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
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
/*******************************************************************************
* Copyright 2020 Intel 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.
*******************************************************************************/
#ifndef GPU_INTEL_MATMUL_GEMM_HPP
#define GPU_INTEL_MATMUL_GEMM_HPP
#include "common/c_types_map.hpp"
#include "common/gemm_utils.hpp"
#include "common/matmul_pd.hpp"
#include "common/memory_desc.hpp"
#include "common/memory_desc_wrapper.hpp"
#include "common/primitive.hpp"
#include "common/primitive_attr_quant.hpp"
#include "common/primitive_desc_iterator.hpp"
#include "gpu/intel/matmul/config.hpp"
#include "gpu/intel/primitive.hpp"
#include "gpu/intel/utils.hpp"
namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace matmul {
struct gemm_t : public primitive_t {
using primitive_t::primitive_t;
struct pd_t : public matmul::pd_t {
using matmul::pd_t::pd_t;
DECLARE_COMMON_PD_T(gemm_pd_ ? gemm_pd_->name() : "gemm_t", gemm_t);
status_t init(impl::engine_t *engine) {
using namespace data_type;
primitive_attr_t gemm_attr;
gemm_attr.fpmath_ = attr()->fpmath_;
gemm_attr.acc_mode_ = attr()->acc_mode_;
gemm_attr.deterministic_ = attr()->deterministic_;
gemm_attr.rounding_mode_ = attr()->rounding_mode_;
gemm_attr.scales_ = attr()->scales_;
gemm_attr.precomputed_reductions_ = attr()->precomputed_reductions_;
gemm_attr.zero_points_ = attr()->zero_points_;
gemm_attr.post_ops_ = attr()->post_ops_;
gemm_attr.dropout_ = attr()->dropout_;
auto a_md = src_md(), b_md = weights_md(), c_md = dst_md(),
bias_md = weights_md(1);
const auto acc_dt = desc()->accum_data_type;
memory_desc_t a_md_reshaped, b_md_reshaped, c_md_reshaped,
bia_md_reshaped;
bool with_bia = bias_md->ndims > 0;
auto orig_dims = a_md->ndims;
// Check for grouped encoding early before reshape attempts
VDISPATCH_MATMUL(
is_dense_format_kind(), VERBOSE_UNSUPPORTED_SPARSE_CFG);
auto maybe_reshape = [&]() -> status_t {
dim_t batch_b_dims = 1;
for (int i = 0; i < b_md->ndims - 2; i++) {
batch_b_dims *= b_md->dims[i];
}
// for batch dim can map broadcast to 2d: eg. 4x1x4096:1x4096x16 -> 4x4096:4096x16
bool reshape_2d = (batch_b_dims == 1 && b_md->ndims > 2);
bool reshape_3d = (a_md->ndims > 3);
bool allow_reshape
= gpu_utils::dev_getenv("GEMM_ALLOW_RESHAPE", true);
// Early exit if not reshaping
if (!allow_reshape || !(reshape_2d || reshape_3d))
return status::success;
int ndims = a_md->ndims;
int reshape_size = reshape_2d ? 2 : 3;
int diff_dims = orig_dims - reshape_size;
// Converts input dims_t to output dims_t after reshaping
auto squash_dims = [](dims_t &out_dims, const dims_t &in_dims,
int in_ndims, int out_ndims) {
int diff_ndims = in_ndims - out_ndims;
gpu_assert(diff_ndims > 0) << "Unexpected squashing dims";
gpu_assert(&out_dims != &in_dims)
<< "Cannot squash in-place";
// Shift dims over
for (int i = 0; i < out_ndims; i++) {
out_dims[i] = in_dims[i + diff_ndims];
}
// Squash dims into 1st dim
for (int i = 0; i < diff_ndims; i++) {
out_dims[0] *= in_dims[i];
}
};
// Convert raw tensors to reshaped tensors
dims_t a_dims, b_dims, c_dims, bia_dims;
squash_dims(a_dims, a_md->dims, ndims, reshape_size);
squash_dims(b_dims, b_md->dims, ndims, reshape_size);
squash_dims(c_dims, c_md->dims, ndims, reshape_size);
squash_dims(bia_dims, bias_md->dims, ndims, reshape_size);
// Cannot reshape if bias is broadcast across a subset of squashed dimensions
bool bcast_ok = IMPLICATION(
with_bia, utils::one_of(bia_dims[0], 1, c_dims[0]));
// 3D reshaping is only possible if A and B batch sizes allow.
