From d0b9326a352ed2fbc3ae66fdef40b4533a2f211d Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 11 Aug 2015 06:22:27 +0000
Subject: [PATCH] Hacks to get nightmare to not break gridsizing
---
src/softmax_layer_kernels.cu | 18 ++++++------------
1 files changed, 6 insertions(+), 12 deletions(-)
diff --git a/src/softmax_layer_kernels.cu b/src/softmax_layer_kernels.cu
index 61dc607..7e13387 100644
--- a/src/softmax_layer_kernels.cu
+++ b/src/softmax_layer_kernels.cu
@@ -4,8 +4,6 @@
#include "blas.h"
}
-#define BLOCK 256
-
__global__ void forward_softmax_layer_kernel(int n, int batch, float *input, float *output)
{
int b = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
@@ -32,21 +30,17 @@
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.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