drivers/gpu/drm/display/drm_dsc_helper.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2018 Intel Corp
   4  *
   5  * Author:
   6  * Manasi Navare <[email protected]>
   7  */
   8
   9 #include <linux/kernel.h>
  10 #include <linux/module.h>
  11 #include <linux/init.h>
  12 #include <linux/errno.h>
  13 #include <linux/byteorder/generic.h>
  14
  15 #include <drm/display/drm_dp_helper.h>
  16 #include <drm/display/drm_dsc_helper.h>
  17 #include <drm/drm_print.h>
  18
  19 /**
  20  * DOC: dsc helpers
  21  *
  22  * VESA specification for DP 1.4 adds a new feature called Display Stream
  23  * Compression (DSC) used to compress the pixel bits before sending it on
  24  * DP/eDP/MIPI DSI interface. DSC is required to be enabled so that the existing
  25  * display interfaces can support high resolutions at higher frames rates uisng
  26  * the maximum available link capacity of these interfaces.
  27  *
  28  * These functions contain some common logic and helpers to deal with VESA
  29  * Display Stream Compression standard required for DSC on Display Port/eDP or
  30  * MIPI display interfaces.
  31  */
  32
  33 /**
  34  * drm_dsc_dp_pps_header_init() - Initializes the PPS Header
  35  * for DisplayPort as per the DP 1.4 spec.
  36  * @pps_header: Secondary data packet header for DSC Picture
  37  *              Parameter Set as defined in &struct dp_sdp_header
  38  *
  39  * DP 1.4 spec defines the secondary data packet for sending the
  40  * picture parameter infoframes from the source to the sink.
  41  * This function populates the SDP header defined in
  42  * &struct dp_sdp_header.
  43  */
  44 void drm_dsc_dp_pps_header_init(struct dp_sdp_header *pps_header)
  45 {
  46         memset(pps_header, 0, sizeof(*pps_header));
  47
  48         pps_header->HB1 = DP_SDP_PPS;
  49         pps_header->HB2 = DP_SDP_PPS_HEADER_PAYLOAD_BYTES_MINUS_1;
  50 }
  51 EXPORT_SYMBOL(drm_dsc_dp_pps_header_init);
  52
  53 /**
  54  * drm_dsc_dp_rc_buffer_size - get rc buffer size in bytes
  55  * @rc_buffer_block_size: block size code, according to DPCD offset 62h
  56  * @rc_buffer_size: number of blocks - 1, according to DPCD offset 63h
  57  *
  58  * return:
  59  * buffer size in bytes, or 0 on invalid input
  60  */
  61 int drm_dsc_dp_rc_buffer_size(u8 rc_buffer_block_size, u8 rc_buffer_size)
  62 {
  63         int size = 1024 * (rc_buffer_size + 1);
  64
  65         switch (rc_buffer_block_size) {
  66         case DP_DSC_RC_BUF_BLK_SIZE_1:
  67                 return 1 * size;
  68         case DP_DSC_RC_BUF_BLK_SIZE_4:
  69                 return 4 * size;
  70         case DP_DSC_RC_BUF_BLK_SIZE_16:
  71                 return 16 * size;
  72         case DP_DSC_RC_BUF_BLK_SIZE_64:
  73                 return 64 * size;
  74         default:
  75                 return 0;
  76         }
  77 }
  78 EXPORT_SYMBOL(drm_dsc_dp_rc_buffer_size);
  79
  80 /**
  81  * drm_dsc_pps_payload_pack() - Populates the DSC PPS
  82  *
  83  * @pps_payload:
  84  * Bitwise struct for DSC Picture Parameter Set. This is defined
  85  * by &struct drm_dsc_picture_parameter_set
  86  * @dsc_cfg:
  87  * DSC Configuration data filled by driver as defined by
  88  * &struct drm_dsc_config
  89  *
  90  * DSC source device sends a picture parameter set (PPS) containing the
  91  * information required by the sink to decode the compressed frame. Driver
  92  * populates the DSC PPS struct using the DSC configuration parameters in
  93  * the order expected by the DSC Display Sink device. For the DSC, the sink
  94  * device expects the PPS payload in big endian format for fields
  95  * that span more than 1 byte.
