| Message ID | 1563786452-22188-3-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 |
On 7/22/19 12:07 PM, Aniket Masule wrote: > Introduced clock scaling using bitrate, preavious > calculations consider only the cycles per mb. > Also, clock scaling is now triggered before every > buffer being queued to the device. This helps in > deciding precise clock cycles required. > > Signed-off-by: Aniket Masule <amasule@codeaurora.org> > --- > drivers/media/platform/qcom/venus/helpers.c | 33 +++++++++++++++++++++++++---- > 1 file changed, 29 insertions(+), 4 deletions(-) > > diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c > index 2c976e4..edf403d 100644 > --- a/drivers/media/platform/qcom/venus/helpers.c > +++ b/drivers/media/platform/qcom/venus/helpers.c > @@ -399,17 +399,26 @@ static int scale_clocks(struct venus_inst *inst) > return ret; > } > > -static unsigned long calculate_vpp_freq(struct venus_inst *inst) > +static unsigned long calculate_inst_freq(struct venus_inst *inst, > + unsigned long filled_len) > { > - unsigned long vpp_freq = 0; > + unsigned long vpp_freq = 0, vsp_freq = 0; > + u64 fps = inst->fps; > 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; > + vsp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vsp_freq; > > - return vpp_freq; > + /* 10 / 7 is overhead factor */ > + if (inst->session_type == VIDC_SESSION_TYPE_ENC) > + vsp_freq += (inst->controls.enc.bitrate * 10) / 7; > + else > + vsp_freq += ((fps * filled_len * 8) * 10) / 7; > + > + return max(vpp_freq, vsp_freq); > } > > static int scale_clocks_v4(struct venus_inst *inst) > @@ -417,13 +426,27 @@ 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 v4l2_m2m_ctx *m2m_ctx = inst->m2m_ctx; > struct clk *clk = core->clks[0]; > struct device *dev = core->dev; > unsigned int i; > unsigned long freq = 0, freq_core1 = 0, freq_core2 = 0; > + unsigned long filled_len = 0; > + struct venus_buffer *buf, *n; > + struct vb2_buffer *vb; > int ret; > > - freq = calculate_vpp_freq(inst); > + mutex_lock(&inst->lock); the lock is not needed and having it will produce a deadlock. > + v4l2_m2m_for_each_src_buf_safe(m2m_ctx, buf, n) { > + vb = &buf->vb.vb2_buf; > + filled_len = max(filled_len, vb2_get_plane_payload(vb, 0)); > + } > + mutex_unlock(&inst->lock); > + > + if (inst->session_type == VIDC_SESSION_TYPE_DEC && !filled_len) > + return 0; > + > + freq = calculate_inst_freq(inst, filled_len); > > if (freq > table[0].freq) > dev_warn(dev, "HW is overloaded, needed: %lu max: %lu\n", > @@ -1093,6 +1116,8 @@ void venus_helper_vb2_buf_queue(struct vb2_buffer *vb) > if (ret) > goto unlock; > > + load_scale_clocks(inst); > + > ret = session_process_buf(inst, vbuf); > if (ret) > return_buf_error(inst, vbuf); >
Hi, On 7/22/19 12:07 PM, Aniket Masule wrote: > Introduced clock scaling using bitrate, preavious > calculations consider only the cycles per mb. > Also, clock scaling is now triggered before every > buffer being queued to the device. This helps in > deciding precise clock cycles required. > > Signed-off-by: Aniket Masule <amasule@codeaurora.org> > --- > drivers/media/platform/qcom/venus/helpers.c | 33 +++++++++++++++++++++++++---- > 1 file changed, 29 insertions(+), 4 deletions(-) > > diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c > index 2c976e4..edf403d 100644 > --- a/drivers/media/platform/qcom/venus/helpers.c > +++ b/drivers/media/platform/qcom/venus/helpers.c > @@ -399,17 +399,26 @@ static int scale_clocks(struct venus_inst *inst) > return ret; > } > > -static unsigned long calculate_vpp_freq(struct venus_inst *inst) > +static unsigned long calculate_inst_freq(struct venus_inst *inst, > + unsigned long filled_len) > { > - unsigned long vpp_freq = 0; > + unsigned long vpp_freq = 0, vsp_freq = 0; > + u64 fps = inst->fps; > 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; > + vsp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vsp_freq; > > - return vpp_freq; > + /* 10 / 7 is overhead factor */ > + if (inst->session_type == VIDC_SESSION_TYPE_ENC) > + vsp_freq += (inst->controls.enc.bitrate * 10) / 7; > + else > + vsp_freq += ((fps * filled_len * 8) * 10) / 7; load_per_instance() already taken into account fps, thus I think fps should be excluded from calculation. > + > + return max(vpp_freq, vsp_freq); > } >
