from __future__ import annotations
import warnings
from dataclasses import dataclass, field
__all__ = ["MemoryInstanceSpec", "MemoryHierarchyConfig"]
@dataclass(frozen=True)
class MemoryInstanceSpec:
"""Specification for a single memory instance in the memory hierarchy.
**API Language:**
:ref:`中文 <MemoryInstanceSpec-cn>` | :ref:`English <MemoryInstanceSpec-en>`
----
.. _MemoryInstanceSpec-cn:
* **中文**
内存层次结构中单个内存级别的规格描述。
定义内存级别的容量（以元素数或字节数计）、每个数据元素的位宽
以及每次访问的能耗（皮焦耳）。多个 ``MemoryInstanceSpec`` 组合
构成一个完整的 :class:`MemoryHierarchyConfig`。
:param capacity: 内存容量（元素数或字节数）
:type capacity: int
:param data_bit: Bit width of each data element
:type data_bit: int
:param memory_cost_pj: Energy cost per access in picojoules
:type memory_cost_pj: float
:rtype: None
Describes the properties of one level in the memory hierarchy, including
capacity, bit width, and access energy cost.
----
.. _MemoryInstanceSpec-en:
* **English**
MemoryInstanceSpec class
:param capacity: Capacity of the memory level (in elements or bytes)
:param data_bit: Bit width of each data element
:param memory_cost_pj: Energy cost per access in picojoules
:type capacity: int
:type data_bit: int
:type memory_cost_pj: float
:return: None
:rtype: None
"""
name: str
size_bits: int
datawidth: int
r_bw: int
w_bw: int
r_cost: float
w_cost: float
_MEMORY_INSTANCES = {
"reg_1b": MemoryInstanceSpec(
"reg_1b", 1, 1, 1, 1, 0.00901 / 7.1 * 2, 0.00901 / 7.1 * 2
),
"reg_16b": MemoryInstanceSpec(
"reg_16b", 16, 16, 16, 16, 0.14416 / 7.1 * 2, 0.14416 / 7.1 * 2
),
"reg_32b": MemoryInstanceSpec(
"reg_32b", 32, 32, 32, 32, 0.14416 * 2 / 7.1 * 2, 0.14416 * 2 / 7.1 * 2
),
"sram_fp_conv_in_s": MemoryInstanceSpec(
"sram_fp_conv_in_s", 18496, 1, 1, 1, 0.44854798, 0.44854798
),
"sram_fp_conv_in_w": MemoryInstanceSpec(
"sram_fp_conv_in_w", 18432, 16, 16, 16, 5.302477904, 5.302477904
),
"sram_fp_conv_out_xi": MemoryInstanceSpec(
"sram_fp_conv_out_xi", 65536, 16, 16, 16, 5.792676768, 5.792676768
),
"sram_fp_soma_u": MemoryInstanceSpec(
"sram_fp_soma_u", 1048576, 16, 256, 256, 92.68282828, 92.68282828
),
"sram_fp_soma_s": MemoryInstanceSpec(
"sram_fp_soma_s", 262144, 1, 16, 16, 7.176767677, 7.176767677
),
"sram_fp_soma_smask": MemoryInstanceSpec(
"sram_fp_soma_smask", 65536, 1, 16, 16, 7.176767677, 7.176767677
),
"sram_bp_conv_in_du": MemoryInstanceSpec(
"sram_bp_conv_in_du", 73984, 16, 16, 16, 5.792676768, 5.792676768
),
"sram_bp_conv_in_w": MemoryInstanceSpec(
"sram_bp_conv_in_w", 18432, 16, 16, 16, 5.302477904, 5.302477904
),
"sram_bp_conv_out_res": MemoryInstanceSpec(
"sram_bp_conv_out_res", 65536, 16, 16, 16, 5.792676768, 5.792676768
),
"sram_bp_grad_in_u": MemoryInstanceSpec(
"sram_bp_grad_in_u", 1048576, 16, 256, 256, 92.68282828, 92.68282828
),
"sram_bp_grad_in_smask": MemoryInstanceSpec(
"sram_bp_grad_in_smask", 65536, 1, 16, 16, 7.176767677, 7.176767677
),
"sram_bp_grad_in_s": MemoryInstanceSpec(
"sram_bp_grad_in_s", 65536, 1, 16, 16, 7.176767677, 7.176767677
),
"sram_bp_grad_out_du": MemoryInstanceSpec(
"sram_bp_grad_out_du", 1048576, 16, 256, 256, 92.68282828, 92.68282828
