• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2012 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 "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * on the rights to use, copy, modify, merge, publish, distribute, sub
8  * license, and/or sell copies of the Software, and to permit persons to whom
9  * the Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include "si_pipe.h"
25 #include "radeon/r600_cs.h"
26 #include "sid.h"
27 #include "gfx9d.h"
28 
29 #include "util/u_index_modify.h"
30 #include "util/u_log.h"
31 #include "util/u_upload_mgr.h"
32 #include "util/u_prim.h"
33 
34 #include "ac_debug.h"
35 
36 /* special primitive types */
37 #define SI_PRIM_RECTANGLE_LIST	PIPE_PRIM_MAX
38 
si_conv_pipe_prim(unsigned mode)39 static unsigned si_conv_pipe_prim(unsigned mode)
40 {
41         static const unsigned prim_conv[] = {
42 		[PIPE_PRIM_POINTS]			= V_008958_DI_PT_POINTLIST,
43 		[PIPE_PRIM_LINES]			= V_008958_DI_PT_LINELIST,
44 		[PIPE_PRIM_LINE_LOOP]			= V_008958_DI_PT_LINELOOP,
45 		[PIPE_PRIM_LINE_STRIP]			= V_008958_DI_PT_LINESTRIP,
46 		[PIPE_PRIM_TRIANGLES]			= V_008958_DI_PT_TRILIST,
47 		[PIPE_PRIM_TRIANGLE_STRIP]		= V_008958_DI_PT_TRISTRIP,
48 		[PIPE_PRIM_TRIANGLE_FAN]		= V_008958_DI_PT_TRIFAN,
49 		[PIPE_PRIM_QUADS]			= V_008958_DI_PT_QUADLIST,
50 		[PIPE_PRIM_QUAD_STRIP]			= V_008958_DI_PT_QUADSTRIP,
51 		[PIPE_PRIM_POLYGON]			= V_008958_DI_PT_POLYGON,
52 		[PIPE_PRIM_LINES_ADJACENCY]		= V_008958_DI_PT_LINELIST_ADJ,
53 		[PIPE_PRIM_LINE_STRIP_ADJACENCY]	= V_008958_DI_PT_LINESTRIP_ADJ,
54 		[PIPE_PRIM_TRIANGLES_ADJACENCY]		= V_008958_DI_PT_TRILIST_ADJ,
55 		[PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY]	= V_008958_DI_PT_TRISTRIP_ADJ,
56 		[PIPE_PRIM_PATCHES]			= V_008958_DI_PT_PATCH,
57 		[SI_PRIM_RECTANGLE_LIST]		= V_008958_DI_PT_RECTLIST
58         };
59 	assert(mode < ARRAY_SIZE(prim_conv));
60 	return prim_conv[mode];
61 }
62 
si_conv_prim_to_gs_out(unsigned mode)63 static unsigned si_conv_prim_to_gs_out(unsigned mode)
64 {
65 	static const int prim_conv[] = {
66 		[PIPE_PRIM_POINTS]			= V_028A6C_OUTPRIM_TYPE_POINTLIST,
67 		[PIPE_PRIM_LINES]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
68 		[PIPE_PRIM_LINE_LOOP]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
69 		[PIPE_PRIM_LINE_STRIP]			= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
70 		[PIPE_PRIM_TRIANGLES]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
71 		[PIPE_PRIM_TRIANGLE_STRIP]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
72 		[PIPE_PRIM_TRIANGLE_FAN]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
73 		[PIPE_PRIM_QUADS]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
74 		[PIPE_PRIM_QUAD_STRIP]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
75 		[PIPE_PRIM_POLYGON]			= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
76 		[PIPE_PRIM_LINES_ADJACENCY]		= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
77 		[PIPE_PRIM_LINE_STRIP_ADJACENCY]	= V_028A6C_OUTPRIM_TYPE_LINESTRIP,
78 		[PIPE_PRIM_TRIANGLES_ADJACENCY]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
79 		[PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY]	= V_028A6C_OUTPRIM_TYPE_TRISTRIP,
80 		[PIPE_PRIM_PATCHES]			= V_028A6C_OUTPRIM_TYPE_POINTLIST,
81 		[SI_PRIM_RECTANGLE_LIST]		= V_028A6C_OUTPRIM_TYPE_TRISTRIP
82 	};
83 	assert(mode < ARRAY_SIZE(prim_conv));
84 
85 	return prim_conv[mode];
86 }
87 
88 /**
89  * This calculates the LDS size for tessellation shaders (VS, TCS, TES).
90  * LS.LDS_SIZE is shared by all 3 shader stages.
91  *
92  * The information about LDS and other non-compile-time parameters is then
93  * written to userdata SGPRs.
94  */
si_emit_derived_tess_state(struct si_context * sctx,const struct pipe_draw_info * info,unsigned * num_patches)95 static void si_emit_derived_tess_state(struct si_context *sctx,
96 				       const struct pipe_draw_info *info,
97 				       unsigned *num_patches)
98 {
99 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
100 	struct si_shader *ls_current;
101 	struct si_shader_selector *ls;
102 	/* The TES pointer will only be used for sctx->last_tcs.
103 	 * It would be wrong to think that TCS = TES. */
104 	struct si_shader_selector *tcs =
105 		sctx->tcs_shader.cso ? sctx->tcs_shader.cso : sctx->tes_shader.cso;
106 	unsigned tess_uses_primid = sctx->ia_multi_vgt_param_key.u.tess_uses_prim_id;
107 	bool has_primid_instancing_bug = sctx->b.chip_class == SI &&
108 					 sctx->b.screen->info.max_se == 1;
109 	unsigned tes_sh_base = sctx->shader_pointers.sh_base[PIPE_SHADER_TESS_EVAL];
110 	unsigned num_tcs_input_cp = info->vertices_per_patch;
111 	unsigned num_tcs_output_cp, num_tcs_inputs, num_tcs_outputs;
112 	unsigned num_tcs_patch_outputs;
113 	unsigned input_vertex_size, output_vertex_size, pervertex_output_patch_size;
114 	unsigned input_patch_size, output_patch_size, output_patch0_offset;
115 	unsigned perpatch_output_offset, lds_size;
116 	unsigned tcs_in_layout, tcs_out_layout, tcs_out_offsets;
117 	unsigned offchip_layout, hardware_lds_size, ls_hs_config;
118 
119 	/* Since GFX9 has merged LS-HS in the TCS state, set LS = TCS. */
120 	if (sctx->b.chip_class >= GFX9) {
121 		if (sctx->tcs_shader.cso)
122 			ls_current = sctx->tcs_shader.current;
123 		else
124 			ls_current = sctx->fixed_func_tcs_shader.current;
125 
126 		ls = ls_current->key.part.tcs.ls;
127 	} else {
128 		ls_current = sctx->vs_shader.current;
129 		ls = sctx->vs_shader.cso;
130 	}
131 
132 	if (sctx->last_ls == ls_current &&
133 	    sctx->last_tcs == tcs &&
134 	    sctx->last_tes_sh_base == tes_sh_base &&
135 	    sctx->last_num_tcs_input_cp == num_tcs_input_cp &&
136 	    (!has_primid_instancing_bug ||
137 	     (sctx->last_tess_uses_primid == tess_uses_primid))) {
138 		*num_patches = sctx->last_num_patches;
139 		return;
140 	}
141 
142 	sctx->last_ls = ls_current;
143 	sctx->last_tcs = tcs;
144 	sctx->last_tes_sh_base = tes_sh_base;
145 	sctx->last_num_tcs_input_cp = num_tcs_input_cp;
146 	sctx->last_tess_uses_primid = tess_uses_primid;
147 
148 	/* This calculates how shader inputs and outputs among VS, TCS, and TES
149 	 * are laid out in LDS. */
150 	num_tcs_inputs = util_last_bit64(ls->outputs_written);
151 
152 	if (sctx->tcs_shader.cso) {
153 		num_tcs_outputs = util_last_bit64(tcs->outputs_written);
154 		num_tcs_output_cp = tcs->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT];
155 		num_tcs_patch_outputs = util_last_bit64(tcs->patch_outputs_written);
156 	} else {
157 		/* No TCS. Route varyings from LS to TES. */
158 		num_tcs_outputs = num_tcs_inputs;
159 		num_tcs_output_cp = num_tcs_input_cp;
160 		num_tcs_patch_outputs = 2; /* TESSINNER + TESSOUTER */
161 	}
162 
163 	input_vertex_size = num_tcs_inputs * 16;
164 	output_vertex_size = num_tcs_outputs * 16;
165 
166 	input_patch_size = num_tcs_input_cp * input_vertex_size;
167 
168 	pervertex_output_patch_size = num_tcs_output_cp * output_vertex_size;
169 	output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16;
170 
171 	/* Ensure that we only need one wave per SIMD so we don't need to check
172 	 * resource usage. Also ensures that the number of tcs in and out
173 	 * vertices per threadgroup are at most 256.
174 	 */
175 	*num_patches = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp) * 4;
176 
177 	/* Make sure that the data fits in LDS. This assumes the shaders only
178 	 * use LDS for the inputs and outputs.
179 	 *
180 	 * While CIK can use 64K per threadgroup, there is a hang on Stoney
181 	 * with 2 CUs if we use more than 32K. The closed Vulkan driver also
182 	 * uses 32K at most on all GCN chips.
183 	 */
184 	hardware_lds_size = 32768;
185 	*num_patches = MIN2(*num_patches, hardware_lds_size / (input_patch_size +
186 	                                                       output_patch_size));
187 
188 	/* Make sure the output data fits in the offchip buffer */
189 	*num_patches = MIN2(*num_patches,
190 			    (sctx->screen->tess_offchip_block_dw_size * 4) /
191 			    output_patch_size);
192 
193 	/* Not necessary for correctness, but improves performance. The
194 	 * specific value is taken from the proprietary driver.
