mirror of
https://github.com/ROCm/jax.git
synced 2025-04-18 04:46:06 +00:00
b9c0658fcf
The idea is that we take all the values that can contain dimension sizes from the descriptor (shape, strides_in, strides_out) and we pass them as 1-d tensor operands. We also pass as an operand the output_shape, so that we can use the hlo.CustomCallOp `indices_of_output_shapes` attribute to tell the shape refinement how to compute the shape of the result. We keep the old descriptor and the ducc_fft registration for the old C++ custom targets for backwards compatibility (for 6 months). That behavior is tested by back_compat_test.py. The one downside of this implementation is that it moves some of the ducc-specific logic from ducc_fft.py (in jaxlib) into fft.py (in jax). This was necessary because that code computes with dimensions that are now dynamic. In JAX we have support for evaluating dynamic shapes and turning them into 1-d tensors. Also added backwards compatibility test for dynamic_ducc_fft and kept the old test for ducc_fft. PiperOrigin-RevId: 541168692
254 lines
9.7 KiB
Python
254 lines
9.7 KiB
Python
# 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 functools import partial
|
|
import math
|
|
from typing import Union, Sequence
|
|
|
|
import numpy as np
|
|
|
|
from jax import lax
|
|
|
|
from jax._src import dispatch
|
|
from jax._src.api import jit, linear_transpose, ShapeDtypeStruct
|
|
from jax._src.core import Primitive, is_constant_shape
|
|
from jax._src.interpreters import ad
|
|
from jax._src.interpreters import batching
|
|
from jax._src.interpreters import mlir
|
|
from jax._src.lib.mlir.dialects import hlo
|
|
from jax._src.lib import xla_client
|
|
from jax._src.lib import ducc_fft
|
|
from jax._src.lib import version as jaxlib_version
|
|
from jax._src.numpy.util import promote_dtypes_complex, promote_dtypes_inexact
|
|
|
|
__all__ = [
|
|
"fft",
|
|
"fft_p",
|
|
]
|
|
|
|
def _str_to_fft_type(s: str) -> xla_client.FftType:
|
|
if s in ("fft", "FFT"):
|
|
return xla_client.FftType.FFT
|
|
elif s in ("ifft", "IFFT"):
|
|
return xla_client.FftType.IFFT
|
|
elif s in ("rfft", "RFFT"):
|
|
return xla_client.FftType.RFFT
|
|
elif s in ("irfft", "IRFFT"):
|
|
return xla_client.FftType.IRFFT
|
|
else:
|
|
raise ValueError(f"Unknown FFT type '{s}'")
|
|
|
|
@partial(jit, static_argnums=(1, 2))
|
|
def fft(x, fft_type: Union[xla_client.FftType, str], fft_lengths: Sequence[int]):
|
|
if isinstance(fft_type, str):
|
|
typ = _str_to_fft_type(fft_type)
|
|
elif isinstance(fft_type, xla_client.FftType):
|
|
typ = fft_type
|
|
else:
|
|
raise TypeError(f"Unknown FFT type value '{fft_type}'")
|
|
|
|
if typ == xla_client.FftType.RFFT:
|
|
if np.iscomplexobj(x):
|
|
raise ValueError("only real valued inputs supported for rfft")
|
|
x, = promote_dtypes_inexact(x)
|
|
else:
|
|
x, = promote_dtypes_complex(x)
|
|
if len(fft_lengths) == 0:
|
|
# XLA FFT doesn't support 0-rank.
|
|
return x
|
|
fft_lengths = tuple(fft_lengths)
|
|
return fft_p.bind(x, fft_type=typ, fft_lengths=fft_lengths)
|
|
|
|
def _fft_impl(x, fft_type, fft_lengths):
|
|
return dispatch.apply_primitive(fft_p, x, fft_type=fft_type, fft_lengths=fft_lengths)
|
|
|
|
_complex_dtype = lambda dtype: (np.zeros((), dtype) + np.zeros((), np.complex64)).dtype
|
|
_real_dtype = lambda dtype: np.finfo(dtype).dtype
|
|
|
|
def fft_abstract_eval(x, fft_type, fft_lengths):
|
|
if len(fft_lengths) > x.ndim:
|
|
raise ValueError(f"FFT input shape {x.shape} must have at least as many "
|
|
f"input dimensions as fft_lengths {fft_lengths}.")
