Dma stuck in .wait() in a design with interleaved hls stream accesses

Support
1

Hi guys,

I have a design where in I am reading from multiple streams in my Vivado HLS code. The dma seems to be stuck in the wait state forever. I am assuming that this is because you cannot have one stream read within the other, is this true? if so, how do I overcome this? I have attached the HLS code below.

#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <hls_math.h>
#include "top_header.hpp"

// --------------------------------------------------------------------
// function to be accelerated in HW wrapped with AXI4-Stream interface
void predict_function(int *length_x, int *tile_size_x, int *remaining_size_x, int *tile_size_sv, int *remaining_size_sv,  AXI_VAL *predict_in_stream, AXI_VAL *supp_vec_stream, AXI_VAL *dual_coef_stream, AXI_VAL *predict_out_stream)
{
#pragma HLS INTERFACE s_axilite port = return bundle = CONTROL_BUS
#pragma HLS INTERFACE s_axilite port = length_x bundle = CONTROL_BUS
#pragma HLS INTERFACE s_axilite port = tile_size_x bundle = CONTROL_BUS
#pragma HLS INTERFACE s_axilite port = remaining_size_x bundle = CONTROL_BUS
#pragma HLS INTERFACE s_axilite port = tile_size_sv bundle = CONTROL_BUS
#pragma HLS INTERFACE s_axilite port = remaining_size_sv bundle = CONTROL_BUS
#pragma HLS INTERFACE axis port = predict_in_stream
#pragma HLS INTERFACE axis port = supp_vec_stream
#pragma HLS INTERFACE axis port = dual_coef_stream
#pragma HLS INTERFACE axis port = predict_out_stream

	// Union used for type conversion
	union
	{
		axi_T packet;
		struct {T f0; } val;
	} 	converter;

    // Hardware buffers
    T input_buf[INPUT_BUFFER_SIZE][FEAT];
    T supp_vecs_buf[INPUT_SVCF_BUFFER_SIZE][FEAT];
    T dual_coef_buf[INPUT_SVCF_BUFFER_SIZE];
    T result_buf[OUTPUT_BUFFER_SIZE];

    // Input and output AXI stream indices
    int is_idx1 = 0;
    int is_idx2 = 0;
    int is_idx3 = 0;
    int l_idx = 0;
    int os_idx = 0;
    T result = 0;

	LOOP_X: for (int l = 0; l < *tile_size_x; l++){
		LOOP_1: for (int m = 0; m < INPUT_BUFFER_SIZE; m++){
			LOOP_2: for (int j = 0; j < FEAT; j++) {
				converter.packet = pop_stream(predict_in_stream[is_idx1++]);
				input_buf[m][j]  = converter.val.f0;
			}

			LOOP_3: for (int k = 0; k < *tile_size_sv; k++){
				LOOP_4: for (int i = 0; i < INPUT_SVCF_BUFFER_SIZE; i++){
					LOOP_5: for (int j = 0; j < FEAT; j++){
						converter.packet = pop_stream(supp_vec_stream[is_idx2++]);
						supp_vecs_buf[i][j] = converter.val.f0;
					}
					converter.packet = pop_stream(dual_coef_stream[is_idx3++]);
					dual_coef_buf[i] = converter.val.f0;
					result += dual_coef_buf[i] * rbf_kernel(supp_vecs_buf[i], input_buf[m]);
				}
			}
			result_buf[l_idx] = result;
			l_idx++;
			result = 0;
		}
	}

    // Stream out value
    LOOP_6: for (int i = 0; i < *length_x; i++){
        converter.val.f0 = result_buf[i];
        predict_out_stream[os_idx++] = push_stream(converter.packet, os_idx == (*length_x));
    }
}

T rbf_kernel(T input1[FEAT], T input2[FEAT]){
    T gamma = 0.05;
    T sum = 0;
    T a[FEAT];

LOOP_12: for (unsigned int i = 0; i < FEAT; i++)
    {
        a[i] = input1[i] - input2[i];
        sum = sum + a[i] * a[i];
    }
    return expf(-gamma * sum);
}

Could it be that the streams predict_in_stream, supp_vec_stream and dual_coef_stream are interfering with each other? is it possible for them to work the way i have written?

2

This isn’t really a PYNQ question. You might be better posting on the Xilinx support forums, or looking for answers there.
https://forums.xilinx.com

Cathal