|
SMAUG
Simulating Machine Learning Applications on gem5-Aladdin
|
Utilities for writing and invoking Aladdin kernels from Operators. More...
#include <stdint.h>#include "gem5/sampling_interface.h"#include <string>#include <utility>#include <memory>#include "smaug/core/globals.h"#include "tracer/trace_logger_aladdin.h"Go to the source code of this file.
Classes | |
| struct | _activation_param_t |
| Parameters to the activation function hardware. More... | |
| struct | ActivationInfo |
| Specifies an activation function and relevant parameters. More... | |
| struct | _SamplingInfo |
| Simulation sampling information maintained by the Operator and passed to the accelerated kernel. More... | |
Namespaces | |
| smaug | |
| The smaug namespace is the parent namespace of all C++ code in SMAUG. | |
Macros | |
| #define | CACHELINE_SIZE 32 |
| #define | LOG_PAGE_SIZE 12 |
| #define | VECTOR_SIZE 8 |
| Vector size used in SMV backends. | |
| #define | VEC128_MASK(input, mask) ((v4fp_t)((v4sfx_t)input & mask)) |
| Apply a elementwise mask to a 128-bit packed single precision FP vector. More... | |
| #define | VEC256_MASK(input, mask) ((v8fp_t)((v8sfx_t)input & mask)) |
| Same as VEC128_MASK, but for 256-bit vectors. More... | |
| #define | ARRAY_1D(TYPE, output_array_name, input_array_name) TYPE* output_array_name = (TYPE*)input_array_name |
| #define | ARRAY_2D(TYPE, output_array_name, input_array_name, DIM_1) TYPE(*output_array_name)[DIM_1] = (TYPE(*)[DIM_1])input_array_name |
| #define | ARRAY_3D(TYPE, output_array_name, input_array_name, DIM_1, DIM_2) |
| #define | ARRAY_4D( TYPE, output_array_name, input_array_name, DIM_1, DIM_2, DIM_3) |
| #define | ARRAY_5D( TYPE, output_array_name, input_array_name, DIM_1, DIM_2, DIM_3, DIM_4) |
| #define | VEC_ARRAY_1D(TYPE, output_array_name, input_array_name) TYPE* output_array_name = (TYPE*)(input_array_name) |
| #define | VEC_ARRAY_2D(TYPE, output_array_name, input_array_name, cols) |
| #define | VEC_ARRAY_3D(TYPE, output_array_name, input_array_name, rows, cols) |
| #define | VEC_ARRAY_4D( TYPE, output_array_name, input_array_name, height, rows, cols) |
| #define | max2(A, B) (((A) > (B)) ? (A) : (B)) |
| #define | max3(e0, e1, e2) max2(max2(e0, e1), e2) |
| #define | max4(e0, e1, e2, e3) max2(max2(e0, e1), max2(e2, e3)) |
| #define | max8(e0, e1, e2, e3, e4, e5, e6, e7) max2(max4(e0, e1, e2, e3), max4(e4, e5, e6, e7)) |
| #define | max9(e0, e1, e2, e3, e4, e5, e6, e7, e8) max2(max8(e0, e1, e2, e3, e4, e5, e6, e7), e8) |
| #define | min2(A, B) (((A) < (B)) ? (A) : (B)) |
| #define | FRAC_CEIL(A, B) ((A) / (B) + ((A) % (B) != 0)) |
| Implements the ceiling function of A/B. | |
| #define | ALWAYS_INLINE |
| We have to disable all function inlining at the global level for Aladdin + LLVM-Tracer to work, but sometimes we do want to force inline functions (otherwise we run into all the issues of function call barriers in Aladdin). More... | |
| #define | ASSERT(x) assert(x) |
| An assertion macro which disables asserts in LLVM-Tracer instrumented code. | |
| #define | ASSUME_ALIGNED(ptr, args...) __builtin_assume_aligned((ptr), args) |
| Tell the compiler to assume a pointer is aligned on some byte boundary. More... | |
| #define | MAYBE_UNUSED __attribute__((__unused__)) |
Typedefs | |
| typedef enum _activation_type | activation_type |
| The activation function to apply to an operator's output in hardware. More... | |
| typedef struct _activation_param_t | activation_param_t |
| Parameters to the activation function hardware. More... | |
| typedef enum _SamplingLevel | SamplingLevel |
| Levels of simulation sampling to apply to certain accelerator kernels. More... | |
| typedef struct _SamplingInfo | SamplingInfo |
| Simulation sampling information maintained by the Operator and passed to the accelerated kernel. | |
| typedef float | fp_t |
| typedef int | sfx_t |
| typedef unsigned | ufx_t |
| typedef uint16_t | fp16_t |
