blob: 780e8961eae92abf53a85e9234f52d28cfd010c0 [file] [log] [blame]
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2014-2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
// Copyright (C) 2015 Navdeep Jaitly <ndjaitly@google.com>
// Copyright (C) 2014 Eric Martin <eric@ericmart.in>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
#define EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H
#if defined(EIGEN_USE_GPU) && defined(EIGEN_GPUCC)
// IWYU pragma: private
#include "./InternalHeaderCheck.h"
namespace Eigen {
template <typename Scalar, typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper,
bool needs_edge_check>
__device__ EIGEN_STRONG_INLINE void EigenContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
const OutputMapper output, Scalar* lhs_shmem,
Scalar* rhs_shmem, const Index m_size,
const Index n_size, const Index k_size) {
const Index m_block_idx = blockIdx.x;
const Index n_block_idx = blockIdx.y;
const Index base_m = 64 * m_block_idx;
const Index base_n = 64 * n_block_idx;
// declare and initialize 64 registers for output 8x8 block
// prefetch registers
Scalar lhs_pf0;
Scalar lhs_pf1;
Scalar lhs_pf2;
Scalar lhs_pf3;
Scalar lhs_pf4;
Scalar lhs_pf5;
Scalar lhs_pf6;
Scalar lhs_pf7;
Scalar rhs_pf0;
Scalar rhs_pf1;
Scalar rhs_pf2;
Scalar rhs_pf3;
Scalar rhs_pf4;
Scalar rhs_pf5;
Scalar rhs_pf6;
Scalar rhs_pf7;
// shared memory is formatted
// (contract idx in block, nocontract idx in block, block idx)
// where block idx is column major. This transposition limits the number of
// bank conflicts when reading the LHS. The core idea is that since the contracting
// index is shared by both sides, then the contracting index should be in threadIdx.x.
// On the LHS, we pad each row inside of each block with an extra element. This makes
// each block 8 rows of 9 elements, which is 72 elements. This gives no bank conflicts
// on writes and very few 2-way conflicts on reads. There is an 8x8 grid of these blocks.
// On the RHS we just add 8 padding elements to the end of each block. This gives no bank
// conflicts on writes and also none on reads.
// storage indices
const Index lhs_store_idx_base = threadIdx.y * 72 + threadIdx.x * 9 + threadIdx.z;
const Index rhs_store_idx_base = threadIdx.y * 72 + threadIdx.z * 8 + threadIdx.x;
const Index lhs_store_idx_0 = lhs_store_idx_base + 576 * 0;
const Index lhs_store_idx_1 = lhs_store_idx_base + 576 * 1;
const Index lhs_store_idx_2 = lhs_store_idx_base + 576 * 2;
const Index lhs_store_idx_3 = lhs_store_idx_base + 576 * 3;
const Index lhs_store_idx_4 = lhs_store_idx_base + 576 * 4;
const Index lhs_store_idx_5 = lhs_store_idx_base + 576 * 5;
const Index lhs_store_idx_6 = lhs_store_idx_base + 576 * 6;
const Index lhs_store_idx_7 = lhs_store_idx_base + 576 * 7;
const Index rhs_store_idx_0 = rhs_store_idx_base + 576 * 0;
const Index rhs_store_idx_1 = rhs_store_idx_base + 576 * 1;
const Index rhs_store_idx_2 = rhs_store_idx_base + 576 * 2;
const Index rhs_store_idx_3 = rhs_store_idx_base + 576 * 3;
const Index rhs_store_idx_4 = rhs_store_idx_base + 576 * 4;
const Index rhs_store_idx_5 = rhs_store_idx_base + 576 * 5;
const Index rhs_store_idx_6 = rhs_store_idx_base + 576 * 6;
const Index rhs_store_idx_7 = rhs_store_idx_base + 576 * 7;
// in the loading code, the following variables are important:
// threadIdx.x: the vertical position in an 8x8 block
// threadIdx.y: the vertical index of the 8x8 block in the grid
// threadIdx.z: the horizontal position in an 8x8 block
// k: the horizontal index of the 8x8 block in the grid
//
// The k parameter is implicit (it was the loop counter for a loop that went
// from 0 to <8, but now that loop is unrolled in the below code.
const Index load_idx_vert = threadIdx.x + 8 * threadIdx.y;
const Index lhs_vert = base_m + load_idx_vert;
#define prefetchIntoRegisters(base_k) \
{ \
lhs_pf0 = conv(0); \
lhs_pf1 = conv(0); \
lhs_pf2 = conv(0); \
lhs_pf3 = conv(0); \
lhs_pf4 = conv(0); \
lhs_pf5 = conv(0); \
lhs_pf6 = conv(0); \
lhs_pf7 = conv(0); \
\
rhs_pf0 = conv(0); \
rhs_pf1 = conv(0); \
rhs_pf2 = conv(0); \
rhs_pf3 = conv(0); \
rhs_pf4 = conv(0); \
rhs_pf5 = conv(0); \
rhs_pf6 = conv(0); \
rhs_pf7 = conv(0); \
\
if (!needs_edge_check || lhs_vert < m_size) { \
const Index lhs_horiz_0 = base_k + threadIdx.z + 0 * 8; \
const Index lhs_horiz_1 = base_k + threadIdx.z + 1 * 8; \
const Index lhs_horiz_2 = base_k + threadIdx.z + 2 * 8; \
const Index lhs_horiz_3 = base_k + threadIdx.z + 3 * 8; \
const Index lhs_horiz_4 = base_k + threadIdx.z + 4 * 8; \
const Index lhs_horiz_5 = base_k + threadIdx.z + 5 * 8; \
const Index lhs_horiz_6 = base_k + threadIdx.z + 6 * 8; \
const Index lhs_horiz_7 = base_k + threadIdx.z + 7 * 8; \
\
if (!needs_edge_check || lhs_horiz_7 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \
lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \
lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \
lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \
lhs_pf5 = lhs(lhs_vert, lhs_horiz_5); \
lhs_pf6 = lhs(lhs_vert, lhs_horiz_6); \
lhs_pf7 = lhs(lhs_vert, lhs_horiz_7); \
} else if (lhs_horiz_6 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \
lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \
lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \
lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \
lhs_pf5 = lhs(lhs_vert, lhs_horiz_5); \
lhs_pf6 = lhs(lhs_vert, lhs_horiz_6); \
} else if (lhs_horiz_5 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \
lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \
lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \
lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \
lhs_pf5 = lhs(lhs_vert, lhs_horiz_5); \
} else if (lhs_horiz_4 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \
lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \
lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \
lhs_pf4 = lhs(lhs_vert, lhs_horiz_4); \
} else if (lhs_horiz_3 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \
lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \
lhs_pf3 = lhs(lhs_vert, lhs_horiz_3); \
} else if (lhs_horiz_2 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \
lhs_pf2 = lhs(lhs_vert, lhs_horiz_2); \
} else if (lhs_horiz_1 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
lhs_pf1 = lhs(lhs_vert, lhs_horiz_1); \
} else if (lhs_horiz_0 < k_size) { \
lhs_pf0 = lhs(lhs_vert, lhs_horiz_0); \
} \
} \
\
const Index rhs_vert = base_k + load_idx_vert; \
if (!needs_edge_check || rhs_vert < k_size) { \
const Index rhs_horiz_0 = base_n + threadIdx.z + 0 * 8; \
const Index rhs_horiz_1 = base_n + threadIdx.z + 1 * 8; \
const Index rhs_horiz_2 = base_n + threadIdx.z + 2 * 8; \
const Index rhs_horiz_3 = base_n + threadIdx.z + 3 * 8; \
const Index rhs_horiz_4 = base_n + threadIdx.z + 4 * 8; \
const Index rhs_horiz_5 = base_n + threadIdx.z + 5 * 8; \
const Index rhs_horiz_6 = base_n + threadIdx.z + 6 * 8; \
const Index rhs_horiz_7 = base_n + threadIdx.z + 7 * 8; \
\
if (rhs_horiz_7 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \
rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \
rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \
rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \
rhs_pf5 = rhs(rhs_vert, rhs_horiz_5); \
rhs_pf6 = rhs(rhs_vert, rhs_horiz_6); \
rhs_pf7 = rhs(rhs_vert, rhs_horiz_7); \
} else if (rhs_horiz_6 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \
rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \
rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \
rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \
rhs_pf5 = rhs(rhs_vert, rhs_horiz_5); \
rhs_pf6 = rhs(rhs_vert, rhs_horiz_6); \
} else if (rhs_horiz_5 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \
rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \
rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \
rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \
rhs_pf5 = rhs(rhs_vert, rhs_horiz_5); \
} else if (rhs_horiz_4 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \
rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \
rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \
rhs_pf4 = rhs(rhs_vert, rhs_horiz_4); \
} else if (rhs_horiz_3 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \
rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \
rhs_pf3 = rhs(rhs_vert, rhs_horiz_3); \
} else if (rhs_horiz_2 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \
rhs_pf2 = rhs(rhs_vert, rhs_horiz_2); \
} else if (rhs_horiz_1 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
rhs_pf1 = rhs(rhs_vert, rhs_horiz_1); \
} else if (rhs_horiz_0 < n_size) { \
rhs_pf0 = rhs(rhs_vert, rhs_horiz_0); \
} \
} \
}
#define writeRegToShmem() \
lhs_shmem[lhs_store_idx_0] = lhs_pf0; \
rhs_shmem[rhs_store_idx_0] = rhs_pf0; \
\
lhs_shmem[lhs_store_idx_1] = lhs_pf1; \
rhs_shmem[rhs_store_idx_1] = rhs_pf1; \
\
lhs_shmem[lhs_store_idx_2] = lhs_pf2; \
rhs_shmem[rhs_store_idx_2] = rhs_pf2; \
\
lhs_shmem[lhs_store_idx_3] = lhs_pf3; \
rhs_shmem[rhs_store_idx_3] = rhs_pf3; \
\
lhs_shmem[lhs_store_idx_4] = lhs_pf4; \
rhs_shmem[rhs_store_idx_4] = rhs_pf4; \
\
lhs_shmem[lhs_store_idx_5] = lhs_pf5; \
rhs_shmem[rhs_store_idx_5] = rhs_pf5; \
\
lhs_shmem[lhs_store_idx_6] = lhs_pf6; \
rhs_shmem[rhs_store_idx_6] = rhs_pf6; \
\
lhs_shmem[lhs_store_idx_7] = lhs_pf7; \
rhs_shmem[rhs_store_idx_7] = rhs_pf7;
// declare and initialize result array
#define res(i, j) _res_##i##j
#define initResultRow(i) \
Scalar res(i, 0) = conv(0); \
Scalar res(i, 1) = conv(0); \
Scalar res(i, 2) = conv(0); \
Scalar res(i, 3) = conv(0); \
Scalar res(i, 4) = conv(0); \
Scalar res(i, 5) = conv(0); \
Scalar res(i, 6) = conv(0); \
Scalar res(i, 7) = conv(0);
internal::scalar_cast_op<int, Scalar> conv;
initResultRow(0);
initResultRow(1);
initResultRow(2);
initResultRow(3);
initResultRow(4);
initResultRow(5);
initResultRow(6);
initResultRow(7);
#undef initResultRow
for (Index base_k = 0; base_k < k_size; base_k += 64) {
// wait for previous iteration to finish with shmem. Despite common sense,
// the code is a bit faster with this here then at bottom of loop
__syncthreads();
prefetchIntoRegisters(base_k);
writeRegToShmem();
#undef prefetchIntoRegisters
#undef writeRegToShmem
// wait for shared mem packing to be done before starting computation
__syncthreads();
// compute 8x8 matrix product by outer product. This involves packing one column
// of LHS and one row of RHS into registers (takes 16 registers).
