mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-12 01:10:17 +00:00
e17c0d8008
* Initial adding of generic reduction * Initial adding of generic reduction ... * Updates to make compiling done * clang-format all files * clang-format some files again * Renaming in profiler/include/profile_reduce.hpp * Updates and make BlockWise cases passed * Updates and make ThreadWise and MultiBlockTwoCall cases passed * Remove the support for MUL and NORM1 reduceOp from the profiler and the device instances * Change to replace the dim0_max_vector_size/dim1_max_vector_size template argument in the device reduce classes * format * adding pooling * added max and average pooling * comment out cout and kernel timing * Tiny simplification in profiler/reduce_profiler.cpp * Add example for reduce_blockwise * Tiny updates * Change to pass the ElementWiseOp from device layer to kernel * Fix the vectorDim and vectorSize in Device layer * Enable vector load on both dim0 and dim1 for Threadwise method * Tiny updates * Change to let the user to pass the preUnaryOp and posUnaryOp * Make pooling example work * split device_reduce_instance into two libraries * Tiny update * Replace nanPropaOpt enum by boolean propagate_nan * Simplification in DeviceReduce layer codes * update build * Change to clarify the difference between ck::half_t and half_float::half * Renaming in all the reduction codes * Add VectorSize as template parameter for device layer * Add BetaIsZero as kernel template and as AccDataType for alpha * print * Small updates for pooling * Updates for host_generic_reduction for reference * Update to make AVG pooling pass * Update to make MAX pooling with indices output pass * fix * add OutDst vector store to threadwise reduction and pooling * tweak * turn off check_indices that caused build issue * refactor pooling * clean up * turn off check_indices for building issue for php-compiler * add more tile size for odd C * tweak conv for odd C * update script * clean up elementwise op * add hack in reduction_operator.hpp to avoid compile error. To fix it, need to use element_wise_op in reduction op * Add OutVectorSize as device and kernel tunable, also update to Elementwise Operations * Move reduce operator mapping to host layer file reduction_operator_mapping.hpp from reduction_operator.hpp * Change to the unary operators * Move the definitions of unary operations to element_wise_operation.hpp * re-org files * Refine in device interfaces and multiblock kernels * Split the reduction configurations into instances for specific methods * Update in getTypeString() of device pool2d * Renaming in host and kernel * Tiny update in profiler/src/profiler.cpp * Uncomment in device_operation/CMakeLists.txt to enable the building of all operations * Make check_indices a templated function to remove some linking issue * Renaming in the profiler reduce module * Add support for double Reduction (but disable MultiblockAtomicAdd for double) * Tiny correction of literal string * Rename DevicePoolFwd to DevicePool2dFwd * Split device_reduce_instance_xxx.cpp files according to the data types to speed up compiling * Add comments for lists of configurations, lists of instances and references of add_reduce_instances_xxx * Remove un-used header file gridwise_generic_reduction_wrapper_common.hpp * Renaming and refining in the Reduction codes * Tiny change in the unary operators * Renaming symbols and files * Renaming symbols in the kernels * Move kernel kernel_set_buffer_value to separate file * Add IndexDataType template parameter for kernels and use int32_t as index data type in device layer * Tiny update in the kernels * Remove definition of sqrtf()/isnan()/abs() for half_t due to some ADL issue * Simplify a helper function in device layer * Tiny adjustment in testing data initialization * Renaming in kernel/device/host * Add two testing scripts for reduction * Refine the Unary operators in element_wise_operation.hpp * Update in the reduce profiler module * Update to the reduction testing scripts * reduce compile parallelism * change CI docker to rocm5.0 * remove unused variables * fix build Co-authored-by: Chao Liu <chao.liu2@amd.com>
Docker script
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
Build ckProfiler
mkdir build && cd build
# Need to Specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
make -j ckProfiler
Profile GEMM kernels
#arg1: tensor operation (gemm=GEMM)
#arg2: data type (0=fp32, 1=fp16)
#arg3: matrix layout (0=NN, 1=NT, 2=TN, 3=TT)
#arg4: verification (0=no, 1=yes)
#arg5: initialization (0=no init, 1=integer value, 2=decimal value)
#arg6: print matrix value (0=no, 1=yes)
#arg7: run kernel # of times (>1)
#arg8 to 13: M, N, K, StrideA, StrideB, StrideC
##################### op datatype layout verify init log repeat M___ N___ K___ StrideA StrideB StrideC
./profiler/ckProfiler gemm 1 1 1 1 0 5 3840 4096 4096 4096 4096 4096
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
....
Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
Profile forward convolution kernels
#arg1: tensor operation (conv=Convolution)
#arg2: data type (0=fp32, 1=fp16)
#arg3: input tensor layout (0=NCHW, 1=NHWC)
#arg4: weight tensor layout (0=KCYX, 1=KYXC)
#arg5: output tensor layout (0=NKHW, 1=NHWK)
#arg6: verification (0=no, 1=yes)
#arg7: initialization (0=no init, 1=integer value, 2=decimal value)
#arg8: print matrix value (0=no, 1=yes)
#arg9: run kernel # of times (>1)
#arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
##################### op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
./profiler/ckProfiler conv 1 1 1 1 1 1 0 5 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
....
Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s