flip dpsoftrast's texture memory layout so it matches the framebuffer orientation...

however, r_viewfbo and r_bloom in combination are still broken on GPUs/dpsoftrast without non-power-of-2 textures due to not matching the position DP's bloom expects them to be located at

git-svn-id: svn://svn.icculus.org/twilight/trunk/darkplaces@11649 d7cf8633-e32d-0410-b094-e92efae38249

diff --git a/dpsoftrast.c b/dpsoftrast.c
index a79a9d363aee62cc8b45416277d572eca511b601..a646db35bd13d32da8f143f7d5b75f028e8ca292 100644 (file)
--- a/dpsoftrast.c
+++ b/dpsoftrast.c
@@ -748,12 +748,12 @@ void DPSOFTRAST_Texture_UpdatePartial(int index, int mip, const unsigned char *p
                DPSOFTRAST_Flush();
        if (pixels)
        {
-               dst = texture->bytes + (blocky * texture->mipmap[0][2] + blockx) * 4;
+               dst = texture->bytes + texture->mipmap[0][1] +(-blocky * texture->mipmap[0][2] + blockx) * 4;
                while (blockheight > 0)
                {
+                       dst -= texture->mipmap[0][2] * 4;
                        memcpy(dst, pixels, blockwidth * 4);
                        pixels += blockwidth * 4;
-                       dst += texture->mipmap[0][2] * 4;
                        blockheight--;
                }
        }
@@ -766,7 +766,16 @@ void DPSOFTRAST_Texture_UpdateFull(int index, const unsigned char *pixels)
        if (texture->binds)
                DPSOFTRAST_Flush();
        if (pixels)
-               memcpy(texture->bytes, pixels, texture->mipmap[0][1]);
+       {
+               int i, stride = texture->mipmap[0][2]*4;
+               unsigned char *dst = texture->bytes + texture->mipmap[0][1];
+               for (i = texture->mipmap[0][3];i > 0;i--)
+               {
+                       dst -= stride;
+                       memcpy(dst, pixels, stride);
+                       pixels += stride;
+               }
+       }
        DPSOFTRAST_Texture_CalculateMipmaps(index);
 }
 int DPSOFTRAST_Texture_GetWidth(int index, int mip)
@@ -1269,9 +1278,10 @@ void DPSOFTRAST_CopyRectangleToTexture(int index, int mip, int tx, int ty, int s
        if (th > sh) th = sh;
        if (tw < 1 || th < 1)
                return;
-       sy1 = sheight - 1 - sy1;
+       sy1 = sheight - sy1 - th;
+       ty1 = theight - ty1 - th;
        for (y = 0;y < th;y++)
-               memcpy(tpixels + ((ty1 + y) * twidth + tx1), spixels + ((sy1 - y) * swidth + sx1), tw*4);
+               memcpy(tpixels + ((ty1 + y) * twidth + tx1), spixels + ((sy1 + y) * swidth + sx1), tw*4);
        if (texture->mipmaps > 1)
                DPSOFTRAST_Texture_CalculateMipmaps(index);
 }
@@ -2293,7 +2303,7 @@ static void DPSOFTRAST_Texture2DBGRA8(DPSOFTRAST_Texture *texture, int mip, floa
        const unsigned char * RESTRICT pixel[4];
        int width = texture->mipmap[mip][2], height = texture->mipmap[mip][3];
        int wrapmask[2] = { width-1, height-1 };
-       pixelbase = (unsigned char *)texture->bytes + texture->mipmap[mip][0];
+       pixelbase = (unsigned char *)texture->bytes + texture->mipmap[mip][0] + texture->mipmap[mip][1] - 4*width;
        if(texture->filter & DPSOFTRAST_TEXTURE_FILTER_LINEAR)
        {
                unsigned int tc[2] = { x * (width<<12) - 2048, y * (height<<12) - 2048};
@@ -2316,10 +2326,10 @@ static void DPSOFTRAST_Texture2DBGRA8(DPSOFTRAST_Texture *texture, int mip, floa
                        tci1[0] &= wrapmask[0];
                        tci1[1] &= wrapmask[1];
