INV / virtual_gpu_driver /src /graphics /rasterizer.py
Fred808's picture
Upload 256 files
7a0c684 verified
Raw
History Blame Contribute Delete
22.5 kB
import struct
import numpy as np
from enum import Enum, auto
from typing import List, Dict, Tuple, Optional
class Rasterizer:
def __init__(self, driver):
self.driver = driver
print("Rasterizer initialized.")
def _compute_edge_function(self, x0, y0, x1, y1, px, py):
"""
Compute edge function for point (px,py) against edge (x0,y0)->(x1,y1)
Positive value means point is on left side of edge
"""
return (px - x0) * (y1 - y0) - (py - y0) * (x1 - x0)
def _is_top_left_edge(self, x0, y0, x1, y1):
"""Check if edge is top or left edge for tie-breaking rules"""
return (y0 == y1 and x0 < x1) or y0 < y1
def _compute_perspective_w(self, barycentric, w0, w1, w2):
"""Compute perspective-correct interpolation weight"""
return 1.0 / (barycentric[0]/w0 + barycentric[1]/w1 + barycentric[2]/w2)
def _interpolate_perspective(self, barycentric, attr0, attr1, attr2, w0, w1, w2):
"""Perspective-correct attribute interpolation"""
w = self._compute_perspective_w(barycentric, w0, w1, w2)
return w * (
attr0 * barycentric[0]/w0 +
attr1 * barycentric[1]/w1 +
attr2 * barycentric[2]/w2
)
def rasterize_triangle(self, v0, v1, v2, framebuffer_width, framebuffer_height,
msaa_samples=1, conservative=False):
"""
Rasterize triangle with modern features:
- Edge function rasterization
- Perspective-correct interpolation
- MSAA support
- Conservative rasterization option
Args:
v0, v1, v2: vertices with (x,y,z,w) coordinates and attributes
msaa_samples: number of MSAA samples (1, 2, 4, or 8)
conservative: use conservative rasterization
"""
fragments = []
# Extract positions and compute W for perspective correction
pos0, pos1, pos2 = v0['position'], v1['position'], v2['position']
w0, w1, w2 = pos0[3], pos1[3], pos2[3]
# Convert to screen space
screen0 = [(pos0[0]/w0 + 1)*0.5*framebuffer_width,
(pos0[1]/w0 + 1)*0.5*framebuffer_height]
screen1 = [(pos1[0]/w1 + 1)*0.5*framebuffer_width,
(pos1[1]/w1 + 1)*0.5*framebuffer_height]
screen2 = [(pos2[0]/w2 + 1)*0.5*framebuffer_width,
(pos2[1]/w2 + 1)*0.5*framebuffer_height]
# Compute bounding box
min_x = max(0, int(min(screen0[0], screen1[0], screen2[0])))
max_x = min(framebuffer_width - 1, int(max(screen0[0], screen1[0], screen2[0])))
min_y = max(0, int(min(screen0[1], screen1[1], screen2[1])))
max_y = min(framebuffer_height - 1, int(max(screen0[1], screen1[1], screen2[1])))
# For conservative rasterization, expand bounding box
if conservative:
min_x -= 1
min_y -= 1
max_x += 1
max_y += 1
# Compute edge functions for triangle edges
def edge01(px, py): return self._compute_edge_function(
screen0[0], screen0[1], screen1[0], screen1[1], px, py)
def edge12(px, py): return self._compute_edge_function(
screen1[0], screen1[1], screen2[0], screen2[1], px, py)
def edge20(px, py): return self._compute_edge_function(
screen2[0], screen2[1], screen0[0], screen0[1], px, py)
# Determine fill rules for edges
is_top_left01 = self._is_top_left_edge(screen0[0], screen0[1],
screen1[0], screen1[1])
is_top_left12 = self._is_top_left_edge(screen1[0], screen1[1],
screen2[0], screen2[1])
is_top_left20 = self._is_top_left_edge(screen2[0], screen2[1],
screen0[0], screen0[1])
# Area of the triangle for barycentric coordinates
area = edge01(screen2[0], screen2[1])
if area <= 0: # Skip back-facing triangles
return []
# MSAA sample positions (for 2x2 grid)
if msaa_samples == 4:
sample_positions = [
(-0.375, -0.375), (0.375, -0.375),
(-0.375, 0.375), (0.375, 0.375)
]
else:
sample_positions = [(0.0, 0.0)]
# Rasterize
for y in range(min_y, max_y + 1):
for x in range(min_x, max_x + 1):
covered_samples = 0
sample_fragments = []
# Test each sample position
for sample_x, sample_y in sample_positions:
px, py = x + sample_x, y + sample_y
