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 gl_shader
*shader
= prog
->_LinkedShaders
[MESA_SHADER_GEOMETRY
];
61 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
62 struct brw_gs_compile c
;
63 memset(&c
, 0, sizeof(c
));
67 c
.prog_data
.include_primitive_id
=
68 (gp
->program
.Base
.InputsRead
& VARYING_BIT_PRIMITIVE_ID
) != 0;
70 c
.prog_data
.invocations
= gp
->program
.Invocations
;
72 assign_gs_binding_table_offsets(brw
->intelScreen
->devinfo
, prog
,
73 &gp
->program
.Base
, &c
.prog_data
);
75 /* Allocate the references to the uniforms that will end up in the
76 * prog_data associated with the compiled program, and which will be freed
79 * Note: param_count needs to be num_uniform_components * 4, since we add
80 * padding around uniform values below vec4 size, so the worst case is that
81 * every uniform is a float which gets padded to the size of a vec4.
83 struct gl_shader
*gs
= prog
->_LinkedShaders
[MESA_SHADER_GEOMETRY
];
84 int param_count
= gp
->program
.Base
.nir
->num_uniforms
* 4;
86 c
.prog_data
.base
.base
.param
=
87 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
88 c
.prog_data
.base
.base
.pull_param
=
89 rzalloc_array(NULL
, const gl_constant_value
*, param_count
);
90 c
.prog_data
.base
.base
.image_param
=
91 rzalloc_array(NULL
, struct brw_image_param
, gs
->NumImages
);
92 c
.prog_data
.base
.base
.nr_params
= param_count
;
93 c
.prog_data
.base
.base
.nr_image_params
= gs
->NumImages
;
95 brw_nir_setup_glsl_uniforms(gp
->program
.Base
.nir
, prog
, &gp
->program
.Base
,
96 &c
.prog_data
.base
.base
, false);
99 c
.prog_data
.static_vertex_count
= !gp
->program
.Base
.nir
? -1 :
100 nir_gs_count_vertices(gp
->program
.Base
.nir
);
104 if (gp
->program
.OutputType
== GL_POINTS
) {
105 /* When the output type is points, the geometry shader may output data
106 * to multiple streams, and EndPrimitive() has no effect. So we
107 * configure the hardware to interpret the control data as stream ID.
109 c
.prog_data
.control_data_format
= GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
;
111 /* We only have to emit control bits if we are using streams */
112 if (prog
->Geom
.UsesStreams
)
113 c
.control_data_bits_per_vertex
= 2;
115 c
.control_data_bits_per_vertex
= 0;
117 /* When the output type is triangle_strip or line_strip, EndPrimitive()
118 * may be used to terminate the current strip and start a new one
119 * (similar to primitive restart), and outputting data to multiple
120 * streams is not supported. So we configure the hardware to interpret
121 * the control data as EndPrimitive information (a.k.a. "cut bits").
123 c
.prog_data
.control_data_format
= GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
;
125 /* We only need to output control data if the shader actually calls
128 c
.control_data_bits_per_vertex
= gp
->program
.UsesEndPrimitive
? 1 : 0;
131 /* There are no control data bits in gen6. */
132 c
.control_data_bits_per_vertex
= 0;
134 /* If it is using transform feedback, enable it */
135 if (prog
->TransformFeedback
.NumVarying
)
136 c
.prog_data
.gen6_xfb_enabled
= true;
138 c
.prog_data
.gen6_xfb_enabled
= false;
140 c
.control_data_header_size_bits
=
141 gp
->program
.VerticesOut
* c
.control_data_bits_per_vertex
;
143 /* 1 HWORD = 32 bytes = 256 bits */
144 c
.prog_data
.control_data_header_size_hwords
=
145 ALIGN(c
.control_data_header_size_bits
, 256) / 256;
147 GLbitfield64 outputs_written
= gp
->program
.Base
.OutputsWritten
;
149 brw_compute_vue_map(brw
->intelScreen
->devinfo
,
150 &c
.prog_data
.base
.vue_map
, outputs_written
,
151 prog
? prog
->SeparateShader
: false);
153 /* Compute the output vertex size.
155 * From the Ivy Bridge PRM, Vol2 Part1 7.2.1.1 STATE_GS - Output Vertex
158 * [0,62] indicating [1,63] 16B units
160 * Specifies the size of each vertex stored in the GS output entry
161 * (following any Control Header data) as a number of 128-bit units
164 * Programming Restrictions: The vertex size must be programmed as a
165 * multiple of 32B units with the following exception: Rendering is
166 * disabled (as per SOL stage state) and the vertex size output by the
169 * If rendering is enabled (as per SOL state) the vertex size must be
170 * programmed as a multiple of 32B units. In other words, the only time
171 * software can program a vertex size with an odd number of 16B units
172 * is when rendering is disabled.
174 * Note: B=bytes in the above text.
176 * It doesn't seem worth the extra trouble to optimize the case where the
177 * vertex size is 16B (especially since this would require special-casing
178 * the GEN assembly that writes to the URB). So we just set the vertex
179 * size to a multiple of 32B (2 vec4's) in all cases.