#define lcol(i) _lcol##i
Scalar lcol(0);
Scalar lcol(1);
Scalar lcol(2);
Scalar lcol(3);
Scalar lcol(4);
Scalar lcol(5);
Scalar lcol(6);
Scalar lcol(7);
#define rrow(j) _rrow##j
Scalar rrow(0);
Scalar rrow(1);
Scalar rrow(2);
Scalar rrow(3);
Scalar rrow(4);
Scalar rrow(5);
Scalar rrow(6);
Scalar rrow(7);
// Now x corresponds to k, y to m, and z to n
const Scalar* lhs_block = &lhs_shmem[threadIdx.x + 9 * threadIdx.y];
const Scalar* rhs_block = &rhs_shmem[threadIdx.x + 8 * threadIdx.z];
#define lhs_element(i, j) lhs_block[72 * ((i) + 8 * (j))]
#define rhs_element(i, j) rhs_block[72 * ((i) + 8 * (j))]
#define loadData(i, j) \
lcol(0) = lhs_element(0, j); \
rrow(0) = rhs_element(i, 0); \
lcol(1) = lhs_element(1, j); \
rrow(1) = rhs_element(i, 1); \
lcol(2) = lhs_element(2, j); \
rrow(2) = rhs_element(i, 2); \
lcol(3) = lhs_element(3, j); \
rrow(3) = rhs_element(i, 3); \
lcol(4) = lhs_element(4, j); \
rrow(4) = rhs_element(i, 4); \
lcol(5) = lhs_element(5, j); \
rrow(5) = rhs_element(i, 5); \
lcol(6) = lhs_element(6, j); \
rrow(6) = rhs_element(i, 6); \
lcol(7) = lhs_element(7, j); \
rrow(7) = rhs_element(i, 7);
#define computeCol(j) \
res(0, j) += lcol(0) * rrow(j); \
res(1, j) += lcol(1) * rrow(j); \
res(2, j) += lcol(2) * rrow(j); \
res(3, j) += lcol(3) * rrow(j); \
res(4, j) += lcol(4) * rrow(j); \
res(5, j) += lcol(5) * rrow(j); \
res(6, j) += lcol(6) * rrow(j); \
res(7, j) += lcol(7) * rrow(j);
#define computePass(i) \
loadData(i, i); \
\
computeCol(0); \
computeCol(1); \
computeCol(2); \
computeCol(3); \
computeCol(4); \
computeCol(5); \
computeCol(6); \
computeCol(7);
computePass(0);
computePass(1);
computePass(2);
computePass(3);
computePass(4);
computePass(5);
computePass(6);
computePass(7);
#undef lcol
#undef rrow
#undef lhs_element
#undef rhs_element
#undef loadData
#undef computeCol
#undef computePass
} // end loop over k
// we've now iterated over all of the large (ie width 64) k blocks and
// accumulated results in registers. At this point thread (x, y, z) contains
// the sum across all big k blocks of the product of little k block of index (x, y)
// with block of index (y, z). To compute the final output, we need to reduce
// the 8 threads over y by summation.
#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000)
#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor(res(i, j), mask)
#else
#define shuffleInc(i, j, mask) res(i, j) += __shfl_xor_sync(0xFFFFFFFF, res(i, j), mask)
#endif
#define reduceRow(i, mask) \
shuffleInc(i, 0, mask); \
shuffleInc(i, 1, mask); \
shuffleInc(i, 2, mask); \
shuffleInc(i, 3, mask); \
shuffleInc(i, 4, mask); \
shuffleInc(i, 5, mask); \
shuffleInc(i, 6, mask); \
shuffleInc(i, 7, mask);
#define reduceMatrix(mask) \
reduceRow(0, mask); \
reduceRow(1, mask); \
reduceRow(2, mask); \
reduceRow(3, mask); \
reduceRow(4, mask); \
reduceRow(5, mask); \
reduceRow(6, mask); \
reduceRow(7, mask);
// actually perform the reduction, now each thread of index (_, y, z)
// contains the correct values in its registers that belong in the output
// block
reduceMatrix(1);
reduceMatrix(2);
reduceMatrix(4);
#undef shuffleInc
#undef reduceRow
#undef reduceMatrix
// now we need to copy the 64 values into main memory. We can't split work
// among threads because all variables are in registers. There's 2 ways
// to do this:
// (1) have 1 thread do 64 writes from registers into global memory
// (2) have 1 thread do 64 writes into shared memory, and then 8 threads
// each do 8 writes into global memory. We can just overwrite the shared
// memory from the problem we just solved.
// (2) is slightly faster than (1) due to less branching and more ILP
// TODO: won't yield much gain, but could just use currently unused shared mem
// and then we won't have to sync
// wait for shared mem to be out of use
__syncthreads();
#define writeResultShmem(i, j) lhs_shmem[i + 8 * threadIdx.y + 64 * threadIdx.z + 512 * j] = res(i, j);
#define writeRow(i) \
writeResultShmem(i, 0); \
writeResultShmem(i, 1); \
writeResultShmem(i, 2); \
writeResultShmem(i, 3); \
writeResultShmem(i, 4); \
writeResultShmem(i, 5); \
writeResultShmem(i, 6); \
writeResultShmem(i, 7);
if (threadIdx.x == 0) {
writeRow(0);
writeRow(1);
writeRow(2);
writeRow(3);
writeRow(4);
writeRow(5);
writeRow(6);
writeRow(7);
}
#undef writeResultShmem
#undef writeRow
const int max_i_write = numext::mini((int)((m_size - base_m - threadIdx.y + 7) / 8), 8);
const int max_j_write = numext::mini((int)((n_size - base_n - threadIdx.z + 7) / 8), 8);
if (threadIdx.x < max_i_write) {
if (max_j_write == 8) {
// TODO: can i trade bank conflicts for coalesced writes?
