2 * Copyright © 2013 Intel Corporation
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * 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 NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
27 * State atom for client-programmable geometry shaders, and support code.
31 #include "brw_context.h"
32 #include "brw_vec4_gs_visitor.h"
33 #include "brw_state.h"
34 #include "brw_ff_gs.h"
38 assign_gs_binding_table_offsets(const struct brw_device_info
*devinfo
,
39 const struct gl_shader_program
*shader_prog
,
40 const struct gl_program
*prog
,
41 struct brw_gs_prog_data
*prog_data
)
43 /* In gen6 we reserve the first BRW_MAX_SOL_BINDINGS entries for transform
46 uint32_t reserved
= devinfo
->gen
== 6 ? BRW_MAX_SOL_BINDINGS
: 0;
48 brw_assign_common_binding_table_offsets(MESA_SHADER_GEOMETRY
, devinfo
,
50 &prog_data
->base
.base
,
55 brw_codegen_gs_prog(struct brw_context
*brw
,
56 struct gl_shader_program
*prog
,
57 struct brw_geometry_program
*gp
,
58 struct brw_gs_prog_key
*key
)
60 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
61 struct brw_gs_compile c
;
62 memset(&c
, 0, sizeof(c
));
66 c
.prog_data
.include_primitive_id
=
67 (gp
->program
.Base
.InputsRead
& VARYING_BIT_PRIMITIVE_ID
) != 0;
69 c
.prog_data
.invocations
= gp
->program
.Invocations
;
71 assign_gs_binding_table_offsets(brw
->intelScreen
->devinfo
, prog
,
72 &gp
->program
.Base
, &c
.prog_data
);
74 /* Allocate the references to the uniforms that will end up in the
75 * prog_data associated with the compiled program, and which will be freed
78 * Note: param_count needs to be num_uniform_components * 4, since we add
79 * padding around uniform values below vec4 size, so the worst case is that
80 * every uniform is a float which gets padded to the size of a vec4.
82 struct gl_shader
*gs
= prog
->_LinkedShaders
[MESA_SHADER_GEOMETRY
];
83 int param_count
= gp
->program
.Base
.nir
->num_uniforms
* 4;
85 c
.prog_data
.base
.base
.param
=
86 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
87 c
.prog_data
.base
.base
.pull_param
=
88 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
89 c
.prog_data
.base
.base
.image_param
=
90 rzalloc_array(NULL
, struct brw_image_param
, gs
->NumImages
);
91 c
.prog_data
.base
.base
.nr_params
= param_count
;
92 c
.prog_data
.base
.base
.nr_image_params
= gs
->NumImages
;
94 brw_nir_setup_glsl_uniforms(gp
->program
.Base
.nir
, prog
, &gp
->program
.Base
,
95 &c
.prog_data
.base
.base
, false);
98 c
.prog_data
.static_vertex_count
= !gp
->program
.Base
.nir
? -1 :
99 nir_gs_count_vertices(gp
->program
.Base
.nir
);
103 if (gp
->program
.OutputType
== GL_POINTS
) {
104 /* When the output type is points, the geometry shader may output data
105 * to multiple streams, and EndPrimitive() has no effect. So we
106 * configure the hardware to interpret the control data as stream ID.
108 c
.prog_data
.control_data_format
= GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
;
110 /* We only have to emit control bits if we are using streams */
111 if (prog
->Geom
.UsesStreams
)
112 c
.control_data_bits_per_vertex
= 2;
114 c
.control_data_bits_per_vertex
= 0;
116 /* When the output type is triangle_strip or line_strip, EndPrimitive()
117 * may be used to terminate the current strip and start a new one
118 * (similar to primitive restart), and outputting data to multiple
119 * streams is not supported. So we configure the hardware to interpret
120 * the control data as EndPrimitive information (a.k.a. "cut bits").
