Eagle517/tge
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
tge/engine/terrain/blender.cc
Metario 9c6ff74f19 added everything
2017-04-17 06:17:10 -06:00

1334 lines
58 KiB
C++
Executable File
Raw Permalink Blame History

//-----------------------------------------------------------------------------
// Torque Game Engine
// Copyright (C) GarageGames.com, Inc.
//-----------------------------------------------------------------------------
#include "terrain/blender.h"
#include "platform/profiler.h"
/*************Configuration Switches*******************************/
// Haven't determined whether the alphaTable is a big win.
// Needs more measurement, must check tuning.
// alphaTable decreases integer math but increases RAM cache usage.
#define USE_ALPHA_TABLE_AT_4 1
#define USE_ALPHA_TABLE_AT_3 1
#define USE_ALPHA_TABLE_AT_2 1
#if defined(TORQUE_SUPPORTS_NASM)
# define BLENDER_USE_ASM
#endif
/*************Explanation*******************************/
// Manifest CONSTANTS mentioned within are defined right below.
// doc block moved to blender.h -- bjg
// Primary Manifest Values
#define TEXELS_PER_SOURCE_EDGE_LOG2 8
#define TEXELS_PER_TARGET_EDGE_LOG2 TerrainTextureMipLevel
#define LUMELS_PER_TILE_EDGE_LOG2 9
#define SQUARES_PER_TILE_EDGE_LOG2 8
#define MAX_MIPMAP_LEVELS 4
#define MAXIMUM_TEXTURES 4
#define SQUARES_PER_MIPMAP_EDGE_LOG2 (10 - TEXELS_PER_TARGET_EDGE_LOG2)
// Derived Manifest Values
#define MAX_TEXELS_PER_SQUARE_EDGE_LOG2 (TEXELS_PER_SOURCE_EDGE_LOG2 - SQUARES_PER_MIPMAP_EDGE_LOG2)
#define MIN_TEXELS_PER_SQUARE_EDGE_LOG2 (MAX_TEXELS_PER_SQUARE_EDGE_LOG2 - MAX_MIPMAP_LEVELS + 1)
#define LUMELS_PER_SQUARE_EDGE_LOG2 (LUMELS_PER_TILE_EDGE_LOG2 - SQUARES_PER_TILE_EDGE_LOG2)
#define MAX_TEXELS_PER_SQUARE_EDGE (1 << MAX_TEXELS_PER_SQUARE_EDGE_LOG2)
#define MAX_TEXELS_PER_SQUARE_LOG2 (MAX_TEXELS_PER_SQUARE_EDGE_LOG2 << 1)
#define MAX_TEXELS_PER_SQUARE (MAX_TEXELS_PER_SQUARE_EDGE*MAX_TEXELS_PER_SQUARE_EDGE)
#define LUMELS_PER_SQUARE_EDGE (1 << LUMELS_PER_SQUARE_EDGE_LOG2)
#define LUMELS_PER_SQUARE (LUMELS_PER_SQUARE_EDGE*LUMELS_PER_SQUARE_EDGE)
#define LUMELS_PER_TILE_EDGE (1 << LUMELS_PER_TILE_EDGE_LOG2)
#define LUMELS_PER_TILE_EDGE_MASK (LUMELS_PER_TILE_EDGE - 1)
#define SQUARES_PER_TILE_EDGE (1 << SQUARES_PER_TILE_EDGE_LOG2)
#define SQUARES_PER_TILE_EDGE_MASK (SQUARES_PER_TILE_EDGE - 1)
#define SQUARES_PER_MIPMAP_EDGE (1 << SQUARES_PER_MIPMAP_EDGE_LOG2)
#define SQUARES_PER_MIPMAP_EDGE_MASK (SQUARES_PER_MIPMAP_EDGE-1)
#define SQUARES_PER_MIPMAP (SQUARES_PER_MIPMAP_EDGE*SQUARES_PER_MIPMAP_EDGE)
#define TEXELS_PER_SOURCE_EDGE (1 << TEXELS_PER_SOURCE_EDGE_LOG2)
#define TEXELS_PER_TARGET_EDGE (1 << TEXELS_PER_TARGET_EDGE_LOG2)
#define TEXELS_PER_SOURCE_BMP (TEXELS_PER_SOURCE_EDGE * TEXELS_PER_SOURCE_EDGE)
#define TEXELS_PER_TARGET_BMP (TEXELS_PER_TARGET_EDGE * TEXELS_PER_TARGET_EDGE)
// Paul's leftovers
#define ALPHA_WID (1 << SQUARES_PER_TILE_EDGE_LOG2)
// -------------------------
// The following defines are for indexing into the source image maps.
// based on endianness, RGBA become ABGR, and we need to change indexing.
#if defined(TORQUE_BIG_ENDIAN)
#define SRC_IS_ABGR 1
#define SRC_OFF_RED 0
#define SRC_OFF_GREEN 1
#define SRC_OFF_BLUE 2
#else
#define SRC_IS_ABGR 0
#define SRC_OFF_RED 0
#define SRC_OFF_GREEN 1
#define SRC_OFF_BLUE 2
#endif
#ifdef __VEC__
vector unsigned int vlumels;
bool Blender::smUseVecBlender = false;
#endif
extern "C"
{
U32 lumels[4];
}
#if defined(BLENDER_USE_ASM)
extern "C"
{
// declarations for the NASM functions
void doSquare4( U32 *dst, int sq_shift, const int *aoff,
const U32 *const *bmp_ptrs, const U8 *const *alpha_ptrs );
void doSquare3( U32 *dst, int sq_shift, const int *aoff,
const U32 *const *bmp_ptrs, const U8 *const *alpha_ptrs );
void doSquare2( U32 *dst, int sq_shift, const int *aoff,
const U32 *const *bmp_ptrs, const U8 *const *alpha_ptrs );
void doLumelPlus1Mip( U16 *dstmip0, U16 *dstmip1, const U32 *srcptr );
void do1x1Lumel( U16 *dstptr, const U32 *srcptr );
void cheatmips( U16 *srcptr, U16 *dstmip0, U16 *dstmip1, int wid );
void cheatmips4x4( const U16 *srcptr, U16 *dstmip0, U16 *dstmip1 );
// global vars used for parameter passing
extern U32 sTargetTexelsPerLumel_log2;
extern U32 sTargetTexelsPerLumel;
extern U32 sTargetTexelsPerLumelDiv2;
extern U32 nextsrcrow;
extern U32 nextdstrow;
extern U32 mip0_dstrowadd;
extern U32 mip1_dstrowadd;
extern U32 minus1srcrowsPlus8;
extern U32 srcrows_x2_MinusTPL;
}
#else /* BLENDER_USE_ASM */
static U8 alphaTable[64 * 256];
class InitAlphaTable
{
public:
// build alpha blending table for C versions...
// This is just a cached pre-calc of some math...
