#include "qa_utils.h"
#include <stdlib.h>
#include <boost/foreach.hpp>
#include <boost/assign/list_of.hpp>
#include <boost/tokenizer.hpp>
//#include <boost/test/unit_test.hpp>
#include <iostream>
#include <vector>
#include <time.h>
#include <math.h>
#include <boost/lexical_cast.hpp>
//#include <volk/volk_runtime.h>
#include <volk/volk_registry.h>
#include <volk/volk.h>
#include <boost/typeof/typeof.hpp>
#include <boost/type_traits.hpp>

float uniform() {
  return 2.0 * ((float) rand() / RAND_MAX - 0.5);	// uniformly (-1, 1)
}

template <class t>
void random_floats (t *buf, unsigned n)
{
  for (unsigned i = 0; i < n; i++)
    buf[i] = uniform ();
}

void load_random_data(void *data, volk_type_t type, unsigned int n) {
    if(type.is_complex) n *= 2;
    if(type.is_float) {
        if(type.size == 8) random_floats<double>((double *)data, n);
        else random_floats<float>((float *)data, n);
    } else {
        float int_max = float(uint64_t(2) << (type.size*8));
        if(type.is_signed) int_max /= 2.0;
        for(int i=0; i<n; i++) {
            float scaled_rand = (((float) (rand() - (RAND_MAX/2))) / static_cast<float>((RAND_MAX/2))) * int_max;
            //man i really don't know how to do this in a more clever way, you have to cast down at some point
            switch(type.size) {
            case 8:
                if(type.is_signed) ((int64_t *)data)[i] = (int64_t) scaled_rand;
                else ((uint64_t *)data)[i] = (uint64_t) scaled_rand;
            break;
            case 4:
                if(type.is_signed) ((int32_t *)data)[i] = (int32_t) scaled_rand;
                else ((uint32_t *)data)[i] = (uint32_t) scaled_rand;
            break;           
            case 2:
                if(type.is_signed) ((int16_t *)data)[i] = (int16_t) scaled_rand;
                else ((uint16_t *)data)[i] = (uint16_t) scaled_rand;
            break;
            case 1:
                if(type.is_signed) ((int8_t *)data)[i] = (int8_t) scaled_rand;
                else ((uint8_t *)data)[i] = (uint8_t) scaled_rand;
            break;
            default:
                throw "load_random_data: no support for data size > 8 or < 1"; //no shenanigans here
            }
        }
    }
}

void *make_aligned_buffer(unsigned int len, unsigned int size) {
  void *buf;
  int ret;
  ret = posix_memalign((void**)&buf, 16, len * size);
  assert(ret == 0);
  memset(buf, 0x00, len*size);
  return buf;
}

void make_buffer_for_signature(std::vector<void *> &buffs, std::vector<volk_type_t> inputsig, unsigned int vlen) {
    BOOST_FOREACH(volk_type_t sig, inputsig) {
        if(!sig.is_scalar) //we don't make buffers for scalars
          buffs.push_back(make_aligned_buffer(vlen, sig.size*(sig.is_complex ? 2 : 1)));
    }
}

static std::vector<std::string> get_arch_list(const int archs[]) {
    std::vector<std::string> archlist;
    int num_archs = archs[0];
    
    //there has got to be a way to query these arches
    for(int i = 0; i < num_archs; i++) {
        switch(archs[i+1]) {
        case (1<<LV_GENERIC):
            archlist.push_back("generic");
            break;
        case (1<<LV_ORC):
            archlist.push_back("orc");
            break;
        case (1<<LV_SSE):
            archlist.push_back("sse");
            break;
        case (1<<LV_SSE2):
            archlist.push_back("sse2");
            break;
        case (1<<LV_SSE3):
            archlist.push_back("sse3");
            break;
        case (1<<LV_SSSE3):
            archlist.push_back("ssse3");
            break;
        case (1<<LV_SSE4_1):
            archlist.push_back("sse4_1");
            break;
        case (1<<LV_SSE4_2):
            archlist.push_back("sse4_2");
            break;
        case (1<<LV_SSE4_A):
            archlist.push_back("sse4_a");
            break;
        case (1<<LV_MMX):
            archlist.push_back("mmx");
            break;
        case (1<<LV_AVX):
            archlist.push_back("avx");
            break;
        default:
            break;
        }
    }
    return archlist;
}

volk_type_t volk_type_from_string(std::string name) {
    volk_type_t type;
    type.is_float = false;
    type.is_scalar = false;
    type.is_complex = false;
    type.is_signed = false;
    type.size = 0;
    type.str = name;
    
    if(name.size() < 2) throw std::string("name too short to be a datatype");
    
