projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: clean up other scratch buffer functions
[mesa.git] / src / gallium / drivers / radeonsi / si_state_shaders.c
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 * Authors:
24 * Christian König <christian.koenig@amd.com>
25 * Marek Olšák <maraeo@gmail.com>
26 */
27
28 #include "si_pipe.h"
29 #include "si_shader.h"
30 #include "sid.h"
31 #include "radeon/r600_cs.h"
32
33 #include "tgsi/tgsi_parse.h"
34 #include "tgsi/tgsi_ureg.h"
35 #include "util/u_memory.h"
36 #include "util/u_simple_shaders.h"
37
38 static void si_set_tesseval_regs(struct si_shader *shader,
39 struct si_pm4_state *pm4)
40 {
41 struct tgsi_shader_info *info = &shader->selector->info;
42 unsigned tes_prim_mode = info->properties[TGSI_PROPERTY_TES_PRIM_MODE];
43 unsigned tes_spacing = info->properties[TGSI_PROPERTY_TES_SPACING];
44 bool tes_vertex_order_cw = info->properties[TGSI_PROPERTY_TES_VERTEX_ORDER_CW];
45 bool tes_point_mode = info->properties[TGSI_PROPERTY_TES_POINT_MODE];
46 unsigned type, partitioning, topology;
47
48 switch (tes_prim_mode) {
49 case PIPE_PRIM_LINES:
50 type = V_028B6C_TESS_ISOLINE;
51 break;
52 case PIPE_PRIM_TRIANGLES:
53 type = V_028B6C_TESS_TRIANGLE;
54 break;
55 case PIPE_PRIM_QUADS:
56 type = V_028B6C_TESS_QUAD;
57 break;
58 default:
59 assert(0);
60 return;
61 }
62
63 switch (tes_spacing) {
64 case PIPE_TESS_SPACING_FRACTIONAL_ODD:
65 partitioning = V_028B6C_PART_FRAC_ODD;
66 break;
67 case PIPE_TESS_SPACING_FRACTIONAL_EVEN:
68 partitioning = V_028B6C_PART_FRAC_EVEN;
69 break;
70 case PIPE_TESS_SPACING_EQUAL:
71 partitioning = V_028B6C_PART_INTEGER;
72 break;
73 default:
74 assert(0);
75 return;
76 }
77
78 if (tes_point_mode)
79 topology = V_028B6C_OUTPUT_POINT;
80 else if (tes_prim_mode == PIPE_PRIM_LINES)
81 topology = V_028B6C_OUTPUT_LINE;
82 else if (tes_vertex_order_cw)
83 /* for some reason, this must be the other way around */
84 topology = V_028B6C_OUTPUT_TRIANGLE_CCW;
85 else
86 topology = V_028B6C_OUTPUT_TRIANGLE_CW;
87
88 si_pm4_set_reg(pm4, R_028B6C_VGT_TF_PARAM,
89 S_028B6C_TYPE(type) |
90 S_028B6C_PARTITIONING(partitioning) |
91 S_028B6C_TOPOLOGY(topology));
92 }
93
94 static void si_shader_ls(struct si_shader *shader)
95 {
96 struct si_pm4_state *pm4;
97 unsigned num_sgprs, num_user_sgprs;
98 unsigned vgpr_comp_cnt;
99 uint64_t va;
100
101 pm4 = shader->pm4 = CALLOC_STRUCT(si_pm4_state);
102 if (pm4 == NULL)
103 return;
104
105 va = shader->bo->gpu_address;
106 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
107
108 /* We need at least 2 components for LS.
109 * VGPR0-3: (VertexID, RelAutoindex, ???, InstanceID). */
110 vgpr_comp_cnt = shader->uses_instanceid ? 3 : 1;
111
112 num_user_sgprs = SI_LS_NUM_USER_SGPR;
113 num_sgprs = shader->num_sgprs;
114 if (num_user_sgprs > num_sgprs) {
115 /* Last 2 reserved SGPRs are used for VCC */
116 num_sgprs = num_user_sgprs + 2;
117 }
118 assert(num_sgprs <= 104);
119
120 si_pm4_set_reg(pm4, R_00B520_SPI_SHADER_PGM_LO_LS, va >> 8);
121 si_pm4_set_reg(pm4, R_00B524_SPI_SHADER_PGM_HI_LS, va >> 40);
122
123 shader->ls_rsrc1 = S_00B528_VGPRS((shader->num_vgprs - 1) / 4) |
124 S_00B528_SGPRS((num_sgprs - 1) / 8) |
125 S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt) |
126 S_00B528_DX10_CLAMP(shader->dx10_clamp_mode);
127 shader->ls_rsrc2 = S_00B52C_USER_SGPR(num_user_sgprs) |
128 S_00B52C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0);
129 }
130
131 static void si_shader_hs(struct si_shader *shader)
132 {
133 struct si_pm4_state *pm4;
134 unsigned num_sgprs, num_user_sgprs;
135 uint64_t va;
136
137 pm4 = shader->pm4 = CALLOC_STRUCT(si_pm4_state);
138 if (pm4 == NULL)
139 return;
140
141 va = shader->bo->gpu_address;
142 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
143
144 num_user_sgprs = SI_TCS_NUM_USER_SGPR;
145 num_sgprs = shader->num_sgprs;
146 /* One SGPR after user SGPRs is pre-loaded with tessellation factor
147 * buffer offset. */
148 if ((num_user_sgprs + 1) > num_sgprs) {
149 /* Last 2 reserved SGPRs are used for VCC */
150 num_sgprs = num_user_sgprs + 1 + 2;
151 }
152 assert(num_sgprs <= 104);
153
154 si_pm4_set_reg(pm4, R_00B420_SPI_SHADER_PGM_LO_HS, va >> 8);
155 si_pm4_set_reg(pm4, R_00B424_SPI_SHADER_PGM_HI_HS, va >> 40);
156 si_pm4_set_reg(pm4, R_00B428_SPI_SHADER_PGM_RSRC1_HS,
157 S_00B428_VGPRS((shader->num_vgprs - 1) / 4) |
158 S_00B428_SGPRS((num_sgprs - 1) / 8) |
159 S_00B428_DX10_CLAMP(shader->dx10_clamp_mode));
160 si_pm4_set_reg(pm4, R_00B42C_SPI_SHADER_PGM_RSRC2_HS,
161 S_00B42C_USER_SGPR(num_user_sgprs) |
162 S_00B42C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0));
163 }
164
165 static void si_shader_es(struct si_shader *shader)
166 {
167 struct si_pm4_state *pm4;
168 unsigned num_sgprs, num_user_sgprs;
169 unsigned vgpr_comp_cnt;
170 uint64_t va;
171
172 pm4 = shader->pm4 = CALLOC_STRUCT(si_pm4_state);
173
174 if (pm4 == NULL)
175 return;
176
177 va = shader->bo->gpu_address;
178 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
179
180 if (shader->selector->type == PIPE_SHADER_VERTEX) {
181 vgpr_comp_cnt = shader->uses_instanceid ? 3 : 0;
182 num_user_sgprs = SI_VS_NUM_USER_SGPR;
183 } else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
184 vgpr_comp_cnt = 3; /* all components are needed for TES */
185 num_user_sgprs = SI_TES_NUM_USER_SGPR;
186 } else
187 unreachable("invalid shader selector type");
188
189 num_sgprs = shader->num_sgprs;
190 /* One SGPR after user SGPRs is pre-loaded with es2gs_offset */
191 if ((num_user_sgprs + 1) > num_sgprs) {
192 /* Last 2 reserved SGPRs are used for VCC */
193 num_sgprs = num_user_sgprs + 1 + 2;
194 }
195 assert(num_sgprs <= 104);
196
197 si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
198 si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, va >> 40);
199 si_pm4_set_reg(pm4, R_00B328_SPI_SHADER_PGM_RSRC1_ES,
200 S_00B328_VGPRS((shader->num_vgprs - 1) / 4) |
201 S_00B328_SGPRS((num_sgprs - 1) / 8) |