181 * The maximum output vertex size is 62*16 = 992 bytes (31 hwords). We
182 * budget that as follows:
184 * 512 bytes for varyings (a varying component is 4 bytes and
185 * gl_MaxGeometryOutputComponents = 128)
186 * 16 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
188 * 16 bytes overhead for gl_Position (we allocate it a slot in the VUE
189 * even if it's not used)
190 * 32 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
191 * whenever clip planes are enabled, even if the shader doesn't
192 * write to gl_ClipDistance)
193 * 16 bytes overhead since the VUE size must be a multiple of 32 bytes
194 * (see below)--this causes up to 1 VUE slot to be wasted
195 * 400 bytes available for varying packing overhead
197 * Worst-case varying packing overhead is 3/4 of a varying slot (12 bytes)
198 * per interpolation type, so this is plenty.
201 unsigned output_vertex_size_bytes
= c
.prog_data
.base
.vue_map
.num_slots
* 16;
202 assert(brw
->gen
== 6 ||
203 output_vertex_size_bytes
<= GEN7_MAX_GS_OUTPUT_VERTEX_SIZE_BYTES
);
204 c
.prog_data
.output_vertex_size_hwords
=
205 ALIGN(output_vertex_size_bytes
, 32) / 32;
207 /* Compute URB entry size. The maximum allowed URB entry size is 32k.
208 * That divides up as follows:
210 * 64 bytes for the control data header (cut indices or StreamID bits)
211 * 4096 bytes for varyings (a varying component is 4 bytes and
212 * gl_MaxGeometryTotalOutputComponents = 1024)
213 * 4096 bytes overhead for VARYING_SLOT_PSIZ (each varying slot is 16
214 * bytes/vertex and gl_MaxGeometryOutputVertices is 256)
215 * 4096 bytes overhead for gl_Position (we allocate it a slot in the VUE
216 * even if it's not used)
217 * 8192 bytes overhead for gl_ClipDistance (we allocate it 2 VUE slots
218 * whenever clip planes are enabled, even if the shader doesn't
219 * write to gl_ClipDistance)
220 * 4096 bytes overhead since the VUE size must be a multiple of 32
221 * bytes (see above)--this causes up to 1 VUE slot to be wasted
222 * 8128 bytes available for varying packing overhead
224 * Worst-case varying packing overhead is 3/4 of a varying slot per
225 * interpolation type, which works out to 3072 bytes, so this would allow
226 * us to accommodate 2 interpolation types without any danger of running
229 * In practice, the risk of running out of URB space is very small, since
230 * the above figures are all worst-case, and most of them scale with the
231 * number of output vertices. So we'll just calculate the amount of space
232 * we need, and if it's too large, fail to compile.
234 * The above is for gen7+ where we have a single URB entry that will hold
235 * all the output. In gen6, we will have to allocate URB entries for every
236 * vertex we emit, so our URB entries only need to be large enough to hold
237 * a single vertex. Also, gen6 does not have a control data header.
239 unsigned output_size_bytes
;
242 c
.prog_data
.output_vertex_size_hwords
* 32 * gp
->program
.VerticesOut
;
243 output_size_bytes
+= 32 * c
.prog_data
.control_data_header_size_hwords
;
245 output_size_bytes
= c
.prog_data
.output_vertex_size_hwords
* 32;
248 /* Broadwell stores "Vertex Count" as a full 8 DWord (32 byte) URB output,
249 * which comes before the control header.
252 output_size_bytes
+= 32;
254 assert(output_size_bytes
>= 1);
255 int max_output_size_bytes
= GEN7_MAX_GS_URB_ENTRY_SIZE_BYTES
;
257 max_output_size_bytes
= GEN6_MAX_GS_URB_ENTRY_SIZE_BYTES
;
258 if (output_size_bytes
> max_output_size_bytes
)
262 /* URB entry sizes are stored as a multiple of 64 bytes in gen7+ and
263 * a multiple of 128 bytes in gen6.
266 c
.prog_data
.base
.urb_entry_size
= ALIGN(output_size_bytes
, 64) / 64;
268 c
.prog_data
.base
.urb_entry_size
= ALIGN(output_size_bytes
, 128) / 128;
270 c
.prog_data
.output_topology
=
271 get_hw_prim_for_gl_prim(gp
->program
.OutputType
);
273 /* The GLSL linker will have already matched up GS inputs and the outputs
274 * of prior stages. The driver does extend VS outputs in some cases, but
275 * only for legacy OpenGL or Gen4-5 hardware, neither of which offer
276 * geometry shader support. So we can safely ignore that.
278 * For SSO pipelines, we use a fixed VUE map layout based on variable
279 * locations, so we can rely on rendezvous-by-location making this work.
281 * However, we need to ignore VARYING_SLOT_PRIMITIVE_ID, as it's not
282 * written by previous stages and shows up via payload magic.
284 GLbitfield64 inputs_read
=
285 gp
->program
.Base
.InputsRead
& ~VARYING_BIT_PRIMITIVE_ID
;
286 brw_compute_vue_map(brw
->intelScreen
->devinfo
,
287 &c
.input_vue_map
, inputs_read
,
288 prog
->SeparateShader
);
290 /* GS inputs are read from the VUE 256 bits (2 vec4's) at a time, so we
291 * need to program a URB read length of ceiling(num_slots / 2).
