test_vta_insn.py 20.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you 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.
17 18 19 20
"""Unit test VTA's instructions """
import tvm
import numpy as np
import topi
21
from tvm.contrib import util
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

import vta
import vta.testing
from vta.testing import simulator


def test_save_load_out():
    """Test save/store output command"""
    def _run(env, remote):
        n = 6
        x = tvm.placeholder(
            (n, n, env.BATCH, env.BLOCK_OUT),
            name="x",
            dtype=env.acc_dtype)
        x_buf = tvm.compute(
            (n, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: x(*i), "x_buf")
        # insert no-op that won't be optimized away
        y_buf = tvm.compute(
            (n, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: x_buf(*i)>>0, "y_buf")
        y = tvm.compute(
            (n, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: y_buf(*i).astype(env.inp_dtype), "y")
        # schedule
        s = tvm.create_schedule(y.op)
        s[x_buf].set_scope(env.acc_scope)
        s[x_buf].pragma(x_buf.op.axis[0], env.dma_copy)
        s[y_buf].set_scope(env.acc_scope)
        s[y_buf].pragma(y_buf.op.axis[0], env.alu)
        s[y].pragma(y.op.axis[0], env.dma_copy)

        # verification
        with vta.build_config():
            m = vta.build(s, [x, y], "ext_dev", env.target_host)

        if not remote:
            return
        temp = util.tempdir()
        m.save(temp.relpath("load_act.o"))
        remote.upload(temp.relpath("load_act.o"))
        f = remote.load_module("load_act.o")
        # verify
        ctx = remote.ext_dev(0)
        x_np = np.random.randint(
            1, 10, size=(n, n, env.BATCH, env.BLOCK_OUT)).astype(x.dtype)
        y_np = x_np.astype(y.dtype)
        x_nd = tvm.nd.array(x_np, ctx)
        y_nd = tvm.nd.empty(y_np.shape, ctx=ctx, dtype=y_np.dtype)
71

72 73
        if env.TARGET in ["sim", "tsim"]:
            simulator.clear_stats()
74

75
        f(x_nd, y_nd)
76

77 78
        np.testing.assert_equal(y_np, y_nd.asnumpy())

79 80 81 82 83
        if env.TARGET in ["sim", "tsim"]:
            sim_stats = simulator.stats()
            print("Save load execution statistics:")
            for k, v in sim_stats.items():
                print("\t{:<16}: {:>16}".format(k, v))
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
    vta.testing.run(_run)


def test_padded_load():
    """Test padded load."""
    def _run(env, remote):
        # declare
        n = 21
        m = 20
        pad_before = [0, 1, 0, 0]
        pad_after = [1, 3, 0, 0]
        x = tvm.placeholder(
            (n, m, env.BATCH, env.BLOCK_OUT),
            name="x",
            dtype=env.acc_dtype)
        x_buf = topi.nn.pad(x, pad_before, pad_after, name="y")
        # insert no-op that won't be optimized away
        y_buf = tvm.compute((n + pad_before[0] + pad_after[0],
                             m + pad_before[1] + pad_after[1],
                             env.BATCH,
                             env.BLOCK_OUT), lambda *i: x_buf(*i)>>0, "y_buf")
        y = tvm.compute((n + pad_before[0] + pad_after[0],
                         m + pad_before[1] + pad_after[1],
                         env.BATCH,
                         env.BLOCK_OUT), lambda *i: y_buf(*i).astype(env.inp_dtype), "y")
        # schedule
        s = tvm.create_schedule(y.op)
        s[x_buf].set_scope(env.acc_scope)
        s[x_buf].pragma(x_buf.op.axis[0], env.dma_copy)
        s[y_buf].set_scope(env.acc_scope)
        s[y_buf].pragma(y_buf.op.axis[0], env.alu)
        s[y].pragma(y.op.axis[0], env.dma_copy)
        # build
        with vta.build_config():
            mod = vta.build(s, [x, y], "ext_dev", env.target_host)

