NEWM / virtual_gpu /vram.py
Factor Studios
Upload 167 files
684cc60 verified
raw
history blame
14.8 kB
import numpy as np
from collections import OrderedDict
from typing import Dict, Any, Optional, Tuple, Union
from dataclasses import dataclass
import time
@dataclass
class MemoryBlock:
"""Represents a block of memory in the symbolic VRAM."""
address: int
size: int
data: Optional[Any]
allocated_time: float
last_accessed: float
class Framebuffer:
"""Represents a 2D drawing surface in VRAM."""
def __init__(self, width: int, height: int, channels: int = 3, dtype=np.uint8):
self.width = width
self.height = height
self.channels = channels
self.dtype = dtype
# Create the pixel buffer symbolically to avoid large allocations
# The actual pixel data will be managed by the MemoryManager
self.pixel_buffer_address: Optional[int] = None
self.pixel_buffer_size: int = width * height * channels * np.dtype(dtype).itemsize
self.pixel_buffer = np.zeros((height, width, channels), dtype=dtype)
self.vram_address: Optional[int] = None # This is the address in the MemoryManager
def resize(self, new_width: int, new_height: int) -> None:
# No actual data to resize, just update symbolic size
self.width = new_width
self.height = new_height
self.pixel_buffer_size = new_width * new_height * self.channels * np.dtype(self.dtype).itemsize
def clear(self, color: Tuple[int, int, int]) -> None:
self.pixel_buffer[:, :] = color
def get_pixel(self, x: int, y: int) -> np.ndarray:
if 0 <= x < self.width and 0 <= y < self.height:
return self.pixel_buffer[y, x]
return np.zeros(self.channels, dtype=self.dtype)
def set_pixel(self, x: int, y: int, color: Tuple[int, int, int]) -> None:
if 0 <= x < self.width and 0 <= y < self.height:
self.pixel_buffer[y, x] = color[:self.channels]
def get_memory_usage(self) -> int:
"""Get the memory usage of this framebuffer in bytes."""
return self.pixel_buffer_size
class MemoryManager:
"""Manages the symbolic 500GB GDDR7 memory space."""
def __init__(self, total_memory_gb: int = 500, block_size_kb: int = 4):
self.total_memory_bytes = total_memory_gb * 1024 * 1024 * 1024 # 500GB
self.block_size_bytes = block_size_kb * 1024 # 4KB blocks
self.total_blocks = self.total_memory_bytes // self.block_size_bytes
# Symbolic memory space - only allocated blocks are stored
self.memory_blocks: Dict[int, MemoryBlock] = {}
# Free block tracking - use a list of free block ranges instead of a set of all blocks
self.free_block_ranges = [(0, self.total_blocks - 1)] # (start_block_id, end_block_id)
self.allocated_blocks = set() # Still track allocated blocks for quick lookup
# Address allocation counter
self.next_address = 0
def allocate_block(self, size_bytes: int) -> Optional[int]:
"""Allocate a block of memory and return its address."""
blocks_needed = (size_bytes + self.block_size_bytes - 1) // self.block_size_bytes
# Find a suitable contiguous block range
for i, (start, end) in enumerate(self.free_block_ranges):
available_blocks = end - start + 1
if available_blocks >= blocks_needed:
# Found a suitable range
base_block_id = start
# Update free_block_ranges
new_start = start + blocks_needed
if new_start <= end:
self.free_block_ranges[i] = (new_start, end)
else:
self.free_block_ranges.pop(i)
# Add to allocated_blocks
for j in range(blocks_needed):
self.allocated_blocks.add(base_block_id + j)
# Create memory block
base_address = base_block_id * self.block_size_bytes
memory_block = MemoryBlock(
address=base_address,
size=size_bytes,
data=bytearray(size_bytes), # Allocate actual bytearray for data
allocated_time=time.time(),
last_accessed=time.time()
)
self.memory_blocks[base_address] = memory_block
return base_address
return None # Out of memory
def deallocate_block(self, address: int) -> bool:
"""Deallocate a block of memory."""
