2 * Copyright © 2011 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 "glsl/ir_uniform.h"
27 #include "program/sampler.h"
32 vec4_instruction::vec4_instruction(vec4_visitor
*v
,
33 enum opcode opcode
, const dst_reg
&dst
,
34 const src_reg
&src0
, const src_reg
&src1
,
37 this->opcode
= opcode
;
42 this->saturate
= false;
43 this->force_writemask_all
= false;
44 this->no_dd_clear
= false;
45 this->no_dd_check
= false;
46 this->writes_accumulator
= false;
47 this->conditional_mod
= BRW_CONDITIONAL_NONE
;
48 this->texture_offset
= 0;
50 this->shadow_compare
= false;
51 this->ir
= v
->base_ir
;
52 this->urb_write_flags
= BRW_URB_WRITE_NO_FLAGS
;
53 this->header_present
= false;
57 this->annotation
= v
->current_annotation
;
61 vec4_visitor::emit(vec4_instruction
*inst
)
63 this->instructions
.push_tail(inst
);
69 vec4_visitor::emit_before(vec4_instruction
*inst
, vec4_instruction
*new_inst
)
71 new_inst
->ir
= inst
->ir
;
72 new_inst
->annotation
= inst
->annotation
;
74 inst
->insert_before(new_inst
);
80 vec4_visitor::emit(enum opcode opcode
, dst_reg dst
,
81 src_reg src0
, src_reg src1
, src_reg src2
)
83 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst
,
89 vec4_visitor::emit(enum opcode opcode
, dst_reg dst
, src_reg src0
, src_reg src1
)
91 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst
, src0
, src1
));
95 vec4_visitor::emit(enum opcode opcode
, dst_reg dst
, src_reg src0
)
97 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst
, src0
));
101 vec4_visitor::emit(enum opcode opcode
, dst_reg dst
)
103 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst
));
107 vec4_visitor::emit(enum opcode opcode
)
109 return emit(new(mem_ctx
) vec4_instruction(this, opcode
, dst_reg()));
114 vec4_visitor::op(const dst_reg &dst, const src_reg &src0) \
116 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
122 vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
123 const src_reg &src1) \
125 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
129 #define ALU2_ACC(op) \
131 vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
132 const src_reg &src1) \
134 vec4_instruction *inst = new(mem_ctx) vec4_instruction(this, \
135 BRW_OPCODE_##op, dst, src0, src1); \
136 inst->writes_accumulator = true; \
142 vec4_visitor::op(const dst_reg &dst, const src_reg &src0, \
143 const src_reg &src1, const src_reg &src2) \
145 assert(brw->gen >= 6); \
146 return new(mem_ctx) vec4_instruction(this, BRW_OPCODE_##op, dst, \
183 /** Gen4 predicated IF. */
185 vec4_visitor::IF(enum brw_predicate predicate
)
187 vec4_instruction
*inst
;
189 inst
= new(mem_ctx
) vec4_instruction(this, BRW_OPCODE_IF
);
190 inst
->predicate
= predicate
;
195 /** Gen6 IF with embedded comparison. */
197 vec4_visitor::IF(src_reg src0
, src_reg src1
,
198 enum brw_conditional_mod condition
)
200 assert(brw
->gen
== 6);
202 vec4_instruction
*inst
;
204 resolve_ud_negate(&src0
);
205 resolve_ud_negate(&src1
);
207 inst
= new(mem_ctx
) vec4_instruction(this, BRW_OPCODE_IF
, dst_null_d(),
209 inst
->conditional_mod
= condition
;
215 * CMP: Sets the low bit of the destination channels with the result
216 * of the comparison, while the upper bits are undefined, and updates
217 * the flag register with the packed 16 bits of the result.
220 vec4_visitor::CMP(dst_reg dst
, src_reg src0
, src_reg src1
,
221 enum brw_conditional_mod condition
)
223 vec4_instruction
*inst
;
225 /* original gen4 does type conversion to the destination type
226 * before before comparison, producing garbage results for floating
230 dst
.type
= src0
.type
;
231 if (dst
.file
== HW_REG
)
232 dst
.fixed_hw_reg
.type
= dst
.type
;
235 resolve_ud_negate(&src0
);
236 resolve_ud_negate(&src1
);
238 inst
= new(mem_ctx
) vec4_instruction(this, BRW_OPCODE_CMP
, dst
, src0
, src1
);
239 inst
->conditional_mod
= condition
;
245 vec4_visitor::SCRATCH_READ(const dst_reg
&dst
, const src_reg
&index
)
247 vec4_instruction
*inst
;
249 inst
= new(mem_ctx
) vec4_instruction(this, SHADER_OPCODE_GEN4_SCRATCH_READ
,
258 vec4_visitor::SCRATCH_WRITE(const dst_reg
&dst
, const src_reg
&src
,
259 const src_reg
&index
)
261 vec4_instruction
*inst
;
263 inst
= new(mem_ctx
) vec4_instruction(this, SHADER_OPCODE_GEN4_SCRATCH_WRITE
,
272 vec4_visitor::emit_dp(dst_reg dst
, src_reg src0
, src_reg src1
, unsigned elements
)
274 static enum opcode dot_opcodes
[] = {
275 BRW_OPCODE_DP2
, BRW_OPCODE_DP3
, BRW_OPCODE_DP4
278 emit(dot_opcodes
[elements
- 2], dst
, src0
, src1
);
282 vec4_visitor::fix_3src_operand(src_reg src
)
284 /* Using vec4 uniforms in SIMD4x2 programs is difficult. You'd like to be
285 * able to use vertical stride of zero to replicate the vec4 uniform, like
287 * g3<0;4,1>:f - [0, 4][1, 5][2, 6][3, 7]
289 * But you can't, since vertical stride is always four in three-source
290 * instructions. Instead, insert a MOV instruction to do the replication so
291 * that the three-source instruction can consume it.
294 /* The MOV is only needed if the source is a uniform or immediate. */
295 if (src
.file
!= UNIFORM
&& src
.file
!= IMM
)
298 if (src
.file
== UNIFORM
&& brw_is_single_value_swizzle(src
.swizzle
))
301 dst_reg expanded
= dst_reg(this, glsl_type::vec4_type
);
302 expanded
.type
= src
.type
;
303 emit(MOV(expanded
, src
));
304 return src_reg(expanded
);
308 vec4_visitor::fix_math_operand(src_reg src
)
310 /* The gen6 math instruction ignores the source modifiers --
311 * swizzle, abs, negate, and at least some parts of the register
312 * region description.
314 * Rather than trying to enumerate all these cases, *always* expand the
315 * operand to a temp GRF for gen6.
317 * For gen7, keep the operand as-is, except if immediate, which gen7 still
321 if (brw
->gen
== 7 && src
.file
!= IMM
)
324 dst_reg expanded
= dst_reg(this, glsl_type::vec4_type
);
325 expanded
.type
= src
.type
;
326 emit(MOV(expanded
, src
));
327 return src_reg(expanded
);
331 vec4_visitor::emit_math1_gen6(enum opcode opcode
, dst_reg dst
, src_reg src
)
333 src
= fix_math_operand(src
);
335 if (brw
->gen
== 6 && dst
.writemask
!= WRITEMASK_XYZW
) {
336 /* The gen6 math instruction must be align1, so we can't do
339 dst_reg temp_dst
= dst_reg(this, glsl_type::vec4_type
);
341 emit(opcode
, temp_dst
, src
);
343 emit(MOV(dst
, src_reg(temp_dst
)));
345 emit(opcode
, dst
, src
);
350 vec4_visitor::emit_math1_gen4(enum opcode opcode
, dst_reg dst
, src_reg src
)
352 vec4_instruction
*inst
= emit(opcode
, dst
, src
);
358 vec4_visitor::emit_math(opcode opcode
, dst_reg dst
, src_reg src
)
361 case SHADER_OPCODE_RCP
:
362 case SHADER_OPCODE_RSQ
:
363 case SHADER_OPCODE_SQRT
:
364 case SHADER_OPCODE_EXP2
:
365 case SHADER_OPCODE_LOG2
:
366 case SHADER_OPCODE_SIN
:
367 case SHADER_OPCODE_COS
:
370 unreachable("not reached: bad math opcode");
374 emit(opcode
, dst
, src
);
375 } else if (brw
->gen
>= 6) {
376 emit_math1_gen6(opcode
, dst
, src
);
378 emit_math1_gen4(opcode
, dst
, src
);
383 vec4_visitor::emit_math2_gen6(enum opcode opcode
,
384 dst_reg dst
, src_reg src0
, src_reg src1
)
386 src0
= fix_math_operand(src0
);
387 src1
= fix_math_operand(src1
);
389 if (brw
->gen
== 6 && dst
.writemask
!= WRITEMASK_XYZW
) {
390 /* The gen6 math instruction must be align1, so we can't do
393 dst_reg temp_dst
= dst_reg(this, glsl_type::vec4_type
);
394 temp_dst
.type
= dst
.type
;
396 emit(opcode
, temp_dst
, src0
, src1
);
398 emit(MOV(dst
, src_reg(temp_dst
)));
400 emit(opcode
, dst
, src0
, src1
);
405 vec4_visitor::emit_math2_gen4(enum opcode opcode
,
406 dst_reg dst
, src_reg src0
, src_reg src1
)
408 vec4_instruction
*inst
= emit(opcode
, dst
, src0
, src1
);
414 vec4_visitor::emit_math(enum opcode opcode
,
415 dst_reg dst
, src_reg src0
, src_reg src1
)
418 case SHADER_OPCODE_POW
:
419 case SHADER_OPCODE_INT_QUOTIENT
:
420 case SHADER_OPCODE_INT_REMAINDER
:
423 unreachable("not reached: unsupported binary math opcode");
427 emit(opcode
, dst
, src0
, src1
);
428 } else if (brw
->gen
>= 6) {
429 emit_math2_gen6(opcode
, dst
, src0
, src1
);
431 emit_math2_gen4(opcode
, dst
, src0
, src1
);
436 vec4_visitor::emit_pack_half_2x16(dst_reg dst
, src_reg src0
)
439 unreachable("ir_unop_pack_half_2x16 should be lowered");
442 assert(dst
.type
== BRW_REGISTER_TYPE_UD
);
443 assert(src0
.type
== BRW_REGISTER_TYPE_F
);
445 /* From the Ivybridge PRM, Vol4, Part3, Section 6.27 f32to16:
447 * Because this instruction does not have a 16-bit floating-point type,
448 * the destination data type must be Word (W).
450 * The destination must be DWord-aligned and specify a horizontal stride
451 * (HorzStride) of 2. The 16-bit result is stored in the lower word of
452 * each destination channel and the upper word is not modified.
454 * The above restriction implies that the f32to16 instruction must use
455 * align1 mode, because only in align1 mode is it possible to specify
456 * horizontal stride. We choose here to defy the hardware docs and emit
457 * align16 instructions.
459 * (I [chadv] did attempt to emit align1 instructions for VS f32to16
460 * instructions. I was partially successful in that the code passed all
461 * tests. However, the code was dubiously correct and fragile, and the
462 * tests were not harsh enough to probe that frailty. Not trusting the
463 * code, I chose instead to remain in align16 mode in defiance of the hw
466 * I've [chadv] experimentally confirmed that, on gen7 hardware and the
467 * simulator, emitting a f32to16 in align16 mode with UD as destination
468 * data type is safe. The behavior differs from that specified in the PRM
469 * in that the upper word of each destination channel is cleared to 0.
472 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
473 src_reg
tmp_src(tmp_dst
);
476 /* Verify the undocumented behavior on which the following instructions
477 * rely. If f32to16 fails to clear the upper word of the X and Y channels,
478 * then the result of the bit-or instruction below will be incorrect.
480 * You should inspect the disasm output in order to verify that the MOV is
481 * not optimized away.
483 emit(MOV(tmp_dst
, src_reg(0x12345678u
)));
486 /* Give tmp the form below, where "." means untouched.
489 * |.|.|0x0000hhhh|0x0000llll|.|.|0x0000hhhh|0x0000llll|
491 * That the upper word of each write-channel be 0 is required for the
492 * following bit-shift and bit-or instructions to work. Note that this
493 * relies on the undocumented hardware behavior mentioned above.
495 tmp_dst
.writemask
= WRITEMASK_XY
;
496 emit(F32TO16(tmp_dst
, src0
));
498 /* Give the write-channels of dst the form:
501 tmp_src
.swizzle
= BRW_SWIZZLE_YYYY
;
502 emit(SHL(dst
, tmp_src
, src_reg(16u)));
504 /* Finally, give the write-channels of dst the form of packHalf2x16's
508 tmp_src
.swizzle
= BRW_SWIZZLE_XXXX
;
509 emit(OR(dst
, src_reg(dst
), tmp_src
));
513 vec4_visitor::emit_unpack_half_2x16(dst_reg dst
, src_reg src0
)
516 unreachable("ir_unop_unpack_half_2x16 should be lowered");
519 assert(dst
.type
== BRW_REGISTER_TYPE_F
);
520 assert(src0
.type
== BRW_REGISTER_TYPE_UD
);
522 /* From the Ivybridge PRM, Vol4, Part3, Section 6.26 f16to32:
524 * Because this instruction does not have a 16-bit floating-point type,
525 * the source data type must be Word (W). The destination type must be
528 * To use W as the source data type, we must adjust horizontal strides,
529 * which is only possible in align1 mode. All my [chadv] attempts at
530 * emitting align1 instructions for unpackHalf2x16 failed to pass the
531 * Piglit tests, so I gave up.
533 * I've verified that, on gen7 hardware and the simulator, it is safe to
534 * emit f16to32 in align16 mode with UD as source data type.
