From 160eddddc4e265d5ee59a38797c30720bf46cd7c Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Sun, 27 May 2018 13:53:42 +0000
Subject: [PATCH] Minor fix
---
src/blas_kernels.cu | 226 ++++++++++++++++++++++++++++++++++++++++++++++++++------
1 files changed, 202 insertions(+), 24 deletions(-)
diff --git a/src/blas_kernels.cu b/src/blas_kernels.cu
index ac537d8..1edbbbd 100644
--- a/src/blas_kernels.cu
+++ b/src/blas_kernels.cu
@@ -23,7 +23,7 @@
dim3 dimGrid((size-1)/BLOCK + 1, n, batch);
dim3 dimBlock(BLOCK, 1, 1);
- scale_bias_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size);
+ scale_bias_kernel<<<dimGrid, dimBlock, 0, get_cuda_stream()>>>(output, biases, n, size);
check_error(cudaPeekAtLastError());
}
@@ -67,7 +67,7 @@
dim3 dimGrid((size-1)/BLOCK + 1, n, batch);
dim3 dimBlock(BLOCK, 1, 1);
- add_bias_kernel<<<dimGrid, dimBlock>>>(output, biases, n, size);
+ add_bias_kernel<<<dimGrid, dimBlock, 0, get_cuda_stream()>>>(output, biases, n, size);
check_error(cudaPeekAtLastError());
}
@@ -140,13 +140,42 @@
}
+__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 * sqrtf(1.F-powf(B2, t)) / (1.F-powf(B1, t)) * m[index] / (sqrtf(v[index]) + eps));
+ //if(index == 0) printf("%f %f %f %f\n", m[index], v[index], (rate * sqrtf(1.F-powf(B2, t)) / (1.F-powf(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());
+}
+
+extern "C" void adam_update_gpu(float *w, float *d, float *m, float *v, float B1, float B2, float eps, float decay, float rate, int n, int batch, int t)
+{
+ scal_ongpu(n, B1, m, 1);
+ scal_ongpu(n, B2, v, 1);
+ axpy_ongpu(n, -decay*batch, w, 1, d, 1);
+
+ axpy_ongpu(n, (1 - B1), d, 1, m, 1);
+ mul_ongpu(n, d, 1, d, 1);
+ axpy_ongpu(n, (1 - B2), d, 1, v, 1);
+
+ adam_gpu(n, w, m, v, B1, B2, rate, eps, t);
+ fill_ongpu(n, 0, d, 1);
+}
+
__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;
if (index >= N) return;
int f = (index/spatial)%filters;
- x[index] = (x[index] - mean[f])/(sqrt(variance[f]) + .000001f);
+ x[index] = (x[index] - mean[f])/(sqrtf(variance[f]) + .000001f);
}
__global__ void normalize_delta_kernel(int N, float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta)
@@ -155,7 +184,7 @@
if (index >= N) return;
int f = (index/spatial)%filters;
- delta[index] = delta[index] * 1./(sqrt(variance[f]) + .000001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch);
+ delta[index] = delta[index] * 1.F/(sqrtf(variance[f]) + .000001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch);
}
extern "C" void normalize_delta_gpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta)
@@ -177,7 +206,7 @@
variance_delta[i] += delta[index]*(x[index] - mean[i]);
}
}
- variance_delta[i] *= -.5 * pow(variance[i] + .000001f, (float)(-3./2.));
+ variance_delta[i] *= -.5 * powf(variance[i] + .000001f, (float)(-3./2.));
}
__global__ void accumulate_kernel(float *x, int n, int groups, float *sum)
@@ -208,13 +237,14 @@
local[id] += (i+id < spatial) ? delta[index] : 0;
}
}
+ __syncthreads();
if(id == 0){
mean_delta[filter] = 0;
for(i = 0; i < threads; ++i){
mean_delta[filter] += local[i];
}
- mean_delta[filter] *= (-1./sqrt(variance[filter] + .000001f));
+ mean_delta[filter] *= (-1.F/sqrtf(variance[filter] + .000001f));
