github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-04-07 03:00:26 -04:00
Code Issues Packages Projects Releases Wiki Activity

amd fix cwsr for gfx11 (#6950)

Browse Source 
* amd cwsr

* ()
This commit is contained in:
nimlgen
2024-10-08 17:44:29 +03:00
committed by GitHub
parent 0d526e251e
commit 137ad5519f
2 changed files with 14 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
tinygrad/runtime/ops_amd.py
View File

@@ -288,7 +288,8 @@ class AMDProgram(HCQProgram):
code = hsa.amd_kernel_code_t.from_address(self.lib_gpu.va_addr + entry_point) # NOTE: this is wrong, it's not this object
assert code.kernel_code_properties & 0x400 == 0x400 # ENABLE_WAVEFRONT_SIZE32
self.rsrc1 = code.compute_pgm_rsrc1
# Set rsrc1.priv=1 on gfx11 to workaround cwsr.
self.rsrc1 = code.compute_pgm_rsrc1 | ((1 << 20) if 110000 <= self.device.target < 120000 else 0)
self.rsrc2 = code.compute_pgm_rsrc2 | (lds_size << 15)
self.prog_addr = self.lib_gpu.va_addr + entry_point + code.kernel_code_entry_byte_offset
@@ -414,20 +415,27 @@ class AMDDevice(HCQCompiled):
engines = self.properties['array_count'] // self.properties['simd_arrays_per_engine']
self.tmpring_size = (wave_scratch_len // 256) << 12 | (self.scratch_len // (wave_scratch_len * engines))
self.compute_queue = self._alloc_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE, 0x100000, ctx_save_restore_size=0x2C02000, eop_buffer_size=0x1000)
# https://gitlab.freedesktop.org/agd5f/linux/-/blob/a1fc9f584c4aaf8bc1ebfa459fc57a3f26a290d8/drivers/gpu/drm/amd/amdkfd/kfd_queue.c#L391
sgrp_size_per_cu, lds_size_per_cu, hwreg_size_per_cu = 0x4000, 0x10000, 0x1000
vgpr_size_per_cu = 0x60000 if self.target in {110000, 110001, 120000, 120001} else 0x40000
wg_data_size = round_up((vgpr_size_per_cu + sgrp_size_per_cu + lds_size_per_cu + hwreg_size_per_cu) * (max_cu_id + 1), mmap.PAGESIZE)
ctl_stack_size = round_up(12 * (max_cu_id + 1) * (max_wave_id + 1) + 8 + 40, mmap.PAGESIZE)
self.compute_queue = self._alloc_queue(kfd.KFD_IOC_QUEUE_TYPE_COMPUTE, 0x100000, ctx_save_restore_size=wg_data_size + ctl_stack_size,
eop_buffer_size=0x1000, ctl_stack_size=ctl_stack_size)
self.sdma_queue = self._alloc_queue(kfd.KFD_IOC_QUEUE_TYPE_SDMA, 0x100000)
super().__init__(device, AMDAllocator(self), AMDRenderer(), AMDCompiler(self.arch), functools.partial(AMDProgram, self),
AMDSignal, AMDComputeQueue, AMDCopyQueue)
def _alloc_queue(self, queue_type, ring_size, ctx_save_restore_size=None, eop_buffer_size=None) -> AMDQueueDesc:
def _alloc_queue(self, queue_type, ring_size, ctx_save_restore_size=None, eop_buffer_size=None, ctl_stack_size=0) -> AMDQueueDesc:
gart = self._gpu_alloc(0x1000, kfd.KFD_IOC_ALLOC_MEM_FLAGS_GTT, uncached=True)
ring = self._gpu_alloc(ring_size, kfd.KFD_IOC_ALLOC_MEM_FLAGS_GTT, uncached=True)
cwsr_ctx = self._gpu_alloc(ctx_save_restore_size, kfd.KFD_IOC_ALLOC_MEM_FLAGS_VRAM) if ctx_save_restore_size else None
eop_buffer = self._gpu_alloc(eop_buffer_size, kfd.KFD_IOC_ALLOC_MEM_FLAGS_VRAM) if eop_buffer_size else None
queue = kfd.AMDKFD_IOC_CREATE_QUEUE(AMDDevice.kfd, ring_base_address=ring.va_addr, ring_size=ring.size, gpu_id=self.gpu_id,
queue_type=queue_type, queue_percentage=kfd.KFD_MAX_QUEUE_PERCENTAGE, queue_priority=kfd.KFD_MAX_QUEUE_PRIORITY,
eop_buffer_address=eop_buffer.va_addr if eop_buffer else 0, eop_buffer_size=eop_buffer.size if eop_buffer else 0,
eop_buffer_address=eop_buffer.va_addr if eop_buffer else 0, eop_buffer_size=eop_buffer.size if eop_buffer else 0, ctl_stack_size=ctl_stack_size,
ctx_save_restore_address=cwsr_ctx.va_addr if cwsr_ctx else 0, ctx_save_restore_size=cwsr_ctx.size if cwsr_ctx else 0,
write_pointer_address=gart.va_addr, read_pointer_address=gart.va_addr + 8)

3
tinygrad/runtime/ops_hip.py
View File

@@ -1,11 +1,12 @@
from __future__ import annotations
import ctypes, functools
from typing import Tuple
from tinygrad.helpers import DEBUG, init_c_var, from_mv, init_c_struct_t
from tinygrad.helpers import DEBUG, init_c_var, from_mv, init_c_struct_t, getenv
from tinygrad.device import Compiled, LRUAllocator, BufferOptions
from tinygrad.runtime.autogen import hip
from tinygrad.runtime.support.compiler_hip import AMDCompiler, disasm
from tinygrad.renderer.cstyle import HIPRenderer
if getenv("IOCTL"): import extra.hip_gpu_driver.hip_ioctl # noqa: F401 # pylint: disable=unused-import
def check(status):
if status != 0: raise RuntimeError(f"HIP Error {status}, {ctypes.string_at(hip.hipGetErrorString(status)).decode()}")