tools/generate-spmm-test.py

#!/usr/bin/env python
# Copyright 2019 Google LLC
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

import argparse
import bisect
import codecs
import os
import sys
import yaml

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from primes import next_prime
import xngen


parser = argparse.ArgumentParser(description='XNNPACK generator')
parser.add_argument("-s", "--spec", metavar="FILE", required=True,
                    help="Spec (YAML) file")
parser.add_argument("-o", "--output", metavar="FILE", required=True,
                    help='Output (C++ source) file')
parser.set_defaults(defines=list())


def indent(text):
  return "\n".join(map(lambda t: "  " + t if t else t, text.splitlines()))


def remove_duplicate_newlines(text):
  filtered_lines = list()
  last_newline = False
  for line in text.splitlines():
    is_newline = len(line.strip()) == 0
    if not is_newline or not last_newline:
      filtered_lines.append(line)
    last_newline = is_newline
  return "\n".join(filtered_lines)


ARCH_TO_MACRO_MAP = {
  "aarch32": "CPUINFO_ARCH_ARM",
  "aarch64": "CPUINFO_ARCH_ARM64",
  "x86": "CPUINFO_ARCH_X86",
  "x86-64": "CPUINFO_ARCH_X86_64",
}

ISA_TO_ARCH_MAP = {
  "neon": ["aarch32", "aarch64"],
  "neonfma": ["aarch32", "aarch64"],
  "neonfp16arith": ["aarch32", "aarch64"],
  "sse": ["x86", "x86-64"],
  "sse2": ["x86", "x86-64"],
  "avx": ["x86", "x86-64"],
  "avx512f": ["x86", "x86-64"],
  "psimd": [],
}

ISA_TO_CHECK_MAP = {
  "neon": "TEST_REQUIRES_ARM_NEON",
  "neonfma": "TEST_REQUIRES_ARM_NEON_FMA",
  "neonfp16arith": "TEST_REQUIRES_ARM_NEON_FP16_ARITH",
  "sse": "TEST_REQUIRES_X86_SSE",
  "sse2": "TEST_REQUIRES_X86_SSE2",
  "avx": "TEST_REQUIRES_X86_AVX",
  "avx512f": "TEST_REQUIRES_X86_AVX512F",
  "psimd": "TEST_REQUIRES_PSIMD",
}


def split_ukernel_name(name):
  common_name, target_name = name.split("__", 1)
  common_parts = common_name.split("_")
  param_spec = common_parts[-1]
  mr, nr = map(int, param_spec.split("x"))
  arch = list()
  isa = None
  for target_part in target_name.split("_"):
    if target_part in ARCH_TO_MACRO_MAP:
      arch = [target_part]
    elif target_part in ISA_TO_ARCH_MAP:
      isa = target_part
  if isa and not arch:
    arch = ISA_TO_ARCH_MAP[isa]
  return mr, nr, arch, isa


TEST_TEMPLATE = """\
TEST(${TEST_NAME}, k_eq_${KBLOCK}) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  SpMMMicrokernelTester()
    .mr(${MR})
    .nr(${NR})
    .m(${MR})
    .n(${NR})
    .k(${KBLOCK})
    .sparsity(0.0f)
    .Test(${", ".join(TEST_ARGS)});
}

$if NR > 1:
  TEST(${TEST_NAME}, k_eq_${KBLOCK}_subtile) {
    $if ISA_CHECK:
      ${ISA_CHECK};
    for (uint32_t n = 1; n <= ${NR}; n++) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR})
        .n(n)
        .k(${KBLOCK})
        .sparsity(0.0f)
        .Test(${", ".join(TEST_ARGS)});
    }
  }

$if IS_PIPELINED:
  TEST(${TEST_NAME}, k_eq_${KBLOCK * 2}) {
    $if ISA_CHECK:
      ${ISA_CHECK};
    SpMMMicrokernelTester()
      .mr(${MR})
      .nr(${NR})
      .m(${MR})
      .n(${NR})
      .k(${KBLOCK * 2})
      .sparsity(0.0f)
      .Test(${", ".join(TEST_ARGS)});
  }

  $if NR > 1:
    TEST(${TEST_NAME}, k_eq_${KBLOCK * 2}_subtile) {
      $if ISA_CHECK:
        ${ISA_CHECK};
      for (uint32_t n = 1; n <= ${NR}; n++) {
        SpMMMicrokernelTester()
          .mr(${MR})
          .nr(${NR})
          .m(${MR})
          .n(n)
          .k(${KBLOCK * 2})
          .sparsity(0.0f)
          .Test(${", ".join(TEST_ARGS)});
      }
    }

$if KBLOCK > 1:
  TEST(${TEST_NAME}, k_lt_${ADJKBLOCK}) {
    $if ISA_CHECK:
      ${ISA_CHECK};
    for (size_t k = 1; k < ${ADJKBLOCK}; k++) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR})
        .n(${NR})
        .k(k)
        .sparsity(0.0f)
        .Test(${", ".join(TEST_ARGS)});
    }
  }

