Bram and LUTs are overutilized #758

sandeep1404 · 2023-04-11T12:21:56Z

sandeep1404
Apr 11, 2023

Hello everyone, I amtrying to synthesize a model for image denoising, My model summary() is as below:

Model: "model_8"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 input_ (InputLayer)            [(None, 128, 128, 3  0           []                               
                                )]                                                                
                                                                                                  
 conv0 (QConv2D)                (None, 128, 128, 16  448         ['input_[0][0]']                 
                                )                                                                 
                                                                                                  
 conv1 (QConv2D)                (None, 128, 128, 16  2320        ['conv0[0][0]']                  
                                )                                                                 
                                                                                                  
 q_activation_57 (QActivation)  (None, 128, 128, 16  0           ['conv1[0][0]']                  
                                )                                                                 
                                                                                                  
 conv2 (QConv2D)                (None, 128, 128, 16  2320        ['q_activation_57[0][0]']        
                                )                                                                 
                                                                                                  
 q_activation_58 (QActivation)  (None, 128, 128, 16  0           ['conv2[0][0]']                  
                                )                                                                 
                                                                                                  
 conv3 (QConv2D)                (None, 128, 128, 16  2320        ['q_activation_58[0][0]']        
                                )                                                                 
                                                                                                  
 q_activation_59 (QActivation)  (None, 128, 128, 16  0           ['conv3[0][0]']                  
                                )                                                                 
                                                                                                  
 add_34 (Add)                   (None, 128, 128, 16  0           ['q_activation_59[0][0]',        
                                )                                 'conv0[0][0]']                  
                                                                                                  
 conv4 (QConv2D)                (None, 128, 128, 16  2320        ['add_34[0][0]']                 
                                )                                                                 
                                                                                                  
 q_activation_60 (QActivation)  (None, 128, 128, 16  0           ['conv4[0][0]']                  
                                )                                                                 
                                                                                                  
 conv5 (QConv2D)                (None, 128, 128, 16  2320        ['q_activation_60[0][0]']        
                                )                                                                 
                                                                                                  
 add_35 (Add)                   (None, 128, 128, 16  0           ['conv5[0][0]',                  
                                )                                 'add_34[0][0]']                 
                                                                                                  
 q_activation_61 (QActivation)  (None, 128, 128, 16  0           ['add_35[0][0]']                 
                                )                                                                 
                                                                                                  
 conv6 (QConv2D)                (None, 128, 128, 16  2320        ['q_activation_61[0][0]']        
                                )                                                                 
                                                                                                  
 q_activation_62 (QActivation)  (None, 128, 128, 16  0           ['conv6[0][0]']                  
                                )                                                                 
                                                                                                  
 conv7 (QConv2D)                (None, 128, 128, 16  272         ['q_activation_62[0][0]']        
                                )                                                                 
                                                                                                  
 q_dense_2 (QDense)             (None, 128, 128, 2)  34          ['conv7[0][0]']                  
                                                                                                  
 q_activation_64 (QActivation)  (None, 128, 128, 2)  0           ['q_dense_2[0][0]']              
                                                                                                  
 q_dense_3 (QDense)             (None, 128, 128, 16  48          ['q_activation_64[0][0]']        
                                )                                                                 
                                                                                                  
 add_36 (Add)                   (None, 128, 128, 16  0           ['conv7[0][0]',                  
                                )                                 'q_activation_61[0][0]']        
                                                                                                  
 activation_1 (Activation)      (None, 128, 128, 16  0           ['q_dense_3[0][0]']              
                                )                                                                 
                                                                                                  
 q_activation_63 (QActivation)  (None, 128, 128, 16  0           ['add_36[0][0]']                 
                                )                                                                 
                                                                                                  
 multiply_1 (Multiply)          (None, 128, 128, 16  0           ['activation_1[0][0]',           
                                )                                 'q_activation_63[0][0]']        
                                                                                                  
 conv8 (QConv2D)                (None, 128, 128, 3)  435         ['multiply_1[0][0]']             
                                                                                                  
 add_37 (Add)                   (None, 128, 128, 3)  0           ['conv8[0][0]',                  
                                                                  'input_[0][0]']                 
                                                                                                  
