github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-02-16 01:26:29 -05:00
Code Issues Packages Projects Releases Wiki Activity

hcq move out program call to base class (#5638)

Browse Source 
* hcq move out program call to base class

* fix
This commit is contained in:
nimlgen
2024-07-23 14:25:38 +03:00
committed by GitHub
parent 7cb67e6fb2
commit a93982ef42
3 changed files with 24 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
tinygrad/runtime/ops_amd.py
View File

@@ -2,9 +2,9 @@ from __future__ import annotations
from typing import Tuple, List, Any, cast
import os, fcntl, ctypes, ctypes.util, functools, re, pathlib, mmap, errno, subprocess, time, array
from dataclasses import dataclass
from tinygrad.device import HCQCompiled, HCQAllocator, HCQBuffer, HWComputeQueue, HWCopyQueue, hcq_profile, \
from tinygrad.device import HCQCompiled, HCQAllocator, HCQBuffer, HWComputeQueue, HWCopyQueue, \
HCQSignal, HCQProgram, Compiler, CompileError, BufferOptions
from tinygrad.helpers import getenv, to_mv, round_up, data64_le, DEBUG, PROFILE, mv_address
from tinygrad.helpers import getenv, to_mv, round_up, data64_le, DEBUG, mv_address
from tinygrad.renderer.cstyle import AMDRenderer
from tinygrad.runtime.support.hip_comgr import compile_hip
import tinygrad.runtime.autogen.kfd as kfd
@@ -311,7 +311,7 @@ class AMDProgram(HCQProgram):
# If required, allocate space for the dispatch packet in the kernargs to pass it to the GPU.
args_alloc_sz = self.kernargs_segment_size + (ctypes.sizeof(hsa.hsa_kernel_dispatch_packet_t) if self.kernel_code_properties & 0x2 else 0)
super().__init__(self.device, kernargs_alloc_size=args_alloc_sz)
super().__init__(self.device, self.name, kernargs_alloc_size=args_alloc_sz)
def __del__(self):
if hasattr(self, 'lib_gpu'): cast(AMDDevice, self.device)._gpu_free(self.lib_gpu)
@@ -321,21 +321,6 @@ class AMDProgram(HCQProgram):
if len(bufs): to_mv(kernargs_ptr, len(bufs) * 8).cast('Q')[:] = array.array('Q', [b.va_addr for b in bufs])
if len(vals): to_mv(kernargs_ptr + len(bufs) * 8, len(vals) * 4).cast('I')[:] = array.array('I', vals)
def __call__(self, *args, global_size:Tuple[int,int,int]=(1,1,1), local_size:Tuple[int,int,int]=(1,1,1), vals:Tuple[int, ...]=(), wait=False):
kernargs_ptr = self.fill_kernargs(args, vals)
q = AMDComputeQueue().wait(self.device.timeline_signal, self.device.timeline_value - 1).memory_barrier()
with hcq_profile(self.device, queue=q, desc=self.name, enabled=wait or PROFILE) as (sig_st, sig_en):
q.exec(self, kernargs_ptr, global_size, local_size)
q.signal(self.device.timeline_signal, self.device.timeline_value).submit(self.device)
self.device.timeline_value += 1
if wait:
self.device.timeline_signal.wait(self.device.timeline_value - 1)
return (sig_en.timestamp - sig_st.timestamp) / 1e6
class AMDAllocator(HCQAllocator):
def __init__(self, device:AMDDevice): super().__init__(device, batch_size=SDMA_MAX_COPY_SIZE)