/* -*- c++ -*- */
/*
 * Copyright 2011 - 2020, 2022 Free Software Foundation, Inc.
 * Copyright 2025 Magnus Lundmark <magnuslundmark@gmail.com>
 *
 * This file is part of VOLK
 *
 * SPDX-License-Identifier: LGPL-3.0-or-later
 */

#include "qa_utils.h"
#include <volk/volk.h>

#include <volk/volk.h>        // for volk_func_desc_t
#include <volk/volk_malloc.h> // for volk_free, volk_m...

#include <assert.h>    // for assert
#include <stdint.h>    // for uint16_t, uint64_t
#include <sys/time.h>  // for CLOCKS_PER_SEC
#include <sys/types.h> // for int16_t, int32_t
#include <chrono>
#include <cmath>    // for sqrt, fabs, abs
#include <cstring>  // for memcpy, memset
#include <ctime>    // for clock
#include <iostream> // for cerr
#include <limits>   // for numeric_limits
#include <map>      // for map, map<>::mappe...
#include <random>
#include <vector> // for vector, _Bit_refe...

#include <fmt/core.h>

// Warmup time for CPU frequency scaling (ms)
static double g_warmup_ms = 2000.0;
static bool g_warmup_done = false;

double volk_test_get_warmup_ms() { return g_warmup_ms; }
void volk_test_set_warmup_ms(double ms) { g_warmup_ms = ms; }
void volk_test_reset_warmup() { g_warmup_done = false; }

template <typename T>
void random_floats(void* buf, unsigned int n, std::default_random_engine& rnd_engine)
{
    T* array = static_cast<T*>(buf);
    std::uniform_real_distribution<T> uniform_dist(T(-1), T(1));
    for (unsigned int i = 0; i < n; i++) {
        array[i] = uniform_dist(rnd_engine);
    }
}

void load_random_data(void* data,
                      volk_type_t type,
                      unsigned int n,
                      const std::vector<float>& float_edge_cases,
                      const std::vector<lv_32fc_t>& complex_edge_cases)
{
    std::random_device rnd_device;
    std::default_random_engine rnd_engine(rnd_device());

    unsigned int edge_case_count = 0;

    // Inject complex edge cases for complex float types
    if (type.is_float && type.is_complex && !complex_edge_cases.empty()) {
        edge_case_count = std::min((unsigned int)complex_edge_cases.size(), n);
        if (type.size == 8) {
            lv_64fc_t* array = static_cast<lv_64fc_t*>(data);
            for (unsigned int i = 0; i < edge_case_count; i++) {
                array[i] = lv_cmake((double)lv_creal(complex_edge_cases[i]),
                                    (double)lv_cimag(complex_edge_cases[i]));
            }
        } else {
            lv_32fc_t* array = static_cast<lv_32fc_t*>(data);
            for (unsigned int i = 0; i < edge_case_count; i++) {
                array[i] = complex_edge_cases[i];
            }
        }
    }
    // Inject float edge cases for non-complex float types
    else if (type.is_float && !type.is_complex && !float_edge_cases.empty()) {
        edge_case_count = std::min((unsigned int)float_edge_cases.size(), n);
        if (type.size == 8) {
            double* array = static_cast<double*>(data);
            for (unsigned int i = 0; i < edge_case_count; i++) {
                array[i] = static_cast<double>(float_edge_cases[i]);
            }
        } else {
            float* array = static_cast<float*>(data);
            for (unsigned int i = 0; i < edge_case_count; i++) {
                array[i] = float_edge_cases[i];
            }
        }
    }

    unsigned int remaining_n = n - edge_case_count;
    if (type.is_complex)
        remaining_n *= 2;

    if (type.is_float) {
        if (type.size == 8) {
            double* array = static_cast<double*>(data);
            random_floats<double>(array + edge_case_count * (type.is_complex ? 2 : 1),
                                  remaining_n,
                                  rnd_engine);
        } else {
            float* array = static_cast<float*>(data);
            random_floats<float>(array + edge_case_count * (type.is_complex ? 2 : 1),
                                 remaining_n,
                                 rnd_engine);
        }
    } else {
        if (type.is_complex)
            n *= 2;
        switch (type.size) {
        case 8:
            if (type.is_signed) {
                std::uniform_int_distribution<int64_t> uniform_dist(
                    std::numeric_limits<int64_t>::min(),
                    std::numeric_limits<int64_t>::max());
                for (unsigned int i = 0; i < n; i++)
                    ((int64_t*)data)[i] = uniform_dist(rnd_engine);
            } else {
                std::uniform_int_distribution<uint64_t> uniform_dist(
                    std::numeric_limits<uint64_t>::min(),
                    std::numeric_limits<uint64_t>::max());
                for (unsigned int i = 0; i < n; i++)
                    ((uint64_t*)data)[i] = uniform_dist(rnd_engine);
            }
            break;
        case 4:
            if (type.is_signed) {
                std::uniform_int_distribution<int32_t> uniform_dist(
                    std::numeric_limits<int32_t>::min(),
                    std::numeric_limits<int32_t>::max());
                for (unsigned int i = 0; i < n; i++)
                    ((int32_t*)data)[i] = uniform_dist(rnd_engine);
            } else {
                std::uniform_int_distribution<uint32_t> uniform_dist(
                    std::numeric_limits<uint32_t>::min(),
                    std::numeric_limits<uint32_t>::max());
                for (unsigned int i = 0; i < n; i++)
                    ((uint32_t*)data)[i] = uniform_dist(rnd_engine);
            }
            break;
        case 2:
            if (type.is_signed) {
                std::uniform_int_distribution<int16_t> uniform_dist(-6, 6);
                for (unsigned int i = 0; i < n; i++)
                    ((int16_t*)data)[i] = uniform_dist(rnd_engine);
            } else {
                std::uniform_int_distribution<uint16_t> uniform_dist(
                    std::numeric_limits<uint16_t>::min(),
                    std::numeric_limits<uint16_t>::max());
                for (unsigned int i = 0; i < n; i++)
                    ((uint16_t*)data)[i] = uniform_dist(rnd_engine);
            }
            break;
        case 1:
            if (type.is_signed) {
                std::uniform_int_distribution<int16_t> uniform_dist(
                    std::numeric_limits<int8_t>::min(),
                    std::numeric_limits<int8_t>::max());
                for (unsigned int i = 0; i < n; i++)
                    ((int8_t*)data)[i] = uniform_dist(rnd_engine);
            } else {
                std::uniform_int_distribution<uint16_t> uniform_dist(
                    std::numeric_limits<uint8_t>::min(),
                    std::numeric_limits<uint8_t>::max());
                for (unsigned int i = 0; i < n; i++)
                    ((uint8_t*)data)[i] = uniform_dist(rnd_engine);
            }
            break;
        default:
            throw "load_random_data: no support for data size > 8 or < 1"; // no
                                                                           // shenanigans
                                                                           // here
        }
    }
}

