340 lines
19 KiB

/*
* This is an extension of the original CUSP csr_vector.h SPMV implementation.
* It is based on the Code and incorporates changes as to cope with the details
* of the StoRM code.
* As this is mostly copy & paste, the original license still applies.
*/
/*
* Copyright 2008-2009 NVIDIA Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#pragma once
#include <cusp/detail/device/spmv/csr_vector.h>
#include <limits>
#include <algorithm>
namespace cusp
{
namespace detail
{
namespace device
{
//////////////////////////////////////////////////////////////////////////////
// CSR SpMV kernels based on a vector model (one warp per row)
//////////////////////////////////////////////////////////////////////////////
//
// spmv_csr_vector_device
// Each row of the CSR matrix is assigned to a warp. The warp computes
// y[i] = A[i,:] * x, i.e. the dot product of the i-th row of A with
// the x vector, in parallel. This division of work implies that
// the CSR index and data arrays (Aj and Ax) are accessed in a contiguous
// manner (but generally not aligned). On GT200 these accesses are
// coalesced, unlike kernels based on the one-row-per-thread division of
// work. Since an entire 32-thread warp is assigned to each row, many
// threads will remain idle when their row contains a small number
// of elements. This code relies on implicit synchronization among
// threads in a warp.
//
// spmv_csr_vector_tex_device
// Same as spmv_csr_vector_tex_device, except that the texture cache is
// used for accessing the x vector.
//
// Note: THREADS_PER_VECTOR must be one of [2,4,8,16,32]
template <typename IndexType, typename ValueType, unsigned int VECTORS_PER_BLOCK, unsigned int THREADS_PER_VECTOR, bool UseCache>
__launch_bounds__(VECTORS_PER_BLOCK * THREADS_PER_VECTOR,1)
__global__ void
storm_cuda_opt_spmv_csr_vector_kernel(const IndexType num_rows, const IndexType * matrixRowIndices, const IndexType * matrixColumnIndicesAndValues, const ValueType * x, ValueType * y)
{
__shared__ volatile ValueType sdata[VECTORS_PER_BLOCK * THREADS_PER_VECTOR + THREADS_PER_VECTOR / 2]; // padded to avoid reduction conditionals
__shared__ volatile IndexType ptrs[VECTORS_PER_BLOCK][2];
const IndexType THREADS_PER_BLOCK = VECTORS_PER_BLOCK * THREADS_PER_VECTOR;
const IndexType thread_id = THREADS_PER_BLOCK * blockIdx.x + threadIdx.x; // global thread index
const IndexType thread_lane = threadIdx.x & (THREADS_PER_VECTOR - 1); // thread index within the vector
const IndexType vector_id = thread_id / THREADS_PER_VECTOR; // global vector index
const IndexType vector_lane = threadIdx.x / THREADS_PER_VECTOR; // vector index within the block
const IndexType num_vectors = VECTORS_PER_BLOCK * gridDim.x; // total number of active vectors
for(IndexType row = vector_id; row < num_rows; row += num_vectors)
{
// use two threads to fetch Ap[row] and Ap[row+1]
// this is considerably faster than the straightforward version
if(thread_lane < 2)
ptrs[vector_lane][thread_lane] = matrixRowIndices[row + thread_lane];
const IndexType row_start = ptrs[vector_lane][0]; //same as: row_start = Ap[row];
const IndexType row_end = ptrs[vector_lane][1]; //same as: row_end = Ap[row+1];
// initialize local sum
ValueType sum = 0;
if (THREADS_PER_VECTOR == 32 && row_end - row_start > 32)
{
// ensure aligned memory access to Aj and Ax
