Merge remote-tracking branch 'upstream/master' into tile_matmul_bf16

.clang-tidy

@@ -17,6 +17,8 @@ Checks: >
  -misc-unconventional-assign-operator,
  -misc-unused-parameters,
  modernize-deprecated-headers,
+ modernize-make-shared,
+ modernize-make-unique,
  modernize-redundant-void-arg,
  modernize-use-bool-literals,
  modernize-use-default-member-init,
@@ -28,6 +30,7 @@ Checks: >
  performance-*,
  -performance-inefficient-string-concatenation,
  -performance-inefficient-vector-operation,
+ -performance-no-int-to-ptr,
  readability-avoid-const-params-in-decls,
  readability-braces-around-statements,
  readability-const-return-type,

.github/workflows/presubmit.yml

@@ -15,11 +15,11 @@ jobs:
  runs-on: ubuntu-20.04
  steps:
  - uses: actions/checkout@v2
- - uses: DoozyX/clang-format-lint-action@v0.11
+ - uses: DoozyX/clang-format-lint-action@v0.12
  with:
  source: '.'
  extensions: 'h,c,cpp'
- clangFormatVersion: 11
+ clangFormatVersion: 12
  check_clang_tidy:
  name: Check clang-tidy
  runs-on: ubuntu-20.04
@@ -28,12 +28,12 @@ jobs:
  - name: Install clang-tidy
  run: |
  sudo apt-get update
- sudo apt-get install llvm-11 clang-11 liblld-11-dev libclang-11-dev clang-tidy-11 ninja-build
+ sudo apt-get install llvm-12 clang-12 liblld-12-dev libclang-12-dev clang-tidy-12 ninja-build
  - name: Run clang-tidy
  run: |
- export CC=clang-11
- export CXX=clang++-11
- export CLANG_TIDY_LLVM_INSTALL_DIR=/usr/lib/llvm-11
+ export CC=clang-12
+ export CXX=clang++-12
+ export CLANG_TIDY_LLVM_INSTALL_DIR=/usr/lib/llvm-12
  ./run-clang-tidy.sh
  check_cmake_file_lists:
  name: Check CMake file lists

.gitignore

@@ -12,7 +12,7 @@ python_bindings/bin/*
 build-64/*
 build-ios/*
 build-osx/*
-build_wasm/*
+build_wasm*/*
 cmake_build*/*
 */build/*
 tmp/*

Makefile

@@ -1169,18 +1169,6 @@ GENERATOR_AOTCPP_TESTS := $(filter-out generator_aotcpp_cleanup_on_error,$(GENER
 # https://github.com/halide/Halide/issues/2084 (only if opencl enabled)
 GENERATOR_AOTCPP_TESTS := $(filter-out generator_aotcpp_buffer_copy,$(GENERATOR_AOTCPP_TESTS))
 
-# https://github.com/halide/Halide/issues/2071
-GENERATOR_AOTCPP_TESTS := $(filter-out generator_aotcpp_user_context,$(GENERATOR_AOTCPP_TESTS))
-
-# https://github.com/halide/Halide/issues/2071
-GENERATOR_AOTCPP_TESTS := $(filter-out generator_aotcpp_argvcall,$(GENERATOR_AOTCPP_TESTS))
-
-# https://github.com/halide/Halide/issues/2071
-GENERATOR_AOTCPP_TESTS := $(filter-out generator_aotcpp_metadata_tester,$(GENERATOR_AOTCPP_TESTS))
-
-# https://github.com/halide/Halide/issues/2071
-GENERATOR_AOTCPP_TESTS := $(filter-out generator_aotcpp_cxx_mangling,$(GENERATOR_AOTCPP_TESTS))
-
 # https://github.com/halide/Halide/issues/2075
 GENERATOR_AOTCPP_TESTS := $(filter-out generator_aotcpp_msan,$(GENERATOR_AOTCPP_TESTS))
 
@@ -1471,6 +1459,8 @@ $(FILTERS_DIR)/metadata_tester_ucon.a: $(BIN_DIR)/metadata_tester.generator
 
