move decomp dtypes

The CUDA graph runner hardcoded sharedMemBytes=0 in the kernel node params, so a custom_kernel
using dynamic shared memory (extern __shared__) failed with CUDA error 700 on graph replay
while the eager path worked. Pass runtime.smem (0 for normal kernels, which use static shared).

Co-authored-by: Claude Opus 4.8 (1M context) <[email protected]>

.github/workflows/test.yml

@@ -327,7 +327,7 @@ jobs:
           llvm: 'true'
       - name: Test openpilot model kernel count and gate usage
         run: |
-          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1468 ALLOWED_GATED_READ_IMAGE=10 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
+          ALLOWED_KERNEL_COUNT=123 ALLOWED_READ_IMAGE=1361 ALLOWED_GATED_READ_IMAGE=55 FLOAT16=1 DEV=CL IMAGE=1 python examples/openpilot/compile3.py https://gitlab.com/commaai/openpilot-lfs.git/gitlab-lfs/objects/cf6376aa9a090f0da26c280ef69eabf9bbdd51d1faac9ed392919c3db69be916
       - name: Test openpilot CL compile fp32 (test correctness)
         run: |
           DEV=CL IMAGE=1 SELFTEST=1 python examples/openpilot/compile3.py https://github.com/haraschax/filedump/raw/refs/heads/master/driving_vision_fp32.onnx
@@ -370,7 +370,6 @@ jobs:
         with:
           key: optim
           deps: testing
-          pydeps: "tensorflow==2.19"
           opencl: 'true'
       #- name: Test Optimization Helpers
       #  run: DEBUG=1 python3 extra/optimization/test_helpers.py
@@ -379,7 +378,7 @@ jobs:
       - name: Test Beam Search
         run: DEV=CL IGNORE_BEAM_CACHE=1 python3 -m pytest extra/optimization/test_beam_search.py
       - name: Test MLPerf stuff
-        run: DEV=CL python -m pytest -n=auto test/external/external_test_optim.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
+        run: DEV=CL python -m pytest -n=auto test/external/external_test_lr_schedule.py test/external/external_test_losses.py test/external/external_test_metrics.py test/external/external_test_datasets.py --durations=20
       - name: DEV=NULL beautiful_mnist_multigpu
         run: DEV=NULL NULL_ALLOW_COPYOUT=1 python examples/beautiful_mnist_multigpu.py
       - name: Test Bert training

examples/mlperf/models/flat_llama.py

@@ -38,7 +38,7 @@ def quantize_fp8(x:Tensor, amax_state:Tensor|None=None):
 
 def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_scale:Tensor|None=None,
            x_fp8:Tensor|None=None, x_new_amax:Tensor|None=None,
-           grad_amax_state:Tensor|None=None) -> tuple[Tensor,...]:
+           grad_amax_state:Tensor|None=None, x_prequant_mx:tuple|None=None) -> tuple[Tensor,...]:
   if not fp8:
     if ASM_GEMM:
       from extra.gemm.cdna_asm_gemm import can_use_asm_gemm, asm_gemm
@@ -47,12 +47,14 @@ def matmul(x:Tensor, w:Tensor, fp8:bool=True, amax_x:Tensor|None=None, w_inv_sca
   assert w_inv_scale is not None, "fp8 matmul requires w_inv_scale (weights must be stored in fp8 with per-tensor scale)"
   if MXFP8:
     from extra.gemm.cdna_asm_gemm import asm_gemm, quantize_mxfp8, mx_pack, can_use_asm_gemm, _mx_block_scale
-    x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
+    if x_prequant_mx is not None: x_q, x_e8, x_si = x_prequant_mx       # fused producer already quantized (2d)
+    else: x_q, x_e8, x_si = quantize_mxfp8(x.reshape(-1, x.shape[-1]))
+    l_shape = x.shape[:-1] if x is not None else x_q.shape[:-1]
     if can_use_asm_gemm(x_q, w.T):
       out = asm_gemm(x_q, w.T, mx=True, mx_scales=(x_si, x_e8, mx_pack(w_inv_scale), w_inv_scale),
-                     mx_w_stored=True).reshape(*x.shape[:-1], w.shape[0])
+                     mx_w_stored=True).reshape(*l_shape, w.shape[0])
     else:
-      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*x.shape[:-1], x.shape[-1])
+      x_phys = (x_q.cast(dtypes.bfloat16) * _mx_block_scale(x_e8)).reshape(*l_shape, x_q.shape[-1])
       out = x_phys @ (w.cast(dtypes.bfloat16) * _mx_block_scale(w_inv_scale)).T
     return out, (amax_x.detach() if amax_x is not None else None), x_q
   if x_fp8 is None:
@@ -214,8 +216,15 @@ class FlatTransformer:
       x_w3, new_amax, *s = matmul(inp, kwargs["w3"], amax_x=kwargs["amax_x3"], w_inv_scale=kwargs["s_3"], grad_amax_state=kwargs["grad_amax_xw3"])
       amaxs.append(new_amax)
       saves.extend([*s, x_w3])
-      out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
-                                 grad_amax_state=kwargs["grad_amax_xout"])
+      if FUSED_SILU_W13 and MXFP8:
+        from extra.llama_kernels.fused_silu_mul_quantize_mxfp8 import fused_silu_mul_quantize_mxfp8
+        aq, ae8, asi = fused_silu_mul_quantize_mxfp8(x_w1.reshape(-1, x_w1.shape[-1]), x_w3.reshape(-1, x_w3.shape[-1]))
+        out, new_amax, *s = matmul(None, kwargs["w2"], x_prequant_mx=(aq, ae8, asi), amax_x=kwargs["amax_x2"],
+                                   w_inv_scale=kwargs["s_2"], grad_amax_state=kwargs["grad_amax_xout"])
+        out = out.reshape(*x_w1.shape[:-1], kwargs["w2"].shape[0])
+      else:
+        out, new_amax, *s = matmul(x_w1.silu() * x_w3, kwargs["w2"], amax_x=kwargs["amax_x2"], w_inv_scale=kwargs["s_2"],
+                                   grad_amax_state=kwargs["grad_amax_xout"])
       amaxs.append(new_amax)
       saves.extend([*s, out])
     else:

examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/profile.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
-export BENCHMARK=5
+export BENCHMARK=${BENCHMARK:-5}
 export EVAL_BS=0
 VIZ=${VIZ:--1} FULL_LAYERS=1 DEBUG=${DEBUG:--0} examples/mlperf/training_submission_v6.0/tinycorp/benchmarks/llama31_8b/implementations/tinybox_8xMI350X/dev_beam.sh
 SRC="AMD"; [[ $DEV == NULL* ]] && SRC="NULL"
-python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"
+[ $BENCHMARK -gt 3 ] && python -m tinygrad.viz.cli -s "$SRC" -t --interval "train @ 2" "train @ 3"

extra/hcq2/hcq2.py

@@ -143,14 +143,17 @@ def make_getaddr(u, device=None):
 def make_ins(op, *srcs):
   return UOp(Ops.INS, dtypes.void, tuple(UOp.const(dtypes.uint32, s) if isinstance(s, int) else s.cast(dtypes.uint32) for s in srcs), op)
 
+def make_patch(buf:UOp, off:sint, val:UOp, dtype=None) -> UOp:
+  dt = dtype or val.dtype
+  return UOp(Ops.SHRINK, buf.dtype.base, (buf, UOp.const(dtypes.int, off), UOp.const(dtypes.int, dt.itemsize))).bitcast(dt).store(val.cast(dt))
+
 def make_cmdbuf(lin, devs, tag):
   blob, patches = b'', []
   for s in (s for ins in lin.src for s in ins.src):
     if s.op is not Ops.CONST: patches.append((len(blob), s))
     blob += struct.pack(f'<{s.dtype.fmt}', s.arg if s.op is Ops.CONST else 0x0)
   buf = UOp.new_buffer(devs, len(blob), dtypes.uint8).rtag(tag)
-  stores = [buf.index(UOp.const(dtypes.int, off), dtype=buf.dtype.ptr()).cast(s.dtype.ptr()).store(s) for off, s in patches]
-  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *stores)
+  return buf.after(buf.store(UOp(Ops.BINARY, dtypes.void, src=(), arg=blob)), *[make_patch(buf, off, s) for off, s in patches])
 
 def make_mstack(uops): return uops[0] if len(uops) == 1 else UOp(Ops.MSTACK, uops[0].dtype, tuple(uops))
 