InitAlphaTable()
{
for (U32 alpha = 0; alpha < 64; ++alpha)
for (U32 pix = 0; pix < 256; ++pix)
{
alphaTable[(alpha << 8) | pix] = (pix * alpha + 32) / 63;
}
}
};
static InitAlphaTable initAlphaTable; // Okay, cheesy
static void doSquare4( U32 *bmp_dest, int sq_shift, const int *alphaOffsets, const U32 *const *bmp_ptrs,
const U8 *const *alpha_ptrs )
{
int squareSize = 1 << sq_shift;
int squareSizeColors = squareSize;
int left_scan_edge_alpha_a;
int left_scan_edge_alpha_b;
int left_scan_edge_alpha_c;
int left_scan_edge_alpha_d;
int delta_left_alpha_a;
int delta_left_alpha_b;
int delta_left_alpha_c;
int delta_left_alpha_d;
int right_scan_edge_alpha_a;
int right_scan_edge_alpha_b;
int right_scan_edge_alpha_c;
int right_scan_edge_alpha_d;
int delta_right_alpha_a;
int delta_right_alpha_b;
int delta_right_alpha_c;
int delta_right_alpha_d;
{
int top_left_alpha = alpha_ptrs[0][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int top_right_alpha = alpha_ptrs[0][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int bot_left_alpha = alpha_ptrs[0][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int bot_right_alpha = alpha_ptrs[0][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_a = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_a = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_a = top_left_alpha;
right_scan_edge_alpha_a = top_right_alpha;
top_left_alpha = alpha_ptrs[1][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
top_right_alpha = alpha_ptrs[1][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_left_alpha = alpha_ptrs[1][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_right_alpha = alpha_ptrs[1][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_b = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_b = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_b = top_left_alpha;
right_scan_edge_alpha_b = top_right_alpha;
top_left_alpha = alpha_ptrs[2][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
top_right_alpha = alpha_ptrs[2][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_left_alpha = alpha_ptrs[2][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_right_alpha = alpha_ptrs[2][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_c = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_c = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_c = top_left_alpha;
right_scan_edge_alpha_c = top_right_alpha;
top_left_alpha = alpha_ptrs[3][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
top_right_alpha = alpha_ptrs[3][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_left_alpha = alpha_ptrs[3][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_right_alpha = alpha_ptrs[3][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_d = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_d = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_d = top_left_alpha;
right_scan_edge_alpha_d = top_right_alpha;
}
U8* sourcePtr_a = (U8*)bmp_ptrs[0];
U8* sourcePtr_b = (U8*)bmp_ptrs[1];
U8* sourcePtr_c = (U8*)bmp_ptrs[2];
U8* sourcePtr_d = (U8*)bmp_ptrs[3];
U8* destPtr = (U8*)bmp_dest;
for(S32 iy = squareSize; --iy >= 0;)
{
int scan_alpha_a = left_scan_edge_alpha_a;
left_scan_edge_alpha_a += delta_left_alpha_a;
int delta_scan_alpha_a = (right_scan_edge_alpha_a - scan_alpha_a) / squareSize;
right_scan_edge_alpha_a += delta_right_alpha_a;
int scan_alpha_b = left_scan_edge_alpha_b;
left_scan_edge_alpha_b += delta_left_alpha_b;
int delta_scan_alpha_b = (right_scan_edge_alpha_b - scan_alpha_b) / squareSize;
right_scan_edge_alpha_b += delta_right_alpha_b;
int scan_alpha_c = left_scan_edge_alpha_c;
left_scan_edge_alpha_c += delta_left_alpha_c;
int delta_scan_alpha_c = (right_scan_edge_alpha_c - scan_alpha_c) / squareSize;
right_scan_edge_alpha_c += delta_right_alpha_c;
int scan_alpha_d = left_scan_edge_alpha_d;
left_scan_edge_alpha_d += delta_left_alpha_d;
int delta_scan_alpha_d = (right_scan_edge_alpha_d - scan_alpha_d) / squareSize;
right_scan_edge_alpha_d += delta_right_alpha_d;
for(S32 ix = squareSizeColors; --ix >= 0;)
{
#if USE_ALPHA_TABLE_AT_4
int six_bit_alpha_a = (scan_alpha_a >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
int six_bit_alpha_b = (scan_alpha_b >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
int six_bit_alpha_c = (scan_alpha_c >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
int six_bit_alpha_d = (scan_alpha_d >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
destPtr[0] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_RED]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_RED]] +
alphaTable[six_bit_alpha_c | sourcePtr_c[SRC_OFF_RED]] +
alphaTable[six_bit_alpha_d | sourcePtr_d[SRC_OFF_RED]];
destPtr[1] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_GREEN]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_GREEN]] +
alphaTable[six_bit_alpha_c | sourcePtr_c[SRC_OFF_GREEN]] +
alphaTable[six_bit_alpha_d | sourcePtr_d[SRC_OFF_GREEN]];
destPtr[2] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_BLUE]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_BLUE]] +
alphaTable[six_bit_alpha_c | sourcePtr_c[SRC_OFF_BLUE]] +
alphaTable[six_bit_alpha_d | sourcePtr_d[SRC_OFF_BLUE]];
#else
#define EIGHT_BITS_SQ_SHIFT_2 (0xff * MAX_TEXELS_PER_SQUARE)
destPtr[0] = (scan_alpha_a * sourcePtr_a[SRC_OFF_RED] +
scan_alpha_b * sourcePtr_b[SRC_OFF_RED] +
scan_alpha_c * sourcePtr_c[SRC_OFF_RED] +
scan_alpha_d * sourcePtr_d[SRC_OFF_RED]) / EIGHT_BITS_SQ_SHIFT_2;
destPtr[1] = (scan_alpha_a * sourcePtr_a[SRC_OFF_GREEN] +
scan_alpha_b * sourcePtr_b[SRC_OFF_GREEN] +
scan_alpha_c * sourcePtr_c[SRC_OFF_GREEN] +
scan_alpha_d * sourcePtr_d[SRC_OFF_GREEN]) / EIGHT_BITS_SQ_SHIFT_2;
destPtr[2] = (scan_alpha_a * sourcePtr_a[SRC_OFF_BLUE] +
scan_alpha_b * sourcePtr_b[SRC_OFF_BLUE] +
scan_alpha_c * sourcePtr_c[SRC_OFF_BLUE] +
scan_alpha_d * sourcePtr_d[SRC_OFF_BLUE]) / EIGHT_BITS_SQ_SHIFT_2;
#endif
scan_alpha_a += delta_scan_alpha_a;
scan_alpha_b += delta_scan_alpha_b;
scan_alpha_c += delta_scan_alpha_c;
scan_alpha_d += delta_scan_alpha_d;
sourcePtr_a += 4;
sourcePtr_b += 4;
sourcePtr_c += 4;
sourcePtr_d += 4;
destPtr += 4;
}
}
}
static void doSquare3( U32 *bmp_dest, int sq_shift, const int *alphaOffsets, const U32 *const *bmp_ptrs,
const U8 *const *alpha_ptrs )
{
int squareSize = 1 << sq_shift;
int squareSizeColors = squareSize;
int left_scan_edge_alpha_a;
int left_scan_edge_alpha_b;
int left_scan_edge_alpha_c;
int delta_left_alpha_a;
int delta_left_alpha_b;
int delta_left_alpha_c;
int right_scan_edge_alpha_a;
int right_scan_edge_alpha_b;
int right_scan_edge_alpha_c;
int delta_right_alpha_a;
int delta_right_alpha_b;
int delta_right_alpha_c;
{
int top_left_alpha = alpha_ptrs[0][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int top_right_alpha = alpha_ptrs[0][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int bot_left_alpha = alpha_ptrs[0][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int bot_right_alpha = alpha_ptrs[0][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_a = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_a = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_a = top_left_alpha;
right_scan_edge_alpha_a = top_right_alpha;
top_left_alpha = alpha_ptrs[1][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
top_right_alpha = alpha_ptrs[1][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_left_alpha = alpha_ptrs[1][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_right_alpha = alpha_ptrs[1][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_b = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_b = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_b = top_left_alpha;
right_scan_edge_alpha_b = top_right_alpha;
