//
// Copyright (c) 2017 The Khronos Group Inc.
//
// 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.
//

#include "common.h"
#include "function_list.h"
#include "test_functions.h"
#include "utility.h"

#include <cinttypes>
#include <cstring>

namespace {

cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
{
    BuildKernelInfo &info = *(BuildKernelInfo *)p;
    auto generator = [](const std::string &kernel_name, const char *builtin,
                        cl_uint vector_size_index) {
        return GetUnaryKernel(kernel_name, builtin, ParameterType::Double,
                              ParameterType::Double, vector_size_index);
    };
    return BuildKernels(info, job_id, generator);
}

// Thread specific data for a worker thread
struct ThreadInfo
{
    // Input and output buffers for the thread
    clMemWrapper inBuf;
    Buffers outBuf;

    float maxError; // max error value. Init to 0.
    double maxErrorValue; // position of the max error value.  Init to 0.

    // Per thread command queue to improve performance
    clCommandQueueWrapper tQueue;
};

struct TestInfo
{
    size_t subBufferSize; // Size of the sub-buffer in elements
    const Func *f; // A pointer to the function info

    // Programs for various vector sizes.
    Programs programs;

    // Thread-specific kernels for each vector size:
    // k[vector_size][thread_id]
    KernelMatrix k;

    // Array of thread specific information
    std::vector<ThreadInfo> tinfo;

    cl_uint threadCount; // Number of worker threads
    cl_uint jobCount; // Number of jobs
    cl_uint step; // step between each chunk and the next.
    cl_uint scale; // stride between individual test values
    float ulps; // max_allowed ulps
    int ftz; // non-zero if running in flush to zero mode

    int isRangeLimited; // 1 if the function is only to be evaluated over a
                        // range
    float half_sin_cos_tan_limit;
    bool relaxedMode; // True if test is running in relaxed mode, false
                      // otherwise.
};

cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
{
    TestInfo *job = (TestInfo *)data;
    size_t buffer_elements = job->subBufferSize;
    size_t buffer_size = buffer_elements * sizeof(cl_double);
    cl_uint scale = job->scale;
    cl_uint base = job_id * (cl_uint)job->step;
    ThreadInfo *tinfo = &(job->tinfo[thread_id]);
    float ulps = job->ulps;
    dptr func = job->f->dfunc;
    cl_int error;
    int ftz = job->ftz;
    bool relaxedMode = job->relaxedMode;

    Force64BitFPUPrecision();

    cl_event e[VECTOR_SIZE_COUNT];
    cl_ulong *out[VECTOR_SIZE_COUNT];
    if (gHostFill)
    {
        // start the map of the output arrays
        for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
        {
            out[j] = (cl_ulong *)clEnqueueMapBuffer(
                tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, CL_MAP_WRITE, 0,
                buffer_size, 0, NULL, e + j, &error);
            if (error || NULL == out[j])
            {
                vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
                           error);
                return error;
            }
        }

        // Get that moving
        if ((error = clFlush(tinfo->tQueue))) vlog("clFlush failed\n");
    }

    // Write the new values to the input array
    cl_double *p = (cl_double *)gIn + thread_id * buffer_elements;
    for (size_t j = 0; j < buffer_elements; j++)
        p[j] = DoubleFromUInt32(base + j * scale);

    if ((error = clEnqueueWriteBuffer(tinfo->tQueue, tinfo->inBuf, CL_FALSE, 0,
                                      buffer_size, p, 0, NULL, NULL)))
    {
        vlog_error("Error: clEnqueueWriteBuffer failed! err: %d\n", error);
        return error;
    }

    for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
    {
        if (gHostFill)
        {
            // Wait for the map to finish
            if ((error = clWaitForEvents(1, e + j)))
            {
                vlog_error("Error: clWaitForEvents failed! err: %d\n", error);
                return error;
            }
            if ((error = clReleaseEvent(e[j])))
            {
                vlog_error("Error: clReleaseEvent failed! err: %d\n", error);
                return error;
            }
        }

        // Fill the result buffer with garbage, so that old results don't carry
        // over
        uint32_t pattern = 0xffffdead;
        if (gHostFill)
        {
            memset_pattern4(out[j], &pattern, buffer_size);
            if ((error = clEnqueueUnmapMemObject(
                     tinfo->tQueue, tinfo->outBuf[j], out[j], 0, NULL, NULL)))
            {
                vlog_error("Error: clEnqueueUnmapMemObject failed! err: %d\n",
                           error);
                return error;
            }
        }
        else
        {
            if ((error = clEnqueueFillBuffer(tinfo->tQueue, tinfo->outBuf[j],
                                             &pattern, sizeof(pattern), 0,
                                             buffer_size, 0, NULL, NULL)))
            {
                vlog_error("Error: clEnqueueFillBuffer failed! err: %d\n",
                           error);
                return error;
            }
        }

