Home
last modified time | relevance | path

Searched defs:Sm75 (Results 1 – 25 of 36) sorted by relevance

12

/external/pytorch/aten/src/ATen/native/transformers/cuda/mem_eff_attention/kernels/
DcutlassB_f16_notaligned_k65536_dropout.cu31 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, true, false, 128, 64, 65536>:… in __launch_bounds__()
88 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, true, false, 64, 64, 65536>::… in __launch_bounds__()
DcutlassB_f16_notaligned_k128_dropout.cu31 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, true, false, 128, 64, 128>::k… in __launch_bounds__()
88 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, true, false, 64, 64, 128>::kN… in __launch_bounds__()
DcutlassB_f16_notaligned_k128.cu31 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, false, false, 128, 64, 128>::… in __launch_bounds__()
88 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, false, false, 64, 64, 128>::k… in __launch_bounds__()
DcutlassB_f16_notaligned_k65536.cu31 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, false, false, 128, 64, 65536>… in __launch_bounds__()
88 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, false, false, 64, 64, 65536>:… in __launch_bounds__()
DcutlassF_f32_notaligned.cu50 AttentionKernel<float, cutlass::arch::Sm75, false, 64, 64, 64, true, true>::kNumThreads, in __launch_bounds__()
107 AttentionKernel<float, cutlass::arch::Sm75, false, 32, 128, 128, true, true>::kNumThreads, in __launch_bounds__()
164 AttentionKernel<float, cutlass::arch::Sm75, false, 32, 128, 65536, true, true>::kNumThreads, in __launch_bounds__()
DcutlassF_f16_notaligned.cu50 … AttentionKernel<cutlass::half_t, cutlass::arch::Sm75, false, 64, 64, 64, true, true>::kNumThreads, in __launch_bounds__()
107 …AttentionKernel<cutlass::half_t, cutlass::arch::Sm75, false, 32, 128, 128, true, true>::kNumThread… in __launch_bounds__()
164 …AttentionKernel<cutlass::half_t, cutlass::arch::Sm75, false, 32, 128, 65536, true, true>::kNumThre… in __launch_bounds__()
DcutlassB_f16_aligned_k65536_dropout.cu31 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, true, true, false, 128, 64, 65536>::… in __launch_bounds__()
107 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, true, true, false, 64, 64, 65536>::k… in __launch_bounds__()
DcutlassB_f16_aligned_k128_dropout.cu31 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, true, true, false, 128, 64, 128>::kN… in __launch_bounds__()
107 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, true, true, false, 64, 64, 128>::kNu… in __launch_bounds__()
DcutlassB_f16_aligned_k65536.cu31 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, true, false, false, 128, 64, 65536>:… in __launch_bounds__()
107 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, true, false, false, 64, 64, 65536>::… in __launch_bounds__()
DcutlassF_f16_aligned.cu50 … AttentionKernel<cutlass::half_t, cutlass::arch::Sm75, true, 64, 64, 64, true, true>::kNumThreads, in __launch_bounds__()
126 …AttentionKernel<cutlass::half_t, cutlass::arch::Sm75, true, 32, 128, 128, true, true>::kNumThreads, in __launch_bounds__()
202 …AttentionKernel<cutlass::half_t, cutlass::arch::Sm75, true, 32, 128, 65536, true, true>::kNumThrea… in __launch_bounds__()
DcutlassF_f32_aligned.cu50 AttentionKernel<float, cutlass::arch::Sm75, true, 64, 64, 64, true, true>::kNumThreads, in __launch_bounds__()
126 AttentionKernel<float, cutlass::arch::Sm75, true, 32, 128, 128, true, true>::kNumThreads, in __launch_bounds__()
202 AttentionKernel<float, cutlass::arch::Sm75, true, 32, 128, 65536, true, true>::kNumThreads, in __launch_bounds__()
DcutlassB_f32_notaligned_k65536_dropout.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, float, false, true, false, 64, 64, 65536>::kNumThread… in __launch_bounds__()
DcutlassB_f16_notaligned_k32.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, false, false, 64, 64, 32>::kN… in __launch_bounds__()
DcutlassB_f16_notaligned_k64.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, false, false, 64, 64, 64>::kN… in __launch_bounds__()
DcutlassB_f32_notaligned_k64_dropout.cu50 … AttentionBackwardKernel<cutlass::arch::Sm75, float, false, true, false, 64, 64, 64>::kNumThreads, in __launch_bounds__()
DcutlassB_f32_notaligned_k128_dropout.cu50 … AttentionBackwardKernel<cutlass::arch::Sm75, float, false, true, false, 64, 64, 128>::kNumThreads, in __launch_bounds__()
DcutlassB_f32_notaligned_k32_dropout.cu50 … AttentionBackwardKernel<cutlass::arch::Sm75, float, false, true, false, 64, 64, 32>::kNumThreads, in __launch_bounds__()
DcutlassB_f32_notaligned_k128.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, float, false, false, false, 64, 64, 128>::kNumThreads, in __launch_bounds__()
DcutlassB_f32_notaligned_k65536.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, float, false, false, false, 64, 64, 65536>::kNumThrea… in __launch_bounds__()
DcutlassB_f16_notaligned_k32_dropout.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, true, false, 64, 64, 32>::kNu… in __launch_bounds__()
DcutlassB_f32_notaligned_k64.cu50 … AttentionBackwardKernel<cutlass::arch::Sm75, float, false, false, false, 64, 64, 64>::kNumThreads, in __launch_bounds__()
DcutlassB_f16_notaligned_k64_dropout.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, false, true, false, 64, 64, 64>::kNu… in __launch_bounds__()
DcutlassB_f32_notaligned_k32.cu50 … AttentionBackwardKernel<cutlass::arch::Sm75, float, false, false, false, 64, 64, 32>::kNumThreads, in __launch_bounds__()
DcutlassB_f16_aligned_k32_dropout.cu50 …AttentionBackwardKernel<cutlass::arch::Sm75, cutlass::half_t, true, true, false, 64, 64, 32>::kNum… in __launch_bounds__()
DcutlassB_f32_aligned_k64.cu50 … AttentionBackwardKernel<cutlass::arch::Sm75, float, true, false, false, 64, 64, 64>::kNumThreads, in __launch_bounds__()

12