2 * Copyright © 2010 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 DEALINGS
25 #include "main/macros.h"
26 #include "brw_context.h"
29 #include "brw_vec4_gs.h"
31 #include "glsl/ir_optimization.h"
32 #include "glsl/glsl_parser_extras.h"
33 #include "main/shaderapi.h"
36 brw_new_shader(struct gl_context
*ctx
, GLuint name
, GLuint type
)
38 struct brw_shader
*shader
;
40 shader
= rzalloc(NULL
, struct brw_shader
);
42 shader
->base
.Type
= type
;
43 shader
->base
.Stage
= _mesa_shader_enum_to_shader_stage(type
);
44 shader
->base
.Name
= name
;
45 _mesa_init_shader(ctx
, &shader
->base
);
51 struct gl_shader_program
*
52 brw_new_shader_program(struct gl_context
*ctx
, GLuint name
)
54 struct gl_shader_program
*prog
= rzalloc(NULL
, struct gl_shader_program
);
57 _mesa_init_shader_program(ctx
, prog
);
63 * Performs a compile of the shader stages even when we don't know
64 * what non-orthogonal state will be set, in the hope that it reflects
65 * the eventual NOS used, and thus allows us to produce link failures.
68 brw_shader_precompile(struct gl_context
*ctx
, struct gl_shader_program
*prog
)
70 struct brw_context
*brw
= brw_context(ctx
);
72 if (brw
->precompile
&& !brw_fs_precompile(ctx
, prog
))
75 if (brw
->precompile
&& !brw_gs_precompile(ctx
, prog
))
78 if (brw
->precompile
&& !brw_vs_precompile(ctx
, prog
))
85 brw_lower_packing_builtins(struct brw_context
*brw
,
86 gl_shader_stage shader_type
,
89 int ops
= LOWER_PACK_SNORM_2x16
90 | LOWER_UNPACK_SNORM_2x16
91 | LOWER_PACK_UNORM_2x16
92 | LOWER_UNPACK_UNORM_2x16
93 | LOWER_PACK_SNORM_4x8
94 | LOWER_UNPACK_SNORM_4x8
95 | LOWER_PACK_UNORM_4x8
96 | LOWER_UNPACK_UNORM_4x8
;
99 /* Gen7 introduced the f32to16 and f16to32 instructions, which can be
100 * used to execute packHalf2x16 and unpackHalf2x16. For AOS code, no
101 * lowering is needed. For SOA code, the Half2x16 ops must be
104 if (shader_type
== MESA_SHADER_FRAGMENT
) {
105 ops
|= LOWER_PACK_HALF_2x16_TO_SPLIT
106 | LOWER_UNPACK_HALF_2x16_TO_SPLIT
;
109 ops
|= LOWER_PACK_HALF_2x16
110 | LOWER_UNPACK_HALF_2x16
;
113 lower_packing_builtins(ir
, ops
);
117 brw_link_shader(struct gl_context
*ctx
, struct gl_shader_program
*shProg
)
119 struct brw_context
*brw
= brw_context(ctx
);
122 for (stage
= 0; stage
< ARRAY_SIZE(shProg
->_LinkedShaders
); stage
++) {
123 struct brw_shader
*shader
=
124 (struct brw_shader
*)shProg
->_LinkedShaders
[stage
];
129 struct gl_program
*prog
=
130 ctx
->Driver
.NewProgram(ctx
, _mesa_shader_stage_to_program(stage
),
134 prog
->Parameters
= _mesa_new_parameter_list();
136 _mesa_copy_linked_program_data((gl_shader_stage
) stage
, shProg
, prog
);
140 /* lower_packing_builtins() inserts arithmetic instructions, so it
141 * must precede lower_instructions().
143 brw_lower_packing_builtins(brw
, (gl_shader_stage
) stage
, shader
->base
.ir
);
144 do_mat_op_to_vec(shader
->base
.ir
);
145 const int bitfield_insert
= brw
->gen
>= 7
146 ? BITFIELD_INSERT_TO_BFM_BFI
148 const int lrp_to_arith
= brw
->gen
< 6 ? LRP_TO_ARITH
: 0;
149 lower_instructions(shader
->base
.ir
,
159 /* Pre-gen6 HW can only nest if-statements 16 deep. Beyond this,
160 * if-statements need to be flattened.
