From 0a326e7afe3e690c0b4cb64bbd0ce4f0603a7d85 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Tue, 07 Aug 2018 23:45:47 +0000
Subject: [PATCH] XNOR-net on CPU AVX2
---
src/network.c | 34 ++
src/demo.c | 1
src/gemm.h | 50 ++
src/layer.c | 2
src/network.h | 1
src/convolutional_layer.c | 217 ++++++++++++++-
src/detector.c | 2
src/gemm.c | 475 ++++++++++++++++++++++++++++++---
src/convolutional_layer.h | 2
src/layer.h | 3
10 files changed, 710 insertions(+), 77 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 554bd42..bbc4807 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -9,7 +9,7 @@
#include <time.h>
#ifdef CUDNN
-#pragma comment(lib, "cudnn.lib")
+#pragma comment(lib, "cudnn.lib")
#endif
#ifdef AI2
@@ -141,7 +141,7 @@
{
#ifdef CUDNN_HALF
- // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0):
+ // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0):
// Tegra X1, Jetson TX1, DRIVE CX, DRIVE PX, Quadro GP100, Tesla P100
// PSEUDO_HALF_CONFIG is required for Tensor Cores - our case!
const cudnnDataType_t data_type = CUDNN_DATA_HALF;
@@ -161,7 +161,7 @@
cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH);
#endif
- // INT8_CONFIG, INT8_EXT_CONFIG, INT8x4_CONFIG and INT8x4_EXT_CONFIG are only supported
+ // INT8_CONFIG, INT8_EXT_CONFIG, INT8x4_CONFIG and INT8x4_EXT_CONFIG are only supported
// on architectures with DP4A support (compute capability 6.1 and later).
//cudnnDataType_t data_type = CUDNN_DATA_INT8;
@@ -188,7 +188,7 @@
int forward_algo = CUDNN_CONVOLUTION_FWD_PREFER_FASTEST;
int backward_algo = CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST;
int backward_filter = CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST;
- if (cudnn_preference == cudnn_smallest)
+ if (cudnn_preference == cudnn_smallest)
{
forward_algo = CUDNN_CONVOLUTION_FWD_NO_WORKSPACE;
backward_algo = CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE;
@@ -221,7 +221,7 @@
0,
&l->bf_algo);
- if (data_type == CUDNN_DATA_HALF)
+ if (data_type == CUDNN_DATA_HALF)
{
// HALF-16 if(data_type == CUDNN_DATA_HALF)
l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
@@ -249,8 +249,8 @@
if (l->bf_algo == CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED) bf = 2;
//printf("Tensor Cores - Backward-filter enabled: l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED \n");
- if (fw == 2 && bd == 2 && bf == 2) printf("TF ");
- else if (fw == 1 && bd == 1 && bf == 1) printf("TH ");
+ //if (fw == 2 && bd == 2 && bf == 2) printf("TF ");
+ //else if (fw == 1 && bd == 1 && bf == 1) printf("TH ");
}
}
#endif
@@ -379,7 +379,7 @@
l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
}
-#ifdef CUDNN
+#ifdef CUDNN
cudnnCreateTensorDescriptor(&l.normDstTensorDesc);
cudnnCreateTensorDescriptor(&l.normDstTensorDescF16);
cudnnCreateTensorDescriptor(&l.normTensorDesc);
@@ -497,7 +497,7 @@
l->workspace_size = get_workspace_size(*l);
#ifdef CUDNN
- // check for excessive memory consumption
+ // check for excessive memory consumption
size_t free_byte;
size_t total_byte;
check_error(cudaMemGetInfo(&free_byte, &total_byte));
@@ -543,6 +543,85 @@
}
}
+void gemm_nn_custom(int M, int N, int K, float ALPHA,
+ float *A, int lda,
+ float *B, int ldb,
+ float *C, int ldc)
+{
+ int i, j, k;
+ for (i = 0; i < M; ++i) {
+ for (k = 0; k < K; ++k) {
+ register float A_PART = ALPHA*A[i*lda + k];
+ //printf("\n weight = %f \n", A_PART);
+ for (j = 0; j < N; ++j) {
+ C[i*ldc + j] += A_PART*B[k*ldb + j];
+ }
+ }
+ }
+}
+
+
+void get_mean_array(float *src, size_t size, size_t filters, float *mean_arr) {
+ size_t i, counter;
+ counter = 0;
+ for (i = 0; i < size; i += size / filters) {
+ mean_arr[counter++] = fabs(src[i]);
+ }
+}
+
+/*
+void float_to_bit(float *src, unsigned char *dst, size_t size) {
+
+ size_t dst_size = size / 8 + 1;
