zdoom/gzdoom-gles
0
0
Fork 0
mirror of https://github.com/ZDoom/gzdoom-gles.git synced 2024-11-25 13:41:05 +00:00
Code Issues Projects Releases Packages Wiki Activity

- replace the old rect drawers with new ones based on render styles

Browse source
This commit is contained in:
Magnus Norddahl 2018-06-03 18:36:37 +02:00
parent 23fce56b5e
commit d425fb2d4a
7 changed files with 603 additions and 1436 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

59
src/polyrenderer/drawers/poly_draw_args.cpp
View file

@ -226,6 +226,7 @@ void RectDrawArgs::SetTexture(FTexture *texture, FRenderStyle style)
void RectDrawArgs::SetTexture(FTexture *texture, uint32_t translationID, FRenderStyle style) void RectDrawArgs::SetTexture(FTexture *texture, uint32_t translationID, FRenderStyle style)
{ {
// Alphatexture overrides translations.
if (translationID != 0xffffffff && translationID != 0 && !(style.Flags & STYLEF_RedIsAlpha)) if (translationID != 0xffffffff && translationID != 0 && !(style.Flags & STYLEF_RedIsAlpha))
{ {
FRemapTable *table = TranslationToTable(translationID); FRemapTable *table = TranslationToTable(translationID);
@ -299,61 +300,63 @@ void RectDrawArgs::Draw(PolyRenderThread *thread, double x0, double x1, double y
thread->DrawQueue->Push<DrawRectCommand>(*this); thread->DrawQueue->Push<DrawRectCommand>(*this);
} }
void RectDrawArgs::SetStyle(FRenderStyle renderstyle, double alpha, uint32_t fillcolor, uint32_t translationID, FTexture *tex, bool fullbright) void RectDrawArgs::SetStyle(const FRenderStyle &renderstyle, double alpha, uint32_t fillcolor, uint32_t translationID, FTexture *tex, bool fullbright)
{ {
SetTexture(tex, translationID, renderstyle); SetTexture(tex, translationID, renderstyle);
SetColor(0xff000000 | fillcolor, fillcolor >> 24);
if (renderstyle == LegacyRenderStyles[STYLE_Normal] || (r_drawfuzz == 0 && renderstyle == LegacyRenderStyles[STYLE_OptFuzzy])) if (renderstyle == LegacyRenderStyles[STYLE_Normal] || (r_drawfuzz == 0 && renderstyle == LegacyRenderStyles[STYLE_OptFuzzy]))
{ {
SetStyle(Translation() ? RectBlendMode::TranslatedAdd : RectBlendMode::TextureAdd, 1.0, 0.0); SetStyle(Translation() ? TriBlendMode::NormalTranslated : TriBlendMode::Normal, alpha);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_Add] && fullbright && alpha == 1.0 && !Translation()) else if (renderstyle == LegacyRenderStyles[STYLE_Add] && fullbright && alpha == 1.0 && !Translation())
{ {
SetStyle(RectBlendMode::TextureAddSrcColor, 1.0, 1.0); SetStyle(TriBlendMode::SrcColor, alpha);
}
else if (renderstyle == LegacyRenderStyles[STYLE_Add])
{
SetStyle(Translation() ? RectBlendMode::TranslatedAdd : RectBlendMode::TextureAdd, alpha, 1.0);
}
else if (renderstyle == LegacyRenderStyles[STYLE_Subtract])
{
SetStyle(Translation() ? RectBlendMode::TranslatedRevSub : RectBlendMode::TextureRevSub, alpha, 1.0);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_SoulTrans]) else if (renderstyle == LegacyRenderStyles[STYLE_SoulTrans])
{ {
SetStyle(Translation() ? RectBlendMode::TranslatedAdd : RectBlendMode::TextureAdd, transsouls, 1.0 - transsouls); SetStyle(Translation() ? TriBlendMode::AddTranslated : TriBlendMode::Add, transsouls);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_Fuzzy] || (r_drawfuzz == 1 && renderstyle == LegacyRenderStyles[STYLE_OptFuzzy])) else if (renderstyle == LegacyRenderStyles[STYLE_Fuzzy] || (r_drawfuzz == 1 && renderstyle == LegacyRenderStyles[STYLE_OptFuzzy]))
{ {
SetColor(0xff000000, 0); SetColor(0xff000000, 0);
SetStyle(RectBlendMode::Fuzz); SetStyle(TriBlendMode::Fuzzy);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_Shadow] || (r_drawfuzz == 2 && renderstyle == LegacyRenderStyles[STYLE_OptFuzzy])) else if (renderstyle == LegacyRenderStyles[STYLE_Shadow] || (r_drawfuzz == 2 && renderstyle == LegacyRenderStyles[STYLE_OptFuzzy]))
{ {
SetStyle(Translation() ? RectBlendMode::TranslatedAdd : RectBlendMode::TextureAdd, 0.0, 160 / 255.0); SetColor(0xff000000, 0);
SetStyle(Translation() ? TriBlendMode::TranslucentStencilTranslated : TriBlendMode::TranslucentStencil, 1.0 - 160 / 255.0);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_TranslucentStencil]) else if (renderstyle == LegacyRenderStyles[STYLE_Stencil])
{ {
SetColor(0xff000000 | fillcolor, fillcolor >> 24); SetStyle(Translation() ? TriBlendMode::StencilTranslated : TriBlendMode::Stencil, alpha);
SetStyle(RectBlendMode::Stencil, alpha, 1.0 - alpha);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_AddStencil]) else if (renderstyle == LegacyRenderStyles[STYLE_Translucent])
{ {
SetColor(0xff000000 | fillcolor, fillcolor >> 24); SetStyle(Translation() ? TriBlendMode::TranslucentTranslated : TriBlendMode::Translucent, alpha);
SetStyle(RectBlendMode::AddStencil, alpha, 1.0); }
else if (renderstyle == LegacyRenderStyles[STYLE_Add])
{
SetStyle(Translation() ? TriBlendMode::AddTranslated : TriBlendMode::Add, alpha);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_Shaded]) else if (renderstyle == LegacyRenderStyles[STYLE_Shaded])
{ {
SetColor(0xff000000 | fillcolor, fillcolor >> 24); SetStyle(Translation() ? TriBlendMode::ShadedTranslated : TriBlendMode::Shaded, alpha);
SetStyle(RectBlendMode::Shaded, alpha, 1.0 - alpha); }
else if (renderstyle == LegacyRenderStyles[STYLE_TranslucentStencil])
{
SetStyle(Translation() ? TriBlendMode::TranslucentStencilTranslated : TriBlendMode::TranslucentStencil, alpha);
}
else if (renderstyle == LegacyRenderStyles[STYLE_Subtract])
{
SetStyle(Translation() ? TriBlendMode::SubtractTranslated : TriBlendMode::Subtract, alpha);
}
else if (renderstyle == LegacyRenderStyles[STYLE_AddStencil])
{
SetStyle(Translation() ? TriBlendMode::AddStencilTranslated : TriBlendMode::AddStencil, alpha);
} }
else if (renderstyle == LegacyRenderStyles[STYLE_AddShaded]) else if (renderstyle == LegacyRenderStyles[STYLE_AddShaded])
{ {
SetColor(0xff000000 | fillcolor, fillcolor >> 24); SetStyle(Translation() ? TriBlendMode::AddShadedTranslated : TriBlendMode::AddShaded, alpha);
SetStyle(RectBlendMode::AddShaded, alpha, 1.0);
}
else
{
SetStyle(Translation() ? RectBlendMode::TranslatedAdd : RectBlendMode::TextureAdd, alpha, 1.0 - alpha);
} }
} }

14
src/polyrenderer/drawers/poly_draw_args.h
View file

@ -180,8 +180,8 @@ public:
void SetTexture(FTexture *texture, FRenderStyle style); void SetTexture(FTexture *texture, FRenderStyle style);
void SetTexture(FTexture *texture, uint32_t translationID, FRenderStyle style); void SetTexture(FTexture *texture, uint32_t translationID, FRenderStyle style);
void SetLight(FSWColormap *basecolormap, uint32_t lightlevel); void SetLight(FSWColormap *basecolormap, uint32_t lightlevel);
void SetStyle(RectBlendMode blendmode, double srcalpha = 1.0, double destalpha = 1.0) { mBlendMode = blendmode; mSrcAlpha = (uint32_t)(srcalpha * 256.0 + 0.5); mDestAlpha = (uint32_t)(destalpha * 256.0 + 0.5); } void SetStyle(TriBlendMode blendmode, double alpha = 1.0) { mBlendMode = blendmode; mAlpha = (uint32_t)(alpha * 256.0 + 0.5); }
void SetStyle(FRenderStyle renderstyle, double alpha, uint32_t fillcolor, uint32_t translationID, FTexture *texture, bool fullbright); void SetStyle(const FRenderStyle &renderstyle, double alpha, uint32_t fillcolor, uint32_t translationID, FTexture *texture, bool fullbright);
void SetColor(uint32_t bgra, uint8_t palindex); void SetColor(uint32_t bgra, uint8_t palindex);
void Draw(PolyRenderThread *thread, double x0, double x1, double y0, double y1, double u0, double u1, double v0, double v1); void Draw(PolyRenderThread *thread, double x0, double x1, double y0, double y1, double u0, double u1, double v0, double v1);
@ -191,10 +191,9 @@ public:
int TextureHeight() const { return mTextureHeight; } int TextureHeight() const { return mTextureHeight; }
const uint8_t *Translation() const { return mTranslation; } const uint8_t *Translation() const { return mTranslation; }
RectBlendMode BlendMode() const { return mBlendMode; } TriBlendMode BlendMode() const { return mBlendMode; }
uint32_t Color() const { return mColor; } uint32_t Color() const { return mColor; }
uint32_t SrcAlpha() const { return mSrcAlpha; } uint32_t Alpha() const { return mAlpha; }
uint32_t DestAlpha() const { return mDestAlpha; }
uint32_t Light() const { return mLight; } uint32_t Light() const { return mLight; }
const uint8_t *BaseColormap() const { return mColormaps; } const uint8_t *BaseColormap() const { return mColormaps; }
@ -225,11 +224,10 @@ private:
int mTextureHeight = 0; int mTextureHeight = 0;
const uint8_t *mTranslation = nullptr; const uint8_t *mTranslation = nullptr;
const uint8_t *mColormaps = nullptr; const uint8_t *mColormaps = nullptr;
RectBlendMode mBlendMode = RectBlendMode::FillOpaque; TriBlendMode mBlendMode = TriBlendMode::Fill;
uint32_t mLight = 0; uint32_t mLight = 0;
uint32_t mColor = 0; uint32_t mColor = 0;
uint32_t mSrcAlpha = 0; uint32_t mAlpha = 0;
uint32_t mDestAlpha = 0;
uint16_t mLightAlpha = 0; uint16_t mLightAlpha = 0;
uint16_t mLightRed = 0; uint16_t mLightRed = 0;
uint16_t mLightGreen = 0; uint16_t mLightGreen = 0;

