diff options
Diffstat (limited to 'drivers/gpu/drm/amd/display/dc/dml/dcn321/dcn321_fpu.c')
| -rw-r--r-- | drivers/gpu/drm/amd/display/dc/dml/dcn321/dcn321_fpu.c | 446 |
1 files changed, 446 insertions, 0 deletions
diff --git a/drivers/gpu/drm/amd/display/dc/dml/dcn321/dcn321_fpu.c b/drivers/gpu/drm/amd/display/dc/dml/dcn321/dcn321_fpu.c index 78408698985b..6e72336b7975 100644 --- a/drivers/gpu/drm/amd/display/dc/dml/dcn321/dcn321_fpu.c +++ b/drivers/gpu/drm/amd/display/dc/dml/dcn321/dcn321_fpu.c @@ -24,6 +24,7 @@ * */ +#include "clk_mgr.h" #include "resource.h" #include "dcn321_fpu.h" #include "dcn32/dcn32_resource.h" @@ -236,3 +237,448 @@ void dcn321_insert_entry_into_table_sorted(struct _vcs_dpi_voltage_scaling_st *t } } +static void remove_entry_from_table_at_index(struct _vcs_dpi_voltage_scaling_st *table, unsigned int *num_entries, + unsigned int index) +{ + int i; + + if (*num_entries == 0) + return; + + for (i = index; i < *num_entries - 1; i++) { + table[i] = table[i + 1]; + } + memset(&table[--(*num_entries)], 0, sizeof(struct _vcs_dpi_voltage_scaling_st)); +} + +static int build_synthetic_soc_states(struct clk_bw_params *bw_params, + struct _vcs_dpi_voltage_scaling_st *table, unsigned int *num_entries) +{ + int i, j; + struct _vcs_dpi_voltage_scaling_st entry = {0}; + + unsigned int max_dcfclk_mhz = 0, max_dispclk_mhz = 0, max_dppclk_mhz = 0, + max_phyclk_mhz = 0, max_dtbclk_mhz = 0, max_fclk_mhz = 0, max_uclk_mhz = 0; + + unsigned int min_dcfclk_mhz = 199, min_fclk_mhz = 299; + + static const unsigned int num_dcfclk_stas = 5; + unsigned int dcfclk_sta_targets[DC__VOLTAGE_STATES] = {199, 615, 906, 1324, 1564}; + + unsigned int num_uclk_dpms = 0; + unsigned int num_fclk_dpms = 0; + unsigned int num_dcfclk_dpms = 0; + + for (i = 0; i < MAX_NUM_DPM_LVL; i++) { + if (bw_params->clk_table.entries[i].dcfclk_mhz > max_dcfclk_mhz) + max_dcfclk_mhz = bw_params->clk_table.entries[i].dcfclk_mhz; + if (bw_params->clk_table.entries[i].fclk_mhz > max_fclk_mhz) + max_fclk_mhz = bw_params->clk_table.entries[i].fclk_mhz; + if (bw_params->clk_table.entries[i].memclk_mhz > max_uclk_mhz) + max_uclk_mhz = bw_params->clk_table.entries[i].memclk_mhz; + if (bw_params->clk_table.entries[i].dispclk_mhz > max_dispclk_mhz) + max_dispclk_mhz = bw_params->clk_table.entries[i].dispclk_mhz; + if (bw_params->clk_table.entries[i].dppclk_mhz > max_dppclk_mhz) + max_dppclk_mhz = bw_params->clk_table.entries[i].dppclk_mhz; + if (bw_params->clk_table.entries[i].phyclk_mhz > max_phyclk_mhz) + max_phyclk_mhz = bw_params->clk_table.entries[i].phyclk_mhz; + if (bw_params->clk_table.entries[i].dtbclk_mhz > max_dtbclk_mhz) + max_dtbclk_mhz = bw_params->clk_table.entries[i].dtbclk_mhz; + + if (bw_params->clk_table.entries[i].memclk_mhz > 0) + num_uclk_dpms++; + if (bw_params->clk_table.entries[i].fclk_mhz > 0) + num_fclk_dpms++; + if (bw_params->clk_table.entries[i].dcfclk_mhz > 0) + num_dcfclk_dpms++; + } + + if (!max_dcfclk_mhz || !max_dispclk_mhz || !max_dtbclk_mhz) + return -1; + + if (max_dppclk_mhz == 0) + max_dppclk_mhz = max_dispclk_mhz; + + if (max_fclk_mhz == 0) + max_fclk_mhz = max_dcfclk_mhz * dcn3_21_soc.pct_ideal_sdp_bw_after_urgent / dcn3_21_soc.pct_ideal_fabric_bw_after_urgent; + + if (max_phyclk_mhz == 0) + max_phyclk_mhz = dcn3_21_soc.clock_limits[0].phyclk_mhz; + + *num_entries = 0; + entry.dispclk_mhz = max_dispclk_mhz; + entry.dscclk_mhz = max_dispclk_mhz / 3; + entry.dppclk_mhz = max_dppclk_mhz; + entry.dtbclk_mhz = max_dtbclk_mhz; + entry.phyclk_mhz = max_phyclk_mhz; + entry.phyclk_d18_mhz = dcn3_21_soc.clock_limits[0].phyclk_d18_mhz; + entry.phyclk_d32_mhz = dcn3_21_soc.clock_limits[0].phyclk_d32_mhz; + + // Insert all the DCFCLK STAs + for (i = 0; i < num_dcfclk_stas; i++) { + entry.dcfclk_mhz = dcfclk_sta_targets[i]; + entry.fabricclk_mhz = 0; + entry.dram_speed_mts = 0; + + dcn321_insert_entry_into_table_sorted(table, num_entries, &entry); + } + + // Insert the max DCFCLK + entry.dcfclk_mhz = max_dcfclk_mhz; + entry.fabricclk_mhz = 0; + entry.dram_speed_mts = 0; + + dcn321_insert_entry_into_table_sorted(table, num_entries, &entry); + + // Insert the UCLK DPMS + for (i = 0; i < num_uclk_dpms; i++) { + entry.dcfclk_mhz = 0; + entry.fabricclk_mhz = 0; + entry.dram_speed_mts = bw_params->clk_table.entries[i].memclk_mhz * 16; + + dcn321_insert_entry_into_table_sorted(table, num_entries, &entry); + } + + // If FCLK is coarse grained, insert individual DPMs. + if (num_fclk_dpms > 2) { + for (i = 0; i < num_fclk_dpms; i++) { + entry.dcfclk_mhz = 0; + entry.fabricclk_mhz = bw_params->clk_table.entries[i].fclk_mhz; + entry.dram_speed_mts = 0; + + dcn321_insert_entry_into_table_sorted(table, num_entries, &entry); + } + } + // If FCLK fine grained, only insert max + else { + entry.dcfclk_mhz = 0; + entry.fabricclk_mhz = max_fclk_mhz; + entry.dram_speed_mts = 0; + + dcn321_insert_entry_into_table_sorted(table, num_entries, &entry); + } + + // At this point, the table contains all "points of interest" based on + // DPMs from PMFW, and STAs. Table is sorted by BW, and all clock + // ratios (by derate, are exact). + + // Remove states that require higher clocks than are supported + for (i = *num_entries - 1; i >= 0 ; i--) { + if (table[i].dcfclk_mhz > max_dcfclk_mhz || + table[i].fabricclk_mhz > max_fclk_mhz || + table[i].dram_speed_mts > max_uclk_mhz * 16) + remove_entry_from_table_at_index(table, num_entries, i); + } + + // At this point, the table only contains supported points of interest + // it could be used as is, but some states may be redundant due to + // coarse grained nature of some clocks, so we want to round up to + // coarse grained DPMs and remove duplicates. + + // Round up UCLKs + for (i = *num_entries - 1; i >= 0 ; i--) { + for (j = 0; j < num_uclk_dpms; j++) { + if (bw_params->clk_table.entries[j].memclk_mhz * 16 >= table[i].dram_speed_mts) { + table[i].dram_speed_mts = bw_params->clk_table.entries[j].memclk_mhz * 16; + break; + } + } + } + + // If FCLK is coarse grained, round up to next