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 "glsl/ir_optimization.h"
26 #include "glsl/nir/glsl_to_nir.h"
27 #include "program/prog_to_nir.h"
34 fs_visitor::emit_nir_code()
36 nir_shader
*nir
= prog
->nir
;
38 /* emit the arrays used for inputs and outputs - load/store intrinsics will
39 * be converted to reads/writes of these arrays
42 if (nir
->num_inputs
> 0) {
43 nir_inputs
= bld
.vgrf(BRW_REGISTER_TYPE_F
, nir
->num_inputs
);
44 nir_setup_inputs(nir
);
47 if (nir
->num_outputs
> 0) {
48 nir_outputs
= bld
.vgrf(BRW_REGISTER_TYPE_F
, nir
->num_outputs
);
49 nir_setup_outputs(nir
);
52 if (nir
->num_uniforms
> 0) {
53 nir_setup_uniforms(nir
);
56 nir_emit_system_values(nir
);
58 nir_globals
= ralloc_array(mem_ctx
, fs_reg
, nir
->reg_alloc
);
59 foreach_list_typed(nir_register
, reg
, node
, &nir
->registers
) {
60 unsigned array_elems
=
61 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
62 unsigned size
= array_elems
* reg
->num_components
;
63 nir_globals
[reg
->index
] = bld
.vgrf(BRW_REGISTER_TYPE_F
, size
);
66 /* get the main function and emit it */
67 nir_foreach_overload(nir
, overload
) {
68 assert(strcmp(overload
->function
->name
, "main") == 0);
69 assert(overload
->impl
);
70 nir_emit_impl(overload
->impl
);
75 fs_visitor::nir_setup_inputs(nir_shader
*shader
)
77 foreach_list_typed(nir_variable
, var
, node
, &shader
->inputs
) {
78 enum brw_reg_type type
= brw_type_for_base_type(var
->type
);
79 fs_reg input
= offset(nir_inputs
, var
->data
.driver_location
);
83 case MESA_SHADER_VERTEX
: {
84 /* Our ATTR file is indexed by VERT_ATTRIB_*, which is the value
85 * stored in nir_variable::location.
87 * However, NIR's load_input intrinsics use a different index - an
88 * offset into a single contiguous array containing all inputs.
89 * This index corresponds to the nir_variable::driver_location field.
91 * So, we need to copy from fs_reg(ATTR, var->location) to
92 * offset(nir_inputs, var->data.driver_location).
94 unsigned components
= var
->type
->without_array()->components();
95 unsigned array_length
= var
->type
->is_array() ? var
->type
->length
: 1;
96 for (unsigned i
= 0; i
< array_length
; i
++) {
97 for (unsigned j
= 0; j
< components
; j
++) {
98 bld
.MOV(retype(offset(input
, components
* i
+ j
), type
),
99 offset(fs_reg(ATTR
, var
->data
.location
+ i
, type
), j
));
104 case MESA_SHADER_GEOMETRY
:
105 case MESA_SHADER_COMPUTE
:
106 unreachable("fs_visitor not used for these stages yet.");
108 case MESA_SHADER_FRAGMENT
:
109 if (var
->data
.location
== VARYING_SLOT_POS
) {
110 reg
= *emit_fragcoord_interpolation(var
->data
.pixel_center_integer
,
111 var
->data
.origin_upper_left
);
112 emit_percomp(bld
, fs_inst(BRW_OPCODE_MOV
, input
, reg
), 0xF);
114 emit_general_interpolation(input
, var
->name
, var
->type
,
115 (glsl_interp_qualifier
) var
->data
.interpolation
,
116 var
->data
.location
, var
->data
.centroid
,
125 fs_visitor::nir_setup_outputs(nir_shader
*shader
)
127 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
129 foreach_list_typed(nir_variable
, var
, node
, &shader
->outputs
) {
130 fs_reg reg
= offset(nir_outputs
, var
->data
.driver_location
);
132 int vector_elements
=
133 var
->type
->is_array() ? var
->type
->fields
.array
->vector_elements
134 : var
->type
->vector_elements
;
136 if (stage
== MESA_SHADER_VERTEX
) {
137 for (int i
= 0; i
< ALIGN(type_size(var
->type
), 4) / 4; i
++) {
138 int output
= var
->data
.location
+ i
;
139 this->outputs
[output
] = offset(reg
, 4 * i
);
140 this->output_components
[output
] = vector_elements
;
142 } else if (var
->data
.index
> 0) {
143 assert(var
->data
.location
== FRAG_RESULT_DATA0
);
144 assert(var
->data
.index
== 1);
145 this->dual_src_output
= reg
;
146 this->do_dual_src
= true;
147 } else if (var
->data
.location
== FRAG_RESULT_COLOR
) {
148 /* Writing gl_FragColor outputs to all color regions. */
149 for (unsigned int i
= 0; i
< MAX2(key
->nr_color_regions
, 1); i
++) {
150 this->outputs
[i
] = reg
;
151 this->output_components
[i
] = 4;
153 } else if (var
->data
.location
== FRAG_RESULT_DEPTH
) {
154 this->frag_depth
= reg
;
155 } else if (var
->data
.location
== FRAG_RESULT_SAMPLE_MASK
) {
156 this->sample_mask
= reg
;
158 /* gl_FragData or a user-defined FS output */
159 assert(var
->data
.location
>= FRAG_RESULT_DATA0
&&
160 var
->data
.location
< FRAG_RESULT_DATA0
+ BRW_MAX_DRAW_BUFFERS
);
162 /* General color output. */
163 for (unsigned int i
= 0; i
< MAX2(1, var
->type
->length
); i
++) {
164 int output
= var
->data
.location
- FRAG_RESULT_DATA0
+ i
;
165 this->outputs
[output
] = offset(reg
, vector_elements
* i
);
166 this->output_components
[output
] = vector_elements
;
173 fs_visitor::nir_setup_uniforms(nir_shader
*shader
)
175 uniforms
= shader
->num_uniforms
;
176 num_direct_uniforms
= shader
->num_direct_uniforms
;
178 /* We split the uniform register file in half. The first half is
179 * entirely direct uniforms. The second half is indirect.