+ memset(dst, 0, dst_size);
+ size_t i, dst_i, dst_shift;
+ for (i = 0; i < size; ++i) {
+ if (src[i] > 0) set_bit(dst, i);
+ }
+}
+*/
+
+void bit_to_float(unsigned char *src, float *dst, size_t size, size_t filters, float *mean_arr) {
+ memset(dst, 0, size *sizeof(float));
+ size_t i, src_i, src_shift;
+
+ for (i = 0; i < size; ++i) {
+ float mean_val = 1;
+ if(mean_arr != NULL) mean_val = fabs(mean_arr[i / (size / filters)]);
+ if(get_bit(src, i)) dst[i] = mean_val;
+ else dst[i] = -mean_val;
+ }
+}
+
+void binary_transpose_align_weights(convolutional_layer *l, size_t ldb_align)
+{
+ int m = l->n;
+ int k = l->size*l->size*l->c;
+ size_t new_ldb = k + (ldb_align - k%ldb_align); // (k / 8 + 1) * 8;
+
+ binarize_weights(l->weights, m, k, l->binary_weights);
+
+ size_t align_weights_size = new_ldb * m;
+ size_t align_bit_weights_size = align_weights_size / 8;// +1;
+ float *align_weights = calloc(align_weights_size, sizeof(float));
+ l->align_bit_weights = calloc(align_bit_weights_size, sizeof(char));
+
+ size_t i, j;
+ // align A without transpose
+ for (i = 0; i < m; ++i) {
+ for (j = 0; j < k; ++j) {
+ align_weights[i*new_ldb + j] = l->binary_weights[i*k + j];
+ }
+ }
+ float_to_bit(align_weights, l->align_bit_weights, align_weights_size);
+
+ l->mean_arr = calloc(l->n, sizeof(float));
+ get_mean_array(align_weights, align_weights_size, l->n, l->mean_arr);
+
+ free(align_weights);
+}
+
+
void forward_convolutional_layer(convolutional_layer l, network_state state)
{
int out_h = convolutional_out_height(l);
@@ -552,7 +631,10 @@
fill_cpu(l.outputs*l.batch, 0, l.output, 1);
if(l.xnor){
- binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
+ if (!l.align_bit_weights) {
+ binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
+ //printf("\n binarize_weights l.align_bit_weights = %p \n", l.align_bit_weights);
+ }
swap_binary(&l);
binarize_cpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input);
state.input = l.binary_input;
@@ -562,15 +644,122 @@
int k = l.size*l.size*l.c;
int n = out_h*out_w;
-
float *a = l.weights;
float *b = state.workspace;
float *c = l.output;
+ static int u = 0;
+ u++;
+
for(i = 0; i < l.batch; ++i){
- im2col_cpu(state.input, l.c, l.h, l.w,
+ im2col_cpu(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(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) {
+ size_t output_size = l.outputs;
+ //float *count_output = calloc(output_size, sizeof(float));
+ //size_t bit_output_size = output_size / 8 + 1;
+ //char *bit_output = calloc(bit_output_size, sizeof(char));
+
+ size_t intput_size = n * k; // (out_h*out_w) X (l.size*l.size*l.c) : after im2col()
+ size_t bit_input_size = intput_size / 8 + 1;
+ //char *bit_input = calloc(bit_input_size, sizeof(char));
+
+ size_t weights_size = k * m; //l.size*l.size*l.c*l.n;
+ size_t bit_weights_size = weights_size / 8 + 1;
+ //char *bit_weights = calloc(bit_weights_size, sizeof(char));
+ //float *mean_arr = calloc(l.n, sizeof(float));
+
+ // test: float->bit->float
+ //get_mean_array(l.weights, weights_size, l.n, mean_arr);
+ //float_to_bit(l.weights, bit_weights, weights_size);
+ //memset(l.weights, 0, weights_size * sizeof(float));
+ //bit_to_float(bit_weights, l.weights, weights_size, l.n, mean_arr); // just for test float->bit->float
+
+ //float_to_bit(b, bit_input, intput_size);
+ //memset(b, 0, intput_size * sizeof(float));
+ //bit_to_float(bit_input, b, intput_size, 1, NULL); // just for test float->bit->float
+
+ // transpose B from NxK to KxN (x-axis (ldb = l.size*l.size*l.c) - should be multiple of 8 bits)
+ {
+ size_t ldb_align = 256;// 8;
+ 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));
+ char *t_bit_input = calloc(t_bit_input_size, sizeof(char));
+
