Initial import: grid-bot — grid trading bot for BTC-USDT on Cifra Markets
This commit is contained in:
Kolp
@@ -0,0 +1,145 @@
|
||||
#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION)
|
||||
// !!!! PLEASE READ !!!!
|
||||
// Minimize (transitively) included headers from _avx*.cc because some of the
|
||||
// functions defined in the headers compiled with platform dependent compiler
|
||||
// options can be reused by other translation units generating illegal
|
||||
// instruction run-time error.
|
||||
|
||||
// Common utilities for writing performance kernels and easy dispatching of
|
||||
// different backends.
|
||||
/*
|
||||
The general workflow shall be as follows, say we want to
|
||||
implement a functionality called void foo(int a, float b).
|
||||
|
||||
In foo.h, do:
|
||||
void foo(int a, float b);
|
||||
|
||||
In foo_avx512.cc, do:
|
||||
void foo__avx512(int a, float b) {
|
||||
[actual avx512 implementation]
|
||||
}
|
||||
|
||||
In foo_avx2.cc, do:
|
||||
void foo__avx2(int a, float b) {
|
||||
[actual avx2 implementation]
|
||||
}
|
||||
|
||||
In foo_avx.cc, do:
|
||||
void foo__avx(int a, float b) {
|
||||
[actual avx implementation]
|
||||
}
|
||||
|
||||
In foo.cc, do:
|
||||
// The base implementation should *always* be provided.
|
||||
void foo__base(int a, float b) {
|
||||
[base, possibly slow implementation]
|
||||
}
|
||||
decltype(foo__base) foo__avx512;
|
||||
decltype(foo__base) foo__avx2;
|
||||
decltype(foo__base) foo__avx;
|
||||
void foo(int a, float b) {
|
||||
// You should always order things by their preference, faster
|
||||
// implementations earlier in the function.
|
||||
AVX512_DO(foo, a, b);
|
||||
AVX2_DO(foo, a, b);
|
||||
AVX_DO(foo, a, b);
|
||||
BASE_DO(foo, a, b);
|
||||
}
|
||||
|
||||
*/
|
||||
// Details: this functionality basically covers the cases for both build time
|
||||
// and run time architecture support.
|
||||
//
|
||||
// During build time:
|
||||
// The build system should provide flags CAFFE2_PERF_WITH_AVX512,
|
||||
// CAFFE2_PERF_WITH_AVX2, and CAFFE2_PERF_WITH_AVX that corresponds to the
|
||||
// __AVX512F__, __AVX512DQ__, __AVX512VL__, __AVX2__, and __AVX__ flags the
|
||||
// compiler provides. Note that we do not use the compiler flags but rely on
|
||||
// the build system flags, because the common files (like foo.cc above) will
|
||||
// always be built without __AVX512F__, __AVX512DQ__, __AVX512VL__, __AVX2__
|
||||
// and __AVX__.
|
||||
// During run time:
|
||||
// we use cpuinfo to identify cpu support and run the proper functions.
|
||||
|
||||
#pragma once
|
||||
#if defined(CAFFE2_PERF_WITH_SVE) || defined(CAFFE2_PERF_WITH_AVX512) || \
|
||||
defined(CAFFE2_PERF_WITH_AVX2) || defined(CAFFE2_PERF_WITH_AVX)
|
||||
#include <cpuinfo.h>
|
||||
#endif
|
||||
|
||||
// DO macros: these should be used in your entry function, similar to foo()
|
||||
// above, that routes implementations based on CPU capability.
|
||||
|
||||
#define BASE_DO(funcname, ...) return funcname##__base(__VA_ARGS__);
|
||||
|
||||
#ifdef CAFFE2_PERF_WITH_SVE
|
||||
#define SVE_DO(funcname, ...) \
|
||||
{ \
|
||||
static const bool isDo = cpuinfo_initialize() && cpuinfo_has_arm_sve(); \
|
||||
if (isDo) { \
|
||||
return funcname##__sve(__VA_ARGS__); \
|
||||
} \
|
||||
}
|
||||
#else // CAFFE2_PERF_WITH_SVE
|
||||
#define SVE_DO(funcname, ...)
|
||||
#endif // CAFFE2_PERF_WITH_SVE
|
||||
|
||||
#ifdef CAFFE2_PERF_WITH_AVX512
|
||||
#define AVX512_DO(funcname, ...) \
|
||||
{ \
|
||||
static const bool isDo = cpuinfo_initialize() && \
|
||||
cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512dq() && \
|
||||
cpuinfo_has_x86_avx512vl(); \
|
||||
if (isDo) { \
|
||||
return funcname##__avx512(__VA_ARGS__); \
|
||||
} \
|
||||
}
|
||||
#else // CAFFE2_PERF_WITH_AVX512
|
||||
#define AVX512_DO(funcname, ...)