181 param_size
[0] = num_direct_uniforms
;
182 if (shader
->num_uniforms
> num_direct_uniforms
)
183 param_size
[num_direct_uniforms
] = shader
->num_uniforms
- num_direct_uniforms
;
185 if (dispatch_width
!= 8)
189 foreach_list_typed(nir_variable
, var
, node
, &shader
->uniforms
) {
190 /* UBO's and atomics don't take up space in the uniform file */
191 if (var
->interface_type
!= NULL
|| var
->type
->contains_atomic())
194 if (strncmp(var
->name
, "gl_", 3) == 0)
195 nir_setup_builtin_uniform(var
);
197 nir_setup_uniform(var
);
200 /* prog_to_nir doesn't create uniform variables; set param up directly. */
201 for (unsigned p
= 0; p
< prog
->Parameters
->NumParameters
; p
++) {
202 for (unsigned int i
= 0; i
< 4; i
++) {
203 stage_prog_data
->param
[4 * p
+ i
] =
204 &prog
->Parameters
->ParameterValues
[p
][i
];
211 fs_visitor::nir_setup_uniform(nir_variable
*var
)
213 int namelen
= strlen(var
->name
);
215 /* The data for our (non-builtin) uniforms is stored in a series of
216 * gl_uniform_driver_storage structs for each subcomponent that
217 * glGetUniformLocation() could name. We know it's been set up in the
218 * same order we'd walk the type, so walk the list of storage and find
219 * anything with our name, or the prefix of a component that starts with
222 unsigned index
= var
->data
.driver_location
;
223 for (unsigned u
= 0; u
< shader_prog
->NumUniformStorage
; u
++) {
224 struct gl_uniform_storage
*storage
= &shader_prog
->UniformStorage
[u
];
226 if (storage
->builtin
)
229 if (strncmp(var
->name
, storage
->name
, namelen
) != 0 ||
230 (storage
->name
[namelen
] != 0 &&
231 storage
->name
[namelen
] != '.' &&
232 storage
->name
[namelen
] != '[')) {
236 unsigned slots
= storage
->type
->component_slots();
237 if (storage
->array_elements
)
238 slots
*= storage
->array_elements
;
240 for (unsigned i
= 0; i
< slots
; i
++) {
241 stage_prog_data
->param
[index
++] = &storage
->storage
[i
];
245 /* Make sure we actually initialized the right amount of stuff here. */
246 assert(var
->data
.driver_location
+ var
->type
->component_slots() == index
);
250 fs_visitor::nir_setup_builtin_uniform(nir_variable
*var
)
252 const nir_state_slot
*const slots
= var
->state_slots
;
253 assert(var
->state_slots
!= NULL
);
255 unsigned uniform_index
= var
->data
.driver_location
;
256 for (unsigned int i
= 0; i
< var
->num_state_slots
; i
++) {
257 /* This state reference has already been setup by ir_to_mesa, but we'll
258 * get the same index back here.
260 int index
= _mesa_add_state_reference(this->prog
->Parameters
,
261 (gl_state_index
*)slots
[i
].tokens
);
263 /* Add each of the unique swizzles of the element as a parameter.
264 * This'll end up matching the expected layout of the
265 * array/matrix/structure we're trying to fill in.
268 for (unsigned int j
= 0; j
< 4; j
++) {
269 int swiz
= GET_SWZ(slots
[i
].swizzle
, j
);
270 if (swiz
== last_swiz
)
274 stage_prog_data
->param
[uniform_index
++] =
275 &prog
->Parameters
->ParameterValues
[index
][swiz
];
281 emit_system_values_block(nir_block
*block
, void *void_visitor
)
283 fs_visitor
*v
= (fs_visitor
*)void_visitor
;
286 nir_foreach_instr(block
, instr
) {
287 if (instr
->type
!= nir_instr_type_intrinsic
)
290 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
291 switch (intrin
->intrinsic
) {
292 case nir_intrinsic_load_vertex_id
:
293 unreachable("should be lowered by lower_vertex_id().");
295 case nir_intrinsic_load_vertex_id_zero_base
:
296 assert(v
->stage
== MESA_SHADER_VERTEX
);
297 reg
= &v
->nir_system_values
[SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
];
298 if (reg
->file
== BAD_FILE
)
299 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
);
302 case nir_intrinsic_load_base_vertex
:
303 assert(v
->stage
== MESA_SHADER_VERTEX
);
304 reg
= &v
->nir_system_values
[SYSTEM_VALUE_BASE_VERTEX
];
305 if (reg
->file
== BAD_FILE
)
306 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_BASE_VERTEX
);
309 case nir_intrinsic_load_instance_id
:
310 assert(v
->stage
== MESA_SHADER_VERTEX
);
311 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INSTANCE_ID
];
312 if (reg
->file
== BAD_FILE
)
313 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_INSTANCE_ID
);
316 case nir_intrinsic_load_sample_pos
:
317 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
318 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
319 if (reg
->file
== BAD_FILE
)
320 *reg
= *v
->emit_samplepos_setup();
323 case nir_intrinsic_load_sample_id
:
324 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
325 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
326 if (reg
->file
== BAD_FILE
)
327 *reg
= *v
->emit_sampleid_setup();
330 case nir_intrinsic_load_sample_mask_in
:
331 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
332 assert(v
->devinfo
->gen
>= 7);
333 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
334 if (reg
->file
== BAD_FILE
)
335 *reg
= fs_reg(retype(brw_vec8_grf(v
->payload
.sample_mask_in_reg
, 0),
336 BRW_REGISTER_TYPE_D
));
348 fs_visitor::nir_emit_system_values(nir_shader
*shader
)
350 nir_system_values
= ralloc_array(mem_ctx
, fs_reg
, SYSTEM_VALUE_MAX
);
351 nir_foreach_overload(shader
, overload
) {
352 assert(strcmp(overload
->function
->name
, "main") == 0);
353 assert(overload
->impl
);
354 nir_foreach_block(overload
->impl
, emit_system_values_block
, this);
359 fs_visitor::nir_emit_impl(nir_function_impl
*impl
)
361 nir_locals
= reralloc(mem_ctx
, nir_locals
, fs_reg
, impl
->reg_alloc
);
362 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
363 unsigned array_elems
=
364 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
365 unsigned size
= array_elems
* reg
->num_components
;
366 nir_locals
[reg
->index
] = bld
.vgrf(BRW_REGISTER_TYPE_F
, size
);
369 nir_emit_cf_list(&impl
->body
);
373 fs_visitor::nir_emit_cf_list(exec_list
*list
)
375 exec_list_validate(list
);
376 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
377 switch (node
->type
) {
379 nir_emit_if(nir_cf_node_as_if(node
));
382 case nir_cf_node_loop
:
383 nir_emit_loop(nir_cf_node_as_loop(node
));
386 case nir_cf_node_block
:
387 nir_emit_block(nir_cf_node_as_block(node
));
391 unreachable("Invalid CFG node block");
397 fs_visitor::nir_emit_if(nir_if
*if_stmt
)
399 /* first, put the condition into f0 */
400 fs_inst
*inst
= bld
.MOV(bld
.null_reg_d(),
401 retype(get_nir_src(if_stmt
->condition
),
402 BRW_REGISTER_TYPE_D
));
403 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
405 bld
.IF(BRW_PREDICATE_NORMAL
);
407 nir_emit_cf_list(&if_stmt
->then_list
);
409 /* note: if the else is empty, dead CF elimination will remove it */
410 bld
.emit(BRW_OPCODE_ELSE
);
412 nir_emit_cf_list(&if_stmt
->else_list
);
414 bld
.emit(BRW_OPCODE_ENDIF
);
416 if (!try_replace_with_sel() && devinfo
->gen
< 6) {