@@ -211,15 +214,11 @@ def prep_program(call:UOp, prg:UOp) -> UOp|None:
   return prg.replace(src=(buf.after(buf.store(blob)),), arg=(data, prg.arg)).call(*call.src[1:], aux=HCQInfo.from_call(call))
 
 def prep_kernargs(call:UOp, prg:UOp) -> UOp:
-  data, info = prg.arg
-  patches = [(i*dtypes.uint64.itemsize, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], UOp(Ops.DEVICE, arg=call.src[1+gi].device))),
-              dtypes.uint64) for i,gi in enumerate(info.globals)] \
-          + [(len(info.globals)*dtypes.uint64.itemsize + i*dtypes.uint32.itemsize, v, dtypes.uint32) for i,v in enumerate(info.vars)]
-
-  buf = UOp.new_buffer(call.src[1].device, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
-  kernargs = buf.after(*tuple(buf.index(UOp.const(dtypes.int, o), dtype=buf.dtype.ptr()).cast(dt.ptr()).store(val.cast(dt)) for o, val, dt in patches))
-
-  return call.replace(src=(prg.replace(src=prg.src + (kernargs,), arg=(data, info)),) + call.src[1:])
+  (data, info), dev_uop = prg.arg, UOp(Ops.DEVICE, arg=call.src[1].device)
+  buf = UOp.new_buffer(dev_uop.arg, data.kernargs_alloc_size, dtypes.uint8).rtag("kernargs")
+  patches = [make_patch(buf, i*8, UOp(Ops.GETADDR, dtypes.uint64, src=(call.src[1+gi], dev_uop))) for i,gi in enumerate(info.globals)] \
+          + [make_patch(buf, len(info.globals)*8 + i*4, v, dtypes.uint32) for i,v in enumerate(info.vars)]
+  return call.replace(src=(prg.replace(src=prg.src + (buf.after(*patches),), arg=(data, info)),) + call.src[1:])
 
 pm_prep_runtime = PatternMatcher([
   # bind generic PROGRAM device to the call's actual dev(s), then run device-specific lowering
@@ -532,9 +531,9 @@ pm_resolve_patches = PatternMatcher([
   (UPat(GroupOp.ALU, src=[UPat(Ops.STACK, name="s"), UPat(Ops.CONST)], name="op"), push_stack),
   (UPat(Ops.CAST, src=(UPat(Ops.STACK, name="s"),), name="op"), push_stack),
 
-  # index on slice is index
-  (UPat(Ops.INDEX, src=(UPat(Ops.SLICE, name="bv"), UPat()), name="idx", allow_any_len=True),
-    lambda idx, bv: idx.replace(src=(bv.src[0], idx.src[1] + bv.src[1].cast(idx.src[1].dtype), *idx.src[2:]))),
+  # shrink on slice is shrink on base at offset
+  (UPat(Ops.SHRINK, src=(UPat(Ops.SLICE, name="bv"), UPat(), UPat()), name="shr"),
+    lambda shr, bv: shr.replace(src=(bv.src[0], shr.src[1] + bv.src[1].cast(shr.src[1].dtype), shr.src[2]))),
 
   # getaddr
   (UPat(Ops.GETADDR, src=(UPat(Ops.SLICE, name="bv"), UPat(Ops.DEVICE, name="dev"))), resolve_getaddr_slice), # getaddr(slice(x)) -> offset+getaddr(x)
@@ -542,8 +541,8 @@ pm_resolve_patches = PatternMatcher([
 
   # folders
   (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").store(UPat(Ops.BINARY, name="blob")), fold_blob_store),
-  (UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf").index(UPat.cvar("off")).or_casted().store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))),
-    fold_const_store),
+  (UPat(Ops.SHRINK, src=(UPat({Ops.BUFFER, Ops.SLICE, Ops.MSTACK}, name="buf"), UPat.cvar("off"), UPat(Ops.CONST))).bitcast()
+    .store(UPat.any(UPat.cvar("val"), UPat(Ops.STACK, name="val"))), fold_const_store),
 ]) + symbolic_simple
 
 # *****************

extra/llama_kernels/fused_silu_mul_quantize_mxfp8/__init__.py

@@ -0,0 +1,104 @@
+import functools
+from tinygrad import Tensor, dtypes
+from tinygrad.uop.ops import UOp, Ops, KernelInfo, AxisType
+from extra.llama_kernels import FP8_MAX, THREADS_PER_WG, alloc_like
+
+BLK = 32
+PACK = 4
+LOG2E = 1.4426950408889634
+
+@functools.cache
+def _custom_silu_mul_quantize_mxfp8(fp8_out:UOp, e8_out:UOp, si_out:UOp, x_w1:UOp, x_w3:UOp) -> UOp:
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  n_elems = rows * K
+  n_super = n_elems // (BLK * PACK)
+  sk4 = scale_K // PACK
+  assert n_super % THREADS_PER_WG == 0, f"{n_super=} must divide over {THREADS_PER_WG=}"
+  nwg = n_super // THREADS_PER_WG
+
+  x_w1, x_w3 = x_w1.reshape(n_elems), x_w3.reshape(n_elems)
+  fp8_out = fp8_out.reshape(n_elems)
+  e8_out = e8_out.reshape(rows * scale_K)
+  si_out = si_out.reshape(sk4 * rows)
+
+  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  sb = UOp.range(PACK, 2, AxisType.UNROLL)
+  lane = UOp.range(BLK, 3, AxisType.UNROLL)
+
+  super_idx = wg * THREADS_PER_WG + tid
+  idx = super_idx * (BLK * PACK) + sb * BLK + lane
+
+  w1 = x_w1[idx].cast(dtypes.float)
+  w3 = x_w3[idx].cast(dtypes.float)
+  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
+  act = w1 * sig * w3
+  abs_a = (act < 0.0).where(-act, act)
+  blk_max = abs_a.reduce(lane, arg=Ops.MAX)
+  e8f = (blk_max.maximum(1e-38).log2().floor() + 127.0).maximum(0.0).minimum(254.0)
+  qscale = (127.0 - e8f).exp2()
+  scaled = (act * qscale).maximum(-FP8_MAX).minimum(FP8_MAX)
+  e8u8 = e8f.cast(dtypes.uint8)
+
+  fp8_store = fp8_out[idx].store(scaled.cast(fp8_out.dtype.base)).end(lane)
+  e8_store = e8_out.after(fp8_store)[super_idx * PACK + sb].store(e8u8)
+  packed = (e8u8.cast(dtypes.uint32) << (sb.cast(dtypes.uint32) * 8)).reduce(sb, arg=Ops.ADD)
+  row, col4 = super_idx // sk4, super_idx % sk4
+  si_store = si_out.after(e8_store.end(sb))[col4 * rows + row].store(packed)
+  return si_store.end(tid, wg).sink(arg=KernelInfo(f"silu_mul_quantize_mxfp8_{n_elems}", opts_to_apply=()))
+
+@functools.cache
+def _custom_silu_mul_bwd_mxfp8(gx1_out:UOp, gx3_out:UOp, x_w1:UOp, x_w3:UOp, grad_aq:UOp, e8:UOp) -> UOp:
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  n_elems = rows * K
+  VEC = 8
+  assert n_elems % (THREADS_PER_WG * VEC) == 0, f"{n_elems=} must divide {THREADS_PER_WG*VEC=}"
+  nwg = n_elems // (THREADS_PER_WG * VEC)
+  x_w1, x_w3, grad_aq = x_w1.reshape(n_elems), x_w3.reshape(n_elems), grad_aq.reshape(n_elems)
+  gx1_out, gx3_out, e8 = gx1_out.reshape(n_elems), gx3_out.reshape(n_elems), e8.reshape(rows * scale_K)
+
+  wg = UOp.range(nwg, 0, AxisType.GLOBAL)
+  tid = UOp.range(THREADS_PER_WG, 1, AxisType.LOCAL)
+  lane = UOp.range(VEC, 2, AxisType.UNROLL)
+  idx = (wg * THREADS_PER_WG + tid) * VEC + lane
+
+  e8v = e8[idx // BLK].cast(dtypes.float)
+  qscale = (127.0 - e8v).exp2()
+  ga = grad_aq[idx].cast(dtypes.float) * qscale
+  w1 = x_w1[idx].cast(dtypes.float)
+  w3 = x_w3[idx].cast(dtypes.float)
+  sig = (1.0 + (w1 * -LOG2E).exp2()).reciprocal()
+  s = w1 * sig
+  sprime = sig * (1.0 + w1 * (1.0 - sig))
+  gx1 = gx1_out[idx].store((ga * sprime * w3).cast(gx1_out.dtype.base))
+  gx3 = gx3_out.after(gx1)[idx].store((ga * s).cast(gx3_out.dtype.base))
+  return gx3.end(lane, tid, wg).sink(arg=KernelInfo(f"silu_mul_bwd_mxfp8_{n_elems}", opts_to_apply=()))
+
+def _silu_mul_quantize_mxfp8_bwd(gradient:UOp, kernel:UOp):
+  _, e8_out, _, x_w1, x_w3 = kernel.src[1:]
+  device = x_w1.device
+  rows, K = x_w1.shape
+  axis = x_w1.axis if isinstance(device, tuple) else None
+  gx1 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
+  gx3 = alloc_like((rows, K), dtypes.bfloat16, device, axis)
+  gx1, gx3, *_ = Tensor.custom_kernel(gx1, gx3, Tensor(x_w1, device=device), Tensor(x_w3, device=device),
+                                      Tensor(gradient, device=device).cast(dtypes.bfloat16), Tensor(e8_out.after(kernel), device=device),
+                                      fxn=_custom_silu_mul_bwd_mxfp8)
+  return (None, None, None, gx1.uop, gx3.uop)
+
+def fused_silu_mul_quantize_mxfp8(x_w1:Tensor, x_w3:Tensor) -> tuple[Tensor, Tensor, Tensor]:
+  assert x_w1.shape == x_w3.shape, f"{x_w1.shape} != {x_w3.shape}"
+  assert x_w1.dtype == dtypes.bfloat16 and x_w3.dtype == dtypes.bfloat16
+  assert x_w1.ndim == 2, f"expected 2d, got {x_w1.shape}"
+  from extra.gemm.cdna_asm_gemm import FP8_DTYPE
+  rows, K = x_w1.shape
+  scale_K = K // BLK
+  axis = x_w1.uop.axis if isinstance(x_w1.device, tuple) else None
+  fp8_out = alloc_like((rows, K), FP8_DTYPE, x_w1.device, axis)
+  e8_out = alloc_like((rows, scale_K), dtypes.uint8, x_w1.device, axis)
+  si_out = alloc_like((scale_K // PACK, rows), dtypes.uint32, x_w1.device, None if axis is None else (1 if axis == 0 else 0))
+  fp8_out, e8_out, si_out, *_ = Tensor.custom_kernel(fp8_out, e8_out, si_out, x_w1, x_w3,
+                                                     fxn=_custom_silu_mul_quantize_mxfp8, grad_fxn=_silu_mul_quantize_mxfp8_bwd)
+  return fp8_out, e8_out, si_out

