1 /*
2  * Copyright 2017 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * on the rights to use, copy, modify, merge, publish, distribute, sub
9  * license, and/or sell copies of the Software, and to permit persons to whom
10  * the Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
23  */
24 
25 /* This file handles register programming of primitive binning. */
26 
27 #include "si_build_pm4.h"
28 #include "sid.h"
29 
30 struct uvec2 {
31    unsigned x, y;
32 };
33 
34 struct si_bin_size_map {
35    unsigned start;
36    unsigned bin_size_x;
37    unsigned bin_size_y;
38 };
39 
40 typedef struct si_bin_size_map si_bin_size_subtable[3][10];
41 
42 /* Find the bin size where sum is >= table[i].start and < table[i + 1].start. */
si_find_bin_size(struct si_screen * sscreen,const si_bin_size_subtable table[],unsigned sum)43 static struct uvec2 si_find_bin_size(struct si_screen *sscreen, const si_bin_size_subtable table[],
44                                      unsigned sum)
45 {
46    unsigned log_num_rb_per_se =
47       util_logbase2_ceil(sscreen->info.num_render_backends / sscreen->info.max_se);
48    unsigned log_num_se = util_logbase2_ceil(sscreen->info.max_se);
49    unsigned i;
50 
51    /* Get the chip-specific subtable. */
52    const struct si_bin_size_map *subtable = &table[log_num_rb_per_se][log_num_se][0];
53 
54    for (i = 0; subtable[i].bin_size_x != 0; i++) {
55       if (sum >= subtable[i].start && sum < subtable[i + 1].start)
56          break;
57    }
58 
59    struct uvec2 size = {subtable[i].bin_size_x, subtable[i].bin_size_y};
60    return size;
61 }
62 
si_get_color_bin_size(struct si_context * sctx,unsigned cb_target_enabled_4bit)63 static struct uvec2 si_get_color_bin_size(struct si_context *sctx, unsigned cb_target_enabled_4bit)
64 {
65    unsigned num_fragments = sctx->framebuffer.nr_color_samples;
66    unsigned sum = 0;
67 
68    /* Compute the sum of all Bpp. */
69    for (unsigned i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
70       if (!(cb_target_enabled_4bit & (0xf << (i * 4))))
71          continue;
72 
73       struct si_texture *tex = (struct si_texture *)sctx->framebuffer.state.cbufs[i]->texture;
74       sum += tex->surface.bpe;
75    }
76 
77    /* Multiply the sum by some function of the number of samples. */
78    if (num_fragments >= 2) {
79       if (si_get_ps_iter_samples(sctx) >= 2)
80          sum *= num_fragments;
81       else
82          sum *= 2;
83    }
84 
85    static const si_bin_size_subtable table[] = {
86       {
87          /* One RB / SE */
88          {
89             /* One shader engine */
90             {0, 128, 128},
91             {1, 64, 128},
92             {2, 32, 128},
93             {3, 16, 128},
94             {17, 0, 0},
95          },
96          {
97             /* Two shader engines */
98             {0, 128, 128},
99             {2, 64, 128},
100             {3, 32, 128},
101             {5, 16, 128},
102             {17, 0, 0},
103          },
104          {
105             /* Four shader engines */
106             {0, 128, 128},
107             {3, 64, 128},
108             {5, 16, 128},
109             {17, 0, 0},
110          },
111       },
112       {
113          /* Two RB / SE */
114          {
115             /* One shader engine */
116             {0, 128, 128},
117             {2, 64, 128},
118             {3, 32, 128},
119             {9, 16, 128},
120             {33, 0, 0},
121          },
