gpu_solver.py

# Copyright 2019 The JAX Authors.
#
# Licensed 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
#
#     https://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.

from collections.abc import Sequence
from functools import partial
import importlib
import math

import jaxlib.mlir.ir as ir
import jaxlib.mlir.dialects.stablehlo as hlo

import numpy as np

from .gpu_common_utils import GpuLibNotLinkedError

from jaxlib import xla_client

from .hlo_helpers import (
    DimensionSize, ShapeTypePair, mk_result_types_and_shapes,
    custom_call, ensure_hlo_s32, hlo_s32, dense_int_array)

try:
  from .cuda import _blas as _cublas  # pytype: disable=import-error
except ImportError:
  for cuda_module_name in ["jax_cuda12_plugin"]:
    try:
      _cublas = importlib.import_module(f"{cuda_module_name}._blas")
    except ImportError:
      _cublas = None
    else:
      break

if _cublas:
  for _name, _value in _cublas.registrations().items():
    xla_client.register_custom_call_target(_name, _value, platform="CUDA")

for cuda_module_name in [".cuda", "jax_cuda12_plugin"]:
  try:
    _cusolver = importlib.import_module(
        f"{cuda_module_name}._solver", package="jaxlib"
    )
  except ImportError:
    _cusolver = None
  else:
    break

if _cusolver:
  for _name, _value in _cusolver.registrations().items():
    # TODO(danfm): Clean up after all legacy custom calls are ported.
    api_version = 1 if _name.endswith("_ffi") else 0
    xla_client.register_custom_call_target(_name, _value, platform="CUDA",
                                           api_version=api_version)

try:
  from .rocm import _blas as _hipblas  # pytype: disable=import-error
except ImportError:
  for rocm_module_name in ["jax_rocm60_plugin"]:
    try:
      _hipblas = importlib.import_module(f"{rocm_module_name}._blas")
    except:
      _hipblas = None
    else:
      break

if _hipblas:
  for _name, _value in _hipblas.registrations().items():
    xla_client.register_custom_call_target(_name, _value, platform="ROCM")

for rocm_module_name in [".rocm", "jax_rocm60_plugin"]:
  try:
    _hipsolver = importlib.import_module(
        f"{rocm_module_name}._solver", package="jaxlib"
    )
  except ImportError:
    _hipsolver = None
  else:
    break

if _hipsolver:
  for _name, _value in _hipsolver.registrations().items():
    # TODO(danfm): Clean up after all legacy custom calls are ported.
    api_version = 1 if _name.endswith("_ffi") else 0
    xla_client.register_custom_call_target(_name, _value, platform="ROCM",
                                           api_version=api_version)

def _real_type(dtype):
  """Returns the real equivalent of 'dtype'."""
  return np.finfo(dtype).dtype


def _geqrf_hlo(platform, gpu_solver, dtype, a):
  """QR decomposition."""
  a_type = ir.RankedTensorType(a.type)
  dims = a_type.shape
  assert len(dims) >= 2
  m, n = dims[-2:]
  batch_dims = tuple(dims[:-2])
  num_bd = len(batch_dims)
  batch = math.prod(batch_dims)

  lwork, opaque = gpu_solver.build_geqrf_descriptor(
      np.dtype(dtype), batch, m, n)

  layout = (num_bd, num_bd + 1) + tuple(range(num_bd - 1, -1, -1))
  i32_type = ir.IntegerType.get_signless(32)
  out = custom_call(
      f"{platform}solver_geqrf",
      result_types=[
        a.type,
        ir.RankedTensorType.get(batch_dims + (min(m, n),), a_type.element_type),
        ir.RankedTensorType.get(batch_dims, i32_type),
        ir.RankedTensorType.get([lwork], a_type.element_type),
      ],
      operands=[a],
      backend_config=opaque,
      operand_layouts=[layout],
      result_layouts=[
        layout,
        tuple(range(num_bd, -1, -1)),
        tuple(range(num_bd - 1, -1, -1)),
        [0],
      ],
      operand_output_aliases={0: 0}).results
  return out[:3]

cuda_geqrf = partial(_geqrf_hlo, "cu", _cusolver)
rocm_geqrf = partial(_geqrf_hlo, "hip", _hipsolver)