195 	 */
196 	*num_patches = MIN2(*num_patches, 40);
197 
198 	if (sctx->b.chip_class == SI) {
199 		/* SI bug workaround, related to power management. Limit LS-HS
200 		 * threadgroups to only one wave.
201 		 */
202 		unsigned one_wave = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp);
203 		*num_patches = MIN2(*num_patches, one_wave);
204 	}
205 
206 	/* The VGT HS block increments the patch ID unconditionally
207 	 * within a single threadgroup. This results in incorrect
208 	 * patch IDs when instanced draws are used.
209 	 *
210 	 * The intended solution is to restrict threadgroups to
211 	 * a single instance by setting SWITCH_ON_EOI, which
212 	 * should cause IA to split instances up. However, this
213 	 * doesn't work correctly on SI when there is no other
214 	 * SE to switch to.
215 	 */
216 	if (has_primid_instancing_bug && tess_uses_primid)
217 		*num_patches = 1;
218 
219 	sctx->last_num_patches = *num_patches;
220 
221 	output_patch0_offset = input_patch_size * *num_patches;
222 	perpatch_output_offset = output_patch0_offset + pervertex_output_patch_size;
223 
224 	/* Compute userdata SGPRs. */
225 	assert(((input_vertex_size / 4) & ~0xff) == 0);
226 	assert(((output_vertex_size / 4) & ~0xff) == 0);
227 	assert(((input_patch_size / 4) & ~0x1fff) == 0);
228 	assert(((output_patch_size / 4) & ~0x1fff) == 0);
229 	assert(((output_patch0_offset / 16) & ~0xffff) == 0);
230 	assert(((perpatch_output_offset / 16) & ~0xffff) == 0);
231 	assert(num_tcs_input_cp <= 32);
232 	assert(num_tcs_output_cp <= 32);
233 
234 	tcs_in_layout = S_VS_STATE_LS_OUT_PATCH_SIZE(input_patch_size / 4) |
235 			S_VS_STATE_LS_OUT_VERTEX_SIZE(input_vertex_size / 4);
236 	tcs_out_layout = output_patch_size / 4;
237 	tcs_out_offsets = (output_patch0_offset / 16) |
238 			  ((perpatch_output_offset / 16) << 16);
239 	offchip_layout = *num_patches |
240 			 (num_tcs_output_cp << 6) |
241 			 (pervertex_output_patch_size * *num_patches << 12);
242 
243 	/* Compute the LDS size. */
244 	lds_size = output_patch0_offset + output_patch_size * *num_patches;
245 
246 	if (sctx->b.chip_class >= CIK) {
247 		assert(lds_size <= 65536);
248 		lds_size = align(lds_size, 512) / 512;
249 	} else {
250 		assert(lds_size <= 32768);
251 		lds_size = align(lds_size, 256) / 256;
252 	}
253 
254 	/* Set SI_SGPR_VS_STATE_BITS. */
255 	sctx->current_vs_state &= C_VS_STATE_LS_OUT_PATCH_SIZE &
256 				  C_VS_STATE_LS_OUT_VERTEX_SIZE;
257 	sctx->current_vs_state |= tcs_in_layout;
258 
259 	if (sctx->b.chip_class >= GFX9) {
260 		unsigned hs_rsrc2 = ls_current->config.rsrc2 |
261 				    S_00B42C_LDS_SIZE(lds_size);
262 
263 		radeon_set_sh_reg(cs, R_00B42C_SPI_SHADER_PGM_RSRC2_HS, hs_rsrc2);
264 
265 		/* Set userdata SGPRs for merged LS-HS. */
266 		radeon_set_sh_reg_seq(cs,
267 				      R_00B430_SPI_SHADER_USER_DATA_LS_0 +
268 				      GFX9_SGPR_TCS_OFFCHIP_LAYOUT * 4, 3);
269 		radeon_emit(cs, offchip_layout);
270 		radeon_emit(cs, tcs_out_offsets);
271 		radeon_emit(cs, tcs_out_layout | (num_tcs_input_cp << 26));
272 	} else {
273 		unsigned ls_rsrc2 = ls_current->config.rsrc2;
274 
275 		si_multiwave_lds_size_workaround(sctx->screen, &lds_size);
276 		ls_rsrc2 |= S_00B52C_LDS_SIZE(lds_size);
277 
278 		/* Due to a hw bug, RSRC2_LS must be written twice with another
279 		 * LS register written in between. */
280 		if (sctx->b.chip_class == CIK && sctx->b.family != CHIP_HAWAII)
281 			radeon_set_sh_reg(cs, R_00B52C_SPI_SHADER_PGM_RSRC2_LS, ls_rsrc2);
282 		radeon_set_sh_reg_seq(cs, R_00B528_SPI_SHADER_PGM_RSRC1_LS, 2);
283 		radeon_emit(cs, ls_current->config.rsrc1);
284 		radeon_emit(cs, ls_rsrc2);
285 
286 		/* Set userdata SGPRs for TCS. */
287 		radeon_set_sh_reg_seq(cs,
288 			R_00B430_SPI_SHADER_USER_DATA_HS_0 + GFX6_SGPR_TCS_OFFCHIP_LAYOUT * 4, 4);
289 		radeon_emit(cs, offchip_layout);
290 		radeon_emit(cs, tcs_out_offsets);
291 		radeon_emit(cs, tcs_out_layout | (num_tcs_input_cp << 26));
292 		radeon_emit(cs, tcs_in_layout);
293 	}
294 
295 	/* Set userdata SGPRs for TES. */
296 	radeon_set_sh_reg_seq(cs, tes_sh_base + SI_SGPR_TES_OFFCHIP_LAYOUT * 4, 2);
297 	radeon_emit(cs, offchip_layout);
298 	radeon_emit(cs, r600_resource(sctx->tess_offchip_ring)->gpu_address >> 16);
299 
300 	ls_hs_config = S_028B58_NUM_PATCHES(*num_patches) |
301 		       S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) |
302 		       S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp);
303 
304 	if (sctx->b.chip_class >= CIK)
305 		radeon_set_context_reg_idx(cs, R_028B58_VGT_LS_HS_CONFIG, 2,
306 					   ls_hs_config);
307 	else
308 		radeon_set_context_reg(cs, R_028B58_VGT_LS_HS_CONFIG,
309 				       ls_hs_config);
310 }
311 
si_num_prims_for_vertices(const struct pipe_draw_info * info)312 static unsigned si_num_prims_for_vertices(const struct pipe_draw_info *info)
313 {
314 	switch (info->mode) {
315 	case PIPE_PRIM_PATCHES:
316 		return info->count / info->vertices_per_patch;
317 	case SI_PRIM_RECTANGLE_LIST:
318 		return info->count / 3;
319 	default:
320 		return u_prims_for_vertices(info->mode, info->count);
321 	}
322 }
323 
324 static unsigned
si_get_init_multi_vgt_param(struct si_screen * sscreen,union si_vgt_param_key * key)325 si_get_init_multi_vgt_param(struct si_screen *sscreen,
326 			    union si_vgt_param_key *key)
327 {
328 	STATIC_ASSERT(sizeof(union si_vgt_param_key) == 4);
329 	unsigned max_primgroup_in_wave = 2;
330 
331 	/* SWITCH_ON_EOP(0) is always preferable. */
332 	bool wd_switch_on_eop = false;
333 	bool ia_switch_on_eop = false;
334 	bool ia_switch_on_eoi = false;
335 	bool partial_vs_wave = false;
336 	bool partial_es_wave = false;
337 
338 	if (key->u.uses_tess) {
339 		/* SWITCH_ON_EOI must be set if PrimID is used. */
340 		if (key->u.tess_uses_prim_id)
341 			ia_switch_on_eoi = true;
342 
343 		/* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
344 		if ((sscreen->info.family == CHIP_TAHITI ||
345 		     sscreen->info.family == CHIP_PITCAIRN ||
346 		     sscreen->info.family == CHIP_BONAIRE) &&
347 		    key->u.uses_gs)
348 			partial_vs_wave = true;
349 
350 		/* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
351 		if (sscreen->has_distributed_tess) {
352 			if (key->u.uses_gs) {
353 				if (sscreen->info.chip_class <= VI)
354 					partial_es_wave = true;
355 
356 				/* GPU hang workaround. */
357 				if (sscreen->info.family == CHIP_TONGA ||
358 				    sscreen->info.family == CHIP_FIJI ||
359 				    sscreen->info.family == CHIP_POLARIS10 ||
360 				    sscreen->info.family == CHIP_POLARIS11 ||
361 				    sscreen->info.family == CHIP_POLARIS12)
362 					partial_vs_wave = true;
363 			} else {
364 				partial_vs_wave = true;
365 			}
366 		}
367 	}
368 
369 	/* This is a hardware requirement. */
370 	if (key->u.line_stipple_enabled ||
371 	    (sscreen->debug_flags & DBG(SWITCH_ON_EOP))) {
372 		ia_switch_on_eop = true;
373 		wd_switch_on_eop = true;
374 	}
375 
376 	if (sscreen->info.chip_class >= CIK) {
377 		/* WD_SWITCH_ON_EOP has no effect on GPUs with less than
378 		 * 4 shader engines. Set 1 to pass the assertion below.
379 		 * The other cases are hardware requirements.
380 		 *
381 		 * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
382 		 * for points, line strips, and tri strips.
