multitarget_test.cpp 27.4 KB
Newer Older
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
/*
 * The MIT License (MIT)
 *
 * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
 *
 * 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 <iostream>
#include <set>
#include <unordered_set>
#include <vector>
#include <random>
#include <cmath>
#include <migraphx/program.hpp>
#include <migraphx/target.hpp>
#include <migraphx/stringutils.hpp>
#include <migraphx/module.hpp>
#include <migraphx/literal.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/shape.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/check_shapes.hpp>
#include <migraphx/functional.hpp>
#include <basic_ops.hpp>
#include <migraphx/compile_options.hpp>
#include <migraphx/register_target.hpp>
umangyadav's avatar
umangyadav committed
43
#include <migraphx/generate.hpp>
44
45
#include "migraphx/partitioner.hpp"
#include "migraphx/target_assignments.hpp"
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
#include "test.hpp"

// check if it is custom_op or run_on_module operator
bool has_target_attr(const migraphx::instruction& ins)
{
    return ins.get_operator().attributes().contains("target");
}

auto nonprefixed_ops()
{
    // ops without prefixes
    static std::unordered_set<std::string> op_map = {
        "select_module", "load", "if", "nonmaxsuppression", "multibroadcast"};
    return op_map;
}

bool is_compiled_gpu_module(const migraphx::module& m)
{
    return std::all_of(m.begin(), m.end(), [](auto ins) {
        auto ins_name = ins.name();
        if(not migraphx::starts_with(ins_name, "@"))
        {
            if(not migraphx::starts_with(ins_name, "gpu::") and
               not migraphx::starts_with(ins_name, "hip::") and
               not migraphx::starts_with(ins_name, "check_context") and
               not migraphx::contains(nonprefixed_ops(), ins_name) and not has_target_attr(ins))
            {
                return false;
            }
        }
        return true;
    });
}

bool is_compiled_fpga_module(const migraphx::module& m)
{
    return std::all_of(m.begin(), m.end(), [](auto ins) {
        auto ins_name = ins.name();
        if(not migraphx::starts_with(ins_name, "@"))
        {
            if(not migraphx::starts_with(ins_name, "fpga::") and
               not migraphx::starts_with(ins_name, "check_context") and
               not migraphx::contains(nonprefixed_ops(), ins_name) and not has_target_attr(ins))
            {
                return false;
            }
        }
        return true;
    });
}

bool is_compiled_cpu_module(const migraphx::module& m)
{
    return std::all_of(m.begin(), m.end(), [](auto ins) {
        auto ins_name = ins.name();
        if(not migraphx::starts_with(ins_name, "@"))
        {
            if(not migraphx::starts_with(ins_name, "cpu::") and
               not migraphx::starts_with(ins_name, "dnnl::") and
               not migraphx::starts_with(ins_name, "check_context") and not has_target_attr(ins) and
               not migraphx::contains(nonprefixed_ops(), ins_name))
            {
                return false;
            }
        }
        return true;
    });
}

bool is_compiled_ref_module(const migraphx::module& m)
{
    return std::all_of(m.begin(), m.end(), [](auto ins) {
        auto ins_name = ins.name();
        if(not migraphx::starts_with(ins_name, "@"))
        {
            if((not migraphx::starts_with(ins_name, "ref::") and
                not migraphx::starts_with(ins_name, "check_context") and
                not has_target_attr(ins)) and
               not migraphx::contains(nonprefixed_ops(), ins_name))
            {
                return false;
            }
        }
        return true;
    });
}

// NOLINT
bool check_compiled_program(const migraphx::program& p,
                            const std::vector<migraphx::target>& targets)
{
    auto mods           = p.get_modules();
    bool check_compiled = true;
    for(const auto* mod : mods)
    {
        for(const auto& ins : *mod)
        {
            if(ins.name() == "run_on_target")
            {
                auto* mod_input = ins.module_inputs().front();
                std::size_t target_id =
                    ins.get_operator().to_value()["target_id"].to<std::size_t>();
                auto target_name = targets.at(target_id).name();
                if(target_name == "gpu")
                    check_compiled &= is_compiled_gpu_module(*mod_input);
                else if(target_name == "cpu")
                    check_compiled &= is_compiled_cpu_module(*mod_input);
                else if(target_name == "fpga")
                    check_compiled &= is_compiled_fpga_module(*mod_input);
                else if(target_name == "ref")
                    check_compiled &= is_compiled_ref_module(*mod_input);
            }
        }
    }
    return check_compiled;
}

