scanline.cpp revision 3225321119408735f16b72b539c9fb7d80683552 
1/**
2 **
3 ** Copyright 2010, The Android Open Source Project
4 **
5 ** Licensed under the Apache License, Version 2.0 (the "License");
6 ** you may not use this file except in compliance with the License.
7 ** You may obtain a copy of the License at
8 **
9 **     http://www.apache.org/licenses/LICENSE-2.0
10 **
11 ** Unless required by applicable law or agreed to in writing, software
12 ** distributed under the License is distributed on an "AS IS" BASIS,
13 ** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 ** See the License for the specific language governing permissions and
15 ** limitations under the License.
16 */
17
18#include <assert.h>
19#include <stdio.h>
20#include <string.h>
21
22#include "src/pixelflinger2/pixelflinger2.h"
23#include "src/pixelflinger2/texture.h"
24#include "src/mesa/main/mtypes.h"
25
26#if !USE_LLVM_SCANLINE
27
28static void Saturate(Vec4<BlendComp_t> * color)
29{
30   color->r = MIN2(MAX2(color->r, 0), 255);
31   color->g = MIN2(MAX2(color->g, 0), 255);
32   color->b = MIN2(MAX2(color->b, 0), 255);
33   color->a = MIN2(MAX2(color->a, 0), 255);
34}
35
36static inline void RGBAIntToRGBAIntx4(unsigned rgba, Vec4<BlendComp_t> * color) __attribute__((always_inline));
37static inline void RGBAIntToRGBAIntx4(unsigned rgba, Vec4<BlendComp_t> * color)
38{
39   color->r = rgba & 0xff;
40   color->g = (rgba >>= 8) & 0xff;
41   color->b = (rgba >>= 8) & 0xff;
42   color->a = (rgba >>= 8);
43}
44
45static inline void RGBAFloatx4ToRGBAIntx4(Vector4 * v, Vec4<BlendComp_t> * color)
46{
47   color->r = v->r * 255;
48   color->g = v->g * 255;
49   color->b = v->b * 255;
50   color->a = v->a * 255;
51}
52
53static inline unsigned RGBAIntx4ToRGBAInt(const Vec4<BlendComp_t> * color);
54static inline unsigned RGBAIntx4ToRGBAInt(const Vec4<BlendComp_t> * color)
55{
56   return color->r | (color->g << 8) | (color->b << 16) | (color->a << 24);
57}
58
59
60
61//static inline Pixel Vector4ToPixelRGBA(const Vector4 * color) __attribute__((always_inline));
62//static inline Pixel Vector4ToPixelRGBA(const Vector4 * color)
63//{
64//    Pixel pixel;
65//#if defined(__ARM_HAVE_NEON) && USE_NEON
66//    int32x4_t  c = vcvtq_s32_f32(vmulq_n_f32(color->f4, 255.0f));
67//    c = vminq_s32(c, vdupq_n_s32(255));
68//    c = vmaxq_s32(c, vdupq_n_s32(0));
69//    pixel.channels[0] = (unsigned char)vgetq_lane_s32(c, 0);
70//    pixel.channels[1] = (unsigned char)vgetq_lane_s32(c, 1);
71//    pixel.channels[2] = (unsigned char)vgetq_lane_s32(c, 2);
72//    pixel.channels[3] = (unsigned char)vgetq_lane_s32(c, 3);
73//#else
74//    pixel.channels[0] = (unsigned char)MIN2(MAX2((short)(color->r * 255), 0), 255);
75//	pixel.channels[1] = (unsigned char)MIN2(MAX2((short)(color->g * 255), 0), 255);
76//	pixel.channels[2] = (unsigned char)MIN2(MAX2((short)(color->b * 255), 0), 255);
77//	pixel.channels[3] = (unsigned char)MIN2(MAX2((short)(color->a * 255), 0), 255);
78//#endif //#if USE_FIXED_POINT
79//	return pixel;
80//}
81
82template<typename T>
83static inline void BlendFactor(const unsigned mode, T & factor, const T & src,
84                               const T & dst, const T & constant, const T & one,
85                               const T & zero, const BlendComp_t & srcA, const BlendComp_t & dstA,
86                               const BlendComp_t & constantA, const BlendComp_t & sOne) __attribute__((always_inline));
87template<typename T>
88static inline void BlendFactor(const unsigned mode, T & factor, const T & src,
