Merge pull request #845 from xmos/beta-float

Add beta float support

README.md

@@ -81,10 +81,10 @@ xf.convert("example_int8_model.tflite", "xcore_optimised_int8_model.tflite", {
 })
 ```
 
-To create a parameters file and a tflite model suitable for loading to flash, use the "xcore-flash-image-file" option.
+To create a parameters file and a tflite model suitable for loading to flash, use the "xcore-weights-file" option.
 ```python
 xf.convert("example_int8_model.tflite", "xcore_optimised_int8_flash_model.tflite", {
-    "xcore-flash-image-file ": "./xcore_params.params",
+    "xcore-weights-file ": "./xcore_params.params",
 })
 ```
 

docs/rst/options.rst

@@ -70,7 +70,7 @@ Name of the file where to place the optimized TFLITE model
 Number of threads to translate for (max=5). Defaults to 1.
 
 
-``xcore-flash-image-file filename``
+``xcore-weights-file filename``
 +++++++++++++++++++++++++++++++++++
 
 File to place the learned parameters in. If this option is not specified,
@@ -86,7 +86,7 @@ be slower but allows large numbers of learned parameters to be used.
 Sets a threshold under which to not place learned parameters in flash. The
 default is set to 96 bytes. If less than 96 bytes, the overhead of lowering to flash is 
 more than the benefit gained. This option is only meaningful if
-``xcore-flash-image-file`` has been used. You can experiment with this
+``xcore-weights-file`` has been used. You can experiment with this
 parameter to get a different trade-off between speed and memory requirements.
 
 ``xcore-reduce-memory true|false``

examples/app_flash_single_model/README.rst

@@ -5,7 +5,7 @@ Please consult `here <../../docs/rst/flow.rst>`_ on how to install the tools.
 
 In order to compile and run this example follow these steps::
 
-  xcore-opt --xcore-flash-image-file=model.params vww_quant.tflite -o model.tflite
+  xcore-opt --xcore-weights-file=model.params vww_quant.tflite -o model.tflite
   mv model.tflite.cpp model.tflite.h src
   xmake
   python -c 'from xmos_ai_tools import xformer as xf; xf.generate_flash(

examples/app_flash_two_models/README.rst

@@ -9,10 +9,10 @@ sets of learned parameters into a single flash image.
 
 In order to compile and run this example follow these steps::
 
-  xcore-opt --xcore-flash-image-file=model1.params \
+  xcore-opt --xcore-weights-file=model1.params \
             --xcore-naming-prefix=model1_ \
             vww_quant1.tflite -o model1.tflite
-  xcore-opt --xcore-flash-image-file=model2.params \
+  xcore-opt --xcore-weights-file=model2.params \
             --xcore-naming-prefix=model2_ \
             vww_quant2.tflite -o model2.tflite
   mv model1.tflite.cpp model1.tflite.h src

examples/app_flash_two_models_one_arena/README.rst

@@ -20,10 +20,10 @@ The differences with ``app_flash_two_models`` example are minimal:
 
 In order to compile and run this example follow these steps::
 
-  xcore-opt --xcore-flash-image-file=model1.params \
+  xcore-opt --xcore-weights-file=model1.params \
             --xcore-naming-prefix=model1_ \
             vww_quant1.tflite -o model1.tflite
-  xcore-opt --xcore-flash-image-file=model2.params \
+  xcore-opt --xcore-weights-file=model2.params \
             --xcore-naming-prefix=model2_ \
             vww_quant2.tflite -o model2.tflite
   mv model1.tflite.cpp model1.tflite.h src

examples/app_mobilenetv2/obtain_and_optimize_mobilenetv2.py