        if not remote:
            return
        temp = util.tempdir()
        mod.save(temp.relpath("padded_load.o"))
        remote.upload(temp.relpath("padded_load.o"))
        f = remote.load_module("padded_load.o")
        # verify
        ctx = remote.ext_dev(0)
        x_np = np.random.randint(1, 2, size=(
            n, m, env.BATCH, env.BLOCK_OUT)).astype(x.dtype)
        y_np = np.zeros((n + pad_before[0] + pad_after[0],
                         m + pad_before[1] + pad_after[1],
                         env.BATCH,
                         env.BLOCK_OUT)).astype(y.dtype)
        y_np[pad_before[0]:pad_before[0] + n,
             pad_before[1]:pad_before[1] + m,
             :] = x_np
        x_nd = tvm.nd.array(x_np, ctx)
        y_nd = tvm.nd.empty(y_np.shape, ctx=ctx, dtype=y_np.dtype)
140

141 142
        if env.TARGET in ["sim", "tsim"]:
            simulator.clear_stats()
143

144
        f(x_nd, y_nd)
145

146 147
        np.testing.assert_equal(y_np, y_nd.asnumpy())

148 149 150 151 152
        if env.TARGET in ["sim", "tsim"]:
            sim_stats = simulator.stats()
            print("Padded load execution statistics:")
            for k, v in sim_stats.items():
                print("\t{:<16}: {:>16}".format(k, v))
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
    vta.testing.run(_run)


def test_gemm():
    """Test GEMM."""
    def _run(env, remote):
        # declare
        o = 4
        n = 1
        m = 4
        x = tvm.placeholder((o, n, env.BATCH, env.BLOCK_IN), name="x", dtype=env.inp_dtype)
        w = tvm.placeholder((m, n, env.BLOCK_OUT, env.BLOCK_IN), name="w", dtype=env.wgt_dtype)
        x_buf = tvm.compute((o, n, env.BATCH, env.BLOCK_IN), lambda *i: x(*i), "x_buf")
        w_buf = tvm.compute((m, n, env.BLOCK_OUT, env.BLOCK_IN), lambda *i: w(*i), "w_buf")
        ko = tvm.reduce_axis((0, n), name="ko")
        ki = tvm.reduce_axis((0, env.BLOCK_IN), name="ki")
        y_gem = tvm.compute(
            (o, m, env.BATCH, env.BLOCK_OUT),
            lambda bo, co, bi, ci:
            tvm.sum(x_buf[bo, ko, bi, ki].astype(env.acc_dtype) *
                    w_buf[co, ko, ci, ki].astype(env.acc_dtype),
                    axis=[ko, ki]),
            name="y_gem")
        y_shf = tvm.compute(
            (o, m, env.BATCH, env.BLOCK_OUT),
            lambda *i: y_gem(*i)>>8,
            name="y_shf")
        y_max = tvm.compute(
            (o, m, env.BATCH, env.BLOCK_OUT),
            lambda *i: tvm.max(y_shf(*i), 0),
            "y_max") #relu
        y_min = tvm.compute(
            (o, m, env.BATCH, env.BLOCK_OUT),
            lambda *i: tvm.min(y_max(*i), (1<<(env.INP_WIDTH-1))-1),
            "y_min") #relu
        y = tvm.compute(
            (o, m, env.BATCH, env.BLOCK_OUT),
            lambda *i: y_min(*i).astype(env.inp_dtype),
            name="y")