top_left_alpha = alpha_ptrs[2][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
top_right_alpha = alpha_ptrs[2][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_left_alpha = alpha_ptrs[2][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_right_alpha = alpha_ptrs[2][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_c = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_c = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_c = top_left_alpha;
right_scan_edge_alpha_c = top_right_alpha;
}
U8* sourcePtr_a = (U8*)bmp_ptrs[0];
U8* sourcePtr_b = (U8*)bmp_ptrs[1];
U8* sourcePtr_c = (U8*)bmp_ptrs[2];
U8* destPtr = (U8*)bmp_dest;
for(S32 iy = squareSize; --iy >= 0;)
{
int scan_alpha_a = left_scan_edge_alpha_a;
left_scan_edge_alpha_a += delta_left_alpha_a;
int delta_scan_alpha_a = (right_scan_edge_alpha_a - scan_alpha_a) / squareSize;
right_scan_edge_alpha_a += delta_right_alpha_a;
int scan_alpha_b = left_scan_edge_alpha_b;
left_scan_edge_alpha_b += delta_left_alpha_b;
int delta_scan_alpha_b = (right_scan_edge_alpha_b - scan_alpha_b) / squareSize;
right_scan_edge_alpha_b += delta_right_alpha_b;
int scan_alpha_c = left_scan_edge_alpha_c;
left_scan_edge_alpha_c += delta_left_alpha_c;
int delta_scan_alpha_c = (right_scan_edge_alpha_c - scan_alpha_c) / squareSize;
right_scan_edge_alpha_c += delta_right_alpha_c;
for(S32 ix = squareSizeColors; --ix >= 0;)
{
#if USE_ALPHA_TABLE_AT_3
int six_bit_alpha_a = (scan_alpha_a >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
int six_bit_alpha_b = (scan_alpha_b >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
int six_bit_alpha_c = (scan_alpha_c >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
destPtr[0] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_RED]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_RED]] +
alphaTable[six_bit_alpha_c | sourcePtr_c[SRC_OFF_RED]];
destPtr[1] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_GREEN]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_GREEN]] +
alphaTable[six_bit_alpha_c | sourcePtr_c[SRC_OFF_GREEN]];
destPtr[2] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_BLUE]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_BLUE]] +
alphaTable[six_bit_alpha_c | sourcePtr_c[SRC_OFF_BLUE]];
#else
#define EIGHT_BITS_SQ_SHIFT_2 (0xff * MAX_TEXELS_PER_SQUARE)
destPtr[0] = (scan_alpha_a * sourcePtr_a[SRC_OFF_RED] +
scan_alpha_b * sourcePtr_b[SRC_OFF_RED] +
scan_alpha_c * sourcePtr_c[SRC_OFF_RED]) / EIGHT_BITS_SQ_SHIFT_2;
destPtr[1] = (scan_alpha_a * sourcePtr_a[SRC_OFF_GREEN] +
scan_alpha_b * sourcePtr_b[SRC_OFF_GREEN] +
scan_alpha_c * sourcePtr_c[SRC_OFF_GREEN]) / EIGHT_BITS_SQ_SHIFT_2;
destPtr[2] = (scan_alpha_a * sourcePtr_a[SRC_OFF_BLUE] +
scan_alpha_b * sourcePtr_b[SRC_OFF_BLUE] +
scan_alpha_c * sourcePtr_c[SRC_OFF_BLUE]) / EIGHT_BITS_SQ_SHIFT_2;
#endif
scan_alpha_a += delta_scan_alpha_a;
scan_alpha_b += delta_scan_alpha_b;
scan_alpha_c += delta_scan_alpha_c;
sourcePtr_a += 4;
sourcePtr_b += 4;
sourcePtr_c += 4;
destPtr += 4;
}
}
}
static void doSquare2( U32 *bmp_dest, int sq_shift, int *alphaOffsets, const U32 *const *bmp_ptrs,
const U8 *const *alpha_ptrs )
{
int squareSize = 1 << sq_shift;
int squareSizeColors = squareSize;
int left_scan_edge_alpha_a;
int left_scan_edge_alpha_b;
int delta_left_alpha_a;
int delta_left_alpha_b;
int right_scan_edge_alpha_a;
int right_scan_edge_alpha_b;
int delta_right_alpha_a;
int delta_right_alpha_b;
{
int top_left_alpha = alpha_ptrs[0][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int top_right_alpha = alpha_ptrs[0][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int bot_left_alpha = alpha_ptrs[0][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
int bot_right_alpha = alpha_ptrs[0][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_a = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_a = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_a = top_left_alpha;
right_scan_edge_alpha_a = top_right_alpha;
top_left_alpha = alpha_ptrs[1][ alphaOffsets[0] ] << MAX_TEXELS_PER_SQUARE_LOG2;
top_right_alpha = alpha_ptrs[1][ alphaOffsets[1] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_left_alpha = alpha_ptrs[1][ alphaOffsets[2] ] << MAX_TEXELS_PER_SQUARE_LOG2;
bot_right_alpha = alpha_ptrs[1][ alphaOffsets[3] ] << MAX_TEXELS_PER_SQUARE_LOG2;
delta_left_alpha_b = (bot_left_alpha - top_left_alpha) / squareSize;
delta_right_alpha_b = (bot_right_alpha - top_right_alpha) / squareSize;
left_scan_edge_alpha_b = top_left_alpha;
right_scan_edge_alpha_b = top_right_alpha;
}
U8* sourcePtr_a = (U8*)bmp_ptrs[0];
U8* sourcePtr_b = (U8*)bmp_ptrs[1];
U8* destPtr = (U8*)bmp_dest;
for(S32 iy = squareSize; --iy >= 0;)
{
int scan_alpha_a = left_scan_edge_alpha_a;
left_scan_edge_alpha_a += delta_left_alpha_a;
int delta_scan_alpha_a = (right_scan_edge_alpha_a - scan_alpha_a) / squareSize;
right_scan_edge_alpha_a += delta_right_alpha_a;
int scan_alpha_b = left_scan_edge_alpha_b;
left_scan_edge_alpha_b += delta_left_alpha_b;
int delta_scan_alpha_b = (right_scan_edge_alpha_b - scan_alpha_b) / squareSize;
right_scan_edge_alpha_b += delta_right_alpha_b;
for(S32 ix = squareSizeColors; --ix >= 0;)
{
#if USE_ALPHA_TABLE_AT_2
int six_bit_alpha_a = (scan_alpha_a >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
int six_bit_alpha_b = (scan_alpha_b >> (MAX_TEXELS_PER_SQUARE_LOG2 - 6)) & 0x3F00;
destPtr[0] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_RED]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_RED]];
destPtr[1] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_GREEN]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_GREEN]];
destPtr[2] = alphaTable[six_bit_alpha_a | sourcePtr_a[SRC_OFF_BLUE]] +
alphaTable[six_bit_alpha_b | sourcePtr_b[SRC_OFF_BLUE]];
#else
#define EIGHT_BITS_SQ_SHIFT_2 (0xff * MAX_TEXELS_PER_SQUARE)
destPtr[0] = (scan_alpha_a * sourcePtr_a[SRC_OFF_RED] +
scan_alpha_b * sourcePtr_b[SRC_OFF_RED] +
EIGHT_BITS_SQ_SHIFT_2 / 2) / EIGHT_BITS_SQ_SHIFT_2;
destPtr[1] = (scan_alpha_a * sourcePtr_a[SRC_OFF_GREEN] +
scan_alpha_b * sourcePtr_b[SRC_OFF_GREEN] +
EIGHT_BITS_SQ_SHIFT_2 / 2) / EIGHT_BITS_SQ_SHIFT_2;
destPtr[2] = (scan_alpha_a * sourcePtr_a[SRC_OFF_BLUE] +
scan_alpha_b * sourcePtr_b[SRC_OFF_BLUE] +
EIGHT_BITS_SQ_SHIFT_2 / 2) / EIGHT_BITS_SQ_SHIFT_2;
#endif
scan_alpha_a += delta_scan_alpha_a;
scan_alpha_b += delta_scan_alpha_b;
sourcePtr_a += 4;
sourcePtr_b += 4;
destPtr += 4;
}
}
}
#endif /* BLENDER_USE_ASM */
// old C extruder
static void extrude5551( const U16 *srcMip, U16 *mip, U32 height, U32 width )
{
const U16 *src = srcMip;
U16 *dst = mip;
U32 stride = width << 1;
for(U32 y = 0; y < height; y++)
{
for(U32 x = 0; x < width; x++)
{
U32 a = src[0];
U32 b = src[1];
U32 c = src[stride];
U32 d = src[stride+1];
#if SRC_IS_ABGR
dst[x] = ((( ((a >> 10) & 0x1f) + ((b >> 10) & 0x1f) + ((c >> 10) & 0x1f) + ((d >> 10) & 0x1F) ) >> 2) << 10) |
((( ((a >> 5) & 0x1f) + ((b >> 5) & 0x1f) + ((c >> 5) & 0x1f) + ((d >> 5) & 0x1F) ) >> 2) << 5) |
((( ((a >> 0) & 0x1F) + ((b >> 0) & 0x1F) + ((c >> 0) & 0x1f) + ((d >> 0) & 0x1f)) >> 2) << 0);
#else
dst[x] = ((((a >> 11) + (b >> 11) + (c >> 11) + (d >> 11)) >> 2) << 11) |
((( ((a >> 6) & 0x1f) + ((b >> 6) & 0x1f) + ((c >> 6) & 0x1f) + ((d >> 6) & 0x1F) ) >> 2) << 6) |
((( ((a >> 1) & 0x1F) + ((b >> 1) & 0x1F) + ((c >> 1) & 0x1f) + ((d >> 1) & 0x1f)) >> 2) << 1);
#endif
src += 2;
}
src += stride;
dst += width;
}
}
/// Take first mip in array, and extrude rest into other entries of array
/// i.e. power is 7 for 128x128, but there should be power+1 entries in the
/// array.
static void extrude( U16 **mips, U32 power )
{
U32 width = 1 << (power - 1);
for ( U32 i = 0; i < power; i++ )
{
extrude5551( mips[i], mips[i+1], width, width );
width >>= 1;
}
}
/// Blend textures!