    //is it a scalar?
    if(name[0] == 's') { 
        type.is_scalar = true;
        name = name.substr(1, name.size()-1);
    }
    
    //get the data size
    int last_size_pos = name.find_last_of("0123456789");
    if(last_size_pos < 0) throw std::string("no size spec in type ").append(name);
    //will throw if malformed
    int size = boost::lexical_cast<int>(name.substr(0, last_size_pos+1));

    assert(((size % 8) == 0) && (size <= 64) && (size != 0));
    type.size = size/8; //in bytes
    
    for(int i=last_size_pos+1; i < name.size(); i++) {
        switch (name[i]) {
        case 'f':
            type.is_float = true;
            break;
        case 'i':
            type.is_signed = true;
            break;
        case 'c':
            type.is_complex = true;
            break;
        case 'u':
            type.is_signed = false;
            break;
        default:
            throw;
        }
    }
    
    return type;
}

static void get_signatures_from_name(std::vector<volk_type_t> &inputsig, 
                                   std::vector<volk_type_t> &outputsig, 
                                   std::string name) {
    boost::char_separator<char> sep("_");
    boost::tokenizer<boost::char_separator<char> > tok(name, sep);
    std::vector<std::string> toked;
    tok.assign(name);
    toked.assign(tok.begin(), tok.end());
    
    assert(toked[0] == "volk");
    toked.erase(toked.begin());

    //ok. we're assuming a string in the form
    //(sig)_(multiplier-opt)_..._(name)_(sig)_(multiplier-opt)_..._(alignment)

    enum { SIDE_INPUT, SIDE_NAME, SIDE_OUTPUT } side = SIDE_INPUT;
    std::string fn_name;
    volk_type_t type;
    BOOST_FOREACH(std::string token, toked) {
        try {
            type = volk_type_from_string(token);
            if(side == SIDE_NAME) side = SIDE_OUTPUT; //if this is the first one after the name...
            
            if(side == SIDE_INPUT) inputsig.push_back(type);
            else outputsig.push_back(type);
        } catch (...){
            if(token[0] == 'x') { //it's a multiplier
                if(side == SIDE_INPUT) assert(inputsig.size() > 0);
                else assert(outputsig.size() > 0);
                int multiplier = boost::lexical_cast<int>(token.substr(1, token.size()-1)); //will throw if invalid
                for(int i=1; i<multiplier; i++) {
                    if(side == SIDE_INPUT) inputsig.push_back(inputsig.back());
                    else outputsig.push_back(outputsig.back());
                }
            }
            else if(side == SIDE_INPUT) { //it's the function name, at least it better be
                side = SIDE_NAME;
                fn_name.append("_");
                fn_name.append(token);
            } 
            else if(side == SIDE_OUTPUT) {
                if(token != toked.back()) throw; //the last token in the name is the alignment
            }
        }
    }
    //we don't need an output signature (some fn's operate on the input data, "in place"), but we do need at least one input!
    assert(inputsig.size() != 0);
}

inline void run_cast_test1(volk_fn_1arg func, std::vector<void *> &buffs, unsigned int vlen, unsigned int iter, std::string arch) {
    while(iter--) func(buffs[0], vlen, arch.c_str());
}

inline void run_cast_test2(volk_fn_2arg func, std::vector<void *> &buffs, unsigned int vlen, unsigned int iter, std::string arch) {
    while(iter--) func(buffs[0], buffs[1], vlen, arch.c_str());
}

inline void run_cast_test3(volk_fn_3arg func, std::vector<void *> &buffs, unsigned int vlen, unsigned int iter, std::string arch) {
    while(iter--) func(buffs[0], buffs[1], buffs[2], vlen, arch.c_str());
}

inline void run_cast_test4(volk_fn_4arg func, std::vector<void *> &buffs, unsigned int vlen, unsigned int iter, std::string arch) {
    while(iter--) func(buffs[0], buffs[1], buffs[2], buffs[3], vlen, arch.c_str());
}

inline void run_cast_test1_s32f(volk_fn_1arg_s32f func, std::vector<void *> &buffs, float scalar, unsigned int vlen, unsigned int iter, std::string arch) {
    while(iter--) func(buffs[0], scalar, vlen, arch.c_str());
}