202 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt) |
203 S_00B328_DX10_CLAMP(shader->dx10_clamp_mode));
204 si_pm4_set_reg(pm4, R_00B32C_SPI_SHADER_PGM_RSRC2_ES,
205 S_00B32C_USER_SGPR(num_user_sgprs) |
206 S_00B32C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0));
207
208 if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
209 si_set_tesseval_regs(shader, pm4);
210 }
211
212 static unsigned si_gs_get_max_stream(struct si_shader *shader)
213 {
214 struct pipe_stream_output_info *so = &shader->selector->so;
215 unsigned max_stream = 0, i;
216
217 if (so->num_outputs == 0)
218 return 0;
219
220 for (i = 0; i < so->num_outputs; i++) {
221 if (so->output[i].stream > max_stream)
222 max_stream = so->output[i].stream;
223 }
224 return max_stream;
225 }
226
227 static void si_shader_gs(struct si_shader *shader)
228 {
229 unsigned gs_vert_itemsize = shader->selector->info.num_outputs * 16;
230 unsigned gs_max_vert_out = shader->selector->gs_max_out_vertices;
231 unsigned gsvs_itemsize = (gs_vert_itemsize * gs_max_vert_out) >> 2;
232 unsigned gs_num_invocations = shader->selector->gs_num_invocations;
233 unsigned cut_mode;
234 struct si_pm4_state *pm4;
235 unsigned num_sgprs, num_user_sgprs;
236 uint64_t va;
237 unsigned max_stream = si_gs_get_max_stream(shader);
238
239 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
240 assert(gsvs_itemsize < (1 << 15));
241
242 pm4 = shader->pm4 = CALLOC_STRUCT(si_pm4_state);
243
244 if (pm4 == NULL)
245 return;
246
247 if (gs_max_vert_out <= 128) {
248 cut_mode = V_028A40_GS_CUT_128;
249 } else if (gs_max_vert_out <= 256) {
250 cut_mode = V_028A40_GS_CUT_256;
251 } else if (gs_max_vert_out <= 512) {
252 cut_mode = V_028A40_GS_CUT_512;
253 } else {
254 assert(gs_max_vert_out <= 1024);
255 cut_mode = V_028A40_GS_CUT_1024;
256 }
257
258 si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE,
259 S_028A40_MODE(V_028A40_GS_SCENARIO_G) |
260 S_028A40_CUT_MODE(cut_mode)|
261 S_028A40_ES_WRITE_OPTIMIZE(1) |
262 S_028A40_GS_WRITE_OPTIMIZE(1));
263
264 si_pm4_set_reg(pm4, R_028A60_VGT_GSVS_RING_OFFSET_1, gsvs_itemsize);
265 si_pm4_set_reg(pm4, R_028A64_VGT_GSVS_RING_OFFSET_2, gsvs_itemsize * ((max_stream >= 2) ? 2 : 1));
266 si_pm4_set_reg(pm4, R_028A68_VGT_GSVS_RING_OFFSET_3, gsvs_itemsize * ((max_stream >= 3) ? 3 : 1));
267
268 si_pm4_set_reg(pm4, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
269 util_bitcount64(shader->selector->inputs_read) * (16 >> 2));
270 si_pm4_set_reg(pm4, R_028AB0_VGT_GSVS_RING_ITEMSIZE, gsvs_itemsize * (max_stream + 1));
271
272 si_pm4_set_reg(pm4, R_028B38_VGT_GS_MAX_VERT_OUT, gs_max_vert_out);
273
274 si_pm4_set_reg(pm4, R_028B5C_VGT_GS_VERT_ITEMSIZE, gs_vert_itemsize >> 2);
275 si_pm4_set_reg(pm4, R_028B60_VGT_GS_VERT_ITEMSIZE_1, (max_stream >= 1) ? gs_vert_itemsize >> 2 : 0);
276 si_pm4_set_reg(pm4, R_028B64_VGT_GS_VERT_ITEMSIZE_2, (max_stream >= 2) ? gs_vert_itemsize >> 2 : 0);
277 si_pm4_set_reg(pm4, R_028B68_VGT_GS_VERT_ITEMSIZE_3, (max_stream >= 3) ? gs_vert_itemsize >> 2 : 0);
278
279 si_pm4_set_reg(pm4, R_028B90_VGT_GS_INSTANCE_CNT,
280 S_028B90_CNT(MIN2(gs_num_invocations, 127)) |
281 S_028B90_ENABLE(gs_num_invocations > 0));
282
283 va = shader->bo->gpu_address;
284 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
285 si_pm4_set_reg(pm4, R_00B220_SPI_SHADER_PGM_LO_GS, va >> 8);
286 si_pm4_set_reg(pm4, R_00B224_SPI_SHADER_PGM_HI_GS, va >> 40);
287
288 num_user_sgprs = SI_GS_NUM_USER_SGPR;
289 num_sgprs = shader->num_sgprs;
290 /* Two SGPRs after user SGPRs are pre-loaded with gs2vs_offset, gs_wave_id */
291 if ((num_user_sgprs + 2) > num_sgprs) {
292 /* Last 2 reserved SGPRs are used for VCC */
293 num_sgprs = num_user_sgprs + 2 + 2;
294 }
295 assert(num_sgprs <= 104);
296
297 si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
298 S_00B228_VGPRS((shader->num_vgprs - 1) / 4) |
299 S_00B228_SGPRS((num_sgprs - 1) / 8) |
300 S_00B228_DX10_CLAMP(shader->dx10_clamp_mode));
301 si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
302 S_00B22C_USER_SGPR(num_user_sgprs) |
303 S_00B22C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0));
304 }
305
306 static void si_shader_vs(struct si_shader *shader)
307 {
308 struct si_pm4_state *pm4;
309 unsigned num_sgprs, num_user_sgprs;
310 unsigned nparams, vgpr_comp_cnt;
311 uint64_t va;
312 unsigned window_space =
313 shader->selector->info.properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
314 bool enable_prim_id = si_vs_exports_prim_id(shader);
315
316 pm4 = shader->pm4 = CALLOC_STRUCT(si_pm4_state);
317
318 if (pm4 == NULL)
319 return;
320
321 /* If this is the GS copy shader, the GS state writes this register.
322 * Otherwise, the VS state writes it.
323 */
324 if (!shader->is_gs_copy_shader) {
325 si_pm4_set_reg(pm4, R_028A40_VGT_GS_MODE,
326 S_028A40_MODE(enable_prim_id ? V_028A40_GS_SCENARIO_A : 0));
327 si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, enable_prim_id);
328 } else
329 si_pm4_set_reg(pm4, R_028A84_VGT_PRIMITIVEID_EN, 0);
330
331 va = shader->bo->gpu_address;
332 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
333
334 if (shader->is_gs_copy_shader) {
335 vgpr_comp_cnt = 0; /* only VertexID is needed for GS-COPY. */
336 num_user_sgprs = SI_GSCOPY_NUM_USER_SGPR;
337 } else if (shader->selector->type == PIPE_SHADER_VERTEX) {
338 vgpr_comp_cnt = shader->uses_instanceid ? 3 : (enable_prim_id ? 2 : 0);
339 num_user_sgprs = SI_VS_NUM_USER_SGPR;
340 } else if (shader->selector->type == PIPE_SHADER_TESS_EVAL) {
341 vgpr_comp_cnt = 3; /* all components are needed for TES */
342 num_user_sgprs = SI_TES_NUM_USER_SGPR;
343 } else
344 unreachable("invalid shader selector type");
345
346 num_sgprs = shader->num_sgprs;
347 if (num_user_sgprs > num_sgprs) {
348 /* Last 2 reserved SGPRs are used for VCC */
349 num_sgprs = num_user_sgprs + 2;
350 }
351 assert(num_sgprs <= 104);
352
353 /* VS is required to export at least one param. */
354 nparams = MAX2(shader->nr_param_exports, 1);
