Hi Maxime,
> -----Original Message-----
> From: Maxime Coquelin <maxime.coque...@redhat.com>
> Sent: Thursday, September 15, 2022 1:15 AM
> To: Vargas, Hernan <hernan.var...@intel.com>; dev@dpdk.org;
> gak...@marvell.com; t...@redhat.com
> Cc: Chautru, Nicolas <nicolas.chau...@intel.com>; Zhang, Qi Z
> <qi.z.zh...@intel.com>
> Subject: Re: [PATCH v2 15/37] baseband/acc100: add workaround for deRM
> corner cases
>
>
>
> On 8/20/22 04:31, Hernan Vargas wrote:
> > Add function to asses if de-ratematch pre-processing should be run in
> > SW for corner cases.
> >
> > Signed-off-by: Hernan Vargas <hernan.var...@intel.com>
> > ---
> > drivers/baseband/acc100/acc100_pmd.h | 13 +++
> > drivers/baseband/acc100/rte_acc100_pmd.c | 103
> ++++++++++++++++++++++-
> > 2 files changed, 114 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/baseband/acc100/acc100_pmd.h
> > b/drivers/baseband/acc100/acc100_pmd.h
> > index 19a1f434bc..c98a182be6 100644
> > --- a/drivers/baseband/acc100/acc100_pmd.h
> > +++ b/drivers/baseband/acc100/acc100_pmd.h
> > @@ -140,6 +140,8 @@
> > /* Constants from K0 computation from 3GPP 38.212 Table 5.4.2.1-2 */
> > #define ACC100_N_ZC_1 66 /* N = 66 Zc for BG 1 */
> > #define ACC100_N_ZC_2 50 /* N = 50 Zc for BG 2 */
> > +#define ACC100_K_ZC_1 22 /* K = 22 Zc for BG 1 */ #define
> > +ACC100_K_ZC_2 10 /* K = 10 Zc for BG 2 */
> > #define ACC100_K0_1_1 17 /* K0 fraction numerator for rv 1 and BG 1 */
> > #define ACC100_K0_1_2 13 /* K0 fraction numerator for rv 1 and BG 2 */
> > #define ACC100_K0_2_1 33 /* K0 fraction numerator for rv 2 and BG 1
> > */ @@ -177,6 +179,16 @@
> > #define ACC100_MS_IN_US (1000)
> > #define ACC100_DDR_TRAINING_MAX (5000)
> >
> > +/* Code rate limitation when padding is required */ #define
> > +ACC100_LIM_03 2 /* 0.03 */ #define ACC100_LIM_09 6 /* 0.09 */
> > +#define ACC100_LIM_14 9 /* 0.14 */ #define ACC100_LIM_21 14 /* 0.21
> > +*/ #define ACC100_LIM_31 20 /* 0.31 */ #define ACC100_MAX_E (128 *
> > +1024 - 2)
> > +
> > +
> > +
> > /* ACC100 DMA Descriptor triplet */
> > struct acc100_dma_triplet {
> > uint64_t address;
> > @@ -572,6 +584,7 @@ struct __rte_cache_aligned acc100_queue {
> > uint8_t *lb_out;
> > rte_iova_t lb_in_addr_iova;
> > rte_iova_t lb_out_addr_iova;
> > + int8_t *derm_buffer; /* interim buffer for de-rm in SDK */
> > struct acc100_device *d;
> > };
> >
> > diff --git a/drivers/baseband/acc100/rte_acc100_pmd.c
> > b/drivers/baseband/acc100/rte_acc100_pmd.c
> > index 1504acfadd..69c0714a37 100644
> > --- a/drivers/baseband/acc100/rte_acc100_pmd.c
> > +++ b/drivers/baseband/acc100/rte_acc100_pmd.c
> > @@ -24,6 +24,10 @@
> > #include "acc100_pmd.h"
> > #include "acc101_pmd.h"
> >
> > +#ifdef RTE_BBDEV_SDK_AVX512
> > +#include <phy_rate_dematching_5gnr.h> #endif
> > +
> > #ifdef RTE_LIBRTE_BBDEV_DEBUG
> > RTE_LOG_REGISTER_DEFAULT(acc100_logtype, DEBUG);
> > #else
> > @@ -898,6 +902,16 @@ acc100_queue_setup(struct rte_bbdev *dev,
> uint16_t queue_id,
> > rte_free(q);
> > return -ENOMEM;
> > }
> > + q->derm_buffer = rte_zmalloc_socket(dev->device->driver->name,
> > + RTE_BBDEV_TURBO_MAX_CB_SIZE * 10,
> > + RTE_CACHE_LINE_SIZE, conf->socket);
> > + if (q->derm_buffer == NULL) {
> > + rte_bbdev_log(ERR, "Failed to allocate derm_buffer
> memory");
> > + rte_free(q->lb_in);
> > + rte_free(q->lb_out);
> > + rte_free(q);
> > + return -ENOMEM;
> > + }
>
> It may make sense to have a common error path to avoid duplication and so
> risk introducing memory leaks when changes will be made.