///
/// - level is between 2 (high detail) and 5 (low detail) inclusive;
/// - x and y are in alpha sized squares (not tex squares)
/// - lmap is light map data, format is 5551 (1555 ABGR)
/// - destmips is a list of 16-bit 5551 RGBA (1555 ABGR) mip-map buffers for the result.
// Note: to use the altivec implementation you must use the -faltivec compiler option and have a G4 or G5 or better
#ifdef __VEC__
void transpose(long size, vector unsigned int *X)
{
long vWidth = size/4;
long i,vi,j,vj;
vector unsigned int R0,R1,R2,R3;
vector unsigned int T0,T1,T2,T3;
vector unsigned int C0,C1,C2,C3;
for (i=0, vi=0; i<vWidth;i++, vi+=size) {
R0 = X[ vi + 0*vWidth + i ];
R1 = X[ vi + 1*vWidth + i ];
R2 = X[ vi + 2*vWidth + i ];
R3 = X[ vi + 3*vWidth + i ];
T0 = vec_mergeh(R0,R2);
T1 = vec_mergeh(R1,R3);
T2 = vec_mergel(R0,R2);
T3 = vec_mergel(R1,R3);
X[ vi + 0*vWidth + i ] = vec_mergeh(T0,T1);
X[ vi + 1*vWidth + i ] = vec_mergel(T0,T1);
X[ vi + 2*vWidth + i ] = vec_mergeh(T2,T3);
X[ vi + 3*vWidth + i ] = vec_mergel(T2,T3);
for (j=i+1, vj=vi+size; j<(vWidth); j++, vj+=size)
{
R0 = X[ vi + 0*vWidth + j ];
R1 = X[ vi + 1*vWidth + j ];
R2 = X[ vi + 2*vWidth + j ];
R3 = X[ vi + 3*vWidth + j ];
C0 = X[ vj + 0*vWidth + i ];
C1 = X[ vj + 1*vWidth + i ];
C2 = X[ vj + 2*vWidth + i ];
C3 = X[ vj + 3*vWidth + i ];
T0 = vec_mergeh(R0,R2);
T1 = vec_mergeh(R1,R3);
T2 = vec_mergel(R0,R2);
T3 = vec_mergel(R1,R3);
X[ vj + 0*vWidth + i ] = vec_mergeh(T0,T1);
X[ vj + 1*vWidth + i ] = vec_mergel(T0,T1);
X[ vj + 2*vWidth + i ] = vec_mergeh(T2,T3);
X[ vj + 3*vWidth + i ] = vec_mergel(T2,T3);
T0 = vec_mergeh(C0,C2);
T1 = vec_mergeh(C1,C3);
T2 = vec_mergel(C0,C2);
T3 = vec_mergel(C1,C3);
X[ vi + 0*vWidth + j ] = vec_mergeh(T0,T1);
X[ vi + 1*vWidth + j ] = vec_mergel(T0,T1);
X[ vi + 2*vWidth + j ] = vec_mergeh(T2,T3);
X[ vi + 3*vWidth + j ] = vec_mergel(T2,T3);
}
}
}
inline vector unsigned int vec_loadAndSplatU32( unsigned int *scalarPtr )
{
register vector unsigned char splatMap = vec_lvsl( 0, scalarPtr );
const register vector unsigned int result = vec_lde( 0, scalarPtr );
splatMap = (vector unsigned char) vec_splat( (vector unsigned int) splatMap, 0 );
return( vec_perm( result, result, splatMap ) );
}
// Move the alpha calculation out of the inner loops of blender_vec() to allow the compiler to work its magic
class AlphaCalc
{
public:
AlphaCalc::AlphaCalc( const U8 * const alphaTable )
: vec_two( (vector unsigned int)( 2 ) ),
index_const( (vector unsigned int)( 0x3F00 ) ),
v255( (vector unsigned int)( 255 ) ),
dstcol_const( (vector unsigned int) (0xf8) ),
#if SRC_IS_ABGR
col_adjust1( (vector unsigned int) (7, 2, 0, 0) ),
col_adjust2( (vector unsigned int) (0, 0, 3, 0) ),
#else
// for some strange and nonexistent altivec processor which uses RGBA5551
col_adjust1( (vector unsigned int) (8, 3, 0, 0) ),
col_adjust2( (vector unsigned int) (0, 0, 2, 0) ),
#endif
globalAlphaTable( alphaTable )
{
}
inline U16 Calc( vector unsigned int hscan_component, vector unsigned int src ) const
{
u_tmp.v = vec_or(vec_and(vec_sr(hscan_component, vec_two), index_const), src);
vector unsigned int dstcol;
unsigned int *sloader = (unsigned int *) &dstcol;
sloader[0] = globalAlphaTable[u_tmp.s[0]];
sloader[1] = globalAlphaTable[u_tmp.s[1]];
sloader[2] = globalAlphaTable[u_tmp.s[2]];
dstcol = vec_add( dstcol, dstcol );
dstcol = vec_min( dstcol, v255 );
// NOTE that on Mac, color order is flipped (ABGR1555 instead of RGBA5551), so:
// 1. we already reversed color order via BIG_ENDIAN indexing above, but
// 2. we need to change the shifts for alpha being the high bit instead of the low.
u_tmp.v = vec_sr(vec_sl(vec_and(dstcol, dstcol_const), col_adjust1), col_adjust2);
return( (unsigned short) (u_tmp.s[0] | u_tmp.s[1] | u_tmp.s[2]) );
}
private:
const vector unsigned int vec_two;
const vector unsigned int index_const;
const vector unsigned int v255;
const vector unsigned int dstcol_const;
const vector unsigned int col_adjust1;
const vector unsigned int col_adjust2;
mutable union {
vector unsigned int v;
U32 s[4];
} u_tmp;
const U8 * const globalAlphaTable;
};
/* Kyle Goodwin's altivec texture blender implementation. There's still a few areas this could be improved,
but given the enormous increase this already grants to performance the minor areas aren't significant.
Unfortunately the code below is highly unreadable for two main reasons, the original blender I was rewriting
was unreadable, and for speed I chose to use exclusively atomic 1-to-1 assembly equivilent vector operations.
The onyl exception to this is the one matrix transpose operation, which could possibly be optimized more.
I often must reference vector values through normal pointers or as part of a union, but since this impedes
parallelization I did it as little as possible. All in all this works out quite well and beings mac
performance in line with pc performance.
*/
inline void Blender::blend_vec( int x, int y, int squaresPerTargetEdge_log2, const U16 *lmap, U16 **destmips )
{
PROFILE_START(ALTIVEC_BLEND);
const U32 squaresPerTargetEdge(1 << squaresPerTargetEdge_log2); // 32 (low detail) to 4 (high detail).
const U32 texelsPerSquareEdge_log2(TEXELS_PER_TARGET_EDGE_LOG2 - squaresPerTargetEdge_log2); // 5 (high detail) to 2 (low detail)
const U32 texelsPerSquareEdge(1 << texelsPerSquareEdge_log2); // == TEXELS_PER_TARGET_EDGE / squaresPerTargetEdge); 4 (low) to 32 (high) detail.