static std::vector<std::string> get_arch_list(volk_func_desc_t desc)
{
    std::vector<std::string> archlist;

    for (size_t i = 0; i < desc.n_impls; i++) {
        archlist.push_back(std::string(desc.impl_names[i]));
    }

    return archlist;
}

template <typename T>
T volk_lexical_cast(const std::string& str)
{
    for (unsigned int c_index = 0; c_index < str.size(); ++c_index) {
        if (str.at(c_index) < '0' || str.at(c_index) > '9') {
            throw "not all numbers!";
        }
    }
    T var;
    std::istringstream iss;
    iss.str(str);
    iss >> var;
    // deal with any error bits that may have been set on the stream
    return var;
}

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
    size_t last_size_pos = name.find_last_of("0123456789");
    if (last_size_pos == std::string::npos) {
        throw std::string("no size spec in type ").append(name);
    }
    // will throw if malformed
    int size = volk_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 (size_t 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 std::string("Error: no such type: '") + name[i] + "'";
        }
    }

    return type;
}

std::vector<std::string> split_signature(const std::string& protokernel_signature)
{
    std::vector<std::string> signature_tokens;
    std::string token;
    for (unsigned int loc = 0; loc < protokernel_signature.size(); ++loc) {
        if (protokernel_signature.at(loc) == '_') {
            // this is a break
            signature_tokens.push_back(token);
            token = "";
        } else {
            token.push_back(protokernel_signature.at(loc));
        }
    }
    // Get the last one to the end of the string
    signature_tokens.push_back(token);
    return signature_tokens;
}

static void get_signatures_from_name(std::vector<volk_type_t>& inputsig,
                                     std::vector<volk_type_t>& outputsig,
                                     std::string name)
{

    std::vector<std::string> toked = split_signature(name);

    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;
    for (unsigned int token_index = 0; token_index < toked.size(); ++token_index) {
        std::string token = toked[token_index];
        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' && (token.size() > 1) &&
                (token[1] > '0' && token[1] < '9')) { // it's a multiplier
                if (side == SIDE_INPUT)
                    assert(inputsig.size() > 0);
                else
                    assert(outputsig.size() > 0);
                int multiplier = volk_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());
}

inline void run_cast_test1_s32fc(volk_fn_1arg_s32fc func,
                                 std::vector<void*>& buffs,
                                 lv_32fc_t 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_s32fc(volk_fn_2arg_s32fc func,
                                 std::vector<void*>& buffs,
                                 lv_32fc_t 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_s32fc(volk_fn_3arg_s32fc func,
                                 std::vector<void*>& buffs,
                                 lv_32fc_t 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* expected,
              t* actual,
              unsigned int vlen,
              float tol,
              bool absolute_mode,
              std::vector<unsigned int>& fail_indices,
              double& max_err)
{
    bool fail = false;
    max_err = 0.0;
    for (unsigned int i = 0; i < vlen; i++) {
        t exp_val = expected[i];
        t act_val = actual[i];

        // Check for special values (NaN, inf)
        bool exp_special = std::isnan(exp_val) || std::isinf(exp_val);
        bool act_special = std::isnan(act_val) || std::isinf(act_val);

        bool this_fail = false;
        double cmp_err = 0.0; // The error metric compared against tol
        if (exp_special || act_special) {
            // For NaN: both must be NaN (NaN != NaN, so use isnan)
            // For inf: both must be same signed infinity
            bool values_match =
                (std::isnan(exp_val) && std::isnan(act_val)) || (exp_val == act_val);
            if (!values_match) {
                this_fail = true;
                cmp_err = std::numeric_limits<double>::infinity();
            }
        } else if (absolute_mode) {
            cmp_err = fabs(exp_val - act_val);
            if (cmp_err > tol) {
                this_fail = true;
            }
        } else {
            // for very small numbers we'll see round off errors due to limited
            // precision. So a special test case...
            if (fabs(exp_val) < 1e-30) {
                cmp_err = fabs(act_val);
                if (cmp_err > tol) {
                    this_fail = true;
                }
            }
            // the primary test is the percent different greater than given tol
            else {
                cmp_err = fabs(exp_val - act_val) / fabs(exp_val);
                if (cmp_err > tol) {
                    this_fail = true;
                }
            }
        }
        if (cmp_err > max_err) {
            max_err = cmp_err;
        }
        if (this_fail) {
            fail = true;
            fail_indices.push_back(i);
        }
    }

