// accelVid.c -- Accelerated video driver manager
//
// Manages registration, detection, and lifecycle of hardware-specific
// video drivers. Drivers register themselves at startup, then the
// manager probes each in order to find matching hardware.
//
// After a chip driver's init() succeeds, the manager fills in
// software fallback implementations for any drawing operations
// the driver left as NULL. This means callers never need to
// check function pointers -- every operation is always callable.
// The fallbacks draw directly to the LFB using simple loops.

#include "accelVid.h"

#include <stdio.h>
#include <stdlib.h>
#include <string.h>

// Maximum number of registered drivers. This is more than enough
// for all chip families we'll ever support.
#define MAX_DRIVERS 32

// ============================================================
// Prototypes -- public API
// ============================================================

AccelDriverT *accelDetect(void);
uint32_t      accelGetCaps(const AccelDriverT *drv);
const char   *accelGetName(const AccelDriverT *drv);
bool          accelInit(AccelDriverT *drv, const AccelModeRequestT *req);
void          accelRegisterDriver(AccelDriverT *drv);
void          accelShutdown(AccelDriverT *drv);

// ============================================================
// Prototypes -- software fallbacks
// ============================================================

static void swBitBlt(AccelDriverT *drv, int32_t srcX, int32_t srcY, int32_t dstX, int32_t dstY, int32_t w, int32_t h);
static void swColorExpand(AccelDriverT *drv, const uint8_t *srcBuf, int32_t srcPitch, int32_t dstX, int32_t dstY, int32_t w, int32_t h, uint32_t fg, uint32_t bg);
static void swHostBlit(AccelDriverT *drv, const uint8_t *srcBuf, int32_t srcPitch, int32_t dstX, int32_t dstY, int32_t w, int32_t h);
static void swLineDraw(AccelDriverT *drv, int32_t x1, int32_t y1, int32_t x2, int32_t y2, uint32_t color);
static void swRectFill(AccelDriverT *drv, int32_t x, int32_t y, int32_t w, int32_t h, uint32_t color);
static void swRectFillPat(AccelDriverT *drv, int32_t x, int32_t y, int32_t w, int32_t h, const uint8_t *pattern, uint32_t fg, uint32_t bg);
static void swSetClip(AccelDriverT *drv, int32_t x, int32_t y, int32_t w, int32_t h);
static void swWaitIdle(AccelDriverT *drv);
static void swInstallFallbacks(AccelDriverT *drv);

// ============================================================
// Inline helpers for software fallbacks
// ============================================================

// Write a pixel at (x, y) in the framebuffer. No bounds checking
// -- the caller must clip before calling.
static inline void swPutPixel(AccelDriverT *drv, int32_t x, int32_t y, uint32_t color) {
    uint8_t *fb  = drv->mode.framebuffer;
    int32_t  bpp = (drv->mode.bpp + 7) / 8;
    uint8_t *dst = fb + y * drv->mode.pitch + x * bpp;

    switch (bpp) {
        case 1:
            *dst = (uint8_t)color;
            break;
        case 2:
            *(uint16_t *)dst = (uint16_t)color;
            break;
        case 4:
            *(uint32_t *)dst = color;
            break;
    }
}

// ============================================================
// Module state
// ============================================================

static AccelDriverT *sDrivers[MAX_DRIVERS];
static int32_t       sDriverCount = 0;

// Software clip rectangle (used by fallbacks when no hardware clip)
static int32_t sClipX = 0;
static int32_t sClipY = 0;
static int32_t sClipW = 0;
static int32_t sClipH = 0;

// ============================================================
// accelDetect
// ============================================================
//
// Iterates all registered drivers and calls detect() on each.
// Returns the first driver that claims the hardware, or NULL
// if no supported hardware is found.
//
// Detection order matters: drivers registered first are tried
// first. This allows callers to prioritize specific drivers
// (e.g. prefer S3 over generic VESA).