476
src/polyrenderer/drawers/poly_drawer32.h
View file

@ -1,476 +0,0 @@
/*
** Polygon Doom software renderer
** Copyright (c) 2016 Magnus Norddahl
**
** This software is provided 'as-is', without any express or implied
** warranty. In no event will the authors be held liable for any damages
** arising from the use of this software.
**
** Permission is granted to anyone to use this software for any purpose,
** including commercial applications, and to alter it and redistribute it
** freely, subject to the following restrictions:
**
** 1. The origin of this software must not be misrepresented; you must not
** claim that you wrote the original software. If you use this software
** in a product, an acknowledgment in the product documentation would be
** appreciated but is not required.
** 2. Altered source versions must be plainly marked as such, and must not be
** misrepresented as being the original software.
** 3. This notice may not be removed or altered from any source distribution.
**
*/
#pragma once
#include "screen_triangle.h"
namespace TriScreenDrawerModes
{
namespace
{
struct BgraColor
{
uint32_t b, g, r, a;
BgraColor() { }
BgraColor(uint32_t c) : b(BPART(c)), g(GPART(c)), r(RPART(c)), a(APART(c)) { }
BgraColor &operator=(uint32_t c) { b = BPART(c); g = GPART(c); r = RPART(c); a = APART(c); return *this; }
operator uint32_t() const { return MAKEARGB(a, r, g, b); }
};
}
template<typename SamplerT, typename FilterModeT>
FORCEINLINE unsigned int Sample32(int32_t u, int32_t v, const uint32_t *texPixels, int texWidth, int texHeight, uint32_t oneU, uint32_t oneV, uint32_t color, const uint32_t *translation)
{
uint32_t texel;
if (SamplerT::Mode == (int)Samplers::Shaded || SamplerT::Mode == (int)Samplers::Stencil || SamplerT::Mode == (int)Samplers::Fill || SamplerT::Mode == (int)Samplers::Fuzz || SamplerT::Mode == (int)Samplers::FogBoundary)
{
return color;
}
else if (SamplerT::Mode == (int)Samplers::Translated)
{
const uint8_t *texpal = (const uint8_t *)texPixels;
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
return translation[texpal[texelX * texHeight + texelY]];
}
else if (FilterModeT::Mode == (int)FilterModes::Nearest)
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
texel = texPixels[texelX * texHeight + texelY];
}
else
{
u -= oneU >> 1;
v -= oneV >> 1;
unsigned int frac_x0 = (((uint32_t)u << 8) >> FRACBITS) * texWidth;
unsigned int frac_x1 = ((((uint32_t)u << 8) + oneU) >> FRACBITS) * texWidth;
unsigned int frac_y0 = (((uint32_t)v << 8) >> FRACBITS) * texHeight;
unsigned int frac_y1 = ((((uint32_t)v << 8) + oneV) >> FRACBITS) * texHeight;
unsigned int x0 = frac_x0 >> FRACBITS;
unsigned int x1 = frac_x1 >> FRACBITS;
unsigned int y0 = frac_y0 >> FRACBITS;
unsigned int y1 = frac_y1 >> FRACBITS;
unsigned int p00 = texPixels[x0 * texHeight + y0];
unsigned int p01 = texPixels[x0 * texHeight + y1];
unsigned int p10 = texPixels[x1 * texHeight + y0];
unsigned int p11 = texPixels[x1 * texHeight + y1];
unsigned int inv_a = (frac_x1 >> (FRACBITS - 4)) & 15;
unsigned int inv_b = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
unsigned int sgreen = (GPART(p00) * (a * b) + GPART(p01) * (inv_a * b) + GPART(p10) * (a * inv_b) + GPART(p11) * (inv_a * inv_b) + 127) >> 8;
unsigned int sblue = (BPART(p00) * (a * b) + BPART(p01) * (inv_a * b) + BPART(p10) * (a * inv_b) + BPART(p11) * (inv_a * inv_b) + 127) >> 8;
unsigned int salpha = (APART(p00) * (a * b) + APART(p01) * (inv_a * b) + APART(p10) * (a * inv_b) + APART(p11) * (inv_a * inv_b) + 127) >> 8;
texel = (salpha << 24) | (sred << 16) | (sgreen << 8) | sblue;
}
if (SamplerT::Mode == (int)Samplers::Skycap)
{
int start_fade = 2; // How fast it should fade out
int alpha_top = clamp(v >> (16 - start_fade), 0, 256);
int alpha_bottom = clamp(((2 << 24) - v) >> (16 - start_fade), 0, 256);
int a = MIN(alpha_top, alpha_bottom);
int inv_a = 256 - a;
uint32_t r = RPART(texel);
uint32_t g = GPART(texel);
uint32_t b = BPART(texel);
uint32_t fg_a = APART(texel);
uint32_t bg_red = RPART(color);
uint32_t bg_green = GPART(color);
uint32_t bg_blue = BPART(color);
r = (r * a + bg_red * inv_a + 127) >> 8;
g = (g * a + bg_green * inv_a + 127) >> 8;
b = (b * a + bg_blue * inv_a + 127) >> 8;
return MAKEARGB(fg_a, r, g, b);
}
else
{
return texel;
}
}
template<typename SamplerT>
FORCEINLINE unsigned int SampleShade32(int32_t u, int32_t v, const uint32_t *texPixels, int texWidth, int texHeight, int x, int y)
{
if (SamplerT::Mode == (int)Samplers::Shaded)
{
const uint8_t *texpal = (const uint8_t *)texPixels;
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = texpal[texelX * texHeight + texelY];
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
return sampleshadeout;
}
else if (SamplerT::Mode == (int)Samplers::Stencil)
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = APART(texPixels[texelX * texHeight + texelY]);
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
return sampleshadeout;
}
else if (SamplerT::Mode == (int)Samplers::Fuzz)
{
using namespace swrenderer;
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = APART(texPixels[texelX * texHeight + texelY]);
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
fixed_t fuzzscale = (200 << FRACBITS) / viewheight;
int scaled_x = (x * fuzzscale) >> FRACBITS;
int fuzz_x = fuzz_random_x_offset[scaled_x % FUZZ_RANDOM_X_SIZE] + fuzzpos;
fixed_t fuzzcount = FUZZTABLE << FRACBITS;
fixed_t fuzz = ((fuzz_x << FRACBITS) + y * fuzzscale) % fuzzcount;
unsigned int alpha = fuzzoffset[fuzz >> FRACBITS];
sampleshadeout = (sampleshadeout * alpha) >> 5;
return sampleshadeout;
}
else
{
return 0;
}
}
FORCEINLINE BgraColor VECTORCALL AddLights(BgraColor material, BgraColor fgcolor, BgraColor dynlight)
{
fgcolor.r = MIN(fgcolor.r + ((material.r * dynlight.r) >> 8), (uint32_t)255);
fgcolor.g = MIN(fgcolor.g + ((material.g * dynlight.g) >> 8), (uint32_t)255);
fgcolor.b = MIN(fgcolor.b + ((material.b * dynlight.b) >> 8), (uint32_t)255);
return fgcolor;
}
FORCEINLINE BgraColor VECTORCALL CalcDynamicLight(const PolyLight *lights, int num_lights, FVector3 worldpos, FVector3 worldnormal, uint32_t dynlightcolor)
{
BgraColor lit = dynlightcolor;
for (int i = 0; i != num_lights; i++)
{
FVector3 lightpos = { lights[i].x, lights[i].y, lights[i].z };
float light_radius = lights[i].radius;
bool is_attenuated = light_radius < 0.0f;
if (is_attenuated)
light_radius = -light_radius;
// L = light-pos
// dist = sqrt(dot(L, L))
// distance_attenuation = 1 - MIN(dist * (1/radius), 1)
FVector3 L = lightpos - worldpos;
float dist2 = L | L;
float rcp_dist = 1.0f / sqrt(dist2);
float dist = dist2 * rcp_dist;
float distance_attenuation = 256.0f - MIN(dist * light_radius, 256.0f);
// The simple light type
float simple_attenuation = distance_attenuation;
// The point light type
// diffuse = max(dot(N,normalize(L)),0) * attenuation
float dotNL = worldnormal | (L * rcp_dist);
float point_attenuation = MAX(dotNL, 0.0f) * distance_attenuation;
uint32_t attenuation = (uint32_t)(is_attenuated ? (int32_t)point_attenuation : (int32_t)simple_attenuation);
BgraColor light_color = lights[i].color;
lit.r += (light_color.r * attenuation) >> 8;
lit.g += (light_color.g * attenuation) >> 8;
lit.b += (light_color.b * attenuation) >> 8;
}
lit.r = MIN(lit.r, (uint32_t)256);
lit.g = MIN(lit.g, (uint32_t)256);
lit.b = MIN(lit.b, (uint32_t)256);
return lit;
}
template<typename ShadeModeT>
FORCEINLINE BgraColor Shade32(BgraColor fgcolor, BgraColor mlight, uint32_t desaturate, uint32_t inv_desaturate, BgraColor shade_fade, BgraColor shade_light, BgraColor dynlight)
{
BgraColor material = fgcolor;
if (ShadeModeT::Mode == (int)ShadeMode::Simple)
{
fgcolor.r = (fgcolor.r * mlight.r) >> 8;
fgcolor.g = (fgcolor.g * mlight.g) >> 8;
fgcolor.b = (fgcolor.b * mlight.b) >> 8;
}
else if (ShadeModeT::Mode == (int)ShadeMode::Advanced)
{
uint32_t intensity = ((fgcolor.r * 77 + fgcolor.g * 143 + fgcolor.b * 37) >> 8) * desaturate;
fgcolor.r = (((shade_fade.r + ((fgcolor.r * inv_desaturate + intensity) >> 8) * mlight.r) >> 8) * shade_light.r) >> 8;
fgcolor.g = (((shade_fade.g + ((fgcolor.g * inv_desaturate + intensity) >> 8) * mlight.g) >> 8) * shade_light.g) >> 8;
fgcolor.b = (((shade_fade.b + ((fgcolor.b * inv_desaturate + intensity) >> 8) * mlight.b) >> 8) * shade_light.b) >> 8;
}
return AddLights(material, fgcolor, dynlight);
}
template<typename BlendT>
FORCEINLINE BgraColor Blend32(BgraColor fgcolor, BgraColor bgcolor, uint32_t ifgcolor, uint32_t ifgshade, uint32_t srcalpha, uint32_t destalpha)
{
if (BlendT::Mode == (int)BlendModes::Opaque)
{
fgcolor.a = 255;
return fgcolor;
}
else if (BlendT::Mode == (int)BlendModes::Masked)
{
return (ifgcolor == 0) ? bgcolor : fgcolor;
}
else if (BlendT::Mode == (int)BlendModes::AddSrcColorOneMinusSrcColor)
{
uint32_t srcred = fgcolor.r + (fgcolor.r >> 7);
uint32_t srcgreen = fgcolor.g + (fgcolor.g >> 7);
uint32_t srcblue = fgcolor.b + (fgcolor.b >> 7);
uint32_t inv_srcred = 256 - srcred;
uint32_t inv_srcgreen = 256 - srcgreen;
uint32_t inv_srcblue = 256 - srcblue;
BgraColor outcolor;
outcolor.r = (fgcolor.r * srcred + bgcolor.r * inv_srcred) >> 8;
outcolor.g = (fgcolor.g * srcgreen + bgcolor.g * inv_srcgreen) >> 8;
outcolor.b = (fgcolor.b * srcblue + bgcolor.b * inv_srcblue) >> 8;
outcolor.a = 255;
return outcolor;
}
else if (BlendT::Mode == (int)BlendModes::Shaded)
{
uint32_t alpha = ifgshade;
uint32_t inv_alpha = 256 - alpha;
BgraColor outcolor;
outcolor.r = (fgcolor.r * alpha + bgcolor.r * inv_alpha) >> 8;
outcolor.g = (fgcolor.g * alpha + bgcolor.g * inv_alpha) >> 8;
outcolor.b = (fgcolor.b * alpha + bgcolor.b * inv_alpha) >> 8;
outcolor.a = 255;
return outcolor;
}
else if (BlendT::Mode == (int)BlendModes::AddClampShaded)
{
uint32_t alpha = ifgshade;
BgraColor outcolor;
outcolor.r = ((fgcolor.r * alpha) >> 8) + bgcolor.r;
outcolor.g = ((fgcolor.g * alpha) >> 8) + bgcolor.g;
outcolor.b = ((fgcolor.b * alpha) >> 8) + bgcolor.b;
outcolor.a = 255;
return outcolor;
}
else
{
uint32_t alpha = APART(ifgcolor);
alpha += alpha >> 7; // 255->256
uint32_t inv_alpha = 256 - alpha;
uint32_t bgalpha = (destalpha * alpha + (inv_alpha << 8) + 128) >> 8;
uint32_t fgalpha = (srcalpha * alpha + 128) >> 8;
fgcolor.r *= fgalpha;
fgcolor.g *= fgalpha;
fgcolor.b *= fgalpha;
bgcolor.r *= bgalpha;
bgcolor.g *= bgalpha;
bgcolor.b *= bgalpha;
BgraColor outcolor;
if (BlendT::Mode == (int)BlendModes::AddClamp)
{
outcolor.r = MIN<uint32_t>((fgcolor.r + bgcolor.r) >> 8, 255);
outcolor.g = MIN<uint32_t>((fgcolor.g + bgcolor.g) >> 8, 255);
outcolor.b = MIN<uint32_t>((fgcolor.b + bgcolor.b) >> 8, 255);
}
else if (BlendT::Mode == (int)BlendModes::SubClamp)
{
outcolor.r = MAX(int32_t(fgcolor.r - bgcolor.r) >> 8, 0);
outcolor.g = MAX(int32_t(fgcolor.g - bgcolor.g) >> 8, 0);
outcolor.b = MAX(int32_t(fgcolor.b - bgcolor.b) >> 8, 0);
}
else if (BlendT::Mode == (int)BlendModes::RevSubClamp)
{
outcolor.r = MAX(int32_t(bgcolor.r - fgcolor.r) >> 8, 0);
outcolor.g = MAX(int32_t(bgcolor.g - fgcolor.g) >> 8, 0);
outcolor.b = MAX(int32_t(bgcolor.b - fgcolor.b) >> 8, 0);
}
outcolor.a = 255;
return outcolor;
}
}
}
template<typename BlendT, typename SamplerT>
class RectScreenDrawer32
{
public:
static void Execute(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
if (SamplerT::Mode == (int)Samplers::Fuzz)
{
Loop<NoShade, NearestFilter>(destOrg, destWidth, destHeight, destPitch, args, thread);
}
else if (args->SimpleShade())
{
Loop<SimpleShade, NearestFilter>(destOrg, destWidth, destHeight, destPitch, args, thread);
}
else
{
Loop<AdvancedShade, NearestFilter>(destOrg, destWidth, destHeight, destPitch, args, thread);
}
}
private:
template<typename ShadeModeT, typename FilterModeT>
FORCEINLINE static void Loop(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
int x0 = clamp((int)(args->X0() + 0.5f), 0, destWidth);
int x1 = clamp((int)(args->X1() + 0.5f), 0, destWidth);
int y0 = clamp((int)(args->Y0() + 0.5f), 0, destHeight);
int y1 = clamp((int)(args->Y1() + 0.5f), 0, destHeight);
if (x1 <= x0 || y1 <= y0)
return;
uint32_t srcalpha = args->SrcAlpha();
uint32_t destalpha = args->DestAlpha();
// Setup step variables
float fstepU = (args->U1() - args->U0()) / (args->X1() - args->X0());
float fstepV = (args->V1() - args->V0()) / (args->Y1() - args->Y0());
uint32_t startU = (int32_t)((args->U0() + (x0 + 0.5f - args->X0()) * fstepU) * 0x1000000);
uint32_t startV = (int32_t)((args->V0() + (y0 + 0.5f - args->Y0()) * fstepV) * 0x1000000);
uint32_t stepU = (int32_t)(fstepU * 0x1000000);
uint32_t stepV = (int32_t)(fstepV * 0x1000000);
// Sampling stuff
uint32_t color = args->Color();
const uint32_t * RESTRICT translation = (const uint32_t *)args->Translation();
const uint32_t * RESTRICT texPixels = (const uint32_t *)args->TexturePixels();
uint32_t texWidth = args->TextureWidth();
uint32_t texHeight = args->TextureHeight();
uint32_t oneU, oneV;
if (SamplerT::Mode != (int)Samplers::Fill)
{
oneU = ((0x800000 + texWidth - 1) / texWidth) * 2 + 1;
oneV = ((0x800000 + texHeight - 1) / texHeight) * 2 + 1;
}
else
{
oneU = 0;
oneV = 0;
}
// Setup light
uint32_t lightpos = args->Light();
lightpos += lightpos >> 7; // 255 -> 256
BgraColor mlight;
BgraColor dynlight = 0;
// Shade constants
int inv_desaturate;
BgraColor shade_fade_lit, shade_light;
int desaturate;
if (ShadeModeT::Mode == (int)ShadeMode::Advanced)
{
uint32_t inv_light = 256 - lightpos;
shade_fade_lit.r = args->ShadeFadeRed() * inv_light;
shade_fade_lit.g = args->ShadeFadeGreen() * inv_light;
shade_fade_lit.b = args->ShadeFadeBlue() * inv_light;
shade_light.r = args->ShadeLightRed();
shade_light.g = args->ShadeLightGreen();
shade_light.b = args->ShadeLightBlue();
desaturate = args->ShadeDesaturate();
inv_desaturate = 256 - desaturate;
mlight.r = lightpos;
mlight.g = lightpos;
mlight.b = lightpos;
}
else
{
inv_desaturate = 0;
shade_fade_lit.r = 0;
shade_fade_lit.g = 0;
shade_fade_lit.b = 0;
shade_light.r = 0;
shade_light.g = 0;
shade_light.b = 0;
desaturate = 0;
mlight.r = lightpos;
mlight.g = lightpos;
mlight.b = lightpos;
}
int count = x1 - x0;
uint32_t posV = startV;
for (int y = y0; y < y1; y++, posV += stepV)
{
int coreBlock = y / 8;
if (coreBlock % thread->num_cores != thread->core)
{
continue;
}
uint32_t *dest = ((uint32_t*)destOrg) + y * destPitch + x0;
uint32_t posU = startU;
for (int i = 0; i < count; i++)
{
// Load bgcolor
BgraColor bgcolor;
if (BlendT::Mode != (int)BlendModes::Opaque)
bgcolor = *dest;
else
bgcolor = 0;
// Sample fgcolor
if (SamplerT::Mode == (int)Samplers::FogBoundary) color = *dest;
unsigned int ifgcolor = Sample32<SamplerT, FilterModeT>(posU, posV, texPixels, texWidth, texHeight, oneU, oneV, color, translation);
unsigned int ifgshade = SampleShade32<SamplerT>(posU, posV, texPixels, texWidth, texHeight, x0 + i, y);
posU += stepU;
// Shade and blend
BgraColor fgcolor = Shade32<ShadeModeT>(ifgcolor, mlight, desaturate, inv_desaturate, shade_fade_lit, shade_light, dynlight);
BgraColor outcolor = Blend32<BlendT>(fgcolor, bgcolor, ifgcolor, ifgshade, srcalpha, destalpha);
// Store result
*dest = outcolor;
dest++;
}
}
}
};