417 no16("Can't support (non-uniform) control flow on SIMD16\n");
422 fs_visitor::nir_emit_loop(nir_loop
*loop
)
424 if (devinfo
->gen
< 6) {
425 no16("Can't support (non-uniform) control flow on SIMD16\n");
428 bld
.emit(BRW_OPCODE_DO
);
430 nir_emit_cf_list(&loop
->body
);
432 bld
.emit(BRW_OPCODE_WHILE
);
436 fs_visitor::nir_emit_block(nir_block
*block
)
438 nir_foreach_instr(block
, instr
) {
439 nir_emit_instr(instr
);
444 fs_visitor::nir_emit_instr(nir_instr
*instr
)
446 const fs_builder abld
= bld
.annotate(NULL
, instr
);
447 this->base_ir
= instr
;
449 switch (instr
->type
) {
450 case nir_instr_type_alu
:
451 nir_emit_alu(abld
, nir_instr_as_alu(instr
));
454 case nir_instr_type_intrinsic
:
455 nir_emit_intrinsic(nir_instr_as_intrinsic(instr
));
458 case nir_instr_type_tex
:
459 nir_emit_texture(nir_instr_as_tex(instr
));
462 case nir_instr_type_load_const
:
463 /* We can hit these, but we do nothing now and use them as
468 case nir_instr_type_jump
:
469 nir_emit_jump(abld
, nir_instr_as_jump(instr
));
473 unreachable("unknown instruction type");
476 this->base_ir
= NULL
;
480 brw_type_for_nir_type(nir_alu_type type
)
483 case nir_type_unsigned
:
484 return BRW_REGISTER_TYPE_UD
;
487 return BRW_REGISTER_TYPE_D
;
489 return BRW_REGISTER_TYPE_F
;
491 unreachable("unknown type");
494 return BRW_REGISTER_TYPE_F
;
498 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr
*instr
,
499 const fs_reg
&result
)
501 if (instr
->src
[0].src
.is_ssa
||
502 !instr
->src
[0].src
.reg
.reg
||
503 !instr
->src
[0].src
.reg
.reg
->parent_instr
)
506 if (instr
->src
[0].src
.reg
.reg
->parent_instr
->type
!=
507 nir_instr_type_intrinsic
)
510 nir_intrinsic_instr
*src0
=
511 nir_instr_as_intrinsic(instr
->src
[0].src
.reg
.reg
->parent_instr
);
513 if (src0
->intrinsic
!= nir_intrinsic_load_front_face
)
516 nir_const_value
*value1
= nir_src_as_const_value(instr
->src
[1].src
);
517 if (!value1
|| fabsf(value1
->f
[0]) != 1.0f
)
520 nir_const_value
*value2
= nir_src_as_const_value(instr
->src
[2].src
);
521 if (!value2
|| fabsf(value2
->f
[0]) != 1.0f
)
524 fs_reg tmp
= vgrf(glsl_type::int_type
);
526 if (devinfo
->gen
>= 6) {
527 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
528 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
530 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
532 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
533 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
535 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
537 * This negation looks like it's safe in practice, because bits 0:4 will
538 * surely be TRIANGLES
541 if (value1
->f
[0] == -1.0f
) {
545 tmp
.type
= BRW_REGISTER_TYPE_W
;
546 tmp
.subreg_offset
= 2;
549 fs_inst
*or_inst
= bld
.OR(tmp
, g0
, fs_reg(0x3f80));
550 or_inst
->src
[1].type
= BRW_REGISTER_TYPE_UW
;
552 tmp
.type
= BRW_REGISTER_TYPE_D
;
553 tmp
.subreg_offset
= 0;
556 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
557 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
559 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
561 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
562 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
564 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
566 * This negation looks like it's safe in practice, because bits 0:4 will
567 * surely be TRIANGLES
570 if (value1
->f
[0] == -1.0f
) {
574 bld
.OR(tmp
, g1_6
, fs_reg(0x3f800000));
576 bld
.AND(retype(result
, BRW_REGISTER_TYPE_D
), tmp
, fs_reg(0xbf800000));
582 fs_visitor::nir_emit_alu(const fs_builder
&bld
, nir_alu_instr
*instr
)
584 struct brw_wm_prog_key
*fs_key
= (struct brw_wm_prog_key
*) this->key
;
587 fs_reg result
= get_nir_dest(instr
->dest
.dest
);
588 result
.type
= brw_type_for_nir_type(nir_op_infos
[instr
->op
].output_type
);
591 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
592 op
[i
] = get_nir_src(instr
->src
[i
].src
);
593 op
[i
].type
= brw_type_for_nir_type(nir_op_infos
[instr
->op
].input_types
[i
]);
594 op
[i
].abs
= instr
->src
[i
].abs
;
595 op
[i
].negate
= instr
->src
[i
].negate
;
598 /* We get a bunch of mov's out of the from_ssa pass and they may still
599 * be vectorized. We'll handle them as a special-case. We'll also
600 * handle vecN here because it's basically the same thing.
608 fs_reg temp
= result
;
609 bool need_extra_copy
= false;
610 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
611 if (!instr
->src
[i
].src
.is_ssa
&&
612 instr
->dest
.dest
.reg
.reg
== instr
->src
[i
].src
.reg
.reg
) {
613 need_extra_copy
= true;
614 temp
= bld
.vgrf(result
.type
, 4);
619 for (unsigned i
= 0; i
< 4; i
++) {
620 if (!(instr
->dest
.write_mask
& (1 << i
)))
623 if (instr
->op
== nir_op_imov
|| instr
->op
== nir_op_fmov
) {
624 inst
= bld
.MOV(offset(temp
, i
),
625 offset(op
[0], instr
->src
[0].swizzle
[i
]));
627 inst
= bld
.MOV(offset(temp
, i
),
628 offset(op
[i
], instr
->src
[i
].swizzle
[0]));
630 inst
->saturate
= instr
->dest
.saturate
;
633 /* In this case the source and destination registers were the same,
634 * so we need to insert an extra set of moves in order to deal with
637 if (need_extra_copy
) {
638 for (unsigned i
= 0; i
< 4; i
++) {
639 if (!(instr
->dest
.write_mask
& (1 << i
)))
642 bld
.MOV(offset(result
, i
), offset(temp
, i
));
651 /* At this point, we have dealt with any instruction that operates on
652 * more than a single channel. Therefore, we can just adjust the source
653 * and destination registers for that channel and emit the instruction.
655 unsigned channel
= 0;
656 if (nir_op_infos
[instr
->op
].output_size
== 0) {
657 /* Since NIR is doing the scalarizing for us, we should only ever see
658 * vectorized operations with a single channel.
660 assert(_mesa_bitcount(instr
->dest
.write_mask
) == 1);
661 channel
= ffs(instr
->dest
.write_mask
) - 1;
663 result
= offset(result
, channel
);
666 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
667 assert(nir_op_infos
[instr
->op
].input_sizes
[i
] < 2);
668 op
[i
] = offset(op
[i
], instr
->src
[i
].swizzle
[channel
]);
674 inst
= bld
.MOV(result
, op
[0]);
675 inst
->saturate
= instr
->dest
.saturate
;
680 bld
.MOV(result
, op
[0]);
684 /* AND(val, 0x80000000) gives the sign bit.
686 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
689 bld
.CMP(bld
.null_reg_f(), op
[0], fs_reg(0.0f
), BRW_CONDITIONAL_NZ
);
691 fs_reg result_int
= retype(result
, BRW_REGISTER_TYPE_UD
);
692 op
[0].type
= BRW_REGISTER_TYPE_UD
;
693 result
.type
= BRW_REGISTER_TYPE_UD
;
694 bld
.AND(result_int
, op
[0], fs_reg(0x80000000u
));
696 inst
= bld
.OR(result_int
, result_int
, fs_reg(0x3f800000u
));
697 inst
->predicate
= BRW_PREDICATE_NORMAL
;
698 if (instr
->dest
.saturate
) {
699 inst
= bld
.MOV(result
, result
);
700 inst
->saturate
= true;
706 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
707 * -> non-negative val generates 0x00000000.
