From: "Chautru, Nicolas" <nicolas.chautru@intel.com>
To: Maxime Coquelin <maxime.coquelin@redhat.com>,
"dev@dpdk.org" <dev@dpdk.org>
Cc: "hemant.agrawal@nxp.com" <hemant.agrawal@nxp.com>,
"david.marchand@redhat.com" <david.marchand@redhat.com>,
"Vargas, Hernan" <hernan.vargas@intel.com>
Subject: RE: [PATCH v3 09/12] baseband/acc: add FFT support to VRB2 variant
Date: Thu, 5 Oct 2023 17:59:04 +0000 [thread overview]
Message-ID: <BY5PR11MB445123214192EC8A9A5031F3F8CAA@BY5PR11MB4451.namprd11.prod.outlook.com> (raw)
In-Reply-To: <21e67536-887e-4a2e-c650-2f056c0d0c6a@redhat.com>
Hi Maxime,
> -----Original Message-----
> From: Maxime Coquelin <maxime.coquelin@redhat.com>
> Sent: Thursday, October 5, 2023 7:35 AM
> To: Chautru, Nicolas <nicolas.chautru@intel.com>; dev@dpdk.org
> Cc: hemant.agrawal@nxp.com; david.marchand@redhat.com; Vargas, Hernan
> <hernan.vargas@intel.com>
> Subject: Re: [PATCH v3 09/12] baseband/acc: add FFT support to VRB2 variant
>
>
>
> On 10/4/23 23:18, Chautru, Nicolas wrote:
> > Hi Maxime,
> >
> >> -----Original Message-----
> >> From: Maxime Coquelin <maxime.coquelin@redhat.com>
> >> Sent: Wednesday, October 4, 2023 12:11 AM
> >> To: Chautru, Nicolas <nicolas.chautru@intel.com>; dev@dpdk.org
> >> Cc: hemant.agrawal@nxp.com; david.marchand@redhat.com; Vargas,
> Hernan
> >> <hernan.vargas@intel.com>
> >> Subject: Re: [PATCH v3 09/12] baseband/acc: add FFT support to VRB2
> >> variant
> >>
> >>
> >>
> >> On 10/3/23 20:20, Chautru, Nicolas wrote:
> >>> Hi Maxime,
> >>>
> >>>> -----Original Message-----
> >>>> From: Maxime Coquelin <maxime.coquelin@redhat.com>
> >>>> Sent: Tuesday, October 3, 2023 7:37 AM
> >>>> To: Chautru, Nicolas <nicolas.chautru@intel.com>; dev@dpdk.org
> >>>> Cc: hemant.agrawal@nxp.com; david.marchand@redhat.com; Vargas,
> >> Hernan
> >>>> <hernan.vargas@intel.com>
> >>>> Subject: Re: [PATCH v3 09/12] baseband/acc: add FFT support to VRB2
> >>>> variant
> >>>>
> >>>>
> >>>>
> >>>> On 9/29/23 18:35, Nicolas Chautru wrote:
> >>>>> Support for the FFT the processing specific to the
> >>>>> VRB2 variant.
> >>>>>
> >>>>> Signed-off-by: Nicolas Chautru <nicolas.chautru@intel.com>
> >>>>> ---
> >>>>> drivers/baseband/acc/rte_vrb_pmd.c | 132
> >>>> ++++++++++++++++++++++++++++-
> >>>>> 1 file changed, 128 insertions(+), 4 deletions(-)
> >>>>>
> >>>>> diff --git a/drivers/baseband/acc/rte_vrb_pmd.c
> >>>>> b/drivers/baseband/acc/rte_vrb_pmd.c
> >>>>> index 93add82947..ce4b90d8e7 100644
> >>>>> --- a/drivers/baseband/acc/rte_vrb_pmd.c
> >>>>> +++ b/drivers/baseband/acc/rte_vrb_pmd.c
> >>>>> @@ -903,6 +903,9 @@ vrb_queue_setup(struct rte_bbdev *dev,
> >>>>> uint16_t
> >>>> queue_id,
> >>>>> ACC_FCW_LD_BLEN : (conf->op_type ==
> >>>> RTE_BBDEV_OP_FFT ?
