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
|
/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
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.
==============================================================================*/
#include "tensorflow/compiler/xla/service/gpu/buffer_comparator.h"
#include <cmath>
#include "absl/strings/str_replace.h"
#include "tensorflow/compiler/xla/service/hlo_parser.h"
#include "tensorflow/compiler/xla/status_macros.h"
namespace xla {
namespace gpu {
static constexpr float kTolerance = 0.1f;
static string GetCompHloText(size_t num_elements) {
// Implements the textual format of the comparison routine, as it's more
// readable.
static constexpr char kF16CompHloText[] = R"(
HloModule CompareF16
MaxF32 {
%lhs = f32[] parameter(0)
%rhs = f32[] parameter(1)
ROOT %max = f32[] maximum(%lhs, %rhs)
}
Canonicalize (aparam: f16[SIZE]) -> f32[SIZE] {
%min_constant = f32[] constant(-65505)
%max_constant = f32[] constant(65505)
%large_constant = f32[] constant(1048576)
%min_values = f32[SIZE] broadcast(%min_constant), dimensions={}
%max_values = f32[SIZE] broadcast(%max_constant), dimensions={}
%large_values = f32[SIZE] broadcast(%large_constant), dimensions={}
%a = f16[SIZE] parameter(0)
%converted = f32[SIZE] convert(%a)
%clamped = f32[SIZE] clamp(%min_values, %converted, %max_values)
// Since the clamp() above already took care of infs, only NaNs will cause
// is-finite() to return false.
%is_finite = pred[SIZE] is-finite(%clamped)
ROOT %result = f32[SIZE] select(%is_finite, %clamped, %large_values)
}
ENTRY MaxDifference {
%one_constant = f32[] constant(1.0)
%zero_constant = f32[] constant(0.0)
%ones = f32[SIZE] broadcast(%one_constant), dimensions={}
%lhs = f16[SIZE] parameter(0)
%rhs = f16[SIZE] parameter(1)
%lhs_canonical = f32[SIZE] call(%lhs), to_apply=Canonicalize
%rhs_canonical = f32[SIZE] call(%rhs), to_apply=Canonicalize
%sub = f32[SIZE] subtract(%lhs_canonical, %rhs_canonical)
%sub_abs = f32[SIZE] abs(%sub)
%lhs_abs = f32[SIZE] abs(%lhs_canonical)
%rhs_abs = f32[SIZE] abs(%rhs_canonical)
%max = f32[SIZE] maximum(%lhs_abs, %rhs_abs)
%denominator = f32[SIZE] add(%max, %ones)
%error = f32[SIZE] divide(%sub_abs, %denominator)
ROOT %max_diff = f32[] reduce(%error, %zero_constant), dimensions={0}, to_apply=MaxF32
})";
return absl::StrReplaceAll(kF16CompHloText,
{{"SIZE", absl::StrCat(num_elements)}});
}
StatusOr<F16BufferComparator> F16BufferComparator::Create(
se::DeviceMemory<Eigen::half> ref_buffer, Compiler* compiler,
DeviceMemoryAllocator* allocator, se::Stream* stream) {
auto stream_exec = stream->parent();
int64 num_elements = ref_buffer.ElementCount();
// One may consider using hlo_runner to do all the compilation and execution.
// However, as of the time hlo_runner doesn't support injection for Compiler*,
// Stream*, or even the allocator. We may revisit this in the future if it
// proves to be a maintenance burden.
