implemented scissoring of triangles and extra necessary blendmode for xonotic hud

git-svn-id: svn://svn.icculus.org/twilight/trunk/darkplaces@10815 d7cf8633-e32d-0410-b094-e92efae38249

diff --git a/dpsoftrast.c b/dpsoftrast.c
index 072fd564466bc209f876be0aa463e75132c07cce..67a111a7e57d929e436a84d5ea32615dc3babb79 100644 (file)
--- a/dpsoftrast.c
+++ b/dpsoftrast.c
@@ -179,7 +179,6 @@ typedef ALIGN(struct DPSOFTRAST_State_Span_s
        int triangle; // triangle this span was generated by
        int x; // framebuffer x coord
        int y; // framebuffer y coord
-       int length; // pixel count
        int startx; // usable range (according to pixelmask)
        int endx; // usable range (according to pixelmask)
        unsigned char *pixelmask; // true for pixels that passed depth test, false for others
@@ -205,6 +204,7 @@ typedef enum DPSOFTRAST_BLENDMODE_e
        DPSOFTRAST_BLENDMODE_MUL2,
        DPSOFTRAST_BLENDMODE_SUBALPHA,
        DPSOFTRAST_BLENDMODE_PSEUDOALPHA,
+       DPSOFTRAST_BLENDMODE_INVADD,
        DPSOFTRAST_BLENDMODE_TOTAL
 }
 DPSOFTRAST_BLENDMODE;
@@ -243,7 +243,7 @@ typedef ATOMIC(struct DPSOFTRAST_State_Thread_s
 
        // derived values (DPSOFTRAST_VALIDATE_FB)
        int fb_colormask;
-       int fb_clearscissor[4];
+       int fb_scissor[4];
        ALIGN(float fb_viewportcenter[4]);
        ALIGN(float fb_viewportscale[4]);
 
@@ -365,10 +365,10 @@ static void DPSOFTRAST_RecalcFB(DPSOFTRAST_State_Thread *thread)
        if (x2 > dpsoftrast.fb_width) x2 = dpsoftrast.fb_width;
        if (y1 < 0) y1 = 0;
        if (y2 > dpsoftrast.fb_height) y2 = dpsoftrast.fb_height;
-       thread->fb_clearscissor[0] = x1;
-       thread->fb_clearscissor[1] = y1;
-       thread->fb_clearscissor[2] = x2 - x1;
-       thread->fb_clearscissor[3] = y2 - y1;
+       thread->fb_scissor[0] = x1;
+       thread->fb_scissor[1] = y1;
+       thread->fb_scissor[2] = x2 - x1;
+       thread->fb_scissor[3] = y2 - y1;
 
