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.
25 * \file brw_vec4_tcs.cpp
27 * Tessellaton control shader specific code derived from the vec4_visitor class.
31 #include "brw_vec4_tcs.h"
35 vec4_tcs_visitor::vec4_tcs_visitor(const struct brw_compiler
*compiler
,
37 const struct brw_tcs_prog_key
*key
,
38 struct brw_tcs_prog_data
*prog_data
,
39 const nir_shader
*nir
,
41 int shader_time_index
)
42 : vec4_visitor(compiler
, log_data
, &key
->tex
, &prog_data
->base
,
43 nir
, mem_ctx
, false, shader_time_index
),
50 vec4_tcs_visitor::nir_setup_system_value_intrinsic(nir_intrinsic_instr
*instr
)
55 vec4_tcs_visitor::make_reg_for_system_value(int location
, const glsl_type
*type
)
62 vec4_tcs_visitor::setup_payload()
66 /* The payload always contains important data in r0, which contains
67 * the URB handles that are passed on to the URB write at the end
72 /* r1.0 - r4.7 may contain the input control point URB handles,
73 * which we use to pull vertex data.
77 /* Push constants may start at r5.0 */
78 reg
= setup_uniforms(reg
);
80 this->first_non_payload_grf
= reg
;
85 vec4_tcs_visitor::emit_prolog()
87 invocation_id
= src_reg(this, glsl_type::uint_type
);
88 emit(TCS_OPCODE_GET_INSTANCE_ID
, dst_reg(invocation_id
));
90 /* HS threads are dispatched with the dispatch mask set to 0xFF.
91 * If there are an odd number of output vertices, then the final
92 * HS instance dispatched will only have its bottom half doing real
93 * work, and so we need to disable the upper half:
95 if (nir
->info
.tcs
.vertices_out
% 2) {
96 emit(CMP(dst_null_d(), invocation_id
,
97 brw_imm_ud(nir
->info
.tcs
.vertices_out
), BRW_CONDITIONAL_L
));
99 /* Matching ENDIF is in emit_thread_end() */
100 emit(IF(BRW_PREDICATE_NORMAL
));
106 vec4_tcs_visitor::emit_thread_end()
108 current_annotation
= "thread end";
110 if (nir
->info
.tcs
.vertices_out
% 2) {
111 emit(BRW_OPCODE_ENDIF
);
114 if (unlikely(INTEL_DEBUG
& DEBUG_SHADER_TIME
))
115 emit_shader_time_end();
117 vec4_instruction
*inst
= emit(VS_OPCODE_URB_WRITE
);
118 inst
->mlen
= 1; /* just the header, no data. */
119 inst
->urb_write_flags
= BRW_URB_WRITE_EOT_COMPLETE
;
124 vec4_tcs_visitor::emit_input_urb_read(const dst_reg
&dst
,
125 const src_reg
&vertex_index
,
126 unsigned base_offset
,
127 const src_reg
&indirect_offset
)
129 vec4_instruction
*inst
;
130 dst_reg
temp(this, glsl_type::ivec4_type
);
131 temp
.type
= dst
.type
;
133 /* Set up the message header to reference the proper parts of the URB */
134 dst_reg header
= dst_reg(this, glsl_type::uvec4_type
);
135 inst
= emit(TCS_OPCODE_SET_INPUT_URB_OFFSETS
, header
, vertex_index
,
137 inst
->force_writemask_all
= true;
139 /* Read into a temporary, ignoring writemasking. */
140 inst
= emit(VEC4_OPCODE_URB_READ
, temp
, src_reg(header
));
141 inst
->offset
= base_offset
;
145 /* Copy the temporary to the destination to deal with writemasking.
147 * Also attempt to deal with gl_PointSize being in the .w component.