==================================================================================================
Total params: 15,157
Trainable params: 15,157
Non-trainable params: 0
__________________________________________________________________________________________________

every layer has a number of parameters less than 4096, my configuration is shown as follows:

import hls4ml

hls4ml.model.optimizer.OutputRoundingSaturationMode.layers = ['Activation']
hls4ml.model.optimizer.OutputRoundingSaturationMode.rounding_mode = 'AP_RND'
hls4ml.model.optimizer.OutputRoundingSaturationMode.saturation_mode = 'AP_SAT'



hls_config = hls4ml.utils.config_from_keras_model(model, granularity='name')
hls_config['Model']['Precision'] = 'ap_fixed<16,6>'

############################################

# Precision = 'ap_fixed<16,6>'
Reuse_Factor = 1000000

# hls_config['Model']['Precision'] = Precision
hls_config['Model']['ReuseFactor'] = Reuse_Factor
hls_config['IOTtype']= 'io_stream'

for Layer in hls_config['LayerName'].keys():
    hls_config['LayerName'][Layer]['Strategy'] = 'resource'
    
hls_config['Model']['Strategy'] = 'Resource'



hls_config['LayerName']['conv0']['ReuseFactor'] = 432
hls_config['LayerName']['conv1']['ReuseFactor'] = 2304
hls_config['LayerName']['conv2']['ReuseFactor'] = 2304
hls_config['LayerName']['conv3']['ReuseFactor'] = 2304
hls_config['LayerName']['conv4']['ReuseFactor'] = 2304
hls_config['LayerName']['conv5']['ReuseFactor'] = 2304
hls_config['LayerName']['conv6']['ReuseFactor'] = 2304
hls_config['LayerName']['conv7']['ReuseFactor'] = 256
hls_config['LayerName']['conv8']['ReuseFactor'] = 432
hls_config['LayerName']['q_dense_2']['ReuseFactor'] = 32
hls_config['LayerName']['q_dense_3']['ReuseFactor'] = 32


print("-----------------------------------")
print("Configuration")
plotting.print_dict(hls_config)
print("-----------------------------------")

cfg_q = hls4ml.converters.create_config(backend='Vivado')
cfg_q['IOType']     = 'io_stream' # Must set this if using CNNs!
cfg_q['HLSConfig']  = hls_config
cfg_q['KerasModel'] = model
cfg_q['OutputDir']  = 'model_zcu104_fpn/hls4ml_prj_pynq_fpnrnet_8bit_qkeras_2'
cfg_q['XilinxPart'] = 'xczu7ev-ffvc1156-2-e'

hls_model = hls4ml.converters.keras_to_hls(cfg_q)