383 		 */
384 		if (sscreen->info.max_se < 4 ||
385 		    key->u.prim == PIPE_PRIM_POLYGON ||
386 		    key->u.prim == PIPE_PRIM_LINE_LOOP ||
387 		    key->u.prim == PIPE_PRIM_TRIANGLE_FAN ||
388 		    key->u.prim == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY ||
389 		    (key->u.primitive_restart &&
390 		     (sscreen->info.family < CHIP_POLARIS10 ||
391 		      (key->u.prim != PIPE_PRIM_POINTS &&
392 		       key->u.prim != PIPE_PRIM_LINE_STRIP &&
393 		       key->u.prim != PIPE_PRIM_TRIANGLE_STRIP))) ||
394 		    key->u.count_from_stream_output)
395 			wd_switch_on_eop = true;
396 
397 		/* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
398 		 * We don't know that for indirect drawing, so treat it as
399 		 * always problematic. */
400 		if (sscreen->info.family == CHIP_HAWAII &&
401 		    key->u.uses_instancing)
402 			wd_switch_on_eop = true;
403 
404 		/* Performance recommendation for 4 SE Gfx7-8 parts if
405 		 * instances are smaller than a primgroup.
406 		 * Assume indirect draws always use small instances.
407 		 * This is needed for good VS wave utilization.
408 		 */
409 		if (sscreen->info.chip_class <= VI &&
410 		    sscreen->info.max_se == 4 &&
411 		    key->u.multi_instances_smaller_than_primgroup)
412 			wd_switch_on_eop = true;
413 
414 		/* Required on CIK and later. */
415 		if (sscreen->info.max_se > 2 && !wd_switch_on_eop)
416 			ia_switch_on_eoi = true;
417 
418 		/* Required by Hawaii and, for some special cases, by VI. */
419 		if (ia_switch_on_eoi &&
420 		    (sscreen->info.family == CHIP_HAWAII ||
421 		     (sscreen->info.chip_class == VI &&
422 		      (key->u.uses_gs || max_primgroup_in_wave != 2))))
423 			partial_vs_wave = true;
424 
425 		/* Instancing bug on Bonaire. */
426 		if (sscreen->info.family == CHIP_BONAIRE && ia_switch_on_eoi &&
427 		    key->u.uses_instancing)
428 			partial_vs_wave = true;
429 
430 		/* If the WD switch is false, the IA switch must be false too. */
431 		assert(wd_switch_on_eop || !ia_switch_on_eop);
432 	}
433 
434 	/* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
435 	if (sscreen->info.chip_class <= VI && ia_switch_on_eoi)
436 		partial_es_wave = true;
437 
438 	return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
439 		S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
440 		S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
441 		S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
442 		S_028AA8_WD_SWITCH_ON_EOP(sscreen->info.chip_class >= CIK ? wd_switch_on_eop : 0) |
443 		/* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */
444 		S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen->info.chip_class == VI ?
445 					     max_primgroup_in_wave : 0) |
446 		S_030960_EN_INST_OPT_BASIC(sscreen->info.chip_class >= GFX9) |
447 		S_030960_EN_INST_OPT_ADV(sscreen->info.chip_class >= GFX9);
448 }
449 
si_init_ia_multi_vgt_param_table(struct si_context * sctx)450 void si_init_ia_multi_vgt_param_table(struct si_context *sctx)
451 {
452 	for (int prim = 0; prim <= SI_PRIM_RECTANGLE_LIST; prim++)
453 	for (int uses_instancing = 0; uses_instancing < 2; uses_instancing++)
454 	for (int multi_instances = 0; multi_instances < 2; multi_instances++)
455 	for (int primitive_restart = 0; primitive_restart < 2; primitive_restart++)
456 	for (int count_from_so = 0; count_from_so < 2; count_from_so++)
457 	for (int line_stipple = 0; line_stipple < 2; line_stipple++)
458 	for (int uses_tess = 0; uses_tess < 2; uses_tess++)
459 	for (int tess_uses_primid = 0; tess_uses_primid < 2; tess_uses_primid++)
460 	for (int uses_gs = 0; uses_gs < 2; uses_gs++) {
461 		union si_vgt_param_key key;
462 
463 		key.index = 0;
464 		key.u.prim = prim;
465 		key.u.uses_instancing = uses_instancing;
466 		key.u.multi_instances_smaller_than_primgroup = multi_instances;
467 		key.u.primitive_restart = primitive_restart;
468 		key.u.count_from_stream_output = count_from_so;
469 		key.u.line_stipple_enabled = line_stipple;
470 		key.u.uses_tess = uses_tess;
471 		key.u.tess_uses_prim_id = tess_uses_primid;
472 		key.u.uses_gs = uses_gs;
473 
474 		sctx->ia_multi_vgt_param[key.index] =
475 			si_get_init_multi_vgt_param(sctx->screen, &key);
476 	}
477 }
478 
si_get_ia_multi_vgt_param(struct si_context * sctx,const struct pipe_draw_info * info,unsigned num_patches)479 static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
480 					  const struct pipe_draw_info *info,
481 					  unsigned num_patches)
482 {
483 	union si_vgt_param_key key = sctx->ia_multi_vgt_param_key;
484 	unsigned primgroup_size;
485 	unsigned ia_multi_vgt_param;
486 
487 	if (sctx->tes_shader.cso) {
488 		primgroup_size = num_patches; /* must be a multiple of NUM_PATCHES */
489 	} else if (sctx->gs_shader.cso) {
490 		primgroup_size = 64; /* recommended with a GS */
491 	} else {
492 		primgroup_size = 128; /* recommended without a GS and tess */
493 	}
494 
495 	key.u.prim = info->mode;
496 	key.u.uses_instancing = info->indirect || info->instance_count > 1;
497 	key.u.multi_instances_smaller_than_primgroup =
498 		info->indirect ||
499 		(info->instance_count > 1 &&
500 		 (info->count_from_stream_output ||
501 		  si_num_prims_for_vertices(info) < primgroup_size));
502 	key.u.primitive_restart = info->primitive_restart;
503 	key.u.count_from_stream_output = info->count_from_stream_output != NULL;
504 
505 	ia_multi_vgt_param = sctx->ia_multi_vgt_param[key.index] |
506 			     S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1);
507 
508 	if (sctx->gs_shader.cso) {
509 		/* GS requirement. */
510 		if (sctx->b.chip_class <= VI &&
511 		    SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
512 			ia_multi_vgt_param |= S_028AA8_PARTIAL_ES_WAVE_ON(1);
513 
514 		/* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
515 		 * The hw doc says all multi-SE chips are affected, but Vulkan
516 		 * only applies it to Hawaii. Do what Vulkan does.
517 		 */
518 		if (sctx->b.family == CHIP_HAWAII &&
519 		    G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param) &&
520 		    (info->indirect ||
521 		     (info->instance_count > 1 &&
522 		      (info->count_from_stream_output ||
523 		       si_num_prims_for_vertices(info) <= 1))))
524 			sctx->b.flags |= SI_CONTEXT_VGT_FLUSH;
525 	}
526 
527 	return ia_multi_vgt_param;
528 }
529 
530 /* rast_prim is the primitive type after GS. */
si_emit_rasterizer_prim_state(struct si_context * sctx)531 static void si_emit_rasterizer_prim_state(struct si_context *sctx)
532 {
533 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
534 	enum pipe_prim_type rast_prim = sctx->current_rast_prim;
535 	struct si_state_rasterizer *rs = sctx->emitted.named.rasterizer;
536 
537 	/* Skip this if not rendering lines. */
538 	if (rast_prim != PIPE_PRIM_LINES &&
539 	    rast_prim != PIPE_PRIM_LINE_LOOP &&
540 	    rast_prim != PIPE_PRIM_LINE_STRIP &&
541 	    rast_prim != PIPE_PRIM_LINES_ADJACENCY &&
542 	    rast_prim != PIPE_PRIM_LINE_STRIP_ADJACENCY)
543 		return;
544 
545 	if (rast_prim == sctx->last_rast_prim &&
546 	    rs->pa_sc_line_stipple == sctx->last_sc_line_stipple)
547 		return;
548 
549 	/* For lines, reset the stipple pattern at each primitive. Otherwise,
550 	 * reset the stipple pattern at each packet (line strips, line loops).
