFF, 0xFF, 4, 4,
+			 0xFF, 0xFF, 5, 5,
+			 0xFF, 0xFF, 6, 6,
+			 0xFF, 0xFF, 7, 7,
+		};
+		__m128i v_dir_sched_move = _mm_loadu_si128((const void *)
+							   data_move);
+		__m128i v_dir_sched_mask_done = _mm_shuffle_epi8(
+							v_dir_sched_mask,
+							v_dir_sched_move);
+
+		v_sev_fid_done = _mm_blendv_epi8(v_sev_fid_done,
+						 v_q_sp_shift,
+						 v_dir_sched_mask_done);
+	}
+
+	/* Unpacks from 2 regs to single u64 for each QE */
+	struct dlb2_enqueue_qe *qe = qm_port->qe4;
+	__m128i v_qe_01_u64s = _mm_unpacklo_epi16(v_sev_fid_done, v_qe_hcws);
+	__m128i v_qe_23_u64s = _mm_unpackhi_epi16(v_sev_fid_done, v_qe_hcws);
+
+	/* QE 0 */
+	__m128i v_qe0 = _mm_alignr_epi8(v_qe_01_u64s, v_evs[0], 8);
+	_mm_storeu_si128((void *)&qe[0], v_qe0);
+	/* QE 1 */
+	__m128i v_qe1 = _mm_blend_epi16(v_qe_01_u64s, _mm_alignr_epi8(v_evs[1],
+								      v_evs[1],
+								      8),
+					0x0F);
+	_mm_storeu_si128((void *)&qe[1], v_qe1);
+	/* QE 2 */
+	__m128i v_qe2 = _mm_alignr_epi8(v_qe_23_u64s, v_evs[2], 8);
+	_mm_storeu_si128((void *)&qe[2], v_qe2);
+	/* QE 3 */
+	__m128i v_qe3 = _mm_blend_epi16(v_qe_23_u64s,
+					_mm_alignr_epi8(v_evs[3], v_evs[3], 8),
+					0x0F);
+	_mm_storeu_si128((void *)&qe[3], v_qe3);
+}
+
+static inline void
 dlb2_construct_token_pop_qe(struct dlb2_port *qm_port, int idx)
 {
 	struct dlb2_cq_pop_qe *qe = (void *)qm_port->qe4;
@@ -2932,10 +3183,11 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 		int j = 0;
 
 		/* Zero-out QEs */
-		qm_port->qe4[0].cmd_byte = 0;
-		qm_port->qe4[1].cmd_byte = 0;
-		qm_port->qe4[2].cmd_byte = 0;
-		qm_port->qe4[3].cmd_byte = 0;
+		_mm_storeu_si128((void *)&qm_port->qe4[0], _mm_setzero_si128());
+		_mm_storeu_si128((void *)&qm_port->qe4[1], _mm_setzero_si128());
+		_mm_storeu_si128((void *)&qm_port->qe4[2], _mm_setzero_si128());
+		_mm_storeu_si128((void *)&qm_port->qe4[3], _mm_setzero_si128());
+
 
 		for (; j < DLB2_NUM_QES_PER_CACHE_LINE && (i + j) < n; j++) {
 			int16_t thresh = qm_port->token_pop_thresh;
@@ -2965,7 +3217,7 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 
 sw_credit_update:
 	/* each release returns one credit */
-	if (!ev_port->outstanding_releases) {
+	if (unlikely(!ev_port->outstanding_releases)) {
 		DLB2_LOG_ERR("%s: Outstanding releases underflowed.\n",
 			     __func__);
 		return;
@@ -3082,7 +3334,7 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 	return 0;
 }
 
-static inline int
+static __rte_noinline int
 dlb2_process_dequeue_qes(struct dlb2_eventdev_port *ev_port,
 			 struct dlb2_port *qm_port,
 			 struct rte_event *events,
@@ -3351,8 +3603,7 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 
 	cq_addr = dlb2_port[qm_port->id][PORT_TYPE(qm_port)].cq_base;
 
