0001
0002
0003
0004
0005
0006
0007
0008
0009
0010
0011
0012
0013
0014
0015
0016
0017
0018
0019
0020
0021
0022
0023
0024
0025
0026 #include "rc_calc_fpu.h"
0027
0028 #include "qp_tables.h"
0029 #include "amdgpu_dm/dc_fpu.h"
0030
0031 #define table_hash(mode, bpc, max_min) ((mode << 16) | (bpc << 8) | max_min)
0032
0033 #define MODE_SELECT(val444, val422, val420) \
0034 (cm == CM_444 || cm == CM_RGB) ? (val444) : (cm == CM_422 ? (val422) : (val420))
0035
0036
0037 #define TABLE_CASE(mode, bpc, max) case (table_hash(mode, BPC_##bpc, max)): \
0038 table = qp_table_##mode##_##bpc##bpc_##max; \
0039 table_size = sizeof(qp_table_##mode##_##bpc##bpc_##max)/sizeof(*qp_table_##mode##_##bpc##bpc_##max); \
0040 break
0041
0042 static int median3(int a, int b, int c)
0043 {
0044 if (a > b)
0045 swap(a, b);
0046 if (b > c)
0047 swap(b, c);
0048 if (a > b)
0049 swap(b, c);
0050
0051 return b;
0052 }
0053
0054 static double dsc_roundf(double num)
0055 {
0056 if (num < 0.0)
0057 num = num - 0.5;
0058 else
0059 num = num + 0.5;
0060
0061 return (int)(num);
0062 }
0063
0064 static void get_qp_set(qp_set qps, enum colour_mode cm, enum bits_per_comp bpc,
0065 enum max_min max_min, float bpp)
0066 {
0067 int mode = MODE_SELECT(444, 422, 420);
0068 int sel = table_hash(mode, bpc, max_min);
0069 int table_size = 0;
0070 int index;
0071 const struct qp_entry *table = NULL;
0072
0073
0074 enum { min = DAL_MM_MIN, max = DAL_MM_MAX };
0075 switch (sel) {
0076 TABLE_CASE(444, 8, max);
0077 TABLE_CASE(444, 8, min);
0078 TABLE_CASE(444, 10, max);
0079 TABLE_CASE(444, 10, min);
0080 TABLE_CASE(444, 12, max);
0081 TABLE_CASE(444, 12, min);
0082 TABLE_CASE(422, 8, max);
0083 TABLE_CASE(422, 8, min);
0084 TABLE_CASE(422, 10, max);
0085 TABLE_CASE(422, 10, min);
0086 TABLE_CASE(422, 12, max);
0087 TABLE_CASE(422, 12, min);
0088 TABLE_CASE(420, 8, max);
0089 TABLE_CASE(420, 8, min);
0090 TABLE_CASE(420, 10, max);
0091 TABLE_CASE(420, 10, min);
0092 TABLE_CASE(420, 12, max);
0093 TABLE_CASE(420, 12, min);
0094 }
0095
0096 if (!table)
0097 return;
0098
0099 index = (bpp - table[0].bpp) * 2;
0100
0101
0102 if (index >= table_size) {
0103 dm_error("ERROR: Requested rc_calc to find a bpp entry that exceeds the table size\n");
0104 return;
0105 }
0106
0107 memcpy(qps, table[index].qps, sizeof(qp_set));
0108 }
0109
0110 static void get_ofs_set(qp_set ofs, enum colour_mode mode, float bpp)
0111 {
0112 int *p = ofs;
0113
0114 if (mode == CM_444 || mode == CM_RGB) {
0115 *p++ = (bpp <= 6) ? (0) : ((((bpp >= 8) && (bpp <= 12))) ? (2) : ((bpp >= 15) ? (10) : ((((bpp > 6) && (bpp < 8))) ? (0 + dsc_roundf((bpp - 6) * (2 / 2.0))) : (2 + dsc_roundf((bpp - 12) * (8 / 3.0))))));