708 * Predicated OR sets 1 if val is positive.
710 bld
.CMP(bld
.null_reg_d(), op
[0], fs_reg(0), BRW_CONDITIONAL_G
);
711 bld
.ASR(result
, op
[0], fs_reg(31));
712 inst
= bld
.OR(result
, result
, fs_reg(1));
713 inst
->predicate
= BRW_PREDICATE_NORMAL
;
717 inst
= bld
.emit(SHADER_OPCODE_RCP
, result
, op
[0]);
718 inst
->saturate
= instr
->dest
.saturate
;
722 inst
= bld
.emit(SHADER_OPCODE_EXP2
, result
, op
[0]);
723 inst
->saturate
= instr
->dest
.saturate
;
727 inst
= bld
.emit(SHADER_OPCODE_LOG2
, result
, op
[0]);
728 inst
->saturate
= instr
->dest
.saturate
;
732 inst
= bld
.emit(SHADER_OPCODE_SIN
, result
, op
[0]);
733 inst
->saturate
= instr
->dest
.saturate
;
737 inst
= bld
.emit(SHADER_OPCODE_COS
, result
, op
[0]);
738 inst
->saturate
= instr
->dest
.saturate
;
742 if (fs_key
->high_quality_derivatives
) {
743 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
745 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
747 inst
->saturate
= instr
->dest
.saturate
;
749 case nir_op_fddx_fine
:
750 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
751 inst
->saturate
= instr
->dest
.saturate
;
753 case nir_op_fddx_coarse
:
754 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
755 inst
->saturate
= instr
->dest
.saturate
;
758 if (fs_key
->high_quality_derivatives
) {
759 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0],
760 fs_reg(fs_key
->render_to_fbo
));
762 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0],
763 fs_reg(fs_key
->render_to_fbo
));
765 inst
->saturate
= instr
->dest
.saturate
;
767 case nir_op_fddy_fine
:
768 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0],
769 fs_reg(fs_key
->render_to_fbo
));
770 inst
->saturate
= instr
->dest
.saturate
;
772 case nir_op_fddy_coarse
:
773 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0],
774 fs_reg(fs_key
->render_to_fbo
));
775 inst
->saturate
= instr
->dest
.saturate
;
780 inst
= bld
.ADD(result
, op
[0], op
[1]);
781 inst
->saturate
= instr
->dest
.saturate
;
785 inst
= bld
.MUL(result
, op
[0], op
[1]);
786 inst
->saturate
= instr
->dest
.saturate
;
790 bld
.MUL(result
, op
[0], op
[1]);
793 case nir_op_imul_high
:
794 case nir_op_umul_high
: {
795 if (devinfo
->gen
>= 7)
796 no16("SIMD16 explicit accumulator operands unsupported\n");
798 struct brw_reg acc
= retype(brw_acc_reg(dispatch_width
), result
.type
);
800 fs_inst
*mul
= bld
.MUL(acc
, op
[0], op
[1]);
801 bld
.MACH(result
, op
[0], op
[1]);
803 /* Until Gen8, integer multiplies read 32-bits from one source, and
804 * 16-bits from the other, and relying on the MACH instruction to
805 * generate the high bits of the result.
807 * On Gen8, the multiply instruction does a full 32x32-bit multiply,
808 * but in order to do a 64x64-bit multiply we have to simulate the
809 * previous behavior and then use a MACH instruction.
811 * FINISHME: Don't use source modifiers on src1.
813 if (devinfo
->gen
>= 8) {
814 assert(mul
->src
[1].type
== BRW_REGISTER_TYPE_D
||
815 mul
->src
[1].type
== BRW_REGISTER_TYPE_UD
);
816 if (mul
->src
[1].type
== BRW_REGISTER_TYPE_D
) {
817 mul
->src
[1].type
= BRW_REGISTER_TYPE_W
;
818 mul
->src
[1].stride
= 2;
820 mul
->src
[1].type
= BRW_REGISTER_TYPE_UW
;
821 mul
->src
[1].stride
= 2;
829 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, result
, op
[0], op
[1]);
832 case nir_op_uadd_carry
: {
833 if (devinfo
->gen
>= 7)
834 no16("SIMD16 explicit accumulator operands unsupported\n");
836 struct brw_reg acc
= retype(brw_acc_reg(dispatch_width
),
837 BRW_REGISTER_TYPE_UD
);
839 bld
.ADDC(bld
.null_reg_ud(), op
[0], op
[1]);
840 bld
.MOV(result
, fs_reg(acc
));
844 case nir_op_usub_borrow
: {
845 if (devinfo
->gen
>= 7)
846 no16("SIMD16 explicit accumulator operands unsupported\n");
848 struct brw_reg acc
= retype(brw_acc_reg(dispatch_width
),
849 BRW_REGISTER_TYPE_UD
);
851 bld
.SUBB(bld
.null_reg_ud(), op
[0], op
[1]);
852 bld
.MOV(result
, fs_reg(acc
));
857 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
863 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
869 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
874 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_Z
);
879 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_NZ
);
883 if (devinfo
->gen
>= 8) {
884 resolve_source_modifiers(&op
[0]);
886 bld
.NOT(result
, op
[0]);
889 if (devinfo
->gen
>= 8) {
890 resolve_source_modifiers(&op
[0]);
891 resolve_source_modifiers(&op
[1]);
893 bld
.XOR(result
, op
[0], op
[1]);
896 if (devinfo
->gen
>= 8) {
897 resolve_source_modifiers(&op
[0]);
898 resolve_source_modifiers(&op
[1]);
900 bld
.OR(result
, op
[0], op
[1]);
903 if (devinfo
->gen
>= 8) {
904 resolve_source_modifiers(&op
[0]);
905 resolve_source_modifiers(&op
[1]);
907 bld
.AND(result
, op
[0], op
[1]);
919 case nir_op_ball_fequal2
:
920 case nir_op_ball_iequal2
:
921 case nir_op_ball_fequal3
:
922 case nir_op_ball_iequal3
:
923 case nir_op_ball_fequal4
:
924 case nir_op_ball_iequal4
:
925 case nir_op_bany_fnequal2
:
926 case nir_op_bany_inequal2
:
927 case nir_op_bany_fnequal3
:
928 case nir_op_bany_inequal3
:
929 case nir_op_bany_fnequal4
:
930 case nir_op_bany_inequal4
:
931 unreachable("Lowered by nir_lower_alu_reductions");
933 case nir_op_fnoise1_1
:
934 case nir_op_fnoise1_2
:
935 case nir_op_fnoise1_3
:
936 case nir_op_fnoise1_4
:
937 case nir_op_fnoise2_1
:
938 case nir_op_fnoise2_2
:
939 case nir_op_fnoise2_3
:
940 case nir_op_fnoise2_4
:
941 case nir_op_fnoise3_1
:
942 case nir_op_fnoise3_2
:
943 case nir_op_fnoise3_3
:
944 case nir_op_fnoise3_4
:
945 case nir_op_fnoise4_1
:
946 case nir_op_fnoise4_2
:
947 case nir_op_fnoise4_3
:
948 case nir_op_fnoise4_4
:
949 unreachable("not reached: should be handled by lower_noise");