const U32 texelsPerSquare_log2(texelsPerSquareEdge_log2 << 1); // 10 (high detail) to 4 (low detail)
const U32 sourceMipMapIndex(MAX_TEXELS_PER_SQUARE_EDGE_LOG2 - texelsPerSquareEdge_log2);
const U32 targetTexelsPerLumel_log2(texelsPerSquareEdge_log2 - LUMELS_PER_SQUARE_EDGE_LOG2);
const U32 targetTexelsPerLumel(1 << targetTexelsPerLumel_log2);
const U32 yStrideThroughTarget(TEXELS_PER_TARGET_EDGE);
const U32 yStrideThroughSquare(texelsPerSquareEdge);
const U32 xStrideAcrossLumels(targetTexelsPerLumel);
const U32 yStrideThroughTargetAcrossLumels(yStrideThroughTarget << targetTexelsPerLumel_log2);
const U32 yStrideThroughSquareAcrossLumels(yStrideThroughSquare << targetTexelsPerLumel_log2);
const U32 *const*const allSourceBitMaps = &bmpdata[sourceMipMapIndex * num_src_bmps];
const AlphaCalc alphaCalc( alphaTable );
// sy & sx index through the SQUAREs of the DESTINATION MIP-MAP
// All Source MIP-MAPs are 2D arrays of squares:
// SQUARE source_mip_map_2D[SQUARES_PER_MIPMAP_EDGE][SQUARES_PER_MIPMAP_EDGE];
// But they are stored as 1D arrays:
// SQUARE source_mip_map_1D[SQUARES_PER_MIPMAP_EDGE*SQUARES_PER_MIPMAP_EDGE];
// therefore the following are equivalent:
// source_mip_map_2D[Y][X]
// source_mip_map_1D[(Y * SQUARES_PER_MIPMAP_EDGE) + X]
// source_mip_map_1D[(Y << SQUARES_PER_MIPMAP_EDGE_LOG2) + X]
// source_mip_map_1D[(Y << SQUARES_PER_MIPMAP_EDGE_LOG2) | X]
// This loop is from [0] through [squaresPerTargetEdge - 1] of the destination
// and from [y] through [y + squaresPerTargetEdge - 1] of the source.
// A single terrain TILE is equivalently:
// SQUARE terrain_tile_2D[SQUARES_PER_TILE_EDGE][SQUARES_PER_TILE_EDGE];
// or
// SQUARE terrain_tile_1D[SQUARES_PER_TILE_EDGE*SQUARES_PER_TILE_EDGE];
// therefore the following are equivalent:
// terrain_tile_2D[Y][X]
// terrain_tile_1D[(Y << SQUARES_PER_MIPMAP_EDGE_LOG2) | X]
// Neither source_mip_map_1D nor terrain_tile_1D appear explicitly.
for ( int yInTarget = 0; yInTarget < squaresPerTargetEdge; ++yInTarget )
{
// This whole section is called "doing 2-dimensional array indexing the hard way"
// yInTile & after_yInTile are the bottom and top of the source square we are actually processing,
// masked to tile size which is what causes the "repeating" effect
const int yInTile((y + yInTarget) & SQUARES_PER_TILE_EDGE_MASK);
const int after_yInTile((yInTile + 1) & SQUARES_PER_TILE_EDGE_MASK);
// yInTile_offset and after_yInTile_offset are the offsets into the terrain_tile_1D format arrays
const int yInTile_offset(yInTile << SQUARES_PER_TILE_EDGE_LOG2);
const int after_yInTile_offset(after_yInTile << SQUARES_PER_TILE_EDGE_LOG2);
// py is the row index in squares into the source_mip_map_2D
const int yInSource(yInTile & SQUARES_PER_MIPMAP_EDGE_MASK);
// yInSource_offset is the offset in squares into the source_mip_map_1D, times the size
// of the squares.
const int yInSource_offset(yInSource << (texelsPerSquare_log2 + SQUARES_PER_MIPMAP_EDGE_LOG2));
// This loop is from [yInTarget][0] through [yInTarget][squaresPerTargetEdge - 1] of the destination
// and from [yInTile][x] through [yInTile][x + squaresPerTargetEdge - 1] of the source.
for ( int xInTarget = 0; xInTarget < squaresPerTargetEdge; xInTarget++ )
{
// xInTile & after_xInTile are the left and right side of the source square we are actually processing,
// masked to tile size which is what causes the "repeating" effect
const int xInTile((x + xInTarget) & SQUARES_PER_TILE_EDGE_MASK);
const int after_xInTile((xInTile + 1) & SQUARES_PER_TILE_EDGE_MASK);
// xInSource is the column index in squares into the source_mip_map_2D
const int xInSource(xInTile & SQUARES_PER_MIPMAP_EDGE_MASK);
// As you can see the GRID is accessed in TILE co-ordinates
const U32 gridflags(GRIDFLAGS( xInTile, yInTile ));
// Cache the source textures at our mip-map level as specified by the GRID-FLAGS
const U32 *sourceSquareBitMaps[MAXIMUM_TEXTURES];
// Cache the Alpha-Maps as specified by the GRID-FLAGS
const U8 *alphaMaps[MAXIMUM_TEXTURES];
// Pre-calculate (U8*) &source_mip_map_2D[yInSource][xInSource] --
// ( (yInSource * SQUARES_PER_MIPMAP_EDGE) + xInSource ) * sizeof(SQUARE)
const int bitmapOffset(yInSource_offset | (xInSource << texelsPerSquare_log2));
int numTexturesToBlend = 0;
for ( int i = 0; i < num_src_bmps; ++i )
{
if ( gridflags & (MATERIALSTART << i) ) // Gridflags tell us which materials are used for this square
{// Cache Alpha maps and bitmaps.
alphaMaps[ numTexturesToBlend ] = alpha_data[ i ];
sourceSquareBitMaps[ numTexturesToBlend++ ] = &allSourceBitMaps[ i ][bitmapOffset];
if ( numTexturesToBlend == MAXIMUM_TEXTURES )
break; // Why? What happens if more than (4) textures should be blended?
}
}
const U32 *bufferToLightFrom = blendbuffer;
if ( numTexturesToBlend < 2 )
{
// don't copy the square over...just leave it and tell
// lighting code to use src bmp as the source instead of
// the blend_buffer;
bufferToLightFrom = sourceSquareBitMaps[ 0 ];
}
else
{
int alphaOffsets[4];
alphaOffsets[0] = yInTile_offset | xInTile; // precalculate offsets for tile-coords[yInTile][xInTile]
alphaOffsets[1] = yInTile_offset | after_xInTile; // and so on for the square bounded by
alphaOffsets[2] = after_yInTile_offset | xInTile; // [yInTile][xInTile] [yInTile][after_xInTile]
alphaOffsets[3] = after_yInTile_offset | after_xInTile;// [after_yInTile][xInTile] [after_yInTile][after_xInTile]
switch( numTexturesToBlend ) // Blend 1 square of the numTexturesToBlend bit-maps into the blend buffer
{
case 2:
doSquare2( blendbuffer, texelsPerSquareEdge_log2, alphaOffsets, sourceSquareBitMaps, alphaMaps );
break;
case 3:
doSquare3( blendbuffer, texelsPerSquareEdge_log2, alphaOffsets, sourceSquareBitMaps, alphaMaps );
break;
default: // more subtle paranoia
doSquare4( blendbuffer, texelsPerSquareEdge_log2, alphaOffsets, sourceSquareBitMaps, alphaMaps );
break;
}
}
// [these comments are making me paranoid -- Ed.]