-	idx = qm_port->cq_idx;
-
+	idx = qm_port->cq_idx_unmasked & qm_port->cq_depth_mask;
 	/* Load the next 4 QEs */
 	addr[0] = (uintptr_t)&cq_addr[idx];
 	addr[1] = (uintptr_t)&cq_addr[(idx +  4) & qm_port->cq_depth_mask];
@@ -3398,6 +3649,272 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 }
 
 static inline void
+_process_deq_qes_vec_impl(struct dlb2_port *qm_port,
+			  struct rte_event *events,
+			  __m128i v_qe_3,
+			  __m128i v_qe_2,
+			  __m128i v_qe_1,
+			  __m128i v_qe_0,
+			  __m128i v_qe_meta,
+			  __m128i v_qe_status,
+			  uint32_t valid_events)
+{
+	/* Look up the event QIDs, using the hardware QIDs to index the
+	 * port's QID mapping.
+	 *
+	 * Each v_qe_[0-4] is just a 16-byte load of the whole QE. It is
+	 * passed along in registers as the QE data is required later.
+	 *
+	 * v_qe_meta is an u32 unpack of all 4x QEs. Aka, it contains one
+	 * 32-bit slice of each QE, so makes up a full SSE register. This
+	 * allows parallel processing of 4x QEs in a single register.
+	 */
+
+	__m128i v_qid_done = {0};
+	int hw_qid0 = _mm_extract_epi8(v_qe_meta, 2);
+	int hw_qid1 = _mm_extract_epi8(v_qe_meta, 6);
+	int hw_qid2 = _mm_extract_epi8(v_qe_meta, 10);
+	int hw_qid3 = _mm_extract_epi8(v_qe_meta, 14);
+
+	int ev_qid0 = qm_port->qid_mappings[hw_qid0];
+	int ev_qid1 = qm_port->qid_mappings[hw_qid1];
+	int ev_qid2 = qm_port->qid_mappings[hw_qid2];
+	int ev_qid3 = qm_port->qid_mappings[hw_qid3];
+
+	v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid0, 2);
+	v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid1, 6);
+	v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid2, 10);
+	v_qid_done = _mm_insert_epi8(v_qid_done, ev_qid3, 14);
+
+	/* Schedule field remapping using byte shuffle
+	 * - Full byte containing sched field handled here (op, rsvd are zero)
+	 * - Note sanitizing the register requires two masking ANDs:
+	 *   1) to strip prio/msg_type from byte for correct shuffle lookup
+	 *   2) to strip any non-sched-field lanes from any results to OR later
+	 * - Final byte result is >> 10 to another byte-lane inside the u32.
+	 *   This makes the final combination OR easier to make the rte_event.
+	 */
+	__m128i v_sched_done;
+	__m128i v_sched_bits;
+	{
+		static const uint8_t sched_type_map[16] = {
+			[DLB2_SCHED_ATOMIC] = RTE_SCHED_TYPE_ATOMIC,
+			[DLB2_SCHED_UNORDERED] = RTE_SCHED_TYPE_PARALLEL,
+			[DLB2_SCHED_ORDERED] = RTE_SCHED_TYPE_ORDERED,
+			[DLB2_SCHED_DIRECTED] = RTE_SCHED_TYPE_ATOMIC,
+		};
+		static const uint8_t sched_and_mask[16] = {
+			0x00, 0x00, 0x00, 0x03,
+			0x00, 0x00, 0x00, 0x03,
+			0x00, 0x00, 0x00, 0x03,
+			0x00, 0x00, 0x00, 0x03,
+		};
+		const __m128i v_sched_map = _mm_loadu_si128(
+					     (const __m128i *)sched_type_map);
+		__m128i v_sched_mask = _mm_loadu_si128(
+					     (const __m128i *)&sched_and_mask);
+		v_sched_bits = _mm_and_si128(v_qe_meta, v_sched_mask);
+		__m128i v_sched_remapped = _mm_shuffle_epi8(v_sched_map,
+							    v_sched_bits);
+		__m128i v_preshift = _mm_and_si128(v_sched_remapped,
+						   v_sched_mask);
+		v_sched_done = _mm_srli_epi32(v_preshift, 10);
+	}
+
+	/* Priority handling
+	 * - QE provides 3 bits of priority
+	 * - Shift << 3 to move to MSBs for byte-prio in rte_event
+	 * - Mask bits to avoid pollution, leaving only 3 prio MSBs in reg
+	 */
+	__m128i v_prio_done;
+	{
+		static const uint8_t prio_mask[16] = {
+			0x00, 0x00, 0x00, 0x07 << 5,
+			0x00, 0x00, 0x00, 0x07 << 5,
+			0x00, 0x00, 0x00, 0x07 << 5,
+			0x00, 0x00, 0x00, 0x07 << 5,
+		};
+		__m128i v_prio_mask  = _mm_loadu_si128(
+						(const __m128i *)prio_mask);