Interpreting Model
Topology:
Layer name: input_, layer type: Input
Layer name: conv0, layer type: QConv2D
Layer name: conv1, layer type: QConv2D
Layer name: q_activation_57, layer type: QActivation
Layer name: conv2, layer type: QConv2D
Layer name: q_activation_58, layer type: QActivation
Layer name: conv3, layer type: QConv2D
Layer name: q_activation_59, layer type: QActivation
Layer name: add_34, layer type: Add
Layer name: conv4, layer type: QConv2D
Layer name: q_activation_60, layer type: QActivation
Layer name: conv5, layer type: QConv2D
Layer name: add_35, layer type: Add
Layer name: q_activation_61, layer type: QActivation
Layer name: conv6, layer type: QConv2D
Layer name: q_activation_62, layer type: QActivation
Layer name: conv7, layer type: QConv2D
Layer name: q_dense_2, layer type: QDense
Layer name: q_activation_64, layer type: QActivation
Layer name: q_dense_3, layer type: QDense
Layer name: add_36, layer type: Add
Layer name: activation_1, layer type: Activation
Layer name: q_activation_63, layer type: QActivation
Layer name: multiply_1, layer type: Multiply
Layer name: conv8, layer type: QConv2D
Layer name: add_37, layer type: Add
-----------------------------------
Configuration
Model
  Precision:         ap_fixed<16,6>
  ReuseFactor:       1000000
  Strategy:          Resource
LayerName
  input_
    Precision
      result:        ap_fixed<16,6>
    Strategy:        resource
  conv0
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     432
    Strategy:        resource
  conv1
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     2304
    Strategy:        resource
  q_activation_57
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  conv2
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     2304
    Strategy:        resource
  q_activation_58
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  conv3
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     2304
    Strategy:        resource
  q_activation_59
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  add_34
    Precision:       ap_fixed<16,6>
    Strategy:        resource
  conv4
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     2304
    Strategy:        resource
  q_activation_60
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  conv5
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     2304
    Strategy:        resource
  add_35
    Precision:       ap_fixed<16,6>
    Strategy:        resource
  q_activation_61
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  conv6
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     2304
    Strategy:        resource
  q_activation_62
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  conv7
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     256
    Strategy:        resource
  q_dense_2
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     32
    Strategy:        resource
  q_activation_64
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  q_dense_3
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     32
    Strategy:        resource
  add_36
    Precision:       ap_fixed<16,6>
    Strategy:        resource
  activation_1
    Precision:       ap_fixed<16,6>
    ReuseFactor:     1
    table_size:      1024
    table_t:         ap_fixed<18,8>
    Strategy:        resource
  q_activation_63
    Precision
      result:        ap_ufixed<8,0>
    ReuseFactor:     1
    Strategy:        resource
  multiply_1
    Precision:       ap_fixed<16,6>
    Strategy:        resource
  conv8
    Precision
      weight:        ap_fixed<8,1>
      bias:          ap_fixed<8,1>
    ReuseFactor:     432
    Strategy:        resource
  add_37
    Precision:       ap_fixed<16,6>
    Strategy:        resource
IOTtype:             io_stream
-----------------------------------
Interpreting Model
Topology:
Layer name: input_, layer type: InputLayer, input shapes: [[None, 128, 128, 3]], output shape: [None, 128, 128, 3]
Layer name: conv0, layer type: QConv2D, input shapes: [[None, 128, 128, 3]], output shape: [None, 128, 128, 16]
Layer name: conv1, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_activation_57, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: conv2, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_activation_58, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: conv3, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_activation_59, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: add_34, layer type: Merge, input shapes: [[None, 128, 128, 16], [None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: conv4, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_activation_60, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: conv5, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: add_35, layer type: Merge, input shapes: [[None, 128, 128, 16], [None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_activation_61, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: conv6, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_activation_62, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: conv7, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_dense_2, layer type: QDense, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 2]
Layer name: q_activation_64, layer type: Activation, input shapes: [[None, 128, 128, 2]], output shape: [None, 128, 128, 2]
Layer name: q_dense_3, layer type: QDense, input shapes: [[None, 128, 128, 2]], output shape: [None, 128, 128, 16]
Layer name: add_36, layer type: Merge, input shapes: [[None, 128, 128, 16], [None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: activation_1, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: q_activation_63, layer type: Activation, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: multiply_1, layer type: Merge, input shapes: [[None, 128, 128, 16], [None, 128, 128, 16]], output shape: [None, 128, 128, 16]
Layer name: conv8, layer type: QConv2D, input shapes: [[None, 128, 128, 16]], output shape: [None, 128, 128, 3]
Layer name: add_37, layer type: Merge, input shapes: [[None, 128, 128, 3], [None, 128, 128, 3]], output shape: [None, 128, 128, 3]
Creating HLS model
WARNING: Config parameter "Precision" overwrites an existing attribute in layer "conv7" (PointwiseConv2D)
WARNING: Config parameter "ReuseFactor" overwrites an existing attribute in layer "conv7" (PointwiseConv2D)
WARNING: Config parameter "Strategy" overwrites an existing attribute in layer "conv7" (PointwiseConv2D)

I was trying to synthesize the above model on zcu104 using Vivado as the backend, but during synthesis and exporting IP, the model is overutilizing the Bram_18k and LUTs, utilization report of the model is shown below:

================================================================
== Utilization Estimates
================================================================
* Summary: 
+-----------------+---------+-------+--------+--------+-----+
|       Name      | BRAM_18K| DSP48E|   FF   |   LUT  | URAM|
+-----------------+---------+-------+--------+--------+-----+
|DSP              |        -|      -|       -|       -|    -|
|Expression       |        -|      -|       0|      40|    -|
|FIFO             |     8541|      -|  159658|  527236|    -|
|Instance         |       22|      3|   35990|   69181|    -|
|Memory           |        -|      -|       -|       -|    -|
|Multiplexer      |        -|      -|       -|      36|    -|
|Register         |        -|      -|       6|       -|    -|
+-----------------+---------+-------+--------+--------+-----+
|Total            |     8563|      3|  195654|  596493|    0|
+-----------------+---------+-------+--------+--------+-----+
|Available        |      624|   1728|  460800|  230400|   96|
+-----------------+---------+-------+--------+--------+-----+
|Utilization (%)  |     1372|   ~0  |      42|     258|    0|
+-----------------+---------+-------+--------+--------+-----+

Any Ideas that you may suggest for reducing the number of BRAM and LUTs by maintaining proper utilization?

