diff options
Diffstat (limited to 'cuda-kernels/scripts')
| -rwxr-xr-x | cuda-kernels/scripts | 118 |
1 files changed, 101 insertions, 17 deletions
diff --git a/cuda-kernels/scripts b/cuda-kernels/scripts index b9a3b50..a24fb72 100755 --- a/cuda-kernels/scripts +++ b/cuda-kernels/scripts @@ -10,11 +10,16 @@ nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp4_256 ./vp4_128>vp4_log128 ./vp4_256>vp4_log256 -grep "kernel_name.*vp_example" -A 589 vp4_log16 >vp4_16_summary -grep "kernel_name.*vp_example" -A 589 vp4_log32 >vp4_32_summary -grep "kernel_name.*vp_example" -A 589 vp4_log64 >vp4_64_summary -grep "kernel_name.*vp_example" -A 589 vp4_log128 >vp4_128_summary -grep "kernel_name.*vp_example" -A 589 vp4_log256 >vp4_256_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log16 >vp4_16_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log32 >vp4_32_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log64 >vp4_64_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log128 >vp4_128_summary +#grep "kernel_name.*convertInt32To" -A 589 vp4_log256 >vp4_256_summary +###grep "kernel_name.*vp_example" -A 589 vp4_log16 >>vp4_16_summary +###grep "kernel_name.*vp_example" -A 589 vp4_log32 >>vp4_32_summary +###grep "kernel_name.*vp_example" -A 589 vp4_log64 >>vp4_64_summary +#grep "kernel_name.*vp_example" -A 589 vp4_log128 >>vp4_128_summary +#grep "kernel_name.*vp_example" -A 589 vp4_log256 >>vp4_256_summary nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp8_16 v8p_genericMatrixMultiply.cu -DSIZE=16 @@ -28,11 +33,16 @@ nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp8_256 ./vp8_128>vp8_log128 ./vp8_256>vp8_log256 -grep "kernel_name.*vp_example" -A 589 vp8_log16 >vp8_16_summary -grep "kernel_name.*vp_example" -A 589 vp8_log32 >vp8_32_summary -grep "kernel_name.*vp_example" -A 589 vp8_log64 >vp8_64_summary -grep "kernel_name.*vp_example" -A 589 vp8_log128 >vp8_128_summary -grep "kernel_name.*vp_example" -A 589 vp8_log256 >vp8_256_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log16 >vp8_16_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log32 >vp8_32_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log64 >vp8_64_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log128 >vp8_128_summary +#grep "kernel_name.*convertInt32To" -A 589 vp8_log256 >vp8_256_summary +##grep "kernel_name.*vp_example" -A 589 vp8_log16 >>vp8_16_summary +##grep "kernel_name.*vp_example" -A 589 vp8_log32 >>vp8_32_summary +##grep "kernel_name.*vp_example" -A 589 vp8_log64 >>vp8_64_summary +#grep "kernel_name.*vp_example" -A 589 vp8_log128 >>vp8_128_summary +#grep "kernel_name.*vp_example" -A 589 vp8_log256 >>vp8_256_summary nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp16_16 v16p_genericMatrixMultiply.cu -DSIZE=16 nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp16_32 v16p_genericMatrixMultiply.cu -DSIZE=32 @@ -45,11 +55,85 @@ nvcc --gpu-architecture=compute_70 --gpu-code=compute_70 -lcudart -g -o vp16_256 ./vp16_128>vp16_log128 ./vp16_256>vp16_log256 -grep "kernel_name.*vp_example" -A 589 vp16_log16 >vp16_16_summary -grep "kernel_name.*vp_example" -A 589 vp16_log32 >vp16_32_summary -grep "kernel_name.*vp_example" -A 589 vp16_log64 >vp16_64_summary -grep "kernel_name.*vp_example" -A 589 vp16_log128 >vp16_128_summary -grep "kernel_name.*vp_example" -A 589 vp16_log256 >vp16_256_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log16 >vp16_16_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log32 >vp16_32_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log64 >vp16_64_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log128 >vp16_128_summary +#grep "kernel_name.*convertInt32To" -A 589 vp16_log256 >vp16_256_summary +##grep "kernel_name.*vp_example" -A 589 vp16_log16 >>vp16_16_summary +##grep "kernel_name.*vp_example" -A 589 vp16_log32 >>vp16_32_summary +##grep "kernel_name.*vp_example" -A 589 vp16_log64 >>vp16_64_summary +#grep "kernel_name.*vp_example" -A 589 vp16_log128 >>vp16_128_summary +#grep "kernel_name.