Running on device: Graphics Device ModelType: AlexNet Kernels: cudnn Input shape: 128x3x224x224 cudnn :updateOutput(): 69.23 cudnn :updateGradInput(): 83.05 cudnn :accGradParameters(): 77.48 cudnn :Forward: 69.23 cudnn :Backward: 160.53 cudnn :TOTAL: 229.76 ModelType: AlexNet Kernels: nn Input shape: 128x3x224x224 nn :updateOutput(): 132.38 nn :updateGradInput(): 112.52 nn :accGradParameters(): 94.82 nn :Forward: 132.38 nn :Backward: 207.34 nn :TOTAL: 339.72 ModelType: AlexNet Kernels: fbfft Input shape: 128x3x224x224 WARNING: Logging before InitGoogleLogging() is written to STDERR I0505 17:26:10.509590 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 31x31 I0505 17:26:11.584348 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=375.39M strategy = (FBMM cuFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 3959.91 time = 9.54ms I0505 17:26:11.584362 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=291.50M strategy = (FBMM FBFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 7185.05 time = 5.26ms I0505 17:26:11.584365 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=291.50M strategy = (FBMM FBFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 7185.05 time = 5.26ms I0505 17:26:11.584368 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 31x31 I0505 17:26:12.499948 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=380.63M strategy = (FBMM cuFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 4061.46 time = 9.30ms I0505 17:26:12.499959 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=229.64M strategy = (FBMM FBFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 8003.12 time = 4.72ms I0505 17:26:12.499963 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=229.64M strategy = (FBMM FBFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 8003.12 time = 4.72ms I0505 17:26:12.499964 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 31x31 I0505 17:26:13.419617 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=378.54M strategy = (FBMM cuFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 3827.26 time = 9.87ms I0505 17:26:13.419628 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=244.32M strategy = (FBMM FBFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 7173.87 time = 5.27ms I0505 17:26:13.419631 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=244.32M strategy = (FBMM FBFFT) (b x p x f) = 128x64x192 (input rows x cols) = 31x31 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 7173.87 time = 5.27ms I0505 17:26:13.432436 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:14.064550 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=297.14M strategy = (Many cuFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2154.02 time = 8.87ms I0505 17:26:14.064561 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=214.96M strategy = (Many FBFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3490.91 time = 5.47ms I0505 17:26:14.064563 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=214.96M strategy = (Many FBFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3490.91 time = 5.47ms I0505 17:26:14.064566 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:14.694813 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=298.98M strategy = (Many cuFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2248.75 time = 8.50ms I0505 17:26:14.694823 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=193.99M strategy = (Many FBFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3870.26 time = 4.94ms I0505 17:26:14.694825 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=193.99M strategy = (Many FBFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3870.26 time = 4.94ms I0505 17:26:14.694828 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:15.291407 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=294.13M strategy = (Many cuFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2257.73 time = 8.46ms I0505 17:26:15.291419 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=236.98M strategy = (Many FBFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3644.29 time = 5.24ms I0505 17:26:15.291421 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=236.98M strategy = (Many FBFFT) (b x p x f) = 128x192x384 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3644.29 time = 5.24ms I0505 17:26:15.305467 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:16.068344 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=364.77M strategy = (Many cuFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2379.12 time = 10.71ms I0505 17:26:16.068354 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=239.08M strategy = (Many FBFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3944.71 time = 6.46ms I0505 17:26:16.068356 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=239.08M strategy = (Many FBFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3944.71 time = 6.46ms I0505 17:26:16.068359 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:16.817615 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=363.86M strategy = (Many cuFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2464.17 time = 10.34ms I0505 17:26:16.817626 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=252.71M strategy = (Many FBFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4110.21 time = 6.20ms I0505 17:26:16.817628 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=252.71M strategy = (Many FBFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4110.21 time = 6.20ms I0505 17:26:16.817631 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:17.527192 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=356.91M strategy = (Many cuFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2496.56 time = 10.21ms I0505 17:26:17.527204 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=297.80M strategy = (Many FBFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3929.13 time = 6.48ms I0505 17:26:17.527205 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=297.80M strategy = (Many FBFFT) (b x p x f) = 