# Compute edge values
e01 = edge01(px, py)
e12 = edge12(px, py)
e20 = edge20(px, py)
# Apply fill rules
inside = (
(e01 > 0 or (e01 == 0 and is_top_left01)) and
(e12 > 0 or (e12 == 0 and is_top_left12)) and
(e20 > 0 or (e20 == 0 and is_top_left20))
)
if inside or (conservative and (e01 >= 0 and e12 >= 0 and e20 >= 0)):
covered_samples += 1
# Compute barycentric coordinates
b0 = e12 / area
b1 = e20 / area
b2 = e01 / area
# Interpolate Z perspectively
z = self._interpolate_perspective(
(b0, b1, b2),
pos0[2], pos1[2], pos2[2],
w0, w1, w2
)
# Interpolate attributes
attributes = {}
for attr in v0['attributes'].keys():
attributes[attr] = self._interpolate_perspective(
(b0, b1, b2),
v0['attributes'][attr],
v1['attributes'][attr],
v2['attributes'][attr],
w0, w1, w2
)
sample_fragments.append({
"x": x,
"y": y,
"sample_x": sample_x,
"sample_y": sample_y,
"depth": z,
"attributes": attributes,
"barycentric": (b0, b1, b2)
})
if covered_samples > 0:
fragment = {
"x": x,
"y": y,
"samples": sample_fragments,
"coverage": covered_samples / len(sample_positions)
}
fragments.append(fragment)
return fragments
class HiZBuffer:
"""Hierarchical Z-buffer for early depth testing"""
def __init__(self, width, height):
self.width = width
self.height = height
self.levels = []
# Build mip chain
current_w, current_h = width, height
while current_w > 0 and current_h > 0:
self.levels.append(np.full((current_h, current_w), 1.0))
current_w //= 2
current_h //= 2
def update_region(self, x, y, z):
"""Update Hi-Z pyramid after depth write"""
level = 0
while level < len(self.levels):
level_x, level_y = x >> level, y >> level
if level_x >= self.levels[level].shape[1] or level_y >= self.levels[level].shape[0]:
break
# Update min depth
self.levels[level][level_y, level_x] = min(
self.levels[level][level_y, level_x], z)
level += 1
def test_region(self, min_x, min_y, max_x, max_y, z):
"""Test if region could be visible (not occluded)"""
# Find appropriate mip level
width = max_x - min_x + 1
height = max_y - min_y + 1
level = max(0, int(np.log2(max(width, height))))
if level >= len(self.levels):
return True
# Scale coordinates to mip level
level_min_x = min_x >> level
level_min_y = min_y >> level
level_max_x = max_x >> level
level_max_y = max_y >> level
# Get min depth in region
min_depth = np.inf
for ly in range(level_min_y, level_max_y + 1):
for lx in range(level_min_x, level_max_x + 1):
if ly < self.levels[level].shape[0] and lx < self.levels[level].shape[1]:
min_depth = min(min_depth, self.levels[level][ly, lx])
return z <= min_depth
def process_fragments(self, fragments, fragment_shader_program, chip_id=0,
early_z=True, hierarchical_z=True):
"""
Process fragments using the fragment shader with early-Z and Hi-Z optimizations
Args:
fragments: List of fragments to process
fragment_shader_program: Shader program to execute
chip_id: GPU chip to use
early_z: Enable early-Z optimization
hierarchical_z: Enable hierarchical Z-buffer
"""
processed_fragments = []
# Initialize Hi-Z buffer if needed
hiz = None
if hierarchical_z:
fb_width = max(f["x"] for f in fragments) + 1
fb_height = max(f["y"] for f in fragments) + 1
hiz = self.HiZBuffer(fb_width, fb_height)
# Group fragments into tiles for better cache coherency
TILE_SIZE = 32
tiles = {}
for fragment in fragments:
tile_x = fragment["x"] // TILE_SIZE
tile_y = fragment["y"] // TILE_SIZE
if (tile_x, tile_y) not in tiles:
tiles[(tile_x, tile_y)] = []
tiles[(tile_x, tile_y)].append(fragment)
# Process tiles
for (tile_x, tile_y), tile_fragments in tiles.items():
# Sort fragments by depth for early-Z efficiency
if early_z:
tile_fragments.sort(key=lambda f: f["samples"][0]["depth"])
# Hi-Z test for entire tile
tile_min_x = tile_x * TILE_SIZE
tile_min_y = tile_y * TILE_SIZE
tile_max_x = min(tile_min_x + TILE_SIZE - 1, fb_width - 1)
tile_max_y = min(tile_min_y + TILE_SIZE - 1, fb_height - 1)