Scalar val0 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 0];
Scalar val1 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 1];
Scalar val2 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 2];
Scalar val3 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 3];
Scalar val4 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 4];
Scalar val5 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 5];
Scalar val6 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 6];
Scalar val7 = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * 7];
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 0) = val0;
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 1) = val1;
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 2) = val2;
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 3) = val3;
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 4) = val4;
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 5) = val5;
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 6) = val6;
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * 7) = val7;
} else {
#pragma unroll 7
for (int j = 0; j < max_j_write; j++) {
Scalar val = lhs_shmem[threadIdx.x + 8 * threadIdx.y + 64 * threadIdx.z + 512 * j];
output(base_m + threadIdx.y + 8 * threadIdx.x, base_n + threadIdx.z + 8 * j) = val;
}
}
}
#undef res
}
template <typename Scalar, typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper>
__global__ void
#if defined(EIGEN_HIPCC)
__launch_bounds__(512, 1)
#else
__launch_bounds__(512)
#endif
EigenContractionKernel(const LhsMapper lhs, const RhsMapper rhs, const OutputMapper output, const Index m_size,
const Index n_size, const Index k_size) {
__shared__ Scalar lhs_shmem[72 * 64];
__shared__ Scalar rhs_shmem[72 * 64];
const Index m_block_idx = blockIdx.x;
const Index n_block_idx = blockIdx.y;
const Index base_m = 64 * m_block_idx;
const Index base_n = 64 * n_block_idx;
if (base_m + 63 < m_size && base_n + 63 < n_size) {
EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, false>(
lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
} else {
EigenContractionKernelInternal<Scalar, Index, LhsMapper, RhsMapper, OutputMapper, true>(
lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size);
}
}
template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
bool CHECK_RHS_BOUNDARY>
__device__ __forceinline__ void EigenFloatContractionKernelInternal16x16(const LhsMapper lhs, const RhsMapper rhs,
const OutputMapper output,
float2 lhs_shmem2[][16],
float2 rhs_shmem2[][8], const Index m_size,
const Index n_size, const Index k_size,
const Index base_m, const Index base_n) {
// prefetch registers
float4 lhs_pf0, rhs_pf0;
float4 results[4];
for (int i = 0; i < 4; i++) {
results[i].x = results[i].y = results[i].z = results[i].w = 0;
}
#define prefetch_lhs(reg, row, col) \
if (!CHECK_LHS_BOUNDARY) { \
if (col < k_size) { \
reg = lhs.template loadPacket<float4, Unaligned>(row, col); \
} \
} else { \
if (col < k_size) { \
if (row + 3 < m_size) { \
reg = lhs.template loadPacket<float4, Unaligned>(row, col); \
} else if (row + 2 < m_size) { \
reg.x = lhs(row + 0, col); \
reg.y = lhs(row + 1, col); \
reg.z = lhs(row + 2, col); \
} else if (row + 1 < m_size) { \
reg.x = lhs(row + 0, col); \
reg.y = lhs(row + 1, col); \
} else if (row < m_size) { \
reg.x = lhs(row + 0, col); \
} \
} \
}
Index lhs_vert = base_m + threadIdx.x * 4;
for (Index k = 0; k < k_size; k += 16) {
lhs_pf0 = internal::pset1<float4>(0);
rhs_pf0 = internal::pset1<float4>(0);
Index lhs_horiz = threadIdx.y + k;
prefetch_lhs(lhs_pf0, lhs_vert, lhs_horiz)
Index rhs_vert = k + (threadIdx.x % 4) * 4;
Index rhs_horiz0 = (threadIdx.x >> 2) + threadIdx.y * 4 + base_n;
if (!CHECK_RHS_BOUNDARY) {
if ((rhs_vert + 3) < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0 = rhs.template loadPacket<float4, Unaligned>(rhs_vert, rhs_horiz0);
} else if (rhs_vert + 2 < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
} else if (rhs_vert + 1 < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
} else if (rhs_vert < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
}
} else {
if (rhs_horiz0 < n_size) {
if ((rhs_vert + 3) < k_size) {
rhs_pf0 = rhs.template loadPacket<float4, Unaligned>(rhs_vert, rhs_horiz0);
} else if ((rhs_vert + 2) < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
} else if ((rhs_vert + 1) < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
} else if (rhs_vert < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
}
}
}
float x1, x2;
// the following can be a bitwise operation..... some day.
if ((threadIdx.x % 8) < 4) {
x1 = rhs_pf0.y;
x2 = rhs_pf0.w;
} else {
x1 = rhs_pf0.x;
x2 = rhs_pf0.z;
}
#if defined(EIGEN_HIPCC) || (defined(EIGEN_CUDA_SDK_VER) && EIGEN_CUDA_SDK_VER < 90000)
x1 = __shfl_xor(x1, 4);
x2 = __shfl_xor(x2, 4);
#else
x1 = __shfl_xor_sync(0xFFFFFFFF, x1, 4);
x2 = __shfl_xor_sync(0xFFFFFFFF, x2, 4);
#endif
if ((threadIdx.x % 8) < 4) {
rhs_pf0.y = x1;
rhs_pf0.w = x2;
} else {
rhs_pf0.x = x1;
rhs_pf0.z = x2;
}
// We have 64 features.
// Row 0 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 0, 1.
// Row 1 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 2, 3.
// ...
// Row 31 -> times (0, 4, 8, 12, 1, 5, 9, 13) for features 62, 63
// Row 32 -> times (2, 6, 10, 14, 3, 7, 11, 15) for features 0, 1
// ...
rhs_shmem2[(threadIdx.x >> 3) + threadIdx.y * 2][threadIdx.x % 8] = make_float2(rhs_pf0.x, rhs_pf0.y);
rhs_shmem2[(threadIdx.x >> 3) + threadIdx.y * 2 + 32][threadIdx.x % 8] = make_float2(rhs_pf0.z, rhs_pf0.w);
// Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), .. (60, 61)
// Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), .. (60, 61)
// ...
// Row 15 (time 15) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), .. (60, 61)
// Row 16 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), .. (62, 63)
// ...
lhs_shmem2[threadIdx.y][threadIdx.x] = make_float2(lhs_pf0.x, lhs_pf0.y);
lhs_shmem2[threadIdx.y + 16][threadIdx.x] = make_float2(lhs_pf0.z, lhs_pf0.w);
#define add_vals(fl1, fl2, fr1, fr2) \
results[0].x += fl1.x * fr1.x; \
results[0].y += fl1.y * fr1.x; \
results[0].z += fl2.x * fr1.x; \
results[0].w += fl2.y * fr1.x; \
\
results[1].x += fl1.x * fr1.y; \
results[1].y += fl1.y * fr1.y; \
results[1].z += fl2.x * fr1.y; \
results[1].w += fl2.y * fr1.y; \
\
results[2].x += fl1.x * fr2.x; \
results[2].y += fl1.y * fr2.x; \
results[2].z += fl2.x * fr2.x; \
results[2].w += fl2.y * fr2.x; \
\
results[3].x += fl1.x * fr2.y; \
results[3].y += fl1.y * fr2.y; \
results[3].z += fl2.x * fr2.y; \
results[3].w += fl2.y * fr2.y;
__syncthreads();
// Do the multiplies.
#pragma unroll
for (int koff = 0; koff < 16; koff++) {
// 32 x threads.
float2 fl1 = lhs_shmem2[koff][threadIdx.x];
float2 fl2 = lhs_shmem2[koff + 16][threadIdx.x];
int start_feature = threadIdx.y * 4;
float2 fr1 = rhs_shmem2[(start_feature >> 1) + 32 * ((koff % 4) / 2)][koff / 4 + (koff % 2) * 4];
float2 fr2 = rhs_shmem2[(start_feature >> 1) + 1 + 32 * ((koff % 4) / 2)][koff / 4 + (koff % 2) * 4];
add_vals(fl1, fl2, fr1, fr2)
}
__syncthreads();
}
#undef prefetch_lhs
#undef add_vals
Index horiz_base = threadIdx.y * 4 + base_n;
if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
for (int i = 0; i < 4; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
} else if (!CHECK_RHS_BOUNDARY) {
// CHECK LHS
if (lhs_vert + 3 < m_size) {
for (int i = 0; i < 4; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
} else if (lhs_vert + 2 < m_size) {
for (int i = 0; i < 4; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
}
} else if (lhs_vert + 1 < m_size) {
for (int i = 0; i < 4; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
}
} else if (lhs_vert < m_size) {
for (int i = 0; i < 4; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
}
}
} else if (!CHECK_LHS_BOUNDARY) {
// CHECK RHS
/*
int ncols_rem = fminf(n_size- horiz_base, 4);
for (int i = 0; i < ncols_rem; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
output(lhs_vert + 3, horiz_base + i) = results[i].w;
}*/
for (int i = 0; i < 4; i++) {
if (horiz_base + i < n_size) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
}
} else {
// CHECK both boundaries.