128x384x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3929.13 time = 6.48ms I0505 17:26:17.534040 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:18.118405 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=265.55M strategy = (Many cuFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2175.58 time = 7.81ms I0505 17:26:18.118417 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=181.40M strategy = (Many FBFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3611.78 time = 4.70ms I0505 17:26:18.118418 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=181.40M strategy = (Many FBFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3611.78 time = 4.70ms I0505 17:26:18.118420 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:18.690752 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=264.50M strategy = (Many cuFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2277.43 time = 7.46ms I0505 17:26:18.690762 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=182.45M strategy = (Many FBFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3922.63 time = 4.33ms I0505 17:26:18.690765 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=182.45M strategy = (Many FBFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3922.63 time = 4.33ms I0505 17:26:18.690774 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x15 I0505 17:26:19.242434 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=261.62M strategy = (Many cuFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 15x16 GReductions(virtual fmas)/s = 2310.74 time = 7.35ms I0505 17:26:19.242445 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=210.76M strategy = (Many FBFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3758.22 time = 4.52ms I0505 17:26:19.242449 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=210.76M strategy = (Many FBFFT) (b x p x f) = 128x256x256 (input rows x cols) = 15x15 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3758.22 time = 4.52ms fbfft :updateOutput(): 44.46 fbfft :updateGradInput(): 46.23 fbfft :accGradParameters(): 43.76 fbfft :Forward: 44.46 fbfft :Backward: 89.99 fbfft :TOTAL: 134.45 ModelType: VGG Model-A Kernels: cudnn Input shape: 64x3x224x224 cudnn :updateOutput(): 337.17 cudnn :updateGradInput(): 396.36 cudnn :accGradParameters(): 350.38 cudnn :Forward: 337.17 cudnn :Backward: 746.74 cudnn :TOTAL: 1083.91 ModelType: VGG Model-A Kernels: nn Input shape: 64x3x224x224 nn :updateOutput(): 344.24 nn :updateGradInput(): 376.29 nn :accGradParameters(): 368.62 nn :Forward: 344.24 nn :Backward: 744.91 nn :TOTAL: 1089.15 ModelType: VGG Model-A Kernels: fbfft Input shape: 64x3x224x224 I0505 17:26:59.683387 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 226x226 I0505 17:27:00.119372 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=1952.88M strategy = (Batch FBFFT) (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 256x256 GReductions(virtual fmas)/s = 211.88 time = 26.66ms I0505 17:27:00.119384 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=1952.88M strategy = (Batch FBFFT) (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 256x256 GReductions(virtual fmas)/s = 211.88 time = 26.66ms I0505 17:27:00.119386 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 226x226 I0505 17:27:00.501482 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=1242.30M strategy = (Batch FBFFT) (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 256x256 GReductions(virtual fmas)/s = 339.35 time = 16.65ms I0505 17:27:00.501493 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=1242.30M strategy = (Batch FBFFT) (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 256x256 GReductions(virtual fmas)/s = 339.35 time = 16.65ms I0505 17:27:00.501497 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 226x226 I0505 17:27:00.925504 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=1241.25M strategy = (Batch FBFFT) (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 256x256 GReductions(virtual fmas)/s = 240.71 time = 23.47ms I0505 17:27:00.925513 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=1241.25M strategy = (Batch FBFFT) (b x p x f) = 64x3x64 (input rows x cols) = 226x226 (filter rows x cols) = 3x3 (common rows x cols) = 256x256 GReductions(virtual fmas)/s = 240.71 time = 23.47ms fbfft :updateOutput(): 353.17 fbfft :updateGradInput(): 400.65 fbfft :accGradParameters(): 329.45 fbfft :Forward: 353.17 fbfft :Backward: 730.10 fbfft :TOTAL: 1083.26 ModelType: OverFeat[fast] Kernels: cudnn Input shape: 128x3x231x231 cudnn :updateOutput(): 234.52 cudnn :updateGradInput(): 302.58 cudnn :accGradParameters(): 264.28 cudnn :Forward: 234.52 cudnn :Backward: 566.86 cudnn :TOTAL: 801.38 ModelType: OverFeat[fast] Kernels: nn Input shape: 128x3x231x231 nn :updateOutput(): 380.10 nn :updateGradInput(): 270.30 nn :accGradParameters(): 226.31 nn :Forward: 380.10 nn :Backward: 496.61 nn :TOTAL: 876.72 ModelType: OverFeat[fast] Kernels: fbfft Input shape: 128x3x231x231 I0505 17:27:50.931972 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 28x28 I0505 17:27:52.513263 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=581.96M strategy = (FBMM cuFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 3998.35 time = 15.42ms I0505 17:27:52.513278 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=429.92M strategy = (FBMM FBFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 6893.10 time = 8.94ms I0505 17:27:52.513280 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=429.92M strategy = (FBMM FBFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 6893.10 time = 8.94ms I0505 17:27:52.513283 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 28x28 I0505 17:27:53.944504 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=513.02M strategy = (Batch cuFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 28x32 GReductions(virtual fmas)/s = 4257.01 time = 14.48ms I0505 17:27:53.944519 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=351.27M strategy = (FBMM FBFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 7197.13 time = 8.57ms