  96  */
  97 void drm_dsc_pps_payload_pack(struct drm_dsc_picture_parameter_set *pps_payload,
  98                                 const struct drm_dsc_config *dsc_cfg)
  99 {
 100         int i;
 101
 102         /* Protect against someone accidentally changing struct size */
 103         BUILD_BUG_ON(sizeof(*pps_payload) !=
 104                      DP_SDP_PPS_HEADER_PAYLOAD_BYTES_MINUS_1 + 1);
 105
 106         memset(pps_payload, 0, sizeof(*pps_payload));
 107
 108         /* PPS 0 */
 109         pps_payload->dsc_version =
 110                 dsc_cfg->dsc_version_minor |
 111                 dsc_cfg->dsc_version_major << DSC_PPS_VERSION_MAJOR_SHIFT;
 112
 113         /* PPS 1, 2 is 0 */
 114
 115         /* PPS 3 */
 116         pps_payload->pps_3 =
 117                 dsc_cfg->line_buf_depth |
 118                 dsc_cfg->bits_per_component << DSC_PPS_BPC_SHIFT;
 119
 120         /* PPS 4 */
 121         pps_payload->pps_4 =
 122                 ((dsc_cfg->bits_per_pixel & DSC_PPS_BPP_HIGH_MASK) >>
 123                  DSC_PPS_MSB_SHIFT) |
 124                 dsc_cfg->vbr_enable << DSC_PPS_VBR_EN_SHIFT |
 125                 dsc_cfg->simple_422 << DSC_PPS_SIMPLE422_SHIFT |
 126                 dsc_cfg->convert_rgb << DSC_PPS_CONVERT_RGB_SHIFT |
 127                 dsc_cfg->block_pred_enable << DSC_PPS_BLOCK_PRED_EN_SHIFT;
 128
 129         /* PPS 5 */
 130         pps_payload->bits_per_pixel_low =
 131                 (dsc_cfg->bits_per_pixel & DSC_PPS_LSB_MASK);
 132
 133         /*
 134          * The DSC panel expects the PPS packet to have big endian format
 135          * for data spanning 2 bytes. Use a macro cpu_to_be16() to convert
 136          * to big endian format. If format is little endian, it will swap
 137          * bytes to convert to Big endian else keep it unchanged.
 138          */
 139
 140         /* PPS 6, 7 */
 141         pps_payload->pic_height = cpu_to_be16(dsc_cfg->pic_height);
 142
 143         /* PPS 8, 9 */
 144         pps_payload->pic_width = cpu_to_be16(dsc_cfg->pic_width);
 145
 146         /* PPS 10, 11 */
 147         pps_payload->slice_height = cpu_to_be16(dsc_cfg->slice_height);
 148
 149         /* PPS 12, 13 */
 150         pps_payload->slice_width = cpu_to_be16(dsc_cfg->slice_width);
 151
 152         /* PPS 14, 15 */
 153         pps_payload->chunk_size = cpu_to_be16(dsc_cfg->slice_chunk_size);
 154
 155         /* PPS 16 */
 156         pps_payload->initial_xmit_delay_high =
 157                 ((dsc_cfg->initial_xmit_delay &
 158                   DSC_PPS_INIT_XMIT_DELAY_HIGH_MASK) >>
 159                  DSC_PPS_MSB_SHIFT);
 160
 161         /* PPS 17 */
 162         pps_payload->initial_xmit_delay_low =
 163                 (dsc_cfg->initial_xmit_delay & DSC_PPS_LSB_MASK);
 164
 165         /* PPS 18, 19 */
 166         pps_payload->initial_dec_delay =
 167                 cpu_to_be16(dsc_cfg->initial_dec_delay);
 168
 169         /* PPS 20 is 0 */
 170
 171         /* PPS 21 */
 172         pps_payload->initial_scale_value =
 173                 dsc_cfg->initial_scale_value;
 174
 175         /* PPS 22, 23 */
 176         pps_payload->scale_increment_interval =
 177                 cpu_to_be16(dsc_cfg->scale_increment_interval);
 178
 179         /* PPS 24 */
 180         pps_payload->scale_decrement_interval_high =
 181                 ((dsc_cfg->scale_decrement_interval &
 182                   DSC_PPS_SCALE_DEC_INT_HIGH_MASK) >>
 183                  DSC_PPS_MSB_SHIFT);
 184
 185         /* PPS 25 */
 186         pps_payload->scale_decrement_interval_low =
