smatz@8091: /* $Id$ */
smatz@8091: 
rubidium@9111: /** @file 32bpp_optimized.cpp Implementation of the optimized 32 bpp blitter. */
rubidium@9111: 
truelight@7553: #include "../stdafx.h"
rubidium@8123: #include "../zoom_func.h"
rubidium@8123: #include "../gfx_func.h"
truelight@7553: #include "../debug.h"
smatz@9597: #include "../core/math_func.hpp"
smatz@9597: #include "../core/alloc_func.hpp"
truelight@7553: #include "32bpp_optimized.hpp"
truelight@7553: 
truelight@7553: static FBlitter_32bppOptimized iFBlitter_32bppOptimized;
truelight@7553: 
smatz@9597: /**
smatz@9597:  * Draws a sprite to a (screen) buffer. It is templated to allow faster operation.
smatz@9597:  *
smatz@9597:  * @param mode blitter mode
smatz@9597:  * @param bp further blitting parameters
smatz@9597:  * @param zoom zoom level at which we are drawing
smatz@9597:  */
smatz@9597: template <BlitterMode mode>
smatz@9597: inline void Blitter_32bppOptimized::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
truelight@7553: {
smatz@9597: 	const SpriteData *src = (const SpriteData *)bp->sprite;
truelight@7553: 
smatz@9597: 	/* src_px : each line begins with uint32 n = 'number of bytes in this line',
smatz@9597: 	 *          then n times is the Colour struct for this line */
smatz@9597: 	const Colour *src_px = (const Colour *)(src->data + src->offset[zoom][0]);
smatz@9597: 	/* src_n  : each line begins with uint32 n = 'number of bytes in this line',
smatz@9597: 	 *          then interleaved stream of 'm' and 'n' channels. 'm' is remap,
smatz@9597: 	 *          'n' is number of bytes with the same alpha channel class */
smatz@9597: 	const uint8  *src_n  = (const uint8  *)(src->data + src->offset[zoom][1]);
smatz@9597: 
smatz@9597: 	/* skip upper lines in src_px and src_n */
smatz@9597: 	for (uint i = bp->skip_top; i != 0; i--) {
smatz@9597: 		src_px = (const Colour *)((const byte *)src_px + *(const uint32 *)src_px);
smatz@9597: 		src_n += *(uint32 *)src_n;
smatz@9597: 	}
smatz@9597: 
smatz@9597: 	/* skip lines in dst */
smatz@9597: 	uint32 *dst = (uint32 *)bp->dst + bp->top * bp->pitch + bp->left;
smatz@9597: 
smatz@9597: 	/* store so we don't have to access it via bp everytime (compiler assumes pointer aliasing) */
smatz@9597: 	const byte *remap = bp->remap;
truelight@7553: 
truelight@7553: 	for (int y = 0; y < bp->height; y++) {
smatz@9597: 		/* next dst line begins here */
smatz@9597: 		uint32 *dst_ln = dst + bp->pitch;
truelight@7553: 
smatz@9597: 		/* next src line begins here */
smatz@9597: 		const Colour *src_px_ln = (const Colour *)((const byte *)src_px + *(const uint32 *)src_px);
smatz@9597: 		src_px++;
truelight@7553: 
smatz@9597: 		/* next src_n line begins here */
smatz@9597: 		const uint8 *src_n_ln = src_n + *(uint32 *)src_n;
smatz@9597: 		src_n += 4;
smatz@9597: 
smatz@9597: 		/* we will end this line when we reach this point */
smatz@9597: 		uint32 *dst_end = dst + bp->skip_left;
smatz@9597: 
smatz@9597: 		/* number of pixels with the same aplha channel class */
smatz@9597: 		uint n;
smatz@9597: 
smatz@9597: 		while (dst < dst_end) {
smatz@9597: 			n = *src_n++;
smatz@9597: 
smatz@9597: 			if (src_px->a == 0) {
smatz@9597: 				dst += n;
smatz@9597: 				src_px ++;
smatz@9597: 				src_n++;
smatz@9597: 			} else {
smatz@9597: 				if (dst + n > dst_end) {
smatz@9597: 					uint d = dst_end - dst;
smatz@9597: 					src_px += d;
smatz@9597: 					src_n += d;
smatz@9597: 
smatz@9597: 					dst = dst_end - bp->skip_left;
smatz@9597: 					dst_end = dst + bp->width;
smatz@9597: 
smatz@9597: 					n = min<uint>(n - d, (uint)bp->width);
smatz@9597: 					goto draw;
smatz@9597: 				}
smatz@9597: 				dst += n;
smatz@9597: 				src_px += n;
