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();
66 emit_percomp(bld
, fs_inst(BRW_OPCODE_MOV
, bld
.dispatch_width(),
68 } else if (var
->data
.location
== VARYING_SLOT_LAYER
) {
69 struct brw_reg reg
= suboffset(interp_reg(VARYING_SLOT_LAYER
, 1), 3);
70 reg
.type
= BRW_REGISTER_TYPE_D
;
71 bld
.emit(FS_OPCODE_CINTERP
, retype(input
, BRW_REGISTER_TYPE_D
), reg
);
72 } else if (var
->data
.location
== VARYING_SLOT_VIEWPORT
) {
73 struct brw_reg reg
= suboffset(interp_reg(VARYING_SLOT_VIEWPORT
, 2), 3);
74 reg
.type
= BRW_REGISTER_TYPE_D
;
75 bld
.emit(FS_OPCODE_CINTERP
, retype(input
, BRW_REGISTER_TYPE_D
), reg
);
77 int location
= var
->data
.location
;
78 emit_general_interpolation(&input
, var
->name
, var
->type
,
79 (glsl_interp_qualifier
) var
->data
.interpolation
,
80 &location
, var
->data
.centroid
,
87 fs_visitor::nir_setup_single_output_varying(fs_reg
*reg
,
88 const glsl_type
*type
,
91 if (type
->is_array() || type
->is_matrix()) {
92 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
93 const unsigned length
= glsl_get_length(type
);
95 for (unsigned i
= 0; i
< length
; i
++) {
96 nir_setup_single_output_varying(reg
, elem_type
, location
);
98 } else if (type
->is_record()) {
99 for (unsigned i
= 0; i
< type
->length
; i
++) {
100 const struct glsl_type
*field_type
= type
->fields
.structure
[i
].type
;
101 nir_setup_single_output_varying(reg
, field_type
, location
);
104 assert(type
->is_scalar() || type
->is_vector());
105 unsigned num_elements
= type
->vector_elements
;
106 if (type
->is_double())
108 for (unsigned count
= 0; count
< num_elements
; count
+= 4) {
109 this->outputs
[*location
] = *reg
;
110 this->output_components
[*location
] = MIN2(4, num_elements
- count
);
111 *reg
= offset(*reg
, bld
, 4);
118 fs_visitor::nir_setup_outputs()
120 if (stage
== MESA_SHADER_TESS_CTRL
)
123 brw_wm_prog_key
*key
= (brw_wm_prog_key
*) this->key
;
125 nir_outputs
= bld
.vgrf(BRW_REGISTER_TYPE_F
, nir
->num_outputs
);
127 nir_foreach_variable(var
, &nir
->outputs
) {
128 fs_reg reg
= offset(nir_outputs
, bld
, var
->data
.driver_location
);
131 case MESA_SHADER_VERTEX
:
132 case MESA_SHADER_TESS_EVAL
:
133 case MESA_SHADER_GEOMETRY
: {
134 unsigned location
= var
->data
.location
;
135 nir_setup_single_output_varying(®
, var
->type
, &location
);
138 case MESA_SHADER_FRAGMENT
:
139 if (key
->force_dual_color_blend
&&
140 var
->data
.location
== FRAG_RESULT_DATA1
) {
141 this->dual_src_output
= reg
;
142 this->do_dual_src
= true;
143 } else if (var
->data
.index
> 0) {
144 assert(var
->data
.location
== FRAG_RESULT_DATA0
);
145 assert(var
->data
.index
== 1);
146 this->dual_src_output
= reg
;
147 this->do_dual_src
= true;
148 } else if (var
->data
.location
== FRAG_RESULT_COLOR
) {
149 /* Writing gl_FragColor outputs to all color regions. */
150 for (unsigned int i
= 0; i
< MAX2(key
->nr_color_regions
, 1); i
++) {
151 this->outputs
[i
] = reg
;
152 this->output_components
[i
] = 4;
154 } else if (var
->data
.location
== FRAG_RESULT_DEPTH
) {
155 this->frag_depth
= reg
;
156 } else if (var
->data
.location
== FRAG_RESULT_STENCIL
) {
157 this->frag_stencil
= reg
;
158 } else if (var
->data
.location
== FRAG_RESULT_SAMPLE_MASK
) {
159 this->sample_mask
= reg
;
161 int vector_elements
= var
->type
->without_array()->vector_elements
;
163 /* gl_FragData or a user-defined FS output */
164 assert(var
->data
.location
>= FRAG_RESULT_DATA0
&&
165 var
->data
.location
< FRAG_RESULT_DATA0
+BRW_MAX_DRAW_BUFFERS
);
167 /* General color output. */
168 for (unsigned int i
= 0; i
< MAX2(1, var
->type
->length
); i
++) {
169 int output
= var
->data
.location
- FRAG_RESULT_DATA0
+ i
;
170 this->outputs
[output
] = offset(reg
, bld
, vector_elements
* i
);
171 this->output_components
[output
] = vector_elements
;
176 unreachable("unhandled shader stage");
182 fs_visitor::nir_setup_uniforms()
184 if (dispatch_width
!= min_dispatch_width
)
187 uniforms
= nir
->num_uniforms
/ 4;
191 emit_system_values_block(nir_block
*block
, fs_visitor
*v
)
195 nir_foreach_instr(instr
, block
) {
196 if (instr
->type
!= nir_instr_type_intrinsic
)
199 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
200 switch (intrin
->intrinsic
) {
201 case nir_intrinsic_load_vertex_id
:
202 unreachable("should be lowered by lower_vertex_id().");
204 case nir_intrinsic_load_vertex_id_zero_base
:
205 assert(v
->stage
== MESA_SHADER_VERTEX
);
206 reg
= &v
->nir_system_values
[SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
];
207 if (reg
->file
== BAD_FILE
)
208 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
);
211 case nir_intrinsic_load_base_vertex
:
212 assert(v
->stage
== MESA_SHADER_VERTEX
);
213 reg
= &v
->nir_system_values
[SYSTEM_VALUE_BASE_VERTEX
];
214 if (reg
->file
== BAD_FILE
)
215 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_BASE_VERTEX
);
218 case nir_intrinsic_load_instance_id
:
219 assert(v
->stage
== MESA_SHADER_VERTEX
);
220 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INSTANCE_ID
];
221 if (reg
->file
== BAD_FILE
)
222 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_INSTANCE_ID
);
225 case nir_intrinsic_load_base_instance
:
226 assert(v
->stage
== MESA_SHADER_VERTEX
);
227 reg
= &v
->nir_system_values
[SYSTEM_VALUE_BASE_INSTANCE
];
228 if (reg
->file
== BAD_FILE
)
229 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_BASE_INSTANCE
);
232 case nir_intrinsic_load_draw_id
:
233 assert(v
->stage
== MESA_SHADER_VERTEX
);
234 reg
= &v
->nir_system_values
[SYSTEM_VALUE_DRAW_ID
];
235 if (reg
->file
== BAD_FILE
)
236 *reg
= *v
->emit_vs_system_value(SYSTEM_VALUE_DRAW_ID
);
239 case nir_intrinsic_load_invocation_id
:
240 if (v
->stage
== MESA_SHADER_TESS_CTRL
)
242 assert(v
->stage
== MESA_SHADER_GEOMETRY
);
243 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
244 if (reg
->file
== BAD_FILE
) {
245 const fs_builder abld
= v
->bld
.annotate("gl_InvocationID", NULL
);
246 fs_reg
g1(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
247 fs_reg iid
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
248 abld
.SHR(iid
, g1
, brw_imm_ud(27u));
253 case nir_intrinsic_load_sample_pos
:
254 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
255 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
256 if (reg
->file
== BAD_FILE
)
257 *reg
= *v
->emit_samplepos_setup();
260 case nir_intrinsic_load_sample_id
:
261 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
262 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
263 if (reg
->file
== BAD_FILE
)
264 *reg
= *v
->emit_sampleid_setup();
267 case nir_intrinsic_load_sample_mask_in
:
268 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
269 assert(v
->devinfo
->gen
>= 7);
270 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
271 if (reg
->file
== BAD_FILE
)
272 *reg
= *v
->emit_samplemaskin_setup();
275 case nir_intrinsic_load_local_invocation_id
:
276 assert(v
->stage
== MESA_SHADER_COMPUTE
);
277 reg
= &v
->nir_system_values
[SYSTEM_VALUE_LOCAL_INVOCATION_ID
];
278 if (reg
->file
== BAD_FILE
)
279 *reg
= *v
->emit_cs_local_invocation_id_setup();
282 case nir_intrinsic_load_work_group_id
:
283 assert(v
->stage
== MESA_SHADER_COMPUTE
);
284 reg
= &v
->nir_system_values
[SYSTEM_VALUE_WORK_GROUP_ID
];
285 if (reg
->file
== BAD_FILE
)
286 *reg
= *v
->emit_cs_work_group_id_setup();
289 case nir_intrinsic_load_helper_invocation
:
290 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
291 reg
= &v
->nir_system_values
[SYSTEM_VALUE_HELPER_INVOCATION
];
292 if (reg
->file
== BAD_FILE
) {
293 const fs_builder abld
=
294 v
->bld
.annotate("gl_HelperInvocation", NULL
);
296 /* On Gen6+ (gl_HelperInvocation is only exposed on Gen7+) the
297 * pixel mask is in g1.7 of the thread payload.
