Linux Media Controller development
 help / color / mirror / Atom feed
* [PATCH v3] media: iris: optimize COMV buffer allocation for VPU3x and VPU4x
@ 2026-05-06  6:12 Vishnu Reddy
  2026-05-06  7:23 ` Vikash Garodia
  0 siblings, 1 reply; 2+ messages in thread
From: Vishnu Reddy @ 2026-05-06  6:12 UTC (permalink / raw)
  To: Vikash Garodia, Dikshita Agarwal, Abhinav Kumar,
	Bryan O'Donoghue, Mauro Carvalho Chehab
  Cc: linux-media, linux-arm-msm, linux-kernel, Vishnu Reddy

The existing iris_vpu_dec_comv_size() used VIDEO_MAX_FRAME (32) as
num_comv count unconditionally when calculating the co-located motion
vector (COMV) buffer size. This resulted in an oversized COMV buffer
allocation throughout decode session, wasting memory regardless of
actual number of buffers required.

For VPU3x and VPU4x platforms, introduce iris_vpu3x_4x_dec_comv_size() to
replace iris_vpu_dec_comv_size(). These derive num_comv dynamically, it
uses inst->fw_min_count once the firmware has reported its buffer
requirements, and fallback to output count during initialization before
firmware has communicated its requirements. This aligns the COMV buffer
size to the actual count needed rather than always allocating with fixed
VIDEO_MAX_FRAME value.

Additionally, during iris_vdec_inst_init(), fw_min_count was initialized
to MIN_BUFFERS instead of 0. This masked the fallback logic and caused the
COMV size calculation to use MIN_BUFFERS even before firmware had reported
its actual requirements. Fix this by initializing fw_min_count to 0.

During testing of 1080p AVC, it reduces the COMV buffer size from 32.89MB
to 6.16MB per decode session, significantly reducing memory consumption.

Signed-off-by: Vishnu Reddy <busanna.reddy@oss.qualcomm.com>
---
Changes in v3:
- Update num_comv from instance data instead of using caps num_comv
  in set_num_comv function to avoid wrong value update during concurrency.
- Link to v2: https://lore.kernel.org/r/20260504-optimize_comv_buffer-v2-1-69379a59e17d@oss.qualcomm.com

Changes in v2:
- Update commit description (Bryan)
- Update hfi comv buffer count value to use actual num_comv count which
  used for buffer calculation to avoid any overhead or fixed values (Vikash)
- Link to v1: https://lore.kernel.org/r/20260421-optimize_comv_buffer-v1-1-7c9a24da3ad3@oss.qualcomm.com
---
 .../platform/qcom/iris/iris_hfi_gen2_command.c     | 14 +++-----------
 .../platform/qcom/iris/iris_platform_common.h      |  1 -
 .../media/platform/qcom/iris/iris_platform_gen2.c  |  1 -
 .../platform/qcom/iris/iris_platform_qcs8300.h     |  1 -
 drivers/media/platform/qcom/iris/iris_vdec.c       |  3 ++-
 drivers/media/platform/qcom/iris/iris_vpu_buffer.c | 22 ++++++++++++++++++++--
 6 files changed, 25 insertions(+), 17 deletions(-)

diff --git a/drivers/media/platform/qcom/iris/iris_hfi_gen2_command.c b/drivers/media/platform/qcom/iris/iris_hfi_gen2_command.c
index 30bfd90d423b..e53b1fca98bd 100644
--- a/drivers/media/platform/qcom/iris/iris_hfi_gen2_command.c
+++ b/drivers/media/platform/qcom/iris/iris_hfi_gen2_command.c
@@ -10,7 +10,6 @@
 
 #define UNSPECIFIED_COLOR_FORMAT 5
 #define NUM_SYS_INIT_PACKETS 8
-#define NUM_COMV_AV1 18
 