952 unreachable("not reached: should be handled by ldexp_to_arith()");
955 inst
= bld
.emit(SHADER_OPCODE_SQRT
, result
, op
[0]);
956 inst
->saturate
= instr
->dest
.saturate
;
960 inst
= bld
.emit(SHADER_OPCODE_RSQ
, result
, op
[0]);
961 inst
->saturate
= instr
->dest
.saturate
;
965 bld
.AND(result
, op
[0], fs_reg(1));
968 bld
.AND(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0], fs_reg(0x3f800000u
));
972 bld
.CMP(result
, op
[0], fs_reg(0.0f
), BRW_CONDITIONAL_NZ
);
975 bld
.CMP(result
, op
[0], fs_reg(0), BRW_CONDITIONAL_NZ
);
979 inst
= bld
.RNDZ(result
, op
[0]);
980 inst
->saturate
= instr
->dest
.saturate
;
984 op
[0].negate
= !op
[0].negate
;
985 fs_reg temp
= vgrf(glsl_type::float_type
);
986 bld
.RNDD(temp
, op
[0]);
988 inst
= bld
.MOV(result
, temp
);
989 inst
->saturate
= instr
->dest
.saturate
;
993 inst
= bld
.RNDD(result
, op
[0]);
994 inst
->saturate
= instr
->dest
.saturate
;
997 inst
= bld
.FRC(result
, op
[0]);
998 inst
->saturate
= instr
->dest
.saturate
;
1000 case nir_op_fround_even
:
1001 inst
= bld
.RNDE(result
, op
[0]);
1002 inst
->saturate
= instr
->dest
.saturate
;
1008 if (devinfo
->gen
>= 6) {
1009 inst
= bld
.emit(BRW_OPCODE_SEL
, result
, op
[0], op
[1]);
1010 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1012 bld
.CMP(bld
.null_reg_d(), op
[0], op
[1], BRW_CONDITIONAL_L
);
1013 inst
= bld
.SEL(result
, op
[0], op
[1]);
1014 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1016 inst
->saturate
= instr
->dest
.saturate
;
1022 if (devinfo
->gen
>= 6) {
1023 inst
= bld
.emit(BRW_OPCODE_SEL
, result
, op
[0], op
[1]);
1024 inst
->conditional_mod
= BRW_CONDITIONAL_GE
;
1026 bld
.CMP(bld
.null_reg_d(), op
[0], op
[1], BRW_CONDITIONAL_GE
);
1027 inst
= bld
.SEL(result
, op
[0], op
[1]);
1028 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1030 inst
->saturate
= instr
->dest
.saturate
;
1033 case nir_op_pack_snorm_2x16
:
1034 case nir_op_pack_snorm_4x8
:
1035 case nir_op_pack_unorm_2x16
:
1036 case nir_op_pack_unorm_4x8
:
1037 case nir_op_unpack_snorm_2x16
:
1038 case nir_op_unpack_snorm_4x8
:
1039 case nir_op_unpack_unorm_2x16
:
1040 case nir_op_unpack_unorm_4x8
:
1041 case nir_op_unpack_half_2x16
:
1042 case nir_op_pack_half_2x16
:
1043 unreachable("not reached: should be handled by lower_packing_builtins");
1045 case nir_op_unpack_half_2x16_split_x
:
1046 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
, result
, op
[0]);
1047 inst
->saturate
= instr
->dest
.saturate
;
1049 case nir_op_unpack_half_2x16_split_y
:
1050 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
, result
, op
[0]);
1051 inst
->saturate
= instr
->dest
.saturate
;
1055 inst
= bld
.emit(SHADER_OPCODE_POW
, result
, op
[0], op
[1]);
1056 inst
->saturate
= instr
->dest
.saturate
;
1059 case nir_op_bitfield_reverse
:
1060 bld
.BFREV(result
, op
[0]);
1063 case nir_op_bit_count
:
1064 bld
.CBIT(result
, op
[0]);
1067 case nir_op_ufind_msb
:
1068 case nir_op_ifind_msb
: {
1069 bld
.FBH(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0]);
1071 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1072 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1073 * subtract the result from 31 to convert the MSB count into an LSB count.
1076 bld
.CMP(bld
.null_reg_d(), result
, fs_reg(-1), BRW_CONDITIONAL_NZ
);
1077 fs_reg
neg_result(result
);
1078 neg_result
.negate
= true;
1079 inst
= bld
.ADD(result
, neg_result
, fs_reg(31));
1080 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1084 case nir_op_find_lsb
:
1085 bld
.FBL(result
, op
[0]);
1088 case nir_op_ubitfield_extract
:
1089 case nir_op_ibitfield_extract
:
1090 bld
.BFE(result
, op
[2], op
[1], op
[0]);
1093 bld
.BFI1(result
, op
[0], op
[1]);
1096 bld
.BFI2(result
, op
[0], op
[1], op
[2]);
1099 case nir_op_bitfield_insert
:
1100 unreachable("not reached: should be handled by "
1101 "lower_instructions::bitfield_insert_to_bfm_bfi");
1104 bld
.SHL(result
, op
[0], op
[1]);
1107 bld
.ASR(result
, op
[0], op
[1]);
1110 bld
.SHR(result
, op
[0], op
[1]);
1113 case nir_op_pack_half_2x16_split
:
1114 bld
.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, result
, op
[0], op
[1]);
1118 inst
= bld
.MAD(result
, op
[2], op
[1], op
[0]);
1119 inst
->saturate
= instr
->dest
.saturate
;
1123 inst
= bld
.LRP(result
, op
[0], op
[1], op
[2]);
1124 inst
->saturate
= instr
->dest
.saturate
;
1128 if (optimize_frontfacing_ternary(instr
, result
))
1131 bld
.CMP(bld
.null_reg_d(), op
[0], fs_reg(0), BRW_CONDITIONAL_NZ
);
1132 inst
= bld
.SEL(result
, op
[1], op
[2]);
1133 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1137 unreachable("unhandled instruction");
1140 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1141 * to sign extend the low bit to 0/~0
1143 if (devinfo
->gen
<= 5 &&
1144 (instr
->instr
.pass_flags
& BRW_NIR_BOOLEAN_MASK
) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE
) {
1145 fs_reg masked
= vgrf(glsl_type::int_type
);
1146 bld
.AND(masked
, result
, fs_reg(1));
1147 masked
.negate
= true;
1148 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), masked
);
1153 fs_reg_for_nir_reg(fs_visitor
*v
, nir_register
*nir_reg
,
1154 unsigned base_offset
, nir_src
*indirect
)
1157 if (nir_reg
->is_global
)
1158 reg
= v
->nir_globals
[nir_reg
->index
];
1160 reg
= v
->nir_locals
[nir_reg
->index
];
1162 reg
= offset(reg
, base_offset
* nir_reg
->num_components
);
1164 int multiplier
= nir_reg
->num_components
* (v
->dispatch_width
/ 8);
1166 reg
.reladdr
= new(v
->mem_ctx
) fs_reg(v
->vgrf(glsl_type::int_type
));
1167 v
->bld
.MUL(*reg
.reladdr
, v
->get_nir_src(*indirect
),
1168 fs_reg(multiplier
));
1175 fs_visitor::get_nir_src(nir_src src
)
1178 assert(src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
);
1179 nir_load_const_instr
