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
24 #include "compiler/glsl/ir.h"
26 #include "brw_fs_surface_builder.h"
28 #include "brw_program.h"
31 using namespace brw::surface_access
;
34 fs_visitor::emit_nir_code()
36 /* emit the arrays used for inputs and outputs - load/store intrinsics will
37 * be converted to reads/writes of these arrays
42 nir_emit_system_values();
44 /* get the main function and emit it */
45 nir_foreach_function(function
, nir
) {
46 assert(strcmp(function
->name
, "main") == 0);
47 assert(function
->impl
);
48 nir_emit_impl(function
->impl
);
53 fs_visitor::nir_setup_inputs()
55 if (stage
!= MESA_SHADER_FRAGMENT
)
58 nir_inputs
= bld
.vgrf(BRW_REGISTER_TYPE_F
, nir
->num_inputs
);
60 nir_foreach_variable(var
, &nir
->inputs
) {
61 fs_reg input
= offset(nir_inputs
, bld
, var
->data
.driver_location
);
64 if (var
->data
.location
== VARYING_SLOT_POS
) {
65 reg
= *emit_fragcoord_interpolation(var
->data
.pixel_center_integer
,
66 var
->data
.origin_upper_left
);
67 emit_percomp(bld
, fs_inst(BRW_OPCODE_MOV
, bld
.dispatch_width(),
69 } else if (var
->data
.location
== VARYING_SLOT_LAYER
) {
70 struct brw_reg reg
= suboffset(interp_reg(VARYING_SLOT_LAYER
, 1), 3);
71 reg
.type
= BRW_REGISTER_TYPE_D
;
72 bld
.emit(FS_OPCODE_CINTERP
, retype(input
, BRW_REGISTER_TYPE_D
), reg
);
73 } else if (var
->data
.location
== VARYING_SLOT_VIEWPORT
) {
74 struct brw_reg reg
= suboffset(interp_reg(VARYING_SLOT_VIEWPORT
, 2), 3);
75 reg
.type
= BRW_REGISTER_TYPE_D
;
76 bld
.emit(FS_OPCODE_CINTERP
, retype(input
, BRW_REGISTER_TYPE_D
), reg
);
78 int location
= var
->data
.location
;
79 emit_general_interpolation(&input
, var
->name
, var
->type
,
80 (glsl_interp_qualifier
) var
->data
.interpolation
,
81 &location
, var
->data
.centroid
,
88 fs_visitor::nir_setup_single_output_varying(fs_reg
*reg
,
89 const glsl_type
*type
,
92 if (type
->is_array() || type
->is_matrix()) {
93 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
94 const unsigned length
= glsl_get_length(type
);
96 for (unsigned i
= 0; i
< length
; i
++) {
97 nir_setup_single_output_varying(reg
, elem_type
, location
);
99 } else if (type
->is_record()) {
100 for (unsigned i
= 0; i
< type
->length
; i
++) {
101 const struct glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
102 nir_setup_single_output_varying(reg
, field_type
, location
);
105 assert(type
->is_scalar() || type
->is_vector());
106 this->outputs
[*location
] = *reg
;
107 this->output_components
[*location
] = type
->vector_elements
;
108 *reg
= offset(*reg
, bld
, 4);
114 fs_visitor::nir_setup_outputs()
116 if (stage
== MESA_SHADER_TESS_CTRL
)
119 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
121 nir_outputs
= bld
.vgrf(BRW_REGISTER_TYPE_F
, nir
->num_outputs
);
123 nir_foreach_variable(var
, &nir
->outputs
) {
124 fs_reg reg
= offset(nir_outputs
, bld
, var
->data
.driver_location
);
127 case MESA_SHADER_VERTEX
:
128 case MESA_SHADER_TESS_EVAL
:
129 case MESA_SHADER_GEOMETRY
: {
130 unsigned location
= var
->data
.location
;
131 nir_setup_single_output_varying(®
, var
->type
, &location
);
134 case MESA_SHADER_FRAGMENT
:
135 if (key
->force_dual_color_blend
&&
136 var
->data
.location
== FRAG_RESULT_DATA1
) {
137 this->dual_src_output
= reg
;
138 this->do_dual_src
= true;
139 } else if (var
->data
.index
> 0) {
140 assert(var
->data
.location
== FRAG_RESULT_DATA0
);
141 assert(var
->data
.index
== 1);
142 this->dual_src_output
= reg
;
143 this->do_dual_src
= true;
144 } else if (var
->data
.location
== FRAG_RESULT_COLOR
) {
145 /* Writing gl_FragColor outputs to all color regions. */
146 for (unsigned int i
= 0; i
< MAX2(key
->nr_color_regions
, 1); i
++) {
147 this->outputs
[i
] = reg
;
148 this->output_components
[i
] = 4;
150 } else if (var
->data
.location
== FRAG_RESULT_DEPTH
) {
151 this->frag_depth
= reg
;
152 } else if (var
->data
.location
== FRAG_RESULT_STENCIL
) {
153 this->frag_stencil
= reg
;
154 } else if (var
->data
.location
== FRAG_RESULT_SAMPLE_MASK
) {
155 this->sample_mask
= reg
;
157 int vector_elements
= var
->type
->without_array()->vector_elements
;
159 /* gl_FragData or a user-defined FS output */
160 assert(var
->data
.location
>= FRAG_RESULT_DATA0
&&
161 var
->data
.location
< FRAG_RESULT_DATA0
+BRW_MAX_DRAW_BUFFERS
);
163 /* General color output. */
164 for (unsigned int i
= 0; i
< MAX2(1, var
->type
->length
); i
++) {
165 int output
= var
->data
.location
- FRAG_RESULT_DATA0
+ i
;
166 this->outputs
[output
] = offset(reg
, bld
, vector_elements
* i
);
167 this->output_components
[output
] = vector_elements
;
172 unreachable("unhandled shader stage");
178 fs_visitor::nir_setup_uniforms()
180 if (dispatch_width
!= 8)
183 uniforms
= nir
->num_uniforms
/ 4;
187 emit_system_values_block(nir_block
*block
, fs_visitor
*v
)
191 nir_foreach_instr(instr
, block
) {
192 if (instr
->type
!= nir_instr_type_intrinsic
)
195 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
196 switch (intrin
->intrinsic
) {
197 case nir_intrinsic_load_vertex_id
:
198 unreachable("should be lowered by lower_vertex_id().");
200 case nir_intrinsic_load_vertex_id_zero_base
:
201 assert(v
->stage
== MESA_SHADER_VERTEX
);
202 reg
= &v
->nir_system_values
[SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
];
203 if (reg
->file
== BAD_FILE
)
204 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
);
207 case nir_intrinsic_load_base_vertex
:
208 assert(v
->stage
== MESA_SHADER_VERTEX
);
209 reg
= &v
->nir_system_values
[SYSTEM_VALUE_BASE_VERTEX
];
210 if (reg
->file
== BAD_FILE
)
211 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_BASE_VERTEX
);
214 case nir_intrinsic_load_instance_id
:
215 assert(v
->stage
== MESA_SHADER_VERTEX
);
216 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INSTANCE_ID
];
217 if (reg
->file
== BAD_FILE
)
218 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_INSTANCE_ID
);
221 case nir_intrinsic_load_base_instance
:
222 assert(v
->stage
== MESA_SHADER_VERTEX
);
223 reg
= &v
->nir_system_values
[SYSTEM_VALUE_BASE_INSTANCE
];
224 if (reg
->file
== BAD_FILE
)
225 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_BASE_INSTANCE
);
228 case nir_intrinsic_load_draw_id
:
229 assert(v
->stage
== MESA_SHADER_VERTEX
);
230 reg
= &v
->nir_system_values
[SYSTEM_VALUE_DRAW_ID
];
231 if (reg
->file
== BAD_FILE
)
232 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_DRAW_ID
);
235 case nir_intrinsic_load_invocation_id
:
236 if (v
->stage
== MESA_SHADER_TESS_CTRL
)
238 assert(v
->stage
== MESA_SHADER_GEOMETRY
);
239 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
240 if (reg
->file
== BAD_FILE
) {
241 const fs_builder abld
= v
->bld
.annotate("gl_InvocationID", NULL
);
242 fs_reg
g1(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
243 fs_reg iid
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
244 abld
.SHR(iid
, g1
, brw_imm_ud(27u));
249 case nir_intrinsic_load_sample_pos
:
250 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
251 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
252 if (reg
->file
== BAD_FILE
)
253 *reg
= *v
->emit_samplepos_setup();
256 case nir_intrinsic_load_sample_id
:
257 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
258 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
259 if (reg
->file
== BAD_FILE
)
260 *reg
= *v
->emit_sampleid_setup();
263 case nir_intrinsic_load_sample_mask_in
:
264 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
265 assert(v
->devinfo
->gen
>= 7);
266 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
267 if (reg
->file
== BAD_FILE
)
268 *reg
= *v
->emit_samplemaskin_setup();
271 case nir_intrinsic_load_local_invocation_id
:
272 assert(v
->stage
== MESA_SHADER_COMPUTE
);
273 reg
= &v
->nir_system_values
[SYSTEM_VALUE_LOCAL_INVOCATION_ID
];
274 if (reg
->file
== BAD_FILE
)
275 *reg
= *v
->emit_cs_local_invocation_id_setup();
278 case nir_intrinsic_load_work_group_id
:
279 assert(v
->stage
== MESA_SHADER_COMPUTE
);
280 reg
= &v
->nir_system_values
[SYSTEM_VALUE_WORK_GROUP_ID
];
281 if (reg
->file
== BAD_FILE
)
282 *reg
= *v
->emit_cs_work_group_id_setup();
285 case nir_intrinsic_load_helper_invocation
:
286 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
287 reg
= &v
->nir_system_values
[SYSTEM_VALUE_HELPER_INVOCATION
];
288 if (reg
->file
== BAD_FILE
) {
289 const fs_builder abld
=
290 v
->bld
.annotate("gl_HelperInvocation", NULL
);
292 /* On Gen6+ (gl_HelperInvocation is only exposed on Gen7+) the
293 * pixel mask is in g1.7 of the thread payload.
295 * We move the per-channel pixel enable bit to the low bit of each
296 * channel by shifting the byte containing the pixel mask by the
297 * vector immediate 0x76543210UV.
299 * The region of <1,8,0> reads only 1 byte (the pixel masks for
300 * subspans 0 and 1) in SIMD8 and an additional byte (the pixel
301 * masks for 2 and 3) in SIMD16.
303 fs_reg shifted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
305 stride(byte_offset(retype(brw_vec1_grf(1, 0),
306 BRW_REGISTER_TYPE_UB
), 28),
308 brw_imm_uv(0x76543210));
310 /* A set bit in the pixel mask means the channel is enabled, but
311 * that is the opposite of gl_HelperInvocation so we need to invert
314 * The negate source-modifier bit of logical instructions on Gen8+
315 * performs 1's complement negation, so we can use that instead of
318 fs_reg inverted
= negate(shifted
);
319 if (v
->devinfo
->gen
< 8) {
320 inverted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
321 abld
.NOT(inverted
, shifted
);
324 /* We then resolve the 0/1 result to 0/~0 boolean values by ANDing
325 * with 1 and negating.