551 	 */
552 	radeon_set_context_reg(cs, R_028A0C_PA_SC_LINE_STIPPLE,
553 		rs->pa_sc_line_stipple |
554 		S_028A0C_AUTO_RESET_CNTL(rast_prim == PIPE_PRIM_LINES ? 1 : 2));
555 
556 	sctx->last_rast_prim = rast_prim;
557 	sctx->last_sc_line_stipple = rs->pa_sc_line_stipple;
558 }
559 
si_emit_vs_state(struct si_context * sctx,const struct pipe_draw_info * info)560 static void si_emit_vs_state(struct si_context *sctx,
561 			     const struct pipe_draw_info *info)
562 {
563 	sctx->current_vs_state &= C_VS_STATE_INDEXED;
564 	sctx->current_vs_state |= S_VS_STATE_INDEXED(!!info->index_size);
565 
566 	if (sctx->num_vs_blit_sgprs) {
567 		/* Re-emit the state after we leave u_blitter. */
568 		sctx->last_vs_state = ~0;
569 		return;
570 	}
571 
572 	if (sctx->current_vs_state != sctx->last_vs_state) {
573 		struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
574 
575 		radeon_set_sh_reg(cs,
576 			sctx->shader_pointers.sh_base[PIPE_SHADER_VERTEX] +
577 			SI_SGPR_VS_STATE_BITS * 4,
578 			sctx->current_vs_state);
579 
580 		sctx->last_vs_state = sctx->current_vs_state;
581 	}
582 }
583 
si_emit_draw_registers(struct si_context * sctx,const struct pipe_draw_info * info,unsigned num_patches)584 static void si_emit_draw_registers(struct si_context *sctx,
585 				   const struct pipe_draw_info *info,
586 				   unsigned num_patches)
587 {
588 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
589 	unsigned prim = si_conv_pipe_prim(info->mode);
590 	unsigned gs_out_prim = si_conv_prim_to_gs_out(sctx->current_rast_prim);
591 	unsigned ia_multi_vgt_param;
592 
593 	ia_multi_vgt_param = si_get_ia_multi_vgt_param(sctx, info, num_patches);
594 
595 	/* Draw state. */
596 	if (ia_multi_vgt_param != sctx->last_multi_vgt_param) {
597 		if (sctx->b.chip_class >= GFX9)
598 			radeon_set_uconfig_reg_idx(cs, R_030960_IA_MULTI_VGT_PARAM, 4, ia_multi_vgt_param);
599 		else if (sctx->b.chip_class >= CIK)
600 			radeon_set_context_reg_idx(cs, R_028AA8_IA_MULTI_VGT_PARAM, 1, ia_multi_vgt_param);
601 		else
602 			radeon_set_context_reg(cs, R_028AA8_IA_MULTI_VGT_PARAM, ia_multi_vgt_param);
603 
604 		sctx->last_multi_vgt_param = ia_multi_vgt_param;
605 	}
606 	if (prim != sctx->last_prim) {
607 		if (sctx->b.chip_class >= CIK)
608 			radeon_set_uconfig_reg_idx(cs, R_030908_VGT_PRIMITIVE_TYPE, 1, prim);
609 		else
610 			radeon_set_config_reg(cs, R_008958_VGT_PRIMITIVE_TYPE, prim);
611 
612 		sctx->last_prim = prim;
613 	}
614 
615 	if (gs_out_prim != sctx->last_gs_out_prim) {
616 		radeon_set_context_reg(cs, R_028A6C_VGT_GS_OUT_PRIM_TYPE, gs_out_prim);
617 		sctx->last_gs_out_prim = gs_out_prim;
618 	}
619 
620 	/* Primitive restart. */
621 	if (info->primitive_restart != sctx->last_primitive_restart_en) {
622 		if (sctx->b.chip_class >= GFX9)
623 			radeon_set_uconfig_reg(cs, R_03092C_VGT_MULTI_PRIM_IB_RESET_EN,
624 					       info->primitive_restart);
625 		else
626 			radeon_set_context_reg(cs, R_028A94_VGT_MULTI_PRIM_IB_RESET_EN,
627 					       info->primitive_restart);
628 
629 		sctx->last_primitive_restart_en = info->primitive_restart;
630 
631 	}
632 	if (info->primitive_restart &&
633 	    (info->restart_index != sctx->last_restart_index ||
634 	     sctx->last_restart_index == SI_RESTART_INDEX_UNKNOWN)) {
635 		radeon_set_context_reg(cs, R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX,
636 				       info->restart_index);
637 		sctx->last_restart_index = info->restart_index;
638 	}
639 }
640 
si_emit_draw_packets(struct si_context * sctx,const struct pipe_draw_info * info,struct pipe_resource * indexbuf,unsigned index_size,unsigned index_offset)641 static void si_emit_draw_packets(struct si_context *sctx,
642 				 const struct pipe_draw_info *info,
643 				 struct pipe_resource *indexbuf,
644 				 unsigned index_size,
645 				 unsigned index_offset)
646 {
647 	struct pipe_draw_indirect_info *indirect = info->indirect;
648 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
649 	unsigned sh_base_reg = sctx->shader_pointers.sh_base[PIPE_SHADER_VERTEX];
650 	bool render_cond_bit = sctx->b.render_cond && !sctx->b.render_cond_force_off;
651 	uint32_t index_max_size = 0;
652 	uint64_t index_va = 0;
653 
654 	if (info->count_from_stream_output) {
655 		struct si_streamout_target *t =
656 			(struct si_streamout_target*)info->count_from_stream_output;
657 		uint64_t va = t->buf_filled_size->gpu_address +
658 			      t->buf_filled_size_offset;
659 
660 		radeon_set_context_reg(cs, R_028B30_VGT_STRMOUT_DRAW_OPAQUE_VERTEX_STRIDE,
661 				       t->stride_in_dw);
662 
663 		radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
664 		radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
665 			    COPY_DATA_DST_SEL(COPY_DATA_REG) |
666 			    COPY_DATA_WR_CONFIRM);
667 		radeon_emit(cs, va);     /* src address lo */
668 		radeon_emit(cs, va >> 32); /* src address hi */
669 		radeon_emit(cs, R_028B2C_VGT_STRMOUT_DRAW_OPAQUE_BUFFER_FILLED_SIZE >> 2);
670 		radeon_emit(cs, 0); /* unused */
671 
672 		radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
673 				      t->buf_filled_size, RADEON_USAGE_READ,
674 				      RADEON_PRIO_SO_FILLED_SIZE);
675 	}
676 
677 	/* draw packet */
678 	if (index_size) {
679 		if (index_size != sctx->last_index_size) {
680 			unsigned index_type;
681 
682 			/* index type */
683 			switch (index_size) {
684 			case 1:
685 				index_type = V_028A7C_VGT_INDEX_8;
686 				break;
687 			case 2:
688 				index_type = V_028A7C_VGT_INDEX_16 |
689 					     (SI_BIG_ENDIAN && sctx->b.chip_class <= CIK ?
690 						      V_028A7C_VGT_DMA_SWAP_16_BIT : 0);
691 				break;
692 			case 4:
693 				index_type = V_028A7C_VGT_INDEX_32 |
694 					     (SI_BIG_ENDIAN && sctx->b.chip_class <= CIK ?
695 						      V_028A7C_VGT_DMA_SWAP_32_BIT : 0);
696 				break;
697 			default:
698 				assert(!"unreachable");
699 				return;
700 			}
701 
702 			if (sctx->b.chip_class >= GFX9) {
703 				radeon_set_uconfig_reg_idx(cs, R_03090C_VGT_INDEX_TYPE,
704 							   2, index_type);
705 			} else {
706 				radeon_emit(cs, PKT3(PKT3_INDEX_TYPE, 0, 0));
707 				radeon_emit(cs, index_type);
708 			}
709 
710 			sctx->last_index_size = index_size;
711 		}
712 
713 		index_max_size = (indexbuf->width0 - index_offset) /
714 				  index_size;
715 		index_va = r600_resource(indexbuf)->gpu_address + index_offset;
716 
717 		radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
718 				      (struct r600_resource *)indexbuf,
719 				      RADEON_USAGE_READ, RADEON_PRIO_INDEX_BUFFER);
720 	} else {
721 		/* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
722 		 * so the state must be re-emitted before the next indexed draw.
