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 "resource.h"
0027 #include "clk_mgr.h"
0028 #include "dcn20/dcn20_resource.h"
0029 #include "dcn302/dcn302_resource.h"
0030
0031 #include "dml/dcn20/dcn20_fpu.h"
0032 #include "dcn302_fpu.h"
0033
0034 struct _vcs_dpi_ip_params_st dcn3_02_ip = {
0035 .use_min_dcfclk = 0,
0036 .clamp_min_dcfclk = 0,
0037 .odm_capable = 1,
0038 .gpuvm_enable = 1,
0039 .hostvm_enable = 0,
0040 .gpuvm_max_page_table_levels = 4,
0041 .hostvm_max_page_table_levels = 4,
0042 .hostvm_cached_page_table_levels = 0,
0043 .pte_group_size_bytes = 2048,
0044 .num_dsc = 5,
0045 .rob_buffer_size_kbytes = 184,
0046 .det_buffer_size_kbytes = 184,
0047 .dpte_buffer_size_in_pte_reqs_luma = 64,
0048 .dpte_buffer_size_in_pte_reqs_chroma = 34,
0049 .pde_proc_buffer_size_64k_reqs = 48,
0050 .dpp_output_buffer_pixels = 2560,
0051 .opp_output_buffer_lines = 1,
0052 .pixel_chunk_size_kbytes = 8,
0053 .pte_enable = 1,
0054 .max_page_table_levels = 2,
0055 .pte_chunk_size_kbytes = 2,
0056 .meta_chunk_size_kbytes = 2,
0057 .writeback_chunk_size_kbytes = 8,
0058 .line_buffer_size_bits = 789504,
0059 .is_line_buffer_bpp_fixed = 0,
0060 .line_buffer_fixed_bpp = 0,
0061 .dcc_supported = true,
0062 .writeback_interface_buffer_size_kbytes = 90,
0063 .writeback_line_buffer_buffer_size = 0,
0064 .max_line_buffer_lines = 12,
0065 .writeback_luma_buffer_size_kbytes = 12,
0066 .writeback_chroma_buffer_size_kbytes = 8,
0067 .writeback_chroma_line_buffer_width_pixels = 4,
0068 .writeback_max_hscl_ratio = 1,
0069 .writeback_max_vscl_ratio = 1,
0070 .writeback_min_hscl_ratio = 1,
0071 .writeback_min_vscl_ratio = 1,
0072 .writeback_max_hscl_taps = 1,
0073 .writeback_max_vscl_taps = 1,
0074 .writeback_line_buffer_luma_buffer_size = 0,
0075 .writeback_line_buffer_chroma_buffer_size = 14643,
0076 .cursor_buffer_size = 8,
0077 .cursor_chunk_size = 2,
0078 .max_num_otg = 5,
0079 .max_num_dpp = 5,
0080 .max_num_wb = 1,
0081 .max_dchub_pscl_bw_pix_per_clk = 4,
0082 .max_pscl_lb_bw_pix_per_clk = 2,
0083 .max_lb_vscl_bw_pix_per_clk = 4,
0084 .max_vscl_hscl_bw_pix_per_clk = 4,
0085 .max_hscl_ratio = 6,
0086 .max_vscl_ratio = 6,
0087 .hscl_mults = 4,
0088 .vscl_mults = 4,
0089 .max_hscl_taps = 8,
0090 .max_vscl_taps = 8,
0091 .dispclk_ramp_margin_percent = 1,
0092 .underscan_factor = 1.11,
0093 .min_vblank_lines = 32,
0094 .dppclk_delay_subtotal = 46,
0095 .dynamic_metadata_vm_enabled = true,
0096 .dppclk_delay_scl_lb_only = 16,
0097 .dppclk_delay_scl = 50,
0098 .dppclk_delay_cnvc_formatter = 27,
0099 .dppclk_delay_cnvc_cursor = 6,
