#include "pxl8_blit.h"

void pxl8_blit_hicolor(u16* fb, u32 fb_width, const u16* sprite, u32 atlas_width,
                       i32 x, i32 y, u32 w, u32 h) {
    u16* dest_base = fb + y * fb_width + x;
    const u16* src_base = sprite;

    for (u32 row = 0; row < h; row++) {
        u16* dest_row = dest_base + row * fb_width;
        const u16* src_row = src_base + row * atlas_width;

        u32 col = 0;
        u32 count2 = w / 2;
        for (u32 i = 0; i < count2; i++) {
            u32 pixels = ((const u32*)src_row)[i];
            if (pixels == 0) {
                col += 2;
                continue;
            }
            u16 s0 = (u16)(pixels);
            u16 s1 = (u16)(pixels >> 16);
            u16 d0 = dest_row[col];
            u16 d1 = dest_row[col + 1];
            u16 m0 = (u16)(-(s0 != 0));
            u16 m1 = (u16)(-(s1 != 0));
            dest_row[col]     = (s0 & m0) | (d0 & ~m0);
            dest_row[col + 1] = (s1 & m1) | (d1 & ~m1);
            col += 2;
        }
        if (w & 1) {
            u16 s = src_row[col];
            u16 d = dest_row[col];
            u16 m = (u16)(-(s != 0));
            dest_row[col] = (s & m) | (d & ~m);
        }
    }
}

void pxl8_blit_indexed(u8* fb, u32 fb_width, const u8* sprite, u32 atlas_width,
                       i32 x, i32 y, u32 w, u32 h) {
    u8* dest_base = fb + y * fb_width + x;
    const u8* src_base = sprite;

    for (u32 row = 0; row < h; row++) {
        u8* dest_row = dest_base + row * fb_width;
        const u8* src_row = src_base + row * atlas_width;

        u32 col = 0;
        u32 count4 = w / 4;
        for (u32 i = 0; i < count4; i++) {
            u32 pixels = ((const u32*)src_row)[i];
            if (pixels == 0) {
                col += 4;
                continue;
            }
            u8 s0 = (u8)(pixels);
            u8 s1 = (u8)(pixels >> 8);
            u8 s2 = (u8)(pixels >> 16);
            u8 s3 = (u8)(pixels >> 24);
            u8 d0 = dest_row[col];
            u8 d1 = dest_row[col + 1];
            u8 d2 = dest_row[col + 2];
            u8 d3 = dest_row[col + 3];
            u8 m0 = (u8)(-(s0 != 0));
            u8 m1 = (u8)(-(s1 != 0));
            u8 m2 = (u8)(-(s2 != 0));
            u8 m3 = (u8)(-(s3 != 0));
            dest_row[col]     = (s0 & m0) | (d0 & ~m0);
            dest_row[col + 1] = (s1 & m1) | (d1 & ~m1);
            dest_row[col + 2] = (s2 & m2) | (d2 & ~m2);
            dest_row[col + 3] = (s3 & m3) | (d3 & ~m3);
            col += 4;
        }
        for (; col < w; col++) {
            u8 s = src_row[col];
            u8 d = dest_row[col];
            u8 m = (u8)(-(s != 0));
            dest_row[col] = (s & m) | (d & ~m);
        }
    }
}