CANN/asc-devkit Fixpipe性能测试

Fixpipe (L0C Egress) Performance Test Example

【免费下载链接】asc-devkit本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言，原生支持C和C++标准规范，主要由类库和语言扩展层构成，提供多层级API，满足多维场景算子开发诉求。项目地址: https://gitcode.com/cann/asc-devkit

Overview

This example tests the performance of the matrix computation egress path. It covers the data path that moves Cube computation results from the L0C Buffer to the L1 Buffer or Unified Buffer (UB).

This is a non-functional performance test. It does not verify computation results and only collects the Fixpipe egress latency.

Supported Products and CANN Versions

Directory Structure

├── fixpipe_perf │ ├── CMakeLists.txt // Build configuration file │ ├── fixpipe_perf.asc // Fixpipe egress performance test implementation and entry point │ ├── perf.sh // Performance test script │ ├── generate_roofline.py // Roofline generation script │ ├── README.md // Example documentation

Example Description

This example uses the runtime parameterSCENARIO_NUMto select different egress paths and data types. Matrix dimensions are passed at runtime through./demo SCENARIO_NUM M K N.

The two egress paths use different interfaces:

The supported test scenarios vary by platform architecture:

Atlas A3/A2 Training/Inference Platform Scenarios

Ascend 950PR/950DT Platform Scenarios

Theoretical bandwidth is calculated based on hardware egress parallelism: the egress unit processes 64 output elements per cycle, so theoretical bandwidth = 64 × sizeof(destination data type) (Byte/cycle). Dual-target mode splits work across two sub-cores in parallel, effectively doubling the parallelism. The L0C Buffer to UB path is supported only on dav-3510 (scenarios 13 and 14). Scenario 14 splits the M×N matrix in the L0C Buffer along the M dimension into two halves and writes them simultaneously to the UB of two Vector cores.

Build and Run

Run the following steps from the root directory of this example to build and run it.

Configure Environment Variables

Configure environment variables based on the installation method of the CANN development kit on your system.

source ${install_path}/cann/set_env.sh

${install_path}is the CANN package installation directory. If no installation directory is specified, the default installation path is/usr/local/Ascend.

Build the Example

Build for Atlas A3/A2 Training/Inference Platform (dav-2201):

mkdir -p build && cd build cmake -DCMAKE_ASC_ARCHITECTURES=dav-2201 .. make -j cd ..

Build for Ascend 950PR/950DT Platform (dav-3510):

mkdir -p build && cd build cmake -DCMAKE_ASC_ARCHITECTURES=dav-3510 .. make -j cd ..

Run the Example

The runtime parameter order isSCENARIO_NUM M K N:

# Atlas A3/A2 Training/Inference Platform examples (scenarios 1, 2) ./build/demo 1 128 64 128 ./build/demo 2 128 64 128 # Ascend 950PR/950DT Platform examples (scenarios 11-14) ./build/demo 11 128 64 128 ./build/demo 12 128 64 128 ./build/demo 13 128 64 128 ./build/demo 14 128 64 128

Matrix dimensions must meet alignment requirements: M and N must be multiples of 16. For dual-target mode split along the M dimension, M must be a multiple of 2. For split along the N dimension, N must be a multiple of 32.

Collecting Performance Data

Use themsproftool to collect detailed performance data:

msprof op build/demo 1 128 64 128

💡 Themsproftool requires CANN Commercial or Community Edition. For details, refer to the msprof Tool Installation Guide.

After the command completes, a folder namedOPPROF_{timestamp}_XXXis generated in the default directory. The performance data folder structure is as follows:

├── dump # Raw performance data ├── ArithmeticUtilization.csv # Cube/vector instruction cycle ratio ├── L2Cache.csv # L2 Cache hit rate, which affects MTE2. Plan data movement logic carefully to increase hit rate ├── Memory.csv # UB, L1 Buffer, and main memory read/write bandwidth ├── MemoryL0.csv # L0A Buffer, L0B Buffer, and L0C Buffer read/write bandwidth ├── MemoryUB.csv # Vector and Scalar to UB read/write bandwidth ├── OpBasicInfo.csv # Operator basic information ├── PipeUtilization.csv # Computation unit and transfer unit latency and ratio ├── ResourceConflictRatio.csv # UB bank group, bank conflict, and resource conflict ratio └── visualize_data.bin # MindStudio Insight presentation file

This example focuses on L0C Buffer egress performance data. View the specific performance data results as follows:

cat ./OPPROF_*/PipeUtilization.csv

Key metrics to monitor:

Performance Test Script

perf.shperforms batch building, runsmsprof op, extracts Fixpipe egress latency, and generates a CSV summary.

# View help ./perf.sh --help # Test scenario 1, using dav-2201 by default ./perf.sh 1 # Test scenario 13, using dav-3510 by default ./perf.sh 13 # Explicitly specify the platform. The platform must match the scenario; otherwise, an error is reported ./perf.sh 1 dav-2201 ./perf.sh 13 dav-3510

The script uses a built-in default shape sequence. Egress performance depends only on M and N. K is fixed at 64 to allow Mmad to produce L0C Buffer data beforehand. The sequence starts from a small M·N value and gradually increases to full capacity. The L0C Buffer size is 128 KB for dav-2201 and 256 KB for dav-3510. The saturation points differ, so the last entry differs between the two architectures:

dav-2201 default shape sequence:

dav-3510 default shape sequence:

The L0C Buffer egress data volume is M × N × sizeof(destination type). K is used only to allow Mmad to produce the M×N result in the L0C Buffer beforehand and is not counted in the egress volume. Adjust K as needed.