327 fs_reg anded
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
328 abld
.AND(anded
, inverted
, brw_imm_uw(1));
330 fs_reg dst
= abld
.vgrf(BRW_REGISTER_TYPE_D
, 1);
331 abld
.MOV(dst
, negate(retype(anded
, BRW_REGISTER_TYPE_D
)));
345 fs_visitor::nir_emit_system_values()
347 nir_system_values
= ralloc_array(mem_ctx
, fs_reg
, SYSTEM_VALUE_MAX
);
348 for (unsigned i
= 0; i
< SYSTEM_VALUE_MAX
; i
++) {
349 nir_system_values
[i
] = fs_reg();
352 nir_foreach_function(function
, nir
) {
353 assert(strcmp(function
->name
, "main") == 0);
354 assert(function
->impl
);
355 nir_foreach_block(block
, function
->impl
) {
356 emit_system_values_block(block
, this);
362 fs_visitor::nir_emit_impl(nir_function_impl
*impl
)
364 nir_locals
= ralloc_array(mem_ctx
, fs_reg
, impl
->reg_alloc
);
365 for (unsigned i
= 0; i
< impl
->reg_alloc
; i
++) {
366 nir_locals
[i
] = fs_reg();
369 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
370 unsigned array_elems
=
371 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
372 unsigned size
= array_elems
* reg
->num_components
;
373 const brw_reg_type reg_type
=
374 reg
->bit_size
== 32 ? BRW_REGISTER_TYPE_F
: BRW_REGISTER_TYPE_DF
;
375 nir_locals
[reg
->index
] = bld
.vgrf(reg_type
, size
);
378 nir_ssa_values
= reralloc(mem_ctx
, nir_ssa_values
, fs_reg
,
381 nir_emit_cf_list(&impl
->body
);
385 fs_visitor::nir_emit_cf_list(exec_list
*list
)
387 exec_list_validate(list
);
388 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
389 switch (node
->type
) {
391 nir_emit_if(nir_cf_node_as_if(node
));
394 case nir_cf_node_loop
:
395 nir_emit_loop(nir_cf_node_as_loop(node
));
398 case nir_cf_node_block
:
399 nir_emit_block(nir_cf_node_as_block(node
));
403 unreachable("Invalid CFG node block");
409 fs_visitor::nir_emit_if(nir_if
*if_stmt
)
411 /* first, put the condition into f0 */
412 fs_inst
*inst
= bld
.MOV(bld
.null_reg_d(),
413 retype(get_nir_src(if_stmt
->condition
),
414 BRW_REGISTER_TYPE_D
));
415 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
417 bld
.IF(BRW_PREDICATE_NORMAL
);
419 nir_emit_cf_list(&if_stmt
->then_list
);
421 /* note: if the else is empty, dead CF elimination will remove it */
422 bld
.emit(BRW_OPCODE_ELSE
);
424 nir_emit_cf_list(&if_stmt
->else_list
);
426 bld
.emit(BRW_OPCODE_ENDIF
);
430 fs_visitor::nir_emit_loop(nir_loop
*loop
)
432 bld
.emit(BRW_OPCODE_DO
);
434 nir_emit_cf_list(&loop
->body
);
436 bld
.emit(BRW_OPCODE_WHILE
);
440 fs_visitor::nir_emit_block(nir_block
*block
)
442 nir_foreach_instr(instr
, block
) {
443 nir_emit_instr(instr
);
448 fs_visitor::nir_emit_instr(nir_instr
*instr
)
450 const fs_builder abld
= bld
.annotate(NULL
, instr
);
452 switch (instr
->type
) {
453 case nir_instr_type_alu
:
454 nir_emit_alu(abld
, nir_instr_as_alu(instr
));
457 case nir_instr_type_intrinsic
:
459 case MESA_SHADER_VERTEX
:
460 nir_emit_vs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
462 case MESA_SHADER_TESS_CTRL
:
463 nir_emit_tcs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
465 case MESA_SHADER_TESS_EVAL
:
466 nir_emit_tes_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
468 case MESA_SHADER_GEOMETRY
:
469 nir_emit_gs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
471 case MESA_SHADER_FRAGMENT
:
472 nir_emit_fs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
474 case MESA_SHADER_COMPUTE
:
475 nir_emit_cs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
478 unreachable("unsupported shader stage");
482 case nir_instr_type_tex
:
483 nir_emit_texture(abld
, nir_instr_as_tex(instr
));
486 case nir_instr_type_load_const
:
487 nir_emit_load_const(abld
, nir_instr_as_load_const(instr
));
490 case nir_instr_type_ssa_undef
:
491 nir_emit_undef(abld
, nir_instr_as_ssa_undef(instr
));
494 case nir_instr_type_jump
:
495 nir_emit_jump(abld
, nir_instr_as_jump(instr
));
499 unreachable("unknown instruction type");
504 * Recognizes a parent instruction of nir_op_extract_* and changes the type to
508 fs_visitor::optimize_extract_to_float(nir_alu_instr
*instr
,
509 const fs_reg
&result
)
511 if (!instr
->src
[0].src
.is_ssa
||
512 !instr
->src
[0].src
.ssa
->parent_instr
)
515 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
518 nir_alu_instr
*src0
=
519 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
521 if (src0
->op
!= nir_op_extract_u8
&& src0
->op
!= nir_op_extract_u16
&&
522 src0
->op
!= nir_op_extract_i8
&& src0
->op
!= nir_op_extract_i16
)
525 nir_const_value
*element
= nir_src_as_const_value(src0
->src
[1].src
);
526 assert(element
!= NULL
);
528 enum opcode extract_op
;
529 if (src0
->op
== nir_op_extract_u16
|| src0
->op
== nir_op_extract_i16
) {
530 assert(element
->u32
[0] <= 1);
531 extract_op
= SHADER_OPCODE_EXTRACT_WORD
;
533 assert(element
->u32
[0] <= 3);
534 extract_op
= SHADER_OPCODE_EXTRACT_BYTE
;
537 fs_reg op0
= get_nir_src(src0
->src
[0].src
);
538 op0
.type
= brw_type_for_nir_type(nir_op_infos
[src0
->op
].input_types
[0]);
539 op0
= offset(op0
, bld
, src0
->src
[0].swizzle
[0]);
541 set_saturate(instr
->dest
.saturate
,
542 bld
.emit(extract_op
, result
, op0
, brw_imm_ud(element
->u32
[0])));
547 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr
*instr
,
548 const fs_reg
&result
)
550 if (!instr
->src
[0].src
.is_ssa
||
551 instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
554 nir_intrinsic_instr
*src0
=
555 nir_instr_as_intrinsic(instr
->src
[0].src
.ssa
->parent_instr
);
557 if (src0
->intrinsic
!= nir_intrinsic_load_front_face
)
560 nir_const_value
*value1
= nir_src_as_const_value(instr
->src
[1].src
);
561 if (!value1
|| fabsf(value1
->f32
[0]) != 1.0f
)
564 nir_const_value
*value2
= nir_src_as_const_value(instr
->src
[2].src
);
565 if (!value2
|| fabsf(value2
->f32
[0]) != 1.0f
)
568 fs_reg tmp
= vgrf(glsl_type::int_type
);
570 if (devinfo
->gen
>= 6) {
571 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
572 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
574 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
576 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
577 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
579 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
581 * This negation looks like it's safe in practice, because bits 0:4 will
582 * surely be TRIANGLES
585 if (value1
->f32
[0] == -1.0f
) {
589 tmp
.type
= BRW_REGISTER_TYPE_W
;
590 tmp
.subreg_offset
= 2;
593 bld
.OR(tmp
, g0
, brw_imm_uw(0x3f80));
595 tmp
.type
= BRW_REGISTER_TYPE_D
;
596 tmp
.subreg_offset
= 0;
599 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
600 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
602 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
604 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
605 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
607 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
609 * This negation looks like it's safe in practice, because bits 0:4 will
610 * surely be TRIANGLES
613 if (value1
->f32
[0] == -1.0f
) {
617 bld
.OR(tmp
, g1_6
, brw_imm_d(0x3f800000));
619 bld
.AND(retype(result
, BRW_REGISTER_TYPE_D
), tmp
, brw_imm_d(0xbf800000));
625 fs_visitor::nir_emit_alu(const fs_builder
&bld
, nir_alu_instr
*instr
)
627 struct brw_wm_prog_key
*fs_key
= (struct brw_wm_prog_key
*) this->key
;
630 fs_reg result
= get_nir_dest(instr
->dest
.dest
);
631 result
.type
= brw_type_for_nir_type(nir_op_infos
[instr
->op
].output_type
);
634 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
635 op
[i
] = get_nir_src(instr
->src
[i
].src
);
636 op
[i
].type
= brw_type_for_nir_type(nir_op_infos
[instr
->op
].input_types
[i
]);
637 op
[i
].abs
= instr
->src
[i
].abs
;
638 op
[i
].negate
= instr
->src
[i
].negate
;
641 /* We get a bunch of mov's out of the from_ssa pass and they may still
642 * be vectorized. We'll handle them as a special-case. We'll also
643 * handle vecN here because it's basically the same thing.
651 fs_reg temp
= result
;
652 bool need_extra_copy
= false;
653 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
654 if (!instr
->src
[i
].src
.is_ssa
&&
655 instr
->dest
.dest
.reg
.reg
== instr
->src
[i
].src
.reg
.reg
) {
656 need_extra_copy
= true;
657 temp
= bld
.vgrf(result
.type
, 4);
662 for (unsigned i
= 0; i
< 4; i
++) {
663 if (!(instr
->dest
.write_mask
& (1 << i
)))
666 if (instr
->op
== nir_op_imov
|| instr
->op
== nir_op_fmov
) {
667 inst
= bld
.MOV(offset(temp
, bld
, i
),
668 offset(op
[0], bld
, instr
->src
[0].swizzle
[i
]));
670 inst
= bld
.MOV(offset(temp
, bld
, i
),
671 offset(op
[i
], bld
, instr
->src
[i
].swizzle
[0]));
673 inst
->saturate
= instr
->dest
.saturate
;
676 /* In this case the source and destination registers were the same,
677 * so we need to insert an extra set of moves in order to deal with
680 if (need_extra_copy
) {
681 for (unsigned i
= 0; i
< 4; i
++) {
682 if (!(instr
->dest
.write_mask
& (1 << i
)))
685 bld
.MOV(offset(result
, bld
, i
), offset(temp
, bld
, i
));
694 /* At this point, we have dealt with any instruction that operates on
695 * more than a single channel. Therefore, we can just adjust the source
696 * and destination registers for that channel and emit the instruction.
698 unsigned channel
= 0;
699 if (nir_op_infos
[instr
->op
].output_size
== 0) {
700 /* Since NIR is doing the scalarizing for us, we should only ever see
701 * vectorized operations with a single channel.
703 assert(_mesa_bitcount(instr
->dest
.write_mask
) == 1);
704 channel
= ffs(instr
->dest
.write_mask
) - 1;
706 result
= offset(result
, bld
, channel
);
709 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
710 assert(nir_op_infos
[instr
->op
].input_sizes
[i
] < 2);
711 op
[i
] = offset(op
[i
], bld
, instr
->src
[i
].swizzle
[channel
]);
717 if (optimize_extract_to_float(instr
, result
))
720 inst
= bld
.MOV(result
, op
[0]);
721 inst
->saturate
= instr
->dest
.saturate
;
726 bld
.MOV(result
, op
[0]);
730 /* AND(val, 0x80000000) gives the sign bit.
732 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
735 bld
.CMP(bld
.null_reg_f(), op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
737 fs_reg result_int
= retype(result
, BRW_REGISTER_TYPE_UD
);
738 op
[0].type
= BRW_REGISTER_TYPE_UD
;
739 result
.type
= BRW_REGISTER_TYPE_UD
;
740 bld
.AND(result_int
, op
[0], brw_imm_ud(0x80000000u
));
742 inst
= bld
.OR(result_int
, result_int
, brw_imm_ud(0x3f800000u
));
743 inst
->predicate
= BRW_PREDICATE_NORMAL
;
744 if (instr
->dest
.saturate
) {
745 inst
= bld
.MOV(result
, result
);
746 inst
->saturate
= true;
752 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
753 * -> non-negative val generates 0x00000000.
754 * Predicated OR sets 1 if val is positive.
756 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_G
);
757 bld
.ASR(result
, op
[0], brw_imm_d(31));
758 inst
= bld
.OR(result
, result
, brw_imm_d(1));
759 inst
->predicate
= BRW_PREDICATE_NORMAL
;
763 inst
= bld
.emit(SHADER_OPCODE_RCP
, result
, op
[0]);
764 inst
->saturate
= instr
->dest
.saturate
;
768 inst
= bld
.emit(SHADER_OPCODE_EXP2
, result
, op
[0]);
769 inst
->saturate
= instr
->dest
.saturate
;
773 inst
= bld
.emit(SHADER_OPCODE_LOG2
, result
, op
[0]);
774 inst
->saturate
= instr
->dest
.saturate
;
778 inst
= bld
.emit(SHADER_OPCODE_SIN
, result
, op
[0]);
779 inst
->saturate
= instr
->dest
.saturate
;
783 inst
= bld
.emit(SHADER_OPCODE_COS
, result
, op
[0]);
784 inst
->saturate
= instr
->dest
.saturate
;
788 if (fs_key
->high_quality_derivatives
) {
789 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
791 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
793 inst
->saturate
= instr
->dest
.saturate
;
795 case nir_op_fddx_fine
:
796 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
797 inst
->saturate
= instr
->dest
.saturate
;
799 case nir_op_fddx_coarse
:
800 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
801 inst
->saturate
= instr
->dest
.saturate
;
804 if (fs_key
->high_quality_derivatives
) {
805 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0],
806 brw_imm_d(fs_key
->render_to_fbo
));
808 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0],
809 brw_imm_d(fs_key
->render_to_fbo
));
811 inst
->saturate
= instr
->dest
.saturate
;
813 case nir_op_fddy_fine
:
814 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0],
815 brw_imm_d(fs_key
->render_to_fbo
));
816 inst
->saturate
= instr
->dest
.saturate
;
818 case nir_op_fddy_coarse
:
819 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0],
820 brw_imm_d(fs_key
->render_to_fbo
));
821 inst
->saturate
= instr
->dest
.saturate
;
826 inst
= bld
.ADD(result
, op
[0], op
[1]);
827 inst
->saturate
= instr
->dest
.saturate
;
831 inst
= bld
.MUL(result
, op
[0], op
[1]);
832 inst
->saturate
= instr
->dest
.saturate
;
836 bld
.MUL(result
, op
[0], op
[1]);
839 case nir_op_imul_high
:
840 case nir_op_umul_high
:
841 bld
.emit(SHADER_OPCODE_MULH
, result
, op
[0], op
[1]);
846 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, result
, op
[0], op
[1]);
849 case nir_op_uadd_carry
:
850 unreachable("Should have been lowered by carry_to_arith().");
852 case nir_op_usub_borrow
:
853 unreachable("Should have been lowered by borrow_to_arith().");
857 /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
858 * appears that our hardware just does the right thing for signed
861 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
865 /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
866 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
868 /* Math instructions don't support conditional mod */
869 inst
= bld
.MOV(bld
.null_reg_d(), result
);
870 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
872 /* Now, we need to determine if signs of the sources are different.
873 * When we XOR the sources, the top bit is 0 if they are the same and 1
874 * if they are different. We can then use a conditional modifier to
875 * turn that into a predicate. This leads us to an XOR.l instruction.
877 * Technically, according to the PRM, you're not allowed to use .l on a
878 * XOR instruction. However, emperical experiments and Curro's reading
879 * of the simulator source both indicate that it's safe.
881 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
882 inst
= bld
.XOR(tmp
, op
[0], op
[1]);
883 inst
->predicate
= BRW_PREDICATE_NORMAL
;
884 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
886 /* If the result of the initial remainder operation is non-zero and the
887 * two sources have different signs, add in a copy of op[1] to get the
888 * final integer modulus value.
890 inst
= bld
.ADD(result
, result
, op
[1]);
891 inst
->predicate
= BRW_PREDICATE_NORMAL
;
898 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
904 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
909 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_Z
);
914 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_NZ
);
918 if (devinfo
->gen
>= 8) {
919 op
[0] = resolve_source_modifiers(op
[0]);
921 bld
.NOT(result
, op
[0]);
924 if (devinfo
->gen
>= 8) {
925 op
[0] = resolve_source_modifiers(op
[0]);
926 op
[1] = resolve_source_modifiers(op
[1]);
928 bld
.XOR(result
, op
[0], op
[1]);
931 if (devinfo
->gen
>= 8) {
932 op
[0] = resolve_source_modifiers(op
[0]);
933 op
[1] = resolve_source_modifiers(op
[1]);
935 bld
.OR(result
, op
[0], op
[1]);
938 if (devinfo
->gen
>= 8) {
939 op
[0] = resolve_source_modifiers(op
[0]);
940 op
[1] = resolve_source_modifiers(op
[1]);
942 bld
.AND(result
, op
[0], op
[1]);
948 case nir_op_ball_fequal2
:
949 case nir_op_ball_iequal2
:
950 case nir_op_ball_fequal3
:
951 case nir_op_ball_iequal3
:
952 case nir_op_ball_fequal4
:
953 case nir_op_ball_iequal4
:
954 case nir_op_bany_fnequal2
:
955 case nir_op_bany_inequal2
:
956 case nir_op_bany_fnequal3
:
957 case nir_op_bany_inequal3
:
958 case nir_op_bany_fnequal4
:
959 case nir_op_bany_inequal4
:
960 unreachable("Lowered by nir_lower_alu_reductions");
962 case nir_op_fnoise1_1
:
963 case nir_op_fnoise1_2
:
964 case nir_op_fnoise1_3
:
965 case nir_op_fnoise1_4
:
966 case nir_op_fnoise2_1
:
967 case nir_op_fnoise2_2
:
968 case nir_op_fnoise2_3
:
969 case nir_op_fnoise2_4
:
970 case nir_op_fnoise3_1
:
971 case nir_op_fnoise3_2
:
972 case nir_op_fnoise3_3
:
973 case nir_op_fnoise3_4
:
974 case nir_op_fnoise4_1
:
975 case nir_op_fnoise4_2
:
976 case nir_op_fnoise4_3
:
977 case nir_op_fnoise4_4
:
978 unreachable("not reached: should be handled by lower_noise");
981 unreachable("not reached: should be handled by ldexp_to_arith()");
984 inst
= bld
.emit(SHADER_OPCODE_SQRT
, result
, op
[0]);
985 inst
->saturate
= instr
->dest
.saturate
;
989 inst
= bld
.emit(SHADER_OPCODE_RSQ
, result
, op
[0]);
990 inst
->saturate
= instr
->dest
.saturate
;
995 bld
.MOV(result
, negate(op
[0]));
999 bld
.CMP(result
, op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
1002 bld
.CMP(result
, op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1006 inst
= bld
.RNDZ(result
, op
[0]);
1007 inst
->saturate
= instr
->dest
.saturate
;
1010 case nir_op_fceil
: {
1011 op
[0].negate
= !op
[0].negate
;
1012 fs_reg temp
= vgrf(glsl_type::float_type
);
1013 bld
.RNDD(temp
, op
[0]);
1015 inst
= bld
.MOV(result
, temp
);
1016 inst
->saturate
= instr
->dest
.saturate
;
1020 inst
= bld
.RNDD(result
, op
[0]);
1021 inst
->saturate
= instr
->dest
.saturate
;
1024 inst
= bld
.FRC(result
, op
[0]);
1025 inst
->saturate
= instr
->dest
.saturate
;
1027 case nir_op_fround_even
:
1028 inst
= bld
.RNDE(result
, op
[0]);
1029 inst
->saturate
= instr
->dest
.saturate
;
1032 case nir_op_fquantize2f16
: {
1033 fs_reg tmp16
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1034 fs_reg tmp32
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1035 fs_reg zero
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1037 /* The destination stride must be at least as big as the source stride. */
1038 tmp16
.type
= BRW_REGISTER_TYPE_W
;
1041 /* Check for denormal */
1042 fs_reg abs_src0
= op
[0];
1043 abs_src0
.abs
= true;
1044 bld
.CMP(bld
.null_reg_f(), abs_src0
, brw_imm_f(ldexpf(1.0, -14)),
1046 /* Get the appropriately signed zero */
1047 bld
.AND(retype(zero
, BRW_REGISTER_TYPE_UD
),
1048 retype(op
[0], BRW_REGISTER_TYPE_UD
),
1049 brw_imm_ud(0x80000000));
1050 /* Do the actual F32 -> F16 -> F32 conversion */
1051 bld
.emit(BRW_OPCODE_F32TO16
, tmp16
, op
[0]);
1052 bld
.emit(BRW_OPCODE_F16TO32
, tmp32
, tmp16
);
1053 /* Select that or zero based on normal status */
1054 inst
= bld
.SEL(result
, zero
, tmp32
);
1055 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1056 inst
->saturate
= instr
->dest
.saturate
;
1063 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1064 inst
->saturate
= instr
->dest
.saturate
;
1070 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1071 inst
->saturate
= instr
->dest
.saturate
;
1074 case nir_op_pack_snorm_2x16
:
1075 case nir_op_pack_snorm_4x8
:
1076 case nir_op_pack_unorm_2x16
:
1077 case nir_op_pack_unorm_4x8
:
1078 case nir_op_unpack_snorm_2x16
:
1079 case nir_op_unpack_snorm_4x8
:
1080 case nir_op_unpack_unorm_2x16
:
1081 case nir_op_unpack_unorm_4x8
:
1082 case nir_op_unpack_half_2x16
:
1083 case nir_op_pack_half_2x16
:
1084 unreachable("not reached: should be handled by lower_packing_builtins");
1086 case nir_op_unpack_half_2x16_split_x
:
1087 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
, result
, op
[0]);
1088 inst
->saturate
= instr
->dest
.saturate
;
1090 case nir_op_unpack_half_2x16_split_y
:
1091 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
, result
, op
[0]);
1092 inst
->saturate
= instr
->dest
.saturate
;
1096 inst
= bld
.emit(SHADER_OPCODE_POW
, result
, op
[0], op
[1]);
1097 inst
->saturate
= instr
->dest
.saturate
;
1100 case nir_op_bitfield_reverse
:
1101 bld
.BFREV(result
, op
[0]);
1104 case nir_op_bit_count
:
1105 bld
.CBIT(result
, op
[0]);
1108 case nir_op_ufind_msb
:
1109 case nir_op_ifind_msb
: {
1110 bld
.FBH(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0]);
1112 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1113 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1114 * subtract the result from 31 to convert the MSB count into an LSB count.