537 dst_reg
tmp_dst(this, glsl_type::uvec2_type
);
538 src_reg
tmp_src(tmp_dst
);
540 tmp_dst
.writemask
= WRITEMASK_X
;
541 emit(AND(tmp_dst
, src0
, src_reg(0xffffu
)));
543 tmp_dst
.writemask
= WRITEMASK_Y
;
544 emit(SHR(tmp_dst
, src0
, src_reg(16u)));
546 dst
.writemask
= WRITEMASK_XY
;
547 emit(F16TO32(dst
, tmp_src
));
551 vec4_visitor::visit_instructions(const exec_list
*list
)
553 foreach_in_list(ir_instruction
, ir
, list
) {
561 type_size(const struct glsl_type
*type
)
566 switch (type
->base_type
) {
569 case GLSL_TYPE_FLOAT
:
571 if (type
->is_matrix()) {
572 return type
->matrix_columns
;
574 /* Regardless of size of vector, it gets a vec4. This is bad
575 * packing for things like floats, but otherwise arrays become a
576 * mess. Hopefully a later pass over the code can pack scalars
577 * down if appropriate.
581 case GLSL_TYPE_ARRAY
:
582 assert(type
->length
> 0);
583 return type_size(type
->fields
.array
) * type
->length
;
584 case GLSL_TYPE_STRUCT
:
586 for (i
= 0; i
< type
->length
; i
++) {
587 size
+= type_size(type
->fields
.structure
[i
].type
);
590 case GLSL_TYPE_SAMPLER
:
591 /* Samplers take up one slot in UNIFORMS[], but they're baked in
595 case GLSL_TYPE_ATOMIC_UINT
:
597 case GLSL_TYPE_IMAGE
:
599 case GLSL_TYPE_ERROR
:
600 case GLSL_TYPE_INTERFACE
:
601 unreachable("not reached");
608 vec4_visitor::virtual_grf_alloc(int size
)
610 if (virtual_grf_array_size
<= virtual_grf_count
) {
611 if (virtual_grf_array_size
== 0)
612 virtual_grf_array_size
= 16;
614 virtual_grf_array_size
*= 2;
615 virtual_grf_sizes
= reralloc(mem_ctx
, virtual_grf_sizes
, int,
616 virtual_grf_array_size
);
617 virtual_grf_reg_map
= reralloc(mem_ctx
, virtual_grf_reg_map
, int,
618 virtual_grf_array_size
);
620 virtual_grf_reg_map
[virtual_grf_count
] = virtual_grf_reg_count
;
621 virtual_grf_reg_count
+= size
;
622 virtual_grf_sizes
[virtual_grf_count
] = size
;
623 return virtual_grf_count
++;
626 src_reg::src_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
631 this->reg
= v
->virtual_grf_alloc(type_size(type
));
633 if (type
->is_array() || type
->is_record()) {
634 this->swizzle
= BRW_SWIZZLE_NOOP
;
636 this->swizzle
= swizzle_for_size(type
->vector_elements
);
639 this->type
= brw_type_for_base_type(type
);
642 dst_reg::dst_reg(class vec4_visitor
*v
, const struct glsl_type
*type
)
647 this->reg
= v
->virtual_grf_alloc(type_size(type
));
649 if (type
->is_array() || type
->is_record()) {
650 this->writemask
= WRITEMASK_XYZW
;
652 this->writemask
= (1 << type
->vector_elements
) - 1;
655 this->type
= brw_type_for_base_type(type
);
658 /* Our support for uniforms is piggy-backed on the struct
659 * gl_fragment_program, because that's where the values actually
660 * get stored, rather than in some global gl_shader_program uniform
664 vec4_visitor::setup_uniform_values(ir_variable
*ir
)
666 int namelen
= strlen(ir
->name
);
668 /* The data for our (non-builtin) uniforms is stored in a series of
669 * gl_uniform_driver_storage structs for each subcomponent that
670 * glGetUniformLocation() could name. We know it's been set up in the same
671 * order we'd walk the type, so walk the list of storage and find anything
672 * with our name, or the prefix of a component that starts with our name.
674 for (unsigned u
= 0; u
< shader_prog
->NumUserUniformStorage
; u
++) {
675 struct gl_uniform_storage
*storage
= &shader_prog
->UniformStorage
[u
];
677 if (strncmp(ir
->name
, storage
->name
, namelen
) != 0 ||
678 (storage
->name
[namelen
] != 0 &&
679 storage
->name
[namelen
] != '.' &&
680 storage
->name
[namelen
] != '[')) {
684 gl_constant_value
*components
= storage
->storage
;
685 unsigned vector_count
= (MAX2(storage
->array_elements
, 1) *
686 storage
->type
->matrix_columns
);
688 for (unsigned s
= 0; s
< vector_count
; s
++) {
689 assert(uniforms
< uniform_array_size
);
690 uniform_vector_size
[uniforms
] = storage
->type
->vector_elements
;
693 for (i
= 0; i
< uniform_vector_size
[uniforms
]; i
++) {
694 stage_prog_data
->param
[uniforms
* 4 + i
] = &components
->f
;
698 static float zero
= 0;
699 stage_prog_data
->param
[uniforms
* 4 + i
] = &zero
;
708 vec4_visitor::setup_uniform_clipplane_values()
710 gl_clip_plane
*clip_planes
= brw_select_clip_planes(ctx
);
712 for (int i
= 0; i
< key
->nr_userclip_plane_consts
; ++i
) {
713 assert(this->uniforms
< uniform_array_size
);
714 this->uniform_vector_size
[this->uniforms
] = 4;
715 this->userplane
[i
] = dst_reg(UNIFORM
, this->uniforms
);
716 this->userplane
[i
].type
= BRW_REGISTER_TYPE_F
;
717 for (int j
= 0; j
< 4; ++j
) {
718 stage_prog_data
->param
[this->uniforms
* 4 + j
] = &clip_planes
[i
][j
];
724 /* Our support for builtin uniforms is even scarier than non-builtin.
725 * It sits on top of the PROG_STATE_VAR parameters that are
726 * automatically updated from GL context state.
729 vec4_visitor::setup_builtin_uniform_values(ir_variable
*ir
)
731 const ir_state_slot
*const slots
= ir
->state_slots
;
732 assert(ir
->state_slots
!= NULL
);
734 for (unsigned int i
= 0; i
< ir
->num_state_slots
; i
++) {
735 /* This state reference has already been setup by ir_to_mesa,
736 * but we'll get the same index back here. We can reference
737 * ParameterValues directly, since unlike brw_fs.cpp, we never
738 * add new state references during compile.
740 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
741 (gl_state_index
*)slots
[i
].tokens
);
742 float *values
= &this->prog
->Parameters
->ParameterValues
[index
][0].f
;
744 assert(this->uniforms
< uniform_array_size
);
745 this->uniform_vector_size
[this->uniforms
] = 0;
746 /* Add each of the unique swizzled channels of the element.
747 * This will end up matching the size of the glsl_type of this field.
750 for (unsigned int j
= 0; j
< 4; j
++) {
751 int swiz
= GET_SWZ(slots
[i
].swizzle
, j
);
754 stage_prog_data
->param
[this->uniforms
* 4 + j
] = &values
[swiz
];
755 assert(this->uniforms
< uniform_array_size
);
756 if (swiz
<= last_swiz
)
757 this->uniform_vector_size
[this->uniforms
]++;
764 vec4_visitor::variable_storage(ir_variable
*var
)
766 return (dst_reg
*)hash_table_find(this->variable_ht
, var
);
770 vec4_visitor::emit_bool_to_cond_code(ir_rvalue
*ir
,
771 enum brw_predicate
*predicate
)
773 ir_expression
*expr
= ir
->as_expression();
775 *predicate
= BRW_PREDICATE_NORMAL
;
779 vec4_instruction
*inst
;
781 assert(expr
->get_num_operands() <= 2);
782 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
783 expr
->operands
[i
]->accept(this);
784 op
[i
] = this->result
;
786 resolve_ud_negate(&op
[i
]);
789 switch (expr
->operation
) {
790 case ir_unop_logic_not
:
791 inst
= emit(AND(dst_null_d(), op
[0], src_reg(1)));
792 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
795 case ir_binop_logic_xor
:
796 inst
= emit(XOR(dst_null_d(), op
[0], op
[1]));
797 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
800 case ir_binop_logic_or
:
801 inst
= emit(OR(dst_null_d(), op
[0], op
[1]));
802 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
805 case ir_binop_logic_and
:
806 inst
= emit(AND(dst_null_d(), op
[0], op
[1]));
807 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
812 emit(CMP(dst_null_d(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
814 inst
= emit(MOV(dst_null_f(), op
[0]));
815 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
821 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
823 inst
= emit(MOV(dst_null_d(), op
[0]));
824 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
828 case ir_binop_all_equal
:
829 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
830 *predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
833 case ir_binop_any_nequal
:
834 inst
= emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
835 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
839 inst
= emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
840 *predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
843 case ir_binop_greater
:
844 case ir_binop_gequal
:
846 case ir_binop_lequal
:
848 case ir_binop_nequal
:
849 emit(CMP(dst_null_d(), op
[0], op
[1],
850 brw_conditional_for_comparison(expr
->operation
)));
854 unreachable("not reached");
861 resolve_ud_negate(&this->result
);
864 vec4_instruction
*inst
= emit(AND(dst_null_d(),
865 this->result
, src_reg(1)));
866 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
868 vec4_instruction
*inst
= emit(MOV(dst_null_d(), this->result
));
869 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
874 * Emit a gen6 IF statement with the comparison folded into the IF
878 vec4_visitor::emit_if_gen6(ir_if
*ir
)
880 ir_expression
*expr
= ir
->condition
->as_expression();
886 assert(expr
->get_num_operands() <= 2);
887 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
888 expr
->operands
[i
]->accept(this);
889 op
[i
] = this->result
;
892 switch (expr
->operation
) {
893 case ir_unop_logic_not
:
894 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_Z
));
897 case ir_binop_logic_xor
:
898 emit(IF(op
[0], op
[1], BRW_CONDITIONAL_NZ
));
901 case ir_binop_logic_or
:
902 temp
= dst_reg(this, glsl_type::bool_type
);
903 emit(OR(temp
, op
[0], op
[1]));
904 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
907 case ir_binop_logic_and
:
908 temp
= dst_reg(this, glsl_type::bool_type
);
909 emit(AND(temp
, op
[0], op
[1]));
910 emit(IF(src_reg(temp
), src_reg(0), BRW_CONDITIONAL_NZ
));
914 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
918 emit(IF(op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
921 case ir_binop_greater
:
922 case ir_binop_gequal
:
924 case ir_binop_lequal
:
926 case ir_binop_nequal
:
927 emit(IF(op
[0], op
[1],
928 brw_conditional_for_comparison(expr
->operation
)));
931 case ir_binop_all_equal
:
932 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
933 emit(IF(BRW_PREDICATE_ALIGN16_ALL4H
));
936 case ir_binop_any_nequal
:
937 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
938 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
942 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
943 emit(IF(BRW_PREDICATE_ALIGN16_ANY4H
));
947 unreachable("not reached");
952 ir
->condition
->accept(this);
954 emit(IF(this->result
, src_reg(0), BRW_CONDITIONAL_NZ
));
958 vec4_visitor::visit(ir_variable
*ir
)
962 if (variable_storage(ir
))
965 switch (ir
->data
.mode
) {
966 case ir_var_shader_in
:
967 reg
= new(mem_ctx
) dst_reg(ATTR
, ir
->data
.location
);
970 case ir_var_shader_out
:
971 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
973 for (int i
= 0; i
< type_size(ir
->type
); i
++) {
974 output_reg
[ir
->data
.location
+ i
] = *reg
;
975 output_reg
[ir
->data
.location
+ i
].reg_offset
= i
;
976 output_reg
[ir
->data
.location
+ i
].type
=
977 brw_type_for_base_type(ir
->type
->get_scalar_type());
978 output_reg_annotation
[ir
->data
.location
+ i
] = ir
->name
;
983 case ir_var_temporary
:
984 reg
= new(mem_ctx
) dst_reg(this, ir
->type
);
988 reg
= new(this->mem_ctx
) dst_reg(UNIFORM
, this->uniforms
);
990 /* Thanks to the lower_ubo_reference pass, we will see only
991 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
992 * variables, so no need for them to be in variable_ht.
994 * Atomic counters take no uniform storage, no need to do
997 if (ir
->is_in_uniform_block() || ir
->type
->contains_atomic())
1000 /* Track how big the whole uniform variable is, in case we need to put a
1001 * copy of its data into pull constants for array access.
1003 assert(this->uniforms
< uniform_array_size
);
1004 this->uniform_size
[this->uniforms
] = type_size(ir
->type
);
1006 if (!strncmp(ir
->name
, "gl_", 3)) {
1007 setup_builtin_uniform_values(ir
);
1009 setup_uniform_values(ir
);
1013 case ir_var_system_value
:
1014 reg
= make_reg_for_system_value(ir
);
1018 unreachable("not reached");
1021 reg
->type
= brw_type_for_base_type(ir
->type
);
1022 hash_table_insert(this->variable_ht
, reg
, ir
);
1026 vec4_visitor::visit(ir_loop
*ir
)
1028 /* We don't want debugging output to print the whole body of the
1029 * loop as the annotation.
