Preliminary state

2026-05-26 10:24:33 +02:00
commit 0ab11cfc5e
950 changed files with 6428 additions and 0 deletions
@@ -0,0 +1,36 @@
+{
+  "median": 0.6731782129991188,
+  "times": [
+    0.6604644770013692,
+    0.6292407609998918,
+    0.6794493079996755,
+    0.7151197740004136,
+    0.6822677139989537,
+    0.6751668389988481,
+    0.6932685800002218,
+    0.6652886089996173,
+    0.6990065699992556,
+    0.6860697520005488,
+    0.6711895869993896,
+    0.6505095659995277,
+    0.6577597320010682,
+    0.6831210750005994,
+    0.6632924360001198,
+    0.6577419900004315,
+    0.6499202390004939,
+    0.6408863569995447,
+    0.6844092320006894,
+    0.6783877489997394,
+    0.6611538839988498,
+    0.6698421969995252,
+    0.6564882740003668,
+    0.7027955690009549,
+    0.6597058009992907,
+    0.7558749809995788,
+    0.6768750350001937,
+    0.6361531940001441,
+    0.71163439799966,
+    0.7025398610003322
+  ],
+  "n_repeats": 30
+}
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:348732c04488ee8dded43d8bd6ca7df2bfe24423c8fa2e7e1f5ce23522f3c9dd
+size 2562
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cafba9f7ae517a80b59cf118c96a301ed33562c1ab6c3404530fed7da3236dc8
+size 1179648
@@ -0,0 +1,17 @@
+{
+    "command": "bench_eval",
+    "network": "resources/dilated_conv_stack_network.mininn",
+    "inputs": [
+        "resources/input_0.bin"
+    ],
+    "n_repeats": 30,
+    "n_warmup": 3,
+    "max_slowdown": 0.95,
+    "description": "speed: stack of three 2-D dilated convolutions (dilation 1/2/4); leaky_relu/ELU/GELU; batch=4, 12x24\n\nScoring: no regular points; the test is marked passed only when the implementation is at least 5 % faster than the reference (max_slowdown=0.95). Bonus: tier 1 = 3 pts for slowdown <= 0.95 (>=5 % speedup), tier 2 = 7 pts for slowdown <= 0.75 (>=25 % speedup); bonuses are cumulative.",
+    "bonus_points": 10,
+    "scoring": {
+        "function": "speed",
+        "tier1_bonus": 3,
+        "tier2_bonus": 7
+    }
+}
@@ -0,0 +1,36 @@
+{
+  "median": 0.6597053400009827,
+  "times": [
+    0.6588008830003673,
+    0.660609797001598,
+    0.7053496530006669,
+    0.661915245000273,
+    0.6636695469987899,
+    0.6757613959998707,
+    0.6935657760004688,
+    0.7083251549993292,
+    0.6651058429997647,
+    0.700552788001005,
+    0.6511960340012592,
+    0.653393098999004,
+    0.6478349190001609,
+    0.6328443050006172,
+    0.6522468030016171,
+    0.7022167730010551,
+    0.7124317390007491,
+    0.6838970240005438,
+    0.6676450960003422,
+    0.6217490859999089,
+    0.651708482999311,
+    0.654718187999606,
+    0.6895673190010712,
+    0.6417726840008982,
+    0.6327202760003274,
+    0.6619786449991807,
+    0.6345275990006485,
+    0.6287327130012272,
+    0.6376377029992,
+    0.648800413000572
+  ],
+  "n_repeats": 30
+}
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f0d3b6424560f4dc8acbdd458b68da146ed7b204ec4f4139ef589b7c8b631f5
+size 655360
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e02ccb8aeae9ebd68989b14a35520ecad23bb8f381ba30ea18528ea46f1a597a
+size 2031
@@ -0,0 +1,17 @@
+{
+    "command": "bench_eval",
+    "network": "resources/residual_dense_mix_network.mininn",
+    "inputs": [
+        "resources/input_0.bin"
+    ],
+    "n_repeats": 30,
+    "n_warmup": 3,
+    "max_slowdown": 0.95,
+    "description": "speed: mixed 2-D residual+dense architecture with leaky_relu/ELU/GELU; batch=4, 8x20\n\nScoring: no regular points; the test is marked passed only when the implementation is at least 5 % faster than the reference (max_slowdown=0.95). Bonus: tier 1 = 3 pts for slowdown <= 0.95 (>=5 % speedup), tier 2 = 7 pts for slowdown <= 0.75 (>=25 % speedup); bonuses are cumulative.",
+    "bonus_points": 10,
+    "scoring": {
+        "function": "speed",