299 * We move the per-channel pixel enable bit to the low bit of each
300 * channel by shifting the byte containing the pixel mask by the
301 * vector immediate 0x76543210UV.
303 * The region of <1,8,0> reads only 1 byte (the pixel masks for
304 * subspans 0 and 1) in SIMD8 and an additional byte (the pixel
305 * masks for 2 and 3) in SIMD16.
307 fs_reg shifted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
309 stride(byte_offset(retype(brw_vec1_grf(1, 0),
310 BRW_REGISTER_TYPE_UB
), 28),
312 brw_imm_v(0x76543210));
314 /* A set bit in the pixel mask means the channel is enabled, but
315 * that is the opposite of gl_HelperInvocation so we need to invert
318 * The negate source-modifier bit of logical instructions on Gen8+
319 * performs 1's complement negation, so we can use that instead of
322 fs_reg inverted
= negate(shifted
);
323 if (v
->devinfo
->gen
< 8) {
324 inverted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
325 abld
.NOT(inverted
, shifted
);
328 /* We then resolve the 0/1 result to 0/~0 boolean values by ANDing
329 * with 1 and negating.
331 fs_reg anded
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
332 abld
.AND(anded
, inverted
, brw_imm_uw(1));
334 fs_reg dst
= abld
.vgrf(BRW_REGISTER_TYPE_D
, 1);
335 abld
.MOV(dst
, negate(retype(anded
, BRW_REGISTER_TYPE_D
)));
349 fs_visitor::nir_emit_system_values()
351 nir_system_values
= ralloc_array(mem_ctx
, fs_reg
, SYSTEM_VALUE_MAX
);
352 for (unsigned i
= 0; i
< SYSTEM_VALUE_MAX
; i
++) {
353 nir_system_values
[i
] = fs_reg();
356 nir_foreach_function(function
, nir
) {
357 assert(strcmp(function
->name
, "main") == 0);
358 assert(function
->impl
);
359 nir_foreach_block(block
, function
->impl
) {
360 emit_system_values_block(block
, this);
366 fs_visitor::nir_emit_impl(nir_function_impl
*impl
)
368 nir_locals
= ralloc_array(mem_ctx
, fs_reg
, impl
->reg_alloc
);
369 for (unsigned i
= 0; i
< impl
->reg_alloc
; i
++) {
370 nir_locals
[i
] = fs_reg();
373 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
374 unsigned array_elems
=
375 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
376 unsigned size
= array_elems
* reg
->num_components
;
377 const brw_reg_type reg_type
=
378 reg
->bit_size
== 32 ? BRW_REGISTER_TYPE_F
: BRW_REGISTER_TYPE_DF
;
379 nir_locals
[reg
->index
] = bld
.vgrf(reg_type
, size
);
382 nir_ssa_values
= reralloc(mem_ctx
, nir_ssa_values
, fs_reg
,
385 nir_emit_cf_list(&impl
->body
);
389 fs_visitor::nir_emit_cf_list(exec_list
*list
)
391 exec_list_validate(list
);
392 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
393 switch (node
->type
) {
395 nir_emit_if(nir_cf_node_as_if(node
));
398 case nir_cf_node_loop
:
399 nir_emit_loop(nir_cf_node_as_loop(node
));
402 case nir_cf_node_block
:
403 nir_emit_block(nir_cf_node_as_block(node
));
407 unreachable("Invalid CFG node block");
413 fs_visitor::nir_emit_if(nir_if
*if_stmt
)
415 /* first, put the condition into f0 */
416 fs_inst
*inst
= bld
.MOV(bld
.null_reg_d(),
417 retype(get_nir_src(if_stmt
->condition
),
418 BRW_REGISTER_TYPE_D
));
419 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
421 bld
.IF(BRW_PREDICATE_NORMAL
);
423 nir_emit_cf_list(&if_stmt
->then_list
);
425 /* note: if the else is empty, dead CF elimination will remove it */
426 bld
.emit(BRW_OPCODE_ELSE
);
428 nir_emit_cf_list(&if_stmt
->else_list
);
430 bld
.emit(BRW_OPCODE_ENDIF
);
434 fs_visitor::nir_emit_loop(nir_loop
*loop
)
436 bld
.emit(BRW_OPCODE_DO
);
438 nir_emit_cf_list(&loop
->body
);
440 bld
.emit(BRW_OPCODE_WHILE
);
444 fs_visitor::nir_emit_block(nir_block
*block
)
446 nir_foreach_instr(instr
, block
) {
447 nir_emit_instr(instr
);
452 fs_visitor::nir_emit_instr(nir_instr
*instr
)
454 const fs_builder abld
= bld
.annotate(NULL
, instr
);
456 switch (instr
->type
) {
457 case nir_instr_type_alu
:
458 nir_emit_alu(abld
, nir_instr_as_alu(instr
));
461 case nir_instr_type_intrinsic
:
463 case MESA_SHADER_VERTEX
:
464 nir_emit_vs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
466 case MESA_SHADER_TESS_CTRL
:
467 nir_emit_tcs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
469 case MESA_SHADER_TESS_EVAL
:
470 nir_emit_tes_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
472 case MESA_SHADER_GEOMETRY
:
473 nir_emit_gs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
475 case MESA_SHADER_FRAGMENT
:
476 nir_emit_fs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
478 case MESA_SHADER_COMPUTE
:
479 nir_emit_cs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
482 unreachable("unsupported shader stage");
486 case nir_instr_type_tex
:
487 nir_emit_texture(abld
, nir_instr_as_tex(instr
));
490 case nir_instr_type_load_const
:
491 nir_emit_load_const(abld
, nir_instr_as_load_const(instr
));
494 case nir_instr_type_ssa_undef
:
495 nir_emit_undef(abld
, nir_instr_as_ssa_undef(instr
));
498 case nir_instr_type_jump
:
499 nir_emit_jump(abld
, nir_instr_as_jump(instr
));
503 unreachable("unknown instruction type");
508 * Recognizes a parent instruction of nir_op_extract_* and changes the type to
512 fs_visitor::optimize_extract_to_float(nir_alu_instr
*instr
,
513 const fs_reg
&result
)
515 if (!instr
->src
[0].src
.is_ssa
||
516 !instr
->src
[0].src
.ssa
->parent_instr
)
519 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
522 nir_alu_instr
*src0
=
523 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
525 if (src0
->op
!= nir_op_extract_u8
&& src0
->op
!= nir_op_extract_u16
&&
526 src0
->op
!= nir_op_extract_i8
&& src0
->op
!= nir_op_extract_i16
)
529 nir_const_value
*element
= nir_src_as_const_value(src0
->src
[1].src