+		__m128i v_prio_shifted = _mm_slli_epi32(v_qe_meta, 3);
+		v_prio_done = _mm_and_si128(v_prio_shifted, v_prio_mask);
+	}
+
+	/* Event Sub/Type handling:
+	 * we want to keep the lower 12 bits of each QE. Shift up by 20 bits
+	 * to get the sub/ev type data into rte_event location, clearing the
+	 * lower 20 bits in the process.
+	 */
+	__m128i v_types_done;
+	{
+		static const uint8_t event_mask[16] = {
+			0x0f, 0x00, 0x00, 0x00,
+			0x0f, 0x00, 0x00, 0x00,
+			0x0f, 0x00, 0x00, 0x00,
+			0x0f, 0x00, 0x00, 0x00,
+		};
+		static const uint8_t sub_event_mask[16] = {
+			0xff, 0x00, 0x00, 0x00,
+			0xff, 0x00, 0x00, 0x00,
+			0xff, 0x00, 0x00, 0x00,
+			0xff, 0x00, 0x00, 0x00,
+		};
+		static const uint8_t flow_mask[16] = {
+			0xff, 0xff, 0x00, 0x00,
+			0xff, 0xff, 0x00, 0x00,
+			0xff, 0xff, 0x00, 0x00,
+			0xff, 0xff, 0x00, 0x00,
+		};
+		__m128i v_event_mask  = _mm_loadu_si128(
+					(const __m128i *)event_mask);
+		__m128i v_sub_event_mask  = _mm_loadu_si128(
+					(const __m128i *)sub_event_mask);
+		__m128i v_flow_mask  = _mm_loadu_si128(
+				       (const __m128i *)flow_mask);
+		__m128i v_sub = _mm_srli_epi32(v_qe_meta, 8);
+		v_sub = _mm_and_si128(v_sub, v_sub_event_mask);
+		__m128i v_type = _mm_and_si128(v_qe_meta, v_event_mask);
+		v_type = _mm_slli_epi32(v_type, 8);
+		v_types_done = _mm_or_si128(v_type, v_sub);
+		v_types_done = _mm_slli_epi32(v_types_done, 20);
+		__m128i v_flow = _mm_and_si128(v_qe_status, v_flow_mask);
+		v_types_done = _mm_or_si128(v_types_done, v_flow);
+	}
+
+	/* Combine QID, Sched and Prio fields, then Shift >> 8 bits to align
+	 * with the rte_event, allowing unpacks to move/blend with payload.
+	 */
+	__m128i v_q_s_p_done;
+	{
+		__m128i v_qid_sched = _mm_or_si128(v_qid_done, v_sched_done);
+		__m128i v_q_s_prio = _mm_or_si128(v_qid_sched, v_prio_done);
+		v_q_s_p_done = _mm_srli_epi32(v_q_s_prio, 8);
+	}
+
+	__m128i v_unpk_ev_23, v_unpk_ev_01, v_ev_2, v_ev_3, v_ev_0, v_ev_1;
+
+	/* Unpack evs into u64 metadata, then indiv events */
+	v_unpk_ev_23 = _mm_unpackhi_epi32(v_types_done, v_q_s_p_done);
+	v_unpk_ev_01 = _mm_unpacklo_epi32(v_types_done, v_q_s_p_done);
+
+	switch (valid_events) {
+	case 4:
+		v_ev_3 = _mm_blend_epi16(v_unpk_ev_23, v_qe_3, 0x0F);
+		v_ev_3 = _mm_alignr_epi8(v_ev_3, v_ev_3, 8);
+		_mm_storeu_si128((__m128i *)&events[3], v_ev_3);
+		/* fallthrough */
+	case 3:
+		v_ev_2 = _mm_unpacklo_epi64(v_unpk_ev_23, v_qe_2);
+		_mm_storeu_si128((__m128i *)&events[2], v_ev_2);
+		/* fallthrough */
+	case 2:
+		v_ev_1 = _mm_blend_epi16(v_unpk_ev_01, v_qe_1, 0x0F);
+		v_ev_1 = _mm_alignr_epi8(v_ev_1, v_ev_1, 8);
+		_mm_storeu_si128((__m128i *)&events[1], v_ev_1);
+		/* fallthrough */
+	case 1:
+		v_ev_0 = _mm_unpacklo_epi64(v_unpk_ev_01, v_qe_0);
+		_mm_storeu_si128((__m128i *)&events[0], v_ev_0);
+	}
+}
+
+static __rte_always_inline int
+dlb2_recv_qe_sparse_vec(struct dlb2_port *qm_port, void *events,
+			uint32_t max_events)
+{
+	/* Using unmasked idx for perf, and masking manually */
+	uint16_t idx = qm_port->cq_idx_unmasked;
+	volatile struct dlb2_dequeue_qe *cq_addr;
+
+	cq_addr = dlb2_port[qm_port->id][PORT_TYPE(qm_port)].cq_base;
+
+	uintptr_t qe_ptr_3 = (uintptr_t)&cq_addr[(idx + 12) &
+						 qm_port->cq_depth_mask];
+	uintptr_t qe_ptr_2 = (uintptr_t)&cq_addr[(idx +  8) &
+						 qm_port->cq_depth_mask];
+	uintptr_t qe_ptr_1 = (uintptr_t)&cq_addr[(idx +  4) &
+						 qm_port->cq_depth_mask];
+	uintptr_t qe_ptr_0 = (uintptr_t)&cq_addr[(idx +  0) &
+						 qm_port->cq_depth_mask];
+
+	/* Load QEs from CQ: use compiler barriers to avoid load reordering */
+	__m128i v_qe_3 = _mm_loadu_si128((const __m128i *)qe_ptr_3);