if address in self.memory_blocks:
memory_block = self.memory_blocks[address]
blocks_to_free = (memory_block.size + self.block_size_bytes - 1) // self.block_size_bytes
base_block_id = address // self.block_size_bytes
for i in range(blocks_to_free):
block_id = base_block_id + i
if block_id in self.allocated_blocks:
self.allocated_blocks.remove(block_id)
# Add back to free_block_ranges (simple merge for now)
self.free_block_ranges.append((block_id, block_id))
self.free_block_ranges.sort() # Keep sorted for efficient merging
del self.memory_blocks[address]
return True
return False
def read_data(self, address: int, size: int) -> Optional[np.ndarray]:
"""Read data from memory."""
if address in self.memory_blocks:
memory_block = self.memory_blocks[address]
if memory_block.data is not None and size <= memory_block.size:
return np.frombuffer(memory_block.data[:size], dtype=np.uint8) # Return as numpy array
return None
def write_data(self, address: int, data: Union[np.ndarray, bytes]) -> bool:
"""Write data to memory."""
if address in self.memory_blocks:
memory_block = self.memory_blocks[address]
if memory_block.data is not None:
if isinstance(data, np.ndarray):
data_bytes = data.tobytes()
elif isinstance(data, bytes):
data_bytes = data
else:
raise TypeError("Data must be a NumPy array or bytes.")
if len(data_bytes) <= memory_block.size:
memory_block.data[:len(data_bytes)] = data_bytes
return True
return False
def get_memory_stats(self) -> Dict[str, Any]:
"""Get memory usage statistics."""
allocated_bytes = sum(block.size for block in self.memory_blocks.values())
free_bytes = self.total_memory_bytes - allocated_bytes
return {
"total_memory_gb": self.total_memory_bytes / (1024**3),
"allocated_bytes": allocated_bytes,
"free_bytes": free_bytes,
"allocated_blocks_count": len(self.allocated_blocks),
"free_block_ranges_count": len(self.free_block_ranges),
"utilization_percent": (allocated_bytes / self.total_memory_bytes) * 100 if self.total_memory_bytes > 0 else 0
}
class VRAM:
"""
Main VRAM class that provides the interface for the 500GB GDDR7 memory.
This class combines the MemoryManager for low-level memory operations
with higher-level abstractions like Framebuffers.
"""
def __init__(self, memory_size_gb: int = 500):
self.memory_manager = MemoryManager(memory_size_gb)
# Cache for frequently accessed data (simulates L1/L2 cache)
self.cache_size = 1000 # Number of cache entries
self.cache = OrderedDict()
# Framebuffer registry
self.framebuffers: Dict[str, Framebuffer] = {}
self.framebuffer_counter = 0
# Texture registry
self.textures: Dict[str, np.ndarray] = {}
self.texture_counter = 0
def create_framebuffer(self, width: int, height: int, channels: int = 3,
name: Optional[str] = None) -> str:
"""Create a new framebuffer and return its ID."""
if name is None:
name = f"framebuffer_{self.framebuffer_counter}"
self.framebuffer_counter += 1
framebuffer = Framebuffer(width, height, channels)
# Allocate memory for the framebuffer
memory_size = framebuffer.get_memory_usage()
address = self.memory_manager.allocate_block(memory_size)
if address is not None:
framebuffer.vram_address = address
self.framebuffers[name] = framebuffer
return name
else:
raise MemoryError("Failed to allocate memory for framebuffer")
def get_framebuffer(self, name: str) -> Optional[Framebuffer]:
"""Get a framebuffer by name."""
return self.framebuffers.get(name)
def delete_framebuffer(self, name: str) -> bool:
"""Delete a framebuffer and free its memory."""
if name in self.framebuffers:
framebuffer = self.framebuffers[name]
if framebuffer.vram_address is not None:
self.memory_manager.deallocate_block(framebuffer.vram_address)
del self.framebuffers[name]
return True
return False
def load_texture(self, texture_data: Union[np.ndarray, bytes], name: Optional[str] = None) -> str:
"""Load texture data into VRAM and return its ID."""
if name is None:
name = f"texture_{self.texture_counter}"
self.texture_counter += 1
size_bytes = 0
if isinstance(texture_data, np.ndarray):
size_bytes = texture_data.nbytes
elif isinstance(texture_data, bytes):
size_bytes = len(texture_data)
else:
raise TypeError("Texture data must be a NumPy array or bytes.")