>
> > q->lb_out_addr_iova = rte_malloc_virt2iova(q->lb_out);
> >
> > /*
> > @@ -918,6 +932,7 @@ acc100_queue_setup(struct rte_bbdev *dev,
> uint16_t
> > queue_id,
> >
> > q_idx = acc100_find_free_queue_idx(dev, conf);
> > if (q_idx == -1) {
> > + rte_free(q->derm_buffer);
> > rte_free(q->lb_in);
> > rte_free(q->lb_out);
> > rte_free(q);
> > @@ -955,6 +970,7 @@ acc100_queue_release(struct rte_bbdev *dev,
> uint16_t q_id)
> > /* Mark the Queue as un-assigned */
> > d->q_assigned_bit_map[q->qgrp_id] &= (0xFFFFFFFF -
> > (1 << q->aq_id));
> > + rte_free(q->derm_buffer);
> > rte_free(q->lb_in);
> > rte_free(q->lb_out);
> > rte_free(q);
> > @@ -3512,10 +3528,42 @@ harq_loopback(struct acc100_queue *q,
> struct rte_bbdev_dec_op *op,
> > return 1;
> > }
> >
> > +/** Assess whether a work around is required for the deRM corner
> > +cases */ static inline bool derm_workaround_required(struct
> > +rte_bbdev_op_ldpc_dec *ldpc_dec, struct acc100_queue *q) {
> > + if (!is_acc100(q))
> > + return false;
> > + int32_t e = ldpc_dec->cb_params.e;
> > + int q_m = ldpc_dec->q_m;
> > + int z_c = ldpc_dec->z_c;
> > + int K = (ldpc_dec->basegraph == 1 ? ACC100_K_ZC_1 :
> ACC100_K_ZC_2)
> > + * z_c;
> > + bool required = false;
>
> Add new line.
>
> > + if (ldpc_dec->basegraph == 1) {
> > + if ((q_m == 4) && (z_c >= 320) && (e * ACC100_LIM_31 > K *
> 64))
> > + required = true;
> > + else if ((e * ACC100_LIM_21 > K * 64))
> > + required = true;
> > + } else {
> > + if (q_m <= 2) {
> > + if ((z_c >= 208) && (e * ACC100_LIM_09 > K * 64))
> > + required = true;
> > + else if ((z_c < 208) && (e * ACC100_LIM_03 > K * 64))
> > + required = true;
> > + } else if (e * ACC100_LIM_14 > K * 64)
> > + required = true;
> > + }
> > + if (required)
> > + rte_bbdev_log(INFO, "Running deRM pre-processing in SW");
>
> Add new line.
>
> > + return required;
> > +}
> > +
> > /** Enqueue one decode operations for ACC100 device in CB mode */
> > static inline int
> > enqueue_ldpc_dec_one_op_cb(struct acc100_queue *q, struct
> rte_bbdev_dec_op *op,
> > - uint16_t total_enqueued_cbs, bool same_op)
> > + uint16_t total_enqueued_cbs, bool same_op,
> > + struct rte_bbdev_queue_data *q_data)
> > {
> > int ret;
> > if (unlikely(check_bit(op->ldpc_dec.op_flags,
> > @@ -3571,6 +3619,57 @@ enqueue_ldpc_dec_one_op_cb(struct
> acc100_queue *q, struct rte_bbdev_dec_op *op,
> > &in_offset, &h_out_offset,
> > &h_out_length, harq_layout);
> > } else {
> > + if (derm_workaround_required(&op->ldpc_dec, q)) {
> > + #ifdef RTE_BBDEV_SDK_AVX512
>
> First, the indentation is not good here.
