Added yolov3-tiny_xnor.cfg

AlexeyAB

2018-08-23 5b6be00d4b1ffd671c20c4c72d2239c924eaa3d4

 src/convolutional_layer.c

@@ -132,6 +132,7 @@
        return most;
    }
    #endif
    if(l.xnor) return (size_t)l.bit_align*l.size*l.size*l.c * sizeof(float);
    return (size_t)l.out_h*l.out_w*l.size*l.size*l.c*sizeof(float);
}

@@ -305,6 +306,10 @@
    if(xnor){
        l.binary_weights = calloc(c*n*size*size, sizeof(float));
        l.binary_input = calloc(l.inputs*l.batch, sizeof(float));

        int align = 8;
        int src_align = l.out_h*l.out_w;
        l.bit_align = src_align + (align - src_align % align);
    }

    if(batch_normalize){
@@ -621,13 +626,13 @@
    free(align_weights);
}


size_t binary_transpose_align_input(int k, int n, float *b, char **t_bit_input, size_t ldb_align)
// further optimizations: im2col_bin() for XNOR, and then transpose_aling_bin()
size_t binary_transpose_align_input(int k, int n, float *b, char **t_bit_input, size_t ldb_align, int bit_align)
{
    size_t new_ldb = k + (ldb_align - k%ldb_align); // (k / 8 + 1) * 8;
    size_t t_intput_size = new_ldb * n;
    size_t t_bit_input_size = t_intput_size / 8;// +1;
    float *t_input = calloc(t_intput_size, sizeof(float));
    //float *t_input = calloc(t_intput_size, sizeof(float));
    //char *
    *t_bit_input = calloc(t_bit_input_size, sizeof(char));

@@ -637,13 +642,19 @@
    //printf("\n align_weights_size = %d, k = %d, m = %d, lda = %d \n", align_weights_size, k, m, k);
    //printf("\n align_bit_weights_size = %d, k = %d, m = %d, new_lda = %d \n", align_bit_weights_size, k, m, new_ldb);

    int blocksize = 64;
    transpose_block_SSE4x4(b, t_input, k, n, n, new_ldb, blocksize);
    int src_size = k * bit_align;
    //printf("\n src_size = %d \n", src_size);

    //printf("\n blocksize = %d \n", blocksize);
    //float_to_bit(b, t_input, src_size);

    float_to_bit(t_input, *t_bit_input, t_intput_size);
    free(t_input);
    // b - [bit_align, k] - [l.bit_align, l.size*l.size*l.c] = src_size
    // t_input - [bit_align, k] - [n', k]
    // t_bit_input - [new_ldb, n] - [k', n]

    //transpose_bin(t_input, *t_bit_input, k, n, bit_align, new_ldb, 8);
    transpose_bin(b, *t_bit_input, k, n, bit_align, new_ldb, 8);

    //free(t_input);

    return t_intput_size;
}
@@ -688,14 +699,19 @@
        //    t_input = calloc(t_intput_size, sizeof(float));
        //    im2col_cpu_custom_transpose(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, t_input, new_ldb);
        //}
        if (l.xnor && l.size == 3 && l.stride == 1 && l.pad == 1) {}
        else
            im2col_cpu_custom(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, b);
        //if (l.xnor && l.size == 3 && l.stride == 1 && l.pad == 1) {}
        //else
        // further optimizations: im2col_bin() for XNOR, and then transpose_aling_bin()
        //im2col_cpu_custom(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, b);


        //gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
        //gemm_nn_custom(m, n, k, 1, a, k, b, n, c, n);
        if (l.xnor) {
        if (l.xnor && (l.stride == 1 && l.pad == 1)) {
            memset(b, 0, l.bit_align*l.size*l.size*l.c * sizeof(float));
            //im2col_cpu_custom_align(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, b, l.bit_align);
            im2col_cpu_custom_bin(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, b, l.bit_align);

            size_t output_size = l.outputs;
            //float *count_output = calloc(output_size, sizeof(float));
            //size_t bit_output_size = output_size / 8 + 1;
@@ -772,7 +788,7 @@
                }
                */


                /*
                if (l.size == 3 && l.stride == 1 && l.pad == 1)
                {
                    //binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
@@ -783,20 +799,25 @@
                        l.binary_weights, state.input, l.output, l.mean_arr);
                }
                else {
                    */

                    //size_t ldb_align = 256; // 256 bit for AVX2
                    int ldb_align = l.lda_align;
                    size_t new_ldb = k + (ldb_align - k%ldb_align);
                    char *t_bit_input = NULL;
                    size_t t_intput_size = binary_transpose_align_input(k, n, b, &t_bit_input, ldb_align);
                    size_t t_intput_size = binary_transpose_align_input(k, n, b, &t_bit_input, ldb_align, l.bit_align);
                    //char *t_bit_input = calloc(new_ldb * n, sizeof(char));    // for im2col_cpu_custom_transpose() only
                    //float_to_bit(t_input, t_bit_input, new_ldb * n);    // for im2col_cpu_custom_transpose() only

                    // 5x times faster than gemm()-float32
                    // further optimizations: accelerate maxpool-layer with OpenMP/AVX
                    gemm_nn_custom_bin_mean_transposed(m, n, k, 1, l.align_bit_weights, new_ldb, t_bit_input, new_ldb, c, n, l.mean_arr);

                    //gemm_nn_custom_bin_mean_transposed(m, n, k, 1, bit_weights, k, t_bit_input, new_ldb, c, n, mean_arr);

                    //free(t_input);
                    free(t_bit_input);
                }
                //}

            }

@@ -819,6 +840,8 @@
            //free(mean_arr);
        }
        else {
            im2col_cpu_custom(state.input, l.c, l.h, l.w, l.size, l.stride, l.pad, b);

            gemm(0, 0, m, n, k, 1, a, k, b, n, 1, c, n);
            // bit-count to float
        }