518
src/polyrenderer/drawers/poly_drawer32_sse2.h
View file

@ -1,518 +0,0 @@
/*
** Polygon Doom software renderer
** Copyright (c) 2016 Magnus Norddahl
**
** This software is provided 'as-is', without any express or implied
** warranty. In no event will the authors be held liable for any damages
** arising from the use of this software.
**
** Permission is granted to anyone to use this software for any purpose,
** including commercial applications, and to alter it and redistribute it
** freely, subject to the following restrictions:
**
** 1. The origin of this software must not be misrepresented; you must not
** claim that you wrote the original software. If you use this software
** in a product, an acknowledgment in the product documentation would be
** appreciated but is not required.
** 2. Altered source versions must be plainly marked as such, and must not be
** misrepresented as being the original software.
** 3. This notice may not be removed or altered from any source distribution.
**
*/
#pragma once
#include "screen_triangle.h"
namespace TriScreenDrawerModes
{
template<typename SamplerT, typename FilterModeT>
FORCEINLINE unsigned int VECTORCALL Sample32(int32_t u, int32_t v, const uint32_t *texPixels, int texWidth, int texHeight, uint32_t oneU, uint32_t oneV, uint32_t color, const uint32_t *translation)
{
uint32_t texel;
if (SamplerT::Mode == (int)Samplers::Shaded || SamplerT::Mode == (int)Samplers::Stencil || SamplerT::Mode == (int)Samplers::Fill || SamplerT::Mode == (int)Samplers::Fuzz || SamplerT::Mode == (int)Samplers::FogBoundary)
{
return color;
}
else if (SamplerT::Mode == (int)Samplers::Translated)
{
const uint8_t *texpal = (const uint8_t *)texPixels;
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
return translation[texpal[texelX * texHeight + texelY]];
}
else if (FilterModeT::Mode == (int)FilterModes::Nearest)
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
texel = texPixels[texelX * texHeight + texelY];
}
else
{
u -= oneU >> 1;
v -= oneV >> 1;
unsigned int frac_x0 = (((uint32_t)u << 8) >> FRACBITS) * texWidth;
unsigned int frac_x1 = ((((uint32_t)u << 8) + oneU) >> FRACBITS) * texWidth;
unsigned int frac_y0 = (((uint32_t)v << 8) >> FRACBITS) * texHeight;
unsigned int frac_y1 = ((((uint32_t)v << 8) + oneV) >> FRACBITS) * texHeight;
unsigned int x0 = frac_x0 >> FRACBITS;
unsigned int x1 = frac_x1 >> FRACBITS;
unsigned int y0 = frac_y0 >> FRACBITS;
unsigned int y1 = frac_y1 >> FRACBITS;
unsigned int p00 = texPixels[x0 * texHeight + y0];
unsigned int p01 = texPixels[x0 * texHeight + y1];
unsigned int p10 = texPixels[x1 * texHeight + y0];
unsigned int p11 = texPixels[x1 * texHeight + y1];
unsigned int inv_a = (frac_x1 >> (FRACBITS - 4)) & 15;
unsigned int inv_b = (frac_y1 >> (FRACBITS - 4)) & 15;
unsigned int a = 16 - inv_a;
unsigned int b = 16 - inv_b;
unsigned int sred = (RPART(p00) * (a * b) + RPART(p01) * (inv_a * b) + RPART(p10) * (a * inv_b) + RPART(p11) * (inv_a * inv_b) + 127) >> 8;
unsigned int sgreen = (GPART(p00) * (a * b) + GPART(p01) * (inv_a * b) + GPART(p10) * (a * inv_b) + GPART(p11) * (inv_a * inv_b) + 127) >> 8;
unsigned int sblue = (BPART(p00) * (a * b) + BPART(p01) * (inv_a * b) + BPART(p10) * (a * inv_b) + BPART(p11) * (inv_a * inv_b) + 127) >> 8;
unsigned int salpha = (APART(p00) * (a * b) + APART(p01) * (inv_a * b) + APART(p10) * (a * inv_b) + APART(p11) * (inv_a * inv_b) + 127) >> 8;
texel = (salpha << 24) | (sred << 16) | (sgreen << 8) | sblue;
}
if (SamplerT::Mode == (int)Samplers::Skycap)
{
int start_fade = 2; // How fast it should fade out
int alpha_top = clamp(v >> (16 - start_fade), 0, 256);
int alpha_bottom = clamp(((2 << 24) - v) >> (16 - start_fade), 0, 256);
int a = MIN(alpha_top, alpha_bottom);
int inv_a = 256 - a;
uint32_t r = RPART(texel);
uint32_t g = GPART(texel);
uint32_t b = BPART(texel);
uint32_t fg_a = APART(texel);
uint32_t bg_red = RPART(color);
uint32_t bg_green = GPART(color);
uint32_t bg_blue = BPART(color);
r = (r * a + bg_red * inv_a + 127) >> 8;
g = (g * a + bg_green * inv_a + 127) >> 8;
b = (b * a + bg_blue * inv_a + 127) >> 8;
return MAKEARGB(fg_a, r, g, b);
}
else
{
return texel;
}
}
template<typename SamplerT>
FORCEINLINE unsigned int VECTORCALL SampleShade32(int32_t u, int32_t v, const uint32_t *texPixels, int texWidth, int texHeight, int x, int y)
{
if (SamplerT::Mode == (int)Samplers::Shaded)
{
const uint8_t *texpal = (const uint8_t *)texPixels;
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = texpal[texelX * texHeight + texelY];
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
return sampleshadeout;
}
else if (SamplerT::Mode == (int)Samplers::Stencil)
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = APART(texPixels[texelX * texHeight + texelY]);
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
return sampleshadeout;
}
else if (SamplerT::Mode == (int)Samplers::Fuzz)
{
using namespace swrenderer;
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = APART(texPixels[texelX * texHeight + texelY]);
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
fixed_t fuzzscale = (200 << FRACBITS) / viewheight;
int scaled_x = (x * fuzzscale) >> FRACBITS;
int fuzz_x = fuzz_random_x_offset[scaled_x % FUZZ_RANDOM_X_SIZE] + fuzzpos;
fixed_t fuzzcount = FUZZTABLE << FRACBITS;
fixed_t fuzz = ((fuzz_x << FRACBITS) + y * fuzzscale) % fuzzcount;
unsigned int alpha = fuzzoffset[fuzz >> FRACBITS];
sampleshadeout = (sampleshadeout * alpha) >> 5;
return sampleshadeout;
}
else
{
return 0;
}
}
FORCEINLINE __m128i VECTORCALL AddLights(__m128i material, __m128i fgcolor, __m128i dynlight)
{
fgcolor = _mm_add_epi16(fgcolor, _mm_srli_epi16(_mm_mullo_epi16(material, dynlight), 8));
fgcolor = _mm_min_epi16(fgcolor, _mm_set1_epi16(255));
return fgcolor;
}
FORCEINLINE __m128i VECTORCALL CalcDynamicLight(const PolyLight *lights, int num_lights, __m128 worldpos, __m128 worldnormal, uint32_t dynlightcolor)
{
__m128i lit = _mm_unpacklo_epi8(_mm_cvtsi32_si128(dynlightcolor), _mm_setzero_si128());
lit = _mm_shuffle_epi32(lit, _MM_SHUFFLE(1, 0, 1, 0));
for (int i = 0; i != num_lights; i++)
{
__m128 m256 = _mm_set1_ps(256.0f);
__m128 mSignBit = _mm_set1_ps(-0.0f);
__m128 lightpos = _mm_loadu_ps(&lights[i].x);
__m128 light_radius = _mm_load_ss(&lights[i].radius);
__m128 is_attenuated = _mm_cmpge_ss(light_radius, _mm_setzero_ps());
is_attenuated = _mm_shuffle_ps(is_attenuated, is_attenuated, _MM_SHUFFLE(0, 0, 0, 0));
light_radius = _mm_andnot_ps(mSignBit, light_radius);
// L = light-pos
// dist = sqrt(dot(L, L))
// distance_attenuation = 1 - MIN(dist * (1/radius), 1)
__m128 L = _mm_sub_ps(lightpos, worldpos);
__m128 dist2 = _mm_mul_ps(L, L);
dist2 = _mm_add_ss(dist2, _mm_add_ss(_mm_shuffle_ps(dist2, dist2, _MM_SHUFFLE(0, 0, 0, 1)), _mm_shuffle_ps(dist2, dist2, _MM_SHUFFLE(0, 0, 0, 2))));
__m128 rcp_dist = _mm_rsqrt_ss(dist2);
__m128 dist = _mm_mul_ss(dist2, rcp_dist);
__m128 distance_attenuation = _mm_sub_ss(m256, _mm_min_ss(_mm_mul_ss(dist, light_radius), m256));
distance_attenuation = _mm_shuffle_ps(distance_attenuation, distance_attenuation, _MM_SHUFFLE(0, 0, 0, 0));
// The simple light type
__m128 simple_attenuation = distance_attenuation;
// The point light type
// diffuse = max(dot(N,normalize(L)),0) * attenuation
__m128 dotNL = _mm_mul_ps(worldnormal, _mm_mul_ps(L, _mm_shuffle_ps(rcp_dist, rcp_dist, _MM_SHUFFLE(0, 0, 0, 0))));
dotNL = _mm_add_ss(dotNL, _mm_add_ss(_mm_shuffle_ps(dotNL, dotNL, _MM_SHUFFLE(0, 0, 0, 1)), _mm_shuffle_ps(dotNL, dotNL, _MM_SHUFFLE(0, 0, 0, 2))));
dotNL = _mm_max_ss(dotNL, _mm_setzero_ps());
__m128 point_attenuation = _mm_mul_ss(dotNL, distance_attenuation);
point_attenuation = _mm_shuffle_ps(point_attenuation, point_attenuation, _MM_SHUFFLE(0, 0, 0, 0));
__m128i attenuation = _mm_cvtps_epi32(_mm_or_ps(_mm_and_ps(is_attenuated, simple_attenuation), _mm_andnot_ps(is_attenuated, point_attenuation)));
attenuation = _mm_packs_epi32(_mm_shuffle_epi32(attenuation, _MM_SHUFFLE(0, 0, 0, 0)), _mm_shuffle_epi32(attenuation, _MM_SHUFFLE(1, 1, 1, 1)));
__m128i light_color = _mm_cvtsi32_si128(lights[i].color);
light_color = _mm_unpacklo_epi8(light_color, _mm_setzero_si128());
light_color = _mm_shuffle_epi32(light_color, _MM_SHUFFLE(1, 0, 1, 0));
lit = _mm_add_epi16(lit, _mm_srli_epi16(_mm_mullo_epi16(light_color, attenuation), 8));
}
return _mm_min_epi16(lit, _mm_set1_epi16(256));
}
template<typename ShadeModeT>
FORCEINLINE __m128i VECTORCALL Shade32(__m128i fgcolor, __m128i mlight, unsigned int ifgcolor0, unsigned int ifgcolor1, int desaturate, __m128i inv_desaturate, __m128i shade_fade, __m128i shade_light, __m128i dynlight)
{
__m128i material = fgcolor;
if (ShadeModeT::Mode == (int)ShadeMode::Simple)
{
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, mlight), 8);
}
else if (ShadeModeT::Mode == (int)ShadeMode::Advanced)
{
int blue0 = BPART(ifgcolor0);
int green0 = GPART(ifgcolor0);
int red0 = RPART(ifgcolor0);
int intensity0 = ((red0 * 77 + green0 * 143 + blue0 * 37) >> 8) * desaturate;
int blue1 = BPART(ifgcolor1);
int green1 = GPART(ifgcolor1);
int red1 = RPART(ifgcolor1);
int intensity1 = ((red1 * 77 + green1 * 143 + blue1 * 37) >> 8) * desaturate;
__m128i intensity = _mm_set_epi16(0, intensity1, intensity1, intensity1, 0, intensity0, intensity0, intensity0);
fgcolor = _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(fgcolor, inv_desaturate), intensity), 8);
fgcolor = _mm_mullo_epi16(fgcolor, mlight);
fgcolor = _mm_srli_epi16(_mm_add_epi16(shade_fade, fgcolor), 8);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, shade_light), 8);
}
return AddLights(material, fgcolor, dynlight);
}
template<typename BlendT>
FORCEINLINE __m128i VECTORCALL Blend32(__m128i fgcolor, __m128i bgcolor, unsigned int ifgcolor0, unsigned int ifgcolor1, unsigned int ifgshade0, unsigned int ifgshade1, uint32_t srcalpha, uint32_t destalpha)
{
if (BlendT::Mode == (int)BlendModes::Opaque)
{
