From 880cf187d87c904f5fe574802ecff99118643f2d Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Fri, 09 Mar 2018 16:44:46 +0000
Subject: [PATCH] Fixed multi-GPU training for Tensor Cores
---
src/network.c | 12 ++++++
src/network.h | 4 ++
src/convolutional_layer.c | 4 +-
src/yolo_console_dll.cpp | 4 +
src/convolutional_kernels.cu | 59 +++++++++++++----------------
5 files changed, 48 insertions(+), 35 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 9d88a88..44b9a0f 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -135,26 +135,24 @@
// More: http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#tensor_ops
const size_t input16_size = l.batch*l.c*l.w*l.h;
- static size_t max_input16_size = input16_size;
- static half* input16 = cuda_make_f16_from_f32_array(NULL, max_input16_size);
-
const size_t output16_size = l.batch*l.out_c*l.out_h*l.out_w;
- static size_t max_output16_size = output16_size;
- static half* output16 = cuda_make_f16_from_f32_array(NULL, max_output16_size);
- if (max_input16_size < input16_size) {
- max_input16_size = input16_size;
- cuda_free((float *)input16);
- input16 = cuda_make_f16_from_f32_array(state.input, max_input16_size);
+ if (*state.net.max_input16_size < input16_size) {
+ //printf("\n input16_size: cur = %zu \t max = %zu \n", input16_size, *state.net.max_input16_size);
+ *state.net.max_input16_size = input16_size;
+ if (*state.net.input16_gpu) cuda_free(*state.net.input16_gpu);
+ *state.net.input16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_input16_size);
}
+ float *input16 = *state.net.input16_gpu;
- if (max_output16_size < output16_size) {
- max_output16_size = output16_size;
- cuda_free((float *)output16);
- output16 = cuda_make_f16_from_f32_array(NULL, max_output16_size);
+ if (*state.net.max_output16_size < output16_size) {
+ *state.net.max_output16_size = output16_size;
+ if (*state.net.output16_gpu) cuda_free(*state.net.output16_gpu);
+ *state.net.output16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_output16_size);
}
+ float *output16 = *state.net.output16_gpu;
- cuda_convert_f32_to_f16(state.input, input16_size, (float *)input16);
+ cuda_convert_f32_to_f16(state.input, input16_size, input16);
//fill_ongpu(output16_size / 2, 0, (float *)output16, 1);
cudnnConvolutionForward(cudnn_handle(),
@@ -171,7 +169,7 @@
l.dstTensorDesc,
output16);
- cuda_convert_f16_to_f32((float *)output16, output16_size, l.output_gpu);
+ cuda_convert_f16_to_f32(output16, output16_size, l.output_gpu);
#else
@@ -238,27 +236,24 @@
#ifdef CUDNN_HALF
const size_t input16_size = l.batch*l.c*l.w*l.h;
- static size_t max_input16_size = input16_size;
- static half* input16 = cuda_make_f16_from_f32_array(NULL, max_input16_size);
-
const size_t delta16_size = l.batch*l.n*l.out_w*l.out_h;
- static size_t max_delta16_size = delta16_size;
- static half* delta16 = cuda_make_f16_from_f32_array(NULL, max_delta16_size);
-
- if (max_input16_size < input16_size) {
- max_input16_size = input16_size;
- cuda_free((float *)input16);
- input16 = cuda_make_f16_from_f32_array(state.input, max_input16_size);
+
+ if (*state.net.max_input16_size < input16_size) {
+ *state.net.max_input16_size = input16_size;
+ if(*state.net.input16_gpu) cuda_free(*state.net.input16_gpu);
+ *state.net.input16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_input16_size);
}
+ float *input16 = *state.net.input16_gpu;
- if (max_delta16_size < delta16_size) {
- max_delta16_size = delta16_size;
- cuda_free((float *)delta16);
- delta16 = cuda_make_f16_from_f32_array(NULL, max_delta16_size);
+ if (*state.net.max_output16_size < delta16_size) {
+ *state.net.max_output16_size = delta16_size;
+ if(*state.net.output16_gpu) cuda_free(*state.net.output16_gpu);
+ *state.net.output16_gpu = (float *)cuda_make_f16_from_f32_array(NULL, *state.net.max_output16_size);
