| Message ID | 20221009184824.457416-1-marijn.suijten@somainline.org |
|---|---|
| Headers | show |
| Series | drm/msm: Fix math issues in MSM DSC implementation | expand |
On 09/10/2022 21:48, Marijn Suijten wrote: > Multiplying a value by 2 and adding 1 to it always results in a value > that is uneven, and that 1 gets truncated immediately when performing > integer division by 2 again. There is no "rounding" possible here. > > After that target_bpp_x16 is used to store a multiplication of > bits_per_pixel by 16 which is only ever read to immediately be divided > by 16 again, and is elided in much the same way. > > Fixes: b9080324d6ca ("drm/msm/dsi: add support for dsc data") > Signed-off-by: Marijn Suijten <marijn.suijten@somainline.org> > Reviewed-by: Abhinav Kumar <quic_abhinavk@quicinc.com> Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org>
On 2022-10-09 20:50:54, Marijn Suijten wrote: Apologies. After attempting to recover from an unexpected interruption _right as I was sending this series_, this patch got sent twice as it only later appeared to also have made its way through in the first round [1], together with the cover letter and first five patches. [1]: https://lore.kernel.org/linux-arm-msm/20221009184824.457416-7-marijn.suijten@somainline.org/ - Marijn > As per the FIXME this code is entirely duplicate with what is already > provided inside drm_dsc_compute_rc_parameters(), and it is yet unknown > why this comment was put in place instead of resolved from the get-go. > Not only does it save on duplication, it would have also spared certain > issues. > > For example, this code from downstream assumed dsc->bits_per_pixel to > contain an integer value, whereas the upstream drm_dsc_config struct has > it with 4 fractional bits. drm_dsc_compute_rc_parameters() already > accounts for this feat, and the sole remaining use of > dsc->bits_per_pixel inside dsi_populate_dsc_params() will be addressed > in a separate patch. > > Fixes: b9080324d6ca ("drm/msm/dsi: add support for dsc data") > Signed-off-by: Marijn Suijten <marijn.suijten@somainline.org> > --- > drivers/gpu/drm/msm/dsi/dsi_host.c | 64 +++--------------------------- > 1 file changed, 6 insertions(+), 58 deletions(-) > > diff --git a/drivers/gpu/drm/msm/dsi/dsi_host.c b/drivers/gpu/drm/msm/dsi/dsi_host.c > index 83cde4d62b68..68c39debc22f 100644 > --- a/drivers/gpu/drm/msm/dsi/dsi_host.c > +++ b/drivers/gpu/drm/msm/dsi/dsi_host.c > @@ -21,6 +21,7 @@ > > #include <video/mipi_display.h> > > +#include <drm/display/drm_dsc_helper.h> > #include <drm/drm_of.h> > > #include "dsi.h" > @@ -1771,14 +1772,6 @@ static char bpg_offset[DSC_NUM_BUF_RANGES] = { > > static int dsi_populate_dsc_params(struct drm_dsc_config *dsc) > { > - int mux_words_size; > - int groups_per_line, groups_total; > - int min_rate_buffer_size; > - int hrd_delay; > - int pre_num_extra_mux_bits, num_extra_mux_bits; > - int slice_bits; > - int data; > - int final_value, final_scale; > int i; > > dsc->rc_model_size = 8192; > @@ -1804,11 +1797,11 @@ static int dsi_populate_dsc_params(struct drm_dsc_config *dsc) > if (dsc->bits_per_pixel != 8) > dsc->initial_offset = 2048; /* bpp = 12 */ > > - mux_words_size = 48; /* bpc == 8/10 */ > - if (dsc->bits_per_component == 12) > - mux_words_size = 64; > + if (dsc->bits_per_component <= 10) > + dsc->mux_word_size = DSC_MUX_WORD_SIZE_8_10_BPC; > + else > + dsc->mux_word_size = DSC_MUX_WORD_SIZE_12_BPC; > > - dsc->mux_word_size = mux_words_size; > dsc->initial_xmit_delay = 512; > dsc->initial_scale_value = 32; > dsc->first_line_bpg_offset = 12; > @@ -1820,52 +1813,7 @@ static int dsi_populate_dsc_params(struct drm_dsc_config *dsc) > dsc->rc_quant_incr_limit0 = 11; > dsc->rc_quant_incr_limit1 = 11; > > - /* FIXME: need to call drm_dsc_compute_rc_parameters() so that rest of > - * params are calculated > - */ > - groups_per_line = DIV_ROUND_UP(dsc->slice_width, 3); > - dsc->slice_chunk_size = DIV_ROUND_UP(dsc->slice_width * dsc->bits_per_pixel, 8); > - > - /* rbs-min */ > - min_rate_buffer_size = dsc->rc_model_size - dsc->initial_offset + > - dsc->initial_xmit_delay * dsc->bits_per_pixel + > - groups_per_line * dsc->first_line_bpg_offset; > - > - hrd_delay = DIV_ROUND_UP(min_rate_buffer_size, dsc->bits_per_pixel); > - > - dsc->initial_dec_delay = hrd_delay - dsc->initial_xmit_delay; > - > - dsc->initial_scale_value = 8 * dsc->rc_model_size / > - (dsc->rc_model_size - dsc->initial_offset); > - > - slice_bits = 8 * dsc->slice_chunk_size * dsc->slice_height; > - > - groups_total = groups_per_line * dsc->slice_height; > - > - data = dsc->first_line_bpg_offset * 2048; > - > - dsc->nfl_bpg_offset = DIV_ROUND_UP(data, (dsc->slice_height - 1)); > - > - pre_num_extra_mux_bits = 3 * (mux_words_size + (4 * dsc->bits_per_component + 4) - 2); > - > - num_extra_mux_bits = pre_num_extra_mux_bits - (mux_words_size - > - ((slice_bits - pre_num_extra_mux_bits) % mux_words_size)); > - > - data = 2048 * (dsc->rc_model_size - dsc->initial_offset + num_extra_mux_bits); > - dsc->slice_bpg_offset = DIV_ROUND_UP(data, groups_total); > - > - data = dsc->initial_xmit_delay * dsc->bits_per_pixel; > - final_value = dsc->rc_model_size - data + num_extra_mux_bits; > - dsc->final_offset = final_value; > - > - final_scale = 8 * dsc->rc_model_size / (dsc->rc_model_size - final_value); > - > - data = (final_scale - 9) * (dsc->nfl_bpg_offset + dsc->slice_bpg_offset); > - dsc->scale_increment_interval = (2048 * dsc->final_offset) / data; > - > - dsc->scale_decrement_interval = groups_per_line / (dsc->initial_scale_value - 8); > - > - return 0; > + return drm_dsc_compute_rc_parameters(dsc); > } > > static int dsi_host_parse_dt(struct msm_dsi_host *msm_host) > -- > 2.38.0 >
On 09/10/2022 21:53, Marijn Suijten wrote: > drm_dsc_config's bits_per_pixel field holds a fractional value with 4 > bits, which all panel drivers should adhere to for > drm_dsc_pps_payload_pack() to generate a valid payload. All code in the > DSI driver here seems to assume that this field doesn't contain any > fractional bits, hence resulting in the wrong values being computed. > Since none of the calculations leave any room for fractional bits or > seem to indicate any possible area of support, disallow such values > altogether. calculate_rc_params() in intel_vdsc.c performs an identical > bitshift to get at this integer value. > > Fixes: b9080324d6ca ("drm/msm/dsi: add support for dsc data") > Signed-off-by: Marijn Suijten <marijn.suijten@somainline.org> Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org> > --- > drivers/gpu/drm/msm/dsi/dsi_host.c | 19 ++++++++++++++----- > 1 file changed, 14 insertions(+), 5 deletions(-) >
Various removals of complex yet unnecessary math, fixing all uses of drm_dsc_config::bits_per_pixel to deal with the fact that this field includes four fractional bits, and finally making sure that range_bpg_offset contains values 6-bits wide to prevent overflows in drm_dsc_pps_payload_pack(). Altogether this series is responsible for solving _all_ Display Stream Compression issues and artifacts on the Sony Tama (sdm845) Akatsuki smartphone (2880x1440p). Changes since v2: - Generalize mux_word_size setting depending on bits_per_component; - Migrate DSI's DSC calculations to drm_dsc_compute_rc_parameters(), implicitly addressing existing math issues; - Disallow any bits_per_component other than 8, until hardcoded values are updated and tested to support such cases. v2: https://lore.kernel.org/linux-arm-msm/20221005181657.784375-1-marijn.suijten@somainline.org/T/#u Changes since v1: - Propagate r-b's, except (obviously) in patches that were (heavily) modified; - Remove accidental debug code in dsi_cmd_dma_add; - Move Range BPG Offset masking out of DCS PPS packing, back into the DSI driver when it is assigned to drm_dsc_config (this series is now strictly focusing on drm/msm again); - Replace modulo-check resulting in conditional increment with DIV_ROUND_UP; - Remove repeated calculation of slice_chunk_size; - Use u16 instead of int when handling bits_per_pixel; - Use DRM_DEV_ERROR instead of pr_err in DSI code; - Also remove redundant target_bpp_x16 variable. v1: https://lore.kernel.org/linux-arm-msm/20221001190807.358691-1-marijn.suijten@somainline.org/T/#u Marijn Suijten (10): drm/msm/dsi: Remove useless math in DSC calculations drm/msm/dsi: Remove repeated calculation of slice_per_intf drm/msm/dsi: Use DIV_ROUND_UP instead of conditional increment on modulo drm/msm/dsi: Reuse earlier computed dsc->slice_chunk_size drm/msm/dsi: Appropriately set dsc->mux_word_size based on bpc drm/msm/dsi: Migrate to drm_dsc_compute_rc_parameters() drm/msm/dsi: Disallow 8 BPC DSC configuration for alternative BPC values drm/msm/dsi: Account for DSC's bits_per_pixel having 4 fractional bits drm/msm/dpu1: Account for DSC's bits_per_pixel having 4 fractional bits drm/msm/dsi: Prevent signed BPG offsets from bleeding into adjacent bits drivers/gpu/drm/msm/disp/dpu1/dpu_hw_dsc.c | 11 +- drivers/gpu/drm/msm/dsi/dsi_host.c | 121 ++++++--------------- 2 files changed, 37 insertions(+), 95 deletions(-) -- 2.38.0