> >>>>> ACC_FCW_FFT_BLEN :
> ACC_FCW_MLDTS_BLEN))));
> >>>>>
> >>>>> + if ((q->d->device_variant == VRB2_VARIANT) && (conf->op_type ==
> >>>> RTE_BBDEV_OP_FFT))
> >>>>> + fcw_len = ACC_FCW_FFT_BLEN_3;
> >>>>> +
> >>>>> for (desc_idx = 0; desc_idx < d->sw_ring_max_depth;
> desc_idx++) {
> >>>>> desc = q->ring_addr + desc_idx;
> >>>>> desc->req.word0 = ACC_DMA_DESC_TYPE; @@ -
> 1323,6
> >>>> +1326,24 @@
> >>>>> vrb_dev_info_get(struct rte_bbdev *dev, struct
> >>>>> rte_bbdev_driver_info
> >>>> *dev_info)
> >>>>> .num_buffers_soft_out = 0,
> >>>>> }
> >>>>> },
> >>>>> + {
> >>>>> + .type = RTE_BBDEV_OP_FFT,
> >>>>> + .cap.fft = {
> >>>>> + .capability_flags =
> >>>>> +
> >>>> RTE_BBDEV_FFT_WINDOWING |
> >>>>> +
> >>>> RTE_BBDEV_FFT_CS_ADJUSTMENT |
> >>>>> +
> >>>> RTE_BBDEV_FFT_DFT_BYPASS |
> >>>>> +
> >>>> RTE_BBDEV_FFT_IDFT_BYPASS |
> >>>>> + RTE_BBDEV_FFT_FP16_INPUT
> >>>> |
> >>>>> +
> >>>> RTE_BBDEV_FFT_FP16_OUTPUT |
> >>>>> +
> >>>> RTE_BBDEV_FFT_POWER_MEAS |
> >>>>> +
> >>>> RTE_BBDEV_FFT_WINDOWING_BYPASS,
> >>>>> + .num_buffers_src =
> >>>>> + 1,
> >>>>> + .num_buffers_dst =
> >>>>> + 1,
> >>>>> + }
> >>>>> + },
> >>>>> RTE_BBDEV_END_OF_CAPABILITIES_LIST()
> >>>>> };
> >>>>>
> >>>>> @@ -3849,6 +3870,47 @@ vrb1_fcw_fft_fill(struct rte_bbdev_fft_op
> >>>>> *op,
> >>>> struct acc_fcw_fft *fcw)
> >>>>> fcw->bypass = 0;
> >>>>> }
> >>>>>
> >>>>> +/* Fill in a frame control word for FFT processing. */ static
> >>>>> +inline void vrb2_fcw_fft_fill(struct rte_bbdev_fft_op *op, struct
> >>>>> +acc_fcw_fft_3 *fcw) {
> >>>>> + fcw->in_frame_size = op->fft.input_sequence_size;
> >>>>> + fcw->leading_pad_size = op->fft.input_leading_padding;
> >>>>> + fcw->out_frame_size = op->fft.output_sequence_size;
> >>>>> + fcw->leading_depad_size = op->fft.output_leading_depadding;
> >>>>> + fcw->cs_window_sel = op->fft.window_index[0] +
> >>>>> + (op->fft.window_index[1] << 8) +
> >>>>> + (op->fft.window_index[2] << 16) +
> >>>>> + (op->fft.window_index[3] << 24);
> >>>>> + fcw->cs_window_sel2 = op->fft.window_index[4] +
> >>>>> + (op->fft.window_index[5] << 8);
> >>>>> + fcw->cs_enable_bmap = op->fft.cs_bitmap;
> >>>>> + fcw->num_antennas = op->fft.num_antennas_log2;
> >>>>> + fcw->idft_size = op->fft.idft_log2;
> >>>>> + fcw->dft_size = op->fft.dft_log2;
> >>>>> + fcw->cs_offset = op->fft.cs_time_adjustment;
> >>>>> + fcw->idft_shift = op->fft.idft_shift;
> >>>>> + fcw->dft_shift = op->fft.dft_shift;
> >>>>> + fcw->cs_multiplier = op->fft.ncs_reciprocal;
> >>>>> + fcw->power_shift = op->fft.power_shift; > + fcw->exp_adj = op-
> >>>>> fft.fp16_exp_adjust;
> >>>>> + fcw->fp16_in = check_bit(op->fft.op_flags,
> >>>> RTE_BBDEV_FFT_FP16_INPUT);
> >>>>> + fcw->fp16_out = check_bit(op->fft.op_flags,
> >>>> RTE_BBDEV_FFT_FP16_OUTPUT);
> >>>>> + fcw->power_en = check_bit(op->fft.op_flags,
> >>>> RTE_BBDEV_FFT_POWER_MEAS);
> >>>>> + if (check_bit(op->fft.op_flags,
> >>>>> + RTE_BBDEV_FFT_IDFT_BYPASS)) {
> >>>>> + if (check_bit(op->fft.op_flags,
> >>>>> + RTE_BBDEV_FFT_WINDOWING_BYPASS))
> >>>>> + fcw->bypass = 2;
> >>>>> + else
> >>>>> + fcw->bypass = 1;
> >>>>> + } else if (check_bit(op->fft.op_flags,
> >>>>> + RTE_BBDEV_FFT_DFT_BYPASS))
> >>>>> + fcw->bypass = 3;
> >>>>> + else
> >>>>> + fcw->bypass = 0;
> >>>>
> >>>> The only difference I see with VRB1 are backed by corresponding
> >>>> op_flags (POWER & FP16), is that correct? If so, it does not make
> >>>> sense to me to have a specific function for VRB2.