);
530 assert(element
!= NULL
);
532 /* Element type to extract.*/
533 const brw_reg_type type
= brw_int_type(
534 src0
->op
== nir_op_extract_u16
|| src0
->op
== nir_op_extract_i16
? 2 : 1,
535 src0
->op
== nir_op_extract_i16
|| src0
->op
== nir_op_extract_i8
);
537 fs_reg op0
= get_nir_src(src0
->src
[0].src
);
538 op0
.type
= brw_type_for_nir_type(
539 (nir_alu_type
)(nir_op_infos
[src0
->op
].input_types
[0] |
540 nir_src_bit_size(src0
->src
[0].src
)));
541 op0
= offset(op0
, bld
, src0
->src
[0].swizzle
[0]);
543 set_saturate(instr
->dest
.saturate
,
544 bld
.MOV(result
, subscript(op0
, type
, element
->u32
[0])));
549 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr
*instr
,
550 const fs_reg
&result
)
552 if (!instr
->src
[0].src
.is_ssa
||
553 instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
556 nir_intrinsic_instr
*src0
=
557 nir_instr_as_intrinsic(instr
->src
[0].src
.ssa
->parent_instr
);
559 if (src0
->intrinsic
!= nir_intrinsic_load_front_face
)
562 nir_const_value
*value1
= nir_src_as_const_value(instr
->src
[1].src
);
563 if (!value1
|| fabsf(value1
->f32
[0]) != 1.0f
)
566 nir_const_value
*value2
= nir_src_as_const_value(instr
->src
[2].src
);
567 if (!value2
|| fabsf(value2
->f32
[0]) != 1.0f
)
570 fs_reg tmp
= vgrf(glsl_type::int_type
);
572 if (devinfo
->gen
>= 6) {
573 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
574 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
576 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
578 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
579 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
581 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
583 * This negation looks like it's safe in practice, because bits 0:4 will
584 * surely be TRIANGLES
587 if (value1
->f32
[0] == -1.0f
) {
591 tmp
.type
= BRW_REGISTER_TYPE_W
;
592 tmp
.subreg_offset
= 2;
595 bld
.OR(tmp
, g0
, brw_imm_uw(0x3f80));
597 tmp
.type
= BRW_REGISTER_TYPE_D
;
598 tmp
.subreg_offset
= 0;
601 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
602 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
604 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
606 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
607 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
609 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
611 * This negation looks like it's safe in practice, because bits 0:4 will
612 * surely be TRIANGLES
615 if (value1
->f32
[0] == -1.0f
) {
619 bld
.OR(tmp
, g1_6
, brw_imm_d(0x3f800000));
621 bld
.AND(retype(result
, BRW_REGISTER_TYPE_D
), tmp
, brw_imm_d(0xbf800000));
627 fs_visitor::nir_emit_alu(const fs_builder
&bld
, nir_alu_instr
*instr
)
629 struct brw_wm_prog_key
*fs_key
= (struct brw_wm_prog_key
*) this->key
;
632 fs_reg result
= get_nir_dest(instr
->dest
.dest
);
633 result
.type
= brw_type_for_nir_type(
634 (nir_alu_type
)(nir_op_infos
[instr
->op
].output_type
|
635 nir_dest_bit_size(instr
->dest
.dest
)));
638 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
639 op
[i
] = get_nir_src(instr
->src
[i
].src
);
640 op
[i
].type
= brw_type_for_nir_type(
641 (nir_alu_type
)(nir_op_infos
[instr
->op
].input_types
[i
] |
642 nir_src_bit_size(instr
->src
[i
].src
)));
643 op
[i
].abs
= instr
->src
[i
].abs
;
644 op
[i
].negate
= instr
->src
[i
].negate
;
647 /* We get a bunch of mov's out of the from_ssa pass and they may still
648 * be vectorized. We'll handle them as a special-case. We'll also
649 * handle vecN here because it's basically the same thing.
657 fs_reg temp
= result
;
658 bool need_extra_copy
= false;
659 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
660 if (!instr
->src
[i
].src
.is_ssa
&&
661 instr
->dest
.dest
.reg
.reg
== instr
->src
[i
].src
.reg
.reg
) {
662 need_extra_copy
= true;
663 temp
= bld
.vgrf(result
.type
, 4);
668 for (unsigned i
= 0; i
< 4; i
++) {
669 if (!(instr
->dest
.write_mask
& (1 << i
)))
672 if (instr
->op
== nir_op_imov
|| instr
->op
== nir_op_fmov
) {
673 inst
= bld
.MOV(offset(temp
, bld
, i
),
674 offset(op
[0], bld
, instr
->src
[0].swizzle
[i
]));
676 inst
= bld
.MOV(offset(temp
, bld
, i
),
677 offset(op
[i
], bld
, instr
->src
[i
].swizzle
[0]));
679 inst
->saturate
= instr
->dest
.saturate
;
682 /* In this case the source and destination registers were the same,
683 * so we need to insert an extra set of moves in order to deal with
686 if (need_extra_copy
) {
687 for (unsigned i
= 0; i
< 4; i
++) {
688 if (!(instr
->dest
.write_mask
& (1 << i
)))
691 bld
.MOV(offset(result
, bld
, i
), offset(temp
, bld
, i
));
700 /* At this point, we have dealt with any instruction that operates on
701 * more than a single channel. Therefore, we can just adjust the source
702 * and destination registers for that channel and emit the instruction.
704 unsigned channel
= 0;
705 if (nir_op_infos
[instr
->op
].output_size
== 0) {
706 /* Since NIR is doing the scalarizing for us, we should only ever see
707 * vectorized operations with a single channel.
709 assert(_mesa_bitcount(instr
->dest
.write_mask
) == 1);
710 channel
= ffs(instr
->dest
.write_mask
) - 1;
712 result
= offset(result
, bld
, channel
);
715 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
716 assert(nir_op_infos
[instr
->op
].input_sizes
[i
] < 2);
717 op
[i
] = offset(op
[i
], bld
, instr
->src
[i
].swizzle
[channel
]);
723 if (optimize_extract_to_float(instr
, result
))
732 inst
= bld
.MOV(result
, op
[0]);
733 inst
->saturate
= instr
->dest
.saturate
;
738 bld
.MOV(result
, op
[0]);
742 if (type_sz(op
[0].type
) < 8) {
743 /* AND(val, 0x80000000) gives the sign bit.