TF_ASSIGN_OR_RETURN(
auto exec, ([&]() -> StatusOr<std::unique_ptr<Executable>> {
HloModuleConfig config;
DebugOptions debug_options;
debug_options.set_xla_backend_optimization_level(2);
config.set_debug_options(debug_options);
TF_ASSIGN_OR_RETURN(
auto module, ParseHloString(GetCompHloText(num_elements), config));
TF_ASSIGN_OR_RETURN(
module,
compiler->RunHloPasses(std::move(module), stream_exec, nullptr));
return compiler->RunBackend(std::move(module), stream_exec, nullptr);
}()));
TF_ASSIGN_OR_RETURN(
auto shaped_buffer, ([&]() -> StatusOr<ScopedShapedBuffer> {
auto device_ordinal = stream_exec->device_ordinal();
TF_ASSIGN_OR_RETURN(
auto owning_buffer,
allocator->Allocate(device_ordinal, ref_buffer.size()));
se::DeviceMemory<Eigen::half> buffer(
owning_buffer.AsDeviceMemoryBase());
stream->ThenMemcpy(&buffer, ref_buffer, ref_buffer.size());
Shape shape = ShapeUtil::MakeShape(xla::F16, {num_elements});
ScopedShapedBuffer ret(shape, shape, allocator, device_ordinal);
ret.set_buffer(std::move(owning_buffer), {});
return std::move(ret);
}()));
return F16BufferComparator(stream, allocator, std::move(exec),
std::move(shaped_buffer));
}
StatusOr<bool> F16BufferComparator::CompareEqualImpl(
se::DeviceMemory<Eigen::half> test_buffer) {
if (ref_buffer_.root_buffer().size() != test_buffer.size()) {
return InternalError("Mismatched buffer size: %d vs %d",
ref_buffer_.root_buffer().size(), test_buffer.size());
}
int64 num_elements = test_buffer.ElementCount();
TF_ASSIGN_OR_RETURN(
auto result_buffer, ([&]() -> StatusOr<ScopedShapedBuffer> {
auto stream_exec = stream_->parent();
Shape shape = ShapeUtil::MakeShape(xla::F16, {num_elements});
auto device_ordinal = stream_exec->device_ordinal();
ShapedBuffer shaped_test_buffer(shape, shape, stream_exec->platform(),
device_ordinal);
shaped_test_buffer.set_buffer(test_buffer, {});
ExecutableRunOptions run_options;
run_options.set_device_ordinal(stream_exec->device_ordinal());
run_options.set_stream(stream_);
run_options.set_allocator(allocator_);
ServiceExecutableRunOptions service_run_options(run_options);
return exec_->ExecuteOnStream(
&service_run_options, {&ref_buffer_, &shaped_test_buffer}, nullptr);
}()));
float result;
CHECK(result_buffer.root_buffer().size() == sizeof(result));
stream_->ThenMemcpy(&result, result_buffer.root_buffer(), sizeof(result));
TF_RETURN_IF_ERROR(stream_->BlockHostUntilDone());
return result < kTolerance;
}
StatusOr<bool> F16BufferComparator::CompareEqual(
se::DeviceMemory<Eigen::half> test_buffer) {
TF_ASSIGN_OR_RETURN(auto result, CompareEqualImpl(test_buffer));
if (result) {
return true;
}
// Host side code that does the same thing, but report some of the
// differences as well.
int64 n = test_buffer.ElementCount();
std::vector<half> host_ref_buffer(n), host_test_buffer(n);
stream_->ThenMemcpy(host_ref_buffer.data(), ref_buffer_.root_buffer(),
ref_buffer_.root_buffer().size());
stream_->ThenMemcpy(host_test_buffer.data(), test_buffer, test_buffer.size());
TF_RETURN_IF_ERROR(stream_->BlockHostUntilDone());
const auto canonicalize = [](float a) -> float {
constexpr float kBigNumer = 1048576.;
constexpr float kMaxFp16Value = 65504.;
if (std::isnan(a)) {
return kBigNumer;
}
if (std::isinf(a)) {
if (a < 0) {
return -(kMaxFp16Value + 1);
}
return kMaxFp16Value + 1;
}
return a;
};
int differences_seen = 0;
for (int64 i = 0; i < n && differences_seen < 10; i++) {
float original_ref = static_cast<float>(host_ref_buffer[i]);
float original_test = static_cast<float>(host_test_buffer[i]);
float ref = canonicalize(original_ref);
float test = canonicalize(original_test);
if (!(std::abs(ref - test) / (std::max(std::abs(ref), std::abs(test)) + 1) <
kTolerance)) {
differences_seen++;
LOG(ERROR) << "Difference at " << i << ": " << original_ref << " vs "
<< original_test;
}
}
return false;
}
} // namespace gpu
} // namespace xla
|