+	rte_compiler_barrier();
+	__m128i v_qe_2 = _mm_loadu_si128((const __m128i *)qe_ptr_2);
+	rte_compiler_barrier();
+	__m128i v_qe_1 = _mm_loadu_si128((const __m128i *)qe_ptr_1);
+	rte_compiler_barrier();
+	__m128i v_qe_0 = _mm_loadu_si128((const __m128i *)qe_ptr_0);
+
+	/* Generate the pkt_shuffle mask;
+	 * - Avoids load in otherwise load-heavy section of code
+	 * - Moves bytes 3,7,11,15 (gen bit bytes) to LSB bytes in XMM
+	 */
+	const uint32_t stat_shuf_bytes = (15 << 24) | (11 << 16) | (7 << 8) | 3;
+	__m128i v_zeros = _mm_setzero_si128();
+	__m128i v_ffff = _mm_cmpeq_epi8(v_zeros, v_zeros);
+	__m128i v_stat_shuf_mask = _mm_insert_epi32(v_ffff, stat_shuf_bytes, 0);
+
+	/* Extract u32 components required from the QE
+	 * - QE[64 to 95 ] for metadata (qid, sched, prio, event type, ...)
+	 * - QE[96 to 127] for status (cq gen bit, error)
+	 *
+	 * Note that stage 1 of the unpacking is re-used for both u32 extracts
+	 */
+	__m128i v_qe_02 = _mm_unpackhi_epi32(v_qe_0, v_qe_2);
+	__m128i v_qe_13 = _mm_unpackhi_epi32(v_qe_1, v_qe_3);
+	__m128i v_qe_status = _mm_unpackhi_epi32(v_qe_02, v_qe_13);
+	__m128i v_qe_meta   = _mm_unpacklo_epi32(v_qe_02, v_qe_13);
+
+	/* Status byte (gen_bit, error) handling:
+	 * - Shuffle to lanes 0,1,2,3, clear all others
+	 * - Shift right by 7 for gen bit to MSB, movemask to scalar
+	 * - Shift right by 2 for error bit to MSB, movemask to scalar
+	 */
+	__m128i v_qe_shuffled = _mm_shuffle_epi8(v_qe_status, v_stat_shuf_mask);
+	__m128i v_qes_shift_gen_bit = _mm_slli_epi32(v_qe_shuffled, 7);
+	int32_t qe_gen_bits = _mm_movemask_epi8(v_qes_shift_gen_bit) & 0xf;
+
+	/* Expected vs Reality of QE Gen bits
+	 * - cq_rolling_mask provides expected bits
+	 * - QE loads, unpacks/shuffle and movemask provides reality
+	 * - XOR of the two gives bitmask of new packets
+	 * - POPCNT to get the number of new events
+	 */
+	uint64_t rolling = qm_port->cq_rolling_mask & 0xF;
+	uint64_t qe_xor_bits = (qe_gen_bits ^ rolling);
+	uint32_t count_new = __builtin_popcount(qe_xor_bits);
+	count_new = RTE_MIN(count_new, max_events);
+	if (!count_new)
+		return 0;
+
+	/* emulate a 128 bit rotate using 2x 64-bit numbers and bit-shifts */
+
+	uint64_t m_rshift = qm_port->cq_rolling_mask >> count_new;
+	uint64_t m_lshift = qm_port->cq_rolling_mask << (64 - count_new);
+	uint64_t m2_rshift = qm_port->cq_rolling_mask_2 >> count_new;
+	uint64_t m2_lshift = qm_port->cq_rolling_mask_2 << (64 - count_new);
+
+	/* shifted out of m2 into MSB of m */
+	qm_port->cq_rolling_mask = (m_rshift | m2_lshift);
+
+	/* shifted out of m "looped back" into MSB of m2 */
+	qm_port->cq_rolling_mask_2 = (m2_rshift | m_lshift);
+
+	/* Prefetch the next QEs - should run as IPC instead of cycles */
+	rte_prefetch0(&cq_addr[(idx + 16) & qm_port->cq_depth_mask]);
+	rte_prefetch0(&cq_addr[(idx + 20) & qm_port->cq_depth_mask]);
+	rte_prefetch0(&cq_addr[(idx + 24) & qm_port->cq_depth_mask]);
+	rte_prefetch0(&cq_addr[(idx + 28) & qm_port->cq_depth_mask]);
+
+	/* Convert QEs from XMM regs to events and store events directly */
+	_process_deq_qes_vec_impl(qm_port, events, v_qe_3, v_qe_2, v_qe_1,
+				  v_qe_0, v_qe_meta, v_qe_status, count_new);
+
+	return count_new;
+}
+
+static inline void
 dlb2_inc_cq_idx(struct dlb2_port *qm_port, int cnt)
 {
 	uint16_t idx = qm_port->cq_idx_unmasked + cnt;
@@ -3414,25 +3931,15 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 		       uint16_t max_num,
 		       uint64_t dequeue_timeout_ticks)
 {
-	uint64_t timeout;
 	uint64_t start_ticks = 0ULL;
 	struct dlb2_port *qm_port;
 	int num = 0;
+	bool use_scalar;
+	uint64_t timeout;
 
