1/*
2 * Copyright 2019 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25
26#include "ac_llvm_cull.h"
27
28#include <llvm-c/Core.h>
29
30struct ac_position_w_info {
31   /* If a primitive intersects the W=0 plane, it causes a reflection
32    * of the determinant used for face culling. Every vertex behind
33    * the W=0 plane negates the determinant, so having 2 vertices behind
34    * the plane has no effect. This is i1 true if the determinant should be
35    * negated.
36    */
37   LLVMValueRef w_reflection;
38
39   /* If we simplify the "-w <= p <= w" view culling equation, we get
40    * "-w <= w", which can't be satisfied when w is negative.
41    * In perspective projection, a negative W means that the primitive
42    * is behind the viewer, but the equation is independent of the type
43    * of projection.
44    *
45    * w_accepted is false when all W are negative and therefore
46    * the primitive is invisible.
47    */
48   LLVMValueRef w_accepted;
49
50   /* The bounding box culling doesn't work and should be skipped when this is true. */
51   LLVMValueRef any_w_negative;
52};
53
54static void ac_analyze_position_w(struct ac_llvm_context *ctx, LLVMValueRef pos[3][4],
55                                  struct ac_position_w_info *w, unsigned num_vertices)
56{
57   LLVMBuilderRef builder = ctx->builder;
58   LLVMValueRef all_w_negative = ctx->i1true;
59
60   w->w_reflection = ctx->i1false;
61   w->any_w_negative = ctx->i1false;
62
63   for (unsigned i = 0; i < num_vertices; i++) {
64      LLVMValueRef neg_w;
65
66      neg_w = LLVMBuildFCmp(builder, LLVMRealOLT, pos[i][3], ctx->f32_0, "");
67      /* If neg_w is true, negate w_reflection. */
68      w->w_reflection = LLVMBuildXor(builder, w->w_reflection, neg_w, "");
69      w->any_w_negative = LLVMBuildOr(builder, w->any_w_negative, neg_w, "");
70      all_w_negative = LLVMBuildAnd(builder, all_w_negative, neg_w, "");
71   }
72   w->w_accepted = LLVMBuildNot(builder, all_w_negative, "");
73}
74
75/* Perform front/back face culling and return true if the primitive is accepted. */
76static LLVMValueRef ac_cull_face(struct ac_llvm_context *ctx, LLVMValueRef pos[3][4],
77                                 struct ac_position_w_info *w, bool cull_front, bool cull_back,
78                                 bool cull_zero_area)
79{
80   LLVMBuilderRef builder = ctx->builder;
81
82   if (cull_front && cull_back)
83      return ctx->i1false;
84
85   if (!cull_front && !cull_back && !cull_zero_area)
86      return ctx->i1true;
87
88   /* Front/back face culling. Also if the determinant == 0, the triangle
89    * area is 0.
90    */
91   LLVMValueRef det_t0 = LLVMBuildFSub(builder, pos[2][0], pos[0][0], "");
92   LLVMValueRef det_t1 = LLVMBuildFSub(builder, pos[1][1], pos[0][1], "");
93   LLVMValueRef det_t2 = LLVMBuildFSub(builder, pos[0][0], pos[1][0], "");
94   LLVMValueRef det_t3 = LLVMBuildFSub(builder, pos[0][1], pos[2][1], "");
95   LLVMValueRef det_p0 = LLVMBuildFMul(builder, det_t0, det_t1, "");
96   LLVMValueRef det_p1 = LLVMBuildFMul(builder, det_t2, det_t3, "");
97   LLVMValueRef det = LLVMBuildFSub(builder, det_p0, det_p1, "");
98
99   /* Negative W negates the determinant. */
100   det = LLVMBuildSelect(builder, w->w_reflection, LLVMBuildFNeg(builder, det, ""), det, "");
101
102   LLVMValueRef accepted = NULL;
103   if (cull_front) {
104      LLVMRealPredicate cond = cull_zero_area ? LLVMRealOGT : LLVMRealOGE;
105      accepted = LLVMBuildFCmp(builder, cond, det, ctx->f32_0, "");
106   } else if (cull_back) {
107      LLVMRealPredicate cond = cull_zero_area ? LLVMRealOLT : LLVMRealOLE;
108      accepted = LLVMBuildFCmp(builder, cond, det, ctx->f32_0, "");
109   } else if (cull_zero_area) {
110      accepted = LLVMBuildFCmp(builder, LLVMRealONE, det, ctx->f32_0, "");
111   }
112   return accepted;
113}
114
115/* Perform view culling and small primitive elimination and return true
116 * if the primitive is accepted and initially_accepted == true. */
117static void cull_bbox(struct ac_llvm_context *ctx, LLVMValueRef pos[3][4],
118                      LLVMValueRef initially_accepted, struct ac_position_w_info *w,
119                      LLVMValueRef vp_scale[2], LLVMValueRef vp_translate[2],
120                      LLVMValueRef small_prim_precision, struct ac_cull_options *options,
121                      ac_cull_accept_func accept_func, void *userdata)
122{
123   LLVMBuilderRef builder = ctx->builder;
124
125   if (!options->cull_view_xy && !options->cull_view_near_z && !options->cull_view_far_z &&
126       !options->cull_small_prims) {
127      if (accept_func)
128         accept_func(ctx, initially_accepted, userdata);
129      return;
130   }
131
132   ac_build_ifcc(ctx, initially_accepted, 10000000);
133   {
134      LLVMValueRef bbox_min[3], bbox_max[3];
135      LLVMValueRef accepted = ctx->i1true;
136
137      /* Compute the primitive bounding box for easy culling. */
138      for (unsigned chan = 0; chan < (options->cull_view_near_z ||
139                                      options->cull_view_far_z ? 3 : 2); chan++) {
140         assert(options->num_vertices >= 2);
141         bbox_min[chan] = ac_build_fmin(ctx, pos[0][chan], pos[1][chan]);
142         bbox_max[chan] = ac_build_fmax(ctx, pos[0][chan], pos[1][chan]);
143
144         if (options->num_vertices == 3) {
145            bbox_min[chan] = ac_build_fmin(ctx, bbox_min[chan], pos[2][chan]);