0100 .dispclk_delay_subtotal = 119,
0101 .dcfclk_cstate_latency = 5.2,
0102 .max_inter_dcn_tile_repeaters = 8,
0103 .max_num_hdmi_frl_outputs = 1,
0104 .odm_combine_4to1_supported = true,
0105
0106 .xfc_supported = false,
0107 .xfc_fill_bw_overhead_percent = 10.0,
0108 .xfc_fill_constant_bytes = 0,
0109 .gfx7_compat_tiling_supported = 0,
0110 .number_of_cursors = 1,
0111 };
0112
0113 struct _vcs_dpi_soc_bounding_box_st dcn3_02_soc = {
0114 .clock_limits = {
0115 {
0116 .state = 0,
0117 .dispclk_mhz = 562.0,
0118 .dppclk_mhz = 300.0,
0119 .phyclk_mhz = 300.0,
0120 .phyclk_d18_mhz = 667.0,
0121 .dscclk_mhz = 405.6,
0122 },
0123 },
0124
0125 .min_dcfclk = 500.0,
0126 .num_states = 1,
0127 .sr_exit_time_us = 26.5,
0128 .sr_enter_plus_exit_time_us = 31,
0129 .urgent_latency_us = 4.0,
0130 .urgent_latency_pixel_data_only_us = 4.0,
0131 .urgent_latency_pixel_mixed_with_vm_data_us = 4.0,
0132 .urgent_latency_vm_data_only_us = 4.0,
0133 .urgent_out_of_order_return_per_channel_pixel_only_bytes = 4096,
0134 .urgent_out_of_order_return_per_channel_pixel_and_vm_bytes = 4096,
0135 .urgent_out_of_order_return_per_channel_vm_only_bytes = 4096,
0136 .pct_ideal_dram_sdp_bw_after_urgent_pixel_only = 80.0,
0137 .pct_ideal_dram_sdp_bw_after_urgent_pixel_and_vm = 60.0,
0138 .pct_ideal_dram_sdp_bw_after_urgent_vm_only = 40.0,
0139 .max_avg_sdp_bw_use_normal_percent = 60.0,
0140 .max_avg_dram_bw_use_normal_percent = 40.0,
0141 .writeback_latency_us = 12.0,
0142 .max_request_size_bytes = 256,
0143 .fabric_datapath_to_dcn_data_return_bytes = 64,
0144 .dcn_downspread_percent = 0.5,
0145 .downspread_percent = 0.38,
0146 .dram_page_open_time_ns = 50.0,
0147 .dram_rw_turnaround_time_ns = 17.5,
0148 .dram_return_buffer_per_channel_bytes = 8192,
0149 .round_trip_ping_latency_dcfclk_cycles = 156,
0150 .urgent_out_of_order_return_per_channel_bytes = 4096,
0151 .channel_interleave_bytes = 256,
0152 .num_banks = 8,
0153 .gpuvm_min_page_size_bytes = 4096,
0154 .hostvm_min_page_size_bytes = 4096,
0155 .dram_clock_change_latency_us = 404,
0156 .dummy_pstate_latency_us = 5,
0157 .writeback_dram_clock_change_latency_us = 23.0,
0158 .return_bus_width_bytes = 64,
0159 .dispclk_dppclk_vco_speed_mhz = 3650,
0160 .xfc_bus_transport_time_us = 20,
0161 .xfc_xbuf_latency_tolerance_us = 4,
0162 .use_urgent_burst_bw = 1,
0163 .do_urgent_latency_adjustment = true,
0164 .urgent_latency_adjustment_fabric_clock_component_us = 1.0,
0165 .urgent_latency_adjustment_fabric_clock_reference_mhz = 1000,
0166 };
0167