+ //printf("\n bit_input_size = %d, n = %d, k = %d, ldb = %d \n", bit_input_size, n, k, n);
+ //printf("\n t_bit_input_size = %d, k = %d, n = %d, new_ldb = %d \n", t_bit_input_size, k, n, new_ldb);
+
+
+ //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);
+
+
+ // transpose and align B
+ int i, j;
+ for (i = 0; i < n; ++i) {
+ for (j = 0; j < k; ++j) {
+ t_input[i*new_ldb + j] = b[j*n + i];
+ }
+ }
+ float_to_bit(t_input, t_bit_input, t_intput_size);
+
+ if (!l.align_bit_weights)
+ {
+ size_t align_weights_size = new_ldb * m;
+ size_t align_bit_weights_size = align_weights_size / 8;// +1;
+ float *align_weights = calloc(align_weights_size, sizeof(float));
+ l.align_bit_weights = calloc(align_bit_weights_size, sizeof(char));
+
+ // align A without transpose
+ for (i = 0; i < m; ++i) {
+ for (j = 0; j < k; ++j) {
+ align_weights[i*new_ldb + j] = a[i*k + j];
+ }
+ }
+ float_to_bit(align_weights, l.align_bit_weights, align_weights_size);
+
+ l.mean_arr = calloc(l.n, sizeof(float));
+ get_mean_array(align_weights, align_weights_size, l.n, l.mean_arr);
+
+ free(align_weights);
+ }
+
+ 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);
+
+ //free(align_bit_weights);
+ }
+
+ // for bit_input: (k * n)
+ //if (u == 8) gemm_nn_custom_bin_mean(m, n, k, 1, bit_weights, k, bit_input, n, c, n, mean_arr); // last xnor layer
+ //else gemm_nn_custom_bin_mean(m, n, k, 1, bit_weights, k, bit_input, n, c, n, NULL);
+
+ //gemm_nn_custom_bin_mean(m, n, k, 1, bit_weights, k, bit_input, n, c, n, mean_arr);
+
+ //printf("\n u = %d \n", u);
+
+ //gemm_nn_custom(m, n, k, 1, a, k, b, n, c, n);
+
+ //int j;
+ //if (u != 8) for (j = 0; j < l.n; ++j) l.biases[j] = l.biases[j] / (mean_arr[j]*2);
+
+ //free(count_output);
+ //free(bit_input);
+ //free(bit_weights);
+ //free(mean_arr);
+ }
+ else {
+ gemm(0, 0, m, n, k, 1, a, k, b, n, 1, c, n);
+ // bit-count to float
+ }
c += n*m;
state.input += l.c*l.h*l.w;
}
@@ -606,7 +795,7 @@
float *im = state.input+i*l.c*l.h*l.w;
- im2col_cpu(im, l.c, l.h, l.w,
+ im2col_cpu(im, l.c, l.h, l.w,
l.size, l.stride, l.pad, b);
gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
diff --git a/src/convolutional_layer.h b/src/convolutional_layer.h
index 6d1e517..dd79c48 100644
--- a/src/convolutional_layer.h
+++ b/src/convolutional_layer.h
@@ -35,6 +35,8 @@
void swap_binary(convolutional_layer *l);
void binarize_weights2(float *weights, int n, int size, char *binary, float *scales);
+void binary_transpose_align_weights(convolutional_layer *l, size_t ldb_align);
+
void backward_convolutional_layer(convolutional_layer layer, network_state state);
void add_bias(float *output, float *biases, int batch, int n, int size);
diff --git a/src/demo.c b/src/demo.c
index 1b6b952..4f8c732 100644
--- a/src/demo.c
+++ b/src/demo.c
@@ -146,6 +146,7 @@
}
//set_batch_network(&net, 1);
fuse_conv_batchnorm(net);
+ calculate_binary_weights(net);
srand(2222222);
if(filename){
diff --git a/src/detector.c b/src/detector.c
index 244b4c3..a816c74 100644
--- a/src/detector.c
+++ b/src/detector.c
@@ -568,6 +568,7 @@
}
//set_batch_network(&net, 1);
fuse_conv_batchnorm(net);
+ calculate_binary_weights(net);
srand(time(0));
list *plist = get_paths(valid_images);
@@ -1094,6 +1095,7 @@
}
//set_batch_network(&net, 1);
fuse_conv_batchnorm(net);
+ calculate_binary_weights(net);
if (net.layers[net.n - 1].classes != names_size) {
printf(" Error: in the file %s number of names %d that isn't equal to classes=%d in the file %s \n",
name_list, names_size, net.layers[net.n - 1].classes, cfgfile);
diff --git a/src/gemm.c b/src/gemm.c
index 2b90b05..ee7fa15 100644
--- a/src/gemm.c
+++ b/src/gemm.c
@@ -5,8 +5,8 @@
#include <stdio.h>
#include <math.h>
-void gemm_bin(int M, int N, int K, float ALPHA,
- char *A, int lda,
+void gemm_bin(int M, int N, int K, float ALPHA,
