MeChris

Add ability to record and replay lists of ASync copies

1

By linking togeather a multiple ASync copies into a single user level transation the opportunity for gather scatter type operations becomes possible. This patch only allows this in the row...

A small number of bugs/issues had crept into the code and so a CI runtime of the C samples as been added to try to prevent this. The issues fixed...

TTL contains code code that looks like the follow // static inline TTL_int_void_sub_tensor_t attribute((overloadable)) TTL_step_buffering(TTL_import_double_const_void_tensor_buffering_t *const db, const TTL_tile_t next_tile); ...snip result.prev_tile = TTL_create_empty_tile(); TTL_step_buffering(&result, first_tile); return result; } ...snip...

TTL contains code code that looks like the follow // ``` static inline TTL_int_void_sub_tensor_t __attribute__((overloadable)) TTL_step_buffering(TTL_import_double_const_void_tensor_buffering_t *const db, const TTL_tile_t next_tile); ...snip result.prev_tile = TTL_create_empty_tile(); TTL_step_buffering(&result, first_tile); return result; }...

Add ability to record and replay lists of ASync copies

1

By linking togeather a multiple ASync copies into a single user level transation the opportunity for gather scatter type operations becomes possible. This patch only allows this in the row...

Add the ability to Align Tensor Rows and Planes

2

Some platforms require that the Rows/Planes of tensors are aligned to a specific value. Add two macros TTL_DEFAULT_INTERNAL_ALIGNMENT and TTL_DEFAULT_EXTERNAL_ALIGNMENT to set these alignement values. The default if undefined is...

The simplex buffering scheme must ensure that the import is complete before the export This is because 'need to understand' is a report with a solution given, but it is...

Certain platforms mandate specific alignment values for tensor rows/planes. To accommodate this, the system shall define two macros: TTL_DEFAULT_INTERNAL_ALIGNMENT: Specifies the default internal alignment for tensors. TTL_DEFAULT_EXTERNAL_ALIGNMENT: Specifies the default...

Allow virtual type methods for TTL_tiler

1

Implement virtual type methods for TTL_tiler to support different tiling strategies without requiring code duplication. This enables more flexible and efficient tiling implementations, especially in OpenCL C++ where virtual methods...

Create a CPP implementation of TTL #37

1

The C implementation of TTL is really an object-orientated CPP implementation written in C. The outcome for the use is good, with the ability to work well with Typed-Tensors, but...