723 		 */
724 		if (sctx->b.chip_class >= CIK)
725 			sctx->last_index_size = -1;
726 	}
727 
728 	if (indirect) {
729 		uint64_t indirect_va = r600_resource(indirect->buffer)->gpu_address;
730 
731 		assert(indirect_va % 8 == 0);
732 
733 		si_invalidate_draw_sh_constants(sctx);
734 
735 		radeon_emit(cs, PKT3(PKT3_SET_BASE, 2, 0));
736 		radeon_emit(cs, 1);
737 		radeon_emit(cs, indirect_va);
738 		radeon_emit(cs, indirect_va >> 32);
739 
740 		radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
741 				      (struct r600_resource *)indirect->buffer,
742 				      RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
743 
744 		unsigned di_src_sel = index_size ? V_0287F0_DI_SRC_SEL_DMA
745 						    : V_0287F0_DI_SRC_SEL_AUTO_INDEX;
746 
747 		assert(indirect->offset % 4 == 0);
748 
749 		if (index_size) {
750 			radeon_emit(cs, PKT3(PKT3_INDEX_BASE, 1, 0));
751 			radeon_emit(cs, index_va);
752 			radeon_emit(cs, index_va >> 32);
753 
754 			radeon_emit(cs, PKT3(PKT3_INDEX_BUFFER_SIZE, 0, 0));
755 			radeon_emit(cs, index_max_size);
756 		}
757 
758 		if (!sctx->screen->has_draw_indirect_multi) {
759 			radeon_emit(cs, PKT3(index_size ? PKT3_DRAW_INDEX_INDIRECT
760 							   : PKT3_DRAW_INDIRECT,
761 					     3, render_cond_bit));
762 			radeon_emit(cs, indirect->offset);
763 			radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
764 			radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
765 			radeon_emit(cs, di_src_sel);
766 		} else {
767 			uint64_t count_va = 0;
768 
769 			if (indirect->indirect_draw_count) {
770 				struct r600_resource *params_buf =
771 					(struct r600_resource *)indirect->indirect_draw_count;
772 
773 				radeon_add_to_buffer_list(
774 					&sctx->b, &sctx->b.gfx, params_buf,
775 					RADEON_USAGE_READ, RADEON_PRIO_DRAW_INDIRECT);
776 
777 				count_va = params_buf->gpu_address + indirect->indirect_draw_count_offset;
778 			}
779 
780 			radeon_emit(cs, PKT3(index_size ? PKT3_DRAW_INDEX_INDIRECT_MULTI :
781 							     PKT3_DRAW_INDIRECT_MULTI,
782 					     8, render_cond_bit));
783 			radeon_emit(cs, indirect->offset);
784 			radeon_emit(cs, (sh_base_reg + SI_SGPR_BASE_VERTEX * 4 - SI_SH_REG_OFFSET) >> 2);
785 			radeon_emit(cs, (sh_base_reg + SI_SGPR_START_INSTANCE * 4 - SI_SH_REG_OFFSET) >> 2);
786 			radeon_emit(cs, ((sh_base_reg + SI_SGPR_DRAWID * 4 - SI_SH_REG_OFFSET) >> 2) |
787 					S_2C3_DRAW_INDEX_ENABLE(1) |
788 					S_2C3_COUNT_INDIRECT_ENABLE(!!indirect->indirect_draw_count));
789 			radeon_emit(cs, indirect->draw_count);
790 			radeon_emit(cs, count_va);
791 			radeon_emit(cs, count_va >> 32);
792 			radeon_emit(cs, indirect->stride);
793 			radeon_emit(cs, di_src_sel);
794 		}
795 	} else {
796 		int base_vertex;
797 
798 		radeon_emit(cs, PKT3(PKT3_NUM_INSTANCES, 0, 0));
799 		radeon_emit(cs, info->instance_count);
800 
801 		/* Base vertex and start instance. */
802 		base_vertex = index_size ? info->index_bias : info->start;
803 
804 		if (sctx->num_vs_blit_sgprs) {
805 			/* Re-emit draw constants after we leave u_blitter. */
806 			si_invalidate_draw_sh_constants(sctx);
807 
808 			/* Blit VS doesn't use BASE_VERTEX, START_INSTANCE, and DRAWID. */
809 			radeon_set_sh_reg_seq(cs, sh_base_reg + SI_SGPR_VS_BLIT_DATA * 4,
810 					      sctx->num_vs_blit_sgprs);
811 			radeon_emit_array(cs, sctx->vs_blit_sh_data,
812 					  sctx->num_vs_blit_sgprs);
813 		} else if (base_vertex != sctx->last_base_vertex ||
814 			   sctx->last_base_vertex == SI_BASE_VERTEX_UNKNOWN ||
815 			   info->start_instance != sctx->last_start_instance ||
816 			   info->drawid != sctx->last_drawid ||
817 			   sh_base_reg != sctx->last_sh_base_reg) {
818 			radeon_set_sh_reg_seq(cs, sh_base_reg + SI_SGPR_BASE_VERTEX * 4, 3);
819 			radeon_emit(cs, base_vertex);
820 			radeon_emit(cs, info->start_instance);
821 			radeon_emit(cs, info->drawid);
822 
823 			sctx->last_base_vertex = base_vertex;
824 			sctx->last_start_instance = info->start_instance;
825 			sctx->last_drawid = info->drawid;
826 			sctx->last_sh_base_reg = sh_base_reg;
827 		}
828 
829 		if (index_size) {
830 			index_va += info->start * index_size;
831 
832 			radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_2, 4, render_cond_bit));
833 			radeon_emit(cs, index_max_size);
834 			radeon_emit(cs, index_va);
835 			radeon_emit(cs, index_va >> 32);
836 			radeon_emit(cs, info->count);
837 			radeon_emit(cs, V_0287F0_DI_SRC_SEL_DMA);
838 		} else {
839 			radeon_emit(cs, PKT3(PKT3_DRAW_INDEX_AUTO, 1, render_cond_bit));
840 			radeon_emit(cs, info->count);
841 			radeon_emit(cs, V_0287F0_DI_SRC_SEL_AUTO_INDEX |
842 				        S_0287F0_USE_OPAQUE(!!info->count_from_stream_output));
843 		}
844 	}
845 }
846 
si_emit_surface_sync(struct r600_common_context * rctx,unsigned cp_coher_cntl)847 static void si_emit_surface_sync(struct r600_common_context *rctx,
848 				 unsigned cp_coher_cntl)
849 {
850 	struct radeon_winsys_cs *cs = rctx->gfx.cs;
851 
852 	if (rctx->chip_class >= GFX9) {
853 		/* Flush caches and wait for the caches to assert idle. */
854 		radeon_emit(cs, PKT3(PKT3_ACQUIRE_MEM, 5, 0));
855 		radeon_emit(cs, cp_coher_cntl);	/* CP_COHER_CNTL */
856 		radeon_emit(cs, 0xffffffff);	/* CP_COHER_SIZE */
857 		radeon_emit(cs, 0xffffff);	/* CP_COHER_SIZE_HI */
858 		radeon_emit(cs, 0);		/* CP_COHER_BASE */
859 		radeon_emit(cs, 0);		/* CP_COHER_BASE_HI */
860 		radeon_emit(cs, 0x0000000A);	/* POLL_INTERVAL */
861 	} else {
862 		/* ACQUIRE_MEM is only required on a compute ring. */
863 		radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0));
864 		radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
865 		radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
866 		radeon_emit(cs, 0);               /* CP_COHER_BASE */
867 		radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
868 	}
869 }
870 
si_emit_cache_flush(struct si_context * sctx)871 void si_emit_cache_flush(struct si_context *sctx)
872 {
873 	struct r600_common_context *rctx = &sctx->b;
874 	struct radeon_winsys_cs *cs = rctx->gfx.cs;
875 	uint32_t cp_coher_cntl = 0;
876 	uint32_t flush_cb_db = rctx->flags & (SI_CONTEXT_FLUSH_AND_INV_CB |
877 					      SI_CONTEXT_FLUSH_AND_INV_DB);
878 
879 	if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB)
880 		sctx->b.num_cb_cache_flushes++;
881 	if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB)
882 		sctx->b.num_db_cache_flushes++;
883 
884 	/* SI has a bug that it always flushes ICACHE and KCACHE if either
885 	 * bit is set. An alternative way is to write SQC_CACHES, but that
886 	 * doesn't seem to work reliably. Since the bug doesn't affect
887 	 * correctness (it only does more work than necessary) and
888 	 * the performance impact is likely negligible, there is no plan
889 	 * to add a workaround for it.
890 	 */
891 
892 	if (rctx->flags & SI_CONTEXT_INV_ICACHE)
893 		cp_coher_cntl |= S_0085F0_SH_ICACHE_ACTION_ENA(1);
894 	if (rctx->flags & SI_CONTEXT_INV_SMEM_L1)
895 		cp_coher_cntl |= S_0085F0_SH_KCACHE_ACTION_ENA(1);
896 
897 	if (rctx->chip_class <= VI) {
898 		if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
899 			cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
900 					 S_0085F0_CB0_DEST_BASE_ENA(1) |
901 					 S_0085F0_CB1_DEST_BASE_ENA(1) |
902 					 S_0085F0_CB2_DEST_BASE_ENA(1) |
903 					 S_0085F0_CB3_DEST_BASE_ENA(1) |
904 					 S_0085F0_CB4_DEST_BASE_ENA(1) |
905 					 S_0085F0_CB5_DEST_BASE_ENA(1) |
906 					 S_0085F0_CB6_DEST_BASE_ENA(1) |
907 					 S_0085F0_CB7_DEST_BASE_ENA(1);
908 
909 			/* Necessary for DCC */
910 			if (rctx->chip_class == VI)
911 				si_gfx_write_event_eop(rctx, V_028A90_FLUSH_AND_INV_CB_DATA_TS,
912 							 0, EOP_DATA_SEL_DISCARD, NULL,
913 							 0, 0, SI_NOT_QUERY);
914 		}
915 		if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_DB)
916 			cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
917 					 S_0085F0_DB_DEST_BASE_ENA(1);
918 	}
919 
920 	if (rctx->flags & SI_CONTEXT_FLUSH_AND_INV_CB) {
921 		/* Flush CMASK/FMASK/DCC. SURFACE_SYNC will wait for idle. */
922 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
923 		radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
924 	}
925 	if (rctx->flags & (SI_CONTEXT_FLUSH_AND_INV_DB |
926 			   SI_CONTEXT_FLUSH_AND_INV_DB_META)) {
927 		/* Flush HTILE. SURFACE_SYNC will wait for idle. */
928 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
929 		radeon_emit(cs, EVENT_TYPE(V_028A90_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
930 	}
931 
932 	/* Wait for shader engines to go idle.
933 	 * VS and PS waits are unnecessary if SURFACE_SYNC is going to wait
934 	 * for everything including CB/DB cache flushes.
935 	 */
936 	if (!flush_cb_db) {
937 		if (rctx->flags & SI_CONTEXT_PS_PARTIAL_FLUSH) {
938 			radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
939 			radeon_emit(cs, EVENT_TYPE(V_028A90_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
940 			/* Only count explicit shader flushes, not implicit ones
941 			 * done by SURFACE_SYNC.
942 			 */
943 			rctx->num_vs_flushes++;
944 			rctx->num_ps_flushes++;
945 		} else if (rctx->flags & SI_CONTEXT_VS_PARTIAL_FLUSH) {
946 			radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
947 			radeon_emit(cs, EVENT_TYPE(V_028A90_VS_PARTIAL_FLUSH) | EVENT_INDEX(4));
948 			rctx->num_vs_flushes++;
949 		}
950 	}
951 
952 	if (rctx->flags & SI_CONTEXT_CS_PARTIAL_FLUSH &&
953 	    sctx->compute_is_busy) {
954 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
955 		radeon_emit(cs, EVENT_TYPE(V_028A90_CS_PARTIAL_FLUSH | EVENT_INDEX(4)));
956 		rctx->num_cs_flushes++;
957 		sctx->compute_is_busy = false;
958 	}
959 
960 	/* VGT state synchronization. */
961 	if (rctx->flags & SI_CONTEXT_VGT_FLUSH) {
962 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
963 		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
964 	}
965 	if (rctx->flags & SI_CONTEXT_VGT_STREAMOUT_SYNC) {
966 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
967 		radeon_emit(cs, EVENT_TYPE(V_028A90_VGT_STREAMOUT_SYNC) | EVENT_INDEX(0));
968 	}
969 
970 	/* GFX9: Wait for idle if we're flushing CB or DB. ACQUIRE_MEM doesn't
971 	 * wait for idle on GFX9. We have to use a TS event.
