From 9207607a595fdf6c2e2cb01473b6e08f0746b8d3 Mon Sep 17 00:00:00 2001
From: iovodov <b@ovdv.ru>
Date: Sat, 21 Apr 2018 20:22:16 +0000
Subject: [PATCH] Fix issue #674: parameter "flip" is added to [net] parameters for classifier and detector
---
src/convolutional_layer.c | 156 +++++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 131 insertions(+), 25 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 798d9c1..9a76bdf 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -8,7 +8,9 @@
#include <stdio.h>
#include <time.h>
+#ifdef CUDNN
#pragma comment(lib, "cudnn.lib")
+#endif
#ifdef AI2
#include "xnor_layer.h"
@@ -135,22 +137,71 @@
#ifdef GPU
#ifdef CUDNN
-void cudnn_convolutional_setup(layer *l)
+void cudnn_convolutional_setup(layer *l, int cudnn_preference)
{
- cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
- cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
- cudnnSetFilter4dDescriptor(l->dweightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
- cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
- cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
- cudnnSetFilter4dDescriptor(l->weightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
- cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
- cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
+#ifdef CUDNN_HALF
+ // 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;
+#else
+ cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
+#endif
+
+#if(CUDNN_MAJOR >= 7)
+ // Tensor Core uses CUDNN_TENSOR_OP_MATH instead of CUDNN_DEFAULT_MATH
+ // For *_ALGO_WINOGRAD_NONFUSED can be used CUDNN_DATA_FLOAT
+ // otherwise Input, Filter and Output descriptors (xDesc, yDesc, wDesc, dxDesc, dyDesc and dwDesc as applicable) have dataType = CUDNN_DATA_HALF
+ // Three techniques for training using Mixed-precision: https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/
+ // 1. Accumulation into FP32
+ // 2. Loss Scaling - required only for: activation gradients. We do not use.
+ // 3. FP32 Master Copy of Weights
+ // More: http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#tensor_ops
+ cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH);
+#endif
+
+ // 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;
+
+ // backward delta
+ cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->dweightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+
+ // forward
+ cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->weightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+
+ // batch norm
+ cudnnSetTensor4dDescriptor(l->normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l->out_c, 1, 1);
+ cudnnSetTensor4dDescriptor(l->normDstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
+
+ cudnnSetTensor4dDescriptor(l->normDstTensorDescF16, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+#if(CUDNN_MAJOR >= 6)
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION, CUDNN_DATA_FLOAT); // cudnn >= 6.0
+#else
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION); // cudnn 5.1
+#endif
+ 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)
+ {
+ forward_algo = CUDNN_CONVOLUTION_FWD_NO_WORKSPACE;
+ backward_algo = CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE;
+ backward_filter = CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE;
+ printf(" CUDNN-slow ");
+ }
+
+ cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
l->srcTensorDesc,
l->weightDesc,
l->convDesc,
l->dstTensorDesc,
- CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
+ forward_algo,
0,
&l->fw_algo);
cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
@@ -158,7 +209,7 @@
l->ddstTensorDesc,
l->convDesc,
l->dsrcTensorDesc,
- CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
+ backward_algo,
0,
&l->bd_algo);
cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
@@ -166,9 +217,41 @@
l->ddstTensorDesc,
l->convDesc,
l->dweightDesc,
- CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
+ backward_filter,
0,
&l->bf_algo);
+
+ if (data_type == CUDNN_DATA_HALF)
+ {
+ // HALF-16 if(data_type == CUDNN_DATA_HALF)
+ l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
+ l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
+ l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
+
+ // FLOAT-32 if(data_type == CUDNN_DATA_FLOAT)
+ //l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED;
+ //l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED;
+ //l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED;
+
+ int fw = 0, bd = 0, bf = 0;
+ if (l->fw_algo == CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM) fw = 1;
+ //printf("Tensor Cores - Forward enabled: l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM \n");
+ if (l->fw_algo == CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED) fw = 2;
+ //printf("Tensor Cores - Forward enabled: l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED \n");
+
+ if (l->bd_algo == CUDNN_CONVOLUTION_BWD_DATA_ALGO_1) bd = 1;
+ //printf("Tensor Cores - Backward-data enabled: l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1 \n");
+ if (l->bd_algo == CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED) bd = 2;
+ //printf("Tensor Cores - Backward-data enabled: l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED \n");
+
+ if (l->bf_algo == CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1) bf = 1;
+ //printf("Tensor Cores - Backward-filter enabled: l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1 \n");
+ 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 ");
+ }
}
#endif
#endif
@@ -260,6 +343,10 @@
}
l.weights_gpu = cuda_make_array(l.weights, c*n*size*size);
+#ifdef CUDNN_HALF
+ 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);
l.biases_gpu = cuda_make_array(l.biases, n);
@@ -292,7 +379,10 @@
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);
cudnnCreateTensorDescriptor(&l.srcTensorDesc);
cudnnCreateTensorDescriptor(&l.dstTensorDesc);
cudnnCreateFilterDescriptor(&l.weightDesc);
@@ -300,7 +390,7 @@
cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
cudnnCreateFilterDescriptor(&l.dweightDesc);
cudnnCreateConvolutionDescriptor(&l.convDesc);
- cudnn_convolutional_setup(&l);
+ cudnn_convolutional_setup(&l, cudnn_fastest);
#endif
}
#endif
@@ -353,6 +443,8 @@
void resize_convolutional_layer(convolutional_layer *l, int w, int h)
{
+ int old_w = l->w;
+ int old_h = l->h;
l->w = w;
l->h = h;
int out_w = convolutional_out_width(*l);
@@ -372,24 +464,38 @@
}
#ifdef GPU
- cuda_free(l->delta_gpu);
- cuda_free(l->output_gpu);
+ if (old_w < w || old_h < h) {
+ cuda_free(l->delta_gpu);
+ cuda_free(l->output_gpu);
- l->delta_gpu = cuda_make_array(l->delta, l->batch*l->outputs);
- l->output_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+ l->delta_gpu = cuda_make_array(l->delta, l->batch*l->outputs);
+ l->output_gpu = cuda_make_array(l->output, l->batch*l->outputs);
- if(l->batch_normalize){
- cuda_free(l->x_gpu);
- cuda_free(l->x_norm_gpu);
+ if (l->batch_normalize) {
+ cuda_free(l->x_gpu);
+ cuda_free(l->x_norm_gpu);
- l->x_gpu = cuda_make_array(l->output, l->batch*l->outputs);
- l->x_norm_gpu = cuda_make_array(l->output, l->batch*l->outputs);
- }
+ l->x_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+ l->x_norm_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+ }
+ }
#ifdef CUDNN
- cudnn_convolutional_setup(l);
+ cudnn_convolutional_setup(l, cudnn_fastest);
#endif
#endif
l->workspace_size = get_workspace_size(*l);
+
+#ifdef CUDNN
+ // check for excessive memory consumption
+ size_t free_byte;
+ size_t total_byte;
+ check_error(cudaMemGetInfo(&free_byte, &total_byte));
+ if (l->workspace_size > free_byte || l->workspace_size >= total_byte / 2) {
+ printf(" used slow CUDNN algo without Workspace! \n");
+ cudnn_convolutional_setup(l, cudnn_smallest);
+ l->workspace_size = get_workspace_size(*l);
+ }
+#endif
}
void add_bias(float *output, float *biases, int batch, int n, int size)
--
Gitblit v1.10.0