Also what's the command for generating a bitstream file for Vivado as the backend, just like if we use vivadoaccelerator as the backend we use hls4ml.templates.VivadoAcceleratorBackend.make_bitfile(hls_model) command to generate the bitstream file, but how to get the bitstream file if we use vivado as the backend is there any command as such, if so please help me in this regard, I am using vivado 2019.2 version. Thank you in Advance.

vloncar · 2023-04-11T13:30:57Z

vloncar
Apr 11, 2023
Maintainer

Unlikely to solve your problem since the activation tensors are simply too large, but you can try:

hls4ml.model.optimizer.get_optimizer('vivadoaccelerator:fifo_depth_optimization').configure(profiling_fifo_depth=100_000) # This will cause even larger use of BRAMs, but is only temporary
hls_config['Flows'] = ['vivadoaccelerator:fifo_depth_optimization']
...
report = model.build(csim=True, synth=True, cosim=True, validation=False, export=True, vsynth=True, reset=True, bitfile=True)

then see what the vivado synthesis report says. This will optimize the sizes of the FIFOs used for activation tensors (which consume the most of BRAMs). You'll likely have to reduce the size of your model (add pooling, get rid of same padding etc)

1 reply

sandeep1404 Apr 11, 2023
Author

Thank you @vloncar for your reply, I will try to synthesize my model again with your suggestions.

sandeep1404 · 2023-04-12T05:21:40Z

sandeep1404
Apr 12, 2023
Author

Hi @vloncar I am using Vivado as the backend and was trying to optimize the FIFO depth to overcome the issue of over utilization. I am using the commands below as you mentioned but I am getting the following key error. I am using hls4ml version 0.6.0

hls_config['Flows'] = ['vivado:fifo_depth_optimization']
hls4ml.model.optimizer.get_optimizer('vivado:fifo_depth_optimization').configure(profiling_fifo_depth=100_000)

even if I use vivadoaccelerator:fifo_depth_optimization instead of vivado:fifo_depth_optimization keeping VivadoAccelrator as backend I am getting the following key error.

KeyError                                  Traceback (most recent call last)
Cell In[37], line 52
     49 print("-----------------------------------")
     51 hls_config['Flows'] = ['vivado:fifo_depth_optimization']
---> 52 hls4ml.model.optimizer.get_optimizer('vivado:fifo_depth_optimization').configure(profiling_fifo_depth=100_000)
     53 # hls4ml.model.optimizer.get_optimizer('vivadoaccelerator:fifo_depth_optimization').configure(profiling_fifo_depth=100_000) # This will cause even larger use of BRAMs, but is only temporary
     54 # hls_config['Flows'] = ['vivadoaccelerator:fifo_depth_optimization']
     57 cfg_q = hls4ml.converters.create_config(backend='Vivado')

File ~/.local/lib/python3.8/site-packages/hls4ml/model/optimizer/optimizer.py:25, in get_optimizer(name)
     24 def get_optimizer(name):
---> 25     return optimizer_map[name]()

KeyError: 'vivado:fifo_depth_optimization'

can you please help me in resolving this issue. Thankyou in advance.

1 reply

vloncar Apr 12, 2023
Maintainer

Use the main branch

sandeep1404 · 2023-04-17T10:08:40Z

sandeep1404
Apr 17, 2023
Author

Hi @vloncar, I tried to use the main branch, and when I tried converting the model it took me 18hrs just to do c-simulation, however, I didn't mention explicitly to do c/RTL simulation while converting the model.

below is the code for that:

 import hls4ml

hls_config = hls4ml.utils.config_from_keras_model(model, granularity='name')
hls_config['Model']['Precision'] = 'ap_fixed<16,6>'
hls_config["Flows"] = ['output_rounding_saturation_mode']
hls4ml.model.optimizer.get_optimizer('output_rounding_saturation_mode').configure(layers=['Activation'])
hls4ml.model.optimizer.get_optimizer('output_rounding_saturation_mode').configure(rounding_mode = 'AP_RND')
hls4ml.model.optimizer.get_optimizer('output_rounding_saturation_mode').configure(saturation_mode = 'AP_SAT')