smatz@9597: 				src_n += n;
smatz@9597: 			}
smatz@9597: 		}
smatz@9597: 
smatz@9597: 		dst -= bp->skip_left;
smatz@9597: 		dst_end -= bp->skip_left;
smatz@9597: 
smatz@9597: 		dst_end += bp->width;
smatz@9597: 
smatz@9597: 		while (dst < dst_end) {
smatz@9597: 			n = min<uint>(*src_n++, (uint)(dst_end - dst));
smatz@9597: 
smatz@9597: 			if (src_px->a == 0) {
smatz@9597: 				dst += n;
smatz@9597: 				src_px++;
smatz@9597: 				src_n++;
truelight@7553: 				continue;
truelight@7553: 			}
truelight@7553: 
smatz@9597: 			draw:;
smatz@9597: 
truelight@7553: 			switch (mode) {
truelight@7553: 				case BM_COLOUR_REMAP:
smatz@9597: 					if (src_px->a == 255) {
smatz@9597: 						do {
smatz@9597: 							uint m = *src_n;
smatz@9597: 							/* In case the m-channel is zero, do not remap this pixel in any way */
smatz@9597: 							if (m == 0) {
smatz@9597: 								*dst = *src_px;
smatz@9597: 							} else {
smatz@9597: 								uint r = remap[m];
smatz@9597: 								if (r != 0) *dst = this->LookupColourInPalette(r);
smatz@9597: 							}
smatz@9597: 							dst++;
smatz@9597: 							src_px++;
smatz@9597: 							src_n++;
smatz@9597: 						} while (--n != 0);
truelight@7553: 					} else {
smatz@9597: 						do {
smatz@9597: 							uint m = *src_n;
smatz@9597: 							if (m == 0) {
smatz@9597: 								*dst = ComposeColourRGBANoCheck(src_px->r, src_px->g, src_px->b, src_px->a, *dst);
smatz@9597: 							} else {
smatz@9597: 								uint r = remap[m];
smatz@9597: 								if (r != 0) *dst = ComposeColourPANoCheck(this->LookupColourInPalette(r), src_px->a, *dst);
smatz@9597: 							}
smatz@9597: 							dst++;
smatz@9597: 							src_px++;
smatz@9597: 							src_n++;
smatz@9597: 						} while (--n != 0);
truelight@7553: 					}
truelight@7553: 					break;
truelight@7553: 
truelight@7553: 				case BM_TRANSPARENT:
truelight@7553: 					/* TODO -- We make an assumption here that the remap in fact is transparency, not some color.
truelight@7553: 					 *  This is never a problem with the code we produce, but newgrfs can make it fail... or at least:
truelight@7553: 					 *  we produce a result the newgrf maker didn't expect ;) */
truelight@7553: 
truelight@7553: 					/* Make the current color a bit more black, so it looks like this image is transparent */
smatz@9597: 					src_n += n;
smatz@9598: 					if (src_px->a == 255) {
smatz@9598: 						src_px += n;
smatz@9598: 						do {
smatz@9598: 							*dst = MakeTransparent(*dst, 3, 4);
smatz@9598: 							dst++;
smatz@9598: 						} while (--n != 0);
smatz@9598: 					} else {
smatz@9598: 						do {
smatz@9598: 							*dst = MakeTransparent(*dst, (256 * 4 - src_px->a), 256 * 4);
smatz@9598: 							dst++;
smatz@9598: 							src_px++;
smatz@9598: 						} while (--n != 0);
smatz@9598: 					}
truelight@7553: 					break;
truelight@7553: 
truelight@7553: 				default:
smatz@9597: 					if (src_px->a == 255) {
smatz@9597: 						/* faster than memcpy(), n is usually low */
smatz@9597: 						src_n += n;
smatz@9597: 						do {
smatz@9597: 							*dst++ = *src_px++;
smatz@9597: 						} while (--n != 0);
smatz@9597: 					} else {
smatz@9597: 						src_n += n;
smatz@9597: 						do {
smatz@9597: 							*dst = ComposeColourRGBANoCheck(src_px->r, src_px->g, src_px->b, src_px->a, *dst);
smatz@9597: 							dst++;
smatz@9597: 							src_px++;
smatz@9597: 						} while (--n != 0);
smatz@9597: 					}
truelight@7553: 					break;
truelight@7553: 			}
truelight@7553: 		}
smatz@9597: 
smatz@9597: 		dst = dst_ln;
smatz@9597: 		src_px = src_px_ln;
smatz@9597: 		src_n  = src_n_ln;
truelight@7553: 	}
truelight@7553: }
truelight@7553: 
smatz@9597: /**
smatz@9597:  * Draws a sprite to a (screen) buffer. Calls adequate templated function.