0116 *p++ = (bpp <= 6) ? (-2) : ((((bpp >= 8) && (bpp <= 12))) ? (0) : ((bpp >= 15) ? (8) : ((((bpp > 6) && (bpp < 8))) ? (-2 + dsc_roundf((bpp - 6) * (2 / 2.0))) : (0 + dsc_roundf((bpp - 12) * (8 / 3.0))))));
0117 *p++ = (bpp <= 6) ? (-2) : ((((bpp >= 8) && (bpp <= 12))) ? (0) : ((bpp >= 15) ? (6) : ((((bpp > 6) && (bpp < 8))) ? (-2 + dsc_roundf((bpp - 6) * (2 / 2.0))) : (0 + dsc_roundf((bpp - 12) * (6 / 3.0))))));
0118 *p++ = (bpp <= 6) ? (-4) : ((((bpp >= 8) && (bpp <= 12))) ? (-2) : ((bpp >= 15) ? (4) : ((((bpp > 6) && (bpp < 8))) ? (-4 + dsc_roundf((bpp - 6) * (2 / 2.0))) : (-2 + dsc_roundf((bpp - 12) * (6 / 3.0))))));
0119 *p++ = (bpp <= 6) ? (-6) : ((((bpp >= 8) && (bpp <= 12))) ? (-4) : ((bpp >= 15) ? (2) : ((((bpp > 6) && (bpp < 8))) ? (-6 + dsc_roundf((bpp - 6) * (2 / 2.0))) : (-4 + dsc_roundf((bpp - 12) * (6 / 3.0))))));
0120 *p++ = (bpp <= 12) ? (-6) : ((bpp >= 15) ? (0) : (-6 + dsc_roundf((bpp - 12) * (6 / 3.0))));
0121 *p++ = (bpp <= 12) ? (-8) : ((bpp >= 15) ? (-2) : (-8 + dsc_roundf((bpp - 12) * (6 / 3.0))));
0122 *p++ = (bpp <= 12) ? (-8) : ((bpp >= 15) ? (-4) : (-8 + dsc_roundf((bpp - 12) * (4 / 3.0))));
0123 *p++ = (bpp <= 12) ? (-8) : ((bpp >= 15) ? (-6) : (-8 + dsc_roundf((bpp - 12) * (2 / 3.0))));
0124 *p++ = (bpp <= 12) ? (-10) : ((bpp >= 15) ? (-8) : (-10 + dsc_roundf((bpp - 12) * (2 / 3.0))));
0125 *p++ = -10;
0126 *p++ = (bpp <= 6) ? (-12) : ((bpp >= 8) ? (-10) : (-12 + dsc_roundf((bpp - 6) * (2 / 2.0))));
0127 *p++ = -12;
0128 *p++ = -12;
0129 *p++ = -12;
0130 } else if (mode == CM_422) {
0131 *p++ = (bpp <= 8) ? (2) : ((bpp >= 10) ? (10) : (2 + dsc_roundf((bpp - 8) * (8 / 2.0))));
0132 *p++ = (bpp <= 8) ? (0) : ((bpp >= 10) ? (8) : (0 + dsc_roundf((bpp - 8) * (8 / 2.0))));
0133 *p++ = (bpp <= 8) ? (0) : ((bpp >= 10) ? (6) : (0 + dsc_roundf((bpp - 8) * (6 / 2.0))));
0134 *p++ = (bpp <= 8) ? (-2) : ((bpp >= 10) ? (4) : (-2 + dsc_roundf((bpp - 8) * (6 / 2.0))));
0135 *p++ = (bpp <= 8) ? (-4) : ((bpp >= 10) ? (2) : (-4 + dsc_roundf((bpp - 8) * (6 / 2.0))));
0136 *p++ = (bpp <= 8) ? (-6) : ((bpp >= 10) ? (0) : (-6 + dsc_roundf((bpp - 8) * (6 / 2.0))));
0137 *p++ = (bpp <= 8) ? (-8) : ((bpp >= 10) ? (-2) : (-8 + dsc_roundf((bpp - 8) * (6 / 2.0))));
0138 *p++ = (bpp <= 8) ? (-8) : ((bpp >= 10) ? (-4) : (-8 + dsc_roundf((bpp - 8) * (4 / 2.0))));
0139 *p++ = (bpp <= 8) ? (-8) : ((bpp >= 10) ? (-6) : (-8 + dsc_roundf((bpp - 8) * (2 / 2.0))));
0140 *p++ = (bpp <= 8) ? (-10) : ((bpp >= 10) ? (-8) : (-10 + dsc_roundf((bpp - 8) * (2 / 2.0))));
0141 *p++ = -10;