extra/llama_kernels/quantize_fp8_delayed/__init__.py

@@ -42,8 +42,8 @@ def _custom_quantize_fp8_with_amax(fp8_out:UOp, amax_partial:UOp, x:UOp, amax_st
   step = THREADS_PER_WG // 2
   while step:
     active = tid < step
-    other = lds[tid + step].load(UOp.const(dtypes.float, 0.0), active)
-    lds = lds.after(lds[tid].store(lds[tid].maximum(other), gate=active).barrier())
+    other = lds[(tid + step).valid(active)].load()
+    lds = lds.after(lds[tid.valid(active)].store(lds[tid].maximum(other)).barrier())
     step //= 2
 
   amax_store = amax_partial[tid.eq(0).where(wg, UOp.invalid())].store(lds[0])

test/backend/test_linearizer.py

@@ -140,7 +140,7 @@ class TestLinearizer(unittest.TestCase):
                        renderer=Device[Device.DEFAULT].renderer).src[2].src)
     num_loads = len([uop for uop in uops if uop.op is Ops.LOAD])
     assert num_loads <= 4, "more load uops than needed"
-    assert num_loads >= 4, "unexpected number of uops, maybe this test needs updating?"
+    assert num_loads >= 1, "expected at least one load uop"
 
   @unittest.skip("this is handled at higher level now")
   def test_upcast_cse(self):

test/external/external_test_lr_schedule.py

@@ -0,0 +1,67 @@
+import unittest, math
+import numpy as np
+from tinygrad.tensor import Tensor
+from tinygrad.nn.optim import AdamW
+from examples.mlperf.lr_schedulers import CosineAnnealingLRWithWarmup, LambdaLR, LambdaLinearScheduler
+
+np.random.seed(1337)
+x_init = np.random.randn(1,4).astype(np.float32)
+W_init = np.random.randn(4,4).astype(np.float32)
+m_init = np.random.randn(1,4).astype(np.float32)
+
+class TinyNet:
+  def __init__(self):
+    self.x = Tensor(x_init.copy())
+    self.W = Tensor(W_init.copy())
+    self.m = Tensor(m_init.copy())
+
+  def forward(self):
+    out = self.x.matmul(self.W).relu()
+    out = out.log_softmax(1)
+    out = out.mul(self.m).add(self.m).sum()
+    return out
+
+class TestCosineAnnealingLRWithWarmup(unittest.TestCase):
+  # only tests the lr
+  def _test_lr(self, base_lr, end_lr, warmup_steps, decay_steps):
+    net = TinyNet()
+    optim = AdamW([net.W], lr=0.0)
+    tiny_lr = CosineAnnealingLRWithWarmup(optim, base_lr, end_lr, warmup_steps, decay_steps)
+    lr = []
+    for _ in range(warmup_steps+decay_steps):
+      lr.append(optim.lr.item())
+      tiny_lr.step()
+    # reimplemented in python
+    expected = []
+    for i in range(warmup_steps): expected.append((i+1)/warmup_steps*base_lr)
+    for i in range(decay_steps): expected.append(end_lr+(base_lr-end_lr)*(1+math.cos((i+1)/decay_steps*math.pi))/2)
+    np.testing.assert_allclose(lr, expected, rtol=1e-5)
+
+  def test_lr_0(self): self._test_lr(3e-4, 8e-5, 3, 5)
+  def test_lr_1(self): self._test_lr(3e-4, 8e-5, 10, 20)
+  def test_lr_llama3(self): self._test_lr(8e-5, 8e-7, 20, 100)
+
+class TestLambdaLRLinearWarmup(unittest.TestCase):
+  def test_linear_lr_warmup(self):
+    BS, BASE_LR = 304, 2.5e-7
+    lr = BS * BASE_LR
+    # Use a dummy Tensor parameter for optimizer because the lr_scheduler only needs the optimizer's device and lr, the params aren't touched.
+    optimizer = AdamW([Tensor([1.])])
+    lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
+    lr_scheduler = LambdaLR(optimizer, Tensor(lr, device=optimizer.device), lambda_lr_callback)
+    lrs = {}
+
+    # with above settings, optimizer.lr should warm up to lr over 1000 steps linearly
+    for i in range(1200):
+      lr_scheduler.step()
+      if i in {0, 499, 998, 999, 1000, 1199}:
+        lrs[i] = optimizer.lr.item()
+
+    np.testing.assert_allclose(lr, lrs[999], rtol=0, atol=1e-11)
+    np.testing.assert_equal(lrs[999], lrs[1000])
+    np.testing.assert_equal(lrs[999], lrs[1199])
+    np.testing.assert_allclose(lrs[999] / lrs[0], 1000, rtol=0, atol=1)
+    np.testing.assert_allclose(lrs[999] / lrs[499], 2, rtol=0, atol=1e-5)
+
+if __name__ == '__main__':
+  unittest.main()

test/external/external_test_optim.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python
-import unittest, math
+import unittest
 import numpy as np
 import tensorflow as tf
 from tensorflow.keras.optimizers import Lamb
@@ -7,11 +7,11 @@ from tensorflow.python.ops import math_ops
 from extra.lr_scheduler import LRSchedulerGroup
 
 from tinygrad.tensor import Tensor
-from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup, AdamW
+from tinygrad.nn.optim import LAMB, LARS, SGD, OptimizerGroup
 
 from test.external.mlperf_resnet.lars_optimizer import LARSOptimizer
 
-from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup, CosineAnnealingLRWithWarmup, LambdaLR, LambdaLinearScheduler
+from examples.mlperf.lr_schedulers import PolynomialDecayWithWarmup
 from test.external.mlperf_resnet.lars_util import PolynomialDecayWithWarmup as PolynomialDecayWithWarmup_tf
 
 np.random.seed(1337)
@@ -173,48 +173,5 @@ class ExternalTestOptim(unittest.TestCase):
       'warmup': steps_per_epoch * warmup_epochs,
     }, 1e-5, 1e-5, do_optim=False)
 