        // Run the kernel
        size_t vectorCount =
            (buffer_elements + sizeValues[j] - 1) / sizeValues[j];
        cl_kernel kernel = job->k[j][thread_id]; // each worker thread has its
                                                 // own copy of the cl_kernel
        cl_program program = job->programs[j];

        if ((error = clSetKernelArg(kernel, 0, sizeof(tinfo->outBuf[j]),
                                    &tinfo->outBuf[j])))
        {
            LogBuildError(program);
            return error;
        }
        if ((error = clSetKernelArg(kernel, 1, sizeof(tinfo->inBuf),
                                    &tinfo->inBuf)))
        {
            LogBuildError(program);
            return error;
        }

        if ((error = clEnqueueNDRangeKernel(tinfo->tQueue, kernel, 1, NULL,
                                            &vectorCount, NULL, 0, NULL, NULL)))
        {
            vlog_error("FAILED -- could not execute kernel\n");
            return error;
        }
    }


    // Get that moving
    if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 2 failed\n");

    if (gSkipCorrectnessTesting) return CL_SUCCESS;

    // Calculate the correctly rounded reference result
    cl_double *r = (cl_double *)gOut_Ref + thread_id * buffer_elements;
    cl_double *s = (cl_double *)p;
    for (size_t j = 0; j < buffer_elements; j++)
        r[j] = (cl_double)func.f_f(s[j]);

    // Read the data back -- no need to wait for the first N-1 buffers but wait
    // for the last buffer. This is an in order queue.
    for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
    {
        cl_bool blocking = (j + 1 < gMaxVectorSizeIndex) ? CL_FALSE : CL_TRUE;
        out[j] = (cl_ulong *)clEnqueueMapBuffer(
            tinfo->tQueue, tinfo->outBuf[j], blocking, CL_MAP_READ, 0,
            buffer_size, 0, NULL, NULL, &error);
        if (error || NULL == out[j])
        {
            vlog_error("Error: clEnqueueMapBuffer %d failed! err: %d\n", j,
                       error);
            return error;
        }
    }

    // Verify data
    cl_ulong *t = (cl_ulong *)r;
    for (size_t j = 0; j < buffer_elements; j++)
    {
        for (auto k = gMinVectorSizeIndex; k < gMaxVectorSizeIndex; k++)
        {
            cl_ulong *q = out[k];

            // If we aren't getting the correctly rounded result
            if (t[j] != q[j])
            {
                cl_double test = ((cl_double *)q)[j];
                long double correct = func.f_f(s[j]);
                float err = Bruteforce_Ulp_Error_Double(test, correct);
                int fail = !(fabsf(err) <= ulps);

                if (fail)
                {
                    if (ftz || relaxedMode)
                    {
                        // retry per section 6.5.3.2
                        if (IsDoubleResultSubnormal(correct, ulps))
                        {
                            fail = fail && (test != 0.0f);
                            if (!fail) err = 0.0f;
                        }

                        // retry per section 6.5.3.3
                        if (IsDoubleSubnormal(s[j]))
                        {
                            long double correct2 = func.f_f(0.0L);
                            long double correct3 = func.f_f(-0.0L);
                            float err2 =
                                Bruteforce_Ulp_Error_Double(test, correct2);
                            float err3 =
                                Bruteforce_Ulp_Error_Double(test, correct3);
                            fail = fail
                                && ((!(fabsf(err2) <= ulps))
                                    && (!(fabsf(err3) <= ulps)));
                            if (fabsf(err2) < fabsf(err)) err = err2;
                            if (fabsf(err3) < fabsf(err)) err = err3;

                            // retry per section 6.5.3.4
                            if (IsDoubleResultSubnormal(correct2, ulps)
                                || IsDoubleResultSubnormal(correct3, ulps))
                            {
                                fail = fail && (test != 0.0f);
                                if (!fail) err = 0.0f;
                            }
                        }
                    }
                }
                if (fabsf(err) > tinfo->maxError)
                {
                    tinfo->maxError = fabsf(err);
                    tinfo->maxErrorValue = s[j];
                }
                if (fail)
                {
                    vlog_error("\nERROR: %s%s: %f ulp error at %.13la "
                               "(0x%16.16" PRIx64 "): *%.13la vs. %.13la\n",
                               job->f->name, sizeNames[k], err,
                               ((cl_double *)gIn)[j], ((cl_ulong *)gIn)[j],
                               ((cl_double *)gOut_Ref)[j], test);
                    return -1;
                }
            }
        }
    }

