2 * Copyright 2012 Advanced Micro Devices, Inc.
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:
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
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.
24 * Christian König <christian.koenig@amd.com>
25 * Marek Olšák <maraeo@gmail.com>
29 #include "si_shader.h"
32 #include "tgsi/tgsi_parse.h"
33 #include "tgsi/tgsi_ureg.h"
34 #include "util/u_memory.h"
35 #include "util/u_simple_shaders.h"
37 static void si_set_tesseval_regs(struct si_shader
*shader
,
38 struct si_pm4_state
*pm4
)
40 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
41 unsigned tes_prim_mode
= info
->properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
42 unsigned tes_spacing
= info
->properties
[TGSI_PROPERTY_TES_SPACING
];
43 bool tes_vertex_order_cw
= info
->properties
[TGSI_PROPERTY_TES_VERTEX_ORDER_CW
];
44 bool tes_point_mode
= info
->properties
[TGSI_PROPERTY_TES_POINT_MODE
];
45 unsigned type
, partitioning
, topology
;
47 switch (tes_prim_mode
) {
49 type
= V_028B6C_TESS_ISOLINE
;
51 case PIPE_PRIM_TRIANGLES
:
52 type
= V_028B6C_TESS_TRIANGLE
;
55 type
= V_028B6C_TESS_QUAD
;
62 switch (tes_spacing
) {
63 case PIPE_TESS_SPACING_FRACTIONAL_ODD
:
64 partitioning
= V_028B6C_PART_FRAC_ODD
;
66 case PIPE_TESS_SPACING_FRACTIONAL_EVEN
:
67 partitioning
= V_028B6C_PART_FRAC_EVEN
;
69 case PIPE_TESS_SPACING_EQUAL
:
70 partitioning
= V_028B6C_PART_INTEGER
;
78 topology
= V_028B6C_OUTPUT_POINT
;
79 else if (tes_prim_mode
== PIPE_PRIM_LINES
)
80 topology
= V_028B6C_OUTPUT_LINE
;
81 else if (tes_vertex_order_cw
)
82 /* for some reason, this must be the other way around */
83 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
85 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
87 si_pm4_set_reg(pm4
, R_028B6C_VGT_TF_PARAM
,
89 S_028B6C_PARTITIONING(partitioning
) |
90 S_028B6C_TOPOLOGY(topology
));
93 static void si_shader_ls(struct si_shader
*shader
)
95 struct si_pm4_state
*pm4
;
96 unsigned num_sgprs
, num_user_sgprs
;
97 unsigned vgpr_comp_cnt
;
100 pm4
= shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
104 va
= shader
->bo
->gpu_address
;
105 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_DATA
);
107 /* We need at least 2 components for LS.
108 * VGPR0-3: (VertexID, RelAutoindex, ???, InstanceID). */
109 vgpr_comp_cnt
= shader
->uses_instanceid
? 3 : 1;
111 num_user_sgprs
= SI_LS_NUM_USER_SGPR
;
112 num_sgprs
= shader
->num_sgprs
;
113 if (num_user_sgprs
> num_sgprs
) {
114 /* Last 2 reserved SGPRs are used for VCC */
115 num_sgprs
= num_user_sgprs
+ 2;
117 assert(num_sgprs
<= 104);
119 si_pm4_set_reg(pm4
, R_00B520_SPI_SHADER_PGM_LO_LS
, va
>> 8);
120 si_pm4_set_reg(pm4
, R_00B524_SPI_SHADER_PGM_HI_LS
, va
>> 40);
122 shader
->ls_rsrc1
= S_00B528_VGPRS((shader
->num_vgprs
- 1) / 4) |
123 S_00B528_SGPRS((num_sgprs
- 1) / 8) |
124 S_00B528_VGPR_COMP_CNT(vgpr_comp_cnt
);
125 shader
->ls_rsrc2
= S_00B52C_USER_SGPR(num_user_sgprs
) |
126 S_00B52C_SCRATCH_EN(shader
->scratch_bytes_per_wave
> 0);
129 static void si_shader_hs(struct si_shader
*shader
)
131 struct si_pm4_state
*pm4
;
132 unsigned num_sgprs
, num_user_sgprs
;
135 pm4
= shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
139 va
= shader
->bo
->gpu_address
;
140 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_DATA
);
142 num_user_sgprs
= SI_TCS_NUM_USER_SGPR
;
143 num_sgprs
= shader
->num_sgprs
;
144 /* One SGPR after user SGPRs is pre-loaded with tessellation factor
146 if ((num_user_sgprs
+ 1) > num_sgprs
) {
147 /* Last 2 reserved SGPRs are used for VCC */
148 num_sgprs
= num_user_sgprs
+ 1 + 2;
150 assert(num_sgprs
<= 104);
152 si_pm4_set_reg(pm4
, R_00B420_SPI_SHADER_PGM_LO_HS
, va
>> 8);
153 si_pm4_set_reg(pm4
, R_00B424_SPI_SHADER_PGM_HI_HS
, va
>> 40);
154 si_pm4_set_reg(pm4
, R_00B428_SPI_SHADER_PGM_RSRC1_HS
,
155 S_00B428_VGPRS((shader
->num_vgprs
- 1) / 4) |
156 S_00B428_SGPRS((num_sgprs
- 1) / 8));
157 si_pm4_set_reg(pm4
, R_00B42C_SPI_SHADER_PGM_RSRC2_HS
,
158 S_00B42C_USER_SGPR(num_user_sgprs
) |
159 S_00B42C_SCRATCH_EN(shader
->scratch_bytes_per_wave
> 0));
162 static void si_shader_es(struct si_shader
*shader
)
164 struct si_pm4_state
*pm4
;
165 unsigned num_sgprs
, num_user_sgprs
;
166 unsigned vgpr_comp_cnt
;
169 pm4
= shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
174 va
= shader
->bo
->gpu_address
;
175 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_DATA
);
177 if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
178 vgpr_comp_cnt
= shader
->uses_instanceid
? 3 : 0;
179 num_user_sgprs
= SI_VS_NUM_USER_SGPR
;
180 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
181 vgpr_comp_cnt
= 3; /* all components are needed for TES */
182 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
186 num_sgprs
= shader
->num_sgprs
;
187 /* One SGPR after user SGPRs is pre-loaded with es2gs_offset */
188 if ((num_user_sgprs
+ 1) > num_sgprs
) {
189 /* Last 2 reserved SGPRs are used for VCC */
190 num_sgprs
= num_user_sgprs
+ 1 + 2;
192 assert(num_sgprs
<= 104);
194 si_pm4_set_reg(pm4
, R_00B320_SPI_SHADER_PGM_LO_ES
, va
>> 8);
195 si_pm4_set_reg(pm4
, R_00B324_SPI_SHADER_PGM_HI_ES
, va
>> 40);
196 si_pm4_set_reg(pm4
, R_00B328_SPI_SHADER_PGM_RSRC1_ES
,
197 S_00B328_VGPRS((shader
->num_vgprs
- 1) / 4) |
198 S_00B328_SGPRS((num_sgprs
- 1) / 8) |
199 S_00B328_VGPR_COMP_CNT(vgpr_comp_cnt
) |
200 S_00B328_DX10_CLAMP(shader
->dx10_clamp_mode
));
201 si_pm4_set_reg(pm4
, R_00B32C_SPI_SHADER_PGM_RSRC2_ES
,
202 S_00B32C_USER_SGPR(num_user_sgprs
