forked from kokkos/kokkos
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Kokkos_SYCL_ParallelReduce_Range.hpp
377 lines (335 loc) · 16 KB
/
Kokkos_SYCL_ParallelReduce_Range.hpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
//@HEADER
// ************************************************************************
//
// Kokkos v. 4.0
// Copyright (2022) National Technology & Engineering
// Solutions of Sandia, LLC (NTESS).
//
// Under the terms of Contract DE-NA0003525 with NTESS,
// the U.S. Government retains certain rights in this software.
//
// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions.
// See https://kokkos.org/LICENSE for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//@HEADER
#ifndef KOKKOS_SYCL_PARALLEL_REDUCE_RANGE_HPP
#define KOKKOS_SYCL_PARALLEL_REDUCE_RANGE_HPP
#include <Kokkos_Macros.hpp>
#include <Kokkos_BitManipulation.hpp>
#include <Kokkos_Parallel_Reduce.hpp>
#include <SYCL/Kokkos_SYCL_WorkgroupReduction.hpp>
#include <vector>
template <class CombinedFunctorReducerType, class... Traits>
class Kokkos::Impl::ParallelReduce<CombinedFunctorReducerType,
Kokkos::RangePolicy<Traits...>,
Kokkos::Experimental::SYCL> {
public:
using Policy = Kokkos::RangePolicy<Traits...>;
using FunctorType = typename CombinedFunctorReducerType::functor_type;
using ReducerType = typename CombinedFunctorReducerType::reducer_type;
private:
using value_type = typename ReducerType::value_type;
using pointer_type = typename ReducerType::pointer_type;
using reference_type = typename ReducerType::reference_type;
using WorkTag = typename Policy::work_tag;
public:
// V - View
template <typename View>
ParallelReduce(const CombinedFunctorReducerType& f, const Policy& p,
const View& v)
: m_functor_reducer(f),
m_policy(p),
m_result_ptr(v.data()),
m_result_ptr_device_accessible(
MemorySpaceAccess<Kokkos::Experimental::SYCLDeviceUSMSpace,
typename View::memory_space>::accessible),
m_scratch_buffers_lock(
p.space().impl_internal_space_instance()->m_mutexScratchSpace) {}
private:
template <typename PolicyType, typename CombinedFunctorReducerWrapper>
sycl::event sycl_direct_launch(
const PolicyType& policy,
const CombinedFunctorReducerWrapper& functor_reducer_wrapper,
const sycl::event& memcpy_event) const {
// Convenience references
const Kokkos::Experimental::SYCL& space = policy.space();
Kokkos::Experimental::Impl::SYCLInternal& instance =
*space.impl_internal_space_instance();
sycl::queue& q = space.sycl_queue();
std::size_t size = policy.end() - policy.begin();
const unsigned int value_count =
m_functor_reducer.get_reducer().value_count();
sycl::ext::intel::device_ptr<value_type> results_ptr = nullptr;
auto host_result_ptr =
(m_result_ptr && !m_result_ptr_device_accessible)
? static_cast<sycl::ext::intel::host_ptr<value_type>>(
instance.scratch_host(sizeof(value_type) * value_count))
: nullptr;
auto device_accessible_result_ptr =
m_result_ptr_device_accessible
? static_cast<sycl::global_ptr<value_type>>(m_result_ptr)
: static_cast<sycl::global_ptr<value_type>>(host_result_ptr);
sycl::event last_reduction_event;
desul::ensure_sycl_lock_arrays_on_device(q);
// If size<=1 we only call init(), the functor and possibly final once
// working with the global scratch memory but don't copy back to
// m_result_ptr yet.
if (size <= 1) {
results_ptr = static_cast<sycl::ext::intel::device_ptr<value_type>>(
instance.scratch_space(sizeof(value_type) * value_count));
auto parallel_reduce_event = q.submit([&](sycl::handler& cgh) {
const auto begin = policy.begin();
#ifndef KOKKOS_IMPL_SYCL_USE_IN_ORDER_QUEUES
cgh.depends_on(memcpy_event);
#else
(void)memcpy_event;
#endif
cgh.single_task([=]() {
const CombinedFunctorReducerType& functor_reducer =
functor_reducer_wrapper.get_functor();
const FunctorType& functor = functor_reducer.get_functor();
const ReducerType& reducer = functor_reducer.get_reducer();
reference_type update = reducer.init(results_ptr);
if (size == 1) {
if constexpr (std::is_void_v<WorkTag>)
functor(begin, update);
else
functor(WorkTag(), begin, update);
}
reducer.final(results_ptr);
if (device_accessible_result_ptr != nullptr)
reducer.copy(device_accessible_result_ptr.get(), results_ptr.get());
});
});
#ifndef KOKKOS_IMPL_SYCL_USE_IN_ORDER_QUEUES
q.ext_oneapi_submit_barrier(
std::vector<sycl::event>{parallel_reduce_event});
#endif
last_reduction_event = parallel_reduce_event;
} else {
// Otherwise (when size > 1), we perform a reduction on the values in all
// workgroups separately, write the workgroup results back to global
// memory and recurse until only one workgroup does the reduction and thus
// gets the final value.
