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

#include <cstring>

#if defined(__APPLE__)
#include <sys/time.h>
#endif

static int BuildKernel(const char *name, int vectorSize, cl_uint kernel_count,
                       cl_kernel *k, cl_program *p, bool relaxedMode)
{
    const char *c[] = { "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
                        "__kernel void math_kernel",
                        sizeNames[vectorSize],
                        "( __global double",
                        sizeNames[vectorSize],
                        "* out, __global double",
                        sizeNames[vectorSize],
                        "* in )\n"
                        "{\n"
                        "   size_t i = get_global_id(0);\n"
                        "   out[i] = ",
                        name,
                        "( in[i] );\n"
                        "}\n" };

    const char *c3[] = {
        "#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n",
        "__kernel void math_kernel",
        sizeNames[vectorSize],
        "( __global double* out, __global double* in)\n"
        "{\n"
        "   size_t i = get_global_id(0);\n"
        "   if( i + 1 < get_global_size(0) )\n"
        "   {\n"
        "       double3 f0 = vload3( 0, in + 3 * i );\n"
        "       f0 = ",
        name,
        "( f0 );\n"
        "       vstore3( f0, 0, out + 3*i );\n"
        "   }\n"
        "   else\n"
        "   {\n"
        "       size_t parity = i & 1;   // Figure out how many elements are "
        "left over after BUFFER_SIZE % (3*sizeof(float)). Assume power of two "
        "buffer size \n"
        "       double3 f0;\n"
        "       switch( parity )\n"
        "       {\n"
        "           case 1:\n"
        "               f0 = (double3)( in[3*i], NAN, NAN ); \n"
        "               break;\n"
        "           case 0:\n"
        "               f0 = (double3)( in[3*i], in[3*i+1], NAN ); \n"
        "               break;\n"
        "       }\n"
        "       f0 = ",
        name,
        "( f0 );\n"
        "       switch( parity )\n"
        "       {\n"
        "           case 0:\n"
        "               out[3*i+1] = f0.y; \n"
        "               // fall through\n"
        "           case 1:\n"
        "               out[3*i] = f0.x; \n"
        "               break;\n"
        "       }\n"
        "   }\n"
        "}\n"
    };

    const char **kern = c;
    size_t kernSize = sizeof(c) / sizeof(c[0]);

    if (sizeValues[vectorSize] == 3)
    {
        kern = c3;
        kernSize = sizeof(c3) / sizeof(c3[0]);
    }

    char testName[32];
    snprintf(testName, sizeof(testName) - 1, "math_kernel%s",
             sizeNames[vectorSize]);

    return MakeKernels(kern, (cl_uint)kernSize, testName, kernel_count, k, p,
                       relaxedMode);
}

typedef struct BuildKernelInfo
{
    cl_uint offset; // the first vector size to build
    cl_uint kernel_count;
    cl_kernel **kernels;
    cl_program *programs;
    const char *nameInCode;
    bool relaxedMode; // Whether to build with -cl-fast-relaxed-math.
} BuildKernelInfo;

static cl_int BuildKernelFn(cl_uint job_id, cl_uint thread_id UNUSED, void *p)
{
    BuildKernelInfo *info = (BuildKernelInfo *)p;
    cl_uint i = info->offset + job_id;
    return BuildKernel(info->nameInCode, i, info->kernel_count,
                       info->kernels[i], info->programs + i, info->relaxedMode);
}