AccelDriverT *accelDetect(void) {
    if (!pciDetect()) {
        fprintf(stderr, "accelVid: PCI bus not detected\n");
        return NULL;
    }

    for (int32_t i = 0; i < sDriverCount; i++) {
        if (sDrivers[i]->detect(sDrivers[i])) {
            printf("accelVid: Detected %s (PCI %02X:%02X.%X, "
                   "vendor=%04X device=%04X)\n",
                   sDrivers[i]->name,
                   sDrivers[i]->pciDev.bus,
                   sDrivers[i]->pciDev.dev,
                   sDrivers[i]->pciDev.func,
                   sDrivers[i]->pciDev.vendorId,
                   sDrivers[i]->pciDev.deviceId);
            return sDrivers[i];
        }
    }

    fprintf(stderr, "accelVid: No supported video hardware found\n");
    return NULL;
}


// ============================================================
// accelGetCaps
// ============================================================

uint32_t accelGetCaps(const AccelDriverT *drv) {
    if (!drv) {
        return 0;
    }

    return drv->caps;
}


// ============================================================
// accelGetName
// ============================================================

const char *accelGetName(const AccelDriverT *drv) {
    if (!drv) {
        return "none";
    }

    return drv->name;
}


// ============================================================
// accelInit
// ============================================================

bool accelInit(AccelDriverT *drv, const AccelModeRequestT *req) {
    if (!drv || !drv->init) {
        return false;
    }

    memset(&drv->mode, 0, sizeof(drv->mode));

    if (!drv->init(drv, req)) {
        fprintf(stderr, "accelVid: Failed to initialize %s\n", drv->name);
        return false;
    }

    printf("accelVid: Initialized %s at %ldx%ldx%ld (pitch=%ld, vram=%luKB)\n",
           drv->name,
           (long)drv->mode.width,
           (long)drv->mode.height,
           (long)drv->mode.bpp,
           (long)drv->mode.pitch,
           (unsigned long)(drv->mode.vramSize / 1024));

    // Report capabilities
    printf("accelVid: Capabilities:");

    if (drv->caps & ACAP_RECT_FILL) {
        printf(" RectFill");
    }
    if (drv->caps & ACAP_RECT_FILL_PAT) {
        printf(" PatFill");
    }
    if (drv->caps & ACAP_BITBLT) {
        printf(" BitBlt");
    }
    if (drv->caps & ACAP_COLOR_EXPAND) {
        printf(" ColorExpand");
    }
    if (drv->caps & ACAP_LINE_DRAW) {
        printf(" LineDraw");
    }
    if (drv->caps & ACAP_HW_CURSOR) {
        printf(" HwCursor");
    }
    if (drv->caps & ACAP_HOST_BLIT) {
        printf(" HostBlit");
    }
    if (drv->caps & ACAP_CLIP) {
        printf(" Clip");
    }
    if (drv->caps & ACAP_TRANSPARENCY) {
        printf(" Transparency");
    }

    printf("\n");

    // Install software fallbacks for any operations the driver
    // didn't implement in hardware
    swInstallFallbacks(drv);

    return true;
}


// ============================================================
// accelRegisterDriver
// ============================================================

void accelRegisterDriver(AccelDriverT *drv) {
    if (sDriverCount >= MAX_DRIVERS) {
        fprintf(stderr, "accelVid: Too many drivers registered (max %d)\n",
                MAX_DRIVERS);
        return;
    }

    sDrivers[sDriverCount++] = drv;
}


// ============================================================
// accelShutdown
// ============================================================

void accelShutdown(AccelDriverT *drv) {
    if (!drv) {
        return;
    }

    if (drv->waitIdle) {
        drv->waitIdle(drv);
    }

    if (drv->showCursor) {
        drv->showCursor(drv, false);
    }

    if (drv->shutdown) {
        drv->shutdown(drv);
    }

    memset(&drv->mode, 0, sizeof(drv->mode));
}


// ============================================================
// Software fallback implementations
// ============================================================
//
// These draw directly to the LFB. They're correct but slow
// (uncached PCI writes). The point isn't performance -- it's
// ensuring every operation is always callable so the caller
// never needs to check for NULL function pointers.


