[PATCH v3 2/7] drm/msm: adreno: add plumbing to generate bandwidth vote table for GMU

Akhil P Oommen quic_akhilpo at quicinc.com
Wed Dec 4 19:15:25 UTC 2024


On 12/4/2024 9:05 PM, Neil Armstrong wrote:
> On 02/12/2024 09:46, Neil Armstrong wrote:
>> On 30/11/2024 22:49, Akhil P Oommen wrote:
>>> On 11/28/2024 3:55 PM, Neil Armstrong wrote:
>>>> The Adreno GPU Management Unit (GMU) can also scale DDR Bandwidth along
>>>> the Frequency and Power Domain level, but by default we leave the
>>>> OPP core scale the interconnect ddr path.
>>>>
>>>> While scaling via the interconnect path was sufficient, newer GPUs
>>>> like the A750 requires specific vote paremeters and bandwidth to
>>>> achieve full functionality.
>>>>
>>>> In order to calculate vote values used by the GPU Management
>>>> Unit (GMU), we need to parse all the possible OPP Bandwidths and
>>>> create a vote value to be sent to the appropriate Bus Control
>>>> Modules (BCMs) declared in the GPU info struct.
>>>>
>>>> This vote value is called IB, while on the othe side the GMU also
>>>> takes another vote called AB which is a 16bit quantized value
>>>> of the bandwidth against the maximum supported bandwidth.
>>>>
>>>> The vote array will then be used to dynamically generate the GMU
>>>> bw_table sent during the GMU power-up.
>>>>
>>>> Signed-off-by: Neil Armstrong <neil.armstrong at linaro.org>
>>>> ---
>>>>   drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 174 ++++++++++++++++++++++
>>>> ++++++++++++
>>>>   drivers/gpu/drm/msm/adreno/a6xx_gmu.h |  14 +++
>>>>   drivers/gpu/drm/msm/adreno/a6xx_gpu.h |   1 +
>>>>   drivers/gpu/drm/msm/adreno/a6xx_hfi.h |   5 +
>>>>   4 files changed, 194 insertions(+)
>>>>
>>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/
>>>> drm/msm/adreno/a6xx_gmu.c
>>>> index
>>>> 14db7376c712d19446b38152e480bd5a1e0a5198..ee2010a01186721dd377f1655fcf05ddaff77131 100644
>>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>>>> @@ -9,6 +9,7 @@
>>>>   #include <linux/pm_domain.h>
>>>>   #include <linux/pm_opp.h>
>>>>   #include <soc/qcom/cmd-db.h>
>>>> +#include <soc/qcom/tcs.h>
>>>>   #include <drm/drm_gem.h>
>>>>   #include "a6xx_gpu.h"
>>>> @@ -1287,6 +1288,131 @@ static int a6xx_gmu_memory_probe(struct
>>>> a6xx_gmu *gmu)
>>>>       return 0;
>>>>   }
>>>> +/**
>>>> + * struct bcm_db - Auxiliary data pertaining to each Bus Clock
>>>> Manager (BCM)
>>>> + * @unit: divisor used to convert bytes/sec bw value to an RPMh msg
>>>> + * @width: multiplier used to convert bytes/sec bw value to an RPMh
>>>> msg
>>>> + * @vcd: virtual clock domain that this bcm belongs to
>>>> + * @reserved: reserved field
>>>> + */
>>>> +struct bcm_db {
>>>> +    __le32 unit;
>>>> +    __le16 width;
>>>> +    u8 vcd;
>>>> +    u8 reserved;
>>>> +};
>>>> +
>>>> +static u64 bcm_div(u64 num, u32 base)
>>>> +{
>>>> +    /* Ensure that small votes aren't lost. */
>>>> +    if (num && num < base)
>>>> +        return 1;
>>>> +
>>>> +    do_div(num, base);
>>>> +
>>>> +    return num;
>>>> +}
>>>> +
>>>> +static int a6xx_gmu_rpmh_bw_votes_init(const struct a6xx_info *info,
>>>> +                       struct a6xx_gmu *gmu)
>>>> +{
>>>> +    const struct bcm_db *bcm_data[GMU_MAX_BCMS] = { 0 };
>>>> +    unsigned int bcm_index, bw_index, bcm_count = 0;
>>>> +
>>>> +    if (!info->bcms)
>>>> +        return 0;
>>>> +
>>>> +    /* Retrieve BCM data from cmd-db */
>>>> +    for (bcm_index = 0; bcm_index < GMU_MAX_BCMS; bcm_index++) {
>>>> +        size_t count;
>>>> +
>>>> +        /* Stop at first unconfigured bcm */
>>>> +        if (!info->bcms[bcm_index].name)
>>>> +            break;
>>>> +
>>>> +        bcm_data[bcm_index] = cmd_db_read_aux_data(
