Age | Commit message (Collapse) | Author |
|
PiperOrigin-RevId: 471584691
|
|
- Rename m1 to samples1
- Separate samples1 from microkernel name by underscore
PiperOrigin-RevId: 471574436
|
|
Add specific unit tests for packing routines.
Microkernel tests are updated to simply pass the primary_tile as kernel height, this result is a nop change, testing kernel sizes smaller than primary tile will be added in a follow-up.
PiperOrigin-RevId: 471142575
|
|
PiperOrigin-RevId: 470890302
|
|
PiperOrigin-RevId: 470331404
|
|
PiperOrigin-RevId: 469816851
|
|
another node
This can happen when a subgraph is split because an operation is not supported by XNNPACK. The FP16 rewriting code does not account for this:
When counting the number of external inputs, we assert that a node's input value which has a fp32_id is an external input, this is not always true, as it could be an external output
PiperOrigin-RevId: 469766360
|
|
Make FILTERBANK-ACCUMULATE microkernels match TFLM audio_frontend semantics
PiperOrigin-RevId: 469635373
|
|
PiperOrigin-RevId: 469514081
|
|
PiperOrigin-RevId: 469506037
|
|
PiperOrigin-RevId: 469480735
|
|
PiperOrigin-RevId: 469456395
|
|
- Remove input and weight offsets which are sequential
- Reduce weight count table from uint16 to uint8. Maximum value is 13.
PiperOrigin-RevId: 469371512
|
|
PiperOrigin-RevId: 469254069
|
|
PiperOrigin-RevId: 469114060
|
|
PiperOrigin-RevId: 469079271
|
|
PiperOrigin-RevId: 469077131
|
|
PiperOrigin-RevId: 468906177
|
|
microkernels
- sort applied to build files and headers.
PiperOrigin-RevId: 468770448
|
|
- Shift by 12 replaces vshlq + vqmovn with a vqshrn_n.
- Shift by 15 replaces vmull + vshlq + vqmovn with vqdmulhq
PiperOrigin-RevId: 468510420
|
|
- Scalar microkernel, test and benchmark
- Input and output are uint32
PiperOrigin-RevId: 468313409
|
|
a uint64 value.
PiperOrigin-RevId: 468301603
|
|
PiperOrigin-RevId: 468263623
|
|
normalization target needs to include "src" directory as it depends on
xnnpack/math.h
PiperOrigin-RevId: 468229416
|
|
PiperOrigin-RevId: 468220740
|
|
PiperOrigin-RevId: 468186161
|
|
PiperOrigin-RevId: 468105144
|
|
ARMv6 SIMD is deceptive because M-profile ARM cores don't support these
instructions until ARMv7
PiperOrigin-RevId: 468071398
|
|
This ensures that our microkernel tests (to be added) can depend on post-operation without depending on :operators.
PiperOrigin-RevId: 468002526
|
|
- input is uint32 and output is uint64
- multiply inputs by uint16 weights
PiperOrigin-RevId: 467435458
|
|
- Scalar C microkernel, test and benchmark
PiperOrigin-RevId: 466519212
|
|
PiperOrigin-RevId: 466191721
|
|
- Add fft size of 1024 benchmark
- Sort filenames in BUILD files
PiperOrigin-RevId: 465786067
|
|
- Apply sort to BUILD and CMakeList.txt
PiperOrigin-RevId: 465785238
|
|
- scalar microkernel, benchmark and unittest
PiperOrigin-RevId: 465672474
|
|
- Change size from 130 to 129.
- Format to 16 entries per row. Was 12.
- vlog to match microkernel name. was log.
- Remove lut from names. table and look up table is redundent.
PiperOrigin-RevId: 465120515
|
|
- scalar microkernel and benchmark
- read uint32 values, write uint16 log of values
PiperOrigin-RevId: 464941876
|
|
CPU case and use them when JIT is enabled
This allows us to actually test and run JIT generated code on emulators, previously there wasn't any generators configured, so no JIT code is generated, and when creating convolution operators we would fall back on the assembly microkernels.
- This adds assertion in convolution tests that specify use_jit(true) to ensure that code is generated by checking code cache size
- guard tests that specify use_jit(true) behind XNN_ENABLE_JIT, we really want those tests to run with JIT code, and we can only have JIT code if it is enabled
- update script to convert assembly to JIT code
- convert both GEMM and IGEMM ld128 microkernels to JIT, we don't convert the 1x8 because they are C code, and right now we don't need them yet (no JIT tests exercise this path), we can add this later
- the JIT code does not specialize on max_mr, that is the same behavior as currently, we don't yet enable microkernels for mr=4 (though we have it), this can also be fixed later
- disable JIT depthwise convolution tests, we don't have JIT dwconv microkernels yet, so those tests were incorrectly passing as they were using the assembly microkernels
PiperOrigin-RevId: 464637714
|
|
PiperOrigin-RevId: 464549548
|
|
-Change channels to batch
-Add _ before lo and hi
-Assert unsigned parameters != 0
-Fix typo in copyright
-Sort filenames in BUILD files
PiperOrigin-RevId: 464129540
|
|
- Rename channels to batch for C version
PiperOrigin-RevId: 463978900
|
|
- Scalar implementation
PiperOrigin-RevId: 463921828
|
|
- Extract declarations of microkernel parameters into microparams.h
- Group and document microkernel parameters
- Rename params-init accordingly
- Make microkernels depend only on microparams.h and not params.h
PiperOrigin-RevId: 463747649
|
|
- Shifts input to the left by specified amount.
- Scalar and Neon implementations
PiperOrigin-RevId: 463734229
|
|
- Remove duplicate channels variable.
- Post increment pointers.
- Sort BUILD file names.
PiperOrigin-RevId: 463470697
|
|
PiperOrigin-RevId: 463282575
|
|
- Returns a single maximum absolute value in an array of int16_t
PiperOrigin-RevId: 463225292
|
|
PiperOrigin-RevId: 463085532
|
|
PiperOrigin-RevId: 462985602
|
|
- Multiplies input by weights, shifts and clamps
PiperOrigin-RevId: 462683877
|