[PATCH v6 1/4] media: venus: Add codec data table

[PATCH v6 1/4] media: venus: Add codec data table

[Aniket Masule]

Add vpp cycles for different types of codec.
It indicates the cycles required by video hardware
to process each macroblock. Add vsp cycles, cycles
required by stream processor. Initialize the codec
data with core resources.

Signed-off-by: Aniket Masule <amasule@codeaurora.org>
---
 drivers/media/platform/qcom/venus/core.c    | 13 +++++++++++++
 drivers/media/platform/qcom/venus/core.h    | 16 +++++++++++++++
 drivers/media/platform/qcom/venus/helpers.c | 30 +++++++++++++++++++++++++++++
 drivers/media/platform/qcom/venus/helpers.h |  1 +
 drivers/media/platform/qcom/venus/vdec.c    |  4 ++++
 drivers/media/platform/qcom/venus/venc.c    |  4 ++++
 6 files changed, 68 insertions(+)

diff --git a/drivers/media/platform/qcom/venus/core.c b/drivers/media/platform/qcom/venus/core.c
index 7393667..ad6bb74 100644
--- a/drivers/media/platform/qcom/venus/core.c
+++ b/drivers/media/platform/qcom/venus/core.c
@@ -473,9 +473,22 @@ static __maybe_unused int venus_runtime_resume(struct device *dev)
 	{  244800, 100000000 },	/* 1920x1080@30 */
 };
 
+static struct codec_freq_data sdm845_codec_freq_data[] =  {
+	{ V4L2_PIX_FMT_H264, VIDC_SESSION_TYPE_ENC, 675, 10 },
+	{ V4L2_PIX_FMT_HEVC, VIDC_SESSION_TYPE_ENC, 675, 10 },
+	{ V4L2_PIX_FMT_VP8, VIDC_SESSION_TYPE_ENC, 675, 10 },
+	{ V4L2_PIX_FMT_MPEG2, VIDC_SESSION_TYPE_DEC, 200, 10 },
+	{ V4L2_PIX_FMT_H264, VIDC_SESSION_TYPE_DEC, 200, 10 },
+	{ V4L2_PIX_FMT_HEVC, VIDC_SESSION_TYPE_DEC, 200, 10 },
+	{ V4L2_PIX_FMT_VP8, VIDC_SESSION_TYPE_DEC, 200, 10 },
+	{ V4L2_PIX_FMT_VP9, VIDC_SESSION_TYPE_DEC, 200, 10 },
+};
+
 static const struct venus_resources sdm845_res = {
 	.freq_tbl = sdm845_freq_table,
 	.freq_tbl_size = ARRAY_SIZE(sdm845_freq_table),
+	.codec_freq_data = sdm845_codec_freq_data,
+	.codec_freq_data_size = ARRAY_SIZE(sdm845_codec_freq_data),
 	.clks = {"core", "iface", "bus" },
 	.clks_num = 3,
 	.max_load = 2563200,
diff --git a/drivers/media/platform/qcom/venus/core.h b/drivers/media/platform/qcom/venus/core.h
index 7a3feb5..b8aef19 100644
--- a/drivers/media/platform/qcom/venus/core.h
+++ b/drivers/media/platform/qcom/venus/core.h
@@ -35,12 +35,21 @@ struct reg_val {
 	u32 value;
 };
 
+struct codec_freq_data {
+	u32 pixfmt;
+	u32 session_type;
+	unsigned long vpp_freq;
+	unsigned long vsp_freq;
+};
+
 struct venus_resources {
 	u64 dma_mask;
 	const struct freq_tbl *freq_tbl;
 	unsigned int freq_tbl_size;
 	const struct reg_val *reg_tbl;
 	unsigned int reg_tbl_size;
+	const struct codec_freq_data *codec_freq_data;
+	unsigned int codec_freq_data_size;
 	const char * const clks[VIDC_CLKS_NUM_MAX];
 	unsigned int clks_num;
 	enum hfi_version hfi_version;
@@ -216,6 +225,12 @@ struct venus_buffer {
 	struct list_head ref_list;
 };
 