        if not remote:
            return

197
        def verify(s, name=None):
198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
            mod = vta.build(s, [x, w, y], "ext_dev", env.target_host)
            temp = util.tempdir()
            mod.save(temp.relpath("gemm.o"))
            remote.upload(temp.relpath("gemm.o"))
            f = remote.load_module("gemm.o")
            # verify
            ctx = remote.ext_dev(0)
            x_np = np.random.randint(
                -128, 128, size=(o, n, env.BATCH, env.BLOCK_IN)).astype(x.dtype)
            w_np = np.random.randint(
                -128, 128, size=(m, n, env.BLOCK_OUT, env.BLOCK_IN)).astype(w.dtype)
            y_np = np.zeros((o, m, env.BATCH, env.BLOCK_OUT)).astype(y.dtype)
            x_nd = tvm.nd.array(x_np, ctx)
            w_nd = tvm.nd.array(w_np, ctx)
            y_nd = tvm.nd.array(y_np, ctx)
            y_np = y_np.astype(env.acc_dtype)
            for b in range(o):
                for i in range(m):
                    for j in range(n):
                        y_np[b,i,:] += np.dot(x_np[b,j,:].astype(env.acc_dtype),
                                              w_np[i,j].T.astype(env.acc_dtype))
            y_np = np.right_shift(y_np, 8)
            y_np = np.clip(y_np, 0, (1<<(env.INP_WIDTH-1))-1).astype(y.dtype)

222
            if env.TARGET in ["sim", "tsim"]:
223
                simulator.clear_stats()
224 225

            f(x_nd, w_nd, y_nd)
226 227 228

            np.testing.assert_equal(y_np, y_nd.asnumpy())

229 230 231 232 233
            if env.TARGET in ["sim", "tsim"]:
                sim_stats = simulator.stats()
                print("GEMM schedule:{} execution statistics:".format(name))
                for k, v in sim_stats.items():
                    print("\t{:<16}: {:>16}".format(k, v))
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
        def test_schedule1():
            # default schedule with no smt
            s = tvm.create_schedule(y.op)
            # set the scope of the SRAM buffers
            s[x_buf].set_scope(env.inp_scope)
            s[w_buf].set_scope(env.wgt_scope)
            s[y_gem].set_scope(env.acc_scope)
            s[y_shf].set_scope(env.acc_scope)
            s[y_max].set_scope(env.acc_scope)
            s[y_min].set_scope(env.acc_scope)
            # set pragmas for DMA transfer and ALU ops
            s[x_buf].compute_at(s[y_gem], ko)
            s[x_buf].pragma(s[x_buf].op.axis[0], env.dma_copy)
            s[w_buf].compute_at(s[y_gem], ko)
            s[w_buf].pragma(s[w_buf].op.axis[0], env.dma_copy)
            s[y_shf].pragma(s[y_shf].op.axis[0], env.alu)
            s[y_max].pragma(s[y_max].op.axis[0], env.alu)
            s[y_min].pragma(s[y_min].op.axis[0], env.alu)
            s[y].pragma(s[y].op.axis[0], env.dma_copy)
            # tensorization
            s[y_gem].reorder(
                ko,
                s[y_gem].op.axis[0],
                s[y_gem].op.axis[1],
                s[y_gem].op.axis[2],
                s[y_gem].op.axis[3],
                ki)
            s[y_gem].tensorize(s[y_gem].op.axis[2], env.gemm)
263
            verify(s, name="default")
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

        def test_smt():
            # test smt schedule
            s = tvm.create_schedule(y.op)
            s[x_buf].set_scope(env.inp_scope)
            s[w_buf].set_scope(env.wgt_scope)
            s[y_gem].set_scope(env.acc_scope)
            s[y_shf].set_scope(env.acc_scope)
            s[y_max].set_scope(env.acc_scope)
            s[y_min].set_scope(env.acc_scope)
            abo, aco, abi, aci = s[y].op.axis
            abo1, abo2 = s[y].split(abo, nparts=2)
            s[y].bind(abo1, tvm.thread_axis("cthread"))
            s[y_gem].compute_at(s[y], abo1)
            s[y_shf].compute_at(s[y], abo1)
            s[y_max].compute_at(s[y], abo1)
            s[y_min].compute_at(s[y], abo1)
            s[y_gem].reorder(
                ko,
                s[y_gem].op.axis[0],
                s[y_gem].op.axis[1],
                s[y_gem].op.axis[2],
                s[y_gem].op.axis[3],
                ki)
            s[y_gem].tensorize(s[y_gem].op.axis[2], env.gemm)
            s[y_shf].pragma(s[y_shf].op.axis[0], env.alu)
            s[y_max].pragma(s[y_max].op.axis[0], env.alu)
            s[y_min].pragma(s[y_min].op.axis[0], env.alu)
            s[x_buf].compute_at(s[y_gem], ko)
            s[x_buf].pragma(s[x_buf].op.axis[0], env.dma_copy)
            s[w_buf].compute_at(s[y_gem], ko)
            s[w_buf].pragma(s[w_buf].op.axis[0], env.dma_copy)
            s[y].pragma(abo2, env.dma_copy)
297
            verify(s, name="smt")
298 299 300 301 302 303 304 305

