Expand Windows GPU support, add native FSR RR, and improve SR interop
Add a native D3D12 FSR Ray Regeneration route for gfx1200/gfx1201, using the game's device and queue with embedded HLSL guide preparation and output composition. Preserve the vkd3d compatibility route for RDNA3 and expose explicit route overrides and diagnostics. Improve Windows runtime compatibility: - Retry rejected application-ID initialization with a project identity. - Defer expensive RR backend creation until feature creation. - Recover from transient submission failures without disabling later frames. - Add shared-buffer export fallbacks through explicit heaps and a private device on the same adapter. - Safely evict old RR descriptor sets when input views change. - Treat requested RR presets as hints when a configured model is available. - Expand SR/RR input formats and handle diffuse/specular hit distances, packed normals, depth-stencil resources, and unused FSR guides. Reduce SR overhead by batching output downloads and format conversion under one completion wait. Optimize strict M enc2/tube attention with bounded FP32 accumulation and the original FP64 fallback for ambiguous rounding cases. Extend RDNA4 tooling with gfx1200/gfx1201 native FP8 RR model builds, capability checks, architecture-aware deployment, and experimental K/M kernel paths. Require matching model and runtime arithmetic families. Add optional embedded FSR shader optimizations keyed by shader hash and exact original SPIR-V contents. Preserve reflected metadata and record optimizer provenance, artifact hashes, and cache identity. Simplify Windows installation with automatic backend dependency downloads, download progress and ETA, backed-up absolute OptiScaler library paths, and recovery from interrupted RR model builds. Add build-kernels.bat and a kernels-only build matrix for all eleven registered RDNA targets, with code-object target validation and SHA256 manifests. Expand conversion probes, numerical stress tests, temporal benchmark tooling, installer/deployment tests, and setup documentation. Document synthetic regression and timing results, experimental build options, and remaining RDNA4 hardware validation requirements.
This commit is contained in:
ayo committed
1 parent
a617c818de
commit
ac4ee6c225
67 files changed
+3147
-326
No files matched your search
@@ -1,4 +1,4 @@
|
||||
"""Build exact gfx11 RR D/E models natively on Windows, without Git Bash.
|
||||
"""Build exact gfx11/gfx12 RR D/E models natively on Windows, without Git Bash.
|
||||
|
||||
Derived weights/shaders remain under ignored build/. Nothing is installed into
|
||||
a game or bundled. Uses the same packers and compiler recipe as the Linux shell.
|
||||
@@ -10,6 +10,7 @@ import os
|
||||
from pathlib import Path
|
||||
import re
|
||||
import shutil
|
||||
import struct
|
||||
import subprocess
|
||||
import sys
|
||||
|
||||
@@ -19,11 +20,33 @@ def sha(path):
|
||||
return hashlib.file_digest(stream, 'sha256').hexdigest()
|
||||
|
||||
|
||||
def require_fp8_matrices(path):
|
||||
"""Reject an RDNA4 candidate that lost native FP8 cooperative matrices."""
|
||||
data = path.read_bytes()
|
||||
if len(data) < 20 or len(data) % 4:
|
||||
raise RuntimeError('Invalid SPIR-V: ' + str(path))
|
||||
words = struct.unpack('<' + 'I' * (len(data) // 4), data)
|
||||
if words[0] != 0x07230203:
|
||||
raise RuntimeError('Invalid SPIR-V magic: ' + str(path))
|
||||
capabilities = set()
|
||||
offset = 5
|
||||
while offset < len(words):
|
||||
size, opcode = words[offset] >> 16, words[offset] & 0xffff
|
||||
if not size or offset + size > len(words):
|
||||
raise RuntimeError('Invalid SPIR-V instruction: ' + str(path))
|
||||
if opcode == 17 and size == 2: # OpCapability
|
||||
capabilities.add(words[offset + 1])
|
||||
offset += size
|
||||
if not {4212, 4213, 6022}.issubset(capabilities):
|
||||
raise RuntimeError('Native FP8 cooperative matrices are absent: ' + str(path))
|
||||
|
||||
|
||||
def main():
|
||||
repo = Path(__file__).resolve().parents[2]
|
||||
parser = argparse.ArgumentParser(description=__doc__)
|
||||
parser.add_argument('--dll', required=True, type=lambda p: Path(p).resolve())
|
||||
parser.add_argument('--preset', choices=('D', 'E'), default='D')
|
||||
parser.add_argument('--gpu-arch', choices=('gfx1101', 'gfx1200', 'gfx1201'), default='gfx1101')