    return fail;
}

template <class t>
bool ccompare(t* expected,
              t* actual,
              unsigned int vlen,
              float tol,
              bool absolute_mode,
              std::vector<unsigned int>& fail_indices,
              double& max_err)
{
    bool fail = false;
    max_err = 0.0;
    for (unsigned int i = 0; i < 2 * vlen; i += 2) {
        t exp_re = expected[i];
        t exp_im = expected[i + 1];
        t act_re = actual[i];
        t act_im = actual[i + 1];

        // Check for special values (NaN, inf) and verify they match
        bool exp_has_special = std::isnan(exp_re) || std::isnan(exp_im) ||
                               std::isinf(exp_re) || std::isinf(exp_im);
        bool act_has_special = std::isnan(act_re) || std::isnan(act_im) ||
                               std::isinf(act_re) || std::isinf(act_im);

        bool this_fail = false;
        double cmp_err = 0.0; // The error metric compared against tol
        if (exp_has_special || act_has_special) {
            // For NaN: both must be NaN (NaN != NaN, so use isnan)
            // For inf: both must be same signed infinity
            bool real_match =
                (std::isnan(exp_re) && std::isnan(act_re)) || (exp_re == act_re);
            bool imag_match =
                (std::isnan(exp_im) && std::isnan(act_im)) || (exp_im == act_im);

            if (!real_match || !imag_match) {
                this_fail = true;
                cmp_err = std::numeric_limits<double>::infinity();
            }
        } else {
            t diff[2] = { exp_re - act_re, exp_im - act_im };
            t err = std::sqrt(diff[0] * diff[0] + diff[1] * diff[1]);
            t norm = std::sqrt(exp_re * exp_re + exp_im * exp_im);

            if (absolute_mode) {
                cmp_err = err;
                if (cmp_err > tol) {
                    this_fail = true;
                }
            } else {
                // for very small numbers we'll see round off errors due to limited
                // precision. So a special test case...
                if (norm < 1e-30) {
                    cmp_err = err;
                    if (cmp_err > tol) {
                        this_fail = true;
                    }
                }
                // the primary test is the percent different greater than given tol
                else {
                    cmp_err = err / norm;
                    if (cmp_err > tol) {
                        this_fail = true;
                    }
                }
            }
        }
        if (cmp_err > max_err) {
            max_err = cmp_err;
        }
        if (this_fail) {
            fail = true;
            fail_indices.push_back(i / 2);
        }
    }

    return fail;
}

template <class t>
bool icompare(t* expected,
              t* actual,
              unsigned int vlen,
              unsigned int tol,
              std::vector<unsigned int>& fail_indices,
              double& max_err)
{
    bool fail = false;
    max_err = 0.0;
    for (unsigned int i = 0; i < vlen; i++) {
        t exp_val = expected[i];
        t act_val = actual[i];
        uint64_t abs_err = (uint64_t)abs(int64_t(exp_val) - int64_t(act_val));
        if ((double)abs_err > max_err) {
            max_err = (double)abs_err;
        }
        if (abs_err > tol) {
            fail = true;
            fail_indices.push_back(i);
        }
    }

    return fail;
}

// Print error table for failed comparisons
// Shows: index, input(s), expected, actual, rel_error, tol
void print_error_table(const std::vector<unsigned int>& fail_indices,
                       const std::vector<void*>& inputs,
                       const std::vector<volk_type_t>& input_sigs,
                       void* expected,
                       void* actual,
                       const volk_type_t& output_sig,
                       float tol,
                       int max_errors = 10)
{
    if (fail_indices.empty())
        return;

    // Print header
    fmt::print("{:>7}", "index");
    for (size_t k = 0; k < input_sigs.size(); k++) {
        fmt::print(" | {:>10}", fmt::format("in{}", k));
    }
    fmt::print(
        " | {:>10} | {:>10} | {:>9} | {:>9}\n", "expected", "actual", "rel_err", "tol");

    // Print separator
    fmt::print("{:-<7}", "");
    for (size_t k = 0; k < input_sigs.size(); k++) {
        fmt::print("-+-{:-<10}", "");
    }
    fmt::print("-+-{:-<10}-+-{:-<10}-+-{:-<9}-+-{:-<9}\n", "", "", "", "");

    int print_count = 0;
    for (unsigned int idx : fail_indices) {
        if (print_count++ >= max_errors) {
            fmt::print("... and {} more errors\n", fail_indices.size() - max_errors);
            break;
        }

        fmt::print("{:>7}", idx);