// copy in the lighting info
// Once again we make with the linear 2D array
const U32 xTexelInTarget(xInTarget << texelsPerSquareEdge_log2);
const U32 yTexelInTarget(yInTarget << texelsPerSquareEdge_log2);
const U32 yTexelInTarget_offset((yTexelInTarget << TEXELS_PER_TARGET_EDGE_LOG2));
U16 *const bits0 = &destmips[0][ yTexelInTarget_offset + xTexelInTarget ];
U16 *const bits1 = &destmips[1][ (yTexelInTarget_offset >> 2) + (xTexelInTarget >> 1) ];
U16 *const bits2 = &destmips[2][ (yTexelInTarget_offset >> 4) + (xTexelInTarget >> 2) ];
const U32 base_xInLightmap(xInTile << LUMELS_PER_SQUARE_EDGE_LOG2);
const U32 base_yInLightmap(yInTile << LUMELS_PER_SQUARE_EDGE_LOG2);
U32 yInLightmap_offset(base_yInLightmap << LUMELS_PER_TILE_EDGE_LOG2);
U32 next_yInLightmap(base_yInLightmap);
U32 yTexelInTargetSquare_offset(0);
U32 yTexelInSquare_offset(0);
for(U32 yLumelInSquare(0); yLumelInSquare < LUMELS_PER_SQUARE_EDGE; ++yLumelInSquare)
{
next_yInLightmap = (next_yInLightmap + 1) & LUMELS_PER_TILE_EDGE_MASK;
const U32 next_yInLightmap_offset(next_yInLightmap << LUMELS_PER_TILE_EDGE_LOG2);
U32 xInLightmap(base_xInLightmap);
U32 xTexelInSquare_offset = 0;
for(U32 xLumelInSquare(0); xLumelInSquare < LUMELS_PER_SQUARE_EDGE; ++xLumelInSquare)
{
const U32 next_xInLightmap((xInLightmap + 1) & LUMELS_PER_TILE_EDGE_MASK);
U32 texelInTargetSquare_offset = yTexelInTargetSquare_offset + xTexelInSquare_offset;
U32 texelInSquare_offset = yTexelInSquare_offset + xTexelInSquare_offset;
unsigned int *loader = (unsigned int *) &vlumels;
loader[0] = lmap[xInLightmap | yInLightmap_offset];
loader[1] = lmap[next_xInLightmap | yInLightmap_offset];
loader[2] = lmap[xInLightmap | next_yInLightmap_offset];
loader[3] = lmap[next_xInLightmap | next_yInLightmap_offset];
// Split the LUMELs into colors
vector unsigned int col[4];
vector unsigned int col_const = (vector unsigned int) (0x1f << 11);
col[2] = vec_and(vlumels, col_const);
col[1] = vec_and(vec_sl(vlumels, (vector unsigned int) (5)), col_const);
col[0] = vec_and(vec_sl(vlumels, (vector unsigned int) (10)), col_const);
vector unsigned int vec_targetTexelsPerLumel_log2 = vec_loadAndSplatU32( &targetTexelsPerLumel_log2 );
transpose(4, col); // transpose the matrix since we were using rows and now we need columns
// One for each color component
vector unsigned int left_component_delta = vec_sr(vec_sub(col[2], col[0]), vec_targetTexelsPerLumel_log2);
vector unsigned int right_component_delta = vec_sr(vec_sub(col[3], col[1]), vec_targetTexelsPerLumel_log2);
vector unsigned int vscan_left_component = col[0];
vector unsigned int vscan_right_component = col[1];
// Now we interpolate the color shifts across the square
for(U32 yTexelInLumel = 0; yTexelInLumel < targetTexelsPerLumel; ++yTexelInLumel)
{
vector unsigned int across_component_delta = vec_sr(vec_sub(vscan_right_component, vscan_left_component), vec_targetTexelsPerLumel_log2);
vector unsigned int hscan_component = vscan_left_component;
vscan_left_component = vec_add(vscan_left_component, left_component_delta);
vscan_right_component = vec_add(vscan_right_component, right_component_delta);
U16 *dstbits = &bits0[ texelInTargetSquare_offset ];
const U8 *srcbits = (U8 *)&bufferToLightFrom[ texelInSquare_offset ];
for(U32 xTexelInLumel = 0; xTexelInLumel < targetTexelsPerLumel; ++xTexelInLumel)
{
vector unsigned int cur_srcbits;
loader = (unsigned int *) &cur_srcbits;
loader[0] = srcbits[0];
loader[1] = srcbits[1];
loader[2] = srcbits[2];
*dstbits++ = alphaCalc.Calc( hscan_component, cur_srcbits );
srcbits += 4;
}
texelInTargetSquare_offset += yStrideThroughTarget;
texelInSquare_offset += yStrideThroughSquare;
}
xTexelInSquare_offset += xStrideAcrossLumels;
xInLightmap = next_xInLightmap;
}
yInLightmap_offset = next_yInLightmap_offset;
yTexelInTargetSquare_offset += yStrideThroughTargetAcrossLumels;
yTexelInSquare_offset += yStrideThroughSquareAcrossLumels;
}
// end of lighting.
}
}
extrude( destmips, TEXELS_PER_TARGET_EDGE_LOG2 );
PROFILE_END();
}
void Blender::blend( int x, int y, int squaresPerTargetEdge_log2, const U16 *lmap, U16 **destmips )
{
if(smUseVecBlender)
blend_vec(x, y, squaresPerTargetEdge_log2, lmap, destmips);
else
blend_c(x, y, squaresPerTargetEdge_log2, lmap, destmips);
}
#else
void Blender::blend( int x, int y, int squaresPerTargetEdge_log2, const U16 *lmap, U16 **destmips )
{
blend_c(x, y, squaresPerTargetEdge_log2, lmap, destmips);
}
#endif
inline void Blender::blend_c( int x, int y, int squaresPerTargetEdge_log2, const U16 *lmap, U16 **destmips )
{
PROFILE_START(Blender);
const int squaresPerTargetEdge(1 << squaresPerTargetEdge_log2); // 32 (low detail) to 4 (high detail).
const int texelsPerSquareEdge_log2(TEXELS_PER_TARGET_EDGE_LOG2 - squaresPerTargetEdge_log2); // 5 (high detail) to 2 (low detail)
const int texelsPerSquareEdge(1 << texelsPerSquareEdge_log2); // == TEXELS_PER_TARGET_EDGE / squaresPerTargetEdge); 4 (low) to 32 (high) detail.
const int texelsPerSquare_log2(texelsPerSquareEdge_log2 << 1); // 10 (high detail) to 4 (low detail)
const int sourceMipMapIndex(MAX_TEXELS_PER_SQUARE_EDGE_LOG2 - texelsPerSquareEdge_log2);
const int targetTexelsPerLumel_log2(texelsPerSquareEdge_log2 - LUMELS_PER_SQUARE_EDGE_LOG2);
const int targetTexelsPerLumel(1 << targetTexelsPerLumel_log2);
const U32 yStrideThroughTarget(TEXELS_PER_TARGET_EDGE);
const U32 yStrideThroughSquare(texelsPerSquareEdge);
const U32 xStrideAcrossLumels(targetTexelsPerLumel);
const U32 yStrideThroughTargetAcrossLumels(yStrideThroughTarget << targetTexelsPerLumel_log2);
const U32 yStrideThroughSquareAcrossLumels(yStrideThroughSquare << targetTexelsPerLumel_log2);
#if defined(BLENDER_USE_ASM)
// These are all secret parameters passed to the assembly language code through statics.
sTargetTexelsPerLumel_log2 = targetTexelsPerLumel_log2;
sTargetTexelsPerLumel = targetTexelsPerLumel;
sTargetTexelsPerLumelDiv2 = targetTexelsPerLumel >> 1;
nextsrcrow = ((yStrideThroughSquare) << 2);
nextdstrow = ((yStrideThroughTarget) << 1);
mip0_dstrowadd = (nextdstrow << 1) - (targetTexelsPerLumel << 1);
mip1_dstrowadd = (nextdstrow >> 1) - (targetTexelsPerLumel);
minus1srcrowsPlus8 = 8 - nextsrcrow;
srcrows_x2_MinusTPL = (nextsrcrow << 1) - (targetTexelsPerLumel << 2);
#endif
const U32 *const*const allSourceBitMaps = &bmpdata[sourceMipMapIndex * num_src_bmps];
// sy & sx index through the SQUAREs of the DESTINATION MIP-MAP
// All Source MIP-MAPs are 2D arrays of squares:
// SQUARE source_mip_map_2D[SQUARES_PER_MIPMAP_EDGE][SQUARES_PER_MIPMAP_EDGE];
// But they are stored as 1D arrays:
// SQUARE source_mip_map_1D[SQUARES_PER_MIPMAP_EDGE*SQUARES_PER_MIPMAP_EDGE];
// therefore the following are equivalent:
// source_mip_map_2D[Y][X]
// source_mip_map_1D[(Y * SQUARES_PER_MIPMAP_EDGE) + X]
// source_mip_map_1D[(Y << SQUARES_PER_MIPMAP_EDGE_LOG2) + X]
// source_mip_map_1D[(Y << SQUARES_PER_MIPMAP_EDGE_LOG2) | X]
// This loop is from [0] through [squaresPerTargetEdge - 1] of the destination
// and from [y] through [y + squaresPerTargetEdge - 1] of the source.