|
|
if fft_type == xla_client.FftType.RFFT:
|
|
if x.shape[-len(fft_lengths):] != fft_lengths:
|
|
raise ValueError(f"RFFT input shape {x.shape} minor dimensions must "
|
|
f"be equal to fft_lengths {fft_lengths}")
|
|
shape = (x.shape[:-len(fft_lengths)] + fft_lengths[:-1]
|
|
+ (fft_lengths[-1] // 2 + 1,))
|
|
dtype = _complex_dtype(x.dtype)
|
|
elif fft_type == xla_client.FftType.IRFFT:
|
|
if x.shape[-len(fft_lengths):-1] != fft_lengths[:-1]:
|
|
raise ValueError(f"IRFFT input shape {x.shape} minor dimensions must "
|
|
"be equal to all except the last fft_length, got "
|
|
f"{fft_lengths=}")
|
|
shape = x.shape[:-len(fft_lengths)] + fft_lengths
|
|
dtype = _real_dtype(x.dtype)
|
|
else:
|
|
if x.shape[-len(fft_lengths):] != fft_lengths:
|
|
raise ValueError(f"FFT input shape {x.shape} minor dimensions must "
|
|
f"be equal to fft_lengths {fft_lengths}")
|
|
shape = x.shape
|
|
dtype = x.dtype
|
|
return x.update(shape=shape, dtype=dtype)
|
|
|
|
def _fft_lowering(ctx, x, *, fft_type, fft_lengths):
|
|
if not is_constant_shape(fft_lengths):
|
|
# TODO: https://github.com/openxla/stablehlo/issues/1366
|
|
raise NotImplementedError("Shape polymorphism for FFT with non-constant fft_length is not implemented for TPU and GPU")
|
|
return [
|
|
hlo.FftOp(x, hlo.FftTypeAttr.get(fft_type.name),
|
|
mlir.dense_int_elements(fft_lengths)).result
|
|
]
|
|
|
|
|
|
def _fft_lowering_cpu(ctx, x, *, fft_type, fft_lengths):
|
|
x_aval, = ctx.avals_in
|
|
if jaxlib_version < (0, 4, 13):
|
|
if any(not is_constant_shape(a.shape) for a in (ctx.avals_in + ctx.avals_out)):
|
|
raise NotImplementedError("Shape polymorphism for custom call is not implemented (fft); b/261671778; try updating your jaxlib.")
|
|
return [ducc_fft.ducc_fft_hlo(x, x_aval.dtype, fft_type=fft_type, # type: ignore
|
|
fft_lengths=fft_lengths)]
|
|
|
|
in_shape = x_aval.shape
|
|
dtype = x_aval.dtype
|
|
out_aval, = ctx.avals_out
|
|
out_shape = out_aval.shape
|
|
|
|
forward = fft_type in (xla_client.FftType.FFT, xla_client.FftType.RFFT)
|
|
ndims = len(in_shape)
|
|
assert len(fft_lengths) >= 1
|
|
assert len(fft_lengths) <= ndims, (fft_lengths, ndims)
|
|
assert len(in_shape) == len(out_shape) == ndims
|
|
|
|
# PocketFft does not allow size 0 dimensions.
|
|
if 0 in in_shape or 0 in out_shape:
|
|
if fft_type == xla_client.FftType.RFFT:
|
|
assert dtype in (np.float32, np.float64), dtype
|
|
out_dtype = np.dtype(np.complex64 if dtype == np.float32 else np.complex128)
|
|
|
|
elif fft_type == xla_client.FftType.IRFFT:
|
|
assert np.issubdtype(dtype, np.complexfloating), dtype
|
|
out_dtype = np.dtype(np.float32 if dtype == np.complex64 else np.float64)
|
|
|
|
else:
|
|
assert np.issubdtype(dtype, np.complexfloating), dtype
|
|
out_dtype = dtype
|
|
|
|
zero = mlir.ir_constant(np.array(0, dtype=out_dtype),
|
|
canonicalize_types=False)
|
|
return [
|
|
mlir.broadcast_in_dim(ctx, zero, out_aval, broadcast_dimensions=[])]
|
|
|
|
strides_in = []
|
|
stride = 1
|
|
for d in reversed(in_shape):
|
|
strides_in.append(stride)
|
|
stride *= d
|
|
strides_in = mlir.shape_tensor(
|
|
mlir.eval_dynamic_shape(ctx, tuple(reversed(strides_in))))
|
|
|
|
strides_out = []
|
|
stride = 1
|
|
for d in reversed(out_shape):
|
|
strides_out.append(stride)
|
|
stride *= d
|
|
strides_out = mlir.shape_tensor(
|
|
mlir.eval_dynamic_shape(ctx, tuple(reversed(strides_out))))
|
|
|
|
# scale = 1. if forward else (1. / np.prod(fft_lengths)) as a f64[1] tensor
|
|
double_type = mlir.ir.RankedTensorType.get((), mlir.ir.F64Type.get())
|
|
size_fft_length_prod = np.prod(fft_lengths) if fft_lengths else 1
