-
Notifications
You must be signed in to change notification settings - Fork 224
/
mha_ck_fa_v2_solver_forward.cpp
268 lines (228 loc) · 10 KB
/
mha_ck_fa_v2_solver_forward.cpp
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
/*******************************************************************************
*
* MIT License
*
* Copyright (c) 2024 Advanced Micro Devices, Inc.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*******************************************************************************/
#include <miopen/mha/solvers.hpp>
#include <miopen/mha/invoke_params.hpp>
#include <miopen/datatype.hpp>
#include <miopen/hipoc_kernel.hpp>
#if MIOPEN_USE_COMPOSABLEKERNEL
#include "ck_tile/ops/fmha.hpp"
#include "ck_tile/ops/fmha_fwd.hpp"
#include "ck/stream_config.hpp"
#endif
MIOPEN_DECLARE_ENV_VAR_BOOL(MIOPEN_DEBUG_FA_CK_V2_FWD)
namespace miopen {
namespace solver {
namespace mha {
#if MIOPEN_USE_COMPOSABLEKERNEL
static std::string Convert(miopenDataType_t dataType)
{
switch(dataType)
{
case miopenHalf: {
return "fp16";
}
case miopenBFloat16: {
return "bfp16";
}
default: {
MIOPEN_THROW("Unsupported datatype provided");
}
}
}
#endif
bool MhaCKFlashAttentionV2Forward::IsApplicable(
[[maybe_unused]] const ExecutionContext& context,
const miopen::mha::ProblemDescription& problem) const
{
#if MIOPEN_USE_COMPOSABLEKERNEL
if(!problem.IsForward())
{
return false;
}
auto deviceName = context.GetStream().GetDeviceName();
if(!StartsWith(deviceName, "gfx94") && deviceName != "gfx90a")
{
return false;
}
const auto& descsFwd = problem.GetDescsForward();
auto [N_k, H_k, S_k, D_k] = miopen::tien<4>(descsFwd.kDesc.GetLengths());
auto [N_stride_k, H_stride_k, S_stride_k, D_stride_k] =
miopen::tien<4>(descsFwd.kDesc.GetStrides());
auto [N_q, H_q, S_q, D_q] = miopen::tien<4>(descsFwd.qDesc.GetLengths());
auto [N_stride_q, H_stride_q, S_stride_q, D_stride_q] =
miopen::tien<4>(descsFwd.qDesc.GetStrides());
auto [N_stride_v, H_stride_v, S_stride_v, D_stride_v] =
miopen::tien<4>(descsFwd.vDesc.GetStrides());
auto [N_stride_o, H_stride_o, S_stride_o, D_stride_o] =
miopen::tien<4>(descsFwd.oDesc.GetStrides());
return !env::disabled(MIOPEN_DEBUG_FA_CK_V2_FWD) //
&& H_q == H_k // Replace with H_q % H_k == 0 once we add support for MQA & GQA.
&& D_q <= 256 //
&& D_q % 8 == 0 //
&& descsFwd.kDesc.IsPacked() //
&& descsFwd.qDesc.IsPacked() //
&& descsFwd.vDesc.IsPacked() //
&& descsFwd.oDesc.IsPacked() //
&& descsFwd.kDesc.GetType() == miopenHalf //
&& descsFwd.qDesc.GetType() == miopenHalf //
&& descsFwd.vDesc.GetType() == miopenHalf //
&& descsFwd.oDesc.GetType() == miopenHalf //
&& D_stride_k == 1 // CK requires D stride as 1.
&& D_stride_q == 1 && D_stride_v == 1 && D_stride_o == 1;
#else
return false;
#endif
}
std::size_t MhaCKFlashAttentionV2Forward::GetWorkspaceSize(
[[maybe_unused]] const ExecutionContext& context,
[[maybe_unused]] const miopen::mha::ProblemDescription& problem) const
{
return 0;
}
ConvSolution
MhaCKFlashAttentionV2Forward::GetSolution([[maybe_unused]] const ExecutionContext& context,
const miopen::mha::ProblemDescription& problem) const
{
#if MIOPEN_USE_COMPOSABLEKERNEL
auto result = ConvSolution{miopenStatusSuccess};
result.workspace_sz = 0;
const miopen::mha::MhaInputDescsForward& descsFwd = problem.GetDescsForward();
auto [N_k, H_k, S_k, D_k] = miopen::tien<4>(descsFwd.kDesc.GetLengths());
auto [N_stride_k, H_stride_k, S_stride_k, D_stride_k] =
miopen::tien<4>(descsFwd.kDesc.GetStrides());
auto [N_q, H_q, S_q, D_q] = miopen::tien<4>(descsFwd.qDesc.GetLengths());
auto [N_stride_q, H_stride_q, S_stride_q, D_stride_q] =
miopen::tien<4>(descsFwd.qDesc.GetStrides());
auto [N_v, H_v, S_v, D_v] = miopen::tien<4>(descsFwd.vDesc.GetLengths());
auto [N_stride_v, H_stride_v, S_stride_v, D_stride_v] =
miopen::tien<4>(descsFwd.vDesc.GetStrides());
auto [N_stride_o, H_stride_o, S_stride_o, D_stride_o] =
miopen::tien<4>(descsFwd.oDesc.GetStrides());
float scale_s = descsFwd.scale;
float scale_p = 1.0;
float scale_o = 1.0;
fmha_fwd_traits fmha_traits;
fmha_traits.hdim_q = D_q;
fmha_traits.hdim_v = D_v;
fmha_traits.data_type = Convert(descsFwd.qDesc.GetType());
fmha_traits.is_group_mode = false;
// is_v_rowmajor relates to the layout of the V tensor. Row major means NHSD, and Col major
// means NHDS.