smatz@9597:  *
smatz@9597:  * @param bp further blitting parameters
smatz@9597:  * @param mode blitter mode
smatz@9597:  * @param zoom zoom level at which we are drawing
smatz@9597:  */
smatz@9486: void Blitter_32bppOptimized::Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomLevel zoom)
smatz@9486: {
smatz@9486: 	switch (mode) {
smatz@9486: 		default: NOT_REACHED();
smatz@9486: 		case BM_NORMAL:       Draw<BM_NORMAL>      (bp, zoom); return;
smatz@9486: 		case BM_COLOUR_REMAP: Draw<BM_COLOUR_REMAP>(bp, zoom); return;
smatz@9486: 		case BM_TRANSPARENT:  Draw<BM_TRANSPARENT> (bp, zoom); return;
smatz@9486: 	}
smatz@9486: }
smatz@9486: 
smatz@9597: /**
smatz@9597:  * Resizes the sprite in a very simple way, takes every n-th pixel and every n-th row
smatz@9597:  *
smatz@9597:  * @param sprite_src sprite to resize
smatz@9597:  * @param zoom resizing scale
smatz@9597:  * @return resized sprite
smatz@9597:  */
smatz@9597: static const SpriteLoader::Sprite *ResizeSprite(const SpriteLoader::Sprite *sprite_src, ZoomLevel zoom)
smatz@9597: {
smatz@9597: 	SpriteLoader::Sprite *sprite = MallocT<SpriteLoader::Sprite>(1);
smatz@9597: 
smatz@9597: 	if (zoom == ZOOM_LVL_NORMAL) {
smatz@9597: 		memcpy(sprite, sprite_src, sizeof(*sprite));
smatz@9597: 		uint size = sprite_src->height * sprite_src->width;
smatz@9597: 		sprite->data = MallocT<SpriteLoader::CommonPixel>(size);
smatz@9597: 		memcpy(sprite->data, sprite_src->data, size * sizeof(SpriteLoader::CommonPixel));
smatz@9597: 		return sprite;
smatz@9597: 	}
smatz@9597: 
smatz@9597: 	sprite->height = UnScaleByZoom(sprite_src->height, zoom);
smatz@9597: 	sprite->width  = UnScaleByZoom(sprite_src->width,  zoom);
smatz@9597: 	sprite->x_offs = UnScaleByZoom(sprite_src->x_offs, zoom);
smatz@9597: 	sprite->y_offs = UnScaleByZoom(sprite_src->y_offs, zoom);
smatz@9597: 
smatz@9597: 	uint size = sprite->height * sprite->width;
smatz@9597: 	SpriteLoader::CommonPixel *dst = sprite->data = CallocT<SpriteLoader::CommonPixel>(size);
smatz@9597: 
smatz@9597: 	const SpriteLoader::CommonPixel *src = (SpriteLoader::CommonPixel *)sprite_src->data;
smatz@9597: 	const SpriteLoader::CommonPixel *src_end = src + sprite_src->height * sprite_src->width;
smatz@9597: 
smatz@9597: 	uint scaled_1 = ScaleByZoom(1, zoom);
smatz@9597: 
smatz@9597: 	for (uint y = 0; y < sprite->height; y++) {
smatz@9597: 		if (src >= src_end) src = src_end - sprite_src->width;
smatz@9597: 
smatz@9597: 		const SpriteLoader::CommonPixel *src_ln = src + sprite_src->width * scaled_1;
smatz@9597: 		for (uint x = 0; x < sprite->width; x++) {
smatz@9597: 			if (src >= src_ln) src = src_ln - 1;
smatz@9597: 			*dst = *src;
smatz@9597: 			dst++;