89                               const T & dst, const T & constant, const T & one,
90                               const T & zero, const BlendComp_t & srcA, const BlendComp_t & dstA,
91                               const BlendComp_t & constantA, const BlendComp_t & sOne)
92{
93   switch (mode) {
94   case 0: // GL_ZERO
95      factor = zero;
96      return;
97   case 1: // GL_ONE
98      factor = one;
99      return;
100   case 2: // GL_SRC_COLOR:
101      factor = src;
102      return;
103   case 3: // GL_ONE_MINUS_SRC_COLOR:
104      factor = one;
105      factor -= src;
106      return;
107   case 4: // GL_DST_COLOR:
108      factor = dst;
109      return;
110   case 5: // GL_ONE_MINUS_DST_COLOR:
111      factor = one;
112      factor -= dst;
113      return;
114   case 6: // GL_SRC_ALPHA:
115      factor = srcA;
116      return;
117   case 7: // GL_ONE_MINUS_SRC_ALPHA:
118      factor = sOne - srcA;
119      return;
120   case 8: // GL_DST_ALPHA:
121      factor = dstA;
122      return;
123   case 9: // GL_ONE_MINUS_DST_ALPHA:
124      factor = sOne - dstA;
125      return;
126   case 10: // GL_SRC_ALPHA_SATURATE: // valid only for source color; src alpha = 1
127      factor = MIN2(srcA, sOne - dstA);
128      return;
129   case 11: // GL_CONSTANT_COLOR:
130      factor = constant;
131      return;
132   case 12: // GL_ONE_MINUS_CONSTANT_COLOR:
133      factor = one;
134      factor -= constant;
135      return;
136   case 13: // GL_CONSTANT_ALPHA:
137      factor = constantA;
138      return;
139   case 14: // GL_ONE_MINUS_CONSTANT_ALPHA:
140      factor = sOne - constantA;
141      return;
142   default:
143      assert(0);
144      return;
145   }
146}
147#endif // #if !USE_LLVM_SCANLINE
148
149unsigned char StencilOp(const unsigned op, unsigned char s, const unsigned char ref)
150{
151   switch (op) {
152   case 0: // GL_ZERO
153      return 0;
154   case 1: // GL_KEEP
155      return s;
156   case 2: // GL_REPLACE
157      return ref;
158   case 3: // GL_INCR
159      if (s < 255)
160         return ++s;
161      return s;
162   case 4: // GL_DECR
163      if (s > 0)
164         return --s;
165      return 0;
166   case 5: // GL_INVERT
167      return ~s;
168   case 6: // GL_INCR_WRAP
169      return ++s;
170   case 7: // GL_DECR_WRAP
171      return --s;
172   default:
173      assert(0);
174      return s;
175   }
176}
177
178template <bool StencilTest, bool DepthTest, bool DepthWrite, bool BlendEnable>
179void ScanLine(const GGLInterface * iface, const VertexOutput * v1, const VertexOutput * v2)
180{
181   GGL_GET_CONST_CONTEXT(ctx, iface);
182   //    assert((unsigned)v1->position.y == (unsigned)v2->position.y);
183   //
184   //    assert(GGL_PIXEL_FORMAT_RGBA_8888 == ctx->frameSurface.format);
185   //    assert(GGL_PIXEL_FORMAT_Z_32 == ctx->depthSurface.format);
186   //    assert(ctx->frameSurface.width == ctx->depthSurface.width);
187   //    assert(ctx->frameSurface.height == ctx->depthSurface.height);
188
189   const unsigned int varyingCount = ctx->glCtx->CurrentProgram->VaryingSlots;
190   const unsigned y = v1->position.y, startX = v1->position.x,
191                      endX = v2->position.x;
192
193   //assert(ctx->frameSurface.width > startX && ctx->frameSurface.width > endX);
194   //assert(ctx->frameSurface.height > y);
195
196   unsigned * frame = (unsigned *)ctx->frameSurface.data
197                      + y * ctx->frameSurface.width + startX;
