lowering.cpp 15.3 KB
Newer Older
1
#include <rocblas.h>
Paul's avatar
Paul committed
2
#include <migraph/gpu/lowering.hpp>
Paul's avatar
Paul committed
3
4
5
#include <migraph/manage_ptr.hpp>
#include <migraph/instruction.hpp>
#include <migraph/operators.hpp>
Paul's avatar
Paul committed
6
#include <migraph/generate.hpp>
Paul's avatar
Paul committed
7
#include <migraph/shape_for_each.hpp>
Paul's avatar
Paul committed
8
9
#include <migraph/gpu/miopen.hpp>
#include <migraph/gpu/hip.hpp>
Paul's avatar
Paul committed
10
#include <migraph/dfor.hpp>
11
#include <migraph/gpu/device/contiguous.hpp>
Paul's avatar
Paul committed
12
#include <migraph/iterator_for.hpp>
Paul's avatar
Paul committed
13
14
#include <migraph/gpu/rocblas.hpp>
#include <migraph/gpu/context.hpp>
Paul's avatar
Paul committed
15
#include <utility>
Paul's avatar
Paul committed
16
17

namespace migraph {
Paul's avatar
Paul committed
18
namespace gpu {
Paul's avatar
Paul committed
19

20
21
22
23
24
25
struct miopen_batch_norm_inference
{
    batch_norm_inference op;

    std::string name() const { return "gpu::batch_norm_inference"; }

Paul's avatar
Paul committed
26
    shape compute_shape(const std::vector<shape>& inputs) const
27
28
29
30
31
32
    {
        check_shapes{inputs, *this}.has(6);
        return op.compute_shape(
            {inputs.at(0), inputs.at(1), inputs.at(2), inputs.at(3), inputs.at(4)});
    }

Paul's avatar
Paul committed
33
34
    argument
    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
35
    {
wsttiger's avatar
wsttiger committed
36
37
        auto x_desc  = make_tensor(args[0].get_shape());
        auto y_desc  = make_tensor(output_shape);
38
        auto bn_desc = make_tensor(args[3].get_shape());
39
40
41
42
43
44
45
46
47
48
49

        float alpha = 1.0, beta = 0.0f;

        miopenBatchNormalizationForwardInference(ctx.handle.get(),
                                                 miopenBatchNormMode_t(op.bn_mode),
                                                 &alpha,
                                                 &beta,
                                                 x_desc.get(),
                                                 args[0].implicit(),
                                                 y_desc.get(),
                                                 args[5].implicit(),
50
                                                 bn_desc.get(),
51
52
                                                 args[1].implicit(),
                                                 args[2].implicit(),
Paul's avatar
Paul committed
53
54
                                                 args[3].implicit(),
                                                 args[4].implicit(),
55
                                                 op.epsilon);
56
57
58
59
60

        return args[5];
    }
};

Paul's avatar
Paul committed
61
62
63
struct miopen_convolution
{
    convolution op;
Paul's avatar
Paul committed
64
    shared<convolution_descriptor> cd;
Paul's avatar
Paul committed
65
    miopenConvFwdAlgorithm_t algo{};
Paul's avatar
Paul committed
66

Paul's avatar
Paul committed
67
    std::string name() const { return "gpu::convolution"; }
Paul's avatar
Paul committed
68
    shape compute_shape(const std::vector<shape>& inputs) const
Paul's avatar
Paul committed
69
    {
Paul's avatar
Paul committed
70
        check_shapes{inputs, *this}.has(4).standard();
Paul's avatar
Paul committed
71
        return op.compute_shape({inputs.at(0), inputs.at(1)});
Paul's avatar
Paul committed
72
    }
Paul's avatar
Paul committed
73
74
    argument
    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
Paul's avatar
Paul committed
75
    {
Paul's avatar
Paul committed
76
77
        auto x_desc = make_tensor(args[0].get_shape());
        auto w_desc = make_tensor(args[1].get_shape());
Paul's avatar
Paul committed
78
79
        auto y_desc = make_tensor(output_shape);