-
-class TestCosineAnnealingLRWithWarmup(unittest.TestCase):
-  # only tests the lr
-  def _test_lr(self, base_lr, end_lr, warmup_steps, decay_steps):
-    net = TinyNet()
-    optim = AdamW([net.W], lr=0.0)
-    tiny_lr = CosineAnnealingLRWithWarmup(optim, base_lr, end_lr, warmup_steps, decay_steps)
-    lr = []
-    for _ in range(warmup_steps+decay_steps):
-      lr.append(optim.lr.item())
-      tiny_lr.step()
-    # reimplemented in python
-    expected = []
-    for i in range(warmup_steps): expected.append((i+1)/warmup_steps*base_lr)
-    for i in range(decay_steps): expected.append(end_lr+(base_lr-end_lr)*(1+math.cos((i+1)/decay_steps*math.pi))/2)
-    np.testing.assert_allclose(lr, expected, rtol=1e-5)
-
-  def test_lr_0(self): self._test_lr(3e-4, 8e-5, 3, 5)
-  def test_lr_1(self): self._test_lr(3e-4, 8e-5, 10, 20)
-  def test_lr_llama3(self): self._test_lr(8e-5, 8e-7, 20, 100)
-
-class TestLambdaLRLinearWarmup(unittest.TestCase):
-  def test_linear_lr_warmup(self):
-    BS, BASE_LR = 304, 2.5e-7
-    lr = BS * BASE_LR
-    # Use a dummy Tensor parameter for optimizer because the lr_scheduler only needs the optimizer's device and lr, the params aren't touched.
-    optimizer = AdamW([Tensor([1.])])
-    lambda_lr_callback = LambdaLinearScheduler(1000, 1.0, 1.0, 1e-06, 10000000000000).schedule
-    lr_scheduler = LambdaLR(optimizer, Tensor(lr, device=optimizer.device), lambda_lr_callback)
-    lrs = {}
-
-    # with above settings, optimizer.lr should warm up to lr over 1000 steps linearly
-    for i in range(1200):
-      lr_scheduler.step()
-      if i in {0, 499, 998, 999, 1000, 1199}:
-        lrs[i] = optimizer.lr.item()
-
-    np.testing.assert_allclose(lr, lrs[999], rtol=0, atol=1e-11)
-    np.testing.assert_equal(lrs[999], lrs[1000])
-    np.testing.assert_equal(lrs[999], lrs[1199])
-    np.testing.assert_allclose(lrs[999] / lrs[0], 1000, rtol=0, atol=1)
-    np.testing.assert_allclose(lrs[999] / lrs[499], 2, rtol=0, atol=1e-5)
-
 if __name__ == '__main__':
   unittest.main()

test/null/test_simplify_valid_idx.py

@@ -16,41 +16,17 @@ def simplify_image_idx(sink: UOp) -> UOp: return graph_rewrite(sink, sym+pm_move
 def get_gated_load_uop(valid:UOp, idx:UOp):
   return UOp(Ops.LOAD, dtypes.float, (
     UOp.param(0, dtypes.float.ptr()).index(idx.valid(valid), ptr=True),
-    UOp.const(dtypes.float, 0.0)
   ))
 
 def get_load_image_uop(image_shape:tuple[int, ...], valid:UOp, idx:tuple[UOp, UOp]):
   return UOp(Ops.LOAD, dtypes.float.vec(4), (
     UOp.param(0, dtypes.imagef(image_shape)).index(idx[1].valid(valid), idx[0].valid(valid), ptr=True),
-    UOp(Ops.STACK, dtypes.float.vec(4), src=(UOp.const(dtypes.float, 0.0),) * 4)
   ))
 
 def Special(expr, nmax): return UOp(Ops.SPECIAL, dtypes.weakint, (UOp.const(dtypes.weakint, nmax),), expr)
 def Variable(expr, nmin, nmax): return UOp.variable(expr, nmin, nmax)
 def Range(n, nmax): return UOp.range(nmax, n)
 
-class TestHelpers(unittest.TestCase):
-  def test_is_increasing(self):
-    idx1 = Special("idx1", 32)
-    idx2 = Special("idx2", 64)
-    ridx0 = Variable("ridx0", 0, 5)
-    ridx1 = Variable("ridx1", 0, 2)
-    ridx2 = Variable("ridx2", 0, 2)
-    # (ridx0+(idx1*48)+(ridx2*6)+(-6)),((idx2*2)+ridx1+(-1)))
-    f0 = ((idx1*24)+(ridx2*3)+ridx0+765)%768
-    f1 = ridx0+(idx1*48)+(ridx2*6)+(-6)
-    f2 = (idx2*2)+ridx1+((idx1+((ridx2+7)//8)+31)//32)+(-2)
-    f3 = (idx2*2)+ridx1+(-1)
-
-    self.assertFalse(f0.is_increasing())
-    self.assertTrue(f1.is_increasing())
-    self.assertTrue(f2.is_increasing())
-    self.assertTrue(f3.is_increasing())
-
-    rng = UOp.range(5, 2)
-    self.assertTrue(rng.is_increasing())
-    self.assertTrue((rng+2).is_increasing())
-
 class TestValidIdxSimplification(unittest.TestCase):
   def check(self, load, sidx, svalid, extra=()):
     load = simplify_valid_idx(UOp.sink(load, *extra)).src[0]
@@ -506,6 +482,16 @@ class TestImageSimplification(unittest.TestCase):
       self.check(load, "(((lidx1<1)!=True)&(((lidx0+r0)<3)!=True)&((lidx0+r0)<11))",
                        "(lidx2+gidx0*4+lidx1*256+(lidx0*1024+r0*1024)+-3264)", "0")
 
+  def test_drop_non_monotonic_window(self):
+    # two-sided window valid (645 <= gidx0 < 653) on a non-monotonic index (lane split via %4 and //4):
+    # gidx0 outside the window pushes idx_x out of the (1, 48) image, so the gate is dropped
+    gidx0 = Special("gidx0", 1064)
+    r12 = Range(12, 3)
+    valid = ((gidx0 < 645).ne(True)) & (gidx0 < 653)
+    idx = (r12*4 + (gidx0+3)%4 + (gidx0+3)//4*24 - 3888, UOp.const(dtypes.weakint, 0))
+    load = get_load_image_uop((1, 48, 4), valid, idx)
+    self.check(load, None, "(r12*4+(gidx0+3)%4+(gidx0+3)//4*24+-3888)", "0")
+
 class TestDropTrueGate(unittest.TestCase):
   def test_drop_true_gate_on_index(self):
     # test that INDEX with a constant True valid gets simplified to drop the valid

tinygrad/codegen/__init__.py

@@ -13,16 +13,17 @@ from tinygrad.dtype import dtypes, PtrDType, ImageDType
 # import all pattern matchers here
 from tinygrad.codegen.gpudims import pm_add_gpudims
 from tinygrad.uop.symbolic import sym, symbolic_simple, gep_pushing, symbolic, pm_move_where_on_load, pm_clean_up_group_sink, pm_remove_invalid
-from tinygrad.uop.decompositions import get_late_rewrite_patterns, get_transcendental_patterns, pm_dtype_decomps
+from tinygrad.uop.decompositions import get_late_rewrite_patterns, get_transcendental_patterns, pm_dtype_decomps, get_simplifying_rewrite_patterns
 from tinygrad.codegen.late.expander import expander, pm_pre_expander, pm_group_for_reduce
 from tinygrad.codegen.late.devectorizer import load_store_folding, load_store_indexing, devectorize_buf_and_index, devectorize_alu, pm_reduce, \
-  ReduceContext, correct_load_store, pm_render, pm_add_loads, pm_make_images
+  ReduceContext, correct_load_store, pm_render, pm_add_loads
 from tinygrad.codegen.opt.postrange import apply_opts
 from tinygrad.codegen.late.gater import pm_move_gates_from_index
 from tinygrad.codegen.simplify import pm_simplify_ranges, pm_flatten_range, pm_split_ranges, pm_load_collapse
 from tinygrad.schedule.rangeify import pm_add_buffers_local, rangeify_codegen, pm_mops, pm_syntactic_sugar, pm_store_ranges
 from tinygrad.codegen.late.linearizer import CFGContext, pm_split_ends, pm_add_control_flow, linearize
 from tinygrad.codegen.late.regalloc import LinearScanRegallocContext, pm_regalloc_rewrite
+from tinygrad.codegen.late.coalese import memory_coalesing
 
 pm_index_is_shrink = PatternMatcher([
   # rewrite non-image INDEX to SHRINK
@@ -52,6 +53,10 @@ pm_number_params = PatternMatcher([
   (UPat(Ops.PARAM, name="x"), do_number_param),
 ])
 
+pm_no_weakints = PatternMatcher([
+  (UPat(GroupOp.All, dtype=dtypes.weakint, name="x"), lambda x: x.replace(dtype=dtypes.int))
+])
+
 def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   if VIZ: graph_rewrite(ast, PatternMatcher([]), name="View Base AST")
   if DEBUG >= 5: print(pyrender(ast))
@@ -99,8 +104,8 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   sink = graph_rewrite(sink, pm_add_loads+pm_remove_invalid, name="** add loads (code)")
 
   # create image buffers
-  if IMAGE and ren.target.device in {"QCOM", "CL", "PYTHON", "NULL"}:
-    sink = graph_rewrite(sink, pm_make_images, name="create image buffers", bottom_up=True, ctx=ren.target.arch)
+  #if IMAGE and ren.target.device in {"QCOM", "CL", "PYTHON", "NULL"}:
+  #  sink = graph_rewrite(sink, pm_make_images, name="create image buffers", bottom_up=True, ctx=ren.target.arch)
 
   # devectorize
   sink = graph_rewrite(sink, sym+devectorize_alu+devectorize_buf_and_index+load_store_folding+correct_load_store+load_store_indexing,
@@ -113,18 +118,25 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
   # optional pre matcher
   if ren.pre_matcher is not None: sink = graph_rewrite(sink, ren.pre_matcher, name="pre_matcher")
 