293 c
.prog_data
.base
.urb_read_length
= (c
.input_vue_map
.num_slots
+ 1) / 2;
295 if (unlikely(INTEL_DEBUG
& DEBUG_GS
))
296 brw_dump_ir("geometry", prog
, gs
, NULL
);
299 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
)
300 st_index
= brw_get_shader_time_index(brw
, prog
, NULL
, ST_GS
);
302 void *mem_ctx
= ralloc_context(NULL
);
303 unsigned program_size
;
305 const unsigned *program
=
306 brw_gs_emit(brw
->intelScreen
->compiler
, brw
, &c
,
307 shader
->Program
->nir
, prog
,
308 mem_ctx
, st_index
, &program_size
, &error_str
);
309 if (program
== NULL
) {
310 ralloc_free(mem_ctx
);
314 /* Scratch space is used for register spilling */
315 if (c
.prog_data
.base
.base
.total_scratch
) {
316 brw_get_scratch_bo(brw
, &stage_state
->scratch_bo
,
317 c
.prog_data
.base
.base
.total_scratch
*
318 brw
->max_gs_threads
);
321 brw_upload_cache(&brw
->cache
, BRW_CACHE_GS_PROG
,
322 &c
.key
, sizeof(c
.key
),
323 program
, program_size
,
324 &c
.prog_data
, sizeof(c
.prog_data
),
325 &stage_state
->prog_offset
, &brw
->gs
.prog_data
);
326 ralloc_free(mem_ctx
);
332 brw_gs_state_dirty(struct brw_context
*brw
)
334 return brw_state_dirty(brw
,
336 BRW_NEW_GEOMETRY_PROGRAM
|
337 BRW_NEW_TRANSFORM_FEEDBACK
);
341 brw_gs_populate_key(struct brw_context
*brw
,
342 struct brw_gs_prog_key
*key
)
344 struct gl_context
*ctx
= &brw
->ctx
;
345 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
346 struct brw_geometry_program
*gp
=
347 (struct brw_geometry_program
*) brw
->geometry_program
;
348 struct gl_program
*prog
= &gp
->program
.Base
;
350 memset(key
, 0, sizeof(*key
));
352 key
->program_string_id
= gp
->id
;
355 brw_populate_sampler_prog_key_data(ctx
, prog
, stage_state
->sampler_count
,
360 brw_upload_gs_prog(struct brw_context
*brw
)
362 struct gl_context
*ctx
= &brw
->ctx
;
363 struct gl_shader_program
**current
= ctx
->_Shader
->CurrentProgram
;
364 struct brw_stage_state
*stage_state
= &brw
->gs
.base
;
365 struct brw_gs_prog_key key
;
366 /* BRW_NEW_GEOMETRY_PROGRAM */
367 struct brw_geometry_program
*gp
=
368 (struct brw_geometry_program
*) brw
->geometry_program
;
370 if (!brw_gs_state_dirty(brw
))
374 /* No geometry shader. Vertex data just passes straight through. */
376 (brw
->ctx
.NewDriverState
& BRW_NEW_TRANSFORM_FEEDBACK
)) {
377 gen6_brw_upload_ff_gs_prog(brw
);
381 /* Other state atoms had better not try to access prog_data, since
382 * there's no GS program.
384 brw
->gs
.prog_data
= NULL
;
385 brw
->gs
.base
.prog_data
= NULL
;
390 brw_gs_populate_key(brw
, &key
);
392 if (!brw_search_cache(&brw
->cache
, BRW_CACHE_GS_PROG
,
394 &stage_state
->prog_offset
, &brw
->gs
.prog_data
)) {
395 bool success
= brw_codegen_gs_prog(brw
, current
[MESA_SHADER_GEOMETRY
],
400 brw
->gs
.base
.prog_data
= &brw
->gs
.prog_data
->base
.base
;
404 brw_gs_precompile(struct gl_context
*ctx
,
405 struct gl_shader_program
*shader_prog
,
406 struct gl_program
*prog
)
408 struct brw_context
*brw
= brw_context(ctx
);
409 struct brw_gs_prog_key key
;
410 uint32_t old_prog_offset
= brw
->gs
.base
.prog_offset
;
411 struct brw_gs_prog_data
*old_prog_data
= brw
->gs
.prog_data
;
414 struct gl_geometry_program
*gp
= (struct gl_geometry_program
*) prog
;
415 struct brw_geometry_program
*bgp
= brw_geometry_program(gp
);
417 memset(&key
, 0, sizeof(key
));
419 brw_setup_tex_for_precompile(brw
, &key
.tex
, prog
);
420 key
.program_string_id
= bgp
->id
;
422 success
= brw_codegen_gs_prog(brw
, shader_prog
, bgp
, &key
);
424 brw
->gs
.base
.prog_offset
= old_prog_offset
;
425 brw
->gs
.prog_data
= old_prog_data
;