@@ -33,7 +33,7 @@
     TFLITE_MODEL_PATH,
     OPT_MODEL_PATH,
     {
-        "xcore-flash-image-file": OPT_PARAMS_PATH,
+        "xcore-weights-file": OPT_PARAMS_PATH,
         "xcore-thread-count": "5",
         "xcore-naming-prefix": NAMING_PREFIX,
         "xcore-op-split-tensor-arena": "True",

examples/app_yolov8_classification/obtain_and_optimize_yolov8_cls.py

@@ -29,7 +29,7 @@
     TFLITE_MODEL_PATH,
     OPT_MODEL_PATH,
     {
-        "xcore-flash-image-file": OPT_PARAMS_PATH,
+        "xcore-weights-file": OPT_PARAMS_PATH,
         "xcore-thread-count": "5",
         "xcore-naming-prefix": NAMING_PREFIX,
     },

python/setup.py

@@ -76,7 +76,7 @@ def finalize_options(self):
 
 # add device lib and headers as package data
 device_files = {
-    root.replace(os.sep, "."): ["*.h", "*.a", "*.make", "*.cmake"]
+    root.replace(os.sep, "."): ["*.h", "*.a", "*.lib", "*.make", "*.cmake"]
     for root, d, f in os.walk(os.path.join("xmos_ai_tools", "runtime"))
 }
 
@@ -107,17 +107,17 @@ def finalize_options(self):
         "Intended Audience :: Science/Research",
         "Programming Language :: Python :: 3",
         "Programming Language :: Python :: 3 :: Only",
-        "Programming Language :: Python :: 3.7",
         "Programming Language :: Python :: 3.8",
         "Programming Language :: Python :: 3.9",
+        "Programming Language :: Python :: 3.10",
         "Topic :: Scientific/Engineering",
         "Topic :: Scientific/Engineering :: Mathematics",
         "Topic :: Scientific/Engineering :: Artificial Intelligence",
         "Topic :: Software Development",
         "Topic :: Software Development :: Libraries",
         "Topic :: Software Development :: Libraries :: Python Modules",
     ],
-    python_requires=">=3.7",
+    python_requires=">=3.8",
     packages=find_namespace_packages(),
     install_requires=REQUIRED_PACKAGES,
     package_data=package_files,

python/xmos_ai_tools/runtime/buildfiles/aitoolslib.cmake

@@ -5,5 +5,9 @@ set(XMOS_AITOOLSLIB_DEFINITIONS
 		"NO_INTERPRETER"
     )
 
-set(XMOS_AITOOLSLIB_LIBRARIES "${CMAKE_CURRENT_LIST_DIR}/../lib/libxtflitemicro.a")
+if(NOT ${CMAKE_SYSTEM_PROCESSOR} STREQUAL XCORE_XS3A)
+	set(XMOS_AITOOLSLIB_LIBRARIES "${CMAKE_CURRENT_LIST_DIR}/../lib/libhost_xtflitemicro.a")
+else()
+	set(XMOS_AITOOLSLIB_LIBRARIES "${CMAKE_CURRENT_LIST_DIR}/../lib/libxtflitemicro.a")
+endif()
 set(XMOS_AITOOLSLIB_INCLUDES "${CMAKE_CURRENT_LIST_DIR}/../include")

python/xmos_ai_tools/xinterpreters/CMakeLists.txt

@@ -3,38 +3,68 @@ cmake_minimum_required(VERSION 3.14)
 set(CMAKE_BUILD_TYPE "Release")
 project(xtflm_python VERSION 1.0.1)
 
-# set host build
-set(X86 ON)
-
 # This variable is ignored on platforms other than Apple
-set(CMAKE_OSX_SYSROOT /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk) 
-
-set(CMAKE_CXX_FLAGS "-std=c++11" CACHE STRING "C++ Compiler Base Flags" FORCE)
-
+set(CMAKE_OSX_SYSROOT /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk)
 
 #**********************
 # Build flags
 #**********************
 if (${CMAKE_SYSTEM_NAME} MATCHES "Windows")
-  set(BUILD_FLAGS "/DTF_LITE_DISABLE_X86_NEON /D__xtflm_conf_h_exists__ /O2 /DNN_USE_REF")
+  set(BUILD_FLAGS "/O2")
 else()
-  set(BUILD_FLAGS "-g -DTF_LITE_DISABLE_X86_NEON -D__xtflm_conf_h_exists__ -O3 -DNN_USE_REF")
+  set(BUILD_FLAGS
+              "-g"
+              "-O3")
 endif()
 
 