IndexType jj = row_start - (row_start & (THREADS_PER_VECTOR - 1)) + thread_lane;
// accumulate local sums
if(jj >= row_start && jj < row_end)
sum += matrixColumnIndicesAndValues[(2 * jj) + 1] * fetch_x<UseCache>(matrixColumnIndicesAndValues[2 * jj], x);
// accumulate local sums
for(jj += THREADS_PER_VECTOR; jj < row_end; jj += THREADS_PER_VECTOR)
sum += matrixColumnIndicesAndValues[(2 * jj) + 1] * fetch_x<UseCache>(matrixColumnIndicesAndValues[2 * jj], x);
}
else
{
// accumulate local sums
for(IndexType jj = row_start + thread_lane; jj < row_end; jj += THREADS_PER_VECTOR)
sum += matrixColumnIndicesAndValues[(2 * jj) + 1] * fetch_x<UseCache>(matrixColumnIndicesAndValues[2 * jj], x);
}
// store local sum in shared memory
sdata[threadIdx.x] = sum;
// reduce local sums to row sum
if (THREADS_PER_VECTOR > 16) sdata[threadIdx.x] = sum = sum + sdata[threadIdx.x + 16];
if (THREADS_PER_VECTOR > 8) sdata[threadIdx.x] = sum = sum + sdata[threadIdx.x + 8];
if (THREADS_PER_VECTOR > 4) sdata[threadIdx.x] = sum = sum + sdata[threadIdx.x + 4];
if (THREADS_PER_VECTOR > 2) sdata[threadIdx.x] = sum = sum + sdata[threadIdx.x + 2];
if (THREADS_PER_VECTOR > 1) sdata[threadIdx.x] = sum = sum + sdata[threadIdx.x + 1];
// first thread writes the result
if (thread_lane == 0)
y[row] = sdata[threadIdx.x];
}
}
template <typename IndexType, typename ValueType, unsigned int ROWS_PER_BLOCK, unsigned int THREADS_PER_ROW, bool Minimize>
__launch_bounds__(ROWS_PER_BLOCK * THREADS_PER_ROW,1)
__global__ void
storm_cuda_opt_vector_reduce_kernel(const IndexType num_rows, const IndexType * nondeterministicChoiceIndices, ValueType * x, const ValueType * y, const ValueType minMaxInitializer)
{
__shared__ volatile ValueType sdata[ROWS_PER_BLOCK * THREADS_PER_ROW + THREADS_PER_ROW / 2]; // padded to avoid reduction conditionals
__shared__ volatile IndexType ptrs[ROWS_PER_BLOCK][2];
const IndexType THREADS_PER_BLOCK = ROWS_PER_BLOCK * THREADS_PER_ROW;
const IndexType thread_id = THREADS_PER_BLOCK * blockIdx.x + threadIdx.x; // global thread index
const IndexType thread_lane = threadIdx.x & (THREADS_PER_ROW - 1); // thread index within the vector
const IndexType vector_id = thread_id / THREADS_PER_ROW; // global vector index
const IndexType vector_lane = threadIdx.x / THREADS_PER_ROW; // vector index within the block
const IndexType num_vectors = ROWS_PER_BLOCK * gridDim.x; // total number of active vectors
for(IndexType row = vector_id; row < num_rows; row += num_vectors)
{
// use two threads to fetch Ap[row] and Ap[row+1]
// this is considerably faster than the straightforward version
if(thread_lane < 2)
ptrs[vector_lane][thread_lane] = nondeterministicChoiceIndices[row + thread_lane];
const IndexType row_start = ptrs[vector_lane][0]; //same as: row_start = Ap[row];
const IndexType row_end = ptrs[vector_lane][1]; //same as: row_end = Ap[row+1];
// initialize local Min/Max
ValueType localMinMaxElement = minMaxInitializer;
if (THREADS_PER_ROW == 32 && row_end - row_start > 32)
{
// ensure aligned memory access to Aj and Ax
IndexType jj = row_start - (row_start & (THREADS_PER_ROW - 1)) + thread_lane;
// accumulate local sums
if(jj >= row_start && jj < row_end) {
if(Minimize) {
localMinMaxElement = (localMinMaxElement > y[jj]) ? y[jj] : localMinMaxElement;
} else {
localMinMaxElement = (localMinMaxElement < y[jj]) ? y[jj] : localMinMaxElement;