DPMs + if (num_fclk_dpms > 2) { + for (i = *num_entries - 1; i >= 0 ; i--) { + for (j = 0; j < num_fclk_dpms; j++) { + if (bw_params->clk_table.entries[j].fclk_mhz >= table[i].fabricclk_mhz) { + table[i].fabricclk_mhz = bw_params->clk_table.entries[j].fclk_mhz; + break; + } + } + } + } + // Otherwise, round up to minimum. + else { + for (i = *num_entries - 1; i >= 0 ; i--) { + if (table[i].fabricclk_mhz < min_fclk_mhz) { + table[i].fabricclk_mhz = min_fclk_mhz; + break; + } + } + } + + // Round DCFCLKs up to minimum + for (i = *num_entries - 1; i >= 0 ; i--) { + if (table[i].dcfclk_mhz < min_dcfclk_mhz) { + table[i].dcfclk_mhz = min_dcfclk_mhz; + break; + } + } + + // Remove duplicate states, note duplicate states are always neighbouring since table is sorted. + i = 0; + while (i < *num_entries - 1) { + if (table[i].dcfclk_mhz == table[i + 1].dcfclk_mhz && + table[i].fabricclk_mhz == table[i + 1].fabricclk_mhz && + table[i].dram_speed_mts == table[i + 1].dram_speed_mts) + remove_entry_from_table_at_index(table, num_entries, i + 1); + else + i++; + } + + // Fix up the state indicies + for (i = *num_entries - 1; i >= 0 ; i--) { + table[i].state = i; + } + + return 0; +} + +static void dcn321_get_optimal_dcfclk_fclk_for_uclk(unsigned int uclk_mts, + unsigned int *optimal_dcfclk, + unsigned int *optimal_fclk) +{ + double bw_from_dram, bw_from_dram1, bw_from_dram2; + + bw_from_dram1 = uclk_mts * dcn3_21_soc.num_chans * + dcn3_21_soc.dram_channel_width_bytes * (dcn3_21_soc.max_avg_dram_bw_use_normal_percent / 100); + bw_from_dram2 = uclk_mts * dcn3_21_soc.num_chans * + dcn3_21_soc.dram_channel_width_bytes * (dcn3_21_soc.max_avg_sdp_bw_use_normal_percent / 100); + + bw_from_dram = (bw_from_dram1 < bw_from_dram2) ? bw_from_dram1 : bw_from_dram2; + + if (optimal_fclk) + *optimal_fclk = bw_from_dram / + (dcn3_21_soc.fabric_datapath_to_dcn_data_return_bytes * (dcn3_21_soc.max_avg_sdp_bw_use_normal_percent / 100)); + + if (optimal_dcfclk) + *optimal_dcfclk = bw_from_dram / + (dcn3_21_soc.return_bus_width_bytes * (dcn3_21_soc.max_avg_sdp_bw_use_normal_percent / 100)); +} + +/** dcn321_update_bw_bounding_box + * This would override some dcn3_2 ip_or_soc initial parameters hardcoded from spreadsheet + * with actual values as per dGPU SKU: + * -with passed few options from dc->config + * -with dentist_vco_frequency from Clk Mgr (currently hardcoded, but might need to get it from PM FW) + * -with passed latency values (passed in ns units) in dc-> bb override for debugging purposes + * -with passed latencies from VBIOS (in 100_ns units) if available for certain dGPU SKU + * -with number of DRAM channels from VBIOS (which differ for certain dGPU SKU of the same ASIC) + * -clocks levels with passed clk_table entries from Clk Mgr as reported by PM FW for different + * clocks (which might differ for certain dGPU SKU of the same ASIC) + */ +void dcn321_update_bw_bounding_box_fpu(struct dc *dc, struct clk_bw_params *bw_params) +{ + dc_assert_fp_enabled(); + if (!IS_FPGA_MAXIMUS_DC(dc->ctx->dce_environment)) { + /* Overrides from dc->config options */ + dcn3_21_ip.clamp_min_dcfclk = dc->config.clamp_min_dcfclk; + + /* Override from passed dc->bb_overrides if available*/ + if ((int)(dcn3_21_soc.sr_exit_time_us * 1000) != dc->bb_overrides.sr_exit_time_ns + && dc->bb_overrides.sr_exit_time_ns) { + dcn3_21_soc.sr_exit_time_us = dc->bb_overrides.sr_exit_time_ns / 1000.0; + } + + if ((int)(dcn3_21_soc.sr_enter_plus_exit_time_us * 1000) + != dc->bb_overrides.sr_enter_plus_exit_time_ns + && dc->bb_overrides.sr_enter_plus_exit_time_ns) { + dcn3_21_soc.sr_enter_plus_exit_time_us = + dc->bb_overrides.sr_enter_plus_exit_time_ns / 1000.0; + } + + if ((int)(dcn3_21_soc.urgent_latency_us * 1000) != dc->bb_overrides.urgent_latency_ns + && dc->bb_overrides.urgent_latency_ns) { + dcn3_21_soc.urgent_latency_us = dc->bb_overrides.urgent_latency_ns / 1000.0; + } + + if ((int)(dcn3_21_soc.dram_clock_change_latency_us * 1000) + != dc->bb_overrides.dram_clock_change_latency_ns + && dc->bb_overrides.dram_clock_change_latency_ns) { + dcn3_21_soc.dram_clock_change_latency_us = + dc->bb_overrides.dram_clock_change_latency_ns / 1000.0; + } + + if ((int)(dcn3_21_soc.dummy_pstate_latency_us * 1000) + != dc->bb_overrides.dummy_clock_change_latency_ns + && dc->bb_overrides.dummy_clock_change_latency_ns) { + dcn3_21_soc.dummy_pstate_latency_us = + dc->bb_overrides.dummy_clock_change_latency_ns / 1000.0; + } + + /* Override from VBIOS if VBIOS bb_info available */ + if (dc->ctx->dc_bios->funcs->get_soc_bb_info) { + struct bp_soc_bb_info bb_info = {0}; + + if (dc->ctx->dc_bios->funcs->get_soc_bb_info(dc->ctx->dc_bios, &bb_info) == BP_RESULT_OK) { + if (bb_info.dram_clock_change_latency_100ns > 0) + dcn3_21_soc.dram_clock_change_latency_us = bb_info.dram_clock_change_latency_100ns * 10; + + if (bb_info.dram_sr_enter_exit_latency_100ns > 0) + dcn3_21_soc.sr_enter_plus_exit_time_us = bb_info.dram_sr_enter_exit_latency_100ns * 10; + + if (bb_info.dram_sr_exit_latency_100ns > 0) + dcn3_21_soc.sr_exit_time_us = bb_info.dram_sr_exit_latency_100ns * 10; + } + } + + /* Override from VBIOS for num_chan */ + if (dc->ctx->dc_bios->vram_info.num_chans) + dcn3_21_soc.num_chans = dc->ctx->dc_bios->vram_info.num_chans; + + if (dc->ctx->dc_bios->vram_info.dram_channel_width_bytes) + dcn3_21_soc.dram_channel_width_bytes = dc->ctx->dc_bios->vram_info.dram_channel_width_bytes; + + } + + /* Override dispclk_dppclk_vco_speed_mhz from Clk Mgr */ + dcn3_21_soc.dispclk_dppclk_vco_speed_mhz = dc->clk_mgr->dentist_vco_freq_khz / 1000.0; + dc->dml.soc.dispclk_dppclk_vco_speed_mhz = dc->clk_mgr->dentist_vco_freq_khz / 1000.0; + + /* Overrides Clock levelsfrom CLK Mgr table entries as reported by PM FW */ + if ((!IS_FPGA_MAXIMUS_DC(dc->ctx->dce_environment)) && (bw_params->clk_table.entries[0].memclk_mhz)) { + if (dc->debug.use_legacy_soc_bb_mechanism) { + unsigned int i = 0, j = 0, num_states = 0; + + unsigned int dcfclk_mhz[DC__VOLTAGE_STATES] = {0}; + unsigned int