) |
203 S_00B32C_SCRATCH_EN(shader
->scratch_bytes_per_wave
> 0));
205 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
206 si_set_tesseval_regs(shader
, pm4
);
209 static unsigned si_gs_get_max_stream(struct si_shader
*shader
)
211 struct pipe_stream_output_info
*so
= &shader
->selector
->so
;
212 unsigned max_stream
= 0, i
;
214 if (so
->num_outputs
== 0)
217 for (i
= 0; i
< so
->num_outputs
; i
++) {
218 if (so
->output
[i
].stream
> max_stream
)
219 max_stream
= so
->output
[i
].stream
;
224 static void si_shader_gs(struct si_shader
*shader
)
226 unsigned gs_vert_itemsize
= shader
->selector
->info
.num_outputs
* 16;
227 unsigned gs_max_vert_out
= shader
->selector
->gs_max_out_vertices
;
228 unsigned gsvs_itemsize
= (gs_vert_itemsize
* gs_max_vert_out
) >> 2;
229 unsigned gs_num_invocations
= shader
->selector
->gs_num_invocations
;
231 struct si_pm4_state
*pm4
;
232 unsigned num_sgprs
, num_user_sgprs
;
234 unsigned max_stream
= si_gs_get_max_stream(shader
);
236 /* The GSVS_RING_ITEMSIZE register takes 15 bits */
237 assert(gsvs_itemsize
< (1 << 15));
239 pm4
= shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
244 if (gs_max_vert_out
<= 128) {
245 cut_mode
= V_028A40_GS_CUT_128
;
246 } else if (gs_max_vert_out
<= 256) {
247 cut_mode
= V_028A40_GS_CUT_256
;
248 } else if (gs_max_vert_out
<= 512) {
249 cut_mode
= V_028A40_GS_CUT_512
;
251 assert(gs_max_vert_out
<= 1024);
252 cut_mode
= V_028A40_GS_CUT_1024
;
255 si_pm4_set_reg(pm4
, R_028A40_VGT_GS_MODE
,
256 S_028A40_MODE(V_028A40_GS_SCENARIO_G
) |
257 S_028A40_CUT_MODE(cut_mode
)|
258 S_028A40_ES_WRITE_OPTIMIZE(1) |
259 S_028A40_GS_WRITE_OPTIMIZE(1));
261 si_pm4_set_reg(pm4
, R_028A60_VGT_GSVS_RING_OFFSET_1
, gsvs_itemsize
);
262 si_pm4_set_reg(pm4
, R_028A64_VGT_GSVS_RING_OFFSET_2
, gsvs_itemsize
* ((max_stream
>= 2) ? 2 : 1));
263 si_pm4_set_reg(pm4
, R_028A68_VGT_GSVS_RING_OFFSET_3
, gsvs_itemsize
* ((max_stream
>= 3) ? 3 : 1));
265 si_pm4_set_reg(pm4
, R_028AAC_VGT_ESGS_RING_ITEMSIZE
,
266 util_bitcount64(shader
->selector
->inputs_read
) * (16 >> 2));
267 si_pm4_set_reg(pm4
, R_028AB0_VGT_GSVS_RING_ITEMSIZE
, gsvs_itemsize
* (max_stream
+ 1));
269 si_pm4_set_reg(pm4
, R_028B38_VGT_GS_MAX_VERT_OUT
, gs_max_vert_out
);
271 si_pm4_set_reg(pm4
, R_028B5C_VGT_GS_VERT_ITEMSIZE
, gs_vert_itemsize
>> 2);
272 si_pm4_set_reg(pm4
, R_028B60_VGT_GS_VERT_ITEMSIZE_1
, (max_stream
>= 1) ? gs_vert_itemsize
>> 2 : 0);
273 si_pm4_set_reg(pm4
, R_028B64_VGT_GS_VERT_ITEMSIZE_2
, (max_stream
>= 2) ? gs_vert_itemsize
>> 2 : 0);
274 si_pm4_set_reg(pm4
, R_028B68_VGT_GS_VERT_ITEMSIZE_3
, (max_stream
>= 3) ? gs_vert_itemsize
>> 2 : 0);
276 si_pm4_set_reg(pm4
, R_028B90_VGT_GS_INSTANCE_CNT
,
277 S_028B90_CNT(MIN2(gs_num_invocations
, 127)) |
278 S_028B90_ENABLE(gs_num_invocations
> 0));
280 va
= shader
->bo
->gpu_address
;
281 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_DATA
);
282 si_pm4_set_reg(pm4
, R_00B220_SPI_SHADER_PGM_LO_GS
, va
>> 8);
283 si_pm4_set_reg(pm4
, R_00B224_SPI_SHADER_PGM_HI_GS
, va
>> 40);
285 num_user_sgprs
= SI_GS_NUM_USER_SGPR
;
286 num_sgprs
= shader
->num_sgprs
;
287 /* Two SGPRs after user SGPRs are pre-loaded with gs2vs_offset, gs_wave_id */
288 if ((num_user_sgprs
+ 2) > num_sgprs
) {
289 /* Last 2 reserved SGPRs are used for VCC */
290 num_sgprs
= num_user_sgprs
+ 2 + 2;
292 assert(num_sgprs
<= 104);
294 si_pm4_set_reg(pm4
, R_00B228_SPI_SHADER_PGM_RSRC1_GS
,
295 S_00B228_VGPRS((shader
->num_vgprs
- 1) / 4) |
296 S_00B228_SGPRS((num_sgprs
- 1) / 8) |
297 S_00B228_DX10_CLAMP(shader
->dx10_clamp_mode
));
298 si_pm4_set_reg(pm4
, R_00B22C_SPI_SHADER_PGM_RSRC2_GS
,
299 S_00B22C_USER_SGPR(num_user_sgprs
) |
300 S_00B22C_SCRATCH_EN(shader
->scratch_bytes_per_wave
> 0));
303 static void si_shader_vs(struct si_shader
*shader
)
305 struct si_pm4_state
*pm4
;
306 unsigned num_sgprs
, num_user_sgprs
;
307 unsigned nparams
, vgpr_comp_cnt
;
309 unsigned window_space
=
310 shader
->selector
->info
.properties
[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION
];
312 pm4
= shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
317 /* If this is the GS copy shader, the GS state writes this register.
318 * Otherwise, the VS state writes it.
320 if (!shader
->is_gs_copy_shader
)
321 si_pm4_set_reg(pm4
, R_028A40_VGT_GS_MODE
, 0);
323 va
= shader
->bo
->gpu_address
;
324 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_DATA
);
326 if (shader
->is_gs_copy_shader
) {
327 vgpr_comp_cnt
= 0; /* only VertexID is needed for GS-COPY. */
328 num_user_sgprs
= SI_GSCOPY_NUM_USER_SGPR
;
329 } else if (shader
->selector
->type
== PIPE_SHADER_VERTEX
) {
330 vgpr_comp_cnt
= shader
->uses_instanceid
? 3 : 0;
331 num_user_sgprs
= SI_VS_NUM_USER_SGPR
;
332 } else if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
) {
333 vgpr_comp_cnt
= 3; /* all components are needed for TES */
334 num_user_sgprs
= SI_TES_NUM_USER_SGPR
;
338 num_sgprs
= shader
->num_sgprs
;
339 if (num_user_sgprs
> num_sgprs
) {
340 /* Last 2 reserved SGPRs are used for VCC */
341 num_sgprs
= num_user_sgprs
+ 2;
343 assert(num_sgprs
<= 104);
345 /* VS is required to export at least one param. */
346 nparams
= MAX2(shader
->nr_param_exports
, 1);
347 si_pm4_set_reg(pm4
, R_0286C4_SPI_VS_OUT_CONFIG
,
348 S_0286C4_VS_EXPORT_COUNT(nparams
- 1));
350 si_pm4_set_reg(pm4
, R_02870C_SPI_SHADER_POS_FORMAT
,
351 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP
) |
352 S_02870C_POS1_EXPORT_FORMAT(shader
->nr_pos_exports
> 1 ?