 	qm_port = &ev_port->qm_port;
+	use_scalar = qm_port->use_scalar;
 
-	/* We have a special implementation for waiting. Wait can be:
-	 * 1) no waiting at all
-	 * 2) busy poll only
-	 * 3) wait for interrupt. If wakeup and poll time
-	 * has expired, then return to caller
-	 * 4) umonitor/umwait repeatedly up to poll time
-	 */
-
-	/* If configured for per dequeue wait, then use wait value provided
-	 * to this API. Otherwise we must use the global
-	 * value from eventdev config time.
-	 */
 	if (!dlb2->global_dequeue_wait)
 		timeout = dequeue_timeout_ticks;
 	else
@@ -3440,35 +3947,41 @@ static int dlb2_num_dir_queues_setup(struct dlb2_eventdev *dlb2)
 
 	start_ticks = rte_get_timer_cycles();
 
+	use_scalar = use_scalar || (max_num & 0x3);
+
 	while (num < max_num) {
 		struct dlb2_dequeue_qe qes[DLB2_NUM_QES_PER_CACHE_LINE];
 		int num_avail;
-
-		/* Copy up to 4 QEs from the current cache line into qes */
-		num_avail = dlb2_recv_qe_sparse(qm_port, qes);
-
-		/* But don't process more than the user requested */
-		num_avail = RTE_MIN(num_avail, max_num - num);
-
-		dlb2_inc_cq_idx(qm_port, num_avail << 2);
-
-		if (num_avail == DLB2_NUM_QES_PER_CACHE_LINE)
-			num += dlb2_process_dequeue_four_qes(ev_port,
-							      qm_port,
-							      &events[num],
-							      &qes[0]);
-		else if (num_avail)
-			num += dlb2_process_dequeue_qes(ev_port,
-							 qm_port,
-							 &events[num],
-							 &qes[0],
-							 num_avail);
-		else if ((timeout == 0) || (num > 0))
-			/* Not waiting in any form, or 1+ events received? */
-			break;
-		else if (dlb2_dequeue_wait(dlb2, ev_port, qm_port,
-					   timeout, start_ticks))
-			break;
+		if (use_scalar) {
+			num_avail = dlb2_recv_qe_sparse(qm_port, qes);
+			num_avail = RTE_MIN(num_avail, max_num - num);
+			dlb2_inc_cq_idx(qm_port, num_avail << 2);
+			if (num_avail == DLB2_NUM_QES_PER_CACHE_LINE)
+				num += dlb2_process_dequeue_four_qes(ev_port,
+								  qm_port,
+								  &events[num],
+								  &qes[0]);
+			else if (num_avail)
+				num += dlb2_process_dequeue_qes(ev_port,
+								qm_port,
+								&events[num],
+								&qes[0],
+								num_avail);
+		} else { /* !use_scalar */
+			num_avail = dlb2_recv_qe_sparse_vec(qm_port,
+							    &events[num],
+							    max_num - num);
+			num += num_avail;
+			dlb2_inc_cq_idx(qm_port, num_avail << 2);
+			DLB2_INC_STAT(ev_port->stats.traffic.rx_ok, num_avail);
+		}
+		if (!num_avail) {
+			if (num > 0)
+				break;
+			else if (dlb2_dequeue_wait(dlb2, ev_port, qm_port,
+						   timeout, start_ticks))
+				break;
+		}
 	}
 