163 lower_if_to_cond_assign(shader
->base
.ir
, 16);
165 do_lower_texture_projection(shader
->base
.ir
);
166 brw_lower_texture_gradients(brw
, shader
->base
.ir
);
167 do_vec_index_to_cond_assign(shader
->base
.ir
);
168 lower_vector_insert(shader
->base
.ir
, true);
169 brw_do_cubemap_normalize(shader
->base
.ir
);
170 brw_do_lower_offset_arrays(shader
->base
.ir
);
171 brw_do_lower_unnormalized_offset(shader
->base
.ir
);
172 lower_noise(shader
->base
.ir
);
173 lower_quadop_vector(shader
->base
.ir
, false);
176 bool output
= stage
== MESA_SHADER_FRAGMENT
;
177 bool temp
= stage
== MESA_SHADER_FRAGMENT
;
178 bool uniform
= false;
180 bool lowered_variable_indexing
=
181 lower_variable_index_to_cond_assign(shader
->base
.ir
,
182 input
, output
, temp
, uniform
);
184 if (unlikely(brw
->perf_debug
&& lowered_variable_indexing
)) {
185 perf_debug("Unsupported form of variable indexing in FS; falling "
186 "back to very inefficient code generation\n");
189 /* FINISHME: Do this before the variable index lowering. */
190 lower_ubo_reference(&shader
->base
, shader
->base
.ir
);
195 if (stage
== MESA_SHADER_FRAGMENT
) {
196 brw_do_channel_expressions(shader
->base
.ir
);
197 brw_do_vector_splitting(shader
->base
.ir
);
200 progress
= do_lower_jumps(shader
->base
.ir
, true, true,
201 true, /* main return */
202 false, /* continue */
206 progress
= do_common_optimization(shader
->base
.ir
, true, true, 32,
207 &ctx
->ShaderCompilerOptions
[stage
])
211 /* Make a pass over the IR to add state references for any built-in
212 * uniforms that are used. This has to be done now (during linking).
213 * Code generation doesn't happen until the first time this shader is
214 * used for rendering. Waiting until then to generate the parameters is
215 * too late. At that point, the values for the built-in uniforms won't
216 * get sent to the shader.
218 foreach_list(node
, shader
->base
.ir
) {
219 ir_variable
*var
= ((ir_instruction
*) node
)->as_variable();
221 if ((var
== NULL
) || (var
->data
.mode
!= ir_var_uniform
)
222 || (strncmp(var
->name
, "gl_", 3) != 0))
225 const ir_state_slot
*const slots
= var
->state_slots
;
226 assert(var
->state_slots
!= NULL
);
228 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
229 _mesa_add_state_reference(prog
->Parameters
,
230 (gl_state_index
*) slots
[i
].tokens
);
234 validate_ir_tree(shader
->base
.ir
);
236 do_set_program_inouts(shader
->base
.ir
, prog
, shader
->base
.Stage
);
238 prog
->SamplersUsed
= shader
->base
.active_samplers
;
239 _mesa_update_shader_textures_used(shProg
, prog
);
241 _mesa_reference_program(ctx
, &shader
->base
.Program
, prog
);
243 brw_add_texrect_params(prog
);
245 /* This has to be done last. Any operation that can cause
246 * prog->ParameterValues to get reallocated (e.g., anything that adds a
247 * program constant) has to happen before creating this linkage.
249 _mesa_associate_uniform_storage(ctx
, shProg
, prog
->Parameters
);
251 _mesa_reference_program(ctx
, &prog
, NULL
);
253 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
255 printf("GLSL IR for linked %s program %d:\n",
256 _mesa_shader_stage_to_string(shader
->base
.Stage
),
258 _mesa_print_ir(shader
->base
.ir
, NULL
);
263 if (ctx
->Shader
.Flags
& GLSL_DUMP
) {
264 for (unsigned i
= 0; i
< shProg
->NumShaders
; i
++) {
265 const struct gl_shader
*sh
= shProg
->Shaders
[i
];
269 printf("GLSL %s shader %d source for linked program %d:\n",
270 _mesa_shader_stage_to_string(sh
->Stage
),
273 printf("%s", sh
->Source
);
278 if (!brw_shader_precompile(ctx
, shProg
))
286 brw_type_for_base_type(const struct glsl_type
*type
)
288 switch (type
->base_type
) {
289 case GLSL_TYPE_FLOAT
:
290 return BRW_REGISTER_TYPE_F
;
293 return BRW_REGISTER_TYPE_D
;
295 return BRW_REGISTER_TYPE_UD
;
296 case GLSL_TYPE_ARRAY
:
297 return brw_type_for_base_type(type
->fields
.array
);
298 case GLSL_TYPE_STRUCT
:
299 case GLSL_TYPE_SAMPLER
:
300 case GLSL_TYPE_ATOMIC_UINT
:
301 /* These should be overridden with the type of the member when
302 * dereferenced into. BRW_REGISTER_TYPE_UD seems like a likely
303 * way to trip up if we don't.