        test_schedule1()
        test_smt()
    vta.testing.run(_run)


def test_alu():
    def _run(env, remote):
306
        def check_alu(tvm_op, np_op=None, use_imm=False, test_name=None):
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
            """Test ALU"""
            m = 8
            n = 8
            imm = np.random.randint(1,5)
            # compute
            a = tvm.placeholder(
                (m, n, env.BATCH, env.BLOCK_OUT),
                name="a",
                dtype=env.acc_dtype)
            a_buf = tvm.compute(
                (m, n, env.BATCH, env.BLOCK_OUT),
                lambda *i: a(*i),
                "a_buf") #DRAM->SRAM
            if use_imm:
                res_buf = tvm.compute(
                    (m, n, env.BATCH, env.BLOCK_OUT),
                    lambda *i: tvm_op(a_buf(*i), imm),
                    "res_buf") #compute
            else:
                b = tvm.placeholder(
                    (m, n, env.BATCH, env.BLOCK_OUT),
                    name="b",
                    dtype=env.acc_dtype)
                b_buf = tvm.compute(
                    (m, n, env.BATCH, env.BLOCK_OUT),
                    lambda *i: b(*i),
                    "b_buf") #DRAM->SRAM
                res_buf = tvm.compute(
                    (m, n, env.BATCH, env.BLOCK_OUT),
                    lambda *i: tvm_op(a_buf(*i), b_buf(*i)),
                    "res_buf") #compute5B
            res = tvm.compute(
                (m, n, env.BATCH, env.BLOCK_OUT),
                lambda *i: res_buf(*i).astype(env.inp_dtype),
                "res") #SRAM->DRAM
            # schedule
            s = tvm.create_schedule(res.op)
            s[a_buf].set_scope(env.acc_scope) # SRAM
            s[a_buf].pragma(a_buf.op.axis[0], env.dma_copy) # DRAM->SRAM
            s[res_buf].set_scope(env.acc_scope) # SRAM
            s[res_buf].pragma(res_buf.op.axis[0], env.alu) # compute
            s[res].pragma(res.op.axis[0], env.dma_copy) # SRAM->DRAM
            if not use_imm:
                s[b_buf].set_scope(env.acc_scope) # SRAM
                s[b_buf].pragma(b_buf.op.axis[0], env.dma_copy) # DRAM->SRAM

            if not remote:
                return

            # build
            with vta.build_config():
                if use_imm:
                    mod = vta.build(s, [a, res], "ext_dev", env.target_host)
                else:
                    mod = vta.build(s, [a, b, res], "ext_dev", env.target_host)
            temp = util.tempdir()
            mod.save(temp.relpath("load_act.o"))
            remote.upload(temp.relpath("load_act.o"))
            f = remote.load_module("load_act.o")
            # verify
            ctx = remote.ext_dev(0)
            a_np = np.random.randint(
                -16, 16, size=(m, n, env.BATCH, env.BLOCK_OUT)).astype(a.dtype)
            if use_imm:
                res_np = np_op(a_np, imm) if np_op else tvm_op(a_np, imm)
            else:
                b_np = np.random.randint(
                    -16, 16, size=(m, n, env.BATCH, env.BLOCK_OUT)).astype(b.dtype)
                res_np = np_op(a_np, b_np) if np_op else tvm_op(a_np, b_np)
            res_np = res_np.astype(res.dtype)
            a_nd = tvm.nd.array(a_np, ctx)
            res_nd = tvm.nd.array(
                np.zeros((m, n, env.BATCH, env.BLOCK_OUT)).astype(res.dtype), ctx)
380

