File size: 11,944 Bytes
2e4c7fe
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
"""Assign each task a DIFFICULTY tier — how hard it is to IMPROVE, not how complex it looks.

Structural scope is the wrong axis. A dense GEMM is a one-liner and essentially unbeatable, because
cuBLAS/CUTLASS already sit at the hardware limit; a five-pass elementwise chain is trivial to describe
and has 5-10x of fusion headroom sitting on the table. So difficulty = HEADROOM x the TECHNIQUE DEPTH
needed to capture it.

    T1  fusion            The best available implementation is several separate passes over memory.
                          The win is doing it in one. Techniques: kernel fusion, coalesced/vectorised
                          access, keeping intermediates in registers. Typically 3-10x available.

    T2  tiling+reduction  Needs shared-memory tiling, warp/block reductions, an online (single-pass)
                          reformulation, or a layout change to make access coalesced.
                          Techniques: smem tiling, warp shuffles, online softmax, swizzles. 2-5x.

    T3  pipelined/MMA     Needs async copy (cp.async / TMA), double buffering, warp specialisation,
                          and tensor-core MMA with correct fragment layouts and bank-conflict-free
                          swizzles. The headroom is real but only reachable this way. 1.5-3x.

    T4  at-roofline       A vendor library already runs this within ~1.5x of the hardware limit.
                          Beating it means out-engineering the vendor's own kernel team. Dense GEMM
                          against cuBLAS, FA-class attention. <1.5x available.

Two signals, combined:

  * EMPIRICAL: what fraction of the roofline the shipped reference already attains. Computed from the
    validated reference metric and the audited roofline. High fraction => little left on the table.

  * STRUCTURAL: whether the reference's inner loop is a VENDOR call (torch matmul -> cuBLAS, SDPA ->
    a FlashAttention kernel, conv -> cuDNN). This matters because a vendor-backed reference IS the
    incumbent an agent has to beat, whereas a hand-written multi-pass reference is not.

Neither signal alone is enough: the roofline constant is not dtype-aware (an int8 GEMM's peak is far
above the bf16 number), so the fraction can badly understate a quantised task's difficulty. The vendor
flag corrects exactly that case.
"""
import ast
import json
import pathlib
import re
import sys

LANE = pathlib.Path(__file__).resolve().parent.parent
H200_TFLOPS, H200_GBPS, LINK_GBPS = 700.0, 4800.0, 50.0

MATMULish = {"matmul", "mm", "bmm", "addmm", "baddbmm", "einsum", "_int_mm", "_scaled_mm",
             "scaled_dot_product_attention", "conv1d", "conv2d", "conv3d", "conv_transpose2d",
             "conv_transpose3d", "linear"}
REDUCE = {"sum", "mean", "amax", "amin", "max", "min", "cumsum", "cumprod", "softmax", "logsumexp",
          "norm", "var", "std", "prod", "argmax", "argsort", "sort", "topk", "rms_norm",
          "layer_norm", "log_softmax", "scatter_add_", "index_add_", "bincount"}
MASKY = {"masked_fill", "masked_fill_", "where", "tril", "triu", "gather", "scatter", "scatter_",
         "index_select", "take_along_dim", "repeat_interleave", "nonzero", "bucketize"}
QUANT = re.compile(r"float8_e[45]m[23]|uint8|int8|int32\b|\.to\(torch\.int|>>|<<|&\s*0x|nibble|"
                   r"e8m0|e2m1|absmax|dequant|qmap", re.I)


