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 | 68 +++++++++++++++++++++++++++++-----
1 files changed, 58 insertions(+), 10 deletions(-)
diff --git a/src/blas_kernels.cu b/src/blas_kernels.cu
index 59ec005..79fc1c1 100644
--- a/src/blas_kernels.cu
+++ b/src/blas_kernels.cu
@@ -140,6 +140,21 @@
}
+__global__ void adam_kernel(int N, float *x, float *m, float *v, float B1, float B2, float rate, float eps, int t)
+{
+ int index = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ if (index >= N) return;
+
+ x[index] = x[index] - (rate * sqrt(1.-pow(B2, t)) / (1.-pow(B1, t)) * m[index] / (sqrt(v[index]) + eps));
+ //if(index == 0) printf("%f %f %f %f\n", m[index], v[index], (rate * sqrt(1.-pow(B2, t)) / (1.-pow(B1, t)) * m[index] / (sqrt(v[index]) + eps)));
+}
+
+extern "C" void adam_gpu(int n, float *x, float *m, float *v, float B1, float B2, float rate, float eps, int t)
+{
+ adam_kernel<<<cuda_gridsize(n), BLOCK>>>(n, x, m, v, B1, B2, rate, eps, t);
+ check_error(cudaPeekAtLastError());
+}
+
__global__ void normalize_kernel(int N, float *x, float *mean, float *variance, int batch, int filters, int spatial)
{
int index = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
@@ -208,6 +223,7 @@
local[id] += (i+id < spatial) ? delta[index] : 0;
}
}
+ __syncthreads();
if(id == 0){
mean_delta[filter] = 0;
@@ -236,6 +252,7 @@
local[id] += (i+id < spatial) ? delta[index]*(x[index] - mean[filter]) : 0;
}
}
+ __syncthreads();
if(id == 0){
variance_delta[filter] = 0;
@@ -431,6 +448,7 @@
local[id] += (i+id < spatial) ? x[index] : 0;
}
}
+ __syncthreads();
if(id == 0){
mean[filter] = 0;
@@ -459,6 +477,7 @@
local[id] += (i+id < spatial) ? pow((x[index] - mean[filter]), 2) : 0;
}
}
+ __syncthreads();
if(id == 0){
variance[filter] = 0;
@@ -528,6 +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;
@@ -693,31 +736,36 @@
}
-__global__ void softmax_kernel(int n, int batch, float *input, float temp, float *output)
+__device__ void softmax_device(int n, float *input, float temp, float *output)
{
- int b = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
- if(b >= batch) return;
-
int i;
float sum = 0;
float largest = -INFINITY;
for(i = 0; i < n; ++i){
- int val = input[i+b*n];
+ int val = input[i];
largest = (val>largest) ? val : largest;
}
for(i = 0; i < n; ++i){
- sum += exp(input[i+b*n]/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+b*n] = exp(input[i+b*n]/temp-sum);
+ output[i] /= sum;
}
}
-extern "C" void softmax_gpu(float *input, int n, int groups, float temp, float *output, cudaStream_t stream)
+__global__ void softmax_kernel(int n, int offset, int batch, float *input, float temp, float *output)
+{
+ int b = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ if(b >= batch) return;
+ softmax_device(n, input + b*offset, temp, output + b*offset);
+}
+
+extern "C" void softmax_gpu(float *input, int n, int offset, int groups, float temp, float *output)
{
int inputs = n;
int batch = groups;
- softmax_kernel<<<cuda_gridsize(batch), BLOCK, 0, stream>>>(inputs, batch, input, temp, output);
+ softmax_kernel<<<cuda_gridsize(batch), BLOCK>>>(inputs, offset, batch, input, temp, output);
check_error(cudaPeekAtLastError());
}
--
Gitblit v1.10.0