381 382
            if env.TARGET in ["sim", "tsim"]:
                simulator.clear_stats()
383

384 385 386 387 388
            if use_imm:
                f(a_nd, res_nd)
            else:
                b_nd = tvm.nd.array(b_np, ctx)
                f(a_nd, b_nd, res_nd)
389

390 391
            np.testing.assert_equal(res_np, res_nd.asnumpy())

392 393 394 395 396
            if env.TARGET in ["sim", "tsim"]:
                sim_stats = simulator.stats()
                print("ALU {} execution statistics:".format(test_name))
                for k, v in sim_stats.items():
                    print("\t{:<16}: {:>16}".format(k, v))
397 398 399 400 401 402 403

        check_alu(lambda x, y: x << y, np.left_shift, use_imm=True, test_name="SHL")
        check_alu(tvm.max, np.maximum, use_imm=True, test_name="MAX")
        check_alu(tvm.max, np.maximum, test_name="MAX")
        check_alu(lambda x, y: x + y, use_imm=True, test_name="ADD")
        check_alu(lambda x, y: x + y, test_name="ADD")
        check_alu(lambda x, y: x >> y, np.right_shift, use_imm=True, test_name="SHR")
404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459

    vta.testing.run(_run)


def test_relu():
    """Test RELU on ALU"""
    def _run(env, remote):
        m = 8
        n = 10
        # compute
        a = tvm.placeholder(
            (m, n, env.BATCH, env.BLOCK_OUT),
            name="a",
            dtype=env.acc_dtype)
        a_buf = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: a(*i),
            "a_buf") # DRAM->SRAM
        max_buf = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: tvm.max(a_buf(*i), 0),
            "res_buf") # relu
        min_buf = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: tvm.min(max_buf(*i), (1<<(env.INP_WIDTH-1))-1),
            "max_buf") # relu
        res = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: min_buf(*i).astype(env.inp_dtype),
            "min_buf") # SRAM->DRAM
        # schedule
        s = tvm.create_schedule(res.op)
        s[a_buf].set_scope(env.acc_scope) # SRAM
        s[a_buf].pragma(a_buf.op.axis[0], env.dma_copy) # DRAM->SRAM
        s[max_buf].set_scope(env.acc_scope) # SRAM
        s[min_buf].set_scope(env.acc_scope) # SRAM
        s[max_buf].pragma(max_buf.op.axis[0], env.alu) # compute
        s[min_buf].pragma(min_buf.op.axis[0], env.alu) # compute
        s[res].pragma(res.op.axis[0], env.dma_copy) # SRAM->DRAM
        # build
        with vta.build_config():
            mod = vta.build(s, [a, res], "ext_dev", env.target_host)
        if not remote:
            return
        temp = util.tempdir()
        mod.save(temp.relpath("load_act.o"))
        remote.upload(temp.relpath("load_act.o"))
        f = remote.load_module("load_act.o")
        # verify
        ctx = remote.ext_dev(0)
        a_np = np.random.randint(
            -256, 256, size=(m, n, env.BATCH, env.BLOCK_OUT)).astype(a.dtype)
        res_np = np.clip(a_np, 0, (1<<(env.INP_WIDTH-1))-1).astype(res.dtype)
        a_nd = tvm.nd.array(a_np, ctx)
        res_nd = tvm.nd.array(
            np.zeros((m, n, env.BATCH, env.BLOCK_OUT)).astype(res.dtype), ctx)
460

461 462
        if env.TARGET in ["sim", "tsim"]:
            simulator.clear_stats()
463