353 V_02870C_SPI_SHADER_4COMP
:
354 V_02870C_SPI_SHADER_NONE
) |
355 S_02870C_POS2_EXPORT_FORMAT(shader
->nr_pos_exports
> 2 ?
356 V_02870C_SPI_SHADER_4COMP
:
357 V_02870C_SPI_SHADER_NONE
) |
358 S_02870C_POS3_EXPORT_FORMAT(shader
->nr_pos_exports
> 3 ?
359 V_02870C_SPI_SHADER_4COMP
:
360 V_02870C_SPI_SHADER_NONE
));
362 si_pm4_set_reg(pm4
, R_00B120_SPI_SHADER_PGM_LO_VS
, va
>> 8);
363 si_pm4_set_reg(pm4
, R_00B124_SPI_SHADER_PGM_HI_VS
, va
>> 40);
364 si_pm4_set_reg(pm4
, R_00B128_SPI_SHADER_PGM_RSRC1_VS
,
365 S_00B128_VGPRS((shader
->num_vgprs
- 1) / 4) |
366 S_00B128_SGPRS((num_sgprs
- 1) / 8) |
367 S_00B128_VGPR_COMP_CNT(vgpr_comp_cnt
) |
368 S_00B128_DX10_CLAMP(shader
->dx10_clamp_mode
));
369 si_pm4_set_reg(pm4
, R_00B12C_SPI_SHADER_PGM_RSRC2_VS
,
370 S_00B12C_USER_SGPR(num_user_sgprs
) |
371 S_00B12C_SO_BASE0_EN(!!shader
->selector
->so
.stride
[0]) |
372 S_00B12C_SO_BASE1_EN(!!shader
->selector
->so
.stride
[1]) |
373 S_00B12C_SO_BASE2_EN(!!shader
->selector
->so
.stride
[2]) |
374 S_00B12C_SO_BASE3_EN(!!shader
->selector
->so
.stride
[3]) |
375 S_00B12C_SO_EN(!!shader
->selector
->so
.num_outputs
) |
376 S_00B12C_SCRATCH_EN(shader
->scratch_bytes_per_wave
> 0));
378 si_pm4_set_reg(pm4
, R_028818_PA_CL_VTE_CNTL
,
379 S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1));
381 si_pm4_set_reg(pm4
, R_028818_PA_CL_VTE_CNTL
,
382 S_028818_VTX_W0_FMT(1) |
383 S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
384 S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
385 S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1));
387 if (shader
->selector
->type
== PIPE_SHADER_TESS_EVAL
)
388 si_set_tesseval_regs(shader
, pm4
);
391 static void si_shader_ps(struct si_shader
*shader
)
393 struct tgsi_shader_info
*info
= &shader
->selector
->info
;
394 struct si_pm4_state
*pm4
;
395 unsigned i
, spi_ps_in_control
;
396 unsigned num_sgprs
, num_user_sgprs
;
397 unsigned spi_baryc_cntl
= 0, spi_ps_input_ena
;
400 pm4
= shader
->pm4
= CALLOC_STRUCT(si_pm4_state
);
405 for (i
= 0; i
< info
->num_inputs
; i
++) {
406 switch (info
->input_semantic_name
[i
]) {
407 case TGSI_SEMANTIC_POSITION
:
408 /* SPI_BARYC_CNTL.POS_FLOAT_LOCATION
410 * 0 -> Position = pixel center (default)
411 * 1 -> Position = pixel centroid
412 * 2 -> Position = at sample position
414 switch (info
->input_interpolate_loc
[i
]) {
415 case TGSI_INTERPOLATE_LOC_CENTROID
:
416 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(1);
418 case TGSI_INTERPOLATE_LOC_SAMPLE
:
419 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_LOCATION(2);
423 if (info
->properties
[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER
] ==
424 TGSI_FS_COORD_PIXEL_CENTER_INTEGER
)
425 spi_baryc_cntl
|= S_0286E0_POS_FLOAT_ULC(1);
430 spi_ps_in_control
= S_0286D8_NUM_INTERP(shader
->nparam
) |
431 S_0286D8_BC_OPTIMIZE_DISABLE(1);
433 si_pm4_set_reg(pm4
, R_0286E0_SPI_BARYC_CNTL
, spi_baryc_cntl
);
434 spi_ps_input_ena
= shader
->spi_ps_input_ena
;
435 /* we need to enable at least one of them, otherwise we hang the GPU */
436 assert(G_0286CC_PERSP_SAMPLE_ENA(spi_ps_input_ena
) ||
437 G_0286CC_PERSP_CENTER_ENA(spi_ps_input_ena
) ||
438 G_0286CC_PERSP_CENTROID_ENA(spi_ps_input_ena
) ||
439 G_0286CC_PERSP_PULL_MODEL_ENA(spi_ps_input_ena
) ||
440 G_0286CC_LINEAR_SAMPLE_ENA(spi_ps_input_ena
) ||
441 G_0286CC_LINEAR_CENTER_ENA(spi_ps_input_ena
) ||
442 G_0286CC_LINEAR_CENTROID_ENA(spi_ps_input_ena
) ||
443 G_0286CC_LINE_STIPPLE_TEX_ENA(spi_ps_input_ena
));
445 si_pm4_set_reg(pm4
, R_0286CC_SPI_PS_INPUT_ENA
, spi_ps_input_ena
);
446 si_pm4_set_reg(pm4
, R_0286D0_SPI_PS_INPUT_ADDR
, spi_ps_input_ena
);
447 si_pm4_set_reg(pm4
, R_0286D8_SPI_PS_IN_CONTROL
, spi_ps_in_control
);
449 si_pm4_set_reg(pm4
, R_028710_SPI_SHADER_Z_FORMAT
, shader
->spi_shader_z_format
);
450 si_pm4_set_reg(pm4
, R_028714_SPI_SHADER_COL_FORMAT
,
451 shader
->spi_shader_col_format
);
452 si_pm4_set_reg(pm4
, R_02823C_CB_SHADER_MASK
, shader
->cb_shader_mask
);
454 va
= shader
->bo
->gpu_address
;
455 si_pm4_add_bo(pm4
, shader
->bo
, RADEON_USAGE_READ
, RADEON_PRIO_SHADER_DATA
);
456 si_pm4_set_reg(pm4
, R_00B020_SPI_SHADER_PGM_LO_PS
, va
>> 8);
457 si_pm4_set_reg(pm4
, R_00B024_SPI_SHADER_PGM_HI_PS
, va
>> 40);
459 num_user_sgprs
= SI_PS_NUM_USER_SGPR
;
460 num_sgprs
= shader
->num_sgprs
;
461 /* One SGPR after user SGPRs is pre-loaded with {prim_mask, lds_offset} */
462 if ((num_user_sgprs
+ 1) > num_sgprs
) {
463 /* Last 2 reserved SGPRs are used for VCC */
464 num_sgprs
= num_user_sgprs
+ 1 + 2;
466 assert(num_sgprs
<= 104);
468 si_pm4_set_reg(pm4
, R_00B028_SPI_SHADER_PGM_RSRC1_PS
,
469 S_00B028_VGPRS((shader
->num_vgprs
- 1) / 4) |
470 S_00B028_SGPRS((num_sgprs
- 1) / 8) |
471 S_00B028_DX10_CLAMP(shader
->dx10_clamp_mode
));
472 si_pm4_set_reg(pm4
, R_00B02C_SPI_SHADER_PGM_RSRC2_PS
,
473 S_00B02C_EXTRA_LDS_SIZE(shader
->lds_size
) |
474 S_00B02C_USER_SGPR(num_user_sgprs
) |
475 S_00B32C_SCRATCH_EN(shader
->scratch_bytes_per_wave
> 0));
478 static void si_shader_init_pm4_state(struct si_shader
*shader
)
482 si_pm4_free_state_simple(shader
->pm4
);
484 switch (shader
->selector
->type
) {
485 case PIPE_SHADER_VERTEX
:
486 if (shader
->key
.vs
.as_ls
)
487 si_shader_ls(shader
);
488 else if (shader
->key
.vs
.as_es
)
489 si_shader_es(shader
);
491 si_shader_vs(shader
);
493 case PIPE_SHADER_TESS_CTRL
:
494 si_shader_hs(shader
);
496 case PIPE_SHADER_TESS_EVAL
:
497 if (shader
->key
.tes
.as_es
)
498 si_shader_es(shader
);
500 si_shader_vs(shader
);
502 case PIPE_SHADER_GEOMETRY
:
503 si_shader_gs(shader
);
504 si_shader_vs(shader
->gs_copy_shader
);
506 case PIPE_SHADER_FRAGMENT