def _geqrf_batched_hlo(platform, gpu_blas, dtype, a):
  """Batched QR decomposition."""
  a_type = ir.RankedTensorType(a.type)
  dims = a_type.shape
  assert len(dims) >= 2
  m, n = dims[-2:]
  batch_dims = tuple(dims[:-2])
  num_bd = len(batch_dims)
  batch = math.prod(batch_dims)

  if not gpu_blas:
    raise GpuLibNotLinkedError()

  lwork, opaque = gpu_blas.build_geqrf_batched_descriptor(
      np.dtype(dtype), batch, m, n)

  layout = (num_bd, num_bd + 1) + tuple(range(num_bd - 1, -1, -1))
  out = custom_call(
      f"{platform}blas_geqrf_batched",
      result_types=[
        a.type,
        ir.RankedTensorType.get(batch_dims + (min(m, n),), a_type.element_type),
        ir.RankedTensorType.get([lwork], ir.IntegerType.get_signless(8)),
        ir.RankedTensorType.get([lwork], ir.IntegerType.get_signless(8)),
      ],
      operands=[a],
      backend_config=opaque,
      operand_layouts=[layout],
      result_layouts=[
        layout,
        tuple(range(num_bd, -1, -1)),
        [0],
        [0],
      ],
      operand_output_aliases={0: 0}
  ).results
  return out[:2]

cuda_geqrf_batched = partial(_geqrf_batched_hlo, "cu", _cublas)
rocm_geqrf_batched = partial(_geqrf_batched_hlo, "hip", _hipblas)


def _csrlsvqr_hlo(platform, gpu_solver, dtype, data,
                  indices, indptr, b, tol, reorder):
  """Sparse solver via QR decomposition. CUDA only."""
  b_type = ir.RankedTensorType(b.type)
  data_type = ir.RankedTensorType(data.type)

  n = b_type.shape[0]
  nnz = data_type.shape[0]
  opaque = gpu_solver.build_csrlsvqr_descriptor(
      np.dtype(dtype), n, nnz, reorder, tol
  )

  out = custom_call(
      f"{platform}solver_csrlsvqr",  # call_target_name
      result_types=[b.type],
      operands=[data, indptr, indices, b],
      backend_config=opaque,  # backend_config
      operand_layouts=[(0,), (0,), (0,), (0,)],  # operand_layouts
      result_layouts=[(0,)]  # result_layouts
  ).results
  return out

cuda_csrlsvqr = partial(_csrlsvqr_hlo, "cu", _cusolver)


def _orgqr_hlo(platform, gpu_solver, dtype, a, tau):
  """Product of elementary Householder reflections."""
  a_type = ir.RankedTensorType(a.type)
  dims = a_type.shape
  assert len(dims) >= 2
  m, n = dims[-2:]
  batch_dims = tuple(dims[:-2])
  num_bd = len(batch_dims)
  batch = math.prod(batch_dims)

  tau_dims = ir.RankedTensorType(tau.type).shape
  assert tau_dims[:-1] == dims[:-2]
  k = tau_dims[-1]

  lwork, opaque = gpu_solver.build_orgqr_descriptor(
      np.dtype(dtype), batch, m, n, k)

  layout = (num_bd, num_bd + 1) + tuple(range(num_bd - 1, -1, -1))
  i32_type = ir.IntegerType.get_signless(32)
  out = custom_call(
      f"{platform}solver_orgqr",
      result_types=[
        a.type,
        ir.RankedTensorType.get(batch_dims, i32_type),
        ir.RankedTensorType.get([lwork], a_type.element_type),
      ],
      operands=[a, tau],
      backend_config=opaque,
      operand_layouts=[
          layout,
          tuple(range(num_bd, -1, -1)),
      ],
      result_layouts=[
        layout,
        tuple(range(num_bd - 1, -1, -1)),
        [0],
      ],
      operand_output_aliases={0: 0}).results
  return out[:2]

cuda_orgqr = partial(_orgqr_hlo, "cu", _cusolver)
rocm_orgqr = partial(_orgqr_hlo, "hip", _hipsolver)