198   const VectorComp_t div = VectorComp_t_CTR(1 / (float)(endX - startX));
199
200   memcpy(ctx->glCtx->CurrentProgram->ValuesVertexOutput, v1, sizeof(*v1));
201   // shader symbols are mapped to gl_shader_program_Values*
202   VertexOutput & vertex(*(VertexOutput*)ctx->glCtx->CurrentProgram->ValuesVertexOutput);
203   VertexOutput vertexDx(*v2);
204
205   vertexDx.position -= v1->position;
206   vertexDx.position *= div;
207   //printf("vertexDx.position.z=%.8g \n", vertexDx.position.z);
208   for (unsigned i = 0; i < varyingCount; i++) {
209      vertexDx.varyings[i] -= v1->varyings[i];
210      vertexDx.varyings[i] *= div;
211   }
212   vertexDx.frontFacingPointCoord -= v1->frontFacingPointCoord;
213   vertexDx.frontFacingPointCoord *= div; // gl_PointCoord, only zw
214   vertexDx.frontFacingPointCoord.y = 0; // gl_FrontFacing not interpolated
215
216#if USE_FORCED_FIXEDPOINT
217   for (unsigned j = 0; j < 4; j++) {
218      for (unsigned i = 0; i < varyingCount; i++) {
219         vertex.varyings[i].i[j] = vertex.varyings[i].f[j] * 65536;
220         vertexDx.varyings[i].i[j] = vertexDx.varyings[i].f[j] * 65536;
221      }
222      vertex.position.i[j] = vertex.position.f[j] * 65536;
223      vertexDx.position.i[j] = vertexDx.position.f[j] * 65536;
224      vertex.frontFacingPointCoord.i[j] = vertex.frontFacingPointCoord.f[j] * 65536;
225   }
226#endif
227
228   int * depth = (int *)ctx->depthSurface.data + y * ctx->frameSurface.width + startX;
229   unsigned char * stencil = (unsigned char *)ctx->stencilSurface.data + y * ctx->frameSurface.width + startX;
230
231#if !USE_LLVM_TEXTURE_SAMPLER
232   extern const GGLContext * textureGGLContext;
233   textureGGLContext = ctx;
234#endif
235
236   // TODO DXL consider inverting gl_FragCoord.y
237
238#if USE_LLVM_SCANLINE
239   typedef void (* ScanLineFunction_t)(VertexOutput * start, VertexOutput * step,
240                                       unsigned * frame, int * depth, unsigned char * stencil,
241                                       GGLContext::ActiveStencilState *, unsigned count);
242
243    ScanLineFunction_t scanLineFunction = (ScanLineFunction_t)
244    ctx->glCtx->CurrentProgram->_LinkedShaders[MESA_SHADER_FRAGMENT]->function;
245   if (endX >= startX) {
246		scanLineFunction(&vertex, &vertexDx, frame, depth, stencil, &ctx->activeStencil, endX - startX + 1);
247   }
248#else
249
250   int z;
251   bool sCmp = true; // default passed, unless failed by stencil test
252   unsigned char s; // masked stored stencil value
253   const unsigned char sMask = ctx->activeStencil.mask;
254   const unsigned char sRef = ctx->activeStencil.ref;
255   const unsigned sFunc = ctx->activeStencil.face ? 0x200 | ctx->backStencil.func :
256                          0x200 | ctx->frontStencil.func;
257   const unsigned ssFail = ctx->activeStencil.face ? ctx->backStencil.sFail :
258                           ctx->frontStencil.sFail;
259   const unsigned sdFail = ctx->activeStencil.face ? ctx->backStencil.dFail :
260                           ctx->frontStencil.dFail;
261   const unsigned sdPass = ctx->activeStencil.face ? ctx->backStencil.dPass :
262                           ctx->frontStencil.dPass;
263
264   for (unsigned x = startX; x <= endX; x++) {
265      //assert(abs((int)(vertex.position.x) - (int)x) < 2);
266      //assert((unsigned)vertex.position.y == y);
267      if (StencilTest) {
268         s = *stencil & sMask;
269         switch (sFunc) {
270         case GL_NEVER:
271            sCmp = false;
272            break;