0168 static void dcn302_get_optimal_dcfclk_fclk_for_uclk(unsigned int uclk_mts,
0169 unsigned int *optimal_dcfclk,
0170 unsigned int *optimal_fclk)
0171 {
0172
0173 double bw_from_dram, bw_from_dram1, bw_from_dram2;
0174
0175 bw_from_dram1 = uclk_mts * dcn3_02_soc.num_chans *
0176 dcn3_02_soc.dram_channel_width_bytes *
0177 (dcn3_02_soc.max_avg_dram_bw_use_normal_percent / 100);
0178 bw_from_dram2 = uclk_mts * dcn3_02_soc.num_chans *
0179 dcn3_02_soc.dram_channel_width_bytes *
0180 (dcn3_02_soc.max_avg_sdp_bw_use_normal_percent / 100);
0181
0182 bw_from_dram = (bw_from_dram1 < bw_from_dram2) ? bw_from_dram1 : bw_from_dram2;
0183
0184 if (optimal_fclk)
0185 *optimal_fclk = bw_from_dram /
0186 (dcn3_02_soc.fabric_datapath_to_dcn_data_return_bytes *
0187 (dcn3_02_soc.max_avg_sdp_bw_use_normal_percent / 100));
0188
0189 if (optimal_dcfclk)
0190 *optimal_dcfclk = bw_from_dram /
0191 (dcn3_02_soc.return_bus_width_bytes *
0192 (dcn3_02_soc.max_avg_sdp_bw_use_normal_percent / 100));
0193 }
0194
0195 void dcn302_fpu_update_bw_bounding_box(struct dc *dc, struct clk_bw_params *bw_params)
0196 {
0197 unsigned int i, j;
0198 unsigned int num_states = 0;
0199
0200 unsigned int dcfclk_mhz[DC__VOLTAGE_STATES] = {0};
0201 unsigned int dram_speed_mts[DC__VOLTAGE_STATES] = {0};
0202 unsigned int optimal_uclk_for_dcfclk_sta_targets[DC__VOLTAGE_STATES] = {0};
0203 unsigned int optimal_dcfclk_for_uclk[DC__VOLTAGE_STATES] = {0};
0204
0205 unsigned int dcfclk_sta_targets[DC__VOLTAGE_STATES] = {694, 875, 1000, 1200};
0206 unsigned int num_dcfclk_sta_targets = 4;
0207 unsigned int num_uclk_states;
0208
0209 dc_assert_fp_enabled();
0210
0211 if (dc->ctx->dc_bios->vram_info.num_chans)
0212 dcn3_02_soc.num_chans = dc->ctx->dc_bios->vram_info.num_chans;
0213
0214 if (dc->ctx->dc_bios->vram_info.dram_channel_width_bytes)
0215 dcn3_02_soc.dram_channel_width_bytes = dc->ctx->dc_bios->vram_info.dram_channel_width_bytes;
0216
0217 dcn3_02_soc.dispclk_dppclk_vco_speed_mhz = dc->clk_mgr->dentist_vco_freq_khz / 1000.0;
0218 dc->dml.soc.dispclk_dppclk_vco_speed_mhz = dc->clk_mgr->dentist_vco_freq_khz / 1000.0;
0219
0220 if (bw_params->clk_table.entries[0].memclk_mhz) {
0221 int max_dcfclk_mhz = 0, max_dispclk_mhz = 0, max_dppclk_mhz = 0, max_phyclk_mhz = 0;
0222
0223 for (i = 0; i < MAX_NUM_DPM_LVL; i++) {
0224 if (bw_params->clk_table.entries[i].dcfclk_mhz > max_dcfclk_mhz)
0225 max_dcfclk_mhz = bw_params->clk_table.entries[i].dcfclk_mhz;
0226 if (bw_params->clk_table.entries[i].dispclk_mhz > max_dispclk_mhz)
0227 max_dispclk_mhz = bw_params->clk_table.entries[i].dispclk_mhz;