 $(BIN_DIR)/$(TARGET)/generator_aot_metadata_tester: $(FILTERS_DIR)/metadata_tester_ucon.a
 
+$(BIN_DIR)/$(TARGET)/generator_aotcpp_metadata_tester: $(FILTERS_DIR)/metadata_tester_ucon.halide_generated.cpp
+
 $(FILTERS_DIR)/multitarget.a: $(BIN_DIR)/multitarget.generator
  @mkdir -p $(@D)
  $(CURDIR)/$< -g multitarget -f "HalideTest::multitarget" $(GEN_AOT_OUTPUTS) -o $(CURDIR)/$(FILTERS_DIR) \
@@ -2131,6 +2121,10 @@ ifneq (,$(findstring clang version 13.0,$(CLANG_VERSION)))
 CLANG_OK=yes
 endif
 
+ifneq (,$(findstring clang version 14.0,$(CLANG_VERSION)))
+CLANG_OK=yes
+endif
+
 ifneq (,$(findstring Apple LLVM version 5.0,$(CLANG_VERSION)))
 CLANG_OK=yes
 endif
@@ -2151,7 +2145,7 @@ $(BUILD_DIR)/clang_ok:
  @exit 1
 endif
 
-ifneq (,$(findstring $(LLVM_VERSION_TIMES_10), 110 111 120 130))
+ifneq (,$(findstring $(LLVM_VERSION_TIMES_10), 110 111 120 130 140))
 LLVM_OK=yes
 endif
 
@@ -2345,15 +2339,15 @@ $(BIN_DIR)/HalideTraceDump: $(ROOT_DIR)/util/HalideTraceDump.cpp $(ROOT_DIR)/uti
 
 # Note: you must have CLANG_FORMAT_LLVM_INSTALL_DIR set for this rule to work.
 # Let's default to the Ubuntu install location.
-CLANG_FORMAT_LLVM_INSTALL_DIR ?= /usr/lib/llvm-11
+CLANG_FORMAT_LLVM_INSTALL_DIR ?= /usr/lib/llvm-12
 
 .PHONY: format
 format:
  @CLANG_FORMAT_LLVM_INSTALL_DIR=$(CLANG_FORMAT_LLVM_INSTALL_DIR) ${ROOT_DIR}/run-clang-format.sh
 
 # Note: you must have CLANG_TIDY_LLVM_INSTALL_DIR set for these rules to work.
 # Let's default to the Ubuntu install location.
-CLANG_TIDY_LLVM_INSTALL_DIR ?= /usr/lib/llvm-11
+CLANG_TIDY_LLVM_INSTALL_DIR ?= /usr/lib/llvm-12
 
 .PHONY: clang-tidy
 clang-tidy:

apps/blur/halide_blur_generator.cpp

@@ -87,7 +87,7 @@ class HalideBlur : public Halide::Generator<HalideBlur> {
 
  blur_y.compute_root()
  .hexagon()
- .prefetch(input, y, 2)
+ .prefetch(input, y, y, 2)
  .split(y, y, yi, 128)
  .parallel(y)
  .vectorize(x, vector_size * 2);

apps/camera_pipe/camera_pipe_generator.cpp

@@ -529,7 +529,7 @@ void CameraPipe::generate() {
  denoised
  .compute_at(processed, yi)
  .store_at(processed, yo)
- .prefetch(input, y, 2)
+ .prefetch(input, y, y, 2)
  .fold_storage(y, 4)
  .tile(x, y, x, y, xi, yi, 2 * vec, 2)
  .vectorize(xi)

apps/fft/fft.cpp

@@ -501,8 +501,7 @@ ComplexFunc fft2d_c2c(ComplexFunc x,
  TwiddleFactorSet twiddle_cache;
 
  // transpose the input to the FFT.
- ComplexFunc xT, x_tiled;
- std::tie(xT, x_tiled) = tiled_transpose(x, N1, target, prefix);
+ auto [xT, x_tiled] = tiled_transpose(x, N1, target, prefix);
 