To test specific shapes, run./build/demo SCENARIO_NUM M K Ndirectly.

After testing, results are saved toperf_data_${timestamp}_scenario${SCENARIO}/perf_result_scenario${SCENARIO}.csv. Rawmsprofdata is saved in thetest_${id}_${M}_${K}_${N}subdirectory under the same directory.

Performance Metrics

perf.shextractsaic_fixpipe_time(us)fromPipeUtilization.csvand calculates bandwidth based on the platform clock frequency and egress volume.

The computed columns in the CSV are as follows:

Performance Metric Calculation Methods

Theaic_fixpipe_time(us)collected bymsprofinPipeUtilization.csvis the egress latency in microseconds.perf.shreads this time column and calculates the cycle count and measured bandwidth based on the platform clock frequency and egress data volume.

Converting Time to Cycles

The clock frequency unit is MHz, which represents cycles per microsecond. No additional conversion is needed:

Cycle = Time(us) × Frequency(MHz)

For example, the Atlas A3/A2 Training/Inference Platform has a clock frequency of 1800 MHz. Ifaic_fixpipe_time(us) = 0.050000:

Cycle = 0.050000 × 1800 = 90.00 cycles

Data Transfer Volume Calculation

perf.shcalculates the egress volume based on the destination data type for each scenario. This value is used asDataSize(bytes)for bandwidth calculation:

Measured Bandwidth Calculation

Bandwidth is output in GB/s. SinceTime(us)is in microseconds,DataSize(bytes) / Time(us)yields MB/s. Dividing by1e3converts to GB/s:

Bandwidth(GB/s) = DataSize(bytes) / Time(us) / 1e3

For example, scenario 11 with shape[128, 64, 128]and destination type half:

DataSize = 128 × 128 × 2 = 32768 bytes Time = 0.050000 us Bandwidth = 32768 / 0.050000 / 1e3 = 655.360 GB/s

Theoretical Latency and Bandwidth Utilization

The "Theoretical Bandwidth (Byte/cycle)" and "Bandwidth Latency (cycle)" in the scenario table can be used to estimate theoretical latency. Theoretical bandwidth is derived from egress parallelism (the egress unit processes 64 output elements per cycle, theoretical bandwidth = 64 × sizeof(destination type), doubled for dual-target mode). The fixed latency represents the base startup overhead for a single transfer. The theoretical transfer time can be estimated as follows:

TransferCycle = DataSize(bytes) / TheoreticalBandwidth(Byte/cycle) TheoryCycle = Latency(cycle) + TransferCycle TheoryTime(us) = TheoryCycle / Frequency(MHz) TheoryBandwidth(GB/s) = DataSize(bytes) / TheoryTime(us) / 1e3

The ratio of measured bandwidth to theoretical bandwidth can be used to evaluate bandwidth utilization:

BandwidthUtilization = MeasuredBandwidth(GB/s) / TheoryBandwidth(GB/s) × 100%

The platform clock frequency is set automatically byperf.shbased on the scenario:

Roofline Analysis

This example providesgenerate_roofline.py, which generates ASCII reports and charts from the CSV output produced byperf.sh.

Python Package Dependencies

generate_roofline.pyuses Python standard libraries to read CSV and generate ASCII reports. To generate PNG/PDF charts, installmatplotlibandnumpy.

python3 -m pip install --user matplotlib numpy

If these dependencies are not installed, the script still generates.txtASCII analysis reports but skips chart generation.

# Automatically find the latest perf_data directory results python3 generate_roofline.py # Specify a CSV file python3 generate_roofline.py --csv perf_data_xxx_scenario11/perf_result_scenario11.csv

The script has built-in egress parallelism for each scenario. You do not need to manually specify peak bandwidth. The first-instruction overhead defaults to the scenario value (20 cycles for dav-2201, 26 cycles for dav-3510), and the clock frequency is automatically selected based on the scenario.

Chart Example

The following is a Roofline chart example generated for scenario 1:

Notes

1. Scenario numbers must match the platform: use scenarios 1, 2 for dav-2201 and scenarios 11-14 for dav-3510.perf.shvalidates the match and reports errors for mismatches.
2. The L0C Buffer to UB path (scenarios 13, 14) is supported only on dav-3510 and uses theFixpipeinterface with the hybrid programming framework.
3. Matrix dimensions must meet alignment requirements: M and N must be multiples of 16. For dual-target mode split along the M dimension, M must be a multiple of 2. For split along the N dimension, N must be a multiple of 32.
4. This is a pure performance test and does not verify computation results. The kernel function does not initialize data in the L0A Buffer or L0B Buffer. It retains a minimal Mmad operation to produce L0C Buffer data beforehand.
5. APipeBarrier<PIPE_ALL>is inserted between the Mmad pre-operation and the egress instruction to prevent pipeline overlap from causing inaccurateaic_fixpipe_timestatistics.

【免费下载链接】asc-devkit本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言，原生支持C和C++标准规范，主要由类库和语言扩展层构成，提供多层级API，满足多维场景算子开发诉求。项目地址: https://gitcode.com/cann/asc-devkit