Hi Stan, On 2019-07-29 17:08, Stanimir Varbanov wrote: > Hi, > > On 7/22/19 12:07 PM, Aniket Masule wrote: >> Introduced clock scaling using bitrate, preavious >> calculations consider only the cycles per mb. >> Also, clock scaling is now triggered before every >> buffer being queued to the device. This helps in >> deciding precise clock cycles required. >> >> Signed-off-by: Aniket Masule <amasule@codeaurora.org> >> --- >> drivers/media/platform/qcom/venus/helpers.c | 33 >> +++++++++++++++++++++++++---- >> 1 file changed, 29 insertions(+), 4 deletions(-) >> >> diff --git a/drivers/media/platform/qcom/venus/helpers.c >> b/drivers/media/platform/qcom/venus/helpers.c >> index 2c976e4..edf403d 100644 >> --- a/drivers/media/platform/qcom/venus/helpers.c >> +++ b/drivers/media/platform/qcom/venus/helpers.c >> @@ -399,17 +399,26 @@ static int scale_clocks(struct venus_inst *inst) >> return ret; >> } >> >> -static unsigned long calculate_vpp_freq(struct venus_inst *inst) >> +static unsigned long calculate_inst_freq(struct venus_inst *inst, >> + unsigned long filled_len) >> { >> - unsigned long vpp_freq = 0; >> + unsigned long vpp_freq = 0, vsp_freq = 0; >> + u64 fps = inst->fps; >> 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; >> + vsp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vsp_freq; >> >> - return vpp_freq; >> + /* 10 / 7 is overhead factor */ >> + if (inst->session_type == VIDC_SESSION_TYPE_ENC) >> + vsp_freq += (inst->controls.enc.bitrate * 10) / 7; >> + else >> + vsp_freq += ((fps * filled_len * 8) * 10) / 7; > > load_per_instance() already taken into account fps, thus I think fps > should be excluded from calculation. > I will correct this, fps is needed for stream processor frequency calculations, also fps is not needed for vpp. >> + >> + return max(vpp_freq, vsp_freq); >> } >> Regards, Aniket
diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c index 2c976e4..edf403d 100644 --- a/drivers/media/platform/qcom/venus/helpers.c +++ b/drivers/media/platform/qcom/venus/helpers.c @@ -399,17 +399,26 @@ static int scale_clocks(struct venus_inst *inst) return ret; } -static unsigned long calculate_vpp_freq(struct venus_inst *inst) +static unsigned long calculate_inst_freq(struct venus_inst *inst, + unsigned long filled_len) { - unsigned long vpp_freq = 0; + unsigned long vpp_freq = 0, vsp_freq = 0; + u64 fps = inst->fps; 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; + vsp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vsp_freq; - return vpp_freq; + /* 10 / 7 is overhead factor */ + if (inst->session_type == VIDC_SESSION_TYPE_ENC) + vsp_freq += (inst->controls.enc.bitrate * 10) / 7; + else + vsp_freq += ((fps * filled_len * 8) * 10) / 7; + + return max(vpp_freq, vsp_freq); } static int scale_clocks_v4(struct venus_inst *inst) @@ -417,13 +426,27 @@ 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 v4l2_m2m_ctx *m2m_ctx = inst->m2m_ctx; struct clk *clk = core->clks[0]; struct device *dev = core->dev; unsigned int i; unsigned long freq = 0, freq_core1 = 0, freq_core2 = 0; + unsigned long filled_len = 0; + struct venus_buffer *buf, *n; + struct vb2_buffer *vb; int ret; - freq = calculate_vpp_freq(inst); + mutex_lock(&inst->lock); + v4l2_m2m_for_each_src_buf_safe(m2m_ctx, buf, n) { + vb = &buf->vb.vb2_buf; + filled_len = max(filled_len, vb2_get_plane_payload(vb, 0)); + } + mutex_unlock(&inst->lock); + + if (inst->session_type == VIDC_SESSION_TYPE_DEC && !filled_len) + return 0; + + freq = calculate_inst_freq(inst, filled_len); if (freq > table[0].freq) dev_warn(dev, "HW is overloaded, needed: %lu max: %lu\n", @@ -1093,6 +1116,8 @@ void venus_helper_vb2_buf_queue(struct vb2_buffer *vb) if (ret) goto unlock; + load_scale_clocks(inst); + ret = session_process_buf(inst, vbuf); if (ret) return_buf_error(inst, vbuf);
Introduced clock scaling using bitrate, preavious calculations consider only the cycles per mb. Also, clock scaling is now triggered before every buffer being queued to the device. This helps in deciding precise clock cycles required. Signed-off-by: Aniket Masule <amasule@codeaurora.org> --- drivers/media/platform/qcom/venus/helpers.c | 33 +++++++++++++++++++++++++---- 1 file changed, 29 insertions(+), 4 deletions(-)