|
||||
parser.add_argument('--output', required=True, type=lambda p: Path(p).resolve())
|
||||
parser.add_argument('--glslang', type=lambda p: Path(p).resolve(), default=repo / '.tools/glslang-16.6.0/bin/glslangValidator.exe')
|
||||
parser.add_argument('--reuse', action='store_true')
|
||||
@@ -38,7 +61,9 @@ def main():
|
||||
for key in tuple(env):
|
||||
if key.startswith('D4R_RR_'):
|
||||
del env[key]
|
||||
env.update(D4R_RR_GFX11='1', D4R_RR_RELAXED='0', PYTHONPATH=str(repo / '.tools/python/vendor'))
|
||||
gfx11 = args.gpu_arch.startswith('gfx11')
|
||||
env.update(D4R_RR_GFX11='1' if gfx11 else '0', D4R_RR_RELAXED='0',
|
||||
PYTHONPATH=str(repo / '.tools/python/vendor'))
|
||||
env['PATH'] = str(args.glslang.parent) + os.pathsep + env['PATH']
|
||||
numpy_version = subprocess.check_output([sys.executable, '-c', 'import numpy; print(numpy.__version__)'], env=env, text=True, timeout=30).strip()
|
||||
sources = {str(Path(__file__).relative_to(repo)).replace('\\', '/'): sha(Path(__file__))}
|
||||
@@ -46,7 +71,8 @@ def main():
|
||||
for path in sorted((repo / directory).rglob('*')):
|
||||
if path.is_file() and path.suffix in ('.py', '.comp', '.glsl', '.h', '.inc', '.sh', '.txt'):
|
||||
sources[path.relative_to(repo).as_posix()] = sha(path)
|
||||
identity = {'schema': 1, 'preset': args.preset, 'architecture': 'gfx1101', 'arithmetic': 'exact-gfx11',
|
||||
identity = {'schema': 1, 'preset': args.preset, 'architecture': args.gpu_arch,
|
||||
'arithmetic': 'exact-gfx11' if gfx11 else 'exact-gfx12',
|
||||
'dll': sha(args.dll), 'glslang': sha(args.glslang), 'python': sha(Path(sys.executable)),
|
||||
'numpy': numpy_version, 'sources': sources}
|
||||
stamp = args.output / 'windows-native-python-cache.json'
|
||||
@@ -91,16 +117,24 @@ def main():
|
||||
python(repo / 'engine/compile_rr.py', '--dll', args.dll, '--preset', args.preset, out,
|
||||
'--ptx-dir', work / 'ptx', '--network-bindings-only', label='bindings')
|
||||
python(rr / 'dll_weights.py', args.dll, work / 'dll', '--model', '0' if args.preset == 'D' else '1', label='weights')
|
||||
manifest = (out / 'rr.txt').read_text().replace(f'preset {4 if args.preset == "D" else 5}\n',
|
||||
f'preset {4 if args.preset == "D" else 5}\narithmetic gfx11\n')
|
||||
defines = ['-DRR_EXACT_CONVERT', '-DRR_GFX11'] + (['-DRR_PRESET_D'] if args.preset == 'D' else [])
|
||||
manifest = (out / 'rr.txt').read_text()
|
||||
arithmetic = 'gfx11' if gfx11 else 'gfx12'
|
||||
manifest = manifest.replace(f'preset {4 if args.preset == "D" else 5}\n',
|
||||
f'preset {4 if args.preset == "D" else 5}\narithmetic {arithmetic}\n')
|
||||
defines = ['-DRR_EXACT_CONVERT'] + (['-DRR_GFX11'] if gfx11 else [])
|
||||
defines += ['-DRR_PRESET_D'] if args.preset == 'D' else []
|
||||
def native(stage, shader, include):
|
||||
nonlocal manifest
|
||||
weights = out / f'weights-{stage}m.bin'
|
||||
wide = int(python(rr / 'widen_weights.py', weights, label='widen-' + stage).splitlines()[-1])
|
||||
run([args.glslang, '--target-env', 'vulkan1.3', '-V', *defines, f'-DW_WIDE={wide}u',
|
||||
widening = []
|
||||
if gfx11:
|
||||
wide = int(python(rr / 'widen_weights.py', weights, label='widen-' + stage).splitlines()[-1])
|
||||
widening = [f'-DW_WIDE={wide}u']
|
||||
run([args.glslang, '--target-env', 'vulkan1.3', '-V', *defines, *widening,
|
||||
'-I' + str(include), '-I' + str(rr), '-o', out / f'{stage}_m.spv', shader], 'compile-' + stage)
|
||||
python(rr / 'spv_exact_convert.py', out / f'{stage}_m.spv', label='exact-' + stage)
|
||||
if not gfx11:
|
||||
require_fp8_matrices(out / f'{stage}_m.spv')
|
||||
pattern = rf'^stage {stage} {stage}\.spv 32 weights-{stage}\.bin (.*)$'
|
||||
def replacement(match):
|
||||
tail = match[1].removesuffix(' noopt')
|
||||
@@ -139,7 +173,7 @@ def main():
|
||||
path.unlink()
|
||||
(out / 'rr.txt').write_text(manifest, newline='\n')
|
||||
stages = [line.split() for line in manifest.splitlines() if line.startswith('stage ')]
|
||||
if len(stages) != 12 or 'arithmetic gfx11' not in manifest:
|
||||
if len(stages) != 12 or f'arithmetic {arithmetic}' not in manifest.splitlines():
|
||||
raise RuntimeError('Incomplete native manifest')
|
||||
for fields in stages:
|
||||
if not (out / fields[2]).is_file() or not (out / fields[4]).is_file():
|
||||
|
||||
Reference in new issue
Block a user