122 c
.prog_data
.control_data_format
= GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
;
124 /* We only need to output control data if the shader actually calls
127 c
.control_data_bits_per_vertex
= gp
->program
.UsesEndPrimitive
? 1 : 0;
130 /* There are no control data bits in gen6. */
131 c
.control_data_bits_per_vertex
= 0;
133 /* If it is using transform feedback, enable it */
134 if (prog
->TransformFeedback
.NumVarying
)
135 c
.prog_data
.gen6_xfb_enabled
= true;
137 c
.prog_data
.gen6_xfb_enabled
= false;
139 c
.control_data_header_size_bits
=
140 gp
->program
.VerticesOut
* c
.control_data_bits_per_vertex
;
142 /* 1 HWORD = 32 bytes = 256 bits */
143 c
.prog_data
.control_data_header_size_hwords
=
144 ALIGN(c
.control_data_header_size_bits
, 256) / 256;
146 GLbitfield64 outputs_written
= gp
->program
.Base
.OutputsWritten
;
148 brw_compute_vue_map(brw
->intelScreen
->devinfo
,
149 &c
.prog_data
.base
.vue_map
, outputs_written
,
150 prog
? prog
->SeparateShader
: false);
152 /* Compute the output vertex size.
154 * From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
157 * [0,62] indicating [1,63] 16B units
159 * Specifies the size of each vertex stored in the GS output entry
160 * (following any Control Header data) as a number of 128-bit units
163 * Programming Restrictions: The vertex size must be programmed as a
164 * multiple of 32B units with the following exception: Rendering is
165 * disabled (as per SOL stage state) and the vertex size output by the
168 * If rendering is enabled (as per SOL state) the vertex size must be
169 * programmed as a multiple of 32B units. In other words, the only time
170 * software can program a vertex size with an odd number of 16B units
171 * is when rendering is disabled.
173 * Note: B=bytes in the above text.
175 * It doesn't seem worth the extra trouble to optimize the case where the
176 * vertex size is 16B (especially since this would require special-casing
177 * the GEN assembly that writes to the URB). So we just set the vertex
178 * size to a multiple of 32B (2 vec4's) in all cases.
180 * The maximum output vertex size is 62*16 = 992 bytes (31 hwords). We
181 * budget that as follows:
183 * 512 bytes for varyings (a varying component is 4 bytes and
184 * gl_MaxGeometryOutputComponents = 128)
185 * 16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
187 * 16 bytes overhead for gl_Position (we allocate it a slot in the VUE
188 * even if it's not used)
189 * 32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
190 * whenever clip planes are enabled, even if the shader doesn't
191 * write to gl_ClipDistance)
192 * 16 bytes overhead since the VUE size must be a multiple of 32 bytes
193 * (see below)--this causes up to 1 VUE slot to be wasted
194 * 400 bytes available for varying packing overhead
196 * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
197 * per interpolation type, so this is plenty.
200 unsigned output_vertex_size_bytes
= c
.prog_data
.base
.vue_map
.num_slots
* 16;
201 assert(brw
->gen
== 6 ||
202 output_vertex_size_bytes
<= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES
);
203 c
.prog_data
.output_vertex_size_hwords
=
204 ALIGN(output_vertex_size_bytes
, 32) / 32;
206 /* Compute URB entry size. The maximum allowed URB entry size is 32k.
207 * That divides up as follows:
209 * 64 bytes for the control data header (cut indices or StreamID bits)
210 * 4096 bytes for varyings (a varying component is 4 bytes and
211 * gl_MaxGeometryTotalOutputComponents = 1024)
212 * 4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
213 * bytes/vertex and gl_MaxGeometryOutputVertices is 256)
214 * 4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
215 * even if it's not used)
216 * 8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
217 * whenever clip planes are enabled, even if the shader doesn't
218 * write to gl_ClipDistance)
219 * 4096 bytes overhead since the VUE size must be a multiple of 32
220 * bytes (see above)--this causes up to 1 VUE slot to be wasted
221 * 8128 bytes available for varying packing overhead
223 * Worst-case varying packing overhead is 3/4 of a varying slot per
224 * interpolation type, which works out to 3072 bytes, so this would allow
225 * us to accommodate 2 interpolation types without any danger of running
228 * In practice, the risk of running out of URB space is very small, since
229 * the above figures are all worst-case, and most of them scale with the
230 * number of output vertices. So we'll just calculate the amount of space
231 * we need, and if it's too large, fail to compile.