def analyse_reference(task):
    """What KIND of kernel is this, structurally? Driven off the AST, not text: the earlier regex
    missed `a @ b.t()` and `(q*s) @ k[b].transpose(1,2)` -- i.e. most of the matmuls in the lane."""
    p = LANE / task / "environment" / "reference.py"
    if not p.exists():
        return None
    src = p.read_text()
    try:
        tree = ast.parse(src)
    except SyntaxError:
        return None
    n_mm, calls = 0, set()
    for n in ast.walk(tree):
        if isinstance(n, ast.BinOp) and isinstance(n.op, ast.MatMult):
            n_mm += 1                                   # the `@` operator IS a cuBLAS call
        elif isinstance(n, ast.Call):
            f = n.func
            nm = f.attr if isinstance(f, ast.Attribute) else (f.id if isinstance(f, ast.Name) else "")
            if nm:
                calls.add(nm)
                if nm in MATMULish:
                    n_mm += 1
    # A vendor kernel only counts as "already at roofline" if it applies to the WHOLE task.
    # These structures mean it does not:
    grouped = bool(re.search(r"\boffsets?\b|\bcounts?\b|group_?(?:size|idx|id)|varm|"
                             r"expert_?(?:idx|id|offsets)|cu_seqlens", src, re.I))
    custom_conv = bool(re.search(r"F\.pad|padding\s*=\s*\(|causal|groups\s*=|feather|blend|"
                                 r"tile|overlap", src, re.I))
    epilogue = bool(re.search(r"silu|gelu|swiglu|geglu|sigmoid|tanh\(|\* *gate|gate *\*", src, re.I))
    return dict(grouped=grouped, custom_conv=custom_conv, epilogue=epilogue, n_mm=n_mm,
                sdpa="scaled_dot_product_attention" in calls,
                conv=any(c.startswith("conv") for c in calls),
                reduce=bool(calls & REDUCE),
                masky=bool(calls & MASKY),
                quant=bool(QUANT.search(src)))


def roofline_us(task):
    v = LANE / task / "tests" / "verify_env.py"
    if not v.exists():
        return None, None
    src = v.read_text()
    if "canonical_work" not in src:
        return None, None
    ns, shapes = {}, None
    try:
        for node in ast.parse(src).body:
            if isinstance(node, (ast.FunctionDef, ast.Assign)):
                try:
                    exec(compile(ast.Module([node], []), "<c>", "exec"), ns)
                except Exception:
                    pass
            if isinstance(node, ast.Assign) and getattr(node.targets[0], "id", "") == "GRADER_SHAPES":
                shapes = ast.literal_eval(node.value)
        metric = "GB/s" if "GB/s" in src else "TFLOP/s"
        big = max(ns["canonical_work"](*s) for s in shapes)
        if metric == "TFLOP/s":
            return metric, big / (H200_TFLOPS * 1e12) * 1e6
        bw = LINK_GBPS if task.startswith("dist-") else H200_GBPS
        return metric, big / (bw * 2 ** 30) * 1e6
    except Exception:
        return None, None


def peak_for(metric, task):
    if metric == "TFLOP/s":
        return H200_TFLOPS
    return (LINK_GBPS if task.startswith("dist-") else H200_GBPS)


# T4 is the highest-stakes label -- it asserts a vendor kernel is already at the hardware limit, i.e.
# that the task is near-unbeatable. The structural classifier gets the other tiers right but is too
# blunt here, so this small set is reviewed by hand and the reason recorded. Auditable by construction.
OVERRIDE = {
    "dist-allgather-gemm-overlap": ("T3", "2-GPU compute/communication overlap: the GEMM is a library "
                                          "call but the overlap schedule is the task"),
    "moe-grouped-gemm-contiguous": ("T3", "grouped GEMM over a contiguous expert layout: no single "
                                          "library call covers it"),
    "moe-grouped-gemm-varm": ("T3", "variable-M grouped GEMM: cuBLAS has no such call"),
    "moe-grouped-swiglu": ("T3", "grouped GEMM with a fused SwiGLU epilogue"),
    "deepseek-mla-vabsorb-outproj": ("T3", "MLA V-absorb folds the value projection into the output "
                                           "projection: a fused GEMM pair, not a plain one"),
    "hunyuan-dualstream-attn-proj": ("T3", "two per-stream output projections plus per-sample gates: "
                                          "a 2-group GEMM with M=1e5 against M=1e2"),
    "spatial-upsample-pixelshuffle3d": ("T2", "pixel-shuffle upsample is a layout transform, not a "
                                              "convolution cuDNN accelerates"),
}


def tier(a, frac):
    """Technique depth required to beat the best AVAILABLE implementation.