  // Compute the DFT of dimension 1 (originally dimension 0).
  ComplexFunc dft1T = fft_dim1(xT,
@@ -516,8 +515,7 @@ ComplexFunc fft2d_c2c(ComplexFunc x,
  &twiddle_cache);
 
  // transpose back.
- ComplexFunc dft1, dft1_tiled;
- std::tie(dft1, dft1_tiled) = tiled_transpose(dft1T, N0, target, prefix);
+ auto [dft1, dft1_tiled] = tiled_transpose(dft1T, N0, target, prefix);
 
  // Compute the DFT of dimension 1.
  ComplexFunc dft = fft_dim1(dft1,
@@ -783,8 +781,7 @@ ComplexFunc fft2d_r2c(Func r,
  int zipped_extent0 = std::min((N1 + 1) / 2, zip_width);
 
  // transpose so we can FFT dimension 0 (by making it dimension 1).
- ComplexFunc unzippedT, unzippedT_tiled;
- std::tie(unzippedT, unzippedT_tiled) = tiled_transpose(zipped_0, zipped_extent0, target, prefix);
+ auto [unzippedT, unzippedT_tiled] = tiled_transpose(zipped_0, zipped_extent0, target, prefix);
 
  // DFT down the columns again (the rows of the original).
  ComplexFunc dftT = fft_dim1(unzippedT,
@@ -949,8 +946,7 @@ Func fft2d_c2r(ComplexFunc c,
  }
 
  // transpose the input.
- ComplexFunc cT, cT_tiled;
- std::tie(cT, cT_tiled) =
+ auto [cT, cT_tiled] =
  tiled_transpose(c_zipped, zipped_extent0, target, prefix);
 
  // Take the inverse DFT of the columns (rows in the final result).
@@ -971,8 +967,7 @@ Func fft2d_c2r(ComplexFunc c,
  }
 
  // transpose so we can take the DFT of the columns again.
- ComplexFunc dft0, dft0_tiled;
- std::tie(dft0, dft0_tiled) = tiled_transpose(dft0T, zip_width, target, prefix, true);
+ auto [dft0, dft0_tiled] = tiled_transpose(dft0T, zip_width, target, prefix, true);
 
  // Unzip the DC and Nyquist DFTs.
  ComplexFunc dft0_unzipped("dft0_unzipped");

apps/hannk/CMakeLists.txt

@@ -20,9 +20,13 @@ find_package(Halide REQUIRED)
 
 # Set up the version of TFLite we expect
 set(TFLITE_VERSION_MAJOR "2" CACHE STRING "Major version of TFLite to assume")
-set(TFLITE_VERSION_MINOR "5" CACHE STRING "Minor version of TFLite to assume")
+set(TFLITE_VERSION_MINOR "6" CACHE STRING "Minor version of TFLite to assume")
 set(TFLITE_VERSION_PATCH "0" CACHE STRING "Patch version of TFLite to assume")
 
+add_compile_definitions(TFLITE_VERSION_MAJOR=${TFLITE_VERSION_MAJOR})
+add_compile_definitions(TFLITE_VERSION_MINOR=${TFLITE_VERSION_MINOR})
+add_compile_definitions(TFLITE_VERSION_PATCH=${TFLITE_VERSION_PATCH})
+
 set(TFLITE_VERSION "${TFLITE_VERSION_MAJOR}.${TFLITE_VERSION_MINOR}.${TFLITE_VERSION_PATCH}")
 
 add_subdirectory(delegate)