>>>> +                        info->bcms[bcm_index].name,
>>>> +                        &count);
>>>> +        if (IS_ERR(bcm_data[bcm_index]))
>>>> +            return PTR_ERR(bcm_data[bcm_index]);
>>>> +
>>>> +        if (!count)
>>>> +            return -EINVAL;
>>>> +
>>>> +        ++bcm_count;
>>>> +    }
>>>> +
>>>> +    /* Generate BCM votes values for each bandwidth & BCM */
>>>> +    for (bw_index = 0; bw_index < gmu->nr_gpu_bws; bw_index++) {
>>>> +        u32 *data = gmu->gpu_ib_votes[bw_index];
>>>> +        u32 bw = gmu->gpu_bw_table[bw_index];
>>>> +
>>>> +        /* Calculations loosely copied from bcm_aggregate() &
>>>> tcs_cmd_gen() */
>>>> +        for (bcm_index = 0; bcm_index < bcm_count; bcm_index++) {
>>>> +            bool commit = false;
>>>> +            u64 peak, vote;
>>>> +            u16 width;
>>>> +            u32 unit;
>>>> +
>>>> +            /* Skip unconfigured BCM */
>>>> +            if (!bcm_data[bcm_index])
>>>> +                continue;
>>>> +
>>>> +            if (bcm_index == bcm_count - 1 ||
>>>> +                (bcm_data[bcm_index + 1] &&
>>>> +                 bcm_data[bcm_index]->vcd != bcm_data[bcm_index +
>>>> 1]->vcd))
>>>> +                commit = true;
>>>> +
>>>> +            if (!bw) {
>>>> +                data[bcm_index] = BCM_TCS_CMD(commit, false, 0, 0);
>>>> +                continue;
>>>> +            }
>>>> +
>>>> +            if (info->bcms[bcm_index].fixed) {
>>>> +                u32 perfmode = 0;
>>>> +
>>>> +                if (bw >= info->bcms[bcm_index].perfmode_bw)
>>>> +                    perfmode = info->bcms[bcm_index].perfmode;
>>>> +
>>>> +                data[bcm_index] = BCM_TCS_CMD(commit, true, 0,
>>>> perfmode);
>>>> +                continue;
>>>> +            }
>>>> +
>>>> +            /* Multiply the bandwidth by the width of the
>>>> connection */
>>>> +            width = le16_to_cpu(bcm_data[bcm_index]->width);
>>>> +            peak = bcm_div((u64)bw * width, info-
>>>> >bcms[bcm_index].buswidth);
>>>> +
>>>> +            /* Input bandwidth value is in KBps, scale the value to
>>>> BCM unit */
>>>> +            unit = le32_to_cpu(bcm_data[bcm_index]->unit);
>>>> +            vote = bcm_div(peak * 1000ULL, unit);
>>>> +
>>>> +            if (vote > BCM_TCS_CMD_VOTE_MASK)
>>>> +                vote = BCM_TCS_CMD_VOTE_MASK;
>>>
>>> use clamp()?
>>
>> Yep, I think I could replace bcm_div with clamp
>>
>>>
>>>> +
>>>> +            data[bcm_index] = BCM_TCS_CMD(commit, true, vote, vote);
>>>> +        }
>>>> +    }
>>>> +
>>>> +    /* Generate AB votes which are a quantitized bandwidth value */
>>>> +    for (bw_index = 0; bw_index < gmu->nr_gpu_bws; bw_index++) {
>>>> +        u64 tmp;
>>>> +
>>>> +        /*
>>>> +         * The AB vote consists of a 16 bit wide quantized level
>>>> +         * against the maximum supported bandwidth.
>>>> +         * Quantization can be calculated as below:
>>>> +         * vote = (bandwidth * 2^16) / max bandwidth
>>>> +         */
>>>> +        tmp = gmu->gpu_bw_table[bw_index] * MAX_AB_VOTE;
>>>> +
>>>> +        /* Divide by the maximum bandwidth to get a quantized value */
>>>> +        gmu->gpu_ab_votes[bw_index] =
>>>> +            bcm_div(tmp, gmu->gpu_bw_table[gmu->nr_gpu_bws - 1]);
>>>> +    }
>>>
>>> So I suppose you are trying to vote AB equal to IB. Aggregation logic
>>> for both are different. So this will make DDR scale very aggressively. A
>>> more reasonable approach would be to vote a % of IB vote (25%?). Ideally
>>> we should measure GPU's bandwidth usage and vote that if you are really
>>> concerned about stability issues. IB is just a floor vote, GPU can
>>> generate way higher traffic.