        DPSOFTRAST_RecalcViewport(thread->viewport, thread->fb_viewportcenter, thread->fb_viewportscale);
 }
@@ -386,7 +386,7 @@ static void DPSOFTRAST_RecalcBlendFunc(DPSOFTRAST_State_Thread *thread)
                {
                #define BLENDFUNC(sfactor, dfactor, blendmode) \
                        case (sfactor<<16)|dfactor: thread->fb_blendmode = blendmode; break;
-               BLENDFUNC(GL_SRC_COLOR, GL_ONE, DPSOFTRAST_BLENDMODE_SUBALPHA)
+               BLENDFUNC(GL_SRC_ALPHA, GL_ONE, DPSOFTRAST_BLENDMODE_SUBALPHA)
                default: thread->fb_blendmode = DPSOFTRAST_BLENDMODE_OPAQUE; break;
                }
        }
@@ -403,7 +403,7 @@ static void DPSOFTRAST_RecalcBlendFunc(DPSOFTRAST_State_Thread *thread)
                BLENDFUNC(GL_DST_COLOR, GL_ZERO, DPSOFTRAST_BLENDMODE_MUL)
                BLENDFUNC(GL_DST_COLOR, GL_SRC_COLOR, DPSOFTRAST_BLENDMODE_MUL2)
                BLENDFUNC(GL_ONE, GL_ONE_MINUS_SRC_ALPHA, DPSOFTRAST_BLENDMODE_PSEUDOALPHA)
-               BLENDFUNC(GL_SRC_COLOR, GL_ONE, DPSOFTRAST_BLENDMODE_SUBALPHA)
+               BLENDFUNC(GL_ONE_MINUS_DST_COLOR, GL_ONE, DPSOFTRAST_BLENDMODE_INVADD)
                default: thread->fb_blendmode = DPSOFTRAST_BLENDMODE_OPAQUE; break;
                }
        }
@@ -900,10 +900,10 @@ static void DPSOFTRAST_Interpret_ClearColor(DPSOFTRAST_State_Thread *thread, con
        unsigned int *p;
        unsigned int c;
        DPSOFTRAST_Validate(thread, DPSOFTRAST_VALIDATE_FB);
-       x1 = thread->fb_clearscissor[0];
-       y1 = thread->fb_clearscissor[1];
-       x2 = thread->fb_clearscissor[0] + thread->fb_clearscissor[2];
-       y2 = thread->fb_clearscissor[1] + thread->fb_clearscissor[3];
+       x1 = thread->fb_scissor[0];
+       y1 = thread->fb_scissor[1];
+       x2 = thread->fb_scissor[0] + thread->fb_scissor[2];
+       y2 = thread->fb_scissor[1] + thread->fb_scissor[3];
        if (y1 < miny1) y1 = miny1;
        if (y2 > maxy2) y2 = maxy2;
        w = x2 - x1;
@@ -946,10 +946,10 @@ static void DPSOFTRAST_Interpret_ClearDepth(DPSOFTRAST_State_Thread *thread, DPS
        unsigned int *p;
        unsigned int c;
        DPSOFTRAST_Validate(thread, DPSOFTRAST_VALIDATE_FB);
-       x1 = thread->fb_clearscissor[0];
-       y1 = thread->fb_clearscissor[1];
-       x2 = thread->fb_clearscissor[0] + thread->fb_clearscissor[2];
-       y2 = thread->fb_clearscissor[1] + thread->fb_clearscissor[3];
+       x1 = thread->fb_scissor[0];
+       y1 = thread->fb_scissor[1];
+       x2 = thread->fb_scissor[0] + thread->fb_scissor[2];
+       y2 = thread->fb_scissor[1] + thread->fb_scissor[3];
        if (y1 < miny1) y1 = miny1;
        if (y2 > maxy2) y2 = maxy2;
        w = x2 - x1;
@@ -2125,6 +2125,21 @@ void DPSOFTRAST_Draw_Span_Finish(DPSOFTRAST_State_Thread *thread, const DPSOFTRA
                        pixel[x*4+3] = d[3];
                }
                break;
+       case DPSOFTRAST_BLENDMODE_INVADD:
+               for (x = startx;x < endx;x++)
+               {
+                       if (!pixelmask[x])
+                               continue;
+                       d[0] = (int)((255.0f-pixel[x*4+2])*in4f[x*4+0] + pixel[x*4+2]);if (d[0] > 255) d[0] = 255;
+                       d[1] = (int)((255.0f-pixel[x*4+1])*in4f[x*4+1] + pixel[x*4+1]);if (d[1] > 255) d[1] = 255;
+                       d[2] = (int)((255.0f-pixel[x*4+0])*in4f[x*4+2] + pixel[x*4+0]);if (d[2] > 255) d[2] = 255;
+                       d[3] = (int)((255.0f-pixel[x*4+3])*in4f[x*4+3] + pixel[x*4+3]);if (d[3] > 255) d[3] = 255;
+                       pixel[x*4+0] = d[0];
+                       pixel[x*4+1] = d[1];
+                       pixel[x*4+2] = d[2];
+                       pixel[x*4+3] = d[3];
+               }
+               break;
        }
 }
 