// A single terrain TILE is equivalently:
// SQUARE terrain_tile_2D[SQUARES_PER_TILE_EDGE][SQUARES_PER_TILE_EDGE];
// or
// SQUARE terrain_tile_1D[SQUARES_PER_TILE_EDGE*SQUARES_PER_TILE_EDGE];
// therefore the following are equivalent:
// terrain_tile_2D[Y][X]
// terrain_tile_1D[(Y << SQUARES_PER_MIPMAP_EDGE_LOG2) | X]
// Neither source_mip_map_1D nor terrain_tile_1D appear explicitly.
for ( int yInTarget = 0; yInTarget < squaresPerTargetEdge; yInTarget++ )
{
// This whole section is called "doing 2-dimensional array indexing the hard way"
// yInTile & after_yInTile are the bottom and top of the source square we are actually processing,
// masked to tile size which is what causes the "repeating" effect
const int yInTile((y + yInTarget) & SQUARES_PER_TILE_EDGE_MASK);
const int after_yInTile((yInTile + 1) & SQUARES_PER_TILE_EDGE_MASK);
// yInTile_offset and after_yInTile_offset are the offsets into the terrain_tile_1D format arrays
const int yInTile_offset(yInTile << SQUARES_PER_TILE_EDGE_LOG2);
const int after_yInTile_offset(after_yInTile << SQUARES_PER_TILE_EDGE_LOG2);
// py is the row index in squares into the source_mip_map_2D
const int yInSource(yInTile & SQUARES_PER_MIPMAP_EDGE_MASK);
// yInSource_offset is the offset in squares into the source_mip_map_1D, times the size
// of the squares.
const int yInSource_offset(yInSource << (texelsPerSquare_log2 + SQUARES_PER_MIPMAP_EDGE_LOG2));
// This loop is from [yInTarget][0] through [yInTarget][squaresPerTargetEdge - 1] of the destination
// and from [yInTile][x] through [yInTile][x + squaresPerTargetEdge - 1] of the source.
for ( int xInTarget = 0; xInTarget < squaresPerTargetEdge; xInTarget++ )
{
// xInTile & after_xInTile are the left and right side of the source square we are actually processing,
// masked to tile size which is what causes the "repeating" effect
const int xInTile((x + xInTarget) & SQUARES_PER_TILE_EDGE_MASK);
const int after_xInTile((xInTile + 1) & SQUARES_PER_TILE_EDGE_MASK);
// xInSource is the column index in squares into the source_mip_map_2D
const int xInSource(xInTile & SQUARES_PER_MIPMAP_EDGE_MASK);
// As you can see the GRID is accessed in TILE co-ordinates
const U32 gridflags(GRIDFLAGS( xInTile, yInTile ));
int numTexturesToBlend = 0;
int alphaOffsets[4];
alphaOffsets[0] = yInTile_offset | xInTile; // precalculate offsets for tile-coords[yInTile][xInTile]
alphaOffsets[1] = yInTile_offset | after_xInTile; // and so on for the square bounded by
alphaOffsets[2] = after_yInTile_offset | xInTile; // [yInTile][xInTile] [yInTile][after_xInTile]
alphaOffsets[3] = after_yInTile_offset | after_xInTile;// [after_yInTile][xInTile] [after_yInTile][after_xInTile]
// Cache the source textures at our mip-map level as specified by the GRID-FLAGS
const U32 *sourceSquareBitMaps[MAXIMUM_TEXTURES];
// Cache the Alpha-Maps as specified by the GRID-FLAGS
const U8 *alphaMaps[MAXIMUM_TEXTURES];
// Pre-calculate (U8*) &source_mip_map_2D[yInSource][xInSource] --
// ( (yInSource * SQUARES_PER_MIPMAP_EDGE) + xInSource ) * sizeof(SQUARE)
const int bitmapOffset(yInSource_offset | (xInSource << texelsPerSquare_log2));
for ( int i = 0; i < num_src_bmps; i++ )
if ( gridflags & (MATERIALSTART << i) ) // Gridflags tell us which materials are used for this square
{// Cache Alpha maps and bitmaps.
alphaMaps[ numTexturesToBlend ] = alpha_data[ i ];
sourceSquareBitMaps[ numTexturesToBlend++ ] = &allSourceBitMaps[ i ][bitmapOffset];
if ( numTexturesToBlend == MAXIMUM_TEXTURES )
break; // Why? What happens if more than (4) textures should be blended?
}
const U32 *bufferToLightFrom = blendbuffer;
switch( numTexturesToBlend ) // Blend 1 square of the numTexturesToBlend bit-maps into the blend buffer
{
case 0:// paranoia
case 1:
// don't copy the square over...just leave it and tell
// lighting code to use src bmp as the source instead of
// the blend_buffer;
bufferToLightFrom = sourceSquareBitMaps[ 0 ];
break;
case 2:
doSquare2( blendbuffer, texelsPerSquareEdge_log2, alphaOffsets, sourceSquareBitMaps, alphaMaps );
break;
case 3:
doSquare3( blendbuffer, texelsPerSquareEdge_log2, alphaOffsets, sourceSquareBitMaps, alphaMaps );
break;
default: // more subtle paranoia
doSquare4( blendbuffer, texelsPerSquareEdge_log2, alphaOffsets, sourceSquareBitMaps, alphaMaps );
break;
}
// [these comments are making me paranoid -- Ed.]
// copy in the lighting info
// Once again we make with the linear 2D array
const U32 xTexelInTarget(xInTarget << texelsPerSquareEdge_log2);
const U32 yTexelInTarget(yInTarget << texelsPerSquareEdge_log2);
const U32 yTexelInTarget_offset((yTexelInTarget << TEXELS_PER_TARGET_EDGE_LOG2));
U16 *const bits0 = &destmips[0][ yTexelInTarget_offset + xTexelInTarget ];
U16 *const bits1 = &destmips[1][ (yTexelInTarget_offset >> 2) + (xTexelInTarget >> 1) ];
U16 *const bits2 = &destmips[2][ (yTexelInTarget_offset >> 4) + (xTexelInTarget >> 2) ];
const U32 base_xInLightmap(xInTile << LUMELS_PER_SQUARE_EDGE_LOG2);
const U32 base_yInLightmap(yInTile << LUMELS_PER_SQUARE_EDGE_LOG2);
U32 yInLightmap_offset(base_yInLightmap << LUMELS_PER_TILE_EDGE_LOG2);
U32 next_yInLightmap(base_yInLightmap);
U32 yTexelInTargetSquare_offset(0);
U32 yTexelInSquare_offset(0);
for(U32 yLumelInSquare(0); yLumelInSquare < LUMELS_PER_SQUARE_EDGE; ++yLumelInSquare)
{
next_yInLightmap = (next_yInLightmap + 1) & LUMELS_PER_TILE_EDGE_MASK;
U32 next_yInLightmap_offset(next_yInLightmap << LUMELS_PER_TILE_EDGE_LOG2);
U32 xInLightmap(base_xInLightmap);
U32 xTexelInSquare_offset = 0;
for(U32 xLumelInSquare(0); xLumelInSquare < LUMELS_PER_SQUARE_EDGE; ++xLumelInSquare)
{
U32 next_xInLightmap((xInLightmap + 1) & LUMELS_PER_TILE_EDGE_MASK);
U32 texelInTargetSquare_offset = yTexelInTargetSquare_offset + xTexelInSquare_offset;
U32 texelInSquare_offset = yTexelInSquare_offset + xTexelInSquare_offset;
// lumels are secret parameters to subroutines
lumels[0] = U32(lmap[xInLightmap | yInLightmap_offset]);
lumels[1] = U32(lmap[next_xInLightmap | yInLightmap_offset]);
lumels[2] = U32(lmap[xInLightmap | next_yInLightmap_offset]);
lumels[3] = U32(lmap[next_xInLightmap | next_yInLightmap_offset]);
PROFILE_START(BlenderInASM);
#if defined(BLENDER_USE_ASM)
if ( targetTexelsPerLumel > 1 )
{
doLumelPlus1Mip( &bits0[ texelInTargetSquare_offset ],
&bits1[ (yTexelInTargetSquare_offset >> 2) + (xTexelInSquare_offset >> 1) ],
&bufferToLightFrom[ texelInSquare_offset ] );
}
else
do1x1Lumel( &bits0[ texelInTargetSquare_offset ], &bufferToLightFrom[ texelInSquare_offset ] );
#else
// Split the LUMELs into colors
U32 col[3][4];
U32 i;
for(i = 0; i < 4; i++)
{
col[2][i] = (lumels[i]) & (0x1f << 11);
col[1][i] = (lumels[i] << 5) & (0x1f << 11);
col[0][i] = (lumels[i] << 10) & (0x1f << 11);
}
// One for each color component
U32 left_component_delta[3];
U32 right_component_delta[3];
U32 vscan_left_component[3];
U32 vscan_right_component[3];
for(i = 0; i < 3; i++)
{
left_component_delta[i] = (col[i][2] - col[i][0]) >> targetTexelsPerLumel_log2;
right_component_delta[i] = (col[i][3] - col[i][1]) >> targetTexelsPerLumel_log2;
vscan_left_component[i] = col[i][0];
vscan_right_component[i] = col[i][1];
}
// Now we interpolate the color shifts across the square
for(U32 yTexelInLumel = 0; yTexelInLumel < targetTexelsPerLumel; yTexelInLumel++)
{
U32 across_component_delta[3];
U32 hscan_component[3];
for(i = 0; i < 3; i++)
{
across_component_delta[i] = (vscan_right_component[i] - vscan_left_component[i]) >> targetTexelsPerLumel_log2;
hscan_component[i] = vscan_left_component[i];
vscan_left_component[i] += left_component_delta[i];
vscan_right_component[i] += right_component_delta[i];
}
U16 *dstbits = &bits0[ texelInTargetSquare_offset ];
const U8 *srcbits = (U8 *)&bufferToLightFrom[ texelInSquare_offset ];
for(U32 xTexelInLumel = 0; xTexelInLumel < targetTexelsPerLumel; xTexelInLumel++)
{
PROFILE_START(BlendInnermost);
U16 dstcol[3];
for(i = 0; i < 3; i++) // Unroll this dumb loop?