 if(DEFINED ENV{CMAKE_ENABLE_DARWIN_TARGET_ARM64})
   set(BUILD_FLAGS "${BUILD_FLAGS} -target arm64-apple-macos11")
 endif()
 
-set(CMAKE_CXX_STANDARD 14)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_FLAGS ${BUILD_FLAGS})
-set(CMAKE_C_FLAGS "${BUILD_FLAGS} -std=c99")
+set(TOP_DIR
+  "${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/lib_tflite_micro")
+
+include(${TOP_DIR}/cmakefiles/xtflm.cmake)
 
 #**********************
-# Build shared library
+# Build host library
 #**********************
+add_library(host_xtflitemicro STATIC)
+set(DEFINTIONS
+    "__xtflm_conf_h_exists__"
+    "NO_INTERPRETER"
+    "TF_LITE_STATIC_MEMORY"
+    "TF_LITE_DISABLE_X86_NEON"
+    "TF_LITE_STRIP_ERROR_STRINGS"
+    "NN_USE_REF"
+)
+list(APPEND DEFINTIONS "FLATBUFFERS_LOCALE_INDEPENDENT=0")
+target_compile_features(host_xtflitemicro PUBLIC cxx_std_17)
+
+target_sources(host_xtflitemicro
+    PRIVATE ${TFLM_KERNEL_SOURCES}
+    PRIVATE ${TFLITE_SOURCES}
+    PRIVATE ${NN_SOURCES}
+    PRIVATE ${XTFLIB_KERNEL_SOURCES}
+)
+
+target_compile_options(host_xtflitemicro PRIVATE ${BUILD_FLAGS})
+target_link_options(host_xtflitemicro PRIVATE ${BUILD_FLAGS})
+target_compile_definitions(host_xtflitemicro PUBLIC
+    ${DEFINTIONS}
+)
 
+target_include_directories(host_xtflitemicro
+    PRIVATE ${ALL_INCLUDES}
+)
+
+set(INSTALL_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../runtime/lib")
+install(TARGETS host_xtflitemicro DESTINATION ${INSTALL_DIR})
+
+#**********************
+# Build shared library
+#**********************
 add_library(xtflm_python SHARED)
 set_target_properties(xtflm_python PROPERTIES VERSION ${PROJECT_VERSION})
 set_target_properties(xtflm_python PROPERTIES PREFIX "")
@@ -44,10 +74,15 @@ elseif (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
   target_link_libraries(xtflm_python stdc++ m pthread)
 endif()
 
-set(TOP_DIR
-  "${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/lib_tflite_micro")
-
-include(${TOP_DIR}/cmakefiles/xtflm.cmake)
+set(DEFINTIONS
+    "__xtflm_conf_h_exists__"
+    "TF_LITE_DISABLE_X86_NEON"
+    "NN_USE_REF"
+)
+target_compile_definitions(xtflm_python PUBLIC ${DEFINTIONS})
+target_compile_features(xtflm_python PUBLIC cxx_std_14)
+target_compile_options(xtflm_python PRIVATE ${BUILD_FLAGS})
+target_link_options(xtflm_python PRIVATE ${BUILD_FLAGS})
 
 target_sources(xtflm_python
   PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}/src/dll_interpreter.cc"

xformer/IR/XCoreOps.td

@@ -132,6 +132,31 @@ def XC_MulOp : XC_Op<"mul", [Pure]> {
   let results = (outs TensorOf<[QI8]> : $output);
 }
 
+def XC_Beta_ActivationF32Op : XC_Op<"beta_activationf32", [Pure]> {
+  let summary = "Beta ActivationF32 op";
+
+  let description = [{Beta ActivationF32 op.}];
+
+  let arguments = (ins
+    TensorOf<[F32]>:$input,
+    I32Attr:$type
+  );
+
+  let results = (outs TensorOf<[F32]> : $output);
+}
+
+def XC_Beta_ConcatF32Op : XC_Op<"beta_concatf32", [Pure]> {
+  let summary = "Beta ConcatF32 op";
+
+  let description = [{Beta ConcatF32 op.}];
+
+  let arguments = (ins
+    Variadic<TensorOf<[F32]>>:$input
+  );
+
+  let results = (outs TensorOf<[F32]> : $output);
+}
+
 def XC_Beta_ConvF32Op : XC_Op<"beta_convf32", [Pure]> {
   let summary = "Beta ConvF32 op";
 
@@ -248,7 +273,7 @@ def XC_LookupOp : XC_Op<"lookup", [Pure]> {
 
   let description = [{Lookup table op.}];
 