}
}
// accumulate local sums
for(jj += THREADS_PER_ROW; jj < row_end; jj += THREADS_PER_ROW)
if(Minimize) {
localMinMaxElement = (localMinMaxElement > y[jj]) ? y[jj] : localMinMaxElement;
} else {
localMinMaxElement = (localMinMaxElement < y[jj]) ? y[jj] : localMinMaxElement;
}
}
else
{
// accumulate local sums
for(IndexType jj = row_start + thread_lane; jj < row_end; jj += THREADS_PER_ROW)
if(Minimize) {
localMinMaxElement = (localMinMaxElement > y[jj]) ? y[jj] : localMinMaxElement;
} else {
localMinMaxElement = (localMinMaxElement < y[jj]) ? y[jj] : localMinMaxElement;
}
}
// store local sum in shared memory
sdata[threadIdx.x] = localMinMaxElement;
// reduce local min/max to row min/max
if (Minimize) {
if (THREADS_PER_ROW > 16) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement > sdata[threadIdx.x + 16]) ? sdata[threadIdx.x + 16] : localMinMaxElement);
if (THREADS_PER_ROW > 8) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement > sdata[threadIdx.x + 8]) ? sdata[threadIdx.x + 8] : localMinMaxElement);
if (THREADS_PER_ROW > 4) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement > sdata[threadIdx.x + 4]) ? sdata[threadIdx.x + 4] : localMinMaxElement);
if (THREADS_PER_ROW > 2) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement > sdata[threadIdx.x + 2]) ? sdata[threadIdx.x + 2] : localMinMaxElement);
if (THREADS_PER_ROW > 1) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement > sdata[threadIdx.x + 1]) ? sdata[threadIdx.x + 1] : localMinMaxElement);
} else {
if (THREADS_PER_ROW > 16) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement < sdata[threadIdx.x + 16]) ? sdata[threadIdx.x + 16] : localMinMaxElement);
if (THREADS_PER_ROW > 8) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement < sdata[threadIdx.x + 8]) ? sdata[threadIdx.x + 8] : localMinMaxElement);
if (THREADS_PER_ROW > 4) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement < sdata[threadIdx.x + 4]) ? sdata[threadIdx.x + 4] : localMinMaxElement);
if (THREADS_PER_ROW > 2) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement < sdata[threadIdx.x + 2]) ? sdata[threadIdx.x + 2] : localMinMaxElement);
if (THREADS_PER_ROW > 1) sdata[threadIdx.x] = localMinMaxElement = ((localMinMaxElement < sdata[threadIdx.x + 1]) ? sdata[threadIdx.x + 1] : localMinMaxElement);
}
// first thread writes the result
if (thread_lane == 0)
x[row] = sdata[threadIdx.x];
}
}
template <bool Minimize, unsigned int THREADS_PER_VECTOR, typename IndexType, typename ValueType>
void __storm_cuda_opt_vector_reduce(const IndexType num_rows, const IndexType * nondeterministicChoiceIndices, ValueType * x, const ValueType * y)
{
ValueType __minMaxInitializer = 0;
if (Minimize) {
__minMaxInitializer = std::numeric_limits<ValueType>::max();
}
const ValueType minMaxInitializer = __minMaxInitializer;
const size_t THREADS_PER_BLOCK = 128;
const size_t VECTORS_PER_BLOCK = THREADS_PER_BLOCK / THREADS_PER_VECTOR;
const size_t MAX_BLOCKS = cusp::detail::device::arch::max_active_blocks(storm_cuda_opt_vector_reduce_kernel<IndexType, ValueType, VECTORS_PER_BLOCK, THREADS_PER_VECTOR, Minimize>, THREADS_PER_BLOCK, (size_t) 0);
const size_t NUM_BLOCKS = std::min<size_t>(MAX_BLOCKS, DIVIDE_INTO(num_rows, VECTORS_PER_BLOCK));
storm_cuda_opt_vector_reduce_kernel<IndexType, ValueType, VECTORS_PER_BLOCK, THREADS_PER_VECTOR, Minimize> <<<NUM_BLOCKS, THREADS_PER_BLOCK>>>
