File size: 14,623 Bytes
4d9b003 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 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 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 | # SPDX-License-Identifier: Apache-2.0
"""C24: SECOND backbone + SECONDFPN neck on the device (PLAN.md section 1.2 row C24; CenterPoint, PointPainting,
TransFusion and BEVFusion use them), built from BN-folded host weights on the C17 builders (``ttaw.ops.conv``).
- :class:`Second`: blocks of 3x3 conv (zero padding 1) + ReLU; the first conv of a block carries the block stride
(mmdet3d ``SECOND``: ``layer_strides`` [1, 2, 2], CenterPoint-tiny [2, 2, 2]). Returns every block output.
- :class:`SecondFPN`: one deblock per block output, each up-sampling to the same grid, + ReLU (mmdet3d
``SECONDFPN``): a ``"conv"`` deblock is a 1x1 stride-1 conv (a :class:`RowLinear`), a ``"conv_transpose"`` deblock
a ConvTranspose with kernel == stride (k = 1 a :class:`RowLinear`, k = 2 ``ttnn.conv_transpose2d``, k >= 3
linear + depth-to-space: C17 ``ConvTranspose2d``, probe P5). The concat of the deblock outputs is NOT built: the
consumer takes the list (C25's K-split shared conv reads the parts, so CenterPoint's 177 MB concat never exists).
Placement: every module output is a DRAM interleaved TILE ``[1, 1, N*H*W, C]`` map (the robust first-port layout:
3x3 convs on DRAM inputs run ttnn's automatic DRAM slicing, probe P14; L1 residency is optimization work).
Precision: :data:`DEFAULT_PRECISION` (C17's ``CONV_PRECISION``, HiFi4 + fp32 accumulation + packer L1
accumulation: without it a conv keeps its partial sums in bf16 between inner blocks) unless a
``ttaw.precision.PrecisionPolicy`` says otherwise for ``<prefix>.block<i>.conv<j>`` / ``<neck>.deblock<i>``; a
precision's ``activations`` dtype (``:a=fp32``) is the dtype of that layer's OUTPUT (bf16 by default), the knob of a
deep plain conv stack whose bf16 activation rounding compounds (S:centerpoint:396-408).
:class:`ConvSpec` records are plain numpy, so a port maps its own weight loader onto them; :func:`second_numpy` /
:func:`second_fpn_numpy` are the fp32 torch oracles of the device modules.
"""
from __future__ import annotations
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Sequence
import numpy as np
from ..ops.conv import CONV_PRECISION, Conv2d, ConvTranspose2d, FeatureMap, is_dram
from ..precision import Precision, PrecisionPolicy
__all__ = ["ConvSpec", "RowLinear", "Second", "SecondFPN", "DEFAULT_PRECISION", "default_policy", "second_numpy",
"second_fpn_numpy", "free_maps"]
DEFAULT_PRECISION = CONV_PRECISION
@dataclass(frozen=True)
class ConvSpec:
"""One conv layer with BN folded: ``y = relu?(conv(x, weight) + bias)``.
``kind="conv"``: weight ``[Cout, Cin, k, k]``; ``kind="conv_transpose"``: weight ``[Cin, Cout, s, s]`` (kernel ==
stride, no padding). ``padding`` None = ``k // 2`` for convs."""
weight: np.ndarray
bias: Optional[np.ndarray]
stride: int = 1
padding: Optional[int] = None
relu: bool = True
kind: str = "conv"
name: str = ""
@property
def kernel(self) -> int:
return int(np.shape(self.weight)[-1])
@property
def in_channels(self) -> int:
return int(np.shape(self.weight)[1 if self.kind == "conv" else 0])
@property
def out_channels(self) -> int:
return int(np.shape(self.weight)[0 if self.kind == "conv" else 1])
def default_policy() -> PrecisionPolicy:
""":data:`DEFAULT_PRECISION` everywhere (bf16 weights)."""
return PrecisionPolicy(default=DEFAULT_PRECISION)
def free_maps(*maps: Any) -> None:
"""Deallocate device feature maps / tensors that still hold their buffer (``force=False``: memory shared with
a view or another owner is left alone; ``ttnn.deallocate`` forces by default, probe P2)."""
import ttnn
for m in maps:
t = m.tensor if isinstance(m, FeatureMap) else m
if t is not None and t.is_allocated():
ttnn.deallocate(t, False)
def _to_dram(fm: FeatureMap) -> FeatureMap:
"""The map in DRAM (moved to DRAM interleaved when an op left it in L1, e.g. ``conv_transpose2d``'s sharded
output; a no-op for DRAM maps)."""