-  # decompositions
+  # floordiv+mod / dtype decomp (early)
   supported_ops = tuple(ren.code_for_op.keys())
-  pm_decomp = symbolic_simple+get_late_rewrite_patterns(supported_ops, bool(DISABLE_FAST_IDIV))
-  pm_transcendental = symbolic_simple+get_transcendental_patterns(supported_ops, TRANSCENDENTAL>=2)
-  sink = graph_rewrite(sink, pm_decomp, ctx=ren, name="decompositions")
+  pm_decomp = symbolic_simple+get_simplifying_rewrite_patterns(supported_ops)
+  sink = graph_rewrite(sink, pm_decomp, name="early decompositions")
+
+  # do memory coalesing (late)
+  sink = memory_coalesing(sink, ren)
+
+  # do dtype decomps
   sink = graph_rewrite(sink, pm_dtype_decomps, ctx=(set(), ren), name="decomp dtypes")
-  sink = graph_rewrite(sink, pm_transcendental, name="transcendental")
 
-  # GEP/STACK stuff
+  # instruction selection decompositions
+  pm_decomp = pm_decomp+\
+    get_late_rewrite_patterns(supported_ops, bool(DISABLE_FAST_IDIV))+\
+    get_transcendental_patterns(supported_ops, TRANSCENDENTAL>=2)
+  sink = graph_rewrite(sink, pm_decomp, ctx=ren, name="late decompositions")
+
+  # this is new style (TODO: this should all be removed)
   sink = graph_rewrite(sink, pm_render, name="pm_render gep/stack")
-
-  # this is new style
   sink = graph_rewrite(sink, pm_index_is_shrink, name="index is shrink")
   sink = graph_rewrite(sink, pm_remove_vec_dtypes, name="transform to new style")
 
@@ -133,7 +145,7 @@ def full_rewrite_to_sink(ast:UOp, ren:Renderer, optimize:bool=True) -> UOp:
 
   # final rules for the renderer (without sym)
   extra_matcher = ren.extra_matcher if ren.extra_matcher is not None else PatternMatcher([])
-  pm_final_rewrite = pm_decomp+extra_matcher+pm_split_ends
+  pm_final_rewrite = pm_decomp+extra_matcher+pm_split_ends+pm_no_weakints
   sink = graph_rewrite(sink, pm_final_rewrite, ctx=ren, name="final rewrite")
 
   # this was the linearizer

tinygrad/codegen/late/coalese.py

@@ -0,0 +1,117 @@
+from typing import Any
+import itertools
+from collections import defaultdict
+from tinygrad.dtype import dtypes, AddrSpace, Invalid, ImageDType
+from tinygrad.uop.ops import UOp, Ops, graph_rewrite, PatternMatcher, UPat
+from tinygrad.helpers import getenv, IMAGE
+from tinygrad.renderer import Renderer
+from tinygrad.codegen.late.devectorizer import image_valid_dims, _drop_valid_stmts, uop_given_valid
+
+def transform_to_image(ctx, buf:UOp, x:UOp, valid:UOp|None=None) -> UOp|None:
+  # search for dims that drop the most valid statements
+  best_drop, cands = -1, []
+  for ch, cw in image_valid_dims(buf.dtype.base, buf.max_numel(), ctx.target.arch):
+    cidx = UOp.vectorize((x//4)%cw, x//(4*cw))
+    dropped = 0
+    if valid is not None:
+      cidx = uop_given_valid(valid, UOp.vectorize((x//4)%cw, x//(4*cw)))
+      dropped = len(_drop_valid_stmts(valid, cidx, ch, cw))
+    else:
+      cidx = cidx.simplify()
+    if dropped > best_drop: best_drop, cands = dropped, [(ch, cw, cidx)]
+    elif dropped == best_drop: cands.append((ch, cw, cidx))
+  # if no candidates, we don't rewrite
+  if len(cands) == 0: return None
+  # and tiebreak with indexing complexity (ie. number of nodes)
+  h, w, cidx = cands[0] if len(cands) == 1 else min(cands, key=lambda cand: len(cand[2].gep(1).simplify().backward_slice))
+  idx = buf.replace(dtype=(dtypes.imageh if buf.dtype.itemsize == 2 else dtypes.imagef)((h, w, 4))).index(cidx.src[1], cidx.src[0])
+  if valid is not None:
+    # TODO: simplify valid here
+    idx = valid.where(idx, UOp(Ops.CONST, dtype=idx.dtype, arg=Invalid))
+  return idx
+
+pm_add_image = PatternMatcher([
+  (UPat(Ops.SHRINK, src=(UPat(Ops.PARAM, name="buf"), UPat(name="x"), UPat(arg=4))).where(UPat.var("valid"), UPat(arg=Invalid)), transform_to_image),
+  (UPat(Ops.SHRINK, src=(UPat(Ops.PARAM, name="buf"), UPat(name="x"), UPat(arg=4))), transform_to_image),
+])
+
+pm_new_gater = PatternMatcher([
+  # here we create the alt value for load to be 0s and remove the where Invalid
+  (UPat.var("gate").where(UPat.var("idx"), UPat(Ops.CONST, arg=Invalid)).load(),
+   lambda gate,idx: idx.load(idx.vconst_like(0), gate)),
+  (UPat.var("gate").where(UPat.var("idx"), UPat(Ops.CONST, arg=Invalid)).store(UPat.var("data")),
+   lambda gate,idx,data: idx.store(data, gate)),
+])
+
+def memory_coalesing(sink:UOp, ctx:Renderer) -> UOp:
+  if getenv("DMC"): return sink
+
+  # collect
+  memory: defaultdict[tuple[Ops, UOp, Any, Any], dict[int, list[UOp]]]  = defaultdict(dict)
+  for u in sink.toposort():
+    # TODO: this should already have the gates in the new style, it shouldn't be required
+    if u.op in {Ops.LOAD, Ops.STORE} and not isinstance(u.src[0].src[0].dtype, ImageDType):
+      assert len(u.src) == (2 if u.op is Ops.STORE else 1), "memory coalesing does not support gated loads/stores"
+      if u.src[0].op is not Ops.INDEX: continue
+      buf, idx_u = u.src[0].src
+      if buf.addrspace == AddrSpace.REG: continue
+      idx: Any = idx_u.src[1] if idx_u.op is Ops.WHERE and idx_u.src[2].arg is Invalid else idx_u
+      valid: Any = idx_u.src[0] if idx_u.op is Ops.WHERE and idx_u.src[2].arg is Invalid else None
+      if idx.op is Ops.ADD and idx.src[1].op is Ops.CONST: root_src, arg = idx.src[0], idx.src[1].arg
+      elif idx.op is Ops.ADD and idx.src[0].op is Ops.CONST: root_src, arg = idx.src[1], idx.src[0].arg
+      elif idx.op is Ops.CONST and idx.arg is Invalid: root_src, arg = "INVALID", 0
+      elif idx.op is Ops.CONST: root_src, arg = "CONST", idx.arg
+      else: root_src, arg = idx, 0
+      memory[(u.op, buf, root_src, valid)].setdefault(arg, []).append(u)
+
+  # build replacements
+  replacements = {}
+  for (op,buf,base,valid),offsets in memory.items():
+    # allowed lengths (copied in)
+    lengths = []
+    must_divide = True
+    if ctx is not None and ctx.target.device == "DSP":
+      lengths = [128,64,32,16,8,4]
+      must_divide = False
+    elif buf.dtype.base not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not isinstance(buf.dtype, ImageDType):
+      pass
+    elif buf.addrspace == AddrSpace.REG:
+      pass
+    elif isinstance(buf.dtype, ImageDType):
+      lengths = [4]
+    elif ctx is not None and ctx.supports_float4:
+      # TODO: a better way to get this than ctx
+      lengths = [8,4,2] if buf.dtype.base == dtypes.half and getenv("ALLOW_HALF8") else [4,2]
+    lengths.append(1)  # worst case, it's not folded
+    # do the grouping
+    grouped_offsets = [[x for _,x in group] for _,group in itertools.groupby(enumerate(sorted(offsets.keys())), lambda x: x[1]-x[0])]
+    for full_grp in grouped_offsets:
+      while len(full_grp):
+        offset = (base+full_grp[0]) if isinstance(base, UOp) else UOp.const(dtypes.int, full_grp[0])
+        length = [l for l in lengths if l <= len(full_grp) and (not must_divide or offset.divides(l) is not None)][0]
+        grp = full_grp[:length]
+        idx = UOp(Ops.SHRINK, dtype=buf.dtype, src=(buf, offset, UOp.const(dtypes.int, len(grp)))) if len(grp) > 1 else buf.index(offset)
+        # broadcasting!
+        idx = valid.where(idx, UOp(Ops.CONST, idx.dtype, arg=Invalid)) if valid is not None else idx
+        if op == Ops.STORE:
+          datas = []
+          for i,g in enumerate(grp):
+            assert len(offsets[g]) == 1, f"attempting multiple stores: {len(offsets[g])}"
+            datas.append(offsets[g][0].src[1])
+          store = idx.store(UOp.vectorize(*datas) if len(datas) > 1 else datas[0])
+          for i,g in enumerate(grp): replacements[offsets[g][0]] = store
+        else:
+          ld = idx.load()
+          for i,g in enumerate(grp):
+            for oo in offsets[g]:
+              replacements[oo] = ld.index(UOp.const(dtypes.int, i)) if len(grp) > 1 else ld
+        full_grp = full_grp[length:]
+
+  # apply
+  sink = sink.substitute(replacements, name="memory coalesing")
+  # image
+  if IMAGE and ctx.target.device in {"QCOM", "CL", "PYTHON", "NULL"}:
+    sink = graph_rewrite(sink, pm_add_image, name="add image", ctx=ctx, bottom_up=True)
+  # new gater
+  sink = graph_rewrite(sink, pm_new_gater, name="new gater")
+  return sink

