/*
 * $Id$
 *
 * $Date$
 * $Revision$
 *
 * (C) 1999 by Hyperion
 * All rights reserved
 *
 * This file is part of the MiniGL library project
 * See the file Licence.txt for more details
 *
 */

// TO DO - optimize all of this byte-reading nonsense!

#include "displaylists.h"
#include "sysinc.h"
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <assert.h>

#include "mgl_profileitems.h"

extern char *format_to_string(GLenum format);
extern char *type_to_string(GLenum type);

static inline uint16 load_16_little(uint16 *addr)
{
	uint16 x;
	__asm volatile ("lhbrx %0, 0, %1" : "=r" (x) : "r" (addr));
	return x;
}

static inline void store_16_little(uint16 *addr, uint16 val)
{
	__asm volatile ("sthbrx %0, 0, %1" :: "r" (val), "r" (addr));
}

static inline uint32 load_32_little(uint32 *addr)
{
	uint32 x;
	__asm volatile ("lwbrx %0, 0, %1" : "=r" (x) : "r" (addr));
	return x;
}

static inline void store_32_little(uint32 *addr, uint32 val)
{
	__asm volatile ("stwbrx %0, 0, %1" :: "r" (val), "r" (addr));
}

#define A_COMPONENT 0
#define R_COMPONENT	1
#define G_COMPONENT 2
#define B_COMPONENT 3


static void _pack_rgb_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{

	PROFILE_ENTRY(FID_PACK_RGB_B)

	uint8 *from = (uint8 *)src;
	uint8 *to = (uint8 *)dst->current;

	/* Note: swap_bytes not applicable to three-byte pixels */
	while (num_pixels--) {
		*to++ = from[R_COMPONENT];
		*to++ = from[G_COMPONENT];
		*to++ = from[B_COMPONENT];

		from += 4;
	}

	PROFILE_EXIT(FID_PACK_RGB_B)

}

static void _pack_rgba_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_RGBA_B)

	uint32 *from = (uint32 *)src;
	uint32 *to = (uint32 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			uint32 temp = load_32_little(from++);
			*to++ = temp >> 8 | temp << 24;
		}
	}
	else {
		while (num_pixels--) {
			*to++ = (*from << 8) | (*from >> 24); from++;
		}
	}

	PROFILE_EXIT(FID_PACK_RGBA_B)

}

static void _pack_abgr_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_ABGR_B)

	uint32 *from = (uint32 *)src;
	uint32 *to = (uint32 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			*to++ = (*from << 8) | (*from >> 24); from++;
		}
	}
	else {
		while (num_pixels--) {
			uint32 temp = load_32_little(from++);
			*to++ = temp >> 8 | temp << 24;
		}
	}

	PROFILE_EXIT(FID_PACK_ABGR_B)

}


static void _pack_argb_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{

	uint32 *from = (uint32 *)src;
	uint32 *to = (uint32 *)dst->current;

	/* Note: assumes BGRA byte reversed */
	if (!pack->swap_bytes) {
		while (num_pixels--) {
			*to++ = load_32_little(from++);
		}
	}
	else {
		while (num_pixels--) {
			*to++ = *from++;
		}
	}
}


static void _pack_a_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_A_B)

	uint8 *from = (uint8 *)src;
	uint8 *to = (uint8 *)dst->current;

	while (num_pixels--) {
		*to++ = from[A_COMPONENT];
		from += 4;
	}

	PROFILE_EXIT(FID_PACK_A_B)
}

static void _pack_l_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_L_B)

	uint8 *from = (uint8 *)src;
	uint8 *to = (uint8 *)dst->current;

	while (num_pixels--) {
		uint32 l = from[R_COMPONENT] + from[G_COMPONENT] + from[B_COMPONENT];
		if (l > 255) {
			l = 255;
		}
		*to++ = l;
		from += 4;
	}

	PROFILE_EXIT(FID_PACK_L_B)
}


static void _pack_la_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_LA_B)

	uint8 *from = (uint8 *)src;
	uint8 *to = (uint8 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			uint32 l = from[R_COMPONENT] + from[G_COMPONENT] + from[B_COMPONENT];
			if (l > 255) {
				l = 255;
			}
			*to++ = from[A_COMPONENT];
			*to++ = l;
			from += 4;
		}
	}
	else {
		while (num_pixels--) {
			uint32 l = from[R_COMPONENT] + from[G_COMPONENT] + from[B_COMPONENT];
			if (l > 255) {
				l = 255;
			}
			*to++ = l;
			*to++ = from[A_COMPONENT];
			from += 4;
		}
	}

	PROFILE_EXIT(FID_PACK_LA_B)