@@ -2259,6 +2274,12 @@ void DPSOFTRAST_Draw_Span_FinishBGRA8(DPSOFTRAST_State_Thread *thread, const DPS
                        dst = _mm_add_epi16(src, _mm_sub_epi16(dst, _mm_srli_epi16(_mm_mullo_epi16(dst, blend), 8)));
                });
                break;
+       case DPSOFTRAST_BLENDMODE_INVADD:
+               FINISHBLEND({
+                       dst = _mm_add_epi16(dst, _mm_mulhi_epi16(_mm_slli_epi16(_mm_sub_epi16(_mm_set1_epi16(255), dst), 4), _mm_slli_epi16(src, 4)));
+               }, {
+                       dst = _mm_add_epi16(dst, _mm_mulhi_epi16(_mm_slli_epi16(_mm_sub_epi16(_mm_set1_epi16(255), dst), 4), _mm_slli_epi16(src, 4)));
+               });
        }
 #endif
 }
@@ -2469,7 +2490,7 @@ void DPSOFTRAST_Draw_Span_Texture2DVaryingBGRA8(DPSOFTRAST_State_Thread *thread,
        // if no texture is bound, just fill it with white
        if (!texture)
        {
-               memset(out4ub + startx*4, 255, span->length*4);
+               memset(out4ub + startx*4, 255, (span->endx - span->startx)*4);
                return;
        }
        mip = triangle->mip[texunitindex];
@@ -2728,7 +2749,7 @@ void DPSOFTRAST_Draw_Span_Texture2DVaryingBGRA8(DPSOFTRAST_State_Thread *thread,
 void DPSOFTRAST_Draw_Span_TextureCubeVaryingBGRA8(const DPSOFTRAST_State_Triangle * RESTRICT triangle, const DPSOFTRAST_State_Span * RESTRICT span, unsigned char * RESTRICT out4ub, int texunitindex, int arrayindex, const float * RESTRICT zf)
 {
        // TODO: IMPLEMENT
-       memset(out4ub, 255, span->length*4);
+       memset(out4ub + span->startx*4, 255, (span->startx - span->endx)*4);
 }
 
 float DPSOFTRAST_SampleShadowmap(const float *vector)
@@ -3185,7 +3206,7 @@ void DPSOFTRAST_PixelShader_Depth_Or_Shadow(DPSOFTRAST_State_Thread *thread, con
        float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
        unsigned char buffer_FragColorbgra8[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
        DPSOFTRAST_Draw_Span_Begin(thread, triangle, span, buffer_z);
-       memset(buffer_FragColorbgra8, 0, span->length*4);
+       memset(buffer_FragColorbgra8 + span->startx*4, 0, (span->endx - span->startx)*4);
        DPSOFTRAST_Draw_Span_FinishBGRA8(thread, triangle, span, buffer_FragColorbgra8);
 }
 
@@ -3407,7 +3428,7 @@ void DPSOFTRAST_PixelShader_FakeLight(DPSOFTRAST_State_Thread *thread, const DPS
        float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
        unsigned char buffer_FragColorbgra8[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
        DPSOFTRAST_Draw_Span_Begin(thread, triangle, span, buffer_z);
-       memset(buffer_FragColorbgra8, 0, span->length*4);
+       memset(buffer_FragColorbgra8 + span->startx*4, 0, (span->endx - span->startx)*4);
        DPSOFTRAST_Draw_Span_FinishBGRA8(thread, triangle, span, buffer_FragColorbgra8);
 }
 
