ADD: update

.gitignore

@@ -2,8 +2,6 @@
 .conda-env/
 .vscode/
 
-example/mnist/data/
-
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
@@ -30,5 +28,7 @@ share/python-wheels/
 MANIFEST
 
 
-example/llama2/checkpoints/stories15M.bin
+examples/mnist/data/
+
+examples/llama2/checkpoints/stories15M.bin
 

CMakeLists.txt

@@ -1,28 +1,48 @@
+########################################################################################################################
+# file:  CMakeLists.txt
+#
+# usage: 
+#   Edit "VARIABLES"-section to suit project requirements.
+#   Build instructions:
+#     cmake . -D CMAKE_TOOLCHAIN_FILE=./riscv-gcc.cmake -S ./ -B ./build/ -G "Unix Makefiles" -D CMAKE_BUILD_TYPE=Debug -D RVV=ON
+#     cmake --build ./build/ --target all
+#     cmake --build ./build/ --target clean
+########################################################################################################################
+cmake_minimum_required(VERSION 3.10)
+
+project(nn LANGUAGES C)
 
 # Options
 option(BUILD_SHARED_LIBS    "Build using shared libraries"      OFF     )
 option(CMAKE_BUILD_TYPE     "Build type"                        Debug   )
 option(STATIC_LINKING       "Link statically"                   TRUE    )
 
+# architecture options
 option(X86                  "Build for x86"                     OFF     )
 option(RISCV                "Build for RISC-V"                  OFF     )
 
+# accelerator options
+option(AVX                  "Use AVX implementation" OFF)
+option(RVV                  "Use RISCV vector extension implementation" OFF)
 
-add_library(target-riscv INTERFACE)
-target_compile_features(target-riscv INTERFACE c_std_11)
-target_compile_definitions(target-riscv INTERFACE RISCV)
+
+add_compile_options(-O1 -Wall -Wextra)
 
 add_library(target-x86 INTERFACE)
 target_compile_features(target-x86 INTERFACE c_std_11)
 target_compile_definitions(target-x86 INTERFACE X86)
 
-target_link_options(target-x86 INTERFACE -static)
-target_link_options(target-riscv INTERFACE -static)
-
 target_compile_options(target-x86 INTERFACE -march=native)
+# target_compile_options(target-x86 INTERFACE -mavx512fp16)
+target_link_options(target-x86 INTERFACE -static)
 target_link_options(target-x86 INTERFACE -march=native)
+target_link_options(target-x86 INTERFACE -fno-common -fno-builtin-printf)
 
 
+add_library(target-riscv INTERFACE)
+target_compile_features(target-riscv INTERFACE c_std_11)
+target_compile_definitions(target-riscv INTERFACE RISCV)
+
 set(WRAP_SPECS_FILE     "htif_wrap.specs")
 set(SPECS_FILE          "htif_nano.specs")
 set(LIBGLOSS_DIR        "$ENV{RISCV}/riscv64-unknown-elf/lib/")
@@ -31,9 +51,22 @@ set(MARCH               "rv64gcv_zfh_zvfh")
 set(MABI                "lp64d")
 set(MCMODEL             "medany")
 
+if (NOT DEFINED LINKER_SCRIPT)
+  set(LINKER_SCRIPT     ${CMAKE_SOURCE_DIR}/toolchain/htif.ld)
+endif()
+
 target_compile_options(target-riscv INTERFACE -fno-common -fno-builtin-printf)
 target_compile_options(target-riscv INTERFACE -march=${MARCH} -mabi=${MABI} -mcmodel=${MCMODEL})
-target_compile_options(target-riscv INTERFACE -Wl,-Map=output.map -specs=${SPECS_FILE} -specs=${WRAP_SPECS_FILE} -T ${CMAKE_SOURCE_DIR}/htif.ld)
-target_link_options(target-x86 INTERFACE -fno-common -fno-builtin-printf)
+target_compile_options(target-riscv INTERFACE -Wl,-Map=output.map -specs=${SPECS_FILE} -specs=${WRAP_SPECS_FILE})
+target_compile_options(target-riscv INTERFACE -T ${LINKER_SCRIPT})
+
+target_link_options(target-riscv INTERFACE -static)
 target_link_options(target-riscv INTERFACE -march=${MARCH} -mabi=${MABI} -mcmodel=${MCMODEL})
-target_link_options(target-riscv INTERFACE -Wl,-Map=output.map -specs=${SPECS_FILE} -specs=${WRAP_SPECS_FILE} -T ${CMAKE_SOURCE_DIR}/htif.ld)
+target_link_options(target-riscv INTERFACE -Wl,-Map=output.map -specs=${SPECS_FILE} -specs=${WRAP_SPECS_FILE})
+target_link_options(target-riscv INTERFACE -T ${LINKER_SCRIPT})
+
+
+add_subdirectory(nn)
+add_subdirectory(tests)
+add_subdirectory(examples)
+