*load
= nir_instr_as_load_const(src
.ssa
->parent_instr
);
1180 fs_reg reg
= bld
.vgrf(BRW_REGISTER_TYPE_D
, src
.ssa
->num_components
);
1182 for (unsigned i
= 0; i
< src
.ssa
->num_components
; ++i
)
1183 bld
.MOV(offset(reg
, i
), fs_reg(load
->value
.i
[i
]));
1187 fs_reg reg
= fs_reg_for_nir_reg(this, src
.reg
.reg
, src
.reg
.base_offset
,
1190 /* to avoid floating-point denorm flushing problems, set the type by
1191 * default to D - instructions that need floating point semantics will set
1192 * this to F if they need to
1194 return retype(reg
, BRW_REGISTER_TYPE_D
);
1199 fs_visitor::get_nir_dest(nir_dest dest
)
1201 return fs_reg_for_nir_reg(this, dest
.reg
.reg
, dest
.reg
.base_offset
,
1206 fs_visitor::emit_percomp(const fs_builder
&bld
, const fs_inst
&inst
,
1209 for (unsigned i
= 0; i
< 4; i
++) {
1210 if (!((wr_mask
>> i
) & 1))
1213 fs_inst
*new_inst
= new(mem_ctx
) fs_inst(inst
);
1214 new_inst
->dst
= offset(new_inst
->dst
, i
);
1215 for (unsigned j
= 0; j
< new_inst
->sources
; j
++)
1216 if (new_inst
->src
[j
].file
== GRF
)
1217 new_inst
->src
[j
] = offset(new_inst
->src
[j
], i
);
1224 fs_visitor::nir_emit_intrinsic(nir_intrinsic_instr
*instr
)
1227 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
1228 dest
= get_nir_dest(instr
->dest
);
1230 bool has_indirect
= false;
1232 switch (instr
->intrinsic
) {
1233 case nir_intrinsic_discard
:
1234 case nir_intrinsic_discard_if
: {
1235 /* We track our discarded pixels in f0.1. By predicating on it, we can
1236 * update just the flag bits that aren't yet discarded. If there's no
1237 * condition, we emit a CMP of g0 != g0, so all currently executing
1238 * channels will get turned off.
1241 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
1242 cmp
= emit(CMP(reg_null_f
, get_nir_src(instr
->src
[0]),
1243 fs_reg(0), BRW_CONDITIONAL_Z
));
1245 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
1246 BRW_REGISTER_TYPE_UW
));
1247 cmp
= emit(CMP(reg_null_f
, some_reg
, some_reg
, BRW_CONDITIONAL_NZ
));
1249 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
1250 cmp
->flag_subreg
= 1;
1252 if (devinfo
->gen
>= 6) {
1253 emit_discard_jump();
1258 case nir_intrinsic_atomic_counter_inc
:
1259 case nir_intrinsic_atomic_counter_dec
:
1260 case nir_intrinsic_atomic_counter_read
: {
1261 unsigned surf_index
= prog_data
->binding_table
.abo_start
+
1262 (unsigned) instr
->const_index
[0];
1263 fs_reg offset
= fs_reg(get_nir_src(instr
->src
[0]));
1265 switch (instr
->intrinsic
) {
1266 case nir_intrinsic_atomic_counter_inc
:
1267 emit_untyped_atomic(BRW_AOP_INC
, surf_index
, dest
, offset
,
1268 fs_reg(), fs_reg());
1270 case nir_intrinsic_atomic_counter_dec
:
1271 emit_untyped_atomic(BRW_AOP_PREDEC
, surf_index
, dest
, offset
,
1272 fs_reg(), fs_reg());
1274 case nir_intrinsic_atomic_counter_read
:
1275 emit_untyped_surface_read(surf_index
, dest
, offset
);
1278 unreachable("Unreachable");
1283 case nir_intrinsic_load_front_face
:
1284 emit(MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
1285 *emit_frontfacing_interpolation()));
1288 case nir_intrinsic_load_vertex_id
:
1289 unreachable("should be lowered by lower_vertex_id()");
1291 case nir_intrinsic_load_vertex_id_zero_base
: {
1292 fs_reg vertex_id
= nir_system_values
[SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
];
1293 assert(vertex_id
.file
!= BAD_FILE
);
1294 dest
.type
= vertex_id
.type
;
1295 emit(MOV(dest
, vertex_id
));
1299 case nir_intrinsic_load_base_vertex
: {
1300 fs_reg base_vertex
= nir_system_values
[SYSTEM_VALUE_BASE_VERTEX
];
1301 assert(base_vertex
.file
!= BAD_FILE
);
1302 dest
.type
= base_vertex
.type
;
1303 emit(MOV(dest
, base_vertex
));
1307 case nir_intrinsic_load_instance_id
: {
1308 fs_reg instance_id
= nir_system_values
[SYSTEM_VALUE_INSTANCE_ID
];
1309 assert(instance_id
.file
!= BAD_FILE
);
1310 dest
.type
= instance_id
.type
;
1311 emit(MOV(dest
, instance_id
));
1315 case nir_intrinsic_load_sample_mask_in
: {
1316 fs_reg sample_mask_in
= nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
1317 assert(sample_mask_in
.file
!= BAD_FILE
);
1318 dest
.type
= sample_mask_in
.type
;
1319 emit(MOV(dest
, sample_mask_in
));
1323 case nir_intrinsic_load_sample_pos
: {
1324 fs_reg sample_pos
= nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
1325 assert(sample_pos
.file
!= BAD_FILE
);
1326 dest
.type
= sample_pos
.type
;
1327 emit(MOV(dest
, sample_pos
));
1328 emit(MOV(offset(dest
, 1), offset(sample_pos
, 1)));
1332 case nir_intrinsic_load_sample_id
: {
1333 fs_reg sample_id
= nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
1334 assert(sample_id
.file
!= BAD_FILE
);
1335 dest
.type
= sample_id
.type
;
1336 emit(MOV(dest
, sample_id
));
1340 case nir_intrinsic_load_uniform_indirect
:
1341 has_indirect
= true;
1343 case nir_intrinsic_load_uniform
: {
1344 unsigned index
= instr
->const_index
[0];
1347 if (index
< num_direct_uniforms
) {
1348 uniform_reg
= fs_reg(UNIFORM
, 0);
1350 uniform_reg
= fs_reg(UNIFORM
, num_direct_uniforms
);
1351 index
-= num_direct_uniforms
;
1354 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
1355 fs_reg src
= offset(retype(uniform_reg
, dest
.type
), index
);
1357 src
.reladdr
= new(mem_ctx
) fs_reg(get_nir_src(instr
->src
[0]));
1360 emit(MOV(dest
, src
));
1361 dest
= offset(dest
, 1);
1366 case nir_intrinsic_load_ubo_indirect
:
1367 has_indirect
= true;
1369 case nir_intrinsic_load_ubo
: {
1370 nir_const_value
*const_index
= nir_src_as_const_value(instr
->src
[0]);
1374 surf_index
= fs_reg(stage_prog_data
->binding_table
.ubo_start
+
1377 /* The block index is not a constant. Evaluate the index expression
1378 * per-channel and add the base UBO index; we have to select a value
1379 * from any live channel.