import ttnn
if is_dram(fm.tensor):
return fm
out = fm.with_tensor(ttnn.to_memory_config(fm.tensor, ttnn.DRAM_MEMORY_CONFIG))
free_maps(fm)
return out
class RowLinear:
"""``y = act(x @ W + b)`` on the rows of a feature map (a 1x1 stride-1 conv): ``[1, 1, R, K]`` TILE ->
``[1, 1, R, N]`` TILE in DRAM, through ``ttnn.linear`` with an explicit compute config and the device grid as
``core_grid`` (which keeps the activation fused in the matmul). ``weight``: ``[K, N]`` (in, out) or a 1x1 conv
weight ``[N, K, 1, 1]``. Device weights are uploaded by the first call (outside any capture)."""
def __init__(self, weight: Any, bias: Any = None, *, activation: Optional[str] = None,
precision: Any = DEFAULT_PRECISION, output_dtype: str = "bfloat16", name: str = "linear"):
w = np.asarray(weight, dtype=np.float32)
if w.ndim == 4:
if w.shape[2:] != (1, 1):
raise ValueError(f"{name}: a conv weight must be 1x1, got {w.shape}")
w = w[:, :, 0, 0].T
if w.ndim != 2:
raise ValueError(f"{name}: weight must be [K, N], got {w.shape}")
if activation not in (None, "relu"):
raise ValueError(f"{name}: activation {activation!r}: None or 'relu'")
self.name = name
self.weight_host = np.ascontiguousarray(w)
self.in_channels, self.out_channels = (int(s) for s in w.shape)
b = np.zeros(self.out_channels, np.float32) if bias is None else np.asarray(bias, np.float32)
if b.shape != (self.out_channels,):
raise ValueError(f"{name}: bias must be ({self.out_channels},), got {b.shape}")
if not (np.all(np.isfinite(w)) and np.all(np.isfinite(b))):
raise ValueError(f"{name}: non-finite weights")
self.bias_host = b
self.activation = activation
self.precision = Precision.parse(precision)
self.output_dtype = output_dtype
self._device: Optional[tuple] = None
def _upload(self, device: Any):
import torch
import ttnn
from ..tensors import ttnn_dtype
wdt = ttnn_dtype(self.precision.weights)
bdt = ttnn.bfloat16 if wdt == ttnn.bfloat8_b else wdt
g = device.compute_with_storage_grid_size()
self._device = (
ttnn.from_torch(torch.from_numpy(self.weight_host), dtype=wdt, layout=ttnn.TILE_LAYOUT, device=device,
memory_config=ttnn.DRAM_MEMORY_CONFIG),
ttnn.from_torch(torch.from_numpy(self.bias_host.reshape(1, -1)), dtype=bdt, layout=ttnn.TILE_LAYOUT,
device=device, memory_config=ttnn.DRAM_MEMORY_CONFIG),
ttnn.CoreGrid(y=int(g.y), x=int(g.x)),
self.precision.compute_kernel_config())
return self._device
def __call__(self, x: Any) -> Any:
"""``x``: a TILE device tensor ``[..., R, K]`` or a :class:`FeatureMap` (returned as one)."""
import ttnn
from ..tensors import ttnn_dtype
t = x.tensor if isinstance(x, FeatureMap) else x
if int(t.shape[-1]) != self.in_channels:
raise ValueError(f"{self.name}: input has {int(t.shape[-1])} channels, expected {self.in_channels}")
weight, bias, grid, cfg = self._device or self._upload(t.device())
y = ttnn.linear(t, weight, bias=bias, activation=self.activation, compute_kernel_config=cfg, core_grid=grid,
memory_config=ttnn.DRAM_MEMORY_CONFIG, dtype=ttnn_dtype(self.output_dtype))
return x.with_tensor(y, self.out_channels) if isinstance(x, FeatureMap) else y
def release(self) -> None:
if self._device is not None:
free_maps(self._device[0], self._device[1])
self._device = None
def describe(self) -> Dict[str, Any]:
return {"name": self.name, "cin": self.in_channels, "cout": self.out_channels, "activation": self.activation,
"precision": self.precision.label, "output_dtype": self.output_dtype}
def _precision(policy: Optional[PrecisionPolicy], name: str) -> Precision:
return policy.resolve(name) if policy is not None else Precision.parse(DEFAULT_PRECISION)
class Second:
"""SECOND backbone: ``blocks[i]`` is the list of :class:`ConvSpec` of block ``i`` (k x k convs, BN folded)."""
def __init__(self, blocks: Sequence[Sequence[ConvSpec]], *, policy: Optional[PrecisionPolicy] = None,
prefix: str = "backbone"):
self.prefix = prefix
self.blocks: List[List[Conv2d]] = []
for i, block in enumerate(blocks):
layers = []
for j, spec in enumerate(block):
if spec.kind != "conv":
raise ValueError(f"{prefix}.block{i}.conv{j}: SECOND layers are convs, got {spec.kind}")
name = f"{prefix}.block{i}.conv{j}"
pad = spec.kernel // 2 if spec.padding is None else int(spec.padding)
prec = _precision(policy, name)
layers.append(Conv2d(spec.weight, spec.bias, stride=spec.stride, padding=pad,
activation="relu" if spec.relu else None, precision=prec,
output_dtype=prec.activations, name=name))
self.blocks.append(layers)
def out_shapes(self, height: int, width: int) -> List[tuple]:
"""``(H, W, C)`` of every block output for an input of ``height x width``."""