@@ -4090,7 +4111,7 @@ void DPSOFTRAST_PixelShader_Refraction(DPSOFTRAST_State_Thread *thread, const DP
        float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
        unsigned char buffer_FragColorbgra8[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
        DPSOFTRAST_Draw_Span_Begin(thread, triangle, span, buffer_z);
-       memset(buffer_FragColorbgra8, 0, span->length*4);
+       memset(buffer_FragColorbgra8 + span->startx*4, 0, (span->endx - span->startx)*4);
        DPSOFTRAST_Draw_Span_FinishBGRA8(thread, triangle, span, buffer_FragColorbgra8);
 }
 
@@ -4108,7 +4129,7 @@ void DPSOFTRAST_PixelShader_Water(DPSOFTRAST_State_Thread *thread, const DPSOFTR
        float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
        unsigned char buffer_FragColorbgra8[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
        DPSOFTRAST_Draw_Span_Begin(thread, triangle, span, buffer_z);
-       memset(buffer_FragColorbgra8, 0, span->length*4);
+       memset(buffer_FragColorbgra8 + span->startx*4, 0, (span->endx - span->startx)*4);
        DPSOFTRAST_Draw_Span_FinishBGRA8(thread, triangle, span, buffer_FragColorbgra8);
 }
 
@@ -4125,7 +4146,7 @@ void DPSOFTRAST_PixelShader_ShowDepth(DPSOFTRAST_State_Thread *thread, const DPS
        float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
        unsigned char buffer_FragColorbgra8[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
        DPSOFTRAST_Draw_Span_Begin(thread, triangle, span, buffer_z);
-       memset(buffer_FragColorbgra8, 0, span->length*4);
+       memset(buffer_FragColorbgra8 + span->startx*4, 0, (span->endx - span->startx)*4);
        DPSOFTRAST_Draw_Span_FinishBGRA8(thread, triangle, span, buffer_FragColorbgra8);
 }
 
@@ -4142,7 +4163,7 @@ void DPSOFTRAST_PixelShader_DeferredGeometry(DPSOFTRAST_State_Thread *thread, co
        float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
        unsigned char buffer_FragColorbgra8[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
        DPSOFTRAST_Draw_Span_Begin(thread, triangle, span, buffer_z);
-       memset(buffer_FragColorbgra8, 0, span->length*4);
+       memset(buffer_FragColorbgra8 + span->startx*4, 0, (span->endx - span->startx)*4);
        DPSOFTRAST_Draw_Span_FinishBGRA8(thread, triangle, span, buffer_FragColorbgra8);
 }
 
@@ -4159,7 +4180,7 @@ void DPSOFTRAST_PixelShader_DeferredLightSource(DPSOFTRAST_State_Thread *thread,
        float buffer_z[DPSOFTRAST_DRAW_MAXSPANLENGTH];
        unsigned char buffer_FragColorbgra8[DPSOFTRAST_DRAW_MAXSPANLENGTH*4];
        DPSOFTRAST_Draw_Span_Begin(thread, triangle, span, buffer_z);
-       memset(buffer_FragColorbgra8, 0, span->length*4);
+       memset(buffer_FragColorbgra8 + span->startx*4, 0, (span->endx - span->startx)*4);
        DPSOFTRAST_Draw_Span_FinishBGRA8(thread, triangle, span, buffer_FragColorbgra8);
 }
 