1381 surf_index
= vgrf(glsl_type::uint_type
);
1382 emit(ADD(surf_index
, get_nir_src(instr
->src
[0]),
1383 fs_reg(stage_prog_data
->binding_table
.ubo_start
)));
1384 emit_uniformize(surf_index
, surf_index
);
1386 /* Assume this may touch any UBO. It would be nice to provide
1387 * a tighter bound, but the array information is already lowered away.
1389 brw_mark_surface_used(prog_data
,
1390 stage_prog_data
->binding_table
.ubo_start
+
1391 shader_prog
->NumUniformBlocks
- 1);
1395 /* Turn the byte offset into a dword offset. */
1396 fs_reg base_offset
= vgrf(glsl_type::int_type
);
1397 emit(SHR(base_offset
, retype(get_nir_src(instr
->src
[1]),
1398 BRW_REGISTER_TYPE_D
),
1401 unsigned vec4_offset
= instr
->const_index
[0] / 4;
1402 for (int i
= 0; i
< instr
->num_components
; i
++)
1403 VARYING_PULL_CONSTANT_LOAD(bld
, offset(dest
, i
), surf_index
,
1404 base_offset
, vec4_offset
+ i
);
1406 fs_reg packed_consts
= vgrf(glsl_type::float_type
);
1407 packed_consts
.type
= dest
.type
;
1409 fs_reg
const_offset_reg((unsigned) instr
->const_index
[0] & ~15);
1410 emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
, packed_consts
,
1411 surf_index
, const_offset_reg
);
1413 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
1414 packed_consts
.set_smear(instr
->const_index
[0] % 16 / 4 + i
);
1416 /* The std140 packing rules don't allow vectors to cross 16-byte
1417 * boundaries, and a reg is 32 bytes.
1419 assert(packed_consts
.subreg_offset
< 32);
1421 emit(MOV(dest
, packed_consts
));
1422 dest
= offset(dest
, 1);
1428 case nir_intrinsic_load_input_indirect
:
1429 has_indirect
= true;
1431 case nir_intrinsic_load_input
: {
1433 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
1434 fs_reg src
= offset(retype(nir_inputs
, dest
.type
),
1435 instr
->const_index
[0] + index
);
1437 src
.reladdr
= new(mem_ctx
) fs_reg(get_nir_src(instr
->src
[0]));
1440 emit(MOV(dest
, src
));
1441 dest
= offset(dest
, 1);
1446 /* Handle ARB_gpu_shader5 interpolation intrinsics
1448 * It's worth a quick word of explanation as to why we handle the full
1449 * variable-based interpolation intrinsic rather than a lowered version
1450 * with like we do for other inputs. We have to do that because the way
1451 * we set up inputs doesn't allow us to use the already setup inputs for
1452 * interpolation. At the beginning of the shader, we go through all of
1453 * the input variables and do the initial interpolation and put it in
1454 * the nir_inputs array based on its location as determined in
1455 * nir_lower_io. If the input isn't used, dead code cleans up and
1456 * everything works fine. However, when we get to the ARB_gpu_shader5
1457 * interpolation intrinsics, we need to reinterpolate the input
1458 * differently. If we used an intrinsic that just had an index it would
1459 * only give us the offset into the nir_inputs array. However, this is
1460 * useless because that value is post-interpolation and we need
1461 * pre-interpolation. In order to get the actual location of the bits
1462 * we get from the vertex fetching hardware, we need the variable.
1464 case nir_intrinsic_interp_var_at_centroid
:
1465 case nir_intrinsic_interp_var_at_sample
:
1466 case nir_intrinsic_interp_var_at_offset
: {
1467 /* in SIMD16 mode, the pixel interpolator returns coords interleaved
1468 * 8 channels at a time, same as the barycentric coords presented in
1469 * the FS payload. this requires a bit of extra work to support.
1471 no16("interpolate_at_* not yet supported in SIMD16 mode.");
1473 fs_reg dst_xy
= vgrf(2);
1475 /* For most messages, we need one reg of ignored data; the hardware
1476 * requires mlen==1 even when there is no payload. in the per-slot
1477 * offset case, we'll replace this with the proper source data.
1479 fs_reg src
= vgrf(glsl_type::float_type
);
1480 int mlen
= 1; /* one reg unless overriden */
1483 switch (instr
->intrinsic
) {
1484 case nir_intrinsic_interp_var_at_centroid
:
1485 inst
= emit(FS_OPCODE_INTERPOLATE_AT_CENTROID
, dst_xy
, src
, fs_reg(0u));
1488 case nir_intrinsic_interp_var_at_sample
: {
1489 /* XXX: We should probably handle non-constant sample id's */
1490 nir_const_value
*const_sample
= nir_src_as_const_value(instr
->src
[0]);
1491 assert(const_sample
);
1492 unsigned msg_data
= const_sample
? const_sample
->i
[0] << 4 : 0;
1493 inst
= emit(FS_OPCODE_INTERPOLATE_AT_SAMPLE
, dst_xy
, src
,
1498 case nir_intrinsic_interp_var_at_offset
: {
1499 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
1502 unsigned off_x
= MIN2((int)(const_offset
->f
[0] * 16), 7) & 0xf;
1503 unsigned off_y
= MIN2((int)(const_offset
->f
[1] * 16), 7) & 0xf;
1505 inst
= emit(FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
, dst_xy
, src
,
1506 fs_reg(off_x
| (off_y
<< 4)));
1508 src
= vgrf(glsl_type::ivec2_type
);
1509 fs_reg offset_src
= retype(get_nir_src(instr
->src
[0]),
1510 BRW_REGISTER_TYPE_F
);
1511 for (int i
= 0; i
< 2; i
++) {
1512 fs_reg temp
= vgrf(glsl_type::float_type
);
1513 emit(MUL(temp
, offset(offset_src
, i
), fs_reg(16.0f
)));
1514 fs_reg itemp
= vgrf(glsl_type::int_type
);
1515 emit(MOV(itemp
, temp
)); /* float to int */
1517 /* Clamp the upper end of the range to +7/16.
1518 * ARB_gpu_shader5 requires that we support a maximum offset
1519 * of +0.5, which isn't representable in a S0.4 value -- if
1520 * we didn't clamp it, we'd end up with -8/16, which is the
1521 * opposite of what the shader author wanted.