inline void run_cast_test2_s32f(volk_fn_2arg_s32f func, std::vector<void *> &buffs, float scalar, unsigned int vlen, unsigned int iter, std::string arch) {
    while(iter--) func(buffs[0], buffs[1], scalar, vlen, arch.c_str());
}

inline void run_cast_test3_s32f(volk_fn_3arg_s32f func, std::vector<void *> &buffs, float scalar, unsigned int vlen, unsigned int iter, std::string arch) {
    while(iter--) func(buffs[0], buffs[1], buffs[2], scalar, vlen, arch.c_str());
}

template <class t>
bool fcompare(t *in1, t *in2, unsigned int vlen, float tol) {
    bool fail = false;
    int print_max_errs = 10;
    for(int i=0; i<vlen; i++) {
        if(((t *)(in1))[i] < 1e-30) continue; //this is a hack: below around here we'll start to get roundoff errors due to limited precision
        if(fabs(((t *)(in1))[i] - ((t *)(in2))[i])/(((t *)in1)[i]) > tol) {
            fail=true;
            if(print_max_errs-- > 0) {
                std::cout << "offset " << i << " in1: " << t(((t *)(in1))[i]) << " in2: " << t(((t *)(in2))[i]) << std::endl;
            }
        }
    }
    
    return fail;
}

template <class t>
bool icompare(t *in1, t *in2, unsigned int vlen, unsigned int tol) {
    bool fail = false;
    int print_max_errs = 10;
    for(int i=0; i<vlen; i++) {
        if(abs(((t *)(in1))[i] - ((t *)(in2))[i]) > tol) {
            fail=true;
            if(print_max_errs-- > 0) {
                std::cout << "offset " << i << " in1: " << static_cast<int>(t(((t *)(in1))[i])) << " in2: " << static_cast<int>(t(((t *)(in2))[i])) << std::endl;
            }
        }
    }
    
    return fail;
}

bool run_volk_tests(const int archs[], void (*manual_func)(), std::string name, float tol, float scalar, int vlen, int iter) {
    std::cout << "RUN_VOLK_TESTS: " << name << std::endl;
    
    //first let's get a list of available architectures for the test
    std::vector<std::string> arch_list = get_arch_list(archs);
    
    if(arch_list.size() < 2) {
        std::cout << "no architectures to test" << std::endl;
        return false;
    }
    
    //now we have to get a function signature by parsing the name
    std::vector<volk_type_t> inputsig, outputsig;
    get_signatures_from_name(inputsig, outputsig, name);
    
    //pull the input scalars into their own vector
    std::vector<volk_type_t> inputsc;
    for(int i=0; i<inputsig.size(); i++) {
        if(inputsig[i].is_scalar) {
            inputsc.push_back(inputsig[i]);
            inputsig.erase(inputsig.begin() + i);
        }
    }

    //for(int i=0; i<inputsig.size(); i++) std::cout << "Input: " << inputsig[i].str << std::endl;
    //for(int i=0; i<outputsig.size(); i++) std::cout << "Output: " << outputsig[i].str << std::endl;
    std::vector<void *> inbuffs;
    std::vector<void *> free_buffs; //this is just a list of void*'s that i'll have to free later.
                                    //we need it because we dupe void*s in test_data below.
    make_buffer_for_signature(inbuffs, inputsig, vlen);
    for(int i=0; i<inbuffs.size(); i++) {
        load_random_data(inbuffs[i], inputsig[i], vlen);   
        free_buffs.push_back(inbuffs[i]);
    }
    
    //ok let's make a vector of vector of void buffers, which holds the input/output vectors for each arch
    std::vector<std::vector<void *> > test_data;
    for(int i=0; i<arch_list.size(); i++) {
        std::vector<void *> arch_buffs;
        for(int j=0; j<outputsig.size(); j++) {
            arch_buffs.push_back(make_aligned_buffer(vlen, outputsig[j].size*(outputsig[j].is_complex ? 2 : 1)));
            free_buffs.push_back(arch_buffs.back());
        }
        for(int j=0; j<inputsig.size(); j++) {
            arch_buffs.push_back(inbuffs[j]);
        }
        test_data.push_back(arch_buffs);
    }
    
    std::vector<volk_type_t> both_sigs;
    both_sigs.insert(both_sigs.end(), outputsig.begin(), outputsig.end());
    both_sigs.insert(both_sigs.end(), inputsig.begin(), inputsig.end());

