From c3c85c64eedfb207d7aba8288c3aef0a63f5ce4a Mon Sep 17 00:00:00 2001 From: George Hotz Date: Fri, 21 Mar 2025 09:24:33 +0800 Subject: [PATCH] simpler --- extra/replay_pkl.py | 2 +- tinygrad/runtime/ops_dsp.py | 4 ++-- tinygrad/spec.py | 2 +- 3 files changed, 4 insertions(+), 4 deletions(-) diff --git a/extra/replay_pkl.py b/extra/replay_pkl.py index 73259f08ab..f49109e7d0 100644 --- a/extra/replay_pkl.py +++ b/extra/replay_pkl.py @@ -87,7 +87,7 @@ if __name__ == "__main__": pass elif knum == 20: # 784x192 * 192x32 -> 784x32 - k.apply_opt(Opt(OptOps.UNROLL, 0, 8)) + k.apply_opt(Opt(OptOps.UNROLL, 0, 4)) k.apply_opt(Opt(OptOps.UPCAST, 1, 32)) k.apply_opt(Opt(OptOps.UPCAST, 0, 4)) elif knum == 35: diff --git a/tinygrad/runtime/ops_dsp.py b/tinygrad/runtime/ops_dsp.py index fcff61895e..ef6f3aca4e 100644 --- a/tinygrad/runtime/ops_dsp.py +++ b/tinygrad/runtime/ops_dsp.py @@ -81,8 +81,8 @@ def add_to_mul(c:UOp, x:UOp): dsp_pm_late = PatternMatcher([ # prefetch L1 - (UPat(Ops.LOAD, dtype=dtypes.uchar.vec(8), name="ld"), - lambda ld: ld.replace(src=ld.src+(UOp(Ops.CUSTOM, dtypes.void, src=(ld.src[0].src[0].src[0].index(ld.src[0].src[0].src[1]+16),), + (UPat(Ops.LOAD, dtype=(dtypes.uchar.vec(4), dtypes.uchar.vec(8)), name="ld"), + lambda ld: ld.replace(src=ld.src+(UOp(Ops.CUSTOM, dtypes.void, src=(ld.src[0].src[0].index(UOp.const(dtypes.int, 16)),), arg="__builtin_HEXAGON_Y2_dcfetch({0});"),)) if ld.src[-1].op is not Ops.CUSTOM else None), (UPat(Ops.CUSTOMI, dtype=dtypes.int.vec(32), name="c")+UPat.var("x"), add_to_mul), diff --git a/tinygrad/spec.py b/tinygrad/spec.py index b66c3d9def..ec3d3a9225 100644 --- a/tinygrad/spec.py +++ b/tinygrad/spec.py @@ -77,7 +77,7 @@ spec = PatternMatcher([ # **** new style load/store **** # INDEX is used in new style load/store - (UPat(Ops.INDEX, src=(UPat((Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL)), UPat())), lambda: True), + (UPat(Ops.INDEX, src=(UPat((Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.INDEX)), UPat())), lambda: True), # LOAD takes a (UPat(Ops.LOAD, src=(UPat((Ops.INDEX, Ops.CAST)),)), lambda: True),