__m128i outcolor = fgcolor;
outcolor = _mm_packus_epi16(outcolor, _mm_setzero_si128());
outcolor = _mm_or_si128(outcolor, _mm_set1_epi32(0xff000000));
return outcolor;
}
else if (BlendT::Mode == (int)BlendModes::Masked)
{
__m128i mask = _mm_cmpeq_epi32(_mm_packus_epi16(fgcolor, _mm_setzero_si128()), _mm_setzero_si128());
mask = _mm_unpacklo_epi8(mask, _mm_setzero_si128());
__m128i outcolor = _mm_or_si128(_mm_and_si128(mask, bgcolor), _mm_andnot_si128(mask, fgcolor));
outcolor = _mm_packus_epi16(outcolor, _mm_setzero_si128());
outcolor = _mm_or_si128(outcolor, _mm_set1_epi32(0xff000000));
return outcolor;
}
else if (BlendT::Mode == (int)BlendModes::AddSrcColorOneMinusSrcColor)
{
__m128i inv_srccolor = _mm_sub_epi16(_mm_set1_epi16(256), _mm_add_epi16(fgcolor, _mm_srli_epi16(fgcolor, 7)));
__m128i outcolor = _mm_add_epi16(fgcolor, _mm_srli_epi16(_mm_mullo_epi16(bgcolor, inv_srccolor), 8));
outcolor = _mm_packus_epi16(outcolor, _mm_setzero_si128());
return outcolor;
}
else if (BlendT::Mode == (int)BlendModes::Shaded)
{
ifgshade0 = (ifgshade0 * srcalpha + 128) >> 8;
ifgshade1 = (ifgshade1 * srcalpha + 128) >> 8;
__m128i alpha = _mm_set_epi16(ifgshade1, ifgshade1, ifgshade1, ifgshade1, ifgshade0, ifgshade0, ifgshade0, ifgshade0);
__m128i inv_alpha = _mm_sub_epi16(_mm_set1_epi16(256), alpha);
fgcolor = _mm_mullo_epi16(fgcolor, alpha);
bgcolor = _mm_mullo_epi16(bgcolor, inv_alpha);
__m128i outcolor = _mm_srli_epi16(_mm_add_epi16(fgcolor, bgcolor), 8);
outcolor = _mm_packus_epi16(outcolor, _mm_setzero_si128());
outcolor = _mm_or_si128(outcolor, _mm_set1_epi32(0xff000000));
return outcolor;
}
else if (BlendT::Mode == (int)BlendModes::AddClampShaded)
{
ifgshade0 = (ifgshade0 * srcalpha + 128) >> 8;
ifgshade1 = (ifgshade1 * srcalpha + 128) >> 8;
__m128i alpha = _mm_set_epi16(ifgshade1, ifgshade1, ifgshade1, ifgshade1, ifgshade0, ifgshade0, ifgshade0, ifgshade0);
fgcolor = _mm_srli_epi16(_mm_mullo_epi16(fgcolor, alpha), 8);
__m128i outcolor = _mm_add_epi16(fgcolor, bgcolor);
outcolor = _mm_packus_epi16(outcolor, _mm_setzero_si128());
outcolor = _mm_or_si128(outcolor, _mm_set1_epi32(0xff000000));
return outcolor;
}
else
{
uint32_t alpha0 = APART(ifgcolor0);
uint32_t alpha1 = APART(ifgcolor1);
alpha0 += alpha0 >> 7; // 255->256
alpha1 += alpha1 >> 7; // 255->256
uint32_t inv_alpha0 = 256 - alpha0;
uint32_t inv_alpha1 = 256 - alpha1;
uint32_t bgalpha0 = (destalpha * alpha0 + (inv_alpha0 << 8) + 128) >> 8;
uint32_t bgalpha1 = (destalpha * alpha1 + (inv_alpha1 << 8) + 128) >> 8;
uint32_t fgalpha0 = (srcalpha * alpha0 + 128) >> 8;
uint32_t fgalpha1 = (srcalpha * alpha1 + 128) >> 8;
__m128i bgalpha = _mm_set_epi16(bgalpha1, bgalpha1, bgalpha1, bgalpha1, bgalpha0, bgalpha0, bgalpha0, bgalpha0);
__m128i fgalpha = _mm_set_epi16(fgalpha1, fgalpha1, fgalpha1, fgalpha1, fgalpha0, fgalpha0, fgalpha0, fgalpha0);
fgcolor = _mm_mullo_epi16(fgcolor, fgalpha);
bgcolor = _mm_mullo_epi16(bgcolor, bgalpha);
__m128i fg_lo = _mm_unpacklo_epi16(fgcolor, _mm_setzero_si128());
__m128i bg_lo = _mm_unpacklo_epi16(bgcolor, _mm_setzero_si128());
__m128i fg_hi = _mm_unpackhi_epi16(fgcolor, _mm_setzero_si128());
__m128i bg_hi = _mm_unpackhi_epi16(bgcolor, _mm_setzero_si128());
__m128i out_lo, out_hi;
if (BlendT::Mode == (int)BlendModes::AddClamp)
{
out_lo = _mm_add_epi32(fg_lo, bg_lo);
out_hi = _mm_add_epi32(fg_hi, bg_hi);
}
else if (BlendT::Mode == (int)BlendModes::SubClamp)
{
out_lo = _mm_sub_epi32(fg_lo, bg_lo);
out_hi = _mm_sub_epi32(fg_hi, bg_hi);
}
else if (BlendT::Mode == (int)BlendModes::RevSubClamp)
{
out_lo = _mm_sub_epi32(bg_lo, fg_lo);
out_hi = _mm_sub_epi32(bg_hi, fg_hi);
}
out_lo = _mm_srai_epi32(out_lo, 8);
out_hi = _mm_srai_epi32(out_hi, 8);
__m128i outcolor = _mm_packs_epi32(out_lo, out_hi);
outcolor = _mm_packus_epi16(outcolor, _mm_setzero_si128());
outcolor = _mm_or_si128(outcolor, _mm_set1_epi32(0xff000000));
return outcolor;
}
}
}
template<typename BlendT, typename SamplerT>
class RectScreenDrawer32
{
public:
static void Execute(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
if (args->SimpleShade())
{
Loop<SimpleShade, NearestFilter>(destOrg, destWidth, destHeight, destPitch, args, thread);
}
else
{
Loop<AdvancedShade, NearestFilter>(destOrg, destWidth, destHeight, destPitch, args, thread);
}
}
private:
template<typename ShadeModeT, typename FilterModeT>
FORCEINLINE static void VECTORCALL Loop(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
int x0 = clamp((int)(args->X0() + 0.5f), 0, destWidth);
int x1 = clamp((int)(args->X1() + 0.5f), 0, destWidth);
int y0 = clamp((int)(args->Y0() + 0.5f), 0, destHeight);
int y1 = clamp((int)(args->Y1() + 0.5f), 0, destHeight);
if (x1 <= x0 || y1 <= y0)
return;
uint32_t srcalpha = args->SrcAlpha();
uint32_t destalpha = args->DestAlpha();
// Setup step variables
float fstepU = (args->U1() - args->U0()) / (args->X1() - args->X0());
float fstepV = (args->V1() - args->V0()) / (args->Y1() - args->Y0());
uint32_t startU = (int32_t)((args->U0() + (x0 + 0.5f - args->X0()) * fstepU) * 0x1000000);
uint32_t startV = (int32_t)((args->V0() + (y0 + 0.5f - args->Y0()) * fstepV) * 0x1000000);
uint32_t stepU = (int32_t)(fstepU * 0x1000000);
uint32_t stepV = (int32_t)(fstepV * 0x1000000);
// Sampling stuff
uint32_t color = args->Color();
const uint32_t * RESTRICT translation = (const uint32_t *)args->Translation();
const uint32_t * RESTRICT texPixels = (const uint32_t *)args->TexturePixels();
uint32_t texWidth = args->TextureWidth();
uint32_t texHeight = args->TextureHeight();
uint32_t oneU, oneV;
if (SamplerT::Mode != (int)Samplers::Fill)
{
oneU = ((0x800000 + texWidth - 1) / texWidth) * 2 + 1;
oneV = ((0x800000 + texHeight - 1) / texHeight) * 2 + 1;
}
else
{
oneU = 0;
oneV = 0;
}
// Shade constants
__m128i inv_desaturate, shade_fade, shade_light;
int desaturate;
if (ShadeModeT::Mode == (int)ShadeMode::Advanced)
{
inv_desaturate = _mm_setr_epi16(256, 256 - args->ShadeDesaturate(), 256 - args->ShadeDesaturate(), 256 - args->ShadeDesaturate(), 256, 256 - args->ShadeDesaturate(), 256 - args->ShadeDesaturate(), 256 - args->ShadeDesaturate());
shade_fade = _mm_set_epi16(args->ShadeFadeAlpha(), args->ShadeFadeRed(), args->ShadeFadeGreen(), args->ShadeFadeBlue(), args->ShadeFadeAlpha(), args->ShadeFadeRed(), args->ShadeFadeGreen(), args->ShadeFadeBlue());
shade_light = _mm_set_epi16(args->ShadeLightAlpha(), args->ShadeLightRed(), args->ShadeLightGreen(), args->ShadeLightBlue(), args->ShadeLightAlpha(), args->ShadeLightRed(), args->ShadeLightGreen(), args->ShadeLightBlue());
desaturate = args->ShadeDesaturate();
}
else
{
inv_desaturate = _mm_setzero_si128();
shade_fade = _mm_setzero_si128();
shade_light = _mm_setzero_si128();
desaturate = 0;
}
// Setup light
uint32_t lightpos = args->Light();
lightpos += lightpos >> 7; // 255 -> 256
__m128i mlight = _mm_set_epi16(256, lightpos, lightpos, lightpos, 256, lightpos, lightpos, lightpos);
__m128i shade_fade_lit;
if (ShadeModeT::Mode == (int)ShadeMode::Advanced)
{
__m128i inv_light = _mm_sub_epi16(_mm_set_epi16(0, 256, 256, 256, 0, 256, 256, 256), mlight);
shade_fade_lit = _mm_mullo_epi16(shade_fade, inv_light);
}
else
{
shade_fade_lit = _mm_setzero_si128();
}
int count = x1 - x0;
int sseCount = count / 2;
uint32_t posV = startV;
for (int y = y0; y < y1; y++, posV += stepV)
{
int coreBlock = y / 8;
if (coreBlock % thread->num_cores != thread->core)
{
continue;
}
uint32_t *dest = ((uint32_t*)destOrg) + y * destPitch + x0;
uint32_t posU = startU;
for (int i = 0; i < sseCount; i++)
{
// Load bgcolor
__m128i bgcolor;
if (BlendT::Mode != (int)BlendModes::Opaque)
bgcolor = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)dest), _mm_setzero_si128());
else
bgcolor = _mm_setzero_si128();
// Sample fgcolor
unsigned int ifgcolor[2], ifgshade[2];
if (SamplerT::Mode == (int)Samplers::FogBoundary) color = dest[0];
ifgcolor[0] = Sample32<SamplerT, FilterModeT>(posU, posV, texPixels, texWidth, texHeight, oneU, oneV, color, translation);
ifgshade[0] = SampleShade32<SamplerT>(posU, posV, texPixels, texWidth, texHeight, x0 + i * 2, y);
posU += stepU;
if (SamplerT::Mode == (int)Samplers::FogBoundary) color = dest[1];
ifgcolor[1] = Sample32<SamplerT, FilterModeT>(posU, posV, texPixels, texWidth, texHeight, oneU, oneV, color, translation);
ifgshade[1] = SampleShade32<SamplerT>(posU, posV, texPixels, texWidth, texHeight, x0 + i * 2 + 1, y);
posU += stepU;
// Shade and blend
__m128i fgcolor = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)ifgcolor), _mm_setzero_si128());
fgcolor = Shade32<ShadeModeT>(fgcolor, mlight, ifgcolor[0], ifgcolor[1], desaturate, inv_desaturate, shade_fade_lit, shade_light, _mm_setzero_si128());
__m128i outcolor = Blend32<BlendT>(fgcolor, bgcolor, ifgcolor[0], ifgcolor[1], ifgshade[0], ifgshade[1], srcalpha, destalpha);
// Store result
_mm_storel_epi64((__m128i*)dest, outcolor);
dest += 2;
}
if (sseCount * 2 != count)
{
// Load bgcolor
__m128i bgcolor;
if (BlendT::Mode != (int)BlendModes::Opaque)
bgcolor = _mm_unpacklo_epi8(_mm_cvtsi32_si128(*dest), _mm_setzero_si128());
else
bgcolor = _mm_setzero_si128();
// Sample fgcolor
unsigned int ifgcolor[2], ifgshade[2];
if (SamplerT::Mode == (int)Samplers::FogBoundary) color = *dest;
ifgcolor[0] = Sample32<SamplerT, FilterModeT>(posU, posV, texPixels, texWidth, texHeight, oneU, oneV, color, translation);
ifgshade[0] = SampleShade32<SamplerT>(posU, posV, texPixels, texWidth, texHeight, x0 + sseCount * 2, y);
ifgcolor[1] = 0;
ifgshade[1] = 0;
posU += stepU;
// Shade and blend
__m128i fgcolor = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i*)ifgcolor), _mm_setzero_si128());
fgcolor = Shade32<ShadeModeT>(fgcolor, mlight, ifgcolor[0], ifgcolor[1], desaturate, inv_desaturate, shade_fade_lit, shade_light, _mm_setzero_si128());
__m128i outcolor = Blend32<BlendT>(fgcolor, bgcolor, ifgcolor[0], ifgcolor[1], ifgshade[0], ifgshade[1], srcalpha, destalpha);
// Store result
*dest = _mm_cvtsi128_si32(outcolor);
}
}
}
};

