add batched_transpose implement

[fangche123]

No description provided.

[ThomasNing]

@fangche123 PTAL at the comments.

[ThomasNing]

Could we also measure the latency and TFLOPs of this operator? Thanks!

[poyenc]

to avoid repeating the target name, can you save the name into a variabble and reuse it later in CMake?

set(TARGET_NAME tile_example_batched_transpose)
add_executable(${TARGET_NAME} EXCLUDE_FROM_ALL batched_transpose.cpp batched_transpose_api.cpp)
target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/)

[ThruptiRajLakshmanaGowda]

This has been addressed

[poyenc]

dim_blocks, dim_block_h and dim_block_w are never used in kernel, suggest to remove them

[ThruptiRajLakshmanaGowda]

Addressed

[poyenc]

you never use dim_blocks in rest of code. suggest to remove it

[ThruptiRajLakshmanaGowda]

Addressed

[poyenc]

Since you are doing part of kernel selection here, I suggest to pass kparam directly to batched_transpose() API.

[ThruptiRajLakshmanaGowda]

This would need us to pass args as well would you suggest me to do that? That would duplicate few parts of the code!

[poyenc]

Suggest to pass kparam directly to batched_transpose() API, so that we won't have to expose implementation detail here. Then we can move those computations into the BATCHED_TRANSPOSE_DISPATCH() macro.

[ThruptiRajLakshmanaGowda]

To find karg we need both args(N, C, H, W) as well as kparam do you suggest to pass both these into batched_transpose()?

[poyenc]

why use pointer instead of reference const transpose_kernel_param_t& here? is kparam nullable?

[ThruptiRajLakshmanaGowda]

Addressed

[poyenc]

Just remove the name of unused parameter. So you won't have to do this.

bool transpose_kernel_is_valid(uint32_t, // 'batch' removed
                               uint32_t height,
                               uint32_t width,
                               const transpose_kernel_param_t* kparam)
{ /* ... */ }

[ThruptiRajLakshmanaGowda]

Addressed

[poyenc]

use reference here: const transpose_kernel_param_t& if kparam is not nullable

[ThruptiRajLakshmanaGowda]

Addressed

[poyenc]

remove no longer used macro

[ThruptiRajLakshmanaGowda]

Addressed

[poyenc]

You always set tile size=16x16 in BATCHED_TRANSPOSE_DISPATCH(). Why do we use other variables: kparam.tile_x/kparam.tile_y to decide the grid size?

[poyenc]

add static_assert() to specify the limitations here

static_assert(kMPerThread == 1 && kNPerThread == 1);

[ThruptiRajLakshmanaGowda]

Addressed

[spolifroni-amd]

Nothing to be reviewed for docs.