#include "commons/commons_tinyrenderer.h"

#include "commons/commons_job_thread_pool.h"
#include "commons/commons_macros.h"
#include "commons/commons_minmax.h"
#include "commons/commons_tga.h"
#include "commons/commons_vec.h"

void API TinyRenderer_Triangle(tga_image& dst, uint2 A, uint2 B, uint2 C, tga_pixel Fill)
{
    struct worker_data
    {
        tga_image* dst;
        tga_pixel Fill;
        uint32_t rowBegin;
        uint32_t rowEnd;
        uint32_t rowLength;
    };

    auto worker = [](void* userdata) {
        worker_data* data = (worker_data*)userdata;

        for (uint32_t y = data->rowBegin; y < data->rowEnd; ++y) {
            for (uint32_t x = 0; x < data->rowLength; ++x) {
                TGA_Set(*data->dst, x, y, data->Fill);
            }
        }
    };

    uint32_t yMin = Min(A[1], B[1], C[1]);
    uint32_t yMax = Max(A[1], B[1], C[1]);
    uint32_t xMin = Min(A[0], B[0], C[0]);
    uint32_t xMax = Max(A[0], B[0], C[0]);

    uint32_t NumRows = yMax - yMin;
    uint32_t RowLength = xMax - xMin;

    uint32_t WorkerCount = JobThreadPool_GetWorkerCount();
    uint32_t RowsPerWorker = NumRows / WorkerCount;

    for (uint32_t i = 0; i < WorkerCount - 1; ++i) {
        worker_data Data {
            .dst = &dst,
            .Fill = Fill,
            .rowBegin = i * RowsPerWorker,
            .rowEnd = (i + 1) * RowsPerWorker,
            .rowLength = RowLength,
        };

        JobThreadPool_Submit(job {
            .fn = worker,
            .data = (void*)&Data,
        });
    }

    {
        worker_data Data {
            .dst = &dst,
            .Fill = Fill,
            .rowBegin = (WorkerCount - 1) * RowsPerWorker,
            .rowEnd = NumRows,
            .rowLength = RowLength,
        };

        JobThreadPool_Submit(job {
            .fn = worker,
            .data = (void*)&Data,
        });
    }

    JobThreadPool_WaitAllDone();
}