// This means no reshaping with partial broadcasting
// or when both tensors are broadcast in different dimensions.
bool a_broadcast = false;
bool b_broadcast = false;
for (int i = 0; i < ndims - 2; i++) {
if (a_md->dims[i] == 1 && b_md->dims[i] > 1)
a_broadcast = true;
if (b_md->dims[i] == 1 && a_md->dims[i] > 1)
b_broadcast = true;
}
bcast_ok = bcast_ok && !(a_broadcast && b_broadcast);
bcast_ok = bcast_ok
&& IMPLICATION(reshape_size == 3,
a_dims[0] == b_dims[0]
|| utils::one_of(
1, a_dims[0], b_dims[0]));
// memory_desc_reshape can fail. If so, fail gracefully so
// we can use batched gemm instead of skipping entirely.
auto safe_reshape
= [](memory_desc_t &out_md, const memory_desc_t &in_md,
int ndims, const dims_t dims) -> bool {
CHECK_BOOL(memory_desc_reshape(out_md, in_md, ndims, dims));
return true;
};
bcast_ok = bcast_ok
&& safe_reshape(
a_md_reshaped, *a_md, reshape_size, a_dims);
bcast_ok = bcast_ok
&& safe_reshape(
b_md_reshaped, *b_md, reshape_size, b_dims);
bcast_ok = bcast_ok
&& safe_reshape(
c_md_reshaped, *c_md, reshape_size, c_dims);
if (with_bia) {
bcast_ok = bcast_ok
&& safe_reshape(bia_md_reshaped, *bias_md,
reshape_size, bia_dims);
}
auto reshaped_post_ops = gemm_attr.post_ops_;
for (int i = 0; i < attr()->post_ops_.len(); i++) {
auto &po = reshaped_post_ops.entry_[i];
if (po.is_binary()) {
const auto &po_desc = po.binary.src1_desc;
auto a_dim = po_desc.dims[po_desc.ndims - reshape_size];
for (int i = po_desc.ndims; i > reshape_size; i--) {
a_dim *= po_desc.dims[po_desc.ndims - i];
}
//post ops cannot be applied if applied on only on a subset of batch dims
if (a_dim != c_dims[0] && a_dim > 1) {
return status::success;
}
auto has_dims = po_desc.ndims > 0;
dims_t po_dims;
if (reshape_2d) {
po_dims[0] = a_dim;
po_dims[1] = has_dims
? po_desc.dims[po_desc.ndims - 1]
: 1;
} else {
po_dims[0] = a_dim;
po_dims[1] = has_dims
? po_desc.dims[po_desc.ndims - 2]
: 1;
po_dims[2] = has_dims
? po_desc.dims[po_desc.ndims - 1]
: 1;
}
memory_desc_t tmp_po_desc;
bcast_ok = bcast_ok
&& safe_reshape(tmp_po_desc, po_desc,
reshape_size, po_dims);
reshaped_post_ops.entry_[i].binary.src1_desc
= tmp_po_desc;
} else if (po.is_prelu()) {
auto mask = po.prelu.mask;
int new_mask = 0;
int batch_idx = reshape_size - 1;
dim_t batch_dim = 1;
dim_t mask_dim = 1;
//get mask for batch dim
for (int i = 0; i < c_md->ndims - batch_idx; i++) {
if (mask >> i & 1) {
//post ops cannot be applied if applied on only on a subset of batch dims
if (new_mask != 0) return status::success;
new_mask |= c_md->dims[i] == 1 ? 0 : 1;
mask_dim *= c_md->dims[i];
}
batch_dim *= c_md->dims[i];
}
//post ops cannot be applied if applied on only on a subset of batch dims
if (batch_dim != mask_dim) return status::success;
//get non-batch part of mask
auto shift = c_md->ndims - batch_idx;
auto non_batch_mask = mask >> shift;
//due to prelu being in axb format, if a reshape is done it
//implies layout is different e.g 1x30x20 -> 30 is innermost dimension
//but 30x20 -> 20 is innermost. Hence reshape does not work if mask
//is applied across more than one dimension.
if (non_batch_mask > 2
|| (non_batch_mask > 0 && new_mask > 0))
return status::success;
new_mask |= non_batch_mask << 1;
reshaped_post_ops.entry_[i].prelu.mask = new_mask;
}
}
// Quantization has a few wrinkles...