149 if (inst
->offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
150 emit(MOV(dst
, swizzle(src_reg(temp
), BRW_SWIZZLE_WWWW
)));
152 emit(MOV(dst
, src_reg(temp
)));
157 vec4_tcs_visitor::emit_output_urb_read(const dst_reg
&dst
,
158 unsigned base_offset
,
159 const src_reg
&indirect_offset
)
161 vec4_instruction
*inst
;
163 /* Set up the message header to reference the proper parts of the URB */
164 dst_reg header
= dst_reg(this, glsl_type::uvec4_type
);
165 inst
= emit(TCS_OPCODE_SET_OUTPUT_URB_OFFSETS
, header
,
166 brw_imm_ud(dst
.writemask
), indirect_offset
);
167 inst
->force_writemask_all
= true;
169 /* Read into a temporary, ignoring writemasking. */
170 vec4_instruction
*read
= emit(VEC4_OPCODE_URB_READ
, dst
, src_reg(header
));
171 read
->offset
= base_offset
;
177 vec4_tcs_visitor::emit_urb_write(const src_reg
&value
,
179 unsigned base_offset
,
180 const src_reg
&indirect_offset
)
185 src_reg
message(this, glsl_type::uvec4_type
, 2);
186 vec4_instruction
*inst
;
188 inst
= emit(TCS_OPCODE_SET_OUTPUT_URB_OFFSETS
, dst_reg(message
),
189 brw_imm_ud(writemask
), indirect_offset
);
190 inst
->force_writemask_all
= true;
191 inst
= emit(MOV(offset(dst_reg(retype(message
, value
.type
)), 1), value
));
192 inst
->force_writemask_all
= true;
194 inst
= emit(TCS_OPCODE_URB_WRITE
, dst_null_f(), message
);
195 inst
->offset
= base_offset
;
201 tesslevel_outer_components(GLenum tes_primitive_mode
)
203 switch (tes_primitive_mode
) {
211 unreachable("Bogus tessellation domain");
217 tesslevel_inner_components(GLenum tes_primitive_mode
)
219 switch (tes_primitive_mode
) {
227 unreachable("Bogus tessellation domain");
233 * Given a normal .xyzw writemask, convert it to a writemask for a vector
234 * that's stored backwards, i.e. .wzyx.
237 writemask_for_backwards_vector(unsigned mask
)
239 unsigned new_mask
= 0;
241 for (int i
= 0; i
< 4; i
++)
242 new_mask
|= ((mask
>> i
) & 1) << (3 - i
);
248 vec4_tcs_visitor::nir_emit_intrinsic(nir_intrinsic_instr
*instr
)
250 switch (instr
->intrinsic
) {
251 case nir_intrinsic_load_invocation_id
:
252 emit(MOV(get_nir_dest(instr
->dest
, BRW_REGISTER_TYPE_UD
),
255 case nir_intrinsic_load_primitive_id
:
256 emit(TCS_OPCODE_GET_PRIMITIVE_ID
,
257 get_nir_dest(instr
->dest
, BRW_REGISTER_TYPE_UD
));
259 case nir_intrinsic_load_patch_vertices_in
:
260 emit(MOV(get_nir_dest(instr
->dest
, BRW_REGISTER_TYPE_D
),
261 brw_imm_d(key
->input_vertices
)));
263 case nir_intrinsic_load_per_vertex_input
: {
264 src_reg indirect_offset
= get_indirect_offset(instr
);
265 unsigned imm_offset
= instr
->const_index
[0];
267 nir_const_value
*vertex_const
= nir_src_as_const_value(instr
->src
[0]);
268 src_reg vertex_index
=
269 vertex_const
? src_reg(brw_imm_ud(vertex_const
->u
[0]))
270 : get_nir_src(instr
->src
[0], BRW_REGISTER_TYPE_UD
, 1);
272 dst_reg dst
= get_nir_dest(instr
->dest
, BRW_REGISTER_TYPE_D
);
273 dst
.writemask
= brw_writemask_for_size(instr
->num_components
);
275 emit_input_urb_read(dst
, vertex_index
, imm_offset
, indirect_offset
);
278 case nir_intrinsic_load_input
:
279 unreachable("nir_lower_io should use load_per_vertex_input intrinsics");
281 case nir_intrinsic_load_output
:
282 case nir_intrinsic_load_per_vertex_output
: {
283 src_reg indirect_offset
= get_indirect_offset(instr
);
284 unsigned imm_offset
= instr
->const_index
[0];;
286 dst_reg dst
= get_nir_dest(instr
->dest
, BRW_REGISTER_TYPE_D
);
287 dst
.writemask
= brw_writemask_for_size(instr
->num_components
);
289 if (imm_offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
290 dst
.type
= BRW_REGISTER_TYPE_F
;
292 /* This is a read of gl_TessLevelInner[], which lives in the
293 * Patch URB header. The layout depends on the domain.