0228 if (bw_params->clk_table.entries[i].dppclk_mhz > max_dppclk_mhz)
0229 max_dppclk_mhz = bw_params->clk_table.entries[i].dppclk_mhz;
0230 if (bw_params->clk_table.entries[i].phyclk_mhz > max_phyclk_mhz)
0231 max_phyclk_mhz = bw_params->clk_table.entries[i].phyclk_mhz;
0232 }
0233 if (!max_dcfclk_mhz)
0234 max_dcfclk_mhz = dcn3_02_soc.clock_limits[0].dcfclk_mhz;
0235 if (!max_dispclk_mhz)
0236 max_dispclk_mhz = dcn3_02_soc.clock_limits[0].dispclk_mhz;
0237 if (!max_dppclk_mhz)
0238 max_dppclk_mhz = dcn3_02_soc.clock_limits[0].dppclk_mhz;
0239 if (!max_phyclk_mhz)
0240 max_phyclk_mhz = dcn3_02_soc.clock_limits[0].phyclk_mhz;
0241
0242 if (max_dcfclk_mhz > dcfclk_sta_targets[num_dcfclk_sta_targets-1]) {
0243
0244 dcfclk_sta_targets[num_dcfclk_sta_targets] = max_dcfclk_mhz;
0245 num_dcfclk_sta_targets++;
0246 } else if (max_dcfclk_mhz < dcfclk_sta_targets[num_dcfclk_sta_targets-1]) {
0247
0248 for (i = 0; i < num_dcfclk_sta_targets; i++) {
0249 if (dcfclk_sta_targets[i] > max_dcfclk_mhz) {
0250 dcfclk_sta_targets[i] = max_dcfclk_mhz;
0251 break;
0252 }
0253 }
0254
0255 num_dcfclk_sta_targets = i + 1;
0256 }
0257
0258 num_uclk_states = bw_params->clk_table.num_entries;
0259
0260
0261 for (i = 0; i < num_uclk_states; i++) {
0262 dcn302_get_optimal_dcfclk_fclk_for_uclk(bw_params->clk_table.entries[i].memclk_mhz * 16,
0263 &optimal_dcfclk_for_uclk[i], NULL);
0264 if (optimal_dcfclk_for_uclk[i] < bw_params->clk_table.entries[0].dcfclk_mhz)
0265 optimal_dcfclk_for_uclk[i] = bw_params->clk_table.entries[0].dcfclk_mhz;
0266 }
0267
0268
0269 for (i = 0; i < num_dcfclk_sta_targets; i++) {
0270 for (j = 0; j < num_uclk_states; j++) {
0271 if (dcfclk_sta_targets[i] < optimal_dcfclk_for_uclk[j]) {
0272 optimal_uclk_for_dcfclk_sta_targets[i] =
0273 bw_params->clk_table.entries[j].memclk_mhz * 16;
0274 break;
0275 }
0276 }
0277 }
0278
0279 i = 0;
0280 j = 0;
0281
0282 while (i < num_dcfclk_sta_targets && j < num_uclk_states && num_states < DC__VOLTAGE_STATES) {
0283 if (dcfclk_sta_targets[i] < optimal_dcfclk_for_uclk[j] && i < num_dcfclk_sta_targets) {
0284 dcfclk_mhz[num_states] = dcfclk_sta_targets[i];
0285 dram_speed_mts[num_states++] = optimal_uclk_for_dcfclk_sta_targets[i++];
0286 } else {
0287 if (j < num_uclk_states && optimal_dcfclk_for_uclk[j] <= max_dcfclk_mhz) {
0288 dcfclk_mhz[num_states] = optimal_dcfclk_for_uclk[j];
0289 dram_speed_mts[num_states++] = bw_params->clk_table.entries[j++].memclk_mhz * 16;
0290 } else {
0291 j = num_uclk_states;
0292 }
0293 }
0294 }
0295
