From ae43c2bc32fbb838bfebeeaf2c2b058ccab5c83c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@burninator.cs.washington.edu>
Date: Thu, 23 Jun 2016 05:31:14 +0000
Subject: [PATCH] hi
---
src/softmax_layer_kernels.cu | 30 ++++++++++++++----------------
1 files changed, 14 insertions(+), 16 deletions(-)
diff --git a/src/softmax_layer_kernels.cu b/src/softmax_layer_kernels.cu
index 61dc607..8feaf89 100644
--- a/src/softmax_layer_kernels.cu
+++ b/src/softmax_layer_kernels.cu
@@ -1,12 +1,14 @@
+#include "cuda_runtime.h"
+#include "curand.h"
+#include "cublas_v2.h"
+
extern "C" {
#include "softmax_layer.h"
#include "cuda.h"
#include "blas.h"
}
-#define BLOCK 256
-
-__global__ void forward_softmax_layer_kernel(int n, int batch, float *input, float *output)
+__global__ void forward_softmax_layer_kernel(int n, int batch, float *input, float temp, float *output)
{
int b = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if(b >= batch) return;
@@ -19,11 +21,11 @@
largest = (val>largest) ? val : largest;
}
for(i = 0; i < n; ++i){
- sum += exp(input[i+b*n]-largest);
+ sum += exp(input[i+b*n]/temp-largest/temp);
}
- sum = (sum != 0) ? largest+log(sum) : largest-100;
+ sum = (sum != 0) ? largest/temp+log(sum) : largest-100;
for(i = 0; i < n; ++i){
- output[i+b*n] = exp(input[i+b*n]-sum);
+ output[i+b*n] = exp(input[i+b*n]/temp-sum);
}
}
@@ -32,21 +34,17 @@
cuda_pull_array(layer.output_gpu, layer.output, layer.inputs*layer.batch);
}
-extern "C" void forward_softmax_layer_gpu(const softmax_layer layer, float *input)
+extern "C" void forward_softmax_layer_gpu(const softmax_layer layer, network_state state)
{
- forward_softmax_layer_kernel<<<cuda_gridsize(layer.batch), BLOCK>>>(layer.inputs, layer.batch, input, layer.output_gpu);
+ int inputs = layer.inputs / layer.groups;
+ int batch = layer.batch * layer.groups;
+ forward_softmax_layer_kernel<<<cuda_gridsize(batch), BLOCK>>>(inputs, batch, state.input, layer.temperature, layer.output_gpu);
check_error(cudaPeekAtLastError());
-
- /*
- cl_read_array(layer.output_cl, layer.output, layer.inputs*layer.batch);
- int z;
- for(z = 0; z < layer.inputs*layer.batch; ++z) printf("%f,",layer.output[z]);
- */
}
-extern "C" void backward_softmax_layer_gpu(const softmax_layer layer, float *delta)
+extern "C" void backward_softmax_layer_gpu(const softmax_layer layer, network_state state)
{
- copy_ongpu(layer.batch*layer.inputs, layer.delta_gpu, 1, delta, 1);
+ axpy_ongpu(layer.batch*layer.inputs, 1, layer.delta_gpu, 1, state.delta, 1);
}
/* This is if you want softmax w/o log-loss classification. You probably don't.
--
Gitblit v1.10.0