by Aniket Masule

[permalink] [raw]

Subject: Re: [PATCH 5/5] media: venus: Update core selection

On 2019-06-17 14:37, Stanimir Varbanov wrote:
> Hi Aniket,
>
> On 6/11/19 9:05 AM, Aniket Masule wrote:
>> Present core assignment is static. Introduced load balancing
>> across the cores. Load on earch core is calculated and core
>> with minimum load is assigned to given instance.
>>
>> Signed-off-by: Aniket Masule <[email protected]>
>> ---
>> drivers/media/platform/qcom/venus/helpers.c | 50
>> +++++++++++++++++++++++++----
>> drivers/media/platform/qcom/venus/helpers.h | 2 +-
>> drivers/media/platform/qcom/venus/vdec.c | 5 +--
>> drivers/media/platform/qcom/venus/venc.c | 4 ++-
>> 4 files changed, 51 insertions(+), 10 deletions(-)
>>
>> diff --git a/drivers/media/platform/qcom/venus/helpers.c
>> b/drivers/media/platform/qcom/venus/helpers.c
>> index edb653e..38d617b 100644
>> --- a/drivers/media/platform/qcom/venus/helpers.c
>> +++ b/drivers/media/platform/qcom/venus/helpers.c
>> @@ -497,6 +497,16 @@ static int load_scale_clocks(struct venus_inst
>> *inst)
>> return scale_clocks_vpu4(inst);
>> }
>>
>> +int set_core_usage(struct venus_inst *inst, u32 usage)
>> +{
>> + const u32 ptype = HFI_PROPERTY_CONFIG_VIDEOCORES_USAGE;
>> + struct hfi_videocores_usage_type cu;
>> +
>> + cu.video_core_enable_mask = usage;
>> +
>> + return hfi_session_set_property(inst, ptype, &cu);
>> +}
>> +
>> static void fill_buffer_desc(const struct venus_buffer *buf,
>> struct hfi_buffer_desc *bd, bool response)
>> {
>> @@ -800,19 +810,47 @@ int venus_helper_set_work_mode(struct venus_inst
>> *inst, u32 mode)
>> }
>> EXPORT_SYMBOL_GPL(venus_helper_set_work_mode);
>>
>> -int venus_helper_set_core_usage(struct venus_inst *inst, u32 usage)
>> +int venus_helper_decide_core(struct venus_inst *inst, u32 cores_max)
>
> I think venus_helper_set_core is better?
>
Sure Stan.
>> {
>> - const u32 ptype = HFI_PROPERTY_CONFIG_VIDEOCORES_USAGE;
>> - struct hfi_videocores_usage_type cu;
>> + struct venus_core *core = inst->core;
>> + u32 min_core_id = 0, core0_load = 0, core1_load = 0;
>> + unsigned long min_load, max_freq, cur_inst_load;
>> + int ret;
>>
>> if (!IS_V4(inst->core))
>> return 0;
>>
>> - cu.video_core_enable_mask = usage;
>> + core0_load = load_per_core(core, VIDC_CORE_ID_1);
>> + core1_load = load_per_core(core, VIDC_CORE_ID_2);
>>
>> - return hfi_session_set_property(inst, ptype, &cu);
>> + min_core_id = core0_load < core1_load ? VIDC_CORE_ID_1 :
>> VIDC_CORE_ID_2;
>> + min_load = min(core0_load, core1_load);
>> +
>> + if (cores_max < VIDC_CORE_ID_1) {
>> + min_core_id = VIDC_CORE_ID_1;
>> + min_load = core0_load;
>> + }
>
> could you please move that fragment just after IS_V4 check and return
> an
> error if cores_max < VIDC_CORE_ID_1.
>
Instead of "if cores_max < VIDC_CORE_ID_1", we need to check if
cores_max < VIDC_CORE_ID_2
and set core the single core as minimum load core. I can't return after
this check immidiately
as it needs to be checked whether load can be accommodated or not.
>> +