apps/hannk/Makefile

@@ -56,7 +56,7 @@ clean:
 # ---------------------- TFLite glue
 
 TFLITE_VERSION_MAJOR ?= 2
-TFLITE_VERSION_MINOR ?= 5
+TFLITE_VERSION_MINOR ?= 6
 TFLITE_VERSION_PATCH ?= 0
 
 TFLITE_VERSION = $(TFLITE_VERSION_MAJOR).$(TFLITE_VERSION_MINOR).$(TFLITE_VERSION_PATCH)
@@ -136,29 +136,37 @@ $(BIN)/%/halide/average_pool_uint8.a: $(GENERATOR_BIN)/pool.generator
  @mkdir -p $(@D)
  $< -g AveragePool -f hannk::average_pool_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
-$(BIN)/%/halide/conv_uint8.a: $(GENERATOR_BIN)/conv.generator
+$(BIN)/%/halide/conv_u8_u8_u8.a: $(GENERATOR_BIN)/conv.generator
  @mkdir -p $(@D)
- $< -g Conv -f hannk::conv_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+ $< -g Conv output.type=uint8 -f hannk::conv_u8_u8_u8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
-$(BIN)/%/halide/conv_r16_uint8.a: $(GENERATOR_BIN)/conv.generator
+$(BIN)/%/halide/conv_u8_u8_i16.a: $(GENERATOR_BIN)/conv.generator
  @mkdir -p $(@D)
- $< -g Conv unroll_reduction=16 -f hannk::conv_r16_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+ $< -g Conv output.type=int16 -f hannk::conv_u8_u8_i16 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
-$(BIN)/%/halide/copy_uint8_uint8.a: $(GENERATOR_BIN)/copy.generator
+$(BIN)/%/halide/conv_r16_u8_u8_u8.a: $(GENERATOR_BIN)/conv.generator
  @mkdir -p $(@D)
- $< -g Copy input.type=uint8 output.type=uint8 -f hannk::copy_uint8_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-no_bounds_query-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+ $< -g Conv unroll_reduction=16 output.type=uint8  -f hannk::conv_r16_u8_u8_u8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
-$(BIN)/%/halide/depthwise_conv_uint8.a: $(GENERATOR_BIN)/depthwise_conv.generator
+$(BIN)/%/halide/conv_r16_u8_u8_i16.a: $(GENERATOR_BIN)/conv.generator
+ @mkdir -p $(@D)
+ $< -g Conv unroll_reduction=16 output.type=int16 -f hannk::conv_r16_u8_u8_i16 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+
+$(BIN)/%/halide/copy_uint8_uint8.a: $(GENERATOR_BIN)/copy.generator
  @mkdir -p $(@D)
- $< -g DepthwiseConv -f hannk::depthwise_conv_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+ $< -g Copy input.type=uint8 output.type=uint8 -f hannk::copy_uint8_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-no_bounds_query-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
 $(BIN)/%/halide/depthwise_conv_broadcast_uint8.a: $(GENERATOR_BIN)/depthwise_conv.generator
  @mkdir -p $(@D)
  $< -g DepthwiseConv inv_depth_multiplier=0 -f hannk::depthwise_conv_broadcast_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
-$(BIN)/%/halide/depthwise_conv_dm1_uint8.a: $(GENERATOR_BIN)/depthwise_conv.generator
+$(BIN)/%/halide/depthwise_conv_uint8.a: $(GENERATOR_BIN)/depthwise_conv.generator
+ @mkdir -p $(@D)
+ $< -g DepthwiseConv inv_depth_multiplier=1 -f hannk::depthwise_conv_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+
+$(BIN)/%/halide/depthwise_conv_shallow_uint8.a: $(GENERATOR_BIN)/depthwise_conv.generator
  @mkdir -p $(@D)
- $< -g DepthwiseConv inv_depth_multiplier=1 -f hannk::depthwise_conv_dm1_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+ $< -g DepthwiseConv inv_depth_multiplier=1 shallow=true -f hannk::depthwise_conv_shallow_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
 $(BIN)/%/halide/elementwise_5xuint8_1xuint8.a: $(GENERATOR_BIN)/elementwise.generator
  @mkdir -p $(@D)
@@ -172,14 +180,6 @@ $(BIN)/%/halide/fill_uint8.a: $(GENERATOR_BIN)/fill.generator
  @mkdir -p $(@D)
  $< -g Fill -f hannk::fill_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-no_asserts-no_bounds_query-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
-$(BIN)/%/halide/fully_connected_uint8_uint8.a: $(GENERATOR_BIN)/fully_connected.generator
- @mkdir -p $(@D)
- $< -g FullyConnected output.type=uint8 -f hannk::fully_connected_uint8_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
-
-$(BIN)/%/halide/fully_connected_uint8_int16.a: $(GENERATOR_BIN)/fully_connected.generator
- @mkdir -p $(@D)
- $< -g FullyConnected output.type=int16 -f hannk::fully_connected_uint8_int16 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
-
 $(BIN)/%/halide/l2_normalization_uint8.a: $(GENERATOR_BIN)/normalizations.generator
  @mkdir -p $(@D)
  $< -g L2Normalization -f hannk::l2_normalization_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-no_bounds_query-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
@@ -204,35 +204,40 @@ $(BIN)/%/halide/tile_conv_filter_uint8.a: $(GENERATOR_BIN)/conv.generator
  @mkdir -p $(@D)
  $< -g TileConvFilter -f hannk::tile_conv_filter_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
 