305 return BRW_REGISTER_TYPE_UD
;
306 case GLSL_TYPE_IMAGE
:
307 return BRW_REGISTER_TYPE_UD
;
309 case GLSL_TYPE_ERROR
:
310 case GLSL_TYPE_INTERFACE
:
311 assert(!"not reached");
315 return BRW_REGISTER_TYPE_F
;
319 brw_conditional_for_comparison(unsigned int op
)
323 return BRW_CONDITIONAL_L
;
324 case ir_binop_greater
:
325 return BRW_CONDITIONAL_G
;
326 case ir_binop_lequal
:
327 return BRW_CONDITIONAL_LE
;
328 case ir_binop_gequal
:
329 return BRW_CONDITIONAL_GE
;
331 case ir_binop_all_equal
: /* same as equal for scalars */
332 return BRW_CONDITIONAL_Z
;
333 case ir_binop_nequal
:
334 case ir_binop_any_nequal
: /* same as nequal for scalars */
335 return BRW_CONDITIONAL_NZ
;
337 assert(!"not reached: bad operation for comparison");
338 return BRW_CONDITIONAL_NZ
;
343 brw_math_function(enum opcode op
)
346 case SHADER_OPCODE_RCP
:
347 return BRW_MATH_FUNCTION_INV
;
348 case SHADER_OPCODE_RSQ
:
349 return BRW_MATH_FUNCTION_RSQ
;
350 case SHADER_OPCODE_SQRT
:
351 return BRW_MATH_FUNCTION_SQRT
;
352 case SHADER_OPCODE_EXP2
:
353 return BRW_MATH_FUNCTION_EXP
;
354 case SHADER_OPCODE_LOG2
:
355 return BRW_MATH_FUNCTION_LOG
;
356 case SHADER_OPCODE_POW
:
357 return BRW_MATH_FUNCTION_POW
;
358 case SHADER_OPCODE_SIN
:
359 return BRW_MATH_FUNCTION_SIN
;
360 case SHADER_OPCODE_COS
:
361 return BRW_MATH_FUNCTION_COS
;
362 case SHADER_OPCODE_INT_QUOTIENT
:
363 return BRW_MATH_FUNCTION_INT_DIV_QUOTIENT
;
364 case SHADER_OPCODE_INT_REMAINDER
:
365 return BRW_MATH_FUNCTION_INT_DIV_REMAINDER
;
367 assert(!"not reached: unknown math function");
373 brw_texture_offset(struct gl_context
*ctx
, ir_constant
*offset
)
375 /* If the driver does not support GL_ARB_gpu_shader5, the offset
378 assert(offset
!= NULL
|| ctx
->Extensions
.ARB_gpu_shader5
);
380 if (!offset
) return 0; /* nonconstant offset; caller will handle it. */
382 signed char offsets
[3];
383 for (unsigned i
= 0; i
< offset
->type
->vector_elements
; i
++)
384 offsets
[i
] = (signed char) offset
->value
.i
[i
];
386 /* Combine all three offsets into a single unsigned dword:
388 * bits 11:8 - U Offset (X component)
389 * bits 7:4 - V Offset (Y component)
390 * bits 3:0 - R Offset (Z component)
392 unsigned offset_bits
= 0;
393 for (unsigned i
= 0; i
< offset
->type
->vector_elements
; i
++) {
394 const unsigned shift
= 4 * (2 - i
);
395 offset_bits
|= (offsets
[i
] << shift
) & (0xF << shift
);
401 brw_instruction_name(enum opcode op
)
405 if (op
< ARRAY_SIZE(opcode_descs
) && opcode_descs
[op
].name
)
406 return opcode_descs
[op
].name
;
409 case FS_OPCODE_FB_WRITE
:
411 case FS_OPCODE_BLORP_FB_WRITE
:
412 return "blorp_fb_write";
414 case SHADER_OPCODE_RCP
:
416 case SHADER_OPCODE_RSQ
:
418 case SHADER_OPCODE_SQRT
:
420 case SHADER_OPCODE_EXP2