// Thread specific data for a worker thread
typedef struct ThreadInfo
{
    cl_mem inBuf; // input buffer for the thread
    cl_mem outBuf[VECTOR_SIZE_COUNT]; // output buffers for the thread
    float maxError; // max error value. Init to 0.
    double maxErrorValue; // position of the max error value.  Init to 0.
    cl_command_queue tQueue; // per thread command queue to improve performance
} ThreadInfo;

typedef struct TestInfo
{
    size_t subBufferSize; // Size of the sub-buffer in elements
    const Func *f; // A pointer to the function info
    cl_program programs[VECTOR_SIZE_COUNT]; // programs for various vector sizes
    cl_kernel
        *k[VECTOR_SIZE_COUNT]; // arrays of thread-specific kernels for each
                               // worker thread:  k[vector_size][thread_id]
    ThreadInfo *
        tinfo; // An array of thread specific information for each worker thread
    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.
} TestInfo;

static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data);

int TestFunc_Double_Double(const Func *f, MTdata d, bool relaxedMode)
{
    TestInfo test_info;
    cl_int error;
    size_t i, j;
    float maxError = 0.0f;
    double maxErrorVal = 0.0;
#if defined(__APPLE__)
    struct timeval time_val;
    gettimeofday(&time_val, NULL);
    double start_time = time_val.tv_sec + 1e-6 * time_val.tv_usec;
    double end_time;
#endif

    logFunctionInfo(f->name, sizeof(cl_double), relaxedMode);
    // Init test_info
    memset(&test_info, 0, sizeof(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));
    if (gWimpyMode)
    {
        test_info.subBufferSize = gWimpyBufferSize
            / (sizeof(cl_double)
               * RoundUpToNextPowerOfTwo(test_info.threadCount));
    }

    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;

    // cl_kernels aren't thread safe, so we make one for each vector size for
    // every thread
    for (i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
    {
        size_t array_size = test_info.threadCount * sizeof(cl_kernel);
        test_info.k[i] = (cl_kernel *)malloc(array_size);
        if (NULL == test_info.k[i])
        {
            vlog_error("Error: Unable to allocate storage for kernels!\n");
            error = CL_OUT_OF_HOST_MEMORY;
            goto exit;
        }
        memset(test_info.k[i], 0, array_size);
    }
    test_info.tinfo =
        (ThreadInfo *)malloc(test_info.threadCount * sizeof(*test_info.tinfo));
    if (NULL == test_info.tinfo)
    {
        vlog_error(
            "Error: Unable to allocate storage for thread specific data.\n");
        error = CL_OUT_OF_HOST_MEMORY;
        goto exit;
    }
    memset(test_info.tinfo, 0,
           test_info.threadCount * sizeof(*test_info.tinfo));
    for (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);
            goto exit;
        }

        for (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);
                goto exit;
            }
        }
        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);
            goto exit;
        }
    }

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

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

        // Accumulate the arithmetic errors
        for (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 (error) goto exit;

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


#if defined(__APPLE__)
    gettimeofday(&time_val, NULL);
    end_time = time_val.tv_sec + 1e-6 * time_val.tv_usec;
#endif

    if (gMeasureTimes)
    {
        // Init input array
        double *p = (double *)gIn;

        if (strstr(f->name, "exp"))
            for (j = 0; j < BUFFER_SIZE / sizeof(double); j++)
                p[j] = (double)genrand_real1(d);
        else if (strstr(f->name, "log"))
            for (j = 0; j < BUFFER_SIZE / sizeof(double); j++)
                p[j] = fabs(DoubleFromUInt32(genrand_int32(d)));
        else
            for (j = 0; j < BUFFER_SIZE / sizeof(double); j++)
                p[j] = DoubleFromUInt32(genrand_int32(d));
        if ((error = clEnqueueWriteBuffer(gQueue, gInBuffer, CL_FALSE, 0,
                                          BUFFER_SIZE, gIn, 0, NULL, NULL)))
        {
            vlog_error("\n*** Error %d in clEnqueueWriteBuffer ***\n", error);
            return error;
        }