dram_speed_mts[DC__VOLTAGE_STATES] = {0}; + unsigned int optimal_uclk_for_dcfclk_sta_targets[DC__VOLTAGE_STATES] = {0}; + unsigned int optimal_dcfclk_for_uclk[DC__VOLTAGE_STATES] = {0}; + + unsigned int dcfclk_sta_targets[DC__VOLTAGE_STATES] = {615, 906, 1324, 1564}; + unsigned int num_dcfclk_sta_targets = 4, num_uclk_states = 0; + unsigned int max_dcfclk_mhz = 0, max_dispclk_mhz = 0, max_dppclk_mhz = 0, max_phyclk_mhz = 0; + + for (i = 0; i < MAX_NUM_DPM_LVL; i++) { + if (bw_params->clk_table.entries[i].dcfclk_mhz > max_dcfclk_mhz) + max_dcfclk_mhz = bw_params->clk_table.entries[i].dcfclk_mhz; + if (bw_params->clk_table.entries[i].dispclk_mhz > max_dispclk_mhz) + max_dispclk_mhz = bw_params->clk_table.entries[i].dispclk_mhz; + if (bw_params->clk_table.entries[i].dppclk_mhz > max_dppclk_mhz) + max_dppclk_mhz = bw_params->clk_table.entries[i].dppclk_mhz; + if (bw_params->clk_table.entries[i].phyclk_mhz > max_phyclk_mhz) + max_phyclk_mhz = bw_params->clk_table.entries[i].phyclk_mhz; + } + if (!max_dcfclk_mhz) + max_dcfclk_mhz = dcn3_21_soc.clock_limits[0].dcfclk_mhz; + if (!max_dispclk_mhz) + max_dispclk_mhz = dcn3_21_soc.clock_limits[0].dispclk_mhz; + if (!max_dppclk_mhz) + max_dppclk_mhz = dcn3_21_soc.clock_limits[0].dppclk_mhz; + if (!max_phyclk_mhz) + max_phyclk_mhz = dcn3_21_soc.clock_limits[0].phyclk_mhz; + + if (max_dcfclk_mhz > dcfclk_sta_targets[num_dcfclk_sta_targets-1]) { + // If max DCFCLK is greater than the max DCFCLK STA target, insert into the DCFCLK STA target array + dcfclk_sta_targets[num_dcfclk_sta_targets] = max_dcfclk_mhz; + num_dcfclk_sta_targets++; + } else if (max_dcfclk_mhz < dcfclk_sta_targets[num_dcfclk_sta_targets-1]) { + // If max DCFCLK is less than the max DCFCLK STA target, cap values and remove duplicates + for (i = 0; i < num_dcfclk_sta_targets; i++) { + if (dcfclk_sta_targets[i] > max_dcfclk_mhz) { + dcfclk_sta_targets[i] = max_dcfclk_mhz; + break; + } + } + // Update size of array since we "removed" duplicates + num_dcfclk_sta_targets = i + 1; + } + + num_uclk_states = bw_params->clk_table.num_entries; + + // Calculate optimal dcfclk for each uclk + for (i = 0; i < num_uclk_states; i++) { + dcn321_get_optimal_dcfclk_fclk_for_uclk(bw_params->clk_table.entries[i].memclk_mhz * 16, + &optimal_dcfclk_for_uclk[i], NULL); + if (optimal_dcfclk_for_uclk[i] < bw_params->clk_table.entries[0].dcfclk_mhz) { + optimal_dcfclk_for_uclk[i] = bw_params->clk_table.entries[0].dcfclk_mhz; + } + } + + // Calculate optimal uclk for each dcfclk sta target + for (i = 0; i < num_dcfclk_sta_targets; i++) { + for (j = 0; j < num_uclk_states; j++) { + if (dcfclk_sta_targets[i] < optimal_dcfclk_for_uclk[j]) { + optimal_uclk_for_dcfclk_sta_targets[i] = + bw_params->clk_table.entries[j].memclk_mhz * 16; + break; + } + } + } + + i = 0; + j = 0; + // create the final dcfclk and uclk table + while (i < num_dcfclk_sta_targets && j < num_uclk_states && num_states < DC__VOLTAGE_STATES) { + if (dcfclk_sta_targets[i] < optimal_dcfclk_for_uclk[j] && i < num_dcfclk_sta_targets) { + dcfclk_mhz[num_states] = dcfclk_sta_targets[i]; + dram_speed_mts[num_states++] = optimal_uclk_for_dcfclk_sta_targets[i++]; + } else { + if (j < num_uclk_states && optimal_dcfclk_for_uclk[j] <= max_dcfclk_mhz) { + dcfclk_mhz[num_states] = optimal_dcfclk_for_uclk[j]; + dram_speed_mts[num_states++] = bw_params->clk_table.entries[j++].memclk_mhz * 16; + } else { + j = num_uclk_states; + } + } + } + + while (i < num_dcfclk_sta_targets && num_states < DC__VOLTAGE_STATES) { + dcfclk_mhz[num_states] = dcfclk_sta_targets[i]; + dram_speed_mts[num_states++] = optimal_uclk_for_dcfclk_sta_targets[i++]; + } + + while (j < num_uclk_states && num_states < DC__VOLTAGE_STATES && + optimal_dcfclk_for_uclk[j] <= max_dcfclk_mhz) { + dcfclk_mhz[num_states] = optimal_dcfclk_for_uclk[j]; + dram_speed_mts[num_states++] = bw_params->clk_table.entries[j++].memclk_mhz * 16; + } + + dcn3_21_soc.num_states = num_states; + for (i = 0; i < dcn3_21_soc.num_states; i++) { + dcn3_21_soc.clock_limits[i].state = i; + dcn3_21_soc.clock_limits[i].dcfclk_mhz = dcfclk_mhz[i]; + dcn3_21_soc.clock_limits[i].fabricclk_mhz = dcfclk_mhz[i]; + + /* Fill all states with max values of all these clocks */ + dcn3_21_soc.clock_limits[i].dispclk_mhz = max_dispclk_mhz; + dcn3_21_soc.clock_limits[i].dppclk_mhz = max_dppclk_mhz; + dcn3_21_soc.clock_limits[i].phyclk_mhz = max_phyclk_mhz; + dcn3_21_soc.clock_limits[i].dscclk_mhz = max_dispclk_mhz / 3; + + /* Populate from bw_params for DTBCLK, SOCCLK */ + if (i > 0) { + if (!bw_params->clk_table.entries[i].dtbclk_mhz) { + dcn3_21_soc.clock_limits[i].dtbclk_mhz = dcn3_21_soc.clock_limits[i-1].dtbclk_mhz; + } else { + dcn3_21_soc.clock_limits[i].dtbclk_mhz = bw_params->clk_table.entries[i].dtbclk_mhz; + } + } else if (bw_params->clk_table.entries[i].dtbclk_mhz) { + dcn3_21_soc.clock_limits[i].dtbclk_mhz = bw_params->clk_table.entries[i].dtbclk_mhz; + } + + if (!bw_params->clk_table.entries[i].socclk_mhz && i > 0) + dcn3_21_soc.clock_limits[i].socclk_mhz = dcn3_21_soc.clock_limits[i-1].socclk_mhz; + else + dcn3_21_soc.clock_limits[i].socclk_mhz = bw_params->clk_table.entries[i].socclk_mhz; + + if (!dram_speed_mts[i] && i > 0) + dcn3_21_soc.clock_limits[i].dram_speed_mts = dcn3_21_soc.clock_limits[i-1].dram_speed_mts; + else + dcn3_21_soc.clock_limits[i].dram_speed_mts = dram_speed_mts[i]; + + /* These clocks cannot come from bw_params, always fill from dcn3_21_soc[0] */ + /* PHYCLK_D18, PHYCLK_D32 */ + dcn3_21_soc.clock_limits[i].phyclk_d18_mhz = dcn3_21_soc.clock_limits[0].phyclk_d18_mhz; + dcn3_21_soc.clock_limits[i].phyclk_d32_mhz = dcn3_21_soc.clock_limits[0].phyclk_d32_mhz; + } + } else { + build_synthetic_soc_states(bw_params, dcn3_21_soc.clock_limits, &dcn3_21_soc.num_states); + } + + /* Re-init DML with updated bb */ + dml_init_instance(&dc->dml, &dcn3_21_soc, &dcn3_21_ip, DML_PROJECT_DCN32); + if (dc->current_state) + dml_init_instance(&dc->current_state->bw_ctx.dml, &dcn3_21_soc, &dcn3_21_ip, DML_PROJECT_DCN32); + } +} + |