From 84d6533cb8112f23a34d3de76435a10f4620f4b8 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Mon, 23 Oct 2017 13:43:03 +0000
Subject: [PATCH] Fixed OpenCV usage in the yolo_console_dll.cpp
---
src/blas_kernels.cu | 35 ++++++++++++++++++++++++++++++++---
1 files changed, 32 insertions(+), 3 deletions(-)
diff --git a/src/blas_kernels.cu b/src/blas_kernels.cu
index 684e66d..79fc1c1 100644
--- a/src/blas_kernels.cu
+++ b/src/blas_kernels.cu
@@ -223,6 +223,7 @@
local[id] += (i+id < spatial) ? delta[index] : 0;
}
}
+ __syncthreads();
if(id == 0){
mean_delta[filter] = 0;
@@ -251,6 +252,7 @@
local[id] += (i+id < spatial) ? delta[index]*(x[index] - mean[filter]) : 0;
}
}
+ __syncthreads();
if(id == 0){
variance_delta[filter] = 0;
@@ -446,6 +448,7 @@
local[id] += (i+id < spatial) ? x[index] : 0;
}
}
+ __syncthreads();
if(id == 0){
mean[filter] = 0;
@@ -474,6 +477,7 @@
local[id] += (i+id < spatial) ? pow((x[index] - mean[filter]), 2) : 0;
}
}
+ __syncthreads();
if(id == 0){
variance[filter] = 0;
@@ -543,6 +547,30 @@
check_error(cudaPeekAtLastError());
}
+__global__ void flatten_kernel(int N, float *x, int spatial, int layers, int batch, int forward, float *out)
+{
+ int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ if(i >= N) return;
+ int in_s = i%spatial;
+ i = i/spatial;
+ int in_c = i%layers;
+ i = i/layers;
+ int b = i;
+
+ int i1 = b*layers*spatial + in_c*spatial + in_s;
+ int i2 = b*layers*spatial + in_s*layers + in_c;
+
+ if (forward) out[i2] = x[i1];
+ else out[i1] = x[i2];
+}
+
+extern "C" void flatten_ongpu(float *x, int spatial, int layers, int batch, int forward, float *out)
+{
+ int size = spatial*batch*layers;
+ flatten_kernel<<<cuda_gridsize(size), BLOCK>>>(size, x, spatial, layers, batch, forward, out);
+ check_error(cudaPeekAtLastError());
+}
+
extern "C" void reorg_ongpu(float *x, int w, int h, int c, int batch, int stride, int forward, float *out)
{
int size = w*h*c*batch;
@@ -718,11 +746,12 @@
largest = (val>largest) ? val : largest;
}
for(i = 0; i < n; ++i){
- sum += exp(input[i]/temp-largest/temp);
+ float e = exp(input[i]/temp - largest/temp);
+ sum += e;
+ output[i] = e;
}
- sum = (sum != 0) ? largest/temp+log(sum) : largest-100;
for(i = 0; i < n; ++i){
- output[i] = exp(input[i]/temp-sum);
+ output[i] /= sum;
}
}
--
Gitblit v1.10.0