        // Print input values
        for (size_t k = 0; k < input_sigs.size(); k++) {
            if (input_sigs[k].is_float) {
                double val = (input_sigs[k].size == 8) ? ((double*)inputs[k])[idx]
                                                       : ((float*)inputs[k])[idx];
                fmt::print(" | {:>10.4f}", val);
            } else {
                int64_t val = 0;
                switch (input_sigs[k].size) {
                case 8:
                    val = input_sigs[k].is_signed ? ((int64_t*)inputs[k])[idx]
                                                  : (int64_t)((uint64_t*)inputs[k])[idx];
                    break;
                case 4:
                    val = input_sigs[k].is_signed ? ((int32_t*)inputs[k])[idx]
                                                  : (int64_t)((uint32_t*)inputs[k])[idx];
                    break;
                case 2:
                    val = input_sigs[k].is_signed ? ((int16_t*)inputs[k])[idx]
                                                  : (int64_t)((uint16_t*)inputs[k])[idx];
                    break;
                case 1:
                    val = input_sigs[k].is_signed ? ((int8_t*)inputs[k])[idx]
                                                  : (int64_t)((uint8_t*)inputs[k])[idx];
                    break;
                }
                fmt::print(" | {:>10}", val);
            }
        }

        // Get expected and actual values, compute relative error
        double exp_val = 0, act_val = 0, rel_err = 0;
        if (output_sig.is_float) {
            if (output_sig.size == 8) {
                exp_val = ((double*)expected)[idx];
                act_val = ((double*)actual)[idx];
            } else {
                exp_val = ((float*)expected)[idx];
                act_val = ((float*)actual)[idx];
            }
            double abs_err = fabs(exp_val - act_val);
            rel_err = (fabs(exp_val) > 1e-30) ? abs_err / fabs(exp_val) : abs_err;
            fmt::print(" | {:>10.4f} | {:>10.4f}", exp_val, act_val);
        } else {
            int64_t exp_i = 0, act_i = 0;
            switch (output_sig.size) {
            case 8:
                exp_i = output_sig.is_signed ? ((int64_t*)expected)[idx]
                                             : (int64_t)((uint64_t*)expected)[idx];
                act_i = output_sig.is_signed ? ((int64_t*)actual)[idx]
                                             : (int64_t)((uint64_t*)actual)[idx];
                break;
            case 4:
                exp_i = output_sig.is_signed ? ((int32_t*)expected)[idx]
                                             : (int64_t)((uint32_t*)expected)[idx];
                act_i = output_sig.is_signed ? ((int32_t*)actual)[idx]
                                             : (int64_t)((uint32_t*)actual)[idx];
                break;
            case 2:
                exp_i = output_sig.is_signed ? ((int16_t*)expected)[idx]
                                             : (int64_t)((uint16_t*)expected)[idx];
                act_i = output_sig.is_signed ? ((int16_t*)actual)[idx]
                                             : (int64_t)((uint16_t*)actual)[idx];
                break;
            case 1:
                exp_i = output_sig.is_signed ? ((int8_t*)expected)[idx]
                                             : (int64_t)((uint8_t*)expected)[idx];
                act_i = output_sig.is_signed ? ((int8_t*)actual)[idx]
                                             : (int64_t)((uint8_t*)actual)[idx];
                break;
            }
            fmt::print(" | {:>10} | {:>10}", exp_i, act_i);
            double abs_err = (double)abs(exp_i - act_i);
            rel_err = (exp_i != 0) ? abs_err / fabs((double)exp_i) : abs_err;
        }

        fmt::print(" | {:>9.1e} | {:>9.1e}\n", rel_err, (double)tol);
    }
}

// Structure to hold failure info for deferred printing
struct fail_info_t {
    std::string arch_name;
    std::vector<unsigned int> fail_indices;
    size_t output_idx;
    size_t arch_index;
    double max_err;
};

class volk_qa_aligned_mem_pool
{
public:
    void* get_new(size_t size)
    {
        size_t alignment = volk_get_alignment();
        void* ptr = volk_malloc(size, alignment);
        memset(ptr, 0x00, size);
        _mems.push_back(ptr);
        return ptr;
    }
    ~volk_qa_aligned_mem_pool()
    {
        for (unsigned int ii = 0; ii < _mems.size(); ++ii) {
            volk_free(_mems[ii]);
        }
    }

private:
    std::vector<void*> _mems;
};

bool run_volk_tests(volk_func_desc_t desc,
                    void (*manual_func)(),
                    std::string name,
                    volk_test_params_t test_params,
                    std::vector<volk_test_results_t>* results,
                    std::string puppet_master_name)
{
    return run_volk_tests(desc,
                          manual_func,
                          name,
                          test_params.tol(),
                          test_params.scalar(),
                          test_params.vlen(),
                          test_params.iter(),
                          results,
                          puppet_master_name,
                          test_params.absolute_mode(),
                          test_params.benchmark_mode(),
                          test_params.float_edge_cases(),
                          test_params.complex_edge_cases());
}

bool run_volk_tests(volk_func_desc_t desc,
                    void (*manual_func)(),
                    std::string name,
                    float tol,
                    lv_32fc_t scalar,
                    unsigned int vlen,
                    unsigned int iter,
                    std::vector<volk_test_results_t>* results,
                    std::string puppet_master_name,
                    bool absolute_mode,
                    bool benchmark_mode,
                    const std::vector<float>& float_edge_cases,
                    const std::vector<lv_32fc_t>& complex_edge_cases)
{
    // Initialize this entry in results vector
    results->push_back(volk_test_results_t());
    results->back().name = name;
    results->back().vlen = vlen;
    results->back().iter = iter;
    fmt::print(
        "\nRUN_VOLK_TESTS: {}(vlen={}, iter={}, tol={:.0e})\n", name, vlen, iter, tol);