def _syevd_hlo(platform, gpu_solver, have_jacobi_solver, dtype, a, *,
               a_shape_vals: tuple[DimensionSize, ...], lower=False):
  """Symmetric (Hermitian) eigendecomposition."""
  a_type = ir.RankedTensorType(a.type)
  assert len(a_shape_vals) >= 2
  m, n = a_shape_vals[-2:]
  assert type(m) is int and type(n) is int and m == n, a_shape_vals
  batch_dims_vals = a_shape_vals[:-2]

  num_bd = len(batch_dims_vals)
  layout = (num_bd, num_bd + 1) + tuple(range(num_bd - 1, -1, -1))

  dynamic_batch_dims = any(type(d) != int for d in batch_dims_vals)
  if dynamic_batch_dims:
    batch_int = -1  # Signals to the kernel that the batch is an operand.
  else:
    batch_int = math.prod(batch_dims_vals)

  if have_jacobi_solver and n <= 32 and not dynamic_batch_dims:
    # We cannot use syevj for dynamic shapes because the workspace size
    # depends on the batch size.
    kernel = f"{platform}solver_syevj"
    lwork, opaque = gpu_solver.build_syevj_descriptor(
        np.dtype(dtype), lower, batch_int, n)
  else:
    kernel = f"{platform}solver_syevd"
    lwork, opaque = gpu_solver.build_syevd_descriptor(
        np.dtype(dtype), lower, batch_int, n)
    # TODO(Ruturaj4): Currently, hipsolverSsyevd sets lwork to 0 if n==0.
    # Remove if this behavior changes in then new ROCm release.
    if n > 0 or platform != "hip":
      assert lwork > 0

  if ir.ComplexType.isinstance(a_type.element_type):
    eigvals_type = ir.ComplexType(a_type.element_type).element_type
  else:
    eigvals_type = a_type.element_type

  i32_type = ir.IntegerType.get_signless(32)
  operands = [a]
  operand_layouts = [layout]
  if dynamic_batch_dims:
    batch_size_val = hlo_s32(1)
    for b_v in batch_dims_vals:
      batch_size_val = hlo.multiply(batch_size_val, ensure_hlo_s32(b_v))
    operands.append(batch_size_val)
    operand_layouts.append(())

  shape_type_pairs: Sequence[ShapeTypePair] = [
      (a_shape_vals, a_type.element_type),
      (batch_dims_vals + (n,), eigvals_type),
      (batch_dims_vals, i32_type),
      ([lwork], a_type.element_type)]
  result_types, result_shapes = mk_result_types_and_shapes(shape_type_pairs)
  out = custom_call(
      kernel,
      result_types=result_types,
      operands=operands,
      backend_config=opaque,
      operand_layouts=operand_layouts,
      result_layouts=[
          layout,
          tuple(range(num_bd, -1, -1)),
          tuple(range(num_bd - 1, -1, -1)),
          [0],
      ],
      operand_output_aliases={0: 0},
      result_shapes=result_shapes).results
  return out[:3]

cuda_syevd = partial(_syevd_hlo, "cu", _cusolver, True)
rocm_syevd = partial(_syevd_hlo, "hip", _hipsolver, True)


def _gesvd_hlo(platform, gpu_solver, have_jacobi_solver, dtype, a,
               full_matrices=True, compute_uv=True):
  """Singular value decomposition."""
  a_type = ir.RankedTensorType(a.type)
  dims = a_type.shape
  assert len(dims) >= 2
  m, n = dims[-2:]
  batch_dims = tuple(dims[:-2])
  num_bd = len(batch_dims)
  b = math.prod(batch_dims)
  if ir.ComplexType.isinstance(a_type.element_type):
    singular_vals_type = ir.ComplexType(a_type.element_type).element_type
  else:
    singular_vals_type = a_type.element_type

  scalar_layout = tuple(range(num_bd - 1, -1, -1))
  vector_layout = (num_bd,) + tuple(range(num_bd - 1, -1, -1))
  i32_type = ir.IntegerType.get_signless(32)