# Allocate memory for the texture
address = self.memory_manager.allocate_block(size_bytes)
if address is not None:
self.memory_manager.write_data(address, texture_data) # Write actual data
self.textures[name] = texture_data # Store actual data for reference
return name
else:
raise MemoryError("Failed to allocate memory for texture")
def get_texture(self, name: str) -> Optional[np.ndarray]:
"""Get texture data by name."""
return self.textures.get(name)
def cache_read(self, address: int, size: int) -> Optional[np.ndarray]:
"""Read data with caching support."""
cache_key = (address, size)
# Check cache first
if cache_key in self.cache:
# Move to end (most recently used)
data = self.cache.pop(cache_key)
self.cache[cache_key] = data
return data.copy()
# Read from memory
data = self.memory_manager.read_data(address, size)
if data is not None:
# Add to cache
if len(self.cache) >= self.cache_size:
# Remove least recently used item
self.cache.popitem(last=False)
self.cache[cache_key] = data.copy()
return data
def transfer_from_ram(self, name: str, data: Union[np.ndarray, bytes],
delay_ms: float = 0.0) -> Optional[str]:
"""Transfer a block of data from RAM to VRAM."""
if isinstance(data, np.ndarray):
size_bytes = data.nbytes
data_to_store = data.flatten()
elif isinstance(data, bytes):
size_bytes = len(data)
data_to_store = np.frombuffer(data, dtype=np.uint8)
else:
raise TypeError("Data must be a NumPy array or bytes.")
# Simulate delay
if delay_ms > 0:
time.sleep(delay_ms / 1000.0)
# Allocate memory in VRAM
address = self.memory_manager.allocate_block(size_bytes)
if address is not None:
# Store data in VRAM
self.memory_manager.write_data(address, data_to_store)
# Register the transferred data as a texture/buffer in VRAM
# For simplicity, we\"ll register it as a texture for now
texture_id = f"ram_transfer_{self.texture_counter}"
self.texture_counter += 1
self.textures[texture_id] = data # Store actual data for reference
print(f"Transferred {size_bytes} bytes from RAM to VRAM at address {address} as {texture_id}")
return texture_id
else:
print(f"Failed to transfer {size_bytes} bytes from RAM to VRAM: Out of VRAM memory.")
return None
def get_stats(self) -> Dict[str, Any]:
"""Get comprehensive VRAM statistics."""
memory_stats = self.memory_manager.get_memory_stats()
framebuffer_memory = sum(fb.get_memory_usage() for fb in self.framebuffers.values())
texture_memory = sum(tex.nbytes for tex in self.textures.values())
return {
**memory_stats,
"framebuffers_count": len(self.framebuffers),
"textures_count": len(self.textures),
"framebuffer_memory_bytes": framebuffer_memory,
"texture_memory_bytes": texture_memory,
"cache_entries": len(self.cache),
"cache_hit_ratio": 0.0 # TODO: Implement cache hit tracking
}
if __name__ == "__main__":
# Test the VRAM module
vram = VRAM(memory_size_gb=1) # Use 1GB for testing
# Create a framebuffer
fb_id = vram.create_framebuffer(1920, 1080, 3)
print(f"Created framebuffer: {fb_id}")
# Get the framebuffer and modify it
fb = vram.get_framebuffer(fb_id)
if fb:
fb.clear((255, 0, 0)) # Clear to red
fb.set_pixel(100, 100, (0, 255, 0)) # Set a green pixel
print(f"Framebuffer size: {fb.width}x{fb.height}")
print(f"Pixel at (100, 100): {fb.get_pixel(100, 100)}")
# Load a test texture
test_texture = np.random.randint(0, 256, (256, 256, 3), dtype=np.uint8)
tex_id = vram.load_texture(test_texture)
print(f"Loaded texture: {tex_id}")
# Test transfer_from_ram
ram_data = b"\x01\x02\x03\x04\x05\x06\x07\x08"
transferred_id = vram.transfer_from_ram("test_ram_data", ram_data, delay_ms=10)
print(f"Transferred RAM data ID: {transferred_id}")
# Print statistics
stats = vram.get_stats()
print(f"VRAM Stats: {stats}")