233 * The above is for gen7+ where we have a single URB entry that will hold
234 * all the output. In gen6, we will have to allocate URB entries for every
235 * vertex we emit, so our URB entries only need to be large enough to hold
236 * a single vertex. Also, gen6 does not have a control data header.
238 unsigned output_size_bytes
;
241 c
.prog_data
.output_vertex_size_hwords
* 32 * gp
->program
.VerticesOut
;
242 output_size_bytes
+= 32 * c
.prog_data
.control_data_header_size_hwords
;
244 output_size_bytes
= c
.prog_data
.output_vertex_size_hwords
* 32;
247 /* Broadwell stores "Vertex Count" as a full 8 DWord (32 byte) URB output,
248 * which comes before the control header.
251 output_size_bytes
+= 32;
253 assert(output_size_bytes
>= 1);
254 int max_output_size_bytes
= GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES
;
256 max_output_size_bytes
= GEN6_MAX_GS_URB_ENTRY_SIZE_BYTES
;
257 if (output_size_bytes
> max_output_size_bytes
)
261 /* URB entry sizes are stored as a multiple of 64 bytes in gen7+ and
262 * a multiple of 128 bytes in gen6.
265 c
.prog_data
.base
.urb_entry_size
= ALIGN(output_size_bytes
, 64) / 64;
267 c
.prog_data
.base
.urb_entry_size
= ALIGN(output_size_bytes
, 128) / 128;
269 c
.prog_data
.output_topology
=
270 get_hw_prim_for_gl_prim(gp
->program
.OutputType
);
272 /* The GLSL linker will have already matched up GS inputs and the outputs
273 * of prior stages. The driver does extend VS outputs in some cases, but
274 * only for legacy OpenGL or Gen4-5 hardware, neither of which offer
275 * geometry shader support. So we can safely ignore that.
277 * For SSO pipelines, we use a fixed VUE map layout based on variable
278 * locations, so we can rely on rendezvous-by-location making this work.
280 * However, we need to ignore VARYING_SLOT_PRIMITIVE_ID, as it's not
281 * written by previous stages and shows up via payload magic.
283 GLbitfield64 inputs_read
=
284 gp
->program
.Base
.InputsRead
& ~VARYING_BIT_PRIMITIVE_ID
;
285 brw_compute_vue_map(brw
->intelScreen
->devinfo
,
286 &c
.input_vue_map
, inputs_read
,
287 prog
->SeparateShader
);
289 /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
290 * need to program a URB read length of ceiling(num_slots / 2).