:
507 si_shader_ps(shader
);
514 /* Compute the key for the hw shader variant */
515 static inline void si_shader_selector_key(struct pipe_context
*ctx
,
516 struct si_shader_selector
*sel
,
517 union si_shader_key
*key
)
519 struct si_context
*sctx
= (struct si_context
*)ctx
;
522 memset(key
, 0, sizeof(*key
));
525 case PIPE_SHADER_VERTEX
:
526 if (sctx
->vertex_elements
)
527 for (i
= 0; i
< sctx
->vertex_elements
->count
; ++i
)
528 key
->vs
.instance_divisors
[i
] =
529 sctx
->vertex_elements
->elements
[i
].instance_divisor
;
531 if (sctx
->tes_shader
)
533 else if (sctx
->gs_shader
) {
535 key
->vs
.es_enabled_outputs
= sctx
->gs_shader
->inputs_read
;
538 case PIPE_SHADER_TESS_CTRL
:
540 sctx
->tes_shader
->info
.properties
[TGSI_PROPERTY_TES_PRIM_MODE
];
542 case PIPE_SHADER_TESS_EVAL
:
543 if (sctx
->gs_shader
) {
545 key
->tes
.es_enabled_outputs
= sctx
->gs_shader
->inputs_read
;
548 case PIPE_SHADER_GEOMETRY
:
550 case PIPE_SHADER_FRAGMENT
: {
551 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
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
;
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
;
563 key
->ps
.color_two_side
= rs
->two_side
;
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
;
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;
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
;
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
)
593 struct si_context
*sctx
= (struct si_context
*)ctx
;
594 union si_shader_key key
;
595 struct si_shader
* shader
= NULL
;
598 si_shader_selector_key(ctx
, sel
, &key
);
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
604 if (likely(sel
->current
&& memcmp(&sel
->current
->key
, &key
, sizeof(key
)) == 0)) {
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
;
612 while (c
&& memcmp(&c
->key
, &key
, sizeof(key
)) != 0) {
618 p
->next_variant
= c
->next_variant
;
624 shader
->next_variant
= sel
->current
;
625 sel
->current
= shader
;
627 shader
= CALLOC(1, sizeof(struct si_shader
));
628 shader
->selector
= sel
;
631 shader
->next_variant
= sel
->current
;
632 sel
->current
= shader
;
633 r
= si_shader_create((struct si_screen
*)ctx
->screen
, sctx
->tm
,
636 R600_ERR("Failed to build shader variant (type=%u) %d\n",
642 si_shader_init_pm4_state(shader
);
644 p_atomic_inc(&sctx
->screen
->b
.num_compilations
);
650 static void *si_create_shader_state(struct pipe_context
*ctx
,
651 const struct pipe_shader_state
*state
,
652 unsigned pipe_shader_type
)
654 struct si_screen
*sscreen
= (struct si_screen
*)ctx
->screen
;
655 struct si_shader_selector
*sel
= CALLOC_STRUCT(si_shader_selector
);
658 sel
->type
= pipe_shader_type
;
659 sel
->tokens
= tgsi_dup_tokens(state
->tokens
);
660 sel
->so
= state
->stream_output
;
661 tgsi_scan_shader(state
->tokens
, &sel
->info
);
662 p_atomic_inc(&sscreen
->b
.num_shaders_created
);
664 switch (pipe_shader_type
) {
665 case PIPE_SHADER_GEOMETRY
:
666 sel
->gs_output_prim
=
667 sel
->info
.properties
[TGSI_PROPERTY_GS_OUTPUT_PRIM
];
668 sel
->gs_max_out_vertices
=
669 sel
->info
.properties
[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES
];
670 sel
->gs_num_invocations
=
671 sel
->info
.properties
[TGSI_PROPERTY_GS_INVOCATIONS
];
673 for (i
= 0; i
< sel
->info
.num_inputs
; i
++) {
674 unsigned name
= sel
->info
.input_semantic_name
[i
];
675 unsigned index
= sel
->info
.input_semantic_index
[i
];
678 case TGSI_SEMANTIC_PRIMID
:
682 1llu << si_shader_io_get_unique_index(name
, index
);
687 case PIPE_SHADER_VERTEX
:
688 case PIPE_SHADER_TESS_CTRL
:
689 for (i
= 0; i
< sel
->info
.num_outputs
; i
++) {
690 unsigned name
= sel
->info
.output_semantic_name
[i
];
691 unsigned index
= sel
->info
.output_semantic_index
[i
];
694 case TGSI_SEMANTIC_TESSINNER
:
695 case TGSI_SEMANTIC_TESSOUTER
:
696 case TGSI_SEMANTIC_PATCH
:
697 sel
->patch_outputs_written
|=
698 1llu << si_shader_io_get_unique_index(name
, index
);
701 sel
->outputs_written
|=
702 1llu << si_shader_io_get_unique_index(name
, index
);
708 if (sscreen
->b
.debug_flags
& DBG_PRECOMPILE
)
709 si_shader_select(ctx
, sel
);
714 static void *si_create_fs_state(struct pipe_context
*ctx
,
715 const struct pipe_shader_state
*state
)
717 return si_create_shader_state(ctx
, state
, PIPE_SHADER_FRAGMENT
);
720 static void *si_create_gs_state(struct pipe_context
*ctx
,
721 const struct pipe_shader_state
*state
)
723 return si_create_shader_state(ctx
, state
, PIPE_SHADER_GEOMETRY
);
726 static void *si_create_vs_state(struct pipe_context
*ctx
,
727 const struct pipe_shader_state
*state
)
729 return si_create_shader_state(ctx
, state
, PIPE_SHADER_VERTEX
);
732 static void *si_create_tcs_state(struct pipe_context
*ctx
,
733 const struct pipe_shader_state
*state
)
735 return si_create_shader_state(ctx
, state
, PIPE_SHADER_TESS_CTRL
);
738 static void *si_create_tes_state(struct pipe_context
*ctx
,
739 const struct pipe_shader_state
*state
)
741 return si_create_shader_state(ctx
, state
, PIPE_SHADER_TESS_EVAL
);
744 static void si_bind_vs_shader(struct pipe_context
*ctx
, void *state
)
746 struct si_context
*sctx
= (struct si_context
*)ctx
;
747 struct si_shader_selector
*sel
= state
;
749 if (sctx
->vs_shader
== sel
|| !sel
)
752 sctx
->vs_shader
= sel
;
753 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
756 static void si_bind_gs_shader(struct pipe_context
*ctx
, void *state
)
758 struct si_context
*sctx
= (struct si_context
*)ctx
;
759 struct si_shader_selector
*sel
= state
;
760 bool enable_changed
= !!sctx
->gs_shader
!= !!sel
;
762 if (sctx
->gs_shader
== sel
)
765 sctx
->gs_shader
= sel
;