+struct clock_data {
+	u32 core_id;
+	unsigned long freq;
+	const struct codec_freq_data *codec_freq_data;
+};
+
 #define to_venus_buffer(ptr)	container_of(ptr, struct venus_buffer, vb)
 
 /**
@@ -275,6 +290,7 @@ struct venus_inst {
 	struct list_head list;
 	struct mutex lock;
 	struct venus_core *core;
+	struct clock_data clk_data;
 	struct list_head dpbbufs;
 	struct list_head internalbufs;
 	struct list_head registeredbufs;
diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c
index 5cad601..7492373 100644
--- a/drivers/media/platform/qcom/venus/helpers.c
+++ b/drivers/media/platform/qcom/venus/helpers.c
@@ -715,6 +715,36 @@ int venus_helper_set_core_usage(struct venus_inst *inst, u32 usage)
 }
 EXPORT_SYMBOL_GPL(venus_helper_set_core_usage);
 
+int venus_helper_init_codec_freq_data(struct venus_inst *inst)
+{
+	const struct codec_freq_data *data;
+	unsigned int i, data_size;
+	u32 pixfmt;
+	int ret = 0;
+
+	if (!IS_V4(inst->core))
+		return 0;
+
+	data = inst->core->res->codec_freq_data;
+	data_size = inst->core->res->codec_freq_data_size;
+	pixfmt = inst->session_type == VIDC_SESSION_TYPE_DEC ?
+			inst->fmt_out->pixfmt : inst->fmt_cap->pixfmt;
+
+	for (i = 0; i < data_size; i++) {
+		if (data[i].pixfmt == pixfmt &&
+		    data[i].session_type == inst->session_type) {
+			inst->clk_data.codec_freq_data = &data[i];
+			break;
+		}
+	}
+
+	if (!inst->clk_data.codec_freq_data)
+		ret = -EINVAL;
+
+	return ret;
+}
+EXPORT_SYMBOL_GPL(venus_helper_init_codec_freq_data);
+
 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/helpers.h b/drivers/media/platform/qcom/venus/helpers.h
index 2475f284..2c13245 100644
--- a/drivers/media/platform/qcom/venus/helpers.h
+++ b/drivers/media/platform/qcom/venus/helpers.h
@@ -41,6 +41,7 @@ int venus_helper_set_output_resolution(struct venus_inst *inst,
 				       unsigned int width, unsigned int height,
 				       u32 buftype);
 int venus_helper_set_work_mode(struct venus_inst *inst, u32 mode);
+int venus_helper_init_codec_freq_data(struct venus_inst *inst);
 int venus_helper_set_core_usage(struct venus_inst *inst, u32 usage);
 int venus_helper_set_num_bufs(struct venus_inst *inst, unsigned int input_bufs,
 			      unsigned int output_bufs,
diff --git a/drivers/media/platform/qcom/venus/vdec.c b/drivers/media/platform/qcom/venus/vdec.c
index 282de21..d037f80 100644
--- a/drivers/media/platform/qcom/venus/vdec.c
+++ b/drivers/media/platform/qcom/venus/vdec.c
@@ -660,6 +660,10 @@ static int vdec_init_session(struct venus_inst *inst)
 	if (ret)
 		goto deinit;
 
+	ret = venus_helper_init_codec_freq_data(inst);
+	if (ret)
+		goto deinit;
+
 	return 0;
 deinit:
 	hfi_session_deinit(inst);
diff --git a/drivers/media/platform/qcom/venus/venc.c b/drivers/media/platform/qcom/venus/venc.c
index 32cff29..cdddc82 100644
--- a/drivers/media/platform/qcom/venus/venc.c
+++ b/drivers/media/platform/qcom/venus/venc.c
@@ -847,6 +847,10 @@ static int venc_init_session(struct venus_inst *inst)
 	if (ret)
 		goto deinit;
 