Paul's avatar
Paul committed
80
        float alpha = 1, beta = 0;
Paul's avatar
Paul committed
81
        miopenConvolutionForward(ctx.handle.get(),
Paul's avatar
Paul committed
82
                                 &alpha,
Paul's avatar
Paul committed
83
                                 x_desc.get(),
Paul's avatar
Paul committed
84
                                 args[0].implicit(),
Paul's avatar
Paul committed
85
                                 w_desc.get(),
Paul's avatar
Paul committed
86
                                 args[1].implicit(),
Paul's avatar
Paul committed
87
                                 cd.get(),
Paul's avatar
Paul committed
88
                                 algo,
Paul's avatar
Paul committed
89
                                 &beta,
Paul's avatar
Paul committed
90
                                 y_desc.get(),
Paul's avatar
Paul committed
91
                                 args[3].implicit(),
Paul's avatar
Paul committed
92
                                 args[2].implicit(),
Paul's avatar
Paul committed
93
94
                                 args[2].get_shape().bytes());
        return args[3];
Paul's avatar
Paul committed
95
    }
Paul's avatar
Paul committed
96

Paul's avatar
Paul committed
97
    shape compile(context& ctx, const shape& output_shape, std::vector<instruction_ref> inputs)
Paul's avatar
Paul committed
98
    {
Paul's avatar
Paul committed
99
        shape workspace_shape{};
Paul's avatar
Paul committed
100
101
102
103
        auto x_desc = make_tensor(inputs[0]->get_shape());
        auto w_desc = make_tensor(inputs[1]->get_shape());
        auto y_desc = make_tensor(output_shape);

Paul's avatar
Paul committed
104
        std::size_t workspace_size = 0;
Paul's avatar
Paul committed
105
        miopenConvolutionForwardGetWorkSpaceSize(
Paul's avatar
Paul committed
106
            ctx.handle.get(), w_desc.get(), x_desc.get(), cd.get(), y_desc.get(), &workspace_size);
Paul's avatar
Paul committed
107
108
        workspace_shape = shape{shape::int8_type, {workspace_size}};

Paul's avatar
Paul committed
109
        auto x         = to_gpu(generate_argument(inputs[0]->get_shape()));
mei-ye's avatar
staging  
mei-ye committed
110
        gpu_sync();
Paul's avatar
Paul committed
111
        auto w         = to_gpu(generate_argument(inputs[1]->get_shape()));
mei-ye's avatar
staging  
mei-ye committed
112
        gpu_sync();
Paul's avatar
Paul committed
113
        auto y         = to_gpu(generate_argument(output_shape));
Paul's avatar
Paul committed
114
        auto workspace = allocate_gpu(workspace_shape);
Paul's avatar
Paul committed
115

Paul's avatar
Paul committed
116
        int algo_count = 1;
Paul's avatar
Paul committed
117
118
119
120
121
122
123
124
125
126
127
128
        miopenConvAlgoPerf_t perf;
        miopenFindConvolutionForwardAlgorithm(ctx.handle.get(),
                                              x_desc.get(),
                                              x.implicit(),
                                              w_desc.get(),
                                              w.implicit(),
                                              cd.get(),
                                              y_desc.get(),
                                              y.implicit(),
                                              1,
                                              &algo_count,
                                              &perf,
Paul's avatar
Paul committed
129
130
                                              workspace.implicit(),
                                              workspace_size,
Paul's avatar
Paul committed
131
132
                                              false);
        algo = perf.fwd_algo;
Paul's avatar
Paul committed
133
134
        return algo == miopenConvolutionFwdAlgoWinograd ? shape{shape::int8_type, {0}}
                                                        : workspace_shape;
Paul's avatar
Paul committed
135
    }
Paul's avatar
Paul committed
136
137
};

Paul's avatar
Paul committed
138
139
140
141
142
struct miopen_pooling
{
    pooling op;
    shared<pooling_descriptor> pd;