// Example: --attr-scales=src:per_ocic:f16:1x128 4x1x4096:1x4096x16
// The src scales tensor has dimensions 1x32
// We have two options since we can't change the scales tensor dimensions:
// (1) Change mask from 6 -> 3 (both remaining dims masked) and change grouping
// -> src:per_ocic:4x128
// (2) Change mask from 6 -> 2 (just K dim masked) and don't change grouping
// -> src:per_dim_1:1x128
// Currently gemmstone only supports (1) so that's what we'll do here.
// TODO: (2) has more optimization potential and is more reusable - implement
// this option in gemmstone.
// Same as squash_dims, but early-outs available if quantization not present
auto squash_quant
= [&](dims_t &out_dims, const quant_entry_t &quant,
const memory_desc_t &qmd) {
if (quant.has_default_values()) return;
squash_dims(out_dims, qmd.dims, ndims, reshape_size);
return;
};
auto squashed_mask = [&](int mask, int diff_dims) -> int {
return mask >> diff_dims;
};
auto reshape_quant = [&](const quant_entry_t &in_entry,
const memory_desc_t &reshaped_md,
const dims_t &qdims,
int diff_dims) -> quant_entry_t {
if (in_entry.has_default_values()) return in_entry;
int new_mask
= squashed_mask(in_entry.get_mask(), diff_dims);
data_type_t dt = in_entry.get_data_type();
dims_t dims {};
int ndims = 0;
if (!in_entry.has_default_groups()) {
ndims = 2;
// Recalculate group sizes to obey (1) above
dims[0] = reshaped_md.dims[reshaped_md.ndims - 2]
/ qdims[reshaped_md.ndims - 2];
dims[1] = reshaped_md.dims[reshaped_md.ndims - 1]
/ qdims[reshaped_md.ndims - 1];
}
quant_entry_t out_entry;
UNUSED_STATUS(out_entry.set(new_mask, dt, ndims, dims,
false, in_entry.get_quantization_mode()));
return out_entry;
};
auto adjust_quant = [&](quant_entries_t &entries, int arg,
const memory_desc_t &md,
const memory_desc_t &reshaped_md,
int diff_dims) -> status_t {
const quant_entry_t &entry = entries.get(arg);
if (entry.is_host_scalar()) return status::success;
memory_desc_t qmd;
CHECK(entry.get_md(qmd, md));
dims_t qdims;
squash_quant(qdims, entry, qmd);
quant_entry_t reshaped_entry = reshape_quant(
entry, reshaped_md, qdims, diff_dims);
CHECK(entries.set(arg, reshaped_entry));
return status::success;
};
// Quantization attributes are unsupported in Gemmstone with any
// non-trivial (neither broadcast nor full dimension) modification
// of groups during reshape.
// TODO: Enable arbitrary reshaped grouped quant.