1031 this->base_ir
= NULL
;
1033 emit(BRW_OPCODE_DO
);
1035 visit_instructions(&ir
->body_instructions
);
1037 emit(BRW_OPCODE_WHILE
);
1041 vec4_visitor::visit(ir_loop_jump
*ir
)
1044 case ir_loop_jump::jump_break
:
1045 emit(BRW_OPCODE_BREAK
);
1047 case ir_loop_jump::jump_continue
:
1048 emit(BRW_OPCODE_CONTINUE
);
1055 vec4_visitor::visit(ir_function_signature
*)
1057 unreachable("not reached");
1061 vec4_visitor::visit(ir_function
*ir
)
1063 /* Ignore function bodies other than main() -- we shouldn't see calls to
1064 * them since they should all be inlined.
1066 if (strcmp(ir
->name
, "main") == 0) {
1067 const ir_function_signature
*sig
;
1070 sig
= ir
->matching_signature(NULL
, &empty
, false);
1074 visit_instructions(&sig
->body
);
1079 vec4_visitor::try_emit_sat(ir_expression
*ir
)
1081 ir_rvalue
*sat_src
= ir
->as_rvalue_to_saturate();
1085 sat_src
->accept(this);
1086 src_reg src
= this->result
;
1088 this->result
= src_reg(this, ir
->type
);
1089 vec4_instruction
*inst
;
1090 inst
= emit(MOV(dst_reg(this->result
), src
));
1091 inst
->saturate
= true;
1097 vec4_visitor::try_emit_mad(ir_expression
*ir
)
1099 /* 3-src instructions were introduced in gen6. */
1103 /* MAD can only handle floating-point data. */
1104 if (ir
->type
->base_type
!= GLSL_TYPE_FLOAT
)
1107 ir_rvalue
*nonmul
= ir
->operands
[1];
1108 ir_expression
*mul
= ir
->operands
[0]->as_expression();
1110 if (!mul
|| mul
->operation
!= ir_binop_mul
) {
1111 nonmul
= ir
->operands
[0];
1112 mul
= ir
->operands
[1]->as_expression();
1114 if (!mul
|| mul
->operation
!= ir_binop_mul
)
1118 nonmul
->accept(this);
1119 src_reg src0
= fix_3src_operand(this->result
);
1121 mul
->operands
[0]->accept(this);
1122 src_reg src1
= fix_3src_operand(this->result
);
1124 mul
->operands
[1]->accept(this);
1125 src_reg src2
= fix_3src_operand(this->result
);
1127 this->result
= src_reg(this, ir
->type
);
1128 emit(BRW_OPCODE_MAD
, dst_reg(this->result
), src0
, src1
, src2
);
1134 vec4_visitor::try_emit_b2f_of_compare(ir_expression
*ir
)
1136 ir_expression
*const cmp
= ir
->operands
[0]->as_expression();
1141 switch (cmp
->operation
) {
1143 case ir_binop_greater
:
1144 case ir_binop_lequal
:
1145 case ir_binop_gequal
:
1146 case ir_binop_equal
:
1147 case ir_binop_nequal
:
1154 cmp
->operands
[0]->accept(this);
1155 const src_reg cmp_src0
= this->result
;
1157 cmp
->operands
[1]->accept(this);
1158 const src_reg cmp_src1
= this->result
;
1160 this->result
= src_reg(this, ir
->type
);
1162 emit(CMP(dst_reg(this->result
), cmp_src0
, cmp_src1
,
1163 brw_conditional_for_comparison(cmp
->operation
)));
1165 /* If the comparison is false, this->result will just happen to be zero.
1167 vec4_instruction
*const inst
= emit(BRW_OPCODE_SEL
, dst_reg(this->result
),
1168 this->result
, src_reg(1.0f
));
1169 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1170 inst
->predicate_inverse
= true;
1176 vec4_visitor::emit_minmax(enum brw_conditional_mod conditionalmod
, dst_reg dst
,
1177 src_reg src0
, src_reg src1
)
1179 vec4_instruction
*inst
;
1181 if (brw
->gen
>= 6) {
1182 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1183 inst
->conditional_mod
= conditionalmod
;
1185 emit(CMP(dst
, src0
, src1
, conditionalmod
));
1187 inst
= emit(BRW_OPCODE_SEL
, dst
, src0
, src1
);
1188 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1193 vec4_visitor::emit_lrp(const dst_reg
&dst
,
1194 const src_reg
&x
, const src_reg
&y
, const src_reg
&a
)
1196 if (brw
->gen
>= 6) {
1197 /* Note that the instruction's argument order is reversed from GLSL
1201 fix_3src_operand(a
), fix_3src_operand(y
), fix_3src_operand(x
)));
1203 /* Earlier generations don't support three source operations, so we
1204 * need to emit x*(1-a) + y*a.
1206 dst_reg y_times_a
= dst_reg(this, glsl_type::vec4_type
);
1207 dst_reg one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
1208 dst_reg x_times_one_minus_a
= dst_reg(this, glsl_type::vec4_type
);
1209 y_times_a
.writemask
= dst
.writemask
;
1210 one_minus_a
.writemask
= dst
.writemask
;
1211 x_times_one_minus_a
.writemask
= dst
.writemask
;
1213 emit(MUL(y_times_a
, y
, a
));
1214 emit(ADD(one_minus_a
, negate(a
), src_reg(1.0f
)));
1215 emit(MUL(x_times_one_minus_a
, x
, src_reg(one_minus_a
)));
1216 emit(ADD(dst
, src_reg(x_times_one_minus_a
), src_reg(y_times_a
)));
1221 vec4_visitor::visit(ir_expression
*ir
)
1223 unsigned int operand
;
1224 src_reg op
[Elements(ir
->operands
)];
1227 vec4_instruction
*inst
;
1229 if (try_emit_sat(ir
))
1232 if (ir
->operation
== ir_binop_add
) {
1233 if (try_emit_mad(ir
))
1237 if (ir
->operation
== ir_unop_b2f
) {
1238 if (try_emit_b2f_of_compare(ir
))
1242 for (operand
= 0; operand
< ir
->get_num_operands(); operand
++) {
1243 this->result
.file
= BAD_FILE
;
1244 ir
->operands
[operand
]->accept(this);
1245 if (this->result
.file
== BAD_FILE
) {
1246 fprintf(stderr
, "Failed to get tree for expression operand:\n");
1247 ir
->operands
[operand
]->fprint(stderr
);
1250 op
[operand
] = this->result
;
1252 /* Matrix expression operands should have been broken down to vector
1253 * operations already.
1255 assert(!ir
->operands
[operand
]->type
->is_matrix());
1258 int vector_elements
= ir
->operands
[0]->type
->vector_elements
;
1259 if (ir
->operands
[1]) {
1260 vector_elements
= MAX2(vector_elements
,
1261 ir
->operands
[1]->type
->vector_elements
);
1264 this->result
.file
= BAD_FILE
;
1266 /* Storage for our result. Ideally for an assignment we'd be using
1267 * the actual storage for the result here, instead.
1269 result_src
= src_reg(this, ir
->type
);
1270 /* convenience for the emit functions below. */
1271 result_dst
= dst_reg(result_src
);
1272 /* If nothing special happens, this is the result. */
1273 this->result
= result_src
;
1274 /* Limit writes to the channels that will be used by result_src later.
1275 * This does limit this temp's use as a temporary for multi-instruction
1278 result_dst
.writemask
= (1 << ir
->type
->vector_elements
) - 1;
1280 switch (ir
->operation
) {
1281 case ir_unop_logic_not
:
1282 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
1283 * ones complement of the whole register, not just bit 0.
1285 emit(XOR(result_dst
, op
[0], src_reg(1)));
1288 op
[0].negate
= !op
[0].negate
;
1289 emit(MOV(result_dst
, op
[0]));
1293 op
[0].negate
= false;
1294 emit(MOV(result_dst
, op
[0]));
1298 if (ir
->type
->is_float()) {
1299 /* AND(val, 0x80000000) gives the sign bit.
1301 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
1304 emit(CMP(dst_null_f(), op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1306 op
[0].type
= BRW_REGISTER_TYPE_UD
;
1307 result_dst
.type
= BRW_REGISTER_TYPE_UD
;
1308 emit(AND(result_dst
, op
[0], src_reg(0x80000000u
)));
1310 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(0x3f800000u
)));
1311 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1313 this->result
.type
= BRW_REGISTER_TYPE_F
;
1315 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
1316 * -> non-negative val generates 0x00000000.
1317 * Predicated OR sets 1 if val is positive.
1319 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_G
));
1321 emit(ASR(result_dst
, op
[0], src_reg(31)));
1323 inst
= emit(OR(result_dst
, src_reg(result_dst
), src_reg(1)));
1324 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1329 emit_math(SHADER_OPCODE_RCP
, result_dst
, op
[0]);
1333 emit_math(SHADER_OPCODE_EXP2
, result_dst
, op
[0]);
1336 emit_math(SHADER_OPCODE_LOG2
, result_dst
, op
[0]);
1340 unreachable("not reached: should be handled by ir_explog_to_explog2");
1342 case ir_unop_sin_reduced
:
1343 emit_math(SHADER_OPCODE_SIN
, result_dst
, op
[0]);
1346 case ir_unop_cos_reduced
:
1347 emit_math(SHADER_OPCODE_COS
, result_dst
, op
[0]);
1352 unreachable("derivatives not valid in vertex shader");
1354 case ir_unop_bitfield_reverse
:
1355 emit(BFREV(result_dst
, op
[0]));
1357 case ir_unop_bit_count
:
1358 emit(CBIT(result_dst
, op
[0]));
1360 case ir_unop_find_msb
: {
1361 src_reg temp
= src_reg(this, glsl_type::uint_type
);
1363 inst
= emit(FBH(dst_reg(temp
), op
[0]));
1364 inst
->dst
.writemask
= WRITEMASK_XYZW
;
1366 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1367 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1368 * subtract the result from 31 to convert the MSB count into an LSB count.
1371 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
1372 temp
.swizzle
= BRW_SWIZZLE_NOOP
;
1373 emit(MOV(result_dst
, temp
));
1375 src_reg src_tmp
= src_reg(result_dst
);
1376 emit(CMP(dst_null_d(), src_tmp
, src_reg(-1), BRW_CONDITIONAL_NZ
));
1378 src_tmp
.negate
= true;
1379 inst
= emit(ADD(result_dst
, src_tmp
, src_reg(31)));
1380 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1383 case ir_unop_find_lsb
:
1384 emit(FBL(result_dst
, op
[0]));
1388 unreachable("not reached: should be handled by lower_noise");
1391 emit(ADD(result_dst
, op
[0], op
[1]));
1394 unreachable("not reached: should be handled by ir_sub_to_add_neg");
1397 if (brw
->gen
< 8 && ir
->type
->is_integer()) {
1398 /* For integer multiplication, the MUL uses the low 16 bits of one of
1399 * the operands (src0 through SNB, src1 on IVB and later). The MACH
1400 * accumulates in the contribution of the upper 16 bits of that
1401 * operand. If we can determine that one of the args is in the low
1402 * 16 bits, though, we can just emit a single MUL.