Paul's avatar
Paul committed
143
    std::string name() const { return "gpu::pooling"; }
Paul's avatar
Paul committed
144
    shape compute_shape(const std::vector<shape>& inputs) const
Paul's avatar
Paul committed
145
    {
Paul's avatar
Paul committed
146
        check_shapes{inputs, *this}.has(2).standard();
Paul's avatar
Paul committed
147
        return op.compute_shape({inputs.at(0)});
Paul's avatar
Paul committed
148
    }
Paul's avatar
Paul committed
149
150
    argument
    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
Paul's avatar
Paul committed
151
    {
Paul's avatar
Paul committed
152
        auto x_desc = make_tensor(args[0].get_shape());
Paul's avatar
Paul committed
153
154
155
156
        auto y_desc = make_tensor(output_shape);

        float alpha = 1, beta = 0;

Paul's avatar
Paul committed
157
        miopenPoolingForward(ctx.handle.get(),
Paul's avatar
Paul committed
158
159
160
                             pd.get(),
                             &alpha,
                             x_desc.get(),
Paul's avatar
Paul committed
161
                             args[0].implicit(),
Paul's avatar
Paul committed
162
163
                             &beta,
                             y_desc.get(),
Paul's avatar
Paul committed
164
                             args[1].implicit(),
Paul's avatar
Paul committed
165
166
167
                             false,
                             nullptr,
                             0);
Paul's avatar
Paul committed
168

Paul's avatar
Paul committed
169
        return args[1];
Paul's avatar
Paul committed
170
171
172
    }
};

Paul's avatar
Paul committed
173
174
struct miopen_add
{
Paul's avatar
Paul committed
175
    std::string name() const { return "gpu::add"; }
Paul's avatar
Paul committed
176
    shape compute_shape(const std::vector<shape>& inputs) const
Paul's avatar
Paul committed
177
    {
Paul's avatar
Paul committed
178
        check_shapes{inputs, *this}.has(3).not_broadcasted();
Paul's avatar
Paul committed
179
        return inputs.at(0);
Paul's avatar
Paul committed
180
181
    }

Paul's avatar
Paul committed
182
183
    argument
    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
Paul's avatar
Paul committed
184
    {
Paul's avatar
Paul committed
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
        float alpha = 1, beta = 0;
        auto a_desc = make_tensor(args[0].get_shape());
        auto b_desc = make_tensor(args[1].get_shape());
        auto c_desc = make_tensor(output_shape);
        miopenOpTensor(ctx.handle.get(),
                       miopenTensorOpAdd,
                       &alpha,
                       a_desc.get(),
                       args[0].implicit(),
                       &alpha,
                       b_desc.get(),
                       args[1].implicit(),
                       &beta,
                       c_desc.get(),
                       args[2].implicit());
        return args[2];
Paul's avatar
Paul committed
201
202
203
    }
};

Paul's avatar
Paul committed
204
205
206
struct miopen_gemm
{
    gemm op;
Paul's avatar
Paul committed
207
    std::string name() const { return "gpu::convolution"; }
Paul's avatar
Paul committed
208
    shape compute_shape(const std::vector<shape>& inputs) const
Paul's avatar
Paul committed
209
    {
Paul's avatar
Paul committed
210
211
        check_shapes{inputs, *this}.has(3);
        return op.compute_shape({inputs.at(0), inputs.at(1)});
Paul's avatar
Paul committed
212
    }
Paul's avatar
Paul committed
213
214
    argument
    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
Paul's avatar
Paul committed
215
    {
216
217
        float alpha     = 1.0f;
        float beta      = 0.0f;
Paul's avatar
Paul committed
218
219
        bool transa     = args[0].get_shape().transposed();
        bool transb     = args[1].get_shape().transposed();
220
221
222
        rocblas_int lda = args[0].get_shape().strides()[transa ? 1 : 0];
        rocblas_int ldb = args[1].get_shape().strides()[transb ? 1 : 0];
        rocblas_int ldc = args[2].get_shape().strides()[0];
223
224
225
        rocblas_int m   = output_shape.lens()[0];
        rocblas_int n   = output_shape.lens()[1];
        rocblas_int k   = args[0].get_shape().lens()[1];
226
        rocblas_sgemm(ctx.rbhandle.get(),
227
228
                      transb ? rocblas_operation_transpose : rocblas_operation_none,
                      transa ? rocblas_operation_transpose : rocblas_operation_none,
229
230
231
232
233
234
235
236
237
238
239
240
                      n,
                      m,
                      k,
                      &alpha,
                      args[1].implicit(),
                      ldb,
                      args[0].implicit(),
                      lda,
                      &beta,
                      args[2].implicit(),
                      ldc);
        return args[2];
Paul's avatar
Paul committed
241
242
243
    }
};