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
TEST_CASE(multitarget_partition_compile)
{
    migraphx::program p;
    auto* mm     = p.get_main_module();
    auto s       = migraphx::shape{migraphx::shape::float_type, {8}};
    auto x_param = mm->add_parameter("x", s);
    auto y_param = mm->add_parameter("y", s);
    auto z_param = mm->add_parameter("z", s);
    auto cpu_ins = mm->add_instruction(migraphx::make_op("add"), x_param, y_param);
    auto gpu_ins = mm->add_instruction(migraphx::make_op("add"), cpu_ins, z_param);
    mm->add_return({gpu_ins});
    p.debug_print();
    migraphx::target_assignments tass;
    tass.insert(tass.begin(), std::make_pair(cpu_ins, 1));
    tass.insert(tass.begin(), std::make_pair(gpu_ins, 0));
    migraphx::partition(p, tass);
    p.debug_print();
    migraphx::compile_options gpu_opts;
    gpu_opts.offload_copy = true;
    p.compile({migraphx::make_target("gpu"), migraphx::make_target("cpu")}, {gpu_opts});
    p.debug_print();
    EXPECT(check_compiled_program(p, {migraphx::make_target("gpu"), migraphx::make_target("cpu")}));
    migraphx::parameter_map params;
    params["x"] = migraphx::fill_argument(s, 1);
    params["y"] = migraphx::fill_argument(s, 2);
    params["z"] = migraphx::fill_argument(s, 3);
    auto result = p.eval(params).back();
    auto gold   = migraphx::fill_argument(s, 6);
    EXPECT(gold == result);
}

194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
TEST_CASE(multitarget_compile_cpu_gpu)
{
    migraphx::program p;
    auto* mm      = p.get_main_module();
    auto* cpu_mod = p.create_module("cpu_mod");
    auto s        = migraphx::shape{migraphx::shape::float_type, {8}};
    auto x_cpu    = cpu_mod->add_parameter("cpu_x", s);
    auto y_cpu    = cpu_mod->add_parameter("cpu_y", s);
    auto cpu_add  = cpu_mod->add_instruction(migraphx::make_op("add"), x_cpu, y_cpu);
    cpu_mod->add_return({cpu_add});

    auto* gpu_mod = p.create_module("gpu_mod");
    auto x_gpu    = gpu_mod->add_parameter("gpu_x", s);
    auto y_gpu    = gpu_mod->add_parameter("gpu_y", s);
    auto gpu_add  = gpu_mod->add_instruction(migraphx::make_op("add"), x_gpu, y_gpu);
    gpu_mod->add_return({gpu_add});

    auto x_param = mm->add_parameter("x", s);
    auto y_param = mm->add_parameter("y", s);
    auto z_param = mm->add_parameter("z", s);
    auto cpu_ins = mm->add_instruction(
        migraphx::make_op("run_on_target", {{"target_id", 1}}), {x_param, y_param}, {cpu_mod});
umangyadav's avatar
umangyadav committed
216
217
    auto cpu_ins_0 =
        mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), cpu_ins);
218
    auto gpu_ins = mm->add_instruction(
umangyadav's avatar
umangyadav committed
219
220
221
222
223
224
225
226
        migraphx::make_op("run_on_target", {{"target_id", 0}}), {cpu_ins_0, z_param}, {gpu_mod});
    auto gpu_ins_0 =
        mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), gpu_ins);

    mm->add_return({gpu_ins_0});
    migraphx::compile_options gpu_opts;
    gpu_opts.offload_copy = true;
    p.compile({migraphx::make_target("gpu"), migraphx::make_target("cpu")}, {gpu_opts});
227
    EXPECT(check_compiled_program(p, {migraphx::make_target("gpu"), migraphx::make_target("cpu")}));
umangyadav's avatar
umangyadav committed
228
    migraphx::parameter_map params;
umangyadav's avatar
umangyadav committed
229
230
231
    params["x"] = migraphx::fill_argument(s, 1);
    params["y"] = migraphx::fill_argument(s, 2);
    params["z"] = migraphx::fill_argument(s, 3);
umangyadav's avatar
umangyadav committed
232
    auto result = p.eval(params).back();
umangyadav's avatar
umangyadav committed
233
234
    auto gold   = migraphx::fill_argument(s, 6);
    EXPECT(gold == result);
235
236
}

