MIGraphX Driver

Applies to Linux

2024-03-08

6 min read time

read

Loads and prints input graph.

<input file>

File to load

--model [resnet50|inceptionv3|alexnet]

Load model

--onnx

Load as onnx

--tf

Load as tensorflow

--migraphx

Load as MIGraphX

--migraphx-json

Load as MIGraphX JSON

--batch [unsigned int] (Default: 1)

For a static model, set batch size. For a dynamic batch model, sets the batch size at runtime.

--nhwc

Treat tensorflow format as nhwc

--skip-unknown-operators

Skip unknown operators when parsing and continue to parse.

--nchw

Treat tensorflow format as nchw

--trim, -t [unsigned int]

Trim instructions from the end (Default: 0)

--input-dim [std::vector<std::string>]

Dim of a parameter (format: “@name d1 d2 dn”)

--dyn-input-dim [std::vector<std::string>]

Set dynamic dimensions of a parameter using JSON formatting (format “@name” “dynamic_dimension_json”)

--default-dyn-dim

Set the default dynamic dimension (format {min:x, max:y, optimals:[o1,o2,…]})

--optimize, -O

Optimize when reading

--apply-pass, -p

Passes to apply to model

--graphviz, -g

Print out a graphviz representation.

--brief

Make the output brief.

--cpp

Print out the program as cpp program.

--json

Print out program as json.

--text

Print out program in text format.

--binary

Print out program in binary format.

--py

Print out program using python API.

--output, -o [std::string]

Output to file.

compile

Compiles and prints input graph.

--fill0 [std::vector<std::string>]

Fill parameter with 0s

--fill1 [std::vector<std::string>]

Fill parameter with 1s

--gpu

Compile on the gpu

--cpu

Compile on the cpu

--ref

Compile on the reference implementation

--enable-offload-copy

Enable implicit offload copying

--disable-fast-math

Disable fast math optimization

--exhaustive-tune

Perform an exhaustive search to find the fastest version of generated kernels for selected backend

--fp16

Quantize for fp16

--int8

Quantize for int8

--fp8

Quantize for Float8E4M3FNUZ type

run

Loads and prints input graph.

--fill0 [std::vector<std::string>]

Fill parameter with 0s

--fill1 [std::vector<std::string>]

Fill parameter with 1s

--gpu

Compile on the gpu

--cpu

Compile on the cpu

--ref

Compile on the reference implementation

--enable-offload-copy

Enable implicit offload copying

--disable-fast-math

Disable fast math optimization

--exhaustive-tune

Perform an exhaustive search to find the fastest version of generated kernels for selected backend

--fp16

Quantize for fp16

--int8

Quantize for int8

--fp8

Quantize for Float8E4M3FNUZ type

perf

Compiles and runs input graph then prints performance report.

--fill0 [std::vector<std::string>]

Fill parameter with 0s

--fill1 [std::vector<std::string>]

Fill parameter with 1s

--gpu

Compile on the gpu

--cpu

Compile on the cpu

--ref

Compile on the reference implementation

--enable-offload-copy

Enable implicit offload copying

--disable-fast-math

Disable fast math optimization

--exhaustive-tune

Perform an exhaustive search to find the fastest version of generated kernels for selected backend

--fp16

Quantize for fp16

--int8

Quantize for int8

--fp8

Quantize for Float8E4M3FNUZ type

--iterations, -n [unsigned int]

Number of iterations to run for perf report (Default: 100)

verify

Runs reference and CPU or GPU implementations and checks outputs for consistency.

--fill0 [std::vector<std::string>]

Fill parameter with 0s

--fill1 [std::vector<std::string>]

Fill parameter with 1s

--gpu

Compile on the gpu

--cpu

Compile on the cpu

--ref

Compile on the reference implementation

--enable-offload-copy

Enable implicit offload copying

--disable-fast-math

Disable fast math optimization

--exhaustive-tune

Perform an exhaustive search to find the fastest version of generated kernels for selected backend

--fp16

Quantize for fp16

--int8

Quantize for int8

--fp8

Quantize for Float8E4M3FNUZ type

--rms-tol [double]

Tolerance for RMS error (Default: 0.001)

--atol [double]

Tolerance for elementwise absolute difference (Default: 0.001)

--rtol [double]

Tolerance for elementwise relative difference (Default: 0.001)

-i, --per-instruction

Verify each instruction

-r, --reduce

Reduce program and verify

roctx

Provides marker information for each operation, allowing MIGraphX to be used with rocprof for performance analysis. This allows user to get GPU-level kernel timing information. An example command line combined with rocprof for tracing purposes is given below:

/opt/rocm/bin/rocprof --hip-trace --roctx-trace --flush-rate 1ms --timestamp on -d <OUTPUT_PATH> --obj-tracking on /opt/rocm/bin/migraphx-driver roctx <ONNX_FILE> <MIGRAPHX_OPTIONS>

After rocprof is run, the output directory will contain trace information for HIP, HCC and ROCTX in seperate .txt files. To understand the interactions between API calls, it is recommended to utilize roctx.py helper script as desribed in dev/tools:rocTX section.

--fill0 [std::vector<std::string>]

Fill parameter with 0s

--fill1 [std::vector<std::string>]

Fill parameter with 1s

--gpu

Compile on the gpu

--cpu

Compile on the cpu

--ref

Compile on the reference implementation

--enable-offload-copy

Enable implicit offload copying

--disable-fast-math

Disable fast math optimization

--exhaustive-tune

Perform an exhaustive search to find the fastest version of generated kernels for selected backend

--fp16

Quantize for fp16

--int8

Quantize for int8

--fp8

Quantize for Float8E4M3FNUZ type