    // vlen_twiddle will increase vlen for malloc and data generation
    // but kernels will still be called with the user provided vlen.
    // This is useful for causing errors in kernels that do bad reads
    const unsigned int vlen_twiddle = 5;
    vlen = vlen + vlen_twiddle;

    const float tol_f = tol;
    const unsigned int tol_i = static_cast<unsigned int>(tol);

    // first let's get a list of available architectures for the test
    std::vector<std::string> arch_list = get_arch_list(desc);

    // Build map from arch name to original index (for impl_alignment lookup)
    std::map<std::string, size_t> arch_to_orig_idx;
    for (size_t i = 0; i < arch_list.size(); i++) {
        arch_to_orig_idx[arch_list[i]] = i;
    }

    // Reorder arch_list to put generic implementations first for consistent output
    // Priority: "generic" first, then other generic_* variants, then everything else
    std::vector<std::string> plain_generic;
    std::vector<std::string> other_generic_impls;
    std::vector<std::string> other_impls;
    for (const auto& arch : arch_list) {
        if (arch == "generic") {
            plain_generic.push_back(arch);
        } else if (arch.find("generic") == 0) { // starts with "generic"
            other_generic_impls.push_back(arch);
        } else {
            other_impls.push_back(arch);
        }
    }
    arch_list.clear();
    arch_list.insert(arch_list.end(), plain_generic.begin(), plain_generic.end());
    arch_list.insert(
        arch_list.end(), other_generic_impls.begin(), other_generic_impls.end());
    arch_list.insert(arch_list.end(), other_impls.begin(), other_impls.end());

    if ((!benchmark_mode) && (arch_list.size() < 2)) {
        std::cout << "no architectures to test" << std::endl;
        return false;
    }

    // something that can hang onto memory and cleanup when this function exits
    volk_qa_aligned_mem_pool mem_pool;

    // now we have to get a function signature by parsing the name
    std::vector<volk_type_t> inputsig, outputsig;
    try {
        get_signatures_from_name(inputsig, outputsig, name);
    } catch (std::exception& error) {
        std::cerr << "Error: unable to get function signature from kernel name"
                  << std::endl;
        std::cerr << "  - " << name << std::endl;
        return false;
    }

    // pull the input scalars into their own vector
    std::vector<volk_type_t> inputsc;
    for (size_t i = 0; i < inputsig.size(); i++) {
        if (inputsig[i].is_scalar) {
            inputsc.push_back(inputsig[i]);
            inputsig.erase(inputsig.begin() + i);
            i -= 1;
        }
    }
    std::vector<void*> inbuffs;
    for (unsigned int inputsig_index = 0; inputsig_index < inputsig.size();
         ++inputsig_index) {
        volk_type_t sig = inputsig[inputsig_index];
        if (!sig.is_scalar) // we don't make buffers for scalars
            inbuffs.push_back(
                mem_pool.get_new(vlen * sig.size * (sig.is_complex ? 2 : 1)));
    }
    for (size_t i = 0; i < inbuffs.size(); i++) {
        load_random_data(
            inbuffs[i], inputsig[i], vlen, float_edge_cases, complex_edge_cases);
    }

    // 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 (size_t i = 0; i < arch_list.size(); i++) {
        std::vector<void*> arch_buffs;
        for (size_t j = 0; j < outputsig.size(); j++) {
            arch_buffs.push_back(mem_pool.get_new(vlen * outputsig[j].size *
                                                  (outputsig[j].is_complex ? 2 : 1)));
        }
        for (size_t j = 0; j < inputsig.size(); j++) {
            void* arch_inbuff = mem_pool.get_new(vlen * inputsig[j].size *
                                                 (inputsig[j].is_complex ? 2 : 1));
            memcpy(arch_inbuff,
                   inbuffs[j],
                   vlen * inputsig[j].size * (inputsig[j].is_complex ? 2 : 1));
            arch_buffs.push_back(arch_inbuff);
        }
        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
    vlen = vlen - vlen_twiddle;
    std::chrono::time_point<std::chrono::system_clock> start, end;
    std::vector<double> profile_times;

    // Warmup to let CPU reach full turbo frequency (only for first kernel)
    const double warmup_target_ms = g_warmup_done ? 0.0 : volk_test_get_warmup_ms();
    {
        // Run a quick test to estimate time per iteration
        start = std::chrono::system_clock::now();
        switch (both_sigs.size()) {
        case 1:
            if (inputsc.size() == 0) {
                run_cast_test1(
                    (volk_fn_1arg)(manual_func), test_data[0], vlen, iter, "generic");
            } else if (inputsc.size() == 1 && inputsc[0].is_float) {
                if (inputsc[0].is_complex) {
                    run_cast_test1_s32fc((volk_fn_1arg_s32fc)(manual_func),
                                         test_data[0],
                                         scalar,
                                         vlen,
                                         iter,
                                         "generic");
                } else {
                    run_cast_test1_s32f((volk_fn_1arg_s32f)(manual_func),
                                        test_data[0],
                                        scalar.real(),
                                        vlen,
                                        iter,
                                        "generic");
                }
            }
            break;
        case 2:
            if (inputsc.size() == 0) {
                run_cast_test2(
                    (volk_fn_2arg)(manual_func), test_data[0], vlen, iter, "generic");
            } else if (inputsc.size() == 1 && inputsc[0].is_float) {
                if (inputsc[0].is_complex) {
                    run_cast_test2_s32fc((volk_fn_2arg_s32fc)(manual_func),
                                         test_data[0],
                                         scalar,
                                         vlen,
                                         iter,
                                         "generic");
                } else {
                    run_cast_test2_s32f((volk_fn_2arg_s32f)(manual_func),
                                        test_data[0],
                                        scalar.real(),
                                        vlen,
                                        iter,
                                        "generic");
                }
            }
            break;
        case 3:
            if (inputsc.size() == 0) {
                run_cast_test3(
                    (volk_fn_3arg)(manual_func), test_data[0], vlen, iter, "generic");
            } else if (inputsc.size() == 1 && inputsc[0].is_float) {
                if (inputsc[0].is_complex) {
                    run_cast_test3_s32fc((volk_fn_3arg_s32fc)(manual_func),
                                         test_data[0],
                                         scalar,
                                         vlen,
                                         iter,
                                         "generic");
                } else {
                    run_cast_test3_s32f((volk_fn_3arg_s32f)(manual_func),
                                        test_data[0],
                                        scalar.real(),
                                        vlen,
                                        iter,
                                        "generic");
                }
            }
            break;
        case 4:
            run_cast_test4(
                (volk_fn_4arg)(manual_func), test_data[0], vlen, iter, "generic");
            break;
        default:
            break;
        }
        end = std::chrono::system_clock::now();
        std::chrono::duration<double> elapsed = end - start;
        double test_time_ms = 1000.0 * elapsed.count();