for (int i = 0; i < 4; i++) {
if (horiz_base + i < n_size) {
if (lhs_vert < m_size) output(lhs_vert, horiz_base + i) = results[i].x;
if (lhs_vert + 1 < m_size) output(lhs_vert + 1, horiz_base + i) = results[i].y;
if (lhs_vert + 2 < m_size) output(lhs_vert + 2, horiz_base + i) = results[i].z;
if (lhs_vert + 3 < m_size) output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
}
}
}
template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper, bool CHECK_LHS_BOUNDARY,
bool CHECK_RHS_BOUNDARY>
__device__ __forceinline__ void EigenFloatContractionKernelInternal(const LhsMapper lhs, const RhsMapper rhs,
const OutputMapper output, float2 lhs_shmem2[][32],
float2 rhs_shmem2[][8], const Index m_size,
const Index n_size, const Index k_size,
const Index base_m, const Index base_n) {
// prefetch registers
float4 lhs_pf0, lhs_pf1, lhs_pf2, lhs_pf3;
float4 rhs_pf0, rhs_pf1;
float4 results[8];
for (int i = 0; i < 8; i++) {
results[i].x = results[i].y = results[i].z = results[i].w = 0;
}
Index lhs_vert = base_m + threadIdx.x * 4 + (threadIdx.y % 4) * 32;
for (Index k = 0; k < k_size; k += 32) {
lhs_pf0 = internal::pset1<float4>(0);
lhs_pf1 = internal::pset1<float4>(0);
lhs_pf2 = internal::pset1<float4>(0);
lhs_pf3 = internal::pset1<float4>(0);
rhs_pf0 = internal::pset1<float4>(0);
rhs_pf1 = internal::pset1<float4>(0);
if (!CHECK_LHS_BOUNDARY) {
if ((threadIdx.y / 4 + k + 24) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
lhs_pf1 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 8));
lhs_pf2 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 16));
lhs_pf3 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 24));
} else if ((threadIdx.y / 4 + k + 16) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
lhs_pf1 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 8));
lhs_pf2 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 16));
} else if ((threadIdx.y / 4 + k + 8) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
lhs_pf1 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 8));
} else if ((threadIdx.y / 4 + k) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
}
} else {
// just CHECK_LHS_BOUNDARY
if (lhs_vert + 3 < m_size) {
if ((threadIdx.y / 4 + k + 24) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
lhs_pf1 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 8));
lhs_pf2 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 16));
lhs_pf3 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 24));
} else if ((threadIdx.y / 4 + k + 16) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
lhs_pf1 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 8));
lhs_pf2 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 16));
} else if ((threadIdx.y / 4 + k + 8) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
lhs_pf1 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k + 8));
} else if ((threadIdx.y / 4 + k) < k_size) {
lhs_pf0 = lhs.template loadPacket<float4, Unaligned>(lhs_vert, (threadIdx.y / 4 + k));
}
} else if (lhs_vert + 2 < m_size) {
if ((threadIdx.y / 4 + k + 24) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
lhs_pf0.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf1.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 8));
lhs_pf1.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k + 8));
lhs_pf2.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 16));
lhs_pf2.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 16));
lhs_pf2.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k + 16));
lhs_pf3.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 24));
lhs_pf3.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 24));
lhs_pf3.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k + 24));
} else if ((threadIdx.y / 4 + k + 16) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
lhs_pf0.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf1.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 8));
lhs_pf1.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k + 8));
lhs_pf2.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 16));
lhs_pf2.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 16));
lhs_pf2.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k + 16));
} else if ((threadIdx.y / 4 + k + 8) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
lhs_pf0.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf1.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 8));
lhs_pf1.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k + 8));
} else if ((threadIdx.y / 4 + k) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
lhs_pf0.z = lhs(lhs_vert + 2, (threadIdx.y / 4 + k));
}
} else if (lhs_vert + 1 < m_size) {
if ((threadIdx.y / 4 + k + 24) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf1.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 8));
lhs_pf2.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 16));
lhs_pf2.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 16));
lhs_pf3.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 24));
lhs_pf3.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 24));
} else if ((threadIdx.y / 4 + k + 16) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf1.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 8));
lhs_pf2.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 16));
lhs_pf2.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 16));
} else if ((threadIdx.y / 4 + k + 8) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf1.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k + 8));
} else if ((threadIdx.y / 4 + k) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf0.y = lhs(lhs_vert + 1, (threadIdx.y / 4 + k));
}
} else if (lhs_vert < m_size) {
if ((threadIdx.y / 4 + k + 24) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf2.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 16));
lhs_pf3.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 24));
} else if ((threadIdx.y / 4 + k + 16) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
lhs_pf2.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 16));
} else if ((threadIdx.y / 4 + k + 8) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
lhs_pf1.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k + 8));
} else if ((threadIdx.y / 4 + k) < k_size) {
lhs_pf0.x = lhs(lhs_vert + 0, (threadIdx.y / 4 + k));
}
}
}
__syncthreads();
Index rhs_vert = k + threadIdx.x * 4;
Index rhs_horiz0 = threadIdx.y * 2 + base_n;
Index rhs_horiz1 = threadIdx.y * 2 + 1 + base_n;
if (!CHECK_RHS_BOUNDARY) {
if ((rhs_vert + 3) < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0 = rhs.template loadPacket<float4, Unaligned>(rhs_vert, rhs_horiz0);
rhs_pf1 = rhs.template loadPacket<float4, Unaligned>(rhs_vert, rhs_horiz1);
} else if (rhs_vert + 2 < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
} else if (rhs_vert + 1 < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
} else if (rhs_vert < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
}
} else {
if (rhs_horiz1 < n_size) {
if ((rhs_vert + 3) < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0 = rhs.template loadPacket<float4, Unaligned>(rhs_vert, rhs_horiz0);
rhs_pf1 = rhs.template loadPacket<float4, Unaligned>(rhs_vert, rhs_horiz1);
} else if (rhs_vert + 2 < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
rhs_pf1.z = rhs(rhs_vert + 2, rhs_horiz1);
} else if (k + threadIdx.x * 4 + 1 < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
rhs_pf1.y = rhs(rhs_vert + 1, rhs_horiz1);
} else if (k + threadIdx.x * 4 < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf1.x = rhs(rhs_vert, rhs_horiz1);
}
} else if (rhs_horiz0 < n_size) {
if ((rhs_vert + 3) < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0 = rhs.template loadPacket<float4, Unaligned>(rhs_vert, rhs_horiz0);
} else if ((rhs_vert + 2) < k_size) {
// just CHECK_RHS_BOUNDARY
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
rhs_pf0.z = rhs(rhs_vert + 2, rhs_horiz0);
} else if ((rhs_vert + 1) < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
rhs_pf0.y = rhs(rhs_vert + 1, rhs_horiz0);
} else if (rhs_vert < k_size) {
rhs_pf0.x = rhs(rhs_vert, rhs_horiz0);
}
}
}
__syncthreads();
// Loaded. Do computation
// Row 0 -> times (0, 4, 8, .. 28) for features 0, 1.