I0505 17:27:53.944520 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=351.27M strategy = (FBMM FBFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 7197.13 time = 8.57ms I0505 17:27:53.944525 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 28x28 I0505 17:27:55.451619 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=510.26M strategy = (Batch cuFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 28x32 GReductions(virtual fmas)/s = 4104.10 time = 15.02ms I0505 17:27:55.451634 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=398.46M strategy = (FBMM FBFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 6566.82 time = 9.39ms I0505 17:27:55.451637 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=398.46M strategy = (FBMM FBFFT) (b x p x f) = 128x96x256 (input rows x cols) = 28x28 (filter rows x cols) = 5x5 (common rows x cols) = 32x32 GReductions(virtual fmas)/s = 6566.82 time = 9.39ms I0505 17:27:55.470185 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:27:57.291265 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=407.11M strategy = (Many cuFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 15x14 GReductions(virtual fmas)/s = 2365.48 time = 12.51ms I0505 17:27:57.291278 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=325.06M strategy = (Many FBFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3410.59 time = 8.68ms I0505 17:27:57.291281 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=325.06M strategy = (Many FBFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3410.59 time = 8.68ms I0505 17:27:57.291285 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:27:59.133831 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=411.04M strategy = (Many cuFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 15x14 GReductions(virtual fmas)/s = 2480.80 time = 11.93ms I0505 17:27:59.133846 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=295.70M strategy = (Many FBFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3747.59 time = 7.90ms I0505 17:27:59.133848 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=295.70M strategy = (Many FBFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3747.59 time = 7.90ms I0505 17:27:59.133852 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:28:00.783583 30142 SpatialConvolutionCuFFTTuner.cpp:115] Found best cufft result Buffer=399.25M strategy = (Many cuFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 15x14 GReductions(virtual fmas)/s = 2531.63 time = 11.69ms I0505 17:28:00.783597 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=383.78M strategy = (Many FBFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3564.11 time = 8.30ms I0505 17:28:00.783601 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=383.78M strategy = (Many FBFFT) (b x p x f) = 128x256x512 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 3564.11 time = 8.30ms I0505 17:28:00.804857 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:28:01.113546 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=948.96M strategy = (Many FBFFT) (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4088.76 time = 28.95ms I0505 17:28:01.113556 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=948.96M strategy = (Many FBFFT) (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4088.76 time = 28.95ms I0505 17:28:01.113559 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:28:01.375329 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=891.29M strategy = (Many FBFFT) (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4334.18 time = 27.31ms I0505 17:28:01.375337 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=891.29M strategy = (Many FBFFT) (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4334.18 time = 27.31ms I0505 17:28:01.375342 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:28:01.661890 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=1301.28M strategy = (Many FBFFT) (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4139.71 time = 28.60ms I0505 17:28:01.661898 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=1301.28M strategy = (Many FBFFT) (b x p x f) = 128x512x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4139.71 time = 28.60ms I0505 17:28:01.716519 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateOutput (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:28:02.401914 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=1629.49M strategy = (Many FBFFT) (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4395.90 time = 53.86ms I0505 17:28:02.401923 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=1629.49M strategy = (Many FBFFT) (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4395.90 time = 53.86ms I0505 17:28:02.401927 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = updateGradInput (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:28:02.909543 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=1628.44M strategy = (Many FBFFT) (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4562.17 time = 51.90ms I0505 17:28:02.909551 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=1628.44M strategy = (Many FBFFT) (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4562.17 time = 51.90ms I0505 17:28:02.909555 30142 SpatialConvolutionCuFFTTuner.cpp:160] START exploring FFT perf for pass = accGradParameters(b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 14x14 I0505 17:28:03.447571 30142 SpatialConvolutionCuFFTTuner.cpp:119] Found best fbfft result Buffer=2451.57M strategy = (Many FBFFT) (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4367.83 time = 54.21ms I0505 17:28:03.447579 30142 SpatialConvolutionCuFFTTuner.cpp:166] Found best result Buffer=2451.57M strategy = (Many FBFFT) (b x p x f) = 128x1024x1024 (input rows x cols) = 14x14 (filter rows x cols) = 3x3 (common rows x cols) = 16x16 GReductions(virtual fmas)/s = 4367.83 time = 54.21ms fbfft :updateOutput(): 139.69 fbfft :updateGradInput(): 136.14 fbfft :accGradParameters(): 131.44 fbfft :Forward: 139.69 fbfft :Backward: 267.57 fbfft :TOTAL: 407.26