122          {
123             /* Two shader engines */
124             {0, 128, 128},
125             {3, 64, 128},
126             {5, 32, 128},
127             {9, 16, 128},
128             {33, 0, 0},
129          },
130          {
131             /* Four shader engines */
132             {0, 256, 256},
133             {2, 128, 256},
134             {3, 128, 128},
135             {5, 64, 128},
136             {9, 16, 128},
137             {33, 0, 0},
138          },
139       },
140       {
141          /* Four RB / SE */
142          {
143             /* One shader engine */
144             {0, 128, 256},
145             {2, 128, 128},
146             {3, 64, 128},
147             {5, 32, 128},
148             {9, 16, 128},
149             {17, 0, 0},
150          },
151          {
152             /* Two shader engines */
153             {0, 256, 256},
154             {2, 128, 256},
155             {3, 128, 128},
156             {5, 64, 128},
157             {9, 32, 128},
158             {17, 16, 128},
159             {33, 0, 0},
160          },
161          {
162             /* Four shader engines */
163             {0, 256, 512},
164             {2, 128, 512},
165             {3, 64, 512},
166             {5, 32, 512},
167             {9, 32, 256},
168             {17, 32, 128},
169             {33, 0, 0},
170          },
171       },
172    };
173 
174    return si_find_bin_size(sctx->screen, table, sum);
175 }
176 
si_get_depth_bin_size(struct si_context * sctx)177 static struct uvec2 si_get_depth_bin_size(struct si_context *sctx)
178 {
179    struct si_state_dsa *dsa = sctx->queued.named.dsa;
180 
181    if (!sctx->framebuffer.state.zsbuf || (!dsa->depth_enabled && !dsa->stencil_enabled)) {
182       /* Return the max size. */
183       struct uvec2 size = {512, 512};
184       return size;
185    }
186 
187    struct si_texture *tex = (struct si_texture *)sctx->framebuffer.state.zsbuf->texture;
188    unsigned depth_coeff = dsa->depth_enabled ? 5 : 0;
189    unsigned stencil_coeff = tex->surface.has_stencil && dsa->stencil_enabled ? 1 : 0;
190    unsigned sum = 4 * (depth_coeff + stencil_coeff) * MAX2(tex->buffer.b.b.nr_samples, 1);
191 
192    static const si_bin_size_subtable table[] = {
193       {
194          // One RB / SE
195          {
196             // One shader engine
197             {0, 64, 512},
198             {2, 64, 256},
199             {4, 64, 128},
200             {7, 32, 128},
201             {13, 16, 128},
202             {49, 0, 0},
203          },
204          {
205             // Two shader engines
206             {0, 128, 512},
207             {2, 64, 512},
208             {4, 64, 256},
209             {7, 64, 128},
210             {13, 32, 128},
211             {25, 16, 128},
212             {49, 0, 0},
213          },
214          {
215             // Four shader engines
216             {0, 256, 512},
217             {2, 128, 512},
218             {4, 64, 512},
219             {7, 64, 256},
220             {13, 64, 128},
221             {25, 16, 128},
222             {49, 0, 0},
223          },
224       },
225       {
226          // Two RB / SE
227          {
228             // One shader engine
229             {0, 128, 512},
230             {2, 64, 512},
231             {4, 64, 256},
232             {7, 64, 128},
233             {13, 32, 128},
234             {25, 16, 128},
235             {97, 0, 0},
236          },
237          {
238             // Two shader engines