// Row 1 -> times (0, 4, 8, .. 28) for features 2, 3.
// ..
// Row 31 -> times (0, 4, 8, .. 28) for features 62, 63
rhs_shmem2[threadIdx.y][threadIdx.x] = make_float2(rhs_pf0.x, rhs_pf1.x);
// Row 32 -> times (1, 5, 9, .. 29) for features 0, 1.
// Row 33 -> times (1, 5, 9, .. 29) for features 2, 3.
// ..
rhs_shmem2[threadIdx.y + 32][threadIdx.x] = make_float2(rhs_pf0.y, rhs_pf1.y);
// Row 64 -> times (2, 6, 10, .. 30) for features 0, 1.
// Row 65 -> times (2, 6, 10, .. 30) for features 2, 3.
rhs_shmem2[threadIdx.y + 64][threadIdx.x] = make_float2(rhs_pf0.z, rhs_pf1.z);
// Row 96 -> times (3, 7, 11, .. 31) for features 0, 1.
// Row 97 -> times (3, 7, 11, .. 31) for features 2, 3.
rhs_shmem2[threadIdx.y + 96][threadIdx.x] = make_float2(rhs_pf0.w, rhs_pf1.w);
// LHS.
// Row 0 (time 0) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), .. (60, 61) .. (124, 125)
// Row 1 (time 1) -> features (0, 1), (4, 5), .. (28, 29), (32, 33), .. (60, 61) .. (124, 125)
// ...
// Row 8 (time 0) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), .. (62, 63) .. (126, 127)
// Row 15 (time 7) -> features (2, 3), (6, 7), .. (30, 31), (34, 35), .. (62, 63) .. (126, 127)
#define add_vals(a_feat1, a_feat2, f1, f2, f3, f4) \
results[0].x += a_feat1.x * f1.x; \
results[1].x += a_feat1.x * f1.y; \
results[2].x += a_feat1.x * f2.x; \
results[3].x += a_feat1.x * f2.y; \
results[4].x += a_feat1.x * f3.x; \
results[5].x += a_feat1.x * f3.y; \
results[6].x += a_feat1.x * f4.x; \
results[7].x += a_feat1.x * f4.y; \
\
results[0].y += a_feat1.y * f1.x; \
results[1].y += a_feat1.y * f1.y; \
results[2].y += a_feat1.y * f2.x; \
results[3].y += a_feat1.y * f2.y; \
results[4].y += a_feat1.y * f3.x; \
results[5].y += a_feat1.y * f3.y; \
results[6].y += a_feat1.y * f4.x; \
results[7].y += a_feat1.y * f4.y; \
\
results[0].z += a_feat2.x * f1.x; \
results[1].z += a_feat2.x * f1.y; \
results[2].z += a_feat2.x * f2.x; \
results[3].z += a_feat2.x * f2.y; \
results[4].z += a_feat2.x * f3.x; \
results[5].z += a_feat2.x * f3.y; \
results[6].z += a_feat2.x * f4.x; \
results[7].z += a_feat2.x * f4.y; \
\
results[0].w += a_feat2.y * f1.x; \
results[1].w += a_feat2.y * f1.y; \
results[2].w += a_feat2.y * f2.x; \
results[3].w += a_feat2.y * f2.y; \
results[4].w += a_feat2.y * f3.x; \
results[5].w += a_feat2.y * f3.y; \
results[6].w += a_feat2.y * f4.x; \
results[7].w += a_feat2.y * f4.y;
lhs_shmem2[threadIdx.y / 4][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf0.x, lhs_pf0.y);
lhs_shmem2[threadIdx.y / 4 + 8][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf1.x, lhs_pf1.y);
lhs_shmem2[threadIdx.y / 4 + 16][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf2.x, lhs_pf2.y);
lhs_shmem2[threadIdx.y / 4 + 24][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf3.x, lhs_pf3.y);
lhs_shmem2[threadIdx.y / 4 + 32][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf0.z, lhs_pf0.w);
lhs_shmem2[threadIdx.y / 4 + 40][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf1.z, lhs_pf1.w);
lhs_shmem2[threadIdx.y / 4 + 48][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf2.z, lhs_pf2.w);
lhs_shmem2[threadIdx.y / 4 + 56][threadIdx.x + (threadIdx.y % 4) * 8] = make_float2(lhs_pf3.z, lhs_pf3.w);
__syncthreads();
// Do the multiplies.