+$(BIN)/%/halide/upsample_channels_uint8.a: $(GENERATOR_BIN)/depthwise_conv.generator
+ @mkdir -p $(@D)
+ $< -g UpsampleChannels -f hannk::upsample_channels_uint8 -o $(BIN)/$*/halide target=$(HL_TARGET)-no_runtime-c_plus_plus_name_mangling -e static_library,assembly,stmt,c_header,llvm_assembly
+
 $(BIN)/%/halide/runtime.a: $(GENERATOR_BIN)/fill.generator
  @mkdir -p $(@D)
  $< -r runtime -o $(BIN)/$*/halide target=$(HL_TARGET)
 
 OPS_HALIDE = \
  $(BIN)/%/halide/add_uint8_uint8.a \
  $(BIN)/%/halide/average_pool_uint8.a \
- $(BIN)/%/halide/conv_uint8.a \
+ $(BIN)/%/halide/conv_u8_u8_u8.a \
+ $(BIN)/%/halide/conv_u8_u8_i16.a \
  $(BIN)/%/halide/copy_uint8_uint8.a \
  $(BIN)/%/halide/depthwise_conv_uint8.a \
  $(BIN)/%/halide/depthwise_conv_broadcast_uint8.a \
- $(BIN)/%/halide/depthwise_conv_dm1_uint8.a \
+ $(BIN)/%/halide/depthwise_conv_shallow_uint8.a \
  $(BIN)/%/halide/elementwise_5xuint8_1xuint8.a \
  $(BIN)/%/halide/elementwise_5xint16_1xuint8int16.a \
  $(BIN)/%/halide/fill_uint8.a \
- $(BIN)/%/halide/fully_connected_uint8_uint8.a \
- $(BIN)/%/halide/fully_connected_uint8_int16.a \
  $(BIN)/%/halide/l2_normalization_uint8.a \
  $(BIN)/%/halide/max_pool_uint8.a \
  $(BIN)/%/halide/mean_uint8.a \
  $(BIN)/%/halide/mul_uint8_uint8_uint8.a \
  $(BIN)/%/halide/softmax_uint8.a \
  $(BIN)/%/halide/tile_conv_filter_uint8.a \
+ $(BIN)/%/halide/upsample_channels_uint8.a \
  $(BIN)/%/halide/runtime.a
 
 OPS_CXXFLAGS = -I$(BIN)/$*
 
 ifneq (,$(findstring arm_dot_prod,$(HL_TARGET)))
-OPS_HALIDE += $(BIN)/%/halide/conv_r16_uint8.a
+OPS_HALIDE += $(BIN)/%/halide/conv_r16_u8_u8_u8.a
+OPS_HALIDE += $(BIN)/%/halide/conv_r16_u8_u8_i16.a
 OPS_CXXFLAGS += -DCONV_R16
 endif
 
