Re: [linux-sunxi] [PATCH v2 2/2] media: cedrus: Add H264 decoding support

["Jernej Škrabec" <jernej.skrabec@gmail.com>]

Dne torek, 27. november 2018 ob 17:30:00 CET je Jernej Škrabec napisal(a):
> Dne torek, 27. november 2018 ob 16:50:28 CET je Maxime Ripard napisal(a):
> > Hi Jernej,
> > 
> > Thanks for your review!
> > 
> > On Sat, Nov 24, 2018 at 09:43:43PM +0100, Jernej Škrabec wrote:
> > > > +enum cedrus_h264_sram_off {
> > > > +	CEDRUS_SRAM_H264_PRED_WEIGHT_TABLE	= 0x000,
> > > > +	CEDRUS_SRAM_H264_FRAMEBUFFER_LIST	= 0x100,
> > > > +	CEDRUS_SRAM_H264_REF_LIST_0		= 0x190,
> > > > +	CEDRUS_SRAM_H264_REF_LIST_1		= 0x199,
> > > > +	CEDRUS_SRAM_H264_SCALING_LIST_8x8	= 0x200,
> > > > +	CEDRUS_SRAM_H264_SCALING_LIST_4x4	= 0x218,
> > > 
> > > I triple checked above address and it should be 0x220. For easier
> > > implementation later, you might want to add second scaling list address
> > > for
> > > 8x8 at 0x210. Then you can do something like:
> > > 
> > > cedrus_h264_write_sram(dev, CEDRUS_SRAM_H264_SCALING_LIST_8x8_0,
> > > 
> > > 			       scaling->scaling_list_8x8[0],
> > > 			       sizeof(scaling->scaling_list_8x8[0]));
> > > 
> > > cedrus_h264_write_sram(dev, CEDRUS_SRAM_H264_SCALING_LIST_8x8_1,
> > > 
> > > 			       scaling->scaling_list_8x8[3],
> > > 			       sizeof(scaling->scaling_list_8x8[0]));
> > > 
> > > cedrus_h264_write_sram(dev, CEDRUS_SRAM_H264_SCALING_LIST_4x4,
> > > 
> > > 			       scaling->scaling_list_4x4,
> > > 			       sizeof(scaling->scaling_list_4x4));
> > > 
> > > I know that it's not implemented here, just FYI.
> > 
> > Ack. I guess I can just leave it out entirely for now, since it's not
> > implemented.
> > 
> > > > +static void cedrus_fill_ref_pic(struct cedrus_ctx *ctx,
> > > > +				struct cedrus_buffer *buf,
> > > > +				unsigned int top_field_order_cnt,
> > > > +				unsigned int bottom_field_order_cnt,
> > > > +				struct cedrus_h264_sram_ref_pic *pic)
> > > > +{
> > > > +	struct vb2_buffer *vbuf = &buf->m2m_buf.vb.vb2_buf;
> > > > +	unsigned int position = buf->codec.h264.position;
> > > > +
> > > > +	pic->top_field_order_cnt = top_field_order_cnt;
> > > > +	pic->bottom_field_order_cnt = bottom_field_order_cnt;
> > > > +	pic->frame_info = buf->codec.h264.pic_type << 8;
> > > > +
> > > > +	pic->luma_ptr = cedrus_buf_addr(vbuf, &ctx->dst_fmt, 0) -
> > > > PHYS_OFFSET;
> > > > +	pic->chroma_ptr = cedrus_buf_addr(vbuf, &ctx->dst_fmt, 1) -
> > > > PHYS_OFFSET;
> > > 
> > > I think subtracting PHYS_OFFSET breaks driver on H3 boards with 2 GiB of
> > > RAM. Isn't that unnecessary anyway due to
> > > 
> > > dev->dev->dma_pfn_offset = PHYS_PFN_OFFSET;
> > > 
> > > in cedrus_hw.c?
> > > 
> > > This comment is meant for all PHYS_OFFSET subtracting in this patch.
> > 
> > PHYS_OFFSET was needed on some older SoCs, and the dma_pfn_offset
> > trick wasn't working, I hacked it and forgot about it. I'll try to
> > figure it out for the next version.