+ char *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
@@ -62,8 +62,8 @@
}
-void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
@@ -71,6 +71,234 @@
gemm_cpu( TA, TB, M, N, K, ALPHA,A,lda, B, ldb,BETA,C,ldc);
}
+
+//--------------------------------------------
+// XNOR bitwise GEMM for binary neural network
+//--------------------------------------------
+
+#include <stdint.h>
+
+static inline unsigned char xnor(unsigned char a, unsigned char b) {
+ //return a == b;
+ return !(a^b);
+}
+
+// INT-32
+static inline uint32_t get_bit_int32(uint32_t const*const src, size_t index) {
+ size_t src_i = index / 32;
+ int src_shift = index % 32;
+ unsigned char val = (src[src_i] & (1 << src_shift)) > 0;
+ return val;
+}
+
+static inline uint32_t xnor_int32(uint32_t a, uint32_t b) {
+ return ~(a^b);
+}
+
+static inline uint64_t xnor_int64(uint64_t a, uint64_t b) {
+ return ~(a^b);
+}
+
+
+static inline uint32_t fill_bit_int32(char src) {
+ if (src == 0) return 0x00000000;
+ else return 0xFFFFFFFF;
+}
+
+static inline uint64_t fill_bit_int64(char src) {
+ if (src == 0) return 0x0000000000000000;
+ else return 0xFFFFFFFFFFFFFFFF;
+}
+
+void binary_int32_printf(uint32_t src) {
+ int i;
+ for (i = 0; i < 32; ++i) {
+ if (src & 1) printf("1");
+ else printf("0");
+ src = src >> 1;
+ }
+ printf("\n");
+}
+
+void binary_int64_printf(uint64_t src) {
+ int i;
+ for (i = 0; i < 64; ++i) {
+ if (src & 1) printf("1");
+ else printf("0");
+ src = src >> 1;
+ }
+ printf("\n");
+}
+
+/*
+void gemm_nn_custom_bin_mean(int M, int N, int K, float ALPHA_UNUSED,
+ unsigned char *A, int lda,
+ unsigned char *B, int ldb,
+ float *C, int ldc, float *mean_arr)
+{
+ int *count_arr = calloc(M*N, sizeof(int));
+
+ int i, j, k;
+ for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
+ for (k = 0; k < K; ++k) { // l.size*l.size*l.c - one filter size [27 - 9216]
+ char a_bit = get_bit(A, i*lda + k);
+
+ for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
+ char b_bit = get_bit(B, k*ldb + j);
+ count_arr[i*ldc + j] += xnor(a_bit, b_bit);
+ }
+ }
+ }
+
+ for (i = 0; i < M; ++i) {
+ float mean_val = mean_arr[i];
+ for (j = 0; j < N; ++j) {
+ C[i*ldc + j] = (2 * count_arr[i*ldc + j] - K) * mean_val;
+ }
+ }
+ free(count_arr);
+}
+*/
+
+/*
+void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
+ unsigned char *A, int lda,
+ unsigned char *B, int ldb,
+ float *C, int ldc, float *mean_arr)
+{
+ int *count_arr = calloc(M*N, sizeof(int));
+
+ int i, j, k;
+ for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
+ for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
+ for (k = 0; k < K; ++k) { // l.size*l.size*l.c - one filter size [27 - 9216]
+ char a_bit = get_bit(A, i*lda + k);
+ char b_bit = get_bit(B, j*ldb + k);
+ count_arr[i*ldc + j] += xnor(a_bit, b_bit);
+ }
+ }
+ }
+
+ for (i = 0; i < M; ++i) {
+ float mean_val = mean_arr[i];
+ for (j = 0; j < N; ++j) {
+ C[i*ldc + j] = (2 * count_arr[i*ldc + j] - K) * mean_val;
+ }
+ }
+ free(count_arr);
+}
+*/
+
+/*
+void gemm_nn_custom_bin_mean(int M, int N, int K, float ALPHA_UNUSED,
+ unsigned char *A, int lda,
+ unsigned char *B, int ldb,
+ float *C, int ldc, float *mean_arr)
+{
+ int *count_arr = calloc(M*N, sizeof(int));
+
+ int i, j, k, h;
+
+#pragma omp parallel for
+ for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
+ for (k = 0; k < K; ++k) { // l.size*l.size*l.c - one filter size [27 - 9216]
+ const char a_bit = get_bit(A, i*lda + k);
+ uint64_t a_bit64 = fill_bit_int64(a_bit);
+ int k_ldb = k*ldb;
+
+ for (j = 0; j < N; j += 64) { // out_h*out_w - one channel output size [169 - 173056]
+ if ((N - j > 64) && (k_ldb % 8 == 0)) {
+ uint64_t b_bit64 = *((uint64_t *)(B + (k_ldb + j) / 8));
+ uint64_t c_bit64 = xnor_int64(a_bit64, b_bit64);
+ //printf("\n %d \n",__builtin_popcountll(c_bit64)); // gcc