#pragma unroll
for (int koff = 0; koff < 32; koff++) {
float2 a3 = lhs_shmem2[koff][threadIdx.x + (threadIdx.y % 4) * 8];
float2 a4 = lhs_shmem2[koff + 32][threadIdx.x + (threadIdx.y % 4) * 8];
// first feature is at (threadIdx.y/4) * 8 last is at start + 8.
int start_feature = (threadIdx.y / 4) * 8;
float2 br1 = rhs_shmem2[start_feature / 2 + (koff % 4) * 32][koff / 4];
float2 br2 = rhs_shmem2[start_feature / 2 + 1 + (koff % 4) * 32][koff / 4];
float2 br3 = rhs_shmem2[start_feature / 2 + 2 + (koff % 4) * 32][koff / 4];
float2 br4 = rhs_shmem2[start_feature / 2 + 3 + (koff % 4) * 32][koff / 4];
add_vals(a3, a4, br1, br2, br3, br4)
}
__syncthreads();
} // end loop over k
#undef add_vals
__syncthreads();
Index horiz_base = (threadIdx.y / 4) * 8 + base_n;
if (!CHECK_LHS_BOUNDARY && !CHECK_RHS_BOUNDARY) {
for (int i = 0; i < 8; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
} else if (!CHECK_RHS_BOUNDARY) {
if (lhs_vert + 3 < m_size) {
for (int i = 0; i < 8; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
} else if (lhs_vert + 2 < m_size) {
for (int i = 0; i < 8; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
}
} else if (lhs_vert + 1 < m_size) {
for (int i = 0; i < 8; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
}
} else if (lhs_vert < m_size) {
for (int i = 0; i < 8; i++) {
output(lhs_vert, horiz_base + i) = results[i].x;
}
}
} else if (!CHECK_LHS_BOUNDARY) {
// CHECK BOUNDARY_B
for (int i = 0; i < 8; i++) {
if (horiz_base + i < n_size) {
output(lhs_vert, horiz_base + i) = results[i].x;
output(lhs_vert + 1, horiz_base + i) = results[i].y;
output(lhs_vert + 2, horiz_base + i) = results[i].z;
output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
}
} else {
// CHECK both boundaries.
for (int i = 0; i < 8; i++) {
if (horiz_base + i < n_size) {
if (lhs_vert < m_size) output(lhs_vert, horiz_base + i) = results[i].x;
if (lhs_vert + 1 < m_size) output(lhs_vert + 1, horiz_base + i) = results[i].y;
if (lhs_vert + 2 < m_size) output(lhs_vert + 2, horiz_base + i) = results[i].z;
if (lhs_vert + 3 < m_size) output(lhs_vert + 3, horiz_base + i) = results[i].w;
}
}
}
}
template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper>
__global__ void
#if defined(EIGEN_HIPCC)
__launch_bounds__(256, 1)
#else
__launch_bounds__(256)
#endif
EigenFloatContractionKernel(const LhsMapper lhs, const RhsMapper rhs, const OutputMapper output, const Index m_size,
const Index n_size, const Index k_size) {
__shared__ float2 lhs_shmem[64 * 32];
__shared__ float2 rhs_shmem[128 * 8];
typedef float2 LHS_MEM[64][32];
typedef float2 RHS_MEM[128][8];
const Index m_block_idx = blockIdx.x;
const Index n_block_idx = blockIdx.y;
const Index base_m = 128 * m_block_idx;
const Index base_n = 64 * n_block_idx;
bool check_rhs = (base_n + 63) >= n_size;
bool check_lhs128 = (base_m + 127) >= m_size;
if (!check_rhs) {
if (!check_lhs128) {
// >= 128 rows left
EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(
lhs, rhs, output, *((LHS_MEM*)lhs_shmem), *((RHS_MEM*)rhs_shmem), m_size, n_size, k_size, base_m, base_n);
} else {
EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(
lhs, rhs, output, *((LHS_MEM*)lhs_shmem), *((RHS_MEM*)rhs_shmem), m_size, n_size, k_size, base_m, base_n);
}
} else {
if (!check_lhs128) {
// >= 128 rows left
EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(
lhs, rhs, output, *((LHS_MEM*)lhs_shmem), *((RHS_MEM*)rhs_shmem), m_size, n_size, k_size, base_m, base_n);
} else {
EigenFloatContractionKernelInternal<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(
lhs, rhs, output, *((LHS_MEM*)lhs_shmem), *((RHS_MEM*)rhs_shmem), m_size, n_size, k_size, base_m, base_n);
}
}
}
template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper>
__global__ void
#if defined(EIGEN_HIPCC)
__launch_bounds__(256, 1)
#else
__launch_bounds__(256)
#endif
EigenFloatContractionKernel16x16(const LhsMapper lhs, const RhsMapper rhs, const OutputMapper output,
const Index m_size, const Index n_size, const Index k_size) {
__shared__ float2 lhs_shmem[32][16];
__shared__ float2 rhs_shmem[64][8];
const Index m_block_idx = blockIdx.x;
const Index n_block_idx = blockIdx.y;
const Index base_m = 64 * m_block_idx;
const Index base_n = 64 * n_block_idx;
if (base_m + 63 < m_size) {
if (base_n + 63 < n_size) {
EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, false>(
lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
} else {
EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, false, true>(
lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
}
} else {
if (base_n + 63 < n_size) {
EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, false>(
lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
} else {
EigenFloatContractionKernelInternal16x16<Index, LhsMapper, RhsMapper, OutputMapper, true, true>(
lhs, rhs, output, lhs_shmem, rhs_shmem, m_size, n_size, k_size, base_m, base_n);
}
}
}
template <typename Indices, typename LeftArgType, typename RightArgType, typename OutputKernelType>
struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice>
: public TensorContractionEvaluatorBase<TensorEvaluator<
const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, GpuDevice> > {
typedef GpuDevice Device;
typedef TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType>, Device> Self;
typedef TensorContractionEvaluatorBase<Self> Base;
typedef TensorContractionOp<Indices, LeftArgType, RightArgType, OutputKernelType> XprType;
typedef std::remove_const_t<typename XprType::Scalar> Scalar;
typedef typename XprType::Index Index;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename PacketType<CoeffReturnType, GpuDevice>::type PacketReturnType;
static constexpr int Layout = TensorEvaluator<LeftArgType, Device>::Layout;
// Most of the code is assuming that both input tensors are ColMajor. If the
// inputs are RowMajor, we will "cheat" by swapping the LHS and RHS:
// If we want to compute A * B = C, where A is LHS and B is RHS, the code
// will pretend B is LHS and A is RHS.