}
}
@@ -236,13 +266,14 @@
local[id] += (i+id < spatial) ? delta[index]*(x[index] - mean[filter]) : 0;
}
}
+ __syncthreads();
if(id == 0){
variance_delta[filter] = 0;
for(i = 0; i < threads; ++i){
variance_delta[filter] += local[i];
}
- variance_delta[filter] *= -.5 * pow(variance[filter] + .000001f, (float)(-3./2.));
+ variance_delta[filter] *= -.5 * powf(variance[filter] + .000001f, (float)(-3./2.));
}
}
@@ -259,7 +290,7 @@
mean_delta[i] += delta[index];
}
}
- mean_delta[i] *= (-1./sqrt(variance[i] + .000001f));
+ mean_delta[i] *= (-1.F/sqrtf(variance[i] + .000001f));
}
extern "C" void mean_delta_gpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
@@ -282,7 +313,7 @@
__global__ void mean_kernel(float *x, int batch, int filters, int spatial, float *mean)
{
- float scale = 1./(batch * spatial);
+ float scale = 1.F/(batch * spatial);
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if (i >= filters) return;
int j,k;
@@ -298,7 +329,7 @@
__global__ void variance_kernel(float *x, float *mean, int batch, int filters, int spatial, float *variance)
{
- float scale = 1./(batch * spatial - 1);
+ float scale = 1.F/(batch * spatial - 1);
int j,k;
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
if (i >= filters) return;
@@ -306,12 +337,44 @@
for(j = 0; j < batch; ++j){
for(k = 0; k < spatial; ++k){
int index = j*filters*spatial + i*spatial + k;
- variance[i] += pow((x[index] - mean[i]), 2);
+ variance[i] += powf((x[index] - mean[i]), 2);
}
}
variance[i] *= scale;
}
+__global__ void reorg_kernel(int N, float *x, int w, int h, int c, int batch, int stride, int forward, float *out)
+{
+ int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ if(i >= N) return;
+ int in_index = i;
+ int in_w = i%w;
+ i = i/w;
+ int in_h = i%h;
+ i = i/h;
+ int in_c = i%c;
+ i = i/c;
+ int b = i%batch;
+
+ int out_c = c/(stride*stride);
+
+ int c2 = in_c % out_c;
+ int offset = in_c / out_c;
+ int w2 = in_w*stride + offset % stride;
+ int h2 = in_h*stride + offset / stride;
+ //printf("%d\n", offset);
+ int out_index = w2 + w*stride*(h2 + h*stride*(c2 + out_c*b));
+
+ // printf("%d %d %d\n", w2, h2, c2);
+ //printf("%d %d\n", in_index, out_index);
+ //if(out_index >= N || out_index < 0) printf("bad bad bad \n");
+
+ if(forward) out[out_index] = x[in_index];
+ else out[in_index] = x[out_index];
+ //if(forward) out[1] = x[1];
+ //else out[0] = x[0];
+}
+
__global__ void axpy_kernel(int N, float ALPHA, float *X, int OFFX, int INCX, float *Y, int OFFY, int INCY)
{
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
@@ -321,7 +384,7 @@
__global__ void pow_kernel(int N, float ALPHA, float *X, int INCX, float *Y, int INCY)
{
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
- if(i < N) Y[i*INCY] = pow(X[i*INCX], ALPHA);
+ if(i < N) Y[i*INCY] = powf(X[i*INCX], ALPHA);
}
__global__ void const_kernel(int N, float ALPHA, float *X, int INCX)
@@ -333,7 +396,15 @@
__global__ void constrain_kernel(int N, float ALPHA, float *X, int INCX)
{
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
- if(i < N) X[i*INCX] = min(ALPHA, max(-ALPHA, X[i*INCX]));
+ if(i < N) X[i*INCX] = fminf(ALPHA, fmaxf(-ALPHA, X[i*INCX]));