+ printf("\n %d \n", __popcnt64(c_bit64)); // msvs
+
+ int h;
+ for (h = 0; h < 64; ++h)
+ if ((c_bit64 >> h) & 1) count_arr[i*ldc + j + h] += 1;
+
+ //binary_int64_printf(a_bit64);
+ //binary_int64_printf(b_bit64);
+ //binary_int64_printf(c_bit64);
+ }
+ else {
+ for (; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
+ char b_bit = get_bit(B, k_ldb + j);
+ if (xnor(a_bit, b_bit)) count_arr[i*ldc + j] += 1;
+ }
+ }
+
+ }
+ }
+ }
+
+ if (mean_arr) {
+ //int K_2 = K / 2;
+ for (i = 0; i < M; ++i) {
+ float mean_val = mean_arr[i];
+ //float mean_val2 = 2 * mean_val;
+ for (j = 0; j < N; ++j) {
+ C[i*ldc + j] = (2 * count_arr[i*ldc + j] - K) * mean_val;
+ //C[i*ldc + j] = (count_arr[i*ldc + j] - K_2) *mean_val2;
+ }
+ }
+ }
+ else {
+ for (i = 0; i < M; ++i) {
+ for (j = 0; j < N; ++j) {
+ C[i*ldc + j] = count_arr[i*ldc + j] - K / 2;
+ }
+ }
+ }
+
+ free(count_arr);
+
+ //getchar();
+}
+*/
+
+
+/*
+void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
+ unsigned char *A, int lda,
+ unsigned char *B, int ldb,
+ float *C, int ldc, float *mean_arr)
+{
+ int i, j, k, h;
+
+#pragma omp parallel for
+ for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
+ float mean_val = mean_arr[i];
+
+ for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
+ int count = 0;
+
+ for (k = 0; k < K; k += 64) { // l.size*l.size*l.c - one filter size [27 - 9216]
+ uint64_t a_bit64 = *((uint64_t *)(A + (i*lda + k) / 8));
+ uint64_t b_bit64 = *((uint64_t *)(B + (j*ldb + k) / 8));
+ uint64_t c_bit64 = xnor_int64(a_bit64, b_bit64);
+
+#ifdef WIN32
+ int tmp_count = __popcnt64(c_bit64);
+#else
+ int tmp_count = __builtin_popcountll(c_bit64);
+#endif
+
+ if (K - k < 64) tmp_count = tmp_count - (64 - (K - k)); // remove extra bits
+ count += tmp_count;
+ //binary_int64_printf(c_bit64);
+ //printf(", count = %d \n\n", tmp_count);
+ }
+
+ C[i*ldc + j] = (2 * count - K) * mean_val;
+ }
+ }
+}
+*/
+
+//----------------------------
+
+
#if (defined(__AVX__) && defined(__x86_64__)) || defined(_WIN64)
#define OSXSAVEFlag (1UL<<27)
@@ -79,8 +307,6 @@
#define CLMULFlag ((1UL<< 1)|AVXFlag|OSXSAVEFlag)
#define VAESFlag ((1UL<<25)|AVXFlag|OSXSAVEFlag)
-#include <stdint.h>
-
#ifdef _WIN64
#include <intrin.h>
#include <ammintrin.h>
@@ -196,6 +422,97 @@
}
}
}
+
+
+// http://graphics.stanford.edu/~seander/bithacks.html
+// https://stackoverflow.com/questions/17354971/fast-counting-the-number-of-set-bits-in-m128i-register
+
+// 2 x faster than popcnt: https://arxiv.org/pdf/1611.07612.pdf
+
+static inline int popcnt128(__m128i n) {
+ const __m128i n_hi = _mm_unpackhi_epi64(n, n);
+#ifdef _MSC_VER
+ return __popcnt64(_mm_cvtsi128_si64(n)) + __popcnt64(_mm_cvtsi128_si64(n_hi));
+#else
+ return __popcntq(_mm_cvtsi128_si64(n)) + __popcntq(_mm_cvtsi128_si64(n_hi));
+#endif
+}
+
+static inline int popcnt256(__m256i n) {
+ return popcnt128(_mm256_extractf128_si256(n, 0)) + popcnt128(_mm256_extractf128_si256(n, 1));
+}
+
+void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
+ unsigned char *A, int lda,
+ unsigned char *B, int ldb,
+ float *C, int ldc, float *mean_arr)
+{
+ __m256i all_1 = _mm256_set1_epi8(255);
+ int i, j, k, h;
+
+ #pragma omp parallel for
+ for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
+ float mean_val = mean_arr[i];
+
+ for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
+ int count = 0;
+ const int bit_step = 256;
+
+ for (k = 0; k < K; k += bit_step) { // l.size*l.size*l.c - one filter size [27 - 9216]
+
+ //__m128i a_bit128 = _mm_loadu_si128((__m128i *)(A + (i*lda + k) / 8));
+ //__m128i b_bit128 = _mm_loadu_si128((__m128i *)(B + (j*ldb + k) / 8));
+ //__m128i xor128 = _mm_xor_si128(a_bit128, b_bit128);
+ //__m128i c_bit128 = _mm_andnot_si128(xor128, all_1);
+ //int tmp_count = popcnt128(c_bit128);