| typedef uint16_t | float16 |
| typedef fp16_t | v16fp_t |
| 16 packed 32-bit floating point values. | |
| typedef fp_t | v8fp_t |
| 8 packed 32-bit floating point values. | |
| typedef fp_t | v4fp_t |
| 4 packed 32-bit floating point values. | |
| typedef fp16_t | v16ph_t |
| 16 packed 16-bit floating point values. | |
| typedef fp16_t | v8ph_t |
| 8 packed 16-bit floating point values. | |
| typedef fp16_t | v4ph_t |
| 4 packed 16-bit floating point values. | |
| typedef sfx_t | v8sfx_t |
| 8 packed 32-bit integer values. | |
| typedef sfx_t | v4sfx_t |
| 4 packed 32-bit integer values. | |
| typedef uint8_t | v8bl_t |
| 8 packed 8-bit bool values. | |
Enumerations | |
| enum | _activation_type { NO_ACTIVATION, RELU, RELU_THRESHOLD, LRELU, ELU, SELU, TANH, HARD_TANH, SIGMOID, SOFTMAX } |
| The activation function to apply to an operator's output in hardware. More... | |
| enum | _SamplingLevel { NoSampling = 0, Low = 1, Medium = 2, High = 3, VeryHigh = 4 } |
| Levels of simulation sampling to apply to certain accelerator kernels. More... | |
Functions | |
| std::string | smaug::getTraceName (int accelIdx) |
| Return the name of the dynamic trace for this accelerator. More... | |
| template<typename Kernel , typename... Args> | |
| void | smaug::invokeKernel (int accelIdx, unsigned reqCode, const Kernel &kernel, Args &&... args) |
| The generic blocking interface for all accelerator kernel functions. More... | |
| template<typename Kernel , typename... Args> | |
| void | smaug::invokeKernel (unsigned reqCode, const Kernel &kernel, Args &&... args) |
| A generic interface for all accelerator kernel functions. More... | |
| template<typename Kernel , typename... Args> | |
| std::unique_ptr< volatile int > | smaug::invokeKernelNoBlock (int accelIdx, unsigned reqCode, const Kernel &kernel, Args &&... args) |
| A generic non-blocking interface to accelerated kernel functions. More... | |
| void | smaug::mapArrayToAccel (unsigned reqCode, const char *arrayName, void *baseAddr, size_t size) |
| Maps an array of data to the accelerator. More... | |
| void | smaug::setArrayMemTypeIfSimulating (unsigned reqCode, const char *arrayName, MemoryType memType) |
| Sets what memory access mechanism the accelerator will use when accessing this array. More... | |
| size_t | next_multiple (size_t request, size_t align) |
| Returns the smallest multiple of align that is >= request. | |
Utilities for writing and invoking Aladdin kernels from Operators.
Definition in file common.h.
| typedef struct _activation_param_t activation_param_t |
Parameters to the activation function hardware.
This is a struct that bridges the C++ class ActivationInfo with the C kernel.
| typedef enum _activation_type activation_type |
The activation function to apply to an operator's output in hardware.
This is a struct that bridges the C++ class ActivationInfo with the C kernel.
| typedef enum _SamplingLevel SamplingLevel |
Levels of simulation sampling to apply to certain accelerator kernels.
It is the responsibility of the kernel writer to translate this qualitative description into a kernel-appropriate modification of the code behavior. For example, Low could mean "only sample the innermost loop", while VeryHigh could be translated to "sample all loop levels". Kernels do not need to support every single sampling level.
| enum _activation_type |
The activation function to apply to an operator's output in hardware.
This is a struct that bridges the C++ class ActivationInfo with the C kernel.
| enum _SamplingLevel |
Levels of simulation sampling to apply to certain accelerator kernels.
It is the responsibility of the kernel writer to translate this qualitative description into a kernel-appropriate modification of the code behavior. For example, Low could mean "only sample the innermost loop", while VeryHigh could be translated to "sample all loop levels". Kernels do not need to support every single sampling level.
1.8.18