>> + cur_inst_load = load_per_instance(inst) *
>> + inst->clk_data.codec_data->vpp_cycles;
>> + max_freq = core->res->freq_tbl[0].freq;
>> +
>> + if ((cur_inst_load + min_load) > max_freq) {
>> + dev_warn(core->dev, "HW is overloaded, needed: %lu max: %lu\n",
>> + cur_inst_load, max_freq);
>> + return -EINVAL;
>> + }
>> +
>> + ret = set_core_usage(inst, min_core_id);
>> +
>> + if (ret)
>> + return ret;
>> +
>> + inst->clk_data.core_id = min_core_id;
>> +
>> + return 0;
>> }
>> -EXPORT_SYMBOL_GPL(venus_helper_set_core_usage);
>> +EXPORT_SYMBOL_GPL(venus_helper_decide_core);
>>
>> int venus_helper_init_codec_data(struct venus_inst *inst)
>> {
>> diff --git a/drivers/media/platform/qcom/venus/helpers.h
>> b/drivers/media/platform/qcom/venus/helpers.h
>> index f9360a8..c41ceb3 100644
>> --- a/drivers/media/platform/qcom/venus/helpers.h
>> +++ b/drivers/media/platform/qcom/venus/helpers.h
>> @@ -42,7 +42,7 @@ int venus_helper_set_output_resolution(struct
>> venus_inst *inst,
>> u32 buftype);
>> int venus_helper_set_work_mode(struct venus_inst *inst, u32 mode);
>> int venus_helper_init_codec_data(struct venus_inst *inst);
>> -int venus_helper_set_core_usage(struct venus_inst *inst, u32 usage);
>> +int venus_helper_decide_core(struct venus_inst *inst, u32 cores_max);
>> int venus_helper_set_num_bufs(struct venus_inst *inst, unsigned int
>> input_bufs,
>> unsigned int output_bufs,
>> unsigned int output2_bufs);
>> diff --git a/drivers/media/platform/qcom/venus/vdec.c
>> b/drivers/media/platform/qcom/venus/vdec.c
>> index 51795fd..9f988ba 100644
>> --- a/drivers/media/platform/qcom/venus/vdec.c
>> +++ b/drivers/media/platform/qcom/venus/vdec.c
>> @@ -544,14 +544,15 @@ static int vdec_output_conf(struct venus_inst
>> *inst)
>> u32 height = inst->out_height;
>> u32 out_fmt, out2_fmt;
>> bool ubwc = false;
>> - u32 ptype;
>> + u32 ptype, cores_max;
>> int ret;
>>
>> ret = venus_helper_set_work_mode(inst, VIDC_WORK_MODE_2);
>> if (ret)
>> return ret;
>>
>> - ret = venus_helper_set_core_usage(inst, VIDC_CORE_ID_1);
>> + cores_max = core_num_max(inst);
>
> please move core_max calculation in the venus_helper_decide_core() here
> and below.
>
Yes Stan.
>> + ret = venus_helper_decide_core(inst, cores_max);
>> if (ret)
>> return ret;
>>
>> diff --git a/drivers/media/platform/qcom/venus/venc.c
>> b/drivers/media/platform/qcom/venus/venc.c
>> index 792cdce..ed39efd 100644
>> --- a/drivers/media/platform/qcom/venus/venc.c
>> +++ b/drivers/media/platform/qcom/venus/venc.c
>> @@ -654,13 +654,15 @@ static int venc_set_properties(struct venus_inst
>> *inst)
>> struct hfi_quantization quant;
>> struct hfi_quantization_range quant_range;
>> u32 ptype, rate_control, bitrate, profile = 0, level = 0;
>> + u32 cores_max;
>> int ret;
>>
>> ret = venus_helper_set_work_mode(inst, VIDC_WORK_MODE_2);
>> if (ret)
>> return ret;
>>
>> - ret = venus_helper_set_core_usage(inst, VIDC_CORE_ID_2);
>> + cores_max = core_num_max(inst);
>> + ret = venus_helper_decide_core(inst, cores_max);
>> if (ret)
>> return ret;
>>
>>