+
+ __m256i a_bit256 = _mm256_loadu_si256((__m256i *)(A + (i*lda + k) / 8));
+ __m256i b_bit256 = _mm256_loadu_si256((__m256i *)(B + (j*ldb + k) / 8));
+ __m256i xor256 = _mm256_xor_si256(a_bit256, b_bit256);
+ __m256i c_bit256 = _mm256_andnot_si256(xor256, all_1); //we can do NOT for wegihts once and do not do this NOT
+ int tmp_count = popcnt256(c_bit256);
+
+ if (K - k < bit_step) tmp_count = tmp_count - (bit_step - (K - k)); // remove extra bits
+ count += tmp_count;
+ //binary_int64_printf(c_bit64);
+ //printf(", count = %d \n\n", tmp_count);
+ }
+
+ C[i*ldc + j] = (2 * count - K) * mean_val;
+ }
+ }
+}
+
+
+void float_to_bit(float *src, unsigned char *dst, size_t size)
+{
+ size_t dst_size = size / 8 + 1;
+ memset(dst, 0, dst_size);
+
+ size_t i;
+ __m128i all128_0 = _mm_set_epi32(0, 0, 0, 0);
+ __m256 all256_0 = _mm256_set1_ps(0);
+ __m256i bits_asc = _mm256_set_epi32(1, 2, 4, 8, 16, 32, 64, 128);
+ //for(i = 0; i < 8; ++i) bits_asc.m256i_i32[i] = 1 << i;
+
+ for (i = 0; i < size; i+=8)
+ {
+ __m256 src256 = _mm256_loadu_ps((__m256i *)(&src[i])); // load 256 bits
+ __m256 result256 = _mm256_cmp_ps(src256, all256_0, _CMP_GT_OS); // compare dst[i] = (float[i] > 0)
+
+ __m256i bits256 = _mm256_castps_si256(result256); // floats to ints32
+ __m256i and256 = _mm256_and_si256(bits256, bits_asc); // bitwise and
+
+ // sum all elements from single and256
+ __m128i tmp128 = _mm_hadd_epi32(_mm256_extractf128_si256(and256, 0), _mm256_extractf128_si256(and256, 1));
+ tmp128 = _mm_hadd_epi32(tmp128, all128_0);
+ tmp128 = _mm_hadd_epi32(tmp128, all128_0);
+
+ dst[i / 8] = tmp128.m128i_i32[0];
+ }
+ // int _mm256_movemask_epi8 (__m256i a)
+}
+
#else
void gemm_nn(int M, int N, int K, float ALPHA,
@@ -213,10 +530,76 @@
}
}
}
+
+void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
+ unsigned char *A, int lda,
+ unsigned char *B, int ldb,
+ float *C, int ldc, float *mean_arr)
+{
+ int i, j, k, h;
+
+#pragma omp parallel for
+ for (i = 0; i < M; ++i) { // l.n - filters [16 - 55 - 1024]
+ float mean_val = mean_arr[i];
+
+ for (j = 0; j < N; ++j) { // out_h*out_w - one channel output size [169 - 173056]
+ int count = 0;
+
+ for (k = 0; k < K; k += 64) { // l.size*l.size*l.c - one filter size [27 - 9216]
+ uint64_t a_bit64 = *((uint64_t *)(A + (i*lda + k) / 8));
+ uint64_t b_bit64 = *((uint64_t *)(B + (j*ldb + k) / 8));
+ uint64_t c_bit64 = xnor_int64(a_bit64, b_bit64);
+
+#ifdef WIN32
+ int tmp_count = __popcnt64(c_bit64);
+#else
+ int tmp_count = __builtin_popcountll(c_bit64);
+#endif
+
+ if (K - k < 64) tmp_count = tmp_count - (64 - (K - k)); // remove extra bits
+ count += tmp_count;
+ //binary_int64_printf(c_bit64);
+ //printf(", count = %d \n\n", tmp_count);
+ }
+
+ C[i*ldc + j] = (2 * count - K) * mean_val;
+ }
+ }
+}
+
+void float_to_bit(float *src, unsigned char *dst, size_t size)
+{
+ size_t dst_size = size / 8 + 1;
+ memset(dst, 0, dst_size);
+
+ size_t i;
+ char *byte_arr = calloc(size, sizeof(char));
+ for (i = 0; i < size; ++i) {
+ if (src[i] > 0) byte_arr[i] = 1;
+ }
+
+ //for (i = 0; i < size; ++i) {
+ // dst[i / 8] |= byte_arr[i] << (i % 8);
+ //}
+
+ for (i = 0; i < size; i += 8) {
+ char dst_tmp = 0;
+ dst_tmp |= byte_arr[i + 0] << 0;
+ dst_tmp |= byte_arr[i + 1] << 1;
+ dst_tmp |= byte_arr[i + 2] << 2;
+ dst_tmp |= byte_arr[i + 3] << 3;
+ dst_tmp |= byte_arr[i + 4] << 4;
+ dst_tmp |= byte_arr[i + 5] << 5;
+ dst_tmp |= byte_arr[i + 6] << 6;
+ dst_tmp |= byte_arr[i + 7] << 7;
+ dst[i / 8] = dst_tmp;
+ }
+ free(byte_arr);
+}
#endif // __x86_64
-void gemm_nt(int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm_nt(int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
@@ -232,8 +615,8 @@
}
}
-void gemm_tn(int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm_tn(int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