1116 bld
.CMP(bld
.null_reg_d(), result
, brw_imm_d(-1), BRW_CONDITIONAL_NZ
);
1118 inst
= bld
.ADD(result
, result
, brw_imm_d(31));
1119 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1120 inst
->src
[0].negate
= true;
1124 case nir_op_find_lsb
:
1125 bld
.FBL(result
, op
[0]);
1128 case nir_op_ubitfield_extract
:
1129 case nir_op_ibitfield_extract
:
1130 unreachable("should have been lowered");
1133 bld
.BFE(result
, op
[2], op
[1], op
[0]);
1136 bld
.BFI1(result
, op
[0], op
[1]);
1139 bld
.BFI2(result
, op
[0], op
[1], op
[2]);
1142 case nir_op_bitfield_insert
:
1143 unreachable("not reached: should have been lowered");
1146 bld
.SHL(result
, op
[0], op
[1]);
1149 bld
.ASR(result
, op
[0], op
[1]);
1152 bld
.SHR(result
, op
[0], op
[1]);
1155 case nir_op_pack_half_2x16_split
:
1156 bld
.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, result
, op
[0], op
[1]);
1160 inst
= bld
.MAD(result
, op
[2], op
[1], op
[0]);
1161 inst
->saturate
= instr
->dest
.saturate
;
1165 inst
= bld
.LRP(result
, op
[0], op
[1], op
[2]);
1166 inst
->saturate
= instr
->dest
.saturate
;
1170 if (optimize_frontfacing_ternary(instr
, result
))
1173 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1174 inst
= bld
.SEL(result
, op
[1], op
[2]);
1175 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1178 case nir_op_extract_u8
:
1179 case nir_op_extract_i8
: {
1180 nir_const_value
*byte
= nir_src_as_const_value(instr
->src
[1].src
);
1181 bld
.emit(SHADER_OPCODE_EXTRACT_BYTE
,
1182 result
, op
[0], brw_imm_ud(byte
->u32
[0]));
1186 case nir_op_extract_u16
:
1187 case nir_op_extract_i16
: {
1188 nir_const_value
*word
= nir_src_as_const_value(instr
->src
[1].src
);
1189 bld
.emit(SHADER_OPCODE_EXTRACT_WORD
,
1190 result
, op
[0], brw_imm_ud(word
->u32
[0]));
1195 unreachable("unhandled instruction");
1198 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1199 * to sign extend the low bit to 0/~0
1201 if (devinfo
->gen
<= 5 &&
1202 (instr
->instr
.pass_flags
& BRW_NIR_BOOLEAN_MASK
) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE
) {
1203 fs_reg masked
= vgrf(glsl_type::int_type
);
1204 bld
.AND(masked
, result
, brw_imm_d(1));
1205 masked
.negate
= true;
1206 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), masked
);
1211 fs_visitor::nir_emit_load_const(const fs_builder
&bld
,
1212 nir_load_const_instr
*instr
)
1214 const brw_reg_type reg_type
=
1215 instr
->def
.bit_size
== 32 ? BRW_REGISTER_TYPE_D
: BRW_REGISTER_TYPE_DF
;
1216 fs_reg reg
= bld
.vgrf(reg_type
, instr
->def
.num_components
);
1218 switch (instr
->def
.bit_size
) {
1220 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1221 bld
.MOV(offset(reg
, bld
, i
), brw_imm_d(instr
->value
.i32
[i
]));
1225 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1226 bld
.MOV(offset(reg
, bld
, i
), brw_imm_df(instr
->value
.f64
[i
]));
1230 unreachable("Invalid bit size");
1233 nir_ssa_values
[instr
->def
.index
] = reg
;
1237 fs_visitor::nir_emit_undef(const fs_builder
&bld
, nir_ssa_undef_instr
*instr
)
1239 const brw_reg_type reg_type
=
1240 instr
->def
.bit_size
== 32 ? BRW_REGISTER_TYPE_D
: BRW_REGISTER_TYPE_DF
;
1241 nir_ssa_values
[instr
->def
.index
] =
1242 bld
.vgrf(reg_type
, instr
->def
.num_components
);
1246 fs_visitor::get_nir_src(nir_src src
)
1250 reg
= nir_ssa_values
[src
.ssa
->index
];
1252 /* We don't handle indirects on locals */
1253 assert(src
.reg
.indirect
== NULL
);
1254 reg
= offset(nir_locals
[src
.reg
.reg
->index
], bld
,
1255 src
.reg
.base_offset
* src
.reg
.reg
->num_components
);
1258 /* to avoid floating-point denorm flushing problems, set the type by
1259 * default to D - instructions that need floating point semantics will set
1260 * this to F if they need to
1262 return retype(reg
, BRW_REGISTER_TYPE_D
);
1266 fs_visitor::get_nir_dest(nir_dest dest
)
1269 const brw_reg_type reg_type
=
1270 dest
.ssa
.bit_size
== 32 ? BRW_REGISTER_TYPE_F
: BRW_REGISTER_TYPE_DF
;
1271 nir_ssa_values
[dest
.ssa
.index
] =
1272 bld
.vgrf(reg_type
, dest
.ssa
.num_components
);
1273 return nir_ssa_values
[dest
.ssa
.index
];
1275 /* We don't handle indirects on locals */
1276 assert(dest
.reg
.indirect
== NULL
);
1277 return offset(nir_locals
[dest
.reg
.reg
->index
], bld
,
1278 dest
.reg
.base_offset
* dest
.reg
.reg
->num_components
);
1283 fs_visitor::get_nir_image_deref(const nir_deref_var
*deref
)
1285 fs_reg
image(UNIFORM
, deref
->var
->data
.driver_location
/ 4,
1286 BRW_REGISTER_TYPE_UD
);
1288 unsigned indirect_max
= 0;
1290 for (const nir_deref
*tail
= &deref
->deref
; tail
->child
;
1291 tail
= tail
->child
) {
1292 const nir_deref_array
*deref_array
= nir_deref_as_array(tail
->child
);
1293 assert(tail
->child
->deref_type
== nir_deref_type_array
);
1294 const unsigned size
= glsl_get_length(tail
->type
);
1295 const unsigned element_size
= type_size_scalar(deref_array
->deref
.type
);
1296 const unsigned base
= MIN2(deref_array
->base_offset
, size
- 1);
1297 image
= offset(image
, bld
, base
* element_size
);
1299 if (deref_array
->deref_array_type
== nir_deref_array_type_indirect
) {
1300 fs_reg tmp
= vgrf(glsl_type::uint_type
);
1302 /* Accessing an invalid surface index with the dataport can result
1303 * in a hang. According to the spec "if the index used to
1304 * select an individual element is negative or greater than or
1305 * equal to the size of the array, the results of the operation
1306 * are undefined but may not lead to termination" -- which is one
1307 * of the possible outcomes of the hang. Clamp the index to
1308 * prevent access outside of the array bounds.
1310 bld
.emit_minmax(tmp
, retype(get_nir_src(deref_array
->indirect
),
1311 BRW_REGISTER_TYPE_UD
),
1312 brw_imm_ud(size
- base
- 1), BRW_CONDITIONAL_L
);
1314 indirect_max
+= element_size
* (tail
->type
->length
- 1);
1316 bld
.MUL(tmp
, tmp
, brw_imm_ud(element_size
* 4));
1317 if (indirect
.file
== BAD_FILE
) {
1320 bld
.ADD(indirect
, indirect
, tmp
);
1325 if (indirect
.file
== BAD_FILE
) {
1328 /* Emit a pile of MOVs to load the uniform into a temporary. The
1329 * dead-code elimination pass will get rid of what we don't use.
1331 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, BRW_IMAGE_PARAM_SIZE
);
1332 for (unsigned j
= 0; j
< BRW_IMAGE_PARAM_SIZE
; j
++) {
1333 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
1334 offset(tmp
, bld
, j
), offset(image
, bld
, j
),
1335 indirect
, brw_imm_ud((indirect_max
+ 1) * 4));
1342 fs_visitor::emit_percomp(const fs_builder
&bld
, const fs_inst
&inst
,
1345 for (unsigned i
= 0; i
< 4; i
++) {
1346 if (!((wr_mask
>> i
) & 1))
1349 fs_inst
*new_inst
= new(mem_ctx
) fs_inst(inst
);
1350 new_inst
->dst
= offset(new_inst
->dst
, bld
, i
);
1351 for (unsigned j
= 0; j
< new_inst
->sources
; j
++)
1352 if (new_inst
->src
[j
].file
== VGRF
)
1353 new_inst
->src
[j
] = offset(new_inst
->src
[j
], bld
, i
);
1360 * Get the matching channel register datatype for an image intrinsic of the
1361 * specified GLSL image type.
1364 get_image_base_type(const glsl_type
*type
)
1366 switch ((glsl_base_type
)type
->sampled_type
) {
1367 case GLSL_TYPE_UINT
:
1368 return BRW_REGISTER_TYPE_UD
;
1370 return BRW_REGISTER_TYPE_D
;
1371 case GLSL_TYPE_FLOAT
:
1372 return BRW_REGISTER_TYPE_F
;
1374 unreachable("Not reached.");
1379 * Get the appropriate atomic op for an image atomic intrinsic.
1382 get_image_atomic_op(nir_intrinsic_op op
, const glsl_type
*type
)
1385 case nir_intrinsic_image_atomic_add
:
1387 case nir_intrinsic_image_atomic_min
:
1388 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1389 BRW_AOP_IMIN
: BRW_AOP_UMIN
);
1390 case nir_intrinsic_image_atomic_max
:
1391 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1392 BRW_AOP_IMAX
: BRW_AOP_UMAX
);
1393 case nir_intrinsic_image_atomic_and
:
1395 case nir_intrinsic_image_atomic_or
:
1397 case nir_intrinsic_image_atomic_xor
:
1399 case nir_intrinsic_image_atomic_exchange
:
1401 case nir_intrinsic_image_atomic_comp_swap
:
1402 return BRW_AOP_CMPWR
;
1404 unreachable("Not reachable.");
1409 emit_pixel_interpolater_send(const fs_builder
&bld
,
1414 glsl_interp_qualifier interpolation
)
1420 if (src
.file
== BAD_FILE
) {
1422 payload
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 1);
1426 mlen
= 2 * bld
.dispatch_width() / 8;
1429 inst
= bld
.emit(opcode
, dst
, payload
, desc
);
1431 /* 2 floats per slot returned */
1432 inst
->regs_written
= 2 * bld
.dispatch_width() / 8;
1433 inst
->pi_noperspective
= interpolation
== INTERP_QUALIFIER_NOPERSPECTIVE
;
1439 * Computes 1 << x, given a D/UD register containing some value x.
1442 intexp2(const fs_builder
&bld
, const fs_reg
&x
)
1444 assert(x
.type
== BRW_REGISTER_TYPE_UD
|| x
.type
== BRW_REGISTER_TYPE_D
);
1446 fs_reg result
= bld
.vgrf(x
.type
, 1);
1447 fs_reg one
= bld
.vgrf(x
.type
, 1);
1449 bld
.MOV(one
, retype(brw_imm_d(1), one
.type
));
1450 bld
.SHL(result
, one
, x
);
1455 fs_visitor::emit_gs_end_primitive(const nir_src
&vertex_count_nir_src
)
1457 assert(stage
== MESA_SHADER_GEOMETRY
);
1459 struct brw_gs_prog_data
*gs_prog_data
=
1460 (struct brw_gs_prog_data
*) prog_data
;
1462 /* We can only do EndPrimitive() functionality when the control data
1463 * consists of cut bits. Fortunately, the only time it isn't is when the
1464 * output type is points, in which case EndPrimitive() is a no-op.
1466 if (gs_prog_data
->control_data_format
!=
1467 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
) {
1471 /* Cut bits use one bit per vertex. */
1472 assert(gs_compile
->control_data_bits_per_vertex
== 1);
1474 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1475 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1477 /* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
1478 * vertex n, 0 otherwise. So all we need to do here is mark bit
1479 * (vertex_count - 1) % 32 in the cut_bits register to indicate that
1480 * EndPrimitive() was called after emitting vertex (vertex_count - 1);
1481 * vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
1483 * Note that if EndPrimitive() is called before emitting any vertices, this
1484 * will cause us to set bit 31 of the control_data_bits register to 1.
1485 * That's fine because:
1487 * - If max_vertices < 32, then vertex number 31 (zero-based) will never be
1488 * output, so the hardware will ignore cut bit 31.
1490 * - If max_vertices == 32, then vertex number 31 is guaranteed to be the
1491 * last vertex, so setting cut bit 31 has no effect (since the primitive
1492 * is automatically ended when the GS terminates).
1494 * - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
1495 * control_data_bits register to 0 when the first vertex is emitted.