766 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
767 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
770 si_shader_change_notify(sctx
);
773 static void si_bind_tcs_shader(struct pipe_context
*ctx
, void *state
)
775 struct si_context
*sctx
= (struct si_context
*)ctx
;
776 struct si_shader_selector
*sel
= state
;
777 bool enable_changed
= !!sctx
->tcs_shader
!= !!sel
;
779 if (sctx
->tcs_shader
== sel
)
782 sctx
->tcs_shader
= sel
;
785 sctx
->last_tcs
= NULL
; /* invalidate derived tess state */
788 static void si_bind_tes_shader(struct pipe_context
*ctx
, void *state
)
790 struct si_context
*sctx
= (struct si_context
*)ctx
;
791 struct si_shader_selector
*sel
= state
;
792 bool enable_changed
= !!sctx
->tes_shader
!= !!sel
;
794 if (sctx
->tes_shader
== sel
)
797 sctx
->tes_shader
= sel
;
798 si_mark_atom_dirty(sctx
, &sctx
->clip_regs
);
799 sctx
->last_rast_prim
= -1; /* reset this so that it gets updated */
801 if (enable_changed
) {
802 si_shader_change_notify(sctx
);
803 sctx
->last_tes_sh_base
= -1; /* invalidate derived tess state */
807 static void si_make_dummy_ps(struct si_context
*sctx
)
809 if (!sctx
->dummy_pixel_shader
) {
810 sctx
->dummy_pixel_shader
=
811 util_make_fragment_cloneinput_shader(&sctx
->b
.b
, 0,
812 TGSI_SEMANTIC_GENERIC
,
813 TGSI_INTERPOLATE_CONSTANT
);
817 static void si_bind_ps_shader(struct pipe_context
*ctx
, void *state
)
819 struct si_context
*sctx
= (struct si_context
*)ctx
;
820 struct si_shader_selector
*sel
= state
;
822 /* skip if supplied shader is one already in use */
823 if (sctx
->ps_shader
== sel
)
826 /* use a dummy shader if binding a NULL shader */
828 si_make_dummy_ps(sctx
);
829 sel
= sctx
->dummy_pixel_shader
;
832 sctx
->ps_shader
= sel
;
835 static void si_delete_shader_selector(struct pipe_context
*ctx
,
836 struct si_shader_selector
*sel
)
838 struct si_context
*sctx
= (struct si_context
*)ctx
;
839 struct si_shader
*p
= sel
->current
, *c
;
844 case PIPE_SHADER_VERTEX
:
846 si_pm4_delete_state(sctx
, ls
, p
->pm4
);
847 else if (p
->key
.vs
.as_es
)
848 si_pm4_delete_state(sctx
, es
, p
->pm4
);
850 si_pm4_delete_state(sctx
, vs
, p
->pm4
);
852 case PIPE_SHADER_TESS_CTRL
:
853 si_pm4_delete_state(sctx
, hs
, p
->pm4
);
855 case PIPE_SHADER_TESS_EVAL
:
856 if (p
->key
.tes
.as_es
)
857 si_pm4_delete_state(sctx
, es
, p
->pm4
);
859 si_pm4_delete_state(sctx
, vs
, p
->pm4
);
861 case PIPE_SHADER_GEOMETRY
:
862 si_pm4_delete_state(sctx
, gs
, p
->pm4
);
863 si_pm4_delete_state(sctx
, vs
, p
->gs_copy_shader
->pm4
);
865 case PIPE_SHADER_FRAGMENT
:
866 si_pm4_delete_state(sctx
, ps
, p
->pm4
);
870 si_shader_destroy(ctx
, p
);
879 static void si_delete_vs_shader(struct pipe_context
*ctx
, void *state
)
881 struct si_context
*sctx
= (struct si_context
*)ctx
;
882 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
884 if (sctx
->vs_shader
== sel
) {
885 sctx
->vs_shader
= NULL
;
888 si_delete_shader_selector(ctx
, sel
);
891 static void si_delete_gs_shader(struct pipe_context
*ctx
, void *state
)
893 struct si_context
*sctx
= (struct si_context
*)ctx
;
894 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
896 if (sctx
->gs_shader
== sel
) {
897 sctx
->gs_shader
= NULL
;
900 si_delete_shader_selector(ctx
, sel
);
903 static void si_delete_ps_shader(struct pipe_context
*ctx
, void *state
)
905 struct si_context
*sctx
= (struct si_context
*)ctx
;
906 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
908 if (sctx
->ps_shader
== sel
) {
909 sctx
->ps_shader
= NULL
;
912 si_delete_shader_selector(ctx
, sel
);
915 static void si_delete_tcs_shader(struct pipe_context
*ctx
, void *state
)
917 struct si_context
*sctx
= (struct si_context
*)ctx
;
918 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
920 if (sctx
->tcs_shader
== sel
) {
921 sctx
->tcs_shader
= NULL
;
924 si_delete_shader_selector(ctx
, sel
);
927 static void si_delete_tes_shader(struct pipe_context
*ctx
, void *state
)
929 struct si_context
*sctx
= (struct si_context
*)ctx
;
930 struct si_shader_selector
*sel
= (struct si_shader_selector
*)state
;
932 if (sctx
->tes_shader
== sel
) {
933 sctx
->tes_shader
= NULL
;
936 si_delete_shader_selector(ctx
, sel
);
939 static void si_update_spi_map(struct si_context
*sctx
)
941 struct si_shader
*ps
= sctx
->ps_shader
->current
;
942 struct si_shader
*vs
= si_get_vs_state(sctx
);
943 struct tgsi_shader_info
*psinfo
= &ps
->selector
->info
;
944 struct tgsi_shader_info
*vsinfo
= &vs
->selector
->info
;
945 struct si_pm4_state
*pm4
= CALLOC_STRUCT(si_pm4_state
);
948 for (i
= 0; i
< psinfo
->num_inputs
; i
++) {
949 unsigned name
= psinfo
->input_semantic_name
[i
];
950 unsigned index
= psinfo
->input_semantic_index
[i
];
951 unsigned interpolate
= psinfo
->input_interpolate
[i
];
952 unsigned param_offset
= ps
->ps_input_param_offset
[i
];
954 if (name
== TGSI_SEMANTIC_POSITION
||
955 name
== TGSI_SEMANTIC_FACE
)
956 /* Read from preloaded VGPRs, not parameters */
962 if (interpolate
== TGSI_INTERPOLATE_CONSTANT
||
963 (interpolate
== TGSI_INTERPOLATE_COLOR
&& sctx
->flatshade
))
964 tmp
|= S_028644_FLAT_SHADE(1);
966 if (name
== TGSI_SEMANTIC_PCOORD
||
967 (name
== TGSI_SEMANTIC_TEXCOORD
&&
968 sctx
->sprite_coord_enable
& (1 << index
))) {
969 tmp
|= S_028644_PT_SPRITE_TEX(1);
972 for (j
= 0; j
< vsinfo
->num_outputs
; j
++) {
973 if (name
== vsinfo
->output_semantic_name
[j
] &&
974 index
== vsinfo
->output_semantic_index
[j
]) {
975 tmp
|= S_028644_OFFSET(vs
->vs_output_param_offset
[j
]);
980 if (j
== vsinfo
->num_outputs
&& !G_028644_PT_SPRITE_TEX(tmp
)) {
981 /* No corresponding output found, load defaults into input.