        // If we haven't reached 500ms yet, calculate how many more iterations we need
        if (test_time_ms < warmup_target_ms) {
            double remaining_ms = warmup_target_ms - test_time_ms;
            unsigned int warmup_iterations =
                (unsigned int)((remaining_ms / test_time_ms) * iter);
            if (warmup_iterations > 0) {
                // Run additional warmup iterations
                switch (both_sigs.size()) {
                case 1:
                    if (inputsc.size() == 0) {
                        run_cast_test1((volk_fn_1arg)(manual_func),
                                       test_data[0],
                                       vlen,
                                       warmup_iterations,
                                       "generic");
                    } else if (inputsc.size() == 1 && inputsc[0].is_float) {
                        if (inputsc[0].is_complex) {
                            run_cast_test1_s32fc((volk_fn_1arg_s32fc)(manual_func),
                                                 test_data[0],
                                                 scalar,
                                                 vlen,
                                                 warmup_iterations,
                                                 "generic");
                        } else {
                            run_cast_test1_s32f((volk_fn_1arg_s32f)(manual_func),
                                                test_data[0],
                                                scalar.real(),
                                                vlen,
                                                warmup_iterations,
                                                "generic");
                        }
                    }
                    break;
                case 2:
                    if (inputsc.size() == 0) {
                        run_cast_test2((volk_fn_2arg)(manual_func),
                                       test_data[0],
                                       vlen,
                                       warmup_iterations,
                                       "generic");
                    } else if (inputsc.size() == 1 && inputsc[0].is_float) {
                        if (inputsc[0].is_complex) {
                            run_cast_test2_s32fc((volk_fn_2arg_s32fc)(manual_func),
                                                 test_data[0],
                                                 scalar,
                                                 vlen,
                                                 warmup_iterations,
                                                 "generic");
                        } else {
                            run_cast_test2_s32f((volk_fn_2arg_s32f)(manual_func),
                                                test_data[0],
                                                scalar.real(),
                                                vlen,
                                                warmup_iterations,
                                                "generic");
                        }
                    }
                    break;
                case 3:
                    if (inputsc.size() == 0) {
                        run_cast_test3((volk_fn_3arg)(manual_func),
                                       test_data[0],
                                       vlen,
                                       warmup_iterations,
                                       "generic");
                    } else if (inputsc.size() == 1 && inputsc[0].is_float) {
                        if (inputsc[0].is_complex) {
                            run_cast_test3_s32fc((volk_fn_3arg_s32fc)(manual_func),
                                                 test_data[0],
                                                 scalar,
                                                 vlen,
                                                 warmup_iterations,
                                                 "generic");
                        } else {
                            run_cast_test3_s32f((volk_fn_3arg_s32f)(manual_func),
                                                test_data[0],
                                                scalar.real(),
                                                vlen,
                                                warmup_iterations,
                                                "generic");
                        }
                    }
                    break;
                case 4:
                    run_cast_test4((volk_fn_4arg)(manual_func),
                                   test_data[0],
                                   vlen,
                                   warmup_iterations,
                                   "generic");
                    break;
                default:
                    break;
                }
            }
        }
        g_warmup_done = true;
    }

    // Reset all test buffers after warmup
    for (size_t i = 0; i < arch_list.size(); i++) {
        for (size_t j = 0; j < outputsig.size(); j++) {
            memset(test_data[i][j],
                   0,
                   vlen * outputsig[j].size * (outputsig[j].is_complex ? 2 : 1));
        }
        // Reload input buffers from original data
        for (size_t j = 0; j < inputsig.size(); j++) {
            memcpy(test_data[i][outputsig.size() + j],
                   inbuffs[j],
                   vlen * inputsig[j].size * (inputsig[j].is_complex ? 2 : 1));
        }
    }

    for (size_t i = 0; i < arch_list.size(); i++) {
        start = std::chrono::system_clock::now();