                }
-               pixel[0] = pixelbase + 4 * (tci[1]*width+tci[0]);
-               pixel[1] = pixelbase + 4 * (tci[1]*width+tci1[0]);
-               pixel[2] = pixelbase + 4 * (tci1[1]*width+tci[0]);
-               pixel[3] = pixelbase + 4 * (tci1[1]*width+tci1[0]);
+               pixel[0] = pixelbase + 4 * (tci[0] - tci[1]*width);
+               pixel[1] = pixelbase + 4 * (tci[0] - tci[1]*width);
+               pixel[2] = pixelbase + 4 * (tci[0] - tci1[1]*width);
+               pixel[3] = pixelbase + 4 * (tci[0] - tci1[1]*width);
                c[0] = (pixel[0][0]*lerp[0]+pixel[1][0]*lerp[1]+pixel[2][0]*lerp[2]+pixel[3][0]*lerp[3])>>24;
                c[1] = (pixel[0][1]*lerp[0]+pixel[1][1]*lerp[1]+pixel[2][1]*lerp[2]+pixel[3][1]*lerp[3])>>24;
                c[2] = (pixel[0][2]*lerp[0]+pixel[1][2]*lerp[1]+pixel[2][2]*lerp[2]+pixel[3][2]*lerp[3])>>24;
@@ -2338,7 +2348,7 @@ static void DPSOFTRAST_Texture2DBGRA8(DPSOFTRAST_Texture *texture, int mip, floa
                        tci[0] &= wrapmask[0];
                        tci[1] &= wrapmask[1];
                }
-               pixel[0] = pixelbase + 4 * (tci[1]*width+tci[0]);
+               pixel[0] = pixelbase + 4 * (tci[0] - tci[1]*width);
                c[0] = pixel[0][0];
                c[1] = pixel[0][1];
                c[2] = pixel[0][2];
@@ -2382,7 +2392,7 @@ static void DPSOFTRAST_Draw_Span_Texture2DVarying(DPSOFTRAST_State_Thread *threa
                return;
        }
        mip = triangle->mip[texunitindex];
-       pixelbase = (unsigned char *)texture->bytes + texture->mipmap[mip][0];
+       pixelbase = (unsigned char *)texture->bytes + texture->mipmap[mip][0] + texture->mipmap[mip][1] - 4*texture->mipmap[mip][2];
        // if this mipmap of the texture is 1 pixel, just fill it with that color
        if (texture->mipmap[mip][1] == 4)
        {
@@ -2404,7 +2414,7 @@ static void DPSOFTRAST_Draw_Span_Texture2DVarying(DPSOFTRAST_State_Thread *threa
        flags = texture->flags;
        tcscale[0] = texture->mipmap[mip][2];
        tcscale[1] = texture->mipmap[mip][3];
-       tciwidth = texture->mipmap[mip][2];
+       tciwidth = -texture->mipmap[mip][2];
        tcimin[0] = 0;
        tcimin[1] = 0;
        tcimax[0] = texture->mipmap[mip][2]-1;
@@ -2569,7 +2579,7 @@ static void DPSOFTRAST_Draw_Span_Texture2DVaryingBGRA8(DPSOFTRAST_State_Thread *
                return;
        }
        mip = triangle->mip[texunitindex];
-       pixelbase = (const unsigned char *)texture->bytes + texture->mipmap[mip][0];
+       pixelbase = (const unsigned char *)texture->bytes + texture->mipmap[mip][0] + texture->mipmap[mip][1] - 4*texture->mipmap[mip][2];
        // if this mipmap of the texture is 1 pixel, just fill it with that color
        if (texture->mipmap[mip][1] == 4)
        {
@@ -2591,7 +2601,7 @@ static void DPSOFTRAST_Draw_Span_Texture2DVaryingBGRA8(DPSOFTRAST_State_Thread *
        if (filter)
                endtc = _mm_sub_ps(endtc, _mm_set1_ps(0.5f));
        endsubtc = _mm_cvtps_epi32(_mm_mul_ps(endtc, _mm_set1_ps(65536.0f)));
-       tcoffset = _mm_add_epi32(_mm_slli_epi32(_mm_shuffle_epi32(tcsize, _MM_SHUFFLE(0, 0, 0, 0)), 18), _mm_set1_epi32(4));
+       tcoffset = _mm_add_epi32(_mm_slli_epi32(_mm_sub_epi32(_mm_setzero_si128(), _mm_shuffle_epi32(tcsize, _MM_SHUFFLE(0, 0, 0, 0))), 18), _mm_set1_epi32(4));
        tcmax = _mm_packs_epi32(tcmask, tcmask);
        for (x = startx;x < endx;)
        {