1523 * This is legal due to ARB_gpu_shader5's quantization
1526 * "Not all values of <offset> may be supported; x and y
1527 * offsets may be rounded to fixed-point values with the
1528 * number of fraction bits given by the
1529 * implementation-dependent constant
1530 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
1533 emit(BRW_OPCODE_SEL
, offset(src
, i
), itemp
, fs_reg(7))
1534 ->conditional_mod
= BRW_CONDITIONAL_L
; /* min(src2, 7) */
1538 inst
= emit(FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
, dst_xy
, src
,
1545 unreachable("Invalid intrinsic");
1549 inst
->regs_written
= 2; /* 2 floats per slot returned */
1550 inst
->pi_noperspective
= instr
->variables
[0]->var
->data
.interpolation
==
1551 INTERP_QUALIFIER_NOPERSPECTIVE
;
1553 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
1554 fs_reg src
= interp_reg(instr
->variables
[0]->var
->data
.location
, j
);
1555 src
.type
= dest
.type
;
1557 emit(FS_OPCODE_LINTERP
, dest
, dst_xy
, src
);
1558 dest
= offset(dest
, 1);
1563 case nir_intrinsic_store_output_indirect
:
1564 has_indirect
= true;
1566 case nir_intrinsic_store_output
: {
1567 fs_reg src
= get_nir_src(instr
->src
[0]);
1569 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
1570 fs_reg new_dest
= offset(retype(nir_outputs
, src
.type
),
1571 instr
->const_index
[0] + index
);
1573 src
.reladdr
= new(mem_ctx
) fs_reg(get_nir_src(instr
->src
[1]));
1575 emit(MOV(new_dest
, src
));
1576 src
= offset(src
, 1);
1582 unreachable("unknown intrinsic");
1587 fs_visitor::nir_emit_texture(nir_tex_instr
*instr
)
1589 unsigned sampler
= instr
->sampler_index
;
1590 fs_reg
sampler_reg(sampler
);
1592 /* FINISHME: We're failing to recompile our programs when the sampler is
1593 * updated. This only matters for the texture rectangle scale parameters
1594 * (pre-gen6, or gen6+ with GL_CLAMP).
1596 int texunit
= prog
->SamplerUnits
[sampler
];
1598 int gather_component
= instr
->component
;
1600 bool is_rect
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_RECT
;
1602 bool is_cube_array
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
1605 int lod_components
= 0, offset_components
= 0;
1607 fs_reg coordinate
, shadow_comparitor
, lod
, lod2
, sample_index
, mcs
, tex_offset
;
1609 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
1610 fs_reg src
= get_nir_src(instr
->src
[i
].src
);
1611 switch (instr
->src
[i
].src_type
) {
1612 case nir_tex_src_bias
:
1613 lod
= retype(src
, BRW_REGISTER_TYPE_F
);
1615 case nir_tex_src_comparitor
:
1616 shadow_comparitor
= retype(src
, BRW_REGISTER_TYPE_F
);
1618 case nir_tex_src_coord
:
1619 switch (instr
->op
) {
1621 case nir_texop_txf_ms
:
1622 coordinate
= retype(src
, BRW_REGISTER_TYPE_D
);
1625 coordinate
= retype(src
, BRW_REGISTER_TYPE_F
);
1629 case nir_tex_src_ddx
:
1630 lod
= retype(src
, BRW_REGISTER_TYPE_F
);
1631 lod_components
= nir_tex_instr_src_size(instr
, i
);
1633 case nir_tex_src_ddy
:
1634 lod2
= retype(src
, BRW_REGISTER_TYPE_F
);
1636 case nir_tex_src_lod
:
1637 switch (instr
->op
) {
1639 lod
= retype(src
, BRW_REGISTER_TYPE_UD
);
1642 lod
= retype(src
, BRW_REGISTER_TYPE_D
);
1645 lod
= retype(src
, BRW_REGISTER_TYPE_F
);
1649 case nir_tex_src_ms_index
:
1650 sample_index
= retype(src
, BRW_REGISTER_TYPE_UD
);
1652 case nir_tex_src_offset
:
1653 tex_offset
= retype(src
, BRW_REGISTER_TYPE_D
);
1654 if (instr
->is_array
)
1655 offset_components
= instr
->coord_components
- 1;
1657 offset_components
= instr
->coord_components
;
1659 case nir_tex_src_projector
:
1660 unreachable("should be lowered");
1662 case nir_tex_src_sampler_offset
: {
1663 /* Figure out the highest possible sampler index and mark it as used */
1664 uint32_t max_used
= sampler
+ instr
->sampler_array_size
- 1;
1665 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
< 8) {
1666 max_used
+= stage_prog_data
->binding_table
.gather_texture_start
;
1668 max_used
+= stage_prog_data
->binding_table
.texture_start
;
1670 brw_mark_surface_used(prog_data
, max_used
);
1672 /* Emit code to evaluate the actual indexing expression */
1673 sampler_reg
= vgrf(glsl_type::uint_type
);
1674 emit(ADD(sampler_reg
, src
, fs_reg(sampler
)));
1675 emit_uniformize(sampler_reg
, sampler_reg
);
1680 unreachable("unknown texture source");
1684 if (instr
->op
== nir_texop_txf_ms
) {
1685 if (devinfo
->gen
>= 7 &&
1686 key_tex
->compressed_multisample_layout_mask
& (1 << sampler
)) {
1687 mcs
= emit_mcs_fetch(coordinate
, instr
->coord_components
, sampler_reg
);
1693 for (unsigned i
= 0; i
< 3; i
++) {
1694 if (instr
->const_offset
[i
] != 0) {
1695 assert(offset_components
== 0);
1696 tex_offset
= fs_reg(brw_texture_offset(instr
->const_offset
, 3));
1701 enum glsl_base_type dest_base_type
;
1702 switch (instr
->dest_type
) {
1703 case nir_type_float
:
1704 dest_base_type
= GLSL_TYPE_FLOAT
;
1707 dest_base_type
= GLSL_TYPE_INT
;
1709 case nir_type_unsigned
:
1710 dest_base_type
= GLSL_TYPE_UINT
;
1713 unreachable("bad type");
1716 const glsl_type
*dest_type
=
1717 glsl_type::get_instance(dest_base_type
, nir_tex_instr_dest_size(instr
),
1720 ir_texture_opcode op
;
1721 switch (instr
->op
) {
1722 case nir_texop_lod
: op
= ir_lod
; break;
1723 case nir_texop_query_levels
: op
= ir_query_levels
; break;
1724 case nir_texop_tex
: op
= ir_tex
; break;
1725 case nir_texop_tg4
: op
= ir_tg4
; break;
1726 case nir_texop_txb
: op
= ir_txb
; break;
1727 case nir_texop_txd
: op
= ir_txd
; break;
1728 case nir_texop_txf
: op
= ir_txf
; break;
1729 case nir_texop_txf_ms
: op
= ir_txf_ms
; break;
1730 case nir_texop_txl
: op
= ir_txl
; break;
1731 case nir_texop_txs
: op
= ir_txs
; break;
1733 unreachable("unknown texture opcode");
1736 emit_texture(op
, dest_type
, coordinate
, instr
->coord_components
,
1737 shadow_comparitor
, lod
, lod2
, lod_components
, sample_index
,
1738 tex_offset
, mcs
, gather_component
,
1739 is_cube_array
, is_rect
, sampler
, sampler_reg
, texunit
);
1741 fs_reg dest
= get_nir_dest(instr
->dest
);
1742 dest
.type
= this->result
.type
;
1743 unsigned num_components
= nir_tex_instr_dest_size(instr
);
1744 emit_percomp(bld
, fs_inst(BRW_OPCODE_MOV
, dest
, this->result
),
1745 (1 << num_components
) - 1);
1749 fs_visitor::nir_emit_jump(const fs_builder
&bld
, nir_jump_instr
*instr
)
1751 switch (instr
->type
) {
1752 case nir_jump_break
:
1753 bld
.emit(BRW_OPCODE_BREAK
);
1755 case nir_jump_continue
:
1756 bld
.emit(BRW_OPCODE_CONTINUE
);
1758 case nir_jump_return
:
1760 unreachable("unknown jump");