295
src/polyrenderer/drawers/poly_drawer8.h
View file

@ -1,295 +0,0 @@
/*
** Polygon Doom software renderer
** Copyright (c) 2016 Magnus Norddahl
**
** This software is provided 'as-is', without any express or implied
** warranty. In no event will the authors be held liable for any damages
** arising from the use of this software.
**
** Permission is granted to anyone to use this software for any purpose,
** including commercial applications, and to alter it and redistribute it
** freely, subject to the following restrictions:
**
** 1. The origin of this software must not be misrepresented; you must not
** claim that you wrote the original software. If you use this software
** in a product, an acknowledgment in the product documentation would be
** appreciated but is not required.
** 2. Altered source versions must be plainly marked as such, and must not be
** misrepresented as being the original software.
** 3. This notice may not be removed or altered from any source distribution.
**
*/
#pragma once
#include "screen_triangle.h"
namespace TriScreenDrawerModes
{
template<typename SamplerT>
FORCEINLINE unsigned int Sample8(int32_t u, int32_t v, const uint8_t *texPixels, int texWidth, int texHeight, uint32_t color, const uint8_t *translation)
{
uint8_t texel;
if (SamplerT::Mode == (int)Samplers::Shaded || SamplerT::Mode == (int)Samplers::Stencil || SamplerT::Mode == (int)Samplers::Fill || SamplerT::Mode == (int)Samplers::Fuzz || SamplerT::Mode == (int)Samplers::FogBoundary)
{
return color;
}
else if (SamplerT::Mode == (int)Samplers::Translated)
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
return translation[texPixels[texelX * texHeight + texelY]];
}
else
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
texel = texPixels[texelX * texHeight + texelY];
}
if (SamplerT::Mode == (int)Samplers::Skycap)
{
int start_fade = 2; // How fast it should fade out
int alpha_top = clamp(v >> (16 - start_fade), 0, 256);
int alpha_bottom = clamp(((2 << 24) - v) >> (16 - start_fade), 0, 256);
int a = MIN(alpha_top, alpha_bottom);
int inv_a = 256 - a;
if (a == 256)
return texel;
uint32_t capcolor = GPalette.BaseColors[color].d;
uint32_t texelrgb = GPalette.BaseColors[texel].d;
uint32_t r = RPART(texelrgb);
uint32_t g = GPART(texelrgb);
uint32_t b = BPART(texelrgb);
uint32_t capcolor_red = RPART(capcolor);
uint32_t capcolor_green = GPART(capcolor);
uint32_t capcolor_blue = BPART(capcolor);
r = (r * a + capcolor_red * inv_a + 127) >> 8;
g = (g * a + capcolor_green * inv_a + 127) >> 8;
b = (b * a + capcolor_blue * inv_a + 127) >> 8;
return RGB256k.All[((r >> 2) << 12) | ((g >> 2) << 6) | (b >> 2)];
}
else
{
return texel;
}
}
template<typename SamplerT>
FORCEINLINE unsigned int SampleShade8(int32_t u, int32_t v, const uint8_t *texPixels, int texWidth, int texHeight, int x, int y)
{
if (SamplerT::Mode == (int)Samplers::Shaded)
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = texPixels[texelX * texHeight + texelY];
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
return sampleshadeout;
}
else if (SamplerT::Mode == (int)Samplers::Stencil)
{
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
return texPixels[texelX * texHeight + texelY] != 0 ? 256 : 0;
}
else if (SamplerT::Mode == (int)Samplers::Fuzz)
{
using namespace swrenderer;
uint32_t texelX = ((((uint32_t)u << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = ((((uint32_t)v << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = (texPixels[texelX * texHeight + texelY] != 0) ? 256 : 0;
fixed_t fuzzscale = (200 << FRACBITS) / viewheight;
int scaled_x = (x * fuzzscale) >> FRACBITS;
int fuzz_x = fuzz_random_x_offset[scaled_x % FUZZ_RANDOM_X_SIZE] + fuzzpos;
fixed_t fuzzcount = FUZZTABLE << FRACBITS;
fixed_t fuzz = ((fuzz_x << FRACBITS) + y * fuzzscale) % fuzzcount;
unsigned int alpha = fuzzoffset[fuzz >> FRACBITS];
sampleshadeout = (sampleshadeout * alpha) >> 5;
return sampleshadeout;
}
else
{
return 0;
}
}
template<typename BlendT>
FORCEINLINE uint8_t ShadeAndBlend8(uint8_t fgcolor, uint8_t bgcolor, uint32_t fgshade, uint32_t lightshade, const uint8_t *colormaps, uint32_t srcalpha, uint32_t destalpha)
{
lightshade = ((256 - lightshade) * NUMCOLORMAPS) & 0xffffff00;
uint8_t shadedfg = colormaps[lightshade + fgcolor];
if (BlendT::Mode == (int)BlendModes::Opaque)
{
return shadedfg;
}
else if (BlendT::Mode == (int)BlendModes::Masked)
{
return (fgcolor != 0) ? shadedfg : bgcolor;
}
else if (BlendT::Mode == (int)BlendModes::AddSrcColorOneMinusSrcColor)
{
int32_t fg_r = GPalette.BaseColors[shadedfg].r;
int32_t fg_g = GPalette.BaseColors[shadedfg].g;
int32_t fg_b = GPalette.BaseColors[shadedfg].b;
int32_t bg_r = GPalette.BaseColors[bgcolor].r;
int32_t bg_g = GPalette.BaseColors[bgcolor].g;
int32_t bg_b = GPalette.BaseColors[bgcolor].b;
int32_t inv_fg_r = 256 - (fg_r + (fg_r >> 7));
int32_t inv_fg_g = 256 - (fg_g + (fg_g >> 7));
int32_t inv_fg_b = 256 - (fg_b + (fg_b >> 7));
fg_r = MIN<int32_t>(fg_r + ((bg_r * inv_fg_r + 127) >> 8), 255);
fg_g = MIN<int32_t>(fg_g + ((bg_g * inv_fg_g + 127) >> 8), 255);
fg_b = MIN<int32_t>(fg_b + ((bg_b * inv_fg_b + 127) >> 8), 255);
shadedfg = RGB256k.All[((fg_r >> 2) << 12) | ((fg_g >> 2) << 6) | (fg_b >> 2)];
return (fgcolor != 0) ? shadedfg : bgcolor;
}
else if (BlendT::Mode == (int)BlendModes::Shaded)
{
fgshade = (fgshade * srcalpha + 128) >> 8;
uint32_t alpha = fgshade;
uint32_t inv_alpha = 256 - fgshade;
int32_t fg_r = GPalette.BaseColors[shadedfg].r;
int32_t fg_g = GPalette.BaseColors[shadedfg].g;
int32_t fg_b = GPalette.BaseColors[shadedfg].b;
int32_t bg_r = GPalette.BaseColors[bgcolor].r;
int32_t bg_g = GPalette.BaseColors[bgcolor].g;
int32_t bg_b = GPalette.BaseColors[bgcolor].b;
fg_r = (fg_r * alpha + bg_r * inv_alpha + 127) >> 8;
fg_g = (fg_g * alpha + bg_g * inv_alpha + 127) >> 8;
fg_b = (fg_b * alpha + bg_b * inv_alpha + 127) >> 8;
shadedfg = RGB256k.All[((fg_r >> 2) << 12) | ((fg_g >> 2) << 6) | (fg_b >> 2)];
return (alpha != 0) ? shadedfg : bgcolor;
}
else if (BlendT::Mode == (int)BlendModes::AddClampShaded)
{
fgshade = (fgshade * srcalpha + 128) >> 8;
uint32_t alpha = fgshade;
int32_t fg_r = GPalette.BaseColors[shadedfg].r;
int32_t fg_g = GPalette.BaseColors[shadedfg].g;
int32_t fg_b = GPalette.BaseColors[shadedfg].b;
int32_t bg_r = GPalette.BaseColors[bgcolor].r;
int32_t bg_g = GPalette.BaseColors[bgcolor].g;
int32_t bg_b = GPalette.BaseColors[bgcolor].b;
fg_r = MIN<int32_t>(bg_r + ((fg_r * alpha + 127) >> 8), 255);
fg_g = MIN<int32_t>(bg_g + ((fg_g * alpha + 127) >> 8), 255);
fg_b = MIN<int32_t>(bg_b + ((fg_b * alpha + 127) >> 8), 255);
shadedfg = RGB256k.All[((fg_r >> 2) << 12) | ((fg_g >> 2) << 6) | (fg_b >> 2)];
return (alpha != 0) ? shadedfg : bgcolor;
}
else
{
int32_t fg_r = GPalette.BaseColors[shadedfg].r;
int32_t fg_g = GPalette.BaseColors[shadedfg].g;
int32_t fg_b = GPalette.BaseColors[shadedfg].b;
int32_t bg_r = GPalette.BaseColors[bgcolor].r;
int32_t bg_g = GPalette.BaseColors[bgcolor].g;
int32_t bg_b = GPalette.BaseColors[bgcolor].b;
if (BlendT::Mode == (int)BlendModes::AddClamp)
{
fg_r = MIN(int32_t(fg_r * srcalpha + bg_r * destalpha + 127) >> 8, 255);
fg_g = MIN(int32_t(fg_g * srcalpha + bg_g * destalpha + 127) >> 8, 255);
fg_b = MIN(int32_t(fg_b * srcalpha + bg_b * destalpha + 127) >> 8, 255);
}
else if (BlendT::Mode == (int)BlendModes::SubClamp)
{
fg_r = MAX(int32_t(fg_r * srcalpha - bg_r * destalpha + 127) >> 8, 0);
fg_g = MAX(int32_t(fg_g * srcalpha - bg_g * destalpha + 127) >> 8, 0);
fg_b = MAX(int32_t(fg_b * srcalpha - bg_b * destalpha + 127) >> 8, 0);
}
else if (BlendT::Mode == (int)BlendModes::RevSubClamp)
{
fg_r = MAX(int32_t(bg_r * srcalpha - fg_r * destalpha + 127) >> 8, 0);
fg_g = MAX(int32_t(bg_g * srcalpha - fg_g * destalpha + 127) >> 8, 0);
fg_b = MAX(int32_t(bg_b * srcalpha - fg_b * destalpha + 127) >> 8, 0);
}
shadedfg = RGB256k.All[((fg_r >> 2) << 12) | ((fg_g >> 2) << 6) | (fg_b >> 2)];
return (fgcolor != 0) ? shadedfg : bgcolor;
}
}
}
template<typename BlendT, typename SamplerT>
class RectScreenDrawer8
{
public:
static void Execute(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
int x0 = clamp((int)(args->X0() + 0.5f), 0, destWidth);
int x1 = clamp((int)(args->X1() + 0.5f), 0, destWidth);
int y0 = clamp((int)(args->Y0() + 0.5f), 0, destHeight);
int y1 = clamp((int)(args->Y1() + 0.5f), 0, destHeight);
if (x1 <= x0 || y1 <= y0)
return;
auto colormaps = args->BaseColormap();
uint32_t srcalpha = args->SrcAlpha();
uint32_t destalpha = args->DestAlpha();
// Setup step variables
float fstepU = (args->U1() - args->U0()) / (args->X1() - args->X0());
float fstepV = (args->V1() - args->V0()) / (args->Y1() - args->Y0());
uint32_t startU = (int32_t)((args->U0() + (x0 + 0.5f - args->X0()) * fstepU) * 0x1000000);
uint32_t startV = (int32_t)((args->V0() + (y0 + 0.5f - args->Y0()) * fstepV) * 0x1000000);
uint32_t stepU = (int32_t)(fstepU * 0x1000000);
uint32_t stepV = (int32_t)(fstepV * 0x1000000);
// Sampling stuff
uint32_t color = args->Color();
const uint8_t * RESTRICT translation = args->Translation();
const uint8_t * RESTRICT texPixels = args->TexturePixels();
uint32_t texWidth = args->TextureWidth();
uint32_t texHeight = args->TextureHeight();
// Setup light
uint32_t lightshade = args->Light();
lightshade += lightshade >> 7; // 255 -> 256
if (SamplerT::Mode == (int)Samplers::Fuzz) lightshade = 256;
int count = x1 - x0;
uint32_t posV = startV;
for (int y = y0; y < y1; y++, posV += stepV)
{
int coreBlock = y / 8;
if (coreBlock % thread->num_cores != thread->core)
{
continue;
}
uint8_t *dest = ((uint8_t*)destOrg) + y * destPitch + x0;
uint32_t posU = startU;
for (int i = 0; i < count; i++)
{
uint8_t bgcolor = *dest;
if (SamplerT::Mode == (int)Samplers::FogBoundary) color = bgcolor;
uint8_t fgcolor = Sample8<SamplerT>(posU, posV, texPixels, texWidth, texHeight, color, translation);
uint32_t fgshade = SampleShade8<SamplerT>(posU, posV, texPixels, texWidth, texHeight, x0 + i, y);
*dest = ShadeAndBlend8<BlendT>(fgcolor, bgcolor, fgshade, lightshade, colormaps, srcalpha, destalpha);
posU += stepU;
dest++;
}
}
}
};