1498 const fs_builder abld
= bld
.annotate("end primitive");
1500 /* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
1501 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1502 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1503 fs_reg mask
= intexp2(abld
, prev_count
);
1504 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1505 * attention to the lower 5 bits of its second source argument, so on this
1506 * architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
1507 * ((vertex_count - 1) % 32).
1509 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1513 fs_visitor::emit_gs_control_data_bits(const fs_reg
&vertex_count
)
1515 assert(stage
== MESA_SHADER_GEOMETRY
);
1516 assert(gs_compile
->control_data_bits_per_vertex
!= 0);
1518 struct brw_gs_prog_data
*gs_prog_data
=
1519 (struct brw_gs_prog_data
*) prog_data
;
1521 const fs_builder abld
= bld
.annotate("emit control data bits");
1522 const fs_builder fwa_bld
= bld
.exec_all();
1524 /* We use a single UD register to accumulate control data bits (32 bits
1525 * for each of the SIMD8 channels). So we need to write a DWord (32 bits)
1528 * Unfortunately, the URB_WRITE_SIMD8 message uses 128-bit (OWord) offsets.
1529 * We have select a 128-bit group via the Global and Per-Slot Offsets, then
1530 * use the Channel Mask phase to enable/disable which DWord within that
1531 * group to write. (Remember, different SIMD8 channels may have emitted
1532 * different numbers of vertices, so we may need per-slot offsets.)
1534 * Channel masking presents an annoying problem: we may have to replicate
1535 * the data up to 4 times:
1537 * Msg = Handles, Per-Slot Offsets, Channel Masks, Data, Data, Data, Data.
1539 * To avoid penalizing shaders that emit a small number of vertices, we
1540 * can avoid these sometimes: if the size of the control data header is
1541 * <= 128 bits, then there is only 1 OWord. All SIMD8 channels will land
1542 * land in the same 128-bit group, so we can skip per-slot offsets.
1544 * Similarly, if the control data header is <= 32 bits, there is only one
1545 * DWord, so we can skip channel masks.
1547 enum opcode opcode
= SHADER_OPCODE_URB_WRITE_SIMD8
;
1549 fs_reg channel_mask
, per_slot_offset
;
1551 if (gs_compile
->control_data_header_size_bits
> 32) {
1552 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
1553 channel_mask
= vgrf(glsl_type::uint_type
);
1556 if (gs_compile
->control_data_header_size_bits
> 128) {
1557 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
;
1558 per_slot_offset
= vgrf(glsl_type::uint_type
);
1561 /* Figure out which DWord we're trying to write to using the formula:
1563 * dword_index = (vertex_count - 1) * bits_per_vertex / 32
1565 * Since bits_per_vertex is a power of two, and is known at compile
1566 * time, this can be optimized to:
1568 * dword_index = (vertex_count - 1) >> (6 - log2(bits_per_vertex))
1570 if (opcode
!= SHADER_OPCODE_URB_WRITE_SIMD8
) {
1571 fs_reg dword_index
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1572 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1573 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1574 unsigned log2_bits_per_vertex
=
1575 _mesa_fls(gs_compile
->control_data_bits_per_vertex
);
1576 abld
.SHR(dword_index
, prev_count
, brw_imm_ud(6u - log2_bits_per_vertex
));
1578 if (per_slot_offset
.file
!= BAD_FILE
) {
1579 /* Set the per-slot offset to dword_index / 4, so that we'll write to
1580 * the appropriate OWord within the control data header.
1582 abld
.SHR(per_slot_offset
, dword_index
, brw_imm_ud(2u));
1585 /* Set the channel masks to 1 << (dword_index % 4), so that we'll
1586 * write to the appropriate DWORD within the OWORD.
1588 fs_reg channel
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1589 fwa_bld
.AND(channel
, dword_index
, brw_imm_ud(3u));
1590 channel_mask
= intexp2(fwa_bld
, channel
);
1591 /* Then the channel masks need to be in bits 23:16. */
1592 fwa_bld
.SHL(channel_mask
, channel_mask
, brw_imm_ud(16u));
1595 /* Store the control data bits in the message payload and send it. */
1597 if (channel_mask
.file
!= BAD_FILE
)
1598 mlen
+= 4; /* channel masks, plus 3 extra copies of the data */
1599 if (per_slot_offset
.file
!= BAD_FILE
)
1602 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
1603 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, mlen
);
1605 sources
[i
++] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1606 if (per_slot_offset
.file
!= BAD_FILE
)
1607 sources
[i
++] = per_slot_offset
;
1608 if (channel_mask
.file
!= BAD_FILE
)
1609 sources
[i
++] = channel_mask
;
1611 sources
[i
++] = this->control_data_bits
;
1614 abld
.LOAD_PAYLOAD(payload
, sources
, mlen
, mlen
);
1615 fs_inst
*inst
= abld
.emit(opcode
, reg_undef
, payload
);
1617 /* We need to increment Global Offset by 256-bits to make room for
1618 * Broadwell's extra "Vertex Count" payload at the beginning of the
1619 * URB entry. Since this is an OWord message, Global Offset is counted
1620 * in 128-bit units, so we must set it to 2.
1622 if (gs_prog_data
->static_vertex_count
== -1)
1627 fs_visitor::set_gs_stream_control_data_bits(const fs_reg
&vertex_count
,
1630 /* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
1632 /* Note: we are calling this *before* increasing vertex_count, so
1633 * this->vertex_count == vertex_count - 1 in the formula above.
1636 /* Stream mode uses 2 bits per vertex */
1637 assert(gs_compile
->control_data_bits_per_vertex
== 2);
1639 /* Must be a valid stream */
1640 assert(stream_id
>= 0 && stream_id
< MAX_VERTEX_STREAMS
);
1642 /* Control data bits are initialized to 0 so we don't have to set any
1643 * bits when sending vertices to stream 0.
1648 const fs_builder abld
= bld
.annotate("set stream control data bits", NULL
);
1650 /* reg::sid = stream_id */
1651 fs_reg sid
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1652 abld
.MOV(sid
, brw_imm_ud(stream_id
));
1654 /* reg:shift_count = 2 * (vertex_count - 1) */
1655 fs_reg shift_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1656 abld
.SHL(shift_count
, vertex_count
, brw_imm_ud(1u));
1658 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1659 * attention to the lower 5 bits of its second source argument, so on this
1660 * architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
1661 * stream_id << ((2 * (vertex_count - 1)) % 32).
1663 fs_reg mask
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1664 abld
.SHL(mask
, sid
, shift_count
);
1665 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1669 fs_visitor::emit_gs_vertex(const nir_src
&vertex_count_nir_src
,
1672 assert(stage
== MESA_SHADER_GEOMETRY
);
1674 struct brw_gs_prog_data
*gs_prog_data
=
1675 (struct brw_gs_prog_data
*) prog_data
;
1677 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1678 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1680 /* Haswell and later hardware ignores the "Render Stream Select" bits
1681 * from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
1682 * and instead sends all primitives down the pipeline for rasterization.
1683 * If the SOL stage is enabled, "Render Stream Select" is honored and
1684 * primitives bound to non-zero streams are discarded after stream output.
1686 * Since the only purpose of primives sent to non-zero streams is to
1687 * be recorded by transform feedback, we can simply discard all geometry
1688 * bound to these streams when transform feedback is disabled.
1690 if (stream_id
> 0 && !nir
->info
.has_transform_feedback_varyings
)
1693 /* If we're outputting 32 control data bits or less, then we can wait
1694 * until the shader is over to output them all. Otherwise we need to
1695 * output them as we go. Now is the time to do it, since we're about to
1696 * output the vertex_count'th vertex, so it's guaranteed that the
1697 * control data bits associated with the (vertex_count - 1)th vertex are
1700 if (gs_compile
->control_data_header_size_bits
> 32) {
1701 const fs_builder abld
=
1702 bld
.annotate("emit vertex: emit control data bits");
1704 /* Only emit control data bits if we've finished accumulating a batch
1705 * of 32 bits. This is the case when:
1707 * (vertex_count * bits_per_vertex) % 32 == 0
1709 * (in other words, when the last 5 bits of vertex_count *
1710 * bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
1711 * integer n (which is always the case, since bits_per_vertex is
1712 * always 1 or 2), this is equivalent to requiring that the last 5-n
1713 * bits of vertex_count are 0:
1715 * vertex_count & (2^(5-n) - 1) == 0
1717 * 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
1720 * vertex_count & (32 / bits_per_vertex - 1) == 0
1722 * TODO: If vertex_count is an immediate, we could do some of this math
1723 * at compile time...
1726 abld
.AND(bld
.null_reg_d(), vertex_count
,
1727 brw_imm_ud(32u / gs_compile
->control_data_bits_per_vertex
- 1u));
1728 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1730 abld
.IF(BRW_PREDICATE_NORMAL
);
1731 /* If vertex_count is 0, then no control data bits have been
1732 * accumulated yet, so we can skip emitting them.
1734 abld
.CMP(bld
.null_reg_d(), vertex_count
, brw_imm_ud(0u),
1735 BRW_CONDITIONAL_NEQ
);
1736 abld
.IF(BRW_PREDICATE_NORMAL
);
1737 emit_gs_control_data_bits(vertex_count
);
1738 abld
.emit(BRW_OPCODE_ENDIF
);
1740 /* Reset control_data_bits to 0 so we can start accumulating a new
1743 * Note: in the case where vertex_count == 0, this neutralizes the
1744 * effect of any call to EndPrimitive() that the shader may have
1745 * made before outputting its first vertex.
1747 inst
= abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
1748 inst
->force_writemask_all
= true;
1749 abld
.emit(BRW_OPCODE_ENDIF
);
1752 emit_urb_writes(vertex_count
);
1754 /* In stream mode we have to set control data bits for all vertices
1755 * unless we have disabled control data bits completely (which we do
1756 * do for GL_POINTS outputs that don't use streams).
1758 if (gs_compile
->control_data_header_size_bits
> 0 &&
1759 gs_prog_data
->control_data_format
==
1760 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
) {
1761 set_gs_stream_control_data_bits(vertex_count
, stream_id
);
1766 fs_visitor::emit_gs_input_load(const fs_reg
&dst
,
1767 const nir_src
&vertex_src
,
1768 unsigned base_offset
,
1769 const nir_src
&offset_src
,
1770 unsigned num_components
)
1772 struct brw_gs_prog_data
*gs_prog_data
= (struct brw_gs_prog_data
*) prog_data
;
1774 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
1775 nir_const_value
*offset_const
= nir_src_as_const_value(offset_src
);
1776 const unsigned push_reg_count
= gs_prog_data
->base
.urb_read_length
* 8;
1778 /* Offset 0 is the VUE header, which contains VARYING_SLOT_LAYER [.y],
1779 * VARYING_SLOT_VIEWPORT [.z], and VARYING_SLOT_PSIZ [.w]. Only
1780 * gl_PointSize is available as a GS input, however, so it must be that.
1782 const bool is_point_size
= (base_offset
== 0);
1784 if (offset_const
!= NULL
&& vertex_const
!= NULL
&&
1785 4 * (base_offset
+ offset_const
->u32
[0]) < push_reg_count
) {
1786 int imm_offset
= (base_offset
+ offset_const
->u32
[0]) * 4 +
1787 vertex_const
->u32
[0] * push_reg_count
;
1788 /* This input was pushed into registers. */
1789 if (is_point_size
) {
1790 /* gl_PointSize comes in .w */
1791 assert(imm_offset
== 0);
1792 bld
.MOV(dst
, fs_reg(ATTR
, imm_offset
+ 3, dst
.type
));
1794 for (unsigned i
= 0; i
< num_components
; i
++) {
1795 bld
.MOV(offset(dst
, bld
, i
),
1796 fs_reg(ATTR
, imm_offset
+ i
, dst
.type
));
1800 /* Resort to the pull model. Ensure the VUE handles are provided. */
1801 gs_prog_data
->base
.include_vue_handles
= true;
1803 unsigned first_icp_handle
= gs_prog_data
->include_primitive_id
? 3 : 2;
1807 /* The vertex index is constant; just select the proper URB handle. */
1809 retype(brw_vec8_grf(first_icp_handle
+ vertex_const
->i32
[0], 0),
1810 BRW_REGISTER_TYPE_UD
);
1812 /* The vertex index is non-constant. We need to use indirect
1813 * addressing to fetch the proper URB handle.
1815 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
1816 * indicating that channel <n> should read the handle from
1817 * DWord <n>. We convert that to bytes by multiplying by 4.
1819 * Next, we convert the vertex index to bytes by multiplying
1820 * by 32 (shifting by 5), and add the two together. This is
1821 * the final indirect byte offset.
1823 fs_reg sequence
= bld
.vgrf(BRW_REGISTER_TYPE_W
, 1);
1824 fs_reg channel_offsets
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1825 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1826 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1827 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1829 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
1830 bld
.MOV(sequence
, fs_reg(brw_imm_v(0x76543210)));
1831 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
1832 bld
.SHL(channel_offsets
, sequence
, brw_imm_ud(2u));
1833 /* Convert vertex_index to bytes (multiply by 32) */
1834 bld
.SHL(vertex_offset_bytes
,
1835 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
1837 bld
.ADD(icp_offset_bytes
, vertex_offset_bytes
, channel_offsets
);
1839 /* Use first_icp_handle as the base offset. There is one register
1840 * of URB handles per vertex, so inform the register allocator that
1841 * we might read up to nir->info.gs.vertices_in registers.
1843 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
1844 fs_reg(brw_vec8_grf(first_icp_handle
, 0)),
1845 fs_reg(icp_offset_bytes
),
1846 brw_imm_ud(nir
->info
.gs
.vertices_in
* REG_SIZE
));
1851 /* Constant indexing - use global offset. */
1852 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
1853 inst
->offset
= base_offset
+ offset_const
->u32
[0];
1854 inst
->base_mrf
= -1;
1856 inst
->regs_written
= num_components
;
1858 /* Indirect indexing - use per-slot offsets as well. */
1859 const fs_reg srcs
[] = { icp_handle
, get_nir_src(offset_src
) };
1860 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
1861 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
1863 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
1864 inst
->offset
= base_offset
;
1865 inst
->base_mrf
= -1;
1867 inst
->regs_written
= num_components
;
1870 if (is_point_size
) {
1871 /* Read the whole VUE header (because of alignment) and read .w. */
1872 fs_reg tmp
= bld
.vgrf(dst
.type
, 4);
1874 inst
->regs_written
= 4;
1875 bld
.MOV(dst
, offset(tmp
, bld
, 3));
1881 fs_visitor::get_indirect_offset(nir_intrinsic_instr
*instr
)
1883 nir_src
*offset_src
= nir_get_io_offset_src(instr
);
1884 nir_const_value
*const_value
= nir_src_as_const_value(*offset_src
);
1887 /* The only constant offset we should find is 0. brw_nir.c's
1888 * add_const_offset_to_base() will fold other constant offsets
1889 * into instr->const_index[0].