273         case GL_LESS:
274            sCmp = sRef < s;
275            break;
276         case GL_EQUAL:
277            sCmp = sRef == s;
278            break;
279         case GL_LEQUAL:
280            sCmp = sRef <= s;
281            break;
282         case GL_GREATER:
283            sCmp = sRef > s;
284            break;
285         case GL_NOTEQUAL:
286            sCmp = sRef != s;
287            break;
288         case GL_GEQUAL:
289            sCmp = sRef >= s;
290            break;
291         case GL_ALWAYS:
292            sCmp = true;
293            break;
294         default:
295            assert(0);
296            break;
297         }
298      }
299
300      if (!StencilTest || sCmp) {
301         z = vertex.position.i[2];
302         if (z & 0x80000000)  // negative float has leading 1
303            z ^= 0x7fffffff;  // bigger negative is smaller
304         bool zCmp = true;
305         if (DepthTest)
306         {
307            switch (0x200 | ctx->bufferState.depthFunc) {
308            case GL_NEVER:
309               zCmp = false;
310               break;
311            case GL_LESS:
312               zCmp = z < *depth;
313               break;
314            case GL_EQUAL:
315               zCmp = z == *depth;
316               break;
317            case GL_LEQUAL:
318               zCmp = z <= *depth;
319               break;
320            case GL_GREATER:
321               zCmp = z > *depth;
322               break;
323            case GL_NOTEQUAL:
324               zCmp = z != *depth;
325               break;
326            case GL_GEQUAL:
327               zCmp = z >= *depth;
328               break;
329            case GL_ALWAYS:
330               zCmp = true;
331               break;
332            default:
333               assert(0);
334               break;
335            }
336         }
337         if (!DepthTest || zCmp) {
338            float * varying = (float *)ctx->glCtx->CurrentProgram->ValuesVertexOutput;
339
340            assert((void *)&(vertex.varyings[0]) == &(varying[2 * 4]));
341            ctx->glCtx->CurrentProgram->_LinkedShaders[MESA_SHADER_FRAGMENT]->function();
342            if (BlendEnable) {
343               BlendComp_t sOne = 255, sZero = 0;
344               Vec4<BlendComp_t> one = sOne, zero = sZero;
345
346               Vec4<BlendComp_t> src;
347//                    if (outputRegDesc.IsInt32Color())
348//                        RGBAIntToRGBAIntx4(vertex.fragColor[0].u[0], &src);
349//                    else if (outputRegDesc.IsVectorType(Float))
350               RGBAFloatx4ToRGBAIntx4(&vertex.fragColor[0], &src);
351//                    else if (outputRegDesc.IsVectorType(Fixed8))
352//                    {
353//                        src.u[0] = vertex.fragColor[0].u[0];
354//                        src.u[1] = vertex.fragColor[0].u[1];
355//                        src.u[2] = vertex.fragColor[0].u[2];
356//                        src.u[3] = vertex.fragColor[0].u[3];
357//                    }
358//                    else
359//                        assert(0);
360
361               Vec4<BlendComp_t> dst;
362               unsigned dc = *frame;
363               dst.r = dc & 255;
364               dst.g = (dc >>= 8) & 255;
365               dst.b = (dc >>= 8) & 255;
366               dst.a = (dc >>= 8) & 255;
367
368               Vec4<BlendComp_t> sf, df;
369
370               BlendFactor(ctx->blendState.scf, sf, src, dst,
371                           ctx->blendState.color, one, zero, src.a, dst.a,