1404 if (ir
->operands
[0]->is_uint16_constant()) {
1406 emit(MUL(result_dst
, op
[0], op
[1]));
1408 emit(MUL(result_dst
, op
[1], op
[0]));
1409 } else if (ir
->operands
[1]->is_uint16_constant()) {
1411 emit(MUL(result_dst
, op
[1], op
[0]));
1413 emit(MUL(result_dst
, op
[0], op
[1]));
1415 struct brw_reg acc
= retype(brw_acc_reg(), result_dst
.type
);
1417 emit(MUL(acc
, op
[0], op
[1]));
1418 emit(MACH(dst_null_d(), op
[0], op
[1]));
1419 emit(MOV(result_dst
, src_reg(acc
)));
1422 emit(MUL(result_dst
, op
[0], op
[1]));
1425 case ir_binop_imul_high
: {
1426 struct brw_reg acc
= retype(brw_acc_reg(), result_dst
.type
);
1428 emit(MUL(acc
, op
[0], op
[1]));
1429 emit(MACH(result_dst
, op
[0], op
[1]));
1433 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
1434 assert(ir
->type
->is_integer());
1435 emit_math(SHADER_OPCODE_INT_QUOTIENT
, result_dst
, op
[0], op
[1]);
1437 case ir_binop_carry
: {
1438 struct brw_reg acc
= retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD
);
1440 emit(ADDC(dst_null_ud(), op
[0], op
[1]));
1441 emit(MOV(result_dst
, src_reg(acc
)));
1444 case ir_binop_borrow
: {
1445 struct brw_reg acc
= retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD
);
1447 emit(SUBB(dst_null_ud(), op
[0], op
[1]));
1448 emit(MOV(result_dst
, src_reg(acc
)));
1452 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
1453 assert(ir
->type
->is_integer());
1454 emit_math(SHADER_OPCODE_INT_REMAINDER
, result_dst
, op
[0], op
[1]);
1458 case ir_binop_greater
:
1459 case ir_binop_lequal
:
1460 case ir_binop_gequal
:
1461 case ir_binop_equal
:
1462 case ir_binop_nequal
: {
1463 emit(CMP(result_dst
, op
[0], op
[1],
1464 brw_conditional_for_comparison(ir
->operation
)));
1465 emit(AND(result_dst
, result_src
, src_reg(0x1)));
1469 case ir_binop_all_equal
:
1470 /* "==" operator producing a scalar boolean. */
1471 if (ir
->operands
[0]->type
->is_vector() ||
1472 ir
->operands
[1]->type
->is_vector()) {
1473 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_Z
));
1474 emit(MOV(result_dst
, src_reg(0)));
1475 inst
= emit(MOV(result_dst
, src_reg(1)));
1476 inst
->predicate
= BRW_PREDICATE_ALIGN16_ALL4H
;
1478 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_Z
));
1479 emit(AND(result_dst
, result_src
, src_reg(0x1)));
1482 case ir_binop_any_nequal
:
1483 /* "!=" operator producing a scalar boolean. */
1484 if (ir
->operands
[0]->type
->is_vector() ||
1485 ir
->operands
[1]->type
->is_vector()) {
1486 emit(CMP(dst_null_d(), op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1488 emit(MOV(result_dst
, src_reg(0)));
1489 inst
= emit(MOV(result_dst
, src_reg(1)));
1490 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1492 emit(CMP(result_dst
, op
[0], op
[1], BRW_CONDITIONAL_NZ
));
1493 emit(AND(result_dst
, result_src
, src_reg(0x1)));
1498 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1499 emit(MOV(result_dst
, src_reg(0)));
1501 inst
= emit(MOV(result_dst
, src_reg(1)));
1502 inst
->predicate
= BRW_PREDICATE_ALIGN16_ANY4H
;
1505 case ir_binop_logic_xor
:
1506 emit(XOR(result_dst
, op
[0], op
[1]));
1509 case ir_binop_logic_or
:
1510 emit(OR(result_dst
, op
[0], op
[1]));
1513 case ir_binop_logic_and
:
1514 emit(AND(result_dst
, op
[0], op
[1]));
1518 assert(ir
->operands
[0]->type
->is_vector());
1519 assert(ir
->operands
[0]->type
== ir
->operands
[1]->type
);
1520 emit_dp(result_dst
, op
[0], op
[1], ir
->operands
[0]->type
->vector_elements
);
1524 emit_math(SHADER_OPCODE_SQRT
, result_dst
, op
[0]);
1527 emit_math(SHADER_OPCODE_RSQ
, result_dst
, op
[0]);
1530 case ir_unop_bitcast_i2f
:
1531 case ir_unop_bitcast_u2f
:
1532 this->result
= op
[0];
1533 this->result
.type
= BRW_REGISTER_TYPE_F
;
1536 case ir_unop_bitcast_f2i
:
1537 this->result
= op
[0];
1538 this->result
.type
= BRW_REGISTER_TYPE_D
;
1541 case ir_unop_bitcast_f2u
:
1542 this->result
= op
[0];
1543 this->result
.type
= BRW_REGISTER_TYPE_UD
;
1554 emit(MOV(result_dst
, op
[0]));
1558 emit(CMP(result_dst
, op
[0], src_reg(0.0f
), BRW_CONDITIONAL_NZ
));
1559 emit(AND(result_dst
, result_src
, src_reg(1)));
1564 emit(RNDZ(result_dst
, op
[0]));
1567 op
[0].negate
= !op
[0].negate
;
1568 inst
= emit(RNDD(result_dst
, op
[0]));
1569 this->result
.negate
= true;
1572 inst
= emit(RNDD(result_dst
, op
[0]));
1575 inst
= emit(FRC(result_dst
, op
[0]));
1577 case ir_unop_round_even
:
1578 emit(RNDE(result_dst
, op
[0]));
1582 emit_minmax(BRW_CONDITIONAL_L
, result_dst
, op
[0], op
[1]);
1585 emit_minmax(BRW_CONDITIONAL_G
, result_dst
, op
[0], op
[1]);
1589 emit_math(SHADER_OPCODE_POW
, result_dst
, op
[0], op
[1]);
1592 case ir_unop_bit_not
:
1593 inst
= emit(NOT(result_dst
, op
[0]));
1595 case ir_binop_bit_and
:
1596 inst
= emit(AND(result_dst
, op
[0], op
[1]));
1598 case ir_binop_bit_xor
:
1599 inst
= emit(XOR(result_dst
, op
[0], op
[1]));
1601 case ir_binop_bit_or
:
1602 inst
= emit(OR(result_dst
, op
[0], op
[1]));
1605 case ir_binop_lshift
:
1606 inst
= emit(SHL(result_dst
, op
[0], op
[1]));
1609 case ir_binop_rshift
:
1610 if (ir
->type
->base_type
== GLSL_TYPE_INT
)
1611 inst
= emit(ASR(result_dst
, op
[0], op
[1]));
1613 inst
= emit(SHR(result_dst
, op
[0], op
[1]));
1617 emit(BFI1(result_dst
, op
[0], op
[1]));
1620 case ir_binop_ubo_load
: {
1621 ir_constant
*uniform_block
= ir
->operands
[0]->as_constant();
1622 ir_constant
*const_offset_ir
= ir
->operands
[1]->as_constant();
1623 unsigned const_offset
= const_offset_ir
? const_offset_ir
->value
.u
[0] : 0;
1626 /* Now, load the vector from that offset. */
1627 assert(ir
->type
->is_vector() || ir
->type
->is_scalar());
1629 src_reg packed_consts
= src_reg(this, glsl_type::vec4_type
);
1630 packed_consts
.type
= result
.type
;
1631 src_reg surf_index
=
1632 src_reg(prog_data
->base
.binding_table
.ubo_start
+ uniform_block
->value
.u
[0]);
1633 if (const_offset_ir
) {
1634 if (brw
->gen
>= 8) {
1635 /* Store the offset in a GRF so we can send-from-GRF. */
1636 offset
= src_reg(this, glsl_type::int_type
);
1637 emit(MOV(dst_reg(offset
), src_reg(const_offset
/ 16)));
1639 /* Immediates are fine on older generations since they'll be moved
1640 * to a (potentially fake) MRF at the generator level.
1642 offset
= src_reg(const_offset
/ 16);
1645 offset
= src_reg(this, glsl_type::uint_type
);
1646 emit(SHR(dst_reg(offset
), op
[1], src_reg(4)));
1649 if (brw
->gen
>= 7) {
1650 dst_reg grf_offset
= dst_reg(this, glsl_type::int_type
);
1651 grf_offset
.type
= offset
.type
;
1653 emit(MOV(grf_offset
, offset
));
1655 emit(new(mem_ctx
) vec4_instruction(this,
1656 VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
1657 dst_reg(packed_consts
),
1659 src_reg(grf_offset
)));
1661 vec4_instruction
*pull
=
1662 emit(new(mem_ctx
) vec4_instruction(this,
1663 VS_OPCODE_PULL_CONSTANT_LOAD
,
1664 dst_reg(packed_consts
),
1667 pull
->base_mrf
= 14;
1671 packed_consts
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1672 packed_consts
.swizzle
+= BRW_SWIZZLE4(const_offset
% 16 / 4,
1673 const_offset
% 16 / 4,
1674 const_offset
% 16 / 4,
1675 const_offset
% 16 / 4);
1677 /* UBO bools are any nonzero int. We store bools as either 0 or 1. */
1678 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
1679 emit(CMP(result_dst
, packed_consts
, src_reg(0u),
1680 BRW_CONDITIONAL_NZ
));
1681 emit(AND(result_dst
, result
, src_reg(0x1)));
1683 emit(MOV(result_dst
, packed_consts
));
1688 case ir_binop_vector_extract
:
1689 unreachable("should have been lowered by vec_index_to_cond_assign");
1692 op
[0] = fix_3src_operand(op
[0]);
1693 op
[1] = fix_3src_operand(op
[1]);
1694 op
[2] = fix_3src_operand(op
[2]);
1695 /* Note that the instruction's argument order is reversed from GLSL
1698 emit(MAD(result_dst
, op
[2], op
[1], op
[0]));
1702 emit_lrp(result_dst
, op
[0], op
[1], op
[2]);
1706 emit(CMP(dst_null_d(), op
[0], src_reg(0), BRW_CONDITIONAL_NZ
));
1707 inst
= emit(BRW_OPCODE_SEL
, result_dst
, op
[1], op
[2]);
1708 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1712 op
[0] = fix_3src_operand(op
[0]);
1713 op
[1] = fix_3src_operand(op
[1]);
1714 op
[2] = fix_3src_operand(op
[2]);
1715 emit(BFI2(result_dst
, op
[0], op
[1], op
[2]));
1718 case ir_triop_bitfield_extract
:
1719 op
[0] = fix_3src_operand(op
[0]);
1720 op
[1] = fix_3src_operand(op
[1]);
1721 op
[2] = fix_3src_operand(op
[2]);
1722 /* Note that the instruction's argument order is reversed from GLSL
1725 emit(BFE(result_dst
, op
[2], op
[1], op
[0]));
1728 case ir_triop_vector_insert
:
1729 unreachable("should have been lowered by lower_vector_insert");
1731 case ir_quadop_bitfield_insert
:
1732 unreachable("not reached: should be handled by "
1733 "bitfield_insert_to_bfm_bfi\n");
1735 case ir_quadop_vector
:
1736 unreachable("not reached: should be handled by lower_quadop_vector");
1738 case ir_unop_pack_half_2x16
:
1739 emit_pack_half_2x16(result_dst
, op
[0]);
1741 case ir_unop_unpack_half_2x16
:
1742 emit_unpack_half_2x16(result_dst
, op
[0]);
1744 case ir_unop_pack_snorm_2x16
:
1745 case ir_unop_pack_snorm_4x8
:
1746 case ir_unop_pack_unorm_2x16
:
1747 case ir_unop_pack_unorm_4x8
:
1748 case ir_unop_unpack_snorm_2x16
:
1749 case ir_unop_unpack_snorm_4x8
:
1750 case ir_unop_unpack_unorm_2x16
:
1751 case ir_unop_unpack_unorm_4x8
:
1752 unreachable("not reached: should be handled by lower_packing_builtins");
1753 case ir_unop_unpack_half_2x16_split_x
:
1754 case ir_unop_unpack_half_2x16_split_y
:
1755 case ir_binop_pack_half_2x16_split
:
1756 case ir_unop_interpolate_at_centroid
:
1757 case ir_binop_interpolate_at_sample
:
1758 case ir_binop_interpolate_at_offset
:
1759 unreachable("not reached: should not occur in vertex shader");
1760 case ir_binop_ldexp
:
1761 unreachable("not reached: should be handled by ldexp_to_arith()");
1767 vec4_visitor::visit(ir_swizzle
*ir
)
1773 /* Note that this is only swizzles in expressions, not those on the left
1774 * hand side of an assignment, which do write masking. See ir_assignment
1778 ir
->val
->accept(this);
1780 assert(src
.file
!= BAD_FILE
);
1782 for (i
= 0; i
< ir
->type
->vector_elements
; i
++) {
1785 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.x
);
1788 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.y
);
1791 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.z
);
1794 swizzle
[i
] = BRW_GET_SWZ(src
.swizzle
, ir
->mask
.w
);
1798 for (; i
< 4; i
++) {
1799 /* Replicate the last channel out. */
1800 swizzle
[i
] = swizzle
[ir
->type
->vector_elements
- 1];
1803 src
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
1809 vec4_visitor::visit(ir_dereference_variable
*ir
)
1811 const struct glsl_type
*type
= ir
->type
;
1812 dst_reg
*reg
= variable_storage(ir
->var
);
1815 fail("Failed to find variable storage for %s\n", ir
->var
->name
);
1816 this->result
= src_reg(brw_null_reg());
1820 this->result
= src_reg(*reg
);
1822 /* System values get their swizzle from the dst_reg writemask */
1823 if (ir
->var
->data
.mode
== ir_var_system_value
)
1826 if (type
->is_scalar() || type
->is_vector() || type
->is_matrix())
1827 this->result
.swizzle
= swizzle_for_size(type
->vector_elements
);
1832 vec4_visitor::compute_array_stride(ir_dereference_array
*ir
)
1834 /* Under normal circumstances array elements are stored consecutively, so
1835 * the stride is equal to the size of the array element.
1837 return type_size(ir
->type
);
1842 vec4_visitor::visit(ir_dereference_array
*ir
)
1844 ir_constant
*constant_index
;
1846 int array_stride
= compute_array_stride(ir
);
1848 constant_index
= ir
->array_index
->constant_expression_value();
1850 ir
->array
->accept(this);
1853 if (constant_index
) {
1854 src
.reg_offset
+= constant_index
->value
.i
[0] * array_stride
;
1856 /* Variable index array dereference. It eats the "vec4" of the
1857 * base of the array and an index that offsets the Mesa register
1860 ir
->array_index
->accept(this);
1864 if (array_stride
== 1) {
1865 index_reg
= this->result
;
1867 index_reg
= src_reg(this, glsl_type::int_type
);
1869 emit(MUL(dst_reg(index_reg
), this->result
, src_reg(array_stride
)));
1873 src_reg temp
= src_reg(this, glsl_type::int_type
);
1875 emit(ADD(dst_reg(temp
), *src
.reladdr
, index_reg
));
1880 src
.reladdr
= ralloc(mem_ctx
, src_reg
);
1881 memcpy(src
.reladdr
, &index_reg
, sizeof(index_reg
));
1884 /* If the type is smaller than a vec4, replicate the last channel out. */
1885 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
1886 src
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1888 src
.swizzle
= BRW_SWIZZLE_NOOP
;
1889 src
.type
= brw_type_for_base_type(ir
->type
);
1895 vec4_visitor::visit(ir_dereference_record
*ir
)
1898 const glsl_type
*struct_type
= ir
->record
->type
;
1901 ir
->record
->accept(this);
1903 for (i
= 0; i
< struct_type
->length
; i
++) {
1904 if (strcmp(struct_type
->fields
.structure
[i
].name
, ir
->field
) == 0)
1906 offset
+= type_size(struct_type
->fields
.structure
[i
].type
);
1909 /* If the type is smaller than a vec4, replicate the last channel out. */
1910 if (ir
->type
->is_scalar() || ir
->type
->is_vector() || ir
->type
->is_matrix())
1911 this->result
.swizzle
= swizzle_for_size(ir
->type
->vector_elements
);
1913 this->result
.swizzle
= BRW_SWIZZLE_NOOP
;
1914 this->result
.type
= brw_type_for_base_type(ir
->type
);
1916 this->result
.reg_offset
+= offset
;
1920 * We want to be careful in assignment setup to hit the actual storage
1921 * instead of potentially using a temporary like we might with the
1922 * ir_dereference handler.