+	ret = venus_helper_init_codec_freq_data(inst);
+	if (ret)
+		goto deinit;
+
 	ret = venc_set_properties(inst);
 	if (ret)
 		goto deinit;
-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project

[PATCH v6 2/4] media: venus: Update clock scaling

[Aniket Masule]

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;
+	}
+
+	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)
-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project

[PATCH v6 3/4] media: venus: Update to bitrate based clock scaling

[Aniket Masule]

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);
+	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);
-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project

[PATCH v6 4/4] media: venus: Update core selection

[Aniket Masule]

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 <amasule@codeaurora.org>
---
 drivers/media/platform/qcom/venus/helpers.c    | 69 +++++++++++++++++++++++---
 drivers/media/platform/qcom/venus/helpers.h    |  2 +-
 drivers/media/platform/qcom/venus/hfi_helper.h |  1 +
 drivers/media/platform/qcom/venus/hfi_parser.h |  5 ++
 drivers/media/platform/qcom/venus/vdec.c       |  2 +-
 drivers/media/platform/qcom/venus/venc.c       |  2 +-
 6 files changed, 72 insertions(+), 9 deletions(-)

diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c
index edf403d..3b6cbbf 100644
--- a/drivers/media/platform/qcom/venus/helpers.c
+++ b/drivers/media/platform/qcom/venus/helpers.c
@@ -26,6 +26,7 @@
 #include "helpers.h"
 #include "hfi_helper.h"
 #include "hfi_venus_io.h"
+#include "hfi_parser.h"
 
 struct intbuf {
 	struct list_head list;
@@ -331,6 +332,24 @@ static u32 load_per_instance(struct venus_inst *inst)
 	return mbs * inst->fps;
 }
 
+static u32 load_per_core(struct venus_core *core, u32 core_id)
+{
+	struct venus_inst *inst = NULL;
+	u32 mbs_per_sec = 0, load = 0;
+
+	mutex_lock(&core->lock);
+	list_for_each_entry(inst, &core->instances, list) {
+		if (inst->clk_data.core_id != core_id)
+			continue;
+
+		mbs_per_sec = load_per_instance(inst);
+		load += mbs_per_sec * inst->clk_data.codec_freq_data->vpp_freq;
+	}
+	mutex_unlock(&core->lock);
+
+	return load;
+}
+
 static u32 load_per_type(struct venus_core *core, u32 session_type)
 {
 	struct venus_inst *inst = NULL;
@@ -505,6 +524,16 @@ static int load_scale_clocks(struct venus_inst *inst)
 	return scale_clocks(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)
 {
@@ -808,19 +837,47 @@ 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_set_core(struct venus_inst *inst)
 {
-	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, core1_load = 0, core2_load = 0;
+	unsigned long min_load, max_freq, cur_inst_load;
+	u32 cores_max;
+	int ret;
 
 	if (!IS_V4(inst->core))
 		return 0;
 
-	cu.video_core_enable_mask = usage;
+	core1_load = load_per_core(core, VIDC_CORE_ID_1);
+	core2_load = load_per_core(core, VIDC_CORE_ID_2);
+	min_core_id = core1_load < core2_load ? VIDC_CORE_ID_1 : VIDC_CORE_ID_2;
+	min_load = min(core1_load, core2_load);
+	cores_max = core_num_max(inst);
 