1891 assert(const_value
->u32
[0] == 0);
1895 return get_nir_src(*offset_src
);
1899 fs_visitor::nir_emit_vs_intrinsic(const fs_builder
&bld
,
1900 nir_intrinsic_instr
*instr
)
1902 assert(stage
== MESA_SHADER_VERTEX
);
1905 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
1906 dest
= get_nir_dest(instr
->dest
);
1908 switch (instr
->intrinsic
) {
1909 case nir_intrinsic_load_vertex_id
:
1910 unreachable("should be lowered by lower_vertex_id()");
1912 case nir_intrinsic_load_vertex_id_zero_base
:
1913 case nir_intrinsic_load_base_vertex
:
1914 case nir_intrinsic_load_instance_id
:
1915 case nir_intrinsic_load_base_instance
:
1916 case nir_intrinsic_load_draw_id
: {
1917 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
1918 fs_reg val
= nir_system_values
[sv
];
1919 assert(val
.file
!= BAD_FILE
);
1920 dest
.type
= val
.type
;
1926 nir_emit_intrinsic(bld
, instr
);
1932 fs_visitor::nir_emit_tcs_intrinsic(const fs_builder
&bld
,
1933 nir_intrinsic_instr
*instr
)
1935 assert(stage
== MESA_SHADER_TESS_CTRL
);
1936 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
1937 struct brw_tcs_prog_data
*tcs_prog_data
=
1938 (struct brw_tcs_prog_data
*) prog_data
;
1941 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
1942 dst
= get_nir_dest(instr
->dest
);
1944 switch (instr
->intrinsic
) {
1945 case nir_intrinsic_load_primitive_id
:
1946 bld
.MOV(dst
, fs_reg(brw_vec1_grf(0, 1)));
1948 case nir_intrinsic_load_invocation_id
:
1949 bld
.MOV(retype(dst
, invocation_id
.type
), invocation_id
);
1951 case nir_intrinsic_load_patch_vertices_in
:
1952 bld
.MOV(retype(dst
, BRW_REGISTER_TYPE_D
),
1953 brw_imm_d(tcs_key
->input_vertices
));
1956 case nir_intrinsic_barrier
: {
1957 if (tcs_prog_data
->instances
== 1)
1960 fs_reg m0
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1961 fs_reg m0_2
= byte_offset(m0
, 2 * sizeof(uint32_t));
1963 const fs_builder fwa_bld
= bld
.exec_all();
1965 /* Zero the message header */
1966 fwa_bld
.MOV(m0
, brw_imm_ud(0u));
1968 /* Copy "Barrier ID" from r0.2, bits 16:13 */
1969 fwa_bld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
1970 brw_imm_ud(INTEL_MASK(16, 13)));
1972 /* Shift it up to bits 27:24. */
1973 fwa_bld
.SHL(m0_2
, m0_2
, brw_imm_ud(11));
1975 /* Set the Barrier Count and the enable bit */
1976 fwa_bld
.OR(m0_2
, m0_2
,
1977 brw_imm_ud(tcs_prog_data
->instances
<< 8 | (1 << 15)));
1979 bld
.emit(SHADER_OPCODE_BARRIER
, bld
.null_reg_ud(), m0
);
1983 case nir_intrinsic_load_input
:
1984 unreachable("nir_lower_io should never give us these.");
1987 case nir_intrinsic_load_per_vertex_input
: {
1988 fs_reg indirect_offset
= get_indirect_offset(instr
);
1989 unsigned imm_offset
= instr
->const_index
[0];
1991 const nir_src
&vertex_src
= instr
->src
[0];
1992 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
1999 /* Emit a MOV to resolve <0,1,0> regioning. */
2000 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2002 retype(brw_vec1_grf(1 + (vertex_const
->i32
[0] >> 3),
2003 vertex_const
->i32
[0] & 7),
2004 BRW_REGISTER_TYPE_UD
));
2005 } else if (tcs_prog_data
->instances
== 1 &&
2006 vertex_src
.is_ssa
&&
2007 vertex_src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
&&
2008 nir_instr_as_intrinsic(vertex_src
.ssa
->parent_instr
)->intrinsic
== nir_intrinsic_load_invocation_id
) {
2009 /* For the common case of only 1 instance, an array index of
2010 * gl_InvocationID means reading g1. Skip all the indirect work.
2012 icp_handle
= retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
);
2014 /* The vertex index is non-constant. We need to use indirect
2015 * addressing to fetch the proper URB handle.
2017 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2019 /* Each ICP handle is a single DWord (4 bytes) */
2020 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2021 bld
.SHL(vertex_offset_bytes
,
2022 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2025 /* Start at g1. We might read up to 4 registers. */
2026 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2027 fs_reg(brw_vec8_grf(1, 0)), vertex_offset_bytes
,
2028 brw_imm_ud(4 * REG_SIZE
));
2031 if (indirect_offset
.file
== BAD_FILE
) {
2032 /* Constant indexing - use global offset. */
2033 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2034 inst
->offset
= imm_offset
;
2036 inst
->base_mrf
= -1;
2037 inst
->regs_written
= instr
->num_components
;
2039 /* Indirect indexing - use per-slot offsets as well. */
2040 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2041 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2042 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2044 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2045 inst
->offset
= imm_offset
;
2046 inst
->base_mrf
= -1;
2048 inst
->regs_written
= instr
->num_components
;
2051 /* Copy the temporary to the destination to deal with writemasking.
2053 * Also attempt to deal with gl_PointSize being in the .w component.
2055 if (inst
->offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
2056 inst
->dst
= bld
.vgrf(dst
.type
, 4);
2057 inst
->regs_written
= 4;
2058 bld
.MOV(dst
, offset(inst
->dst
, bld
, 3));
2063 case nir_intrinsic_load_output
:
2064 case nir_intrinsic_load_per_vertex_output
: {
2065 fs_reg indirect_offset
= get_indirect_offset(instr
);
2066 unsigned imm_offset
= instr
->const_index
[0];
2069 if (indirect_offset
.file
== BAD_FILE
) {
2070 /* Replicate the patch handle to all enabled channels */
2071 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2072 bld
.MOV(patch_handle
,
2073 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2075 if (imm_offset
== 0) {
2076 /* This is a read of gl_TessLevelInner[], which lives in the
2077 * Patch URB header. The layout depends on the domain.
2079 dst
.type
= BRW_REGISTER_TYPE_F
;
2080 switch (tcs_key
->tes_primitive_mode
) {
2082 /* DWords 3-2 (reversed) */
2083 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
2085 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, patch_handle
);
2088 inst
->base_mrf
= -1;
2089 inst
->regs_written
= 4;
2091 /* dst.xy = tmp.wz */
2092 bld
.MOV(dst
, offset(tmp
, bld
, 3));
2093 bld
.MOV(offset(dst
, bld
, 1), offset(tmp
, bld
, 2));
2097 /* DWord 4; hardcode offset = 1 and regs_written = 1 */
2098 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, patch_handle
);
2101 inst
->base_mrf
= -1;
2102 inst
->regs_written
= 1;
2105 /* All channels are undefined. */
2108 unreachable("Bogus tessellation domain");
2110 } else if (imm_offset
== 1) {
2111 /* This is a read of gl_TessLevelOuter[], which lives in the
2112 * Patch URB header. The layout depends on the domain.
2114 dst
.type
= BRW_REGISTER_TYPE_F
;
2116 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
2117 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, patch_handle
);
2120 inst
->base_mrf
= -1;
2121 inst
->regs_written
= 4;
2123 /* Reswizzle: WZYX */
2125 offset(tmp
, bld
, 3),
2126 offset(tmp
, bld
, 2),
2127 offset(tmp
, bld
, 1),
2128 offset(tmp
, bld
, 0),
2131 unsigned num_components
;
2132 switch (tcs_key
->tes_primitive_mode
) {
2140 /* Isolines are not reversed; swizzle .zw -> .xy */
2141 srcs
[0] = offset(tmp
, bld
, 2);
2142 srcs
[1] = offset(tmp
, bld
, 3);
2146 unreachable("Bogus tessellation domain");
2148 bld
.LOAD_PAYLOAD(dst
, srcs
, num_components
, 0);
2150 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, patch_handle
);
2151 inst
->offset
= imm_offset
;
2153 inst
->base_mrf
= -1;
2154 inst
->regs_written
= instr
->num_components
;
2157 /* Indirect indexing - use per-slot offsets as well. */
2158 const fs_reg srcs
[] = {
2159 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2162 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2163 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2165 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2166 inst
->offset
= imm_offset
;
2168 inst
->base_mrf
= -1;
2169 inst
->regs_written
= instr
->num_components
;
2174 case nir_intrinsic_store_output
:
2175 case nir_intrinsic_store_per_vertex_output
: {
2176 fs_reg value
= get_nir_src(instr
->src
[0]);
2177 fs_reg indirect_offset
= get_indirect_offset(instr
);
2178 unsigned imm_offset
= instr
->const_index
[0];
2179 unsigned swiz
= BRW_SWIZZLE_XYZW
;
2180 unsigned mask
= instr
->const_index
[1];
2181 unsigned header_regs
= 0;
2183 srcs
[header_regs
++] = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
);
2185 if (indirect_offset
.file
!= BAD_FILE
) {
2186 srcs
[header_regs
++] = indirect_offset
;
2187 } else if (!is_passthrough_shader
) {
2188 if (imm_offset
== 0) {
2189 value
.type
= BRW_REGISTER_TYPE_F
;
2191 mask
&= (1 << tesslevel_inner_components(tcs_key
->tes_primitive_mode
)) - 1;
2193 /* This is a write to gl_TessLevelInner[], which lives in the
2194 * Patch URB header. The layout depends on the domain.
2196 switch (tcs_key
->tes_primitive_mode
) {
2198 /* gl_TessLevelInner[].xy lives at DWords 3-2 (reversed).
2199 * We use an XXYX swizzle to reverse put .xy in the .wz
2200 * channels, and use a .zw writemask.
2202 mask
= writemask_for_backwards_vector(mask
);
2203 swiz
= BRW_SWIZZLE4(0, 0, 1, 0);
2206 /* gl_TessLevelInner[].x lives at DWord 4, so we set the
2207 * writemask to X and bump the URB offset by 1.
2212 /* Skip; gl_TessLevelInner[] doesn't exist for isolines. */
2215 unreachable("Bogus tessellation domain");
2217 } else if (imm_offset
== 1) {
2218 /* This is a write to gl_TessLevelOuter[] which lives in the
2219 * Patch URB Header at DWords 4-7. However, it's reversed, so
2220 * instead of .xyzw we have .wzyx.
2222 value
.type
= BRW_REGISTER_TYPE_F
;
2224 mask
&= (1 << tesslevel_outer_components(tcs_key
->tes_primitive_mode
)) - 1;
2226 if (tcs_key
->tes_primitive_mode
== GL_ISOLINES
) {
2227 /* Isolines .xy should be stored in .zw, in order. */
2228 swiz
= BRW_SWIZZLE4(0, 0, 0, 1);
2231 /* Other domains are reversed; store .wzyx instead of .xyzw */
2232 swiz
= BRW_SWIZZLE_WZYX
;
2233 mask
= writemask_for_backwards_vector(mask
);
2241 unsigned num_components
= _mesa_fls(mask
);
2244 if (mask
!= WRITEMASK_XYZW
) {
2245 srcs
[header_regs
++] = brw_imm_ud(mask
<< 16);
2246 opcode
= indirect_offset
.file
!= BAD_FILE
?
2247 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2248 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2250 opcode
= indirect_offset
.file
!= BAD_FILE
?
2251 SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2252 SHADER_OPCODE_URB_WRITE_SIMD8
;
2255 for (unsigned i
= 0; i
< num_components
; i
++) {
2256 if (mask
& (1 << i
))
2257 srcs
[header_regs
+ i
] = offset(value
, bld
, BRW_GET_SWZ(swiz
, i
));
2260 unsigned mlen
= header_regs
+ num_components
;
2263 bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2264 bld
.LOAD_PAYLOAD(payload
, srcs
, mlen
, header_regs
);
2266 fs_inst
*inst
= bld
.emit(opcode
, bld
.null_reg_ud(), payload
);
2267 inst
->offset
= imm_offset
;
2269 inst
->base_mrf
= -1;
2274 nir_emit_intrinsic(bld
, instr
);
2280 fs_visitor::nir_emit_tes_intrinsic(const fs_builder
&bld
,
2281 nir_intrinsic_instr
*instr
)
2283 assert(stage
== MESA_SHADER_TESS_EVAL
);
2284 struct brw_tes_prog_data
*tes_prog_data
= (struct brw_tes_prog_data
*) prog_data
;
2287 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2288 dest
= get_nir_dest(instr
->dest
);
2290 switch (instr
->intrinsic
) {
2291 case nir_intrinsic_load_primitive_id
:
2292 bld
.MOV(dest
, fs_reg(brw_vec1_grf(0, 1)));
2294 case nir_intrinsic_load_tess_coord
:
2295 /* gl_TessCoord is part of the payload in g1-3 */
2296 for (unsigned i
= 0; i
< 3; i
++) {
2297 bld
.MOV(offset(dest
, bld
, i
), fs_reg(brw_vec8_grf(1 + i
, 0)));
2301 case nir_intrinsic_load_tess_level_outer
:
2302 /* When the TES reads gl_TessLevelOuter, we ensure that the patch header
2303 * appears as a push-model input. So, we can simply use the ATTR file
2304 * rather than issuing URB read messages. The data is stored in the
2305 * high DWords in reverse order - DWord 7 contains .x, DWord 6 contains
2308 switch (tes_prog_data
->domain
) {
2309 case BRW_TESS_DOMAIN_QUAD
:
2310 for (unsigned i
= 0; i
< 4; i
++)
2311 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2313 case BRW_TESS_DOMAIN_TRI
:
2314 for (unsigned i
= 0; i
< 3; i
++)
2315 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2317 case BRW_TESS_DOMAIN_ISOLINE
:
2318 for (unsigned i
= 0; i
< 2; i
++)
2319 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2324 case nir_intrinsic_load_tess_level_inner
:
2325 /* When the TES reads gl_TessLevelInner, we ensure that the patch header
2326 * appears as a push-model input. So, we can simply use the ATTR file
2327 * rather than issuing URB read messages.
2329 switch (tes_prog_data
->domain
) {
2330 case BRW_TESS_DOMAIN_QUAD
:
2331 bld
.MOV(dest
, component(fs_reg(ATTR
, 0), 3));
2332 bld
.MOV(offset(dest
, bld
, 1), component(fs_reg(ATTR
, 0), 2));
2334 case BRW_TESS_DOMAIN_TRI
:
2335 bld
.MOV(dest
, component(fs_reg(ATTR
, 0), 4));
2337 case BRW_TESS_DOMAIN_ISOLINE
:
2338 /* ignore - value is undefined */
2343 case nir_intrinsic_load_input
:
2344 case nir_intrinsic_load_per_vertex_input
: {
2345 fs_reg indirect_offset
= get_indirect_offset(instr
);
2346 unsigned imm_offset
= instr
->const_index
[0];
2349 if (indirect_offset
.file
== BAD_FILE
) {
2350 /* Arbitrarily only push up to 32 vec4 slots worth of data,
2351 * which is 16 registers (since each holds 2 vec4 slots).