> > 
> > > > +static void _cedrus_write_ref_list(struct cedrus_ctx *ctx,
> > > > +				   struct cedrus_run *run,
> > > > +				   const u8 *ref_list, u8 num_ref,
> > > > +				   enum cedrus_h264_sram_off sram)
> > > > +{
> > > > +	const struct v4l2_ctrl_h264_decode_param *decode =
> > > > run->h264.decode_param; +	struct vb2_queue *cap_q =
> > > > &ctx->fh.m2m_ctx->cap_q_ctx.q;
> > > > +	struct cedrus_dev *dev = ctx->dev;
> > > > +	u32 sram_array[CEDRUS_MAX_REF_IDX / sizeof(u32)];
> > > > +	unsigned int size, i;
> > > > +
> > > > +	memset(sram_array, 0, sizeof(sram_array));
> > > > +
> > > > +	for (i = 0; i < num_ref; i += 4) {
> > > > +		unsigned int j;
> > > > +
> > > > +		for (j = 0; j < 4; j++) {
> > > 
> > > I don't think you have to complicate with two loops here.
> > > cedrus_h264_write_sram() takes void* and it aligns to 4 anyway. So as
> > > long
> > > input buffer is multiple of 4 (u8[CEDRUS_MAX_REF_IDX] qualifies for
> > > that),
> > > you can use single for loop with "u8 sram_array[CEDRUS_MAX_REF_IDX]".
> > > This should make code much more readable.
> > 
> > This wasn't really about the alignment, but in order to get the
> > offsets in the u32 and the array more easily.
> > 
> > Breaking out the loop will make that computation less easy on the eye,
> > so I guess it's very subjective.
> 
> For some strange reason, code below fixes decoding issue from one of my test
> samples. This is what I actually meant with 1 loop approach:
> 
> static void _cedrus_write_ref_list(struct cedrus_ctx *ctx,
> 				   struct cedrus_run *run,
> 				   const u8 *ref_list, u8 num_ref,
> 				   enum cedrus_h264_sram_off sram)
> {
> 	const struct v4l2_ctrl_h264_decode_param *decode = run->h264.decode_param;
> 	struct vb2_queue *cap_q = &ctx->fh.m2m_ctx->cap_q_ctx.q;
> 	struct cedrus_dev *dev = ctx->dev;
> 	u8 sram_array[CEDRUS_MAX_REF_IDX];
> 	unsigned int i;
> 
> 	memset(sram_array, 0, sizeof(sram_array));
> 	num_ref = min(num_ref, (u8)CEDRUS_MAX_REF_IDX);
> 
> 	for (i = 0; i < num_ref; i++) {
> 		const struct v4l2_h264_dpb_entry *dpb;
> 		const struct cedrus_buffer *cedrus_buf;
> 		const struct vb2_v4l2_buffer *ref_buf;
> 		unsigned int position;
> 		int buf_idx;
> 		u8 dpb_idx;
> 
> 		dpb_idx = ref_list[i];
> 		dpb = &decode->dpb[dpb_idx];
> 
> 		if (!(dpb->flags & V4L2_H264_DPB_ENTRY_FLAG_ACTIVE))
> 			continue;
> 
> 		buf_idx = vb2_find_tag(cap_q, dpb->tag, 0);
> 		if (buf_idx < 0)
> 			continue;
> 
> 		ref_buf = to_vb2_v4l2_buffer(ctx->dst_bufs[buf_idx]);
> 		cedrus_buf = vb2_v4l2_to_cedrus_buffer(ref_buf);
> 		position = cedrus_buf->codec.h264.position;
> 
> 		sram_array[i] |= position << 1;
> 		if (ref_buf->field == V4L2_FIELD_BOTTOM)
> 			sram_array[i] |= BIT(0);
> 	}
> 
> 	cedrus_h264_write_sram(dev, sram, &sram_array, num_ref);
> }
> 
> IMO this code is easier to read.