-	return hfi_session_set_property(inst, ptype, &cu);
+	if (cores_max < VIDC_CORE_ID_2) {
+		min_core_id = VIDC_CORE_ID_1;
+		min_load = core1_load;
+	}
+
+	cur_inst_load = load_per_instance(inst) *
+		inst->clk_data.codec_freq_data->vpp_freq;
+	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_set_core);
 
 int venus_helper_init_codec_freq_data(struct venus_inst *inst)
 {
diff --git a/drivers/media/platform/qcom/venus/helpers.h b/drivers/media/platform/qcom/venus/helpers.h
index 2c13245..1034111 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_freq_data(struct venus_inst *inst);
-int venus_helper_set_core_usage(struct venus_inst *inst, u32 usage);
+int venus_helper_set_core(struct venus_inst *inst);
 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/hfi_helper.h b/drivers/media/platform/qcom/venus/hfi_helper.h
index 34ea503..f3d1018 100644
--- a/drivers/media/platform/qcom/venus/hfi_helper.h
+++ b/drivers/media/platform/qcom/venus/hfi_helper.h
@@ -559,6 +559,7 @@ struct hfi_bitrate {
 #define HFI_CAPABILITY_LCU_SIZE				0x14
 #define HFI_CAPABILITY_HIER_P_HYBRID_NUM_ENH_LAYERS	0x15
 #define HFI_CAPABILITY_MBS_PER_SECOND_POWERSAVE		0x16
+#define HFI_CAPABILITY_MAX_VIDEOCORES		0x2B
 
 struct hfi_capability {
 	u32 capability_type;
diff --git a/drivers/media/platform/qcom/venus/hfi_parser.h b/drivers/media/platform/qcom/venus/hfi_parser.h
index 3e931c7..264e6dd 100644
--- a/drivers/media/platform/qcom/venus/hfi_parser.h
+++ b/drivers/media/platform/qcom/venus/hfi_parser.h
@@ -107,4 +107,9 @@ static inline u32 frate_step(struct venus_inst *inst)
 	return cap_step(inst, HFI_CAPABILITY_FRAMERATE);
 }
 
+static inline u32 core_num_max(struct venus_inst *inst)
+{
+	return cap_max(inst, HFI_CAPABILITY_MAX_VIDEOCORES);
+}
+
 #endif
diff --git a/drivers/media/platform/qcom/venus/vdec.c b/drivers/media/platform/qcom/venus/vdec.c
index d037f80..620e060 100644
--- a/drivers/media/platform/qcom/venus/vdec.c
+++ b/drivers/media/platform/qcom/venus/vdec.c
@@ -551,7 +551,7 @@ static int vdec_output_conf(struct venus_inst *inst)
 	if (ret)
 		return ret;
 
-	ret = venus_helper_set_core_usage(inst, VIDC_CORE_ID_1);
+	ret = venus_helper_set_core(inst);
 	if (ret)
 		return ret;
 
diff --git a/drivers/media/platform/qcom/venus/venc.c b/drivers/media/platform/qcom/venus/venc.c
index cdddc82..28e76cc 100644
--- a/drivers/media/platform/qcom/venus/venc.c
+++ b/drivers/media/platform/qcom/venus/venc.c
@@ -660,7 +660,7 @@ static int venc_set_properties(struct venus_inst *inst)
 	if (ret)
 		return ret;
 
-	ret = venus_helper_set_core_usage(inst, VIDC_CORE_ID_2);
+	ret = venus_helper_set_core(inst);
 	if (ret)
 		return ret;
 
-- 
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project

Re: [PATCH v6 3/4] media: venus: Update to bitrate based clock scaling

[Stanimir Varbanov]

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);
> 

-- 
regards,
Stan

Re: [PATCH v6 2/4] media: venus: Update clock scaling

[Stanimir Varbanov]

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)
> 

-- 
regards,
Stan

Re: [PATCH v6 4/4] media: venus: Update core selection

[Stanimir Varbanov]