if hierarchical_z:
min_depth = min(s["depth"] for f in tile_fragments for s in f["samples"])
if not hiz.test_region(tile_min_x, tile_min_y, tile_max_x, tile_max_y, min_depth):
continue
# Process fragments in tile
for fragment in tile_fragments:
# Early-Z test (per sample)
if early_z:
depth_test_passed = False
for sample in fragment["samples"]:
if self._depth_test(sample["depth"], fragment["x"], fragment["y"]):
depth_test_passed = True
break
if not depth_test_passed:
continue
# Execute fragment shader
processed_samples = []
for sample in fragment["samples"]:
color = self._execute_fragment_shader(sample, fragment_shader_program, chip_id)
processed_sample = {
"sample_x": sample["sample_x"],
"sample_y": sample["sample_y"],
"depth": sample["depth"],
"color": color
}
processed_samples.append(processed_sample)
# Update Hi-Z buffer
if hierarchical_z:
hiz.update_region(fragment["x"], fragment["y"], sample["depth"])
processed_fragment = {
"x": fragment["x"],
"y": fragment["y"],
"samples": processed_samples,
"coverage": fragment["coverage"]
}
processed_fragments.append(processed_fragment)
return processed_fragments
def _execute_fragment_shader(self, fragment, fragment_shader_program, chip_id):
"""
Simulate execution of a fragment shader for a single fragment.
"""
# In a real implementation, this would dispatch the shader instructions
# to an available SM and execute them using the SM's cores.
# For simulation, just return a dummy color based on fragment position
r = (fragment["x"] % 256) / 255.0
g = (fragment["y"] % 256) / 255.0
b = fragment["depth"]
a = 1.0
return (r, g, b, a)
def _depth_test(self, fragment_depth: float, x: int, y: int,
depth_func=lambda a,b: a < b) -> bool:
"""
Test fragment depth against depth buffer
Args:
fragment_depth: Fragment's depth value
x, y: Fragment coordinates
depth_func: Depth comparison function
Returns:
bool: True if fragment passes depth test
"""
depth_buffer_index = y * self.framebuffer_width + x
current_depth = self.depth_buffer[depth_buffer_index]
return depth_func(fragment_depth, current_depth)
def depth_test(self, fragments: List[Dict], depth_buffer_bytes: bytes,
framebuffer_width: int,
depth_func: str = 'LESS',
depth_write: bool = True,
stencil_enabled: bool = False) -> Tuple[List[Dict], bytes]:
"""
Perform depth and optional stencil testing on fragments
Args:
fragments: List of fragments to test
depth_buffer_bytes: Current depth buffer
framebuffer_width: Width of framebuffer
depth_func: Depth comparison function ('LESS', 'LEQUAL', etc)
depth_write: Whether to write passing fragments to depth buffer
stencil_enabled: Whether to perform stencil testing
Returns:
Tuple of (passed fragments, modified depth buffer)
"""
self.framebuffer_width = framebuffer_width
# Set up depth comparison function
depth_funcs = {
'NEVER': lambda a,b: False,
'LESS': lambda a,b: a < b,
'EQUAL': lambda a,b: abs(a - b) < 1e-6,
'LEQUAL': lambda a,b: a <= b,
'GREATER': lambda a,b: a > b,
'NOTEQUAL': lambda a,b: abs(a - b) >= 1e-6,
'GEQUAL': lambda a,b: a >= b,
'ALWAYS': lambda a,b: True
}
depth_compare = depth_funcs[depth_func]
# Unpack depth buffer
self.depth_buffer = []
if depth_buffer_bytes:
for i in range(0, len(depth_buffer_bytes), 4):
depth = struct.unpack("f", bytes(bytearray(depth_buffer_bytes[i:i+4])))[0]
self.depth_buffer.append(depth)
else:
self.depth_buffer = [1.0] * (framebuffer_width * framebuffer_width)
passed_fragments = []
for fragment in fragments:
x, y = fragment["x"], fragment["y"]
passed_samples = []
for sample in fragment["samples"]:
if self._depth_test(sample["depth"], x, y, depth_compare):
passed_samples.append(sample)
# Write depth if enabled
if depth_write:
depth_idx = y * framebuffer_width + x
self.depth_buffer[depth_idx] = sample["depth"]
if passed_samples:
fragment = fragment.copy()
fragment["samples"] = passed_samples