2353 const unsigned max_push_slots
= 32;
2354 if (imm_offset
< max_push_slots
) {
2355 fs_reg src
= fs_reg(ATTR
, imm_offset
/ 2, dest
.type
);
2356 for (int i
= 0; i
< instr
->num_components
; i
++) {
2357 bld
.MOV(offset(dest
, bld
, i
),
2358 component(src
, 4 * (imm_offset
% 2) + i
));
2360 tes_prog_data
->base
.urb_read_length
=
2361 MAX2(tes_prog_data
->base
.urb_read_length
,
2362 DIV_ROUND_UP(imm_offset
+ 1, 2));
2364 /* Replicate the patch handle to all enabled channels */
2365 const fs_reg srcs
[] = {
2366 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)
2368 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2369 bld
.LOAD_PAYLOAD(patch_handle
, srcs
, ARRAY_SIZE(srcs
), 0);
2371 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dest
, patch_handle
);
2373 inst
->offset
= imm_offset
;
2374 inst
->base_mrf
= -1;
2375 inst
->regs_written
= instr
->num_components
;
2378 /* Indirect indexing - use per-slot offsets as well. */
2379 const fs_reg srcs
[] = {
2380 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2383 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2384 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2386 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dest
, payload
);
2388 inst
->offset
= imm_offset
;
2389 inst
->base_mrf
= -1;
2390 inst
->regs_written
= instr
->num_components
;
2395 nir_emit_intrinsic(bld
, instr
);
2401 fs_visitor::nir_emit_gs_intrinsic(const fs_builder
&bld
,
2402 nir_intrinsic_instr
*instr
)
2404 assert(stage
== MESA_SHADER_GEOMETRY
);
2405 fs_reg indirect_offset
;
2408 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2409 dest
= get_nir_dest(instr
->dest
);
2411 switch (instr
->intrinsic
) {
2412 case nir_intrinsic_load_primitive_id
:
2413 assert(stage
== MESA_SHADER_GEOMETRY
);
2414 assert(((struct brw_gs_prog_data
*)prog_data
)->include_primitive_id
);
2415 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
2416 retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD
));
2419 case nir_intrinsic_load_input
:
2420 unreachable("load_input intrinsics are invalid for the GS stage");
2422 case nir_intrinsic_load_per_vertex_input
:
2423 emit_gs_input_load(dest
, instr
->src
[0], instr
->const_index
[0],
2424 instr
->src
[1], instr
->num_components
);
2427 case nir_intrinsic_emit_vertex_with_counter
:
2428 emit_gs_vertex(instr
->src
[0], instr
->const_index
[0]);
2431 case nir_intrinsic_end_primitive_with_counter
:
2432 emit_gs_end_primitive(instr
->src
[0]);
2435 case nir_intrinsic_set_vertex_count
:
2436 bld
.MOV(this->final_gs_vertex_count
, get_nir_src(instr
->src
[0]));
2439 case nir_intrinsic_load_invocation_id
: {
2440 fs_reg val
= nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
2441 assert(val
.file
!= BAD_FILE
);
2442 dest
.type
= val
.type
;
2448 nir_emit_intrinsic(bld
, instr
);
2454 fs_visitor::nir_emit_fs_intrinsic(const fs_builder
&bld
,
2455 nir_intrinsic_instr
*instr
)
2457 assert(stage
== MESA_SHADER_FRAGMENT
);
2458 struct brw_wm_prog_data
*wm_prog_data
=
2459 (struct brw_wm_prog_data
*) prog_data
;
2460 const struct brw_wm_prog_key
*wm_key
= (const struct brw_wm_prog_key
*) key
;
2463 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2464 dest
= get_nir_dest(instr
->dest
);
2466 switch (instr
->intrinsic
) {
2467 case nir_intrinsic_load_front_face
:
2468 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
2469 *emit_frontfacing_interpolation());
2472 case nir_intrinsic_load_sample_pos
: {
2473 fs_reg sample_pos
= nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
2474 assert(sample_pos
.file
!= BAD_FILE
);
2475 dest
.type
= sample_pos
.type
;
2476 bld
.MOV(dest
, sample_pos
);
2477 bld
.MOV(offset(dest
, bld
, 1), offset(sample_pos
, bld
, 1));
2481 case nir_intrinsic_load_helper_invocation
:
2482 case nir_intrinsic_load_sample_mask_in
:
2483 case nir_intrinsic_load_sample_id
: {
2484 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
2485 fs_reg val
= nir_system_values
[sv
];
2486 assert(val
.file
!= BAD_FILE
);
2487 dest
.type
= val
.type
;
2492 case nir_intrinsic_discard
:
2493 case nir_intrinsic_discard_if
: {
2494 /* We track our discarded pixels in f0.1. By predicating on it, we can
2495 * update just the flag bits that aren't yet discarded. If there's no
2496 * condition, we emit a CMP of g0 != g0, so all currently executing
2497 * channels will get turned off.
2500 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
2501 cmp
= bld
.CMP(bld
.null_reg_f(), get_nir_src(instr
->src
[0]),
2502 brw_imm_d(0), BRW_CONDITIONAL_Z
);
2504 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
2505 BRW_REGISTER_TYPE_UW
));
2506 cmp
= bld
.CMP(bld
.null_reg_f(), some_reg
, some_reg
, BRW_CONDITIONAL_NZ
);
2508 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
2509 cmp
->flag_subreg
= 1;
2511 if (devinfo
->gen
>= 6) {
2512 emit_discard_jump();
2517 case nir_intrinsic_interp_var_at_centroid
:
2518 case nir_intrinsic_interp_var_at_sample
:
2519 case nir_intrinsic_interp_var_at_offset
: {
2520 /* Handle ARB_gpu_shader5 interpolation intrinsics
2522 * It's worth a quick word of explanation as to why we handle the full
2523 * variable-based interpolation intrinsic rather than a lowered version
2524 * with like we do for other inputs. We have to do that because the way
2525 * we set up inputs doesn't allow us to use the already setup inputs for
2526 * interpolation. At the beginning of the shader, we go through all of
2527 * the input variables and do the initial interpolation and put it in
2528 * the nir_inputs array based on its location as determined in
2529 * nir_lower_io. If the input isn't used, dead code cleans up and
2530 * everything works fine. However, when we get to the ARB_gpu_shader5
2531 * interpolation intrinsics, we need to reinterpolate the input
2532 * differently. If we used an intrinsic that just had an index it would
2533 * only give us the offset into the nir_inputs array. However, this is
2534 * useless because that value is post-interpolation and we need
2535 * pre-interpolation. In order to get the actual location of the bits
2536 * we get from the vertex fetching hardware, we need the variable.
2538 wm_prog_data
->pulls_bary
= true;
2540 fs_reg dst_xy
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 2);
2541 const glsl_interp_qualifier interpolation
=
2542 (glsl_interp_qualifier
) instr
->variables
[0]->var
->data
.interpolation
;
2544 switch (instr
->intrinsic
) {
2545 case nir_intrinsic_interp_var_at_centroid
:
2546 emit_pixel_interpolater_send(bld
,
2547 FS_OPCODE_INTERPOLATE_AT_CENTROID
,
2554 case nir_intrinsic_interp_var_at_sample
: {
2555 if (!wm_key
->multisample_fbo
) {
2556 /* From the ARB_gpu_shader5 specification:
2557 * "If multisample buffers are not available, the input varying
2558 * will be evaluated at the center of the pixel."
2560 emit_pixel_interpolater_send(bld
,
2561 FS_OPCODE_INTERPOLATE_AT_CENTROID
,
2569 nir_const_value
*const_sample
= nir_src_as_const_value(instr
->src
[0]);
2572 unsigned msg_data
= const_sample
->i32
[0] << 4;
2574 emit_pixel_interpolater_send(bld
,
2575 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
2578 brw_imm_ud(msg_data
),
2581 const fs_reg sample_src
= retype(get_nir_src(instr
->src
[0]),
2582 BRW_REGISTER_TYPE_UD
);
2584 if (nir_src_is_dynamically_uniform(instr
->src
[0])) {
2585 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
2586 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
2587 bld
.exec_all().group(1, 0)
2588 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
2589 emit_pixel_interpolater_send(bld
,
2590 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
2596 /* Make a loop that sends a message to the pixel interpolater
2597 * for the sample number in each live channel. If there are
2598 * multiple channels with the same sample number then these
2599 * will be handled simultaneously with a single interation of
2602 bld
.emit(BRW_OPCODE_DO
);
2604 /* Get the next live sample number into sample_id_reg */
2605 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
2607 /* Set the flag register so that we can perform the send
2608 * message on all channels that have the same sample number
2610 bld
.CMP(bld
.null_reg_ud(),
2611 sample_src
, sample_id
,
2612 BRW_CONDITIONAL_EQ
);
2613 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
2614 bld
.exec_all().group(1, 0)
2615 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
2617 emit_pixel_interpolater_send(bld
,
2618 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
2623 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
2625 /* Continue the loop if there are any live channels left */
2626 set_predicate_inv(BRW_PREDICATE_NORMAL
,
2628 bld
.emit(BRW_OPCODE_WHILE
));
2635 case nir_intrinsic_interp_var_at_offset
: {
2636 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
2639 unsigned off_x
= MIN2((int)(const_offset
->f32
[0] * 16), 7) & 0xf;
2640 unsigned off_y
= MIN2((int)(const_offset
->f32
[1] * 16), 7) & 0xf;
2642 emit_pixel_interpolater_send(bld
,
2643 FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
,
2646 brw_imm_ud(off_x
| (off_y
<< 4)),
2649 fs_reg src
= vgrf(glsl_type::ivec2_type
);
2650 fs_reg offset_src
= retype(get_nir_src(instr
->src
[0]),
2651 BRW_REGISTER_TYPE_F
);
2652 for (int i
= 0; i
< 2; i
++) {
2653 fs_reg temp
= vgrf(glsl_type::float_type
);
2654 bld
.MUL(temp
, offset(offset_src
, bld
, i
), brw_imm_f(16.0f
));
2655 fs_reg itemp
= vgrf(glsl_type::int_type
);
2656 bld
.MOV(itemp
, temp
); /* float to int */
2658 /* Clamp the upper end of the range to +7/16.
2659 * ARB_gpu_shader5 requires that we support a maximum offset
2660 * of +0.5, which isn't representable in a S0.4 value -- if
2661 * we didn't clamp it, we'd end up with -8/16, which is the
2662 * opposite of what the shader author wanted.
2664 * This is legal due to ARB_gpu_shader5's quantization
2667 * "Not all values of <offset> may be supported; x and y
2668 * offsets may be rounded to fixed-point values with the
2669 * number of fraction bits given by the
2670 * implementation-dependent constant
2671 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
2673 set_condmod(BRW_CONDITIONAL_L
,
2674 bld
.SEL(offset(src
, bld
, i
), itemp
, brw_imm_d(7)));
2677 const enum opcode opcode
= FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
;
2678 emit_pixel_interpolater_send(bld
,
2689 unreachable("Invalid intrinsic");
2692 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
2693 fs_reg src
= interp_reg(instr
->variables
[0]->var
->data
.location
, j
);
2694 src
.type
= dest
.type
;
2696 bld
.emit(FS_OPCODE_LINTERP
, dest
, dst_xy
, src
);
2697 dest
= offset(dest
, bld
, 1);
2702 nir_emit_intrinsic(bld
, instr
);
2708 fs_visitor::nir_emit_cs_intrinsic(const fs_builder
&bld
,
2709 nir_intrinsic_instr
*instr
)
2711 assert(stage
== MESA_SHADER_COMPUTE
);
2712 struct brw_cs_prog_data
*cs_prog_data
=
2713 (struct brw_cs_prog_data
*) prog_data
;
2716 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2717 dest
= get_nir_dest(instr
->dest
);
2719 switch (instr
->intrinsic
) {
2720 case nir_intrinsic_barrier
:
2722 cs_prog_data
->uses_barrier
= true;
2725 case nir_intrinsic_load_local_invocation_id
:
2726 case nir_intrinsic_load_work_group_id
: {
2727 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
2728 fs_reg val
= nir_system_values
[sv
];
2729 assert(val
.file
!= BAD_FILE
);
2730 dest
.type
= val
.type
;
2731 for (unsigned i
= 0; i
< 3; i
++)
2732 bld
.MOV(offset(dest
, bld
, i
), offset(val
, bld
, i
));
2736 case nir_intrinsic_load_num_work_groups
: {
2737 const unsigned surface
=
2738 cs_prog_data
->binding_table
.work_groups_start
;
2740 cs_prog_data
->uses_num_work_groups
= true;
2742 fs_reg surf_index
= brw_imm_ud(surface
);
2743 brw_mark_surface_used(prog_data
, surface
);
2745 /* Read the 3 GLuint components of gl_NumWorkGroups */
2746 for (unsigned i
= 0; i
< 3; i
++) {
2747 fs_reg read_result
=
2748 emit_untyped_read(bld
, surf_index
,
2750 1 /* dims */, 1 /* size */,
2751 BRW_PREDICATE_NONE
);
2752 read_result
.type
= dest
.type
;
2753 bld
.MOV(dest
, read_result
);
2754 dest
= offset(dest
, bld
, 1);
2759 case nir_intrinsic_shared_atomic_add
:
2760 nir_emit_shared_atomic(bld
, BRW_AOP_ADD
, instr
);
2762 case nir_intrinsic_shared_atomic_imin
:
2763 nir_emit_shared_atomic(bld
, BRW_AOP_IMIN
, instr
);
2765 case nir_intrinsic_shared_atomic_umin
:
2766 nir_emit_shared_atomic(bld
, BRW_AOP_UMIN
, instr
);
2768 case nir_intrinsic_shared_atomic_imax
:
2769 nir_emit_shared_atomic(bld
, BRW_AOP_IMAX
, instr
);
2771 case nir_intrinsic_shared_atomic_umax
:
2772 nir_emit_shared_atomic(bld
, BRW_AOP_UMAX
, instr
);
2774 case nir_intrinsic_shared_atomic_and
:
2775 nir_emit_shared_atomic(bld
, BRW_AOP_AND
, instr
);
2777 case nir_intrinsic_shared_atomic_or
:
2778 nir_emit_shared_atomic(bld
, BRW_AOP_OR
, instr
);
2780 case nir_intrinsic_shared_atomic_xor
:
2781 nir_emit_shared_atomic(bld
, BRW_AOP_XOR
, instr
);
2783 case nir_intrinsic_shared_atomic_exchange
:
2784 nir_emit_shared_atomic(bld
, BRW_AOP_MOV
, instr
);
2786 case nir_intrinsic_shared_atomic_comp_swap
:
2787 nir_emit_shared_atomic(bld
, BRW_AOP_CMPWR
, instr
);
2790 case nir_intrinsic_load_shared
: {
2791 assert(devinfo
->gen
>= 7);
2793 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
2795 /* Get the offset to read from */
2797 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
2799 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0]);
2801 offset_reg
= vgrf(glsl_type::uint_type
);
2803 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
2804 brw_imm_ud(instr
->const_index
[0]));
2807 /* Read the vector */
2808 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, offset_reg
,
2810 instr
->num_components
,
2811 BRW_PREDICATE_NONE
);
2812 read_result
.type
= dest
.type
;
2813 for (int i
= 0; i
< instr
->num_components
; i
++)
2814 bld
.MOV(offset(dest
, bld
, i
), offset(read_result
, bld
, i
));
2819 case nir_intrinsic_store_shared
: {
2820 assert(devinfo
->gen
>= 7);
2823 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
2826 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
2829 unsigned writemask
= instr
->const_index
[1];
2831 /* Combine groups of consecutive enabled channels in one write
2832 * message. We use ffs to find the first enabled channel and then ffs on
2833 * the bit-inverse, down-shifted writemask to determine the length of
2834 * the block of enabled bits.
2837 unsigned first_component
= ffs(writemask
) - 1;
2838 unsigned length
= ffs(~(writemask
>> first_component
)) - 1;
2841 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
2843 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0] +
2844 4 * first_component
);
2846 offset_reg
= vgrf(glsl_type::uint_type
);
2848 retype(get_nir_src(instr
->src
[1]), BRW_REGISTER_TYPE_UD
),
2849 brw_imm_ud(instr
->const_index
[0] + 4 * first_component
));
2852 emit_untyped_write(bld
, surf_index
, offset_reg
,
2853 offset(val_reg
, bld
, first_component
),
2854 1 /* dims */, length
,
2855 BRW_PREDICATE_NONE
);
2857 /* Clear the bits in the writemask that we just wrote, then try
2858 * again to see if more channels are left.