745 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
748 bld
.CMP(bld
.null_reg_f(), op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
750 fs_reg result_int
= retype(result
, BRW_REGISTER_TYPE_UD
);
751 op
[0].type
= BRW_REGISTER_TYPE_UD
;
752 result
.type
= BRW_REGISTER_TYPE_UD
;
753 bld
.AND(result_int
, op
[0], brw_imm_ud(0x80000000u
));
755 inst
= bld
.OR(result_int
, result_int
, brw_imm_ud(0x3f800000u
));
756 inst
->predicate
= BRW_PREDICATE_NORMAL
;
757 if (instr
->dest
.saturate
) {
758 inst
= bld
.MOV(result
, result
);
759 inst
->saturate
= true;
762 /* For doubles we do the same but we need to consider:
764 * - 2-src instructions can't operate with 64-bit immediates
765 * - The sign is encoded in the high 32-bit of each DF
766 * - CMP with DF requires special handling in SIMD16
767 * - We need to produce a DF result.
770 /* 2-src instructions can't have 64-bit immediates, so put 0.0 in
771 * a register and compare with that.
773 fs_reg tmp
= vgrf(glsl_type::double_type
);
774 bld
.MOV(tmp
, brw_imm_df(0.0));
776 /* A direct DF CMP using the flag register (null dst) won't work in
777 * SIMD16 because the CMP will be split in two by lower_simd_width,
778 * resulting in two CMP instructions with the same dst (NULL),
779 * leading to dead code elimination of the first one. In SIMD8,
780 * however, there is no need to split the CMP and we can save some
783 fs_reg dst_tmp
= vgrf(glsl_type::double_type
);
784 bld
.CMP(dst_tmp
, op
[0], tmp
, BRW_CONDITIONAL_NZ
);
786 /* In SIMD16 we want to avoid using a NULL dst register with DF CMP,
787 * so we store the result of the comparison in a vgrf instead and
788 * then we generate a UD comparison from that that won't have to
789 * be split by lower_simd_width. This is what NIR does to handle
790 * double comparisons in the general case.
792 if (bld
.dispatch_width() == 16 ) {
793 fs_reg dst_tmp_ud
= retype(dst_tmp
, BRW_REGISTER_TYPE_UD
);
794 bld
.MOV(dst_tmp_ud
, subscript(dst_tmp
, BRW_REGISTER_TYPE_UD
, 0));
795 bld
.CMP(bld
.null_reg_ud(),
796 dst_tmp_ud
, brw_imm_ud(0), BRW_CONDITIONAL_NZ
);
799 /* Get the high 32-bit of each double component where the sign is */
800 fs_reg result_int
= retype(result
, BRW_REGISTER_TYPE_UD
);
801 bld
.MOV(result_int
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1));
803 /* Get the sign bit */
804 bld
.AND(result_int
, result_int
, brw_imm_ud(0x80000000u
));
806 /* Add 1.0 to the sign, predicated to skip the case of op[0] == 0.0 */
807 inst
= bld
.OR(result_int
, result_int
, brw_imm_ud(0x3f800000u
));
808 inst
->predicate
= BRW_PREDICATE_NORMAL
;
810 /* Convert from 32-bit float to 64-bit double */
811 result
.type
= BRW_REGISTER_TYPE_DF
;
812 inst
= bld
.MOV(result
, retype(result_int
, BRW_REGISTER_TYPE_F
));
814 if (instr
->dest
.saturate
) {
815 inst
= bld
.MOV(result
, result
);
816 inst
->saturate
= true;
823 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
824 * -> non-negative val generates 0x00000000.
825 * Predicated OR sets 1 if val is positive.
827 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
828 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_G
);
829 bld
.ASR(result
, op
[0], brw_imm_d(31));
830 inst
= bld
.OR(result
, result
, brw_imm_d(1));
831 inst
->predicate
= BRW_PREDICATE_NORMAL
;
835 inst
= bld
.emit(SHADER_OPCODE_RCP
, result
, op
[0]);
836 inst
->saturate
= instr
->dest
.saturate
;
840 inst
= bld
.emit(SHADER_OPCODE_EXP2
, result
, op
[0]);
841 inst
->saturate
= instr
->dest
.saturate
;
845 inst
= bld
.emit(SHADER_OPCODE_LOG2
, result
, op
[0]);
846 inst
->saturate
= instr
->dest
.saturate
;
850 inst
= bld
.emit(SHADER_OPCODE_SIN
, result
, op
[0]);
851 inst
->saturate
= instr
->dest
.saturate
;
855 inst
= bld
.emit(SHADER_OPCODE_COS
, result
, op
[0]);
856 inst
->saturate
= instr
->dest
.saturate
;
860 if (fs_key
->high_quality_derivatives
) {
861 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
863 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
865 inst
->saturate
= instr
->dest
.saturate
;
867 case nir_op_fddx_fine
:
868 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
869 inst
->saturate
= instr
->dest
.saturate
;
871 case nir_op_fddx_coarse
:
872 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
873 inst
->saturate
= instr
->dest
.saturate
;
876 if (fs_key
->high_quality_derivatives
) {
877 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
879 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
881 inst
->saturate
= instr
->dest
.saturate
;
883 case nir_op_fddy_fine
:
884 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
885 inst
->saturate
= instr
->dest
.saturate
;
887 case nir_op_fddy_coarse
:
888 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
889 inst
->saturate
= instr
->dest
.saturate
;
893 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
895 inst
= bld
.ADD(result
, op
[0], op
[1]);
896 inst
->saturate
= instr
->dest
.saturate
;
900 inst
= bld
.MUL(result
, op
[0], op
[1]);
901 inst
->saturate
= instr
->dest
.saturate
;
905 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
906 bld
.MUL(result
, op
[0], op
[1]);
909 case nir_op_imul_high
:
910 case nir_op_umul_high
:
911 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
912 bld
.emit(SHADER_OPCODE_MULH
, result
, op
[0], op
[1]);
917 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
918 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, result
, op
[0], op
[1]);
921 case nir_op_uadd_carry
:
922 unreachable("Should have been lowered by carry_to_arith().");
924 case nir_op_usub_borrow
:
925 unreachable("Should have been lowered by borrow_to_arith().");
929 /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
930 * appears that our hardware just does the right thing for signed
933 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
934 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
938 /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
939 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
941 /* Math instructions don't support conditional mod */
942 inst
= bld
.MOV(bld
.null_reg_d(), result
);
943 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
945 /* Now, we need to determine if signs of the sources are different.
946 * When we XOR the sources, the top bit is 0 if they are the same and 1
947 * if they are different. We can then use a conditional modifier to
948 * turn that into a predicate. This leads us to an XOR.l instruction.
950 * Technically, according to the PRM, you're not allowed to use .l on a
951 * XOR instruction. However, emperical experiments and Curro's reading
952 * of the simulator source both indicate that it's safe.
954 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
955 inst
= bld
.XOR(tmp
, op
[0], op
[1]);
956 inst
->predicate
= BRW_PREDICATE_NORMAL
;
957 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
959 /* If the result of the initial remainder operation is non-zero and the
960 * two sources have different signs, add in a copy of op[1] to get the
961 * final integer modulus value.