972 	 */
973 	if (sctx->b.chip_class >= GFX9 && flush_cb_db) {
974 		uint64_t va;
975 		unsigned tc_flags, cb_db_event;
976 
977 		/* Set the CB/DB flush event. */
978 		switch (flush_cb_db) {
979 		case SI_CONTEXT_FLUSH_AND_INV_CB:
980 			cb_db_event = V_028A90_FLUSH_AND_INV_CB_DATA_TS;
981 			break;
982 		case SI_CONTEXT_FLUSH_AND_INV_DB:
983 			cb_db_event = V_028A90_FLUSH_AND_INV_DB_DATA_TS;
984 			break;
985 		default:
986 			/* both CB & DB */
987 			cb_db_event = V_028A90_CACHE_FLUSH_AND_INV_TS_EVENT;
988 		}
989 
990 		/* These are the only allowed combinations. If you need to
991 		 * do multiple operations at once, do them separately.
992 		 * All operations that invalidate L2 also seem to invalidate
993 		 * metadata. Volatile (VOL) and WC flushes are not listed here.
994 		 *
995 		 * TC    | TC_WB         = writeback & invalidate L2 & L1
996 		 * TC    | TC_WB | TC_NC = writeback & invalidate L2 for MTYPE == NC
997 		 *         TC_WB | TC_NC = writeback L2 for MTYPE == NC
998 		 * TC            | TC_NC = invalidate L2 for MTYPE == NC
999 		 * TC    | TC_MD         = writeback & invalidate L2 metadata (DCC, etc.)
1000 		 * TCL1                  = invalidate L1
1001 		 */
1002 		tc_flags = 0;
1003 
1004 		if (rctx->flags & SI_CONTEXT_INV_L2_METADATA) {
1005 			tc_flags = EVENT_TC_ACTION_ENA |
1006 				   EVENT_TC_MD_ACTION_ENA;
1007 		}
1008 
1009 		/* Ideally flush TC together with CB/DB. */
1010 		if (rctx->flags & SI_CONTEXT_INV_GLOBAL_L2) {
1011 			/* Writeback and invalidate everything in L2 & L1. */
1012 			tc_flags = EVENT_TC_ACTION_ENA |
1013 				   EVENT_TC_WB_ACTION_ENA;
1014 
1015 			/* Clear the flags. */
1016 			rctx->flags &= ~(SI_CONTEXT_INV_GLOBAL_L2 |
1017 					 SI_CONTEXT_WRITEBACK_GLOBAL_L2 |
1018 					 SI_CONTEXT_INV_VMEM_L1);
1019 			sctx->b.num_L2_invalidates++;
1020 		}
1021 
1022 		/* Do the flush (enqueue the event and wait for it). */
1023 		va = sctx->wait_mem_scratch->gpu_address;
1024 		sctx->wait_mem_number++;
1025 
1026 		si_gfx_write_event_eop(rctx, cb_db_event, tc_flags,
1027 					 EOP_DATA_SEL_VALUE_32BIT,
1028 					 sctx->wait_mem_scratch, va,
1029 					 sctx->wait_mem_number, SI_NOT_QUERY);
1030 		si_gfx_wait_fence(rctx, va, sctx->wait_mem_number, 0xffffffff);
1031 	}
1032 
1033 	/* Make sure ME is idle (it executes most packets) before continuing.
1034 	 * This prevents read-after-write hazards between PFP and ME.
1035 	 */
1036 	if (cp_coher_cntl ||
1037 	    (rctx->flags & (SI_CONTEXT_CS_PARTIAL_FLUSH |
1038 			    SI_CONTEXT_INV_VMEM_L1 |
1039 			    SI_CONTEXT_INV_GLOBAL_L2 |
1040 			    SI_CONTEXT_WRITEBACK_GLOBAL_L2))) {
1041 		radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
1042 		radeon_emit(cs, 0);
1043 	}
1044 
1045 	/* SI-CI-VI only:
1046 	 *   When one of the CP_COHER_CNTL.DEST_BASE flags is set, SURFACE_SYNC
1047 	 *   waits for idle, so it should be last. SURFACE_SYNC is done in PFP.
1048 	 *
1049 	 * cp_coher_cntl should contain all necessary flags except TC flags
1050 	 * at this point.
1051 	 *
1052 	 * SI-CIK don't support L2 write-back.
1053 	 */
1054 	if (rctx->flags & SI_CONTEXT_INV_GLOBAL_L2 ||
1055 	    (rctx->chip_class <= CIK &&
1056 	     (rctx->flags & SI_CONTEXT_WRITEBACK_GLOBAL_L2))) {
1057 		/* Invalidate L1 & L2. (L1 is always invalidated on SI)
1058 		 * WB must be set on VI+ when TC_ACTION is set.
1059 		 */
1060 		si_emit_surface_sync(rctx, cp_coher_cntl |
1061 				     S_0085F0_TC_ACTION_ENA(1) |
1062 				     S_0085F0_TCL1_ACTION_ENA(1) |
1063 				     S_0301F0_TC_WB_ACTION_ENA(rctx->chip_class >= VI));
1064 		cp_coher_cntl = 0;
1065 		sctx->b.num_L2_invalidates++;
1066 	} else {
1067 		/* L1 invalidation and L2 writeback must be done separately,
1068 		 * because both operations can't be done together.
1069 		 */
1070 		if (rctx->flags & SI_CONTEXT_WRITEBACK_GLOBAL_L2) {
1071 			/* WB = write-back
1072 			 * NC = apply to non-coherent MTYPEs
1073 			 *      (i.e. MTYPE <= 1, which is what we use everywhere)
1074 			 *
1075 			 * WB doesn't work without NC.
1076 			 */
1077 			si_emit_surface_sync(rctx, cp_coher_cntl |
1078 					     S_0301F0_TC_WB_ACTION_ENA(1) |
1079 					     S_0301F0_TC_NC_ACTION_ENA(1));
1080 			cp_coher_cntl = 0;
1081 			sctx->b.num_L2_writebacks++;
1082 		}
1083 		if (rctx->flags & SI_CONTEXT_INV_VMEM_L1) {
1084 			/* Invalidate per-CU VMEM L1. */
1085 			si_emit_surface_sync(rctx, cp_coher_cntl |
1086 					     S_0085F0_TCL1_ACTION_ENA(1));
1087 			cp_coher_cntl = 0;
1088 		}
1089 	}
1090 
1091 	/* If TC flushes haven't cleared this... */
1092 	if (cp_coher_cntl)
1093 		si_emit_surface_sync(rctx, cp_coher_cntl);
1094 
1095 	if (rctx->flags & SI_CONTEXT_START_PIPELINE_STATS) {
1096 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1097 		radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_START) |
1098 			        EVENT_INDEX(0));
1099 	} else if (rctx->flags & SI_CONTEXT_STOP_PIPELINE_STATS) {
1100 		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
1101 		radeon_emit(cs, EVENT_TYPE(V_028A90_PIPELINESTAT_STOP) |
1102 			        EVENT_INDEX(0));
1103 	}
1104 
1105 	rctx->flags = 0;
1106 }
1107 
si_get_draw_start_count(struct si_context * sctx,const struct pipe_draw_info * info,unsigned * start,unsigned * count)1108 static void si_get_draw_start_count(struct si_context *sctx,
1109 				    const struct pipe_draw_info *info,
1110 				    unsigned *start, unsigned *count)
1111 {
1112 	struct pipe_draw_indirect_info *indirect = info->indirect;
1113 
1114 	if (indirect) {
1115 		unsigned indirect_count;
1116 		struct pipe_transfer *transfer;
1117 		unsigned begin, end;
1118 		unsigned map_size;
1119 		unsigned *data;
1120 
1121 		if (indirect->indirect_draw_count) {
1122 			data = pipe_buffer_map_range(&sctx->b.b,
1123 					indirect->indirect_draw_count,
1124 					indirect->indirect_draw_count_offset,
1125 					sizeof(unsigned),
1126 					PIPE_TRANSFER_READ, &transfer);
1127 
1128 			indirect_count = *data;
1129 
1130 			pipe_buffer_unmap(&sctx->b.b, transfer);
1131 		} else {
1132 			indirect_count = indirect->draw_count;
1133 		}
1134 
1135 		if (!indirect_count) {
1136 			*start = *count = 0;
1137 			return;
1138 		}
1139 
1140 		map_size = (indirect_count - 1) * indirect->stride + 3 * sizeof(unsigned);
1141 		data = pipe_buffer_map_range(&sctx->b.b, indirect->buffer,
1142 					     indirect->offset, map_size,
1143 					     PIPE_TRANSFER_READ, &transfer);
1144 
1145 		begin = UINT_MAX;
1146 		end = 0;
1147 
1148 		for (unsigned i = 0; i < indirect_count; ++i) {
1149 			unsigned count = data[0];
1150 			unsigned start = data[2];
1151 
1152 			if (count > 0) {
1153 				begin = MIN2(begin, start);
1154 				end = MAX2(end, start + count);
1155 			}
1156 
1157 			data += indirect->stride / sizeof(unsigned);
1158 		}
1159 
1160 		pipe_buffer_unmap(&sctx->b.b, transfer);
1161 
1162 		if (begin < end) {
1163 			*start = begin;
1164 			*count = end - begin;
1165 		} else {
1166 			*start = *count = 0;
1167 		}
1168 	} else {
1169 		*start = info->start;
1170 		*count = info->count;
1171 	}
1172 }
1173 
si_emit_all_states(struct si_context * sctx,const struct pipe_draw_info * info,unsigned skip_atom_mask)1174 static void si_emit_all_states(struct si_context *sctx, const struct pipe_draw_info *info,
1175 			       unsigned skip_atom_mask)
1176 {
1177 	/* Emit state atoms. */
1178 	unsigned mask = sctx->dirty_atoms & ~skip_atom_mask;
1179 	while (mask) {
1180 		struct r600_atom *atom = sctx->atoms.array[u_bit_scan(&mask)];
1181 
1182 		atom->emit(&sctx->b, atom);
1183 	}
1184 	sctx->dirty_atoms &= skip_atom_mask;
1185 
1186 	/* Emit states. */
1187 	mask = sctx->dirty_states;
1188 	while (mask) {
1189 		unsigned i = u_bit_scan(&mask);
1190 		struct si_pm4_state *state = sctx->queued.array[i];
1191 
1192 		if (!state || sctx->emitted.array[i] == state)
1193 			continue;
1194 
1195 		si_pm4_emit(sctx, state);
1196 		sctx->emitted.array[i] = state;
1197 	}
1198 	sctx->dirty_states = 0;
1199 
1200 	/* Emit draw states. */
1201 	unsigned num_patches = 0;
1202 
1203 	si_emit_rasterizer_prim_state(sctx);
1204 	if (sctx->tes_shader.cso)
1205 		si_emit_derived_tess_state(sctx, info, &num_patches);
1206 	si_emit_vs_state(sctx, info);
1207 	si_emit_draw_registers(sctx, info, num_patches);
1208 }
1209 
si_draw_vbo(struct pipe_context * ctx,const struct pipe_draw_info * info)1210 void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *info)
1211 {
1212 	struct si_context *sctx = (struct si_context *)ctx;
1213 	struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1214 	struct pipe_resource *indexbuf = info->index.resource;
1215 	unsigned dirty_tex_counter;
1216 	enum pipe_prim_type rast_prim;
1217 	unsigned index_size = info->index_size;
1218 	unsigned index_offset = info->indirect ? info->start * index_size : 0;
1219 
1220 	if (likely(!info->indirect)) {
1221 		/* SI-CI treat instance_count==0 as instance_count==1. There is
1222 		 * no workaround for indirect draws, but we can at least skip
1223 		 * direct draws.