 	qm_port->owed_tokens += num;
diff --git a/drivers/event/dlb2/dlb2_priv.h b/drivers/event/dlb2/dlb2_priv.h
index ad663a3..5693448 100644
--- a/drivers/event/dlb2/dlb2_priv.h
+++ b/drivers/event/dlb2/dlb2_priv.h
@@ -199,9 +199,9 @@ enum dlb2_enqueue_type {
 /* hw-specific format - do not change */
 
 struct dlb2_event_type {
-	uint8_t major:4;
-	uint8_t unused:4;
-	uint8_t sub;
+	uint16_t major:4;
+	uint16_t unused:4;
+	uint16_t sub:8;
 };
 
 union dlb2_opaque_data {
@@ -345,6 +345,12 @@ struct dlb2_port {
 	uint16_t cq_idx_unmasked;
 	uint16_t cq_depth_mask;
 	uint16_t gen_bit_shift;
+	uint64_t cq_rolling_mask; /*
+				   * rotate to always have right expected
+				   * gen bits
+				   */
+	uint64_t cq_rolling_mask_2;
+	void *cq_addr_cached; /* avoid multiple refs */
 	enum dlb2_port_state state;
 	enum dlb2_configuration_state config_state;
 	int num_mapped_qids;
@@ -354,6 +360,7 @@ struct dlb2_port {
 	struct dlb2_cq_pop_qe *consume_qe;
 	struct dlb2_eventdev *dlb2; /* back ptr */
 	struct dlb2_eventdev_port *ev_port; /* back ptr */
+	bool use_scalar; /* force usage of scalar code */
 };
 
 /* Per-process per-port mmio and memory pointers */
@@ -507,9 +514,9 @@ struct dlb2_queue {
 	uint32_t num_qid_inflights; /* User config */
 	uint32_t num_atm_inflights; /* User config */
 	enum dlb2_configuration_state config_state;
-	int sched_type; /* LB queue only */
-	uint32_t id;
-	bool is_directed;
+	int  sched_type; /* LB queue only */
+	uint8_t id;
+	bool	 is_directed;
 };
 
 struct dlb2_eventdev_queue {
-- 
1.7.10


^ permalink raw reply	[flat|nested] 9+ messages in thread

end of thread, other threads:[~2021-05-04  8:29 UTC | newest]

Thread overview: 9+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-03-17 17:02 [dpdk-dev] [PATCH] event/dlb2: Optimize Dequeue Operations Timothy McDaniel
2021-03-21 11:10 ` Jerin Jacob
2021-04-13 20:30 ` [dpdk-dev] [PATCH v2 0/1] Optimize DLB Dequeue Timothy McDaniel
2021-04-13 20:30   ` [dpdk-dev] [PATCH v2 1/1] event/dlb: optimize Dequeue Operations Timothy McDaniel
2021-04-29  7:20   ` [dpdk-dev] [PATCH v2 0/1] Optimize DLB Dequeue Jerin Jacob
2021-04-29 13:45     ` McDaniel, Timothy
2021-05-01 19:07 ` [dpdk-dev] [PATCH v3 0/1] Optimize DLB2 Dequeue Operations McDaniel, Timothy
2021-05-01 19:07   ` [dpdk-dev] [PATCH v3 1/1] event/dlb2: optimize " McDaniel, Timothy
2021-05-04  8:29     ` Jerin Jacob

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).