292 c
.prog_data
.base
.urb_read_length
= (c
.input_vue_map
.num_slots
+ 1) / 2;
294 if (unlikely(INTEL_DEBUG
& DEBUG_GS
))
295 brw_dump_ir("geometry", prog
, gs
, NULL
);
298 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
299 st_index
= brw_get_shader_time_index(brw
, prog
, NULL
, ST_GS
);
301 void *mem_ctx
= ralloc_context(NULL
);
302 unsigned program_size
;
303 const unsigned *program
=
304 brw_gs_emit(brw
, prog
, &c
, mem_ctx
, st_index
, &program_size
);
305 if (program
== NULL
) {
306 ralloc_free(mem_ctx
);
310 /* Scratch space is used for register spilling */
311 if (c
.prog_data
.base
.base
.total_scratch
) {
312 brw_get_scratch_bo(brw
, &stage_state
->scratch_bo
,
313 c
.prog_data
.base
.base
.total_scratch
*
314 brw
->max_gs_threads
);
317 brw_upload_cache(&brw
->cache
, BRW_CACHE_GS_PROG
,
318 &c
.key
, sizeof(c
.key
),
319 program
, program_size
,
320 &c
.prog_data
, sizeof(c
.prog_data
),
321 &stage_state
->prog_offset
, &brw
->gs
.prog_data
);
322 ralloc_free(mem_ctx
);
328 brw_gs_state_dirty(struct brw_context
*brw
)
330 return brw_state_dirty(brw
,
332 BRW_NEW_GEOMETRY_PROGRAM
|
333 BRW_NEW_TRANSFORM_FEEDBACK
);
337 brw_gs_populate_key(struct brw_context
*brw
,
338 struct brw_gs_prog_key
*key
)
340 struct gl_context
*ctx
= &brw
->ctx
;
341 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
342 struct brw_geometry_program
*gp
=
343 (struct brw_geometry_program
*) brw
->geometry_program
;
344 struct gl_program
*prog
= &gp
->program
.Base
;
346 memset(key
, 0, sizeof(*key
));
348 key
->program_string_id
= gp
->id
;
351 brw_populate_sampler_prog_key_data(ctx
, prog
, stage_state
->sampler_count
,
356 brw_upload_gs_prog(struct brw_context
*brw
)
358 struct gl_context
*ctx
= &brw
->ctx
;
359 struct gl_shader_program
**current
= ctx
->_Shader
->CurrentProgram
;
360 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
361 struct brw_gs_prog_key key
;
362 /* BRW_NEW_GEOMETRY_PROGRAM */
363 struct brw_geometry_program
*gp
=
364 (struct brw_geometry_program
*) brw
->geometry_program
;
366 if (!brw_gs_state_dirty(brw
))
370 /* No geometry shader. Vertex data just passes straight through. */
372 (brw
->ctx
.NewDriverState
& BRW_NEW_TRANSFORM_FEEDBACK
)) {
373 gen6_brw_upload_ff_gs_prog(brw
);
377 /* Other state atoms had better not try to access prog_data, since
378 * there's no GS program.
380 brw
->gs
.prog_data
= NULL
;
381 brw
->gs
.base
.prog_data
= NULL
;
386 brw_gs_populate_key(brw
, &key
);
388 if (!brw_search_cache(&brw
->cache
, BRW_CACHE_GS_PROG
,
390 &stage_state
->prog_offset
, &brw
->gs
.prog_data
)) {
391 bool success
= brw_codegen_gs_prog(brw
, current
[MESA_SHADER_GEOMETRY
],
396 brw
->gs
.base
.prog_data
= &brw
->gs
.prog_data
->base
.base
;
400 brw_gs_precompile(struct gl_context
*ctx
,
401 struct gl_shader_program
*shader_prog
,
402 struct gl_program
*prog
)
404 struct brw_context
*brw
= brw_context(ctx
);
405 struct brw_gs_prog_key key
;
406 uint32_t old_prog_offset
= brw
->gs
.base
.prog_offset
;
407 struct brw_gs_prog_data
*old_prog_data
= brw
->gs
.prog_data
;
410 struct gl_geometry_program
*gp
= (struct gl_geometry_program
*) prog
;
411 struct brw_geometry_program
*bgp
= brw_geometry_program(gp
);
413 memset(&key
, 0, sizeof(key
));
415 brw_setup_tex_for_precompile(brw
, &key
.tex
, prog
);
416 key
.program_string_id
= bgp
->id
;
418 success
= brw_codegen_gs_prog(brw
, shader_prog
, bgp
, &key
);
420 brw
->gs
.base
.prog_offset
= old_prog_offset
;
421 brw
->gs
.prog_data
= old_prog_data
;