- 07 8月, 2021 2 次提交
-
-
由 deqiangc 提交于
* Add a new test case for conv operator (#329) This new test case is based on issue #329 to increase coverage on optimized kernel's data precision for conv operator. In this test, input, output and filter are all 8 bits and filter tensor is of dimension 8x3x3x3 with different scales per output channel. TESTED= local test with x86 and HiFi4. * Move large test data variable into its own folder and use one flat top namespace instead of nested namespace. BUG=195779890
-
由 Artem Tsvetkov 提交于
-
- 06 8月, 2021 1 次提交
-
-
由 Nat Jeffries 提交于
-
- 28 7月, 2021 3 次提交
-
-
由 cad-audio 提交于
* REF_CODE_REFACTOR: hard_swish Refactoring the reference code for hard_swish operator. BUG=refactoring existing code. * Remove uint8 support. * fix formatting. Co-authored-by: NAdvait Jain <advaitjain@google.com>
-
由 cad-audio 提交于
* REF_CODE_REFACTOR: logicalAND and logicalOR Refactoring the reference code for logicalAND and logicalOR operators. BUG=refactoring existing code. * Fix formatting. * fix bad merge (copyright year). Co-authored-by: NAdvait Jain <advaitjain@google.com> Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com>
-
由 cad-audio 提交于
* REF_CODE_REFACTOR: logistic Refactoring the reference code for logistic operator. BUG=refactoring existing code. * Fix formatting Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com> Co-authored-by: NAdvait Jain <advaitjain@google.com>
-
- 27 7月, 2021 1 次提交
-
-
由 cad-audio 提交于
* REF_CODE_REFACTOR: activations.cc Refactoring the reference code for activations operators. BUG=refactoring existing code. * Remove uint8 support. * Ran `test_code_style.sh --fix_formatting` * Fix the build. Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com> Co-authored-by: NAdvait Jain <advaitjain@google.com>
-
- 09 7月, 2021 2 次提交
-
-
由 Shlomi Regev 提交于
- Save ~4KB of RAM by implementing a lightweight version of the vector, which only accesses integer values, without implicit conversion. - Change Micro kernels to access flexbuffers as vectors instead of maps, which is recommended by the flexbuffers doc for efficiency. Since the values in the vector are ordered alphabetically by their keys, the kernels can access them by index instead. - Revert detection_postprocess to the native flexbuffers API. The LiteVector API doesn't support IsNull() and I prefer not to support two custom flexbuffer APIs. Co-authored-by: NNat Jeffries <natmjeffries@gmail.com>
-
由 Jens Elofsson 提交于
* Remove MICROLITE_CC_KERNELS_SRCS from the MICROLITE_CC_SRCS list. Stop the kernel sources from being compiled into both the core objects and kernel objects. * Add kernel sources to list_library_sources * Fix project generation presubmit Co-authored-by: NNat Jeffries <natmjeffries@gmail.com>
-
- 02 7月, 2021 1 次提交
-
-
由 Nat Jeffries 提交于
Previously, some sources were built before THIRD_PARTY_TARGETS was downloaded, meaning dependencies in those sources on flatbuffers, gemmlowp, etc could be missed. An alternative to this "dependency for every source" approach could be to add $(THIRD_PARTY_TARGETS) as the first dependency in $(MICROLITE_LIB_PATH), but this relies on Make dependency ordering which is ill-advised Background: https://stackoverflow.com/questions/9159960/order-of-processing-components-in-makefile
-
- 30 6月, 2021 1 次提交
-
-
由 Nat Jeffries 提交于
-
- 25 6月, 2021 1 次提交
-
-
由 Shlomi Regev 提交于
- Add external lib references to Hexagon FFT functions
-
- 22 6月, 2021 2 次提交
-
-
由 Advait Jain 提交于
* share code with the cmsis implementation * groundwork needed to add in Xtensa implementation BUG=https://github.com/tensorflow/tflite-micro/issues/205
-
由 Ryan Kuester 提交于
* Remove lite-specific code from copy of SPACE_TO_DEPTH Remove the bulk of lite-specific code from the micro implementation of operator SPACE_TO_DEPTH. - Flatten namespace - Don't resize output tensors - Remove type other than int8 and float32 - Don't use gtest * Port operator SPACE_TO_DEPTH from lite Port the SPACE_TO_DEPTH operator from lite to micro. Add the operator and test to the build. Co-authored-by: NPete Warden <pete@petewarden.com>
-
- 18 6月, 2021 1 次提交
-
-
由 Fredrik Knutsson 提交于
* Put location variables before target .inc's are included They can be useful in the target .inc Change-Id: I0ee3f77f79be272f4dc3502fb4a38017d8162fa1 * Add ARMC6 compiler support for Corstone-300 target * Use fromelf instead of objcopy for armclang toolchain when generating a binary * Add ARMC6 linker and build flags to the Corstone-330 target * Add RETARGET macro in patch script to avoid undefined symbol build error for ARMC6. * Use a reduced set of Cortex-M CPU's for easier maintenance. Change-Id: Id9a20d57fa4fa0f1339f44523417e2dabfe7e152 * Review comment - Override exit symbol only for GCC Change-Id: I93628c92ee352f36c7e7dd99351e4a73c29a8d30 * Review comment - Change how to pass linker options for ARMC6 Change-Id: Iacec5a6df6902bc8a14f460d63ae917039f982cc * Review comment - correct the upmerge Change-Id: I9a250f6a444336f9b95428a126a94a18f40060dc Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com>
-
- 17 6月, 2021 1 次提交
-
-
由 Advait Jain 提交于
For the Xtensa docker container, we were getting an error message along the lines of python not found. And the result was that the specialization was happening incorrectly (i.e. the specialization for the kernels was failing since it relies on python since #160). #182 is likely the reason why this error passed the CI for #160 but started failing after.