// ============================================================
// swBitBlt
// ============================================================
//
// Screen-to-screen blit via the LFB. Handles overlapping regions
// by choosing copy direction.

static void swBitBlt(AccelDriverT *drv, int32_t srcX, int32_t srcY, int32_t dstX, int32_t dstY, int32_t w, int32_t h) {
    if (w <= 0 || h <= 0) {
        return;
    }

    uint8_t *fb    = drv->mode.framebuffer;
    int32_t  pitch = drv->mode.pitch;
    int32_t  bpp   = (drv->mode.bpp + 7) / 8;
    int32_t  rowBytes = w * bpp;

    if (dstY < srcY || (dstY == srcY && dstX <= srcX)) {
        // Copy forward (top to bottom, left to right)
        for (int32_t row = 0; row < h; row++) {
            uint8_t *src = fb + (srcY + row) * pitch + srcX * bpp;
            uint8_t *dst = fb + (dstY + row) * pitch + dstX * bpp;
            memmove(dst, src, rowBytes);
        }
    } else {
        // Copy backward (bottom to top)
        for (int32_t row = h - 1; row >= 0; row--) {
            uint8_t *src = fb + (srcY + row) * pitch + srcX * bpp;
            uint8_t *dst = fb + (dstY + row) * pitch + dstX * bpp;
            memmove(dst, src, rowBytes);
        }
    }
}


// ============================================================
// swColorExpand
// ============================================================
//
// Monochrome-to-color expansion via the LFB. Each 1-bit in srcBuf
// becomes the fg color, each 0-bit becomes the bg color.

static void swColorExpand(AccelDriverT *drv, const uint8_t *srcBuf, int32_t srcPitch, int32_t dstX, int32_t dstY, int32_t w, int32_t h, uint32_t fg, uint32_t bg) {
    if (w <= 0 || h <= 0) {
        return;
    }

    uint8_t *fb    = drv->mode.framebuffer;
    int32_t  pitch = drv->mode.pitch;
    int32_t  bpp   = (drv->mode.bpp + 7) / 8;

    for (int32_t row = 0; row < h; row++) {
        const uint8_t *mono = srcBuf + row * srcPitch;
        uint8_t       *dst  = fb + (dstY + row) * pitch + dstX * bpp;

        for (int32_t col = 0; col < w; col++) {
            int32_t  byteIdx = col / 8;
            int32_t  bitIdx  = 7 - (col % 8);
            uint32_t color   = (mono[byteIdx] >> bitIdx) & 1 ? fg : bg;

            switch (bpp) {
                case 1:
                    dst[col] = (uint8_t)color;
                    break;
                case 2:
                    ((uint16_t *)dst)[col] = (uint16_t)color;
                    break;
                case 4:
                    ((uint32_t *)dst)[col] = color;
                    break;
            }
        }
    }
}


// ============================================================
// swHostBlit
// ============================================================
//
// CPU-to-screen blit via the LFB. Just a memcpy per scanline.

static void swHostBlit(AccelDriverT *drv, const uint8_t *srcBuf, int32_t srcPitch, int32_t dstX, int32_t dstY, int32_t w, int32_t h) {
    if (w <= 0 || h <= 0) {
        return;
    }

    uint8_t *fb       = drv->mode.framebuffer;
    int32_t  pitch    = drv->mode.pitch;
    int32_t  bpp      = (drv->mode.bpp + 7) / 8;
    int32_t  rowBytes = w * bpp;

    for (int32_t row = 0; row < h; row++) {
        const uint8_t *src = srcBuf + row * srcPitch;
        uint8_t       *dst = fb + (dstY + row) * pitch + dstX * bpp;
        memcpy(dst, src, rowBytes);
    }
}


// ============================================================
// swLineDraw
// ============================================================
//
// Bresenham line draw via the LFB.

static void swLineDraw(AccelDriverT *drv, int32_t x1, int32_t y1, int32_t x2, int32_t y2, uint32_t color) {
    int32_t dx  = abs(x2 - x1);
    int32_t dy  = abs(y2 - y1);
    int32_t sx  = (x1 < x2) ? 1 : -1;
    int32_t sy  = (y1 < y2) ? 1 : -1;
    int32_t err = dx - dy;
    int32_t x   = x1;
    int32_t y   = y1;

    for (;;) {
        if (x >= sClipX && x < sClipX + sClipW &&
            y >= sClipY && y < sClipY + sClipH) {
            swPutPixel(drv, x, y, color);
        }

        if (x == x2 && y == y2) {
            break;
        }

        int32_t e2 = 2 * err;

        if (e2 > -dy) {
            err -= dy;
            x   += sx;
        }

        if (e2 < dx) {
            err += dx;
            y   += sy;
        }
    }
}


// ============================================================
// swRectFill
// ============================================================
//
// Solid rectangle fill via the LFB.