620
src/polyrenderer/drawers/screen_triangle.cpp
View file

@ -36,12 +36,6 @@
#include "poly_triangle.h" #include "poly_triangle.h"
#include "swrenderer/drawers/r_draw_rgba.h" #include "swrenderer/drawers/r_draw_rgba.h"
#include "screen_triangle.h" #include "screen_triangle.h"
#ifndef NO_SSE
#include "poly_drawer32_sse2.h"
#else
#include "poly_drawer32.h"
#endif
#include "poly_drawer8.h"
#include "x86.h" #include "x86.h"
static void SortVertices(const TriDrawTriangleArgs *args, ShadedTriVertex **sortedVertices) static void SortVertices(const TriDrawTriangleArgs *args, ShadedTriVertex **sortedVertices)
@ -1284,6 +1278,516 @@ void DrawSpan8(int y, int x0, int x1, const TriDrawTriangleArgs *args)
} }
} }
template<typename ModeT>
void DrawRect8(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
int x0 = clamp((int)(args->X0() + 0.5f), 0, destWidth);
int x1 = clamp((int)(args->X1() + 0.5f), 0, destWidth);
int y0 = clamp((int)(args->Y0() + 0.5f), 0, destHeight);
int y1 = clamp((int)(args->Y1() + 0.5f), 0, destHeight);
if (x1 <= x0 || y1 <= y0)
return;
const uint8_t *colormaps, *texPixels, *translation;
int texWidth, texHeight;
uint32_t fillcolor;
int alpha;
uint32_t light;
texPixels = args->TexturePixels();
translation = args->Translation();
texWidth = args->TextureWidth();
texHeight = args->TextureHeight();
fillcolor = args->Color();
alpha = args->Alpha();
colormaps = args->BaseColormap();
light = args->Light();
light += light >> 7; // 255 -> 256
light = ((256 - light) * NUMCOLORMAPS) & 0xffffff00;
float fstepU = (args->U1() - args->U0()) / (args->X1() - args->X0());
float fstepV = (args->V1() - args->V0()) / (args->Y1() - args->Y0());
uint32_t startU = (int32_t)((args->U0() + (x0 + 0.5f - args->X0()) * fstepU) * 0x1000000);
uint32_t startV = (int32_t)((args->V0() + (y0 + 0.5f - args->Y0()) * fstepV) * 0x1000000);
uint32_t stepU = (int32_t)(fstepU * 0x1000000);
uint32_t stepV = (int32_t)(fstepV * 0x1000000);
uint32_t posV = startV;
int num_cores = thread->num_cores;
int skip = thread->skipped_by_thread(y0);
posV += skip * stepV;
stepV *= num_cores;
for (int y = y0 + skip; y < y1; y += num_cores, posV += stepV)
{
uint8_t *destLine = ((uint8_t*)destOrg) + y * destPitch;
uint32_t posU = startU;
for (int x = x0; x < x1; x++)
{
int fg = 0;
int fgalpha = 255;
if (ModeT::SWFlags & SWSTYLEF_Fill)
{
fg = fillcolor;
}
else if (ModeT::BlendOp != STYLEOP_Fuzz)
{
uint32_t texelX = (((posU << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = (((posV << 8) >> 16) * texHeight) >> 16;
fg = texPixels[texelX * texHeight + texelY];
if (ModeT::SWFlags & SWSTYLEF_Translated)
fg = translation[fg];
fgalpha = (fg != 0) ? 255 : 0;
}
if (ModeT::BlendOp == STYLEOP_Fuzz)
{
using namespace swrenderer;
uint32_t texelX = (((posU << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = (((posV << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = (texPixels[texelX * texHeight + texelY] != 0) ? 256 : 0;
fixed_t fuzzscale = (200 << FRACBITS) / viewheight;
int scaled_x = (x * fuzzscale) >> FRACBITS;
int fuzz_x = fuzz_random_x_offset[scaled_x % FUZZ_RANDOM_X_SIZE] + fuzzpos;
fixed_t fuzzcount = FUZZTABLE << FRACBITS;
fixed_t fuzz = ((fuzz_x << FRACBITS) + y * fuzzscale) % fuzzcount;
unsigned int alpha = fuzzoffset[fuzz >> FRACBITS];
sampleshadeout = (sampleshadeout * alpha) >> 5;
uint32_t a = 256 - sampleshadeout;
uint32_t dest = GPalette.BaseColors[destLine[x]].d;
uint32_t r = (RPART(dest) * a) >> 8;
uint32_t g = (GPART(dest) * a) >> 8;
uint32_t b = (BPART(dest) * a) >> 8;
destLine[x] = RGB256k.All[((r >> 2) << 12) | ((g >> 2) << 6) | (b >> 2)];
}
else
{
if ((ModeT::Flags & STYLEF_ColorIsFixed) && !(ModeT::SWFlags & SWSTYLEF_Fill))
{
if (ModeT::Flags & STYLEF_RedIsAlpha)
fgalpha = fg;
fg = fillcolor;
}
if (!(ModeT::Flags & STYLEF_Alpha1))
{
fgalpha = (fgalpha * alpha) >> 8;
}
uint8_t shadedfg = colormaps[light + fg];
if (ModeT::BlendSrc == STYLEALPHA_One && ModeT::BlendDest == STYLEALPHA_Zero)
{
destLine[x] = shadedfg;
}
else if (ModeT::BlendSrc == STYLEALPHA_One && ModeT::BlendDest == STYLEALPHA_One)
{
uint32_t src = GPalette.BaseColors[shadedfg];
uint32_t dest = GPalette.BaseColors[destLine[x]];
if (ModeT::BlendOp == STYLEOP_Add)
{
uint32_t out_r = MIN<uint32_t>(RPART(dest) + RPART(src), 255);
uint32_t out_g = MIN<uint32_t>(GPART(dest) + GPART(src), 255);
uint32_t out_b = MIN<uint32_t>(BPART(dest) + BPART(src), 255);
destLine[x] = RGB256k.All[((out_r >> 2) << 12) | ((out_g >> 2) << 6) | (out_b >> 2)];
}
else if (ModeT::BlendOp == STYLEOP_RevSub)
{
uint32_t out_r = MAX<uint32_t>(RPART(dest) - RPART(src), 0);
uint32_t out_g = MAX<uint32_t>(GPART(dest) - GPART(src), 0);
uint32_t out_b = MAX<uint32_t>(BPART(dest) - BPART(src), 0);
destLine[x] = RGB256k.All[((out_r >> 2) << 12) | ((out_g >> 2) << 6) | (out_b >> 2)];
}
else //if (ModeT::BlendOp == STYLEOP_Sub)
{
uint32_t out_r = MAX<uint32_t>(RPART(src) - RPART(dest), 0);
uint32_t out_g = MAX<uint32_t>(GPART(src) - GPART(dest), 0);
uint32_t out_b = MAX<uint32_t>(BPART(src) - BPART(dest), 0);
destLine[x] = RGB256k.All[((out_r >> 2) << 12) | ((out_g >> 2) << 6) | (out_b >> 2)];
}
}
else if (ModeT::SWFlags & SWSTYLEF_SrcColorOneMinusSrcColor)
{
uint32_t src = GPalette.BaseColors[shadedfg];
uint32_t dest = GPalette.BaseColors[destLine[x]];
uint32_t sfactor_r = RPART(src); sfactor_r += sfactor_r >> 7; // 255 -> 256
uint32_t sfactor_g = GPART(src); sfactor_g += sfactor_g >> 7; // 255 -> 256
uint32_t sfactor_b = BPART(src); sfactor_b += sfactor_b >> 7; // 255 -> 256
uint32_t sfactor_a = fgalpha; sfactor_a += sfactor_a >> 7; // 255 -> 256
uint32_t dfactor_r = 256 - sfactor_r;
uint32_t dfactor_g = 256 - sfactor_g;
uint32_t dfactor_b = 256 - sfactor_b;
uint32_t out_r = (RPART(dest) * dfactor_r + RPART(src) * sfactor_r + 128) >> 8;
uint32_t out_g = (GPART(dest) * dfactor_g + GPART(src) * sfactor_g + 128) >> 8;
uint32_t out_b = (BPART(dest) * dfactor_b + BPART(src) * sfactor_b + 128) >> 8;
destLine[x] = RGB256k.All[((out_r >> 2) << 12) | ((out_g >> 2) << 6) | (out_b >> 2)];
}
else if (ModeT::BlendSrc == STYLEALPHA_Src && ModeT::BlendDest == STYLEALPHA_InvSrc && fgalpha == 255)
{
destLine[x] = shadedfg;
}
else if (ModeT::BlendSrc != STYLEALPHA_Src || ModeT::BlendDest != STYLEALPHA_InvSrc || fgalpha != 0)
{
uint32_t src = GPalette.BaseColors[shadedfg];
uint32_t dest = GPalette.BaseColors[destLine[x]];
uint32_t sfactor = fgalpha; sfactor += sfactor >> 7; // 255 -> 256
uint32_t dfactor = 256 - sfactor;
uint32_t src_r = RPART(src) * sfactor;
uint32_t src_g = GPART(src) * sfactor;
uint32_t src_b = BPART(src) * sfactor;
uint32_t dest_r = RPART(dest);
uint32_t dest_g = GPART(dest);
uint32_t dest_b = BPART(dest);
if (ModeT::BlendDest == STYLEALPHA_One)
{
dest_r <<= 8;
dest_g <<= 8;
dest_b <<= 8;
}
else
{
uint32_t dfactor = 256 - sfactor;
dest_r *= dfactor;
dest_g *= dfactor;
dest_b *= dfactor;
}
uint32_t out_r, out_g, out_b;
if (ModeT::BlendOp == STYLEOP_Add)
{
if (ModeT::BlendDest == STYLEALPHA_One)
{
out_r = MIN<int32_t>((dest_r + src_r + 128) >> 8, 255);
out_g = MIN<int32_t>((dest_g + src_g + 128) >> 8, 255);
out_b = MIN<int32_t>((dest_b + src_b + 128) >> 8, 255);
}
else
{
out_r = (dest_r + src_r + 128) >> 8;
out_g = (dest_g + src_g + 128) >> 8;
out_b = (dest_b + src_b + 128) >> 8;
}
}
else if (ModeT::BlendOp == STYLEOP_RevSub)
{
out_r = MAX<int32_t>(static_cast<int32_t>(dest_r - src_r + 128) >> 8, 0);
out_g = MAX<int32_t>(static_cast<int32_t>(dest_g - src_g + 128) >> 8, 0);
out_b = MAX<int32_t>(static_cast<int32_t>(dest_b - src_b + 128) >> 8, 0);
}
else //if (ModeT::BlendOp == STYLEOP_Sub)
{
out_r = MAX<int32_t>(static_cast<int32_t>(src_r - dest_r + 128) >> 8, 0);
out_g = MAX<int32_t>(static_cast<int32_t>(src_g - dest_g + 128) >> 8, 0);
out_b = MAX<int32_t>(static_cast<int32_t>(src_b - dest_b + 128) >> 8, 0);
}
destLine[x] = RGB256k.All[((out_r >> 2) << 12) | ((out_g >> 2) << 6) | (out_b >> 2)];
}
}
posU += stepU;
}
}
}
template<typename ModeT, typename OptT>
void DrawRectOpt32(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
int x0 = clamp((int)(args->X0() + 0.5f), 0, destWidth);
int x1 = clamp((int)(args->X1() + 0.5f), 0, destWidth);
int y0 = clamp((int)(args->Y0() + 0.5f), 0, destHeight);
int y1 = clamp((int)(args->Y1() + 0.5f), 0, destHeight);
if (x1 <= x0 || y1 <= y0)
return;
const uint32_t *texPixels, *translation;
int texWidth, texHeight;
uint32_t fillcolor;
int alpha;
uint32_t light;
uint32_t shade_fade_r, shade_fade_g, shade_fade_b, shade_light_r, shade_light_g, shade_light_b, desaturate, inv_desaturate;
texPixels = (const uint32_t*)args->TexturePixels();
translation = (const uint32_t*)args->Translation();
texWidth = args->TextureWidth();
texHeight = args->TextureHeight();
fillcolor = args->Color();
alpha = args->Alpha();
light = args->Light();
light += light >> 7; // 255 -> 256
if (OptT::Flags & SWOPT_ColoredFog)
{
shade_fade_r = args->ShadeFadeRed();
shade_fade_g = args->ShadeFadeGreen();
shade_fade_b = args->ShadeFadeBlue();
shade_light_r = args->ShadeLightRed();
shade_light_g = args->ShadeLightGreen();
shade_light_b = args->ShadeLightBlue();
desaturate = args->ShadeDesaturate();
inv_desaturate = 256 - desaturate;
}
float fstepU = (args->U1() - args->U0()) / (args->X1() - args->X0());
float fstepV = (args->V1() - args->V0()) / (args->Y1() - args->Y0());
uint32_t startU = (int32_t)((args->U0() + (x0 + 0.5f - args->X0()) * fstepU) * 0x1000000);
uint32_t startV = (int32_t)((args->V0() + (y0 + 0.5f - args->Y0()) * fstepV) * 0x1000000);
uint32_t stepU = (int32_t)(fstepU * 0x1000000);
uint32_t stepV = (int32_t)(fstepV * 0x1000000);
uint32_t posV = startV;
int num_cores = thread->num_cores;
int skip = thread->skipped_by_thread(y0);
posV += skip * stepV;
stepV *= num_cores;
for (int y = y0 + skip; y < y1; y += num_cores, posV += stepV)
{
uint32_t *destLine = ((uint32_t*)destOrg) + y * destPitch;
uint32_t posU = startU;
for (int x = x0; x < x1; x++)
{
uint32_t fg = 0;
if (ModeT::SWFlags & SWSTYLEF_Fill)
{
fg = fillcolor;
}
else if (ModeT::SWFlags & SWSTYLEF_FogBoundary)
{
fg = destLine[x];
}
else if (ModeT::BlendOp != STYLEOP_Fuzz)
{
uint32_t texelX = (((posU << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = (((posV << 8) >> 16) * texHeight) >> 16;
if (ModeT::SWFlags & SWSTYLEF_Translated)
{
fg = translation[((const uint8_t*)texPixels)[texelX * texHeight + texelY]];
}
else if (ModeT::Flags & STYLEF_RedIsAlpha)
{
fg = ((const uint8_t*)texPixels)[texelX * texHeight + texelY];
}
else
{
fg = texPixels[texelX * texHeight + texelY];
}
}
if (ModeT::BlendOp == STYLEOP_Fuzz)
{
using namespace swrenderer;
uint32_t texelX = (((posU << 8) >> 16) * texWidth) >> 16;
uint32_t texelY = (((posV << 8) >> 16) * texHeight) >> 16;
unsigned int sampleshadeout = APART(texPixels[texelX * texHeight + texelY]);
sampleshadeout += sampleshadeout >> 7; // 255 -> 256
fixed_t fuzzscale = (200 << FRACBITS) / viewheight;
int scaled_x = (x * fuzzscale) >> FRACBITS;
int fuzz_x = fuzz_random_x_offset[scaled_x % FUZZ_RANDOM_X_SIZE] + fuzzpos;
fixed_t fuzzcount = FUZZTABLE << FRACBITS;
fixed_t fuzz = ((fuzz_x << FRACBITS) + y * fuzzscale) % fuzzcount;
unsigned int alpha = fuzzoffset[fuzz >> FRACBITS];
sampleshadeout = (sampleshadeout * alpha) >> 5;
uint32_t a = 256 - sampleshadeout;
uint32_t dest = destLine[x];
uint32_t out_r = (RPART(dest) * a) >> 8;
uint32_t out_g = (GPART(dest) * a) >> 8;
uint32_t out_b = (BPART(dest) * a) >> 8;
destLine[x] = MAKEARGB(255, out_r, out_g, out_b);
}
else
{
if ((ModeT::Flags & STYLEF_ColorIsFixed) && !(ModeT::SWFlags & SWSTYLEF_Fill))
{
if (ModeT::Flags & STYLEF_RedIsAlpha)
fg = (fg << 24) | (fillcolor & 0x00ffffff);
else
fg = (fg & 0xff000000) | (fillcolor & 0x00ffffff);
}
uint32_t fgalpha = fg >> 24;
if (!(ModeT::Flags & STYLEF_Alpha1))
{
fgalpha = (fgalpha * alpha) >> 8;
}
int lightshade = light;
uint32_t lit_r = 0, lit_g = 0, lit_b = 0;
uint32_t shadedfg_r, shadedfg_g, shadedfg_b;
if (OptT::Flags & SWOPT_ColoredFog)
{
uint32_t fg_r = RPART(fg);
uint32_t fg_g = GPART(fg);
uint32_t fg_b = BPART(fg);
uint32_t intensity = ((fg_r * 77 + fg_g * 143 + fg_b * 37) >> 8) * desaturate;
shadedfg_r = (((shade_fade_r + ((fg_r * inv_desaturate + intensity) >> 8) * lightshade) >> 8) * shade_light_r) >> 8;
shadedfg_g = (((shade_fade_g + ((fg_g * inv_desaturate + intensity) >> 8) * lightshade) >> 8) * shade_light_g) >> 8;
shadedfg_b = (((shade_fade_b + ((fg_b * inv_desaturate + intensity) >> 8) * lightshade) >> 8) * shade_light_b) >> 8;
}
else
{
shadedfg_r = (RPART(fg) * lightshade) >> 8;
shadedfg_g = (GPART(fg) * lightshade) >> 8;
shadedfg_b = (BPART(fg) * lightshade) >> 8;
}
if (ModeT::BlendSrc == STYLEALPHA_One && ModeT::BlendDest == STYLEALPHA_Zero)
{
destLine[x] = MAKEARGB(255, shadedfg_r, shadedfg_g, shadedfg_b);
}
else if (ModeT::BlendSrc == STYLEALPHA_One && ModeT::BlendDest == STYLEALPHA_One)
{
uint32_t dest = destLine[x];
if (ModeT::BlendOp == STYLEOP_Add)
{
uint32_t out_r = MIN<uint32_t>(RPART(dest) + shadedfg_r, 255);
uint32_t out_g = MIN<uint32_t>(GPART(dest) + shadedfg_g, 255);