umangyadav's avatar
umangyadav committed
237
TEST_CASE(single_target_multi_compile)
238
239
240
{
    migraphx::program p;
    migraphx::shape boxes_s{migraphx::shape::float_type, {1, 6, 4}};
241
242
    auto* mm         = p.get_main_module();
    auto boxes_param = mm->add_parameter("boxes", boxes_s);
243

244
245
    auto* gpu_mod        = p.create_module("gpu_mod");
    auto boxes_param_gpu = gpu_mod->add_parameter("boxes_param_gpu", boxes_s);
246
247
    migraphx::shape scores_s{migraphx::shape::float_type, {1, 1, 6}};
    std::vector<float> scores_vec = {0.9, 0.75, 0.6, 0.95, 0.5, 0.3};
248
249
250
251
252
253
254
255
256
257
258
259
260
    auto scores_l                 = gpu_mod->add_literal(migraphx::literal(scores_s, scores_vec));
    auto max_out_l                = gpu_mod->add_literal(int64_t{4});
    auto iou_threshold            = gpu_mod->add_literal(0.5f);
    auto score_threshold          = gpu_mod->add_literal(0.0f);
    auto r                        = gpu_mod->add_instruction(
        migraphx::make_op("nonmaxsuppression",
                          {{"center_point_box", true}, {"use_dyn_output", true}}),
        boxes_param_gpu,
        scores_l,
        max_out_l,
        iou_threshold,
        score_threshold);
    gpu_mod->add_return({r});
261

262
263
264
265
266
267
268
    auto run_on_gpu = mm->add_instruction(
        migraphx::make_op("run_on_target", {{"target_id", 0}}), {boxes_param}, {gpu_mod});
    auto run_on_gpu_0 =
        mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_on_gpu);
    mm->add_return({run_on_gpu_0});

    // compile using multi-target compilation path
umangyadav's avatar
umangyadav committed
269
270
    migraphx::compile_options gpu_opts;
    gpu_opts.offload_copy = true;
271
272
273
274
275
    // need to add "ref" to avoid ambigious call to "compile()"
    p.compile({migraphx::make_target("gpu"), migraphx::make_target("ref")}, {gpu_opts});
    EXPECT(check_compiled_program(p, {migraphx::make_target("gpu"), migraphx::make_target("ref")}));
    // eval
    migraphx::parameter_map params;
umangyadav's avatar
umangyadav committed
276
    std::vector<float> boxes_vec  = {0.5, 0.5,  1.0, 1.0, 0.5, 0.6,  1.0, 1.0, 0.5, 0.4,   1.0, 1.0,
277
                                    0.5, 10.5, 1.0, 1.0, 0.5, 10.6, 1.0, 1.0, 0.5, 100.5, 1.0, 1.0};
umangyadav's avatar
umangyadav committed
278
279
280
281
282
283
    params["boxes"]               = migraphx::argument(boxes_s, boxes_vec.data());
    auto output                   = p.eval(params).back();
    std::vector<int64_t> gold_vec = {0, 0, 3, 0, 0, 0, 0, 0, 5};
    auto gold =
        migraphx::argument(migraphx::shape{migraphx::shape::int64_type, {3, 3}}, gold_vec.data());
    EXPECT(output == gold);
284
285
286
287
288
289
290
291
292
293
294
295
}