963 inst
= bld
.ADD(result
, result
, op
[1]);
964 inst
->predicate
= BRW_PREDICATE_NORMAL
;
972 fs_reg dest
= result
;
973 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
974 dest
= bld
.vgrf(BRW_REGISTER_TYPE_DF
, 1);
976 brw_conditional_mod cond
;
979 cond
= BRW_CONDITIONAL_L
;
982 cond
= BRW_CONDITIONAL_GE
;
985 cond
= BRW_CONDITIONAL_Z
;
988 cond
= BRW_CONDITIONAL_NZ
;
991 unreachable("bad opcode");
993 bld
.CMP(dest
, op
[0], op
[1], cond
);
994 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
995 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1002 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1003 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1008 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1009 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1013 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1014 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_Z
);
1018 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1019 bld
.CMP(result
, op
[0], op
[1], BRW_CONDITIONAL_NZ
);
1023 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1024 if (devinfo
->gen
>= 8) {
1025 op
[0] = resolve_source_modifiers(op
[0]);
1027 bld
.NOT(result
, op
[0]);
1030 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1031 if (devinfo
->gen
>= 8) {
1032 op
[0] = resolve_source_modifiers(op
[0]);
1033 op
[1] = resolve_source_modifiers(op
[1]);
1035 bld
.XOR(result
, op
[0], op
[1]);
1038 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1039 if (devinfo
->gen
>= 8) {
1040 op
[0] = resolve_source_modifiers(op
[0]);
1041 op
[1] = resolve_source_modifiers(op
[1]);
1043 bld
.OR(result
, op
[0], op
[1]);
1046 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1047 if (devinfo
->gen
>= 8) {
1048 op
[0] = resolve_source_modifiers(op
[0]);
1049 op
[1] = resolve_source_modifiers(op
[1]);
1051 bld
.AND(result
, op
[0], op
[1]);
1057 case nir_op_ball_fequal2
:
1058 case nir_op_ball_iequal2
:
1059 case nir_op_ball_fequal3
:
1060 case nir_op_ball_iequal3
:
1061 case nir_op_ball_fequal4
:
1062 case nir_op_ball_iequal4
:
1063 case nir_op_bany_fnequal2
:
1064 case nir_op_bany_inequal2
:
1065 case nir_op_bany_fnequal3
:
1066 case nir_op_bany_inequal3
:
1067 case nir_op_bany_fnequal4
:
1068 case nir_op_bany_inequal4
:
1069 unreachable("Lowered by nir_lower_alu_reductions");
1071 case nir_op_fnoise1_1
:
1072 case nir_op_fnoise1_2
:
1073 case nir_op_fnoise1_3
:
1074 case nir_op_fnoise1_4
:
1075 case nir_op_fnoise2_1
:
1076 case nir_op_fnoise2_2
:
1077 case nir_op_fnoise2_3
:
1078 case nir_op_fnoise2_4
:
1079 case nir_op_fnoise3_1
:
1080 case nir_op_fnoise3_2
:
1081 case nir_op_fnoise3_3
:
1082 case nir_op_fnoise3_4
:
1083 case nir_op_fnoise4_1
:
1084 case nir_op_fnoise4_2
:
1085 case nir_op_fnoise4_3
:
1086 case nir_op_fnoise4_4
:
1087 unreachable("not reached: should be handled by lower_noise");
1090 unreachable("not reached: should be handled by ldexp_to_arith()");
1093 inst
= bld
.emit(SHADER_OPCODE_SQRT
, result
, op
[0]);
1094 inst
->saturate
= instr
->dest
.saturate
;
1098 inst
= bld
.emit(SHADER_OPCODE_RSQ
, result
, op
[0]);
1099 inst
->saturate
= instr
->dest
.saturate
;
1104 bld
.MOV(result
, negate(op
[0]));
1108 bld
.CMP(result
, op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
1111 /* two-argument instructions can't take 64-bit immediates */
1112 fs_reg zero
= vgrf(glsl_type::double_type
);
1113 bld
.MOV(zero
, brw_imm_df(0.0));
1114 /* A SIMD16 execution needs to be split in two instructions, so use
1115 * a vgrf instead of the flag register as dst so instruction splitting
1118 fs_reg tmp
= vgrf(glsl_type::double_type
);
1119 bld
.CMP(tmp
, op
[0], zero
, BRW_CONDITIONAL_NZ
);
1120 bld
.MOV(result
, subscript(tmp
, BRW_REGISTER_TYPE_UD
, 0));
1124 bld
.CMP(result
, op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1128 inst
= bld
.RNDZ(result
, op
[0]);
1129 inst
->saturate
= instr
->dest
.saturate
;
1132 case nir_op_fceil
: {
1133 op
[0].negate
= !op
[0].negate
;
1134 fs_reg temp
= vgrf(glsl_type::float_type
);
1135 bld
.RNDD(temp
, op
[0]);
1137 inst
= bld
.MOV(result
, temp
);
1138 inst
->saturate
= instr
->dest
.saturate
;
1142 inst
= bld
.RNDD(result
, op
[0]);
1143 inst
->saturate
= instr
->dest
.saturate
;
1146 inst
= bld
.FRC(result
, op
[0]);
1147 inst
->saturate
= instr
->dest
.saturate
;
1149 case nir_op_fround_even
:
1150 inst
= bld
.RNDE(result
, op
[0]);
1151 inst
->saturate
= instr
->dest
.saturate
;
1154 case nir_op_fquantize2f16
: {
1155 fs_reg tmp16
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1156 fs_reg tmp32
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1157 fs_reg zero
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1159 /* The destination stride must be at least as big as the source stride. */
1160 tmp16
.type
= BRW_REGISTER_TYPE_W
;
1163 /* Check for denormal */
1164 fs_reg abs_src0
= op
[0];
1165 abs_src0
.abs
= true;
1166 bld
.CMP(bld
.null_reg_f(), abs_src0
, brw_imm_f(ldexpf(1.0, -14)),
1168 /* Get the appropriately signed zero */
1169 bld
.AND(retype(zero
, BRW_REGISTER_TYPE_UD
),
1170 retype(op
[0], BRW_REGISTER_TYPE_UD
),
1171 brw_imm_ud(0x80000000));
1172 /* Do the actual F32 -> F16 -> F32 conversion */
1173 bld
.emit(BRW_OPCODE_F32TO16
, tmp16
, op
[0]);
1174 bld
.emit(BRW_OPCODE_F16TO32
, tmp32
, tmp16
);
1175 /* Select that or zero based on normal status */
1176 inst
= bld
.SEL(result
, zero
, tmp32
);
1177 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1178 inst
->saturate
= instr
->dest
.saturate
;
1184 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1186 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1187 inst
->saturate
= instr
->dest
.saturate
;
1192 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1194 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1195 inst
->saturate
= instr
->dest
.saturate
;
1198 case nir_op_pack_snorm_2x16
:
1199 case nir_op_pack_snorm_4x8
:
1200 case nir_op_pack_unorm_2x16
:
1201 case nir_op_pack_unorm_4x8
:
1202 case nir_op_unpack_snorm_2x16
:
1203 case nir_op_unpack_snorm_4x8
:
1204 case nir_op_unpack_unorm_2x16
:
1205 case nir_op_unpack_unorm_4x8
:
1206 case nir_op_unpack_half_2x16
:
1207 case nir_op_pack_half_2x16
:
1208 unreachable("not reached: should be handled by lower_packing_builtins");
1210 case nir_op_unpack_half_2x16_split_x
:
1211 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
, result
, op
[0]);
1212 inst
->saturate
= instr
->dest
.saturate
;
1214 case nir_op_unpack_half_2x16_split_y
:
1215 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
, result
, op
[0]);
1216 inst
->saturate
= instr
->dest
.saturate
;
1219 case nir_op_pack_double_2x32_split
:
1220 /* Optimize the common case where we are re-packing a double with
1221 * the result of a previous double unpack. In this case we can take the
1222 * 32-bit value to use in the re-pack from the original double and bypass
1223 * the unpack operation.
