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_compile_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
,
59 struct brw_gs_compile_output
*output
)
61 struct brw_gs_compile c
;
62 memset(&c
, 0, sizeof(c
));
66 /* We get the bind map as input in the output struct...*/
67 c
.prog_data
.base
.base
.map_entries
= output
->prog_data
.base
.base
.map_entries
;
68 memcpy(c
.prog_data
.base
.base
.bind_map
, output
->prog_data
.base
.base
.bind_map
,
69 sizeof(c
.prog_data
.base
.base
.bind_map
));
71 c
.prog_data
.include_primitive_id
=
72 (gp
->program
.Base
.InputsRead
& VARYING_BIT_PRIMITIVE_ID
) != 0;
74 c
.prog_data
.invocations
= gp
->program
.Invocations
;
76 assign_gs_binding_table_offsets(brw
->intelScreen
->devinfo
, prog
,
77 &gp
->program
.Base
, &c
.prog_data
);
79 /* Allocate the references to the uniforms that will end up in the
80 * prog_data associated with the compiled program, and which will be freed
83 * Note: param_count needs to be num_uniform_components * 4, since we add
84 * padding around uniform values below vec4 size, so the worst case is that
85 * every uniform is a float which gets padded to the size of a vec4.
87 struct gl_shader
*gs
= prog
->_LinkedShaders
[MESA_SHADER_GEOMETRY
];
88 int param_count
= gp
->program
.Base
.nir
->num_uniforms
* 4;
90 c
.prog_data
.base
.base
.param
=
91 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
92 c
.prog_data
.base
.base
.pull_param
=
93 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
94 c
.prog_data
.base
.base
.image_param
=
95 rzalloc_array(NULL
, struct brw_image_param
, gs
->NumImages
);
96 c
.prog_data
.base
.base
.nr_params
= param_count
;
97 c
.prog_data
.base
.base
.nr_image_params
= gs
->NumImages
;
99 brw_nir_setup_glsl_uniforms(gp
->program
.Base
.nir
, prog
, &gp
->program
.Base
,
100 &c
.prog_data
.base
.base
, false);
103 c
.prog_data
.static_vertex_count
= !gp
->program
.Base
.nir
? -1 :
104 nir_gs_count_vertices(gp
->program
.Base
.nir
);
108 if (gp
->program
.OutputType
== GL_POINTS
) {
109 /* When the output type is points, the geometry shader may output data
110 * to multiple streams, and EndPrimitive() has no effect. So we
111 * configure the hardware to interpret the control data as stream ID.
113 c
.prog_data
.control_data_format
= GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
;
115 /* We only have to emit control bits if we are using streams */
116 if (prog
->Geom
.UsesStreams
)
117 c
.control_data_bits_per_vertex
= 2;
119 c
.control_data_bits_per_vertex
= 0;
121 /* When the output type is triangle_strip or line_strip, EndPrimitive()
122 * may be used to terminate the current strip and start a new one
123 * (similar to primitive restart), and outputting data to multiple
124 * streams is not supported. So we configure the hardware to interpret
125 * the control data as EndPrimitive information (a.k.a. "cut bits").
127 c
.prog_data
.control_data_format
= GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
;
129 /* We only need to output control data if the shader actually calls
132 c
.control_data_bits_per_vertex
= gp
->program
.UsesEndPrimitive
? 1 : 0;
135 /* There are no control data bits in gen6. */
136 c
.control_data_bits_per_vertex
= 0;
138 /* If it is using transform feedback, enable it */
139 if (prog
->TransformFeedback
.NumVarying
)
140 c
.prog_data
.gen6_xfb_enabled
= true;
142 c
.prog_data
.gen6_xfb_enabled
= false;
144 c
.control_data_header_size_bits
=
145 gp
->program
.VerticesOut
* c
.control_data_bits_per_vertex
;
147 /* 1 HWORD = 32 bytes = 256 bits */
148 c
.prog_data
.control_data_header_size_hwords
=
149 ALIGN(c
.control_data_header_size_bits
, 256) / 256;
151 GLbitfield64 outputs_written
= gp
->program
.Base
.OutputsWritten
;
153 brw_compute_vue_map(brw
->intelScreen
->devinfo
,
154 &c
.prog_data
.base
.vue_map
, outputs_written
,
155 prog
? prog
->SeparateShader
: false);
157 /* Compute the output vertex size.
159 * From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
162 * [0,62] indicating [1,63] 16B units
164 * Specifies the size of each vertex stored in the GS output entry
165 * (following any Control Header data) as a number of 128-bit units
168 * Programming Restrictions: The vertex size must be programmed as a
169 * multiple of 32B units with the following exception: Rendering is
170 * disabled (as per SOL stage state) and the vertex size output by the
173 * If rendering is enabled (as per SOL state) the vertex size must be
174 * programmed as a multiple of 32B units. In other words, the only time
175 * software can program a vertex size with an odd number of 16B units
176 * is when rendering is disabled.
178 * Note: B=bytes in the above text.