> >>>
> >>> There are more changes but these are only formally enabled in the
> >>> next stepping hence some of the related code is not included yet.
> >>> More generally
> >> the FCW and IP is different from VRB1 implementation.
> >>
> >> Currently, the code is almost identical so vrb1 implementation should
> >> be reused. If some later changes makes the two implementations
> >> diverge, then we can consider having a dedicated function for VRB2 at that
> time.
> >
> > If I may, I believe this is best as-is notably for patches and support.
> > The functions are fairly small (not much code overlap quantitatively)
> > and the underlying IP is different (with more differences we can
> > enable over time). I don’t think it would help anyone really to try to make
> them coexist for a small period of time.
> > Does that sound fair?
>
> I disagree, as I explained the code currently is almost identical, so just share the
> code.
>
> You will diverge, if *really* necessary, when it will make more sense to have
> two separate functions. For now it is not the case in my opinion.
OK I had another look. I can share a common descriptor generation function. For the FCW generation these are just different structures and sizes, different prototype, I really don't think it would make sense to try to artificially generate them together.
Updating in new v5 this week.
>
> Thanks,
> Maxime
>
> >
> >
> >>
> >>>>
> >>>>> +}
> >>>>> +
> >>>>> static inline int
> >>>>> vrb1_dma_desc_fft_fill(struct rte_bbdev_fft_op *op,
> >>>>> struct acc_dma_req_desc *desc, @@ -3882,6
> +3944,58 @@
> >>>>> vrb1_dma_desc_fft_fill(struct rte_bbdev_fft_op
> >>>> *op,
> >>>>> return 0;
> >>>>> }
> >>>>>
> >>>>> +static inline int
> >>>>> +vrb2_dma_desc_fft_fill(struct rte_bbdev_fft_op *op,
> >>>>> + struct acc_dma_req_desc *desc,
> >>>>> + struct rte_mbuf *input, struct rte_mbuf *output, struct
> >>>> rte_mbuf *win_input,
> >>>>> + struct rte_mbuf *pwr, uint32_t *in_offset, uint32_t
> >>>> *out_offset,
> >>>>> + uint32_t *win_offset, uint32_t *pwr_offset) {
> >>>>> + bool pwr_en = check_bit(op->fft.op_flags,
> >>>> RTE_BBDEV_FFT_POWER_MEAS);
> >>>>> + bool win_en = check_bit(op->fft.op_flags,
> >>>> RTE_BBDEV_FFT_DEWINDOWING);
> >>>>> + int num_cs = 0, i, bd_idx = 1;
> >>>>> +
> >>>>> + /* FCW already done */
> >>>>> + acc_header_init(desc);
> >>>>> +
> >>>>> + RTE_SET_USED(win_input);
> >>>>> + RTE_SET_USED(win_offset);
> >>>>> +
> >>>>> + desc->data_ptrs[bd_idx].address = rte_pktmbuf_iova_offset(input,
> >>>> *in_offset);
> >>>>> + desc->data_ptrs[bd_idx].blen = op->fft.input_sequence_size *
> >>>> ACC_IQ_SIZE;
> >>>>> + desc->data_ptrs[bd_idx].blkid = ACC_DMA_BLKID_IN;
> >>>>> + desc->data_ptrs[bd_idx].last = 1;
> >>>>> + desc->data_ptrs[bd_idx].dma_ext = 0;
> >>>>> + bd_idx++;
> >>>>> +
> >>>>> + desc->data_ptrs[bd_idx].address =
> >>>>> +rte_pktmbuf_iova_offset(output,
> >>>> *out_offset);
> >>>>> + desc->data_ptrs[bd_idx].blen = op->fft.output_sequence_size *