+}
+
+__global__ void supp_kernel(int N, float ALPHA, float *X, int INCX)
+{
+ int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ if(i < N) {
+ if((X[i*INCX] * X[i*INCX]) < (ALPHA * ALPHA)) X[i*INCX] = 0;
+ }
}
__global__ void scal_kernel(int N, float ALPHA, float *X, int INCX)
@@ -370,7 +441,7 @@
extern "C" void normalize_gpu(float *x, float *mean, float *variance, int batch, int filters, int spatial)
{
size_t N = batch*filters*spatial;
- normalize_kernel<<<cuda_gridsize(N), BLOCK>>>(N, x, mean, variance, batch, filters, spatial);
+ normalize_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, x, mean, variance, batch, filters, spatial);
check_error(cudaPeekAtLastError());
}
@@ -391,6 +462,7 @@
local[id] += (i+id < spatial) ? x[index] : 0;
}
}
+ __syncthreads();
if(id == 0){
mean[filter] = 0;
@@ -416,9 +488,10 @@
for(i = 0; i < spatial; i += threads){
int index = j*spatial*filters + filter*spatial + i + id;
- local[id] += (i+id < spatial) ? pow((x[index] - mean[filter]), 2) : 0;
+ local[id] += (i+id < spatial) ? powf((x[index] - mean[filter]), 2) : 0;
}
}
+ __syncthreads();
if(id == 0){
variance[filter] = 0;
@@ -431,13 +504,13 @@
extern "C" void fast_mean_gpu(float *x, int batch, int filters, int spatial, float *mean)
{
- fast_mean_kernel<<<filters, BLOCK>>>(x, batch, filters, spatial, mean);
+ fast_mean_kernel<<<filters, BLOCK, 0, get_cuda_stream()>>>(x, batch, filters, spatial, mean);
check_error(cudaPeekAtLastError());
}
extern "C" void fast_variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *variance)
{
- fast_variance_kernel<<<filters, BLOCK>>>(x, mean, batch, filters, spatial, variance);
+ fast_variance_kernel<<<filters, BLOCK, 0, get_cuda_stream() >>>(x, mean, batch, filters, spatial, variance);
check_error(cudaPeekAtLastError());
}
@@ -461,13 +534,13 @@
extern "C" void pow_ongpu(int N, float ALPHA, float * X, int INCX, float * Y, int INCY)
{
- pow_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, INCX, Y, INCY);
+ pow_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream() >>>(N, ALPHA, X, INCX, Y, INCY);
check_error(cudaPeekAtLastError());
}
extern "C" void axpy_ongpu_offset(int N, float ALPHA, float * X, int OFFX, int INCX, float * Y, int OFFY, int INCY)
{
- axpy_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, OFFX, INCX, Y, OFFY, INCY);
+ axpy_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, ALPHA, X, OFFX, INCX, Y, OFFY, INCY);
check_error(cudaPeekAtLastError());
}
@@ -484,13 +557,44 @@
extern "C" void copy_ongpu_offset(int N, float * X, int OFFX, int INCX, float * Y, int OFFY, int INCY)
{
- copy_kernel<<<cuda_gridsize(N), BLOCK>>>(N, X, OFFX, INCX, Y, OFFY, INCY);
+ copy_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, X, OFFX, INCX, Y, OFFY, INCY);
+ 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, 0, get_cuda_stream()>>>(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;
+ reorg_kernel<<<cuda_gridsize(size), BLOCK, 0, get_cuda_stream()>>>(size, x, w, h, c, batch, stride, forward, out);
check_error(cudaPeekAtLastError());
}
extern "C" void mask_ongpu(int N, float * X, float mask_num, float * mask)
{
- mask_kernel<<<cuda_gridsize(N), BLOCK>>>(N, X, mask_num, mask);