auto scratch_flags =
static_cast<sycl::ext::intel::device_ptr<unsigned int>>(
instance.scratch_flags(sizeof(unsigned int)));
auto reduction_lambda_factory =
[&](sycl::local_accessor<value_type> local_mem,
sycl::local_accessor<unsigned int> num_teams_done,
sycl::ext::intel::device_ptr<value_type> results_ptr,
int values_per_thread) {
const auto begin = policy.begin();
auto lambda = [=](sycl::nd_item<1> item) {
const auto n_wgroups = item.get_group_range()[0];
const auto wgroup_size = item.get_local_range()[0];
const auto local_id = item.get_local_linear_id();
const auto global_id =
wgroup_size * item.get_group_linear_id() * values_per_thread +
local_id;
const CombinedFunctorReducerType& functor_reducer =
functor_reducer_wrapper.get_functor();
const FunctorType& functor = functor_reducer.get_functor();
const ReducerType& reducer = functor_reducer.get_reducer();
using index_type = typename Policy::index_type;
const auto upper_bound = std::min<index_type>(
global_id + values_per_thread * wgroup_size, size);
if constexpr (!SYCLReduction::use_shuffle_based_algorithm<
ReducerType>) {
reference_type update =
reducer.init(&local_mem[local_id * value_count]);
for (index_type id = global_id; id < upper_bound;
id += wgroup_size) {
if constexpr (std::is_void_v<WorkTag>)
functor(id + begin, update);
else
functor(WorkTag(), id + begin, update);
}
item.barrier(sycl::access::fence_space::local_space);
SYCLReduction::workgroup_reduction<>(
item, local_mem, results_ptr, device_accessible_result_ptr,
value_count, reducer, false, std::min(size, wgroup_size));
if (local_id == 0) {
sycl::atomic_ref<unsigned, sycl::memory_order::acq_rel,
sycl::memory_scope::device,
sycl::access::address_space::global_space>
scratch_flags_ref(*scratch_flags);
num_teams_done[0] = ++scratch_flags_ref;
}
item.barrier(sycl::access::fence_space::local_space);
if (num_teams_done[0] == n_wgroups) {
if (local_id == 0) *scratch_flags = 0;
if (local_id >= n_wgroups)
reducer.init(&local_mem[local_id * value_count]);
else {
reducer.copy(&local_mem[local_id * value_count],
&results_ptr[local_id * value_count]);
for (unsigned int id = local_id + wgroup_size;
id < n_wgroups; id += wgroup_size) {
reducer.join(&local_mem[local_id * value_count],
&results_ptr[id * value_count]);
}
}
SYCLReduction::workgroup_reduction<>(
item, local_mem, results_ptr,
device_accessible_result_ptr, value_count, reducer, true,
std::min(n_wgroups, wgroup_size));
}
} else {
value_type local_value;
reference_type update = reducer.init(&local_value);
for (index_type id = global_id; id < upper_bound;
id += wgroup_size) {
if constexpr (std::is_void_v<WorkTag>)
functor(id + begin, update);
else
functor(WorkTag(), id + begin, update);
}
SYCLReduction::workgroup_reduction<>(
item, local_mem, local_value, results_ptr,
device_accessible_result_ptr, reducer, false,
std::min(size, wgroup_size));
if (local_id == 0) {
sycl::atomic_ref<unsigned, sycl::memory_order::acq_rel,
sycl::memory_scope::device,
sycl::access::address_space::global_space>
scratch_flags_ref(*scratch_flags);
num_teams_done[0] = ++scratch_flags_ref;
}
item.barrier(sycl::access::fence_space::local_space);
if (num_teams_done[0] == n_wgroups) {
if (local_id == 0) *scratch_flags = 0;
if (local_id >= n_wgroups)
reducer.init(&local_value);
else {
local_value = results_ptr[local_id];
for (unsigned int id = local_id + wgroup_size;
id < n_wgroups; id += wgroup_size) {
reducer.join(&local_value, &results_ptr[id]);
}
}
SYCLReduction::workgroup_reduction<>(
item, local_mem, local_value, results_ptr,
device_accessible_result_ptr, reducer, true,
std::min(n_wgroups, wgroup_size));
}
}
};
return lambda;
};
auto parallel_reduce_event = q.submit([&](sycl::handler& cgh) {
sycl::local_accessor<unsigned int> num_teams_done(1, cgh);
auto dummy_reduction_lambda =
reduction_lambda_factory({1, cgh}, num_teams_done, nullptr, 1);
static sycl::kernel kernel = [&] {
sycl::kernel_id functor_kernel_id =
sycl::get_kernel_id<decltype(dummy_reduction_lambda)>();
auto kernel_bundle =
sycl::get_kernel_bundle<sycl::bundle_state::executable>(
q.get_context(), std::vector{functor_kernel_id});
return kernel_bundle.get_kernel(functor_kernel_id);
}();
auto multiple = kernel.get_info<sycl::info::kernel_device_specific::
preferred_work_group_size_multiple>(
q.get_device());
// FIXME_SYCL The code below queries the kernel for the maximum subgroup
// size but it turns out that this is not accurate and choosing a larger
// subgroup size gives better peformance (and is what the oneAPI
// reduction algorithm does).