TEST_CASE(multitarget_compile_if_then_else)
{
    migraphx::program p;
    auto* mm = p.get_main_module();
    migraphx::shape cond_s{migraphx::shape::bool_type};
    auto cond = mm->add_parameter("cond", cond_s);
    migraphx::shape ds{migraphx::shape::float_type, {2, 3}};
    auto x = mm->add_parameter("x", ds);
    auto y = mm->add_parameter("y", ds);

umangyadav's avatar
umangyadav committed
296
297
298
299
300
    auto* then_mod = p.create_module("if_gpu_mod");
    std::vector<float> data1(ds.elements(), 1);
    auto l1    = then_mod->add_literal(migraphx::literal(ds, data1));
    auto gpu_x = then_mod->add_parameter("gpu_x", ds);
    auto a1    = then_mod->add_instruction(migraphx::make_op("add"), gpu_x, l1);
301
302
    then_mod->add_return({a1});

umangyadav's avatar
umangyadav committed
303
304
305
306
307
    auto* else_mod = p.create_module("else_cpu_mod");
    std::vector<float> data2(ds.elements(), 2);
    auto l2    = else_mod->add_literal(migraphx::literal(ds, data2));
    auto cpu_y = else_mod->add_parameter("cpu_y", ds);
    auto a2    = else_mod->add_instruction(migraphx::make_op("mul"), cpu_y, l2);
308
309
310
311
    else_mod->add_return({a2});

    auto* run_on_cpu_mod = p.create_module("run_on_cpu");
    auto run_cpu_ins     = run_on_cpu_mod->add_instruction(
umangyadav's avatar
umangyadav committed
312
313
314
315
        migraphx::make_op("run_on_target", {{"target_id", 1}}), {y}, {else_mod});
    auto run_cpu_ins_0 = run_on_cpu_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_cpu_ins);
    run_on_cpu_mod->add_return({run_cpu_ins_0});
316
317
318

    auto* run_on_gpu_mod = p.create_module("run_on_gpu");
    auto run_gpu_ins     = run_on_gpu_mod->add_instruction(
umangyadav's avatar
umangyadav committed
319
320
321
322
        migraphx::make_op("run_on_target", {{"target_id", 0}}), {x}, {then_mod});
    auto run_gpu_ins_0 = run_on_gpu_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_gpu_ins);
    run_on_gpu_mod->add_return({run_gpu_ins_0});
323
324
325
326
327
328

    auto ret =
        mm->add_instruction(migraphx::make_op("if"), {cond}, {run_on_gpu_mod, run_on_cpu_mod});
    auto r = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), ret);
    mm->add_return({r});
    // compile
umangyadav's avatar
umangyadav committed
329
330
331
    migraphx::compile_options gpu_opts;
    gpu_opts.offload_copy = true;
    p.compile({migraphx::make_target("gpu"), migraphx::make_target("cpu")}, {gpu_opts});
332
    EXPECT(check_compiled_program(p, {migraphx::make_target("gpu"), migraphx::make_target("cpu")}));
umangyadav's avatar
umangyadav committed
333
    migraphx::parameter_map params;
umangyadav's avatar
umangyadav committed
334
335
    params["x"] = migraphx::fill_argument(ds, 2);
    params["y"] = migraphx::fill_argument(ds, 3);
umangyadav's avatar
umangyadav committed
336
337
338
339
    for(bool cond_val : {true, false})
    {
        params["cond"] = migraphx::argument(cond_s, &cond_val);
        auto result    = p.eval(params).back();
umangyadav's avatar
umangyadav committed
340
341
        auto gold      = migraphx::fill_argument(ds, (cond_val ? 3 : 6));
        EXPECT(gold == result);
umangyadav's avatar
umangyadav committed
342
    }
343
344
}