295 switch (key
->tes_primitive_mode
) {
297 /* DWords 3-2 (reversed); use offset 0 and WZYX swizzle. */
298 dst_reg
tmp(this, glsl_type::vec4_type
);
299 emit_output_urb_read(tmp
, 0, src_reg());
300 emit(MOV(writemask(dst
, WRITEMASK_XY
),
301 swizzle(src_reg(tmp
), BRW_SWIZZLE_WZYX
)));
305 /* DWord 4; use offset 1 but normal swizzle/writemask. */
306 emit_output_urb_read(writemask(dst
, WRITEMASK_X
), 1, src_reg());
309 /* All channels are undefined. */
312 unreachable("Bogus tessellation domain");
314 } else if (imm_offset
== 1 && indirect_offset
.file
== BAD_FILE
) {
315 dst
.type
= BRW_REGISTER_TYPE_F
;
317 /* This is a read of gl_TessLevelOuter[], which lives in the
318 * high 4 DWords of the Patch URB header, in reverse order.
320 switch (key
->tes_primitive_mode
) {
322 dst
.writemask
= WRITEMASK_XYZW
;
325 dst
.writemask
= WRITEMASK_XYZ
;
328 dst
.writemask
= WRITEMASK_XY
;
331 unreachable("Bogus tessellation domain");
334 dst_reg
tmp(this, glsl_type::vec4_type
);
335 emit_output_urb_read(tmp
, 1, src_reg());
336 emit(MOV(dst
, swizzle(src_reg(tmp
), BRW_SWIZZLE_WZYX
)));
338 emit_output_urb_read(dst
, imm_offset
, indirect_offset
);
342 case nir_intrinsic_store_output
:
343 case nir_intrinsic_store_per_vertex_output
: {
344 src_reg value
= get_nir_src(instr
->src
[0]);
345 unsigned mask
= instr
->const_index
[1];
346 unsigned swiz
= BRW_SWIZZLE_XYZW
;
348 src_reg indirect_offset
= get_indirect_offset(instr
);
349 unsigned imm_offset
= instr
->const_index
[0];
351 if (imm_offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
352 value
.type
= BRW_REGISTER_TYPE_F
;
354 mask
&= (1 << tesslevel_inner_components(key
->tes_primitive_mode
)) - 1;
356 /* This is a write to gl_TessLevelInner[], which lives in the
357 * Patch URB header. The layout depends on the domain.
359 switch (key
->tes_primitive_mode
) {
361 /* gl_TessLevelInner[].xy lives at DWords 3-2 (reversed).
362 * We use an XXYX swizzle to reverse put .xy in the .wz
363 * channels, and use a .zw writemask.
365 swiz
= BRW_SWIZZLE4(0, 0, 1, 0);
366 mask
= writemask_for_backwards_vector(mask
);
369 /* gl_TessLevelInner[].x lives at DWord 4, so we set the
370 * writemask to X and bump the URB offset by 1.
375 /* Skip; gl_TessLevelInner[] doesn't exist for isolines. */
378 unreachable("Bogus tessellation domain");
380 } else if (imm_offset
== 1 && indirect_offset
.file
== BAD_FILE
) {
381 value
.type
= BRW_REGISTER_TYPE_F
;
383 mask
&= (1 << tesslevel_outer_components(key
->tes_primitive_mode
)) - 1;
385 /* This is a write to gl_TessLevelOuter[] which lives in the
386 * Patch URB Header at DWords 4-7. However, it's reversed, so
387 * instead of .xyzw we have .wzyx.