982 * Don't set any other bits.
983 * (FLAT_SHADE=1 completely changes behavior) */
984 tmp
= S_028644_OFFSET(0x20);
988 R_028644_SPI_PS_INPUT_CNTL_0
+ param_offset
* 4,
991 if (name
== TGSI_SEMANTIC_COLOR
&&
992 ps
->key
.ps
.color_two_side
) {
993 name
= TGSI_SEMANTIC_BCOLOR
;
999 si_pm4_set_state(sctx
, spi
, pm4
);
1002 /* Initialize state related to ESGS / GSVS ring buffers */
1003 static void si_init_gs_rings(struct si_context
*sctx
)
1005 unsigned esgs_ring_size
= 128 * 1024;
1006 unsigned gsvs_ring_size
= 64 * 1024 * 1024;
1008 assert(!sctx
->gs_rings
);
1009 sctx
->gs_rings
= CALLOC_STRUCT(si_pm4_state
);
1011 sctx
->esgs_ring
= pipe_buffer_create(sctx
->b
.b
.screen
, PIPE_BIND_CUSTOM
,
1012 PIPE_USAGE_DEFAULT
, esgs_ring_size
);
1014 sctx
->gsvs_ring
= pipe_buffer_create(sctx
->b
.b
.screen
, PIPE_BIND_CUSTOM
,
1015 PIPE_USAGE_DEFAULT
, gsvs_ring_size
);
1017 if (sctx
->b
.chip_class
>= CIK
) {
1018 si_pm4_set_reg(sctx
->gs_rings
, R_030900_VGT_ESGS_RING_SIZE
,
1019 esgs_ring_size
/ 256);
1020 si_pm4_set_reg(sctx
->gs_rings
, R_030904_VGT_GSVS_RING_SIZE
,
1021 gsvs_ring_size
/ 256);
1023 si_pm4_set_reg(sctx
->gs_rings
, R_0088C8_VGT_ESGS_RING_SIZE
,
1024 esgs_ring_size
/ 256);
1025 si_pm4_set_reg(sctx
->gs_rings
, R_0088CC_VGT_GSVS_RING_SIZE
,
1026 gsvs_ring_size
/ 256);
1029 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_VERTEX
, SI_RING_ESGS
,
1030 sctx
->esgs_ring
, 0, esgs_ring_size
,
1031 true, true, 4, 64, 0);
1032 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_GEOMETRY
, SI_RING_ESGS
,
1033 sctx
->esgs_ring
, 0, esgs_ring_size
,
1034 false, false, 0, 0, 0);
1035 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_VERTEX
, SI_RING_GSVS
,
1036 sctx
->gsvs_ring
, 0, gsvs_ring_size
,
1037 false, false, 0, 0, 0);
1040 static void si_update_gs_rings(struct si_context
*sctx
)
1042 unsigned gs_vert_itemsize
= sctx
->gs_shader
->info
.num_outputs
* 16;
1043 unsigned gs_max_vert_out
= sctx
->gs_shader
->gs_max_out_vertices
;
1044 unsigned gsvs_itemsize
= gs_vert_itemsize
* gs_max_vert_out
;
1047 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_GEOMETRY
, SI_RING_GSVS
,
1048 sctx
->gsvs_ring
, gsvs_itemsize
,
1049 64, true, true, 4, 16, 0);
1051 offset
= gsvs_itemsize
* 64;
1052 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_GEOMETRY
, SI_RING_GSVS_1
,
1053 sctx
->gsvs_ring
, gsvs_itemsize
,
1054 64, true, true, 4, 16, offset
);
1056 offset
= (gsvs_itemsize
* 2) * 64;
1057 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_GEOMETRY
, SI_RING_GSVS_2
,
1058 sctx
->gsvs_ring
, gsvs_itemsize
,
1059 64, true, true, 4, 16, offset
);
1061 offset
= (gsvs_itemsize
* 3) * 64;
1062 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_GEOMETRY
, SI_RING_GSVS_3
,
1063 sctx
->gsvs_ring
, gsvs_itemsize
,
1064 64, true, true, 4, 16, offset
);
1068 * @returns 1 if \p sel has been updated to use a new scratch buffer and 0
1071 static unsigned si_update_scratch_buffer(struct si_context
*sctx
,
1072 struct si_shader_selector
*sel
)
1074 struct si_shader
*shader
;
1075 uint64_t scratch_va
= sctx
->scratch_buffer
->gpu_address
;
1080 shader
= sel
->current
;
1082 /* This shader doesn't need a scratch buffer */
1083 if (shader
->scratch_bytes_per_wave
== 0)
1086 /* This shader is already configured to use the current
1087 * scratch buffer. */
1088 if (shader
->scratch_bo
== sctx
->scratch_buffer
)
1091 assert(sctx
->scratch_buffer
);
1093 si_shader_apply_scratch_relocs(sctx
, shader
, scratch_va
);
1095 /* Replace the shader bo with a new bo that has the relocs applied. */
1096 si_shader_binary_upload(sctx
->screen
, shader
);
1098 /* Update the shader state to use the new shader bo. */
1099 si_shader_init_pm4_state(shader
);
1101 r600_resource_reference(&shader
->scratch_bo
, sctx
->scratch_buffer
);
1106 static unsigned si_get_current_scratch_buffer_size(struct si_context
*sctx
)
1108 if (!sctx
->scratch_buffer
)
1111 return sctx
->scratch_buffer
->b
.b
.width0
;
1114 static unsigned si_get_scratch_buffer_bytes_per_wave(struct si_context
*sctx
,
1115 struct si_shader_selector
*sel
)
1120 return sel
->current
->scratch_bytes_per_wave
;
1123 static unsigned si_get_max_scratch_bytes_per_wave(struct si_context
*sctx
)
1127 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
, sctx
->ps_shader
));
1128 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
, sctx
->gs_shader
));
1129 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
, sctx
->vs_shader
));
1130 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
, sctx
->tcs_shader
));
1131 bytes
= MAX2(bytes
, si_get_scratch_buffer_bytes_per_wave(sctx
, sctx
->tes_shader
));
1135 static void si_update_spi_tmpring_size(struct si_context
*sctx
)
1137 unsigned current_scratch_buffer_size
=
1138 si_get_current_scratch_buffer_size(sctx
);
1139 unsigned scratch_bytes_per_wave
=
1140 si_get_max_scratch_bytes_per_wave(sctx
);
1141 unsigned scratch_needed_size
= scratch_bytes_per_wave
*
1142 sctx
->scratch_waves
;
1144 if (scratch_needed_size
> 0) {
1146 if (scratch_needed_size
> current_scratch_buffer_size
) {
1147 /* Create a bigger scratch buffer */
1148 pipe_resource_reference(
1149 (struct pipe_resource
**)&sctx
->scratch_buffer
,
1152 sctx
->scratch_buffer
=
1153 si_resource_create_custom(&sctx
->screen
->b
.b
,
1154 PIPE_USAGE_DEFAULT
, scratch_needed_size
);
1157 /* Update the shaders, so they are using the latest scratch. The
1158 * scratch buffer may have been changed since these shaders were
1159 * last used, so we still need to try to update them, even if
1160 * they require scratch buffers smaller than the current size.