umangyadav's avatar
umangyadav committed
345
// TODO : FPGA compilation is broken right now, below test mentions fpga but doesn't compile for it
346
347
348
349
350
351
352
TEST_CASE(multitarget_compile_nested_if_then_else)
{
    std::unordered_map<std::size_t, std::size_t> counter_map = {{0, 0}, {1, 0}};
    migraphx::shape ds{migraphx::shape::float_type, {2, 3}};
    migraphx::program p;
    auto* mm = p.get_main_module();
    migraphx::shape cond_s{migraphx::shape::bool_type};
umangyadav's avatar
umangyadav committed
353
354
    auto cond_0             = mm->add_parameter("cond_0", cond_s);
    auto cond_1             = mm->add_parameter("cond_1", cond_s);
355
356
357
358
359
360
361
362
    auto x                  = mm->add_parameter("x", ds);
    auto y                  = mm->add_parameter("y", ds);
    auto z                  = mm->add_parameter("z", ds);
    auto create_test_module = [&](migraphx::program& prog,
                                  const std::vector<migraphx::instruction_ref>& inputs,
                                  std::size_t tid) {
        std::string mod_name =
            "target_" + std::to_string(tid) + "_" + std::to_string(counter_map[tid]++);
umangyadav's avatar
umangyadav committed
363
364
365
366
367
368
369
370
371
        auto* test_mod = prog.create_module(mod_name);
        std::vector<float> data(ds.elements(), -1);
        auto l1               = test_mod->add_literal(migraphx::literal(ds, data));
        auto test_mod_param_0 = test_mod->add_parameter(mod_name + "_param_0", ds);
        auto test_mod_param_1 = test_mod->add_parameter(mod_name + "_param_1", ds);
        auto test_mod_param_2 = test_mod->add_parameter(mod_name + "_param_2", ds);
        auto ins1 = test_mod->add_instruction(migraphx::make_op("add"), test_mod_param_0, l1);
        auto ins2 = test_mod->add_instruction(migraphx::make_op("mul"), ins1, test_mod_param_1);
        auto ins3 = test_mod->add_instruction(migraphx::make_op("sub"), ins2, test_mod_param_2);
372
373
        test_mod->add_return({ins3});
        auto* run_on_target_mod = prog.create_module("run_on_" + mod_name);
umangyadav's avatar
umangyadav committed
374
375
376
377
378
        auto run_ins            = run_on_target_mod->add_instruction(
            migraphx::make_op("run_on_target", {{"target_id", tid}}), inputs, {test_mod});
        auto run_ins_0 = run_on_target_mod->add_instruction(
            migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_ins);
        run_on_target_mod->add_return({run_ins_0});
379
380
381
382
383
384
385
386
387
388
389
390
391
        return run_on_target_mod;
    };

    // create nested module with multiple targets.
    // then_mod has one instruction that runs a module on "ref" and another instruction that
    // creates nested modules using "If" that runs on "cpu" and "gpu"
    auto* ref_mod = p.create_module("ref_mod");
    auto ref_x    = ref_mod->add_parameter("ref_x", ds);
    auto ref_y    = ref_mod->add_parameter("ref_y", ds);
    auto ref_add  = ref_mod->add_instruction(migraphx::make_op("add"), ref_x, ref_y);
    ref_mod->add_return({ref_add});

    auto* then_mod        = p.create_module("then_mod");
umangyadav's avatar
umangyadav committed
392
393
394
395
396
397
398
399
    auto then_mod_cond    = then_mod->add_parameter("then_mod_cond", cond_s);
    auto then_mod_param_0 = then_mod->add_parameter("then_mod_param_0", ds);
    auto then_mod_param_1 = then_mod->add_parameter("then_mod_param_1", ds);
    auto then_mod_param_2 = then_mod->add_parameter("then_mod_param_2", ds);
    auto then_mod_ref_ins =
        then_mod->add_instruction(migraphx::make_op("run_on_target", {{"target_id", 3}}),
                                  {then_mod_param_0, then_mod_param_1},
                                  {ref_mod});
400
401
    auto then_mod_ref_ins_0 = then_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), then_mod_ref_ins);
umangyadav's avatar
umangyadav committed
402
    auto then_mod_if = then_mod->add_instruction(
403
        migraphx::make_op("if"),
umangyadav's avatar
umangyadav committed
404
405
406
407
408
409
410
411
412
413
414
415
        {then_mod_cond,
         then_mod_param_0,
         then_mod_param_1,
         then_mod_param_2,
         then_mod_ref_ins_0,
         then_mod_param_1,
         then_mod_param_2},
        {create_test_module(p, {then_mod_param_0, then_mod_param_1, then_mod_param_2}, 1),
         create_test_module(p, {then_mod_ref_ins_0, then_mod_param_1, then_mod_param_2}, 0)});
    auto then_mod_if_0 =
        then_mod->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), then_mod_if);
    then_mod->add_return({then_mod_if_0});
416
417
418
419
420
421
422
423
424
425