1224 		 */
1225 		if (unlikely(!info->instance_count))
1226 			return;
1227 
1228 		/* Handle count == 0. */
1229 		if (unlikely(!info->count &&
1230 			     (index_size || !info->count_from_stream_output)))
1231 			return;
1232 	}
1233 
1234 	if (unlikely(!sctx->vs_shader.cso)) {
1235 		assert(0);
1236 		return;
1237 	}
1238 	if (unlikely(!sctx->ps_shader.cso && (!rs || !rs->rasterizer_discard))) {
1239 		assert(0);
1240 		return;
1241 	}
1242 	if (unlikely(!!sctx->tes_shader.cso != (info->mode == PIPE_PRIM_PATCHES))) {
1243 		assert(0);
1244 		return;
1245 	}
1246 
1247 	/* Recompute and re-emit the texture resource states if needed. */
1248 	dirty_tex_counter = p_atomic_read(&sctx->b.screen->dirty_tex_counter);
1249 	if (unlikely(dirty_tex_counter != sctx->b.last_dirty_tex_counter)) {
1250 		sctx->b.last_dirty_tex_counter = dirty_tex_counter;
1251 		sctx->framebuffer.dirty_cbufs |=
1252 			((1 << sctx->framebuffer.state.nr_cbufs) - 1);
1253 		sctx->framebuffer.dirty_zsbuf = true;
1254 		si_mark_atom_dirty(sctx, &sctx->framebuffer.atom);
1255 		si_update_all_texture_descriptors(sctx);
1256 	}
1257 
1258 	si_decompress_textures(sctx, u_bit_consecutive(0, SI_NUM_GRAPHICS_SHADERS));
1259 
1260 	/* Set the rasterization primitive type.
1261 	 *
1262 	 * This must be done after si_decompress_textures, which can call
1263 	 * draw_vbo recursively, and before si_update_shaders, which uses
1264 	 * current_rast_prim for this draw_vbo call. */
1265 	if (sctx->gs_shader.cso)
1266 		rast_prim = sctx->gs_shader.cso->gs_output_prim;
1267 	else if (sctx->tes_shader.cso) {
1268 		if (sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_POINT_MODE])
1269 			rast_prim = PIPE_PRIM_POINTS;
1270 		else
1271 			rast_prim = sctx->tes_shader.cso->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
1272 	} else
1273 		rast_prim = info->mode;
1274 
1275 	if (rast_prim != sctx->current_rast_prim) {
1276 		bool old_is_poly = sctx->current_rast_prim >= PIPE_PRIM_TRIANGLES;
1277 		bool new_is_poly = rast_prim >= PIPE_PRIM_TRIANGLES;
1278 		if (old_is_poly != new_is_poly) {
1279 			sctx->scissors.dirty_mask = (1 << SI_MAX_VIEWPORTS) - 1;
1280 			si_mark_atom_dirty(sctx, &sctx->scissors.atom);
1281 		}
1282 
1283 		sctx->current_rast_prim = rast_prim;
1284 		sctx->do_update_shaders = true;
1285 	}
1286 
1287 	if (sctx->tes_shader.cso &&
1288 	    sctx->screen->has_ls_vgpr_init_bug) {
1289 		/* Determine whether the LS VGPR fix should be applied.
1290 		 *
1291 		 * It is only required when num input CPs > num output CPs,
1292 		 * which cannot happen with the fixed function TCS. We should
1293 		 * also update this bit when switching from TCS to fixed
1294 		 * function TCS.
1295 		 */
1296 		struct si_shader_selector *tcs = sctx->tcs_shader.cso;
1297 		bool ls_vgpr_fix =
1298 			tcs &&
1299 			info->vertices_per_patch >
1300 			tcs->info.properties[TGSI_PROPERTY_TCS_VERTICES_OUT];
1301 
1302 		if (ls_vgpr_fix != sctx->ls_vgpr_fix) {
1303 			sctx->ls_vgpr_fix = ls_vgpr_fix;
1304 			sctx->do_update_shaders = true;
1305 		}
1306 	}
1307 
1308 	if (sctx->gs_shader.cso) {
1309 		/* Determine whether the GS triangle strip adjacency fix should
1310 		 * be applied. Rotate every other triangle if
1311 		 * - triangle strips with adjacency are fed to the GS and
1312 		 * - primitive restart is disabled (the rotation doesn't help
1313 		 *   when the restart occurs after an odd number of triangles).
1314 		 */
1315 		bool gs_tri_strip_adj_fix =
1316 			!sctx->tes_shader.cso &&
1317 			info->mode == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY &&
1318 			!info->primitive_restart;
1319 
1320 		if (gs_tri_strip_adj_fix != sctx->gs_tri_strip_adj_fix) {
1321 			sctx->gs_tri_strip_adj_fix = gs_tri_strip_adj_fix;
1322 			sctx->do_update_shaders = true;
1323 		}
1324 	}
1325 
1326 	if (sctx->do_update_shaders && !si_update_shaders(sctx))
1327 		return;
1328 
1329 	if (index_size) {
1330 		/* Translate or upload, if needed. */
1331 		/* 8-bit indices are supported on VI. */
1332 		if (sctx->b.chip_class <= CIK && index_size == 1) {
1333 			unsigned start, count, start_offset, size, offset;
1334 			void *ptr;
1335 
1336 			si_get_draw_start_count(sctx, info, &start, &count);
1337 			start_offset = start * 2;
1338 			size = count * 2;
1339 
1340 			indexbuf = NULL;
1341 			u_upload_alloc(ctx->stream_uploader, start_offset,
1342 				       size,
1343 				       si_optimal_tcc_alignment(sctx, size),
1344 				       &offset, &indexbuf, &ptr);
1345 			if (!indexbuf)
1346 				return;
1347 
1348 			util_shorten_ubyte_elts_to_userptr(&sctx->b.b, info, 0, 0,
1349 							   index_offset + start,
1350 							   count, ptr);
1351 
1352 			/* info->start will be added by the drawing code */
1353 			index_offset = offset - start_offset;
1354 			index_size = 2;
1355 		} else if (info->has_user_indices) {
1356 			unsigned start_offset;
1357 
1358 			assert(!info->indirect);
1359 			start_offset = info->start * index_size;
1360 
1361 			indexbuf = NULL;
1362 			u_upload_data(ctx->stream_uploader, start_offset,
1363 				      info->count * index_size,
1364 				      sctx->screen->info.tcc_cache_line_size,
1365 				      (char*)info->index.user + start_offset,
1366 				      &index_offset, &indexbuf);
1367 			if (!indexbuf)
1368 				return;
1369 
1370 			/* info->start will be added by the drawing code */
1371 			index_offset -= start_offset;
1372 		} else if (sctx->b.chip_class <= CIK &&
1373 			   r600_resource(indexbuf)->TC_L2_dirty) {
1374 			/* VI reads index buffers through TC L2, so it doesn't
1375 			 * need this. */
1376 			sctx->b.flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
1377 			r600_resource(indexbuf)->TC_L2_dirty = false;
1378 		}
1379 	}
1380 
1381 	if (info->indirect) {
1382 		struct pipe_draw_indirect_info *indirect = info->indirect;
1383 
1384 		/* Add the buffer size for memory checking in need_cs_space. */
1385 		si_context_add_resource_size(ctx, indirect->buffer);
1386 
1387 		/* Indirect buffers use TC L2 on GFX9, but not older hw. */
1388 		if (sctx->b.chip_class <= VI) {
1389 			if (r600_resource(indirect->buffer)->TC_L2_dirty) {
1390 				sctx->b.flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
1391 				r600_resource(indirect->buffer)->TC_L2_dirty = false;
1392 			}
1393 
1394 			if (indirect->indirect_draw_count &&
1395 			    r600_resource(indirect->indirect_draw_count)->TC_L2_dirty) {
1396 				sctx->b.flags |= SI_CONTEXT_WRITEBACK_GLOBAL_L2;
1397 				r600_resource(indirect->indirect_draw_count)->TC_L2_dirty = false;
1398 			}
1399 		}
1400 	}
1401 
1402 	si_need_cs_space(sctx);
1403 
1404 	/* Since we've called r600_context_add_resource_size for vertex buffers,
1405 	 * this must be called after si_need_cs_space, because we must let
1406 	 * need_cs_space flush before we add buffers to the buffer list.