239             {0, 256, 512},
240             {2, 128, 512},
241             {4, 64, 512},
242             {7, 64, 256},
243             {13, 64, 128},
244             {25, 32, 128},
245             {49, 16, 128},
246             {97, 0, 0},
247          },
248          {
249             // Four shader engines
250             {0, 512, 512},
251             {2, 256, 512},
252             {4, 128, 512},
253             {7, 64, 512},
254             {13, 64, 256},
255             {25, 64, 128},
256             {49, 16, 128},
257             {97, 0, 0},
258          },
259       },
260       {
261          // Four RB / SE
262          {
263             // One shader engine
264             {0, 256, 512},
265             {2, 128, 512},
266             {4, 64, 512},
267             {7, 64, 256},
268             {13, 64, 128},
269             {25, 32, 128},
270             {49, 16, 128},
271             {193, 0, 0},
272          },
273          {
274             // Two shader engines
275             {0, 512, 512},
276             {2, 256, 512},
277             {4, 128, 512},
278             {7, 64, 512},
279             {13, 64, 256},
280             {25, 64, 128},
281             {49, 32, 128},
282             {97, 16, 128},
283             {193, 0, 0},
284          },
285          {
286             // Four shader engines
287             {0, 512, 512},
288             {4, 256, 512},
289             {7, 128, 512},
290             {13, 64, 512},
291             {25, 32, 512},
292             {49, 32, 256},
293             {97, 16, 128},
294             {193, 0, 0},
295          },
296       },
297    };
298 
299    return si_find_bin_size(sctx->screen, table, sum);
300 }
301 
gfx10_get_bin_sizes(struct si_context * sctx,unsigned cb_target_enabled_4bit,struct uvec2 * color_bin_size,struct uvec2 * depth_bin_size)302 static void gfx10_get_bin_sizes(struct si_context *sctx, unsigned cb_target_enabled_4bit,
303                                 struct uvec2 *color_bin_size, struct uvec2 *depth_bin_size)
304 {
305    const unsigned ZsTagSize = 64;
306    const unsigned ZsNumTags = 312;
307    const unsigned CcTagSize = 1024;
308    const unsigned CcReadTags = 31;
309    const unsigned FcTagSize = 256;
310    const unsigned FcReadTags = 44;
311 
312    const unsigned num_rbs = sctx->screen->info.num_render_backends;
313    const unsigned num_pipes = MAX2(num_rbs, sctx->screen->info.num_sdp_interfaces);
314 
315    const unsigned depthBinSizeTagPart =
316       ((ZsNumTags * num_rbs / num_pipes) * (ZsTagSize * num_pipes));
317    const unsigned colorBinSizeTagPart =
318       ((CcReadTags * num_rbs / num_pipes) * (CcTagSize * num_pipes));
319    const unsigned fmaskBinSizeTagPart =
320       ((FcReadTags * num_rbs / num_pipes) * (FcTagSize * num_pipes));
321 
322    const unsigned minBinSizeX = 128;
323    const unsigned minBinSizeY = 64;
324 
325    const unsigned num_fragments = sctx->framebuffer.nr_color_samples;
326    const unsigned num_samples = sctx->framebuffer.nr_samples;
327    const bool ps_iter_sample = si_get_ps_iter_samples(sctx) >= 2;
328 
329    /* Calculate cColor and cFmask(if applicable) */
330    unsigned cColor = 0;
331    unsigned cFmask = 0;
332    bool has_fmask = false;
333 
334    for (unsigned i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
335       if (!sctx->framebuffer.state.cbufs[i])
336          continue;
337 
