commaai
/
tinygrad
mirror of https://github.com/commaai/tinygrad.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

qcom optimize speed (#6345)

This commit is contained in:
nimlgen 2024-09-03 19:31:37 +03:00 committed by GitHub
parent 97c8b32a7b
commit 951c4973de
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tinygrad/runtime/graph/hcq.py
View File

@ -77,7 +77,7 @@ class HCQGraph(MultiGraphRunner):
# For compute, in case of NV, optimize when only 1 same-queue dependency exists, since NV chains 2+ executions in this case,
# eliminating dependency need.
dname = enqueue_dev.dname.split(":", 1)[0]
can_opt = (dname == "AMD" or (dname == "NV" and len(sync_signals) == 0 and len(opt_deps) == 1 and id(opt_deps[0][0]) == id(out_signal)))
can_opt = dname in {"AMD", "QCOM"} or (dname == "NV" and len(sync_signals) == 0 and len(opt_deps) == 1 and id(opt_deps[0][0]) == id(out_signal))
if can_opt or isinstance(ji.prg, BufferXfer): opt_deps = [x for x in opt_deps if id(x[0]) != id(out_signal)]
# Enable necessary signals in the schedule by setting the signal value.

2
tinygrad/runtime/ops_qcom.py
View File

@ -112,7 +112,7 @@ class QCOMComputeQueue(HWComputeQueue):
int(math.ceil(global_size[0])), int(math.ceil(global_size[1])), int(math.ceil(global_size[2])))
self.reg(adreno.REG_A6XX_SP_CS_CTRL_REG0,
adreno.A6XX_SP_CS_CTRL_REG0_THREADSIZE(adreno.THREAD64) | adreno.A6XX_SP_CS_CTRL_REG0_HALFREGFOOTPRINT(prg.hregs_count)
| adreno.A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT(prg.fregs_count) | adreno.A6XX_SP_CS_CTRL_REG0_BRANCHSTACK(prg.branch_stack),
| adreno.A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT(prg.fregs_count) | adreno.A6XX_SP_CS_CTRL_REG0_BRANCHSTACK(prg.branch_stack // 2),
adreno.A6XX_SP_CS_UNKNOWN_A9B1_UNK5 | adreno.A6XX_SP_CS_UNKNOWN_A9B1_UNK6 | adreno.A6XX_SP_CS_UNKNOWN_A9B1_SHARED_SIZE(prg.shared_size),
0, prg.prg_offset, *data64_le(prg.lib_gpu.va_addr), adreno.A6XX_SP_CS_PVT_MEM_PARAM_MEMSIZEPERITEM(prg.pvtmem_size_per_item),
*data64_le(prg.device._stack.va_addr), adreno.A6XX_SP_CS_PVT_MEM_SIZE_TOTALPVTMEMSIZE(prg.pvtmem_size_total))