355 si_pm4_set_reg(pm4, R_0286C4_SPI_VS_OUT_CONFIG,
356 S_0286C4_VS_EXPORT_COUNT(nparams - 1));
357
358 si_pm4_set_reg(pm4, R_02870C_SPI_SHADER_POS_FORMAT,
359 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
360 S_02870C_POS1_EXPORT_FORMAT(shader->nr_pos_exports > 1 ?
361 V_02870C_SPI_SHADER_4COMP :
362 V_02870C_SPI_SHADER_NONE) |
363 S_02870C_POS2_EXPORT_FORMAT(shader->nr_pos_exports > 2 ?
364 V_02870C_SPI_SHADER_4COMP :
365 V_02870C_SPI_SHADER_NONE) |
366 S_02870C_POS3_EXPORT_FORMAT(shader->nr_pos_exports > 3 ?
367 V_02870C_SPI_SHADER_4COMP :
368 V_02870C_SPI_SHADER_NONE));
369
370 si_pm4_set_reg(pm4, R_00B120_SPI_SHADER_PGM_LO_VS, va >> 8);
371 si_pm4_set_reg(pm4, R_00B124_SPI_SHADER_PGM_HI_VS, va >> 40);
372 si_pm4_set_reg(pm4, R_00B128_SPI_SHADER_PGM_RSRC1_VS,
373 S_00B128_VGPRS((shader->num_vgprs - 1) / 4) |
374 S_00B128_SGPRS((num_sgprs - 1) / 8) |
375 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt) |
376 S_00B128_DX10_CLAMP(shader->dx10_clamp_mode));
377 si_pm4_set_reg(pm4, R_00B12C_SPI_SHADER_PGM_RSRC2_VS,
378 S_00B12C_USER_SGPR(num_user_sgprs) |
379 S_00B12C_SO_BASE0_EN(!!shader->selector->so.stride[0]) |
380 S_00B12C_SO_BASE1_EN(!!shader->selector->so.stride[1]) |
381 S_00B12C_SO_BASE2_EN(!!shader->selector->so.stride[2]) |
382 S_00B12C_SO_BASE3_EN(!!shader->selector->so.stride[3]) |
383 S_00B12C_SO_EN(!!shader->selector->so.num_outputs) |
384 S_00B12C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0));
385 if (window_space)
386 si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
387 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1));
388 else
389 si_pm4_set_reg(pm4, R_028818_PA_CL_VTE_CNTL,
390 S_028818_VTX_W0_FMT(1) |
391 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
392 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
393 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
394
395 if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
396 si_set_tesseval_regs(shader, pm4);
397 }
398
399 static void si_shader_ps(struct si_shader *shader)
400 {
401 struct tgsi_shader_info *info = &shader->selector->info;
402 struct si_pm4_state *pm4;
403 unsigned i, spi_ps_in_control;
404 unsigned num_sgprs, num_user_sgprs;
405 unsigned spi_baryc_cntl = 0;
406 uint64_t va;
407
408 pm4 = shader->pm4 = CALLOC_STRUCT(si_pm4_state);
409
410 if (pm4 == NULL)
411 return;
412
413 for (i = 0; i < info->num_inputs; i++) {
414 switch (info->input_semantic_name[i]) {
415 case TGSI_SEMANTIC_POSITION:
416 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
417 * Possible vaules:
418 * 0 -> Position = pixel center (default)
419 * 1 -> Position = pixel centroid
420 * 2 -> Position = at sample position
421 */
422 switch (info->input_interpolate_loc[i]) {
423 case TGSI_INTERPOLATE_LOC_CENTROID:
424 spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(1);
425 break;
426 case TGSI_INTERPOLATE_LOC_SAMPLE:
427 spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2);
428 break;
429 }
430
431 if (info->properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] ==
432 TGSI_FS_COORD_PIXEL_CENTER_INTEGER)
433 spi_baryc_cntl |= S_0286E0_POS_FLOAT_ULC(1);
434 break;
435 }
436 }
437
438 spi_ps_in_control = S_0286D8_NUM_INTERP(shader->nparam) |
439 S_0286D8_BC_OPTIMIZE_DISABLE(1);
440
441 si_pm4_set_reg(pm4, R_0286E0_SPI_BARYC_CNTL, spi_baryc_cntl);
442 si_pm4_set_reg(pm4, R_0286D8_SPI_PS_IN_CONTROL, spi_ps_in_control);
443
444 si_pm4_set_reg(pm4, R_028710_SPI_SHADER_Z_FORMAT, shader->spi_shader_z_format);
445 si_pm4_set_reg(pm4, R_028714_SPI_SHADER_COL_FORMAT,
446 shader->spi_shader_col_format);
447 si_pm4_set_reg(pm4, R_02823C_CB_SHADER_MASK, shader->cb_shader_mask);
448
449 va = shader->bo->gpu_address;
450 si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_USER_SHADER);
451 si_pm4_set_reg(pm4, R_00B020_SPI_SHADER_PGM_LO_PS, va >> 8);
452 si_pm4_set_reg(pm4, R_00B024_SPI_SHADER_PGM_HI_PS, va >> 40);
453
454 num_user_sgprs = SI_PS_NUM_USER_SGPR;
455 num_sgprs = shader->num_sgprs;
456 /* One SGPR after user SGPRs is pre-loaded with {prim_mask, lds_offset} */
457 if ((num_user_sgprs + 1) > num_sgprs) {
458 /* Last 2 reserved SGPRs are used for VCC */
459 num_sgprs = num_user_sgprs + 1 + 2;
460 }
461 assert(num_sgprs <= 104);
462
463 si_pm4_set_reg(pm4, R_00B028_SPI_SHADER_PGM_RSRC1_PS,
464 S_00B028_VGPRS((shader->num_vgprs - 1) / 4) |
465 S_00B028_SGPRS((num_sgprs - 1) / 8) |
466 S_00B028_DX10_CLAMP(shader->dx10_clamp_mode));
467 si_pm4_set_reg(pm4, R_00B02C_SPI_SHADER_PGM_RSRC2_PS,
468 S_00B02C_EXTRA_LDS_SIZE(shader->lds_size) |
469 S_00B02C_USER_SGPR(num_user_sgprs) |
470 S_00B32C_SCRATCH_EN(shader->scratch_bytes_per_wave > 0));
471 }
472
473 static void si_shader_init_pm4_state(struct si_shader *shader)
474 {
475
476 if (shader->pm4)
477 si_pm4_free_state_simple(shader->pm4);
478
479 switch (shader->selector->type) {
480 case PIPE_SHADER_VERTEX:
481 if (shader->key.vs.as_ls)
482 si_shader_ls(shader);
483 else if (shader->key.vs.as_es)
484 si_shader_es(shader);
485 else
486 si_shader_vs(shader);
487 break;
488 case PIPE_SHADER_TESS_CTRL:
489 si_shader_hs(shader);
490 break;
491 case PIPE_SHADER_TESS_EVAL:
492 if (shader->key.tes.as_es)
493 si_shader_es(shader);
494 else
495 si_shader_vs(shader);
496 break;
497 case PIPE_SHADER_GEOMETRY:
498 si_shader_gs(shader);
499 si_shader_vs(shader->gs_copy_shader);
500 break;
501 case PIPE_SHADER_FRAGMENT:
502 si_shader_ps(shader);
503 break;
504 default:
505 assert(0);
506 }
507 }
508
509 /* Compute the key for the hw shader variant */
510 static inline void si_shader_selector_key(struct pipe_context *ctx,
511 struct si_shader_selector *sel,
512 union si_shader_key *key)
513 {
514 struct si_context *sctx = (struct si_context *)ctx;
515 unsigned i;
516
517 memset(key, 0, sizeof(*key));
518
519 switch (sel->type) {
520 case PIPE_SHADER_VERTEX:
521 if (sctx->vertex_elements)
522 for (i = 0; i < sctx->vertex_elements->count; ++i)
523 key->vs.instance_divisors[i] =
524 sctx->vertex_elements->elements[i].instance_divisor;
525
526 if (sctx->tes_shader)
527 key->vs.as_ls = 1;
528 else if (sctx->gs_shader) {
529 key->vs.as_es = 1;
530 key->vs.es_enabled_outputs = sctx->gs_shader->inputs_read;
531 }
532
533 if (!sctx->gs_shader && sctx->ps_shader &&
534 sctx->ps_shader->info.uses_primid)
535 key->vs.export_prim_id = 1;
536 break;
537 case PIPE_SHADER_TESS_CTRL:
538 key->tcs.prim_mode =
539 sctx->tes_shader->info.properties[TGSI_PROPERTY_TES_PRIM_MODE];
540 break;
541 case PIPE_SHADER_TESS_EVAL:
542 if (sctx->gs_shader) {
543 key->tes.as_es = 1;
544 key->tes.es_enabled_outputs = sctx->gs_shader->inputs_read;
545 } else if (sctx->ps_shader && sctx->ps_shader->info.uses_primid)
546 key->tes.export_prim_id = 1;
547 break;
548 case PIPE_SHADER_GEOMETRY:
549 break;
550 case PIPE_SHADER_FRAGMENT: {
551 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
552
553 if (sel->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS])
554 key->ps.last_cbuf = MAX2(sctx->framebuffer.state.nr_cbufs, 1) - 1;
555 key->ps.export_16bpc = sctx->framebuffer.export_16bpc;
556
557 if (rs) {
558 bool is_poly = (sctx->current_rast_prim >= PIPE_PRIM_TRIANGLES &&
559 sctx->current_rast_prim <= PIPE_PRIM_POLYGON) ||
560 sctx->current_rast_prim >= PIPE_PRIM_TRIANGLES_ADJACENCY;
561 bool is_line = !is_poly && sctx->current_rast_prim != PIPE_PRIM_POINTS;
562
563 key->ps.color_two_side = rs->two_side;
564
565 if (sctx->queued.named.blend) {
566 key->ps.alpha_to_one = sctx->queued.named.blend->alpha_to_one &&
567 rs->multisample_enable &&
568 !sctx->framebuffer.cb0_is_integer;
569 }
570
571 key->ps.poly_stipple = rs->poly_stipple_enable && is_poly;
572 key->ps.poly_line_smoothing = ((is_poly && rs->poly_smooth) ||
573 (is_line && rs->line_smooth)) &&
574 sctx->framebuffer.nr_samples <= 1;
575 }
576
577 key->ps.alpha_func = PIPE_FUNC_ALWAYS;
578 /* Alpha-test should be disabled if colorbuffer 0 is integer. */
579 if (sctx->queued.named.dsa &&
580 !sctx->framebuffer.cb0_is_integer)
581 key->ps.alpha_func = sctx->queued.named.dsa->alpha_func;
582 break;
583 }
584 default:
585 assert(0);
586 }
587 }
588
589 /* Select the hw shader variant depending on the current state. */
590 static int si_shader_select(struct pipe_context *ctx,
591 struct si_shader_selector *sel)
592 {
593 struct si_context *sctx = (struct si_context *)ctx;
594 union si_shader_key key;
595 struct si_shader * shader = NULL;
596 int r;
597
598 si_shader_selector_key(ctx, sel, &key);
599
600 /* Check if we don't need to change anything.