-  let arguments = (ins TensorOf<[QI8]> : $input, TensorOf<[I8]> : $lut, I32Attr : $thread_count);
+  let arguments = (ins TensorOf<[QI8]> : $input, TensorOf<[I8]> : $lut);
 
   let results = (outs TensorOf<[QI8]> : $output);
 }

xformer/Test/loadconstantop.mlir

@@ -1,6 +1,6 @@
-// RUN: xcore-opt --mlir-io %s --xcore-apply-loadconstantop-patterns --xcore-flash-image-file=/dev/null --xcore-load-externally-if-larger=0 | FileCheck %s
+// RUN: xcore-opt --mlir-io %s --xcore-apply-loadconstantop-patterns --xcore-weights-file=/dev/null --xcore-load-externally-if-larger=0 | FileCheck %s
 
-// RUN: xcore-opt --mlir-io %s --xcore-apply-loadconstantop-patterns --xcore-flash-image-file=/dev/null --xcore-load-externally-if-larger=16 | FileCheck %s -check-prefix=LARGER-CHECK
+// RUN: xcore-opt --mlir-io %s --xcore-apply-loadconstantop-patterns --xcore-weights-file=/dev/null --xcore-load-externally-if-larger=16 | FileCheck %s -check-prefix=LARGER-CHECK
 
 // CHECK-LABEL: valid
 func.func @valid(%arg0: tensor<?x4x8x1x!quant.uniform<i8:f32, 0.0078160231932997704>>) -> tensor<?x32x!quant.uniform<i8:f32, 0.037329975515604019:-13>> attributes {tf.entry_function = {inputs = "flatten_input", outputs = "Identity"}} {

xformer/Test/loadflashop.mlir

@@ -1,4 +1,4 @@
-// RUN: xcore-opt --mlir-io %s --xcore-write-flash-image --xcore-flash-image-file=/dev/null | FileCheck %s
+// RUN: xcore-opt --mlir-io %s --xcore-write-flash-image --xcore-weights-file=/dev/null | FileCheck %s
 
 // CHECK-LABEL: valid
 func.func @valid(%arg0: tensor<?x4x8x1x!quant.uniform<i8:f32, 0.0078160231932997704>>) -> tensor<?x32x!quant.uniform<i8:f32, 0.037329975515604019:-13>> attributes {tf.entry_function = {inputs = "flatten_input", outputs = "Identity"}} {

xformer/Transforms/ApplyLoadConstantOpPatterns.cpp

@@ -30,7 +30,12 @@ struct ApplyLoadConstantOpPatterns
   void runOnOperation() override;
 };
 
+static int totalSize_ = 0;
+
 bool shouldBeLoadedExternally(Attribute values) {
+  if (totalSize_ > maxLoadExternalSizeOption) {
+    return false;
+  }
   // values might be UnitAttr or BoolAttr which are too small to be loaded
   // externally anyway
   auto totalSizeInBits = 0;
@@ -40,6 +45,7 @@ bool shouldBeLoadedExternally(Attribute values) {
         (valuesAttr.getNumElements() *
          valuesAttr.getType().getElementType().getIntOrFloatBitWidth());
   }
+  totalSize_ += totalSizeInBits / CHAR_BIT;
   return totalSizeInBits / CHAR_BIT > loadExternallyIfLargerOption;
 }
 
@@ -56,7 +62,7 @@ bool isNotUsedByLoadConstantOp(Value result) {
 
 void ApplyLoadConstantOpPatterns::runOnOperation() {
   func::FuncOp f = getOperation();
-  if (flashImageFilenameOption.empty()) {
+  if (weightsFilenameOption.empty()) {
     f.emitError("Flash image file option should be provided to run this pass!");
     signalPassFailure();
     return;

xformer/Transforms/ApplyXCPatterns.cpp

@@ -30,6 +30,8 @@ struct ApplyXCPatterns
   void runOnOperation() override;
 };
 
+bool isBetaFloatEnabled() { return enableBetaFloatOption; }
+
 StringAttr getPaddingPlan(PatternRewriter &rewriter, TFL::PadOp padOp) {
   DenseIntElementsAttr paddingAttr;
   if (!matchPattern(padOp.getPadding(), m_Constant(&paddingAttr))) {
@@ -83,8 +85,17 @@ IntegerAttr getPadValue(PatternRewriter &rewriter, Value inputVal) {
   return rewriter.getI32IntegerAttr(padValue);
 }
 
-IntegerAttr getThreadCount(PatternRewriter &rewriter) {