    // create nested else_mod with multiple targets.
    // else_mod has one instruction that runs a module on "fpga" and another instruction that
    // creates nested modules using "If" that runs on "cpu" and "gpu"
    auto* fpga_mod = p.create_module("fpga_mod");
    auto fpga_x    = fpga_mod->add_parameter("fpga_x", ds);
    auto fpga_y    = fpga_mod->add_parameter("fpga_y", ds);
    auto fpga_add  = fpga_mod->add_instruction(migraphx::make_op("add"), fpga_x, fpga_y);
    fpga_mod->add_return({fpga_add});

umangyadav's avatar
umangyadav committed
426
427
428
429
430
431
432
433
434
    auto* else_mod        = p.create_module("else_mod");
    auto else_mod_cond    = else_mod->add_parameter("else_mod_cond", cond_s);
    auto else_mod_param_0 = else_mod->add_parameter("else_mod_param_0", ds);
    auto else_mod_param_1 = else_mod->add_parameter("else_mod_param_1", ds);
    auto else_mod_param_2 = else_mod->add_parameter("else_mod_param_2", ds);
    auto else_mod_fpga_ins =
        else_mod->add_instruction(migraphx::make_op("run_on_target", {{"target_id", 2}}),
                                  {else_mod_param_0, else_mod_param_2},
                                  {fpga_mod});
435
436
437
    auto else_mod_fpga_ins_0 = else_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), else_mod_fpga_ins);

umangyadav's avatar
umangyadav committed
438
439
440
441
442
443
444
445
446
447
448
449
450
451
    auto else_mod_if = else_mod->add_instruction(
        migraphx::make_op("if"),
        {else_mod_cond,
         else_mod_fpga_ins_0,
         else_mod_param_0,
         else_mod_param_1,
         else_mod_param_2,
         else_mod_param_1,
         else_mod_param_0},
        {create_test_module(p, {else_mod_fpga_ins_0, else_mod_param_0, else_mod_param_1}, 0),
         create_test_module(p, {else_mod_param_2, else_mod_param_1, else_mod_param_0}, 1)});
    auto else_mod_if_0 =
        else_mod->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), else_mod_if);
    else_mod->add_return({else_mod_if_0});
452
453

    // Create nested and multi-target main module using "If"
umangyadav's avatar
umangyadav committed
454
455
    auto main_if_ins = mm->add_instruction(
        migraphx::make_op("if"), {cond_0, cond_1, x, y, z, cond_1, x, y, z}, {then_mod, else_mod});
456
457
458
459
    auto r = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), main_if_ins);
    mm->add_return({r});

    // compile
umangyadav's avatar
umangyadav committed
460
461
462
    migraphx::compile_options gpu_opts;
    gpu_opts.offload_copy = true;

463
464
    p.compile({migraphx::make_target("gpu"),
               migraphx::make_target("cpu"),
umangyadav's avatar
umangyadav committed
465
466
467
               migraphx::make_target("ref"),
               migraphx::make_target("ref")},
              {gpu_opts});
468
469
470
    EXPECT(check_compiled_program(p,
                                  {migraphx::make_target("gpu"),
                                   migraphx::make_target("cpu"),
umangyadav's avatar
umangyadav committed
471
                                   migraphx::make_target("ref"),
472
                                   migraphx::make_target("ref")}));
umangyadav's avatar
umangyadav committed
473
474
    // do evaluation using different conditions
    migraphx::parameter_map params;
umangyadav's avatar
umangyadav committed
475
476
477
478
479
480
    float x_i   = 2.0;
    float y_i   = 3.0;
    float z_i   = 4.0;
    params["x"] = migraphx::fill_argument(ds, x_i);
    params["y"] = migraphx::fill_argument(ds, y_i);
    params["z"] = migraphx::fill_argument(ds, z_i);
umangyadav's avatar
umangyadav committed
481
482
483
484
485
486
487
488
489
490
491
    // cover all paths with different combination of conditions
    std::vector<std::pair<bool, bool>> test_conds = {
        {true, true}, {true, false}, {false, true}, {false, false}};
    for(auto [cond_val_0, cond_val_1] : test_conds)
    {
        params["cond_0"] = migraphx::argument(cond_s, &cond_val_0);
        params["cond_1"] = migraphx::argument(cond_s, &cond_val_1);
        auto result      = p.eval(params).back();
        // main has one instruction that is : if_then_else
        // then mod is doing : {tmp = x+y; (cond) ? (((x-1)*y)-z)  : (((tmp-1)*y)-z);}
        // else mod is doing : {tmp = x+z; (cond) ? (((tmp-1)*x)-y) : (((z-1)*y)-x);}
umangyadav's avatar
umangyadav committed
492
        float gold_i = -1.0;
umangyadav's avatar
umangyadav committed
493
494
        if(cond_val_0)
        {
umangyadav's avatar
umangyadav committed
495
496
            float tmp_i = x_i + y_i;
            gold_i      = (cond_val_1) ? (((x_i - 1) * y_i) - z_i) : (((tmp_i - 1) * y_i) - z_i);
umangyadav's avatar
umangyadav committed
497
498
499
        }
        else
        {
umangyadav's avatar
umangyadav committed
500
501
            float tmp_i = x_i + z_i;
            gold_i      = (cond_val_1) ? (((tmp_i - 1) * x_i) - y_i) : (((z_i - 1) * y_i) - x_i);
umangyadav's avatar
umangyadav committed
502
        }
umangyadav's avatar
umangyadav committed
503
504
        auto gold = migraphx::fill_argument(ds, gold_i);
        EXPECT(gold == result);
umangyadav's avatar
umangyadav committed
505
    }
506
507
}