1225 for (int i
= 0; i
< 2; i
++) {
1226 if (instr
->src
[i
].src
.is_ssa
)
1229 const nir_instr
*parent_instr
= instr
->src
[i
].src
.ssa
->parent_instr
;
1230 if (parent_instr
->type
== nir_instr_type_alu
)
1233 const nir_alu_instr
*alu_parent
= nir_instr_as_alu(parent_instr
);
1234 if (alu_parent
->op
== nir_op_unpack_double_2x32_split_x
||
1235 alu_parent
->op
== nir_op_unpack_double_2x32_split_y
)
1238 if (!alu_parent
->src
[0].src
.is_ssa
)
1241 op
[i
] = get_nir_src(alu_parent
->src
[0].src
);
1242 op
[i
] = offset(retype(op
[i
], BRW_REGISTER_TYPE_DF
), bld
,
1243 alu_parent
->src
[0].swizzle
[channel
]);
1244 if (alu_parent
->op
== nir_op_unpack_double_2x32_split_y
)
1245 op
[i
] = subscript(op
[i
], BRW_REGISTER_TYPE_UD
, 1);
1247 op
[i
] = subscript(op
[i
], BRW_REGISTER_TYPE_UD
, 0);
1249 bld
.emit(FS_OPCODE_PACK
, result
, op
[0], op
[1]);
1252 case nir_op_unpack_double_2x32_split_x
:
1253 case nir_op_unpack_double_2x32_split_y
: {
1254 /* Optimize the common case where we are unpacking from a double we have
1255 * previously packed. In this case we can just bypass the pack operation
1256 * and source directly from its arguments.
1258 unsigned index
= (instr
->op
== nir_op_unpack_double_2x32_split_x
) ? 0 : 1;
1259 if (instr
->src
[0].src
.is_ssa
) {
1260 nir_instr
*parent_instr
= instr
->src
[0].src
.ssa
->parent_instr
;
1261 if (parent_instr
->type
== nir_instr_type_alu
) {
1262 nir_alu_instr
*alu_parent
= nir_instr_as_alu(parent_instr
);
1263 if (alu_parent
->op
== nir_op_pack_double_2x32_split
&&
1264 alu_parent
->src
[index
].src
.is_ssa
) {
1265 op
[0] = retype(get_nir_src(alu_parent
->src
[index
].src
),
1266 BRW_REGISTER_TYPE_UD
);
1268 offset(op
[0], bld
, alu_parent
->src
[index
].swizzle
[channel
]);
1269 bld
.MOV(result
, op
[0]);
1275 if (instr
->op
== nir_op_unpack_double_2x32_split_x
)
1276 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 0));
1278 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1));
1283 inst
= bld
.emit(SHADER_OPCODE_POW
, result
, op
[0], op
[1]);
1284 inst
->saturate
= instr
->dest
.saturate
;
1287 case nir_op_bitfield_reverse
:
1288 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1289 bld
.BFREV(result
, op
[0]);
1292 case nir_op_bit_count
:
1293 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1294 bld
.CBIT(result
, op
[0]);
1297 case nir_op_ufind_msb
:
1298 case nir_op_ifind_msb
: {
1299 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1300 bld
.FBH(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0]);
1302 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1303 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1304 * subtract the result from 31 to convert the MSB count into an LSB count.
1306 bld
.CMP(bld
.null_reg_d(), result
, brw_imm_d(-1), BRW_CONDITIONAL_NZ
);
1308 inst
= bld
.ADD(result
, result
, brw_imm_d(31));
1309 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1310 inst
->src
[0].negate
= true;
1314 case nir_op_find_lsb
:
1315 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1316 bld
.FBL(result
, op
[0]);
1319 case nir_op_ubitfield_extract
:
1320 case nir_op_ibitfield_extract
:
1321 unreachable("should have been lowered");
1324 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1325 bld
.BFE(result
, op
[2], op
[1], op
[0]);
1328 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1329 bld
.BFI1(result
, op
[0], op
[1]);
1332 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1333 bld
.BFI2(result
, op
[0], op
[1], op
[2]);
1336 case nir_op_bitfield_insert
:
1337 unreachable("not reached: should have been lowered");
1340 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1341 bld
.SHL(result
, op
[0], op
[1]);
1344 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1345 bld
.ASR(result
, op
[0], op
[1]);
1348 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1349 bld
.SHR(result
, op
[0], op
[1]);
1352 case nir_op_pack_half_2x16_split
:
1353 bld
.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, result
, op
[0], op
[1]);
1357 inst
= bld
.MAD(result
, op
[2], op
[1], op
[0]);
1358 inst
->saturate
= instr
->dest
.saturate
;
1362 inst
= bld
.LRP(result
, op
[0], op
[1], op
[2]);
1363 inst
->saturate
= instr
->dest
.saturate
;
1367 if (optimize_frontfacing_ternary(instr
, result
))
1370 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1371 inst
= bld
.SEL(result
, op
[1], op
[2]);
1372 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1375 case nir_op_extract_u8
:
1376 case nir_op_extract_i8
: {
1377 const brw_reg_type type
= brw_int_type(1, instr
->op
== nir_op_extract_i8
);
1378 nir_const_value
*byte
= nir_src_as_const_value(instr
->src
[1].src
);
1379 assert(byte
!= NULL
);
1380 bld
.MOV(result
, subscript(op
[0], type
, byte
->u32
[0]));
1384 case nir_op_extract_u16
:
1385 case nir_op_extract_i16
: {
1386 const brw_reg_type type
= brw_int_type(2, instr
->op
== nir_op_extract_i16
);
1387 nir_const_value
*word
= nir_src_as_const_value(instr
->src
[1].src
);
1388 assert(word
!= NULL
);
1389 bld
.MOV(result
, subscript(op
[0], type
, word
->u32
[0]));
1394 unreachable("unhandled instruction");
1397 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1398 * to sign extend the low bit to 0/~0
1400 if (devinfo
->gen
<= 5 &&
1401 (instr
->instr
.pass_flags
& BRW_NIR_BOOLEAN_MASK
) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE
) {
1402 fs_reg masked
= vgrf(glsl_type::int_type
);
1403 bld
.AND(masked
, result
, brw_imm_d(1));
1404 masked
.negate
= true;
1405 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), masked
);
1410 fs_visitor::nir_emit_load_const(const fs_builder
&bld
,
1411 nir_load_const_instr
*instr
)
1413 const brw_reg_type reg_type
=
1414 instr
->def
.bit_size
== 32 ? BRW_REGISTER_TYPE_D
: BRW_REGISTER_TYPE_DF
;
1415 fs_reg reg
= bld
.vgrf(reg_type
, instr
->def
.num_components
);
1417 switch (instr
->def
.bit_size
) {
1419 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1420 bld
.MOV(offset(reg
, bld
, i
), brw_imm_d(instr
->value
.i32
[i
]));
1424 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1425 bld
.MOV(offset(reg
, bld
, i
), brw_imm_df(instr
->value
.f64
[i
]));
1429 unreachable("Invalid bit size");
1432 nir_ssa_values
[instr
->def
.index
] = reg
;
1436 fs_visitor::nir_emit_undef(const fs_builder
&bld
, nir_ssa_undef_instr
*instr
)
1438 const brw_reg_type reg_type
=
1439 instr
->def
.bit_size
== 32 ? BRW_REGISTER_TYPE_D
: BRW_REGISTER_TYPE_DF
;
1440 nir_ssa_values
[instr
->def
.index
] =
1441 bld
.vgrf(reg_type
, instr
->def
.num_components
);
1445 fs_visitor::get_nir_src(const nir_src
&src
)
1449 reg
= nir_ssa_values
[src
.ssa
->index
];
1451 /* We don't handle indirects on locals */
1452 assert(src
.reg
.indirect
== NULL
);
1453 reg
= offset(nir_locals
[src
.reg
.reg
->index
], bld
,
1454 src
.reg
.base_offset
* src
.reg
.reg
->num_components
);
1457 /* to avoid floating-point denorm flushing problems, set the type by
1458 * default to D - instructions that need floating point semantics will set
1459 * this to F if they need to
1461 return retype(reg
, BRW_REGISTER_TYPE_D
);
1465 * Return an IMM for constants; otherwise call get_nir_src() as normal.