244
245
246
struct miopen_contiguous
{
    contiguous op;
Paul's avatar
Paul committed
247
    std::string name() const { return "gpu::contiguous"; }
Paul's avatar
Paul committed
248
    shape compute_shape(const std::vector<shape>& inputs) const
249
250
251
252
    {
        check_shapes{inputs, *this}.has(2);
        return op.compute_shape({inputs.at(0)});
    }
Paul's avatar
Paul committed
253
    argument compute(context&, shape output_shape, const std::vector<argument>& args) const
254
    {
Paul's avatar
Paul committed
255
256
        assert(output_shape == args[1].get_shape());
        assert(output_shape.standard());
Paul's avatar
Paul committed
257
        (void)output_shape;
258
        device::contiguous(args.at(1), args.at(0));
259
        return args.at(1);
260
261
262
    }
};

Paul's avatar
Paul committed
263
264
265
struct miopen_relu
{
    shared<activation_descriptor> ad;
Paul's avatar
Paul committed
266
    std::string name() const { return "gpu::relu"; }
Paul's avatar
Paul committed
267
    shape compute_shape(const std::vector<shape>& inputs) const
Paul's avatar
Paul committed
268
    {
Paul's avatar
Paul committed
269
        check_shapes{inputs, *this}.has(2).not_broadcasted();
Paul's avatar
Paul committed
270
        return inputs.at(1);
Paul's avatar
Paul committed
271
272
    }

Paul's avatar
Paul committed
273
274
    argument
    compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
Paul's avatar
Paul committed
275
276
    {
        float alpha = 1, beta = 0;
Paul's avatar
Paul committed
277
        auto x_desc = make_tensor(args[0].get_shape());
Paul's avatar
Paul committed
278
        auto y_desc = make_tensor(output_shape);
Paul's avatar
Paul committed
279
        miopenActivationForward(ctx.handle.get(),
Paul's avatar
Paul committed
280
281
282
                                ad.get(),
                                &alpha,
                                x_desc.get(),
Paul's avatar
Paul committed
283
                                args[0].implicit(),
Paul's avatar
Paul committed
284
285
                                &beta,
                                y_desc.get(),
Paul's avatar
Paul committed
286
                                args[1].implicit());
Paul's avatar
Paul committed
287

Paul's avatar
Paul committed
288
        return args[1];
Paul's avatar
Paul committed
289
290
291
    }
};

Paul's avatar
Paul committed
292
293
struct miopen_apply
{
Paul's avatar
Paul committed
294
    program* prog = nullptr;
Paul's avatar
Paul committed
295
    context ctx{};
Paul's avatar
Paul committed
296

Paul's avatar
Paul committed
297
298
299
300
301
302
303
    void check_shape(shape x, instruction_ref i)
    {
        assert(x == i->get_shape());
        (void)x;
        (void)i;
    }