 187                 (dsc_cfg->scale_decrement_interval & DSC_PPS_LSB_MASK);
 188
 189         /* PPS 26[7:0], PPS 27[7:5] RESERVED */
 190
 191         /* PPS 27 */
 192         pps_payload->first_line_bpg_offset =
 193                 dsc_cfg->first_line_bpg_offset;
 194
 195         /* PPS 28, 29 */
 196         pps_payload->nfl_bpg_offset =
 197                 cpu_to_be16(dsc_cfg->nfl_bpg_offset);
 198
 199         /* PPS 30, 31 */
 200         pps_payload->slice_bpg_offset =
 201                 cpu_to_be16(dsc_cfg->slice_bpg_offset);
 202
 203         /* PPS 32, 33 */
 204         pps_payload->initial_offset =
 205                 cpu_to_be16(dsc_cfg->initial_offset);
 206
 207         /* PPS 34, 35 */
 208         pps_payload->final_offset = cpu_to_be16(dsc_cfg->final_offset);
 209
 210         /* PPS 36 */
 211         pps_payload->flatness_min_qp = dsc_cfg->flatness_min_qp;
 212
 213         /* PPS 37 */
 214         pps_payload->flatness_max_qp = dsc_cfg->flatness_max_qp;
 215
 216         /* PPS 38, 39 */
 217         pps_payload->rc_model_size = cpu_to_be16(dsc_cfg->rc_model_size);
 218
 219         /* PPS 40 */
 220         pps_payload->rc_edge_factor = DSC_RC_EDGE_FACTOR_CONST;
 221
 222         /* PPS 41 */
 223         pps_payload->rc_quant_incr_limit0 =
 224                 dsc_cfg->rc_quant_incr_limit0;
 225
 226         /* PPS 42 */
 227         pps_payload->rc_quant_incr_limit1 =
 228                 dsc_cfg->rc_quant_incr_limit1;
 229
 230         /* PPS 43 */
 231         pps_payload->rc_tgt_offset = DSC_RC_TGT_OFFSET_LO_CONST |
 232                 DSC_RC_TGT_OFFSET_HI_CONST << DSC_PPS_RC_TGT_OFFSET_HI_SHIFT;
 233
 234         /* PPS 44 - 57 */
 235         for (i = 0; i < DSC_NUM_BUF_RANGES - 1; i++)
 236                 pps_payload->rc_buf_thresh[i] =
 237                         dsc_cfg->rc_buf_thresh[i];
 238
 239         /* PPS 58 - 87 */
 240         /*
 241          * For DSC sink programming the RC Range parameter fields
 242          * are as follows: Min_qp[15:11], max_qp[10:6], offset[5:0]
 243          */
 244         for (i = 0; i < DSC_NUM_BUF_RANGES; i++) {
 245                 pps_payload->rc_range_parameters[i] =
 246                         cpu_to_be16((dsc_cfg->rc_range_params[i].range_min_qp <<
 247                                      DSC_PPS_RC_RANGE_MINQP_SHIFT) |
 248                                     (dsc_cfg->rc_range_params[i].range_max_qp <<
 249                                      DSC_PPS_RC_RANGE_MAXQP_SHIFT) |
 250                                     (dsc_cfg->rc_range_params[i].range_bpg_offset));
 251         }
 252
 253         /* PPS 88 */
 254         pps_payload->native_422_420 = dsc_cfg->native_422 |
 255                 dsc_cfg->native_420 << DSC_PPS_NATIVE_420_SHIFT;
 256
 257         /* PPS 89 */
 258         pps_payload->second_line_bpg_offset =
 259                 dsc_cfg->second_line_bpg_offset;
 260
 261         /* PPS 90, 91 */
 262         pps_payload->nsl_bpg_offset =
 263                 cpu_to_be16(dsc_cfg->nsl_bpg_offset);
 264
 265         /* PPS 92, 93 */
 266         pps_payload->second_line_offset_adj =
 267                 cpu_to_be16(dsc_cfg->second_line_offset_adj);
 268
 269         /* PPS 94 - 127 are O */
 270 }
 271 EXPORT_SYMBOL(drm_dsc_pps_payload_pack);
 272
 273 /**
 274  * drm_dsc_compute_rc_parameters() - Write rate control
 275  * parameters to the dsc configuration defined in
 276  * &struct drm_dsc_config in accordance with the DSC 1.2
 277  * specification. Some configuration fields must be present
 278  * beforehand.