}


static void _pack_i_b(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_I_B)

	uint8 *from = (uint8 *)src;
	uint8 *to = (uint8 *)dst->current;

	while (num_pixels--) {
		*to++ = from[R_COMPONENT];
		from += 4;
	}

	PROFILE_EXIT(FID_PACK_I_B)

}

static void _pack_rgb_us565(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_RGB_US565)

	uint8 *from = (uint8 *)src;
	uint16 *to = (uint16 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			store_16_little(to, ((from[R_COMPONENT] & 0xf8) << 8)
								| ((from[G_COMPONENT] & 0xfc) << 3)
								| ((from[B_COMPONENT] & 0xf8) >> 3));
			from += 4;
			to++;
		}
	}
	else {
		while (num_pixels--) {
			*to++ = ((from[R_COMPONENT] & 0xf8) << 8)
		 		  | ((from[G_COMPONENT] & 0xfc) << 3)
		   		  | ((from[B_COMPONENT] & 0xf8) >> 3);
			from += 4;
		}
	}

	PROFILE_EXIT(FID_PACK_RGB_US565)

}

static void _pack_rgb_us565r(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_RGB_US565R)

	uint8 *from = (uint8 *)src;
	uint16 *to = (uint16 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			store_16_little(to, ((from[B_COMPONENT] & 0xf8) << 8)
								| ((from[G_COMPONENT] & 0xfc) << 3)
								| ((from[R_COMPONENT] & 0xf8) >> 3)
			);
			from += 4;
			to++;
		}
	}
	else {
		while (num_pixels--) {
			*to++ = ((from[B_COMPONENT] & 0xf8) << 8)
			      | ((from[G_COMPONENT] & 0xfc) << 3)
			      | ((from[R_COMPONENT] & 0xf8) >> 3);
			from += 4;
		}
	}

	PROFILE_ENTRY(FID_PACK_RGB_US565R)

}

static void _pack_rgba_us4444(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_RGBA_US4444)

	uint8 *from = (uint8 *)src;
	uint16 *to = (uint16 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			store_16_little(to, ((from[R_COMPONENT] & 0xf0) << 8)
								| ((from[G_COMPONENT] & 0xf0) << 4)
								| ((from[B_COMPONENT] & 0xf0)     )
								| ((from[A_COMPONENT] & 0xf0) >> 4));
			from += 4;
			to++;
		}
	}
	else {
		while (num_pixels--) {
			*to++ = ((from[R_COMPONENT] & 0xf0) << 8)
						| ((from[G_COMPONENT] & 0xf0) << 4)
						| ((from[B_COMPONENT] & 0xf0)     )
						| ((from[A_COMPONENT] & 0xf0) >> 4);
			from += 4;
		}
	}

	PROFILE_EXIT(FID_PACK_RGBA_US4444)

}


static void _pack_rgba_us4444r(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_RGBA_US4444R)

	uint8 *from = (uint8 *)src;
	uint16 *to = (uint16 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			store_16_little(to, ((from[A_COMPONENT] & 0xf0) << 8)
								| ((from[B_COMPONENT] & 0xf0) << 4)
								| ((from[G_COMPONENT] & 0xf0)     )
								| ((from[R_COMPONENT] & 0xf0) >> 4)
			);
			from += 4;
			to++;
		}
	}
	else {
		while (num_pixels--) {
			*to++ = ((from[A_COMPONENT] & 0xf0) << 8)
						| ((from[B_COMPONENT] & 0xf0) << 4)
						| ((from[G_COMPONENT] & 0xf0)     )
						| ((from[R_COMPONENT] & 0xf0) >> 4);
			from += 4;
		}
	}

	PROFILE_EXIT(FID_PACK_RGBA_US4444R)
}


static void _pack_rgba_us1555(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_RGBA_US1555)

	uint8 *from = (uint8 *)src;
	uint16 *to = (uint16 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			store_16_little(to, ((from[A_COMPONENT] & 0x80) << 8)
							| ((from[R_COMPONENT] & 0xf8) << 7)
							| ((from[G_COMPONENT] & 0xf8) << 2)
							| ((from[B_COMPONENT] & 0xf8) >> 3)
			);
			from += 4;
			to++;
		}
	}
	else {
		while (num_pixels--) {
			*to++ = ((from[A_COMPONENT] & 0x80) << 8)
			      | ((from[R_COMPONENT] & 0xf8) << 7)
			      | ((from[G_COMPONENT] & 0xf8) << 2)
			      | ((from[B_COMPONENT] & 0xf8) >> 3);
			from += 4;
		}
	}

	PROFILE_EXIT(FID_PACK_RGBA_US1555)