1468 fs_visitor::get_nir_src_imm(const nir_src
&src
)
1470 nir_const_value
*val
= nir_src_as_const_value(src
);
1471 return val
? fs_reg(brw_imm_d(val
->i32
[0])) : get_nir_src(src
);
1475 fs_visitor::get_nir_dest(const nir_dest
&dest
)
1478 const brw_reg_type reg_type
=
1479 dest
.ssa
.bit_size
== 32 ? BRW_REGISTER_TYPE_F
: BRW_REGISTER_TYPE_DF
;
1480 nir_ssa_values
[dest
.ssa
.index
] =
1481 bld
.vgrf(reg_type
, dest
.ssa
.num_components
);
1482 return nir_ssa_values
[dest
.ssa
.index
];
1484 /* We don't handle indirects on locals */
1485 assert(dest
.reg
.indirect
== NULL
);
1486 return offset(nir_locals
[dest
.reg
.reg
->index
], bld
,
1487 dest
.reg
.base_offset
* dest
.reg
.reg
->num_components
);
1492 fs_visitor::get_nir_image_deref(const nir_deref_var
*deref
)
1494 fs_reg
image(UNIFORM
, deref
->var
->data
.driver_location
/ 4,
1495 BRW_REGISTER_TYPE_UD
);
1497 unsigned indirect_max
= 0;
1499 for (const nir_deref
*tail
= &deref
->deref
; tail
->child
;
1500 tail
= tail
->child
) {
1501 const nir_deref_array
*deref_array
= nir_deref_as_array(tail
->child
);
1502 assert(tail
->child
->deref_type
== nir_deref_type_array
);
1503 const unsigned size
= glsl_get_length(tail
->type
);
1504 const unsigned element_size
= type_size_scalar(deref_array
->deref
.type
);
1505 const unsigned base
= MIN2(deref_array
->base_offset
, size
- 1);
1506 image
= offset(image
, bld
, base
* element_size
);
1508 if (deref_array
->deref_array_type
== nir_deref_array_type_indirect
) {
1509 fs_reg tmp
= vgrf(glsl_type::uint_type
);
1511 /* Accessing an invalid surface index with the dataport can result
1512 * in a hang. According to the spec "if the index used to
1513 * select an individual element is negative or greater than or
1514 * equal to the size of the array, the results of the operation
1515 * are undefined but may not lead to termination" -- which is one
1516 * of the possible outcomes of the hang. Clamp the index to
1517 * prevent access outside of the array bounds.
1519 bld
.emit_minmax(tmp
, retype(get_nir_src(deref_array
->indirect
),
1520 BRW_REGISTER_TYPE_UD
),
1521 brw_imm_ud(size
- base
- 1), BRW_CONDITIONAL_L
);
1523 indirect_max
+= element_size
* (tail
->type
->length
- 1);
1525 bld
.MUL(tmp
, tmp
, brw_imm_ud(element_size
* 4));
1526 if (indirect
.file
== BAD_FILE
) {
1529 bld
.ADD(indirect
, indirect
, tmp
);
1534 if (indirect
.file
== BAD_FILE
) {
1537 /* Emit a pile of MOVs to load the uniform into a temporary. The
1538 * dead-code elimination pass will get rid of what we don't use.
1540 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, BRW_IMAGE_PARAM_SIZE
);
1541 for (unsigned j
= 0; j
< BRW_IMAGE_PARAM_SIZE
; j
++) {
1542 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
1543 offset(tmp
, bld
, j
), offset(image
, bld
, j
),
1544 indirect
, brw_imm_ud((indirect_max
+ 1) * 4));
1551 fs_visitor::emit_percomp(const fs_builder
&bld
, const fs_inst
&inst
,
1554 for (unsigned i
= 0; i
< 4; i
++) {
1555 if (!((wr_mask
>> i
) & 1))
1558 fs_inst
*new_inst
= new(mem_ctx
) fs_inst(inst
);
1559 new_inst
->dst
= offset(new_inst
->dst
, bld
, i
);
1560 for (unsigned j
= 0; j
< new_inst
->sources
; j
++)
1561 if (new_inst
->src
[j
].file
== VGRF
)
1562 new_inst
->src
[j
] = offset(new_inst
->src
[j
], bld
, i
);
1569 * Get the matching channel register datatype for an image intrinsic of the
1570 * specified GLSL image type.
1573 get_image_base_type(const glsl_type
*type
)
1575 switch ((glsl_base_type
)type
->sampled_type
) {
1576 case GLSL_TYPE_UINT
:
1577 return BRW_REGISTER_TYPE_UD
;
1579 return BRW_REGISTER_TYPE_D
;
1580 case GLSL_TYPE_FLOAT
:
1581 return BRW_REGISTER_TYPE_F
;
1583 unreachable("Not reached.");
1588 * Get the appropriate atomic op for an image atomic intrinsic.
1591 get_image_atomic_op(nir_intrinsic_op op
, const glsl_type
*type
)
1594 case nir_intrinsic_image_atomic_add
:
1596 case nir_intrinsic_image_atomic_min
:
1597 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1598 BRW_AOP_IMIN
: BRW_AOP_UMIN
);
1599 case nir_intrinsic_image_atomic_max
:
1600 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1601 BRW_AOP_IMAX
: BRW_AOP_UMAX
);
1602 case nir_intrinsic_image_atomic_and
:
1604 case nir_intrinsic_image_atomic_or
:
1606 case nir_intrinsic_image_atomic_xor
:
1608 case nir_intrinsic_image_atomic_exchange
:
1610 case nir_intrinsic_image_atomic_comp_swap
:
1611 return BRW_AOP_CMPWR
;
1613 unreachable("Not reachable.");
1618 emit_pixel_interpolater_send(const fs_builder
&bld
,
1623 glsl_interp_qualifier interpolation
)
1629 if (src
.file
== BAD_FILE
) {
1631 payload
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 1);
1635 mlen
= 2 * bld
.dispatch_width() / 8;
1638 inst
= bld
.emit(opcode
, dst
, payload
, desc
);
1640 /* 2 floats per slot returned */
1641 inst
->regs_written
= 2 * bld
.dispatch_width() / 8;
1642 inst
->pi_noperspective
= interpolation
== INTERP_QUALIFIER_NOPERSPECTIVE
;
1648 * Computes 1 << x, given a D/UD register containing some value x.
1651 intexp2(const fs_builder
&bld
, const fs_reg
&x
)
1653 assert(x
.type
== BRW_REGISTER_TYPE_UD
|| x
.type
== BRW_REGISTER_TYPE_D
);
1655 fs_reg result
= bld
.vgrf(x
.type
, 1);
1656 fs_reg one
= bld
.vgrf(x
.type
, 1);
1658 bld
.MOV(one
, retype(brw_imm_d(1), one
.type
));
1659 bld
.SHL(result
, one
, x
);
1664 fs_visitor::emit_gs_end_primitive(const nir_src
&vertex_count_nir_src
)
1666 assert(stage
== MESA_SHADER_GEOMETRY
);
1668 struct brw_gs_prog_data
*gs_prog_data
=
1669 (struct brw_gs_prog_data
*) prog_data
;
1671 if (gs_compile
->control_data_header_size_bits
== 0)
1674 /* We can only do EndPrimitive() functionality when the control data
1675 * consists of cut bits. Fortunately, the only time it isn't is when the
1676 * output type is points, in which case EndPrimitive() is a no-op.