    //now run the test
    clock_t start, end;
    for(int i = 0; i < arch_list.size(); i++) {
        start = clock();

        switch(both_sigs.size()) {
            case 1:
                if(inputsc.size() == 0) {
                    run_cast_test1((volk_fn_1arg)(manual_func), test_data[i], vlen, iter, arch_list[i]); 
                } else if(inputsc.size() == 1 && inputsc[0].is_float) {
                    run_cast_test1_s32f((volk_fn_1arg_s32f)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
                } else throw "unsupported 1 arg function >1 scalars";
                break;
            case 2:
                if(inputsc.size() == 0) {
                    run_cast_test2((volk_fn_2arg)(manual_func), test_data[i], vlen, iter, arch_list[i]);
                } else if(inputsc.size() == 1 && inputsc[0].is_float) {
                    run_cast_test2_s32f((volk_fn_2arg_s32f)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
                } else throw "unsupported 2 arg function >1 scalars";
                break;
            case 3:
                if(inputsc.size() == 0) {
                    run_cast_test3((volk_fn_3arg)(manual_func), test_data[i], vlen, iter, arch_list[i]);
                } else if(inputsc.size() == 1 && inputsc[0].is_float) {
                    run_cast_test3_s32f((volk_fn_3arg_s32f)(manual_func), test_data[i], scalar, vlen, iter, arch_list[i]);
                } else throw "unsupported 3 arg function >1 scalars";
                break;
            case 4:
                run_cast_test4((volk_fn_4arg)(manual_func), test_data[i], vlen, iter, arch_list[i]);
                break;
            default:
                throw "no function handler for this signature";
                break;
        }
        
        end = clock();
        std::cout << arch_list[i] << " completed in " << (double)(end-start)/(double)CLOCKS_PER_SEC << "s" << std::endl;
    }
    //and now compare each output to the generic output
    //first we have to know which output is the generic one, they aren't in order...
    int generic_offset=0;
    for(int i=0; i<arch_list.size(); i++) 
        if(arch_list[i] == "generic") generic_offset=i;

    //now compare
    //if(outputsig.size() == 0) outputsig = inputsig; //a hack, i know
    
    bool fail = false;
    bool fail_global = false;
    for(int i=0; i<arch_list.size(); i++) {
        if(i != generic_offset) {
            for(int j=0; j<both_sigs.size(); j++) {
                if(both_sigs[j].is_float) {
                    if(both_sigs[j].size == 8) {
                        fail = fcompare((double *) test_data[generic_offset][j], (double *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                    } else {
                        fail = fcompare((float *) test_data[generic_offset][j], (float *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                    }
                } else {
                    //i could replace this whole switch statement with a memcmp if i wasn't interested in printing the outputs where they differ
                    switch(both_sigs[j].size) {
                    case 8:
                        if(both_sigs[j].is_signed) {
                            fail = icompare((int64_t *) test_data[generic_offset][j], (int64_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        } else {
                            fail = icompare((uint64_t *) test_data[generic_offset][j], (uint64_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        }
                        break;
                    case 4:
                        if(both_sigs[j].is_signed) {
                            fail = icompare((int32_t *) test_data[generic_offset][j], (int32_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        } else {
                            fail = icompare((uint32_t *) test_data[generic_offset][j], (uint32_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        }
                        break;
                    case 2:
                        if(both_sigs[j].is_signed) {
                            fail = icompare((int16_t *) test_data[generic_offset][j], (int16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        } else {
                            fail = icompare((uint16_t *) test_data[generic_offset][j], (uint16_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        }
                        break;
                    case 1:
                        if(both_sigs[j].is_signed) {
                            fail = icompare((int8_t *) test_data[generic_offset][j], (int8_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        } else {
                            fail = icompare((uint8_t *) test_data[generic_offset][j], (uint8_t *) test_data[i][j], vlen*(both_sigs[j].is_complex ? 2 : 1), tol);
                        }
                        break;
                    default:
                        fail=1;
                    }
                }
                if(fail) {
                    fail_global = true;
                    std::cout << name << ": fail on arch " << arch_list[i] << std::endl;
                }
                //fail = memcmp(outbuffs[generic_offset], outbuffs[i], outputsig[0].size * vlen * (outputsig[0].is_complex ? 2:1));
            }
        }
    }

    BOOST_FOREACH(void *buf, free_buffs) {
        free(buf);
    }

    return fail_global;
}