  # NVIDIA's batched Jacobi solver supports a maximum matrix size of 32x32, but
  # the unbatched solver has no such limit. The unbatched solver appears to
  # outperform gesvd for small-moderate matrices, e.g., see:
  # https://developer.download.nvidia.com/video/gputechconf/gtc/2019/presentation/s9226-fast-singular-value-decomposition-on-gpus-v2.pdf
  # slide 5.
  if have_jacobi_solver and m <= 1024 and n <= 1024:
    # The gesvdjbatched kernel doesn't support "econ" mode. We will use that
    # kernel only if b > 1 and m <= 32 and n <= 32.
    econ = not full_matrices and (b <= 1 or m > 32 or n > 32)
    lwork, opaque = gpu_solver.build_gesvdj_descriptor(
        np.dtype(dtype), b, m, n, compute_uv, 1 if econ else 0)
    k = min(m, n)
    matrix_layout = (num_bd, num_bd + 1) + tuple(range(num_bd - 1, -1, -1))
    _, s, u, v, info, _ = custom_call(
        f"{platform}solver_gesvdj",
        result_types=[
          a.type,
          ir.RankedTensorType.get(batch_dims + (min(m, n),), singular_vals_type),
          ir.RankedTensorType.get(batch_dims + (m, k if econ else m),
                                  a_type.element_type),
          ir.RankedTensorType.get(batch_dims + (n, k if econ else n),
                                  a_type.element_type),
          ir.RankedTensorType.get(batch_dims, i32_type),
          ir.RankedTensorType.get([lwork], a_type.element_type),
        ],
        operands=[a],
        backend_config=opaque,
        operand_layouts=[matrix_layout],
        result_layouts=[
            matrix_layout,
            vector_layout,
            matrix_layout,
            matrix_layout,
            scalar_layout,
            [0],
        ],
        operand_output_aliases={0: 0}).results
    vt = hlo.transpose(
        v,
        dense_int_array(np.array(tuple(range(num_bd)) + (num_bd + 1, num_bd))))
    if np.issubdtype(dtype, np.complexfloating):
      vt = hlo.complex(hlo.real(vt), hlo.negate(hlo.imag(vt)))
    if not full_matrices and not econ:
      u = hlo.slice(
          u,
          dense_int_array(np.zeros([len(dims)], np.int64)),
          dense_int_array(np.array(batch_dims + (m, min(m, n)))),
          dense_int_array(np.ones([len(dims)], np.int64)))
      vt = hlo.slice(
          vt,
          dense_int_array(np.zeros([len(dims)], np.int64)),
          dense_int_array(np.array(batch_dims + (min(m, n), n))),
          dense_int_array(np.ones([len(dims)], np.int64)))
  elif m < n:
    lwork, opaque = gpu_solver.build_gesvd_descriptor(
        np.dtype(dtype), b, n, m, compute_uv, full_matrices)
    k = n if full_matrices else m
    matrix_layout = (num_bd + 1, num_bd) + tuple(range(num_bd - 1, -1, -1))
    _, s, vt, u, info, _ = custom_call(
        f"{platform}solver_gesvd",
        result_types=[
          a.type,
          ir.RankedTensorType.get(batch_dims + (min(m, n),), singular_vals_type),
          ir.RankedTensorType.get(batch_dims + (k, n), a_type.element_type),
          ir.RankedTensorType.get(batch_dims + (m, m), a_type.element_type),
          ir.RankedTensorType.get(batch_dims, i32_type),
          ir.RankedTensorType.get([lwork], a_type.element_type),
        ],
        operands=[a],
        backend_config=opaque,
        operand_layouts=[matrix_layout],
        result_layouts=[
          matrix_layout,
          vector_layout,
          matrix_layout,
          matrix_layout,
          scalar_layout,
          [0],
        ],
        operand_output_aliases={0: 0}).results
  else:
    lwork, opaque = gpu_solver.build_gesvd_descriptor(
        np.dtype(dtype), b, m, n, compute_uv, full_matrices)
    k = m if full_matrices else n
    matrix_layout = (num_bd, num_bd + 1) + tuple(range(num_bd - 1, -1, -1))
    _, s, u, vt, info, _ = custom_call(
        f"{platform}solver_gesvd",
        result_types=[
          a.type,
          ir.RankedTensorType.get(batch_dims + (min(m, n),), singular_vals_type),
          ir.RankedTensorType.get(batch_dims + (m, k), a_type.element_type),
          ir.RankedTensorType.get(batch_dims + (n, n), a_type.element_type),
          ir.RankedTensorType.get(batch_dims, i32_type),
          ir.RankedTensorType.get([lwork], a_type.element_type),
        ],
        operands=[a],
        backend_config=opaque,
        operand_layouts=[matrix_layout],
        result_layouts=[
          matrix_layout,
          vector_layout,
          matrix_layout,
          matrix_layout,
          scalar_layout,
          [0],
        ],
        operand_output_aliases={0: 0}).results
  return s, u, vt, info