601 * This path is also used for most shaders that don't need multiple
602 * variants, it will cost just a computation of the key and this
603 * test. */
604 if (likely(sel->current && memcmp(&sel->current->key, &key, sizeof(key)) == 0)) {
605 return 0;
606 }
607
608 /* lookup if we have other variants in the list */
609 if (sel->num_shaders > 1) {
610 struct si_shader *p = sel->current, *c = p->next_variant;
611
612 while (c && memcmp(&c->key, &key, sizeof(key)) != 0) {
613 p = c;
614 c = c->next_variant;
615 }
616
617 if (c) {
618 p->next_variant = c->next_variant;
619 shader = c;
620 }
621 }
622
623 if (shader) {
624 shader->next_variant = sel->current;
625 sel->current = shader;
626 } else {
627 shader = CALLOC(1, sizeof(struct si_shader));
628 shader->selector = sel;
629 shader->key = key;
630
631 shader->next_variant = sel->current;
632 sel->current = shader;
633 r = si_shader_create((struct si_screen*)ctx->screen, sctx->tm,
634 shader);
635 if (unlikely(r)) {
636 R600_ERR("Failed to build shader variant (type=%u) %d\n",
637 sel->type, r);
638 sel->current = NULL;
639 FREE(shader);
640 return r;
641 }
642 si_shader_init_pm4_state(shader);
643 sel->num_shaders++;
644 p_atomic_inc(&sctx->screen->b.num_compilations);
645 }
646
647 return 0;
648 }
649
650 static void *si_create_shader_selector(struct pipe_context *ctx,
651 const struct pipe_shader_state *state)
652 {
653 struct si_screen *sscreen = (struct si_screen *)ctx->screen;
654 struct si_shader_selector *sel = CALLOC_STRUCT(si_shader_selector);
655 int i;
656
657 if (!sel)
658 return NULL;
659
660 sel->tokens = tgsi_dup_tokens(state->tokens);
661 if (!sel->tokens) {
662 FREE(sel);
663 return NULL;
664 }
665
666 sel->so = state->stream_output;
667 tgsi_scan_shader(state->tokens, &sel->info);
668 sel->type = util_pipe_shader_from_tgsi_processor(sel->info.processor);
669 p_atomic_inc(&sscreen->b.num_shaders_created);
670
671 /* First set which opcode uses which (i,j) pair. */
672 if (sel->info.uses_persp_opcode_interp_centroid)
673 sel->info.uses_persp_centroid = true;
674
675 if (sel->info.uses_linear_opcode_interp_centroid)
676 sel->info.uses_linear_centroid = true;
677
678 if (sel->info.uses_persp_opcode_interp_offset ||
679 sel->info.uses_persp_opcode_interp_sample)
680 sel->info.uses_persp_center = true;
681
682 if (sel->info.uses_linear_opcode_interp_offset ||
683 sel->info.uses_linear_opcode_interp_sample)
684 sel->info.uses_linear_center = true;
685
686 /* Determine if the shader has to use a conditional assignment when
687 * emulating force_persample_interp.
688 */
689 sel->forces_persample_interp_for_persp =
690 sel->info.uses_persp_center +
691 sel->info.uses_persp_centroid +
692 sel->info.uses_persp_sample >= 2;
693
694 sel->forces_persample_interp_for_linear =
695 sel->info.uses_linear_center +
696 sel->info.uses_linear_centroid +
697 sel->info.uses_linear_sample >= 2;
698
699 switch (sel->type) {
700 case PIPE_SHADER_GEOMETRY:
701 sel->gs_output_prim =
702 sel->info.properties[TGSI_PROPERTY_GS_OUTPUT_PRIM];
703 sel->gs_max_out_vertices =
704 sel->info.properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES];
705 sel->gs_num_invocations =
706 sel->info.properties[TGSI_PROPERTY_GS_INVOCATIONS];
707 sel->gsvs_itemsize = sel->info.num_outputs * 16 *
708 sel->gs_max_out_vertices;
709
710 for (i = 0; i < sel->info.num_inputs; i++) {
711 unsigned name = sel->info.input_semantic_name[i];
712 unsigned index = sel->info.input_semantic_index[i];
713
714 switch (name) {
715 case TGSI_SEMANTIC_PRIMID:
716 break;
717 default:
718 sel->inputs_read |=
719 1llu << si_shader_io_get_unique_index(name, index);
720 }
721 }
722 break;
723
724 case PIPE_SHADER_VERTEX:
725 case PIPE_SHADER_TESS_CTRL:
726 for (i = 0; i < sel->info.num_outputs; i++) {
727 unsigned name = sel->info.output_semantic_name[i];
728 unsigned index = sel->info.output_semantic_index[i];
729
730 switch (name) {
731 case TGSI_SEMANTIC_TESSINNER:
732 case TGSI_SEMANTIC_TESSOUTER:
733 case TGSI_SEMANTIC_PATCH:
734 sel->patch_outputs_written |=
735 1llu << si_shader_io_get_unique_index(name, index);
736 break;
737 default:
738 sel->outputs_written |=
739 1llu << si_shader_io_get_unique_index(name, index);
740 }
741 }
742 break;
743 case PIPE_SHADER_FRAGMENT:
744 for (i = 0; i < sel->info.num_outputs; i++) {
745 unsigned name = sel->info.output_semantic_name[i];
746 unsigned index = sel->info.output_semantic_index[i];
747
748 if (name == TGSI_SEMANTIC_COLOR)
749 sel->ps_colors_written |= 1 << index;
750 }
751 break;
752 }
753
754 if (sscreen->b.debug_flags & DBG_PRECOMPILE)
755 if (si_shader_select(ctx, sel)) {
756 fprintf(stderr, "radeonsi: can't create a shader\n");
757 tgsi_free_tokens(sel->tokens);
758 FREE(sel);
759 return NULL;
760 }
761
762 return sel;
763 }
764
765 /**
766 * Normally, we only emit 1 viewport and 1 scissor if no shader is using
767 * the VIEWPORT_INDEX output, and emitting the other viewports and scissors
768 * is delayed. When a shader with VIEWPORT_INDEX appears, this should be
769 * called to emit the rest.