338       struct si_texture *tex = (struct si_texture *)sctx->framebuffer.state.cbufs[i]->texture;
339       const unsigned mmrt = num_fragments == 1 ? 1 : (ps_iter_sample ? num_fragments : 2);
340 
341       cColor += tex->surface.bpe * mmrt;
342       if (num_samples >= 2 /* if FMASK is bound */) {
343          const unsigned fragmentsLog2 = util_logbase2(num_fragments);
344          const unsigned samplesLog2 = util_logbase2(num_samples);
345 
346          static const unsigned cFmaskMrt[4 /* fragments */][5 /* samples */] = {
347             {0, 1, 1, 1, 2}, /* fragments = 1 */
348             {0, 1, 1, 2, 4}, /* fragments = 2 */
349             {0, 1, 1, 4, 8}, /* fragments = 4 */
350             {0, 1, 2, 4, 8}  /* fragments = 8 */
351          };
352          cFmask += cFmaskMrt[fragmentsLog2][samplesLog2];
353          has_fmask = true;
354       }
355    }
356    cColor = MAX2(cColor, 1u);
357 
358    const unsigned colorLog2Pixels = util_logbase2(colorBinSizeTagPart / cColor);
359    const unsigned colorBinSizeX = 1 << ((colorLog2Pixels + 1) / 2); /* round up width */
360    const unsigned colorBinSizeY = 1 << (colorLog2Pixels / 2);       /* round down height */
361 
362    unsigned binSizeX = colorBinSizeX;
363    unsigned binSizeY = colorBinSizeY;
364 
365    if (has_fmask) {
366       cFmask = MAX2(cFmask, 1u);
367 
368       const unsigned fmaskLog2Pixels = util_logbase2(fmaskBinSizeTagPart / cFmask);
369       const unsigned fmaskBinSizeX = 1 << ((fmaskLog2Pixels + 1) / 2); /* round up width */
370       const unsigned fmaskBinSizeY = 1 << (fmaskLog2Pixels / 2);       /* round down height */
371 
372       /* use the smaller of the Color vs. Fmask bin sizes */
373       if (fmaskLog2Pixels < colorLog2Pixels) {
374          binSizeX = fmaskBinSizeX;
375          binSizeY = fmaskBinSizeY;
376       }
377    }
378 
379    /* Return size adjusted for minimum bin size */
380    color_bin_size->x = MAX2(binSizeX, minBinSizeX);
381    color_bin_size->y = MAX2(binSizeY, minBinSizeY);
382 
383    if (!sctx->framebuffer.state.zsbuf) {
384       /* Set to max sizes when no depth buffer is bound. */
385       depth_bin_size->x = 512;
386       depth_bin_size->y = 512;
387    } else {
388       struct si_texture *zstex = (struct si_texture *)sctx->framebuffer.state.zsbuf->texture;
389       struct si_state_dsa *dsa = sctx->queued.named.dsa;
390 
391       const unsigned cPerDepthSample = dsa->depth_enabled ? 5 : 0;
392       const unsigned cPerStencilSample = dsa->stencil_enabled ? 1 : 0;
393       const unsigned cDepth =
394          (cPerDepthSample + cPerStencilSample) * MAX2(zstex->buffer.b.b.nr_samples, 1);
395 
396       const unsigned depthLog2Pixels = util_logbase2(depthBinSizeTagPart / MAX2(cDepth, 1u));
397       unsigned depthBinSizeX = 1 << ((depthLog2Pixels + 1) / 2);
398       unsigned depthBinSizeY = 1 << (depthLog2Pixels / 2);
399 
400       depth_bin_size->x = MAX2(depthBinSizeX, minBinSizeX);
401       depth_bin_size->y = MAX2(depthBinSizeY, minBinSizeY);
402    }
403 }
404 
si_emit_dpbb_disable(struct si_context * sctx)405 static void si_emit_dpbb_disable(struct si_context *sctx)
406 {
407    unsigned initial_cdw = sctx->gfx_cs->current.cdw;
408 
409    if (sctx->chip_class >= GFX10) {
410       struct uvec2 bin_size = {};
411       struct uvec2 bin_size_extend = {};
412 
413       bin_size.x = 128;
414       bin_size.y = sctx->framebuffer.min_bytes_per_pixel <= 4 ? 128 : 64;
415 
416       if (bin_size.x >= 32)