(num_rows, nondeterministicChoiceIndices, x, y, minMaxInitializer);
}
template <bool Minimize, typename IndexType, typename ValueType>
void storm_cuda_opt_vector_reduce(const IndexType num_rows, const IndexType num_entries, const IndexType * nondeterministicChoiceIndices, ValueType * x, const ValueType * y)
{
const IndexType rows_per_group = num_entries / num_rows;
if (rows_per_group <= 2) { __storm_cuda_opt_vector_reduce<Minimize, 2>(num_rows, nondeterministicChoiceIndices, x, y); return; }
if (rows_per_group <= 4) { __storm_cuda_opt_vector_reduce<Minimize, 4>(num_rows, nondeterministicChoiceIndices, x, y); return; }
if (rows_per_group <= 8) { __storm_cuda_opt_vector_reduce<Minimize, 8>(num_rows, nondeterministicChoiceIndices, x, y); return; }
if (rows_per_group <= 16) { __storm_cuda_opt_vector_reduce<Minimize,16>(num_rows, nondeterministicChoiceIndices, x, y); return; }
__storm_cuda_opt_vector_reduce<Minimize,32>(num_rows, nondeterministicChoiceIndices, x, y);
}
template <bool UseCache, unsigned int THREADS_PER_VECTOR, typename IndexType, typename ValueType>
void __storm_cuda_opt_spmv_csr_vector(const IndexType num_rows, const IndexType * matrixRowIndices, const IndexType * matrixColumnIndicesAndValues, const ValueType* x, ValueType* y)
{
const size_t THREADS_PER_BLOCK = 128;
const size_t VECTORS_PER_BLOCK = THREADS_PER_BLOCK / THREADS_PER_VECTOR;
const size_t MAX_BLOCKS = cusp::detail::device::arch::max_active_blocks(storm_cuda_opt_spmv_csr_vector_kernel<IndexType, ValueType, VECTORS_PER_BLOCK, THREADS_PER_VECTOR, UseCache>, THREADS_PER_BLOCK, (size_t) 0);
const size_t NUM_BLOCKS = std::min<size_t>(MAX_BLOCKS, DIVIDE_INTO(num_rows, VECTORS_PER_BLOCK));
if (UseCache)
bind_x(x);
storm_cuda_opt_spmv_csr_vector_kernel<IndexType, ValueType, VECTORS_PER_BLOCK, THREADS_PER_VECTOR, UseCache> <<<NUM_BLOCKS, THREADS_PER_BLOCK>>>
(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y);
if (UseCache)
unbind_x(x);
}
template <typename IndexType, typename ValueType>
void storm_cuda_opt_spmv_csr_vector(const IndexType num_rows, const IndexType num_entries, const IndexType * matrixRowIndices, const IndexType * matrixColumnIndicesAndValues, const ValueType* x, ValueType* y)
{
const IndexType nnz_per_row = num_entries / num_rows;
if (nnz_per_row <= 2) { __storm_cuda_opt_spmv_csr_vector<false, 2>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
if (nnz_per_row <= 4) { __storm_cuda_opt_spmv_csr_vector<false, 4>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
if (nnz_per_row <= 8) { __storm_cuda_opt_spmv_csr_vector<false, 8>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
if (nnz_per_row <= 16) { __storm_cuda_opt_spmv_csr_vector<false,16>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
__storm_cuda_opt_spmv_csr_vector<false,32>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y);
}
template <typename IndexType, typename ValueType>
void storm_cuda_opt_spmv_csr_vector_tex(const IndexType num_rows, const IndexType num_entries, const IndexType * matrixRowIndices, const IndexType * matrixColumnIndicesAndValues, const ValueType* x, ValueType* y)
{
const IndexType nnz_per_row = num_entries / num_rows;
if (nnz_per_row <= 2) { __storm_cuda_opt_spmv_csr_vector<true, 2>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