}
+ float *delta16 = *state.net.output16_gpu;
- cuda_convert_f32_to_f16(state.input, input16_size, (float *)input16);
- cuda_convert_f32_to_f16(l.delta_gpu, delta16_size, (float *)delta16);
+ cuda_convert_f32_to_f16(state.input, input16_size, input16);
+ cuda_convert_f32_to_f16(l.delta_gpu, delta16_size, delta16);
// convert input: state.input (x), l.delta_gpu (y) from fp32 to fp16
// get output: l.weight_updates_gpu (dw) and convert it to fp32 (ONLY if it is fp16)
@@ -305,7 +300,7 @@
l.dsrcTensorDesc,
input16); // state.delta);
- cuda_convert_f16_to_f32((float *)input16, input16_size, state.delta);
+ cuda_convert_f16_to_f32(input16, input16_size, state.delta);
if (l.binary || l.xnor) swap_binary(&l);
if (l.xnor) gradient_array_ongpu(original_input, l.batch*l.c*l.h*l.w, HARDTAN, state.delta);
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 377b898..7c0c00b 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -305,8 +305,8 @@
l.weights_gpu = cuda_make_array(l.weights, c*n*size*size);
#ifdef CUDNN_HALF
- l.weights_gpu16 = cuda_make_array(l.weights, c*n*size*size / 2);
- l.weight_updates_gpu16 = cuda_make_array(l.weight_updates, c*n*size*size / 2);
+ l.weights_gpu16 = cuda_make_array(NULL, c*n*size*size / 2); //cuda_make_array(l.weights, c*n*size*size / 2);
+ l.weight_updates_gpu16 = cuda_make_array(NULL, c*n*size*size / 2); //cuda_make_array(l.weight_updates, c*n*size*size / 2);
#endif
l.weight_updates_gpu = cuda_make_array(l.weight_updates, c*n*size*size);
diff --git a/src/network.c b/src/network.c
index d23468d..964d3e8 100644
--- a/src/network.c
+++ b/src/network.c
@@ -140,6 +140,11 @@
#ifdef GPU
net.input_gpu = calloc(1, sizeof(float *));
net.truth_gpu = calloc(1, sizeof(float *));
+
+ net.input16_gpu = calloc(1, sizeof(float *));
+ net.output16_gpu = calloc(1, sizeof(float *));
+ net.max_input16_size = calloc(1, sizeof(size_t));
+ net.max_output16_size = calloc(1, sizeof(size_t));
#endif
return net;
}
@@ -622,6 +627,13 @@
if (*net.truth_gpu) cuda_free(*net.truth_gpu);
if (net.input_gpu) free(net.input_gpu);
if (net.truth_gpu) free(net.truth_gpu);
+
+ if (*net.input16_gpu) cuda_free(*net.input16_gpu);
+ if (*net.output16_gpu) cuda_free(*net.output16_gpu);
+ if (net.input16_gpu) free(net.input16_gpu);
+ if (net.output16_gpu) free(net.output16_gpu);
+ if (net.max_input16_size) free(net.max_input16_size);
+ if (net.max_output16_size) free(net.max_output16_size);
#else
free(net.workspace);
#endif
diff --git a/src/network.h b/src/network.h
index 6f4123a..2d28e81 100644
--- a/src/network.h
+++ b/src/network.h
@@ -64,6 +64,10 @@
#ifdef GPU
float **input_gpu;
float **truth_gpu;
+ float **input16_gpu;
+ float **output16_gpu;
+ size_t *max_input16_size;
+ size_t *max_output16_size;
int wait_stream;
#endif
} network;
diff --git a/src/yolo_console_dll.cpp b/src/yolo_console_dll.cpp
index f08d531..4a8310a 100644
--- a/src/yolo_console_dll.cpp
+++ b/src/yolo_console_dll.cpp
@@ -26,17 +26,19 @@
#include "opencv2/videoio/videoio.hpp"
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_MAJOR)""CVAUX_STR(CV_VERSION_MINOR)""CVAUX_STR(CV_VERSION_REVISION)
#pragma comment(lib, "opencv_world" OPENCV_VERSION ".lib")
+#ifdef TRACK_OPTFLOW
#pragma comment(lib, "opencv_cudaoptflow" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_cudaimgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
+#endif // TRACK_OPTFLOW
#else
#define OPENCV_VERSION CVAUX_STR(CV_VERSION_EPOCH)""CVAUX_STR(CV_VERSION_MAJOR)""CVAUX_STR(CV_VERSION_MINOR)
#pragma comment(lib, "opencv_core" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_imgproc" OPENCV_VERSION ".lib")
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
-#endif
+#endif // CV_VERSION_EPOCH
class track_kalman {
public:
--
Gitblit v1.10.0