        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) {
                if (inputsc[0].is_complex) {
                    run_cast_test1_s32fc((volk_fn_1arg_s32fc)(manual_func),
                                         test_data[i],
                                         scalar,
                                         vlen,
                                         iter,
                                         arch_list[i]);
                } else {
                    run_cast_test1_s32f((volk_fn_1arg_s32f)(manual_func),
                                        test_data[i],
                                        scalar.real(),
                                        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) {
                if (inputsc[0].is_complex) {
                    run_cast_test2_s32fc((volk_fn_2arg_s32fc)(manual_func),
                                         test_data[i],
                                         scalar,
                                         vlen,
                                         iter,
                                         arch_list[i]);
                } else {
                    run_cast_test2_s32f((volk_fn_2arg_s32f)(manual_func),
                                        test_data[i],
                                        scalar.real(),
                                        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) {
                if (inputsc[0].is_complex) {
                    run_cast_test3_s32fc((volk_fn_3arg_s32fc)(manual_func),
                                         test_data[i],
                                         scalar,
                                         vlen,
                                         iter,
                                         arch_list[i]);
                } else {
                    run_cast_test3_s32f((volk_fn_3arg_s32f)(manual_func),
                                        test_data[i],
                                        scalar.real(),
                                        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 = std::chrono::system_clock::now();
        std::chrono::duration<double> elapsed_seconds = end - start;
        double arch_time = 1000.0 * elapsed_seconds.count();

        volk_test_time_t result;
        result.name = arch_list[i];
        result.time = arch_time;
        result.units = "ms";
        result.pass = true;
        results->back().results[result.name] = result;

        profile_times.push_back(arch_time);
    }

    // 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...
    size_t generic_offset = 0;
    for (size_t i = 0; i < arch_list.size(); i++) {
        if (arch_list[i] == "generic") {
            generic_offset = i;
        }
    }

    // Just in case a kernel wrote to OOB memory, use the twiddled vlen
    vlen = vlen + vlen_twiddle;
    bool fail;
    bool fail_global = false;
    std::vector<bool> arch_results;

    // Collect input buffers for error reporting (inputs are after outputs in test_data)
    std::vector<void*> input_buffs;
    for (size_t k = outputsig.size(); k < both_sigs.size(); k++) {
        input_buffs.push_back(test_data[generic_offset][k]);
    }

    // Collect failures for deferred printing (after timing summary)
    std::vector<fail_info_t> failures;

    // Track max error per architecture (absolute or relative depending on mode)
    std::vector<double> arch_max_err(arch_list.size(), 0.0);

    for (size_t i = 0; i < arch_list.size(); i++) {
        fail = false;
        if (i != generic_offset) {
            for (size_t j = 0; j < outputsig.size(); j++) {
                std::vector<unsigned int> fail_indices;
                double max_err = 0.0;
                if (both_sigs[j].is_float) {
                    if (both_sigs[j].size == 8) {
                        if (both_sigs[j].is_complex) {
                            fail = ccompare((double*)test_data[generic_offset][j],
                                            (double*)test_data[i][j],
                                            vlen,
                                            tol_f,
                                            absolute_mode,
                                            fail_indices,
                                            max_err);
                        } else {
                            fail = fcompare((double*)test_data[generic_offset][j],
                                            (double*)test_data[i][j],
                                            vlen,
                                            tol_f,
                                            absolute_mode,
                                            fail_indices,
                                            max_err);
                        }
                    } else {
                        if (both_sigs[j].is_complex) {
                            fail = ccompare((float*)test_data[generic_offset][j],
                                            (float*)test_data[i][j],
                                            vlen,
                                            tol_f,
                                            absolute_mode,
                                            fail_indices,
                                            max_err);
                        } else {
                            fail = fcompare((float*)test_data[generic_offset][j],
                                            (float*)test_data[i][j],
                                            vlen,
                                            tol_f,
                                            absolute_mode,
                                            fail_indices,
                                            max_err);
                        }
                    }
                } 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_i,
                                            fail_indices,
                                            max_err);
                        } 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_i,
                                            fail_indices,
                                            max_err);
                        }
                        break;
                    case 4:
                        if (both_sigs[j].is_complex) {
                            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_i,
                                                fail_indices,
                                                max_err);
                            } 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_i,
                                                fail_indices,
                                                max_err);
                            }
                        } else {
                            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_i,
                                                fail_indices,
                                                max_err);
                            } 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_i,
                                                fail_indices,
                                                max_err);
                            }
                        }
                        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_i,
                                            fail_indices,
                                            max_err);
                        } 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_i,
                                            fail_indices,
                                            max_err);
                        }
                        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_i,
                                            fail_indices,
                                            max_err);
                        } 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_i,
                                            fail_indices,
                                            max_err);
                        }
                        break;
                    default:
                        fail = 1;
                    }
                }
                // Track max error for this arch across all outputs
                if (max_err > arch_max_err[i]) {
                    arch_max_err[i] = max_err;
                }
                if (fail) {
                    volk_test_time_t* result = &results->back().results[arch_list[i]];
                    result->pass = false;
                    fail_global = true;
                    // Store failure info for later printing
                    fail_info_t fi;
                    fi.max_err = max_err;
                    fi.arch_name = arch_list[i];
                    fi.fail_indices = fail_indices;
                    fi.output_idx = j;
                    fi.arch_index = i;
                    failures.push_back(fi);
                }
            }
        }
        arch_results.push_back(!fail);
    }

    double best_time_a = std::numeric_limits<double>::max();
    double best_time_u = std::numeric_limits<double>::max();
    std::string best_arch_a = "generic";
    std::string best_arch_u = "generic";
    for (size_t i = 0; i < arch_list.size(); i++) {
        // Look up alignment using original index (before reordering)
        size_t orig_idx = arch_to_orig_idx[arch_list[i]];
        bool requires_alignment = desc.impl_alignment[orig_idx];

        if ((profile_times[i] < best_time_u) && arch_results[i] && !requires_alignment) {
            best_time_u = profile_times[i];
            best_arch_u = arch_list[i];
        }
        if ((profile_times[i] < best_time_a) && arch_results[i]) {
            best_time_a = profile_times[i];
            best_arch_a = arch_list[i];
        }
    }