umangyadav's avatar
umangyadav committed
508
// TODO : FPGA compilation is broken right now, below test mentions fpga but doesn't compile for it
509
510
511
512
513
514
515
516
TEST_CASE(multitarget_select_module)
{
    migraphx::program p;
    // create batch submodules
    auto create_submodule = [&](std::size_t batch_size, const std::string& module_name) {
        auto* submod = p.create_module(module_name);
        migraphx::shape sm_shape{migraphx::shape::float_type, {batch_size, 4}};
        auto sm_input = submod->add_parameter("data", sm_shape);
umangyadav's avatar
umangyadav committed
517
518
        migraphx::shape lit_s{migraphx::shape{migraphx::shape::float_type, {1}}};
        auto literal_ins = submod->add_literal(migraphx::literal{lit_s, {6}});
519
520
521
522
        auto broadcast_lit =
            submod->add_instruction(migraphx::make_op("multibroadcast"), literal_ins, sm_input);
        auto add_ins0 = submod->add_instruction(migraphx::make_op("add"), sm_input, broadcast_lit);
        auto add_ins1 = submod->add_instruction(migraphx::make_op("add"), add_ins0, broadcast_lit);
umangyadav's avatar
umangyadav committed
523
        submod->add_return({add_ins1});
524
525
526
527
528
529
530
531
        return submod;
    };
    auto* batch1 = create_submodule(1, "batch_1");
    auto* batch2 = create_submodule(2, "batch_2");
    auto* batch3 = create_submodule(3, "batch_3");
    auto* batch4 = create_submodule(4, "batch_4");

    auto* run_cpu_mod = p.create_module("cpu_mod");
umangyadav's avatar
umangyadav committed
532
533
    auto cpu_param =
        run_cpu_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {1, 4}});
534
535
    auto run_cpu_ins = run_cpu_mod->add_instruction(
        migraphx::make_op("run_on_target", {{"target_id", 1}}), {cpu_param}, {batch1});
umangyadav's avatar
umangyadav committed
536
537
538
    auto run_cpu_ins_0 = run_cpu_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_cpu_ins);
    run_cpu_mod->add_return({run_cpu_ins_0});
539
540

    auto* run_gpu_mod = p.create_module("gpu_mod");
umangyadav's avatar
umangyadav committed
541
542
    auto gpu_param =
        run_gpu_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {2, 4}});
543
544
    auto run_gpu_ins = run_gpu_mod->add_instruction(
        migraphx::make_op("run_on_target", {{"target_id", 0}}), {gpu_param}, {batch2});
umangyadav's avatar
umangyadav committed
545
546
547
    auto run_gpu_ins_0 = run_gpu_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_gpu_ins);
    run_gpu_mod->add_return({run_gpu_ins_0});
548
549