{
U32 index = (hscan_component[i] >> 2) & 0x3F00;
dstcol[i] = alphaTable[index | srcbits[i]];
hscan_component[i] += across_component_delta[i];
}
const U16 max = 255;
dstcol[0] += dstcol[0];
dstcol[1] += dstcol[1];
dstcol[2] += dstcol[2];
dstcol[0] = (dstcol[0] > max) ? max : dstcol[0];
dstcol[1] = (dstcol[1] > max) ? max : dstcol[1];
dstcol[2] = (dstcol[2] > max) ? max : dstcol[2];
#if SRC_IS_ABGR
// NOTE that on Mac, color order is flipped (ABGR1555 instead of RGBA5551), so:
// 1. we already reversed color order via BIG_ENDIAN indexing above, but
// 2. we need to change the shifts for alpha being the high bit instead of the low.
*dstbits++ = ((dstcol[0] & 0xf8) << 7) | ((dstcol[1] & 0xf8) << 2) | ((dstcol[2] & 0xf8) >> 3);
#else
*dstbits++ = ((dstcol[0] & 0xf8) << 8) | ((dstcol[1] & 0xf8) << 3) | ((dstcol[2] & 0xf8) >> 2);
#endif
srcbits += 4;
PROFILE_END();
}
texelInTargetSquare_offset += yStrideThroughTarget;
texelInSquare_offset += yStrideThroughSquare;
}
#endif
PROFILE_END();
xTexelInSquare_offset += xStrideAcrossLumels;
xInLightmap = next_xInLightmap;
}
yInLightmap_offset = next_yInLightmap_offset;
yTexelInTargetSquare_offset += yStrideThroughTargetAcrossLumels;
yTexelInSquare_offset += yStrideThroughSquareAcrossLumels;
}
// end of lighting.
}
}
#if defined(BLENDER_USE_ASM)
if ( targetTexelsPerLumel > 1)
{
cheatmips( destmips[1], destmips[2], destmips[3], 64 );
cheatmips( destmips[3], destmips[4], destmips[5], 16 );
cheatmips4x4( destmips[5], destmips[6], destmips[7] );
}
else
#endif
extrude( destmips, TEXELS_PER_TARGET_EDGE_LOG2 );
PROFILE_END();
}
/// dc - note that this takes an 24bit texture and places it in 32bit space,
/// prob for better alignment/access.
void Blender::addSourceTexture( int textureIndex, const U8 **bmps )
{
int texelsPerSquareEdge = MAX_TEXELS_PER_SQUARE_EDGE;
for ( int mipLevel = 0; mipLevel < num_mip_levels; mipLevel++, texelsPerSquareEdge >>= 1)
{
U32 *dst = bmpdata[ mipLevel * num_src_bmps + textureIndex ];
const U8 *rowPtr = bmps[ mipLevel ];
const U32 texelStride(3);
const U32 colStride(texelsPerSquareEdge * texelStride);
const U32 blkStride(colStride * SQUARES_PER_MIPMAP_EDGE);
const U32 rowStride(texelsPerSquareEdge * blkStride);
// copy the bmp data over, changing the format so each block
// is contiguous.
for ( int row = 0; row < SQUARES_PER_MIPMAP_EDGE; row++, rowPtr += rowStride)
{
const U8* colPtr = rowPtr;
for ( int col = 0; col < SQUARES_PER_MIPMAP_EDGE; col++, colPtr += colStride)
{
const U8 *blkPtr = colPtr;
for ( int py = 0; py < texelsPerSquareEdge; py++, blkPtr += blkStride)
{
const U8* texelPtr = blkPtr;
for ( int px = 0; px < texelsPerSquareEdge; px++, texelPtr += texelStride )
#if SRC_IS_ABGR
*dst++ = (U32(texelPtr[0]) << 24) | (U32(texelPtr[1]) << 16) | (U32(texelPtr[2]) << 8);
#else
*dst++ = texelPtr[0] | (U32(texelPtr[1]) << 8) | (U32(texelPtr[2]) << 16);
#endif
}
}
}
}
}
#define CACHE_ROUND_SHIFT 12
#define CACHE_ROUND_ADJUST ((1 << CACHE_ROUND_SHIFT) - 1)
#define CACHE_ROUND_MASK (~CACHE_ROUND_ADJUST)
#define DWORD_STAGGER 0
static U32 *round_to_cache_start( U32 *ptr )
{
return ( (U32 *) ((dsize_t(ptr) + CACHE_ROUND_ADJUST) & CACHE_ROUND_MASK) );
}
Blender::Blender( int num_src, int num_mips, U8 **alphas )
{
int bmps_size = MAX_TEXELS_PER_SQUARE; // blending buffer (1 square)
int mip_size = TEXELS_PER_SOURCE_BMP;
int i, j;
alpha_data = new U8*[num_src];
for (i = 0; i < num_src; i++)
alpha_data[i] = alphas[i];
num_src_bmps = num_src;
num_mip_levels = num_mips;
bmpdata = new U32*[ num_src * num_mips ];
for ( i = 0; i < num_mips; i++ )
{
bmps_size += (mip_size + DWORD_STAGGER) * num_src;
mip_size >>= 2;
}
bmp_alloc_ptr = new U32[ bmps_size + CACHE_ROUND_ADJUST ];
U32 *bmps = round_to_cache_start( bmp_alloc_ptr );
// buffer that we'll be blending into, and lighting out of.
blendbuffer = bmps;
U32 *curbmp = blendbuffer + MAX_TEXELS_PER_SQUARE;
int bmp_size = TEXELS_PER_SOURCE_BMP;
int bmpnum = 0;
// initialize pointers into buffer for source textures.
for ( j = 0; j < num_mips; j++ )
{
for ( i = 0; i < num_src; i++ )
{
bmpdata[ bmpnum ] = curbmp;
// U32 *bptr = curbmp;
curbmp += (bmp_size + DWORD_STAGGER);
bmpnum++;
}
bmp_size >>= 2;
}
}
Blender::~Blender()
{
if ( bmp_alloc_ptr )
delete [] bmp_alloc_ptr;
if ( bmpdata )
delete [] bmpdata;
if ( alpha_data )
delete [] alpha_data;
}
Reference in New Issue View Git Blame Copy Permalink