cuda_gesvd = partial(_gesvd_hlo, "cu", _cusolver, True)
rocm_gesvd = partial(_gesvd_hlo, "hip", _hipsolver, False)


def _sytrd_hlo(platform, gpu_solver, dtype, a, *, lower):
  """sytrd: Reduction of a symmetric (Hermitian) matrix to tridiagonal form."""
  a_type = ir.RankedTensorType(a.type)
  dims = a_type.shape
  assert len(dims) >= 2
  m, n = dims[-2:]
  assert m == n, (m, n)
  batch_dims = tuple(dims[:-2])
  num_bd = len(batch_dims)
  b = 1
  for d in batch_dims:
    b *= d

  lwork, opaque = gpu_solver.build_sytrd_descriptor(dtype, lower, b, n)
  if np.issubdtype(dtype, np.floating):
    diag_type = a_type.element_type
  elif dtype == np.complex64:
    diag_type = ir.F32Type.get()
  elif dtype == np.complex128:
    diag_type = ir.F64Type.get()
  else:
    raise NotImplementedError(f"Unsupported dtype {dtype}")

  layout = (num_bd, num_bd + 1) + tuple(range(num_bd - 1, -1, -1))
  i32_type = ir.IntegerType.get_signless(32)
  a, d, e, taus, info, _ = custom_call(
      f"{platform}solver_sytrd",
      result_types=[
        a.type,
        ir.RankedTensorType.get(batch_dims + (n,), diag_type),
        ir.RankedTensorType.get(batch_dims + (n - 1,), diag_type),
        ir.RankedTensorType.get(batch_dims + (n - 1,), a_type.element_type),
        ir.RankedTensorType.get(batch_dims, i32_type),
        ir.RankedTensorType.get([lwork], a_type.element_type),
      ],
      operands=[a],
      backend_config=opaque,
      operand_layouts=[layout],
      result_layouts=[
        layout,
        (num_bd,) + tuple(range(num_bd - 1, -1, -1)),
        (num_bd,) + tuple(range(num_bd - 1, -1, -1)),
        (num_bd,) + tuple(range(num_bd - 1, -1, -1)),
        tuple(range(num_bd - 1, -1, -1)),
        [0],
      ],
      operand_output_aliases={0: 0},
  ).results
  # Workaround for NVIDIA partners bug #3865118: sytrd returns an incorrect "1"
  # in the first element of the superdiagonal in the `a` matrix in the
  # lower=False case. The correct result is returned in the `e` vector so we can
  # simply copy it back to where it needs to be:
  intattr = lambda xs: ir.DenseIntElementsAttr.get(np.asarray(xs, np.int64))
  intarrattr = lambda xs: dense_int_array(np.asarray(xs, np.int64))
  if not lower and platform == "cu" and m > 1:
    start = (0,) * len(batch_dims) + (0,)
    end = batch_dims + (1,)
    s = hlo.slice(
        e, intarrattr(start), intarrattr(end), intarrattr([1] * len(start)))
    s_type = ir.RankedTensorType.get(batch_dims + (1, 1), diag_type)
    s = hlo.broadcast_in_dim(s_type, s, intarrattr(range(len(dims) - 1)))
    # The diagonals are always real; convert to complex if needed.
    s = hlo.convert(
        ir.RankedTensorType.get(s_type.shape, a_type.element_type), s)
    offsets = tuple(hlo.constant(intattr(i))
                    for i in ((0,) * len(batch_dims) + (0, 1)))
    a = hlo.dynamic_update_slice(a, s, offsets)

  return a, d, e, taus, info

cuda_sytrd = partial(_sytrd_hlo, "cu", _cusolver)
rocm_sytrd = partial(_sytrd_hlo, "hip", _hipsolver)