:
422 case SHADER_OPCODE_LOG2
:
424 case SHADER_OPCODE_POW
:
426 case SHADER_OPCODE_INT_QUOTIENT
:
428 case SHADER_OPCODE_INT_REMAINDER
:
430 case SHADER_OPCODE_SIN
:
432 case SHADER_OPCODE_COS
:
435 case SHADER_OPCODE_TEX
:
437 case SHADER_OPCODE_TXD
:
439 case SHADER_OPCODE_TXF
:
441 case SHADER_OPCODE_TXL
:
443 case SHADER_OPCODE_TXS
:
447 case SHADER_OPCODE_TXF_CMS
:
449 case SHADER_OPCODE_TXF_UMS
:
451 case SHADER_OPCODE_TXF_MCS
:
453 case SHADER_OPCODE_TG4
:
455 case SHADER_OPCODE_TG4_OFFSET
:
458 case SHADER_OPCODE_GEN4_SCRATCH_READ
:
459 return "gen4_scratch_read";
460 case SHADER_OPCODE_GEN4_SCRATCH_WRITE
:
461 return "gen4_scratch_write";
462 case SHADER_OPCODE_GEN7_SCRATCH_READ
:
463 return "gen7_scratch_read";
470 case FS_OPCODE_PIXEL_X
:
472 case FS_OPCODE_PIXEL_Y
:
475 case FS_OPCODE_CINTERP
:
477 case FS_OPCODE_LINTERP
:
480 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
:
481 return "uniform_pull_const";
482 case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7
:
483 return "uniform_pull_const_gen7";
484 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD
:
485 return "varying_pull_const";
486 case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7
:
487 return "varying_pull_const_gen7";
489 case FS_OPCODE_MOV_DISPATCH_TO_FLAGS
:
490 return "mov_dispatch_to_flags";
491 case FS_OPCODE_DISCARD_JUMP
:
492 return "discard_jump";
494 case FS_OPCODE_SET_SIMD4X2_OFFSET
:
495 return "set_simd4x2_offset";
497 case FS_OPCODE_PACK_HALF_2x16_SPLIT
:
498 return "pack_half_2x16_split";
499 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
:
500 return "unpack_half_2x16_split_x";
501 case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
:
502 return "unpack_half_2x16_split_y";
504 case FS_OPCODE_PLACEHOLDER_HALT
:
505 return "placeholder_halt";
507 case VS_OPCODE_URB_WRITE
:
508 return "vs_urb_write";
509 case VS_OPCODE_PULL_CONSTANT_LOAD
:
510 return "pull_constant_load";
511 case VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
:
512 return "pull_constant_load_gen7";
513 case VS_OPCODE_UNPACK_FLAGS_SIMD4X2
:
514 return "unpack_flags_simd4x2";
516 case GS_OPCODE_URB_WRITE
:
517 return "gs_urb_write";
518 case GS_OPCODE_THREAD_END
:
519 return "gs_thread_end";
520 case GS_OPCODE_SET_WRITE_OFFSET
:
521 return "set_write_offset";
522 case GS_OPCODE_SET_VERTEX_COUNT
:
523 return "set_vertex_count";
524 case GS_OPCODE_SET_DWORD_2_IMMED
:
525 return "set_dword_2_immed";
526 case GS_OPCODE_PREPARE_CHANNEL_MASKS
:
527 return "prepare_channel_masks";
528 case GS_OPCODE_SET_CHANNEL_MASKS
:
529 return "set_channel_masks";
532 /* Yes, this leaks. It's in debug code, it should never occur, and if
533 * it does, you should just add the case to the list above.