    This is a structural judgement, deliberately not an empirical one. Measuring the shipped reference
    cannot answer it: the references are intentionally slow fp32/fp64 specs, so `int8-w8a8-gemm` shows
    12.9x of apparent headroom purely because its reference multiplies in float64. And
    torch.compile(max-autotune) is not a usable proxy either -- it fails to trace most of these
    references (exec'd source, closures, data-dependent control flow) and barely helps where it does.
    """
    if a is None:
        return "T2", "unanalysable reference; defaulted"
    compute = a["n_mm"] > 0 or a["sdpa"] or a["conv"]

    if not compute:
        if a["reduce"] or a["masky"]:
            return "T2", "bandwidth kernel with a reduction/gather: shared-memory tiling and a warp reduction"
        return "T1", "elementwise/bandwidth chain: the win is fusing the passes into one"

    # --- compute-bound: does a library kernel apply to the WHOLE task, or only to a piece? ---
    if a["quant"]:
        return "T3", ("quantised matmul: cuBLAS will not fuse the dequant, so the MMA pipeline is "
                      "hand-written -- async copy, double buffering, fragment layouts")
    if a["grouped"]:
        return "T3", ("grouped / variable-M GEMM: there is no single library call for it, so the "
                      "per-group tiling and scheduling are the task")
    if a["sdpa"] and not a["masky"]:
        return "T4", "plain attention: an FA-class kernel applies directly and already sits near roofline"
    if a["masky"] or a["sdpa"]:
        return "T3", ("attention with custom masking/sparsity: no library kernel applies as-is, so the "
                      "tiled online-softmax pipeline is written by hand")
    if a["conv"]:
        if a["custom_conv"]:
            return "T3", ("convolution with custom padding/grouping/tiling: cuDNN's path for this case "
                          "is not the fast one, so the tiled kernel is the task")
        return "T4", "plain convolution: cuDNN applies directly and already runs near roofline"
    if a["epilogue"]:
        return "T3", "GEMM with a fused activation/gate epilogue: needs a hand-written MMA pipeline"
    if a["n_mm"] <= 2 and not a["reduce"]:
        return "T4", "dense GEMM: cuBLAS/CUTLASS are already at the hardware limit"
    return "T3", "multi-GEMM block: needs a hand-written, pipelined MMA sequence to beat"


def load_measured():
    """Reference achieved metrics from the validation sweeps."""
    out = {}
    for f in pathlib.Path("/tmp").glob("*out_*.txt"):
        try:
            for line in f.read_text().splitlines():
                m = re.match(r"^(\S+)\s+stub\[.*?\]\s+ref\[([0-9.]+)\s", line)
                if m:
                    out[m.group(1)] = float(m.group(2))
        except Exception:
            pass
    return out


def main():
    meas = load_measured()
    rows = []
    for d in sorted(p for p in LANE.iterdir() if p.is_dir() and not p.name.startswith("_")):
        t = d.name
        if not (d / "task.toml").exists():
            continue
        metric, rus = roofline_us(t)
        got = meas.get(t)
        a = analyse_reference(t)
        frac = None
        if metric and got:
            frac = min(got / peak_for(metric, t), 1.0)
        tr, why = OVERRIDE.get(t) or tier(a, frac)
        rows.append(dict(name=t, tier=tr, why=why, frac=round(frac, 4) if frac else None,
                         analysis=a, metric=metric, roofline_us=rus, ref_metric=got))
    (LANE / "_factory" / "difficulty.json").write_text(json.dumps(rows, indent=2) + "\n")
    from collections import Counter
    c = Counter(r["tier"] for r in rows)
    print(f"scored {len(rows)} tasks   measured={sum(1 for r in rows if r['frac'] is not None)}")
    for k in ("T1", "T2", "T3", "T4"):
        print(f"  {k}: {c[k]}")
    return rows


if __name__ == "__main__":
    rows = main()
    for anchor in sys.argv[1:]:
        for r in rows:
            if r["name"] == anchor:
                print(f"\n{r['name']}: {r['tier']}  ({r['why']})")