smatz@9597: 			src += scaled_1;
smatz@9597: 		}
smatz@9597: 
smatz@9597: 		src = src_ln;
smatz@9597: 	}
smatz@9597: 
smatz@9597: 	return sprite;
smatz@9597: }
smatz@9597: 
truelight@7553: Sprite *Blitter_32bppOptimized::Encode(SpriteLoader::Sprite *sprite, Blitter::AllocatorProc *allocator)
truelight@7553: {
smatz@9597: 	/* streams of pixels (a, r, g, b channels)
smatz@9597: 	 *
smatz@9597: 	 * stored in separated stream so data are always aligned on 4B boundary */
smatz@9597: 	Colour *dst_px_orig[ZOOM_LVL_COUNT];
smatz@9597: 
smatz@9597: 	/* interleaved stream of 'm' channel and 'n' channel
smatz@9597: 	 * 'n' is number if following pixels with the same alpha channel class
smatz@9597: 	 * there are 3 classes: 0, 255, others
smatz@9597: 	 *
smatz@9597: 	 * it has to be stored in one stream so fewer registers are used -
smatz@9597: 	 * x86 has problems with register allocation even with this solution */
smatz@9597: 	uint8  *dst_n_orig[ZOOM_LVL_COUNT];
smatz@9597: 
smatz@9597: 	/* lengths of streams */
smatz@9597: 	uint32 lengths[ZOOM_LVL_COUNT][2];
smatz@9597: 
smatz@9597: 	for (ZoomLevel z = ZOOM_LVL_BEGIN; z < ZOOM_LVL_END; z++) {
smatz@9597: 		const SpriteLoader::Sprite *src_orig = ResizeSprite(sprite, z);
smatz@9597: 
smatz@9597: 		uint size = src_orig->height * src_orig->width;
smatz@9597: 
smatz@9597: 		dst_px_orig[z] = CallocT<Colour>(size + src_orig->height * 2);
smatz@9597: 		dst_n_orig[z]  = CallocT<uint8>(size * 2 + src_orig->height * 4 * 2);
smatz@9597: 
smatz@9597: 		uint32 *dst_px_ln = (uint32 *)dst_px_orig[z];
smatz@9597: 		uint32 *dst_n_ln  = (uint32 *)dst_n_orig[z];
smatz@9597: 
smatz@9597: 		const SpriteLoader::CommonPixel *src = (const SpriteLoader::CommonPixel *)src_orig->data;
smatz@9597: 
smatz@9597: 		for (uint y = src_orig->height; y > 0; y--) {
smatz@9597: 			Colour *dst_px = (Colour *)(dst_px_ln + 1);
smatz@9597: 			uint8 *dst_n = (uint8 *)(dst_n_ln + 1);
smatz@9597: 
smatz@9597: 			uint8 *dst_len = dst_n++;
smatz@9597: 
smatz@9597: 			uint last = 3;
smatz@9597: 			int len = 0;
smatz@9597: 
smatz@9597: 			for (uint x = src_orig->width; x > 0; x--) {
smatz@9597: 				uint8 a = src->a;
smatz@9597: 				uint t = a > 0 && a < 255 ? 1 : a;
smatz@9597: 
smatz@9597: 				if (last != t || len == 255) {
smatz@9597: 					if (last != 3) {
smatz@9597: 						*dst_len = len;
smatz@9597: 						dst_len = dst_n++;
smatz@9597: 					}
smatz@9597: 					len = 0;
smatz@9597: 				}
smatz@9597: 
smatz@9597: 				last = t;
smatz@9597: 				len++;
smatz@9597: 
smatz@9597: 				if (a != 0) {
smatz@9597: 					dst_px->a = a;
smatz@9597: 					*dst_n = src->m;
smatz@9597: 					if (src->m != 0) {
smatz@9597: 						/* Pre-convert the mapping channel to a RGB value */