@@ -4223,23 +4244,23 @@ void DPSOFTRAST_Draw_ProcessSpans(DPSOFTRAST_State_Thread *thread)
                        depthslope = (int)(wslope*DPSOFTRAST_DEPTHSCALE);
                        depth = (int)(w*DPSOFTRAST_DEPTHSCALE - DPSOFTRAST_DEPTHOFFSET*(thread->polygonoffset[1] + fabs(wslope)*thread->polygonoffset[0]));
                        depthpixel = dpsoftrast.fb_depthpixels + span->y * dpsoftrast.fb_width + span->x;
+                       startx = span->startx;
+                       endx = span->endx;
                        switch(thread->fb_depthfunc)
                        {
                        default:
-                       case GL_ALWAYS:  for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = true; break;
-                       case GL_LESS:    for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] < d; break;
-                       case GL_LEQUAL:  for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] <= d; break;
-                       case GL_EQUAL:   for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] == d; break;
-                       case GL_GEQUAL:  for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] >= d; break;
-                       case GL_GREATER: for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = depthpixel[x] > d; break;
-                       case GL_NEVER:   for (x = 0, d = depth;x < span->length;x++, d += depthslope) pixelmask[x] = false; break;
+                       case GL_ALWAYS:  for (x = startx, d = depth + depthslope*startx;x < endx;x++, d += depthslope) pixelmask[x] = true; break;
+                       case GL_LESS:    for (x = startx, d = depth + depthslope*startx;x < endx;x++, d += depthslope) pixelmask[x] = depthpixel[x] < d; break;
+                       case GL_LEQUAL:  for (x = startx, d = depth + depthslope*startx;x < endx;x++, d += depthslope) pixelmask[x] = depthpixel[x] <= d; break;
+                       case GL_EQUAL:   for (x = startx, d = depth + depthslope*startx;x < endx;x++, d += depthslope) pixelmask[x] = depthpixel[x] == d; break;
+                       case GL_GEQUAL:  for (x = startx, d = depth + depthslope*startx;x < endx;x++, d += depthslope) pixelmask[x] = depthpixel[x] >= d; break;
+                       case GL_GREATER: for (x = startx, d = depth + depthslope*startx;x < endx;x++, d += depthslope) pixelmask[x] = depthpixel[x] > d; break;
+                       case GL_NEVER:   for (x = startx, d = depth + depthslope*startx;x < endx;x++, d += depthslope) pixelmask[x] = false; break;
                        }
                        //colorpixel = dpsoftrast.fb_colorpixels[0] + (span->y * dpsoftrast.fb_width + span->x) * 4;;
-                       //for (x = 0;x < span->length;x++)
+                       //for (x = startx;x < endx;x++)
                        //      colorpixel[x] = (depthpixel[x] & 0xFF000000) ? (0x00FF0000) : (depthpixel[x] & 0x00FF0000);
                        // if there is no color buffer, skip pixel shader
-                       startx = 0;
-                       endx = span->length;
                        while (startx < endx && !pixelmask[startx])
                                startx++;
                        while (endx > startx && !pixelmask[endx-1])
@@ -4265,10 +4286,8 @@ void DPSOFTRAST_Draw_ProcessSpans(DPSOFTRAST_State_Thread *thread)
                        // if there is no color buffer, skip pixel shader
                        if (dpsoftrast.fb_colorpixels[0] && thread->fb_colormask)
                        {
-                               memset(pixelmask, 1, span->length);
+                               memset(pixelmask + span->startx, 1, span->endx - span->startx);
                                span->pixelmask = pixelmask;
-                               span->startx = 0;
-                               span->endx = span->length;
                                DPSOFTRAST_ShaderModeTable[thread->shader_mode].Span(thread, triangle, span);
                        }
                }
@@ -4282,11 +4301,8 @@ static void DPSOFTRAST_Interpret_Draw(DPSOFTRAST_State_Thread *thread, DPSOFTRAS
 {
 #ifdef SSE2_PRESENT
        int cullface = thread->cullface;
-       int width = dpsoftrast.fb_width;