-
- 16 6月, 2021 2 次提交
-
-
由 Nat Jeffries 提交于
* Separate core, kernel and third party objects. Add a second optimization level in the Makefile to enable different levels between kernels and the rest of the TFLM code. This results in a smaller binary with minimal performance impact compared with using a single optimization level. Allows the use of implicit patterns to compile all sources, choosing different flags for core, kernel and third party sources. The following measurements are taken using the hexagon toolchain + hexagon-size and hexagon-sim. For the keyword benchmark using -O2: text data bss dec 58140 37639 46612 142391 Cycles: 1700364 For the keyword benchmark using -O2 for kernels and -Oz for framework: text data bss dec 52796 37623 46612 137031 Cycles: 1759664 * Make the optimization level log an error. Remove OPTIMIZATION_LEVEL setting for bluepill since core framework now automatically is compiled with -Os. * Remove section that builds bluepill with -Os since default uses -Os. * Disable -Werror=vla in order to pass stm32 bare lib presubmit. * Chnage order so that -Wno-vla takes priority over -Wvla Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com>
-
由 Advait Jain 提交于
* Add Hexagon optimized kernels. * Hexagon optimized kernels copied from https://source.codeaurora.org/quic/embedded_ai/tensorflow at 2d052806c211144875c89315a4fc6f1393064cf6 * Changed the include paths and directory structure a bit. * Modified Makefile to allow optimized kernels to be in a separate directory * Path to the Hexagon lib is now specified on the command line. Verified that the optimized kernels are properly built and linked with: ``` make -f tensorflow/lite/micro/tools/make/Makefile TARGET=hexagon OPTIMIZED_KERNEL_DIR=hexagon OPTIMIZED_KERNEL_DIR_PREFIX=third_party HEXAGON_TFLM_LIB=~/Qualcomm/tflm_google/hexagon_tflm_core.a -j8 run_keyword_benchmark ``` Gives: ``` KeywordRunNIerations(1) took 52608 ticks (52 ms) ``` Whereas reference kernels with: ``` make -f tensorflow/lite/micro/tools/make/Makefile TARGET=hexagon -j8 run_keyword_benchmark ``` Gives: ``` KeywordRunNIerations(1) took 110248 ticks (110 ms) ``` BUG=http://b/190754463 * enable code style checks for third_party as well. * clang-formatted the hexagon kernels. * Rename to keep scope focused on what we currently need.
-
- 02 6月, 2021 1 次提交
-
-
由 Advait Jain 提交于
* The TFLM Makefile globs all the shared TFL/TFLM code. * This allows us to move the explicit list of sources and headers to the sync script since that is where we determine what code needs to be sync'd from upstream TF. With this change, we are ready to have the tflite_micro repository be the source of truth for all TFLM code and sync only the shared TFL/TFLM code from the tensorflow repo. Bug: http://b/182914089
-
- 27 5月, 2021 1 次提交
-
-
由 TFLM-bot 提交于
-
- 25 5月, 2021 1 次提交
-
-
由 rsun-bdti 提交于
* Port TFL kernel Gather to TFL Micro * Remove TODO comment in micro/kernel/gather.cc Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com>
-
- 23 5月, 2021 1 次提交
-
-
由 TFLM-bot 提交于
-
- 21 5月, 2021 1 次提交
-
-
由 TFLM-bot 提交于
Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com>
-
- 06 5月, 2021 1 次提交
-
-
由 TFLM-bot 提交于
-
- 04 5月, 2021 1 次提交
-
-
由 github-actions[bot] 提交于
Co-authored-by: NAdvait Jain <advaitjain@users.noreply.github.com>
-
- 29 4月, 2021 1 次提交
-
-
由 Advait Jain 提交于
-
- 21 4月, 2021 1 次提交
-
-
由 Advait Jain 提交于
-
- 10 4月, 2021 1 次提交
-
-
由 Advait Jain 提交于
Confirmed that the following command passes: ``` ./tensorflow/lite/micro/tools/ci_build/test_all_new.sh GITHUB_PRESUBMIT ```
-