),
"sram_wg_conv_in_s": MemoryInstanceSpec(
"sram_wg_conv_in_s", 4624, 1, 1, 1, 0.44854798, 0.44854798
),
"sram_wg_conv_in_du": MemoryInstanceSpec(
"sram_wg_conv_in_du", 65536, 16, 16, 16, 5.792676768, 5.792676768
),
"sram_wg_conv_out_dw": MemoryInstanceSpec(
"sram_wg_conv_out_dw", 18432, 16, 16, 16, 5.302477904, 5.302477904
),
"sram_2MB": MemoryInstanceSpec("sram_2MB", 1048576, 16, 256, 256, 46.976, 45.96),
"sram_6MB": MemoryInstanceSpec(
"sram_6MB", 50331648, 16, 256, 256, 458.1830181, 375.0422185
),
"dram": MemoryInstanceSpec("dram", 134217728 * 64, 16, 64, 64, 1300.0, 1300.0),
}
[文档]
@dataclass
class MemoryHierarchyConfig:
r"""
**API Language:**
:ref:`中文 <MemoryHierarchyConfig-cn>` | :ref:`English <MemoryHierarchyConfig-en>`
----
.. _MemoryHierarchyConfig-cn:
* **中文**
NeuroMC v1 硬件预设配置类。
:param preset_name: 预设名称
:type preset_name: str
:param technology_nm: 工艺节点（纳米）
----
.. _MemoryHierarchyConfig-en:
* **English**
MemoryHierarchyConfig class
:return: None
:rtype: None
"""
preset_name: str = "neuromc_like_v1"
technology_nm: int = 32
level_order: tuple[str, str, str, str] = ("dram", "sram", "reg", "noc")
memory_instances: dict[str, MemoryInstanceSpec] = field(
default_factory=lambda: dict(_MEMORY_INSTANCES)
)
zero_dram_in_paper_energy: bool = True
zero_noc_in_paper_energy: bool = True
zero_sram_high_directions: bool = True
[文档]
@classmethod
def neuromc_like_v1(cls, memory_model: str | None = None):
if memory_model is not None:
warnings.warn(
"memory_model is deprecated and ignored by exact NeuroMC v2.",
DeprecationWarning,
stacklevel=2,
)
return cls(
preset_name="neuromc_like_v1",
technology_nm=32,
level_order=("dram", "sram", "reg", "noc"),
memory_instances=dict(_MEMORY_INSTANCES),
zero_dram_in_paper_energy=True,
zero_noc_in_paper_energy=True,
zero_sram_high_directions=True,
)
[文档]
def copy(self):
return MemoryHierarchyConfig(
preset_name=self.preset_name,
technology_nm=self.technology_nm,
level_order=tuple(self.level_order),
memory_instances=dict(self.memory_instances),
zero_dram_in_paper_energy=self.zero_dram_in_paper_energy,
zero_noc_in_paper_energy=self.zero_noc_in_paper_energy,
zero_sram_high_directions=self.zero_sram_high_directions,
)
[文档]
def validate(self):
if self.preset_name != "neuromc_like_v1":
raise ValueError(
f"Only preset_name='neuromc_like_v1' is supported, got {self.preset_name}."
)
if self.technology_nm != 32:
raise ValueError("Exact NeuroMC v1 only supports 32nm hardware constants.")
required = {
"reg_1b",
"reg_16b",
"reg_32b",
"sram_fp_conv_in_s",
"sram_fp_conv_in_w",
"sram_fp_conv_out_xi",
"sram_fp_soma_u",
"sram_fp_soma_s",
"sram_fp_soma_smask",
"sram_bp_conv_in_du",
"sram_bp_conv_in_w",
"sram_bp_conv_out_res",
"sram_bp_grad_in_u",
"sram_bp_grad_in_s",
"sram_bp_grad_in_smask",
"sram_bp_grad_out_du",
"sram_wg_conv_in_s",
"sram_wg_conv_in_du",
"sram_wg_conv_out_dw",
"sram_2MB",
"sram_6MB",
"dram",
}
missing = sorted(required.difference(self.memory_instances.keys()))
if missing:
raise ValueError(f"Missing required NeuroMC memory instances: {missing}.")
