sirhcm/tinygrad
1
0
Fork 0
forked from tinygrad/tinygrad
Code Pull Requests 2 Actions 8 Activity

amd: trace all instructions (#12831)

Browse Source
This commit is contained in:
nimlgen
2025-10-21 20:52:24 +08:00
committed by GitHub
parent cdc72556a1
commit 1ad6598963
2 changed files with 13 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
extra/sqtt/roc.py
+7 -3
View File
@@ -27,14 +27,18 @@ class _ROCParseCtx:
self.disasms[prog.base + addr] = info
self.addr2prg[prog.base + addr] = prog
def next_sqtt(self): return next(self.sqtt_evs, None)
def next_sqtt(self):
x = next(self.sqtt_evs, None)
self.active_se = x.se if x is not None else None
return x
def find_program(self, addr): return self.addr2prg[addr]
def on_occupancy_ev(self, ev):
if DEBUG >= 4: print("OCC", ev.time, ev.cu, ev.simd, ev.wave_id, ev.start)
if DEBUG >= 4: print("OCC", ev.time, self.active_se, ev.cu, ev.simd, ev.wave_id, ev.start)
def on_wave_ev(self, ev):
if DEBUG >= 4: print("WAVE", ev.wave_id, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
if DEBUG >= 4: print("WAVE", ev.wave_id, self.active_se, ev.cu, ev.simd, ev.contexts, ev.begin_time, ev.end_time)
asm = {}
for j in range(ev.instructions_size):
tinygrad/runtime/ops_amd.py
+6 -5
View File
@@ -20,6 +20,7 @@ from tinygrad.runtime.support.system import System, PCIIfaceBase, PCIAllocationM
from tinygrad.runtime.support.usb import ASM24Controller, USBMMIOInterface
if getenv("IOCTL"): import extra.hip_gpu_driver.hip_ioctl # noqa: F401 # pylint: disable=unused-import
SQTT = getenv("SQTT", 0)
EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
WAIT_REG_MEM_FUNCTION_EQ = 3 # ==
WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
@@ -254,9 +255,9 @@ class AMDComputeQueue(HWQueue):
if (10,0,0) <= prg.dev.target < (11,0,0): self.wreg(self.gc.mmCP_COHER_START_DELAY, 0x20)
self.wreg(self.gc.regCOMPUTE_RESTART_X, 0, 0, 0)
self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE0, 0xFFFFFFFF, 0xFFFFFFFF)
self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE2, 0xFFFFFFFF, 0xFFFFFFFF)
if prg.dev.target >= (11,0,0): self.wreg(self.gc.regCOMPUTE_STATIC_THREAD_MGMT_SE4, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF)
for i in range(8 if prg.dev.target >= (11,0,0) else 4):
semask = (prg.dev.sqtt_itrace_se_mask >> i) & 0b1 if prg.dev.sqtt_enabled and SQTT >= 2 else 0xffffffff
self.wreg(getattr(self.gc, f'regCOMPUTE_STATIC_THREAD_MGMT_SE{i}'), semask)
self.wreg(self.gc.regCOMPUTE_USER_DATA_0, *user_regs)
self.wreg(self.gc.regCOMPUTE_RESOURCE_LIMITS, 0)
@@ -812,7 +813,7 @@ class AMDDevice(HCQCompiled):
self._ensure_has_local_memory(128) # set default scratch size to 128 bytes per thread
# SQTT is disabled by default because of runtime overhead and big file sizes (~200mb to Tensor.full() two 4096x4096 tensors and matmul them)
self.sqtt_enabled = PROFILE and bool(getenv("SQTT", 0))
self.sqtt_enabled = PROFILE and SQTT > 0
if self.sqtt_enabled:
if self.target[0] < 11: raise RuntimeError(f'SQ Thread Tracing is not supported on gc:{self.target}')
if not self.is_am() and (ppfeaturemask:=int(FileIOInterface('/sys/module/amdgpu/parameters/ppfeaturemask', os.O_RDONLY).read(), 16))&0x8000:
@@ -821,7 +822,7 @@ class AMDDevice(HCQCompiled):
"For more information read https://github.com/tinygrad/tinygrad/blob/master/extra/sqtt/README.md")
SQTT_BUFFER_SIZE = getenv("SQTT_BUFFER_SIZE", 256) # in mb, per shader engine
self.sqtt_buffers = [self.allocator.alloc(SQTT_BUFFER_SIZE*1024*1024, BufferSpec(nolru=True)) for _ in range(self.se_cnt)]
self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", 2) # -1 enable all, 0 disable all, >0 bitmask for where to enable instruction tracing
self.sqtt_itrace_se_mask = getenv("SQTT_ITRACE_SE_MASK", -1 if SQTT >= 2 else (1 << 1)) # se bitmask: -1 enable all, 0 disable all
self.sqtt_next_cmd_id = itertools.count(0)
cast(AMDComputeQueue, self.hw_compute_queue_t()).sqtt_start(self.sqtt_buffers, self.sqtt_itrace_se_mask).submit(self)