  $if NR > 1:
    TEST(${TEST_NAME}, k_lt_${ADJKBLOCK}_subtile) {
      $if ISA_CHECK:
        ${ISA_CHECK};
      for (size_t k = 1; k < ${ADJKBLOCK}; k++) {
        for (uint32_t n = 1; n <= ${NR}; n++) {
          SpMMMicrokernelTester()
            .mr(${MR})
            .nr(${NR})
            .m(${MR})
            .n(n)
            .k(k)
            .sparsity(0.0f)
            .Test(${", ".join(TEST_ARGS)});
        }
      }
    }

TEST(${TEST_NAME}, k_gt_${ADJKBLOCK}) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (size_t k = ${ADJKBLOCK + 1}; k < ${KBLOCK * 10 if KBLOCK == 1 else KBLOCK * 2}; k++) {
    SpMMMicrokernelTester()
      .mr(${MR})
      .nr(${NR})
      .m(${MR})
      .n(${NR})
      .k(k)
      .sparsity(0.0f)
      .Test(${", ".join(TEST_ARGS)});
  }
}

$if NR > 1:
  TEST(${TEST_NAME}, k_gt_${KBLOCK}_subtile) {
    $if ISA_CHECK:
      ${ISA_CHECK};
    for (size_t k = ${ADJKBLOCK + 1}; k < ${10 if KBLOCK == 1 else KBLOCK * 2}; k++) {
      for (uint32_t n = 1; n <= ${NR}; n++) {
        SpMMMicrokernelTester()
          .mr(${MR})
          .nr(${NR})
          .m(${MR})
          .n(n)
          .k(k)
          .sparsity(0.0f)
          .Test(${", ".join(TEST_ARGS)});
      }
    }
  }

$if KBLOCK > 1:
  TEST(${TEST_NAME}, k_div_${KBLOCK}) {
    $if ISA_CHECK:
      ${ISA_CHECK};
    for (size_t k = ${ADJKBLOCK + KBLOCK}; k <= ${KBLOCK * 10}; k += ${KBLOCK}) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR})
        .n(${NR})
        .k(k)
        .sparsity(0.0f)
        .Test(${", ".join(TEST_ARGS)});
    }
  }

  $if NR > 1:
    TEST(${TEST_NAME}, k_div_${KBLOCK}_subtile) {
      $if ISA_CHECK:
        ${ISA_CHECK};
      for (size_t k = ${ADJKBLOCK + KBLOCK}; k <= ${KBLOCK * 10}; k += ${KBLOCK}) {
        for (uint32_t n = 1; n <= ${NR}; n++) {
          SpMMMicrokernelTester()
            .mr(${MR})
            .nr(${NR})
            .m(${MR})
            .n(n)
            .k(k)
            .sparsity(0.0f)
            .Test(${", ".join(TEST_ARGS)});
        }
      }
    }

TEST(${TEST_NAME}, n_gt_${NR}) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t n = ${NR + 1}; n < ${max(10, NR * 2)}; n++) {
    for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR})
        .n(n)
        .k(k)
        .sparsity(0.0f)
        .Test(${", ".join(TEST_ARGS)});
    }
  }
}

$if NR > 1:
  TEST(${TEST_NAME}, n_div_${NR}) {
    $if ISA_CHECK:
      ${ISA_CHECK};
    for (uint32_t n = ${2 * NR}; n <= ${3 * NR}; n += ${NR}) {
      for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
        SpMMMicrokernelTester()
          .mr(${MR})
          .nr(${NR})
          .m(${MR})
          .n(n)
          .k(k)
          .Test(${", ".join(TEST_ARGS)});
      }
    }
  }

TEST(${TEST_NAME}, m_lt_${MR}) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t m = ${1}; m < ${MR}; m++) {
    for (uint32_t n = 1; n < ${max(10, NR * 5)}; n += ${NR + 1}) {
      for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
        SpMMMicrokernelTester()
          .mr(${MR})
          .nr(${NR})
          .m(m)
          .n(n)
          .k(k)
          .sparsity(0.0f)
          .Test(${", ".join(TEST_ARGS)});
      }
    }
  }
}

TEST(${TEST_NAME}, m_div_${MR}) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t m = ${MR * 2}; m <= ${MR * 3}; m += ${MR}) {
    for (uint32_t n = 1; n < ${max(10, NR * 5)}; n += ${NR + 1}) {
      for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
        SpMMMicrokernelTester()
          .mr(${MR})
          .nr(${NR})
          .m(m)
          .n(n)
          .k(k)
          .sparsity(0.0f)
          .Test(${", ".join(TEST_ARGS)});
      }
    }
  }
}

TEST(${TEST_NAME}, m_gt_${MR}) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t m = ${MR + 1}; m < ${MR * 2}; m++) {
    for (uint32_t n = 1; n < ${max(10, NR * 5)}; n += ${NR + 1}) {
      for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
        SpMMMicrokernelTester()
          .mr(${MR})
          .nr(${NR})
          .m(m)
          .n(n)
          .k(k)
          .sparsity(0.0f)
          .Test(${", ".join(TEST_ARGS)});
      }
    }
  }
}