> >>>> ACC_IQ_SIZE;
> >>>>> + desc->data_ptrs[bd_idx].blkid = ACC_DMA_BLKID_OUT_HARD;
> >>>>> + desc->data_ptrs[bd_idx].last = pwr_en ? 0 : 1;
> >>>>> + desc->data_ptrs[bd_idx].dma_ext = 0;
> >>>>> + desc->m2dlen = win_en ? 3 : 2;
> >>>>> + desc->d2mlen = pwr_en ? 2 : 1;
> >>>>> + desc->ib_ant_offset = op->fft.input_sequence_size;
> >>>>> + desc->num_ant = op->fft.num_antennas_log2 - 3;
> >>>>> +
> >>>>> + for (i = 0; i < RTE_BBDEV_MAX_CS; i++)
> >>>>> + if (check_bit(op->fft.cs_bitmap, 1 << i))
> >>>>> + num_cs++;
> >>>>> + desc->num_cs = num_cs;
> >>>>> +
> >>>>> + if (pwr_en && pwr) {
> >>>>> + bd_idx++;
> >>>>> + desc->data_ptrs[bd_idx].address =
> >>>> rte_pktmbuf_iova_offset(pwr, *pwr_offset);
> >>>>> + desc->data_ptrs[bd_idx].blen = num_cs * (1 << op-
> >>>>> fft.num_antennas_log2) * 4;
> >>>>> + desc->data_ptrs[bd_idx].blkid = ACC_DMA_BLKID_OUT_SOFT;
> >>>>> + desc->data_ptrs[bd_idx].last = 1;
> >>>>> + desc->data_ptrs[bd_idx].dma_ext = 0;
> >>>>> + }
> >>>>> + desc->ob_cyc_offset = op->fft.output_sequence_size;
> >>>>> + desc->ob_ant_offset = op->fft.output_sequence_size * num_cs;
> >>>>> + desc->op_addr = op;
> >>>>> + return 0;
> >>>>> +}
> >>>>>
> >>>>> /** Enqueue one FFT operation for device. */
> >>>>> static inline int
> >>>>> @@ -3889,22 +4003,32 @@ vrb_enqueue_fft_one_op(struct acc_queue
> >> *q,
> >>>> struct rte_bbdev_fft_op *op,
> >>>>> uint16_t total_enqueued_cbs)
> >>>>> {
> >>>>> union acc_dma_desc *desc;
> >>>>> - struct rte_mbuf *input, *output;
> >>>>> - uint32_t in_offset, out_offset;
> >>>>> + struct rte_mbuf *input, *output, *pwr, *win;
> >>>>> + uint32_t in_offset, out_offset, pwr_offset, win_offset;
> >>>>> struct acc_fcw_fft *fcw;
> >>>>>
> >>>>> desc = acc_desc(q, total_enqueued_cbs);
> >>>>> input = op->fft.base_input.data;
> >>>>> output = op->fft.base_output.data;
> >>>>> + pwr = op->fft.power_meas_output.data;
> >>>>> + win = op->fft.dewindowing_input.data;
> >>>>> in_offset = op->fft.base_input.offset;
> >>>>> out_offset = op->fft.base_output.offset;
> >>>>> + pwr_offset = op->fft.power_meas_output.offset;
> >>>>> + win_offset = op->fft.dewindowing_input.offset;
> >>>>>
> >>>>> fcw = (struct acc_fcw_fft *) (q->fcw_ring +
> >>>>> ((q->sw_ring_head + total_enqueued_cbs) & q-
> >>>>> sw_ring_wrap_mask)
> >>>>> * ACC_MAX_FCW_SIZE);
> >>>>>
> >>>>> - vrb1_fcw_fft_fill(op, fcw);
> >>>>> - vrb1_dma_desc_fft_fill(op, &desc->req, input, output, &in_offset,
> >>>> &out_offset);
> >>>>> + if (q->d->device_variant == VRB1_VARIANT) {
> >>>>> + vrb1_fcw_fft_fill(op, fcw);
> >>>>> + vrb1_dma_desc_fft_fill(op, &desc->req, input, output,
> >>>> &in_offset, &out_offset);
> >>>>> + } else {
> >>>>> + vrb2_fcw_fft_fill(op, (struct acc_fcw_fft_3 *) fcw);
> >>>>> + vrb2_dma_desc_fft_fill(op, &desc->req, input, output, win,
> >>>> pwr,
> >>>>> + &in_offset, &out_offset, &win_offset,
> >>>> &pwr_offset);
> >>>>> + }
> >>>>> #ifdef RTE_LIBRTE_BBDEV_DEBUG