2860 writemask
&= (15 << (first_component
+ length
));
2867 nir_emit_intrinsic(bld
, instr
);
2873 fs_visitor::nir_emit_intrinsic(const fs_builder
&bld
, nir_intrinsic_instr
*instr
)
2876 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2877 dest
= get_nir_dest(instr
->dest
);
2879 switch (instr
->intrinsic
) {
2880 case nir_intrinsic_atomic_counter_inc
:
2881 case nir_intrinsic_atomic_counter_dec
:
2882 case nir_intrinsic_atomic_counter_read
: {
2883 /* Get the arguments of the atomic intrinsic. */
2884 const fs_reg offset
= get_nir_src(instr
->src
[0]);
2885 const unsigned surface
= (stage_prog_data
->binding_table
.abo_start
+
2886 instr
->const_index
[0]);
2889 /* Emit a surface read or atomic op. */
2890 switch (instr
->intrinsic
) {
2891 case nir_intrinsic_atomic_counter_read
:
2892 tmp
= emit_untyped_read(bld
, brw_imm_ud(surface
), offset
, 1, 1);
2895 case nir_intrinsic_atomic_counter_inc
:
2896 tmp
= emit_untyped_atomic(bld
, brw_imm_ud(surface
), offset
, fs_reg(),
2897 fs_reg(), 1, 1, BRW_AOP_INC
);
2900 case nir_intrinsic_atomic_counter_dec
:
2901 tmp
= emit_untyped_atomic(bld
, brw_imm_ud(surface
), offset
, fs_reg(),
2902 fs_reg(), 1, 1, BRW_AOP_PREDEC
);
2906 unreachable("Unreachable");
2909 /* Assign the result. */
2910 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
), tmp
);
2912 /* Mark the surface as used. */
2913 brw_mark_surface_used(stage_prog_data
, surface
);
2917 case nir_intrinsic_image_load
:
2918 case nir_intrinsic_image_store
:
2919 case nir_intrinsic_image_atomic_add
:
2920 case nir_intrinsic_image_atomic_min
:
2921 case nir_intrinsic_image_atomic_max
:
2922 case nir_intrinsic_image_atomic_and
:
2923 case nir_intrinsic_image_atomic_or
:
2924 case nir_intrinsic_image_atomic_xor
:
2925 case nir_intrinsic_image_atomic_exchange
:
2926 case nir_intrinsic_image_atomic_comp_swap
: {
2927 using namespace image_access
;
2929 /* Get the referenced image variable and type. */
2930 const nir_variable
*var
= instr
->variables
[0]->var
;
2931 const glsl_type
*type
= var
->type
->without_array();
2932 const brw_reg_type base_type
= get_image_base_type(type
);
2934 /* Get some metadata from the image intrinsic. */
2935 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
2936 const unsigned arr_dims
= type
->sampler_array
? 1 : 0;
2937 const unsigned surf_dims
= type
->coordinate_components() - arr_dims
;
2938 const unsigned format
= var
->data
.image
.format
;
2940 /* Get the arguments of the image intrinsic. */
2941 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
2942 const fs_reg addr
= retype(get_nir_src(instr
->src
[0]),
2943 BRW_REGISTER_TYPE_UD
);
2944 const fs_reg src0
= (info
->num_srcs
>= 3 ?
2945 retype(get_nir_src(instr
->src
[2]), base_type
) :
2947 const fs_reg src1
= (info
->num_srcs
>= 4 ?
2948 retype(get_nir_src(instr
->src
[3]), base_type
) :
2952 /* Emit an image load, store or atomic op. */
2953 if (instr
->intrinsic
== nir_intrinsic_image_load
)
2954 tmp
= emit_image_load(bld
, image
, addr
, surf_dims
, arr_dims
, format
);
2956 else if (instr
->intrinsic
== nir_intrinsic_image_store
)
2957 emit_image_store(bld
, image
, addr
, src0
, surf_dims
, arr_dims
,
2958 var
->data
.image
.write_only
? GL_NONE
: format
);
2961 tmp
= emit_image_atomic(bld
, image
, addr
, src0
, src1
,
2962 surf_dims
, arr_dims
, info
->dest_components
,
2963 get_image_atomic_op(instr
->intrinsic
, type
));
2965 /* Assign the result. */
2966 for (unsigned c
= 0; c
< info
->dest_components
; ++c
)
2967 bld
.MOV(offset(retype(dest
, base_type
), bld
, c
),
2968 offset(tmp
, bld
, c
));
2972 case nir_intrinsic_memory_barrier_atomic_counter
:
2973 case nir_intrinsic_memory_barrier_buffer
:
2974 case nir_intrinsic_memory_barrier_image
:
2975 case nir_intrinsic_memory_barrier
: {
2976 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 16 / dispatch_width
);
2977 bld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
)
2982 case nir_intrinsic_group_memory_barrier
:
2983 case nir_intrinsic_memory_barrier_shared
:
2984 /* We treat these workgroup-level barriers as no-ops. This should be
2985 * safe at present and as long as:
2987 * - Memory access instructions are not subsequently reordered by the
2988 * compiler back-end.
2990 * - All threads from a given compute shader workgroup fit within a
2991 * single subslice and therefore talk to the same HDC shared unit
2992 * what supposedly guarantees ordering and coherency between threads
2993 * from the same workgroup. This may change in the future when we
2994 * start splitting workgroups across multiple subslices.
2996 * - The context is not in fault-and-stream mode, which could cause
2997 * memory transactions (including to SLM) prior to the barrier to be
2998 * replayed after the barrier if a pagefault occurs. This shouldn't
2999 * be a problem up to and including SKL because fault-and-stream is
3000 * not usable due to hardware issues, but that's likely to change in
3005 case nir_intrinsic_shader_clock
: {
3006 /* We cannot do anything if there is an event, so ignore it for now */
3007 fs_reg shader_clock
= get_timestamp(bld
);
3008 const fs_reg srcs
[] = { shader_clock
.set_smear(0), shader_clock
.set_smear(1) };
3010 bld
.LOAD_PAYLOAD(dest
, srcs
, ARRAY_SIZE(srcs
), 0);
3014 case nir_intrinsic_image_size
: {
3015 /* Get the referenced image variable and type. */
3016 const nir_variable
*var
= instr
->variables
[0]->var
;
3017 const glsl_type
*type
= var
->type
->without_array();
3019 /* Get the size of the image. */
3020 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
3021 const fs_reg size
= offset(image
, bld
, BRW_IMAGE_PARAM_SIZE_OFFSET
);
3023 /* For 1DArray image types, the array index is stored in the Z component.
3024 * Fix this by swizzling the Z component to the Y component.
3026 const bool is_1d_array_image
=
3027 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_1D
&&
3028 type
->sampler_array
;
3030 /* For CubeArray images, we should count the number of cubes instead
3031 * of the number of faces. Fix it by dividing the (Z component) by 6.
3033 const bool is_cube_array_image
=
3034 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
3035 type
->sampler_array
;
3037 /* Copy all the components. */
3038 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
3039 for (unsigned c
= 0; c
< info
->dest_components
; ++c
) {
3040 if ((int)c
>= type
->coordinate_components()) {
3041 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3043 } else if (c
== 1 && is_1d_array_image
) {
3044 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3045 offset(size
, bld
, 2));
3046 } else if (c
== 2 && is_cube_array_image
) {
3047 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
,
3048 offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3049 offset(size
, bld
, c
), brw_imm_d(6));
3051 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3052 offset(size
, bld
, c
));
3059 case nir_intrinsic_image_samples
:
3060 /* The driver does not support multi-sampled images. */
3061 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), brw_imm_d(1));
3064 case nir_intrinsic_load_uniform
: {
3065 /* Offsets are in bytes but they should always be multiples of 4 */
3066 assert(instr
->const_index
[0] % 4 == 0);
3068 fs_reg
src(UNIFORM
, instr
->const_index
[0] / 4, dest
.type
);
3070 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3072 /* Offsets are in bytes but they should always be multiples of 4 */
3073 assert(const_offset
->u32
[0] % 4 == 0);
3074 src
.reg_offset
= const_offset
->u32
[0] / 4;
3076 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3077 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
3080 fs_reg indirect
= retype(get_nir_src(instr
->src
[0]),
3081 BRW_REGISTER_TYPE_UD
);
3083 /* We need to pass a size to the MOV_INDIRECT but we don't want it to
3084 * go past the end of the uniform. In order to keep the n'th
3085 * component from running past, we subtract off the size of all but
3086 * one component of the vector.
3088 assert(instr
->const_index
[1] >= instr
->num_components
* 4);
3089 unsigned read_size
= instr
->const_index
[1] -
3090 (instr
->num_components
- 1) * 4;
3092 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3093 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
3094 offset(dest
, bld
, j
), offset(src
, bld
, j
),
3095 indirect
, brw_imm_ud(read_size
));
3101 case nir_intrinsic_load_ubo
: {
3102 nir_const_value
*const_index
= nir_src_as_const_value(instr
->src
[0]);
3106 const unsigned index
= stage_prog_data
->binding_table
.ubo_start
+
3107 const_index
->u32
[0];
3108 surf_index
= brw_imm_ud(index
);
3109 brw_mark_surface_used(prog_data
, index
);
3111 /* The block index is not a constant. Evaluate the index expression
3112 * per-channel and add the base UBO index; we have to select a value
3113 * from any live channel.
3115 surf_index
= vgrf(glsl_type::uint_type
);
3116 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
3117 brw_imm_ud(stage_prog_data
->binding_table
.ubo_start
));
3118 surf_index
= bld
.emit_uniformize(surf_index
);
3120 /* Assume this may touch any UBO. It would be nice to provide
3121 * a tighter bound, but the array information is already lowered away.
3123 brw_mark_surface_used(prog_data
,
3124 stage_prog_data
->binding_table
.ubo_start
+
3125 nir
->info
.num_ubos
- 1);
3128 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3129 if (const_offset
== NULL
) {
3130 fs_reg base_offset
= retype(get_nir_src(instr
->src
[1]),
3131 BRW_REGISTER_TYPE_UD
);
3133 for (int i
= 0; i
< instr
->num_components
; i
++)
3134 VARYING_PULL_CONSTANT_LOAD(bld
, offset(dest
, bld
, i
), surf_index
,
3135 base_offset
, i
* 4);
3137 fs_reg packed_consts
= vgrf(glsl_type::float_type
);
3138 packed_consts
.type
= dest
.type
;
3140 struct brw_reg const_offset_reg
= brw_imm_ud(const_offset
->u32
[0] & ~15);
3141 bld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
, packed_consts
,
3142 surf_index
, const_offset_reg
);
3144 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
3145 packed_consts
.set_smear(const_offset
->u32
[0] % 16 / 4 + i
);
3147 /* The std140 packing rules don't allow vectors to cross 16-byte
3148 * boundaries, and a reg is 32 bytes.
3150 assert(packed_consts
.subreg_offset
< 32);
3152 bld
.MOV(dest
, packed_consts
);
3153 dest
= offset(dest
, bld
, 1);
3159 case nir_intrinsic_load_ssbo
: {
3160 assert(devinfo
->gen
>= 7);
3162 nir_const_value
*const_uniform_block
=
3163 nir_src_as_const_value(instr
->src
[0]);
3166 if (const_uniform_block
) {
3167 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3168 const_uniform_block
->u32
[0];
3169 surf_index
= brw_imm_ud(index
);
3170 brw_mark_surface_used(prog_data
, index
);
3172 surf_index
= vgrf(glsl_type::uint_type
);
3173 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
3174 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3176 /* Assume this may touch any UBO. It would be nice to provide
3177 * a tighter bound, but the array information is already lowered away.
3179 brw_mark_surface_used(prog_data
,
3180 stage_prog_data
->binding_table
.ssbo_start
+
3181 nir
->info
.num_ssbos
- 1);
3185 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3187 offset_reg
= brw_imm_ud(const_offset
->u32
[0]);
3189 offset_reg
= get_nir_src(instr
->src
[1]);
3192 /* Read the vector */
3193 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, offset_reg
,
3195 instr
->num_components
,
3196 BRW_PREDICATE_NONE
);
3197 read_result
.type
= dest
.type
;
3198 for (int i
= 0; i
< instr
->num_components
; i
++)
3199 bld
.MOV(offset(dest
, bld
, i
), offset(read_result
, bld
, i
));
3204 case nir_intrinsic_load_input
: {
3206 if (stage
== MESA_SHADER_VERTEX
) {
3207 src
= fs_reg(ATTR
, instr
->const_index
[0], dest
.type
);
3209 src
= offset(retype(nir_inputs
, dest
.type
), bld
,
3210 instr
->const_index
[0]);
3213 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3214 assert(const_offset
&& "Indirect input loads not allowed");
3215 src
= offset(src
, bld
, const_offset
->u32
[0]);
3217 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3218 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
3223 case nir_intrinsic_store_ssbo
: {
3224 assert(devinfo
->gen
>= 7);
3228 nir_const_value
*const_uniform_block
=
3229 nir_src_as_const_value(instr
->src
[1]);
3230 if (const_uniform_block
) {
3231 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3232 const_uniform_block
->u32
[0];
3233 surf_index
= brw_imm_ud(index
);
3234 brw_mark_surface_used(prog_data
, index
);
3236 surf_index
= vgrf(glsl_type::uint_type
);
3237 bld
.ADD(surf_index
, get_nir_src(instr
->src
[1]),
3238 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3240 brw_mark_surface_used(prog_data
,
3241 stage_prog_data
->binding_table
.ssbo_start
+
3242 nir
->info
.num_ssbos
- 1);
3246 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
3249 unsigned writemask
= instr
->const_index
[0];
3251 /* Combine groups of consecutive enabled channels in one write
3252 * message. We use ffs to find the first enabled channel and then ffs on
3253 * the bit-inverse, down-shifted writemask to determine the length of
3254 * the block of enabled bits.
3257 unsigned first_component
= ffs(writemask
) - 1;
3258 unsigned length
= ffs(~(writemask
>> first_component
)) - 1;
3261 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[2]);
3263 offset_reg
= brw_imm_ud(const_offset
->u32
[0] + 4 * first_component
);
3265 offset_reg
= vgrf(glsl_type::uint_type
);
3267 retype(get_nir_src(instr
->src
[2]), BRW_REGISTER_TYPE_UD
),
3268 brw_imm_ud(4 * first_component
));
3271 emit_untyped_write(bld
, surf_index
, offset_reg
,
3272 offset(val_reg
, bld
, first_component
),
3273 1 /* dims */, length
,
3274 BRW_PREDICATE_NONE
);
3276 /* Clear the bits in the writemask that we just wrote, then try
3277 * again to see if more channels are left.