if (nnz_per_row <= 4) { __storm_cuda_opt_spmv_csr_vector<true, 4>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
if (nnz_per_row <= 8) { __storm_cuda_opt_spmv_csr_vector<true, 8>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
if (nnz_per_row <= 16) { __storm_cuda_opt_spmv_csr_vector<true,16>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y); return; }
__storm_cuda_opt_spmv_csr_vector<true,32>(num_rows, matrixRowIndices, matrixColumnIndicesAndValues, x, y);
}
// NON-OPT
template <bool UseCache, unsigned int THREADS_PER_VECTOR, typename IndexType, typename ValueType>
void __storm_cuda_spmv_csr_vector(const IndexType num_rows, const IndexType * matrixRowIndices, const IndexType * matrixColumnIndices, const ValueType * matrixValues, const ValueType* x, ValueType* y)
{
const size_t THREADS_PER_BLOCK = 128;
const size_t VECTORS_PER_BLOCK = THREADS_PER_BLOCK / THREADS_PER_VECTOR;
const size_t MAX_BLOCKS = cusp::detail::device::arch::max_active_blocks(spmv_csr_vector_kernel<IndexType, ValueType, VECTORS_PER_BLOCK, THREADS_PER_VECTOR, UseCache>, THREADS_PER_BLOCK, (size_t) 0);
const size_t NUM_BLOCKS = std::min<size_t>(MAX_BLOCKS, DIVIDE_INTO(num_rows, VECTORS_PER_BLOCK));
if (UseCache)
bind_x(x);
spmv_csr_vector_kernel<IndexType, ValueType, VECTORS_PER_BLOCK, THREADS_PER_VECTOR, UseCache> <<<NUM_BLOCKS, THREADS_PER_BLOCK>>>
(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y);
if (UseCache)
unbind_x(x);
}
template <typename IndexType, typename ValueType>
void storm_cuda_spmv_csr_vector(const IndexType num_rows, const IndexType num_entries, const IndexType * matrixRowIndices, const IndexType * matrixColumnIndices, const ValueType * matrixValues, const ValueType* x, ValueType* y)
{
const IndexType nnz_per_row = num_entries / num_rows;
if (nnz_per_row <= 2) { __storm_cuda_spmv_csr_vector<false, 2>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
if (nnz_per_row <= 4) { __storm_cuda_spmv_csr_vector<false, 4>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
if (nnz_per_row <= 8) { __storm_cuda_spmv_csr_vector<false, 8>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
if (nnz_per_row <= 16) { __storm_cuda_spmv_csr_vector<false,16>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
__storm_cuda_spmv_csr_vector<false,32>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y);
}
template <typename IndexType, typename ValueType>
void storm_cuda_spmv_csr_vector_tex(const IndexType num_rows, const IndexType num_entries, const IndexType * matrixRowIndices, const IndexType * matrixColumnIndices, const ValueType * matrixValues, const ValueType* x, ValueType* y)
{
const IndexType nnz_per_row = num_entries / num_rows;
if (nnz_per_row <= 2) { __storm_cuda_spmv_csr_vector<true, 2>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
if (nnz_per_row <= 4) { __storm_cuda_spmv_csr_vector<true, 4>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
if (nnz_per_row <= 8) { __storm_cuda_spmv_csr_vector<true, 8>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
if (nnz_per_row <= 16) { __storm_cuda_spmv_csr_vector<true,16>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y); return; }
__storm_cuda_spmv_csr_vector<true,32>(num_rows, matrixRowIndices, matrixColumnIndices, matrixValues, x, y);
}
} // end namespace device
} // end namespace detail
} // end namespace cusp