Emily/add-cost-model-make-rule

Makefile

@@ -2,7 +2,7 @@ JOBS ?= all
 
 THIS := $(dir $(realpath $(lastword $(MAKEFILE_LIST))))
 
-.PHONY: default fast all clean artifact
+.PHONY: default fast all clean artifact low_level_representation pipeline cost_model
 
 default: fast
 
@@ -20,7 +20,10 @@ low_level_representation: maybe_update_xdsl_commit
 pipeline: maybe_update_xdsl_commit
 	snakemake --cores $(JOBS) --rerun-incomplete pipeline
 
+cost_model: maybe_update_xdsl_commit
+	snakemake --cores $(JOBS) --rerun-incomplete cost_model
+
 clean:
-	snakemake --delete-all-output --rerun-incomplete fast all pipeline low_level_representation
+	snakemake --delete-all-output --rerun-incomplete fast all pipeline low_level_representation cost_model
 
 include ./Makefile.xdsl

Snakefile

@@ -139,6 +139,16 @@ TESTSET_FAST = [
     *expand("sum/8x8xf16/{variant}", variant=["baseline", "linalg_xdsl"]),
 ]
 
+TESTSET_COST_MODEL = [
+    *expand(
+        "matmul_transb/1x{K}x{N}xf64/linalg_xdsl",
+        # reduction dimension = K
+        K = list(range(30,40)),
+        # row dimension = N
+        N = list(range(1,10)),
+    ),
+]
+
 TESTSET_LOW_LEVEL_REPRESENTATION = [
     *expand(
         "{kernel}/40x{N}xf32/{variant}",
@@ -243,6 +253,7 @@ def select_test_set_profiles(wildcards) -> list[str]:
         "all": sorted(set(TESTSET_ALL)),
         "low_level_representation": sorted(set(TESTSET_LOW_LEVEL_REPRESENTATION)),
         "pipeline": sorted(set(TESTSET_PIPELINE)),
+        "cost_model": sorted(set(TESTSET_COST_MODEL)),
     }
     name = wildcards.testset
     if name not in sets:
@@ -257,6 +268,7 @@ def select_test_set_regalloc_jsons(wildcards) -> list[str]:
         "all": sorted(set(TESTSET_ALL)),
         "low_level_representation": sorted(set(TESTSET_LOW_LEVEL_REPRESENTATION)),
         "pipeline": sorted(set(TESTSET_PIPELINE)),
+        "cost_model": sorted(set(TESTSET_COST_MODEL)),
     }
     name = wildcards.testset
     if name not in sets:
@@ -314,6 +326,13 @@ rule all:
         "results/pivoted_fpu.all.csv",
         "results/pivoted_ipc.all.csv",
 
+rule cost_model:
+    input:
+        "results/kernels.cost_model.csv",
+        "results/pivoted.cost_model.csv",
+        "results/pivoted_fpu.cost_model.csv",
+        "results/pivoted_ipc.cost_model.csv",
+        "results/regalloc.cost_model.csv",
 
 ###########################################################
 # Rules

kernels/matmul_transb/main.c

@@ -30,6 +30,9 @@ int main() {
     int thiscore = snrt_cluster_core_idx();
     if (thiscore != 0) return 0;
 
+    // warm up instruction
+    matmul_transb(local_x, local_y, local_z);
+
     snrt_fpu_fence();
     (void)snrt_mcycle();
     matmul_transb(local_x, local_y, local_z);

results/kernels.csv

@@ -28,12 +28,12 @@ matmul,1x20x5xf64,linalg_5_xdsl,191,967,964,2.81981981981982,0.0,100,111,313,0.5
 matmul,4x16x8xf64,baseline,2495,3293,3290,2.9941520467836256,1.4991334488734835,480,513,1536,0.20561122244488977,0.4568121104185218,1123,865,577,0.4501002004008016,0,33,1.0,1.0,1,0.0,1123,0.7975852272727273,285,0,0,0.11422845691382766,0,799,0.0,0.5643286573146293,0.0
 matmul,4x16x8xf64,linalg,2694,3483,3480,2.9941520467836256,1.4745484400656814,512,513,1536,0.19042316258351893,0.44415584415584414,1155,898,609,0.4287305122494432,0,33,1.0,1.0,1,3.4,1155,0.73520050922979,416,17,5,0.1544172234595397,69,790,0.0,0.5831477357089829,0.0
 matmul,4x16x8xf64,linalg_xdsl,708,1493,1490,2.811418685121107,0.0,512,578,1625,0.8163841807909604,0.996551724137931,580,0,0,0.8192090395480226,0,0,5.37037037037037,5.37037037037037,1,0.0,108,0.5869565217391305,76,0,0,0.10734463276836158,0,786,0.0,0.9265536723163842,0.0
-matmul_transb,4x16x16xf32,baseline,3386,4184,4181,2.539660056657224,1.4921875,0,706,1793,0.20850561134081513,0.3935340022296544,1794,1528,1024,0.5298287064382753,0,64,1.0,1.0,1,0.0,1794,0.5561066336019839,1432,0,0,0.42291789722386297,0,799,0.0,0.9527466036621383,0.0
-matmul_transb,4x16x16xf32,linalg,5038,5831,5828,2.9970731707317073,1.4995663486556807,0,1025,3072,0.203453751488686,0.4569772625947392,2243,1729,1153,0.44521635569670504,0,65,1.0,1.0,1,3.4,2243,0.8036546040845575,548,17,5,0.10877332274712187,69,794,0.0,0.553989678443827,0.0
-matmul_transb,4x16x16xf32,snitch_stream,845,1636,1633,2.7429906542056073,0.0,0,642,1761,0.7597633136094675,0.9067796610169492,708,0,0,0.8378698224852071,0,64,2.0823529411764703,2.0823529411764707,1,0.0,340,0.7296137339055794,126,0,0,0.14911242603550295,0,792,0.0,0.98698224852071,0.0
-matmul_transb,4x16x16xf32,snrt,849,1612,1609,2.648367952522255,0.0,0,674,1785,0.7938751472320377,0.9519774011299436,708,0,0,0.833922261484099,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.8924731182795699,40,0,0,0.04711425206124853,0,764,0.0,0.8810365135453475,0.0
-matmul_transb,4x16x16xf64,linalg,5142,5967,5964,2.9970731707317073,1.4995663486556807,1024,1025,3072,0.19933877868533645,0.4569772625947392,2243,1729,1153,0.43621159082069233,0,65,1.0,1.0,1,3.4,2243,0.7769310703152061,644,17,5,0.12524309607156747,133,826,0.0,0.5614546868922599,0.0
-matmul_transb,4x16x16xf64,linalg_xdsl,1295,2124,2121,2.815424610051993,0.0,1024,1154,3249,0.8911196911196911,0.9982698961937716,1156,0,0,0.8926640926640926,0,0,5.452830188679246,5.452830188679245,1,0.0,212,0.6794871794871795,100,0,0,0.07722007722007722,0,830,0.0,0.9698841698841698,0.0
+matmul_transb,4x16x16xf32,baseline,3246,7408,7405,2.539660056657224,1.5,0,706,1793,0.2174984596426371,0.3935340022296544,1794,1536,1024,0.5526802218114603,0,64,1.0,1.0,1,0.0,1794,0.5559343043074063,1433,0,0,0.4414664202094886,0,4163,0.0,0.9941466420209488,0.0
+matmul_transb,4x16x16xf32,linalg,4935,10759,10756,2.9970731707317073,1.4995663486556807,0,1025,3072,0.2077001013171226,0.4569772625947392,2243,1729,1153,0.45450861195542047,0,65,1.0,1.0,1,3.4,2243,0.8033667621776505,549,17,5,0.11124620060790273,69,5825,0.0,0.5657548125633232,0.0
+matmul_transb,4x16x16xf32,snitch_stream,786,2416,2413,2.7429906542056073,0.0,0,642,1761,0.816793893129771,0.9067796610169492,708,0,0,0.9007633587786259,0,64,2.0823529411764703,2.0823529411764707,1,0.0,340,0.728051391862955,127,0,0,0.16157760814249364,0,1631,0.0,1.0623409669211195,0.0
+matmul_transb,4x16x16xf32,snrt,772,2409,2406,2.648367952522255,0.0,0,674,1785,0.8730569948186528,0.9519774011299436,708,0,0,0.917098445595855,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.8900804289544236,41,0,0,0.05310880829015544,0,1638,0.0,0.9702072538860104,0.0
+matmul_transb,4x16x16xf64,linalg,5025,10973,10970,2.9970731707317073,1.4995663486556807,1024,1025,3072,0.20398009950248755,0.4569772625947392,2243,1729,1153,0.4463681592039801,0,65,1.0,1.0,1,3.4,2243,0.7766620498614959,645,17,5,0.12835820895522387,133,5949,0.0,0.574726368159204,0.0
+matmul_transb,4x16x16xf64,linalg_xdsl,1246,3342,3339,2.815424610051993,0.0,1024,1154,3249,0.9261637239165329,0.9982698961937716,1156,0,0,0.92776886035313,0,0,5.452830188679246,5.452830188679245,1,0.0,212,0.6773162939297125,101,0,0,0.08105939004815409,0,2097,0.0,1.008828250401284,0.0
 pooling_nchw_max_d1_s2_3x3,4x4xf64,baseline,584,1328,1325,0.995575221238938,1.1226415094339623,0,226,225,0.386986301369863,0.6330532212885154,357,119,106,0.6113013698630136,0,25,1.0,1.0,1,0.0,357,0.9153846153846154,33,0,0,0.05650684931506849,0,745,0.0,0.6678082191780821,0.0
 pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg,484,1242,1239,0.993103448275862,1.0909090909090908,0,145,144,0.29958677685950413,0.5823293172690763,249,96,88,0.5144628099173554,0,16,1.0,1.0,1,0.0,249,0.8498293515358362,44,0,0,0.09090909090909091,32,759,0.0,0.6053719008264463,0.0
 pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg_xdsl,275,1018,1015,0.9943820224719101,0.0,0,178,177,0.6472727272727272,0.9888888888888889,180,0,0,0.6545454545454545,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5283018867924528,50,0,0,0.18181818181818182,0,744,0.0,0.8363636363636364,0.0

results/kernels.fast.csv

@@ -28,12 +28,12 @@ matmul,1x20x5xf64,linalg_5_xdsl,191,967,964,2.81981981981982,0.0,100,111,313,0.5
 matmul,4x16x8xf64,baseline,2495,3293,3290,2.9941520467836256,1.4991334488734835,480,513,1536,0.20561122244488977,0.4568121104185218,1123,865,577,0.4501002004008016,0,33,1.0,1.0,1,0.0,1123,0.7975852272727273,285,0,0,0.11422845691382766,0,799,0.0,0.5643286573146293,0.0
 matmul,4x16x8xf64,linalg,2694,3483,3480,2.9941520467836256,1.4745484400656814,512,513,1536,0.19042316258351893,0.44415584415584414,1155,898,609,0.4287305122494432,0,33,1.0,1.0,1,3.4,1155,0.73520050922979,416,17,5,0.1544172234595397,69,790,0.0,0.5831477357089829,0.0
 matmul,4x16x8xf64,linalg_xdsl,708,1493,1490,2.811418685121107,0.0,512,578,1625,0.8163841807909604,0.996551724137931,580,0,0,0.8192090395480226,0,0,5.37037037037037,5.37037037037037,1,0.0,108,0.5869565217391305,76,0,0,0.10734463276836158,0,786,0.0,0.9265536723163842,0.0
-matmul_transb,4x16x16xf32,baseline,3386,4184,4181,2.539660056657224,1.4921875,0,706,1793,0.20850561134081513,0.3935340022296544,1794,1528,1024,0.5298287064382753,0,64,1.0,1.0,1,0.0,1794,0.5561066336019839,1432,0,0,0.42291789722386297,0,799,0.0,0.9527466036621383,0.0
-matmul_transb,4x16x16xf32,linalg,5038,5831,5828,2.9970731707317073,1.4995663486556807,0,1025,3072,0.203453751488686,0.4569772625947392,2243,1729,1153,0.44521635569670504,0,65,1.0,1.0,1,3.4,2243,0.8036546040845575,548,17,5,0.10877332274712187,69,794,0.0,0.553989678443827,0.0
-matmul_transb,4x16x16xf32,snitch_stream,845,1636,1633,2.7429906542056073,0.0,0,642,1761,0.7597633136094675,0.9067796610169492,708,0,0,0.8378698224852071,0,64,2.0823529411764703,2.0823529411764707,1,0.0,340,0.7296137339055794,126,0,0,0.14911242603550295,0,792,0.0,0.98698224852071,0.0
-matmul_transb,4x16x16xf32,snrt,849,1612,1609,2.648367952522255,0.0,0,674,1785,0.7938751472320377,0.9519774011299436,708,0,0,0.833922261484099,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.8924731182795699,40,0,0,0.04711425206124853,0,764,0.0,0.8810365135453475,0.0
-matmul_transb,4x16x16xf64,linalg,5142,5967,5964,2.9970731707317073,1.4995663486556807,1024,1025,3072,0.19933877868533645,0.4569772625947392,2243,1729,1153,0.43621159082069233,0,65,1.0,1.0,1,3.4,2243,0.7769310703152061,644,17,5,0.12524309607156747,133,826,0.0,0.5614546868922599,0.0
-matmul_transb,4x16x16xf64,linalg_xdsl,1295,2124,2121,2.815424610051993,0.0,1024,1154,3249,0.8911196911196911,0.9982698961937716,1156,0,0,0.8926640926640926,0,0,5.452830188679246,5.452830188679245,1,0.0,212,0.6794871794871795,100,0,0,0.07722007722007722,0,830,0.0,0.9698841698841698,0.0
+matmul_transb,4x16x16xf32,baseline,3246,7408,7405,2.539660056657224,1.5,0,706,1793,0.2174984596426371,0.3935340022296544,1794,1536,1024,0.5526802218114603,0,64,1.0,1.0,1,0.0,1794,0.5559343043074063,1433,0,0,0.4414664202094886,0,4163,0.0,0.9941466420209488,0.0
+matmul_transb,4x16x16xf32,linalg,4935,10759,10756,2.9970731707317073,1.4995663486556807,0,1025,3072,0.2077001013171226,0.4569772625947392,2243,1729,1153,0.45450861195542047,0,65,1.0,1.0,1,3.4,2243,0.8033667621776505,549,17,5,0.11124620060790273,69,5825,0.0,0.5657548125633232,0.0
+matmul_transb,4x16x16xf32,snitch_stream,786,2416,2413,2.7429906542056073,0.0,0,642,1761,0.816793893129771,0.9067796610169492,708,0,0,0.9007633587786259,0,64,2.0823529411764703,2.0823529411764707,1,0.0,340,0.728051391862955,127,0,0,0.16157760814249364,0,1631,0.0,1.0623409669211195,0.0
+matmul_transb,4x16x16xf32,snrt,772,2409,2406,2.648367952522255,0.0,0,674,1785,0.8730569948186528,0.9519774011299436,708,0,0,0.917098445595855,0,32,2.1325301204819276,2.1325301204819276,1,0.0,332,0.8900804289544236,41,0,0,0.05310880829015544,0,1638,0.0,0.9702072538860104,0.0
+matmul_transb,4x16x16xf64,linalg,5025,10973,10970,2.9970731707317073,1.4995663486556807,1024,1025,3072,0.20398009950248755,0.4569772625947392,2243,1729,1153,0.4463681592039801,0,65,1.0,1.0,1,3.4,2243,0.7766620498614959,645,17,5,0.12835820895522387,133,5949,0.0,0.574726368159204,0.0
+matmul_transb,4x16x16xf64,linalg_xdsl,1246,3342,3339,2.815424610051993,0.0,1024,1154,3249,0.9261637239165329,0.9982698961937716,1156,0,0,0.92776886035313,0,0,5.452830188679246,5.452830188679245,1,0.0,212,0.6773162939297125,101,0,0,0.08105939004815409,0,2097,0.0,1.008828250401284,0.0
 pooling_nchw_max_d1_s2_3x3,4x4xf64,baseline,584,1328,1325,0.995575221238938,1.1226415094339623,0,226,225,0.386986301369863,0.6330532212885154,357,119,106,0.6113013698630136,0,25,1.0,1.0,1,0.0,357,0.9153846153846154,33,0,0,0.05650684931506849,0,745,0.0,0.6678082191780821,0.0
 pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg,484,1242,1239,0.993103448275862,1.0909090909090908,0,145,144,0.29958677685950413,0.5823293172690763,249,96,88,0.5144628099173554,0,16,1.0,1.0,1,0.0,249,0.8498293515358362,44,0,0,0.09090909090909091,32,759,0.0,0.6053719008264463,0.0
 pooling_nchw_max_d1_s2_3x3,4x4xf64,linalg_xdsl,275,1018,1015,0.9943820224719101,0.0,0,178,177,0.6472727272727272,0.9888888888888889,180,0,0,0.6545454545454545,0,0,3.214285714285714,3.2142857142857144,1,0.0,56,0.5283018867924528,50,0,0,0.18181818181818182,0,744,0.0,0.8363636363636364,0.0

results/kernels.low_level_representation.csv

@@ -1,13 +1,13 @@
 test,params,impl,cycles,end,end_fpss,fpss_avg_fpu_latency,fpss_avg_load_latency,fpss_fpu_fmadd_issues,fpss_fpu_issues,fpss_fpu_latency,fpss_fpu_occupancy,fpss_fpu_rel_occupancy,fpss_issues,fpss_load_latency,fpss_loads,fpss_occupancy,fpss_section_latency,fpss_stores,fseq_fpu_yield,fseq_yield,section,snitch_avg_load_latency,snitch_fseq_offloads,snitch_fseq_rel_offloads,snitch_issues,snitch_load_latency,snitch_loads,snitch_occupancy,snitch_stores,start,tend,total_ipc,tstart
-matmul_transb,1x12x40xf32,snitch_stream,484,1267,1264,2.6739130434782608,0,0,322,861,0.6652892561983471,0.8846153846153846,364,0,0,0.7520661157024794,0,40,1.7009345794392525,1.7009345794392523,1,0,214,0.7086092715231788,88,0,0,0.18181818181818182,0,784,0.0,0.9338842975206612,0.0
-matmul_transb,1x16x40xf32,snitch_stream,573,1386,1383,2.7388059701492535,0,0,402,1101,0.7015706806282722,0.9054054054054054,444,0,0,0.774869109947644,0,40,2.074766355140187,2.074766355140187,1,0,214,0.7062706270627063,89,0,0,0.15532286212914484,0,814,0.0,0.9301919720767888,0.0
-matmul_transb,1x20x16xf32,snitch_stream,326,1132,1129,2.768041237113402,0,0,194,537,0.5950920245398773,0.9150943396226415,212,0,0,0.6503067484662577,0,16,2.409090909090909,2.409090909090909,1,0,88,0.5751633986928104,65,0,0,0.19938650306748465,0,807,0.0,0.8496932515337423,0.0
-matmul_transb,1x20x24xf32,snitch_stream,431,1220,1217,2.7758620689655173,0,0,290,805,0.6728538283062645,0.9177215189873418,316,0,0,0.7331786542923434,0,24,2.430769230769231,2.4307692307692306,1,0,130,0.6403940886699507,73,0,0,0.16937354988399073,0,790,0.0,0.9025522041763341,0.0
-matmul_transb,1x20x32xf32,snitch_stream,547,1360,1357,2.7797927461139897,0,0,386,1073,0.7056672760511883,0.919047619047619,420,0,0,0.7678244972577697,0,32,2.441860465116279,2.441860465116279,1,0,172,0.6798418972332015,81,0,0,0.1480804387568556,0,814,0.0,0.9159049360146253,0.0
-matmul_transb,1x20x40xf32,snitch_stream,653,1476,1473,2.7821576763485476,0,0,482,1341,0.7381316998468607,0.9198473282442748,524,0,0,0.8024502297090352,0,40,2.448598130841121,2.4485981308411215,1,0,214,0.7062706270627063,89,0,0,0.1362940275650842,0,824,0.0,0.9387442572741195,0.0
-matmul_transb,1x20x8xf32,snitch_stream,215,983,980,2.7448979591836733,0,0,98,269,0.4558139534883721,0.9074074074074074,108,0,0,0.5023255813953489,0,8,2.347826086956522,2.347826086956522,1,0,46,0.44660194174757284,57,0,0,0.2651162790697674,0,769,0.0,0.7674418604651163,0.0
-matmul_transb,1x4x40xf32,snitch_stream,380,1133,1130,2.3518518518518516,0,0,162,381,0.4263157894736842,0.7941176470588235,204,0,0,0.5368421052631579,0,40,0.9532710280373832,0.9532710280373832,1,0,214,0.7328767123287672,78,0,0,0.20526315789473684,0,754,0.0,0.7421052631578947,0.0
-matmul_transb,1x8x40xf32,snitch_stream,412,1180,1177,2.5661157024793386,0,0,242,621,0.587378640776699,0.852112676056338,284,0,0,0.6893203883495146,0,40,1.327102803738318,1.3271028037383177,1,0,214,0.7086092715231788,88,0,0,0.21359223300970873,0,769,0.0,0.9029126213592233,0.0