0142 *p++ = (bpp <= 6) ? (-12) : ((bpp >= 7) ? (-10) : (-12 + dsc_roundf((bpp - 6) * (2.0 / 1))));
0143 *p++ = -12;
0144 *p++ = -12;
0145 *p++ = -12;
0146 } else {
0147 *p++ = (bpp <= 6) ? (2) : ((bpp >= 8) ? (10) : (2 + dsc_roundf((bpp - 6) * (8 / 2.0))));
0148 *p++ = (bpp <= 6) ? (0) : ((bpp >= 8) ? (8) : (0 + dsc_roundf((bpp - 6) * (8 / 2.0))));
0149 *p++ = (bpp <= 6) ? (0) : ((bpp >= 8) ? (6) : (0 + dsc_roundf((bpp - 6) * (6 / 2.0))));
0150 *p++ = (bpp <= 6) ? (-2) : ((bpp >= 8) ? (4) : (-2 + dsc_roundf((bpp - 6) * (6 / 2.0))));
0151 *p++ = (bpp <= 6) ? (-4) : ((bpp >= 8) ? (2) : (-4 + dsc_roundf((bpp - 6) * (6 / 2.0))));
0152 *p++ = (bpp <= 6) ? (-6) : ((bpp >= 8) ? (0) : (-6 + dsc_roundf((bpp - 6) * (6 / 2.0))));
0153 *p++ = (bpp <= 6) ? (-8) : ((bpp >= 8) ? (-2) : (-8 + dsc_roundf((bpp - 6) * (6 / 2.0))));
0154 *p++ = (bpp <= 6) ? (-8) : ((bpp >= 8) ? (-4) : (-8 + dsc_roundf((bpp - 6) * (4 / 2.0))));
0155 *p++ = (bpp <= 6) ? (-8) : ((bpp >= 8) ? (-6) : (-8 + dsc_roundf((bpp - 6) * (2 / 2.0))));
0156 *p++ = (bpp <= 6) ? (-10) : ((bpp >= 8) ? (-8) : (-10 + dsc_roundf((bpp - 6) * (2 / 2.0))));
0157 *p++ = -10;
0158 *p++ = (bpp <= 4) ? (-12) : ((bpp >= 5) ? (-10) : (-12 + dsc_roundf((bpp - 4) * (2 / 1.0))));
0159 *p++ = -12;
0160 *p++ = -12;
0161 *p++ = -12;
0162 }
0163 }
0164
0165 void _do_calc_rc_params(struct rc_params *rc,
0166 enum colour_mode cm,
0167 enum bits_per_comp bpc,
0168 u16 drm_bpp,
0169 bool is_navite_422_or_420,
0170 int slice_width,
0171 int slice_height,
0172 int minor_version)
0173 {
0174 float bpp;
0175 float bpp_group;
0176 float initial_xmit_delay_factor;
0177 int padding_pixels;
0178 int i;
0179
0180 dc_assert_fp_enabled();
0181
0182 bpp = ((float)drm_bpp / 16.0);
0183
0184
0185
0186 if (is_navite_422_or_420)
0187 bpp /= 2.0;
0188
0189 rc->rc_quant_incr_limit0 = ((bpc == BPC_8) ? 11 : (bpc == BPC_10 ? 15 : 19)) - ((minor_version == 1 && cm == CM_444) ? 1 : 0);
0190 rc->rc_quant_incr_limit1 = ((bpc == BPC_8) ? 11 : (bpc == BPC_10 ? 15 : 19)) - ((minor_version == 1 && cm == CM_444) ? 1 : 0);
0191
0192 bpp_group = MODE_SELECT(bpp, bpp * 2.0, bpp * 2.0);
0193
0194 switch (cm) {
0195 case CM_420:
0196 rc->initial_fullness_offset = (bpp >= 6) ? (2048) : ((bpp <= 4) ? (6144) : ((((bpp > 4) && (bpp <= 5))) ? (6144 - dsc_roundf((bpp - 4) * (512))) : (5632 - dsc_roundf((bpp - 5) * (3584)))));
0197 rc->first_line_bpg_offset = median3(0, (12 + (int) (0.09 * min(34, slice_height - 8))), (int)((3 * bpc * 3) - (3 * bpp_group)));
0198 rc->second_line_bpg_offset = median3(0, 12, (int)((3 * bpc * 3) - (3 * bpp_group)));
0199 break;
0200 case CM_422:
0201 rc->initial_fullness_offset = (bpp >= 8) ? (2048) : ((bpp <= 7) ? (5632) : (5632 - dsc_roundf((bpp - 7) * (3584))));
0202 rc->first_line_bpg_offset = median3(0, (12 + (int) (0.09 * min(34, slice_height - 8))), (int)((3 * bpc * 4) - (3 * bpp_group)));
0203 rc->second_line_bpg_offset = 0;
0204 break;
0205 case CM_444:
0206 case CM_RGB:
0207 rc->initial_fullness_offset = (bpp >= 12) ? (2048) : ((bpp <= 8) ? (6144) : ((((bpp > 8) && (bpp <= 10))) ? (6144 - dsc_roundf((bpp - 8) * (512 / 2))) : (5632 - dsc_roundf((bpp - 10) * (3584 / 2)))));
0208 rc->first_line_bpg_offset = median3(0, (12 + (int) (0.09 * min(34, slice_height - 8))), (int)(((3 * bpc + (cm == CM_444 ? 0 : 2)) * 3) - (3 * bpp_group)));
0209 rc->second_line_bpg_offset = 0;
0210 break;
0211 }
0212
0213 initial_xmit_delay_factor = (cm == CM_444 || cm == CM_RGB) ? 1.0 : 2.0;
0214 rc->initial_xmit_delay = dsc_roundf(8192.0/2.0/bpp/initial_xmit_delay_factor);
0215
0216 if (cm == CM_422 || cm == CM_420)
0217 slice_width /= 2;
0218
0219 padding_pixels = ((slice_width % 3) != 0) ? (3 - (slice_width % 3)) * (rc->initial_xmit_delay / slice_width) : 0;
0220 if (3 * bpp_group >= (((rc->initial_xmit_delay + 2) / 3) * (3 + (cm == CM_422)))) {
0221 if ((rc->initial_xmit_delay + padding_pixels) % 3 == 1)
0222 rc->initial_xmit_delay++;
0223 }
0224
0225 rc->flatness_min_qp = ((bpc == BPC_8) ? (3) : ((bpc == BPC_10) ? (7) : (11))) - ((minor_version == 1 && cm == CM_444) ? 1 : 0);
0226 rc->flatness_max_qp = ((bpc == BPC_8) ? (12) : ((bpc == BPC_10) ? (16) : (20))) - ((minor_version == 1 && cm == CM_444) ? 1 : 0);
0227 rc->flatness_det_thresh = 2 << (bpc - 8);
0228
0229 get_qp_set(rc->qp_min, cm, bpc, DAL_MM_MIN, bpp);
0230 get_qp_set(rc->qp_max, cm, bpc, DAL_MM_MAX, bpp);
0231 if (cm == CM_444 && minor_version == 1) {
0232 for (i = 0; i < QP_SET_SIZE; ++i) {
0233 rc->qp_min[i] = rc->qp_min[i] > 0 ? rc->qp_min[i] - 1 : 0;
0234 rc->qp_max[i] = rc->qp_max[i] > 0 ? rc->qp_max[i] - 1 : 0;
0235 }
0236 }
0237 get_ofs_set(rc->ofs, cm, bpp);
0238
0239
0240 rc->rc_model_size = 8192;
0241 rc->rc_edge_factor = 6;
0242 rc->rc_tgt_offset_hi = 3;
0243 rc->rc_tgt_offset_lo = 3;
0244
0245 rc->rc_buf_thresh[0] = 896;
0246 rc->rc_buf_thresh[1] = 1792;
0247 rc->rc_buf_thresh[2] = 2688;
0248 rc->rc_buf_thresh[3] = 3584;
0249 rc->rc_buf_thresh[4] = 4480;
0250 rc->rc_buf_thresh[5] = 5376;
0251 rc->rc_buf_thresh[6] = 6272;
0252 rc->rc_buf_thresh[7] = 6720;
0253 rc->rc_buf_thresh[8] = 7168;
0254 rc->rc_buf_thresh[9] = 7616;
0255 rc->rc_buf_thresh[10] = 7744;
0256 rc->rc_buf_thresh[11] = 7872;
0257 rc->rc_buf_thresh[12] = 8000;
0258 rc->rc_buf_thresh[13] = 8064;
0259 }
0260