180 * It doesn't seem worth the extra trouble to optimize the case where the
181 * vertex size is 16B (especially since this would require special-casing
182 * the GEN assembly that writes to the URB). So we just set the vertex
183 * size to a multiple of 32B (2 vec4's) in all cases.
185 * The maximum output vertex size is 62*16 = 992 bytes (31 hwords). We
186 * budget that as follows:
188 * 512 bytes for varyings (a varying component is 4 bytes and
189 * gl_MaxGeometryOutputComponents = 128)
190 * 16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
192 * 16 bytes overhead for gl_Position (we allocate it a slot in the VUE
193 * even if it's not used)
194 * 32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
195 * whenever clip planes are enabled, even if the shader doesn't
196 * write to gl_ClipDistance)
197 * 16 bytes overhead since the VUE size must be a multiple of 32 bytes
198 * (see below)--this causes up to 1 VUE slot to be wasted
199 * 400 bytes available for varying packing overhead
201 * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
202 * per interpolation type, so this is plenty.
205 unsigned output_vertex_size_bytes
= c
.prog_data
.base
.vue_map
.num_slots
* 16;
206 assert(brw
->gen
== 6 ||
207 output_vertex_size_bytes
<= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES
);
208 c
.prog_data
.output_vertex_size_hwords
=
209 ALIGN(output_vertex_size_bytes
, 32) / 32;
211 /* Compute URB entry size. The maximum allowed URB entry size is 32k.
212 * That divides up as follows:
214 * 64 bytes for the control data header (cut indices or StreamID bits)
215 * 4096 bytes for varyings (a varying component is 4 bytes and
216 * gl_MaxGeometryTotalOutputComponents = 1024)
217 * 4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
218 * bytes/vertex and gl_MaxGeometryOutputVertices is 256)
219 * 4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
220 * even if it's not used)
221 * 8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
222 * whenever clip planes are enabled, even if the shader doesn't
223 * write to gl_ClipDistance)
224 * 4096 bytes overhead since the VUE size must be a multiple of 32
225 * bytes (see above)--this causes up to 1 VUE slot to be wasted
226 * 8128 bytes available for varying packing overhead
228 * Worst-case varying packing overhead is 3/4 of a varying slot per
229 * interpolation type, which works out to 3072 bytes, so this would allow
230 * us to accommodate 2 interpolation types without any danger of running
233 * In practice, the risk of running out of URB space is very small, since
234 * the above figures are all worst-case, and most of them scale with the
235 * number of output vertices. So we'll just calculate the amount of space
236 * we need, and if it's too large, fail to compile.
238 * The above is for gen7+ where we have a single URB entry that will hold
239 * all the output. In gen6, we will have to allocate URB entries for every
240 * vertex we emit, so our URB entries only need to be large enough to hold
241 * a single vertex. Also, gen6 does not have a control data header.
243 unsigned output_size_bytes
;
246 c
.prog_data
.output_vertex_size_hwords
* 32 * gp
->program
.VerticesOut
;
247 output_size_bytes
+= 32 * c
.prog_data
.control_data_header_size_hwords
;
249 output_size_bytes
= c
.prog_data
.output_vertex_size_hwords
* 32;
252 /* Broadwell stores "Vertex Count" as a full 8 DWord (32 byte) URB output,
253 * which comes before the control header.
256 output_size_bytes
+= 32;
258 assert(output_size_bytes
>= 1);
259 int max_output_size_bytes
= GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES
;
261 max_output_size_bytes
= GEN6_MAX_GS_URB_ENTRY_SIZE_BYTES
;
262 if (output_size_bytes
> max_output_size_bytes
)
266 /* URB entry sizes are stored as a multiple of 64 bytes in gen7+ and
267 * a multiple of 128 bytes in gen6.
270 c
.prog_data
.base
.urb_entry_size
= ALIGN(output_size_bytes
, 64) / 64;
272 c
.prog_data
.base
.urb_entry_size
= ALIGN(output_size_bytes
, 128) / 128;
274 c
.prog_data
.output_topology
=
275 get_hw_prim_for_gl_prim(gp
->program
.OutputType
);
277 /* The GLSL linker will have already matched up GS inputs and the outputs
278 * of prior stages. The driver does extend VS outputs in some cases, but
279 * only for legacy OpenGL or Gen4-5 hardware, neither of which offer
280 * geometry shader support. So we can safely ignore that.
282 * For SSO pipelines, we use a fixed VUE map layout based on variable
283 * locations, so we can rely on rendezvous-by-location making this work.
285 * However, we need to ignore VARYING_SLOT_PRIMITIVE_ID, as it's not
286 * written by previous stages and shows up via payload magic.
288 GLbitfield64 inputs_read
=
289 gp
->program
.Base
.InputsRead
& ~VARYING_BIT_PRIMITIVE_ID
;
290 brw_compute_vue_map(brw
->intelScreen
->devinfo
,
291 &c
.input_vue_map
, inputs_read
,
292 prog
->SeparateShader
);
294 /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
295 * need to program a URB read length of ceiling(num_slots / 2).