############################################

# Precision = 'ap_fixed<16,6>'
Reuse_Factor = 1000000

# hls_config['Model']['Precision'] = Precision
hls_config['Model']['ReuseFactor'] = Reuse_Factor
hls_config['IOTtype']= 'io_stream'

for Layer in hls_config['LayerName'].keys():
    hls_config['LayerName'][Layer]['Strategy'] = 'resource'
    
hls_config['Model']['Strategy'] = 'Resource'



hls_config['LayerName']['conv0']['ReuseFactor'] = 432
hls_config['LayerName']['conv1']['ReuseFactor'] = 2304
hls_config['LayerName']['conv2']['ReuseFactor'] = 2304
hls_config['LayerName']['conv3']['ReuseFactor'] = 2304
hls_config['LayerName']['conv4']['ReuseFactor'] = 2304
hls_config['LayerName']['conv5']['ReuseFactor'] = 2304
hls_config['LayerName']['conv6']['ReuseFactor'] = 2304
hls_config['LayerName']['conv7']['ReuseFactor'] = 256
hls_config['LayerName']['conv8']['ReuseFactor'] = 432


print("-----------------------------------")
print("Configuration")
plotting.print_dict(hls_config)
print("-----------------------------------")

hls_config['Flows'] = ['vivado:fifo_depth_optimization']
hls4ml.model.optimizer.get_optimizer('vivado:fifo_depth_optimization').configure(profiling_fifo_depth=100_000)



hls_model = hls4ml.converters.convert_from_keras_model(model, 
                                                       io_type='io_stream',
                                                       hls_config=hls_config,
                                                       output_dir='model_zcu104_fpnrnet/hls4ml_prj_pynq_fpnrnet_32bit_1',
                                                       part='xczu7ev-ffvc1156-2-e',
                                                       backend='Vivado'
                                                       )

after running the above cell it took almost 18hrs to finish the execution of the cell and the output is below:

INFO: Saved inference results to file: tb_data/rtl_cosim_results.log
INFO:
Report time       : Monday 12 April 2021 10:30:31 PM IST.
Solution          : solution1.
Simulation tool   : xsim.

+----------+----------+-----------------------------------------------+-----------------------------------------------+
|          |          |                    Latency                    |                    Interval                   |
+   RTL    +  Status  +-----------------------------------------------+-----------------------------------------------+
|          |          |      min      |      avg      |      max      |      min      |      avg      |      max      |
+----------+----------+-----------------------------------------------+-----------------------------------------------+
|      VHDL|        NA|             NA|             NA|             NA|             NA|             NA|             NA|
|   Verilog|      Pass|             NA|             NA|             NA|             NA|             NA|             NA|
+----------+----------+-----------------------------------------------+-----------------------------------------------+

***** C/RTL SIMULATION COMPLETED IN 18h34m13s *****
INFO: [HLS 200-112] Total elapsed time: 67327.7 seconds; peak allocated memory: 1.208 GB.
INFO: [Common 17-206] Exiting vivado_hls at Thu Apr 13 17:02:52 2023...
Vivado synthesis report not found.
Implementation report not found.
Timing report not found.
[hls4ml] - FIFO optimization completed
Writing HLS project
Done

I don't know why it performed C/RTL simulation while conversion of the model and took too long 18hrs of time.

And one more thing is when I want to generate bit stream file, by the following command using the main branch, I was getting the following error;

hls_model.build(csim=False, synth=True, cosim=False, validation=False, export=True, vsynth=True, reset=False, bitfile=True)

TypeError                                 Traceback (most recent call last)
Cell In[78], line 1
----> 1 hls_model.build(csim=False, synth=True, cosim=False, validation=False, export=True, vsynth=True, reset=False, bitfile=True)

File ~/.local/lib/python3.8/site-packages/hls4ml/model/graph.py:854, in ModelGraph.build(self, **kwargs)
    850 if not os.path.exists(self.config.get_output_dir()):
    851     # Assume the project wasn't written before
    852     self.write()
--> 854 return self.config.backend.build(self, **kwargs)

TypeError: build() got an unexpected keyword argument 'bitfile'

can you please help me in resolving this issue? Thankyou in advance.