> >>>>> rte_memdump(stderr, "FCW", &desc->req.fcw_fft,
> >>>>> sizeof(desc->req.fcw_fft));
> >>>
> >
next prev parent reply other threads:[~2023-10-05 17:59 UTC|newest]
Thread overview: 42+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-09-29 16:35 [PATCH v3 00/12] VRB2 bbdev PMD introduction Nicolas Chautru
2023-09-29 16:35 ` [PATCH v3 01/12] bbdev: add FFT window width member in driver info Nicolas Chautru
2023-09-29 16:35 ` [PATCH v3 02/12] baseband/acc: add FFT window width in the VRB PMD Nicolas Chautru
2023-10-03 11:52 ` Maxime Coquelin
2023-10-03 19:06 ` Chautru, Nicolas
2023-10-04 7:55 ` Maxime Coquelin
2023-09-29 16:35 ` [PATCH v3 03/12] baseband/acc: remove the 4G SO capability for VRB1 Nicolas Chautru
2023-10-03 12:04 ` Maxime Coquelin
2023-09-29 16:35 ` [PATCH v3 04/12] baseband/acc: allocate FCW memory separately Nicolas Chautru
2023-10-03 12:51 ` Maxime Coquelin
2023-09-29 16:35 ` [PATCH v3 05/12] baseband/acc: add support for MLD operation Nicolas Chautru
2023-09-29 16:35 ` [PATCH v3 06/12] baseband/acc: refactor to allow unified driver extension Nicolas Chautru
2023-10-03 13:14 ` Maxime Coquelin
2023-10-03 18:54 ` Chautru, Nicolas
2023-10-04 7:35 ` Maxime Coquelin
2023-10-04 21:28 ` Chautru, Nicolas
2023-10-05 14:31 ` Maxime Coquelin
2023-10-05 15:00 ` Chautru, Nicolas
2023-09-29 16:35 ` [PATCH v3 07/12] baseband/acc: adding VRB2 device variant Nicolas Chautru
2023-10-03 13:41 ` Maxime Coquelin
2023-09-29 16:35 ` [PATCH v3 08/12] baseband/acc: add FEC capabilities for the VRB2 variant Nicolas Chautru
2023-10-03 14:28 ` Maxime Coquelin
2023-10-04 21:11 ` Chautru, Nicolas
2023-10-05 14:36 ` Maxime Coquelin
2023-09-29 16:35 ` [PATCH v3 09/12] baseband/acc: add FFT support to " Nicolas Chautru
2023-10-03 14:36 ` Maxime Coquelin
2023-10-03 18:20 ` Chautru, Nicolas
2023-10-04 7:11 ` Maxime Coquelin
2023-10-04 21:18 ` Chautru, Nicolas
2023-10-05 14:34 ` Maxime Coquelin
2023-10-05 17:59 ` Chautru, Nicolas [this message]
2023-10-06 12:05 ` Maxime Coquelin
2023-10-06 20:25 ` Chautru, Nicolas
2023-09-29 16:35 ` [PATCH v3 10/12] baseband/acc: add MLD support in " Nicolas Chautru
2023-10-03 15:12 ` Maxime Coquelin
2023-10-03 18:12 ` Chautru, Nicolas
2023-09-29 16:35 ` [PATCH v3 11/12] baseband/acc: add support for VRB2 engine error detection Nicolas Chautru
2023-10-03 15:16 ` Maxime Coquelin
2023-10-03 17:22 ` Chautru, Nicolas
2023-10-03 17:26 ` Maxime Coquelin
2023-09-29 16:35 ` [PATCH v3 12/12] baseband/acc: add configure helper for VRB2 Nicolas Chautru
2023-10-03 15:30 ` Maxime Coquelin
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=BY5PR11MB445123214192EC8A9A5031F3F8CAA@BY5PR11MB4451.namprd11.prod.outlook.com \
--to=nicolas.chautru@intel.com \
--cc=david.marchand@redhat.com \
--cc=dev@dpdk.org \
--cc=hemant.agrawal@nxp.com \
--cc=hernan.vargas@intel.com \
--cc=maxime.coquelin@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
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).