From: Dillon Varone <[email protected]> [WHY] Prefetch calculations did not guarantee that bandwidth required in mode support was less than mode programming which can cause failures.
[HOW] Fix bandwidth calculations to assume fixed times for OTO schedule, and choose which schedule to use based on time to fetch pixel data. Reviewed-by: Jun Lei <[email protected]> Signed-off-by: Dillon Varone <[email protected]> Signed-off-by: Zaeem Mohamed <[email protected]> --- .../src/dml2_core/dml2_core_dcn4_calcs.c | 47 ++++++++++++------- .../src/dml2_core/dml2_core_shared_types.h | 5 ++ 2 files changed, 36 insertions(+), 16 deletions(-) diff --git a/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_dcn4_calcs.c b/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_dcn4_calcs.c index e2c45e498664..b158b2781239 100644 --- a/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_dcn4_calcs.c +++ b/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_dcn4_calcs.c @@ -5056,6 +5056,8 @@ static bool CalculatePrefetchSchedule(struct dml2_core_internal_scratch *scratch s->trip_to_mem = 0.0; *p->Tvm_trips = 0.0; *p->Tr0_trips = 0.0; + s->Tvm_no_trip_oto = 0.0; + s->Tr0_no_trip_oto = 0.0; s->Tvm_trips_rounded = 0.0; s->Tr0_trips_rounded = 0.0; s->max_Tsw = 0.0; @@ -5293,31 +5295,38 @@ static bool CalculatePrefetchSchedule(struct dml2_core_internal_scratch *scratch s->Lsw_oto = math_ceil2(4.0 * math_max2(s->prefetch_sw_bytes / s->prefetch_bw_oto / s->LineTime, s->min_Lsw_oto), 1.0) / 4.0; if (p->display_cfg->gpuvm_enable == true) { - s->Tvm_oto = math_max3( - *p->Tvm_trips, + s->Tvm_no_trip_oto = math_max2( *p->Tno_bw + vm_bytes * p->HostVMInefficiencyFactor / s->prefetch_bw_oto, s->LineTime / 4.0); + s->Tvm_oto = math_max2( + *p->Tvm_trips, + s->Tvm_no_trip_oto); #ifdef __DML_VBA_DEBUG__ dml2_printf("DML::%s: Tvm_oto max0 = %f\n", __func__, *p->Tvm_trips); dml2_printf("DML::%s: Tvm_oto max1 = %f\n", __func__, *p->Tno_bw + vm_bytes * p->HostVMInefficiencyFactor / s->prefetch_bw_oto); dml2_printf("DML::%s: Tvm_oto max2 = %f\n", __func__, s->LineTime / 4.0); #endif } else { + s->Tvm_no_trip_oto = s->Tvm_trips_rounded; s->Tvm_oto = s->Tvm_trips_rounded; } if ((p->display_cfg->gpuvm_enable == true || p->setup_for_tdlut || dcc_mrq_enable)) { - s->Tr0_oto = math_max3( - *p->Tr0_trips, + s->Tr0_no_trip_oto = math_max2( (p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor + p->meta_row_bytes + tdlut_row_bytes) / s->prefetch_bw_oto, s->LineTime / 4.0); + s->Tr0_oto = math_max2( + *p->Tr0_trips, + s->Tr0_no_trip_oto); #ifdef __DML_VBA_DEBUG__ dml2_printf("DML::%s: Tr0_oto max0 = %f\n", __func__, *p->Tr0_trips); dml2_printf("DML::%s: Tr0_oto max1 = %f\n", __func__, (p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor + p->meta_row_bytes + tdlut_row_bytes) / s->prefetch_bw_oto); dml2_printf("DML::%s: Tr0_oto max2 = %f\n", __func__, s->LineTime / 4); #endif - } else - s->Tr0_oto = (s->LineTime - s->Tvm_oto) / 4.0; + } else { + s->Tr0_no_trip_oto = (s->LineTime - s->Tvm_oto) / 4.0; + s->Tr0_oto = s->Tr0_no_trip_oto; + } s->Tvm_oto_lines = math_ceil2(4.0 * s->Tvm_oto / s->LineTime, 1) / 4.0; s->Tr0_oto_lines = math_ceil2(4.0 * s->Tr0_oto / s->LineTime, 1) / 4.0; @@ -5595,6 +5604,9 @@ static bool CalculatePrefetchSchedule(struct dml2_core_internal_scratch *scratch dml2_printf("DML::%s: Tvm_equ = %f\n", __func__, s->Tvm_equ); dml2_printf("DML::%s: Tr0_equ = %f\n", __func__, s->Tr0_equ); #endif + // Lsw = dst_y_prefetch - (dst_y_per_vm_vblank + 2*dst_y_per_row_vblank) + s->Lsw_equ = s->dst_y_prefetch_equ - math_ceil2(4.0 * (s->Tvm_equ + 2 * s->Tr0_equ) / s->LineTime, 1.0) / 4.0; + // Use the more stressful prefetch schedule if (s->dst_y_prefetch_oto < s->dst_y_prefetch_equ) { *p->dst_y_prefetch = s->dst_y_prefetch_oto; @@ -5603,25 +5615,28 @@ static bool CalculatePrefetchSchedule(struct dml2_core_internal_scratch *scratch *p->dst_y_per_vm_vblank = math_ceil2(4.0 * s->TimeForFetchingVM / s->LineTime, 1.0) / 4.0; *p->dst_y_per_row_vblank = math_ceil2(4.0 * s->TimeForFetchingRowInVBlank / s->LineTime, 1.0) / 4.0; + s->dst_y_per_vm_no_trip_vblank = math_ceil2(4.0 * s->Tvm_no_trip_oto / s->LineTime, 1.0) / 4.0; + s->dst_y_per_row_no_trip_vblank = math_ceil2(4.0 * s->Tr0_no_trip_oto / s->LineTime, 1.0) / 4.0; #ifdef __DML_VBA_DEBUG__ dml2_printf("DML::%s: Using oto scheduling for prefetch\n", __func__); #endif - } else { *p->dst_y_prefetch = s->dst_y_prefetch_equ; s->TimeForFetchingVM = s->Tvm_equ; s->TimeForFetchingRowInVBlank = s->Tr0_equ; - *p->dst_y_per_vm_vblank = math_ceil2(4.0 * s->TimeForFetchingVM / s->LineTime, 1.0) / 4.0; - *p->dst_y_per_row_vblank = math_ceil2(4.0 * s->TimeForFetchingRowInVBlank / s->LineTime, 1.0) / 4.0; + *p->dst_y_per_vm_vblank = math_ceil2(4.0 * s->TimeForFetchingVM / s->LineTime, 1.0) / 4.0; + *p->dst_y_per_row_vblank = math_ceil2(4.0 * s->TimeForFetchingRowInVBlank / s->LineTime, 1.0) / 4.0; + s->dst_y_per_vm_no_trip_vblank = *p->dst_y_per_vm_vblank; + s->dst_y_per_row_no_trip_vblank = *p->dst_y_per_row_vblank; #ifdef __DML_VBA_DEBUG__ dml2_printf("DML::%s: Using equ bw scheduling for prefetch\n", __func__); #endif } - // Lsw = dst_y_prefetch - (dst_y_per_vm_vblank + 2*dst_y_per_row_vblank) - s->LinesToRequestPrefetchPixelData = *p->dst_y_prefetch - *p->dst_y_per_vm_vblank - 2 * *p->dst_y_per_row_vblank; // Lsw + /* take worst case Lsw to calculate bandwidth requirement regardless of schedule */ + s->LinesToRequestPrefetchPixelData = math_min2(s->Lsw_equ, s->Lsw_oto); // Lsw s->cursor_prefetch_bytes = (unsigned int)math_max2(p->cursor_bytes_per_chunk, 4 * p->cursor_bytes_per_line); *p->prefetch_cursor_bw = p->num_cursors * s->cursor_prefetch_bytes / (s->LinesToRequestPrefetchPixelData * s->LineTime); @@ -5741,13 +5756,13 @@ static bool CalculatePrefetchSchedule(struct dml2_core_internal_scratch *scratch if (vm_bytes == 0) { prefetch_vm_bw = 0; - } else if (*p->dst_y_per_vm_vblank > 0) { + } else if (s->dst_y_per_vm_no_trip_vblank > 0) { #ifdef __DML_VBA_DEBUG__ dml2_printf("DML::%s: HostVMInefficiencyFactor = %f\n", __func__, p->HostVMInefficiencyFactor); dml2_printf("DML::%s: dst_y_per_vm_vblank = %f\n", __func__, *p->dst_y_per_vm_vblank); dml2_printf("DML::%s: LineTime = %f\n", __func__, s->LineTime); #endif - prefetch_vm_bw = vm_bytes * p->HostVMInefficiencyFactor / (*p->dst_y_per_vm_vblank * s->LineTime); + prefetch_vm_bw = vm_bytes * p->HostVMInefficiencyFactor / (s->dst_y_per_vm_no_trip_vblank * s->LineTime); #ifdef __DML_VBA_DEBUG__ dml2_printf("DML::%s: prefetch_vm_bw = %f\n", __func__, prefetch_vm_bw); #endif @@ -5759,8 +5774,8 @@ static bool CalculatePrefetchSchedule(struct dml2_core_internal_scratch *scratch if (p->PixelPTEBytesPerRow == 0 && tdlut_row_bytes == 0) { prefetch_row_bw = 0; - } else if (*p->dst_y_per_row_vblank > 0) { - prefetch_row_bw = (p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor + tdlut_row_bytes) / (*p->dst_y_per_row_vblank * s->LineTime); + } else if (s->dst_y_per_row_no_trip_vblank > 0) { + prefetch_row_bw = (p->PixelPTEBytesPerRow * p->HostVMInefficiencyFactor + tdlut_row_bytes) / (s->dst_y_per_row_no_trip_vblank * s->LineTime); #ifdef __DML_VBA_DEBUG__ dml2_printf("DML::%s: PixelPTEBytesPerRow = %u\n", __func__, p->PixelPTEBytesPerRow); @@ -10738,7 +10753,7 @@ static bool dml_core_mode_programming(struct dml2_core_calcs_mode_programming_ex mode_lib->mp.UrgentLatency, mode_lib->mp.TripToMemory, !dml_is_phantom_pipe(&display_cfg->plane_descriptors[k]) && display_cfg->stream_descriptors[display_cfg->plane_descriptors[k].stream_index].timing.drr_config.enabled ? - get_g6_temp_read_blackout_us(&mode_lib->soc, (unsigned int)(mode_lib->ms.uclk_freq_mhz * 1000), in_out_params->min_clk_index) : 0.0); + get_g6_temp_read_blackout_us(&mode_lib->soc, (unsigned int)(mode_lib->mp.uclk_freq_mhz * 1000), in_out_params->min_clk_index) : 0.0); myPipe->Dppclk = mode_lib->mp.Dppclk[k]; myPipe->Dispclk = mode_lib->mp.Dispclk; diff --git a/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_shared_types.h b/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_shared_types.h index 13961c2eb634..cbdfbd5a0bde 100644 --- a/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_shared_types.h +++ b/drivers/gpu/drm/amd/display/dc/dml2/dml21/src/dml2_core/dml2_core_shared_types.h @@ -1187,11 +1187,15 @@ struct dml2_core_calcs_CalculatePrefetchSchedule_locals { double prefetch_bw_oto; double Tvm_oto; double Tr0_oto; + double Tvm_no_trip_oto; + double Tr0_no_trip_oto; double Tvm_oto_lines; double Tr0_oto_lines; double dst_y_prefetch_oto; double TimeForFetchingVM; double TimeForFetchingRowInVBlank; + double dst_y_per_vm_no_trip_vblank; + double dst_y_per_row_no_trip_vblank; double LinesToRequestPrefetchPixelData; unsigned int HostVMDynamicLevelsTrips; double trip_to_mem; @@ -1199,6 +1203,7 @@ struct dml2_core_calcs_CalculatePrefetchSchedule_locals { double Tr0_trips_rounded; double max_Tsw; double Lsw_oto; + double Lsw_equ; double Tpre_rounded; double prefetch_bw_equ; double Tvm_equ; -- 2.34.1