417          bin_size_extend.x = util_logbase2(bin_size.x) - 5;
418       if (bin_size.y >= 32)
419          bin_size_extend.y = util_logbase2(bin_size.y) - 5;
420 
421       radeon_opt_set_context_reg(
422          sctx, R_028C44_PA_SC_BINNER_CNTL_0, SI_TRACKED_PA_SC_BINNER_CNTL_0,
423          S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_NEW_SC) |
424             S_028C44_BIN_SIZE_X(bin_size.x == 16) | S_028C44_BIN_SIZE_Y(bin_size.y == 16) |
425             S_028C44_BIN_SIZE_X_EXTEND(bin_size_extend.x) |
426             S_028C44_BIN_SIZE_Y_EXTEND(bin_size_extend.y) | S_028C44_DISABLE_START_OF_PRIM(1) |
427             S_028C44_FLUSH_ON_BINNING_TRANSITION(sctx->last_binning_enabled != 0));
428    } else {
429       radeon_opt_set_context_reg(
430          sctx, R_028C44_PA_SC_BINNER_CNTL_0, SI_TRACKED_PA_SC_BINNER_CNTL_0,
431          S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) |
432             S_028C44_DISABLE_START_OF_PRIM(1) |
433             S_028C44_FLUSH_ON_BINNING_TRANSITION((sctx->family == CHIP_VEGA12 ||
434                                                   sctx->family == CHIP_VEGA20 ||
435                                                   sctx->family >= CHIP_RAVEN2) &&
436                                                  sctx->last_binning_enabled != 0));
437    }
438 
439    unsigned db_dfsm_control =
440       sctx->chip_class >= GFX10 ? R_028038_DB_DFSM_CONTROL : R_028060_DB_DFSM_CONTROL;
441    radeon_opt_set_context_reg(
442       sctx, db_dfsm_control, SI_TRACKED_DB_DFSM_CONTROL,
443       S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF) | S_028060_POPS_DRAIN_PS_ON_OVERLAP(1));
444    if (initial_cdw != sctx->gfx_cs->current.cdw)
445       sctx->context_roll = true;
446 
447    sctx->last_binning_enabled = false;
448 }
449 
si_emit_dpbb_state(struct si_context * sctx)450 void si_emit_dpbb_state(struct si_context *sctx)
451 {
452    struct si_screen *sscreen = sctx->screen;
453    struct si_state_blend *blend = sctx->queued.named.blend;
454    struct si_state_dsa *dsa = sctx->queued.named.dsa;
455    unsigned db_shader_control = sctx->ps_db_shader_control;
456 
457    assert(sctx->chip_class >= GFX9);
458 
459    if (!sscreen->dpbb_allowed || sctx->dpbb_force_off) {
460       si_emit_dpbb_disable(sctx);
461       return;
462    }
463 
464    bool ps_can_kill =
465       G_02880C_KILL_ENABLE(db_shader_control) || G_02880C_MASK_EXPORT_ENABLE(db_shader_control) ||
466       G_02880C_COVERAGE_TO_MASK_ENABLE(db_shader_control) || blend->alpha_to_coverage;
467 
468    bool db_can_reject_z_trivially = !G_02880C_Z_EXPORT_ENABLE(db_shader_control) ||
469                                     G_02880C_CONSERVATIVE_Z_EXPORT(db_shader_control) ||
470                                     G_02880C_DEPTH_BEFORE_SHADER(db_shader_control);
471 
472    /* Disable DPBB when it's believed to be inefficient. */
473    if (sscreen->info.num_render_backends > 4 && ps_can_kill && db_can_reject_z_trivially &&
474        sctx->framebuffer.state.zsbuf && dsa->db_can_write) {
475       si_emit_dpbb_disable(sctx);
476       return;
477    }
478 
479    /* Compute the bin size. */
480    /* TODO: We could also look at enabled pixel shader outputs. */
481    unsigned cb_target_enabled_4bit =
482       sctx->framebuffer.colorbuf_enabled_4bit & blend->cb_target_enabled_4bit;
483    struct uvec2 color_bin_size, depth_bin_size;
484 
485    if (sctx->chip_class >= GFX10) {