1162 if (si_update_scratch_buffer(sctx
, sctx
->ps_shader
))
1163 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
->current
->pm4
);
1164 if (si_update_scratch_buffer(sctx
, sctx
->gs_shader
))
1165 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
->current
->pm4
);
1166 if (si_update_scratch_buffer(sctx
, sctx
->tcs_shader
))
1167 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
->current
->pm4
);
1169 /* VS can be bound as LS, ES, or VS. */
1170 if (sctx
->tes_shader
) {
1171 if (si_update_scratch_buffer(sctx
, sctx
->vs_shader
))
1172 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
->current
->pm4
);
1173 } else if (sctx
->gs_shader
) {
1174 if (si_update_scratch_buffer(sctx
, sctx
->vs_shader
))
1175 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
->current
->pm4
);
1177 if (si_update_scratch_buffer(sctx
, sctx
->vs_shader
))
1178 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
->current
->pm4
);
1181 /* TES can be bound as ES or VS. */
1182 if (sctx
->gs_shader
) {
1183 if (si_update_scratch_buffer(sctx
, sctx
->tes_shader
))
1184 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
->current
->pm4
);
1186 if (si_update_scratch_buffer(sctx
, sctx
->tes_shader
))
1187 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
->current
->pm4
);
1191 /* The LLVM shader backend should be reporting aligned scratch_sizes. */
1192 assert((scratch_needed_size
& ~0x3FF) == scratch_needed_size
&&
1193 "scratch size should already be aligned correctly.");
1195 sctx
->spi_tmpring_size
= S_0286E8_WAVES(sctx
->scratch_waves
) |
1196 S_0286E8_WAVESIZE(scratch_bytes_per_wave
>> 10);
1199 static void si_init_tess_factor_ring(struct si_context
*sctx
)
1201 assert(!sctx
->tf_state
);
1202 sctx
->tf_state
= CALLOC_STRUCT(si_pm4_state
);
1204 sctx
->tf_ring
= pipe_buffer_create(sctx
->b
.b
.screen
, PIPE_BIND_CUSTOM
,
1206 32768 * sctx
->screen
->b
.info
.max_se
);
1207 sctx
->b
.clear_buffer(&sctx
->b
.b
, sctx
->tf_ring
, 0,
1208 sctx
->tf_ring
->width0
, fui(0), false);
1209 assert(((sctx
->tf_ring
->width0
/ 4) & C_030938_SIZE
) == 0);
1211 if (sctx
->b
.chip_class
>= CIK
) {
1212 si_pm4_set_reg(sctx
->tf_state
, R_030938_VGT_TF_RING_SIZE
,
1213 S_030938_SIZE(sctx
->tf_ring
->width0
/ 4));
1214 si_pm4_set_reg(sctx
->tf_state
, R_030940_VGT_TF_MEMORY_BASE
,
1215 r600_resource(sctx
->tf_ring
)->gpu_address
>> 8);
1217 si_pm4_set_reg(sctx
->tf_state
, R_008988_VGT_TF_RING_SIZE
,
1218 S_008988_SIZE(sctx
->tf_ring
->width0
/ 4));
1219 si_pm4_set_reg(sctx
->tf_state
, R_0089B8_VGT_TF_MEMORY_BASE
,
1220 r600_resource(sctx
->tf_ring
)->gpu_address
>> 8);
1222 si_pm4_add_bo(sctx
->tf_state
, r600_resource(sctx
->tf_ring
),
1223 RADEON_USAGE_READWRITE
, RADEON_PRIO_SHADER_RESOURCE_RW
);
1224 si_pm4_bind_state(sctx
, tf_ring
, sctx
->tf_state
);
1226 si_set_ring_buffer(&sctx
->b
.b
, PIPE_SHADER_TESS_CTRL
,
1227 SI_RING_TESS_FACTOR
, sctx
->tf_ring
, 0,
1228 sctx
->tf_ring
->width0
, false, false, 0, 0, 0);
1230 sctx
->b
.flags
|= SI_CONTEXT_VGT_FLUSH
;
1234 * This is used when TCS is NULL in the VS->TCS->TES chain. In this case,
1235 * VS passes its outputs to TES directly, so the fixed-function shader only
1236 * has to write TESSOUTER and TESSINNER.
1238 static void si_generate_fixed_func_tcs(struct si_context
*sctx
)
1240 struct ureg_src const0
, const1
;
1241 struct ureg_dst tessouter
, tessinner
;
1242 struct ureg_program
*ureg
= ureg_create(TGSI_PROCESSOR_TESS_CTRL
);
1245 return; /* if we get here, we're screwed */
1247 assert(!sctx
->fixed_func_tcs_shader
);
1249 ureg_DECL_constant2D(ureg
, 0, 1, SI_DRIVER_STATE_CONST_BUF
);
1250 const0
= ureg_src_dimension(ureg_src_register(TGSI_FILE_CONSTANT
, 0),
1251 SI_DRIVER_STATE_CONST_BUF
);
1252 const1
= ureg_src_dimension(ureg_src_register(TGSI_FILE_CONSTANT
, 1),
1253 SI_DRIVER_STATE_CONST_BUF
);
1255 tessouter
= ureg_DECL_output(ureg
, TGSI_SEMANTIC_TESSOUTER
, 0);
1256 tessinner
= ureg_DECL_output(ureg
, TGSI_SEMANTIC_TESSINNER
, 0);
1258 ureg_MOV(ureg
, tessouter
, const0
);
1259 ureg_MOV(ureg
, tessinner
, const1
);
1262 sctx
->fixed_func_tcs_shader
=
1263 ureg_create_shader_and_destroy(ureg
, &sctx
->b
.b
);
1264 assert(sctx
->fixed_func_tcs_shader
);
1267 static void si_update_vgt_shader_config(struct si_context
*sctx
)
1269 /* Calculate the index of the config.