uint32_t out_b = MIN<uint32_t>(BPART(dest) + shadedfg_b, 255);
destLine[x] = MAKEARGB(255, out_r, out_g, out_b);
}
else if (ModeT::BlendOp == STYLEOP_RevSub)
{
uint32_t out_r = MAX<uint32_t>(RPART(dest) - shadedfg_r, 0);
uint32_t out_g = MAX<uint32_t>(GPART(dest) - shadedfg_g, 0);
uint32_t out_b = MAX<uint32_t>(BPART(dest) - shadedfg_b, 0);
destLine[x] = MAKEARGB(255, out_r, out_g, out_b);
}
else //if (ModeT::BlendOp == STYLEOP_Sub)
{
uint32_t out_r = MAX<uint32_t>(shadedfg_r - RPART(dest), 0);
uint32_t out_g = MAX<uint32_t>(shadedfg_g - GPART(dest), 0);
uint32_t out_b = MAX<uint32_t>(shadedfg_b - BPART(dest), 0);
destLine[x] = MAKEARGB(255, out_r, out_g, out_b);
}
}
else if (ModeT::SWFlags & SWSTYLEF_SrcColorOneMinusSrcColor)
{
uint32_t dest = destLine[x];
uint32_t sfactor_r = shadedfg_r; sfactor_r += sfactor_r >> 7; // 255 -> 256
uint32_t sfactor_g = shadedfg_g; sfactor_g += sfactor_g >> 7; // 255 -> 256
uint32_t sfactor_b = shadedfg_b; sfactor_b += sfactor_b >> 7; // 255 -> 256
uint32_t sfactor_a = fgalpha; sfactor_a += sfactor_a >> 7; // 255 -> 256
uint32_t dfactor_r = 256 - sfactor_r;
uint32_t dfactor_g = 256 - sfactor_g;
uint32_t dfactor_b = 256 - sfactor_b;
uint32_t out_r = (RPART(dest) * dfactor_r + shadedfg_r * sfactor_r + 128) >> 8;
uint32_t out_g = (GPART(dest) * dfactor_g + shadedfg_g * sfactor_g + 128) >> 8;
uint32_t out_b = (BPART(dest) * dfactor_b + shadedfg_b * sfactor_b + 128) >> 8;
destLine[x] = MAKEARGB(255, out_r, out_g, out_b);
}
else if (ModeT::BlendSrc == STYLEALPHA_Src && ModeT::BlendDest == STYLEALPHA_InvSrc && fgalpha == 255)
{
destLine[x] = MAKEARGB(255, shadedfg_r, shadedfg_g, shadedfg_b);
}
else if (ModeT::BlendSrc != STYLEALPHA_Src || ModeT::BlendDest != STYLEALPHA_InvSrc || fgalpha != 0)
{
uint32_t dest = destLine[x];
uint32_t sfactor = fgalpha; sfactor += sfactor >> 7; // 255 -> 256
uint32_t src_r = shadedfg_r * sfactor;
uint32_t src_g = shadedfg_g * sfactor;
uint32_t src_b = shadedfg_b * sfactor;
uint32_t dest_r = RPART(dest);
uint32_t dest_g = GPART(dest);
uint32_t dest_b = BPART(dest);
if (ModeT::BlendDest == STYLEALPHA_One)
{
dest_r <<= 8;
dest_g <<= 8;
dest_b <<= 8;
}
else
{
uint32_t dfactor = 256 - sfactor;
dest_r *= dfactor;
dest_g *= dfactor;
dest_b *= dfactor;
}
uint32_t out_r, out_g, out_b;
if (ModeT::BlendOp == STYLEOP_Add)
{
if (ModeT::BlendDest == STYLEALPHA_One)
{
out_r = MIN<int32_t>((dest_r + src_r + 128) >> 8, 255);
out_g = MIN<int32_t>((dest_g + src_g + 128) >> 8, 255);
out_b = MIN<int32_t>((dest_b + src_b + 128) >> 8, 255);
}
else
{
out_r = (dest_r + src_r + 128) >> 8;
out_g = (dest_g + src_g + 128) >> 8;
out_b = (dest_b + src_b + 128) >> 8;
}
}
else if (ModeT::BlendOp == STYLEOP_RevSub)
{
out_r = MAX<int32_t>(static_cast<int32_t>(dest_r - src_r + 128) >> 8, 0);
out_g = MAX<int32_t>(static_cast<int32_t>(dest_g - src_g + 128) >> 8, 0);
out_b = MAX<int32_t>(static_cast<int32_t>(dest_b - src_b + 128) >> 8, 0);
}
else //if (ModeT::BlendOp == STYLEOP_Sub)
{
out_r = MAX<int32_t>(static_cast<int32_t>(src_r - dest_r + 128) >> 8, 0);
out_g = MAX<int32_t>(static_cast<int32_t>(src_g - dest_g + 128) >> 8, 0);
out_b = MAX<int32_t>(static_cast<int32_t>(src_b - dest_b + 128) >> 8, 0);
}
destLine[x] = MAKEARGB(255, out_r, out_g, out_b);
}
}
posU += stepU;
}
}
}
template<typename ModeT>
void DrawRect32(const void *destOrg, int destWidth, int destHeight, int destPitch, const RectDrawArgs *args, PolyTriangleThreadData *thread)
{
using namespace TriScreenDrawerModes;
if (args->SimpleShade())
DrawRectOpt32<ModeT, DrawerOptF>(destOrg, destWidth, destHeight, destPitch, args, thread);
else
DrawRectOpt32<ModeT, DrawerOptCF>(destOrg, destWidth, destHeight, destPitch, args, thread);
}
void(*ScreenTriangle::SpanDrawers8[])(int, int, int, const TriDrawTriangleArgs *) = void(*ScreenTriangle::SpanDrawers8[])(int, int, int, const TriDrawTriangleArgs *) =
{ {
&DrawSpan8<TriScreenDrawerModes::StyleOpaque>, &DrawSpan8<TriScreenDrawerModes::StyleOpaque>,
@ -1350,58 +1854,66 @@ void(*ScreenTriangle::SpanDrawers32[])(int, int, int, const TriDrawTriangleArgs
void(*ScreenTriangle::RectDrawers8[])(const void *, int, int, int, const RectDrawArgs *, PolyTriangleThreadData *) = void(*ScreenTriangle::RectDrawers8[])(const void *, int, int, int, const RectDrawArgs *, PolyTriangleThreadData *) =
{ {
&RectScreenDrawer8<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureOpaque &DrawRect8<TriScreenDrawerModes::StyleOpaque>,
&RectScreenDrawer8<TriScreenDrawerModes::MaskedBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureMasked &DrawRect8<TriScreenDrawerModes::StyleSkycap>,
&RectScreenDrawer8<TriScreenDrawerModes::AddClampBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureAdd &DrawRect8<TriScreenDrawerModes::StyleFogBoundary>,
&RectScreenDrawer8<TriScreenDrawerModes::SubClampBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureSub &DrawRect8<TriScreenDrawerModes::StyleSrcColor>,
&RectScreenDrawer8<TriScreenDrawerModes::RevSubClampBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureRevSub &DrawRect8<TriScreenDrawerModes::StyleFill>,
&RectScreenDrawer8<TriScreenDrawerModes::AddSrcColorBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureAddSrcColor &DrawRect8<TriScreenDrawerModes::StyleNormal>,
&RectScreenDrawer8<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedOpaque &DrawRect8<TriScreenDrawerModes::StyleFuzzy>,
&RectScreenDrawer8<TriScreenDrawerModes::MaskedBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedMasked &DrawRect8<TriScreenDrawerModes::StyleStencil>,
&RectScreenDrawer8<TriScreenDrawerModes::AddClampBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedAdd &DrawRect8<TriScreenDrawerModes::StyleTranslucent>,
&RectScreenDrawer8<TriScreenDrawerModes::SubClampBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedSub &DrawRect8<TriScreenDrawerModes::StyleAdd>,
&RectScreenDrawer8<TriScreenDrawerModes::RevSubClampBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedRevSub &DrawRect8<TriScreenDrawerModes::StyleShaded>,
&RectScreenDrawer8<TriScreenDrawerModes::AddSrcColorBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedAddSrcColor &DrawRect8<TriScreenDrawerModes::StyleTranslucentStencil>,
&RectScreenDrawer8<TriScreenDrawerModes::ShadedBlend, TriScreenDrawerModes::ShadedSampler>::Execute, // Shaded &DrawRect8<TriScreenDrawerModes::StyleShadow>,
&RectScreenDrawer8<TriScreenDrawerModes::AddClampShadedBlend, TriScreenDrawerModes::ShadedSampler>::Execute, // AddShaded &DrawRect8<TriScreenDrawerModes::StyleSubtract>,
&RectScreenDrawer8<TriScreenDrawerModes::ShadedBlend, TriScreenDrawerModes::StencilSampler>::Execute, // Stencil &DrawRect8<TriScreenDrawerModes::StyleAddStencil>,
&RectScreenDrawer8<TriScreenDrawerModes::AddClampShadedBlend, TriScreenDrawerModes::StencilSampler>::Execute, // AddStencil &DrawRect8<TriScreenDrawerModes::StyleAddShaded>,
&RectScreenDrawer8<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillOpaque &DrawRect8<TriScreenDrawerModes::StyleOpaqueTranslated>,
&RectScreenDrawer8<TriScreenDrawerModes::AddClampBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillAdd &DrawRect8<TriScreenDrawerModes::StyleSrcColorTranslated>,
&RectScreenDrawer8<TriScreenDrawerModes::SubClampBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillSub &DrawRect8<TriScreenDrawerModes::StyleNormalTranslated>,
&RectScreenDrawer8<TriScreenDrawerModes::RevSubClampBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillRevSub &DrawRect8<TriScreenDrawerModes::StyleStencilTranslated>,
&RectScreenDrawer8<TriScreenDrawerModes::AddSrcColorBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillAddSrcColor &DrawRect8<TriScreenDrawerModes::StyleTranslucentTranslated>,
&RectScreenDrawer8<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::SkycapSampler>::Execute, // Skycap &DrawRect8<TriScreenDrawerModes::StyleAddTranslated>,
&RectScreenDrawer8<TriScreenDrawerModes::ShadedBlend, TriScreenDrawerModes::FuzzSampler>::Execute, // Fuzz &DrawRect8<TriScreenDrawerModes::StyleShadedTranslated>,
&RectScreenDrawer8<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::FogBoundarySampler>::Execute // FogBoundary &DrawRect8<TriScreenDrawerModes::StyleTranslucentStencilTranslated>,
&DrawRect8<TriScreenDrawerModes::StyleShadowTranslated>,
&DrawRect8<TriScreenDrawerModes::StyleSubtractTranslated>,
&DrawRect8<TriScreenDrawerModes::StyleAddStencilTranslated>,
&DrawRect8<TriScreenDrawerModes::StyleAddShadedTranslated>
}; };
void(*ScreenTriangle::RectDrawers32[])(const void *, int, int, int, const RectDrawArgs *, PolyTriangleThreadData *) = void(*ScreenTriangle::RectDrawers32[])(const void *, int, int, int, const RectDrawArgs *, PolyTriangleThreadData *) =
{ {
&RectScreenDrawer32<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureOpaque &DrawRect32<TriScreenDrawerModes::StyleOpaque>,
&RectScreenDrawer32<TriScreenDrawerModes::MaskedBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureMasked &DrawRect32<TriScreenDrawerModes::StyleSkycap>,
&RectScreenDrawer32<TriScreenDrawerModes::AddClampBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureAdd &DrawRect32<TriScreenDrawerModes::StyleFogBoundary>,
&RectScreenDrawer32<TriScreenDrawerModes::SubClampBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureSub &DrawRect32<TriScreenDrawerModes::StyleSrcColor>,
&RectScreenDrawer32<TriScreenDrawerModes::RevSubClampBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureRevSub &DrawRect32<TriScreenDrawerModes::StyleFill>,
&RectScreenDrawer32<TriScreenDrawerModes::AddSrcColorBlend, TriScreenDrawerModes::TextureSampler>::Execute, // TextureAddSrcColor &DrawRect32<TriScreenDrawerModes::StyleNormal>,
&RectScreenDrawer32<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedOpaque &DrawRect32<TriScreenDrawerModes::StyleFuzzy>,
&RectScreenDrawer32<TriScreenDrawerModes::MaskedBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedMasked &DrawRect32<TriScreenDrawerModes::StyleStencil>,
&RectScreenDrawer32<TriScreenDrawerModes::AddClampBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedAdd &DrawRect32<TriScreenDrawerModes::StyleTranslucent>,
&RectScreenDrawer32<TriScreenDrawerModes::SubClampBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedSub &DrawRect32<TriScreenDrawerModes::StyleAdd>,
&RectScreenDrawer32<TriScreenDrawerModes::RevSubClampBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedRevSub &DrawRect32<TriScreenDrawerModes::StyleShaded>,
&RectScreenDrawer32<TriScreenDrawerModes::AddSrcColorBlend, TriScreenDrawerModes::TranslatedSampler>::Execute, // TranslatedAddSrcColor &DrawRect32<TriScreenDrawerModes::StyleTranslucentStencil>,
&RectScreenDrawer32<TriScreenDrawerModes::ShadedBlend, TriScreenDrawerModes::ShadedSampler>::Execute, // Shaded &DrawRect32<TriScreenDrawerModes::StyleShadow>,
&RectScreenDrawer32<TriScreenDrawerModes::AddClampShadedBlend, TriScreenDrawerModes::ShadedSampler>::Execute, // AddShaded &DrawRect32<TriScreenDrawerModes::StyleSubtract>,
&RectScreenDrawer32<TriScreenDrawerModes::ShadedBlend, TriScreenDrawerModes::StencilSampler>::Execute, // Stencil &DrawRect32<TriScreenDrawerModes::StyleAddStencil>,
&RectScreenDrawer32<TriScreenDrawerModes::AddClampShadedBlend, TriScreenDrawerModes::StencilSampler>::Execute, // AddStencil &DrawRect32<TriScreenDrawerModes::StyleAddShaded>,
&RectScreenDrawer32<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillOpaque &DrawRect32<TriScreenDrawerModes::StyleOpaqueTranslated>,
&RectScreenDrawer32<TriScreenDrawerModes::AddClampBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillAdd &DrawRect32<TriScreenDrawerModes::StyleSrcColorTranslated>,
&RectScreenDrawer32<TriScreenDrawerModes::SubClampBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillSub &DrawRect32<TriScreenDrawerModes::StyleNormalTranslated>,
&RectScreenDrawer32<TriScreenDrawerModes::RevSubClampBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillRevSub &DrawRect32<TriScreenDrawerModes::StyleStencilTranslated>,
&RectScreenDrawer32<TriScreenDrawerModes::AddSrcColorBlend, TriScreenDrawerModes::FillSampler>::Execute, // FillAddSrcColor &DrawRect32<TriScreenDrawerModes::StyleTranslucentTranslated>,
&RectScreenDrawer32<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::SkycapSampler>::Execute, // Skycap &DrawRect32<TriScreenDrawerModes::StyleAddTranslated>,
&RectScreenDrawer32<TriScreenDrawerModes::ShadedBlend, TriScreenDrawerModes::FuzzSampler>::Execute, // Fuzz &DrawRect32<TriScreenDrawerModes::StyleShadedTranslated>,
&RectScreenDrawer32<TriScreenDrawerModes::OpaqueBlend, TriScreenDrawerModes::FogBoundarySampler>::Execute, // FogBoundary &DrawRect32<TriScreenDrawerModes::StyleTranslucentStencilTranslated>,
&DrawRect32<TriScreenDrawerModes::StyleShadowTranslated>,
&DrawRect32<TriScreenDrawerModes::StyleSubtractTranslated>,
&DrawRect32<TriScreenDrawerModes::StyleAddStencilTranslated>,
&DrawRect32<TriScreenDrawerModes::StyleAddShadedTranslated>
}; };
int ScreenTriangle::FuzzStart = 0; int ScreenTriangle::FuzzStart = 0;