372                           ctx->blendState.color.a, sOne);
373               if (ctx->blendState.scf != ctx->blendState.saf)
374                  BlendFactor(ctx->blendState.saf, sf.a, src.a, dst.a,
375                              ctx->blendState.color.a, sOne, sZero, src.a, dst.a,
376                              ctx->blendState.color.a, sOne);
377               BlendFactor(ctx->blendState.dcf, df, src, dst,
378                           ctx->blendState.color, one, zero, src.a, dst.a,
379                           ctx->blendState.color.a, sOne);
380               if (ctx->blendState.dcf != ctx->blendState.daf)
381                  BlendFactor(ctx->blendState.daf, df.a, src.a, dst.a,
382                              ctx->blendState.color.a, sOne, sZero, src.a, dst.a,
383                              ctx->blendState.color.a, sOne);
384
385               Vec4<BlendComp_t> sfs(sf), dfs(df);
386               sfs.LShr(7);
387               sf += sfs;
388               dfs.LShr(7);
389               df += dfs;
390
391               src *= sf;
392               dst *= df;
393               Vec4<BlendComp_t> res(src);
394               switch (ctx->blendState.ce + GL_FUNC_ADD) {
395               case GL_FUNC_ADD:
396                  res += dst;
397                  break;
398               case GL_FUNC_SUBTRACT:
399                  res -= dst;
400                  break;
401               case GL_FUNC_REVERSE_SUBTRACT:
402                  res = dst;
403                  res -= src;
404                  break;
405               default:
406                  assert(0);
407                  break;
408               }
409               if (ctx->blendState.ce != ctx->blendState.ae)
410                  switch (ctx->blendState.ce + GL_FUNC_ADD) {
411                  case GL_FUNC_ADD:
412                     res.a = src.a + dst.a;
413                     break;
414                  case GL_FUNC_SUBTRACT:
415                     res.a = src.a - dst.a;
416                     break;
417                  case GL_FUNC_REVERSE_SUBTRACT:
418                     res.a = dst.a - src.a;
419                     break;
420                  default:
421                     assert(0);
422                     break;
423                  }
424
425               res.AShr(8);
426               Saturate(&res);
427               *frame = RGBAIntx4ToRGBAInt(&res);
428            } else {
429//                    if (outputRegDesc.IsInt32Color())
430//                        *frame = vertex.fragColor[0].u[0];
431//                    else if (outputRegDesc.IsVectorType(Float))
432               {
433                  Vec4<BlendComp_t> src;
434                  RGBAFloatx4ToRGBAIntx4(&vertex.fragColor[0], &src);
435                  Saturate(&src);
436                  *frame = RGBAIntx4ToRGBAInt(&src);
437               }
438//                    else if (outputRegDesc.IsVectorType(Fixed16))
439//                    {
440//                        Vec4<BlendComp_t> & src = (Vec4<BlendComp_t> &)vertex.fragColor[0];
441//                        src.r = (src.r * 255 >> 16);
442//                        src.g = (src.g * 255 >> 16);
443//                        src.b = (src.b * 255 >> 16);
444//                        src.a = (src.a * 255 >> 16);
445//                        Saturate(&src);
446//                        *frame = RGBAIntx4ToRGBAInt(&src);
447//                    }
448//                    else if (outputRegDesc.IsVectorType(Fixed8))