770 */
771 static void si_update_viewports_and_scissors(struct si_context *sctx)
772 {
773 struct tgsi_shader_info *info = si_get_vs_info(sctx);
774
775 if (!info || !info->writes_viewport_index)
776 return;
777
778 if (sctx->scissors.dirty_mask)
779 si_mark_atom_dirty(sctx, &sctx->scissors.atom);
780 if (sctx->viewports.dirty_mask)
781 si_mark_atom_dirty(sctx, &sctx->viewports.atom);
782 }
783
784 static void si_bind_vs_shader(struct pipe_context *ctx, void *state)
785 {
786 struct si_context *sctx = (struct si_context *)ctx;
787 struct si_shader_selector *sel = state;
788
789 if (sctx->vs_shader == sel || !sel)
790 return;
791
792 sctx->vs_shader = sel;
793 si_mark_atom_dirty(sctx, &sctx->clip_regs);
794 si_update_viewports_and_scissors(sctx);
795 }
796
797 static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
798 {
799 struct si_context *sctx = (struct si_context *)ctx;
800 struct si_shader_selector *sel = state;
801 bool enable_changed = !!sctx->gs_shader != !!sel;
802
803 if (sctx->gs_shader == sel)
804 return;
805
806 sctx->gs_shader = sel;
807 si_mark_atom_dirty(sctx, &sctx->clip_regs);
808 sctx->last_rast_prim = -1; /* reset this so that it gets updated */
809
810 if (enable_changed)
811 si_shader_change_notify(sctx);
812 si_update_viewports_and_scissors(sctx);
813 }
814
815 static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
816 {
817 struct si_context *sctx = (struct si_context *)ctx;
818 struct si_shader_selector *sel = state;
819 bool enable_changed = !!sctx->tcs_shader != !!sel;
820
821 if (sctx->tcs_shader == sel)
822 return;
823
824 sctx->tcs_shader = sel;
825
826 if (enable_changed)
827 sctx->last_tcs = NULL; /* invalidate derived tess state */
828 }
829
830 static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
831 {
832 struct si_context *sctx = (struct si_context *)ctx;
833 struct si_shader_selector *sel = state;
834 bool enable_changed = !!sctx->tes_shader != !!sel;
835
836 if (sctx->tes_shader == sel)
837 return;
838
839 sctx->tes_shader = sel;
840 si_mark_atom_dirty(sctx, &sctx->clip_regs);
841 sctx->last_rast_prim = -1; /* reset this so that it gets updated */
842
843 if (enable_changed) {
844 si_shader_change_notify(sctx);
845 sctx->last_tes_sh_base = -1; /* invalidate derived tess state */
846 }
847 si_update_viewports_and_scissors(sctx);
848 }
849
850 static void si_make_dummy_ps(struct si_context *sctx)
851 {
852 if (!sctx->dummy_pixel_shader) {
853 sctx->dummy_pixel_shader =
854 util_make_fragment_cloneinput_shader(&sctx->b.b, 0,
855 TGSI_SEMANTIC_GENERIC,
856 TGSI_INTERPOLATE_CONSTANT);
857 }
858 }
859
860 static void si_bind_ps_shader(struct pipe_context *ctx, void *state)
861 {
862 struct si_context *sctx = (struct si_context *)ctx;
863 struct si_shader_selector *sel = state;
864
865 /* skip if supplied shader is one already in use */
866 if (sctx->ps_shader == sel)
867 return;
868
869 /* use a dummy shader if binding a NULL shader */
870 if (!sel) {
871 si_make_dummy_ps(sctx);
872 sel = sctx->dummy_pixel_shader;
873 }
874
875 sctx->ps_shader = sel;
876 si_mark_atom_dirty(sctx, &sctx->cb_target_mask);
877 }
878
879 static void si_delete_shader_selector(struct pipe_context *ctx, void *state)
880 {
881 struct si_context *sctx = (struct si_context *)ctx;
882 struct si_shader_selector *sel = (struct si_shader_selector *)state;
883 struct si_shader *p = sel->current, *c;
884 struct si_shader_selector **current_shader[SI_NUM_SHADERS] = {
885 [PIPE_SHADER_VERTEX] = &sctx->vs_shader,
886 [PIPE_SHADER_TESS_CTRL] = &sctx->tcs_shader,
887 [PIPE_SHADER_TESS_EVAL] = &sctx->tes_shader,
888 [PIPE_SHADER_GEOMETRY] = &sctx->gs_shader,
889 [PIPE_SHADER_FRAGMENT] = &sctx->ps_shader,
890 };
891
892 if (*current_shader[sel->type] == sel)
893 *current_shader[sel->type] = NULL;
894
895 while (p) {
896 c = p->next_variant;
897 switch (sel->type) {
898 case PIPE_SHADER_VERTEX:
899 if (p->key.vs.as_ls)
900 si_pm4_delete_state(sctx, ls, p->pm4);
901 else if (p->key.vs.as_es)
902 si_pm4_delete_state(sctx, es, p->pm4);
903 else
904 si_pm4_delete_state(sctx, vs, p->pm4);
905 break;
906 case PIPE_SHADER_TESS_CTRL:
907 si_pm4_delete_state(sctx, hs, p->pm4);
908 break;
909 case PIPE_SHADER_TESS_EVAL:
910 if (p->key.tes.as_es)
911 si_pm4_delete_state(sctx, es, p->pm4);
912 else
913 si_pm4_delete_state(sctx, vs, p->pm4);
914 break;
915 case PIPE_SHADER_GEOMETRY:
916 si_pm4_delete_state(sctx, gs, p->pm4);
917 si_pm4_delete_state(sctx, vs, p->gs_copy_shader->pm4);
918 break;
919 case PIPE_SHADER_FRAGMENT:
920 si_pm4_delete_state(sctx, ps, p->pm4);
921 break;
922 }
923
924 si_shader_destroy(p);
925 free(p);
926 p = c;
927 }
928
929 free(sel->tokens);
930 free(sel);
931 }
932
933 static void si_emit_spi_map(struct si_context *sctx, struct r600_atom *atom)
934 {
935 struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
936 struct si_shader *ps = sctx->ps_shader->current;
937 struct si_shader *vs = si_get_vs_state(sctx);
938 struct tgsi_shader_info *psinfo = &ps->selector->info;
939 struct tgsi_shader_info *vsinfo = &vs->selector->info;
940 unsigned i, j, tmp, num_written = 0;
941
942 if (!ps->nparam)
943 return;
944
945 radeon_set_context_reg_seq(cs, R_028644_SPI_PS_INPUT_CNTL_0, ps->nparam);
946
947 for (i = 0; i < psinfo->num_inputs; i++) {
948 unsigned name = psinfo->input_semantic_name[i];
949 unsigned index = psinfo->input_semantic_index[i];
950 unsigned interpolate = psinfo->input_interpolate[i];
951 unsigned param_offset = ps->ps_input_param_offset[i];
952
953 if (name == TGSI_SEMANTIC_POSITION ||
954 name == TGSI_SEMANTIC_FACE)
955 /* Read from preloaded VGPRs, not parameters */
956 continue;
957
958 bcolor:
959 tmp = 0;
960
961 if (interpolate == TGSI_INTERPOLATE_CONSTANT ||
962 (interpolate == TGSI_INTERPOLATE_COLOR && sctx->flatshade))
963 tmp |= S_028644_FLAT_SHADE(1);
964
965 if (name == TGSI_SEMANTIC_PCOORD ||
966 (name == TGSI_SEMANTIC_TEXCOORD &&
967 sctx->sprite_coord_enable & (1 << index))) {
968 tmp |= S_028644_PT_SPRITE_TEX(1);
969 }
970
971 for (j = 0; j < vsinfo->num_outputs; j++) {
972 if (name == vsinfo->output_semantic_name[j] &&
973 index == vsinfo->output_semantic_index[j]) {
974 tmp |= S_028644_OFFSET(vs->vs_output_param_offset[j]);
975 break;
976 }
977 }
978
979 if (name == TGSI_SEMANTIC_PRIMID)
980 /* PrimID is written after the last output. */
981 tmp |= S_028644_OFFSET(vs->vs_output_param_offset[vsinfo->num_outputs]);
982 else if (j == vsinfo->num_outputs && !G_028644_PT_SPRITE_TEX(tmp)) {
983 /* No corresponding output found, load defaults into input.
984 * Don't set any other bits.
985 * (FLAT_SHADE=1 completely changes behavior) */
986 tmp = S_028644_OFFSET(0x20);
987 }
988
989 assert(param_offset == num_written);
990 radeon_emit(cs, tmp);
991 num_written++;
992
993 if (name == TGSI_SEMANTIC_COLOR &&
994 ps->key.ps.color_two_side) {
995 name = TGSI_SEMANTIC_BCOLOR;
996 param_offset++;
997 goto bcolor;
998 }
999 }
1000 assert(ps->nparam == num_written);
1001 }
1002
1003 static void si_emit_spi_ps_input(struct si_context *sctx, struct r600_atom *atom)
1004 {
1005 struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
1006 struct si_shader *ps = sctx->ps_shader->current;
1007 unsigned input_ena = ps->spi_ps_input_ena;
1008
1009 /* we need to enable at least one of them, otherwise we hang the GPU */
1010 assert(G_0286CC_PERSP_SAMPLE_ENA(input_ena) ||
1011 G_0286CC_PERSP_CENTER_ENA(input_ena) ||
1012 G_0286CC_PERSP_CENTROID_ENA(input_ena) ||
1013 G_0286CC_PERSP_PULL_MODEL_ENA(input_ena) ||
1014 G_0286CC_LINEAR_SAMPLE_ENA(input_ena) ||
1015 G_0286CC_LINEAR_CENTER_ENA(input_ena) ||
1016 G_0286CC_LINEAR_CENTROID_ENA(input_ena) ||
1017 G_0286CC_LINE_STIPPLE_TEX_ENA(input_ena));
1018
1019 if (sctx->force_persample_interp) {
1020 unsigned num_persp = G_0286CC_PERSP_SAMPLE_ENA(input_ena) +
1021 G_0286CC_PERSP_CENTER_ENA(input_ena) +
1022 G_0286CC_PERSP_CENTROID_ENA(input_ena);
1023 unsigned num_linear = G_0286CC_LINEAR_SAMPLE_ENA(input_ena) +
1024 G_0286CC_LINEAR_CENTER_ENA(input_ena) +
1025 G_0286CC_LINEAR_CENTROID_ENA(input_ena);
1026
1027 /* If only one set of (i,j) coordinates is used, we can disable
1028 * CENTER/CENTROID, enable SAMPLE and it will load SAMPLE coordinates
1029 * where CENTER/CENTROID are expected, effectively forcing per-sample
1030 * interpolation.