297 c
.prog_data
.base
.urb_read_length
= (c
.input_vue_map
.num_slots
+ 1) / 2;
299 void *mem_ctx
= ralloc_context(NULL
);
300 unsigned program_size
;
301 const unsigned *program
=
302 brw_gs_emit(brw
, prog
, &c
, mem_ctx
, &program_size
);
303 if (program
== NULL
) {
304 ralloc_free(mem_ctx
);
308 output
->mem_ctx
= mem_ctx
;
309 output
->program
= program
;
310 output
->program_size
= program_size
;
311 memcpy(&output
->prog_data
, &c
.prog_data
,
312 sizeof(output
->prog_data
));
318 brw_codegen_gs_prog(struct brw_context
*brw
,
319 struct gl_shader_program
*prog
,
320 struct brw_geometry_program
*gp
,
321 struct brw_gs_prog_key
*key
)
323 struct brw_gs_compile_output output
;
324 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
326 if (brw_compile_gs_prog(brw
, prog
, gp
, key
, &output
))
329 if (output
.prog_data
.base
.base
.total_scratch
) {
330 brw_get_scratch_bo(brw
, &stage_state
->scratch_bo
,
331 output
.prog_data
.base
.base
.total_scratch
*
332 brw
->max_gs_threads
);
335 brw_upload_cache(&brw
->cache
, BRW_CACHE_GS_PROG
,
337 output
.program
, output
.program_size
,
338 &output
.prog_data
, sizeof(output
.prog_data
),
339 &stage_state
->prog_offset
, &brw
->gs
.prog_data
);
340 ralloc_free(output
.mem_ctx
);
346 brw_gs_state_dirty(struct brw_context
*brw
)
348 return brw_state_dirty(brw
,
350 BRW_NEW_GEOMETRY_PROGRAM
|
351 BRW_NEW_TRANSFORM_FEEDBACK
);
355 brw_gs_populate_key(struct brw_context
*brw
,
356 struct brw_gs_prog_key
*key
)
358 struct gl_context
*ctx
= &brw
->ctx
;
359 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
360 struct brw_geometry_program
*gp
=
361 (struct brw_geometry_program
*) brw
->geometry_program
;
362 struct gl_program
*prog
= &gp
->program
.Base
;
364 memset(key
, 0, sizeof(*key
));
366 key
->program_string_id
= gp
->id
;
369 brw_populate_sampler_prog_key_data(ctx
, prog
, stage_state
->sampler_count
,
374 brw_upload_gs_prog(struct brw_context
*brw
)
376 struct gl_context
*ctx
= &brw
->ctx
;
377 struct gl_shader_program
**current
= ctx
->_Shader
->CurrentProgram
;
378 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
379 struct brw_gs_prog_key key
;
380 /* BRW_NEW_GEOMETRY_PROGRAM */
381 struct brw_geometry_program
*gp
=
382 (struct brw_geometry_program
*) brw
->geometry_program
;
384 if (!brw_gs_state_dirty(brw
))
388 /* No geometry shader. Vertex data just passes straight through. */
390 (brw
->ctx
.NewDriverState
& BRW_NEW_TRANSFORM_FEEDBACK
)) {
391 gen6_brw_upload_ff_gs_prog(brw
);
395 /* Other state atoms had better not try to access prog_data, since
396 * there's no GS program.
398 brw
->gs
.prog_data
= NULL
;
399 brw
->gs
.base
.prog_data
= NULL
;
404 brw_gs_populate_key(brw
, &key
);
406 if (!brw_search_cache(&brw
->cache
, BRW_CACHE_GS_PROG
,
408 &stage_state
->prog_offset
, &brw
->gs
.prog_data
)) {
409 bool success
= brw_codegen_gs_prog(brw
, current
[MESA_SHADER_GEOMETRY
],
414 brw
->gs
.base
.prog_data
= &brw
->gs
.prog_data
->base
.base
;
418 brw_gs_precompile(struct gl_context
*ctx
,
419 struct gl_shader_program
*shader_prog
,
420 struct gl_program
*prog
)
422 struct brw_context
*brw
= brw_context(ctx
);
423 struct brw_gs_prog_key key
;
424 uint32_t old_prog_offset
= brw
->gs
.base
.prog_offset
;
425 struct brw_gs_prog_data
*old_prog_data
= brw
->gs
.prog_data
;
428 struct gl_geometry_program
*gp
= (struct gl_geometry_program
*) prog
;
429 struct brw_geometry_program
*bgp
= brw_geometry_program(gp
);
431 memset(&key
, 0, sizeof(key
));
433 brw_setup_tex_for_precompile(brw
, &key
.tex
, prog
);
434 key
.program_string_id
= bgp
->id
;
436 success
= brw_codegen_gs_prog(brw
, shader_prog
, bgp
, &key
);
438 brw
->gs
.base
.prog_offset
= old_prog_offset
;
439 brw
->gs
.prog_data
= old_prog_data
;