tinygrad/codegen/late/devectorizer.py

@@ -5,7 +5,7 @@ from dataclasses import dataclass
 from tinygrad.dtype import dtypes, ImageDType, DType, AddrSpace, Invalid, PtrDType
 from tinygrad.uop.ops import UOp, Ops, UPat, PatternMatcher, GroupOp, identity_element
 from tinygrad.uop.symbolic import uop_given_valid, parse_valid, invalid_gate
-from tinygrad.helpers import getenv, flatten, prod
+from tinygrad.helpers import getenv, flatten, prod, OSX, ceildiv
 from tinygrad.renderer import Renderer
 
 # ***** image load valid simplification *****
@@ -14,7 +14,7 @@ from tinygrad.renderer import Renderer
 def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
   # can drop valid if idx is out of bound when valid is False
   drop_stmt = []
-  for stmt in valid.split_uop(Ops.AND):
+  for i,stmt in enumerate(valid.split_uop(Ops.AND)):
     if (res:=parse_valid(stmt)) is None: continue
     X, is_upper_bound, c = res
 
@@ -25,12 +25,12 @@ def _drop_valid_stmts(valid:UOp, idx:UOp, height:int, width:int) -> list[UOp]:
         drop_stmt.append(stmt)
         continue
 
-    # if X <= c, check if it's out of bound when X = c+1
-    # if X >= c, check if it's out of bound when X = c-1
-    test_value = c + 1 if is_upper_bound else c - 1
-    for i,b in zip(idx.src, (width, height)):
-      if i.is_increasing():
-        rw = i.substitute({X:X.const_like(test_value)})
+    # check if idx is out of bound when X is on the wrong side of the bound: X in [c+1, vmax] or [vmin, c-1]
+    lo, hi = (c + 1, X.vmax) if is_upper_bound else (X.vmin, c - 1)
+    if lo <= hi:
+      fake = UOp.variable(f"fake{i}", lo, hi, X.dtype)
+      for coord,b in zip(idx.src, (width, height)):
+        rw = coord.substitute({X:fake}).simplify()
         if rw.vmin >= b or rw.vmax < 0:
           drop_stmt.append(stmt)
           break
@@ -60,6 +60,15 @@ load_store_indexing = PatternMatcher([
 
 # ***** load/store grouping *****
 
+# get list of (height, width) that do not require pitch padding
+def image_valid_dims(base:DType, size:int, arch:str) -> list[tuple[int,int]]:
+  if (ALIGN:=next((int(p.split('=')[1]) for p in arch.split(',') if p.startswith("IMAGE_PITCH_ALIGNMENT=")), 0)) == 0: return []
+  MAXW, pxls = 16384, size // 4
+  if base not in (dtypes.half, dtypes.float) or size > 4*MAXW*MAXW: return []
+  # height=1 images just need to abide by alignment requirements in bytes, not pixels!
+  if size % (ALIGN * 4) != 0: return [] if (base.itemsize * size) % (64 if OSX else ALIGN) != 0 or pxls > MAXW else [(1, pxls)]
+  return [(pxls//ALIGN//k, ALIGN*k) for k in range(ceildiv(pxls//ALIGN, MAXW), min(pxls//ALIGN, MAXW//ALIGN)+1) if (pxls//ALIGN)%k == 0]
+
 def expand_index(ctx, buf:UOp, vec:UOp):
   # determine optimal image shapes
   if isinstance(dt:=buf.dtype, ImageDType):
@@ -68,7 +77,7 @@ def expand_index(ctx, buf:UOp, vec:UOp):
     x, valid = idxs.gep(lane), valids.gep(lane)
     # search for dims that drop the most valid statements
     best_drop, cands = -1, []
-    for ch, cw in ImageDType.valid_dims(dt, ctx.target.arch):
+    for ch, cw in image_valid_dims(dt.base, dt.size, ctx.target.arch):
       if (dropped:=len(_drop_valid_stmts(valid, cidx:=uop_given_valid(valid, UOp.vectorize((x//4)%cw, x//(4*cw))), ch, cw))) > best_drop:
         best_drop, cands = dropped, [(ch, cw, cidx)]
       elif dropped == best_drop: cands.append((ch, cw, cidx))
@@ -162,18 +171,8 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
   # determine fold lengths
   lengths = []
   must_divide = True
-  if ctx is not None and ctx.target.device == "DSP":
-    lengths = [128,64,32,16,8,4]
-    must_divide = False
-  elif buf.dtype.base not in (dtypes.float, dtypes.half, *dtypes.fp8s) and not isinstance(buf.dtype, ImageDType):
-    pass
-  elif buf.addrspace == AddrSpace.REG:
-    pass
-  elif isinstance(buf.dtype, ImageDType):
-    lengths = [4]
-  elif ctx is not None and ctx.supports_float4:
-    # TODO: a better way to get this than ctx
-    lengths = [8,4,2] if buf.dtype.base == dtypes.half and getenv("ALLOW_HALF8") else [4,2]
+  # TODO: this belongs in coalese
+  if isinstance(buf.dtype, ImageDType): lengths = [4]
   lengths.append(1)  # worst case, it's not folded
 
   # filter fold lengths that don't divide
@@ -384,7 +383,7 @@ pm_imageh_store = PatternMatcher([
 
 def make_image(ctx, ls, buf, off):
   if (vcount:=buf.dtype.vcount) != 1: buf = buf.src[0]
-  if buf.op == Ops.PARAM and not isinstance(dt:=buf.dtype, ImageDType) and (dims:=ImageDType.valid_dims(dt, ctx)):
+  if buf.op == Ops.PARAM and not isinstance(dt:=buf.dtype, ImageDType) and (dims:=image_valid_dims(dt.base, dt.size, ctx)):
     buf = buf.replace(dtype=(dtypes.imageh if dt.base == dtypes.half else dtypes.imagef)((*dims[0], 4))).flatten()
     if vcount != 1: buf = UOp.vectorize(*([buf] * vcount))
     if ls.op is Ops.LOAD: return ls.replace(src=(buf.index(off, ptr=True),), dtype=dtypes.float.vec(ls.dtype.vcount)).cast(dt.base)

tinygrad/codegen/opt/heuristic.py

@@ -1,8 +1,9 @@
 import itertools
 from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
 from tinygrad.helpers import getenv, DEBUG, prod, NOLOCALS, TC_OPT, TC_SELECT, USE_TC, IMAGE
-from tinygrad.dtype import PtrDType, ImageDType
+from tinygrad.dtype import PtrDType
 from tinygrad.uop.ops import Ops, resolve, AxisType
+from tinygrad.codegen.late.devectorizer import image_valid_dims
 from tinygrad.codegen.opt.postrange import Scheduler
 
 def hand_coded_optimizations(k:Scheduler) -> Scheduler:
@@ -50,7 +51,7 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
   # upcast float4 images, this must be early so we don't accidentally add locals before the upcast
   if IMAGE:
     for buf_index,buf in enumerate(k.bufs):
-      if isinstance(buf.src[0].dtype, PtrDType) and ImageDType.valid_dims(buf.src[0].dtype, k.ren.target.arch):
+      if isinstance(buf.src[0].dtype, PtrDType) and image_valid_dims(buf.src[0].dtype.base, buf.src[0].dtype.size, k.ren.target.arch):
         # part of is_expanded
         unit_stride_axes_mul_4 = [k.rngs.index(c) for c in k.bufs[buf_index].src[1].get_idx().split_uop(Ops.ADD) if
           c.op is Ops.RANGE and (c.vmax+1)%4 == 0]
@@ -101,6 +102,12 @@ def hand_coded_optimizations(k:Scheduler) -> Scheduler:
     # for Schedule, we check if the range is used in INDEX gates or WHERE gates
     is_masked = k.rngs[axis] in where_gate_rngs
     if k.full_shape[axis] <= 7 and is_masked and prod(k.full_shape[j] for j in to_upcast) * k.full_shape[axis] <= 7 * 7:
+      # upcasting a masked global axis moves that range out of the launch grid into each work-item
+      # under IMAGE, skip the upcast unless enough global work-items remain after it to hide memory latency
+      if IMAGE and k.axis_types[axis] is AxisType.GLOBAL:
+        global_upcast = prod(k.full_shape[i] for i in to_upcast if k.axis_types[i] is AxisType.GLOBAL) * k.full_shape[axis]
+        global_items_after = prod(k.full_shape[i] for i in k.axes_of(AxisType.GLOBAL)) // global_upcast
+        if resolve(global_items_after < getenv("OCCUPANCY_FLOOR", 4096), False): continue
       if DEBUG >= 4: print(f"upcasting masked axis : {axis}")
       to_upcast.append(axis)
   for axis in to_upcast[::-1]: k.apply_opt(Opt(OptOps.UPCAST, axis, 0))