0 replies

sandeep1404 · 2023-04-18T10:40:40Z

sandeep1404
Apr 18, 2023
Author

Hi can anyone look into this issue and help me in resolving the issue, I got stuck and have no idea how to proceed. Thank you in advance.

0 replies

vloncar · 2023-04-18T10:53:47Z

vloncar
Apr 18, 2023
Maintainer

there is no bitfile generation in Vivado backend. you need to use the accelerator backend.

0 replies

sandeep1404 · 2023-04-18T11:06:13Z

sandeep1404
Apr 18, 2023
Author

Hi @vloncar Thanks for the reply, I just want to port my model on zcu104, but if I use accelerator as a backend zcu104 doesn't support it since it was not listed in the supported board files, so I am using vivado as a backend and the part number is for zcu104. So is there any way where we can port the models on zcu104 by generating a bit stream file, and why does the model conversion take too much time nearly to 19hrs when I include fifo_depth_optimization any idea about it?

0 replies

vloncar · 2023-04-18T11:21:05Z

vloncar
Apr 18, 2023
Maintainer

There is PR #752 that adds support for it. You can look into those changes and follow them up or just use that branch. As for the fifo depth optimization, look into the logs. First the model is created with large fifo depths, this is then synthesized, co-simulation is ran and the actual occupancy of the fifos is extracted. then the model is re-synthesized with those depths, giving you the final result. this obviously takes more time. the co-simulation can also take a significant amount of time depending on the number of samples in the dataset you provided (if you provided no dataset, it will likely work with a single sample of zeros, producing incorrect fifo depths).

0 replies

sandeep1404 · 2023-04-18T16:08:03Z

sandeep1404
Apr 18, 2023
Author

Hi @vloncar, Thanks for the reply, can I know how to use that branch, do I need to do pip install, if so can you please provide me with the branch link, I am not able to get it. Thanks in advance

1 reply

vloncar Apr 18, 2023
Maintainer

It took you only 16 minutes to give up? 😃 Here. As for the branch name, you can get it from the PR itself, just look under the title.

sandeep1404 · 2023-04-19T07:20:05Z

sandeep1404
Apr 19, 2023
Author

Hi @vloncar Thanks for the reply, Sorry I am really confused on what to do, when I try to generate bit stream using accelerator as the backend using the main branch, I was getting the following attribute error

hls4ml.templates.VivadoAcceleratorBackend.make_bitfile(hls_model)

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[58], line 1
----> 1 hls4ml.templates.VivadoAcceleratorBackend.make_bitfile(hls_model)

AttributeError: module 'hls4ml' has no attribute 'templates'

sorry for troubling you by asking many issues, I tried to figure it out, but I am not sure how to resolve this.

0 replies

Bram and LUTs are overutilized #758

Uh oh!

Uh oh!

sandeep1404 Apr 11, 2023

Replies: 9 comments · 3 replies

Uh oh!

vloncar Apr 11, 2023 Maintainer

Uh oh!

sandeep1404 Apr 11, 2023 Author

Uh oh!

Uh oh!

sandeep1404 Apr 12, 2023 Author

Uh oh!

vloncar Apr 12, 2023 Maintainer

Uh oh!

Uh oh!

sandeep1404 Apr 17, 2023 Author

Uh oh!

sandeep1404 Apr 18, 2023 Author

Uh oh!

vloncar Apr 18, 2023 Maintainer

Uh oh!

sandeep1404 Apr 18, 2023 Author

Uh oh!

vloncar Apr 18, 2023 Maintainer

Uh oh!

Uh oh!

sandeep1404 Apr 18, 2023 Author

Uh oh!

vloncar Apr 18, 2023 Maintainer

Uh oh!

sandeep1404 Apr 19, 2023 Author

sandeep1404
Apr 11, 2023

Replies: 9 comments 3 replies

vloncar
Apr 11, 2023
Maintainer

sandeep1404 Apr 11, 2023
Author

sandeep1404
Apr 12, 2023
Author

vloncar Apr 12, 2023
Maintainer

sandeep1404
Apr 17, 2023
Author

sandeep1404
Apr 18, 2023
Author

vloncar
Apr 18, 2023
Maintainer

sandeep1404
Apr 18, 2023
Author

vloncar
Apr 18, 2023
Maintainer

sandeep1404
Apr 18, 2023
Author

vloncar Apr 18, 2023
Maintainer

sandeep1404
Apr 19, 2023
Author