fragment["coverage"] = len(passed_samples) / len(fragment["samples"])
passed_fragments.append(fragment)
# Pack modified depth buffer
modified_depth_buffer = b''.join(
[struct.pack("f", d) for d in self.depth_buffer])
return passed_fragments, modified_depth_buffer
class BlendMode(Enum):
"""Blend modes for color blending"""
ZERO = auto()
ONE = auto()
SRC_COLOR = auto()
ONE_MINUS_SRC_COLOR = auto()
DST_COLOR = auto()
ONE_MINUS_DST_COLOR = auto()
SRC_ALPHA = auto()
ONE_MINUS_SRC_ALPHA = auto()
DST_ALPHA = auto()
ONE_MINUS_DST_ALPHA = auto()
class BlendOp(Enum):
"""Blend operations"""
ADD = auto()
SUBTRACT = auto()
REVERSE_SUBTRACT = auto()
MIN = auto()
MAX = auto()
def _blend_factor(self, mode: BlendMode, src_color, dst_color) -> np.ndarray:
"""Calculate blend factor based on mode"""
if mode == self.BlendMode.ZERO:
return np.zeros(4)
elif mode == self.BlendMode.ONE:
return np.ones(4)
elif mode == self.BlendMode.SRC_COLOR:
return src_color
elif mode == self.BlendMode.ONE_MINUS_SRC_COLOR:
return 1.0 - src_color
elif mode == self.BlendMode.DST_COLOR:
return dst_color
elif mode == self.BlendMode.ONE_MINUS_DST_COLOR:
return 1.0 - dst_color
elif mode == self.BlendMode.SRC_ALPHA:
return np.full(4, src_color[3])
elif mode == self.BlendMode.ONE_MINUS_SRC_ALPHA:
return np.full(4, 1.0 - src_color[3])
elif mode == self.BlendMode.DST_ALPHA:
return np.full(4, dst_color[3])
elif mode == self.BlendMode.ONE_MINUS_DST_ALPHA:
return np.full(4, 1.0 - dst_color[3])
def _blend_operation(self, op: BlendOp, src: np.ndarray, dst: np.ndarray) -> np.ndarray:
"""Apply blend operation"""
if op == self.BlendOp.ADD:
return src + dst
elif op == self.BlendOp.SUBTRACT:
return src - dst
elif op == self.BlendOp.REVERSE_SUBTRACT:
return dst - src
elif op == self.BlendOp.MIN:
return np.minimum(src, dst)
elif op == self.BlendOp.MAX:
return np.maximum(src, dst)
def write_to_framebuffer(self, fragments: List[Dict], color_buffer: bytearray,
framebuffer_width: int,
blend_enable: bool = True,
src_blend: BlendMode = BlendMode.SRC_ALPHA,
dst_blend: BlendMode = BlendMode.ONE_MINUS_SRC_ALPHA,
blend_op: BlendOp = BlendOp.ADD) -> bytearray:
"""
Write fragments to framebuffer with MSAA resolve and blending
Args:
fragments: List of fragments to write
color_buffer: Current framebuffer contents
framebuffer_width: Width of framebuffer
blend_enable: Whether to enable blending
src_blend: Source blend factor
dst_blend: Destination blend factor
blend_op: Blend operation
Returns:
Modified color buffer
"""
for fragment in fragments:
x, y = fragment["x"], fragment["y"]
buffer_index = (y * framebuffer_width + x) * 4
# Read current framebuffer color
dst_color = np.array([
color_buffer[buffer_index] / 255.0,
color_buffer[buffer_index + 1] / 255.0,
color_buffer[buffer_index + 2] / 255.0,
color_buffer[buffer_index + 3] / 255.0
])
# Resolve MSAA samples
if len(fragment["samples"]) > 1:
# Weight colors by coverage
src_color = np.zeros(4)
total_weight = 0.0
for sample in fragment["samples"]:
weight = fragment["coverage"] / len(fragment["samples"])
src_color += np.array(sample["color"]) * weight
total_weight += weight
if total_weight > 0:
src_color /= total_weight
else:
src_color = np.array(fragment["samples"][0]["color"])
# Apply blending if enabled
if blend_enable:
src_factor = self._blend_factor(src_blend, src_color, dst_color)
dst_factor = self._blend_factor(dst_blend, src_color, dst_color)
final_color = self._blend_operation(
blend_op,
src_color * src_factor,
dst_color * dst_factor
)
else:
final_color = src_color
# Clamp and convert to 8-bit
final_color = np.clip(final_color, 0.0, 1.0)
color_buffer[buffer_index] = int(final_color[0] * 255)
color_buffer[buffer_index + 1] = int(final_color[1] * 255)
color_buffer[buffer_index + 2] = int(final_color[2] * 255)
color_buffer[buffer_index + 3] = int(final_color[3] * 255)
return color_buffer