1270 * 0 = VS, 1 = VS+GS, 2 = VS+Tess, 3 = VS+Tess+GS */
1271 unsigned index
= 2*!!sctx
->tes_shader
+ !!sctx
->gs_shader
;
1272 struct si_pm4_state
**pm4
= &sctx
->vgt_shader_config
[index
];
1275 uint32_t stages
= 0;
1277 *pm4
= CALLOC_STRUCT(si_pm4_state
);
1279 if (sctx
->tes_shader
) {
1280 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
1283 if (sctx
->gs_shader
)
1284 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
1286 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
1288 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
1289 } else if (sctx
->gs_shader
) {
1290 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
1292 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
1295 si_pm4_set_reg(*pm4
, R_028B54_VGT_SHADER_STAGES_EN
, stages
);
1297 si_pm4_bind_state(sctx
, vgt_shader_config
, *pm4
);
1300 static void si_update_so(struct si_context
*sctx
, struct si_shader_selector
*shader
)
1302 struct pipe_stream_output_info
*so
= &shader
->so
;
1303 uint32_t enabled_stream_buffers_mask
= 0;
1306 for (i
= 0; i
< so
->num_outputs
; i
++)
1307 enabled_stream_buffers_mask
|= (1 << so
->output
[i
].output_buffer
) << (so
->output
[i
].stream
* 4);
1308 sctx
->b
.streamout
.enabled_stream_buffers_mask
= enabled_stream_buffers_mask
;
1309 sctx
->b
.streamout
.stride_in_dw
= shader
->so
.stride
;
1312 void si_update_shaders(struct si_context
*sctx
)
1314 struct pipe_context
*ctx
= (struct pipe_context
*)sctx
;
1315 struct si_state_rasterizer
*rs
= sctx
->queued
.named
.rasterizer
;
1317 /* Update stages before GS. */
1318 if (sctx
->tes_shader
) {
1319 if (!sctx
->tf_state
)
1320 si_init_tess_factor_ring(sctx
);
1323 si_shader_select(ctx
, sctx
->vs_shader
);
1324 si_pm4_bind_state(sctx
, ls
, sctx
->vs_shader
->current
->pm4
);
1326 if (sctx
->tcs_shader
) {
1327 si_shader_select(ctx
, sctx
->tcs_shader
);
1328 si_pm4_bind_state(sctx
, hs
, sctx
->tcs_shader
->current
->pm4
);
1330 if (!sctx
->fixed_func_tcs_shader
)
1331 si_generate_fixed_func_tcs(sctx
);
1332 si_shader_select(ctx
, sctx
->fixed_func_tcs_shader
);
1333 si_pm4_bind_state(sctx
, hs
,
1334 sctx
->fixed_func_tcs_shader
->current
->pm4
);
1337 si_shader_select(ctx
, sctx
->tes_shader
);
1338 if (sctx
->gs_shader
) {
1340 si_pm4_bind_state(sctx
, es
, sctx
->tes_shader
->current
->pm4
);
1343 si_pm4_bind_state(sctx
, vs
, sctx
->tes_shader
->current
->pm4
);
1344 si_update_so(sctx
, sctx
->tes_shader
);
1346 } else if (sctx
->gs_shader
) {
1348 si_shader_select(ctx
, sctx
->vs_shader
);
1349 si_pm4_bind_state(sctx
, es
, sctx
->vs_shader
->current
->pm4
);
1352 si_shader_select(ctx
, sctx
->vs_shader
);
1353 si_pm4_bind_state(sctx
, vs
, sctx
->vs_shader
->current
->pm4
);
1354 si_update_so(sctx
, sctx
->vs_shader
);
1358 if (sctx
->gs_shader
) {
1359 si_shader_select(ctx
, sctx
->gs_shader
);
1360 si_pm4_bind_state(sctx
, gs
, sctx
->gs_shader
->current
->pm4
);
1361 si_pm4_bind_state(sctx
, vs
, sctx
->gs_shader
->current
->gs_copy_shader
->pm4
);
1362 si_update_so(sctx
, sctx
->gs_shader
);
1364 if (!sctx
->gs_rings
)
1365 si_init_gs_rings(sctx
);
1367 if (sctx
->emitted
.named
.gs_rings
!= sctx
->gs_rings
)
1368 sctx
->b
.flags
|= SI_CONTEXT_VGT_FLUSH
;
1369 si_pm4_bind_state(sctx
, gs_rings
, sctx
->gs_rings
);
1371 si_update_gs_rings(sctx
);
1373 si_pm4_bind_state(sctx
, gs_rings
, NULL
);
1374 si_pm4_bind_state(sctx
, gs
, NULL
);
1375 si_pm4_bind_state(sctx
, es
, NULL
);
1378 si_update_vgt_shader_config(sctx
);
1380 si_shader_select(ctx
, sctx
->ps_shader
);
1382 if (!sctx
->ps_shader
->current
) {
1383 struct si_shader_selector
*sel
;
1385 /* use a dummy shader if compiling the shader (variant) failed */
1386 si_make_dummy_ps(sctx
);
1387 sel
= sctx
->dummy_pixel_shader
;
1388 si_shader_select(ctx
, sel
);
1389 sctx
->ps_shader
->current
= sel
->current
;
1392 si_pm4_bind_state(sctx
, ps
, sctx
->ps_shader
->current
->pm4
);
1394 if (si_pm4_state_changed(sctx
, ps
) || si_pm4_state_changed(sctx
, vs
) ||
1395 sctx
->sprite_coord_enable
!= rs
->sprite_coord_enable
||
1396 sctx
->flatshade
!= rs
->flatshade
) {
1397 sctx
->sprite_coord_enable
= rs
->sprite_coord_enable
;
1398 sctx
->flatshade
= rs
->flatshade
;
1399 si_update_spi_map(sctx
);
1402 if (si_pm4_state_changed(sctx
, ps
) || si_pm4_state_changed(sctx
, vs
) ||
1403 si_pm4_state_changed(sctx
, gs
)) {
1404 si_update_spi_tmpring_size(sctx
);
1407 if (sctx
->ps_db_shader_control
!= sctx
->ps_shader
->current
->db_shader_control
) {
1408 sctx
->ps_db_shader_control
= sctx
->ps_shader
->current
->db_shader_control
;
1409 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
1412 if (sctx
->smoothing_enabled
!= sctx
->ps_shader
->current
->key
.ps
.poly_line_smoothing
) {
1413 sctx
->smoothing_enabled
= sctx
->ps_shader
->current
->key
.ps
.poly_line_smoothing
;
1414 si_mark_atom_dirty(sctx
, &sctx
->msaa_config
);
1416 if (sctx
->b
.chip_class
== SI
)
1417 si_mark_atom_dirty(sctx
, &sctx
->db_render_state
);
1421 void si_init_shader_functions(struct si_context
*sctx
)
1423 sctx
->b
.b
.create_vs_state
= si_create_vs_state
;
1424 sctx
->b
.b
.create_tcs_state
= si_create_tcs_state
;
1425 sctx
->b
.b
.create_tes_state
= si_create_tes_state
;
1426 sctx
->b
.b
.create_gs_state
= si_create_gs_state
;
1427 sctx
->b
.b
.create_fs_state
= si_create_fs_state
;
1429 sctx
->b
.b
.bind_vs_state
= si_bind_vs_shader
;
1430 sctx
->b
.b
.bind_tcs_state
= si_bind_tcs_shader
;
1431 sctx
->b
.b
.bind_tes_state
= si_bind_tes_shader
;
1432 sctx
->b
.b
.bind_gs_state
= si_bind_gs_shader
;
1433 sctx
->b
.b
.bind_fs_state
= si_bind_ps_shader
;
1435 sctx
->b
.b
.delete_vs_state
= si_delete_vs_shader
;
1436 sctx
->b
.b
.delete_tcs_state
= si_delete_tcs_shader
;
1437 sctx
->b
.b
.delete_tes_state
= si_delete_tes_shader
;
1438 sctx
->b
.b
.delete_gs_state
= si_delete_gs_shader
;
1439 sctx
->b
.b
.delete_fs_state
= si_delete_ps_shader
;