 #define SYS_INIT_PKT_SIZE (sizeof(struct iris_hfi_header) + \
 	NUM_SYS_INIT_PACKETS * (sizeof(struct iris_hfi_packet) + sizeof(u32)))
@@ -1207,18 +1206,11 @@ static u32 iris_hfi_gen2_buf_type_from_driver(u32 domain, enum iris_buffer_type
 
 static int iris_set_num_comv(struct iris_inst *inst)
 {
-	struct platform_inst_caps *caps;
+	u32 num_comv = inst->buffers[BUF_OUTPUT].min_count;
 	struct iris_core *core = inst->core;
-	u32 num_comv;
 
-	caps = core->iris_platform_data->inst_caps;
-
-	/*
-	 * AV1 needs more comv buffers than other codecs.
-	 * Update accordingly.
-	 */
-	num_comv = (inst->codec == V4L2_PIX_FMT_AV1) ?
-				NUM_COMV_AV1 : caps->num_comv;
+	if (inst->fw_min_count)
+		num_comv = inst->fw_min_count;
 
 	return core->hfi_ops->session_set_property(inst,
 						   HFI_PROP_COMV_BUFFER_COUNT,
diff --git a/drivers/media/platform/qcom/iris/iris_platform_common.h b/drivers/media/platform/qcom/iris/iris_platform_common.h
index 5a489917580e..2cda8cbba8d6 100644
--- a/drivers/media/platform/qcom/iris/iris_platform_common.h
+++ b/drivers/media/platform/qcom/iris/iris_platform_common.h
@@ -95,7 +95,6 @@ struct platform_inst_caps {
 	u32 mb_cycles_vpp;
 	u32 mb_cycles_fw;
 	u32 mb_cycles_fw_vpp;
-	u32 num_comv;
 	u32 max_frame_rate;
 	u32 max_operating_rate;
 };
diff --git a/drivers/media/platform/qcom/iris/iris_platform_gen2.c b/drivers/media/platform/qcom/iris/iris_platform_gen2.c
index 5da90d47f9c6..80222fb9da7b 100644
--- a/drivers/media/platform/qcom/iris/iris_platform_gen2.c
+++ b/drivers/media/platform/qcom/iris/iris_platform_gen2.c
@@ -751,7 +751,6 @@ static struct platform_inst_caps platform_inst_cap_sm8550 = {
 	.mb_cycles_vpp = 200,
 	.mb_cycles_fw = 489583,
 	.mb_cycles_fw_vpp = 66234,
-	.num_comv = 0,
 	.max_frame_rate = MAXIMUM_FPS,
 	.max_operating_rate = MAXIMUM_FPS,
 };
diff --git a/drivers/media/platform/qcom/iris/iris_platform_qcs8300.h b/drivers/media/platform/qcom/iris/iris_platform_qcs8300.h
index 61025f1e965b..3cfecae80d1e 100644
--- a/drivers/media/platform/qcom/iris/iris_platform_qcs8300.h
+++ b/drivers/media/platform/qcom/iris/iris_platform_qcs8300.h
@@ -15,7 +15,6 @@ static struct platform_inst_caps platform_inst_cap_qcs8300 = {
 	.mb_cycles_vpp = 200,
 	.mb_cycles_fw = 326389,
 	.mb_cycles_fw_vpp = 44156,
-	.num_comv = 0,
 	.max_frame_rate = MAXIMUM_FPS,
 	.max_operating_rate = MAXIMUM_FPS,
 };
diff --git a/drivers/media/platform/qcom/iris/iris_vdec.c b/drivers/media/platform/qcom/iris/iris_vdec.c
index 719217399a30..bab5c66df2d3 100644
--- a/drivers/media/platform/qcom/iris/iris_vdec.c
+++ b/drivers/media/platform/qcom/iris/iris_vdec.c
@@ -24,7 +24,7 @@ int iris_vdec_inst_init(struct iris_inst *inst)
 	inst->fmt_src = kzalloc_obj(*inst->fmt_src);
 	inst->fmt_dst = kzalloc_obj(*inst->fmt_dst);
 
