onednn-src 0.1.13

/*******************************************************************************
* Copyright 2023 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/
#ifndef GPU_INTEL_UTILS_HPP
#define GPU_INTEL_UTILS_HPP

#include <iostream>

#include "common/cpp_compat.hpp"
#include "common/utils.hpp"
#include "gpu/intel/compute/device_info.hpp"

#define VCHECK_KERNEL(stat, msg, ...) \
    VCHECK(common, create, check, runtime, stat, msg, ##__VA_ARGS__);

#if defined(__GNUC__)
// clang-format off
// Defines dump() method for debugging purposes, to pretty print the object.
#define XE_DEFINE_DUMP() \
    __attribute__((noinline)) \
    __attribute__((used)) \
    void dump() const { \
        printf("%s\n", str().c_str()); \
    }
// clang-format on
#else
#define XE_DEFINE_DUMP()
#endif

// Uncomment this when aborting on ir_assert is desired:
// #define GPU_ABORT_ON_ERROR

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {

enum { OCL_BUFFER_ALIGNMENT = 128 };

const char *get_kernel_header(const std::string &name);
const char *get_kernel_source(const char *name);

#define MAX_NDIMS 6
#define MAX_POST_OPS_SUPPORTED 32

using dim_idx_t = uint32_t;
namespace dim_idx {

constexpr dim_idx_t invalid = static_cast<dim_idx_t>(-1);

inline char as_tag(dim_idx_t idx, bool is_outer = false) {
    return (is_outer ? 'A' : 'a') + static_cast<char>(idx);
}

} // namespace dim_idx

namespace gpu_utils {

class error_stream_t {
public:
    error_stream_t(const char *file, int line, const char *assert_msg) {
        out_ << "Assertion " << assert_msg << " failed at " << file << ":"
             << line << std::endl;
    }

    // This is to be able use a steam object in short-circuit evaluation with
    // booleans, see below.
    operator bool() const { return true; }

    template <typename T>
    error_stream_t &operator<<(const T &t) {
        out_ << t;
        return *this;
    }

    ~error_stream_t() noexcept(false) {
        if (cpp_compat::uncaught_exceptions()) return;
        std::cout << out_.str() << std::endl;
#ifdef GPU_ABORT_ON_ERROR
        std::abort();
#else
        throw std::runtime_error(out_.str());
#endif
    }

private:
    ostringstream_t out_;
};

#if !defined(NDEBUG) || defined(DNNL_DEV_MODE)
#define gpu_assert(cond) \
    !(cond) \
            && dnnl::impl::gpu::intel::gpu_utils::error_stream_t( \
                    __FILE__, __LINE__, #cond)
#else
#define gpu_assert(cond) \
    (false) && !(cond) \
            && dnnl::impl::gpu::intel::gpu_utils::error_stream_t( \
                    __FILE__, __LINE__, #cond)
#endif

#define gpu_error_not_expected() gpu_assert(false) << "Not expected. "
#define gpu_except_not_implemented(msg) \
    throw std::runtime_error(std::string(msg) + std::string(" at ") \
            + std::string(__FILENAME__) + std::string(":") \
            + std::to_string(__LINE__))

template <typename out_type, typename in_type,
        typename std::enable_if<!std::is_fundamental<out_type>::value
                || !std::is_fundamental<in_type>::value>::type>
inline bool validate_into(in_type in) {
    return true;
}
template <typename out_type, typename in_type,
        typename std::enable_if<std::is_fundamental<out_type>::value
                && std::is_fundamental<in_type>::value>::type>
inline bool validate_into(in_type in) {
    const double in_compare = static_cast<double>(in);
    const double out_max
            = static_cast<double>(std::numeric_limits<out_type>::max());
    const double out_lowest
            = static_cast<double>(std::numeric_limits<out_type>::lowest());
    return in_compare <= out_max && in_compare >= out_lowest;
}
template <typename out_type>
inline bool validate_into(bool b) {
    return std::is_integral<out_type>::value;
}