146            bbox_max[chan] = ac_build_fmax(ctx, bbox_max[chan], pos[2][chan]);
147         }
148      }
149
150      /* View culling. */
151      if (options->cull_view_xy || options->cull_view_near_z || options->cull_view_far_z) {
152         for (unsigned chan = 0; chan < 3; chan++) {
153            LLVMValueRef visible;
154
155            if ((options->cull_view_xy && chan <= 1) || (options->cull_view_near_z && chan == 2)) {
156               float t = chan == 2 && options->use_halfz_clip_space ? 0 : -1;
157               visible = LLVMBuildFCmp(builder, LLVMRealOGE, bbox_max[chan],
158                                       LLVMConstReal(ctx->f32, t), "");
159               accepted = LLVMBuildAnd(builder, accepted, visible, "");
160            }
161
162            if ((options->cull_view_xy && chan <= 1) || (options->cull_view_far_z && chan == 2)) {
163               visible = LLVMBuildFCmp(builder, LLVMRealOLE, bbox_min[chan], ctx->f32_1, "");
164               accepted = LLVMBuildAnd(builder, accepted, visible, "");
165            }
166         }
167      }
168
169      /* Small primitive elimination. */
170      if (options->cull_small_prims) {
171         /* Assuming a sample position at (0.5, 0.5), if we round
172          * the bounding box min/max extents and the results of
173          * the rounding are equal in either the X or Y direction,
174          * the bounding box does not intersect the sample.
175          *
176          * See these GDC slides for pictures:
177          * https://frostbite-wp-prd.s3.amazonaws.com/wp-content/uploads/2016/03/29204330/GDC_2016_Compute.pdf
178          */
179         LLVMValueRef min, max, not_equal[2], visible;
180
181         for (unsigned chan = 0; chan < 2; chan++) {
182            /* Convert the position to screen-space coordinates. */
183            min = ac_build_fmad(ctx, bbox_min[chan], vp_scale[chan], vp_translate[chan]);
184            max = ac_build_fmad(ctx, bbox_max[chan], vp_scale[chan], vp_translate[chan]);
185            /* Scale the bounding box according to the precision of
186             * the rasterizer and the number of MSAA samples. */
187            min = LLVMBuildFSub(builder, min, small_prim_precision, "");
188            max = LLVMBuildFAdd(builder, max, small_prim_precision, "");
189
190            /* Determine if the bbox intersects the sample point.
191             * It also works for MSAA, but vp_scale, vp_translate,
192             * and small_prim_precision are computed differently.
193             */
194            min = ac_build_round(ctx, min);
195            max = ac_build_round(ctx, max);
196            not_equal[chan] = LLVMBuildFCmp(builder, LLVMRealONE, min, max, "");
197         }
198         visible = LLVMBuildAnd(builder, not_equal[0], not_equal[1], "");
199         accepted = LLVMBuildAnd(builder, accepted, visible, "");
200      }
201
202      /* Disregard the bounding box culling if any W is negative because the code
203       * doesn't work with that.
204       */
205      accepted = LLVMBuildOr(builder, accepted, w->any_w_negative, "");
206
207      if (accept_func)
208         accept_func(ctx, accepted, userdata);
209   }
210   ac_build_endif(ctx, 10000000);
211}
212
213/**
214 * Return i1 true if the primitive is accepted (not culled).
215 *
216 * \param pos                   Vertex positions 3x vec4
217 * \param initially_accepted    AND'ed with the result. Some computations can be
218 *                              skipped if this is false.
219 * \param vp_scale              Viewport scale XY.
220 *                              For MSAA, multiply them by the number of samples.
221 * \param vp_translate          Viewport translation XY.
222 *                              For MSAA, multiply them by the number of samples.
223 * \param small_prim_precision  Precision of small primitive culling. This should
224 *                              be the same as or greater than the precision of
225 *                              the rasterizer. Set to num_samples / 2^subpixel_bits.
226 *                              subpixel_bits are defined by the quantization mode.
227 * \param options               See ac_cull_options.
228 * \param accept_func           Callback invoked in the inner-most branch where the primitive is accepted.
229 */
230void ac_cull_primitive(struct ac_llvm_context *ctx, LLVMValueRef pos[3][4],
231                       LLVMValueRef initially_accepted, LLVMValueRef vp_scale[2],
232                       LLVMValueRef vp_translate[2], LLVMValueRef small_prim_precision,
233                       struct ac_cull_options *options, ac_cull_accept_func accept_func,
234                       void *userdata)
235{
236   struct ac_position_w_info w;
237   ac_analyze_position_w(ctx, pos, &w, options->num_vertices);
238
239   /* W culling. */
240   LLVMValueRef accepted = options->cull_w ? w.w_accepted : ctx->i1true;
241   accepted = LLVMBuildAnd(ctx->builder, accepted, initially_accepted, "");
242
243   /* Face culling. */
244   accepted = LLVMBuildAnd(
245      ctx->builder, accepted,
246      ac_cull_face(ctx, pos, &w, options->cull_front, options->cull_back, options->cull_zero_area),
247      "");
248
249   /* View culling and small primitive elimination. */
250   cull_bbox(ctx, pos, accepted, &w, vp_scale, vp_translate, small_prim_precision, options,
251             accept_func, userdata);
252}
253