-	inst->fw_min_count = MIN_BUFFERS;
+	inst->fw_min_count = 0;
 
 	f = inst->fmt_src;
 	f->type = V4L2_BUF_TYPE_VIDEO_OUTPUT_MPLANE;
@@ -258,6 +258,7 @@ int iris_vdec_s_fmt(struct iris_inst *inst, struct v4l2_format *f)
 		/* Update capture format based on new ip w/h */
 		output_fmt->fmt.pix_mp.width = ALIGN(f->fmt.pix_mp.width, 128);
 		output_fmt->fmt.pix_mp.height = ALIGN(f->fmt.pix_mp.height, 32);
+		inst->buffers[BUF_OUTPUT].min_count = iris_vpu_buf_count(inst, BUF_OUTPUT);
 		inst->buffers[BUF_OUTPUT].size = iris_get_buffer_size(inst, BUF_OUTPUT);
 
 		inst->crop.left = 0;
diff --git a/drivers/media/platform/qcom/iris/iris_vpu_buffer.c b/drivers/media/platform/qcom/iris/iris_vpu_buffer.c
index 9270422c1601..7ac6d9e49584 100644
--- a/drivers/media/platform/qcom/iris/iris_vpu_buffer.c
+++ b/drivers/media/platform/qcom/iris/iris_vpu_buffer.c
@@ -731,6 +731,24 @@ static u32 iris_vpu_dec_comv_size(struct iris_inst *inst)
 	u32 height = f->fmt.pix_mp.height;
 	u32 width = f->fmt.pix_mp.width;
 
+	if (inst->codec == V4L2_PIX_FMT_H264)
+		return hfi_buffer_comv_h264d(width, height, num_comv);
+	else if (inst->codec == V4L2_PIX_FMT_HEVC)
+		return hfi_buffer_comv_h265d(width, height, num_comv);
+
+	return 0;
+}
+
+static u32 iris_vpu3x_4x_dec_comv_size(struct iris_inst *inst)
+{
+	u32 num_comv = inst->buffers[BUF_OUTPUT].min_count;
+	struct v4l2_format *f = inst->fmt_src;
+	u32 height = f->fmt.pix_mp.height;
+	u32 width = f->fmt.pix_mp.width;
+
+	if (inst->fw_min_count)
+		num_comv = inst->fw_min_count;
+
 	if (inst->codec == V4L2_PIX_FMT_H264)
 		return hfi_buffer_comv_h264d(width, height, num_comv);
 	else if (inst->codec == V4L2_PIX_FMT_HEVC)
@@ -2025,7 +2043,7 @@ u32 iris_vpu_buf_size(struct iris_inst *inst, enum iris_buffer_type buffer_type)
 
 	static const struct iris_vpu_buf_type_handle dec_internal_buf_type_handle[] = {
 		{BUF_BIN,         iris_vpu_dec_bin_size             },
-		{BUF_COMV,        iris_vpu_dec_comv_size            },
+		{BUF_COMV,        iris_vpu3x_4x_dec_comv_size       },
 		{BUF_NON_COMV,    iris_vpu_dec_non_comv_size        },
 		{BUF_LINE,        iris_vpu_dec_line_size            },
 		{BUF_PERSIST,     iris_vpu_dec_persist_size         },
@@ -2098,7 +2116,7 @@ u32 iris_vpu4x_buf_size(struct iris_inst *inst, enum iris_buffer_type buffer_typ
 
 	static const struct iris_vpu_buf_type_handle dec_internal_buf_type_handle[] = {
 		{BUF_BIN,         iris_vpu_dec_bin_size         },
-		{BUF_COMV,        iris_vpu_dec_comv_size        },
+		{BUF_COMV,        iris_vpu3x_4x_dec_comv_size   },
 		{BUF_NON_COMV,    iris_vpu_dec_non_comv_size    },
 		{BUF_LINE,        iris_vpu4x_dec_line_size      },
 		{BUF_PERSIST,     iris_vpu4x_dec_persist_size   },

---
base-commit: 254f49634ee16a731174d2ae34bc50bd5f45e731
change-id: 20260421-optimize_comv_buffer-ae7107673609

Best regards,
-- 
Vishnu Reddy <busanna.reddy@oss.qualcomm.com>


^ permalink raw reply related	[flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-05-06  7:24 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-05-06  6:12 [PATCH v3] media: iris: optimize COMV buffer allocation for VPU3x and VPU4x Vishnu Reddy
2026-05-06  7:23 ` Vikash Garodia

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox