diffcb 0.1.13__tar.gz → 0.1.14__tar.gz

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round42_fp64_diag_dcb.csv

@@ -0,0 +1,17 @@
+phase,n,ms,note
+precompute_fft_f32,50000,0.5517,histogram+rfft float32
+precompute_fft_f64,50000,0.5468,histogram+rfft float64 (C_mc)
+mode_count_from_C_single,50000,0.2722,irfft+sign_detect float64 single call
+mode_count_from_C_batch2,50000,0.3697,irfft+sign_detect float64 batch B=2 (one trisection step)
+trisection_16_rounds,50000,7.4257,16x mode_count_from_C_batch float64
+_refine_hcrit,50000,1.7154,_refine_hcrit float64
+full_forward_f64,50000,74.5462,DCBLayer.forward float64 (reference)
+full_forward_f32,50000,237.1963,DCBLayer.forward float32
+measured_subtotal,50000,10.2396,pre_f32+pre_f64+trisection+refine
+forward_batched_total,16,50000,868.6899,total forward_batched fp64 K=16 n=50K
+serial_total,16,50000,3913.1577,K sequential forward calls fp64
+speedup,16,50000,4.5047,serial/batch speedup ratio
+precompute_fft_batch_f32,16,50000,23.4957,K histograms + batched rfft float32
+precompute_fft_batch_f64,16,50000,31.4592,K histograms + batched rfft float64
+mode_count_K_batch_1round,16,50000,2.8570,batched irfft+sign K=16 B=2 (one round)
+trisection_16R_estimate,16,50000,45.7114,estimated 16 rounds trisection K=16

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round42_fp64_diag_ops.csv

@@ -0,0 +1,13 @@
+op,size,fp32_ms,fp64_ms,ratio
+rfft,16384,0.012288,0.149504,12.1667
+irfft,16384,0.017408,0.155648,8.9412
+exp,16384,0.010240,0.026400,2.5781
+mul,16384,0.008192,0.009216,1.1250
+sum,16384,0.008192,0.012288,1.5000
+histogram,16384,0.019456,0.028960,1.4885
+rfft,32768,0.020480,0.061440,3.0000
+irfft,32768,0.026624,0.065536,2.4615
+exp,32768,0.034944,0.017408,0.4982
+mul,32768,0.008192,0.027648,3.3750
+sum,32768,0.012192,0.020480,1.6798
+histogram,32768,0.019520,0.029696,1.5213

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round42_fp64_diag_summary.txt

@@ -0,0 +1,79 @@
+======================================================================
+Round 42 -- RTX 4060 Laptop fp64 Diagnosis Summary
+======================================================================
+PyTorch        : 2.5.1+cu121
+GPU            : NVIDIA GeForce RTX 4060 Laptop GPU
+Compute cap    : sm_89
+VRAM           : 8.6 GB
+
+-- Test 1: Isolated fp64 vs fp32 penalty for core operations --------
+          op      size     fp32_ms     fp64_ms     ratio
+-------------------------------------------------------
+        rfft    16,384    0.012288    0.149504   12.1667x
+       irfft    16,384    0.017408    0.155648    8.9412x
+         exp    16,384    0.010240    0.026400    2.5781x
+         mul    16,384    0.008192    0.009216    1.1250x
+         sum    16,384    0.008192    0.012288    1.5000x
+   histogram    16,384    0.019456    0.028960    1.4885x
+        rfft    32,768    0.020480    0.061440    3.0000x
+       irfft    32,768    0.026624    0.065536    2.4615x
+         exp    32,768    0.034944    0.017408    0.4982x
+         mul    32,768    0.008192    0.027648    3.3750x
+         sum    32,768    0.012192    0.020480    1.6798x
+   histogram    32,768    0.019520    0.029696    1.5213x
+
+  Worst fp64 penalty: rfft at size=16,384  ratio=12.1667x
+  RTX 4060 fp64 TFLOPS ~ 0.25 (1/32 of fp32 8 TFLOPS = compute bottleneck)
+
+-- Test 2: DCB forward phase breakdown (n=50K bimodal, fp64) --------
+  Full forward fp64 : 74.546 ms
+  Full forward fp32 : 237.196 ms
+  fp64/fp32 ratio   : 0.31x
+  (Round 38 baseline was 134ms -- this run should be similar)
+
+  precompute_fft fp64 : 0.547 ms  (0.7% of forward)
+  trisection 16R fp64 : 7.426 ms  (10.0% of forward)  = 0.464 ms/round
+  _refine_hcrit fp64  : 1.715 ms  (2.3% of forward)
+
+-- Test 3: forward_batched breakdown (K=16, n=50K, fp64) ------------
+  serial   K=16: 3913.158 ms  (244.572 ms/dist)
+  batched  K=16: 868.690 ms  (54.293 ms/dist)
+  speedup      : 4.50x  (P100 baseline: 10.7x at K=16)
+
+  forward_batched_total                   : 868.6899 ms
+  serial_total                            : 3913.1577 ms
+  speedup                                 : 4.5047 ms
+  precompute_fft_batch_f32                : 23.4957 ms
+  precompute_fft_batch_f64                : 31.4592 ms
+  mode_count_K_batch_1round               : 2.8570 ms
+  trisection_16R_estimate                 : 45.7114 ms
+
+-- Test 4: Memory bandwidth saturation ------------------------------
+  fp32 copy 100MB: 0.9216 ms  -> 227.6 GB/s  (83.7% of 272 GB/s spec)
+  fp64 copy 100MB: 0.9185 ms  -> 228.3 GB/s  (83.9% of 272 GB/s spec)
+  fp64/fp32 copy ratio: 1.00x
+  -> Memory-bound: fp64 copy is fast -> bottleneck is COMPUTE, not BW
+    (RTX 4060 fp64 compute = 0.25 TFLOPS = 1/32 of fp32)
+
+-- Test 5: fp32 drop-in accuracy ------------------------------------
+  fp64 batch K=16: 868.690 ms  (reference)
+  fp32 batch K=16: 272.758 ms  speedup=3.18x
+  fp32 accuracy: max |err%|=0.0001%  mean=0.0000%
+  fp32 on fp64 data accuracy: max |err%|=0.0000%  mean=0.0000%
+  fp32 PASS (<0.1%): YES
+
+-- Practical Recommendation -----------------------------------------
+RTX 4060 Laptop fp64 TFLOPS = 0.25 TFLOPS (1/32 of 8 TFLOPS fp32).
+The irfft inner loop in mode_count_from_C / mode_count_K_batch
+is executed in fp64 (complex128) for numerical correctness.
+On RTX 4060 this gives ~30x compute penalty per irfft vs fp32.
+
+fp32 forward_batched is accurate to <0.1% -- ACCEPTABLE for training.
+RECOMMENDATION: Use fp32 internally for CUDA training (fft_dtype=float32)
+and only verify final mode count in fp64 (C_mc separate compute).
+This is already the design of DCBLayer (fft_dtype=float32 default)
+but forward_batched up-casts C_batch to complex128 for mode counting.
+A CUDA-fp32 mode counting path (with 1e-6 relative threshold) could
+recover the 30x penalty while staying within <0.1% accuracy.
+
+======================================================================

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round43_c1_benchmark.csv

@@ -0,0 +1,13 @@
+n,K,device,t_serial_ms,t_batch_ms,speedup
+50000,1,cuda,238.132,116.65,2.041
+50000,4,cuda,979.639,318.312,3.078
+50000,8,cuda,2035.184,550.501,3.697
+50000,16,cuda,4050.137,1037.984,3.902
+50000,32,cuda,7676.015,2268.065,3.384
+50000,64,cuda,15506.934,4283.204,3.62
+100000,1,cuda,198.963,115.238,1.727
+100000,4,cuda,761.1,305.93,2.488
+100000,8,cuda,1477.925,553.355,2.671
+100000,16,cuda,2940.617,1048.924,2.803
+100000,32,cuda,5840.391,2175.663,2.684
+100000,64,cuda,11846.297,4287.322,2.763

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round43_c1_benchmark.log

@@ -0,0 +1,37 @@
+Device: cuda
+GPU: NVIDIA GeForce RTX 4060 Laptop GPU
+
+n=  50000 K=  1 | serial=  238.13ms  batch(C1)=  116.65ms  speedup= 2.04x
+n=  50000 K=  4 | serial=  979.64ms  batch(C1)=  318.31ms  speedup= 3.08x
+n=  50000 K=  8 | serial= 2035.18ms  batch(C1)=  550.50ms  speedup= 3.70x
+n=  50000 K= 16 | serial= 4050.14ms  batch(C1)= 1037.98ms  speedup= 3.90x
+n=  50000 K= 32 | serial= 7676.02ms  batch(C1)= 2268.06ms  speedup= 3.38x
+n=  50000 K= 64 | serial=15506.93ms  batch(C1)= 4283.20ms  speedup= 3.62x
+n= 100000 K=  1 | serial=  198.96ms  batch(C1)=  115.24ms  speedup= 1.73x
+n= 100000 K=  4 | serial=  761.10ms  batch(C1)=  305.93ms  speedup= 2.49x
+n= 100000 K=  8 | serial= 1477.93ms  batch(C1)=  553.35ms  speedup= 2.67x
+n= 100000 K= 16 | serial= 2940.62ms  batch(C1)= 1048.92ms  speedup= 2.80x
+n= 100000 K= 32 | serial= 5840.39ms  batch(C1)= 2175.66ms  speedup= 2.68x
+n= 100000 K= 64 | serial=11846.30ms  batch(C1)= 4287.32ms  speedup= 2.76x
+
+CSV: results/round43_c1_benchmark.csv
+Round 43 C1 Benchmark �� forward vs forward_batched (vectorised _refine_hcrit)
+Device: cuda (NVIDIA GeForce RTX 4060 Laptop GPU)
+Warmup=2, best-of-5
+
+       n    K    serial_ms     batch_ms    speedup
+----------------------------------------------------
+   50000    1       238.13       116.65       2.04x
+   50000    4       979.64       318.31       3.08x
+   50000    8      2035.18       550.50       3.70x
+   50000   16      4050.14      1037.98       3.90x
+   50000   32      7676.02      2268.07       3.38x
+   50000   64     15506.93      4283.20       3.62x
+  100000    1       198.96       115.24       1.73x
+  100000    4       761.10       305.93       2.49x
+  100000    8      1477.92       553.36       2.67x
+  100000   16      2940.62      1048.92       2.80x
+  100000   32      5840.39      2175.66       2.68x
+  100000   64     11846.30      4287.32       2.76x
+
+Summary: results/round43_c1_summary.txt

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round43_c1_summary.txt

@@ -0,0 +1,18 @@
+Round 43 C1 Benchmark �� forward vs forward_batched (vectorised _refine_hcrit)
+Device: cuda (NVIDIA GeForce RTX 4060 Laptop GPU)
+Warmup=2, best-of-5
+
+       n    K    serial_ms     batch_ms    speedup
+----------------------------------------------------
+   50000    1       238.13       116.65       2.04x
+   50000    4       979.64       318.31       3.08x
+   50000    8      2035.18       550.50       3.70x
+   50000   16      4050.14      1037.98       3.90x
+   50000   32      7676.02      2268.07       3.38x
+   50000   64     15506.93      4283.20       3.62x
+  100000    1       198.96       115.24       1.73x
+  100000    4       761.10       305.93       2.49x
+  100000    8      1477.92       553.36       2.67x
+  100000   16      2940.62      1048.92       2.80x
+  100000   32      5840.39      2175.66       2.68x
+  100000   64     11846.30      4287.32       2.76x

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round43_correctness.log

@@ -0,0 +1,124 @@
+    dist        n    K  seed |     h_single      h_batch   rel_err%   PASS
+--------------------------------------------------------------------------
+ bimodal     1000    1     0 |     1.756220     1.756220     0.0000   PASS
+ bimodal     1000    4     0 |     1.756220     1.756220     0.0000   PASS
+ bimodal     1000    8     0 |     1.756220     1.756220     0.0000   PASS
+ bimodal     1000   16     0 |     1.756220     1.756220     0.0000   PASS
+ bimodal     1000   32     0 |     1.756220     1.756220     0.0000   PASS
+ bimodal     1000    1     1 |     1.716072     1.716072     0.0000   PASS
+ bimodal     1000    4     1 |     1.716072     1.716072     0.0000   PASS
+ bimodal     1000    8     1 |     1.716072     1.716072     0.0000   PASS
+ bimodal     1000   16     1 |     1.716072     1.716072     0.0000   PASS
+ bimodal     1000   32     1 |     1.716072     1.716072     0.0000   PASS
+ bimodal     1000    1     2 |     1.827619     1.827619     0.0000   PASS
+ bimodal     1000    4     2 |     1.827619     1.827619     0.0000   PASS
+ bimodal     1000    8     2 |     1.827619     1.827619     0.0000   PASS
+ bimodal     1000   16     2 |     1.827619     1.827619     0.0000   PASS
+ bimodal     1000   32     2 |     1.827619     1.827619     0.0000   PASS
+ bimodal    10000    1     0 |     1.834665     1.834665     0.0000   PASS
+ bimodal    10000    4     0 |     1.834665     1.834665     0.0000   PASS
+ bimodal    10000    8     0 |     1.834665     1.834665     0.0000   PASS
+ bimodal    10000   16     0 |     1.834665     1.834665     0.0000   PASS
+ bimodal    10000   32     0 |     1.834665     1.834665     0.0000   PASS
+ bimodal    10000    1     1 |     1.835204     1.835204     0.0000   PASS
+ bimodal    10000    4     1 |     1.835204     1.835204     0.0000   PASS
+ bimodal    10000    8     1 |     1.835204     1.835204     0.0000   PASS
+ bimodal    10000   16     1 |     1.835204     1.835204     0.0000   PASS
+ bimodal    10000   32     1 |     1.835204     1.835204     0.0000   PASS
+ bimodal    10000    1     2 |     1.827390     1.827390     0.0000   PASS
+ bimodal    10000    4     2 |     1.827390     1.827390     0.0000   PASS
+ bimodal    10000    8     2 |     1.827390     1.827390     0.0000   PASS
+ bimodal    10000   16     2 |     1.827390     1.827390     0.0000   PASS
+ bimodal    10000   32     2 |     1.827390     1.827390     0.0000   PASS
+ bimodal    50000    1     0 |     1.808084     1.808088     0.0002   PASS
+ bimodal    50000    4     0 |     1.808084     1.808088     0.0002   PASS
+ bimodal    50000    8     0 |     1.808084     1.808088     0.0002   PASS
+ bimodal    50000   16     0 |     1.808084     1.808088     0.0002   PASS
+ bimodal    50000   32     0 |     1.808084     1.808088     0.0002   PASS
+ bimodal    50000    1     1 |     1.828867     1.828869     0.0001   PASS
+ bimodal    50000    4     1 |     1.828867     1.828869     0.0001   PASS
+ bimodal    50000    8     1 |     1.828867     1.828869     0.0001   PASS
+ bimodal    50000   16     1 |     1.828867     1.828869     0.0001   PASS
+ bimodal    50000   32     1 |     1.828867     1.828869     0.0001   PASS
+ bimodal    50000    1     2 |     1.812422     1.812425     0.0002   PASS
+ bimodal    50000    4     2 |     1.812422     1.812425     0.0002   PASS
+ bimodal    50000    8     2 |     1.812422     1.812425     0.0002   PASS
+ bimodal    50000   16     2 |     1.812422     1.812425     0.0002   PASS
+ bimodal    50000   32     2 |     1.812422     1.812425     0.0002   PASS
+ bimodal   100000    1     0 |     1.825266     1.825258     0.0004   PASS
+ bimodal   100000    4     0 |     1.825266     1.825258     0.0004   PASS
+ bimodal   100000    8     0 |     1.825266     1.825258     0.0004   PASS
+ bimodal   100000   16     0 |     1.825266     1.825258     0.0004   PASS
+ bimodal   100000   32     0 |     1.825266     1.825258     0.0004   PASS
+ bimodal   100000    1     1 |     1.833408     1.833405     0.0001   PASS
+ bimodal   100000    4     1 |     1.833408     1.833405     0.0001   PASS
+ bimodal   100000    8     1 |     1.833408     1.833405     0.0001   PASS
+ bimodal   100000   16     1 |     1.833408     1.833405     0.0001   PASS
+ bimodal   100000   32     1 |     1.833408     1.833405     0.0001   PASS
+ bimodal   100000    1     2 |     1.820311     1.820317     0.0003   PASS
+ bimodal   100000    4     2 |     1.820311     1.820317     0.0003   PASS
+ bimodal   100000    8     2 |     1.820311     1.820317     0.0003   PASS
+ bimodal   100000   16     2 |     1.820311     1.820317     0.0003   PASS
+ bimodal   100000   32     2 |     1.820311     1.820317     0.0003   PASS
+gaussian     1000    1     0 |     0.314169     0.314169     0.0000   PASS
+gaussian     1000    4     0 |     0.314169     0.314169     0.0000   PASS
+gaussian     1000    8     0 |     0.314169     0.314169     0.0000   PASS
+gaussian     1000   16     0 |     0.314169     0.314169     0.0000   PASS
+gaussian     1000   32     0 |     0.314169     0.314169     0.0000   PASS
+gaussian     1000    1     1 |     0.278207     0.278207     0.0000   PASS
+gaussian     1000    4     1 |     0.278207     0.278207     0.0000   PASS
+gaussian     1000    8     1 |     0.278207     0.278207     0.0000   PASS
+gaussian     1000   16     1 |     0.278207     0.278207     0.0000   PASS
+gaussian     1000   32     1 |     0.278207     0.278207     0.0000   PASS
+gaussian     1000    1     2 |     0.244192     0.244192     0.0000   PASS
+gaussian     1000    4     2 |     0.244192     0.244192     0.0000   PASS
+gaussian     1000    8     2 |     0.244192     0.244192     0.0000   PASS
+gaussian     1000   16     2 |     0.244192     0.244192     0.0000   PASS
+gaussian     1000   32     2 |     0.244192     0.244192     0.0000   PASS
+gaussian    10000    1     0 |     0.169495     0.169495     0.0000   PASS
+gaussian    10000    4     0 |     0.169495     0.169495     0.0000   PASS
+gaussian    10000    8     0 |     0.169495     0.169495     0.0000   PASS
+gaussian    10000   16     0 |     0.169495     0.169495     0.0000   PASS
+gaussian    10000   32     0 |     0.169495     0.169495     0.0000   PASS
+gaussian    10000    1     1 |     0.123838     0.123838     0.0000   PASS
+gaussian    10000    4     1 |     0.123838     0.123838     0.0000   PASS
+gaussian    10000    8     1 |     0.123838     0.123838     0.0000   PASS
+gaussian    10000   16     1 |     0.123838     0.123838     0.0000   PASS
+gaussian    10000   32     1 |     0.123838     0.123838     0.0000   PASS
+gaussian    10000    1     2 |     0.142834     0.142834     0.0000   PASS
+gaussian    10000    4     2 |     0.142834     0.142834     0.0000   PASS
+gaussian    10000    8     2 |     0.142834     0.142834     0.0000   PASS
+gaussian    10000   16     2 |     0.142834     0.142834     0.0000   PASS
+gaussian    10000   32     2 |     0.142834     0.142834     0.0000   PASS
+gaussian    50000    1     0 |     0.246755     0.246772     0.0068   PASS
+gaussian    50000    4     0 |     0.246755     0.246772     0.0068   PASS
+gaussian    50000    8     0 |     0.246755     0.246772     0.0068   PASS
+gaussian    50000   16     0 |     0.246755     0.246772     0.0068   PASS
+gaussian    50000   32     0 |     0.246755     0.246772     0.0068   PASS
+gaussian    50000    1     1 |     0.149850     0.149896     0.0306   PASS
+gaussian    50000    4     1 |     0.149850     0.149896     0.0306   PASS
+gaussian    50000    8     1 |     0.149850     0.149896     0.0306   PASS
+gaussian    50000   16     1 |     0.149850     0.149896     0.0306   PASS
+gaussian    50000   32     1 |     0.149850     0.149896     0.0306   PASS
+gaussian    50000    1     2 |     0.121115     0.121048     0.0556   PASS
+gaussian    50000    4     2 |     0.121115     0.121048     0.0556   PASS
+gaussian    50000    8     2 |     0.121115     0.121048     0.0556   PASS
+gaussian    50000   16     2 |     0.121115     0.121048     0.0556   PASS
+gaussian    50000   32     2 |     0.121115     0.121048     0.0556   PASS
+gaussian   100000    1     0 |     0.221107     0.221059     0.0216   PASS
+gaussian   100000    4     0 |     0.221107     0.221059     0.0216   PASS
+gaussian   100000    8     0 |     0.221107     0.221059     0.0216   PASS
+gaussian   100000   16     0 |     0.221107     0.221059     0.0216   PASS
+gaussian   100000   32     0 |     0.221107     0.221059     0.0216   PASS
+gaussian   100000    1     1 |     0.110631     0.110631     0.0000   PASS
+gaussian   100000    4     1 |     0.110631     0.110631     0.0000   PASS
+gaussian   100000    8     1 |     0.110631     0.110631     0.0000   PASS
+gaussian   100000   16     1 |     0.110631     0.110631     0.0000   PASS
+gaussian   100000   32     1 |     0.110631     0.110631     0.0000   PASS
+gaussian   100000    1     2 |     0.284689     0.284646     0.0150   PASS
+gaussian   100000    4     2 |     0.284689     0.284646     0.0150   PASS
+gaussian   100000    8     2 |     0.284689     0.284646     0.0150   PASS
+gaussian   100000   16     2 |     0.284689     0.284646     0.0150   PASS
+gaussian   100000   32     2 |     0.284689     0.284646     0.0150   PASS
+
+ALL CELLS PASSED (120 cells, tol=0.1%)

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round44_b5_benchmark.csv

@@ -0,0 +1,11 @@
+n,K,device,batched_min_ms,batched_mean_ms,sequential_min_ms,sequential_mean_ms,speedup_min,speedup_mean
+1000,1,cpu,85.72,96.11,81.23,88.52,0.948,0.921
+1000,2,cpu,164.69,178.75,186.52,200.78,1.133,1.123
+1000,4,cpu,337.49,486.25,359.4,396.35,1.065,0.815
+1000,8,cpu,683.1,718.5,758.21,770.42,1.11,1.072
+1000,16,cpu,1452.8,1548.19,1561.97,1597.4,1.075,1.032
+50000,1,cpu,968.08,988.68,1723.52,1945.68,1.78,1.968
+50000,2,cpu,1939.14,1965.48,3261.56,4119.5,1.682,2.096
+50000,4,cpu,3493.4,3742.31,6772.61,7883.17,1.939,2.107
+50000,8,cpu,7558.26,7868.59,13422.82,13693.85,1.776,1.74
+50000,16,cpu,14008.71,15191.58,29186.0,29992.16,2.083,1.974

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round44_b5_benchmark_remote.csv

@@ -0,0 +1,11 @@
+n,K,device,batched_min_ms,batched_mean_ms,sequential_min_ms,sequential_mean_ms,speedup_min,speedup_mean
+1000,1,cuda,16.51,200.03,284.35,341.52,17.221,1.707
+1000,2,cuda,405.14,461.78,623.17,768.84,1.538,1.665
+1000,4,cuda,389.6,655.61,393.17,408.28,1.009,0.623
+1000,8,cuda,726.44,788.72,781.72,868.58,1.076,1.101
+1000,16,cuda,1499.95,1533.1,1534.83,1562.35,1.023,1.019
+50000,1,cuda,134.6,141.92,241.63,253.91,1.795,1.789
+50000,2,cuda,207.84,210.4,456.98,491.81,2.199,2.338
+50000,4,cuda,363.56,364.94,902.96,966.07,2.484,2.647
+50000,8,cuda,669.32,677.79,1911.62,2006.5,2.856,2.96
+50000,16,cuda,1303.99,1313.56,3915.4,3993.36,3.003,3.04

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round44_b5_summary.txt

@@ -0,0 +1,26 @@
+HPC-B5 Training Throughput Benchmark
+Device: cpu   PyTorch: 2.12.0
+Warmup runs: 3   Timed runs: 5   Metric: best-of-5
+
+n=1000
+    K   batched(ms)   sequential(ms)   speedup
+  K= 1          85.7             81.2      0.95x
+  K= 2         164.7            186.5      1.13x
+  K= 4         337.5            359.4      1.06x
+  K= 8         683.1            758.2      1.11x
+  K=16        1452.8           1562.0      1.07x
+
+n=50000
+    K   batched(ms)   sequential(ms)   speedup
+  K= 1         968.1           1723.5      1.78x
+  K= 2        1939.1           3261.6      1.68x
+  K= 4        3493.4           6772.6      1.94x
+  K= 8        7558.3          13422.8      1.78x
+  K=16       14008.7          29186.0      2.08x
+
+Key result (n=50K, best-of speedup over K sequential backward calls):
+  K= 1: 1.78x
+  K= 2: 1.68x
+  K= 4: 1.94x
+  K= 8: 1.78x
+  K=16: 2.08x

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round44_b5_summary_remote.txt

@@ -0,0 +1,26 @@
+HPC-B5 Training Throughput Benchmark
+Device: cuda   PyTorch: 2.5.1+cu121
+Warmup runs: 3   Timed runs: 5   Metric: best-of-5
+
+n=1000
+    K   batched(ms)   sequential(ms)   speedup
+  K= 1          16.5            284.4     17.22x
+  K= 2         405.1            623.2      1.54x
+  K= 4         389.6            393.2      1.01x
+  K= 8         726.4            781.7      1.08x
+  K=16        1500.0           1534.8      1.02x
+
+n=50000
+    K   batched(ms)   sequential(ms)   speedup
+  K= 1         134.6            241.6      1.79x
+  K= 2         207.8            457.0      2.20x
+  K= 4         363.6            903.0      2.48x
+  K= 8         669.3           1911.6      2.86x
+  K=16        1304.0           3915.4      3.00x
+
+Key result (n=50K, best-of speedup over K sequential backward calls):
+  K= 1: 1.79x
+  K= 2: 2.20x
+  K= 4: 2.48x
+  K= 8: 2.86x
+  K=16: 3.00x

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round44_benchmark.log

@@ -0,0 +1,101 @@
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=0.00e+00 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=4.47e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.12e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.24e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.13e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.26e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.10e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.23e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.20e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=4.79e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=3.40e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.25e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.21e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=4.55e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.15e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.18e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.22e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.32e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.16e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=3.25e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.08e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=3.45e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.17e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=3.52e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=4.74e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.28e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.27e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.31e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.07e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.17e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=3.44e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.30e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=4.70e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.29e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=3.23e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.19e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=1.14e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=3.54e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+D:\Anaconda3\lib\site-packages\torch\autograd\function.py:307: UserWarning: DCB IFT denominator |\u2202M\u0303/\u2202h|=2.16e-04 < 0.01. Gradient may be large. Use safe_backward=True to clamp.
+  return user_fn(self, *args)
+[B5-bench] Device: cuda
+[B5-bench] PyTorch: 2.5.1+cu121
+
+n=  1000  use_fft=False
+    K   batched_min     seq_min   speedup
+  K= 1  batched=    16.5ms  seq=   284.3ms  speedup=17.22x
+  K= 2  batched=   405.1ms  seq=   623.2ms  speedup=1.54x
+  K= 4  batched=   389.6ms  seq=   393.2ms  speedup=1.01x
+  K= 8  batched=   726.4ms  seq=   781.7ms  speedup=1.08x
+  K=16  batched=  1500.0ms  seq=  1534.8ms  speedup=1.02x
+
+n= 50000  use_fft=True
+    K   batched_min     seq_min   speedup
+  K= 1  batched=   134.6ms  seq=   241.6ms  speedup=1.80x
+  K= 2  batched=   207.8ms  seq=   457.0ms  speedup=2.20x
+  K= 4  batched=   363.6ms  seq=   903.0ms  speedup=2.48x
+  K= 8  batched=   669.3ms  seq=  1911.6ms  speedup=2.86x
+  K=16  batched=  1304.0ms  seq=  3915.4ms  speedup=3.00x
+
+CSV written: C:\Users\lenovo\dcb_exp\results\round44_b5_benchmark.csv
+Summary written: C:\Users\lenovo\dcb_exp\results\round44_b5_summary.txt
+
+[B5-bench] Done.

diffcb-0.1.14/02_projects/01_dcb_proposal/04_analysis/results/round44_gradcheck.log

@@ -0,0 +1,92 @@
+[B5] Device: cuda
+[B5] PyTorch: 2.5.1+cu121
+
+============================================================
+PART 1: Forward correctness (forward_batched_grad vs forward_batched)
+============================================================
+  [PASS] n=  500 bimodal  K=1 seed=0 max_rel=0.000000%
+  [PASS] n=  500 bimodal  K=1 seed=1 max_rel=0.000000%
+  [PASS] n=  500 bimodal  K=1 seed=2 max_rel=0.000000%
+  [PASS] n=  500 bimodal  K=2 seed=0 max_rel=0.000000%
+  [PASS] n=  500 bimodal  K=2 seed=1 max_rel=0.000007%
+  [PASS] n=  500 bimodal  K=2 seed=2 max_rel=0.000007%
+  [PASS] n=  500 bimodal  K=4 seed=0 max_rel=0.000000%
+  [PASS] n=  500 bimodal  K=4 seed=1 max_rel=0.000022%
+  [PASS] n=  500 bimodal  K=4 seed=2 max_rel=0.000007%
+  [PASS] n=  500 bimodal  K=8 seed=0 max_rel=0.000042%
+  [PASS] n=  500 bimodal  K=8 seed=1 max_rel=0.000027%
+  [PASS] n=  500 bimodal  K=8 seed=2 max_rel=0.000007%
+  [PASS] n=  500 gaussian K=1 seed=0 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=1 seed=1 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=1 seed=2 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=2 seed=0 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=2 seed=1 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=2 seed=2 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=4 seed=0 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=4 seed=1 max_rel=0.000000%
+  [PASS] n=  500 gaussian K=4 seed=2 max_rel=0.000132%
+  [PASS] n=  500 gaussian K=8 seed=0 max_rel=0.000167%
+  [PASS] n=  500 gaussian K=8 seed=1 max_rel=0.000204%
+  [PASS] n=  500 gaussian K=8 seed=2 max_rel=0.000132%
+  [PASS] n= 2000 bimodal  K=1 seed=0 max_rel=0.000000%
+  [PASS] n= 2000 bimodal  K=1 seed=1 max_rel=0.000000%
+  [PASS] n= 2000 bimodal  K=1 seed=2 max_rel=0.000000%
+  [PASS] n= 2000 bimodal  K=2 seed=0 max_rel=0.000007%
+  [PASS] n= 2000 bimodal  K=2 seed=1 max_rel=0.000007%
+  [PASS] n= 2000 bimodal  K=2 seed=2 max_rel=0.000007%
+  [PASS] n= 2000 bimodal  K=4 seed=0 max_rel=0.000007%
+  [PASS] n= 2000 bimodal  K=4 seed=1 max_rel=0.000007%
+  [PASS] n= 2000 bimodal  K=4 seed=2 max_rel=0.000035%
+  [PASS] n= 2000 bimodal  K=8 seed=0 max_rel=0.000028%
+  [PASS] n= 2000 bimodal  K=8 seed=1 max_rel=0.000021%
+  [PASS] n= 2000 bimodal  K=8 seed=2 max_rel=0.000035%
+  [PASS] n= 2000 gaussian K=1 seed=0 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=1 seed=1 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=1 seed=2 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=2 seed=0 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=2 seed=1 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=2 seed=2 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=4 seed=0 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=4 seed=1 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=4 seed=2 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=8 seed=0 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=8 seed=1 max_rel=0.000000%
+  [PASS] n= 2000 gaussian K=8 seed=2 max_rel=0.000215%
+
+Part 1 RESULT: ALL PASS (forward_batched_grad agrees with forward_batched within 0.1%)
+
+============================================================
+PART 2: Gradient correctness �� cosine(IFT grad, FD grad)
+============================================================
+  [PASS] K=1 k=0 cosine=0.9994  rel_norm_err=0.0718
+  [PASS] K=2 k=0 cosine=0.9994  rel_norm_err=0.0722
+  [PASS] K=2 k=1 cosine=0.9994  rel_norm_err=0.0344
+  [PASS] K=4 k=0 cosine=0.9994  rel_norm_err=0.0725
+  [PASS] K=4 k=1 cosine=0.9994  rel_norm_err=0.0362
+  [PASS] K=4 k=2 cosine=0.9997  rel_norm_err=0.0238
+  [PASS] K=4 k=3 cosine=0.9992  rel_norm_err=0.1125
+
+Part 2 RESULT: ALL PASS (cosine >= 0.9 for all K=1,2,4)
+
+============================================================
+PART 3: Smoke test backward on cuda
+============================================================
+  K=4, n=1000, device=cuda
+  h_crits: [1.6533485651016235, 1.6996427774429321, 1.6819535493850708, 1.6654480695724487]
+  forward+backward wall time: 561.1 ms
+  [PASS] xs[0].grad  norm=0.1055  min=-7.4008e-03  max=3.3091e-03
+  [PASS] xs[1].grad  norm=0.1763  min=-1.2266e-02  max=5.4856e-03
+  [PASS] xs[2].grad  norm=0.1174  min=-8.0658e-03  max=3.6058e-03
+  [PASS] xs[3].grad  norm=0.1089  min=-3.4354e-03  max=7.6481e-03
+
+Part 3 RESULT: PASS (backward ran, grads are finite)
+
+============================================================
+SUMMARY
+============================================================
+  Part 1 (forward correctness):  PASS
+  Part 2 (gradient cosine):       PASS
+  Part 3 (smoke backward):        PASS
+  Min cosine similarity (K=1,2,4): 0.9992
+
+OVERALL: ALL PASS