0296 while (i < num_dcfclk_sta_targets && num_states < DC__VOLTAGE_STATES) {
0297 dcfclk_mhz[num_states] = dcfclk_sta_targets[i];
0298 dram_speed_mts[num_states++] = optimal_uclk_for_dcfclk_sta_targets[i++];
0299 }
0300
0301 while (j < num_uclk_states && num_states < DC__VOLTAGE_STATES &&
0302 optimal_dcfclk_for_uclk[j] <= max_dcfclk_mhz) {
0303 dcfclk_mhz[num_states] = optimal_dcfclk_for_uclk[j];
0304 dram_speed_mts[num_states++] = bw_params->clk_table.entries[j++].memclk_mhz * 16;
0305 }
0306
0307 dcn3_02_soc.num_states = num_states;
0308 for (i = 0; i < dcn3_02_soc.num_states; i++) {
0309 dcn3_02_soc.clock_limits[i].state = i;
0310 dcn3_02_soc.clock_limits[i].dcfclk_mhz = dcfclk_mhz[i];
0311 dcn3_02_soc.clock_limits[i].fabricclk_mhz = dcfclk_mhz[i];
0312 dcn3_02_soc.clock_limits[i].dram_speed_mts = dram_speed_mts[i];
0313
0314
0315 dcn3_02_soc.clock_limits[i].dispclk_mhz = max_dispclk_mhz;
0316 dcn3_02_soc.clock_limits[i].dppclk_mhz = max_dppclk_mhz;
0317 dcn3_02_soc.clock_limits[i].phyclk_mhz = max_phyclk_mhz;
0318
0319 if (!bw_params->clk_table.entries[i].dtbclk_mhz && i > 0)
0320 dcn3_02_soc.clock_limits[i].dtbclk_mhz = dcn3_02_soc.clock_limits[i-1].dtbclk_mhz;
0321 else
0322 dcn3_02_soc.clock_limits[i].dtbclk_mhz = bw_params->clk_table.entries[i].dtbclk_mhz;
0323 if (!bw_params->clk_table.entries[i].socclk_mhz && i > 0)
0324 dcn3_02_soc.clock_limits[i].socclk_mhz = dcn3_02_soc.clock_limits[i-1].socclk_mhz;
0325 else
0326 dcn3_02_soc.clock_limits[i].socclk_mhz = bw_params->clk_table.entries[i].socclk_mhz;
0327
0328
0329 dcn3_02_soc.clock_limits[i].phyclk_d18_mhz = dcn3_02_soc.clock_limits[0].phyclk_d18_mhz;
0330 dcn3_02_soc.clock_limits[i].dscclk_mhz = dcn3_02_soc.clock_limits[0].dscclk_mhz;
0331 }
0332
0333 dml_init_instance(&dc->dml, &dcn3_02_soc, &dcn3_02_ip, DML_PROJECT_DCN30);
0334 if (dc->current_state)
0335 dml_init_instance(&dc->current_state->bw_ctx.dml, &dcn3_02_soc, &dcn3_02_ip, DML_PROJECT_DCN30);
0336 }
0337 }
0338
0339 void dcn302_fpu_init_soc_bounding_box(struct bp_soc_bb_info bb_info)
0340 {
0341
0342 dc_assert_fp_enabled();
0343
0344 if (bb_info.dram_clock_change_latency_100ns > 0)
0345 dcn3_02_soc.dram_clock_change_latency_us =
0346 bb_info.dram_clock_change_latency_100ns * 10;
0347
0348 if (bb_info.dram_sr_enter_exit_latency_100ns > 0)
0349 dcn3_02_soc.sr_enter_plus_exit_time_us =
0350 bb_info.dram_sr_enter_exit_latency_100ns * 10;
0351
0352 if (bb_info.dram_sr_exit_latency_100ns > 0)
0353 dcn3_02_soc.sr_exit_time_us =
0354 bb_info.dram_sr_exit_latency_100ns * 10;
0355 }
0356
0357