static void swRectFill(AccelDriverT *drv, int32_t x, int32_t y, int32_t w, int32_t h, uint32_t color) {
    if (w <= 0 || h <= 0) {
        return;
    }

    uint8_t *fb    = drv->mode.framebuffer;
    int32_t  pitch = drv->mode.pitch;
    int32_t  bpp   = (drv->mode.bpp + 7) / 8;

    for (int32_t row = 0; row < h; row++) {
        uint8_t *dst = fb + (y + row) * pitch + x * bpp;

        switch (bpp) {
            case 1:
                memset(dst, (uint8_t)color, w);
                break;
            case 2: {
                uint16_t *dst16 = (uint16_t *)dst;
                for (int32_t col = 0; col < w; col++) {
                    dst16[col] = (uint16_t)color;
                }
                break;
            }
            case 4: {
                uint32_t *dst32 = (uint32_t *)dst;
                for (int32_t col = 0; col < w; col++) {
                    dst32[col] = color;
                }
                break;
            }
        }
    }
}


// ============================================================
// swRectFillPat
// ============================================================
//
// 8x8 monochrome pattern fill via the LFB. The pattern is 8 bytes,
// one bit per pixel, MSB-first, row 0 first. Each 1-bit gets the
// fg color, each 0-bit gets the bg color. The pattern tiles across
// the destination rectangle with alignment to screen coordinates
// (so patterns line up across adjacent fills).

static void swRectFillPat(AccelDriverT *drv, int32_t x, int32_t y, int32_t w, int32_t h, const uint8_t *pattern, uint32_t fg, uint32_t bg) {
    if (w <= 0 || h <= 0) {
        return;
    }

    uint8_t *fb    = drv->mode.framebuffer;
    int32_t  pitch = drv->mode.pitch;
    int32_t  bpp   = (drv->mode.bpp + 7) / 8;

    for (int32_t row = 0; row < h; row++) {
        uint8_t  patRow = pattern[(y + row) & 7];
        uint8_t *dst    = fb + (y + row) * pitch + x * bpp;

        for (int32_t col = 0; col < w; col++) {
            int32_t  patBit = 7 - ((x + col) & 7);
            uint32_t color  = (patRow >> patBit) & 1 ? fg : bg;

            switch (bpp) {
                case 1:
                    dst[col] = (uint8_t)color;
                    break;
                case 2:
                    ((uint16_t *)dst)[col] = (uint16_t)color;
                    break;
                case 4:
                    ((uint32_t *)dst)[col] = color;
                    break;
            }
        }
    }
}


// ============================================================
// swSetClip
// ============================================================
//
// Software clip rectangle for fallback line drawing.

static void swSetClip(AccelDriverT *drv, int32_t x, int32_t y, int32_t w, int32_t h) {
    (void)drv;
    sClipX = x;
    sClipY = y;
    sClipW = w;
    sClipH = h;
}


// ============================================================
// swWaitIdle
// ============================================================
//
// No-op -- software operations complete synchronously.

static void swWaitIdle(AccelDriverT *drv) {
    (void)drv;
}


// ============================================================
// swInstallFallbacks
// ============================================================
//
// Fills in software implementations for any NULL function
// pointers in the driver struct. Called by accelInit() after
// the chip driver's init() succeeds. This guarantees that
// every drawing operation is always callable.

static void swInstallFallbacks(AccelDriverT *drv) {
    int32_t count = 0;

    if (!drv->waitIdle) {
        drv->waitIdle = swWaitIdle;
    }

    if (!drv->setClip) {
        drv->setClip = swSetClip;
        count++;
    }

    if (!drv->rectFill) {
        drv->rectFill = swRectFill;
        count++;
    }

    if (!drv->bitBlt) {
        drv->bitBlt = swBitBlt;
        count++;
    }

    if (!drv->hostBlit) {
        drv->hostBlit = swHostBlit;
        count++;
    }

    if (!drv->colorExpand) {
        drv->colorExpand = swColorExpand;
        count++;
    }

    if (!drv->rectFillPat) {
        drv->rectFillPat = swRectFillPat;
        count++;
    }

    if (!drv->lineDraw) {
        drv->lineDraw = swLineDraw;
        count++;
    }

    // Initialize the software clip rect to full screen
    sClipX = 0;
    sClipY = 0;
    sClipW = drv->mode.width;
    sClipH = drv->mode.height;

    if (count > 0) {
        printf("accelVid: %ld operation(s) using software fallback\n",
               (long)count);
    }
}