smatz@9597: 						uint32 colour = this->LookupColourInPalette(src->m);
smatz@9597: 						dst_px->r = GB(colour, 16, 8);
smatz@9597: 						dst_px->g = GB(colour, 8,  8);
smatz@9597: 						dst_px->b = GB(colour, 0,  8);
smatz@9597: 					} else {
smatz@9597: 						dst_px->r = src->r;
smatz@9597: 						dst_px->g = src->g;
smatz@9597: 						dst_px->b = src->b;
smatz@9597: 					}
smatz@9597: 					dst_px++;
smatz@9597: 					dst_n++;
smatz@9597: 				} else if (len == 1) {
smatz@9597: 					dst_px++;
smatz@9597: 					*dst_n = src->m;
smatz@9597: 					dst_n++;
smatz@9597: 				}
smatz@9597: 
smatz@9597: 				src++;
smatz@9597: 			}
smatz@9597: 
smatz@9597: 			if (last != 3) {
smatz@9597: 				*dst_len = len;
smatz@9597: 			}
smatz@9597: 
smatz@9597: 			dst_px = (Colour *)AlignPtr(dst_px, 4);
smatz@9597: 			dst_n  = (uint8 *)AlignPtr(dst_n, 4);
smatz@9597: 
smatz@9597: 			*dst_px_ln = (uint8 *)dst_px - (uint8 *)dst_px_ln;
smatz@9597: 			*dst_n_ln  = (uint8 *)dst_n  - (uint8 *)dst_n_ln;
smatz@9597: 
smatz@9597: 			dst_px_ln = (uint32 *)dst_px;
smatz@9597: 			dst_n_ln =  (uint32 *)dst_n;
smatz@9597: 		}
smatz@9597: 
smatz@9597: 		lengths[z][0] = (byte *)dst_px_ln - (byte *)dst_px_orig[z]; // all are aligned to 4B boundary
smatz@9597: 		lengths[z][1] = (byte *)dst_n_ln  - (byte *)dst_n_orig[z];
smatz@9597: 
smatz@9597: 		free(src_orig->data);
smatz@9597: 		free((void *)src_orig);
smatz@9597: 	}
smatz@9597: 
smatz@9597: 	uint len = 0; // total length of data
smatz@9597: 	for (ZoomLevel z = ZOOM_LVL_BEGIN; z < ZOOM_LVL_END; z++) {
smatz@9597: 		len += lengths[z][0] + lengths[z][1];
smatz@9597: 	}
smatz@9597: 
smatz@9597: 	Sprite *dest_sprite = (Sprite *)allocator(sizeof(*dest_sprite) + sizeof(SpriteData) + len);
truelight@7553: 
truelight@7553: 	dest_sprite->height = sprite->height;
truelight@7553: 	dest_sprite->width  = sprite->width;
truelight@7553: 	dest_sprite->x_offs = sprite->x_offs;
truelight@7553: 	dest_sprite->y_offs = sprite->y_offs;
truelight@7553: 
smatz@9597: 	SpriteData *dst = (SpriteData *)dest_sprite->data;
truelight@7553: 
smatz@9597: 	for (ZoomLevel z = ZOOM_LVL_BEGIN; z < ZOOM_LVL_END; z++) {
smatz@9597: 		dst->offset[z][0] = z == ZOOM_LVL_BEGIN ? 0 : lengths[z - 1][1] + dst->offset[z - 1][1];
smatz@9597: 		dst->offset[z][1] = lengths[z][0] + dst->offset[z][0];
smatz@9597: 
smatz@9597: 		memcpy(dst->data + dst->offset[z][0], dst_px_orig[z], lengths[z][0]);
smatz@9597: 		memcpy(dst->data + dst->offset[z][1], dst_n_orig[z],  lengths[z][1]);
smatz@9597: 
smatz@9597: 		free(dst_px_orig[z]);
smatz@9597: 		free(dst_n_orig[z]);
truelight@7553: 	}
smatz@9597: 
truelight@7553: 	return dest_sprite;
truelight@7553: }