You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
psblas3/cuda/spgpu/kernels/axy_base.cuh

211 lines
5.2 KiB
Plaintext

/*
* spGPU - Sparse matrices on GPU library.
*
* Copyright (C) 2010 - 2014
* Salvatore Filippone - University of Rome Tor Vergata
* Davide Barbieri - University of Rome Tor Vergata
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* version 3 as published by the Free Software Foundation.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*/
#define PRE_CONCAT(A, B) A ## B
#define CONCAT(A, B) PRE_CONCAT(A, B)
#undef GEN_SPGPU_FUNC_NAME
#define GEN_SPGPU_FUNC_NAME(x) CONCAT(CONCAT(spgpu,x),axy)
#define GEN_SPGPU_MFUNC_NAME(x) CONCAT(CONCAT(spgpu,x),maxy)
#define GEN_SPGPU_FUNC_NAME_2(x) CONCAT(CONCAT(spgpu,x),axypbz)
#define GEN_SPGPU_MFUNC_NAME_2(x) CONCAT(CONCAT(spgpu,x),maxypbz)
#define GEN_SPGPU_SCAL_NAME(x) CONCAT(CONCAT(spgpu,x),scal)
#define BLOCK_SIZE 256
// Define:
//#define VALUE_TYPE
//#define TYPE_SYMBOL
#include "mathbase.cuh"
__global__ void
CONCAT(GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL),_kern)
(VALUE_TYPE *z, int n, VALUE_TYPE alpha, VALUE_TYPE* x, VALUE_TYPE* y)
{
int id = threadIdx.x + BLOCK_SIZE*blockIdx.x;
if (id < n)
{
// Since z, x and y are accessed with the same offset by the same thread,
// and the write to z follows the x and y reads, x, y and z can share the same base address (in-place computing).
z[id] = CONCAT(VALUE_TYPE, _mul)(alpha, CONCAT(VALUE_TYPE, _mul)(x[id], y[id]));
}
}
void
CONCAT(GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL),_)
(spgpuHandle_t handle, VALUE_TYPE *z, int n, VALUE_TYPE alpha, VALUE_TYPE* x, VALUE_TYPE* y)
{
int msize = (n+BLOCK_SIZE-1)/BLOCK_SIZE;
dim3 block(BLOCK_SIZE);
dim3 grid(msize);
CONCAT(GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL),_kern)<<<grid, block, 0, handle->currentStream>>>(z, n, alpha, x, y);
}
void
GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL)
(spgpuHandle_t handle,
__device VALUE_TYPE *z,
int n,
VALUE_TYPE alpha,
__device VALUE_TYPE *x,
__device VALUE_TYPE *y)
{
int maxNForACall = max(handle->maxGridSizeX, BLOCK_SIZE*handle->maxGridSizeX);
while (n > maxNForACall) //managing large vectors
{
CONCAT(GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL),_)
(handle, z, maxNForACall, alpha, x, y);
x = x + maxNForACall;
y = y + maxNForACall;
z = z + maxNForACall;
n -= maxNForACall;
}
CONCAT(GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL),_) (handle, z, n, alpha, x, y);
cudaCheckError("CUDA error on axy");
}
void
GEN_SPGPU_MFUNC_NAME(TYPE_SYMBOL)
(spgpuHandle_t handle,
__device VALUE_TYPE *z,
int n,
VALUE_TYPE alpha,
__device VALUE_TYPE* x,
__device VALUE_TYPE *y,
int count,
int pitch)
{
for (int i=0; i<count; i++)
{
GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL)(handle, z, n, alpha, x, y);
x += pitch;
y += pitch;
z += pitch;
}
}
__global__ void
CONCAT(GEN_SPGPU_FUNC_NAME_2(TYPE_SYMBOL),_kern)
(VALUE_TYPE *w, int n, VALUE_TYPE beta, VALUE_TYPE* z, VALUE_TYPE alpha, VALUE_TYPE* x, VALUE_TYPE* y)
{
int id = threadIdx.x + BLOCK_SIZE*blockIdx.x;
if (id < n)
{
// Since w, x and y and z are accessed with the same offset by the same thread,
// and the write to z follows the x, y and z reads, x, y, z and w can share the same base address (in-place computing).
w[id] = CONCAT(VALUE_TYPE, _fma) (
alpha,
CONCAT(VALUE_TYPE, _mul)(x[id],y[id]),
CONCAT(VALUE_TYPE, _mul)(beta,z[id]));
}
}
void
CONCAT(GEN_SPGPU_FUNC_NAME_2(TYPE_SYMBOL),_)
(spgpuHandle_t handle, VALUE_TYPE *w, int n, VALUE_TYPE beta, VALUE_TYPE* z, VALUE_TYPE alpha, VALUE_TYPE* x, VALUE_TYPE* y)
{
int msize = (n+BLOCK_SIZE-1)/BLOCK_SIZE;
dim3 block(BLOCK_SIZE);
dim3 grid(msize);
CONCAT(GEN_SPGPU_FUNC_NAME_2(TYPE_SYMBOL),_kern)
<<<grid, block, 0, handle->currentStream>>>(w, n, beta, z, alpha, x, y);
}
void
GEN_SPGPU_FUNC_NAME_2(TYPE_SYMBOL)
(spgpuHandle_t handle,
__device VALUE_TYPE *w,
int n,
VALUE_TYPE beta,
__device VALUE_TYPE *z,
VALUE_TYPE alpha,
__device VALUE_TYPE* x,
__device VALUE_TYPE *y
)
{
if (CONCAT(VALUE_TYPE, _isZero(alpha)))
{
GEN_SPGPU_SCAL_NAME(TYPE_SYMBOL)
(handle, w, n, beta, z);
}
else if (CONCAT(VALUE_TYPE, _isZero(beta))) {
GEN_SPGPU_FUNC_NAME(TYPE_SYMBOL)
(handle, w, n, alpha, x, y);
}
else {
int maxNForACall = max(handle->maxGridSizeX, BLOCK_SIZE*handle->maxGridSizeX);
while (n > maxNForACall) //managing large vectors
{
CONCAT(GEN_SPGPU_FUNC_NAME_2(TYPE_SYMBOL),_)
(handle, w, maxNForACall, beta, z, alpha, x, y);
x = x + maxNForACall;
y = y + maxNForACall;
z = z + maxNForACall;
w = w + maxNForACall;
n -= maxNForACall;
}
CONCAT(GEN_SPGPU_FUNC_NAME_2(TYPE_SYMBOL),_)
(handle, w, n, beta, z, alpha, x, y);
}
cudaCheckError("CUDA error on axypbz");
}
void
GEN_SPGPU_MFUNC_NAME_2(TYPE_SYMBOL)
(spgpuHandle_t handle,
__device VALUE_TYPE *w,
int n,
VALUE_TYPE beta,
__device VALUE_TYPE *z,
VALUE_TYPE alpha,
__device VALUE_TYPE* x,
__device VALUE_TYPE *y,
int count,
int pitch)
{
for (int i=0; i<count; i++)
{
GEN_SPGPU_FUNC_NAME_2(TYPE_SYMBOL)(handle, w, n, beta, z, alpha, x, y);
x += pitch;
y += pitch;
z += pitch;
w += pitch;
}
}