@@ -248,8 +631,8 @@
}
}
-void gemm_tt(int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm_tt(int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float *C, int ldc)
{
@@ -266,8 +649,8 @@
}
-void gemm_cpu(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm_cpu(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
@@ -300,21 +683,21 @@
#include <math.h>
-void gemm_ongpu(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A_gpu, int lda,
+void gemm_ongpu(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A_gpu, int lda,
float *B_gpu, int ldb,
float BETA,
float *C_gpu, int ldc)
{
cublasHandle_t handle = blas_handle();
cudaError_t stream_status = cublasSetStream(handle, get_cuda_stream());
- cudaError_t status = cublasSgemm(handle, (TB ? CUBLAS_OP_T : CUBLAS_OP_N),
+ cudaError_t status = cublasSgemm(handle, (TB ? CUBLAS_OP_T : CUBLAS_OP_N),
(TA ? CUBLAS_OP_T : CUBLAS_OP_N), N, M, K, &ALPHA, B_gpu, ldb, A_gpu, lda, &BETA, C_gpu, ldc);
check_error(status);
}
-void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
@@ -435,38 +818,38 @@
int test_gpu_blas()
{
/*
- test_gpu_accuracy(0,0,10,576,75);
+ test_gpu_accuracy(0,0,10,576,75);
- test_gpu_accuracy(0,0,17,10,10);
- test_gpu_accuracy(1,0,17,10,10);
- test_gpu_accuracy(0,1,17,10,10);
- test_gpu_accuracy(1,1,17,10,10);
+ test_gpu_accuracy(0,0,17,10,10);
+ test_gpu_accuracy(1,0,17,10,10);
+ test_gpu_accuracy(0,1,17,10,10);
+ test_gpu_accuracy(1,1,17,10,10);
- test_gpu_accuracy(0,0,1000,10,100);
- test_gpu_accuracy(1,0,1000,10,100);
- test_gpu_accuracy(0,1,1000,10,100);
- test_gpu_accuracy(1,1,1000,10,100);
+ test_gpu_accuracy(0,0,1000,10,100);
+ test_gpu_accuracy(1,0,1000,10,100);
+ test_gpu_accuracy(0,1,1000,10,100);
+ test_gpu_accuracy(1,1,1000,10,100);
- test_gpu_accuracy(0,0,10,10,10);
+ test_gpu_accuracy(0,0,10,10,10);
- time_ongpu(0,0,64,2916,363);
- time_ongpu(0,0,64,2916,363);
- time_ongpu(0,0,64,2916,363);
- time_ongpu(0,0,192,729,1600);
- time_ongpu(0,0,384,196,1728);
- time_ongpu(0,0,256,196,3456);
- time_ongpu(0,0,256,196,2304);
- time_ongpu(0,0,128,4096,12544);
- time_ongpu(0,0,128,4096,4096);
+ time_ongpu(0,0,64,2916,363);
+ time_ongpu(0,0,64,2916,363);
+ time_ongpu(0,0,64,2916,363);
+ time_ongpu(0,0,192,729,1600);
+ time_ongpu(0,0,384,196,1728);
+ time_ongpu(0,0,256,196,3456);
+ time_ongpu(0,0,256,196,2304);
+ time_ongpu(0,0,128,4096,12544);
+ time_ongpu(0,0,128,4096,4096);
*/
- time_ongpu(0,0,64,75,12544);
- time_ongpu(0,0,64,75,12544);
- time_ongpu(0,0,64,75,12544);
- time_ongpu(0,0,64,576,12544);
- time_ongpu(0,0,256,2304,784);
- time_ongpu(1,1,2304,256,784);
- time_ongpu(0,0,512,4608,196);
- time_ongpu(1,1,4608,512,196);
+ time_ongpu(0,0,64,75,12544);
+ time_ongpu(0,0,64,75,12544);
+ time_ongpu(0,0,64,75,12544);
+ time_ongpu(0,0,64,576,12544);
+ time_ongpu(0,0,256,2304,784);
+ time_ongpu(1,1,2304,256,784);
+ time_ongpu(0,0,512,4608,196);
+ time_ongpu(1,1,4608,512,196);
return 0;
}
diff --git a/src/gemm.h b/src/gemm.h
index f0231bf..8664d8e 100644
--- a/src/gemm.h
+++ b/src/gemm.h
@@ -1,32 +1,60 @@
#ifndef GEMM_H
#define GEMM_H
-void gemm_bin(int M, int N, int K, float ALPHA,
- char *A, int lda,
+static inline void set_bit(unsigned char *const dst, size_t index) {
+ size_t dst_i = index / 8;
+ int dst_shift = index % 8;
+ dst[dst_i] |= 1 << dst_shift;
+}
+
+static inline unsigned char get_bit(unsigned char const*const src, size_t index) {
+ size_t src_i = index / 8;
+ int src_shift = index % 8;
+ unsigned char val = (src[src_i] & (1 << src_shift)) > 0;
+ return val;
+}
+
+void float_to_bit(float *src, unsigned char *dst, size_t size);
+
+void gemm_nn_custom_bin_mean_transposed(int M, int N, int K, float ALPHA_UNUSED,
+ unsigned char *A, int lda,
+ unsigned char *B, int ldb,