        // Run the kernels
        for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
        {
            size_t vectorSize = sizeof(cl_double) * sizeValues[j];
            size_t localCount = (BUFFER_SIZE + vectorSize - 1)
                / vectorSize; // BUFFER_SIZE / vectorSize  rounded up
            if ((error = clSetKernelArg(test_info.k[j][0], 0,
                                        sizeof(gOutBuffer[j]), &gOutBuffer[j])))
            {
                LogBuildError(test_info.programs[j]);
                goto exit;
            }
            if ((error = clSetKernelArg(test_info.k[j][0], 1, sizeof(gInBuffer),
                                        &gInBuffer)))
            {
                LogBuildError(test_info.programs[j]);
                goto exit;
            }

            double sum = 0.0;
            double bestTime = INFINITY;
            for (i = 0; i < PERF_LOOP_COUNT; i++)
            {
                uint64_t startTime = GetTime();
                if ((error = clEnqueueNDRangeKernel(gQueue, test_info.k[j][0],
                                                    1, NULL, &localCount, NULL,
                                                    0, NULL, NULL)))
                {
                    vlog_error("FAILED -- could not execute kernel\n");
                    goto exit;
                }

                // Make sure OpenCL is done
                if ((error = clFinish(gQueue)))
                {
                    vlog_error("Error %d at clFinish\n", error);
                    goto exit;
                }

                uint64_t endTime = GetTime();
                double time = SubtractTime(endTime, startTime);
                sum += time;
                if (time < bestTime) bestTime = time;
            }

            if (gReportAverageTimes) bestTime = sum / PERF_LOOP_COUNT;
            double clocksPerOp = bestTime * (double)gDeviceFrequency
                * gComputeDevices * gSimdSize * 1e6
                / (BUFFER_SIZE / sizeof(double));
            vlog_perf(clocksPerOp, LOWER_IS_BETTER, "clocks / element", "%sD%s",
                      f->name, sizeNames[j]);
        }
    }

    if (!gSkipCorrectnessTesting) vlog("\t%8.2f @ %a", maxError, maxErrorVal);

#if defined(__APPLE__)
    vlog("\t(%2.2f seconds)", end_time - start_time);
#endif
    vlog("\n");

exit:
    // Release
    for (i = gMinVectorSizeIndex; i < gMaxVectorSizeIndex; i++)
    {
        clReleaseProgram(test_info.programs[i]);
        if (test_info.k[i])
        {
            for (j = 0; j < test_info.threadCount; j++)
                clReleaseKernel(test_info.k[i][j]);

            free(test_info.k[i]);
        }
    }
    if (test_info.tinfo)
    {
        for (i = 0; i < test_info.threadCount; i++)
        {
            clReleaseMemObject(test_info.tinfo[i].inBuf);
            for (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
                clReleaseMemObject(test_info.tinfo[i].outBuf[j]);
            clReleaseCommandQueue(test_info.tinfo[i].tQueue);
        }

        free(test_info.tinfo);
    }

    return error;
}

static cl_int Test(cl_uint job_id, cl_uint thread_id, void *data)
{
    const TestInfo *job = (const 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_uint j, k;
    cl_int error;
    int ftz = job->ftz;

    Force64BitFPUPrecision();

    // start the map of the output arrays
    cl_event e[VECTOR_SIZE_COUNT];
    cl_ulong *out[VECTOR_SIZE_COUNT];
    for (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 (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 (j = gMinVectorSizeIndex; j < gMaxVectorSizeIndex; j++)
    {
        // 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;
        memset_pattern4(out[j], &pattern, buffer_size);
        if ((error = clEnqueueUnmapMemObject(tinfo->tQueue, tinfo->outBuf[j],
                                             out[j], 0, NULL, NULL)))
        {
            vlog_error("Error: clEnqueueMapBuffer 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 (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. This is
    // an in order queue.
    for (j = gMinVectorSizeIndex; j + 1 < gMaxVectorSizeIndex; j++)
    {
        out[j] = (cl_ulong *)clEnqueueMapBuffer(
            tinfo->tQueue, tinfo->outBuf[j], CL_FALSE, 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;
        }
    }
    // Wait for the last buffer
    out[j] = (cl_ulong *)clEnqueueMapBuffer(tinfo->tQueue, tinfo->outBuf[j],
                                            CL_TRUE, 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 (j = 0; j < buffer_elements; j++)
    {
        for (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)
                    {
                        // 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%16.16llx): *%.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 (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;
}