shapes = []
for block in self.blocks:
for layer in block:
height, width = layer.out_hw(height, width)
shapes.append((height, width, block[-1].out_channels))
return shapes
def __call__(self, x: FeatureMap) -> List[FeatureMap]:
"""Run every block on ``x`` (kept); returns the block outputs (DRAM). Inner layer outputs are freed."""
outs: List[FeatureMap] = []
cur = x
for block in self.blocks:
for layer in block:
nxt = _to_dram(layer(cur))
if cur is not x and all(cur is not o for o in outs):
free_maps(cur)
cur = nxt
outs.append(cur)
return outs
def release(self) -> None:
for block in self.blocks:
for layer in block:
layer.release()
def describe(self) -> List[List[Dict[str, Any]]]:
return [[layer.describe() for layer in block] for block in self.blocks]
class SecondFPN:
"""SECONDFPN neck: ``deblocks[i]`` (a :class:`ConvSpec`, ReLU after) up-samples block output ``i``."""
def __init__(self, deblocks: Sequence[ConvSpec], *, policy: Optional[PrecisionPolicy] = None,
prefix: str = "neck"):
self.prefix = prefix
self.deblocks: List[Any] = []
self.strides: List[int] = []
for i, spec in enumerate(deblocks):
name = f"{prefix}.deblock{i}"
act = "relu" if spec.relu else None
prec = _precision(policy, name)
if spec.kind == "conv":
if spec.kernel != 1 or spec.stride != 1:
raise ValueError(f"{name}: a conv deblock must be 1x1 stride 1 (k={spec.kernel}, s={spec.stride})")
layer: Any = RowLinear(spec.weight, spec.bias, activation=act, precision=prec,
output_dtype=prec.activations, name=name)
self.strides.append(1)
elif spec.kind == "conv_transpose":
if spec.kernel != spec.stride:
raise ValueError(f"{name}: ConvTranspose deblocks need kernel == stride (k={spec.kernel}, "
f"s={spec.stride})")
if spec.stride == 1:
layer = RowLinear(np.asarray(spec.weight, np.float32)[:, :, 0, 0], spec.bias, activation=act,
precision=prec, output_dtype=prec.activations, name=name)
else:
layer = ConvTranspose2d(spec.weight, spec.bias, stride=spec.stride, activation=act,
precision=prec, output_dtype=prec.activations, name=name)
self.strides.append(spec.stride)
else:
raise ValueError(f"{name}: unknown kind {spec.kind!r}")
self.deblocks.append(layer)
def __call__(self, blocks: Sequence[FeatureMap]) -> List[FeatureMap]:
"""The deblock outputs (DRAM interleaved TILE maps of one grid); the inputs are kept."""
if len(blocks) != len(self.deblocks):
raise ValueError(f"{self.prefix}: {len(blocks)} inputs for {len(self.deblocks)} deblocks")
outs = []
for layer, x in zip(self.deblocks, blocks):
outs.append(_to_dram(layer(x)))
return outs
def release(self) -> None:
for d in self.deblocks:
d.release()
def describe(self) -> List[Dict[str, Any]]:
return [d.describe() for d in self.deblocks]
# ------------------------------------------------------------------------------------------- host oracles
def _apply_torch(spec: ConvSpec, x):
"""One layer in torch fp32 on an NCHW tensor."""
import torch
import torch.nn.functional as F
w = torch.from_numpy(np.ascontiguousarray(spec.weight, dtype=np.float32))
b = None if spec.bias is None else torch.from_numpy(np.ascontiguousarray(spec.bias, dtype=np.float32))
if spec.kind == "conv":
pad = spec.kernel // 2 if spec.padding is None else spec.padding
y = F.conv2d(x, w, b, stride=spec.stride, padding=pad)
else:
y = F.conv_transpose2d(x, w, b, stride=spec.stride)
return torch.relu(y) if spec.relu else y
def second_numpy(blocks: Sequence[Sequence[ConvSpec]], x_nchw: Any) -> List[np.ndarray]:
"""fp32 oracle of :class:`Second` (torch): NCHW in, NCHW block outputs out."""
import torch
with torch.no_grad():
x = torch.as_tensor(np.asarray(x_nchw, dtype=np.float32))
outs = []
for block in blocks:
for spec in block:
x = _apply_torch(spec, x)
outs.append(x.numpy().copy())
return outs
def second_fpn_numpy(deblocks: Sequence[ConvSpec], blocks_nchw: Sequence[Any]) -> List[np.ndarray]:
"""fp32 oracle of :class:`SecondFPN` (torch): NCHW in and out."""
import torch
with torch.no_grad():
return [_apply_torch(s, torch.as_tensor(np.asarray(x, dtype=np.float32))).numpy().copy()
for s, x in zip(deblocks, blocks_nchw)]
|