docs/Tensor-Creation.md

@@ -0,0 +1,30 @@
+# Tensor Creation
+
+A set of factory functions are available for creating a tensor object. These factory functions configure the shape, data type, device and other properties of the new tensor, and optionally populate them according to specific algorithms.
+
+## Factory Functions
+
+A *factory function* is a function that produces a new tensor. There are many factory functions available, which differ in the way they initialize a new tensor before returning it. All factory functions adhere to the following general “schema”:
+
+```c
+Tensor *NN_<function-name>(<ndim>, <shape>, <datatype>, <tensor-options>)
+```
+
+### Available Factory Functions
+
+The following factory functions are available at the time of this writing:
+
+**tensor**: Returns a tensor with uninitialized values or preallocated buffer.
+
+**zeros**: Returns a tensor filled with all zeros.
+
+**ones**: Returns a tensor filled with all ones.
+
+**full**: Returns a tensor filled with a single value.
+
+**rand**: Returns a tensor filled with values drawn from a uniform distribution on [0, 1).
+
+**randint**: Returns a tensor with integers randomly drawn from an interval.
+
+**arange**: Returns a tensor with a sequence of integers.
+

examples/CMakeLists.txt

@@ -0,0 +1,3 @@
+add_subdirectory(fast-depth)
+add_subdirectory(llama2)
+add_subdirectory(stereo-block-matching)

examples/fast-depth/CMakeLists.txt

@@ -2,44 +2,22 @@ cmake_minimum_required(VERSION 3.10)
 
 project(fast-depth LANGUAGES C)
 
-
 add_executable(fast-depth main.c)
-target_include_directories(fast-depth PUBLIC inc ../utility)
+
+target_include_directories(fast-depth PUBLIC ./)
+target_include_directories(fast-depth PUBLIC ../utility)
+
 
 target_compile_features(fast-depth INTERFACE c_std_11)
 
 if (X86)
-    message("fast-depth: building for x86")
-    target_link_libraries(fast-depth PUBLIC target-x86)
-
+  message(STATUS "fast-depth: Building for x86")
+  target_link_libraries(fast-depth PUBLIC target-x86)
 elseif (RISCV)
-    message("fast-depth: building for RISC-V")
-    # CMake toolchain definition for RISC-V GCC toolchain
-    set(CMAKE_SYSTEM_NAME "Generic" CACHE STRING "")
-    set(CMAKE_SYSTEM_PROCESSOR "riscv" CACHE STRING "")
-
-    set(TOOLCHAIN_PREFIX    "riscv64-unknown-elf-")
-
-    set(CMAKE_C_COMPILER    "${TOOLCHAIN_PREFIX}gcc")
-    set(CMAKE_ASM_COMPILER  "${TOOLCHAIN_PREFIX}gcc")
-    set(CMAKE_CXX_COMPILER  "${TOOLCHAIN_PREFIX}g++")
-    set(CMAKE_AR            "${TOOLCHAIN_PREFIX}ar")
-    set(CMAKE_LINKER        "{TOOLCHAIN_PREFIX}ld")
-    set(CMAKE_OBJCOPY       "${TOOLCHAIN_PREFIX}objcopy")
-    set(CMAKE_SIZE          "${TOOLCHAIN_PREFIX}size")
-    set(CMAKE_STRIP         "${TOOLCHAIN_PREFIX}ld")
-
-    target_link_libraries(fast-depth PUBLIC target-riscv)
+  message(STATUS "fast-depth: Building for RISC-V")
+  target_link_libraries(fast-depth PUBLIC target-riscv)
 endif ()
 
-add_compile_options(-O3 -Wall -Wextra)
+target_compile_options(fast-depth PRIVATE -O3 -Wall -Wextra)
 
-target_compile_options(fast-depth PRIVATE -u _printf_float)
-
-add_subdirectory(../../ ./build/)
-
-add_subdirectory(../../nn/ ./build/nn)
 target_link_libraries(fast-depth PUBLIC nn)
-
-target_link_libraries(fast-depth PUBLIC m)
-

examples/fast-depth/main.c

@@ -16,7 +16,7 @@
 #include "termimg.h"
 