}

static void _pack_rgba_us1555r(GLcontext context, int num_pixels, void *src, GLimage_info *dst, GLpixel_store *pack)
{
	PROFILE_ENTRY(FID_PACK_RGBA_US1555R)

	uint8 *from = (uint8 *)src;
	uint16 *to = (uint16 *)dst->current;

	if (pack->swap_bytes) {
		while (num_pixels--) {
			store_16_little(to, ((from[A_COMPONENT] & 0x80) >> 7)
							| ((from[B_COMPONENT] & 0xf8) << 8)
							| ((from[G_COMPONENT] & 0xf8) << 3)
							| ((from[R_COMPONENT] & 0xf8) >> 2));
			from += 4;
			to++;
		}
	}
	else {
		while (num_pixels--) {
			*to++ =((from[A_COMPONENT] & 0x80) >> 7)
					| ((from[B_COMPONENT] & 0xf8) << 8)
					| ((from[G_COMPONENT] & 0xf8) << 3)
					| ((from[R_COMPONENT] & 0xf8) >> 2);
			from += 4;
		}
	}

	PROFILE_EXIT(FID_PACK_RGBA_US1555R)

}

GLpack mglPackers[] =
{
	{GL_RGB,             GL_UNSIGNED_BYTE,              _pack_rgb_b,        3},
	{3,                  GL_UNSIGNED_BYTE,              _pack_rgb_b,        3},
	{GL_RGBA,            GL_UNSIGNED_BYTE,              _pack_rgba_b,       4},
	{4,                  GL_UNSIGNED_BYTE,              _pack_rgba_b,       4},
	{GL_RGBA,            GL_UNSIGNED_INT_8_8_8_8_REV,              _pack_abgr_b,       4},
	/* Note: BGRA with unsigned int reversed is ARGB */
	{GL_BGRA,            GL_UNSIGNED_INT_8_8_8_8_REV,   _pack_argb_b,       4},
	{GL_ALPHA,           GL_UNSIGNED_BYTE,              _pack_a_b,          1},
	{GL_LUMINANCE,       GL_UNSIGNED_BYTE,              _pack_l_b,          1},
	{1,                  GL_UNSIGNED_BYTE,              _pack_l_b,          1},
	{GL_LUMINANCE_ALPHA, GL_UNSIGNED_BYTE,              _pack_la_b,         2},
	{2,                  GL_UNSIGNED_BYTE,              _pack_la_b,         2},
	{GL_INTENSITY,       GL_UNSIGNED_BYTE,              _pack_i_b,          1},
	{GL_RGB5,            GL_UNSIGNED_SHORT_5_6_5,       _pack_rgb_us565,    2},
	{GL_RGB,             GL_UNSIGNED_SHORT_5_6_5,       _pack_rgb_us565,    2},
	{GL_RGB5,            GL_UNSIGNED_SHORT_5_6_5_REV,   _pack_rgb_us565r,   2},
	{GL_RGB,             GL_UNSIGNED_SHORT_5_6_5_REV,   _pack_rgb_us565r,   2},
	{GL_RGB4,            GL_UNSIGNED_SHORT_4_4_4_4,     _pack_rgba_us4444,  2},
	{GL_RGB,             GL_UNSIGNED_SHORT_4_4_4_4,     _pack_rgba_us4444,  2},
	{GL_RGB4,            GL_UNSIGNED_SHORT_4_4_4_4_REV, _pack_rgba_us4444r, 2},
	{GL_RGB,             GL_UNSIGNED_SHORT_4_4_4_4_REV, _pack_rgba_us4444r, 2},
	{GL_RGB5_A1,         GL_UNSIGNED_SHORT_5_5_5_1,     _pack_rgba_us1555,  2},
	{GL_RGB,             GL_UNSIGNED_SHORT_5_5_5_1,     _pack_rgba_us1555,  2},
	{GL_RGB5_A1,         GL_UNSIGNED_SHORT_1_5_5_5_REV, _pack_rgba_us1555r, 2},
	{GL_RGB,             GL_UNSIGNED_SHORT_1_5_5_5_REV, _pack_rgba_us1555r, 2},
	{GL_NONE,            GL_NONE,                       NULL,               0},
};

_glPackFn MGLSelectPacker(GLenum format, GLenum type, GLuint *pixelStride)
{
	int i = 0;
	while (mglPackers[i].format != GL_NONE) {
		if (format == mglPackers[i].format && type == mglPackers[i].type) {
			if (pixelStride) {
				*pixelStride = mglPackers[i].stride;
			}
			return mglPackers[i].pack;
		}
		i++;
	}

	IExec->DebugPrintF(
		"Missing pack support for %s/%s\n",
		format_to_string(format), type_to_string(type)
	);

	return NULL;
}