|
|
size_fft_lengths, = mlir.eval_dynamic_shape_as_vals(ctx, (size_fft_length_prod,))
|
|
size_fft_lengths = hlo.ConvertOp(double_type, size_fft_lengths)
|
|
one = mlir.ir_constant(np.float64(1.), canonicalize_types=False)
|
|
scale = one if forward else hlo.DivOp(one, size_fft_lengths)
|
|
scale = hlo.ReshapeOp(
|
|
mlir.ir.RankedTensorType.get((1,), mlir.ir.F64Type.get()),
|
|
scale).result
|
|
|
|
in_shape = mlir.shape_tensor(mlir.eval_dynamic_shape(ctx, in_shape))
|
|
out_shape = mlir.shape_tensor(mlir.eval_dynamic_shape(ctx, out_shape))
|
|
in_shape = in_shape if fft_type != xla_client.FftType.IRFFT else out_shape
|
|
|
|
result_type = mlir.aval_to_ir_type(out_aval)
|
|
return [ducc_fft.dynamic_ducc_fft_hlo(
|
|
result_type, x,
|
|
input_dtype=x_aval.dtype, ndims=ndims, input_shape=in_shape,
|
|
strides_in=strides_in, strides_out=strides_out, scale=scale,
|
|
fft_type=fft_type, fft_lengths=fft_lengths, result_shape=out_shape)]
|
|
|
|
def _naive_rfft(x, fft_lengths):
|
|
y = fft(x, xla_client.FftType.FFT, fft_lengths)
|
|
n = fft_lengths[-1]
|
|
return y[..., : n//2 + 1]
|
|
|
|
@partial(jit, static_argnums=1)
|
|
def _rfft_transpose(t, fft_lengths):
|
|
# The transpose of RFFT can't be expressed only in terms of irfft. Instead of
|
|
# manually building up larger twiddle matrices (which would increase the
|
|
# asymptotic complexity and is also rather complicated), we rely JAX to
|
|
# transpose a naive RFFT implementation.
|
|
dummy_shape = t.shape[:-len(fft_lengths)] + fft_lengths
|
|
dummy_primal = ShapeDtypeStruct(dummy_shape, _real_dtype(t.dtype))
|
|
transpose = linear_transpose(
|
|
partial(_naive_rfft, fft_lengths=fft_lengths), dummy_primal)
|
|
result, = transpose(t)
|
|
assert result.dtype == _real_dtype(t.dtype), (result.dtype, t.dtype)
|
|
return result
|
|
|
|
def _irfft_transpose(t, fft_lengths):
|
|
# The transpose of IRFFT is the RFFT of the cotangent times a scaling
|
|
# factor and a mask. The mask scales the cotangent for the Hermitian
|
|
# symmetric components of the RFFT by a factor of two, since these components
|
|
# are de-duplicated in the RFFT.
|
|
x = fft(t, xla_client.FftType.RFFT, fft_lengths)
|
|
n = x.shape[-1]
|
|
is_odd = fft_lengths[-1] % 2
|
|
full = partial(lax.full_like, t, dtype=x.dtype)
|
|
mask = lax.concatenate(
|
|
[full(1.0, shape=(1,)),
|
|
full(2.0, shape=(n - 2 + is_odd,)),
|
|
full(1.0, shape=(1 - is_odd,))],
|
|
dimension=0)
|
|
scale = 1 / math.prod(fft_lengths)
|
|
out = scale * lax.expand_dims(mask, range(x.ndim - 1)) * x
|
|
assert out.dtype == _complex_dtype(t.dtype), (out.dtype, t.dtype)
|
|
# Use JAX's convention for complex gradients
|
|
# https://github.com/google/jax/issues/6223#issuecomment-807740707
|
|
return lax.conj(out)
|
|
|
|
def _fft_transpose_rule(t, operand, fft_type, fft_lengths):
|
|
if fft_type == xla_client.FftType.RFFT:
|
|
result = _rfft_transpose(t, fft_lengths)
|
|
elif fft_type == xla_client.FftType.IRFFT:
|
|
result = _irfft_transpose(t, fft_lengths)
|
|
else:
|
|
result = fft(t, fft_type, fft_lengths)
|
|
return result,
|
|
|
|
def _fft_batching_rule(batched_args, batch_dims, fft_type, fft_lengths):
|
|
x, = batched_args
|
|
bd, = batch_dims
|
|
x = batching.moveaxis(x, bd, 0)
|
|
return fft(x, fft_type, fft_lengths), 0
|
|
|
|
fft_p = Primitive('fft')
|
|
fft_p.def_impl(_fft_impl)
|
|
fft_p.def_abstract_eval(fft_abstract_eval)
|
|
mlir.register_lowering(fft_p, _fft_lowering)
|
|
ad.deflinear2(fft_p, _fft_transpose_rule)
|
|
batching.primitive_batchers[fft_p] = _fft_batching_rule
|
|
mlir.register_lowering(fft_p, _fft_lowering_cpu, platform='cpu')
|