+ mask_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream() >>>(N, X, mask_num, mask);
check_error(cudaPeekAtLastError());
}
@@ -509,13 +613,19 @@
extern "C" void scal_ongpu(int N, float ALPHA, float * X, int INCX)
{
- scal_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, INCX);
+ scal_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, ALPHA, X, INCX);
+ check_error(cudaPeekAtLastError());
+}
+
+extern "C" void supp_ongpu(int N, float ALPHA, float * X, int INCX)
+{
+ supp_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, INCX);
check_error(cudaPeekAtLastError());
}
extern "C" void fill_ongpu(int N, float ALPHA, float * X, int INCX)
{
- fill_kernel<<<cuda_gridsize(N), BLOCK>>>(N, ALPHA, X, INCX);
+ fill_kernel<<<cuda_gridsize(N), BLOCK, 0, get_cuda_stream()>>>(N, ALPHA, X, INCX);
check_error(cudaPeekAtLastError());
}
@@ -550,7 +660,7 @@
if(sample < 1) sample = 1;
int size = batch * minw * minh * minc;
- shortcut_kernel<<<cuda_gridsize(size), BLOCK>>>(size, minw, minh, minc, stride, sample, batch, w1, h1, c1, add, w2, h2, c2, out);
+ shortcut_kernel<<<cuda_gridsize(size), BLOCK, 0, get_cuda_stream()>>>(size, minw, minh, minc, stride, sample, batch, w1, h1, c1, add, w2, h2, c2, out);
check_error(cudaPeekAtLastError());
}
@@ -594,6 +704,7 @@
}
+
__global__ void weighted_sum_kernel(int n, float *a, float *b, float *s, float *c)
{
int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
@@ -637,3 +748,70 @@
mult_add_into_kernel<<<cuda_gridsize(num), BLOCK>>>(num, a, b, c);
check_error(cudaPeekAtLastError());
}
+
+
+__device__ void softmax_device(int n, float *input, float temp, float *output)
+{
+ int i;
+ float sum = 0;
+ float largest = -INFINITY;
+ for(i = 0; i < n; ++i){
+ int val = input[i];
+ largest = (val>largest) ? val : largest;
+ }
+ for(i = 0; i < n; ++i){
+ float e = exp(input[i]/temp - largest/temp);
+ sum += e;
+ output[i] = e;
+ }
+ for(i = 0; i < n; ++i){
+ output[i] /= sum;
+ }
+}
+
+__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, get_cuda_stream()>>>(inputs, offset, batch, input, temp, output);
+ check_error(cudaPeekAtLastError());
+}
+
+
+__global__ void upsample_kernel(size_t N, float *x, int w, int h, int c, int batch, int stride, int forward, float scale, float *out)
+{
+ size_t i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ if (i >= N) return;
+ int out_index = i;
+ int out_w = i % (w*stride);
+ i = i / (w*stride);
+ int out_h = i % (h*stride);
+ i = i / (h*stride);
+ int out_c = i%c;
+ i = i / c;
+ int b = i%batch;
+
+ int in_w = out_w / stride;
+ int in_h = out_h / stride;
+ int in_c = out_c;
+
+ int in_index = b*w*h*c + in_c*w*h + in_h*w + in_w;
+
+
+ if (forward) out[out_index] += scale * x[in_index];
+ else atomicAdd(x + in_index, scale * out[out_index]);
+}
+
+extern "C" void upsample_gpu(float *in, int w, int h, int c, int batch, int stride, int forward, float scale, float *out)
+{
+ size_t size = w*h*c*batch*stride*stride;
+ upsample_kernel << <cuda_gridsize(size), BLOCK >> >(size, in, w, h, c, batch, stride, forward, scale, out);
+ check_error(cudaPeekAtLastError());
+}
\ No newline at end of file
--
Gitblit v1.10.0