57
src/polyrenderer/drawers/screen_triangle.h
View file

@ -136,34 +136,6 @@ enum class TriBlendMode
AddShadedTranslated AddShadedTranslated
}; };
enum class RectBlendMode
{
TextureOpaque,
TextureMasked,
TextureAdd,
TextureSub,
TextureRevSub,
TextureAddSrcColor,
TranslatedOpaque,
TranslatedMasked,
TranslatedAdd,
TranslatedSub,
TranslatedRevSub,
TranslatedAddSrcColor,
Shaded,
AddShaded,
Stencil,
AddStencil,
FillOpaque,
FillAdd,
FillSub,
FillRevSub,
FillAddSrcColor,
Skycap,
Fuzz,
FogBoundary
};
class ScreenTriangle class ScreenTriangle
{ {
public: public:
@ -219,35 +191,6 @@ namespace TriScreenDrawerModes
struct StyleAddStencilTranslated { static const int BlendOp = STYLEOP_Add, BlendSrc = STYLEALPHA_Src, BlendDest = STYLEALPHA_One, Flags = STYLEF_ColorIsFixed, SWFlags = SWSTYLEF_Translated; }; struct StyleAddStencilTranslated { static const int BlendOp = STYLEOP_Add, BlendSrc = STYLEALPHA_Src, BlendDest = STYLEALPHA_One, Flags = STYLEF_ColorIsFixed, SWFlags = SWSTYLEF_Translated; };
struct StyleAddShadedTranslated { static const int BlendOp = STYLEOP_Add, BlendSrc = STYLEALPHA_Src, BlendDest = STYLEALPHA_One, Flags = STYLEF_RedIsAlpha | STYLEF_ColorIsFixed, SWFlags = SWSTYLEF_Translated; }; struct StyleAddShadedTranslated { static const int BlendOp = STYLEOP_Add, BlendSrc = STYLEALPHA_Src, BlendDest = STYLEALPHA_One, Flags = STYLEF_RedIsAlpha | STYLEF_ColorIsFixed, SWFlags = SWSTYLEF_Translated; };
enum class BlendModes { Opaque, Masked, AddClamp, SubClamp, RevSubClamp, AddSrcColorOneMinusSrcColor, Shaded, AddClampShaded };
struct OpaqueBlend { static const int Mode = (int)BlendModes::Opaque; };
struct MaskedBlend { static const int Mode = (int)BlendModes::Masked; };
struct AddClampBlend { static const int Mode = (int)BlendModes::AddClamp; };
struct SubClampBlend { static const int Mode = (int)BlendModes::SubClamp; };
struct RevSubClampBlend { static const int Mode = (int)BlendModes::RevSubClamp; };
struct AddSrcColorBlend { static const int Mode = (int)BlendModes::AddSrcColorOneMinusSrcColor; };
struct ShadedBlend { static const int Mode = (int)BlendModes::Shaded; };
struct AddClampShadedBlend { static const int Mode = (int)BlendModes::AddClampShaded; };
enum class FilterModes { Nearest, Linear };
struct NearestFilter { static const int Mode = (int)FilterModes::Nearest; };
struct LinearFilter { static const int Mode = (int)FilterModes::Linear; };
enum class ShadeMode { None, Simple, Advanced };
struct NoShade { static const int Mode = (int)ShadeMode::None; };
struct SimpleShade { static const int Mode = (int)ShadeMode::Simple; };
struct AdvancedShade { static const int Mode = (int)ShadeMode::Advanced; };
enum class Samplers { Texture, Fill, Shaded, Stencil, Translated, Skycap, Fuzz, FogBoundary };
struct TextureSampler { static const int Mode = (int)Samplers::Texture; };
struct FillSampler { static const int Mode = (int)Samplers::Fill; };
struct ShadedSampler { static const int Mode = (int)Samplers::Shaded; };
struct StencilSampler { static const int Mode = (int)Samplers::Stencil; };
struct TranslatedSampler { static const int Mode = (int)Samplers::Translated; };
struct SkycapSampler { static const int Mode = (int)Samplers::Skycap; };
struct FuzzSampler { static const int Mode = (int)Samplers::Fuzz; };
struct FogBoundarySampler { static const int Mode = (int)Samplers::FogBoundary; };
enum SWOptFlags enum SWOptFlags
{ {
SWOPT_DynLights = 1, SWOPT_DynLights = 1,