 // load the weight data block from the model.bin file
-INCLUDE_FILE(".rodata", "../input.bin", model_input);
+INCLUDE_FILE(".rodata", "./input.bin", model_input);
 extern uint8_t model_input_data[];
 extern size_t model_input_start[];
 extern size_t model_input_end[];

examples/fast-depth/model.h

@@ -6,7 +6,7 @@
 
 
 // load the weight data block from the model.bin file
-INCLUDE_FILE(".rodata", "../model.bin", model_weight);
+INCLUDE_FILE(".rodata", "./model.bin", model_weight);
 extern uint8_t model_weight_data[];
 extern size_t model_weight_start[];
 extern size_t model_weight_end[];

examples/llama2/CMakeLists.txt

@@ -2,44 +2,21 @@ cmake_minimum_required(VERSION 3.10)
 
 project(llama2 LANGUAGES C)
 
-
 add_executable(llama2 main.c)
-target_include_directories(llama2 PUBLIC inc)
+
+target_include_directories(llama2 PUBLIC ./)
+
 
 target_compile_features(llama2 INTERFACE c_std_11)
 
 if (X86)
-    message("llama2: building for x86")
-    target_link_libraries(llama2 PUBLIC target-x86)
-
+  message(STATUS "llama2: Building for x86")
 elseif (RISCV)
-    message("llama2: building for RISC-V")
-    # CMake toolchain definition for RISC-V GCC toolchain
-    set(CMAKE_SYSTEM_NAME "Generic" CACHE STRING "")
-    set(CMAKE_SYSTEM_PROCESSOR "riscv" CACHE STRING "")
-
-    set(TOOLCHAIN_PREFIX    "riscv64-unknown-elf-")
-
-    set(CMAKE_C_COMPILER    "${TOOLCHAIN_PREFIX}gcc")
-    set(CMAKE_ASM_COMPILER  "${TOOLCHAIN_PREFIX}gcc")
-    set(CMAKE_CXX_COMPILER  "${TOOLCHAIN_PREFIX}g++")
-    set(CMAKE_AR            "${TOOLCHAIN_PREFIX}ar")
-    set(CMAKE_LINKER        "{TOOLCHAIN_PREFIX}ld")
-    set(CMAKE_OBJCOPY       "${TOOLCHAIN_PREFIX}objcopy")
-    set(CMAKE_SIZE          "${TOOLCHAIN_PREFIX}size")
-    set(CMAKE_STRIP         "${TOOLCHAIN_PREFIX}ld")
-
-    target_link_libraries(llama2 PUBLIC target-riscv)
+  message(STATUS "llama2: Building for RISC-V")
+  target_link_libraries(llama2 PUBLIC target-riscv)
 endif ()
 
-add_compile_options(-O3 -Wall -Wextra)
+target_compile_options(llama2 PRIVATE -O3 -Wall -Wextra)
 
-target_compile_options(llama2 PRIVATE -u _printf_float)
-
-add_subdirectory(../../ ./build/)
-
-add_subdirectory(../../nn/ ./build/nn)
 target_link_libraries(llama2 PUBLIC nn)
 
-target_link_libraries(llama2 PUBLIC m)
-

examples/llama2/README.md

@@ -10,9 +10,9 @@ wget -P checkpoints/ https://huggingface.co/karpathy/tinyllamas/resolve/main/sto
 
 # Performance Benchmark
 
-Native impl
+Native impl (150479)
 
-```
+```bash
 [100%] Built target llama2
 Llama 2: a small transformer model for text generation
 forward taking 458394741 cycles
@@ -29,4 +29,33 @@ forward taking 458942021 cycles
  there
 forward taking 459044968 cycles
  was
-```
+```
+
+Replace matmul and softmax with JIT tensors
+
+```bash
+[100%] Built target llama2
+Llama 2: a small transformer model for text generation
+forward taking 11942828 cycles
+Once
+forward taking 12093307 cycles
+ upon
+forward taking 12196991 cycles
+ a
+forward taking 12292760 cycles
+ time
+forward taking 12398778 cycles
+,
+forward taking 12490108 cycles
+ there
+forward taking 12593055 cycles
+ was
+```
+
+
+replacing float arrays to tensors
+```
+forward taking 11274729 cycles
+```
+
+