 279  *
 280  * @vdsc_cfg:
 281  * DSC Configuration data partially filled by driver
 282  */
 283 int drm_dsc_compute_rc_parameters(struct drm_dsc_config *vdsc_cfg)
 284 {
 285         unsigned long groups_per_line = 0;
 286         unsigned long groups_total = 0;
 287         unsigned long num_extra_mux_bits = 0;
 288         unsigned long slice_bits = 0;
 289         unsigned long hrd_delay = 0;
 290         unsigned long final_scale = 0;
 291         unsigned long rbs_min = 0;
 292
 293         if (vdsc_cfg->native_420 || vdsc_cfg->native_422) {
 294                 /* Number of groups used to code each line of a slice */
 295                 groups_per_line = DIV_ROUND_UP(vdsc_cfg->slice_width / 2,
 296                                                DSC_RC_PIXELS_PER_GROUP);
 297
 298                 /* chunksize in Bytes */
 299                 vdsc_cfg->slice_chunk_size = DIV_ROUND_UP(vdsc_cfg->slice_width / 2 *
 300                                                           vdsc_cfg->bits_per_pixel,
 301                                                           (8 * 16));
 302         } else {
 303                 /* Number of groups used to code each line of a slice */
 304                 groups_per_line = DIV_ROUND_UP(vdsc_cfg->slice_width,
 305                                                DSC_RC_PIXELS_PER_GROUP);
 306
 307                 /* chunksize in Bytes */
 308                 vdsc_cfg->slice_chunk_size = DIV_ROUND_UP(vdsc_cfg->slice_width *
 309                                                           vdsc_cfg->bits_per_pixel,
 310                                                           (8 * 16));
 311         }
 312
 313         if (vdsc_cfg->convert_rgb)
 314                 num_extra_mux_bits = 3 * (vdsc_cfg->mux_word_size +
 315                                           (4 * vdsc_cfg->bits_per_component + 4)
 316                                           - 2);
 317         else if (vdsc_cfg->native_422)
 318                 num_extra_mux_bits = 4 * vdsc_cfg->mux_word_size +
 319                         (4 * vdsc_cfg->bits_per_component + 4) +
 320                         3 * (4 * vdsc_cfg->bits_per_component) - 2;
 321         else
 322                 num_extra_mux_bits = 3 * vdsc_cfg->mux_word_size +
 323                         (4 * vdsc_cfg->bits_per_component + 4) +
 324                         2 * (4 * vdsc_cfg->bits_per_component) - 2;
 325         /* Number of bits in one Slice */
 326         slice_bits = 8 * vdsc_cfg->slice_chunk_size * vdsc_cfg->slice_height;
 327
 328         while ((num_extra_mux_bits > 0) &&
 329                ((slice_bits - num_extra_mux_bits) % vdsc_cfg->mux_word_size))
 330                 num_extra_mux_bits--;
 331
 332         if (groups_per_line < vdsc_cfg->initial_scale_value - 8)
 333                 vdsc_cfg->initial_scale_value = groups_per_line + 8;
 334
 335         /* scale_decrement_interval calculation according to DSC spec 1.11 */
 336         if (vdsc_cfg->initial_scale_value > 8)
 337                 vdsc_cfg->scale_decrement_interval = groups_per_line /
 338                         (vdsc_cfg->initial_scale_value - 8);
 339         else
 340                 vdsc_cfg->scale_decrement_interval = DSC_SCALE_DECREMENT_INTERVAL_MAX;
 341
 342         vdsc_cfg->final_offset = vdsc_cfg->rc_model_size -
 343                 (vdsc_cfg->initial_xmit_delay *