1925 get_assignment_lhs(ir_dereference
*ir
, vec4_visitor
*v
)
1927 /* The LHS must be a dereference. If the LHS is a variable indexed array
1928 * access of a vector, it must be separated into a series conditional moves
1929 * before reaching this point (see ir_vec_index_to_cond_assign).
1931 assert(ir
->as_dereference());
1932 ir_dereference_array
*deref_array
= ir
->as_dereference_array();
1934 assert(!deref_array
->array
->type
->is_vector());
1937 /* Use the rvalue deref handler for the most part. We'll ignore
1938 * swizzles in it and write swizzles using writemask, though.
1941 return dst_reg(v
->result
);
1945 vec4_visitor::emit_block_move(dst_reg
*dst
, src_reg
*src
,
1946 const struct glsl_type
*type
,
1947 enum brw_predicate predicate
)
1949 if (type
->base_type
== GLSL_TYPE_STRUCT
) {
1950 for (unsigned int i
= 0; i
< type
->length
; i
++) {
1951 emit_block_move(dst
, src
, type
->fields
.structure
[i
].type
, predicate
);
1956 if (type
->is_array()) {
1957 for (unsigned int i
= 0; i
< type
->length
; i
++) {
1958 emit_block_move(dst
, src
, type
->fields
.array
, predicate
);
1963 if (type
->is_matrix()) {
1964 const struct glsl_type
*vec_type
;
1966 vec_type
= glsl_type::get_instance(GLSL_TYPE_FLOAT
,
1967 type
->vector_elements
, 1);
1969 for (int i
= 0; i
< type
->matrix_columns
; i
++) {
1970 emit_block_move(dst
, src
, vec_type
, predicate
);
1975 assert(type
->is_scalar() || type
->is_vector());
1977 dst
->type
= brw_type_for_base_type(type
);
1978 src
->type
= dst
->type
;
1980 dst
->writemask
= (1 << type
->vector_elements
) - 1;
1982 src
->swizzle
= swizzle_for_size(type
->vector_elements
);
1984 vec4_instruction
*inst
= emit(MOV(*dst
, *src
));
1985 inst
->predicate
= predicate
;
1992 /* If the RHS processing resulted in an instruction generating a
1993 * temporary value, and it would be easy to rewrite the instruction to
1994 * generate its result right into the LHS instead, do so. This ends
1995 * up reliably removing instructions where it can be tricky to do so
1996 * later without real UD chain information.
1999 vec4_visitor::try_rewrite_rhs_to_dst(ir_assignment
*ir
,
2002 vec4_instruction
*pre_rhs_inst
,
2003 vec4_instruction
*last_rhs_inst
)
2005 /* This could be supported, but it would take more smarts. */
2009 if (pre_rhs_inst
== last_rhs_inst
)
2010 return false; /* No instructions generated to work with. */
2012 /* Make sure the last instruction generated our source reg. */
2013 if (src
.file
!= GRF
||
2014 src
.file
!= last_rhs_inst
->dst
.file
||
2015 src
.reg
!= last_rhs_inst
->dst
.reg
||
2016 src
.reg_offset
!= last_rhs_inst
->dst
.reg_offset
||
2020 last_rhs_inst
->predicate
!= BRW_PREDICATE_NONE
)
2023 /* Check that that last instruction fully initialized the channels
2024 * we want to use, in the order we want to use them. We could
2025 * potentially reswizzle the operands of many instructions so that
2026 * we could handle out of order channels, but don't yet.
2029 for (unsigned i
= 0; i
< 4; i
++) {
2030 if (dst
.writemask
& (1 << i
)) {
2031 if (!(last_rhs_inst
->dst
.writemask
& (1 << i
)))
2034 if (BRW_GET_SWZ(src
.swizzle
, i
) != i
)
2039 /* Success! Rewrite the instruction. */
2040 last_rhs_inst
->dst
.file
= dst
.file
;
2041 last_rhs_inst
->dst
.reg
= dst
.reg
;
2042 last_rhs_inst
->dst
.reg_offset
= dst
.reg_offset
;
2043 last_rhs_inst
->dst
.reladdr
= dst
.reladdr
;
2044 last_rhs_inst
->dst
.writemask
&= dst
.writemask
;
2050 vec4_visitor::visit(ir_assignment
*ir
)
2052 dst_reg dst
= get_assignment_lhs(ir
->lhs
, this);
2053 enum brw_predicate predicate
= BRW_PREDICATE_NONE
;
2055 if (!ir
->lhs
->type
->is_scalar() &&
2056 !ir
->lhs
->type
->is_vector()) {
2057 ir
->rhs
->accept(this);
2058 src_reg src
= this->result
;
2060 if (ir
->condition
) {
2061 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2064 /* emit_block_move doesn't account for swizzles in the source register.
2065 * This should be ok, since the source register is a structure or an
2066 * array, and those can't be swizzled. But double-check to be sure.
2068 assert(src
.swizzle
==
2069 (ir
->rhs
->type
->is_matrix()
2070 ? swizzle_for_size(ir
->rhs
->type
->vector_elements
)
2071 : BRW_SWIZZLE_NOOP
));
2073 emit_block_move(&dst
, &src
, ir
->rhs
->type
, predicate
);
2077 /* Now we're down to just a scalar/vector with writemasks. */
2080 vec4_instruction
*pre_rhs_inst
, *last_rhs_inst
;
2081 pre_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2083 ir
->rhs
->accept(this);
2085 last_rhs_inst
= (vec4_instruction
*)this->instructions
.get_tail();
2087 src_reg src
= this->result
;
2090 int first_enabled_chan
= 0;
2093 assert(ir
->lhs
->type
->is_vector() ||
2094 ir
->lhs
->type
->is_scalar());
2095 dst
.writemask
= ir
->write_mask
;
2097 for (int i
= 0; i
< 4; i
++) {
2098 if (dst
.writemask
& (1 << i
)) {
2099 first_enabled_chan
= BRW_GET_SWZ(src
.swizzle
, i
);
2104 /* Swizzle a small RHS vector into the channels being written.
2106 * glsl ir treats write_mask as dictating how many channels are
2107 * present on the RHS while in our instructions we need to make
2108 * those channels appear in the slots of the vec4 they're written to.
2110 for (int i
= 0; i
< 4; i
++) {
2111 if (dst
.writemask
& (1 << i
))
2112 swizzles
[i
] = BRW_GET_SWZ(src
.swizzle
, src_chan
++);
2114 swizzles
[i
] = first_enabled_chan
;
2116 src
.swizzle
= BRW_SWIZZLE4(swizzles
[0], swizzles
[1],
2117 swizzles
[2], swizzles
[3]);
2119 if (try_rewrite_rhs_to_dst(ir
, dst
, src
, pre_rhs_inst
, last_rhs_inst
)) {
2123 if (ir
->condition
) {
2124 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2127 for (i
= 0; i
< type_size(ir
->lhs
->type
); i
++) {
2128 vec4_instruction
*inst
= emit(MOV(dst
, src
));
2129 inst
->predicate
= predicate
;
2137 vec4_visitor::emit_constant_values(dst_reg
*dst
, ir_constant
*ir
)
2139 if (ir
->type
->base_type
== GLSL_TYPE_STRUCT
) {
2140 foreach_in_list(ir_constant
, field_value
, &ir
->components
) {
2141 emit_constant_values(dst
, field_value
);
2146 if (ir
->type
->is_array()) {
2147 for (unsigned int i
= 0; i
< ir
->type
->length
; i
++) {
2148 emit_constant_values(dst
, ir
->array_elements
[i
]);
2153 if (ir
->type
->is_matrix()) {
2154 for (int i
= 0; i
< ir
->type
->matrix_columns
; i
++) {
2155 float *vec
= &ir
->value
.f
[i
* ir
->type
->vector_elements
];
2157 for (int j
= 0; j
< ir
->type
->vector_elements
; j
++) {
2158 dst
->writemask
= 1 << j
;
2159 dst
->type
= BRW_REGISTER_TYPE_F
;
2161 emit(MOV(*dst
, src_reg(vec
[j
])));
2168 int remaining_writemask
= (1 << ir
->type
->vector_elements
) - 1;
2170 for (int i
= 0; i
< ir
->type
->vector_elements
; i
++) {
2171 if (!(remaining_writemask
& (1 << i
)))
2174 dst
->writemask
= 1 << i
;
2175 dst
->type
= brw_type_for_base_type(ir
->type
);
2177 /* Find other components that match the one we're about to
2178 * write. Emits fewer instructions for things like vec4(0.5,
2181 for (int j
= i
+ 1; j
< ir
->type
->vector_elements
; j
++) {
2182 if (ir
->type
->base_type
== GLSL_TYPE_BOOL
) {
2183 if (ir
->value
.b
[i
] == ir
->value
.b
[j
])
2184 dst
->writemask
|= (1 << j
);
2186 /* u, i, and f storage all line up, so no need for a
2187 * switch case for comparing each type.
2189 if (ir
->value
.u
[i
] == ir
->value
.u
[j
])
2190 dst
->writemask
|= (1 << j
);
2194 switch (ir
->type
->base_type
) {
2195 case GLSL_TYPE_FLOAT
:
2196 emit(MOV(*dst
, src_reg(ir
->value
.f
[i
])));
2199 emit(MOV(*dst
, src_reg(ir
->value
.i
[i
])));
2201 case GLSL_TYPE_UINT
:
2202 emit(MOV(*dst
, src_reg(ir
->value
.u
[i
])));
2204 case GLSL_TYPE_BOOL
:
2205 emit(MOV(*dst
, src_reg(ir
->value
.b
[i
])));
2208 unreachable("Non-float/uint/int/bool constant");
2211 remaining_writemask
&= ~dst
->writemask
;
2217 vec4_visitor::visit(ir_constant
*ir
)
2219 dst_reg dst
= dst_reg(this, ir
->type
);
2220 this->result
= src_reg(dst
);
2222 emit_constant_values(&dst
, ir
);
2226 vec4_visitor::visit_atomic_counter_intrinsic(ir_call
*ir
)
2228 ir_dereference
*deref
= static_cast<ir_dereference
*>(
2229 ir
->actual_parameters
.get_head());
2230 ir_variable
*location
= deref
->variable_referenced();
2231 unsigned surf_index
= (prog_data
->base
.binding_table
.abo_start
+
2232 location
->data
.atomic
.buffer_index
);
2234 /* Calculate the surface offset */
2235 src_reg
offset(this, glsl_type::uint_type
);
2236 ir_dereference_array
*deref_array
= deref
->as_dereference_array();
2238 deref_array
->array_index
->accept(this);
2240 src_reg
tmp(this, glsl_type::uint_type
);
2241 emit(MUL(dst_reg(tmp
), this->result
, ATOMIC_COUNTER_SIZE
));
2242 emit(ADD(dst_reg(offset
), tmp
, location
->data
.atomic
.offset
));
2244 offset
= location
->data
.atomic
.offset
;
2247 /* Emit the appropriate machine instruction */
2248 const char *callee
= ir
->callee
->function_name();
2249 dst_reg dst
= get_assignment_lhs(ir
->return_deref
, this);
2251 if (!strcmp("__intrinsic_atomic_read", callee
)) {
2252 emit_untyped_surface_read(surf_index
, dst
, offset
);
2254 } else if (!strcmp("__intrinsic_atomic_increment", callee
)) {
2255 emit_untyped_atomic(BRW_AOP_INC
, surf_index
, dst
, offset
,
2256 src_reg(), src_reg());
2258 } else if (!strcmp("__intrinsic_atomic_predecrement", callee
)) {
2259 emit_untyped_atomic(BRW_AOP_PREDEC
, surf_index
, dst
, offset
,
2260 src_reg(), src_reg());
2265 vec4_visitor::visit(ir_call
*ir
)
2267 const char *callee
= ir
->callee
->function_name();
2269 if (!strcmp("__intrinsic_atomic_read", callee
) ||
2270 !strcmp("__intrinsic_atomic_increment", callee
) ||
2271 !strcmp("__intrinsic_atomic_predecrement", callee
)) {
2272 visit_atomic_counter_intrinsic(ir
);
2274 unreachable("Unsupported intrinsic.");
2279 vec4_visitor::emit_mcs_fetch(ir_texture
*ir
, src_reg coordinate
, uint32_t sampler
)
2281 vec4_instruction
*inst
= new(mem_ctx
) vec4_instruction(this, SHADER_OPCODE_TXF_MCS
);
2284 inst
->dst
= dst_reg(this, glsl_type::uvec4_type
);
2285 inst
->dst
.writemask
= WRITEMASK_XYZW
;
2287 inst
->src
[1] = src_reg(sampler
);
2289 /* parameters are: u, v, r, lod; lod will always be zero due to api restrictions */
2290 int param_base
= inst
->base_mrf
;
2291 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2292 int zero_mask
= 0xf & ~coord_mask
;
2294 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2297 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2301 return src_reg(inst
->dst
);
2305 vec4_visitor::visit(ir_texture
*ir
)
2308 _mesa_get_sampler_uniform_value(ir
->sampler
, shader_prog
, prog
);
2310 /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
2311 * emitting anything other than setting up the constant result.
