The SDK provides debug and compiler options for the user to configure the compiler behavior.
This document provides a description of the debug and compile options that are
set using the SetFlag function.
GetInfo returns some information about the SDK. The flags for GetInfo are
also provided here.
These options are the same for both the C API and the Python API.
debug: set what outputs the API will print during its execution, a string containing zero or more of the following characters:
- 'b': print basic messages
- 'w': print warnings
- 'c': print memory allocation addresses in external memory
- 'C': print workload distribution accross clusters
- 'n': print list of layers
- 's': print partitioning strategy
- 'S': print info about DLA instructions run by the simulator (only with fpgaid 'sim')
- 'p': print list of operations
- 'd': print accelerator debug registers
- 'a': print memory access pattern. All load/store lengths and addresses
- 'r': print wrong results only. This needs to be used with SetFlag('options', 's')`.
- 'R': print all results returned by the accelerator. This needs to be used with
SetFlag('options', 's'). - 'i': dump accelerator instructions into instr.txt
- 'I': print intermediate code instructions
- 't': print dynamic instruction trace
- 'T': print verbose dynamic instruction trace
- 'o': print thnets debuging
- 'Q': print variable fixpoint selected of each layer
- 'J': dump the partitioning information into partition.json
options: configures compile options of the SDK, a string containing zero or more of the following characters:
- 'v': compile for varfp hardware, if hardware is not present and thus not auto-detected, same as has_varfp=1
- 'V': enable variable fix-point precision. Default precision is Q8.8, same as varfp=1
- 'M': enable multi-MM vertical padding optimization, same as multicu_vpadding=1
- 'C': no batch mode for multiple clusters, same as clusterbatchmode=1
- 'B': sandbox mode. Reads DLA instruction from a text file, same as sandbox=1
- 'k': try kernel_stationary optimization if possible, same as convalgo=1 [Paper reference]
- 'r': try kernel_stationary option if possible, maxpool will reshape output, same as convalgo=2
- 'K': try kernel_stationary option if possible, with yPxp ordering (instead of Pyxp), maxpool will reshape output, same as convalgo=3
- 'f': force one of kernel stationary modes above, same as force_algo=1
- 'd': run tests with deterministic input and weights instead of random, same as rand=0
- 'P': progressive load instead of DLA banks switching, same as progload=1
- 'w': wait one second for hardware to finish instead of polling the DLA, same as hwwait=1
- 's': run the software version for comparing the accelerator output, same as comparesw=1
- 'S': do not run DLA. Only run software version, same as runsw=1
- 'Q': do not run DLA. Only run software version using float precision and save quantization metrics: variable fix-point for inputs, intermediate activations and outputs, same as runsw=2
- 'i': measure time to load the initial data into DLA, same as initmeasure=1
- 'L': profile each layer separately: run each layer in the model individually (measure execution time of each layer). This option only works with the Run function, same as breaklayer=2
- 'z': profile each layer separately: run entire model and output each layer's output (only to check output of each layer), same as breaklayer=1
- 't': save input and output of the inference in a file, same as temporary_save=1
- 'm': MV level batch (batchsize must be a multiple of 4), same as mvbatch=1
- 'a': compile, run and check which option works better. loop{ compile, run, save_best_choice }, same as compile_profile=1
- 'x': enable prune outplanes if they are zero, same as prune=1
- 'n': disable computes optimizations (useful for debugging), same as optimize_computes=0
- 'l': merge computes creating loops, same as gen_loops=1
reset: reset flags to the default values
bitfile: immediately upload the given bitfile
fpgaid: FPGA to use, 510, 511, 852 or sim, default is -1, which will use the first FPGA found
firstcluster: First cluster to use
nclusters: Number of clusters to use
nfpgas: Number of FPGAs to use
imgs_per_cluster: Per-cluster batch size (number of inputs processed by each cluster)
clustersbatchmode: If 1, multiple clusters process different parts of the same input (no clusters-level batching)
bufsmode: 0=one program for each of the two buffers, 1=parametric offset, 2=program patching at each run
comparesw: Compare DLA output to reference software output
progload: Progressive loading of maps instead of banks switching
convalgo: Convolution algorithm: 0 non-kernel stationary (default), 1 Pyxp, 2 Pyxp maxpool-reshapes, 3 yPxp maxpool-reshapes
force_algo: Force convolution kernel stationary mode instead of leaving it to automatic choice
compile_profile: True: loop{compile, run, save_best_choice} False: compile (only)
depthconvalgo: Depth-convolution algorithm: 0 add reshaping layer, 1 pixel-by-pixel reshaping, 2 row-by-row reshaping inside compute, 3 create reshaping object before store object
gen_loops: Generate hardware loops to group similar operations, when possible
multicu_vpadding: Vertical padding optimization for convolutions
optimize_computes: Optimize compute objects by applying double buffering, removing double loads, etc... useful to disable for simpler debugging, default 1
prune: Prune output planes
blockingmode: Blocking mode API operation, putinput/getresult will wait instead of failing when the device is busy
mvbatch: Distribute inputs among MVs, imgs_per_cluster must be a multiple of 4
int8mode: Select int8 mode: 0=no, 1=int8 hw, 2=auto, 3=16->8, 4=8->16
hwlinear: Run linear layers in hardware (default True)
addroff: Set starting address on DLA
runsw: 0: Standard operation, 1: Run reference software instead of DLA, 2: Run reference software in floating-point mode
seed: Set seed for random values to get deterministic results at each run using random values
rand: Use random values for input, weights and bias for random tests, default True
nlayers: Number of layers to parse (-1=all, default)
lremove: Remove first N layers
dumpdir: Dump all writes to DLA memory to this directory
has_varfp: Compile for variable-fixed-point hardware
varfp: Use variable fixed-point weights if 1
sandbox: Read code from instr_.txt
breaklayer: 0 (normal operation), 1 run entire model, but output all intermediate results, 2 run each layer individually
initmeasure: Measure the initial loading time
hwwait: Wait 1 second after inference start instead of polling the hardware to detect the end of inference
max_instruction_size: Max instructions memory size
temporary_save: Cache testing data on disk
no_rearrange: Skip output rearrangement
heterogeneous: Run DLA-unsupported layers on CPU also in the middle of the network
Gets information from the DLA or return flags that have been set with SetFlag; additionally to the flags specified above, these names can be passed:
hwtime: returns the number of milliseconds taken by the Micron DLA hardware for the processing only, returned as a float.
addroff: get the last address of the memory map, returned as an int64. Used to attach a new memory map using this addr_off.
freq: the Micron DLA hardware frequency, returned as an int.
maxcluster: the maximum number of clusters in Micron DLA hardware, returned as an int.
maxfpga: the maximum number of FPGAs available, returned as an int.
version: the version number of the SDK, returned as a string.
build: the build hash, returned as a string.
hwversion: the DLA hardware version, returned as a string.
hwbuild: the DLA hardware build, return as a string.
hwdebug: print the DLA registers.
temperature: print the DLA temperature in Celsius, returned as a float.
busy_comp: returns if DLA is busy processing an input buffer, returned as an int.
inshapes: semi-colon separated input shapes in the format size0xsize1...sizeN
outshapes: semi-colon separated output shapes in the format size0xsize1...sizeN
innames: semi-colon separated inputs names
outnames: semi-colon separated outputs names
print_flags: print the currently set flags
help: print the available SetFlag and GetInfo options