1031 */
1032 if (num_persp == 1) {
1033 input_ena &= C_0286CC_PERSP_CENTER_ENA;
1034 input_ena &= C_0286CC_PERSP_CENTROID_ENA;
1035 input_ena |= G_0286CC_PERSP_SAMPLE_ENA(1);
1036 }
1037 if (num_linear == 1) {
1038 input_ena &= C_0286CC_LINEAR_CENTER_ENA;
1039 input_ena &= C_0286CC_LINEAR_CENTROID_ENA;
1040 input_ena |= G_0286CC_LINEAR_SAMPLE_ENA(1);
1041 }
1042
1043 /* If at least 2 sets of coordinates are used, we can't use this
1044 * trick and have to select SAMPLE using a conditional assignment
1045 * in the shader with "force_persample_interp" being a shader constant.
1046 */
1047 }
1048
1049 radeon_set_context_reg_seq(cs, R_0286CC_SPI_PS_INPUT_ENA, 2);
1050 radeon_emit(cs, input_ena);
1051 radeon_emit(cs, input_ena);
1052
1053 if (ps->selector->forces_persample_interp_for_persp ||
1054 ps->selector->forces_persample_interp_for_linear)
1055 radeon_set_sh_reg(cs, R_00B030_SPI_SHADER_USER_DATA_PS_0 +
1056 SI_SGPR_PS_STATE_BITS * 4,
1057 sctx->force_persample_interp);
1058 }
1059
1060 /**
1061 * Writing CONFIG or UCONFIG VGT registers requires VGT_FLUSH before that.
1062 */
1063 static void si_init_config_add_vgt_flush(struct si_context *sctx)
1064 {
1065 if (sctx->init_config_has_vgt_flush)
1066 return;
1067
1068 si_pm4_cmd_begin(sctx->init_config, PKT3_EVENT_WRITE);
1069 si_pm4_cmd_add(sctx->init_config, EVENT_TYPE(V_028A90_VGT_FLUSH) | EVENT_INDEX(0));
1070 si_pm4_cmd_end(sctx->init_config, false);
1071 sctx->init_config_has_vgt_flush = true;
1072 }
1073
1074 /* Initialize state related to ESGS / GSVS ring buffers */
1075 static void si_init_gs_rings(struct si_context *sctx)
1076 {
1077 unsigned esgs_ring_size = 128 * 1024;
1078 unsigned gsvs_ring_size = 60 * 1024 * 1024;
1079
1080 assert(!sctx->esgs_ring && !sctx->gsvs_ring);
1081
1082 sctx->esgs_ring = pipe_buffer_create(sctx->b.b.screen, PIPE_BIND_CUSTOM,
1083 PIPE_USAGE_DEFAULT, esgs_ring_size);
1084 if (!sctx->esgs_ring)
1085 return;
1086
1087 sctx->gsvs_ring = pipe_buffer_create(sctx->b.b.screen, PIPE_BIND_CUSTOM,
1088 PIPE_USAGE_DEFAULT, gsvs_ring_size);
1089 if (!sctx->gsvs_ring) {
1090 pipe_resource_reference(&sctx->esgs_ring, NULL);
1091 return;
1092 }
1093
1094 si_init_config_add_vgt_flush(sctx);
1095
1096 /* Append these registers to the init config state. */
1097 if (sctx->b.chip_class >= CIK) {
1098 if (sctx->b.chip_class >= VI) {
1099 /* The maximum sizes are 63.999 MB on VI, because
1100 * the register fields only have 18 bits. */
1101 assert(esgs_ring_size / 256 < (1 << 18));
1102 assert(gsvs_ring_size / 256 < (1 << 18));
1103 }
1104 si_pm4_set_reg(sctx->init_config, R_030900_VGT_ESGS_RING_SIZE,
1105 esgs_ring_size / 256);
1106 si_pm4_set_reg(sctx->init_config, R_030904_VGT_GSVS_RING_SIZE,
1107 gsvs_ring_size / 256);
1108 } else {
1109 si_pm4_set_reg(sctx->init_config, R_0088C8_VGT_ESGS_RING_SIZE,
1110 esgs_ring_size / 256);
1111 si_pm4_set_reg(sctx->init_config, R_0088CC_VGT_GSVS_RING_SIZE,
1112 gsvs_ring_size / 256);
1113 }
1114
1115 /* Flush the context to re-emit the init_config state.
1116 * This is done only once in a lifetime of a context.
1117 */
1118 si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
1119 sctx->b.initial_gfx_cs_size = 0; /* force flush */
1120 si_context_gfx_flush(sctx, RADEON_FLUSH_ASYNC, NULL);
1121
1122 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_VERTEX, SI_RING_ESGS,
1123 sctx->esgs_ring, 0, esgs_ring_size,
1124 true, true, 4, 64, 0);
1125 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_GEOMETRY, SI_RING_ESGS,
1126 sctx->esgs_ring, 0, esgs_ring_size,
1127 false, false, 0, 0, 0);
1128 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_VERTEX, SI_RING_GSVS,
1129 sctx->gsvs_ring, 0, gsvs_ring_size,
1130 false, false, 0, 0, 0);
1131 }
1132
1133 static void si_update_gs_rings(struct si_context *sctx)
1134 {
1135 unsigned gsvs_itemsize = sctx->gs_shader->gsvs_itemsize;
1136 uint64_t offset;
1137
1138 if (gsvs_itemsize == sctx->last_gsvs_itemsize)
1139 return;
1140
1141 sctx->last_gsvs_itemsize = gsvs_itemsize;
1142
1143 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_GEOMETRY, SI_RING_GSVS,
1144 sctx->gsvs_ring, gsvs_itemsize,
1145 64, true, true, 4, 16, 0);
1146
1147 offset = gsvs_itemsize * 64;
1148 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_GEOMETRY, SI_RING_GSVS_1,
1149 sctx->gsvs_ring, gsvs_itemsize,
1150 64, true, true, 4, 16, offset);
1151
1152 offset = (gsvs_itemsize * 2) * 64;
1153 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_GEOMETRY, SI_RING_GSVS_2,
1154 sctx->gsvs_ring, gsvs_itemsize,
1155 64, true, true, 4, 16, offset);
1156
1157 offset = (gsvs_itemsize * 3) * 64;
1158 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_GEOMETRY, SI_RING_GSVS_3,
1159 sctx->gsvs_ring, gsvs_itemsize,
1160 64, true, true, 4, 16, offset);
1161 }
1162
1163 /**
1164 * @returns 1 if \p sel has been updated to use a new scratch buffer
1165 * 0 if not
1166 * < 0 if there was a failure
1167 */
1168 static int si_update_scratch_buffer(struct si_context *sctx,
1169 struct si_shader_selector *sel)
1170 {
1171 struct si_shader *shader;
1172 uint64_t scratch_va = sctx->scratch_buffer->gpu_address;
1173 int r;
1174
1175 if (!sel)
1176 return 0;
1177
1178 shader = sel->current;
1179
1180 /* This shader doesn't need a scratch buffer */
1181 if (shader->scratch_bytes_per_wave == 0)
1182 return 0;
1183
1184 /* This shader is already configured to use the current
1185 * scratch buffer. */
1186 if (shader->scratch_bo == sctx->scratch_buffer)
1187 return 0;
1188
1189 assert(sctx->scratch_buffer);
1190
1191 si_shader_apply_scratch_relocs(sctx, shader, scratch_va);
1192
1193 /* Replace the shader bo with a new bo that has the relocs applied. */
1194 r = si_shader_binary_upload(sctx->screen, shader);
1195 if (r)
1196 return r;
1197
1198 /* Update the shader state to use the new shader bo. */
1199 si_shader_init_pm4_state(shader);
1200
1201 r600_resource_reference(&shader->scratch_bo, sctx->scratch_buffer);
1202
1203 return 1;
1204 }
1205
1206 static unsigned si_get_current_scratch_buffer_size(struct si_context *sctx)
1207 {
1208 return sctx->scratch_buffer ? sctx->scratch_buffer->b.b.width0 : 0;
1209 }
1210
1211 static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_shader_selector *sel)
1212 {
1213 return sel ? sel->current->scratch_bytes_per_wave : 0;
1214 }
1215
1216 static unsigned si_get_max_scratch_bytes_per_wave(struct si_context *sctx)
1217 {
1218 unsigned bytes = 0;
1219
1220 bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->ps_shader));
1221 bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->gs_shader));
1222 bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->vs_shader));
1223 bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->tcs_shader));
1224 bytes = MAX2(bytes, si_get_scratch_buffer_bytes_per_wave(sctx->tes_shader));
1225 return bytes;
1226 }
1227
1228 static bool si_update_spi_tmpring_size(struct si_context *sctx)
1229 {
1230 unsigned current_scratch_buffer_size =
1231 si_get_current_scratch_buffer_size(sctx);
1232 unsigned scratch_bytes_per_wave =
1233 si_get_max_scratch_bytes_per_wave(sctx);
1234 unsigned scratch_needed_size = scratch_bytes_per_wave *
1235 sctx->scratch_waves;
1236 int r;
1237
1238 if (scratch_needed_size > 0) {
1239 if (scratch_needed_size > current_scratch_buffer_size) {
1240 /* Create a bigger scratch buffer */
1241 pipe_resource_reference(
1242 (struct pipe_resource**)&sctx->scratch_buffer,
1243 NULL);
1244
1245 sctx->scratch_buffer =
1246 si_resource_create_custom(&sctx->screen->b.b,
1247 PIPE_USAGE_DEFAULT, scratch_needed_size);
1248 if (!sctx->scratch_buffer)
1249 return false;
1250 sctx->emit_scratch_reloc = true;
1251 }
1252
1253 /* Update the shaders, so they are using the latest scratch. The
1254 * scratch buffer may have been changed since these shaders were
1255 * last used, so we still need to try to update them, even if
1256 * they require scratch buffers smaller than the current size.