fmha_traits.is_v_rowmajor = true;
fmha_traits.mask_type = mask_enum::no_mask;
fmha_traits.has_lse = false;
fmha_traits.do_fp8_static_quant = false;
fmha_traits.has_dropout = false;
fmha_traits.bias_type = bias_enum::no_bias;
fmha_fwd_args fmha_args;
fmha_args.hdim_q = D_q;
fmha_args.hdim_v = D_v;
fmha_args.seqlen_k = S_k;
fmha_args.seqlen_q = S_q;
fmha_args.max_seqlen_q = S_q;
fmha_args.nhead_q = H_q;
fmha_args.nhead_k = H_k;
fmha_args.batch = N_q;
fmha_args.stride_q = S_stride_q;
fmha_args.stride_k = S_stride_k;
fmha_args.stride_v = S_stride_v;
fmha_args.stride_o = S_stride_o;
fmha_args.stride_bias = 0;
fmha_args.stride_randval = S_q;
fmha_args.nhead_stride_q = H_stride_q;
fmha_args.nhead_stride_k = H_stride_k;
fmha_args.nhead_stride_v = H_stride_v;
fmha_args.nhead_stride_o = H_stride_o;
fmha_args.nhead_stride_lse = S_q;
fmha_args.nhead_stride_bias = 0;
fmha_args.nhead_stride_randval = S_q * S_k;
fmha_args.batch_stride_q = N_stride_q;
fmha_args.batch_stride_k = N_stride_k;
fmha_args.batch_stride_v = N_stride_v;
fmha_args.batch_stride_o = N_stride_o;
fmha_args.batch_stride_lse = H_q * S_q;
fmha_args.batch_stride_bias = 0;
fmha_args.batch_stride_randval = H_q * S_q * S_k;
// These are used for group mode, and we are in batch right now.
fmha_args.seqstart_q_ptr = nullptr;
fmha_args.seqstart_k_ptr = nullptr;
// Batch does not support padding, and we aren't using kvcache yet.
fmha_args.seqlen_k_ptr = nullptr;
fmha_args.s_randval = false;
fmha_args.scale_s = scale_s;
fmha_args.scale_p = scale_p;
fmha_args.scale_o = scale_o;
fmha_args.window_size_left = 0;
fmha_args.window_size_right = 0;
fmha_args.mask_type = static_cast<ck_tile::index_t>(fmha_traits.mask_type);
result.invoker_factory = [=](const std::vector<Kernel>&) {
return [=](const Handle& handle_, const AnyInvokeParams& raw_params) {
decltype(auto) params = raw_params.CastTo<miopen::mha::InvokeParams>();
const auto& dataFwd = params.GetDataForward();
fmha_fwd_traits fmha_runtime_traits = fmha_traits;
fmha_fwd_args fmha_runtime_args = fmha_args;
fmha_runtime_args.q_ptr = dataFwd.qData;
fmha_runtime_args.k_ptr = dataFwd.kData;
fmha_runtime_args.v_ptr = dataFwd.vData;
fmha_runtime_args.o_ptr = dataFwd.oData;
fmha_runtime_args.rand_val_ptr = nullptr;
fmha_runtime_args.bias_ptr = nullptr;
fmha_runtime_args.lse_ptr = nullptr;
// Top-left causal mask
if(dataFwd.mask == miopenMhaMask_t::miopenMhaMaskCausal)
{
fmha_runtime_traits.mask_type = mask_enum::mask_top_left;
fmha_runtime_args.mask_type =
static_cast<ck_tile::index_t>(mask_enum::mask_top_left);
fmha_runtime_args.window_size_left = -1;
fmha_runtime_args.window_size_right = 0;
}
fmha_runtime_traits.has_dropout = false;
float probability = 0;
// TODO : Change API to take in probability value as host side value instead of device
// pointer to match CK API. Calling a blocking hipMemcpy will cause issues with stream,
// and isn't async.
fmha_runtime_args.p_drop = probability;
fmha_runtime_args.drop_seed_offset =
std::make_pair(dataFwd.dropoutSeedData, dataFwd.dropoutOffsetData);
// Create stream_config, and set it to not time kernel.
ck_tile::stream_config stream_config;
stream_config.stream_id_ = handle_.GetStream();
{
HipEventProfiler profiler(handle_);
fmha_fwd(fmha_runtime_traits, fmha_runtime_args, stream_config);
}
};
};
return result;
#else
return ConvSolution{miopenStatusNotImplemented};
#endif
}
bool MhaCKFlashAttentionV2Forward::MayNeedWorkspace() const { return false; }
} // namespace mha
} // namespace solver
} // namespace miopen