micro/third-party/pico-tflmicro/PATCHES.md
This tree is vendored from
raspberrypi/pico-tflmicro (TFLM
under Apache 2.0, the Pico wrapper code MIT). It is not pristine — it carries
a small, platform-specific (RP2350) performance patch plus the usual
subdirectory-friendly trimming. Anything depending on TFLM being agnostic should
go through the tflm CMake INTERFACE target in the top-level CMakeLists.txt,
not reach in here.
Gated behind the SPELLING_TINY_MULTICORE compile definition (set by the parent
build; ON by default). It splits the two hottest int8 kernels across both
Cortex-M33 cores for a measured ~1.6x inference speedup with bit-identical
output. Files touched:
src/third_party/cmsis_nn/Include/spelling_multicore.h — added. The
core-1 persistent-worker dispatch API (spelling_mc_dispatch_task() etc.).src/third_party/cmsis_nn/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c
— patched to partition the output channels across the two cores under
SPELLING_TINY_MULTICORE (spelling_mm_dsp_rows, __not_in_flash_func).src/third_party/cmsis_nn/Source/ConvolutionFunctions/arm_depthwise_conv_3x3_s8.c
— patched to partition the output rows across the two cores
(spelling_dw3x3_rows).When SPELLING_TINY_MULTICORE is undefined these fall back to the upstream
single-core scalar/SIMD paths, so the patch is inert in a single-core build.
Because this is platform-specific, it is the one piece of "platform code" that
could not be lifted out into examples/rp2350/ — it is physically interleaved
with the CMSIS-NN kernels. A port to another platform would simply leave
SPELLING_TINY_MULTICORE undefined.
Upstream TFLM has no GELU kernel (the schema, TfLiteGeluParams, and the
ParseOpData switch case all exist upstream, but no micro implementation).
The Phase-B TTS decoder graph (s16x8: int16 activations, int8 weights) needs
GELU and int16 TRANSPOSE, so:
src/tensorflow/lite/micro/kernels/gelu.cpp — added. LUT-based int8
(256-entry) and int16 (513-entry interpolated) paths via the shared
LUTPopulate/LUTLookup helpers, plus a float path. Mirrors how full
TFLite lowers GELU.src/tensorflow/lite/micro/kernels/micro_ops.h — Register_GELU() decl.src/tensorflow/lite/micro/micro_mutable_op_resolver.h — AddGelu().src/tensorflow/lite/core/api/flatbuffer_conversions.{h,cpp} — standalone
ParseGelu() (reads the approximate flag) for the micro resolver.src/tensorflow/lite/micro/kernels/transpose.cpp — added the
kTfLiteInt16 case (pure data movement, same reference kernel).CMakeLists.txt — added gelu.cpp to the kernel sources.If upstream ever adds a GELU micro kernel, prefer theirs on the next sync.
The upstream top-level project() / pico_sdk_init() calls and the
example/test subdirectory blocks were removed so this builds as an
add_subdirectory() of the parent (which does both). See the comment at the top
of CMakeLists.txt.