2313 if (ir
->op
== ir_tg4
) {
2314 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2315 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2316 if (swiz
== SWIZZLE_ZERO
|| swiz
== SWIZZLE_ONE
) {
2317 dst_reg
result(this, ir
->type
);
2318 this->result
= src_reg(result
);
2319 emit(MOV(result
, src_reg(swiz
== SWIZZLE_ONE
? 1.0f
: 0.0f
)));
2324 /* Should be lowered by do_lower_texture_projection */
2325 assert(!ir
->projector
);
2327 /* Should be lowered */
2328 assert(!ir
->offset
|| !ir
->offset
->type
->is_array());
2330 /* Generate code to compute all the subexpression trees. This has to be
2331 * done before loading any values into MRFs for the sampler message since
2332 * generating these values may involve SEND messages that need the MRFs.
2335 if (ir
->coordinate
) {
2336 ir
->coordinate
->accept(this);
2337 coordinate
= this->result
;
2340 src_reg shadow_comparitor
;
2341 if (ir
->shadow_comparitor
) {
2342 ir
->shadow_comparitor
->accept(this);
2343 shadow_comparitor
= this->result
;
2346 bool has_nonconstant_offset
= ir
->offset
&& !ir
->offset
->as_constant();
2347 src_reg offset_value
;
2348 if (has_nonconstant_offset
) {
2349 ir
->offset
->accept(this);
2350 offset_value
= src_reg(this->result
);
2353 const glsl_type
*lod_type
= NULL
, *sample_index_type
= NULL
;
2354 src_reg lod
, dPdx
, dPdy
, sample_index
, mcs
;
2357 lod
= src_reg(0.0f
);
2358 lod_type
= glsl_type::float_type
;
2363 ir
->lod_info
.lod
->accept(this);
2365 lod_type
= ir
->lod_info
.lod
->type
;
2367 case ir_query_levels
:
2369 lod_type
= glsl_type::int_type
;
2372 ir
->lod_info
.sample_index
->accept(this);
2373 sample_index
= this->result
;
2374 sample_index_type
= ir
->lod_info
.sample_index
->type
;
2376 if (brw
->gen
>= 7 && key
->tex
.compressed_multisample_layout_mask
& (1<<sampler
))
2377 mcs
= emit_mcs_fetch(ir
, coordinate
, sampler
);
2382 ir
->lod_info
.grad
.dPdx
->accept(this);
2383 dPdx
= this->result
;
2385 ir
->lod_info
.grad
.dPdy
->accept(this);
2386 dPdy
= this->result
;
2388 lod_type
= ir
->lod_info
.grad
.dPdx
->type
;
2398 case ir_tex
: opcode
= SHADER_OPCODE_TXL
; break;
2399 case ir_txl
: opcode
= SHADER_OPCODE_TXL
; break;
2400 case ir_txd
: opcode
= SHADER_OPCODE_TXD
; break;
2401 case ir_txf
: opcode
= SHADER_OPCODE_TXF
; break;
2402 case ir_txf_ms
: opcode
= SHADER_OPCODE_TXF_CMS
; break;
2403 case ir_txs
: opcode
= SHADER_OPCODE_TXS
; break;
2404 case ir_tg4
: opcode
= has_nonconstant_offset
2405 ? SHADER_OPCODE_TG4_OFFSET
: SHADER_OPCODE_TG4
; break;
2406 case ir_query_levels
: opcode
= SHADER_OPCODE_TXS
; break;
2408 unreachable("TXB is not valid for vertex shaders.");
2410 unreachable("LOD is not valid for vertex shaders.");
2412 unreachable("Unrecognized tex op");
2415 vec4_instruction
*inst
= new(mem_ctx
) vec4_instruction(this, opcode
);
2417 if (ir
->offset
!= NULL
&& ir
->op
!= ir_txf
)
2418 inst
->texture_offset
= brw_texture_offset(ctx
, ir
->offset
->as_constant());
2420 /* Stuff the channel select bits in the top of the texture offset */
2421 if (ir
->op
== ir_tg4
)
2422 inst
->texture_offset
|= gather_channel(ir
, sampler
) << 16;
2424 /* The message header is necessary for:
2427 * - Gather channel selection
2428 * - Sampler indices too large to fit in a 4-bit value.
2430 inst
->header_present
=
2431 brw
->gen
< 5 || inst
->texture_offset
!= 0 || ir
->op
== ir_tg4
||
2434 inst
->mlen
= inst
->header_present
+ 1; /* always at least one */
2435 inst
->dst
= dst_reg(this, ir
->type
);
2436 inst
->dst
.writemask
= WRITEMASK_XYZW
;
2437 inst
->shadow_compare
= ir
->shadow_comparitor
!= NULL
;
2439 inst
->src
[1] = src_reg(sampler
);
2441 /* MRF for the first parameter */
2442 int param_base
= inst
->base_mrf
+ inst
->header_present
;
2444 if (ir
->op
== ir_txs
|| ir
->op
== ir_query_levels
) {
2445 int writemask
= brw
->gen
== 4 ? WRITEMASK_W
: WRITEMASK_X
;
2446 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, writemask
), lod
));
2448 /* Load the coordinate */
2449 /* FINISHME: gl_clamp_mask and saturate */
2450 int coord_mask
= (1 << ir
->coordinate
->type
->vector_elements
) - 1;
2451 int zero_mask
= 0xf & ~coord_mask
;
2453 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, coord_mask
),
2456 if (zero_mask
!= 0) {
2457 emit(MOV(dst_reg(MRF
, param_base
, ir
->coordinate
->type
, zero_mask
),
2460 /* Load the shadow comparitor */
2461 if (ir
->shadow_comparitor
&& ir
->op
!= ir_txd
&& (ir
->op
!= ir_tg4
|| !has_nonconstant_offset
)) {
2462 emit(MOV(dst_reg(MRF
, param_base
+ 1, ir
->shadow_comparitor
->type
,
2464 shadow_comparitor
));
2468 /* Load the LOD info */
2469 if (ir
->op
== ir_tex
|| ir
->op
== ir_txl
) {
2471 if (brw
->gen
>= 5) {
2472 mrf
= param_base
+ 1;
2473 if (ir
->shadow_comparitor
) {
2474 writemask
= WRITEMASK_Y
;
2475 /* mlen already incremented */
2477 writemask
= WRITEMASK_X
;
2480 } else /* brw->gen == 4 */ {
2482 writemask
= WRITEMASK_W
;
2484 emit(MOV(dst_reg(MRF
, mrf
, lod_type
, writemask
), lod
));
2485 } else if (ir
->op
== ir_txf
) {
2486 emit(MOV(dst_reg(MRF
, param_base
, lod_type
, WRITEMASK_W
), lod
));
2487 } else if (ir
->op
== ir_txf_ms
) {
2488 emit(MOV(dst_reg(MRF
, param_base
+ 1, sample_index_type
, WRITEMASK_X
),
2491 /* MCS data is in the first channel of `mcs`, but we need to get it into
2492 * the .y channel of the second vec4 of params, so replicate .x across
2493 * the whole vec4 and then mask off everything except .y
2495 mcs
.swizzle
= BRW_SWIZZLE_XXXX
;
2496 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::uint_type
, WRITEMASK_Y
),
2499 } else if (ir
->op
== ir_txd
) {
2500 const glsl_type
*type
= lod_type
;
2502 if (brw
->gen
>= 5) {
2503 dPdx
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2504 dPdy
.swizzle
= BRW_SWIZZLE4(SWIZZLE_X
,SWIZZLE_X
,SWIZZLE_Y
,SWIZZLE_Y
);
2505 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XZ
), dPdx
));
2506 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_YW
), dPdy
));
2509 if (ir
->type
->vector_elements
== 3 || ir
->shadow_comparitor
) {
2510 dPdx
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2511 dPdy
.swizzle
= BRW_SWIZZLE_ZZZZ
;
2512 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_X
), dPdx
));
2513 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_Y
), dPdy
));
2516 if (ir
->shadow_comparitor
) {
2517 emit(MOV(dst_reg(MRF
, param_base
+ 2,
2518 ir
->shadow_comparitor
->type
, WRITEMASK_Z
),
2519 shadow_comparitor
));
2522 } else /* brw->gen == 4 */ {
2523 emit(MOV(dst_reg(MRF
, param_base
+ 1, type
, WRITEMASK_XYZ
), dPdx
));
2524 emit(MOV(dst_reg(MRF
, param_base
+ 2, type
, WRITEMASK_XYZ
), dPdy
));
2527 } else if (ir
->op
== ir_tg4
&& has_nonconstant_offset
) {
2528 if (ir
->shadow_comparitor
) {
2529 emit(MOV(dst_reg(MRF
, param_base
, ir
->shadow_comparitor
->type
, WRITEMASK_W
),
2530 shadow_comparitor
));
2533 emit(MOV(dst_reg(MRF
, param_base
+ 1, glsl_type::ivec2_type
, WRITEMASK_XY
),
2541 /* fixup num layers (z) for cube arrays: hardware returns faces * layers;
2542 * spec requires layers.
2544 if (ir
->op
== ir_txs
) {
2545 glsl_type
const *type
= ir
->sampler
->type
;
2546 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
2547 type
->sampler_array
) {
2548 emit_math(SHADER_OPCODE_INT_QUOTIENT
,
2549 writemask(inst
->dst
, WRITEMASK_Z
),
2550 src_reg(inst
->dst
), src_reg(6));
2554 if (brw
->gen
== 6 && ir
->op
== ir_tg4
) {
2555 emit_gen6_gather_wa(key
->tex
.gen6_gather_wa
[sampler
], inst
->dst
);
2558 swizzle_result(ir
, src_reg(inst
->dst
), sampler
);
2562 * Apply workarounds for Gen6 gather with UINT/SINT
2565 vec4_visitor::emit_gen6_gather_wa(uint8_t wa
, dst_reg dst
)
2570 int width
= (wa
& WA_8BIT
) ? 8 : 16;
2571 dst_reg dst_f
= dst
;
2572 dst_f
.type
= BRW_REGISTER_TYPE_F
;
2574 /* Convert from UNORM to UINT */
2575 emit(MUL(dst_f
, src_reg(dst_f
), src_reg((float)((1 << width
) - 1))));
2576 emit(MOV(dst
, src_reg(dst_f
)));
2579 /* Reinterpret the UINT value as a signed INT value by
2580 * shifting the sign bit into place, then shifting back
2583 emit(SHL(dst
, src_reg(dst
), src_reg(32 - width
)));
2584 emit(ASR(dst
, src_reg(dst
), src_reg(32 - width
)));
2589 * Set up the gather channel based on the swizzle, for gather4.
2592 vec4_visitor::gather_channel(ir_texture
*ir
, uint32_t sampler
)
2594 ir_constant
*chan
= ir
->lod_info
.component
->as_constant();
2595 int swiz
= GET_SWZ(key
->tex
.swizzles
[sampler
], chan
->value
.i
[0]);
2597 case SWIZZLE_X
: return 0;
2599 /* gather4 sampler is broken for green channel on RG32F --
2600 * we must ask for blue instead.
2602 if (key
->tex
.gather_channel_quirk_mask
& (1<<sampler
))
2605 case SWIZZLE_Z
: return 2;
2606 case SWIZZLE_W
: return 3;
2608 unreachable("Not reached"); /* zero, one swizzles handled already */
2613 vec4_visitor::swizzle_result(ir_texture
*ir
, src_reg orig_val
, uint32_t sampler
)
2615 int s
= key
->tex
.swizzles
[sampler
];
2617 this->result
= src_reg(this, ir
->type
);
2618 dst_reg
swizzled_result(this->result
);
2620 if (ir
->op
== ir_query_levels
) {
2621 /* # levels is in .w */
2622 orig_val
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
2623 emit(MOV(swizzled_result
, orig_val
));
2627 if (ir
->op
== ir_txs
|| ir
->type
== glsl_type::float_type
2628 || s
== SWIZZLE_NOOP
|| ir
->op
== ir_tg4
) {
2629 emit(MOV(swizzled_result
, orig_val
));
2634 int zero_mask
= 0, one_mask
= 0, copy_mask
= 0;
2635 int swizzle
[4] = {0};
2637 for (int i
= 0; i
< 4; i
++) {
2638 switch (GET_SWZ(s
, i
)) {
2640 zero_mask
|= (1 << i
);
2643 one_mask
|= (1 << i
);
2646 copy_mask
|= (1 << i
);
2647 swizzle
[i
] = GET_SWZ(s
, i
);
2653 orig_val
.swizzle
= BRW_SWIZZLE4(swizzle
[0], swizzle
[1], swizzle
[2], swizzle
[3]);
2654 swizzled_result
.writemask
= copy_mask
;
2655 emit(MOV(swizzled_result
, orig_val
));
2659 swizzled_result
.writemask
= zero_mask
;
2660 emit(MOV(swizzled_result
, src_reg(0.0f
)));
2664 swizzled_result
.writemask
= one_mask
;
2665 emit(MOV(swizzled_result
, src_reg(1.0f
)));
2670 vec4_visitor::visit(ir_return
*)
2672 unreachable("not reached");
2676 vec4_visitor::visit(ir_discard
*)
2678 unreachable("not reached");
2682 vec4_visitor::visit(ir_if
*ir
)
2684 /* Don't point the annotation at the if statement, because then it plus
2685 * the then and else blocks get printed.