-       int miny1 = thread->miny1;
-       int maxy1 = thread->maxy1;
-       int miny2 = thread->miny2;
-       int maxy2 = thread->maxy2;
+       int minx, maxx, miny, maxy;
+       int miny1, maxy1, miny2, maxy2;
        __m128i fbmin, fbmax;
        __m128 viewportcenter, viewportscale;
        int firstvertex = command->firstvertex;
@@ -4310,6 +4326,13 @@ static void DPSOFTRAST_Interpret_Draw(DPSOFTRAST_State_Thread *thread, DPSOFTRAS
        __m128 screen[4];
        DPSOFTRAST_State_Triangle *triangle;
        DPSOFTRAST_Texture *texture;
+       DPSOFTRAST_ValidateQuick(thread, DPSOFTRAST_VALIDATE_DRAW);
+       miny = thread->fb_scissor[1];
+       maxy = thread->fb_scissor[1] + thread->fb_scissor[3];
+       miny1 = bound(miny, thread->miny1, maxy);
+       maxy1 = bound(miny, thread->maxy1, maxy);
+       miny2 = bound(miny, thread->miny2, maxy);
+       maxy2 = bound(miny, thread->maxy2, maxy);
        if ((command->starty >= maxy1 || command->endy <= miny1) && (command->starty >= maxy2 || command->endy <= miny2))
        {
                if (!ATOMIC_DECREMENT(command->refcount))
@@ -4319,9 +4342,10 @@ static void DPSOFTRAST_Interpret_Draw(DPSOFTRAST_State_Thread *thread, DPSOFTRAS
                }
                return;
        }
-       DPSOFTRAST_ValidateQuick(thread, DPSOFTRAST_VALIDATE_DRAW);
-       fbmin = _mm_setr_epi16(0, miny1, 0, miny1, 0, miny1, 0, miny1);
-       fbmax = _mm_sub_epi16(_mm_setr_epi16(width, maxy2, width, maxy2, width, maxy2, width, maxy2), _mm_set1_epi16(1));
+       minx = thread->fb_scissor[0];
+       maxx = thread->fb_scissor[0] + thread->fb_scissor[2];
+       fbmin = _mm_setr_epi16(minx, miny1, minx, miny1, minx, miny1, minx, miny1);
+       fbmax = _mm_sub_epi16(_mm_setr_epi16(maxx, maxy2, maxx, maxy2, maxx, maxy2, maxx, maxy2), _mm_set1_epi16(1));
        viewportcenter = _mm_load_ps(thread->fb_viewportcenter);
        viewportscale = _mm_load_ps(thread->fb_viewportscale);
        screen[3] = _mm_setzero_ps();
@@ -4644,19 +4668,23 @@ static void DPSOFTRAST_Interpret_Draw(DPSOFTRAST_State_Thread *thread, DPSOFTRAS
                                int startx, endx, offset;
                                startx = _mm_cvtss_si32(xcoords);
                                endx = _mm_cvtss_si32(_mm_movehl_ps(xcoords, xcoords));
-                               if (startx < 0) startx = 0;
-                               if (endx > dpsoftrast.fb_width) endx = dpsoftrast.fb_width;
+                               if (startx < minx) 
+                               {
+                                       if (startx < 0) startx = 0;
+                                       startx += (minx-startx)&~(DPSOFTRAST_DRAW_MAXSPANLENGTH-1);
+                               }
+                               if (endx > maxx) endx = maxx;
                                if (startx >= endx) continue;
-                               for (offset = startx; offset < endx;)
+                               for (offset = startx; offset < endx;offset += DPSOFTRAST_DRAW_MAXSPANLENGTH)
                                {
                                        DPSOFTRAST_State_Span *span = &thread->spans[thread->numspans];
                                        span->triangle = thread->numtriangles;
                                        span->x = offset;
                                        span->y = y;
-                                       span->length = endx - offset;
-                                       if (span -> length > DPSOFTRAST_DRAW_MAXSPANLENGTH)
-                                               span -> length = DPSOFTRAST_DRAW_MAXSPANLENGTH;
-                                       offset += span->length;
+                                       span->startx = max(minx - offset, 0);
+                                       span->endx = min(endx - offset, DPSOFTRAST_DRAW_MAXSPANLENGTH);
+                                       if (span->startx >= span->endx)
+                                               continue; 
                                        if (++thread->numspans >= DPSOFTRAST_DRAW_MAXSPANS)
                                                DPSOFTRAST_Draw_ProcessSpans(thread);
                                }