@@ -377,7 +382,7 @@ HANNK_INTERNAL_DELEGATE_DEPS = \
 
 $(BIN)/%/$(BENCHMARK_OUT): benchmark.cpp $(INTERPRETER_DEPS) $(TFLITE_PARSER_DEPS) $(UTIL_DEPS) util/file_util.h
  @mkdir -p $(@D)
- $(CXX-$*) $(CXXFLAGS-$*) $(BENCHMARK_HEXAGON_FLAGS) $(APP_CXXFLAGS) -Wl,--whole-archive $(filter %.cpp %.o %.a,$^) -Wl,--no-whole-archive -o $@ $(LDFLAGS-$*)
+ $(CXX-$*) $(CXXFLAGS-$*) $(BENCHMARK_HEXAGON_FLAGS) $(APP_CXXFLAGS) $(filter %.cpp %.o %.a,$^) -o $@ $(LDFLAGS-$*)
 
 
 # To build for Android, use `HL_TARGET=arm-64-android make compare_vs_tflite`

apps/hannk/README.md

@@ -10,7 +10,7 @@ All of the [TensorFlow hosted models](https://www.tensorflow.org/lite/guide/host
 are working and producing good performance.
 
 ### Benchmarks
-The comparison data below was produced with TensorFlow v.2.4.0 (the latest release as of this writing):
+The comparison data below was produced with TensorFlow v.2.5.0 (the latest release as of this writing):
 
 x86 OSX laptop w/ AVX2:
 
@@ -28,14 +28,14 @@ Qualcomm Snapdragon 855 A76 core (Pixel 4):
 
 | Network | TFlite (ms)|Halide (ms)| Speedup
 | ---- | ---- | ---- | ---- |
-| inception_v1_224_quant | 24.7 | 25.9 | 0.95 |
-| inception_v2_224_quant | 49.8 | 34.2 | 1.46 |
-| inception_v3_quant | 97 | 89 | 1.09 |
-| inception_v4_299_quant | 198 | 186.7 | 1.06 |
-| mobilenet_v1_0.25_128_quant | 0.97 | 0.72 | 1.34 |
+| inception_v1_224_quant | 24.7 | 25.0 | 0.99 |
+| inception_v2_224_quant | 49.8 | 33.5 | 1.49 |
+| inception_v3_quant | 97 | 87.6 | 1.11 |
+| inception_v4_299_quant | 198 | 183.4 | 1.09 |
+| mobilenet_v1_0.25_128_quant | 0.97 | 0.63 | 1.54 |
 | mobilenet_v1_1.0_128_quant |4.64 | 4.44 | 1.05 |
 | mobilenet_v1_1.0_224_quant | 12.9 | 11.6 | 1.11 |
-| mobilenet_v2_1.0_224_quant | 11.8 | 9.89 | 1.19 |
+| mobilenet_v2_1.0_224_quant | 11.8 | 9.72 | 1.21 |
 
 ### Planned but still TODO
 - More op support

apps/hannk/benchmark.cpp

@@ -21,7 +21,7 @@ void run_benchmark(const std::string &filename, const InterpreterOptions &option
  std::vector<char> buffer = read_entire_file(filename);
  std::unique_ptr<OpGroup> model = parse_tflite_model_from_buffer(buffer.data());
 
- if (options.verbose) {
+ if (options.verbosity >= 1) {
  model->dump(std::cout);
  }
 
@@ -49,7 +49,7 @@ __attribute__((visibility("default"))) int main(int argc, char **argv) {
 
  for (int i = 1; i < argc; i++) {
  if (!strcmp(argv[i], "--verbose")) {
- options.verbose = true;
+ options.verbosity = 1;
  continue;
  }
  if (!strcmp(argv[i], "--trace")) {
@@ -62,7 +62,7 @@ __attribute__((visibility("default"))) int main(int argc, char **argv) {
  }
  }
 
- if (options.verbose && options.trace) {
+ if (options.verbosity > 0 && options.trace) {
  HLOG(ERROR) << "You cannot specify --trace and --verbose at the same time.\n";
  exit(-1);
  }

apps/hannk/compare_vs_tflite.cpp

@@ -25,9 +25,6 @@ int main(int argc, char **argv) {
  runner.run(f);
  halide_profiler_report(nullptr);
  halide_profiler_reset();
- std::cout << "\n";
  }
-
- std::cout << "Done!\n";
  return 0;
 }