Hi,

On 7/22/19 12:07 PM, 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 <amasule@codeaurora.org>
> ---
>  drivers/media/platform/qcom/venus/helpers.c    | 69 +++++++++++++++++++++++---
>  drivers/media/platform/qcom/venus/helpers.h    |  2 +-
>  drivers/media/platform/qcom/venus/hfi_helper.h |  1 +
>  drivers/media/platform/qcom/venus/hfi_parser.h |  5 ++
>  drivers/media/platform/qcom/venus/vdec.c       |  2 +-
>  drivers/media/platform/qcom/venus/venc.c       |  2 +-
>  6 files changed, 72 insertions(+), 9 deletions(-)
> 
> diff --git a/drivers/media/platform/qcom/venus/helpers.c b/drivers/media/platform/qcom/venus/helpers.c
> index edf403d..3b6cbbf 100644
> --- a/drivers/media/platform/qcom/venus/helpers.c
> +++ b/drivers/media/platform/qcom/venus/helpers.c
> @@ -26,6 +26,7 @@
>  #include "helpers.h"
>  #include "hfi_helper.h"
>  #include "hfi_venus_io.h"
> +#include "hfi_parser.h"
>  
>  struct intbuf {
>  	struct list_head list;
> @@ -331,6 +332,24 @@ static u32 load_per_instance(struct venus_inst *inst)
>  	return mbs * inst->fps;
>  }
>  
> +static u32 load_per_core(struct venus_core *core, u32 core_id)
> +{
> +	struct venus_inst *inst = NULL;
> +	u32 mbs_per_sec = 0, load = 0;
> +
> +	mutex_lock(&core->lock);
> +	list_for_each_entry(inst, &core->instances, list) {
> +		if (inst->clk_data.core_id != core_id)
> +			continue;
> +
> +		mbs_per_sec = load_per_instance(inst);
> +		load += mbs_per_sec * inst->clk_data.codec_freq_data->vpp_freq;
> +	}
> +	mutex_unlock(&core->lock);
> +
> +	return load;
> +}
> +
>  static u32 load_per_type(struct venus_core *core, u32 session_type)
>  {
>  	struct venus_inst *inst = NULL;
> @@ -505,6 +524,16 @@ static int load_scale_clocks(struct venus_inst *inst)
>  	return scale_clocks(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)
>  {
> @@ -808,19 +837,47 @@ 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_set_core(struct venus_inst *inst)
>  {
> -	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, core1_load = 0, core2_load = 0;
> +	unsigned long min_load, max_freq, cur_inst_load;
> +	u32 cores_max;
> +	int ret;
>  
>  	if (!IS_V4(inst->core))
>  		return 0;
>  
> -	cu.video_core_enable_mask = usage;
> +	core1_load = load_per_core(core, VIDC_CORE_ID_1);
> +	core2_load = load_per_core(core, VIDC_CORE_ID_2);
> +	min_core_id = core1_load < core2_load ? VIDC_CORE_ID_1 : VIDC_CORE_ID_2;
> +	min_load = min(core1_load, core2_load);
> +	cores_max = core_num_max(inst);
>  
> -	return hfi_session_set_property(inst, ptype, &cu);
> +	if (cores_max < VIDC_CORE_ID_2) {
> +		min_core_id = VIDC_CORE_ID_1;
> +		min_load = core1_load;
> +	}
> +
> +	cur_inst_load = load_per_instance(inst) *
> +		inst->clk_data.codec_freq_data->vpp_freq;
> +	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);

We have a problem here. Lets assume that we have only one running
decoder session and the code above decides that it should be handled by
core2, but core2 clocks presently are enabled only if there is an
encoder session (see DT subnodes), thus we select core2 but without
enabling core2 clocks and power domain.

> +	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_set_core);
>  


-- 
regards,
Stan

Re: [PATCH v6 3/4] media: venus: Update to bitrate based clock scaling

[Stanimir Varbanov]

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);
>  }
>  


-- 
regards,
Stan

Re: [PATCH v6 2/4] media: venus: Update clock scaling

[amasule]

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

Re: [PATCH v6 3/4] media: venus: Update to bitrate based clock scaling

[amasule]

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

Re: [PATCH v6 4/4] media: venus: Update core selection

[amasule]