 344                  vdsc_cfg->bits_per_pixel + 8) / 16 + num_extra_mux_bits;
 345
 346         if (vdsc_cfg->final_offset >= vdsc_cfg->rc_model_size) {
 347                 DRM_DEBUG_KMS("FinalOfs < RcModelSze for this InitialXmitDelay\n");
 348                 return -ERANGE;
 349         }
 350
 351         final_scale = (vdsc_cfg->rc_model_size * 8) /
 352                 (vdsc_cfg->rc_model_size - vdsc_cfg->final_offset);
 353         if (vdsc_cfg->slice_height > 1)
 354                 /*
 355                  * NflBpgOffset is 16 bit value with 11 fractional bits
 356                  * hence we multiply by 2^11 for preserving the
 357                  * fractional part
 358                  */
 359                 vdsc_cfg->nfl_bpg_offset = DIV_ROUND_UP((vdsc_cfg->first_line_bpg_offset << 11),
 360                                                         (vdsc_cfg->slice_height - 1));
 361         else
 362                 vdsc_cfg->nfl_bpg_offset = 0;
 363
 364         /* Number of groups used to code the entire slice */
 365         groups_total = groups_per_line * vdsc_cfg->slice_height;
 366
 367         /* slice_bpg_offset is 16 bit value with 11 fractional bits */
 368         vdsc_cfg->slice_bpg_offset = DIV_ROUND_UP(((vdsc_cfg->rc_model_size -
 369                                                     vdsc_cfg->initial_offset +
 370                                                     num_extra_mux_bits) << 11),
 371                                                   groups_total);
 372
 373         if (final_scale > 9) {
 374                 /*
 375                  * ScaleIncrementInterval =
 376                  * finaloffset/((NflBpgOffset + SliceBpgOffset)*8(finalscale - 1.125))
 377                  * as (NflBpgOffset + SliceBpgOffset) has 11 bit fractional value,
 378                  * we need divide by 2^11 from pstDscCfg values
 379                  */
 380                 vdsc_cfg->scale_increment_interval =
 381                                 (vdsc_cfg->final_offset * (1 << 11)) /
 382                                 ((vdsc_cfg->nfl_bpg_offset +
 383                                 vdsc_cfg->slice_bpg_offset) *
 384                                 (final_scale - 9));
 385         } else {
 386                 /*
 387                  * If finalScaleValue is less than or equal to 9, a value of 0 should
 388                  * be used to disable the scale increment at the end of the slice
 389                  */
 390                 vdsc_cfg->scale_increment_interval = 0;
 391         }
 392
 393         /*
 394          * DSC spec mentions that bits_per_pixel specifies the target
 395          * bits/pixel (bpp) rate that is used by the encoder,
 396          * in steps of 1/16 of a bit per pixel
 397          */
 398         rbs_min = vdsc_cfg->rc_model_size - vdsc_cfg->initial_offset +
 399                 DIV_ROUND_UP(vdsc_cfg->initial_xmit_delay *
 400                              vdsc_cfg->bits_per_pixel, 16) +
 401                 groups_per_line * vdsc_cfg->first_line_bpg_offset;
 402
 403         hrd_delay = DIV_ROUND_UP((rbs_min * 16), vdsc_cfg->bits_per_pixel);
 404         vdsc_cfg->rc_bits = (hrd_delay * vdsc_cfg->bits_per_pixel) / 16;
 405         vdsc_cfg->initial_dec_delay = hrd_delay - vdsc_cfg->initial_xmit_delay;
 406
 407         return 0;
 408 }
 409 EXPORT_SYMBOL(drm_dsc_compute_rc_parameters);