1257 */
1258 r = si_update_scratch_buffer(sctx, sctx->ps_shader);
1259 if (r < 0)
1260 return false;
1261 if (r == 1)
1262 si_pm4_bind_state(sctx, ps, sctx->ps_shader->current->pm4);
1263
1264 r = si_update_scratch_buffer(sctx, sctx->gs_shader);
1265 if (r < 0)
1266 return false;
1267 if (r == 1)
1268 si_pm4_bind_state(sctx, gs, sctx->gs_shader->current->pm4);
1269
1270 r = si_update_scratch_buffer(sctx, sctx->tcs_shader);
1271 if (r < 0)
1272 return false;
1273 if (r == 1)
1274 si_pm4_bind_state(sctx, hs, sctx->tcs_shader->current->pm4);
1275
1276 /* VS can be bound as LS, ES, or VS. */
1277 r = si_update_scratch_buffer(sctx, sctx->vs_shader);
1278 if (r < 0)
1279 return false;
1280 if (r == 1) {
1281 if (sctx->tes_shader)
1282 si_pm4_bind_state(sctx, ls, sctx->vs_shader->current->pm4);
1283 else if (sctx->gs_shader)
1284 si_pm4_bind_state(sctx, es, sctx->vs_shader->current->pm4);
1285 else
1286 si_pm4_bind_state(sctx, vs, sctx->vs_shader->current->pm4);
1287 }
1288
1289 /* TES can be bound as ES or VS. */
1290 r = si_update_scratch_buffer(sctx, sctx->tes_shader);
1291 if (r < 0)
1292 return false;
1293 if (r == 1) {
1294 if (sctx->gs_shader)
1295 si_pm4_bind_state(sctx, es, sctx->tes_shader->current->pm4);
1296 else
1297 si_pm4_bind_state(sctx, vs, sctx->tes_shader->current->pm4);
1298 }
1299 }
1300
1301 /* The LLVM shader backend should be reporting aligned scratch_sizes. */
1302 assert((scratch_needed_size & ~0x3FF) == scratch_needed_size &&
1303 "scratch size should already be aligned correctly.");
1304
1305 sctx->spi_tmpring_size = S_0286E8_WAVES(sctx->scratch_waves) |
1306 S_0286E8_WAVESIZE(scratch_bytes_per_wave >> 10);
1307 return true;
1308 }
1309
1310 static void si_init_tess_factor_ring(struct si_context *sctx)
1311 {
1312 assert(!sctx->tf_ring);
1313
1314 sctx->tf_ring = pipe_buffer_create(sctx->b.b.screen, PIPE_BIND_CUSTOM,
1315 PIPE_USAGE_DEFAULT,
1316 32768 * sctx->screen->b.info.max_se);
1317 if (!sctx->tf_ring)
1318 return;
1319
1320 assert(((sctx->tf_ring->width0 / 4) & C_030938_SIZE) == 0);
1321
1322 si_init_config_add_vgt_flush(sctx);
1323
1324 /* Append these registers to the init config state. */
1325 if (sctx->b.chip_class >= CIK) {
1326 si_pm4_set_reg(sctx->init_config, R_030938_VGT_TF_RING_SIZE,
1327 S_030938_SIZE(sctx->tf_ring->width0 / 4));
1328 si_pm4_set_reg(sctx->init_config, R_030940_VGT_TF_MEMORY_BASE,
1329 r600_resource(sctx->tf_ring)->gpu_address >> 8);
1330 } else {
1331 si_pm4_set_reg(sctx->init_config, R_008988_VGT_TF_RING_SIZE,
1332 S_008988_SIZE(sctx->tf_ring->width0 / 4));
1333 si_pm4_set_reg(sctx->init_config, R_0089B8_VGT_TF_MEMORY_BASE,
1334 r600_resource(sctx->tf_ring)->gpu_address >> 8);
1335 }
1336
1337 /* Flush the context to re-emit the init_config state.
1338 * This is done only once in a lifetime of a context.
1339 */
1340 si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
1341 sctx->b.initial_gfx_cs_size = 0; /* force flush */
1342 si_context_gfx_flush(sctx, RADEON_FLUSH_ASYNC, NULL);
1343
1344 si_set_ring_buffer(&sctx->b.b, PIPE_SHADER_TESS_CTRL,
1345 SI_RING_TESS_FACTOR, sctx->tf_ring, 0,
1346 sctx->tf_ring->width0, false, false, 0, 0, 0);
1347 }
1348
1349 /**
1350 * This is used when TCS is NULL in the VS->TCS->TES chain. In this case,
1351 * VS passes its outputs to TES directly, so the fixed-function shader only
1352 * has to write TESSOUTER and TESSINNER.
1353 */
1354 static void si_generate_fixed_func_tcs(struct si_context *sctx)
1355 {
1356 struct ureg_src const0, const1;
1357 struct ureg_dst tessouter, tessinner;
1358 struct ureg_program *ureg = ureg_create(TGSI_PROCESSOR_TESS_CTRL);
1359
1360 if (!ureg)
1361 return; /* if we get here, we're screwed */
1362
1363 assert(!sctx->fixed_func_tcs_shader);
1364
1365 ureg_DECL_constant2D(ureg, 0, 1, SI_DRIVER_STATE_CONST_BUF);
1366 const0 = ureg_src_dimension(ureg_src_register(TGSI_FILE_CONSTANT, 0),
1367 SI_DRIVER_STATE_CONST_BUF);
1368 const1 = ureg_src_dimension(ureg_src_register(TGSI_FILE_CONSTANT, 1),
1369 SI_DRIVER_STATE_CONST_BUF);
1370
1371 tessouter = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSOUTER, 0);
1372 tessinner = ureg_DECL_output(ureg, TGSI_SEMANTIC_TESSINNER, 0);
1373
1374 ureg_MOV(ureg, tessouter, const0);
1375 ureg_MOV(ureg, tessinner, const1);
1376 ureg_END(ureg);
1377
1378 sctx->fixed_func_tcs_shader =
1379 ureg_create_shader_and_destroy(ureg, &sctx->b.b);
1380 }
1381
1382 static void si_update_vgt_shader_config(struct si_context *sctx)
1383 {
1384 /* Calculate the index of the config.