    // Unaligned implementations (alignment == 0) work on any memory alignment.
    // If an unaligned impl is faster than all aligned impls, use it for both.
    if (best_time_u < best_time_a) {
        best_time_a = best_time_u;
        best_arch_a = best_arch_u;
    }

    // Calculate total data transferred (bytes read + written) for throughput display
    size_t bytes_per_call = 0;
    for (size_t j = 0; j < outputsig.size(); j++) {
        bytes_per_call += outputsig[j].size * (outputsig[j].is_complex ? 2 : 1) * vlen;
    }
    for (size_t j = 0; j < inputsig.size(); j++) {
        bytes_per_call += inputsig[j].size * (inputsig[j].is_complex ? 2 : 1) * vlen;
    }
    double total_mb = (bytes_per_call * iter) / 1e6; // Total megabytes transferred

    // Get generic timing for speedup calculation
    double generic_time = 0.0;
    for (size_t i = 0; i < arch_list.size(); i++) {
        if (arch_list[i] == "generic") {
            generic_time = profile_times[i];
            break;
        }
    }

    // Column widths for results table
    constexpr int w_arch = 26;
    constexpr int w_time = 14;
    constexpr int w_tput = 14;
    constexpr int w_speedup = 8;
    constexpr int w_err = 10;

    // Column header depends on error mode
    // Integer outputs always use absolute comparison, so show "max_abs" for them
    bool has_int_output = false;
    for (const auto& sig : outputsig) {
        if (!sig.is_float) {
            has_int_output = true;
            break;
        }
    }
    const char* err_col = (absolute_mode || has_int_output) ? "max_abs" : "max_rel";

    // Helper for adaptive decimal places based on magnitude
    auto format_time = [](double ms) -> std::string {
        return fmt::format("{:.2f} ms", ms);
    };

    auto format_throughput = [](double mbps) -> std::string {
        return fmt::format("{:.1f} MB/s", mbps);
    };

    // Print table header
    fmt::print("{:<{}} | {:>{}} | {:>{}} | {:>{}} | {:>{}} |\n",
               "arch",
               w_arch,
               "time",
               w_time,
               "throughput",
               w_tput,
               "speedup",
               w_speedup,
               err_col,
               w_err);
    fmt::print("{:-<{}}-+-{:-<{}}-+-{:-<{}}-+-{:-<{}}-+-{:-<{}}-+\n",
               "",
               w_arch,
               "",
               w_time,
               "",
               w_tput,
               "",
               w_speedup,
               "",
               w_err);

    // Print each architecture row
    for (size_t i = 0; i < arch_list.size(); i++) {
        double time_seconds = profile_times[i] / 1000.0;
        double throughput_mbps = total_mb / time_seconds;

        std::string time_str = format_time(profile_times[i]);
        std::string tput_str = format_throughput(throughput_mbps);
        std::string speedup_str;
        if (arch_list[i] == "generic" || generic_time <= 0) {
            speedup_str = "-";
        } else {
            double speedup = generic_time / profile_times[i];
            speedup_str = fmt::format("{:.2f}x", speedup);
        }
        std::string err_str =
            (arch_list[i] == "generic") ? "-" : fmt::format("{:.1e}", arch_max_err[i]);
        std::string win_str =
            (arch_list[i] == best_arch_a || arch_list[i] == best_arch_u) ? " *" : "";

        fmt::print("{:<{}} | {:>{}} | {:>{}} | {:>{}} | {:>{}} |{}\n",
                   arch_list[i],
                   w_arch,
                   time_str,
                   w_time,
                   tput_str,
                   w_tput,
                   speedup_str,
                   w_speedup,
                   err_str,
                   w_err,
                   win_str);
    }

    // Print best arch summary (left-aligned, ":" at arch column width)
    auto print_best_line = [&](const char* label, const std::string& arch, double time) {
        std::string speedup_str;
        if (arch != "generic" && generic_time > 0) {
            speedup_str = fmt::format(" ({:.2f}x)", generic_time / time);
        }
        fmt::print("{:<{}} {}{}\n", label, w_arch, arch, speedup_str);
    };

    print_best_line("Best aligned arch          |", best_arch_a, best_time_a);
    print_best_line("Best unaligned arch        |", best_arch_u, best_time_u);

    // Print failure details after timing summary
    for (const auto& fi : failures) {
        fmt::print("\n{}: fail on arch {}\n", name, fi.arch_name);
        print_error_table(fi.fail_indices,
                          input_buffs,
                          inputsig,
                          test_data[generic_offset][fi.output_idx],
                          test_data[fi.arch_index][fi.output_idx],
                          outputsig[fi.output_idx],
                          tol_f);
    }

    fmt::print("{:-<88}\n", "");

    if (puppet_master_name == "NULL") {
        results->back().config_name = name;
    } else {
        results->back().config_name = puppet_master_name;
    }
    results->back().best_arch_a = best_arch_a;
    results->back().best_arch_u = best_arch_u;

    return fail_global;
}