535 asprintf(&fallback
, "op%d", op
);
541 backend_instruction::is_tex()
543 return (opcode
== SHADER_OPCODE_TEX
||
544 opcode
== FS_OPCODE_TXB
||
545 opcode
== SHADER_OPCODE_TXD
||
546 opcode
== SHADER_OPCODE_TXF
||
547 opcode
== SHADER_OPCODE_TXF_CMS
||
548 opcode
== SHADER_OPCODE_TXF_UMS
||
549 opcode
== SHADER_OPCODE_TXF_MCS
||
550 opcode
== SHADER_OPCODE_TXL
||
551 opcode
== SHADER_OPCODE_TXS
||
552 opcode
== SHADER_OPCODE_LOD
||
553 opcode
== SHADER_OPCODE_TG4
||
554 opcode
== SHADER_OPCODE_TG4_OFFSET
);
558 backend_instruction::is_math()
560 return (opcode
== SHADER_OPCODE_RCP
||
561 opcode
== SHADER_OPCODE_RSQ
||
562 opcode
== SHADER_OPCODE_SQRT
||
563 opcode
== SHADER_OPCODE_EXP2
||
564 opcode
== SHADER_OPCODE_LOG2
||
565 opcode
== SHADER_OPCODE_SIN
||
566 opcode
== SHADER_OPCODE_COS
||
567 opcode
== SHADER_OPCODE_INT_QUOTIENT
||
568 opcode
== SHADER_OPCODE_INT_REMAINDER
||
569 opcode
== SHADER_OPCODE_POW
);
573 backend_instruction::is_control_flow()
577 case BRW_OPCODE_WHILE
:
579 case BRW_OPCODE_ELSE
:
580 case BRW_OPCODE_ENDIF
:
581 case BRW_OPCODE_BREAK
:
582 case BRW_OPCODE_CONTINUE
:
590 backend_instruction::can_do_source_mods()
593 case BRW_OPCODE_ADDC
:
595 case BRW_OPCODE_BFI1
:
596 case BRW_OPCODE_BFI2
:
597 case BRW_OPCODE_BFREV
:
598 case BRW_OPCODE_CBIT
:
601 case BRW_OPCODE_SUBB
:
609 backend_instruction::can_do_saturate()
619 case BRW_OPCODE_F16TO32
:
620 case BRW_OPCODE_F32TO16
:
621 case BRW_OPCODE_LINE
:
624 case BRW_OPCODE_MACH
:
626 case BRW_OPCODE_MATH
:
630 case BRW_OPCODE_RNDD
:
631 case BRW_OPCODE_RNDE
:
632 case BRW_OPCODE_RNDU
:
633 case BRW_OPCODE_RNDZ
:
637 case FS_OPCODE_LINTERP
:
638 case SHADER_OPCODE_COS
:
639 case SHADER_OPCODE_EXP2
:
640 case SHADER_OPCODE_LOG2
:
641 case SHADER_OPCODE_POW
:
642 case SHADER_OPCODE_RCP
:
643 case SHADER_OPCODE_RSQ
:
644 case SHADER_OPCODE_SIN
:
645 case SHADER_OPCODE_SQRT
:
653 backend_instruction::has_side_effects() const
656 case SHADER_OPCODE_UNTYPED_ATOMIC
:
664 backend_visitor::dump_instructions()
667 foreach_list(node
, &this->instructions
) {
668 backend_instruction
*inst
= (backend_instruction
*)node
;
669 printf("%d: ", ip
++);
670 dump_instruction(inst
);
676 * Sets up the starting offsets for the groups of binding table entries
677 * commong to all pipeline stages.
679 * Unused groups are initialized to 0xd0d0d0d0 to make it obvious that they're
680 * unused but also make sure that addition of small offsets to them will
681 * trigger some of our asserts that surface indices are < BRW_MAX_SURFACES.
684 backend_visitor::assign_common_binding_table_offsets(uint32_t next_binding_table_offset
)
686 int num_textures
= _mesa_fls(prog
->SamplersUsed
);
688 stage_prog_data
->binding_table
.texture_start
= next_binding_table_offset
;
689 next_binding_table_offset
+= num_textures
;
692 stage_prog_data
->binding_table
.ubo_start
= next_binding_table_offset
;
693 next_binding_table_offset
+= shader
->base
.NumUniformBlocks
;
695 stage_prog_data
->binding_table
.ubo_start
= 0xd0d0d0d0;
698 if (INTEL_DEBUG
& DEBUG_SHADER_TIME
) {
699 stage_prog_data
->binding_table
.shader_time_start
= next_binding_table_offset
;
700 next_binding_table_offset
++;
702 stage_prog_data
->binding_table
.shader_time_start
= 0xd0d0d0d0;
705 if (prog
->UsesGather
) {
706 stage_prog_data
->binding_table
.gather_texture_start
= next_binding_table_offset
;
707 next_binding_table_offset
+= num_textures
;
709 stage_prog_data
->binding_table
.gather_texture_start
= 0xd0d0d0d0;
712 if (shader_prog
&& shader_prog
->NumAtomicBuffers
) {
713 stage_prog_data
->binding_table
.abo_start
= next_binding_table_offset
;
714 next_binding_table_offset
+= shader_prog
->NumAtomicBuffers
;
716 stage_prog_data
->binding_table
.abo_start
= 0xd0d0d0d0;
719 /* This may or may not be used depending on how the compile goes. */
720 stage_prog_data
->binding_table
.pull_constants_start
= next_binding_table_offset
;
721 next_binding_table_offset
++;
723 assert(next_binding_table_offset
<= BRW_MAX_SURFACES
);
725 /* prog_data->base.binding_table.size will be set by brw_mark_surface_used. */