+ float *C, int ldc, float *mean_arr);
+
+
+//void gemm_nn_custom_bin_mean(int M, int N, int K, float ALPHA_UNUSED,
+ //unsigned char *A, int lda,
+ //unsigned char *B, int ldb,
+ //float *C, int ldc, float *mean_arr)
+
+
+
+void gemm_bin(int M, int N, int K, float ALPHA,
+ char *A, int lda,
float *B, int ldb,
float *C, int ldc);
-
-void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A, int lda,
+
+void gemm(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc);
-void gemm_cpu(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm_cpu(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc);
#ifdef GPU
-void gemm_ongpu(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A_gpu, int lda,
+void gemm_ongpu(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A_gpu, int lda,
float *B_gpu, int ldb,
float BETA,
float *C_gpu, int ldc);
-void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
- float *A, int lda,
+void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
+ float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc);
diff --git a/src/layer.c b/src/layer.c
index 3b47917..caa9c4c 100644
--- a/src/layer.c
+++ b/src/layer.c
@@ -33,6 +33,8 @@
if (l.scale_updates) free(l.scale_updates);
if (l.weights) free(l.weights);
if (l.weight_updates) free(l.weight_updates);
+ if (l.weights) free(l.align_bit_weights);
+ if (l.weights) free(l.mean_arr);
if (l.delta) free(l.delta);
if (l.output) free(l.output);
if (l.squared) free(l.squared);
diff --git a/src/layer.h b/src/layer.h
index 8a58c92..224f77a 100644
--- a/src/layer.h
+++ b/src/layer.h
@@ -179,6 +179,9 @@
float *weights;
float *weight_updates;
+ char *align_bit_weights;
+ float *mean_arr;
+
float *col_image;
int * input_layers;
int * input_sizes;
diff --git a/src/network.c b/src/network.c
index 3df837d..63b76a8 100644
--- a/src/network.c
+++ b/src/network.c
@@ -222,7 +222,7 @@
{
#ifdef GPU
if (gpu_index >= 0) return get_network_output_gpu(net);
-#endif
+#endif
int i;
for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
return net.layers[i].output;
@@ -366,7 +366,7 @@
/*
layer *l = net->layers + i;
cudnn_convolutional_setup(l, cudnn_fastest);
- // check for excessive memory consumption
+ // check for excessive memory consumption
size_t free_byte;
size_t total_byte;
check_error(cudaMemGetInfo(&free_byte, &total_byte));
@@ -520,7 +520,7 @@
if(l.type == CONVOLUTIONAL){
prev = visualize_convolutional_layer(l, buff, prev);
}
- }
+ }
}
void top_predictions(network net, int k, int *index)
@@ -684,7 +684,7 @@
}
}
free(X);
- return pred;
+ return pred;
}
matrix network_predict_data(network net, data test)
@@ -707,7 +707,7 @@
}
}
free(X);
- return pred;
+ return pred;
}
void print_network(network net)
@@ -749,7 +749,7 @@
printf("%5d %5d\n%5d %5d\n", a, b, c, d);
float num = pow((abs(b - c) - 1.), 2.);
float den = b + c;
- printf("%f\n", num/den);
+ printf("%f\n", num/den);
}
float network_accuracy(network net, data d)
@@ -847,3 +847,25 @@
}
}
}
+
+
+
+void calculate_binary_weights(network net)
+{
+ int j;
+ for (j = 0; j < net.n; ++j) {
+ layer *l = &net.layers[j];
+
+ if (l->type == CONVOLUTIONAL) {
+ //printf(" Merges Convolutional-%d and batch_norm \n", j);
+
+ if (l->xnor) {
+ //printf("\n %d \n", j);
+ size_t ldb_align = 256; // 256bit for AVX2
+ binary_transpose_align_weights(l, ldb_align);
+ }
+ }
+ }
+ //printf("\n calculate_binary_weights Done! \n");
+
+}
\ No newline at end of file
diff --git a/src/network.h b/src/network.h
index f80f58d..3b07dbb 100644
--- a/src/network.h
+++ b/src/network.h
@@ -151,6 +151,7 @@
int get_network_nuisance(network net);
int get_network_background(network net);
YOLODLL_API void fuse_conv_batchnorm(network net);
+YOLODLL_API void calculate_binary_weights(network net);
#ifdef __cplusplus
}
--
Gitblit v1.10.0