>>
>> I think this should be optimized better in a different patchset, so I
>> would
>> like to make the simplest vote possible here to retain functionnality.
>>
>> So if I understand I should divide this vote value by 4 ? Downstram
>> uses 25% by default
>> when no AB was calculated, but what does that mean exactly ?
>>
>> Is there a counter somewhere to measure the GPU traffic ?
> 
> 
> What if we also specified opp-avg-kBps for each OPP, define it to the
> downstream "qcom,bus-min"
> and use this value for the AB vote ? and use 25% as fallback like
> downstream.

Fixed AB vote for each opp is a temporary measure. So I feel that we
should not mess with DT for this.

-Akhil.

> 
> Neil
> 
>>
>>>
>>>> +
>>>> +    return 0;
>>>> +}
>>>> +
>>>>   /* Return the 'arc-level' for the given frequency */
>>>>   static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
>>>>                          unsigned long freq)
>>>> @@ -1390,12 +1516,15 @@ static int
>>>> a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
>>>>    * The GMU votes with the RPMh for itself and on behalf of the GPU
>>>> but we need
>>>>    * to construct the list of votes on the CPU and send it over.
>>>> Query the RPMh
>>>>    * voltage levels and build the votes
>>>> + * The GMU can also vote for DDR interconnects, use the OPP
>>>> bandwidth entries
>>>> + * and BCM parameters to build the votes.
>>>>    */
>>>>   static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
>>>>   {
>>>>       struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu,
>>>> gmu);
>>>>       struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
>>>> +    const struct a6xx_info *info = adreno_gpu->info->a6xx;
>>>>       struct msm_gpu *gpu = &adreno_gpu->base;
>>>>       int ret;
>>>> @@ -1407,6 +1536,10 @@ static int a6xx_gmu_rpmh_votes_init(struct
>>>> a6xx_gmu *gmu)
>>>>       ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
>>>>           gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
>>>> +    /* Build the interconnect votes */
>>>> +    if (info->bcms && gmu->nr_gpu_bws > 1)
>>>> +        ret |= a6xx_gmu_rpmh_bw_votes_init(info, gmu);
>>>> +
>>>>       return ret;
>>>>   }
>>>> @@ -1442,10 +1575,43 @@ static int a6xx_gmu_build_freq_table(struct
>>>> device *dev, unsigned long *freqs,
>>>>       return index;
>>>>   }
>>>> +static int a6xx_gmu_build_bw_table(struct device *dev, unsigned
>>>> long *bandwidths,
>>>> +        u32 size)
>>>> +{
>>>> +    int count = dev_pm_opp_get_opp_count(dev);
>>>
>>> I am less concerned about this now since you are not voting real AB BW.
>>
>> Sorry I don't understand
>>
>> Thanks,
>> Neil
>>
>>>
>>> -Akhil.
>>>
>>>> +    struct dev_pm_opp *opp;
>>>> +    int i, index = 0;
>>>> +    unsigned int bandwidth = 1;
>>>> +
>>>> +    /*
>>>> +     * The OPP table doesn't contain the "off" bandwidth level so
>>>> we need to
>>>> +     * add 1 to the table size to account for it
>>>> +     */
>>>> +
>>>> +    if (WARN(count + 1 > size,
>>>> +        "The GMU bandwidth table is being truncated\n"))
>>>> +        count = size - 1;
>>>> +
>>>> +    /* Set the "off" bandwidth */
>>>> +    bandwidths[index++] = 0;
>>>> +
>>>> +    for (i = 0; i < count; i++) {
>>>> +        opp = dev_pm_opp_find_bw_ceil(dev, &bandwidth, 0);
>>>> +        if (IS_ERR(opp))
>>>> +            break;
>>>> +
>>>> +        dev_pm_opp_put(opp);
>>>> +        bandwidths[index++] = bandwidth++;
>>>> +    }
>>>> +
>>>> +    return index;
>>>> +}
>>>> +
>>>>   static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
>>>>   {
>>>>       struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu,
>>>> gmu);
>>>>       struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
>>>> +    const struct a6xx_info *info = adreno_gpu->info->a6xx;
>>>>       struct msm_gpu *gpu = &adreno_gpu->base;
>>>>       int ret = 0;
>>>> @@ -1472,6 +1638,14 @@ static int a6xx_gmu_pwrlevels_probe(struct
>>>> a6xx_gmu *gmu)
>>>>       gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
>>>> +    /*
>>>> +     * The GMU also handles GPU Interconnect Votes so build a list
>>>> +     * of DDR bandwidths from the GPU OPP table
>>>> +     */
>>>> +    if (info->bcms)
>>>> +        gmu->nr_gpu_bws = a6xx_gmu_build_bw_table(&gpu->pdev->dev,
>>>> +            gmu->gpu_bw_table, ARRAY_SIZE(gmu->gpu_bw_table));
>>>> +
>>>>       /* Build the list of RPMh votes that we'll send to the GMU */
>>>>       return a6xx_gmu_rpmh_votes_init(gmu);
>>>>   }
>>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h b/drivers/gpu/
>>>> drm/msm/adreno/a6xx_gmu.h
>>>> index
>>>> 88f18ea6a38a08b5b171709e5020010947a5d347..bdfc106cb3a578c90d7cd84f7d4fe228d761a994 100644
>>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
>>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.h
>>>> @@ -21,6 +21,15 @@ struct a6xx_gmu_bo {
>>>>   #define GMU_MAX_GX_FREQS    16
>>>>   #define GMU_MAX_CX_FREQS    4
>>>> +#define GMU_MAX_BCMS        3
>>>> +
>>>> +struct a6xx_bcm {
>>>> +    char *name;
>>>> +    unsigned int buswidth;
>>>> +    bool fixed;
>>>> +    unsigned int perfmode;
>>>> +    unsigned int perfmode_bw;
>>>> +};
>>>>   /*
>>>>    * These define the different GMU wake up options - these define
>>>> how both the
>>>> @@ -85,6 +94,11 @@ struct a6xx_gmu {
>>>>       unsigned long gpu_freqs[GMU_MAX_GX_FREQS];
>>>>       u32 gx_arc_votes[GMU_MAX_GX_FREQS];
>>>> +    int nr_gpu_bws;
>>>> +    unsigned long gpu_bw_table[GMU_MAX_GX_FREQS];
>>>> +    u32 gpu_ib_votes[GMU_MAX_GX_FREQS][GMU_MAX_BCMS];
>>>> +    u16 gpu_ab_votes[GMU_MAX_GX_FREQS];
>>>> +
>>>>       int nr_gmu_freqs;
>>>>       unsigned long gmu_freqs[GMU_MAX_CX_FREQS];
>>>>       u32 cx_arc_votes[GMU_MAX_CX_FREQS];
>>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h b/drivers/gpu/
>>>> drm/msm/adreno/a6xx_gpu.h
>>>> index
>>>> 4aceffb6aae89c781facc2a6e4a82b20b341b6cb..9201a53dd341bf432923ffb44947e015208a3d02 100644
>>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gpu.h
>>>> @@ -44,6 +44,7 @@ struct a6xx_info {
>>>>       u32 gmu_chipid;
>>>>       u32 gmu_cgc_mode;
>>>>       u32 prim_fifo_threshold;
>>>> +    const struct a6xx_bcm *bcms;
>>>>   };
>>>>   struct a6xx_gpu {
>>>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_hfi.h b/drivers/gpu/
>>>> drm/msm/adreno/a6xx_hfi.h
>>>> index
>>>> 528110169398f69f16443a29a1594d19c36fb595..52ba4a07d7b9a709289acd244a751ace9bdaab5d 100644
>>>> --- a/drivers/gpu/drm/msm/adreno/a6xx_hfi.h
>>>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_hfi.h
>>>> @@ -173,6 +173,11 @@ struct a6xx_hfi_gx_bw_perf_vote_cmd {
>>>>       u32 bw;
>>>>   };
>>>> +#define AB_VOTE_MASK        GENMASK(31, 16)
>>>> +#define MAX_AB_VOTE        (FIELD_MAX(AB_VOTE_MASK) - 1)
>>>> +#define AB_VOTE(vote)        FIELD_PREP(AB_VOTE_MASK, (vote))
>>>> +#define AB_VOTE_ENABLE        BIT(8)
>>>> +
>>>>   #define HFI_H2F_MSG_PREPARE_SLUMBER 33
>>>>   struct a6xx_hfi_prep_slumber_cmd {
>>>>
>>>
>>
> 



More information about the Freedreno mailing list