Re: [PATCH] media: iris: optimize COMV buffer allocation for VPU3x and VPU4x
From: Vishnu Reddy
Date: Fri Apr 24 2026 - 05:17:59 EST
On 4/22/2026 1:12 PM, Vikash Garodia wrote:
>
>
> On 4/21/2026 2:31 PM, Bryan O'Donoghue wrote:
>> On 21/04/2026 07:41, Vishnu Reddy wrote:
>>> The existing iris_vpu_dec_comv_size() used VIDEO_MAX_FRAME (32) as
>>> num_comv count unconditionally when calculating the COMV buffer size.
>>> This resulted in an oversized COMV buffer allocation throughout decode
>>> session, wasting memory regardless of actual number of buffers required.
>>
>> You should define what a COMV buffer is before talking about how you are
>> changing it, i.e. define the term Co-located Motion Vector (CMOV) and then
>> use the abbreviation CMOV liberally as you wish.
>>
>>> For VPU3x and VPU4x platforms, introduce iris_vpu3x_4x_dec_comv_size() to
>>> replace iris_vpu_dec_comv_size(). It derives num_comv dynamically, it
>>
>> "These derive num_cmove dynamically"
>>
>>> uses inst->fw_min_count once the firmware has reported its minimum buffer
>>> requirements, and fallback to inst->buffers[BUF_OUTPUT].min_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.
>>
>> Cool nice fix.
>>
>
> Indeed, a good saving.
>
>>>
>>> Signed-off-by: Vishnu Reddy <busanna.reddy@xxxxxxxxxxxxxxxx>
>>> ---
>>> drivers/media/platform/qcom/iris/iris_vdec.c | 2 +-
>>> drivers/media/platform/qcom/iris/iris_vpu_buffer.c | 24 ++++++++++++
>>> +++++++---
>>> drivers/media/platform/qcom/iris/iris_vpu_buffer.h | 1 +
>>> 3 files changed, 23 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/media/platform/qcom/iris/iris_vdec.c b/drivers/
>>> media/platform/qcom/iris/iris_vdec.c
>>> index 719217399a30..f433065e08b2 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;
>>> diff --git a/drivers/media/platform/qcom/iris/iris_vpu_buffer.c b/
>>> drivers/media/platform/qcom/iris/iris_vpu_buffer.c
>>> index 9270422c1601..57237543b229 100644
>>> --- a/drivers/media/platform/qcom/iris/iris_vpu_buffer.c
>>> +++ b/drivers/media/platform/qcom/iris/iris_vpu_buffer.c
>>> @@ -731,6 +731,23 @@ 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)
>>> +{
>>> + struct v4l2_format *f = inst->fmt_src;
>>> + u32 height = f->fmt.pix_mp.height;
>>> + u32 width = f->fmt.pix_mp.width;
>>> + u32 num_comv;
>>> +
>>> + num_comv = inst->fw_min_count ? inst->fw_min_count : inst-
>>> >buffers[BUF_OUTPUT].min_count;
>>
>> Please just if/else this though its far easier to read/understand that way.
>>
>>> +
>>> if (inst->codec == V4L2_PIX_FMT_H264)
>>> return hfi_buffer_comv_h264d(width, height, num_comv);
>>> else if (inst->codec == V4L2_PIX_FMT_HEVC)
>>> @@ -739,7 +756,8 @@ static u32 iris_vpu_dec_comv_size(struct iris_inst *inst)
>>> if (inst->fw_caps[DRAP].value)
>>> return 0;
>>> else
>>> - return hfi_buffer_comv_av1d(width, height, num_comv);
>>> + return hfi_buffer_comv_av1d(width, height,
>>> + num_comv + AV1D_COMV_BUFFER_OVERHEAD);
>>> }
>>>
>>> return 0;
>>> @@ -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 },
>>> diff --git a/drivers/media/platform/qcom/iris/iris_vpu_buffer.h b/
>>> drivers/media/platform/qcom/iris/iris_vpu_buffer.h
>>> index 12640eb5ed8c..7a9cc1c92da3 100644
>>> --- a/drivers/media/platform/qcom/iris/iris_vpu_buffer.h
>>> +++ b/drivers/media/platform/qcom/iris/iris_vpu_buffer.h
>>> @@ -110,6 +110,7 @@ struct iris_inst;
>>> #define MAX_PE_NBR_DATA_LCU16_LINE_BUFFER_SIZE 96
>>> #define AV1D_NUM_HW_PIC_BUF 16
>>> #define AV1D_NUM_FRAME_HEADERS 16
>>> +#define AV1D_COMV_BUFFER_OVERHEAD 7
>>
>> Whats this ? Why is there a new seven byte overhead ? Does it represent a
>> header, an alignment ?
>
> Vishnu, pls check if we need to add this as initial count was 18 [1] ? What if
> initial count was 11 [2], and post DRC, fw_min_count would be 11 too for AV1d,
> so the overhead can be avoided.
>
> [1]
> https://elixir.bootlin.com/linux/v7.0-rc7/source/drivers/media/platform/qcom/iris/iris_hfi_gen2_command.c#L1220
>
> [2]
> https://elixir.bootlin.com/linux/v7.0-rc7/source/drivers/media/platform/qcom/iris/iris_vpu_buffer.c#L2157
>
Thanks for the information, will check and update in next revision.
> Regards,
> Vikash
>>
>> An overhead can mean anything.
>>
>>> #define SIZE_AV1D_SEQUENCE_HEADER 768
>>> #define SIZE_AV1D_METADATA 512
>>> #define SIZE_AV1D_FRAME_HEADER 1280
>>>
>>> ---
>>> base-commit: 4fbeef21f5387234111b5d52924e77757626faa5
>>> change-id: 20260421-optimize_comv_buffer-ae7107673609
>>>
>>> Best regards,
>>> --
>>> Vishnu Reddy <busanna.reddy@xxxxxxxxxxxxxxxx>
>>>
>>
>>
>