-  return rewriter.getI32IntegerAttr(threadCountOption);
+IntegerAttr getActivationType(PatternRewriter &rewriter, Operation *op) {
+  // TODO: Refactor to use shared header file for enum
+  if (isa<TFL::EluOp>(op)) {
+    return rewriter.getI32IntegerAttr(0);
+  } else if (isa<TFL::LogisticOp>(op)) {
+    return rewriter.getI32IntegerAttr(1);
+  } else if (isa<TFL::TanhOp>(op)) {
+    return rewriter.getI32IntegerAttr(2);
+  } else {
+    llvm_unreachable("Unsupported op!");
+  }
 }
 
 DenseElementsAttr getLookupTable(PatternRewriter &rewriter, Operation *op) {

xformer/Transforms/ConvPatterns.td

@@ -42,26 +42,26 @@ Pat<(TFL_DepthwiseConv2DOp: $output TensorOf<[QI8]>:$input, TensorOf<[QI8]>:$f,
               (IsConstOp $f),
               ]>;
 
-// TODO: Special case, we only optimize conv with filter width 5, filter height
-// 2, and stride height 3
+// Special case, we only optimize conv with filter width 3, filter height
+// 2, and stride height 2
 def Hasfw5fh2
     : Constraint<CPred<"$0.getType().cast<ShapedType>().getRank() == 4 && "
-                       "$0.getType().cast<ShapedType>().getDimSize(1) == 5 && "
+                       "$0.getType().cast<ShapedType>().getDimSize(1) == 3 && "
                        "$0.getType().cast<ShapedType>().getDimSize(2) == 2">>;
 
 // F32 TFL_Conv2D() -> XC_Beta_ConvF32()
 def :
-Pat<(TFL_Conv2DOp: $output TensorOf<[F32]>:$input, TensorOf<[F32]>:$f, TensorOf<[F32]>:$b, $dh, $dw, $faf, $wf, ConstantAttr<I32Attr, "3">, ConstantAttr<I32Attr, "1">),
+Pat<(TFL_Conv2DOp: $output TensorOf<[F32]>:$input, TensorOf<[F32]>:$f, TensorOf<[F32]>:$b, $dh, $dw, $faf, $wf, ConstantAttr<I32Attr, "2">, ConstantAttr<I32Attr, "1">),
           (XC_Beta_ConvF32Op $input, $f, $b),
-          [(Hasfw5fh2 $f)]>;
+          [(Hasfw5fh2 $f), (isBetaFloatEnabled)]>;
 
-// F32 TFL_TransposeConv2D() -> XC_Beta_TransposeConvF32()
+// // F32 TFL_TransposeConv2D() -> XC_Beta_TransposeConvF32()
 def :
-Pat<(TFL_TransposeConvOp: $output $outshape, TensorOf<[F32]>:$f, TensorOf<[F32]>:$input, TensorOf<[F32]>:$b, $wf, ConstantAttr<I32Attr, "3">, ConstantAttr<I32Attr, "1">, $faf),
+Pat<(TFL_TransposeConvOp: $output $outshape, TensorOf<[F32]>:$f, TensorOf<[F32]>:$input, TensorOf<[F32]>:$b, $wf, ConstantAttr<I32Attr, "2">, ConstantAttr<I32Attr, "1">, $faf),
           (XC_Beta_TransposeConvF32Op $input, $f, $b),
-          [(Hasfw5fh2 $f)]>;
+          [(Hasfw5fh2 $f), (isBetaFloatEnabled)]>;
 
-// F32 TFL_FullyConnected() -> XC_Beta_FcF32()
+// // F32 TFL_FullyConnected() -> XC_Beta_FcF32()
 def :
 Pat<(TFL_FullyConnectedOp: $output TensorOf<[F32]>:$input, TensorOf<[F32]>:$f, $b, $faf, $wf, $knd, $aqi),
-          (XC_Beta_FcF32Op $input, $f)>;
+          (XC_Beta_FcF32Op $input, $f), [(isBetaFloatEnabled)]>;

xformer/Transforms/Options.h

@@ -9,9 +9,12 @@
 namespace mlir {
 namespace xcore {
 
+extern llvm::cl::opt<bool> enableBetaFloatOption;
 extern llvm::cl::opt<unsigned> threadCountOption;
-extern llvm::cl::opt<std::string> flashImageFilenameOption;
+extern llvm::cl::opt<std::string> weightsFilenameOption;
 extern llvm::cl::opt<unsigned> loadExternallyIfLargerOption;
+extern llvm::cl::opt<bool> tileLoadOption;
+extern llvm::cl::opt<unsigned> maxLoadExternalSizeOption;
 extern llvm::cl::opt<double> convQuantErrorThresholdOption;
 extern llvm::cl::opt<bool> convForceErrorCheckOption;
 extern llvm::cl::opt<unsigned> convMultiplierFactorOption;