typedef std::conditional_t<Layout == static_cast<int>(ColMajor), LeftArgType, RightArgType> EvalLeftArgType;
typedef std::conditional_t<Layout == static_cast<int>(ColMajor), RightArgType, LeftArgType> EvalRightArgType;
static constexpr int LDims =
internal::array_size<typename TensorEvaluator<EvalLeftArgType, Device>::Dimensions>::value;
static constexpr int RDims =
internal::array_size<typename TensorEvaluator<EvalRightArgType, Device>::Dimensions>::value;
static constexpr int ContractDims = internal::array_size<Indices>::value;
typedef array<Index, LDims> left_dim_mapper_t;
typedef array<Index, RDims> right_dim_mapper_t;
typedef array<Index, ContractDims> contract_t;
typedef array<Index, LDims - ContractDims> left_nocontract_t;
typedef array<Index, RDims - ContractDims> right_nocontract_t;
static constexpr int NumDims = LDims + RDims - 2 * ContractDims;
typedef DSizes<Index, NumDims> Dimensions;
// typedefs needed in evalTo
typedef std::remove_const_t<typename EvalLeftArgType::Scalar> LhsScalar;
typedef std::remove_const_t<typename EvalRightArgType::Scalar> RhsScalar;
typedef TensorEvaluator<EvalLeftArgType, Device> LeftEvaluator;
typedef TensorEvaluator<EvalRightArgType, Device> RightEvaluator;
typedef typename LeftEvaluator::Dimensions LeftDimensions;
typedef typename RightEvaluator::Dimensions RightDimensions;
TensorEvaluator(const XprType& op, const Device& device) : Base(op, device) {
EIGEN_STATIC_ASSERT((internal::is_same<OutputKernelType, const NoOpOutputKernel>::value),
GPU_TENSOR_CONTRACTION_DOES_NOT_SUPPORT_OUTPUT_KERNELS);
}
// We need to redefine this method to make nvcc happy
EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data) {
this->m_leftImpl.evalSubExprsIfNeeded(NULL);
this->m_rightImpl.evalSubExprsIfNeeded(NULL);
if (data) {
evalTo(data);
return false;
} else {
this->m_result = static_cast<Scalar*>(this->m_device.allocate(this->dimensions().TotalSize() * sizeof(Scalar)));
evalTo(this->m_result);
return true;
}
}
void evalTo(Scalar* buffer) const {
if (this->m_lhs_inner_dim_contiguous) {
if (this->m_rhs_inner_dim_contiguous) {
if (this->m_rhs_inner_dim_reordered) {
evalTyped<true, true, true, Unaligned>(buffer);
} else {
evalTyped<true, true, false, Unaligned>(buffer);
}
} else {
if (this->m_rhs_inner_dim_reordered) {
evalTyped<true, false, true, Unaligned>(buffer);
} else {
evalTyped<true, false, false, Unaligned>(buffer);
}
}
} else {
if (this->m_rhs_inner_dim_contiguous) {
if (this->m_rhs_inner_dim_reordered) {
evalTyped<false, true, true, Unaligned>(buffer);
} else {
evalTyped<false, true, false, Unaligned>(buffer);
}
} else {
if (this->m_rhs_inner_dim_reordered) {
evalTyped<false, false, true, Unaligned>(buffer);
} else {
evalTyped<false, false, false, Unaligned>(buffer);
}
}
}
}
template <typename LhsScalar, typename RhsScalar, typename Index, typename LhsMapper, typename RhsMapper,
typename OutputMapper>
struct LaunchKernels {
static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k,
const GpuDevice& device) {
const Index m_blocks = (m + 63) / 64;
const Index n_blocks = (n + 63) / 64;
const dim3 num_blocks(m_blocks, n_blocks, 1);
const dim3 block_size(8, 8, 8);
LAUNCH_GPU_KERNEL((EigenContractionKernel<Scalar, Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks,
block_size, 0, device, lhs, rhs, output, m, n, k);
}
};
template <typename Index, typename LhsMapper, typename RhsMapper, typename OutputMapper>
struct LaunchKernels<float, float, Index, LhsMapper, RhsMapper, OutputMapper> {
static void Run(const LhsMapper& lhs, const RhsMapper& rhs, const OutputMapper& output, Index m, Index n, Index k,
const GpuDevice& device) {
if (m < 768 || n < 768) {
const Index m_blocks = (m + 63) / 64;
const Index n_blocks = (n + 63) / 64;
const dim3 num_blocks(m_blocks, n_blocks, 1);
const dim3 block_size(16, 16, 1);
LAUNCH_GPU_KERNEL((EigenFloatContractionKernel16x16<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks,
block_size, 0, device, lhs, rhs, output, m, n, k);
} else {
const Index m_blocks = (m + 127) / 128;
const Index n_blocks = (n + 63) / 64;
const dim3 num_blocks(m_blocks, n_blocks, 1);
const dim3 block_size(8, 32, 1);
LAUNCH_GPU_KERNEL((EigenFloatContractionKernel<Index, LhsMapper, RhsMapper, OutputMapper>), num_blocks,
block_size, 0, device, lhs, rhs, output, m, n, k);
}
}
};
template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
void evalTyped(Scalar* buffer) const {
// columns in left side, rows in right side
const Index k = this->m_k_size;
EIGEN_UNUSED_VARIABLE(k)
// rows in left side
const Index m = this->m_i_size;
// columns in right side
const Index n = this->m_j_size;
// zero out the result buffer (which must be of size at least m * n * sizeof(Scalar))
this->m_device.fill(buffer, buffer + m * n, Scalar(0));
typedef internal::TensorContractionInputMapper<LhsScalar, Index, internal::Lhs, LeftEvaluator, left_nocontract_t,
contract_t, 4, lhs_inner_dim_contiguous, false, Unaligned>
LhsMapper;
typedef internal::TensorContractionInputMapper<RhsScalar, Index, internal::Rhs, RightEvaluator, right_nocontract_t,
contract_t, 4, rhs_inner_dim_contiguous, rhs_inner_dim_reordered,
Unaligned>
RhsMapper;
typedef internal::blas_data_mapper<Scalar, Index, ColMajor> OutputMapper;
// initialize data mappers
LhsMapper lhs(this->m_leftImpl, this->m_left_nocontract_strides, this->m_i_strides,
this->m_left_contracting_strides, this->m_k_strides);
RhsMapper rhs(this->m_rightImpl, this->m_right_nocontract_strides, this->m_j_strides,
this->m_right_contracting_strides, this->m_k_strides);
OutputMapper output(buffer, m);
#if defined(EIGEN_USE_HIP)
setGpuSharedMemConfig(hipSharedMemBankSizeEightByte);
#else
setGpuSharedMemConfig(cudaSharedMemBankSizeEightByte);
#endif
LaunchKernels<LhsScalar, RhsScalar, Index, LhsMapper, RhsMapper, OutputMapper>::Run(lhs, rhs, output, m, n, k,
this->m_device);
}
};
} // end namespace Eigen
#endif // EIGEN_USE_GPU and EIGEN_GPUCC
#endif // EIGEN_CXX11_TENSOR_TENSOR_CONTRACTION_GPU_H