2687 this->base_ir
= ir
->condition
;
2689 if (brw
->gen
== 6) {
2692 enum brw_predicate predicate
;
2693 emit_bool_to_cond_code(ir
->condition
, &predicate
);
2694 emit(IF(predicate
));
2697 visit_instructions(&ir
->then_instructions
);
2699 if (!ir
->else_instructions
.is_empty()) {
2700 this->base_ir
= ir
->condition
;
2701 emit(BRW_OPCODE_ELSE
);
2703 visit_instructions(&ir
->else_instructions
);
2706 this->base_ir
= ir
->condition
;
2707 emit(BRW_OPCODE_ENDIF
);
2711 vec4_visitor::visit(ir_emit_vertex
*)
2713 unreachable("not reached");
2717 vec4_visitor::visit(ir_end_primitive
*)
2719 unreachable("not reached");
2723 vec4_visitor::emit_untyped_atomic(unsigned atomic_op
, unsigned surf_index
,
2724 dst_reg dst
, src_reg offset
,
2725 src_reg src0
, src_reg src1
)
2729 /* Set the atomic operation offset. */
2730 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), offset
));
2733 /* Set the atomic operation arguments. */
2734 if (src0
.file
!= BAD_FILE
) {
2735 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src0
));
2739 if (src1
.file
!= BAD_FILE
) {
2740 emit(MOV(brw_writemask(brw_uvec_mrf(8, mlen
, 0), WRITEMASK_X
), src1
));
2744 /* Emit the instruction. Note that this maps to the normal SIMD8
2745 * untyped atomic message on Ivy Bridge, but that's OK because
2746 * unused channels will be masked out.
2748 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_ATOMIC
, dst
,
2749 src_reg(atomic_op
), src_reg(surf_index
));
2755 vec4_visitor::emit_untyped_surface_read(unsigned surf_index
, dst_reg dst
,
2758 /* Set the surface read offset. */
2759 emit(MOV(brw_writemask(brw_uvec_mrf(8, 0, 0), WRITEMASK_X
), offset
));
2761 /* Emit the instruction. Note that this maps to the normal SIMD8
2762 * untyped surface read message, but that's OK because unused
2763 * channels will be masked out.
2765 vec4_instruction
*inst
= emit(SHADER_OPCODE_UNTYPED_SURFACE_READ
,
2766 dst
, src_reg(surf_index
));
2772 vec4_visitor::emit_ndc_computation()
2774 /* Get the position */
2775 src_reg pos
= src_reg(output_reg
[VARYING_SLOT_POS
]);
2777 /* Build ndc coords, which are (x/w, y/w, z/w, 1/w) */
2778 dst_reg ndc
= dst_reg(this, glsl_type::vec4_type
);
2779 output_reg
[BRW_VARYING_SLOT_NDC
] = ndc
;
2781 current_annotation
= "NDC";
2782 dst_reg ndc_w
= ndc
;
2783 ndc_w
.writemask
= WRITEMASK_W
;
2784 src_reg pos_w
= pos
;
2785 pos_w
.swizzle
= BRW_SWIZZLE4(SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
, SWIZZLE_W
);
2786 emit_math(SHADER_OPCODE_RCP
, ndc_w
, pos_w
);
2788 dst_reg ndc_xyz
= ndc
;
2789 ndc_xyz
.writemask
= WRITEMASK_XYZ
;
2791 emit(MUL(ndc_xyz
, pos
, src_reg(ndc_w
)));
2795 vec4_visitor::emit_psiz_and_flags(struct brw_reg reg
)
2798 ((prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) ||
2799 key
->userclip_active
|| brw
->has_negative_rhw_bug
)) {
2800 dst_reg header1
= dst_reg(this, glsl_type::uvec4_type
);
2801 dst_reg header1_w
= header1
;
2802 header1_w
.writemask
= WRITEMASK_W
;
2804 emit(MOV(header1
, 0u));
2806 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
2807 src_reg psiz
= src_reg(output_reg
[VARYING_SLOT_PSIZ
]);
2809 current_annotation
= "Point size";
2810 emit(MUL(header1_w
, psiz
, src_reg((float)(1 << 11))));
2811 emit(AND(header1_w
, src_reg(header1_w
), 0x7ff << 8));
2814 if (key
->userclip_active
) {
2815 current_annotation
= "Clipping flags";
2816 dst_reg flags0
= dst_reg(this, glsl_type::uint_type
);
2817 dst_reg flags1
= dst_reg(this, glsl_type::uint_type
);
2819 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST0
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
2820 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags0
, src_reg(0));
2821 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags0
)));
2823 emit(CMP(dst_null_f(), src_reg(output_reg
[VARYING_SLOT_CLIP_DIST1
]), src_reg(0.0f
), BRW_CONDITIONAL_L
));
2824 emit(VS_OPCODE_UNPACK_FLAGS_SIMD4X2
, flags1
, src_reg(0));
2825 emit(SHL(flags1
, src_reg(flags1
), src_reg(4)));
2826 emit(OR(header1_w
, src_reg(header1_w
), src_reg(flags1
)));
2829 /* i965 clipping workaround:
2830 * 1) Test for -ve rhw
2832 * set ndc = (0,0,0,0)
2835 * Later, clipping will detect ucp[6] and ensure the primitive is
2836 * clipped against all fixed planes.
2838 if (brw
->has_negative_rhw_bug
) {
2839 src_reg ndc_w
= src_reg(output_reg
[BRW_VARYING_SLOT_NDC
]);
2840 ndc_w
.swizzle
= BRW_SWIZZLE_WWWW
;
2841 emit(CMP(dst_null_f(), ndc_w
, src_reg(0.0f
), BRW_CONDITIONAL_L
));
2842 vec4_instruction
*inst
;
2843 inst
= emit(OR(header1_w
, src_reg(header1_w
), src_reg(1u << 6)));
2844 inst
->predicate
= BRW_PREDICATE_NORMAL
;
2845 inst
= emit(MOV(output_reg
[BRW_VARYING_SLOT_NDC
], src_reg(0.0f
)));
2846 inst
->predicate
= BRW_PREDICATE_NORMAL
;
2849 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), src_reg(header1
)));
2850 } else if (brw
->gen
< 6) {
2851 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_UD
), 0u));
2853 emit(MOV(retype(reg
, BRW_REGISTER_TYPE_D
), src_reg(0)));
2854 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_PSIZ
) {
2855 emit(MOV(brw_writemask(reg
, WRITEMASK_W
),
2856 src_reg(output_reg
[VARYING_SLOT_PSIZ
])));
2858 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_LAYER
) {
2859 emit(MOV(retype(brw_writemask(reg
, WRITEMASK_Y
), BRW_REGISTER_TYPE_D
),
2860 src_reg(output_reg
[VARYING_SLOT_LAYER
])));
2862 if (prog_data
->vue_map
.slots_valid
& VARYING_BIT_VIEWPORT
) {
2863 emit(MOV(retype(brw_writemask(reg
, WRITEMASK_Z
), BRW_REGISTER_TYPE_D
),
2864 src_reg(output_reg
[VARYING_SLOT_VIEWPORT
])));
2870 vec4_visitor::emit_clip_distances(dst_reg reg
, int offset
)
2872 /* From the GLSL 1.30 spec, section 7.1 (Vertex Shader Special Variables):
2874 * "If a linked set of shaders forming the vertex stage contains no
2875 * static write to gl_ClipVertex or gl_ClipDistance, but the
2876 * application has requested clipping against user clip planes through
2877 * the API, then the coordinate written to gl_Position is used for
2878 * comparison against the user clip planes."
2880 * This function is only called if the shader didn't write to
2881 * gl_ClipDistance. Accordingly, we use gl_ClipVertex to perform clipping
2882 * if the user wrote to it; otherwise we use gl_Position.
2884 gl_varying_slot clip_vertex
= VARYING_SLOT_CLIP_VERTEX
;
2885 if (!(prog_data
->vue_map
.slots_valid
& VARYING_BIT_CLIP_VERTEX
)) {
2886 clip_vertex
= VARYING_SLOT_POS
;
2889 for (int i
= 0; i
+ offset
< key
->nr_userclip_plane_consts
&& i
< 4;
2891 reg
.writemask
= 1 << i
;
2893 src_reg(output_reg
[clip_vertex
]),
2894 src_reg(this->userplane
[i
+ offset
])));
2899 vec4_visitor::emit_generic_urb_slot(dst_reg reg
, int varying
)
2901 assert (varying
< VARYING_SLOT_MAX
);
2902 reg
.type
= output_reg
[varying
].type
;
2903 current_annotation
= output_reg_annotation
[varying
];
2904 /* Copy the register, saturating if necessary */
2905 vec4_instruction
*inst
= emit(MOV(reg
,
2906 src_reg(output_reg
[varying
])));
2907 if ((varying
== VARYING_SLOT_COL0
||
2908 varying
== VARYING_SLOT_COL1
||
2909 varying
== VARYING_SLOT_BFC0
||
2910 varying
== VARYING_SLOT_BFC1
) &&
2911 key
->clamp_vertex_color
) {
2912 inst
->saturate
= true;
2917 vec4_visitor::emit_urb_slot(int mrf
, int varying
)
2919 struct brw_reg hw_reg
= brw_message_reg(mrf
);
2920 dst_reg reg
= dst_reg(MRF
, mrf
);
2921 reg
.type
= BRW_REGISTER_TYPE_F
;
2924 case VARYING_SLOT_PSIZ
:
2925 /* PSIZ is always in slot 0, and is coupled with other flags. */
2926 current_annotation
= "indices, point width, clip flags";
2927 emit_psiz_and_flags(hw_reg
);
2929 case BRW_VARYING_SLOT_NDC
:
2930 current_annotation
= "NDC";
2931 emit(MOV(reg
, src_reg(output_reg
[BRW_VARYING_SLOT_NDC
])));
2933 case VARYING_SLOT_POS
:
2934 current_annotation
= "gl_Position";
2935 emit(MOV(reg
, src_reg(output_reg
[VARYING_SLOT_POS
])));
2937 case VARYING_SLOT_EDGE
:
2938 /* This is present when doing unfilled polygons. We're supposed to copy
2939 * the edge flag from the user-provided vertex array
2940 * (glEdgeFlagPointer), or otherwise we'll copy from the current value
2941 * of that attribute (starts as 1.0f). This is then used in clipping to
2942 * determine which edges should be drawn as wireframe.
2944 current_annotation
= "edge flag";
2945 emit(MOV(reg
, src_reg(dst_reg(ATTR
, VERT_ATTRIB_EDGEFLAG
,
2946 glsl_type::float_type
, WRITEMASK_XYZW
))));
2948 case BRW_VARYING_SLOT_PAD
:
2949 /* No need to write to this slot */
2952 emit_generic_urb_slot(reg
, varying
);
2958 align_interleaved_urb_mlen(struct brw_context
*brw
, int mlen
)
2960 if (brw
->gen
>= 6) {
2961 /* URB data written (does not include the message header reg) must
2962 * be a multiple of 256 bits, or 2 VS registers. See vol5c.5,
2963 * section 5.4.3.2.2: URB_INTERLEAVED.
2965 * URB entries are allocated on a multiple of 1024 bits, so an
2966 * extra 128 bits written here to make the end align to 256 is
2969 if ((mlen
% 2) != 1)
2978 * Generates the VUE payload plus the necessary URB write instructions to
2981 * The VUE layout is documented in Volume 2a.
2984 vec4_visitor::emit_vertex()
2986 /* MRF 0 is reserved for the debugger, so start with message header
2991 /* In the process of generating our URB write message contents, we
2992 * may need to unspill a register or load from an array. Those
2993 * reads would use MRFs 14-15.
2995 int max_usable_mrf
= 13;
2997 /* The following assertion verifies that max_usable_mrf causes an
2998 * even-numbered amount of URB write data, which will meet gen6's
2999 * requirements for length alignment.
3001 assert ((max_usable_mrf
- base_mrf
) % 2 == 0);
3003 /* First mrf is the g0-based message header containing URB handles and
3006 emit_urb_write_header(mrf
++);
3009 emit_ndc_computation();
3012 /* Lower legacy ff and ClipVertex clipping to clip distances */
3013 if (key
->userclip_active
&& !prog
->UsesClipDistanceOut
) {
3014 current_annotation
= "user clip distances";
3016 output_reg
[VARYING_SLOT_CLIP_DIST0
] = dst_reg(this, glsl_type::vec4_type
);
3017 output_reg
[VARYING_SLOT_CLIP_DIST1
] = dst_reg(this, glsl_type::vec4_type
);
3019 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST0
], 0);
3020 emit_clip_distances(output_reg
[VARYING_SLOT_CLIP_DIST1
], 4);
3023 /* We may need to split this up into several URB writes, so do them in a
3027 bool complete
= false;
3029 /* URB offset is in URB row increments, and each of our MRFs is half of
3030 * one of those, since we're doing interleaved writes.
3032 int offset
= slot
/ 2;
3035 for (; slot
< prog_data
->vue_map
.num_slots
; ++slot
) {
3036 emit_urb_slot(mrf
++, prog_data
->vue_map
.slot_to_varying
[slot
]);
3038 /* If this was max_usable_mrf, we can't fit anything more into this
3041 if (mrf
> max_usable_mrf
) {
3047 complete
= slot
>= prog_data
->vue_map
.num_slots
;
3048 current_annotation
= "URB write";
3049 vec4_instruction
*inst
= emit_urb_write_opcode(complete
);
3050 inst
->base_mrf
= base_mrf
;
3051 inst
->mlen
= align_interleaved_urb_mlen(brw
, mrf
- base_mrf
);
3052 inst
->offset
+= offset
;
3058 vec4_visitor::get_scratch_offset(vec4_instruction
*inst
,
3059 src_reg
*reladdr
, int reg_offset
)
3061 /* Because we store the values to scratch interleaved like our
3062 * vertex data, we need to scale the vec4 index by 2.