Hi,

On 2019-07-29 15:16, Stanimir Varbanov wrote:
> Hi,
> 
> On 7/22/19 12:07 PM, 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 <amasule@codeaurora.org>
>> ---
>>  drivers/media/platform/qcom/venus/helpers.c    | 69 
>> +++++++++++++++++++++++---
>>  drivers/media/platform/qcom/venus/helpers.h    |  2 +-
>>  drivers/media/platform/qcom/venus/hfi_helper.h |  1 +
>>  drivers/media/platform/qcom/venus/hfi_parser.h |  5 ++
>>  drivers/media/platform/qcom/venus/vdec.c       |  2 +-
>>  drivers/media/platform/qcom/venus/venc.c       |  2 +-
>>  6 files changed, 72 insertions(+), 9 deletions(-)
>> 
>> diff --git a/drivers/media/platform/qcom/venus/helpers.c 
>> b/drivers/media/platform/qcom/venus/helpers.c
>> index edf403d..3b6cbbf 100644
>> --- a/drivers/media/platform/qcom/venus/helpers.c
>> +++ b/drivers/media/platform/qcom/venus/helpers.c
>> @@ -26,6 +26,7 @@
>>  #include "helpers.h"
>>  #include "hfi_helper.h"
>>  #include "hfi_venus_io.h"
>> +#include "hfi_parser.h"
>> 
>>  struct intbuf {
>>  	struct list_head list;
>> @@ -331,6 +332,24 @@ static u32 load_per_instance(struct venus_inst 
>> *inst)
>>  	return mbs * inst->fps;
>>  }
>> 
>> +static u32 load_per_core(struct venus_core *core, u32 core_id)
>> +{
>> +	struct venus_inst *inst = NULL;
>> +	u32 mbs_per_sec = 0, load = 0;
>> +
>> +	mutex_lock(&core->lock);
>> +	list_for_each_entry(inst, &core->instances, list) {
>> +		if (inst->clk_data.core_id != core_id)
>> +			continue;
>> +
>> +		mbs_per_sec = load_per_instance(inst);
>> +		load += mbs_per_sec * inst->clk_data.codec_freq_data->vpp_freq;
>> +	}
>> +	mutex_unlock(&core->lock);
>> +
>> +	return load;
>> +}
>> +
>>  static u32 load_per_type(struct venus_core *core, u32 session_type)
>>  {
>>  	struct venus_inst *inst = NULL;
>> @@ -505,6 +524,16 @@ static int load_scale_clocks(struct venus_inst 
>> *inst)
>>  	return scale_clocks(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)
>>  {
>> @@ -808,19 +837,47 @@ 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_set_core(struct venus_inst *inst)
>>  {
>> -	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, core1_load = 0, core2_load = 0;
>> +	unsigned long min_load, max_freq, cur_inst_load;
>> +	u32 cores_max;
>> +	int ret;
>> 
>>  	if (!IS_V4(inst->core))
>>  		return 0;
>> 
>> -	cu.video_core_enable_mask = usage;
>> +	core1_load = load_per_core(core, VIDC_CORE_ID_1);
>> +	core2_load = load_per_core(core, VIDC_CORE_ID_2);
>> +	min_core_id = core1_load < core2_load ? VIDC_CORE_ID_1 : 
>> VIDC_CORE_ID_2;
>> +	min_load = min(core1_load, core2_load);
>> +	cores_max = core_num_max(inst);
>> 
>> -	return hfi_session_set_property(inst, ptype, &cu);
>> +	if (cores_max < VIDC_CORE_ID_2) {
>> +		min_core_id = VIDC_CORE_ID_1;
>> +		min_load = core1_load;
>> +	}
>> +
>> +	cur_inst_load = load_per_instance(inst) *
>> +		inst->clk_data.codec_freq_data->vpp_freq;
>> +	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);
> 
> We have a problem here. Lets assume that we have only one running
> decoder session and the code above decides that it should be handled by
> core2, but core2 clocks presently are enabled only if there is an
> encoder session (see DT subnodes), thus we select core2 but without
> enabling core2 clocks and power domain.
> 
I will send a separate patch for core selection, once power domain issue
is fixed.

>> +	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_set_core);
>> 

Regards,
Aniket