diff --git a/gl_rmain.c b/gl_rmain.c
index 5a3b32932f4458541e0449c7f52d2485c29c0820..d19028177c83af85e771c79d58044839c1012416 100644 (file)
--- a/gl_rmain.c
+++ b/gl_rmain.c
@@ -6168,6 +6168,14 @@ static void R_Bloom_StartFrame(void)
        r_fb.screentexcoord2f[6] = 0;
        r_fb.screentexcoord2f[7] = 0;
 
+       if(r_fb.fbo) 
+       {
+               for (i = 1;i < 8;i += 2)
+               {
+                       r_fb.screentexcoord2f[i] += 1 - (float)(viewheight + r_refdef.view.y) / (float)r_fb.screentextureheight;
+               }
+       }
+
        // set up a texcoord array for the reduced resolution bloom image
        // (which will be additive blended over the screen image)
        r_fb.bloomtexcoord2f[0] = 0;
@@ -6191,20 +6199,17 @@ static void R_Bloom_StartFrame(void)
        case RENDERPATH_D3D9:
        case RENDERPATH_D3D10:
        case RENDERPATH_D3D11:
+               for (i = 0;i < 4;i++)
                {
-                       int i;
-                       for (i = 0;i < 4;i++)
-                       {
-                               r_fb.screentexcoord2f[i*2+0] += 0.5f / (float)r_fb.screentexturewidth;
-                               r_fb.screentexcoord2f[i*2+1] += 0.5f / (float)r_fb.screentextureheight;
-                               r_fb.bloomtexcoord2f[i*2+0] += 0.5f / (float)r_fb.bloomtexturewidth;
-                               r_fb.bloomtexcoord2f[i*2+1] += 0.5f / (float)r_fb.bloomtextureheight;
-                       }
+                       r_fb.screentexcoord2f[i*2+0] += 0.5f / (float)r_fb.screentexturewidth;
+                       r_fb.screentexcoord2f[i*2+1] += 0.5f / (float)r_fb.screentextureheight;
+                       r_fb.bloomtexcoord2f[i*2+0] += 0.5f / (float)r_fb.bloomtexturewidth;
+                       r_fb.bloomtexcoord2f[i*2+1] += 0.5f / (float)r_fb.bloomtextureheight;
                }
                break;
        }
 
-       R_Viewport_InitOrtho(&r_fb.bloomviewport, &identitymatrix, r_refdef.view.x, (r_fb.bloomfbo[0] ? r_fb.bloomtextureheight : vid.height) - r_fb.bloomheight - r_refdef.view.y, r_fb.bloomwidth, r_fb.bloomheight, 0, 0, 1, 1, -10, 100, NULL);
+       R_Viewport_InitOrtho(&r_fb.bloomviewport, &identitymatrix, 0, 0, r_fb.bloomwidth, r_fb.bloomheight, 0, 0, 1, 1, -10, 100, NULL);
 
        if (r_fb.fbo)
                r_refdef.view.clear = true;
@@ -6318,14 +6323,14 @@ static void R_Bloom_MakeTexture(void)
                        xoffset /= (float)r_fb.bloomtexturewidth;
                        yoffset /= (float)r_fb.bloomtextureheight;
                        // compute a texcoord array with the specified x and y offset
-                       r_fb.offsettexcoord2f[0] = xoffset+0;
-                       r_fb.offsettexcoord2f[1] = yoffset+(float)r_fb.bloomheight / (float)r_fb.bloomtextureheight;
-                       r_fb.offsettexcoord2f[2] = xoffset+(float)r_fb.bloomwidth / (float)r_fb.bloomtexturewidth;
-                       r_fb.offsettexcoord2f[3] = yoffset+(float)r_fb.bloomheight / (float)r_fb.bloomtextureheight;
-                       r_fb.offsettexcoord2f[4] = xoffset+(float)r_fb.bloomwidth / (float)r_fb.bloomtexturewidth;
-                       r_fb.offsettexcoord2f[5] = yoffset+0;
-                       r_fb.offsettexcoord2f[6] = xoffset+0;
-                       r_fb.offsettexcoord2f[7] = yoffset+0;
+                       r_fb.offsettexcoord2f[0] = xoffset+r_fb.bloomtexcoord2f[0];
+                       r_fb.offsettexcoord2f[1] = yoffset+r_fb.bloomtexcoord2f[1];
+                       r_fb.offsettexcoord2f[2] = xoffset+r_fb.bloomtexcoord2f[2];
+                       r_fb.offsettexcoord2f[3] = yoffset+r_fb.bloomtexcoord2f[3];
+                       r_fb.offsettexcoord2f[4] = xoffset+r_fb.bloomtexcoord2f[4];
+                       r_fb.offsettexcoord2f[5] = yoffset+r_fb.bloomtexcoord2f[5];
+                       r_fb.offsettexcoord2f[6] = xoffset+r_fb.bloomtexcoord2f[6];
+                       r_fb.offsettexcoord2f[7] = yoffset+r_fb.bloomtexcoord2f[7];
                        // this r value looks like a 'dot' particle, fading sharply to
                        // black at the edges
                        // (probably not realistic but looks good enough)

diff --git a/vid_shared.c b/vid_shared.c
index 7bea96c259f4c46167aaf7334f27c0ae3b289b5e..0859617f5f906c45e51aa189785376f40785f298 100644 (file)
--- a/vid_shared.c
+++ b/vid_shared.c
@@ -2041,7 +2041,7 @@ void VID_Soft_SharedSetup(void)
        vid.support.arb_vertex_buffer_object = true;
        vid.support.ext_blend_subtract = true;
        vid.support.ext_draw_range_elements = true;
-       vid.support.ext_framebuffer_object = false; // FIXME actually dpsoftrast has code for this, but everything is downside up then
+       vid.support.ext_framebuffer_object = true;
        vid.support.ext_texture_3d = true;
        //vid.support.ext_texture_compression_s3tc = true;
        vid.support.ext_texture_filter_anisotropic = true;