    for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
    {
        if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
                                             out[j], 0, NULL, NULL)))
        {
            vlog_error("Error: clEnqueueUnmapMemObject %d failed 2! err: %d\n",
                       j, error);
            return error;
        }
    }

    if ((error = clFlush(tinfo->tQueue))) vlog("clFlush 3 failed\n");


    if (0 == (base & 0x0fffffff))
    {
        if (gVerboseBruteForce)
        {
            vlog("base:%14u step:%10u scale:%10zd buf_elements:%10u ulps:%5.3f "
                 "ThreadCount:%2u\n",
                 base, job->step, buffer_elements, job->scale, job->ulps,
                 job->threadCount);
        }
        else
        {
            vlog(".");
        }
        fflush(stdout);
    }

    return CL_SUCCESS;
}

} // anonymous namespace

int TestFunc_Double_Double(const Func *f, MTdata d, bool relaxedMode)
{
    TestInfo test_info{};
    cl_int error;
    float maxError = 0.0f;
    double maxErrorVal = 0.0;

    logFunctionInfo(f->name, sizeof(cl_double), relaxedMode);
    // Init test_info
    test_info.threadCount = GetThreadCount();
    test_info.subBufferSize = BUFFER_SIZE
        / (sizeof(cl_double) * RoundUpToNextPowerOfTwo(test_info.threadCount));
    test_info.scale = getTestScale(sizeof(cl_double));

    test_info.step = (cl_uint)test_info.subBufferSize * test_info.scale;
    if (test_info.step / test_info.subBufferSize != test_info.scale)
    {
        // there was overflow
        test_info.jobCount = 1;
    }
    else
    {
        test_info.jobCount = (cl_uint)((1ULL << 32) / test_info.step);
    }

    test_info.f = f;
    test_info.ulps = f->double_ulps;
    test_info.ftz = f->ftz || gForceFTZ;
    test_info.relaxedMode = relaxedMode;

    test_info.tinfo.resize(test_info.threadCount);
    for (cl_uint i = 0; i < test_info.threadCount; i++)
    {
        cl_buffer_region region = {
            i * test_info.subBufferSize * sizeof(cl_double),
            test_info.subBufferSize * sizeof(cl_double)
        };
        test_info.tinfo[i].inBuf =
            clCreateSubBuffer(gInBuffer, CL_MEM_READ_ONLY,
                              CL_BUFFER_CREATE_TYPE_REGION, &region, &error);
        if (error || NULL == test_info.tinfo[i].inBuf)
        {
            vlog_error("Error: Unable to create sub-buffer of gInBuffer for "
                       "region {%zd, %zd}\n",
                       region.origin, region.size);
            return error;
        }

        for (auto j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
        {
            test_info.tinfo[i].outBuf[j] = clCreateSubBuffer(
                gOutBuffer[j], CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION,
                &region, &error);
            if (error || NULL == test_info.tinfo[i].outBuf[j])
            {
                vlog_error("Error: Unable to create sub-buffer of "
                           "gOutBuffer[%d] for region {%zd, %zd}\n",
                           (int)j, region.origin, region.size);
                return error;
            }
        }
        test_info.tinfo[i].tQueue =
            clCreateCommandQueue(gContext, gDevice, 0, &error);
        if (NULL == test_info.tinfo[i].tQueue || error)
        {
            vlog_error("clCreateCommandQueue failed. (%d)\n", error);
            return error;
        }
    }

    // Init the kernels
    BuildKernelInfo build_info{ test_info.threadCount, test_info.k,
                                test_info.programs, f->nameInCode,
                                relaxedMode };
    if ((error = ThreadPool_Do(BuildKernelFn,
                               gMaxVectorSizeIndex - gMinVectorSizeIndex,
                               &build_info)))
        return error;

    // Run the kernels
    if (!gSkipCorrectnessTesting)
    {
        error = ThreadPool_Do(Test, test_info.jobCount, &test_info);
        if (error) return error;

        // Accumulate the arithmetic errors
        for (cl_uint i = 0; i < test_info.threadCount; i++)
        {
            if (test_info.tinfo[i].maxError > maxError)
            {
                maxError = test_info.tinfo[i].maxError;
                maxErrorVal = test_info.tinfo[i].maxErrorValue;
            }
        }

        if (gWimpyMode)
            vlog("Wimp pass");
        else
            vlog("passed");

        vlog("\t%8.2f @ %a", maxError, maxErrorVal);
    }

    vlog("\n");

    return CL_SUCCESS;
}