+        "tier1_bonus": 3,
+        "tier2_bonus": 7
+    }
+}
@@ -0,0 +1,36 @@
+{
+  "median": 0.6708022339998934,
+  "times": [
+    0.669449609998992,
+    0.645968004999304,
+    0.6869760080007836,
+    0.6653184660008264,
+    0.6711289119994035,
+    0.6666175419995852,
+    0.6564001719998487,
+    0.6763312660004885,
+    0.6714280679989315,
+    0.6847440599995025,
+    0.6704755560003832,
+    0.6575861059991439,
+    0.6796879159992386,
+    0.6594827870012523,
+    0.7487746679998963,
+    0.6494830270003149,
+    0.6802418199986278,
+    0.7007190820004325,
+    0.6683371039998747,
+    0.681551567000497,
+    0.6629623690005246,
+    0.6854776170002879,
+    0.7160350009999092,
+    0.6702585629991518,
+    0.667205087000184,
+    0.6629004050009826,
+    0.6747479939986079,
+    0.692501079000067,
+    0.6739553940005862,
+    0.6577685189986369
+  ],
+  "n_repeats": 30
+}
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b4c043533b3c953c96725cb40b3e05ee59946fdf3e35a45172da58f048cf637
+size 49152
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15d3f6d073e57ccb5e2810be8e9725a3c65ebe00dd832e13d7e2589b449b80c2
+size 13558
@@ -0,0 +1,17 @@
+{
+    "command": "bench_eval",
+    "network": "resources/transformer_two_blocks_network.mininn",
+    "inputs": [
+        "resources/input_0.bin"
+    ],
+    "n_repeats": 30,
+    "n_warmup": 3,
+    "max_slowdown": 0.95,
+    "description": "speed: two stacked pre-norm single-head transformer blocks; seq=6, d_model=8, d_head=4, d_ff=16\n\nScoring: no regular points; the test is marked passed only when the implementation is at least 5 % faster than the reference (max_slowdown=0.95). Bonus: tier 1 = 3 pts for slowdown <= 0.95 (>=5 % speedup), tier 2 = 7 pts for slowdown <= 0.75 (>=25 % speedup); bonuses are cumulative.",
+    "bonus_points": 10,
+    "scoring": {
+        "function": "speed",
+        "tier1_bonus": 3,
+        "tier2_bonus": 7
+    }
+}
@@ -0,0 +1,36 @@
+{
+  "median": 0.6913382860002457,
+  "times": [
+    0.7181488250007533,
+    0.6882181509990914,
+    0.6892177329991682,
+    0.7106628969995654,
+    0.7126837310006522,
+    0.691078888001357,
+    0.6915976839991345,
+    0.7045463059985195,
+    0.6723891719993844,
+    0.7171977689995401,
+    0.6693990150015452,
+    0.6795068999999785,
+    0.6964747769998212,
+    0.692497801999707,
+    0.7017947509993974,
+    0.6824454929992498,
+    0.6821788209999795,
+    0.7133113889995002,
+    0.7371664179991058,
+    0.673557535999862,
+    0.6838336380005785,
+    0.6884392770007253,
+    0.6956874999996217,
+    0.6593486749989097,
+    0.6807764959994529,
+    0.7030594869993365,
+    0.6746473509992938,
+    0.6918940119994659,
+    0.6577691439997579,
+    0.7121864170003391
+  ],
+  "n_repeats": 30
+}
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:152e3e4273bf3cfa924be9dee559b1c0e4cdda904411e5c1fd3d13a8b4f34d96
+size 2359296
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdbdfc127c2e75eb4f5ad6482320ea4427e0b3b48eb5e4b856b4b7d23cca7d49
+size 12058
@@ -0,0 +1,17 @@
+{
+    "command": "bench_eval",
+    "network": "resources/unet_1d_three_level_network.mininn",
+    "inputs": [
+        "resources/input_0.bin"
+    ],
+    "n_repeats": 30,
+    "n_warmup": 3,
+    "max_slowdown": 0.95,
+    "description": "speed: 3-level 2D UNet (encoder/bottleneck/decoder with skip connections); batch=4, 24x24\n\nScoring: no regular points; the test is marked passed only when the implementation is at least 5 % faster than the reference (max_slowdown=0.95). Bonus: tier 1 = 3 pts for slowdown <= 0.95 (>=5 % speedup), tier 2 = 7 pts for slowdown <= 0.75 (>=25 % speedup); bonuses are cumulative.",
+    "bonus_points": 10,
+    "scoring": {
+        "function": "speed",
+        "tier1_bonus": 3,
+        "tier2_bonus": 7
+    }
+}