auto safe_bcast_quant = [&](const quant_entry_t &entry,
const dims_t orig_dims,
dims_t reshape_dims) -> bool {
int full_tensor_mask = ((1 << c_md->ndims) - 1);
int per_oc_mask = (1 << 1);
if (!reshape_2d || entry.has_default_values()
|| utils::one_of(entry.get_mask(), 0, per_oc_mask,
full_tensor_mask))
return true;
if (utils::one_of(orig_dims[diff_dims], reshape_dims[0], 1))
return true;
return false;
};
bcast_ok = bcast_ok
&& safe_bcast_quant(gemm_attr.scales_.get(DNNL_ARG_SRC),
a_md->dims, a_dims);
bcast_ok = bcast_ok
&& safe_bcast_quant(
gemm_attr.scales_.get(DNNL_ARG_WEIGHTS),
b_md->dims, b_dims);
bcast_ok = bcast_ok
&& safe_bcast_quant(
gemm_attr.zero_points_.get(DNNL_ARG_SRC),
a_md->dims, a_dims);
bcast_ok = bcast_ok
&& safe_bcast_quant(
gemm_attr.zero_points_.get(DNNL_ARG_WEIGHTS),
b_md->dims, b_dims);
if (!bcast_ok) return status::success;
scales_t reshaped_scales = gemm_attr.scales_;
zero_points_t reshaped_zp = gemm_attr.zero_points_;
precomputed_reductions_t reshaped_pr
= gemm_attr.precomputed_reductions_;
CHECK(adjust_quant(reshaped_scales, DNNL_ARG_SRC, *a_md,
a_md_reshaped, diff_dims));
CHECK(adjust_quant(reshaped_scales, DNNL_ARG_WEIGHTS, *b_md,
b_md_reshaped, diff_dims));
CHECK(adjust_quant(reshaped_scales, DNNL_ARG_DST, *c_md,
c_md_reshaped, diff_dims));
CHECK(adjust_quant(reshaped_zp, DNNL_ARG_SRC, *a_md,
a_md_reshaped, diff_dims));
CHECK(adjust_quant(reshaped_zp, DNNL_ARG_WEIGHTS, *b_md,
b_md_reshaped, diff_dims));
CHECK(adjust_quant(reshaped_zp, DNNL_ARG_DST, *c_md,
c_md_reshaped, diff_dims));
CHECK(adjust_quant(reshaped_pr, DNNL_ARG_SRC, *a_md,
a_md_reshaped, diff_dims));
CHECK(adjust_quant(reshaped_pr, DNNL_ARG_WEIGHTS, *b_md,
b_md_reshaped, diff_dims));
// Reshaping successful - lock in changes
a_md = &a_md_reshaped;
b_md = &b_md_reshaped;
c_md = &c_md_reshaped;
if (with_bia) bias_md = &bia_md_reshaped;
gemm_attr.scales_ = std::move(reshaped_scales);
gemm_attr.zero_points_ = std::move(reshaped_zp);
gemm_attr.precomputed_reductions_ = std::move(reshaped_pr);
gemm_attr.post_ops_ = reshaped_post_ops;
return status::success;
};
VDISPATCH_MATMUL_SC(maybe_reshape(), VERBOSE_IMPL_HEURISTIC_FAIL,
"2D/3D reshaping");
// We create a gemm_pd and resolve 'any' desc by querying gemm_pd
VDISPATCH_MATMUL_SC(create_gemm_pd(gemm_pd_, engine, a_md, b_md,
c_md, bias_md, acc_dt, &gemm_attr),
VERBOSE_PRIMITIVE_CREATION_FAIL, "gemm");
VDISPATCH_MATMUL_SC(set_default_params(), VERBOSE_UNSUPPORTED_TAG);
VDISPATCH_MATMUL_SC(attr_.set_default_formats(dst_md(0)),
VERBOSE_UNSUPPORTED_POSTOP);
init_scratchpad();
return status::success;
}
std::shared_ptr<primitive_desc_t> gemm_pd_;
private:
// We cannot change the number of dimensions in the input mds.
// Grab the gemm_pd_ mds (which should have resolved any tags
// by now) and unsquash the squashed dims
status_t set_default_params() {
auto update_md
= [](memory_desc_t &md,
const memory_desc_t &reshaped_md) -> status_t {
if (md.ndims == reshaped_md.ndims) {
md = reshaped_md;
} else {
CHECK(memory_desc_reshape(
md, reshaped_md, md.ndims, md.dims));
}
return status::success;
};
CHECK(update_md(src_md_, *gemm_pd_->arg_md(DNNL_ARG_SRC_0)));
CHECK(update_md(weights_md_, *gemm_pd_->arg_md(DNNL_ARG_SRC_1)));
CHECK(update_md(dst_md_, *gemm_pd_->arg_md(DNNL_ARG_DST)));
if (with_bias())
CHECK(update_md(bias_md_, *gemm_pd_->arg_md(DNNL_ARG_BIAS)));
return status::success;
}
void init_scratchpad() {
auto scratchpad = scratchpad_registry().registrar();
scratchpad.book(memory_tracking::names::key_nested,
gemm_pd_->scratchpad_registry());
}
};
status_t init(impl::engine_t *engine) override {
return create_nested_primitive(gemm_, pd()->gemm_pd_, engine);
}
status_t execute(const exec_ctx_t &ctx) const override;
private:
const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
std::shared_ptr<impl::primitive_t> gemm_;
};
} // namespace matmul
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl
#endif