Paul's avatar
Paul committed
304
305
    void apply()
    {
Paul's avatar
Paul committed
306
307
        for(auto it = prog->begin(); it != prog->end(); it++)
        {
Paul's avatar
Paul committed
308
            auto s = it->get_shape();
Paul's avatar
Paul committed
309
310
            if(it->op.name() == "convolution")
            {
Paul's avatar
Paul committed
311
                check_shape(s, apply_convolution(it));
Paul's avatar
Paul committed
312
313
314
            }
            else if(it->op.name() == "activation")
            {
Paul's avatar
Paul committed
315
                check_shape(s, apply_activation(it));
Paul's avatar
Paul committed
316
            }
Paul's avatar
Paul committed
317
318
            else if(it->op.name() == "pooling")
            {
Paul's avatar
Paul committed
319
                check_shape(s, apply_pooling(it));
Paul's avatar
Paul committed
320
            }
Paul's avatar
Paul committed
321
322
            else if(it->op.name() == "add")
            {
Paul's avatar
Paul committed
323
                check_shape(s, apply_add(it));
Paul's avatar
Paul committed
324
            }
Paul's avatar
Paul committed
325
326
            else if(it->op.name() == "gemm")
            {
Paul's avatar
Paul committed
327
                check_shape(s, apply_gemm(it));
Paul's avatar
Paul committed
328
            }
329
330
            else if(it->op.name() == "contiguous")
            {
Paul's avatar
Paul committed
331
                check_shape(s, apply_contiguous(it));
332
            }
333
334
            else if(it->op.name() == "batch_norm_inference")
            {
Paul's avatar
Paul committed
335
                check_shape(s, apply_batch_norm_inference(it));
336
            }
Paul's avatar
Paul committed
337
338
339
        }
    }

Paul's avatar
Paul committed
340
    instruction_ref insert_allocation(instruction_ref ins, const shape& s, std::string tag = "")
Paul's avatar
Paul committed
341
    {
Paul's avatar
Paul committed
342
        if(ins == --prog->end())
Paul's avatar
Paul committed
343
344
345
346
347
        {
            return prog->add_parameter("output", s);
        }
        else
        {
Paul's avatar
Paul committed
348
            auto is     = prog->add_outline(s);
Paul's avatar
Paul committed
349
            auto result = prog->insert_instruction(ins, hip_allocate{std::move(tag)}, is);
Paul's avatar
Paul committed
350
351
352
353
            return result;
        }
    }

Paul's avatar
Paul committed
354
    instruction_ref apply_convolution(instruction_ref ins)
Paul's avatar
Paul committed
355
    {
Paul's avatar
Paul committed
356
        auto&& op = any_cast<convolution>(ins->op);
Paul's avatar
Paul committed
357

Paul's avatar
Paul committed
358
        auto conv = miopen_convolution{op, make_conv(op)};
Paul's avatar
Paul committed
359
        auto ws   = conv.compile(ctx, ins->result, ins->arguments);
Paul's avatar
Paul committed
360

361
        auto workspace = insert_allocation(ins, ws, "workspace");
Paul's avatar
Paul committed
362
        auto output    = insert_allocation(ins, ins->result);
Paul's avatar
Paul committed
363

Paul's avatar
Paul committed
364
        return prog->replace_instruction(
Paul's avatar
Paul committed
365
            ins, conv, ins->arguments.at(0), ins->arguments.at(1), workspace, output);
Paul's avatar
Paul committed
366
367
    }

Paul's avatar
Paul committed
368
    instruction_ref apply_pooling(instruction_ref ins)
Paul's avatar
Paul committed
369
370
371
372
373
    {
        auto&& op   = any_cast<pooling>(ins->op);
        auto pd     = make_pooling(op);
        auto output = insert_allocation(ins, ins->result);

Paul's avatar
Paul committed
374
        return prog->replace_instruction(
Paul's avatar
Paul committed
375
            ins, miopen_pooling{op, std::move(pd)}, ins->arguments.at(0), output);
Paul's avatar
Paul committed
376
377
    }