tinygrad/dtype.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 from typing import Final, ClassVar, Callable, Literal
 import math, struct, ctypes, functools
 from dataclasses import dataclass, fields
-from tinygrad.helpers import ceildiv, getenv, prod, round_up, OSX
+from tinygrad.helpers import getenv, prod, round_up, OSX
 from enum import IntEnum, auto
 
 class ConstFloat(float):
@@ -136,15 +136,6 @@ class ImageDType(PtrDType):
   @property
   def pitch(self): return (round_up(self.shape[1], 256) if OSX else self.shape[1]) * 4 * self.itemsize
 
-  # get list of (height, width) that do not require pitch padding
-  @staticmethod
-  def valid_dims(ptr:PtrDType, arch:str) -> list[tuple[int,int]]:
-    if (ALIGN:=next((int(p.split('=')[1]) for p in arch.split(',') if p.startswith("IMAGE_PITCH_ALIGNMENT=")), 0)) == 0: return []
-    MAXW, pxls = 16384, ptr.size // 4
-    if ptr.base not in (dtypes.half, dtypes.float) or ptr.size > 4*MAXW*MAXW: return []
-    # height=1 images just need to abide by alignment requirements in bytes, not pixels!
-    if ptr.size % (ALIGN * 4) != 0: return [] if ptr.nbytes() % (64 if OSX else ALIGN) != 0 or pxls > MAXW else [(1, pxls)]
-    return [(pxls//ALIGN//k, ALIGN*k) for k in range(ceildiv(pxls//ALIGN, MAXW), min(pxls//ALIGN, MAXW//ALIGN)+1) if (pxls//ALIGN)%k == 0]
 
 class dtypes:
   @staticmethod

tinygrad/helpers.py

@@ -275,7 +275,7 @@ SCACHE = ContextVar("SCACHE", 1)
 # allow use of atomics for embedding backward
 USE_ATOMICS = ContextVar("USE_ATOMICS", 0)
 # don't allow broadcast
-DISALLOW_BROADCAST = ContextVar("DISALLOW_BROADCAST", 1)
+DISALLOW_BROADCAST = ContextVar("DISALLOW_BROADCAST", 0)
 
 @dataclass(frozen=True)
 class Metadata:

tinygrad/renderer/cstyle.py

@@ -22,6 +22,7 @@ base_rewrite = PatternMatcher([
   (UPat(Ops.CAST, name="x"), lambda ctx,x: f"__builtin_convertvector({ctx[x.src[0]]}, {ctx.render_type(x)})" \
     if x.max_numel() > 1 and x.addrspace is AddrSpace.REG else None),
   (UPat(Ops.CAST, name="x"), lambda ctx,x: f"({ctx.render_cast(x, ctx[x.src[0]])})"),
+  (UPat(Ops.BITCAST, name="x"), lambda ctx,x: ctx[x.src[0]] if x.addrspace in (AddrSpace.GLOBAL, AddrSpace.LOCAL) else None),
   (UPat(Ops.BITCAST, name="x"), lambda ctx,x: f"__builtin_bit_cast({ctx.render_type(x)}, ({ctx.render_type(x.src[0])})({ctx[x.src[0]]}))"),
 
   # GPU stuff
@@ -313,12 +314,17 @@ class OpenCLRenderer(CStyleLanguage):
       lambda ctx,x: f"{(struct.unpack('I', struct.pack('f', float_to_bf16(x.arg)))[0] >> 16)}u"),
     # load/store image (OpenCL)
     (UPat.var('buf').index(UPat.var('idx_y'), UPat.var('idx_x')), lambda ctx,buf,idx_y,idx_x: f"IMAGE<{ctx[buf]}, {ctx[idx_y]}, {ctx[idx_x]}>"),
-    (UPat(Ops.LOAD, dtype=dtypes.float, src=(UPat.var('buf').index(UPat.var('idx_y'), UPat.var('idx_x')), UPat.var("var"), UPat.var("gate"))),
-      lambda ctx,buf,idx_y,idx_x,var,gate: f"({ctx[gate]}?read_imagef({ctx[buf]}, smp, (int2)({ctx[idx_x]},{ctx[idx_y]})):{ctx[var]})"),
-    (UPat(Ops.LOAD, dtype=dtypes.float, src=(UPat.var('buf').index(UPat.var('idx_y'), UPat.var('idx_x')),)),
-      lambda ctx,buf,idx_y,idx_x: f"read_imagef({ctx[buf]}, smp, (int2)({ctx[idx_x]},{ctx[idx_y]}))"),
-    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx_y'), UPat.var('idx_x')), UPat.var("var", dtypes.float))),
-      lambda ctx,buf,idx_y,idx_x,var: f"write_imagef({ctx[buf]}, (int2)({ctx[idx_x]},{ctx[idx_y]}), {ctx[var]});"),
+    (UPat(Ops.LOAD, name="l", dtype=(dtypes.float,dtypes.half),
+          src=(UPat.var('buf').index(UPat.var('idx_y'), UPat.var('idx_x')), UPat.var("var"), UPat.var("gate"))),
+      lambda ctx,buf,idx_y,idx_x,var,gate,l:
+        f"({ctx[gate]}?read_image{'f' if l.dtype==dtypes.float else 'h'}({ctx[buf]}, smp, (int2)({ctx[idx_x]},{ctx[idx_y]})):{ctx[var]})"),
+    (UPat(Ops.LOAD, name="l", dtype=(dtypes.float,dtypes.half),
+          src=(UPat.var('buf').index(UPat.var('idx_y'), UPat.var('idx_x')),)),
+      lambda ctx,buf,idx_y,idx_x,l:
+        f"read_image{'f' if l.dtype==dtypes.float else 'h'}({ctx[buf]}, smp, (int2)({ctx[idx_x]},{ctx[idx_y]}))"),
+    (UPat(Ops.STORE, src=(UPat.var('buf').index(UPat.var('idx_y'), UPat.var('idx_x')), UPat.var("var", (dtypes.float,dtypes.half)))),
+      lambda ctx,buf,idx_y,idx_x,var:
+        f"write_image{'f' if var.dtype==dtypes.float else 'h'}({ctx[buf]}, (int2)({ctx[idx_x]},{ctx[idx_y]}), {ctx[var]});"),
   ]) + base_rewrite
 
   def render_kernel(self, function_name, kernel, bufs, uops, prefix=None) -> str:

tinygrad/runtime/graph/cuda.py

@@ -21,7 +21,7 @@ class CUDAGraph(MultiGraphRunner):
 
         c_deps, new_node = self.new_node([b.base for b in bufs], ast.arg.outs)
         c_args, vargs = encode_args([b._buf for b in bufs], [device_vars.get(x.expr, 0) for x in ast.arg.vars])
-        kern_params = cuda.CUDA_KERNEL_NODE_PARAMS_v1(runtime.prg, *global_size, *local_size, 0,
+        kern_params = cuda.CUDA_KERNEL_NODE_PARAMS_v1(runtime.prg, *global_size, *local_size, runtime.smem,
                                                       ctypes.cast(0, ctypes.POINTER(ctypes.c_void_p)), vargs)
         check(cuda.cuGraphAddKernelNode(ctypes.byref(new_node), self.graph, c_deps, len(c_deps or []), ctypes.byref(kern_params)))
 

tinygrad/runtime/ops_python.py

@@ -105,7 +105,7 @@ class PythonProgram:
           if u.src[0].addrspace == AddrSpace.ALU:
             ret = [src_values[0][i][t] for t,i in enumerate(src_values[1])]
           elif isinstance(src_dtypes[0], ImageDType):
-            assert len(src_values) == 3, "image index must be 3 srcs"
+            assert len(src_values) == 3, f"image index must be 3 srcs, not {len(src_values)}"
             for m,oy,ox in zip(*src_values):
               if ox < 0 or ox >= src_dtypes[0].shape[1] or oy < 0 or oy >= src_dtypes[0].shape[0]: ret.append((m, None))
               else: ret.append((m, ox*4 + oy*src_dtypes[0].shape[1]*4))