486       gfx10_get_bin_sizes(sctx, cb_target_enabled_4bit, &color_bin_size, &depth_bin_size);
487    } else {
488       color_bin_size = si_get_color_bin_size(sctx, cb_target_enabled_4bit);
489       depth_bin_size = si_get_depth_bin_size(sctx);
490    }
491 
492    unsigned color_area = color_bin_size.x * color_bin_size.y;
493    unsigned depth_area = depth_bin_size.x * depth_bin_size.y;
494 
495    struct uvec2 bin_size = color_area < depth_area ? color_bin_size : depth_bin_size;
496 
497    if (!bin_size.x || !bin_size.y) {
498       si_emit_dpbb_disable(sctx);
499       return;
500    }
501 
502    /* Enable DFSM if it's preferred. */
503    unsigned punchout_mode = V_028060_FORCE_OFF;
504    bool disable_start_of_prim = true;
505    bool zs_eqaa_dfsm_bug =
506       sctx->chip_class == GFX9 && sctx->framebuffer.state.zsbuf &&
507       sctx->framebuffer.nr_samples != MAX2(1, sctx->framebuffer.state.zsbuf->texture->nr_samples);
508 
509    if (sscreen->dfsm_allowed && !zs_eqaa_dfsm_bug && cb_target_enabled_4bit &&
510        !G_02880C_KILL_ENABLE(db_shader_control) &&
511        /* These two also imply that DFSM is disabled when PS writes to memory. */
512        !G_02880C_EXEC_ON_HIER_FAIL(db_shader_control) &&
513        !G_02880C_EXEC_ON_NOOP(db_shader_control) &&
514        G_02880C_Z_ORDER(db_shader_control) == V_02880C_EARLY_Z_THEN_LATE_Z) {
515       punchout_mode = V_028060_AUTO;
516       disable_start_of_prim = (cb_target_enabled_4bit & blend->blend_enable_4bit) != 0;
517    }
518 
519    /* Tunable parameters. */
520    unsigned fpovs_per_batch = 63; /* allowed range: [0, 255], 0 = unlimited */
521 
522    /* Emit registers. */
523    struct uvec2 bin_size_extend = {};
524    if (bin_size.x >= 32)
525       bin_size_extend.x = util_logbase2(bin_size.x) - 5;
526    if (bin_size.y >= 32)
527       bin_size_extend.y = util_logbase2(bin_size.y) - 5;
528 
529    unsigned initial_cdw = sctx->gfx_cs->current.cdw;
530    radeon_opt_set_context_reg(
531       sctx, R_028C44_PA_SC_BINNER_CNTL_0, SI_TRACKED_PA_SC_BINNER_CNTL_0,
532       S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) | S_028C44_BIN_SIZE_X(bin_size.x == 16) |
533          S_028C44_BIN_SIZE_Y(bin_size.y == 16) | S_028C44_BIN_SIZE_X_EXTEND(bin_size_extend.x) |
534          S_028C44_BIN_SIZE_Y_EXTEND(bin_size_extend.y) |
535          S_028C44_CONTEXT_STATES_PER_BIN(sscreen->pbb_context_states_per_bin - 1) |
536          S_028C44_PERSISTENT_STATES_PER_BIN(sscreen->pbb_persistent_states_per_bin - 1) |
537          S_028C44_DISABLE_START_OF_PRIM(disable_start_of_prim) |
538          S_028C44_FPOVS_PER_BATCH(fpovs_per_batch) | S_028C44_OPTIMAL_BIN_SELECTION(1) |
539          S_028C44_FLUSH_ON_BINNING_TRANSITION((sctx->family == CHIP_VEGA12 ||
540                                                sctx->family == CHIP_VEGA20 ||
541                                                sctx->family >= CHIP_RAVEN2) &&
542                                               sctx->last_binning_enabled != 1));
543 
544    unsigned db_dfsm_control =
545       sctx->chip_class >= GFX10 ? R_028038_DB_DFSM_CONTROL : R_028060_DB_DFSM_CONTROL;
546    radeon_opt_set_context_reg(
547       sctx, db_dfsm_control, SI_TRACKED_DB_DFSM_CONTROL,
548       S_028060_PUNCHOUT_MODE(punchout_mode) | S_028060_POPS_DRAIN_PS_ON_OVERLAP(1));
549    if (initial_cdw != sctx->gfx_cs->current.cdw)
550       sctx->context_roll = true;
551 
552    sctx->last_binning_enabled = true;
553 }
554