Paul's avatar
Paul committed
378
    instruction_ref apply_activation(instruction_ref ins)
Paul's avatar
Paul committed
379
380
    {
        auto&& op = any_cast<activation>(ins->op);
Paul's avatar
Paul committed
381
382
        auto ad   = make_relu();
        if(op.mode == "relu")
Paul's avatar
Paul committed
383
384
        {
            auto output = insert_allocation(ins, ins->result);
Paul's avatar
Paul committed
385
            return prog->replace_instruction(
Paul's avatar
Paul committed
386
                ins, miopen_relu{std::move(ad)}, ins->arguments.at(0), output);
Paul's avatar
Paul committed
387
        }
Paul's avatar
Paul committed
388
        return ins;
Paul's avatar
Paul committed
389
    }
Paul's avatar
Paul committed
390

Paul's avatar
Paul committed
391
    instruction_ref apply_add(instruction_ref ins)
Paul's avatar
Paul committed
392
393
    {
        auto output = insert_allocation(ins, ins->result);
Paul's avatar
Paul committed
394
        return prog->replace_instruction(
Paul's avatar
Paul committed
395
            ins, miopen_add{}, ins->arguments.at(0), ins->arguments.at(1), output);
Paul's avatar
Paul committed
396
    }
Paul's avatar
Paul committed
397

Paul's avatar
Paul committed
398
    instruction_ref apply_gemm(instruction_ref ins)
Paul's avatar
Paul committed
399
    {
Paul's avatar
Paul committed
400
        auto&& op   = any_cast<gemm>(ins->op);
Paul's avatar
Paul committed
401
        auto output = insert_allocation(ins, ins->result);
Paul's avatar
Paul committed
402
        return prog->replace_instruction(
Paul's avatar
Paul committed
403
            ins, miopen_gemm{op}, ins->arguments.at(0), ins->arguments.at(1), output);
Paul's avatar
Paul committed
404
    }
405

Paul's avatar
Paul committed
406
    instruction_ref apply_contiguous(instruction_ref ins)
407
408
409
    {
        auto&& op   = any_cast<contiguous>(ins->op);
        auto output = insert_allocation(ins, ins->result);
Paul's avatar
Paul committed
410
        return prog->replace_instruction(ins, miopen_contiguous{op}, ins->arguments.at(0), output);
411
    }
412

Paul's avatar
Paul committed
413
    instruction_ref apply_batch_norm_inference(instruction_ref ins)
414
    {
wsttiger's avatar
wsttiger committed
415
416
        auto&& op       = any_cast<batch_norm_inference>(ins->op);
        auto output     = insert_allocation(ins, ins->result);
417
        shape old_shape = ins->arguments.at(1)->get_shape();
wsttiger's avatar
wsttiger committed
418
        std::vector<int64_t> new_shape{1, static_cast<int64_t>(old_shape.elements()), 1, 1};
Paul's avatar
Paul committed
419
420
        auto reshape_op = reshape{new_shape};
        std::vector<instruction_ref> reshapes;
Paul's avatar
Paul committed
421
422
423
424
        std::transform(ins->arguments.begin() + 1,
                       ins->arguments.end(),
                       std::back_inserter(reshapes),
                       [&](auto i) { return prog->insert_instruction(ins, reshape_op, i); });
Paul's avatar
Paul committed
425
        return prog->replace_instruction(ins,
Paul's avatar
Paul committed
426
427
428
429
430
431
432
                                         miopen_batch_norm_inference{op},
                                         ins->arguments.at(0),
                                         reshapes[0],
                                         reshapes[1],
                                         reshapes[2],
                                         reshapes[3],
                                         output);
433
    }
Paul's avatar
Paul committed
434
435
};

Paul's avatar
Paul committed
436
void lowering::apply(program& p) const { miopen_apply{&p, ctx}.apply(); }
Paul's avatar
Paul committed
437

Paul's avatar
Paul committed
438
} // namespace gpu
Paul's avatar
Paul committed
439

Paul's avatar
Paul committed
440
} // namespace migraph