package tensorflow.profiler.roofline_model
Get desktop application:
View/edit binary Protocol Buffers messages
The record type which describes the scope this record captures.
Used in:
INVALID_RECORD_TYPE = 0
ALL = 1
Captures the entire profiling duration including incomplete steps.
AVERAGE_STEP = 2
Captures the average of all complete steps.
PER_STEP = 3
Captures a single step.
ALL_HW = 4
Same as ALL but the performance metrics (FLOPS and memory bandwidth) are derived from the hardware performance counters.
message
roofline_model.proto:30A database of RooflineModel records. Next ID: 17
optional string device_type = 1
The device type.
optional bool megacore = 12
Whether megacore is used.
optional bool has_cmem = 8
Whether the device has shared CMEM.
optional bool has_merged_vmem = 15
Whether the device has merged VMEM.
optional double peak_flop_rate = 2
Peak flop rate in GFLOP/s.
optional double peak_hbm_bw = 9
Peak HBM bandwidth in GiB/s
optional double peak_cmem_read_bw = 10
Peak CMEM read bandwidth in GiB/s
optional double peak_cmem_write_bw = 11
Peak CMEM write bandwidth in GiB/s
optional double peak_vmem_read_bw = 13
Peak VMEM read bandwidth in GiB/s
optional double peak_vmem_write_bw = 14
Peak VMEM write bandwidth in GiB/s
repeated roofline_model_record = 5
All RooflineModel records, one for each HLO operation.
optional diagnostics = 7
Error and warning messages for diagnosing profiling issues.
optional double time_scale_multiplier = 16
The weighted average time scale multiplier over all device op metrics. It is calculated by the total time of each op if running on default pstate divided by actual total op time.
message
roofline_model.proto:77There is one RooflineModelRecord for each HLO operation profiled. Next ID: 46
Used in:
optional RecordType record_type = 18
The record type.
optional uint32 step_num = 19
Step number when record type is PER_STEP. Otherwise, invalid.
optional uint64 rank = 1
The rank by self time
optional uint64 hlo_module_id = 35
The hlo module id of the op
optional string hlo_category = 17
The HLO category name.
optional string hlo_name = 2
The HLO operation name.
optional int64 occurrences = 3
Number of occurrences of the operation.
optional double total_time_in_us = 4
Total "accumulated" time in micro-seconds that the operation took. If this operation has any children operations, the "accumulated" time includes the time spent inside children.
optional double total_time_per_core_in_us = 20
Total time per core in micro-seconds.
optional double total_time_in_percentage = 21
Total time as fraction of the total program time.
optional double avg_time_in_us = 5
Average "accumulated" time in micro-seconds that each occurrence of the operation took.
optional double total_self_time_in_us = 6
Total "self" time in micro-seconds that the operation took. If this operation has any children operations, the "self" time doesn't include the time spent inside children.
optional double avg_self_time_in_us = 7
Average "self" time in micro-seconds that the operation took.
optional double total_self_time_as_fraction = 8
Percentage of the total "accumulated" time that was caused by DMA stall.
optional double cumulative_total_self_time_as_fraction = 9
optional double dma_stall_fraction = 10
optional double measured_flop_rate = 13
Number of total floating-point operations (FLOPs) performed per second normalized to the bf16 peak performance.
optional double model_flop_rate = 38
Numbef or total floating point operations (FLOPs) performed per second for the op.
optional double measured_memory_bw = 14
Number of total bytes (including both read and write) accessed per second.
optional double hbm_bw = 27
HBM bandwidth in GiB/s (including both read and write).
optional double cmem_read_bw = 28
CMEM read bandwidth in GiB/s.
optional double cmem_write_bw = 29
CMEM write bandwidth in GiB/s.
optional double vmem_read_bw = 39
VMEM read bandwidth in GiB/s.
optional double vmem_write_bw = 40
VMEM write bandwidth in GiB/s.
optional double operational_intensity = 15
Overall operational intensity in FLOP/Byte.
optional double hbm_operational_intensity = 30
Operational intensity based on HBM in FLOP/Byte.
optional double cmem_read_operational_intensity = 31
Operational intensity based on CMEM read in FLOP/Byte.
optional double cmem_write_operational_intensity = 32
Operational intensity based on CMEM write in FLOP/Byte.
optional double vmem_read_operational_intensity = 41
Operational intensity based on VMEM read in FLOP/Byte.
optional double vmem_write_operational_intensity = 42
Operational intensity based on VMEM write in FLOP/Byte.
optional double bottleneck_operational_intensity = 33
Operational intensity based on the bottleneck resource in FLOP/Byte.
optional string bound_by = 16
Whether this operation is "Compute", "HBM", "CMEM Read", "CMEM Write" bound, according to the Roofline Model.
optional double optimal_flop_rate = 22
The optimal flop rate calculated as (operational intensity) * (peak memory bw)
optional double roofline_efficiency = 34
Roofline efficiency.
optional double flop_rate_relative_to_hw_limit = 24
Percentage of measured flop rate relative to the hardware limit.
optional double memory_bw_relative_to_hw_limit = 25
Percentage of measured memory bandwidth relative to the hardware limit.
optional bool include_infeed_outfeed = 26
Whether the record is calculated including infeed and outfeed ops.
optional bool apply_time_scale_multiplier = 44
Whether the device metrics (eg. flops utilization) is calculated with by applying the time scale multiplier.
optional uint64 flops = 36
Deprecated. Use flops_v2 instead.
optional double flops_v2 = 45
Flops for the record
optional uint64 bytes_accessed = 37
Bytes accessed for the record
optional source_info = 43
Infrmation about the corresponding source code.