Measure the shipped default path in the VAE conv benchmark.

The *_def columns passed DEFAULT_TILE explicitly, so they bypassed _pick_tile
and no longer described what a caller without a tile argument actually gets.
Drop the tile argument, record the tile that _pick_tile returned in a new
tile_def column, and let the best-of-candidates start from that pick so the
figure's outer bar stays an upper bound. The plotter derives its footnote
ratios from the data instead of carrying the previous run's constants.

Refreshed fair_baseline.json and the figure on the current tree. Weighted over
the VAE call mix, the default path now sits at 1.068x of hipBLASLt on the 1024
path (10.35 ms of GEMM, was 11.51) and 1.284x on the two 1328 layers (7.12 ms,
was 7.99).

One regression the refresh exposes, unrelated to the tile picker: the
256x256x4x4 candidate has become 2.3-2.9x slower than 128x128x2x4 (e.g.
192->192 @512x512, 290 -> 835 us for the whole path) though it still compiles
and matches torch. It used to win 7 of the 18 shapes, so the best-of-candidates
upper bound drops on the 384->192 family (105.9 -> 132.1 us of GEMM on
384->192 @256x256).

Author: amd-nprotaso

tests/perf/qwenimage_vae_conv/bench_fair.py

@@ -7,6 +7,10 @@
 Each of the 18 shapes runs in a fresh process so JIT / autotune state cannot
 leak. Writes ``fair_baseline.json`` next to this script.
 
+The ``*_def`` columns are the shipped default path (no ``tile`` argument, so
+``_pick_tile`` decides); ``*_best`` is the fastest of that pick and the tile
+candidates below.
+
 Usage::
 
     python tests/perf/qwenimage_vae_conv/bench_fair.py
@@ -60,6 +64,7 @@
     ("fly_def", float),
     ("gemm_def", float),
     ("tr_def", float),
+    ("tile_def", str),
     ("fly_best", float),
     ("gemm_best", float),
     ("tr_best", float),
@@ -126,7 +131,7 @@ def run_one(sid, cin, cout, hin, stride, padding, freq, path) -> None:
     import torch.nn.functional as F
     from torch.profiler import ProfilerActivity, profile
 
-    from kernels.conv.conv3d_implicit import DEFAULT_TILE, conv3d_implicit
+    from kernels.conv.conv3d_implicit import DEFAULT_TILE, _pick_tile, conv3d_implicit
 
     cin, cout, hin, stride, padding, freq = map(int, (cin, cout, hin, stride, padding, freq))
     candidates = [
@@ -158,11 +163,13 @@ def fly(tile):
     def time(call):
         return gpu(torch, profile, ProfilerActivity, call)
 
-    check(fly(tuple(DEFAULT_TILE)))
-    t_def, per_def = time(lambda: fly(tuple(DEFAULT_TILE)))
+    # "default" is the shipped path: no tile argument, so _pick_tile decides.
+    check(fly(None))
+    t_def, per_def = time(lambda: fly(None))
     g_def, tr_def = split_fly(per_def)
-    best = (g_def, t_def, tr_def, tuple(DEFAULT_TILE))
-    for tile in candidates[1:]:
+    tile_def = _pick_tile(M, N, 1, x4.device)
+    best = (g_def, t_def, tr_def, tile_def)
+    for tile in candidates:
         try:
             check(fly(tile))
             t, per = time(lambda: fly(tile))
@@ -207,7 +214,7 @@ def unfold_mm(wt=w2):
     t_mio, _ = time(lambda: F.conv2d(x4, w4, bias=bbf, stride=stride, padding=padding))
     print(
         f"RESULT\t{sid}\t{cin}\t{cout}\t{hin}\t{stride}\t{freq}\t{path}\t{M}\t{N}\t{K}\t{flops:.6e}\t"
-        f"{t_def:.2f}\t{g_def:.2f}\t{tr_def:.2f}\t"
+        f"{t_def:.2f}\t{g_def:.2f}\t{tr_def:.2f}\t{'x'.join(str(v) for v in tile_def)}\t"
         f"{t_best:.2f}\t{g_best:.2f}\t{tr_best:.2f}\t{'x'.join(str(v) for v in tile_best)}\t"
         f"{t_mm:.2f}\t{t_im2col:.2f}\t{t_unfold_mm:.2f}\t{t_mio:.2f}"
     )
@@ -219,7 +226,8 @@ def drive() -> None:
     rows = []
     print(
         f"{'shape':18s} {'x':>3s} {'M':>8s} {'K':>5s} | {'hipBLASLt':>9s} {'FlyGEMM':>8s} {'比':>5s} "
-        f"{'BLASt T/s':>9s} {'Fly T/s':>8s} | {'im2col':>7s} {'unfold+mm':>10s} {'Fly 全':>7s} {'MIOpen':>8s} {'最优tile':>13s}"
+        f"{'BLASt T/s':>9s} {'Fly T/s':>8s} | {'im2col':>7s} {'unfold+mm':>10s} {'Fly 全':>7s} {'MIOpen':>8s} "
+        f"{'默认tile':>13s} {'最优tile':>13s}"
     )
     for sp in SHAPES:
         proc = subprocess.run(
@@ -240,7 +248,7 @@ def drive() -> None:
             f"{shp:18s} {r['freq']:3d} {r['M']:8d} {r['K']:5d} | {r['mm']:9.1f} {r['gemm_best']:8.1f} "
             f"{r['mm'] / r['gemm_best']:4.2f}x {r['flops'] / r['mm'] / 1e6:9.0f} "
             f"{r['flops'] / r['gemm_best'] / 1e6:8.0f} | {r['im2col']:7.1f} {r['unfold_mm']:10.1f} "
-            f"{r['fly_best']:7.1f} {r['mio']:8.1f} {r['tile']:>13s}",
+            f"{r['fly_best']:7.1f} {r['mio']:8.1f} {r['tile_def']:>13s} {r['tile']:>13s}",
             flush=True,
         )