> 
> > > > +			const struct v4l2_h264_dpb_entry *dpb;
> > > > +			const struct cedrus_buffer *cedrus_buf;
> > > > +			const struct vb2_v4l2_buffer *ref_buf;
> > > > +			unsigned int position;
> > > > +			int buf_idx;
> > > > +			u8 ref_idx = i + j;
> > > > +			u8 dpb_idx;
> > > > +
> > > > +			if (ref_idx >= num_ref)
> > > > +				break;
> > > > +
> > > > +			dpb_idx = ref_list[ref_idx];
> > > > +			dpb = &decode->dpb[dpb_idx];
> > > > +
> > > > +			if (!(dpb->flags & V4L2_H264_DPB_ENTRY_FLAG_ACTIVE))
> > > > +				continue;
> > > > +
> > > > +			buf_idx = vb2_find_tag(cap_q, dpb->tag, 0);
> > > > +			if (buf_idx < 0)
> > > > +				continue;
> > > > +
> > > > +			ref_buf = to_vb2_v4l2_buffer(ctx->dst_bufs[buf_idx]);
> > > > +			cedrus_buf = vb2_v4l2_to_cedrus_buffer(ref_buf);
> > > > +			position = cedrus_buf->codec.h264.position;
> > > > +
> > > > +			sram_array[i] |= position << (j * 8 + 1);
> > > > +			if (ref_buf->field == V4L2_FIELD_BOTTOM)
> > > 
> > > You newer set above flag to buffer so this will be always false.
> > 
> > As far as I know, the field is supposed to be set by the userspace.
> 
> How? I thought that only flags at queueing buffers can be set and there is
> no bottom/top flag.
> 
> > > > +	// sequence parameters
> > > > +	reg = BIT(19);
> > > 
> > > This one can be inferred from sps->chroma_format_idc.
> > 
> > I'll look into this
> 
> I'm using this:
> reg |= (sps->chroma_format_idc & 0x7) << 19;
> 
> Although I can't tell if I tested anything else than 1 there (same as it was
> before).
> 
> > > > +	reg |= (sps->pic_width_in_mbs_minus1 & 0xff) << 8;
> > > > +	reg |= sps->pic_height_in_map_units_minus1 & 0xff;
> > > > +	if (sps->flags & V4L2_H264_SPS_FLAG_FRAME_MBS_ONLY)
> > > > +		reg |= BIT(18);
> > > > +	if (sps->flags & V4L2_H264_SPS_FLAG_MB_ADAPTIVE_FRAME_FIELD)
> > > > +		reg |= BIT(17);
> > > > +	if (sps->flags & V4L2_H264_SPS_FLAG_DIRECT_8X8_INFERENCE)
> > > > +		reg |= BIT(16);
> > > > +	cedrus_write(dev, VE_H264_FRAME_SIZE, reg);
> > > > +
> > > > +	// slice parameters
> > > > +	reg = 0;
> > > > +	/*
> > > > +	 * FIXME: This bit marks all the frames as references. This
> > > > +	 * should probably be set based on nal_ref_idc, but the libva
> > > > +	 * doesn't pass that information along, so this is not always
> > > > +	 * available. We should find something else, maybe change the
> > > > +	 * kernel UAPI somehow?
> > > > +	 */
> > > > +	reg |= BIT(12);
> > > 
> > > I really think you should use nal_ref_idc here as it is in
> > > specification.
> > > You can still fake the data from libva backend. I don't think that any
> > > driver needs this for anything else than check if it is 0 or not.
> > 
> > Yeah, Tomasz suggested the same thing as a reply to the cover letter,
> > I'll change that in the next version.
> > 
> > > > +	reg |= (slice->slice_type & 0xf) << 8;
> > > > +	reg |= slice->cabac_init_idc & 0x3;
> > > > +	reg |= BIT(5);
> > > > +	if (slice->flags & V4L2_H264_SLICE_FLAG_FIELD_PIC)
> > > > +		reg |= BIT(4);
> > > > +	if (slice->flags & V4L2_H264_SLICE_FLAG_BOTTOM_FIELD)
> > > > +		reg |= BIT(3);
> > > > +	if (slice->flags & V4L2_H264_SLICE_FLAG_DIRECT_SPATIAL_MV_PRED)
> > > > +		reg |= BIT(2);
> > > > +	cedrus_write(dev, VE_H264_SLICE_HDR, reg);
> > > > +
> > > > +	reg = 0;
> > > 
> > > You might want to set bit 12 here, which enables active reference
> > > picture
> > > override. However, I'm not completely sure about that.
> > 
> > Did you find some videos that were broken because of this?
> 
> No, not really. That's why I don't really know if it is needed or not.

I found a flag in specs: num_ref_idx_active_override_flag
I guess VAAPI always give correct value, so this doesn't need to be set.

Best regards,
Jernej