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

Neil Armstrong neil.armstrong at linaro.org
Mon Dec 2 08:46:25 UTC 2024


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 ?

> 
>> +
>> +	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