464
        f(a_nd, res_nd)
465

466 467
        np.testing.assert_equal(res_np, res_nd.asnumpy())

468 469 470 471 472
        if env.TARGET in ["sim", "tsim"]:
            sim_stats = simulator.stats()
            print("Relu execution statistics:")
            for k, v in sim_stats.items():
                print("\t{:<16}: {:>16}".format(k, v))
473

474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529
    vta.testing.run(_run)


def test_shift_and_scale():
    """Test shift and scale on ALU"""
    def _run(env, remote):
        m = 2
        n = 8
        imm_shift = np.random.randint(0,8)
        imm_scale = np.random.randint(1,5)
        # compute
        a = tvm.placeholder(
            (m, n, env.BATCH, env.BLOCK_OUT),
            name="a", dtype=env.acc_dtype)
        a_buf = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: a(*i),
            "a_buf") # DRAM->SRAM
        res_shift = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: a_buf(*i)+imm_shift,
            "res_shift") # compute
        res_scale = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: res_shift(*i)>>imm_scale,
            "res_scale") # compute
        res = tvm.compute(
            (m, n, env.BATCH, env.BLOCK_OUT),
            lambda *i: res_scale(*i).astype(env.inp_dtype),
            "res") # SRAM->DRAM
        # schedule
        s = tvm.create_schedule(res.op)
        s[a_buf].set_scope(env.acc_scope) # SRAM
        s[res_shift].set_scope(env.acc_scope) # SRAM
        s[res_scale].set_scope(env.acc_scope) # SRAM
        s[a_buf].pragma(a_buf.op.axis[0], env.dma_copy) # DRAM->SRAM
        s[res_shift].pragma(res_shift.op.axis[0], env.alu) # compute
        s[res_scale].pragma(res_scale.op.axis[0], env.alu) # compute
        s[res].pragma(res.op.axis[0], env.dma_copy) # SRAM->DRAM
        # build
        mod = vta.build(s, [a, res], "ext_dev", env.target_host)
        if not remote:
            return
        temp = util.tempdir()
        mod.save(temp.relpath("load_act.o"))
        remote.upload(temp.relpath("load_act.o"))
        f = remote.load_module("load_act.o")
        # verify
        ctx = remote.ext_dev(0)
        a_np = np.random.randint(
            -10, 10, size=(m, n, env.BATCH, env.BLOCK_OUT)).astype(a.dtype)
        res_np = np.right_shift((a_np + imm_shift), imm_scale)
        res_np = res_np.astype(res.dtype)
        a_nd = tvm.nd.array(a_np, ctx)
        res_nd = tvm.nd.array(
            np.zeros((m, n, env.BATCH, env.BLOCK_OUT)).astype(res.dtype), ctx)
530

531 532
        if env.TARGET in ["sim", "tsim"]:
            simulator.clear_stats()
533

534
        f(a_nd, res_nd)
535

536 537
        np.testing.assert_equal(res_np, res_nd.asnumpy())

538 539 540 541 542
        if env.TARGET in ["sim", "tsim"]:
            sim_stats = simulator.stats()
            print("Shift and scale execution statistics:")
            for k, v in sim_stats.items():
                print("\t{:<16}: {:>16}".format(k, v))
543

544 545
    vta.testing.run(_run)

546 547 548 549 550 551 552 553 554 555 556 557 558

def test_runtime_array():
    def _run(env, remote):
        n = 100
        ctx = remote.ext_dev(0)
        x_np = np.random.randint(
            1, 10, size=(n, n, env.BATCH, env.BLOCK_OUT)).astype("int8")
        x_nd = tvm.nd.array(x_np, ctx)
        np.testing.assert_equal(x_np, x_nd.asnumpy())

    vta.testing.run(_run)


559
if __name__ == "__main__":
560
    test_runtime_array()
561
    test_save_load_out()
562
    test_padded_load()
563
    test_gemm()
564
    test_alu()
565 566
    test_relu()
    test_shift_and_scale()