1407 	 */
1408 	if (!si_upload_vertex_buffer_descriptors(sctx))
1409 		return;
1410 
1411 	/* Vega10/Raven scissor bug workaround. This must be done before VPORT
1412 	 * scissor registers are changed. There is also a more efficient but
1413 	 * more involved alternative workaround.
1414 	 */
1415 	if ((sctx->b.family == CHIP_VEGA10 || sctx->b.family == CHIP_RAVEN) &&
1416 	    si_is_atom_dirty(sctx, &sctx->scissors.atom)) {
1417 		sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH;
1418 		si_emit_cache_flush(sctx);
1419 	}
1420 
1421 	/* Use optimal packet order based on whether we need to sync the pipeline. */
1422 	if (unlikely(sctx->b.flags & (SI_CONTEXT_FLUSH_AND_INV_CB |
1423 				      SI_CONTEXT_FLUSH_AND_INV_DB |
1424 				      SI_CONTEXT_PS_PARTIAL_FLUSH |
1425 				      SI_CONTEXT_CS_PARTIAL_FLUSH))) {
1426 		/* If we have to wait for idle, set all states first, so that all
1427 		 * SET packets are processed in parallel with previous draw calls.
1428 		 * Then upload descriptors, set shader pointers, and draw, and
1429 		 * prefetch at the end. This ensures that the time the CUs
1430 		 * are idle is very short. (there are only SET_SH packets between
1431 		 * the wait and the draw)
1432 		 */
1433 		struct r600_atom *shader_pointers = &sctx->shader_pointers.atom;
1434 		unsigned masked_atoms = 1u << shader_pointers->id;
1435 
1436 		if (unlikely(sctx->b.flags & SI_CONTEXT_FLUSH_FOR_RENDER_COND))
1437 			masked_atoms |= 1u << sctx->b.render_cond_atom.id;
1438 
1439 		/* Emit all states except shader pointers and render condition. */
1440 		si_emit_all_states(sctx, info, masked_atoms);
1441 		si_emit_cache_flush(sctx);
1442 
1443 		/* <-- CUs are idle here. */
1444 		if (!si_upload_graphics_shader_descriptors(sctx))
1445 			return;
1446 
1447 		/* Set shader pointers after descriptors are uploaded. */
1448 		if (si_is_atom_dirty(sctx, shader_pointers))
1449 			shader_pointers->emit(&sctx->b, NULL);
1450 		if (si_is_atom_dirty(sctx, &sctx->b.render_cond_atom))
1451 			sctx->b.render_cond_atom.emit(&sctx->b, NULL);
1452 		sctx->dirty_atoms = 0;
1453 
1454 		si_emit_draw_packets(sctx, info, indexbuf, index_size, index_offset);
1455 		/* <-- CUs are busy here. */
1456 
1457 		/* Start prefetches after the draw has been started. Both will run
1458 		 * in parallel, but starting the draw first is more important.
1459 		 */
1460 		if (sctx->b.chip_class >= CIK && sctx->prefetch_L2_mask)
1461 			cik_emit_prefetch_L2(sctx);
1462 	} else {
1463 		/* If we don't wait for idle, start prefetches first, then set
1464 		 * states, and draw at the end.
1465 		 */
1466 		if (sctx->b.flags)
1467 			si_emit_cache_flush(sctx);
1468 
1469 		if (sctx->b.chip_class >= CIK && sctx->prefetch_L2_mask)
1470 			cik_emit_prefetch_L2(sctx);
1471 
1472 		if (!si_upload_graphics_shader_descriptors(sctx))
1473 			return;
1474 
1475 		si_emit_all_states(sctx, info, 0);
1476 		si_emit_draw_packets(sctx, info, indexbuf, index_size, index_offset);
1477 	}
1478 
1479 	if (unlikely(sctx->current_saved_cs)) {
1480 		si_trace_emit(sctx);
1481 		si_log_draw_state(sctx, sctx->b.log);
1482 	}
1483 
1484 	/* Workaround for a VGT hang when streamout is enabled.
1485 	 * It must be done after drawing. */
1486 	if ((sctx->b.family == CHIP_HAWAII ||
1487 	     sctx->b.family == CHIP_TONGA ||
1488 	     sctx->b.family == CHIP_FIJI) &&
1489 	    si_get_strmout_en(sctx)) {
1490 		sctx->b.flags |= SI_CONTEXT_VGT_STREAMOUT_SYNC;
1491 	}
1492 
1493 	if (unlikely(sctx->decompression_enabled)) {
1494 		sctx->b.num_decompress_calls++;
1495 	} else {
1496 		sctx->b.num_draw_calls++;
1497 		if (sctx->framebuffer.state.nr_cbufs > 1)
1498 			sctx->b.num_mrt_draw_calls++;
1499 		if (info->primitive_restart)
1500 			sctx->b.num_prim_restart_calls++;
1501 		if (G_0286E8_WAVESIZE(sctx->spi_tmpring_size))
1502 			sctx->b.num_spill_draw_calls++;
1503 	}
1504 	if (index_size && indexbuf != info->index.resource)
1505 		pipe_resource_reference(&indexbuf, NULL);
1506 }
1507 
si_draw_rectangle(struct blitter_context * blitter,void * vertex_elements_cso,blitter_get_vs_func get_vs,int x1,int y1,int x2,int y2,float depth,unsigned num_instances,enum blitter_attrib_type type,const union blitter_attrib * attrib)1508 void si_draw_rectangle(struct blitter_context *blitter,
1509 		       void *vertex_elements_cso,
1510 		       blitter_get_vs_func get_vs,
1511 		       int x1, int y1, int x2, int y2,
1512 		       float depth, unsigned num_instances,
1513 		       enum blitter_attrib_type type,
1514 		       const union blitter_attrib *attrib)
1515 {
1516 	struct pipe_context *pipe = util_blitter_get_pipe(blitter);
1517 	struct si_context *sctx = (struct si_context*)pipe;
1518 
1519 	/* Pack position coordinates as signed int16. */
1520 	sctx->vs_blit_sh_data[0] = (uint32_t)(x1 & 0xffff) |
1521 				   ((uint32_t)(y1 & 0xffff) << 16);
1522 	sctx->vs_blit_sh_data[1] = (uint32_t)(x2 & 0xffff) |
1523 				   ((uint32_t)(y2 & 0xffff) << 16);
1524 	sctx->vs_blit_sh_data[2] = fui(depth);
1525 
1526 	switch (type) {
1527 	case UTIL_BLITTER_ATTRIB_COLOR:
1528 		memcpy(&sctx->vs_blit_sh_data[3], attrib->color,
1529 		       sizeof(float)*4);
1530 		break;
1531 	case UTIL_BLITTER_ATTRIB_TEXCOORD_XY:
1532 	case UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW:
1533 		memcpy(&sctx->vs_blit_sh_data[3], &attrib->texcoord,
1534 		       sizeof(attrib->texcoord));
1535 		break;
1536 	case UTIL_BLITTER_ATTRIB_NONE:;
1537 	}
1538 
1539 	pipe->bind_vs_state(pipe, si_get_blit_vs(sctx, type, num_instances));
1540 
1541 	struct pipe_draw_info info = {};
1542 	info.mode = SI_PRIM_RECTANGLE_LIST;
1543 	info.count = 3;
1544 	info.instance_count = num_instances;
1545 
1546 	/* Don't set per-stage shader pointers for VS. */
1547 	sctx->shader_pointers_dirty &= ~SI_DESCS_SHADER_MASK(VERTEX);
1548 	sctx->vertex_buffer_pointer_dirty = false;
1549 
1550 	si_draw_vbo(pipe, &info);
1551 }
1552 
si_trace_emit(struct si_context * sctx)1553 void si_trace_emit(struct si_context *sctx)
1554 {
1555 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
1556 	uint64_t va = sctx->current_saved_cs->trace_buf->gpu_address;
1557 	uint32_t trace_id = ++sctx->current_saved_cs->trace_id;
1558 
1559 	radeon_emit(cs, PKT3(PKT3_WRITE_DATA, 3, 0));
1560 	radeon_emit(cs, S_370_DST_SEL(V_370_MEMORY_SYNC) |
1561 		    S_370_WR_CONFIRM(1) |
1562 		    S_370_ENGINE_SEL(V_370_ME));
1563 	radeon_emit(cs, va);
1564 	radeon_emit(cs, va >> 32);
1565 	radeon_emit(cs, trace_id);
1566 	radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
1567 	radeon_emit(cs, AC_ENCODE_TRACE_POINT(trace_id));
1568 
1569 	if (sctx->b.log)
1570 		u_log_flush(sctx->b.log);
1571 }
1572