Some fixes for AVX support on CPU

AlexeyAB

2018-08-13 f92b20580a21663c5db9eb8608f8cabd7adbeb10

 src/convolutional_layer.c

@@ -44,7 +44,7 @@
        }
        mean = mean / size;
        for(i = 0; i < size; ++i){
            binary[f*size + i] = (weights[f*size + i] > 0) ? mean : -mean;
            binary[f*size + i] = (weights[f*size + i] > 0) ? mean: -mean;
        }
    }
}
@@ -621,7 +621,7 @@
    free(align_weights);
}


// 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)
{
    size_t new_ldb = k + (ldb_align - k%ldb_align); // (k / 8 + 1) * 8;
@@ -688,8 +688,10 @@
        //    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);
        // 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);
@@ -772,18 +774,28 @@
                */

                /*
                if (l.size == 3 && l.stride == 1 && l.pad == 1) {
                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);
                    //printf("\n mean = %f \n", l.mean_arr[0]);

                    convolution_2d(l.w, l.h, l.size, l.n, l.c, l.pad, l.stride,
                        l.weights, state.input, l.output);
                        //l.weights, state.input, l.output, l.mean_arr);
                        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);
                    //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);