3279 writemask
&= (15 << (first_component
+ length
));
3284 case nir_intrinsic_store_output
: {
3285 fs_reg src
= get_nir_src(instr
->src
[0]);
3286 fs_reg new_dest
= offset(retype(nir_outputs
, src
.type
), bld
,
3287 instr
->const_index
[0]);
3289 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3290 assert(const_offset
&& "Indirect output stores not allowed");
3291 new_dest
= offset(new_dest
, bld
, const_offset
->u32
[0]);
3293 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3294 bld
.MOV(offset(new_dest
, bld
, j
), offset(src
, bld
, j
));
3299 case nir_intrinsic_ssbo_atomic_add
:
3300 nir_emit_ssbo_atomic(bld
, BRW_AOP_ADD
, instr
);
3302 case nir_intrinsic_ssbo_atomic_imin
:
3303 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMIN
, instr
);
3305 case nir_intrinsic_ssbo_atomic_umin
:
3306 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMIN
, instr
);
3308 case nir_intrinsic_ssbo_atomic_imax
:
3309 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMAX
, instr
);
3311 case nir_intrinsic_ssbo_atomic_umax
:
3312 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMAX
, instr
);
3314 case nir_intrinsic_ssbo_atomic_and
:
3315 nir_emit_ssbo_atomic(bld
, BRW_AOP_AND
, instr
);
3317 case nir_intrinsic_ssbo_atomic_or
:
3318 nir_emit_ssbo_atomic(bld
, BRW_AOP_OR
, instr
);
3320 case nir_intrinsic_ssbo_atomic_xor
:
3321 nir_emit_ssbo_atomic(bld
, BRW_AOP_XOR
, instr
);
3323 case nir_intrinsic_ssbo_atomic_exchange
:
3324 nir_emit_ssbo_atomic(bld
, BRW_AOP_MOV
, instr
);
3326 case nir_intrinsic_ssbo_atomic_comp_swap
:
3327 nir_emit_ssbo_atomic(bld
, BRW_AOP_CMPWR
, instr
);
3330 case nir_intrinsic_get_buffer_size
: {
3331 nir_const_value
*const_uniform_block
= nir_src_as_const_value(instr
->src
[0]);
3332 unsigned ssbo_index
= const_uniform_block
? const_uniform_block
->u32
[0] : 0;
3333 int reg_width
= dispatch_width
/ 8;
3336 fs_reg source
= brw_imm_d(0);
3338 int mlen
= 1 * reg_width
;
3340 /* A resinfo's sampler message is used to get the buffer size.
3341 * The SIMD8's writeback message consists of four registers and
3342 * SIMD16's writeback message consists of 8 destination registers
3343 * (two per each component), although we are only interested on the
3344 * first component, where resinfo returns the buffer size for
3347 int regs_written
= 4 * mlen
;
3348 fs_reg src_payload
= fs_reg(VGRF
, alloc
.allocate(mlen
),
3349 BRW_REGISTER_TYPE_UD
);
3350 bld
.LOAD_PAYLOAD(src_payload
, &source
, 1, 0);
3351 fs_reg buffer_size
= fs_reg(VGRF
, alloc
.allocate(regs_written
),
3352 BRW_REGISTER_TYPE_UD
);
3353 const unsigned index
= prog_data
->binding_table
.ssbo_start
+ ssbo_index
;
3354 fs_inst
*inst
= bld
.emit(FS_OPCODE_GET_BUFFER_SIZE
, buffer_size
,
3355 src_payload
, brw_imm_ud(index
));
3356 inst
->header_size
= 0;
3358 inst
->regs_written
= regs_written
;
3360 bld
.MOV(retype(dest
, buffer_size
.type
), buffer_size
);
3362 brw_mark_surface_used(prog_data
, index
);
3367 unreachable("unknown intrinsic");
3372 fs_visitor::nir_emit_ssbo_atomic(const fs_builder
&bld
,
3373 int op
, nir_intrinsic_instr
*instr
)
3376 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3377 dest
= get_nir_dest(instr
->dest
);
3380 nir_const_value
*const_surface
= nir_src_as_const_value(instr
->src
[0]);
3381 if (const_surface
) {
3382 unsigned surf_index
= stage_prog_data
->binding_table
.ssbo_start
+
3383 const_surface
->u32
[0];
3384 surface
= brw_imm_ud(surf_index
);
3385 brw_mark_surface_used(prog_data
, surf_index
);
3387 surface
= vgrf(glsl_type::uint_type
);
3388 bld
.ADD(surface
, get_nir_src(instr
->src
[0]),
3389 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3391 /* Assume this may touch any SSBO. This is the same we do for other
3392 * UBO/SSBO accesses with non-constant surface.
3394 brw_mark_surface_used(prog_data
,
3395 stage_prog_data
->binding_table
.ssbo_start
+
3396 nir
->info
.num_ssbos
- 1);
3399 fs_reg offset
= get_nir_src(instr
->src
[1]);
3400 fs_reg data1
= get_nir_src(instr
->src
[2]);
3402 if (op
== BRW_AOP_CMPWR
)
3403 data2
= get_nir_src(instr
->src
[3]);
3405 /* Emit the actual atomic operation operation */
3407 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
3409 1 /* dims */, 1 /* rsize */,
3411 BRW_PREDICATE_NONE
);
3412 dest
.type
= atomic_result
.type
;
3413 bld
.MOV(dest
, atomic_result
);
3417 fs_visitor::nir_emit_shared_atomic(const fs_builder
&bld
,
3418 int op
, nir_intrinsic_instr
*instr
)
3421 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3422 dest
= get_nir_dest(instr
->dest
);
3424 fs_reg surface
= brw_imm_ud(GEN7_BTI_SLM
);
3425 fs_reg offset
= get_nir_src(instr
->src
[0]);
3426 fs_reg data1
= get_nir_src(instr
->src
[1]);
3428 if (op
== BRW_AOP_CMPWR
)
3429 data2
= get_nir_src(instr
->src
[2]);
3431 /* Emit the actual atomic operation operation */
3433 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
3435 1 /* dims */, 1 /* rsize */,
3437 BRW_PREDICATE_NONE
);
3438 dest
.type
= atomic_result
.type
;
3439 bld
.MOV(dest
, atomic_result
);
3443 fs_visitor::nir_emit_texture(const fs_builder
&bld
, nir_tex_instr
*instr
)
3445 unsigned texture
= instr
->texture_index
;
3446 unsigned sampler
= instr
->sampler_index
;
3448 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
3450 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture
);
3451 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(sampler
);
3453 int lod_components
= 0;
3455 /* The hardware requires a LOD for buffer textures */
3456 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
3457 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_d(0);
3459 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
3460 fs_reg src
= get_nir_src(instr
->src
[i
].src
);
3461 switch (instr
->src
[i
].src_type
) {
3462 case nir_tex_src_bias
:
3463 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
3465 case nir_tex_src_comparitor
:
3466 srcs
[TEX_LOGICAL_SRC_SHADOW_C
] = retype(src
, BRW_REGISTER_TYPE_F
);
3468 case nir_tex_src_coord
:
3469 switch (instr
->op
) {
3471 case nir_texop_txf_ms
:
3472 case nir_texop_samples_identical
:
3473 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_D
);
3476 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_F
);
3480 case nir_tex_src_ddx
:
3481 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
3482 lod_components
= nir_tex_instr_src_size(instr
, i
);
3484 case nir_tex_src_ddy
:
3485 srcs
[TEX_LOGICAL_SRC_LOD2
] = retype(src
, BRW_REGISTER_TYPE_F
);
3487 case nir_tex_src_lod
:
3488 switch (instr
->op
) {
3490 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_UD
);
3493 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_D
);
3496 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
3500 case nir_tex_src_ms_index
:
3501 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = retype(src
, BRW_REGISTER_TYPE_UD
);
3504 case nir_tex_src_offset
: {
3505 nir_const_value
*const_offset
=
3506 nir_src_as_const_value(instr
->src
[i
].src
);
3508 unsigned header_bits
= brw_texture_offset(const_offset
->i32
, 3);
3509 if (header_bits
!= 0)
3510 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
] = brw_imm_ud(header_bits
);
3512 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
] =
3513 retype(src
, BRW_REGISTER_TYPE_D
);
3518 case nir_tex_src_projector
:
3519 unreachable("should be lowered");
3521 case nir_tex_src_texture_offset
: {
3522 /* Figure out the highest possible texture index and mark it as used */
3523 uint32_t max_used
= texture
+ instr
->texture_array_size
- 1;
3524 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
< 8) {
3525 max_used
+= stage_prog_data
->binding_table
.gather_texture_start
;
3527 max_used
+= stage_prog_data
->binding_table
.texture_start
;
3529 brw_mark_surface_used(prog_data
, max_used
);
3531 /* Emit code to evaluate the actual indexing expression */
3532 fs_reg tmp
= vgrf(glsl_type::uint_type
);
3533 bld
.ADD(tmp
, src
, brw_imm_ud(texture
));
3534 srcs
[TEX_LOGICAL_SRC_SURFACE
] = bld
.emit_uniformize(tmp
);
3538 case nir_tex_src_sampler_offset
: {
3539 /* Emit code to evaluate the actual indexing expression */
3540 fs_reg tmp
= vgrf(glsl_type::uint_type
);
3541 bld
.ADD(tmp
, src
, brw_imm_ud(sampler
));
3542 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = bld
.emit_uniformize(tmp
);
3547 unreachable("unknown texture source");
3551 if (instr
->op
== nir_texop_txf_ms
||
3552 instr
->op
== nir_texop_samples_identical
) {
3553 if (devinfo
->gen
>= 7 &&
3554 key_tex
->compressed_multisample_layout_mask
& (1 << texture
)) {
3555 srcs
[TEX_LOGICAL_SRC_MCS
] =
3556 emit_mcs_fetch(srcs
[TEX_LOGICAL_SRC_COORDINATE
],
3557 instr
->coord_components
,
3558 srcs
[TEX_LOGICAL_SRC_SURFACE
]);
3560 srcs
[TEX_LOGICAL_SRC_MCS
] = brw_imm_ud(0u);
3564 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(instr
->coord_components
);
3565 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(lod_components
);
3567 if (instr
->op
== nir_texop_query_levels
) {
3568 /* textureQueryLevels() is implemented in terms of TXS so we need to
3569 * pass a valid LOD argument.
3571 assert(srcs
[TEX_LOGICAL_SRC_LOD
].file
== BAD_FILE
);
3572 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_ud(0u);
3576 switch (instr
->op
) {
3578 opcode
= SHADER_OPCODE_TEX_LOGICAL
;
3581 opcode
= FS_OPCODE_TXB_LOGICAL
;
3584 opcode
= SHADER_OPCODE_TXL_LOGICAL
;
3587 opcode
= SHADER_OPCODE_TXD_LOGICAL
;
3590 opcode
= SHADER_OPCODE_TXF_LOGICAL
;
3592 case nir_texop_txf_ms
:
3593 if ((key_tex
->msaa_16
& (1 << sampler
)))
3594 opcode
= SHADER_OPCODE_TXF_CMS_W_LOGICAL
;
3596 opcode
= SHADER_OPCODE_TXF_CMS_LOGICAL
;
3598 case nir_texop_query_levels
:
3600 opcode
= SHADER_OPCODE_TXS_LOGICAL
;
3603 opcode
= SHADER_OPCODE_LOD_LOGICAL
;
3606 if (srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
!= BAD_FILE
&&
3607 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
!= IMM
)
3608 opcode
= SHADER_OPCODE_TG4_OFFSET_LOGICAL
;
3610 opcode
= SHADER_OPCODE_TG4_LOGICAL
;
3612 case nir_texop_texture_samples
: {
3613 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
3615 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
, 4);
3616 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_SAMPLEINFO
, tmp
,
3617 bld
.vgrf(BRW_REGISTER_TYPE_D
, 1),
3618 srcs
[TEX_LOGICAL_SRC_SURFACE
],
3619 srcs
[TEX_LOGICAL_SRC_SURFACE
]);
3621 inst
->header_size
= 1;
3622 inst
->base_mrf
= -1;
3623 inst
->regs_written
= 4 * (dispatch_width
/ 8);
3625 /* Pick off the one component we care about */
3629 case nir_texop_samples_identical
: {
3630 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
3632 /* If mcs is an immediate value, it means there is no MCS. In that case
3633 * just return false.
3635 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BRW_IMMEDIATE_VALUE
) {
3636 bld
.MOV(dst
, brw_imm_ud(0u));
3637 } else if ((key_tex
->msaa_16
& (1 << sampler
))) {
3638 fs_reg tmp
= vgrf(glsl_type::uint_type
);
3639 bld
.OR(tmp
, srcs
[TEX_LOGICAL_SRC_MCS
],
3640 offset(srcs
[TEX_LOGICAL_SRC_MCS
], bld
, 1));
3641 bld
.CMP(dst
, tmp
, brw_imm_ud(0u), BRW_CONDITIONAL_EQ
);
3643 bld
.CMP(dst
, srcs
[TEX_LOGICAL_SRC_MCS
], brw_imm_ud(0u),
3644 BRW_CONDITIONAL_EQ
);
3649 unreachable("unknown texture opcode");
3652 fs_reg dst
= bld
.vgrf(brw_type_for_nir_type(instr
->dest_type
), 4);
3653 fs_inst
*inst
= bld
.emit(opcode
, dst
, srcs
, ARRAY_SIZE(srcs
));
3655 const unsigned dest_size
= nir_tex_instr_dest_size(instr
);
3656 if (devinfo
->gen
>= 9 &&
3657 instr
->op
!= nir_texop_tg4
&& instr
->op
!= nir_texop_query_levels
) {
3658 unsigned write_mask
= instr
->dest
.is_ssa
?
3659 nir_ssa_def_components_read(&instr
->dest
.ssa
):
3660 (1 << dest_size
) - 1;
3661 assert(write_mask
!= 0); /* dead code should have been eliminated */
3662 inst
->regs_written
= _mesa_fls(write_mask
) * dispatch_width
/ 8;
3664 inst
->regs_written
= 4 * dispatch_width
/ 8;
3667 if (srcs
[TEX_LOGICAL_SRC_SHADOW_C
].file
!= BAD_FILE
)
3668 inst
->shadow_compare
= true;
3670 if (srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
== IMM
)
3671 inst
->offset
= srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].ud
;
3673 if (instr
->op
== nir_texop_tg4
) {
3674 if (instr
->component
== 1 &&
3675 key_tex
->gather_channel_quirk_mask
& (1 << texture
)) {
3676 /* gather4 sampler is broken for green channel on RG32F --
3677 * we must ask for blue instead.
3679 inst
->offset
|= 2 << 16;
3681 inst
->offset
|= instr
->component
<< 16;
3684 if (devinfo
->gen
== 6)
3685 emit_gen6_gather_wa(key_tex
->gen6_gather_wa
[texture
], dst
);
3689 for (unsigned i
= 0; i
< dest_size
; i
++)
3690 nir_dest
[i
] = offset(dst
, bld
, i
);
3692 bool is_cube_array
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
3695 if (instr
->op
== nir_texop_query_levels
) {
3696 /* # levels is in .w */
3697 nir_dest
[0] = offset(dst
, bld
, 3);
3698 } else if (instr
->op
== nir_texop_txs
&& dest_size
>= 3 &&
3699 (devinfo
->gen
< 7 || is_cube_array
)) {
3700 fs_reg depth
= offset(dst
, bld
, 2);
3701 fs_reg fixed_depth
= vgrf(glsl_type::int_type
);
3703 if (is_cube_array
) {
3704 /* fixup #layers for cube map arrays */
3705 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, fixed_depth
, depth
, brw_imm_d(6));
3706 } else if (devinfo
->gen
< 7) {
3707 /* Gen4-6 return 0 instead of 1 for single layer surfaces. */
3708 bld
.emit_minmax(fixed_depth
, depth
, brw_imm_d(1), BRW_CONDITIONAL_GE
);
3711 nir_dest
[2] = fixed_depth
;
3714 bld
.LOAD_PAYLOAD(get_nir_dest(instr
->dest
), nir_dest
, dest_size
, 0);
3718 fs_visitor::nir_emit_jump(const fs_builder
&bld
, nir_jump_instr
*instr
)
3720 switch (instr
->type
) {
3721 case nir_jump_break
:
3722 bld
.emit(BRW_OPCODE_BREAK
);
3724 case nir_jump_continue
:
3725 bld
.emit(BRW_OPCODE_CONTINUE
);
3727 case nir_jump_return
:
3729 unreachable("unknown jump");