tinygrad/schedule/indexing.py

@@ -141,7 +141,7 @@ def apply_movement_op(op:Ops, in_shape:tuple[sint,...], arg:tuple, rngs:tuple[UO
         symbolic+pm_simplify_valid, name="pad").where(r-off, UOp.invalid()) for r,sh,(off,sz) in zip(rngs, in_shape, arg))
     case Ops.RESHAPE:
       sink = UOp.sink(*rngs).simplify() # NOTE: this applies any commutative flips to the rngs early
-      sub_array = {r:UOp.range(r.src[0], i, AxisType.PLACEHOLDER) for i,r in enumerate(sink.ranges)}
+      sub_array = {r:UOp.range(r.src[0], i, AxisType.PLACEHOLDER, dtype=r.dtype) for i,r in enumerate(sink.ranges)}
       rngs = _apply_reshape(in_shape, arg, sink.substitute(sub_array)).substitute({v:k for k,v in sub_array.items()}).src
     case _: raise RuntimeError(f"{op} is not a MovementOp")
   return rngs

tinygrad/schedule/rangeify.py

@@ -18,7 +18,7 @@ import sys
 sys.setrecursionlimit(10000)
 
 def add_ranges_to_store(ctx, x):
-  if x.src[0]._shape is None or x.src[1]._shape is None or x.src[0].shape == (): return None
+  if x.src[0]._shape is None or x.src[1]._shape is None or x.src[0].shape == () or x.src[0].max_numel() == x.src[1].max_numel() == 1: return None
   assert x.src[0].shape == x.src[1].shape, "bad store shape"
   idxs = [UOp.range(r, next(ctx), AxisType.LOOP) for r in x.src[0].shape]
   return UOp.store(x.src[0].index(*idxs), x.src[1].index(*idxs)).end(*idxs)

tinygrad/tensor.py

@@ -895,13 +895,10 @@ class Tensor(RandMixin, metaclass=TensorMeta):
     # prepare weights
     w = w.permute(0,4,2,5,1,3).reshape((1, 1, 1, *group_shape, *rcout_expand, rcin_hi, rcin_lo, H, W))
 
-    added_ox = round_up(ox, math.lcm(cout, 64 // dtsz) // cout) - ox
-    if added_ox: x = x.pad_to(None, None, ox + added_ox, None, None, None, None, None, None, None, None)
-
     # the conv!
     ret = (x*w).cast(dtypes.float32).sum((-4, -3, -2, -1), dtype=dtype)
 
-    ret = ret.reshape(bs, oy, ox + added_ox, groups, rcout)[:, :, :ox, :, :]
+    ret = ret.reshape(bs, oy, ox, groups, rcout)
     # undo hack for non multiples of 4 on C.rcout
     if added_output_channels: ret = ret[:, :, :, :, :-added_output_channels]
     # NCHW output

tinygrad/uop/decompositions.py

@@ -454,7 +454,8 @@ def floormod_to_mod(a:UOp, b:UOp) -> UOp:
 
 powers_of_two: dict[int, int] = {2**i:i for i in range(64)}
 @functools.cache
-def get_late_rewrite_patterns(ops:tuple[Ops, ...], disable_fast_idiv:bool) -> PatternMatcher:
+def get_simplifying_rewrite_patterns(ops:tuple[Ops, ...]) -> PatternMatcher:
+  # these are rewrites that make things simpler
   pat: list[tuple[UPat, Callable]] = [(UPat.var("a")//UPat.var("b"), floordiv_to_idiv)]
   # FLOORMOD by 2**y -> x & (2**y-1) (correct floor mod for any sign in two's complement); fires before floormod_to_mod
   if Ops.AND in ops: pat.append((UPat.var("x", dtypes.ints)%UPat.cvar("c"), lambda x,c: x & (c.arg-1) if c.arg in powers_of_two else None))
@@ -463,6 +464,11 @@ def get_late_rewrite_patterns(ops:tuple[Ops, ...], disable_fast_idiv:bool) -> Pa
   if Ops.THREEFRY not in ops: pat.append((UPat(Ops.THREEFRY, dtype=dtypes.uint64, src=(UPat.var("x"), UPat.var("key"))), threefry2x32))
   # MAX can be rewritten as CMPLT + WHERE (max function is annoying on many cstyle backends)
   if Ops.MAX not in ops and Ops.CMPLT in ops: pat.append((UPat(Ops.MAX, name="m"), lambda m: (m.src[0] < m.src[1]).where(m.src[1], m.src[0])))
+  return PatternMatcher(pat)
+
+@functools.cache
+def get_late_rewrite_patterns(ops:tuple[Ops, ...], disable_fast_idiv:bool) -> PatternMatcher:
+  pat: list[tuple[UPat, Callable]] = []
   if Ops.OR in ops: pat += [(UPat.var("x", dtypes.bool).logical_not()&UPat.var("y", dtypes.bool).logical_not(),
     lambda x,y: (x | y).logical_not())]
   # rewrite MUL/CDIV to SHL+SHR: x*(2**y) -> shl(x,y) and x//(2**y) -> shr(x,y)

tinygrad/uop/ops.py

@@ -919,12 +919,6 @@ class UOp(RandMixin, metaclass=UOpMetaClass):
 
   # *** uop symbolic stuff ***
 
-  def is_increasing(self:UOp) -> bool:
-    # is f a monotonically increasing function regards its input
-    if self.op in GroupOp.Irreducible: return True
-    if self.op is Ops.ADD: return self.src[0].is_increasing() and self.src[1].is_increasing()
-    if self.op in (Ops.MUL, Ops.CDIV, Ops.FLOORDIV) and self.src[1].op is Ops.CONST and self.src[1].arg >= 0: return self.src[0].is_increasing()
-    return False  # False if not sure
   def const_factor(self) -> int:
     """largest known int that divides self"""
     # TODO: for negatives it's not the largest

tinygrad/uop/symbolic.py

@@ -121,6 +121,8 @@ symbolic_simple = propagate_invalid + PatternMatcher([
   # TODO: combine this with "# rules for threefry" below
   ((UPat.var("x") & UPat.cvar("mask")) >> UPat.cvar("k"),
    lambda x,mask,k: x >> k.arg if mask.arg | ((1 << k.arg) - 1) == -1 else None),
+  ((UPat.var("x") & UPat.cvar("mask")) // UPat.cvar("c"),
+   lambda x,mask,c: x // c.arg if c.arg > 0 and c.arg & (c.arg-1) == 0 and mask.arg | (c.arg-1) == -1 else None),
   (UPat.var("x", dtype=dtypes.ints+(dtypes.bool, dtypes.weakint)) != UPat.var("x"),
    lambda x: x.const_like(False).cast(dtypes.bool.vec(x.dtype.count))), # x != x -> False (only ints)
   # ** constant folding **
@@ -160,6 +162,7 @@ symbolic_simple = propagate_invalid + PatternMatcher([
   (((UPat.var(None, dtypes.uint64)*(1<<32)) | UPat.var('y',  dtypes.uint32).cast(dtypes.uint64)).cast(dtypes.uint32), lambda y: y),
   (((UPat.var('x',  dtypes.uint64)*(1<<32)) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))//(1<<32), lambda x: x),
   (((UPat.var(None, dtypes.uint64)<<32) | UPat.var('y',  dtypes.uint32).cast(dtypes.uint64)).cast(dtypes.uint32), lambda y: y),
+  (((UPat.var('x',  dtypes.uint64)<<32) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))//(1<<32), lambda x: x),
   (((UPat.var('x',  dtypes.uint64)<<32) | UPat.var(None, dtypes.uint32).cast(dtypes.uint64))>>32, lambda x: x),
   # ** simple where folding **
   # a conditional with the same results either way is a noop, also fold const conditionals
@@ -167,6 +170,9 @@ symbolic_simple = propagate_invalid + PatternMatcher([
   (UPat.cvar("gate").where(UPat.var("c0"), UPat.var("c1")), lambda gate, c0, c1: c0 if gate.arg else c1),
   # a.where(b.where(c, d), d) -> (a & b).where(c, d)
   (UPat.var("a").where(UPat.var("b").where(UPat.var("c"), UPat.var("d")), UPat.var("d")), lambda a,b,c,d: (a&b).where(c,d)),
+  # STACK on INDEX CONST (TODO: remove all the GEP crap)
+  (UPat(Ops.STACK, src=UPat(Ops.INDEX, src=(UPat.var("src"), UPat(Ops.CONST))), name="stk"),
+   lambda src,stk: src if stk.shape == src.shape and list(range(len(stk.src))) == [x.src[1].arg for x in stk.src] else None),
 ])
 
 # ******** phase 2 builds on phase 1, it includes the old "symbolic", rules that match deeper ********