*vp_example" -A 589 vp16_log256 >>vp16_256_summary -mv vp* VPlog/ -make clean +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log16 > kernel_log_vp4_log16 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log32 > kernel_log_vp4_log32 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log64 > kernel_log_vp4_log64 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log128 > kernel_log_vp4_log128 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp4_log256 > kernel_log_vp4_log256 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log16 > kernel_log_vp8_log16 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log32 > kernel_log_vp8_log32 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log64 > kernel_log_vp8_log64 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log128 > kernel_log_vp8_log128 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp8_log256 > kernel_log_vp8_log256 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log16 > kernel_log_vp16_log16 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log32 > kernel_log_vp16_log32 +###sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log64 > kernel_log_vp16_log64 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log128 > kernel_log_vp16_log128 +####sed -n '/bind.*to.*kernel.*vp_example/,$p' vp16_log256 > kernel_log_vp16_log256 +### +###grep memtocore:Icnt kernel_log_vp4_log16>memtocore_vp4_log16 +###grep coretomem:Icnt kernel_log_vp4_log16>coretomem_vp4_log16 +###grep MEM_TXN kernel_log_vp4_log16> gen_mem_vp4_log16 +###grep memtocore:Icnt kernel_log_vp4_log32>memtocore_vp4_log32 +###grep coretomem:Icnt kernel_log_vp4_log32>coretomem_vp4_log32 +###grep MEM_TXN kernel_log_vp4_log32> gen_mem_vp4_log32 +###grep memtocore:Icnt kernel_log_vp4_log64>memtocore_vp4_log64 +###grep coretomem:Icnt kernel_log_vp4_log64>coretomem_vp4_log64 +###grep MEM_TXN kernel_log_vp4_log64> gen_mem_vp4_log64 +####grep memtocore:Icnt kernel_log_vp4_log128>memtocore_vp4_log128 +####grep coretomem:Icnt kernel_log_vp4_log128>coretomem_vp4_log128 +####grep MEM_TXN kernel_log_vp4_log128> gen_mem_vp4_log128 +####grep memtocore:Icnt kernel_log_vp4_log256>memtocore_vp4_log256 +####grep coretomem:Icnt kernel_log_vp4_log256>coretomem_vp4_log256 +####grep MEM_TXN kernel_log_vp4_log256> gen_mem_vp4_log256 +### +###grep memtocore:Icnt kernel_log_vp8_log16>memtocore_vp8_log16 +###grep coretomem:Icnt kernel_log_vp8_log16>coretomem_vp8_log16 +###grep MEM_TXN kernel_log_vp8_log16> gen_mem_vp8_log16 +###grep memtocore:Icnt kernel_log_vp8_log32>memtocore_vp8_log32 +###grep coretomem:Icnt kernel_log_vp8_log32>coretomem_vp8_log32 +###grep MEM_TXN kernel_log_vp8_log32> gen_mem_vp8_log32 +###grep memtocore:Icnt kernel_log_vp8_log64>memtocore_vp8_log64 +###grep coretomem:Icnt kernel_log_vp8_log64>coretomem_vp8_log64 +###grep MEM_TXN kernel_log_vp8_log64> gen_mem_vp8_log64 +####grep memtocore:Icnt kernel_log_vp8_log128>memtocore_vp8_log128 +####grep coretomem:Icnt kernel_log_vp8_log128>coretomem_vp8_log128 +####grep MEM_TXN kernel_log_vp8_log128> gen_mem_vp8_log128 +####grep memtocore:Icnt kernel_log_vp8_log256>memtocore_vp8_log256 +####grep coretomem:Icnt kernel_log_vp8_log256>coretomem_vp8_log256 +####grep MEM_TXN kernel_log_vp8_log256> gen_mem_vp8_log256 +### +###grep memtocore:Icnt kernel_log_vp16_log16>memtocore_vp16_log16 +###grep coretomem:Icnt kernel_log_vp16_log16>coretomem_vp16_log16 +###grep MEM_TXN kernel_log_vp16_log16> gen_mem_vp16_log16 +###grep memtocore:Icnt kernel_log_vp16_log32>memtocore_vp16_log32 +###grep coretomem:Icnt kernel_log_vp16_log32>coretomem_vp16_log32 +###grep MEM_TXN kernel_log_vp16_log32> gen_mem_vp16_log32 +###grep memtocore:Icnt kernel_log_vp16_log64>memtocore_vp16_log64 +###grep coretomem:Icnt kernel_log_vp16_log64>coretomem_vp16_log64 +###grep MEM_TXN kernel_log_vp16_log64> gen_mem_vp16_log64 +####grep memtocore:Icnt kernel_log_vp16_log128>memtocore_vp16_log128 +####grep coretomem:Icnt kernel_log_vp16_log128>coretomem_vp16_log128 +####grep MEM_TXN kernel_log_vp16_log128> gen_mem_vp16_log128 +####grep memtocore:Icnt kernel_log_vp16_log256>memtocore_vp16_log256 +####grep coretomem:Icnt kernel_log_vp16_log256>coretomem_vp16_log256 +####grep MEM_TXN kernel_log_vp16_log256> gen_mem_vp16_log256 +### +###mv kernel_* MemTraffic/ +###mv memtocore* MemTraffic/ +###mv coretomem* MemTraffic/ +###mv gen_mem* MemTraffic/ +### +###mv vp* VPlog/ +###make clean |