| Message ID | 1563786452-22188-2-git-send-email-amasule@codeaurora.org (mailing list archive) |
|---|---|
| State | Not Applicable, archived |
| Delegated to: | Andy Gross |
| Headers | show |
| Series | media: venus: Update clock scaling and core selection | expand |
Hi Aniket, On 7/22/19 12:07 PM, Aniket Masule wrote: > Current clock scaling calculations are same for vpu4 and > previous versions. For vpu4, Clock scaling calculations > are updated with cycles/mb. This helps in getting precise > clock required. > > Signed-off-by: Aniket Masule <amasule@codeaurora.org> > --- > drivers/media/platform/qcom/venus/helpers.c | 91 +++++++++++++++++++++++++++-- > 1 file changed, 87 insertions(+), 4 deletions(-) > > diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c > index 7492373..2c976e4 100644 > --- a/drivers/media/platform/qcom/venus/helpers.c > +++ b/drivers/media/platform/qcom/venus/helpers.c > @@ -348,8 +348,9 @@ static u32 load_per_type(struct venus_core *core, u32 session_type) > return mbs_per_sec; > } > > -static int load_scale_clocks(struct venus_core *core) > +static int scale_clocks(struct venus_inst *inst) > { > + struct venus_core *core = inst->core; > const struct freq_tbl *table = core->res->freq_tbl; > unsigned int num_rows = core->res->freq_tbl_size; > unsigned long freq = table[0].freq; > @@ -398,6 +399,89 @@ static int load_scale_clocks(struct venus_core *core) > return ret; > } > > +static unsigned long calculate_vpp_freq(struct venus_inst *inst) > +{ > + unsigned long vpp_freq = 0; > + u32 mbs_per_sec; > + > + mbs_per_sec = load_per_instance(inst); > + vpp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vpp_freq; > + /* 21 / 20 is overhead factor */ > + vpp_freq += vpp_freq / 20; > + > + return vpp_freq; > +} > + > +static int scale_clocks_v4(struct venus_inst *inst) > +{ > + struct venus_core *core = inst->core; > + const struct freq_tbl *table = core->res->freq_tbl; > + unsigned int num_rows = core->res->freq_tbl_size; > + struct clk *clk = core->clks[0]; > + struct device *dev = core->dev; > + unsigned int i; > + unsigned long freq = 0, freq_core1 = 0, freq_core2 = 0; > + int ret; > + > + freq = calculate_vpp_freq(inst); > + > + if (freq > table[0].freq) > + dev_warn(dev, "HW is overloaded, needed: %lu max: %lu\n", > + freq, table[0].freq); > + ... > + for (i = 0; i < num_rows; i++) { > + if (freq > table[i].freq) > + break; > + freq = table[i].freq; > + } The above code snippet will select the biggest table[0].freq always. Infact do we need to "normalize" the calculated freq to the table of possible clock rates? I think tjat should be made after sum all needed frequencies for all cores. > + > + inst->clk_data.freq = freq; > + > + mutex_lock(&core->lock); > + list_for_each_entry(inst, &core->instances, list) { > + if (inst->clk_data.core_id == VIDC_CORE_ID_1) { > + freq_core1 += inst->clk_data.freq; > + } else if (inst->clk_data.core_id == VIDC_CORE_ID_2) { > + freq_core2 += inst->clk_data.freq; > + } else if (inst->clk_data.core_id == VIDC_CORE_ID_3) { > + freq_core1 += inst->clk_data.freq; > + freq_core2 += inst->clk_data.freq; > + } > + } > + mutex_unlock(&core->lock); > + > + freq = max(freq_core1, freq_core2); > + > + ret = clk_set_rate(clk, freq); > + if (ret) > + goto err; > + > + ret = clk_set_rate(core->core0_clk, freq); > + if (ret) > + goto err; > + > + ret = clk_set_rate(core->core1_clk, freq); > + if (ret) > + goto err; This