inline int dev_getenv(const char *name, int default_value) {
#ifdef DNNL_DEV_MODE
    return getenv_int(name, default_value);
#else
    return default_value;
#endif
}

inline bool dev_getenv(const char *s, bool def) {
    return dev_getenv(s, def ? 1 : 0) != 0;
}

inline std::string dev_getenv(const char *s, const std::string &def) {
#ifdef DNNL_DEV_MODE
    char buf[1024];
    int ret = getenv(s, buf, sizeof(buf));
    if (ret > 0) return buf;
    return def;
#else
    return def;
#endif
}

// Input is a comma separate list containing gpu_arch and optionally eu_count.
inline compute::gpu_arch_t dev_getenv(const char *s, compute::gpu_arch_t arch,
        int *eu_count = nullptr, int *max_wg_size = nullptr) {
#ifdef DNNL_DEV_MODE
    char buf[1024];
    int ret = getenv(s, buf, sizeof(buf));
    if (ret > 0) {
        char *arch_str = buf, *eu_str = nullptr;
        for (int i = 0; i < ret; i++) {
            if (buf[i] == ',') {
                buf[i] = 0;
                if (i < ret - 1) { eu_str = &buf[i + 1]; }
                break;
            }
        }
        arch = compute::str2gpu_arch(arch_str);
        if (eu_count && eu_str) { *eu_count = atoi(eu_str); }
        if (max_wg_size) {
            // Assume maximum wg size is basically the number of threads
            // available in a subslice with simd_size 16
            const int max_eus_per_wg
                    = compute::device_info_t::max_eus_per_wg(arch);
            const int simd_size = 16;
            const int thr_per_eu = utils::rnd_down_pow2(
                    compute::device_info_t::threads_per_eu(arch));
            *max_wg_size = simd_size * max_eus_per_wg * thr_per_eu;
        }
    }
    return arch;
#else
    return arch;
#endif
}

inline bool to_bool(const std::string &s) {
    if (s == "0" || s == "false") return false;
    return true;
}

inline std::vector<std::string> split(const std::string &s,
        const std::string &delimiter = std::string(1, ' ')) {
    size_t beg = 0;
    size_t end = 0;
    std::vector<std::string> ret;
    do {
        end = s.find(delimiter, beg);
        size_t len
                = (end == std::string::npos) ? std::string::npos : (end - beg);
        ret.push_back(s.substr(beg, len));
        beg = end + delimiter.size();
    } while (end != std::string::npos);
    return ret;
}

inline std::string join(
        const std::string &delimiter, const std::vector<std::string> &parts) {
    ostringstream_t oss;
    bool is_first = true;
    for (auto &p : parts) {
        if (!is_first) oss << delimiter;
        oss << p;
        is_first = false;
    }
    return oss.str();
}

bool is_jit_dump_enabled();
status_t dump_kernel_binary(
        const std::vector<uint8_t> &binary, const std::string &name);

// -------------------------------------------------------------
//  Backend      | Compound ID
// -------------------------------------------------------------
//  OpenCL       | <backend_t::opencl, cl_device, 0>
//  Level Zero   | <backend_t::ze, uuid[0-63], uuid[64-127]>
using device_id_t = std::tuple<int, uint64_t, uint64_t>;

struct device_id_hash_t {
    size_t operator()(const device_id_t &id) const {
        size_t result = 0;
        result = hash_combine(result, std::get<0>(id));
        result = hash_combine(result, std::get<1>(id));
        result = hash_combine(result, std::get<2>(id));
        return result;
    }
};

} // namespace gpu_utils

template <typename out_type, typename in_type>
inline out_type into(in_type in) {
    gpu_assert(gpu_utils::validate_into<out_type>(in))
            << "Value " << in << " cannot be converted into type "
            << typeid(out_type).name();
    return static_cast<out_type>(in);
}

} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl
#endif