449//                    {
450//                        Vec4<BlendComp_t> & src = (Vec4<BlendComp_t> &)vertex.fragColor[0];
451//                        Saturate(&src);
452//                        *frame = RGBAIntx4ToRGBAInt(&src);
453//                    }
454//                    else
455//                        assert(0);
456            }
457
458            if (DepthWrite)
459               *depth = z;
460            if (StencilTest)
461               *stencil = StencilOp(sdPass, s, sRef);
462         } else if (StencilTest)
463            *stencil = StencilOp(sdFail, s, sRef);
464      } else if (StencilTest)
465         *stencil = StencilOp(ssFail, s, sRef);
466
467      frame++;
468      depth++;
469      stencil++;
470
471#if USE_FORCED_FIXEDPOINT
472      for (unsigned j = 0; j < 4; j++) {
473         if (ctx->glCtx->Shader.CurrentProgram->FragmentProgram->UsesFragCoord)
474            vertex.position.i[j] += vertexDx.position.i[j];
475         for (unsigned i = 0; i < varyingCount; i++)
476            vertex.varyings[i].i[j] += vertexDx.varyings[i].i[j];
477      }
478      vertex.position.i[2] += vertexDx.position.i[2];
479      if (ctx->glCtx->Shader.CurrentProgram->FragmentProgram->UsesPointCoord) {
480         vertex.frontFacingPointCoord.i[2] = vertexDx.frontFacingPointCoord.i[2];
481         vertex.frontFacingPointCoord.i[3] = vertexDx.frontFacingPointCoord.i[3];
482      }
483#else
484        if (ctx->glCtx->CurrentProgram->UsesFragCoord)
485   vertex.position += vertexDx.position;
486        else if (ctx->bufferState.depthTest)
487   vertex.position.z += vertexDx.position.z;
488
489   for (unsigned i = 0; i < varyingCount; i++)
490      vertex.varyings[i] += vertexDx.varyings[i];
491   if (ctx->glCtx->CurrentProgram->UsesPointCoord)
492   {
493      vertex.frontFacingPointCoord.z += vertexDx.frontFacingPointCoord.z;
494      vertex.frontFacingPointCoord.w += vertexDx.frontFacingPointCoord.w;
495   }
496#endif // #if USE_FORCED_FIXEDPOINT
497   }
498
499#endif // #if USE_LLVM_SCANLINE
500
501#if !USE_LLVM_TEXTURE_SAMPLER
502   textureGGLContext = NULL;
503#endif
504}
505
506static void PickScanLine(GGLInterface * iface)
507{
508   GGL_GET_CONTEXT(ctx, iface);
509
510   ctx->interface.ScanLine = NULL;
511   if (ctx->bufferState.stencilTest) {
512      if (ctx->bufferState.depthTest) {
513         if (ctx->blendState.enable)
514            ctx->interface.ScanLine = ScanLine<true, true, true, true>;
515         else
516            ctx->interface.ScanLine = ScanLine<true, true, true, false>;
517      } else {
518         if (ctx->blendState.enable)
519            ctx->interface.ScanLine = ScanLine<true, false, false, true>;
520         else
521            ctx->interface.ScanLine = ScanLine<true, false, false, false>;
522      }
523   } else {
524      if (ctx->bufferState.depthTest) {
525         if (ctx->blendState.enable)
526            ctx->interface.ScanLine = ScanLine<false, true, true, true>;
527         else
528            ctx->interface.ScanLine = ScanLine<false, true, true, false>;
529      } else {
530         if (ctx->blendState.enable)
531            ctx->interface.ScanLine = ScanLine<false, false, false, true>;
532         else
533            ctx->interface.ScanLine = ScanLine<false, false, false, false>;
534      }
535   }
536
537   assert(ctx->interface.ScanLine);
538}
539
540void InitializeScanLineFunctions(GGLInterface * iface)
541{
542   GGL_GET_CONTEXT(ctx, iface);
543   ctx->PickScanLine = PickScanLine;
544}
545