#ifndef KOKKOS_ARCH_INTEL_GPU
auto max =
kernel
.get_info<sycl::info::kernel_device_specific::work_group_size>(
q.get_device());
#else
auto max =
q.get_device().get_info<sycl::info::device::max_work_group_size>();
#endif
auto max_local_memory =
q.get_device().get_info<sycl::info::device::local_mem_size>();
// The workgroup size is computed as the minimum of
// - the smallest power of two not less than the total work size
// - the largest power of two not exceeding the largest multiple of the
// recommended workgroup size not exceeding the maximum workgroup size
// - the largest power of two such that we don't use more than 99% (as a
// safe-guard) of the available local memory.
const auto wgroup_size = std::min(
{Kokkos::bit_ceil(size),
Kokkos::bit_floor(static_cast<size_t>(max / multiple) * multiple),
Kokkos::bit_floor(static_cast<size_t>(max_local_memory * .99) /
(sizeof(value_type) * value_count))});
// FIXME_SYCL Find a better way to determine a good limit for the
// maximum number of work groups, also see
// https://github.com/intel/llvm/blob/756ba2616111235bba073e481b7f1c8004b34ee6/sycl/source/detail/reduction.cpp#L51-L62
size_t max_work_groups =
2 *
q.get_device().get_info<sycl::info::device::max_compute_units>();
int values_per_thread = 1;
size_t n_wgroups = (size + wgroup_size - 1) / wgroup_size;
while (n_wgroups > max_work_groups) {
values_per_thread *= 2;
n_wgroups = ((size + values_per_thread - 1) / values_per_thread +
wgroup_size - 1) /
wgroup_size;
}
results_ptr = static_cast<sycl::ext::intel::device_ptr<value_type>>(
instance.scratch_space(sizeof(value_type) * value_count *
n_wgroups));
sycl::local_accessor<value_type> local_mem(
sycl::range<1>(wgroup_size) * value_count, cgh);
#ifndef KOKKOS_IMPL_SYCL_USE_IN_ORDER_QUEUES
cgh.depends_on(memcpy_event);
#else
(void)memcpy_event;
#endif
auto reduction_lambda = reduction_lambda_factory(
local_mem, num_teams_done, results_ptr, values_per_thread);
cgh.parallel_for(
sycl::nd_range<1>(n_wgroups * wgroup_size, wgroup_size),
reduction_lambda);
});
#ifndef KOKKOS_IMPL_SYCL_USE_IN_ORDER_QUEUES
q.ext_oneapi_submit_barrier(
std::vector<sycl::event>{parallel_reduce_event});
#endif
last_reduction_event = parallel_reduce_event;
}
// At this point, the reduced value is written to the entry in results_ptr
// and all that is left is to copy it back to the given result pointer if
// necessary.
// Using DeepCopy instead of fence+memcpy turned out to be up to 2x slower.
if (host_result_ptr) {
space.fence(
"Kokkos::Impl::ParallelReduce<SYCL, RangePolicy>::execute: result "
"not device-accessible");
std::memcpy(m_result_ptr, host_result_ptr,
sizeof(*m_result_ptr) * value_count);
}
return last_reduction_event;
}
public:
void execute() const {
Kokkos::Experimental::Impl::SYCLInternal& instance =
*m_policy.space().impl_internal_space_instance();
using IndirectKernelMem =
Kokkos::Experimental::Impl::SYCLInternal::IndirectKernelMem;
IndirectKernelMem& indirectKernelMem = instance.get_indirect_kernel_mem();
auto functor_reducer_wrapper =
Experimental::Impl::make_sycl_function_wrapper(m_functor_reducer,
indirectKernelMem);
sycl::event event =
sycl_direct_launch(m_policy, functor_reducer_wrapper,
functor_reducer_wrapper.get_copy_event());
functor_reducer_wrapper.register_event(event);
}
private:
const CombinedFunctorReducerType m_functor_reducer;
const Policy m_policy;
const pointer_type m_result_ptr;
const bool m_result_ptr_device_accessible;
// Only let one ParallelReduce instance at a time use the host scratch memory.
// The constructor acquires the mutex which is released in the destructor.
std::scoped_lock<std::mutex> m_scratch_buffers_lock;
};
#endif /* KOKKOS_SYCL_PARALLEL_REDUCE_RANGE_HPP */