1385 * 0 = VS, 1 = VS+GS, 2 = VS+Tess, 3 = VS+Tess+GS */
1386 unsigned index = 2*!!sctx->tes_shader + !!sctx->gs_shader;
1387 struct si_pm4_state **pm4 = &sctx->vgt_shader_config[index];
1388
1389 if (!*pm4) {
1390 uint32_t stages = 0;
1391
1392 *pm4 = CALLOC_STRUCT(si_pm4_state);
1393
1394 if (sctx->tes_shader) {
1395 stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) |
1396 S_028B54_HS_EN(1);
1397
1398 if (sctx->gs_shader)
1399 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) |
1400 S_028B54_GS_EN(1) |
1401 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
1402 else
1403 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS);
1404 } else if (sctx->gs_shader) {
1405 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) |
1406 S_028B54_GS_EN(1) |
1407 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER);
1408 }
1409
1410 si_pm4_set_reg(*pm4, R_028B54_VGT_SHADER_STAGES_EN, stages);
1411 }
1412 si_pm4_bind_state(sctx, vgt_shader_config, *pm4);
1413 }
1414
1415 static void si_update_so(struct si_context *sctx, struct si_shader_selector *shader)
1416 {
1417 struct pipe_stream_output_info *so = &shader->so;
1418 uint32_t enabled_stream_buffers_mask = 0;
1419 int i;
1420
1421 for (i = 0; i < so->num_outputs; i++)
1422 enabled_stream_buffers_mask |= (1 << so->output[i].output_buffer) << (so->output[i].stream * 4);
1423 sctx->b.streamout.enabled_stream_buffers_mask = enabled_stream_buffers_mask;
1424 sctx->b.streamout.stride_in_dw = shader->so.stride;
1425 }
1426
1427 bool si_update_shaders(struct si_context *sctx)
1428 {
1429 struct pipe_context *ctx = (struct pipe_context*)sctx;
1430 struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
1431 int r;
1432
1433 /* Update stages before GS. */
1434 if (sctx->tes_shader) {
1435 if (!sctx->tf_ring) {
1436 si_init_tess_factor_ring(sctx);
1437 if (!sctx->tf_ring)
1438 return false;
1439 }
1440
1441 /* VS as LS */
1442 r = si_shader_select(ctx, sctx->vs_shader);
1443 if (r)
1444 return false;
1445 si_pm4_bind_state(sctx, ls, sctx->vs_shader->current->pm4);
1446
1447 if (sctx->tcs_shader) {
1448 r = si_shader_select(ctx, sctx->tcs_shader);
1449 if (r)
1450 return false;
1451 si_pm4_bind_state(sctx, hs, sctx->tcs_shader->current->pm4);
1452 } else {
1453 if (!sctx->fixed_func_tcs_shader) {
1454 si_generate_fixed_func_tcs(sctx);
1455 if (!sctx->fixed_func_tcs_shader)
1456 return false;
1457 }
1458
1459 r = si_shader_select(ctx, sctx->fixed_func_tcs_shader);
1460 if (r)
1461 return false;
1462 si_pm4_bind_state(sctx, hs,
1463 sctx->fixed_func_tcs_shader->current->pm4);
1464 }
1465
1466 r = si_shader_select(ctx, sctx->tes_shader);
1467 if (r)
1468 return false;
1469
1470 if (sctx->gs_shader) {
1471 /* TES as ES */
1472 si_pm4_bind_state(sctx, es, sctx->tes_shader->current->pm4);
1473 } else {
1474 /* TES as VS */
1475 si_pm4_bind_state(sctx, vs, sctx->tes_shader->current->pm4);
1476 si_update_so(sctx, sctx->tes_shader);
1477 }
1478 } else if (sctx->gs_shader) {
1479 /* VS as ES */
1480 r = si_shader_select(ctx, sctx->vs_shader);
1481 if (r)
1482 return false;
1483 si_pm4_bind_state(sctx, es, sctx->vs_shader->current->pm4);
1484 } else {
1485 /* VS as VS */
1486 r = si_shader_select(ctx, sctx->vs_shader);
1487 if (r)
1488 return false;
1489 si_pm4_bind_state(sctx, vs, sctx->vs_shader->current->pm4);
1490 si_update_so(sctx, sctx->vs_shader);
1491 }
1492
1493 /* Update GS. */
1494 if (sctx->gs_shader) {
1495 r = si_shader_select(ctx, sctx->gs_shader);
1496 if (r)
1497 return false;
1498 si_pm4_bind_state(sctx, gs, sctx->gs_shader->current->pm4);
1499 si_pm4_bind_state(sctx, vs, sctx->gs_shader->current->gs_copy_shader->pm4);
1500 si_update_so(sctx, sctx->gs_shader);
1501
1502 if (!sctx->gsvs_ring) {
1503 si_init_gs_rings(sctx);
1504 if (!sctx->gsvs_ring)
1505 return false;
1506 }
1507
1508 si_update_gs_rings(sctx);
1509 } else {
1510 si_pm4_bind_state(sctx, gs, NULL);
1511 si_pm4_bind_state(sctx, es, NULL);
1512 }
1513
1514 si_update_vgt_shader_config(sctx);
1515
1516 r = si_shader_select(ctx, sctx->ps_shader);
1517 if (r)
1518 return false;
1519 si_pm4_bind_state(sctx, ps, sctx->ps_shader->current->pm4);
1520
1521 if (si_pm4_state_changed(sctx, ps) || si_pm4_state_changed(sctx, vs) ||
1522 sctx->sprite_coord_enable != rs->sprite_coord_enable ||
1523 sctx->flatshade != rs->flatshade) {
1524 sctx->sprite_coord_enable = rs->sprite_coord_enable;
1525 sctx->flatshade = rs->flatshade;
1526 si_mark_atom_dirty(sctx, &sctx->spi_map);
1527 }
1528
1529 if (si_pm4_state_changed(sctx, ps) ||
1530 sctx->force_persample_interp != rs->force_persample_interp) {
1531 sctx->force_persample_interp = rs->force_persample_interp;
1532 si_mark_atom_dirty(sctx, &sctx->spi_ps_input);
1533 }
1534
1535 if (si_pm4_state_changed(sctx, ls) ||
1536 si_pm4_state_changed(sctx, hs) ||
1537 si_pm4_state_changed(sctx, es) ||
1538 si_pm4_state_changed(sctx, gs) ||
1539 si_pm4_state_changed(sctx, vs) ||
1540 si_pm4_state_changed(sctx, ps)) {
1541 if (!si_update_spi_tmpring_size(sctx))
1542 return false;
1543 }
1544
1545 if (sctx->ps_db_shader_control != sctx->ps_shader->current->db_shader_control) {
1546 sctx->ps_db_shader_control = sctx->ps_shader->current->db_shader_control;
1547 si_mark_atom_dirty(sctx, &sctx->db_render_state);
1548 }
1549
1550 if (sctx->smoothing_enabled != sctx->ps_shader->current->key.ps.poly_line_smoothing) {
1551 sctx->smoothing_enabled = sctx->ps_shader->current->key.ps.poly_line_smoothing;
1552 si_mark_atom_dirty(sctx, &sctx->msaa_config);
1553
1554 if (sctx->b.chip_class == SI)
1555 si_mark_atom_dirty(sctx, &sctx->db_render_state);
1556 }
1557 return true;
1558 }
1559
1560 void si_init_shader_functions(struct si_context *sctx)
1561 {
1562 si_init_atom(sctx, &sctx->spi_map, &sctx->atoms.s.spi_map, si_emit_spi_map);
1563 si_init_atom(sctx, &sctx->spi_ps_input, &sctx->atoms.s.spi_ps_input, si_emit_spi_ps_input);
1564
1565 sctx->b.b.create_vs_state = si_create_shader_selector;
1566 sctx->b.b.create_tcs_state = si_create_shader_selector;
1567 sctx->b.b.create_tes_state = si_create_shader_selector;
1568 sctx->b.b.create_gs_state = si_create_shader_selector;
1569 sctx->b.b.create_fs_state = si_create_shader_selector;
1570
1571 sctx->b.b.bind_vs_state = si_bind_vs_shader;
1572 sctx->b.b.bind_tcs_state = si_bind_tcs_shader;
1573 sctx->b.b.bind_tes_state = si_bind_tes_shader;
1574 sctx->b.b.bind_gs_state = si_bind_gs_shader;
1575 sctx->b.b.bind_fs_state = si_bind_ps_shader;
1576
1577 sctx->b.b.delete_vs_state = si_delete_shader_selector;
1578 sctx->b.b.delete_tcs_state = si_delete_shader_selector;
1579 sctx->b.b.delete_tes_state = si_delete_shader_selector;
1580 sctx->b.b.delete_gs_state = si_delete_shader_selector;
1581 sctx->b.b.delete_fs_state = si_delete_shader_selector;
1582 }