+matmul_transb,1x12x40xf32,snitch_stream,436,1730,1727,2.6739130434782608,0,0,322,861,0.7385321100917431,0.8846153846153846,364,0,0,0.8348623853211009,0,40,1.7009345794392525,1.7009345794392523,1,0,214,0.7062706270627063,89,0,0,0.20412844036697247,0,1295,0.0,1.0389908256880733,0.0
+matmul_transb,1x16x40xf32,snitch_stream,515,1895,1892,2.7388059701492535,0,0,402,1101,0.7805825242718447,0.9054054054054054,444,0,0,0.8621359223300971,0,40,2.074766355140187,2.074766355140187,1,0,214,0.7039473684210527,90,0,0,0.17475728155339806,0,1381,0.0,1.036893203883495,0.0
+matmul_transb,1x20x16xf32,snitch_stream,278,1411,1408,2.768041237113402,0,0,194,537,0.697841726618705,0.9150943396226415,212,0,0,0.762589928057554,0,16,2.409090909090909,2.409090909090909,1,0,88,0.5751633986928104,65,0,0,0.23381294964028776,0,1134,0.0,0.9964028776978417,0.0
+matmul_transb,1x20x24xf32,snitch_stream,386,1621,1618,2.7758620689655173,0,0,290,805,0.7512953367875648,0.9177215189873418,316,0,0,0.8186528497409327,0,24,2.430769230769231,2.4307692307692306,1,0,130,0.6403940886699507,73,0,0,0.18911917098445596,0,1236,0.0,1.0077720207253886,0.0
+matmul_transb,1x20x32xf32,snitch_stream,489,1843,1840,2.7797927461139897,0,0,386,1073,0.7893660531697342,0.919047619047619,420,0,0,0.8588957055214724,0,32,2.441860465116279,2.441860465116279,1,0,172,0.6771653543307087,82,0,0,0.16768916155419222,0,1355,0.0,1.0265848670756648,0.0
+matmul_transb,1x20x40xf32,snitch_stream,595,2065,2062,2.7821576763485476,0,0,482,1341,0.8100840336134454,0.9198473282442748,524,0,0,0.880672268907563,0,40,2.448598130841121,2.4485981308411215,1,0,214,0.7039473684210527,90,0,0,0.15126050420168066,0,1471,0.0,1.0319327731092436,0.0
+matmul_transb,1x20x8xf32,snitch_stream,172,1186,1183,2.7448979591836733,0,0,98,269,0.5697674418604651,0.9074074074074074,108,0,0,0.627906976744186,0,8,2.347826086956522,2.347826086956522,1,0,46,0.44660194174757284,57,0,0,0.3313953488372093,0,1015,0.0,0.9593023255813953,0.0
+matmul_transb,1x4x40xf32,snitch_stream,311,1469,1466,2.3518518518518516,0,0,162,381,0.5209003215434084,0.7941176470588235,204,0,0,0.6559485530546624,0,40,0.9532710280373832,0.9532710280373832,1,0,214,0.7303754266211604,79,0,0,0.2540192926045016,0,1159,0.0,0.909967845659164,0.0
+matmul_transb,1x8x40xf32,snitch_stream,356,1563,1560,2.5661157024793386,0,0,242,621,0.6797752808988764,0.852112676056338,284,0,0,0.797752808988764,0,40,1.327102803738318,1.3271028037383177,1,0,214,0.7062706270627063,89,0,0,0.25,0,1208,0.0,1.047752808988764,0.0
 relu,16x40xf32,snitch_stream,371,1151,1148,0.9969040247678018,0,0,323,322,0.8706199460916442,0.9938461538461538,325,0,0,0.876010781671159,0,0,46.42857142857143,46.42857142857143,1,0,7,0.28,18,0,0,0.04851752021563342,0,781,0.0,0.9245283018867924,0.0
 relu,24x40xf32,snitch_stream,531,1331,1328,0.9979296066252588,0,0,483,482,0.9096045197740112,0.9958762886597938,485,0,0,0.9133709981167608,0,0,69.28571428571428,69.28571428571429,1,0,7,0.28,18,0,0,0.03389830508474576,0,801,0.0,0.9472693032015066,0.0
 relu,32x40xf32,snitch_stream,688,1502,1499,0.9984447900466563,0,0,643,642,0.934593023255814,0.9968992248062015,645,0,0,0.9375,0,0,92.14285714285715,92.14285714285714,1,0,7,0.2916666666666667,17,0,0,0.024709302325581394,0,815,0.0,0.9622093023255814,0.0