TEST(${TEST_NAME}, qmin) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t n = 1; n < ${max(10, NR * 5)}; n += ${NR + 1}) {
    for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR * 2})
        .n(n)
        .k(k)
        .sparsity(0.0f)
        .qmin(128)
        .Test(${", ".join(TEST_ARGS)});
    }
  }
}

TEST(${TEST_NAME}, qmax) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t n = 1; n < ${max(10, NR * 5)}; n += ${NR + 1}) {
    for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR * 2})
        .n(n)
        .k(k)
        .sparsity(0.0f)
        .qmax(128)
        .Test(${", ".join(TEST_ARGS)});
    }
  }
}

TEST(${TEST_NAME}, half_sparse) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t n = 1; n < ${max(10, NR * 5)}; n += ${NR + 1}) {
    for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR * 2})
        .n(n)
        .k(k)
        .sparsity(0.5f)
        .Test(${", ".join(TEST_ARGS)});
    }
  }
}

TEST(${TEST_NAME}, zero_weights) {
  $if ISA_CHECK:
    ${ISA_CHECK};
  for (uint32_t n = 1; n < ${max(10, NR * 5)}; n += ${NR + 1}) {
    for (size_t k = 1; k <= ${KBLOCK * 5}; k += ${KBLOCK + 1}) {
      SpMMMicrokernelTester()
        .mr(${MR})
        .nr(${NR})
        .m(${MR * 2})
        .n(n)
        .k(k)
        .sparsity(1.0f)
        .Test(${", ".join(TEST_ARGS)});
    }
  }
}
"""


def generate_test_cases(ukernel, mr, nr, k_block, is_pipelined, isa):
  """Generates all tests cases for a GEMM micro-kernel.

  Args:
    ukernel: C name of the micro-kernel function.
    mr: MR parameter of the GEMM micro-kernel.
    nr: NR parameter of the GEMM micro-kernel.
    k_block: Number of K values processed per one iteration of the main loop of
             the micro-kernel.
    is_pipelined: Indicates if the micro-kernel is implemented with software
                  pipelining. Additional test cases are generated for software
                  pipelined micro-kernels to separately test prologue + epiloque
                  of the pipelined loop and iteration of the pipelined loop.
    isa: instruction set required to run the micro-kernel. Generated unit test
         will skip execution if the host processor doesn't support this ISA.

  Returns:
    Code for the test case.
  """
  _, test_name = ukernel.split("_", 1)
  _, datatype, ukernel_type, _ = ukernel.split("_", 3)
  test_args = [ukernel]
  if not isa or isa == "psimd":
    test_args.append("SpMMMicrokernelTester::Variant::Scalar")
  return xngen.preprocess(TEST_TEMPLATE, {
      "TEST_NAME": test_name.upper().replace("UKERNEL_", ""),
      "TEST_ARGS": test_args,
      "UKERNEL_TYPE": ukernel_type.upper(),
      "DATATYPE": datatype,
      "MR": mr,
      "NR": nr,
      "KBLOCK": k_block,
      "ADJKBLOCK": 2 * k_block if is_pipelined else k_block,
      "IS_PIPELINED": is_pipelined,
      "ISA_CHECK": ISA_TO_CHECK_MAP.get(isa, ""),
      "next_prime": next_prime,
    })


def main(args):
  options = parser.parse_args(args)

  with codecs.open(options.spec, "r", encoding="utf-8") as spec_file:
    spec_yaml = yaml.safe_load(spec_file)
    if not isinstance(spec_yaml, list):
      raise ValueError("expected a list of micro-kernels in the spec")

    tests = """\
// Copyright 2019 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
//
// Auto-generated file. Do not edit!
//   Specification: {specification}
//   Generator: {generator}


#include <cpuinfo.h>
#include <gtest/gtest.h>

#include <xnnpack/spmm.h>
#include <xnnpack/isa-checks.h>

#include "spmm-microkernel-tester.h"
""".format(specification=options.spec, generator=sys.argv[0])

    for ukernel_spec in spec_yaml:
      name = ukernel_spec["name"]
      k_block = int(ukernel_spec["k-block"])
      pipelined = bool(ukernel_spec.get("pipelined", False))
      mr, nr, arch, isa = split_ukernel_name(name)

      # specification can override architecture
      arch = ukernel_spec.get("arch", arch)

      test_case = generate_test_cases(name, mr, nr, k_block, pipelined, isa)
      test_case = remove_duplicate_newlines(test_case)
      tests += "\n\n"
      if arch:
        guard_macro = " || ".join(map(ARCH_TO_MACRO_MAP.get, arch))
        tests += "#if %s\n" % guard_macro
        tests += indent(test_case) + "\n"
        tests += "#endif  // %s\n" % guard_macro
      elif isa == "psimd":
        guard_macro = "!CPUINFO_ARCH_ASMJS && !CPUINFO_ARCH_WASM"
        tests += "#if %s\n" % guard_macro
        tests += indent(test_case) + "\n"
        tests += "#endif  // %s\n" % guard_macro
      else:
        tests += test_case

    with codecs.open(options.output, "w", encoding="utf-8") as output_file:
      output_file.write(tests)


if __name__ == "__main__":
  main(sys.argv[1:])