|
||||
#endif // CAFFE2_PERF_WITH_AVX512
|
||||
|
||||
#ifdef CAFFE2_PERF_WITH_AVX2
|
||||
#define AVX2_DO(funcname, ...) \
|
||||
{ \
|
||||
static const bool isDo = cpuinfo_initialize() && cpuinfo_has_x86_avx2(); \
|
||||
if (isDo) { \
|
||||
return funcname##__avx2(__VA_ARGS__); \
|
||||
} \
|
||||
}
|
||||
#define AVX2_FMA_DO(funcname, ...) \
|
||||
{ \
|
||||
static const bool isDo = cpuinfo_initialize() && cpuinfo_has_x86_avx2() && \
|
||||
cpuinfo_has_x86_fma3(); \
|
||||
if (isDo) { \
|
||||
return funcname##__avx2_fma(__VA_ARGS__); \
|
||||
} \
|
||||
}
|
||||
#else // CAFFE2_PERF_WITH_AVX2
|
||||
#define AVX2_DO(funcname, ...)
|
||||
#define AVX2_FMA_DO(funcname, ...)
|
||||
#endif // CAFFE2_PERF_WITH_AVX2
|
||||
|
||||
#ifdef CAFFE2_PERF_WITH_AVX
|
||||
#define AVX_DO(funcname, ...) \
|
||||
{ \
|
||||
static const bool isDo = cpuinfo_initialize() && cpuinfo_has_x86_avx(); \
|
||||
if (isDo) { \
|
||||
return funcname##__avx(__VA_ARGS__); \
|
||||
} \
|
||||
}
|
||||
#define AVX_F16C_DO(funcname, ...) \
|
||||
{ \
|
||||
static const bool isDo = cpuinfo_initialize() && cpuinfo_has_x86_avx() && \
|
||||
cpuinfo_has_x86_f16c(); \
|
||||
if (isDo) { \
|
||||
return funcname##__avx_f16c(__VA_ARGS__); \
|
||||
} \
|
||||
}
|
||||
#else // CAFFE2_PERF_WITH_AVX
|
||||
#define AVX_DO(funcname, ...)
|
||||
#define AVX_F16C_DO(funcname, ...)
|
||||
#endif // CAFFE2_PERF_WITH_AVX
|
||||
|
||||
#else
|
||||
#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined."
|
||||
#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION)
|
||||
+62
@@ -0,0 +1,62 @@
|
||||
#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION)
|
||||
#pragma once
|
||||
|
||||
#include <cstdint>
|
||||
|
||||
namespace caffe2 {
|
||||
|
||||
// clang-format off
|
||||
/**
|
||||
* Embedding lookup with reduction.
|
||||
*
|
||||
* `input` of size data_size * block_size
|
||||
* `indices` of size index_size
|
||||
* `offsets` of size output_size
|
||||
* `weights` nullptr or array of size index_size
|
||||
* `out` of size output_size * block_size
|
||||
*
|
||||
* Behavior is roughly equivalent to pseudocode:
|
||||
*
|
||||
* pos = 0
|
||||
* for (i = 0..output_size-1)
|
||||
* for (k = 0..block_size-1)
|
||||
* out[i*block_size + k] = 0
|
||||
* start_offset = offsets[i]
|
||||
* end_offset = offsets[i+1]
|
||||
* length = end_offset - start_offset
|
||||
* for (j = start_offset..end_offset-1)
|
||||
* for (k = 0..block_size-1)
|
||||
* out[i*block_size + k] += input[indices[pos]*block_size + k] *
|
||||
* (weights ? weights[IS_WEIGHT_POSITIONAL ? j - start_offset : pos] : 1.0)
|
||||
* pos += 1
|
||||
* if (normalize_weights && length > 0)
|
||||
* for (k = 0..block_size-1)
|
||||
* out[i*block_size + k] /= length
|
||||
*
|
||||
* TODO: make this API also take "offsets" rather than "lengths" to match the
|
||||
* API for PyTorch's EmbeddingBag
|
||||
*/
|
||||
// clang-format on
|
||||
template <
|
||||
typename IndexType,
|
||||
typename InType,
|
||||
typename OutType,
|
||||
bool IS_WEIGHT_POSITIONAL = false>
|
||||
void EmbeddingLookupIdx(
|
||||
const std::int64_t block_size,
|
||||
const std::int64_t output_size,
|
||||
const std::int64_t index_size,
|
||||
const std::int64_t data_size,
|
||||
const InType* input,
|
||||
const IndexType* indices,
|
||||
const IndexType* offsets,
|
||||
const float* weights, // optional, can be null for non-weighted sum
|
||||
const float* scale_bias, // optional scale & bias params for uint8 input
|
||||
bool normalize_by_lengths,
|
||||
OutType* out);
|
||||
|
||||
} // namespace caffe2
|
||||
|
||||
#else
|
||||
#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined."
|
||||
#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION)
|
||||
Reference in New Issue
Block a user