is duplicated in both scale_clocks and scale_clocks_v4, and could be a common function. > + > + return 0; > + > +err: > + dev_err(dev, "failed to set clock rate %lu (%d)\n", freq, ret); > + return ret; > +} > + > +static int load_scale_clocks(struct venus_inst *inst) > +{ > + if (IS_V4(inst->core)) > + return scale_clocks_v4(inst); > + > + if (inst->state == INST_START) > + return 0; I'm still not sure about this check of the instance state. If we look into load_per_instance() it already doing similar check : !(inst->state >= INST_INIT && inst->state < INST_STOP) > + > + return scale_clocks(inst); > +} > + > static void fill_buffer_desc(const struct venus_buffer *buf, > struct hfi_buffer_desc *bd, bool response) > { > @@ -1053,7 +1137,7 @@ void venus_helper_vb2_stop_streaming(struct vb2_queue *q) > > venus_helper_free_dpb_bufs(inst); > > - load_scale_clocks(core); > + load_scale_clocks(inst); > INIT_LIST_HEAD(&inst->registeredbufs); > } > > @@ -1070,7 +1154,6 @@ void venus_helper_vb2_stop_streaming(struct vb2_queue *q) > > int venus_helper_vb2_start_streaming(struct venus_inst *inst) > { > - struct venus_core *core = inst->core; > int ret; > > ret = intbufs_alloc(inst); > @@ -1081,7 +1164,7 @@ int venus_helper_vb2_start_streaming(struct venus_inst *inst) > if (ret) > goto err_bufs_free; > > - load_scale_clocks(core); > + load_scale_clocks(inst); > > ret = hfi_session_load_res(inst); > if (ret) >
Hi Stan, On 2019-07-29 14:05, Stanimir Varbanov wrote: > Hi Aniket, > > On 7/22/19 12:07 PM, Aniket Masule wrote: >> Current clock scaling calculations are same for vpu4 and >> previous versions. For vpu4, Clock scaling calculations >> are updated with cycles/mb. This helps in getting precise >> clock required. >> >> Signed-off-by: Aniket Masule <amasule@codeaurora.org> >> --- >> drivers/media/platform/qcom/venus/helpers.c | 91 >> +++++++++++++++++++++++++++-- >> 1 file changed, 87 insertions(+), 4 deletions(-) >> >> diff --git a/drivers/media/platform/qcom/venus/helpers.c >> b/drivers/media/platform/qcom/venus/helpers.c >> index 7492373..2c976e4 100644 >> --- a/drivers/media/platform/qcom/venus/helpers.c >> +++ b/drivers/media/platform/qcom/venus/helpers.c >> @@ -348,8 +348,9 @@ static u32 load_per_type(struct venus_core *core, >> u32 session_type) >> return mbs_per_sec; >> } >> >> -static int load_scale_clocks(struct venus_core *core) >> +static int scale_clocks(struct venus_inst *inst) >> { >> + struct venus_core *core = inst->core; >> const struct freq_tbl *table = core->res->freq_tbl; >> unsigned int num_rows = core->res->freq_tbl_size; >> unsigned long freq = table[0].freq; >> @@ -398,6 +399,89 @@ static int load_scale_clocks(struct venus_core >> *core) >> return ret; >> } >> >> +static unsigned long calculate_vpp_freq(struct venus_inst *inst) >> +{ >> + unsigned long vpp_freq = 0; >> + u32 mbs_per_sec; >> + >> + mbs_per_sec = load_per_instance(inst); >> + vpp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vpp_freq; >> + /* 21 / 20 is overhead factor */ >> + vpp_freq += vpp_freq / 20; >> + >> + return vpp_freq; >> +} >> + >> +static int scale_clocks_v4(struct venus_inst *inst) >> +{ >> + struct venus_core *core = inst->core; >> + const struct freq_tbl *table = core->res->freq_tbl; >> + unsigned int num_rows = core->res->freq_tbl_size; >> + struct clk *clk = core->clks[0]; >> + struct device *dev = core->dev; >> + unsigned int i; >> + unsigned long freq = 0, freq_core1 = 0, freq_core2 = 0; >> + int ret; >> + >> + freq = calculate_vpp_freq(inst); >> + >> + if (freq > table[0].freq) >> + dev_warn(dev, "HW is overloaded, needed: %lu max: %lu\n", >> + freq, table[0].freq); >> + > > ... > >> + for (i = 0; i < num_rows; i++) { >> + if (freq > table[i].freq) >> + break; >> + freq = table[i].freq; >> + } > > The above code snippet will select the biggest table[0].freq always. > Infact do we need to "normalize" the calculated freq to the table of > possible clock rates? I think tjat should be made after sum all needed > frequencies for all cores. > Will correct this. Normalizing after sum of all needed frequencies would be more precise. >> + >> + inst->clk_data.freq = freq; >> + >> + mutex_lock(&core->lock); >> + list_for_each_entry(inst, &core->instances, list) { >> + if (inst->clk_data.core_id == VIDC_CORE_ID_1) { >> + freq_core1 += inst->clk_data.freq; >> + } else if (inst->clk_data.core_id == VIDC_CORE_ID_2) { >> + freq_core2 += inst->clk_data.freq; >> + } else if (inst->clk_data.core_id == VIDC_CORE_ID_3) { >> + freq_core1 += inst->clk_data.freq; >> + freq_core2 += inst->clk_data.freq; >> + } >> + } >> + mutex_unlock(&core->lock); >> + >> + freq = max(freq_core1, freq_core2); >> + >> + ret = clk_set_rate(clk, freq); >> + if (ret) >> + goto err; >> + >> + ret = clk_set_rate(core->core0_clk, freq); >> + if (ret) >> + goto err; >> + >> + ret = clk_set_rate(core->core1_clk, freq); >> + if (ret) >> + goto err; > > This is duplicated in both scale_clocks and scale_clocks_v4, and could > be a common function. > I will move this to common function. >> + >> + return 0; >> + >> +err: >> + dev_err(dev, "failed to set clock rate %lu (%d)\n", freq, ret); >> + return ret; >> +} >> + >> +static int load_scale_clocks(struct venus_inst *inst) >> +{ >> + if (IS_V4(inst->core)) >> + return scale_clocks_v4(inst); >> + >> + if (inst->state == INST_START) >> + return 0; > > I'm still not sure about this check of the instance state. > > If we look into load_per_instance() it already doing similar check : > > !(inst->state >= INST_INIT && inst->state < INST_STOP) > > This check would just make sure instance state for scaling, not a mandatory check. I can remove the check. >> + >> + return scale_clocks(inst); >> +} >> + >> static void fill_buffer_desc(const struct venus_buffer *buf, >> struct hfi_buffer_desc *bd, bool response) >> { >> @@ -1053,7 +1137,7 @@ void venus_helper_vb2_stop_streaming(struct >> vb2_queue *q) >> >> venus_helper_free_dpb_bufs(inst); >> >> - load_scale_clocks(core); >> + load_scale_clocks(inst); >> INIT_LIST_HEAD(&inst->registeredbufs); >> } >> >> @@ -1070,7 +1154,6 @@ void venus_helper_vb2_stop_streaming(struct >> vb2_queue *q) >> >> int venus_helper_vb2_start_streaming(struct venus_inst *inst) >> { >> - struct venus_core *core = inst->core; >> int ret; >> >> ret = intbufs_alloc(inst); >> @@ -1081,7 +1164,7 @@ int venus_helper_vb2_start_streaming(struct >> venus_inst *inst) >> if (ret) >> goto err_bufs_free; >> >> - load_scale_clocks(core); >> + load_scale_clocks(inst); >> >> ret = hfi_session_load_res(inst); >> if (ret) >> Regards, Aniket
diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c index 7492373..2c976e4 100644 --- a/drivers/media/platform/qcom/venus/helpers.c +++ b/drivers/media/platform/qcom/venus/helpers.c @@ -348,8 +348,9 @@ static u32 load_per_type(struct venus_core *core, u32 session_type) return mbs_per_sec; } -static int load_scale_clocks(struct venus_core *core) +static int scale_clocks(struct venus_inst *inst) { + struct venus_core *core = inst->core; const struct freq_tbl *table = core->res->freq_tbl; unsigned int num_rows = core->res->freq_tbl_size; unsigned long freq = table[0].freq; @@ -398,6 +399,89 @@ static int load_scale_clocks(struct venus_core *core) return ret; } +static unsigned long calculate_vpp_freq(struct venus_inst *inst) +{ + unsigned long vpp_freq = 0; + u32 mbs_per_sec; + + mbs_per_sec = load_per_instance(inst); + vpp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vpp_freq; + /* 21 / 20 is overhead factor */ + vpp_freq += vpp_freq / 20; + + return vpp_freq; +} + +static int scale_clocks_v4(struct venus_inst *inst) +{ + struct venus_core *core = inst->core; + const struct freq_tbl *table = core->res->freq_tbl; + unsigned int num_rows = core->res->freq_tbl_size; + struct clk *clk = core->clks[0]; + struct device *dev = core->dev; + unsigned int i; + unsigned long freq = 0, freq_core1 = 0, freq_core2 = 0; + int ret; + + freq = calculate_vpp_freq(inst); + + if (freq > table[0].freq) + dev_warn(dev, "HW is overloaded, needed: %lu max: %lu\n", + freq, table[0].freq); + + for (i = 0; i < num_rows; i++) { + if (freq > table[i].freq) + break; + freq = table[i].freq; + } + + inst->clk_data.freq = freq; + + mutex_lock(&core->lock); + list_for_each_entry(inst, &core->instances, list) { + if (inst->clk_data.core_id == VIDC_CORE_ID_1) { + freq_core1 += inst->clk_data.freq; + } else if (inst->clk_data.core_id == VIDC_CORE_ID_2) { + freq_core2 += inst->clk_data.freq; + } else if (inst->clk_data.core_id == VIDC_CORE_ID_3) { + freq_core1 += inst->clk_data.freq; + freq_core2 += inst->clk_data.freq; + } + } + mutex_unlock(&core->lock); + + freq = max(freq_core1, freq_core2); + + ret = clk_set_rate(clk, freq); + if (ret) + goto err; + + ret = clk_set_rate(core->core0_clk, freq); + if (ret) + goto err; + + ret = clk_set_rate(core->core1_clk, freq); + if (ret) + goto err; + + return 0; + +err: + dev_err(dev, "failed to set clock rate %lu (%d)\n", freq, ret); + return ret; +} + +static int load_scale_clocks(struct venus_inst *inst) +{ + if (IS_V4(inst->core)) + return scale_clocks_v4(inst); + + if (inst->state == INST_START) + return 0; + + return scale_clocks(inst); +} + static void fill_buffer_desc(const struct venus_buffer *buf, struct hfi_buffer_desc *bd, bool response) { @@ -1053,7 +1137,7 @@ void venus_helper_vb2_stop_streaming(struct vb2_queue *q) venus_helper_free_dpb_bufs(inst); - load_scale_clocks(core); + load_scale_clocks(inst); INIT_LIST_HEAD(&inst->registeredbufs); } @@ -1070,7 +1154,6 @@ void venus_helper_vb2_stop_streaming(struct vb2_queue *q) int venus_helper_vb2_start_streaming(struct venus_inst *inst) { - struct venus_core *core = inst->core; int ret; ret = intbufs_alloc(inst); @@ -1081,7 +1164,7 @@ int venus_helper_vb2_start_streaming(struct venus_inst *inst) if (ret) goto err_bufs_free; - load_scale_clocks(core); + load_scale_clocks(inst); ret = hfi_session_load_res(inst); if (ret)
Current clock scaling calculations are same for vpu4 and previous versions. For vpu4, Clock scaling calculations are updated with cycles/mb. This helps in getting precise clock required. Signed-off-by: Aniket Masule <amasule@codeaurora.org> --- drivers/media/platform/qcom/venus/helpers.c | 91 +++++++++++++++++++++++++++-- 1 file changed, 87 insertions(+), 4 deletions(-)