This discussion has been locked.

You can no longer post new replies to this discussion. If you have a question you can start a new discussion

Problem with Compiling Pure C with Larger Arrays in Keil (goal: compiling CNN)

YasmineO over 1 year ago

* CONTEXT: Hi! I wrote a CNN inference model (MobileNetV3-Small) in bare metal C and verified its correctness (outputs match PyTorch). I did this on my local machine in Visual Studio Code.
I am trying to simulate this C program on ARM Cortex M4 with FPU, and I am using Keil uVision to compile my C program. I am doing this as part of my thesis project where I will be characterizing the performance/energy before and after adding a specialized custom hardware unit.

I am using my group's RTL simulation infrastructure to run the compiled program (obtained from Keil) on the ARM Cortex M4 core/peripherals and visualize the cycle by cycle info on a waveform viewer for validation and debugging.

* PROBLEM: I am still learning Keil and the proper way to write code for embedded systems, and I've been having a lot of trouble with running even a simple convolution with larger inputs. It seems that whenever the input arrays get even a little large, the SRAM data inputs (as can be seen on the waveform) become undefined/stop updating and the code runs into memory faults. When the program gets stuck in mem faults instructions, it can also be seen in the disassembly. Below, I show the problem with a simple example (simple array eg.) as well as a snippet of the actual code I am aiming to run (1 layer of the network). I would really appreciate any guidance with code guidelines or environment setup I should pay attention to in order to get my target layer working. Thank you so much in advance!

*EXAMPLES:

*Project config, various settings:

*Simple array eg:

The code below works for in and out sizes of less of 30. Once the arrays have size 30, the mem write data becomes undefined.

SRAM_DIN looks good (defined values), with size 5 here.

Breaks with size 30: see the red undefined xxx data getting written in SRAM_DIN.

*Eg target layer from the network: 1 Conv layer:

- Note: header file "bneck_config.h" contains definitions:

float ifmap_buf[3072] = {0, 1, 2,..,3071};

const float conv0_kernels [432] = { 5.5238e-01,2.0333e-01, ..};

#include "bneck_config.h"

// #include <inttypes.h>
#include <stdio.h>
#include <stdlib.h>

//NOTE: To avoid issues with malloc, all buffers to be used were declared as arrays with the max size that will be needed

//ifmap_buf is a float[3072] array in the .h file
float ofmap_buf [3072];  //32x32x3: 3 channels of 32x32 each
float conv_to_sum_buf [90112];

void convolution2D(float* channel_input, int inputSize, 
                    int kernelSize, float* kernel, int stride, 
                    float* channel_output) {
    
    // Calculate the output size
    int padding = (kernelSize - 1) / 2;
    int outputSize = (inputSize + 2 * padding - kernelSize) / stride + 1;

    for (int i = -padding; i < inputSize - padding; i += stride) {
        for (int j = -padding; j < inputSize - padding; j += stride) {

            // Apply convolution at each position
            float sum = 0.0;
            for (int ki = 0; ki < kernelSize; ki++) {
                for (int kj = 0; kj < kernelSize; kj++) {
                    // Check boundaries to handle padding
                    if (i + ki >= 0 && i + ki < inputSize && j + kj >= 0 && j + kj < inputSize) {
                        sum += channel_input[(i + ki)* inputSize + (j + kj)] * kernel[ki * kernelSize + kj];
                    }
                }
            }

            // Store the result in the output channel
            int output_i = (i + padding) / stride;
            int output_j = (j + padding) / stride;
            channel_output[output_i * outputSize + output_j] = sum;
            
        }
    }
}

//Depthwise convolution
//1 kernel with depth = nb_channels
void Depthwise(float* ifmap, int nb_channels, int inputSize, 
                int kernelSize, float* kernels, int stride, 
               float* ofmaps) {
    for (int i = 0; i < nb_channels; i++) {
        //int padding = 1;
        int padding = (kernelSize - 1) / 2;
        int outputSize = (inputSize + 2 * padding - kernelSize) / stride + 1;
        convolution2D(&(ifmap[i * inputSize * inputSize]), inputSize, kernelSize, &(kernels[i*kernelSize * kernelSize]), stride, &(ofmaps[i * outputSize * outputSize]));
    }
}

//Conv2D layer
void Conv2D_layer(float* ifmap, int nb_channels, int inputSize, 
                int kernelSize, float* kernels, int nb_kernels,
                 int stride,  float* ofmap) {
    //nb output channels = nb kernels
    float sum; 
    int padding = (kernelSize - 1) / 2; 
    int dw_size = (inputSize + 2 * padding - kernelSize) / stride + 1; 

    //float* conv_to_sum = malloc(90112);       
            
                                     
    for (int output_channel = 0; output_channel < nb_kernels; output_channel++) {

       Depthwise(ifmap, nb_channels, inputSize, kernelSize,  &(kernels[output_channel*nb_channels*kernelSize*kernelSize]), stride,(float*)conv_to_sum_buf);
       
                for (int y = 0; y < dw_size; y++) { 
            for (int x = 0; x < dw_size; x++) {
                sum = 0;
                for (int channel = 0; channel < nb_channels; channel++){
                    sum = sum + ((float*)conv_to_sum_buf)[channel * dw_size * dw_size  + y * dw_size + x];
                }
                ofmap[output_channel * dw_size * dw_size + y * dw_size + x] = sum;
            }
           
        }
            }
        
    
}


int main (void) {
    

    Conv2D_layer((float*) ifmap_buf, 3, 32, 3, (float*)conv0_kernels, 16, 1, (float*)ofmap_buf); 


    while(1){};

    }