    auto* run_fpga_mod = p.create_module("fpga_mod");
umangyadav's avatar
umangyadav committed
550
551
    auto fpga_param =
        run_fpga_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {3, 4}});
552
553
    auto run_fpga_ins = run_fpga_mod->add_instruction(
        migraphx::make_op("run_on_target", {{"target_id", 2}}), {fpga_param}, {batch3});
umangyadav's avatar
umangyadav committed
554
555
556
    auto run_fpga_ins_0 = run_fpga_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_fpga_ins);
    run_fpga_mod->add_return({run_fpga_ins_0});
557
558

    auto* run_ref_mod = p.create_module("ref_mod");
umangyadav's avatar
umangyadav committed
559
560
    auto ref_param =
        run_ref_mod->add_parameter("data", migraphx::shape{migraphx::shape::float_type, {4, 4}});
561
562
    auto run_ref_ins = run_ref_mod->add_instruction(
        migraphx::make_op("run_on_target", {{"target_id", 3}}), {ref_param}, {batch4});
umangyadav's avatar
umangyadav committed
563
564
565
    auto run_ref_ins_0 = run_ref_mod->add_instruction(
        migraphx::make_op("get_tuple_elem", {{"index", 0}}), run_ref_ins);
    run_ref_mod->add_return({run_ref_ins_0});
566

umangyadav's avatar
umangyadav committed
567
568
569
    auto* mm = p.get_main_module();
    migraphx::shape dyn_s{migraphx::shape::float_type, {{1, 4}, {4, 4}}};
    auto input                              = mm->add_parameter("data", dyn_s);
570
571
572
573
574
575
576
577
578
    std::vector<migraphx::shape> sub_shapes = {};
    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {4, 4}}});
    sub_shapes.push_back(migraphx::shape{migraphx::shape::float_type, {{1, 4}, {4, 4}}});
    migraphx::shape out_attr = migraphx::shape{sub_shapes};
    auto sm_ins              = mm->add_instruction(
        migraphx::make_op("select_module", {{"output_dyn_shapes", migraphx::to_value(out_attr)}}),
        {input},
        {run_cpu_mod, run_gpu_mod, run_fpga_mod, run_ref_mod});
    auto ret0 = mm->add_instruction(migraphx::make_op("get_tuple_elem", {{"index", 0}}), sm_ins);
umangyadav's avatar
umangyadav committed
579
    mm->add_return({ret0});
580
    // compile
umangyadav's avatar
umangyadav committed
581
582
    migraphx::compile_options gpu_opts;
    gpu_opts.offload_copy = true;
583
584
    p.compile({migraphx::make_target("gpu"),
               migraphx::make_target("cpu"),
umangyadav's avatar
umangyadav committed
585
586
587
               migraphx::make_target("ref"),
               migraphx::make_target("ref")},
              {gpu_opts});
588
589
590
    EXPECT(check_compiled_program(p,
                                  {migraphx::make_target("gpu"),
                                   migraphx::make_target("cpu"),
umangyadav's avatar
umangyadav committed
591
                                   migraphx::make_target("ref"),
592
                                   migraphx::make_target("ref")}));
umangyadav's avatar
umangyadav committed
593
594
595
596
597
    // program does the 12+x where x has dynamic shape {{1, 4}, {4, 4}}
    for(const size_t bs : {1, 2, 3, 4})
    {
        migraphx::shape arg_shape{migraphx::shape::float_type, {bs, 4}};
        migraphx::parameter_map params;
umangyadav's avatar
umangyadav committed
598
599
600
601
602
603
604
605
        params["data"] = migraphx::generate_argument(arg_shape, arg_shape.elements());
        std::vector<float> input_data;
        params["data"].visit([&](const auto& vec) { input_data.assign(vec.begin(), vec.end()); });
        std::transform(input_data.begin(), input_data.end(), input_data.begin(), [](const auto& i) {
            return i + 12.0;
        });
        auto result = p.eval(params).back();
        EXPECT(migraphx::argument(arg_shape, input_data.data()) == result);
umangyadav's avatar
umangyadav committed
606
    }
607
608
609
}

int main(int argc, const char* argv[]) { test::run(argc, argv); }