[MXNET-537] add_n(dense, csr, dense) = dense and add_n([dense, csr, rsp]*, dense, [dense, csr, rsp]*) = dense on CPU & GPU

[haojin2]

Description

As title

Checklist

Essentials

• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http:https://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• support for add_n(dense, csr, dense) = dense on CPU & GPU
• support for add_n(any combination longer than 4 with at least one dense storage) = dense on CPU & GPU
• unit tests

Comments

Also comes with a optimized GPU kernel for elemwise_add/sub(dns, csr)/elemwise_add/sub(csr, dns), for benchmark results please see comments.

[haojin2]

@eric-haibin-lin

[haojin2]

Benchmark results for warp-optimized GPU kernel for elemwise_add/sub(dense, csr):
([density%] [new speedup of write inplace] [old speedup of write inplace])
([density%] [new speedup of write to] [old speedup of write to])
1.00% 4.3422253664233255 1.1133807433946643
1.00% 1.8064753025920386 1.1127745540337441
0.50% 8.719801584535675 1.2989243065699914
0.50% 2.2845434302137857 1.2954892083078022
0.10% 51.95314630061374 1.4716730306016637
0.10% 2.878010179453661 1.4621131161544634
0.05% 90.41164608500259 1.4950209892594164
0.05% 2.9590177057354445 1.494533324414405
0.01% 165.45560871876663 1.5066270652228635
0.01% 2.9965883337578574 1.4932449464071242
benchmark script is the same one used in #10550

[haojin2]

Benchmark result for add_n(dense, csr, dense) = dense:
([density%] [speedup])
CPU:
1.00% 1.1282194997074237
0.50% 1.1686160529139418
0.10% 1.1909255730224886
0.05% 1.1970586102280831
0.01% 1.202483677804412
GPU:
1.00% 1.1627124767202126
0.50% 1.2392510678426578
0.10% 1.3169708612264934
0.05% 1.3275811285384644
0.01 % 1.3358768672033845
benchmark script:

import mxnet as mx
import sys
import os
import scipy
import numpy as np
from mxnet.test_utils import rand_ndarray, assert_almost_equal
import time

def measure_cost(repeat, a, b, c, out=None):
    # start bench
    start = time.time()
    results = []
    for i in range(repeat):
        results.append(mx.nd.sparse.add_n(a, b, c, out=out))
    for result in results:
        result.wait_to_read()
    end = time.time()
    diff = end - start
    return diff / repeat

def measure_fallback(repeat, a):
    # start bench
    start = time.time()
    results = []
    for i in range(repeat):
        results.append(a.tostype('default'))
    for result in results:
        result.wait_to_read()
    end = time.time()
    diff = end - start
    return diff / repeat

def main():
    shape = (128, 1000000)
    dns = np.random.uniform(size=shape)
    # context = mx.gpu(0)
    context = mx.cpu()
    mx_dns1 = mx.nd.array(dns, ctx=context)
    mx_dns2 = mx.nd.array(dns, ctx=context)
    for density in [0.01, 0.005, 0.001, 0.0005, 0.0001]:
        mx_csr = rand_ndarray(shape=shape, stype='csr', density=density).as_in_context(context)
        mx_csr_dns = mx_csr.tostype('default')
        sparse_cost = 0.0
        dns_cost = 0.0
        mx.nd.waitall()
        #warmup
        check = mx.nd.sparse.add_n(mx_dns1, mx_csr, mx_dns2)
        dns1 = dns + mx_csr_dns.asnumpy() + dns
        assert_almost_equal(check.asnumpy(), dns1, atol=1e-5, rtol=1e-4)
        mx.nd.waitall()
        for i in range(20):
            sparse_cost += measure_cost(5, mx_dns1, mx_csr, mx_dns2)
            dns_cost += measure_cost(5, mx_dns1, mx_csr_dns, mx_dns2)
        print("%.2f %%" % (density*100), dns_cost / sparse_cost)


if __name__ == "__main__":
    main()

[haojin2]

Benchmark result for add_n(more than 4 inputs with at least 1 dense) = dense (the combination being benchmarked here is add_n(dense, csr, dense, rsp, dense) = dense):
([density%] [speedup])
CPU:
1.00% 1.4248320861874664
0.50% 1.4591373125830511
0.10% 1.487516900293522
0.05% 1.4891773584928327
0.01% 1.4833875047500007
GPU:
1.00% 1.5829503717448206
0.50% 1.612348854910054
0.10% 1.6657770987040201
0.05% 1.6743607944367647
0.01% 1.6844786052948375
Benchmark script:

import mxnet as mx
import sys
import os
import scipy
import numpy as np
from mxnet.test_utils import rand_ndarray, assert_almost_equal
import time

def measure_cost(repeat, a, b, c, d, e, out=None):
    # start bench
    start = time.time()
    results = []
    for i in range(repeat):
        results.append(mx.nd.sparse.add_n(a, b, c, d, e, out=out))
    for result in results:
        result.wait_to_read()
    end = time.time()
    diff = end - start
    return diff / repeat

def measure_fallback(repeat, a):
    # start bench
    start = time.time()
    results = []
    for i in range(repeat):
        results.append(a.tostype('default'))
    for result in results:
        result.wait_to_read()
    end = time.time()
    diff = end - start
    return diff / repeat

def main():
    shape = (1000000, 128)
    dns = np.random.uniform(size=shape)
    context = mx.gpu(0)
    # context = mx.cpu()
    mx_dns1 = mx.nd.array(dns, ctx=context)
    mx_dns2 = mx.nd.array(dns, ctx=context)
    mx_dns3 = mx.nd.array(dns, ctx=context)
    for density in [0.01, 0.005, 0.001, 0.0005, 0.0001]:
        mx_csr = rand_ndarray(shape=shape, stype='csr', density=density).as_in_context(context)
        mx_csr_dns = mx_csr.tostype('default')
        mx_rsp = rand_ndarray(shape=shape, stype='row_sparse', density=density).as_in_context(context)
        mx_rsp_dns = mx_rsp.tostype('default')
        sparse_cost = 0.0
        dns_cost = 0.0
        mx.nd.waitall()
        #warmup
        check = mx.nd.sparse.add_n(mx_dns1, mx_csr, mx_rsp, mx_dns2, mx_dns3)
        dns1 = dns + mx_csr_dns.asnumpy() + mx_rsp_dns.asnumpy() + dns + dns
        assert_almost_equal(check.asnumpy(), dns1, atol=1e-5, rtol=1e-4)
        mx.nd.waitall()
        for i in range(20):
            sparse_cost += measure_cost(5, mx_dns1, mx_csr, mx_dns2, mx_rsp, mx_dns3)
            dns_cost += measure_cost(5, mx_dns1, mx_csr_dns, mx_dns2, mx_rsp_dns, mx_dns3)
        print("%.2f %%" % (density*100), dns_cost / sparse_cost)


if __name__ == "__main__":
    main()

[eric-haibin-lin]

Suggest use a const var with meaningful name instead of 32

[haojin2]

@eric-haibin-lin build passed, should be good for merge.