>
> Also, my understanding is that this code will get built only if Flexran SDK is
> available. Flexran SDK is proprietary, and so it is not possible to have this
> code exercised by the upstream CI.
>
> Code under RTE_BBDEV_SDK_AVX512 should be dropped IMO.
We provide a subset of the intel SDKs to the community. More generally these
functions could be replaced with other versions (including from other
companies).
>
> > + struct rte_bbdev_op_ldpc_dec *dec = &op-
> >ldpc_dec;
> > + /* Checking input size is matching with E */
> > + if (dec->input.data->data_len < dec->cb_params.e) {
> > + rte_bbdev_log(ERR,
> > + "deRM: Input size
> mismatch");
> > + return -EFAULT;
> > + }
> > + /* Run first deRM processing in SW */
> > + struct bblib_rate_dematching_5gnr_request
> derm_req;
> > + struct bblib_rate_dematching_5gnr_response
> derm_resp;
> > + uint8_t *in = rte_pktmbuf_mtod_offset(dec-
> >input.data,
> > + uint8_t *, in_offset);
>
> Don't mix declarations & code.
>
> > + derm_req.p_in = (int8_t *) in;
> > + derm_req.p_harq = (int8_t *) q->derm_buffer;
> > + derm_req.base_graph = dec->basegraph;
> > + derm_req.zc = dec->z_c;
> > + derm_req.ncb = dec->n_cb;
> > + derm_req.e = dec->cb_params.e;
> > + if (derm_req.e > ACC100_MAX_E) {
> > + rte_bbdev_log(WARNING,
> > + "deRM: E %d > %d max",
> > + derm_req.e,
> ACC100_MAX_E);
> > + derm_req.e = ACC100_MAX_E;
> > + }
> > + derm_req.k0 = 0; /* Actual output from SDK */
> > + derm_req.isretx = false;
> > + derm_req.rvid = dec->rv_index;
> > + derm_req.modulation_order = dec->q_m;
> > + derm_req.start_null_index =
> > + (dec->basegraph == 1 ? 22 : 10)
> > + * dec->z_c - 2 * dec->z_c
> > + - dec->n_filler;
> > + derm_req.num_of_null = dec->n_filler;
> > + bblib_rate_dematching_5gnr(&derm_req,
> &derm_resp);
> > + /* Force back the HW DeRM */
> > + dec->q_m = 1;
> > + dec->cb_params.e = dec->n_cb - dec->n_filler;
> > + dec->rv_index = 0;
> > + rte_memcpy(in, q->derm_buffer, dec->cb_params.e);
> > + /* Capture counter when pre-processing is used */
> > + q_data->queue_stats.enqueue_warn_count++;
> > + #else
> > + RTE_SET_USED(q_data);
> > + rte_bbdev_log(WARNING,
> > + "Corner case may require deRM pre-
> processing in SDK"
> > + );
> > + #endif
> > + }
> > +
> > struct acc100_fcw_ld *fcw;
> > uint32_t seg_total_left;
>
> Don't mix declarations & code.
>
> > fcw = &desc->req.fcw_ld;
> > @@ -4322,7 +4421,7 @@ acc100_enqueue_ldpc_dec_cb(struct
> rte_bbdev_queue_data *q_data,
> > ops[i]->ldpc_dec.n_cb, ops[i]->ldpc_dec.q_m,
> > ops[i]->ldpc_dec.n_filler, ops[i]-
> >ldpc_dec.cb_params.e,
> > same_op);
> > - ret = enqueue_ldpc_dec_one_op_cb(q, ops[i], i, same_op);
> > + ret = enqueue_ldpc_dec_one_op_cb(q, ops[i], i, same_op,
> q_data);
> > if (ret < 0) {
> > acc100_enqueue_invalid(q_data);
> > break;