3064 int message_header_scale
= 2;
3066 /* Pre-gen6, the message header uses byte offsets instead of vec4
3067 * (16-byte) offset units.
3070 message_header_scale
*= 16;
3073 src_reg index
= src_reg(this, glsl_type::int_type
);
3075 emit_before(inst
, ADD(dst_reg(index
), *reladdr
, src_reg(reg_offset
)));
3076 emit_before(inst
, MUL(dst_reg(index
),
3077 index
, src_reg(message_header_scale
)));
3081 return src_reg(reg_offset
* message_header_scale
);
3086 vec4_visitor::get_pull_constant_offset(vec4_instruction
*inst
,
3087 src_reg
*reladdr
, int reg_offset
)
3090 src_reg index
= src_reg(this, glsl_type::int_type
);
3092 emit_before(inst
, ADD(dst_reg(index
), *reladdr
, src_reg(reg_offset
)));
3094 /* Pre-gen6, the message header uses byte offsets instead of vec4
3095 * (16-byte) offset units.
3098 emit_before(inst
, MUL(dst_reg(index
), index
, src_reg(16)));
3102 } else if (brw
->gen
>= 8) {
3103 /* Store the offset in a GRF so we can send-from-GRF. */
3104 src_reg offset
= src_reg(this, glsl_type::int_type
);
3105 emit_before(inst
, MOV(dst_reg(offset
), src_reg(reg_offset
)));
3108 int message_header_scale
= brw
->gen
< 6 ? 16 : 1;
3109 return src_reg(reg_offset
* message_header_scale
);
3114 * Emits an instruction before @inst to load the value named by @orig_src
3115 * from scratch space at @base_offset to @temp.
3117 * @base_offset is measured in 32-byte units (the size of a register).
3120 vec4_visitor::emit_scratch_read(vec4_instruction
*inst
,
3121 dst_reg temp
, src_reg orig_src
,
3124 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3125 src_reg index
= get_scratch_offset(inst
, orig_src
.reladdr
, reg_offset
);
3127 emit_before(inst
, SCRATCH_READ(temp
, index
));
3131 * Emits an instruction after @inst to store the value to be written
3132 * to @orig_dst to scratch space at @base_offset, from @temp.
3134 * @base_offset is measured in 32-byte units (the size of a register).
3137 vec4_visitor::emit_scratch_write(vec4_instruction
*inst
, int base_offset
)
3139 int reg_offset
= base_offset
+ inst
->dst
.reg_offset
;
3140 src_reg index
= get_scratch_offset(inst
, inst
->dst
.reladdr
, reg_offset
);
3142 /* Create a temporary register to store *inst's result in.
3144 * We have to be careful in MOVing from our temporary result register in
3145 * the scratch write. If we swizzle from channels of the temporary that
3146 * weren't initialized, it will confuse live interval analysis, which will
3147 * make spilling fail to make progress.
3149 src_reg temp
= src_reg(this, glsl_type::vec4_type
);
3150 temp
.type
= inst
->dst
.type
;
3151 int first_writemask_chan
= ffs(inst
->dst
.writemask
) - 1;
3153 for (int i
= 0; i
< 4; i
++)
3154 if (inst
->dst
.writemask
& (1 << i
))
3157 swizzles
[i
] = first_writemask_chan
;
3158 temp
.swizzle
= BRW_SWIZZLE4(swizzles
[0], swizzles
[1],
3159 swizzles
[2], swizzles
[3]);
3161 dst_reg dst
= dst_reg(brw_writemask(brw_vec8_grf(0, 0),
3162 inst
->dst
.writemask
));
3163 vec4_instruction
*write
= SCRATCH_WRITE(dst
, temp
, index
);
3164 write
->predicate
= inst
->predicate
;
3165 write
->ir
= inst
->ir
;
3166 write
->annotation
= inst
->annotation
;
3167 inst
->insert_after(write
);
3169 inst
->dst
.file
= temp
.file
;
3170 inst
->dst
.reg
= temp
.reg
;
3171 inst
->dst
.reg_offset
= temp
.reg_offset
;
3172 inst
->dst
.reladdr
= NULL
;
3176 * We can't generally support array access in GRF space, because a
3177 * single instruction's destination can only span 2 contiguous
3178 * registers. So, we send all GRF arrays that get variable index
3179 * access to scratch space.
3182 vec4_visitor::move_grf_array_access_to_scratch()
3184 int scratch_loc
[this->virtual_grf_count
];
3186 for (int i
= 0; i
< this->virtual_grf_count
; i
++) {
3187 scratch_loc
[i
] = -1;
3190 /* First, calculate the set of virtual GRFs that need to be punted
3191 * to scratch due to having any array access on them, and where in
3194 foreach_in_list(vec4_instruction
, inst
, &instructions
) {
3195 if (inst
->dst
.file
== GRF
&& inst
->dst
.reladdr
&&
3196 scratch_loc
[inst
->dst
.reg
] == -1) {
3197 scratch_loc
[inst
->dst
.reg
] = c
->last_scratch
;
3198 c
->last_scratch
+= this->virtual_grf_sizes
[inst
->dst
.reg
];
3201 for (int i
= 0 ; i
< 3; i
++) {
3202 src_reg
*src
= &inst
->src
[i
];
3204 if (src
->file
== GRF
&& src
->reladdr
&&
3205 scratch_loc
[src
->reg
] == -1) {
3206 scratch_loc
[src
->reg
] = c
->last_scratch
;
3207 c
->last_scratch
+= this->virtual_grf_sizes
[src
->reg
];
3212 /* Now, for anything that will be accessed through scratch, rewrite
3213 * it to load/store. Note that this is a _safe list walk, because
3214 * we may generate a new scratch_write instruction after the one
3217 foreach_in_list_safe(vec4_instruction
, inst
, &instructions
) {
3218 /* Set up the annotation tracking for new generated instructions. */
3220 current_annotation
= inst
->annotation
;
3222 if (inst
->dst
.file
== GRF
&& scratch_loc
[inst
->dst
.reg
] != -1) {
3223 emit_scratch_write(inst
, scratch_loc
[inst
->dst
.reg
]);
3226 for (int i
= 0 ; i
< 3; i
++) {
3227 if (inst
->src
[i
].file
!= GRF
|| scratch_loc
[inst
->src
[i
].reg
] == -1)
3230 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3232 emit_scratch_read(inst
, temp
, inst
->src
[i
],
3233 scratch_loc
[inst
->src
[i
].reg
]);
3235 inst
->src
[i
].file
= temp
.file
;
3236 inst
->src
[i
].reg
= temp
.reg
;
3237 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
3238 inst
->src
[i
].reladdr
= NULL
;
3244 * Emits an instruction before @inst to load the value named by @orig_src
3245 * from the pull constant buffer (surface) at @base_offset to @temp.
3248 vec4_visitor::emit_pull_constant_load(vec4_instruction
*inst
,
3249 dst_reg temp
, src_reg orig_src
,
3252 int reg_offset
= base_offset
+ orig_src
.reg_offset
;
3253 src_reg index
= src_reg(prog_data
->base
.binding_table
.pull_constants_start
);
3254 src_reg offset
= get_pull_constant_offset(inst
, orig_src
.reladdr
, reg_offset
);
3255 vec4_instruction
*load
;
3257 if (brw
->gen
>= 7) {
3258 dst_reg grf_offset
= dst_reg(this, glsl_type::int_type
);
3259 grf_offset
.type
= offset
.type
;
3260 emit_before(inst
, MOV(grf_offset
, offset
));
3262 load
= new(mem_ctx
) vec4_instruction(this,
3263 VS_OPCODE_PULL_CONSTANT_LOAD_GEN7
,
3264 temp
, index
, src_reg(grf_offset
));
3266 load
= new(mem_ctx
) vec4_instruction(this, VS_OPCODE_PULL_CONSTANT_LOAD
,
3267 temp
, index
, offset
);
3268 load
->base_mrf
= 14;
3271 emit_before(inst
, load
);
3275 * Implements array access of uniforms by inserting a
3276 * PULL_CONSTANT_LOAD instruction.
3278 * Unlike temporary GRF array access (where we don't support it due to
3279 * the difficulty of doing relative addressing on instruction
3280 * destinations), we could potentially do array access of uniforms
3281 * that were loaded in GRF space as push constants. In real-world
3282 * usage we've seen, though, the arrays being used are always larger
3283 * than we could load as push constants, so just always move all
3284 * uniform array access out to a pull constant buffer.
3287 vec4_visitor::move_uniform_array_access_to_pull_constants()
3289 int pull_constant_loc
[this->uniforms
];
3291 for (int i
= 0; i
< this->uniforms
; i
++) {
3292 pull_constant_loc
[i
] = -1;
3295 /* Walk through and find array access of uniforms. Put a copy of that
3296 * uniform in the pull constant buffer.
3298 * Note that we don't move constant-indexed accesses to arrays. No
3299 * testing has been done of the performance impact of this choice.
3301 foreach_in_list_safe(vec4_instruction
, inst
, &instructions
) {
3302 for (int i
= 0 ; i
< 3; i
++) {
3303 if (inst
->src
[i
].file
!= UNIFORM
|| !inst
->src
[i
].reladdr
)
3306 int uniform
= inst
->src
[i
].reg
;
3308 /* If this array isn't already present in the pull constant buffer,
3311 if (pull_constant_loc
[uniform
] == -1) {
3312 const float **values
= &stage_prog_data
->param
[uniform
* 4];
3314 pull_constant_loc
[uniform
] = stage_prog_data
->nr_pull_params
/ 4;
3316 assert(uniform
< uniform_array_size
);
3317 for (int j
= 0; j
< uniform_size
[uniform
] * 4; j
++) {
3318 stage_prog_data
->pull_param
[stage_prog_data
->nr_pull_params
++]
3323 /* Set up the annotation tracking for new generated instructions. */
3325 current_annotation
= inst
->annotation
;
3327 dst_reg temp
= dst_reg(this, glsl_type::vec4_type
);
3329 emit_pull_constant_load(inst
, temp
, inst
->src
[i
],
3330 pull_constant_loc
[uniform
]);
3332 inst
->src
[i
].file
= temp
.file
;
3333 inst
->src
[i
].reg
= temp
.reg
;
3334 inst
->src
[i
].reg_offset
= temp
.reg_offset
;
3335 inst
->src
[i
].reladdr
= NULL
;
3339 /* Now there are no accesses of the UNIFORM file with a reladdr, so
3340 * no need to track them as larger-than-vec4 objects. This will be
3341 * relied on in cutting out unused uniform vectors from push
3344 split_uniform_registers();
3348 vec4_visitor::resolve_ud_negate(src_reg
*reg
)
3350 if (reg
->type
!= BRW_REGISTER_TYPE_UD
||
3354 src_reg temp
= src_reg(this, glsl_type::uvec4_type
);
3355 emit(BRW_OPCODE_MOV
, dst_reg(temp
), *reg
);
3359 vec4_visitor::vec4_visitor(struct brw_context
*brw
,
3360 struct brw_vec4_compile
*c
,
3361 struct gl_program
*prog
,
3362 const struct brw_vec4_prog_key
*key
,
3363 struct brw_vec4_prog_data
*prog_data
,
3364 struct gl_shader_program
*shader_prog
,
3365 gl_shader_stage stage
,
3369 shader_time_shader_type st_base
,
3370 shader_time_shader_type st_written
,
3371 shader_time_shader_type st_reset
)
3372 : backend_visitor(brw
, shader_prog
, prog
, &prog_data
->base
, stage
),
3375 prog_data(prog_data
),
3376 sanity_param_count(0),
3378 first_non_payload_grf(0),
3379 need_all_constants_in_pull_buffer(false),
3380 debug_flag(debug_flag
),
3381 no_spills(no_spills
),
3383 st_written(st_written
),
3386 this->mem_ctx
= mem_ctx
;
3387 this->failed
= false;
3389 this->base_ir
= NULL
;
3390 this->current_annotation
= NULL
;
3391 memset(this->output_reg_annotation
, 0, sizeof(this->output_reg_annotation
));
3393 this->variable_ht
= hash_table_ctor(0,
3394 hash_table_pointer_hash
,
3395 hash_table_pointer_compare
);
3397 this->virtual_grf_start
= NULL
;
3398 this->virtual_grf_end
= NULL
;
3399 this->virtual_grf_sizes
= NULL
;
3400 this->virtual_grf_count
= 0;
3401 this->virtual_grf_reg_map
= NULL
;
3402 this->virtual_grf_reg_count
= 0;
3403 this->virtual_grf_array_size
= 0;
3404 this->live_intervals_valid
= false;
3406 this->max_grf
= brw
->gen
>= 7 ? GEN7_MRF_HACK_START
: BRW_MAX_GRF
;
3410 /* Initialize uniform_array_size to at least 1 because pre-gen6 VS requires
3411 * at least one. See setup_uniforms() in brw_vec4.cpp.
3413 this->uniform_array_size
= 1;
3415 this->uniform_array_size
= MAX2(stage_prog_data
->nr_params
, 1);
3418 this->uniform_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3419 this->uniform_vector_size
= rzalloc_array(mem_ctx
, int, this->uniform_array_size
);
3422 vec4_visitor::~vec4_visitor()
3424 hash_table_dtor(this->variable_ht
);
3429 vec4_visitor::fail(const char *format
, ...)
3439 va_start(va
, format
);
3440 msg
= ralloc_vasprintf(mem_ctx
, format
, va
);
3442 msg
= ralloc_asprintf(mem_ctx
, "vec4 compile failed: %s\n", msg
);
3444 this->fail_msg
= msg
;
3447 fprintf(stderr
, "%s", msg
);
3451 } /* namespace brw */