389 swiz
= BRW_SWIZZLE_WZYX
;
390 mask
= writemask_for_backwards_vector(mask
);
393 emit_urb_write(swizzle(value
, swiz
), mask
,
394 imm_offset
, indirect_offset
);
398 case nir_intrinsic_barrier
: {
399 dst_reg header
= dst_reg(this, glsl_type::uvec4_type
);
400 emit(TCS_OPCODE_CREATE_BARRIER_HEADER
, header
);
401 emit(SHADER_OPCODE_BARRIER
, dst_null_ud(), src_reg(header
));
406 vec4_visitor::nir_emit_intrinsic(instr
);
411 extern "C" const unsigned *
412 brw_compile_tcs(const struct brw_compiler
*compiler
,
415 const struct brw_tcs_prog_key
*key
,
416 struct brw_tcs_prog_data
*prog_data
,
417 const nir_shader
*src_shader
,
418 int shader_time_index
,
419 unsigned *final_assembly_size
,
422 const struct brw_device_info
*devinfo
= compiler
->devinfo
;
423 struct brw_vue_prog_data
*vue_prog_data
= &prog_data
->base
;
424 const bool is_scalar
= compiler
->scalar_stage
[MESA_SHADER_TESS_CTRL
];
426 nir_shader
*nir
= nir_shader_clone(mem_ctx
, src_shader
);
427 nir
= brw_nir_apply_sampler_key(nir
, devinfo
, &key
->tex
, is_scalar
);
428 nir
= brw_postprocess_nir(nir
, compiler
->devinfo
, is_scalar
);
430 prog_data
->instances
= DIV_ROUND_UP(nir
->info
.tcs
.vertices_out
, 2);
432 brw_compute_tess_vue_map(&vue_prog_data
->vue_map
,
433 nir
->info
.outputs_written
,
434 nir
->info
.patch_outputs_written
);
436 /* Compute URB entry size. The maximum allowed URB entry size is 32k.
437 * That divides up as follows:
439 * 32 bytes for the patch header (tessellation factors)
440 * 480 bytes for per-patch varyings (a varying component is 4 bytes and
441 * gl_MaxTessPatchComponents = 120)
442 * 16384 bytes for per-vertex varyings (a varying component is 4 bytes,
443 * gl_MaxPatchVertices = 32 and
444 * gl_MaxTessControlOutputComponents = 128)
446 * 15808 bytes left for varying packing overhead
448 const int num_per_patch_slots
= vue_prog_data
->vue_map
.num_per_patch_slots
;
449 const int num_per_vertex_slots
= vue_prog_data
->vue_map
.num_per_vertex_slots
;
450 unsigned output_size_bytes
= 0;
451 /* Note that the patch header is counted in num_per_patch_slots. */
452 output_size_bytes
+= num_per_patch_slots
* 16;
453 output_size_bytes
+= nir
->info
.tcs
.vertices_out
* num_per_vertex_slots
* 16;
455 assert(output_size_bytes
>= 1);
456 if (output_size_bytes
> GEN7_MAX_HS_URB_ENTRY_SIZE_BYTES
)
459 /* URB entry sizes are stored as a multiple of 64 bytes. */
460 vue_prog_data
->urb_entry_size
= ALIGN(output_size_bytes
, 64) / 64;
462 struct brw_vue_map input_vue_map
;
463 brw_compute_vue_map(devinfo
, &input_vue_map
,
464 nir
->info
.inputs_read
& ~VARYING_BIT_PRIMITIVE_ID
,
467 /* HS does not use the usual payload pushing from URB to GRFs,
468 * because we don't have enough registers for a full-size payload, and
469 * the hardware is broken on Haswell anyway.
471 vue_prog_data
->urb_read_length
= 0;
473 if (unlikely(INTEL_DEBUG
& DEBUG_TCS
)) {
474 fprintf(stderr
, "TCS Input ");
475 brw_print_vue_map(stderr
, &input_vue_map
);
476 fprintf(stderr
, "TCS Output ");
477 brw_print_vue_map(stderr
, &vue_prog_data
->vue_map
);
480 vec4_tcs_visitor
v(compiler
, log_data
, key
, prog_data
,
481 nir
, mem_ctx
, shader_time_index
);
484 *error_str
= ralloc_strdup(mem_ctx
, v
.fail_msg
);
488 if (unlikely(INTEL_DEBUG
& DEBUG_TCS
))
489 v
.dump_instructions();
491 return brw_vec4_generate_assembly(compiler
, log_data
, mem_ctx
, nir
,
492 &prog_data
->base
, v
.cfg
,
493 final_assembly_size
);
497 } /* namespace brw */