1678 if (gs_prog_data
->control_data_format
!=
1679 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
) {
1683 /* Cut bits use one bit per vertex. */
1684 assert(gs_compile
->control_data_bits_per_vertex
== 1);
1686 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1687 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1689 /* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
1690 * vertex n, 0 otherwise. So all we need to do here is mark bit
1691 * (vertex_count - 1) % 32 in the cut_bits register to indicate that
1692 * EndPrimitive() was called after emitting vertex (vertex_count - 1);
1693 * vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
1695 * Note that if EndPrimitive() is called before emitting any vertices, this
1696 * will cause us to set bit 31 of the control_data_bits register to 1.
1697 * That's fine because:
1699 * - If max_vertices < 32, then vertex number 31 (zero-based) will never be
1700 * output, so the hardware will ignore cut bit 31.
1702 * - If max_vertices == 32, then vertex number 31 is guaranteed to be the
1703 * last vertex, so setting cut bit 31 has no effect (since the primitive
1704 * is automatically ended when the GS terminates).
1706 * - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
1707 * control_data_bits register to 0 when the first vertex is emitted.
1710 const fs_builder abld
= bld
.annotate("end primitive");
1712 /* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
1713 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1714 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1715 fs_reg mask
= intexp2(abld
, prev_count
);
1716 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1717 * attention to the lower 5 bits of its second source argument, so on this
1718 * architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
1719 * ((vertex_count - 1) % 32).
1721 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1725 fs_visitor::emit_gs_control_data_bits(const fs_reg
&vertex_count
)
1727 assert(stage
== MESA_SHADER_GEOMETRY
);
1728 assert(gs_compile
->control_data_bits_per_vertex
!= 0);
1730 struct brw_gs_prog_data
*gs_prog_data
=
1731 (struct brw_gs_prog_data
*) prog_data
;
1733 const fs_builder abld
= bld
.annotate("emit control data bits");
1734 const fs_builder fwa_bld
= bld
.exec_all();
1736 /* We use a single UD register to accumulate control data bits (32 bits
1737 * for each of the SIMD8 channels). So we need to write a DWord (32 bits)
1740 * Unfortunately, the URB_WRITE_SIMD8 message uses 128-bit (OWord) offsets.
1741 * We have select a 128-bit group via the Global and Per-Slot Offsets, then
1742 * use the Channel Mask phase to enable/disable which DWord within that
1743 * group to write. (Remember, different SIMD8 channels may have emitted
1744 * different numbers of vertices, so we may need per-slot offsets.)
1746 * Channel masking presents an annoying problem: we may have to replicate
1747 * the data up to 4 times:
1749 * Msg = Handles, Per-Slot Offsets, Channel Masks, Data, Data, Data, Data.
1751 * To avoid penalizing shaders that emit a small number of vertices, we
1752 * can avoid these sometimes: if the size of the control data header is
1753 * <= 128 bits, then there is only 1 OWord. All SIMD8 channels will land
1754 * land in the same 128-bit group, so we can skip per-slot offsets.
1756 * Similarly, if the control data header is <= 32 bits, there is only one
1757 * DWord, so we can skip channel masks.
1759 enum opcode opcode
= SHADER_OPCODE_URB_WRITE_SIMD8
;
1761 fs_reg channel_mask
, per_slot_offset
;
1763 if (gs_compile
->control_data_header_size_bits
> 32) {
1764 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
1765 channel_mask
= vgrf(glsl_type::uint_type
);
1768 if (gs_compile
->control_data_header_size_bits
> 128) {
1769 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
;
1770 per_slot_offset
= vgrf(glsl_type::uint_type
);
1773 /* Figure out which DWord we're trying to write to using the formula:
1775 * dword_index = (vertex_count - 1) * bits_per_vertex / 32
1777 * Since bits_per_vertex is a power of two, and is known at compile
1778 * time, this can be optimized to:
1780 * dword_index = (vertex_count - 1) >> (6 - log2(bits_per_vertex))
1782 if (opcode
!= SHADER_OPCODE_URB_WRITE_SIMD8
) {
1783 fs_reg dword_index
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1784 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1785 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1786 unsigned log2_bits_per_vertex
=
1787 _mesa_fls(gs_compile
->control_data_bits_per_vertex
);
1788 abld
.SHR(dword_index
, prev_count
, brw_imm_ud(6u - log2_bits_per_vertex
));
1790 if (per_slot_offset
.file
!= BAD_FILE
) {
1791 /* Set the per-slot offset to dword_index / 4, so that we'll write to
1792 * the appropriate OWord within the control data header.
1794 abld
.SHR(per_slot_offset
, dword_index
, brw_imm_ud(2u));
1797 /* Set the channel masks to 1 << (dword_index % 4), so that we'll
1798 * write to the appropriate DWORD within the OWORD.
1800 fs_reg channel
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1801 fwa_bld
.AND(channel
, dword_index
, brw_imm_ud(3u));
1802 channel_mask
= intexp2(fwa_bld
, channel
);
1803 /* Then the channel masks need to be in bits 23:16. */
1804 fwa_bld
.SHL(channel_mask
, channel_mask
, brw_imm_ud(16u));
1807 /* Store the control data bits in the message payload and send it. */
1809 if (channel_mask
.file
!= BAD_FILE
)
1810 mlen
+= 4; /* channel masks, plus 3 extra copies of the data */
1811 if (per_slot_offset
.file
!= BAD_FILE
)
1814 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
1815 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, mlen
);
1817 sources
[i
++] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1818 if (per_slot_offset
.file
!= BAD_FILE
)
1819 sources
[i
++] = per_slot_offset
;
1820 if (channel_mask
.file
!= BAD_FILE
)
1821 sources
[i
++] = channel_mask
;
1823 sources
[i
++] = this->control_data_bits
;
1826 abld
.LOAD_PAYLOAD(payload
, sources
, mlen
, mlen
);
1827 fs_inst
*inst
= abld
.emit(opcode
, reg_undef
, payload
);
1829 /* We need to increment Global Offset by 256-bits to make room for
1830 * Broadwell's extra "Vertex Count" payload at the beginning of the
1831 * URB entry. Since this is an OWord message, Global Offset is counted
1832 * in 128-bit units, so we must set it to 2.
1834 if (gs_prog_data
->static_vertex_count
== -1)
1839 fs_visitor::set_gs_stream_control_data_bits(const fs_reg
&vertex_count
,
1842 /* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
1844 /* Note: we are calling this *before* increasing vertex_count, so
1845 * this->vertex_count == vertex_count - 1 in the formula above.
1848 /* Stream mode uses 2 bits per vertex */
1849 assert(gs_compile
->control_data_bits_per_vertex
== 2);
1851 /* Must be a valid stream */
1852 assert(stream_id
>= 0 && stream_id
< MAX_VERTEX_STREAMS
);
1854 /* Control data bits are initialized to 0 so we don't have to set any
1855 * bits when sending vertices to stream 0.
1860 const fs_builder abld
= bld
.annotate("set stream control data bits", NULL
);
1862 /* reg::sid = stream_id */
1863 fs_reg sid
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1864 abld
.MOV(sid
, brw_imm_ud(stream_id
));
1866 /* reg:shift_count = 2 * (vertex_count - 1) */
1867 fs_reg shift_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1868 abld
.SHL(shift_count
, vertex_count
, brw_imm_ud(1u));
1870 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1871 * attention to the lower 5 bits of its second source argument, so on this
1872 * architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
1873 * stream_id << ((2 * (vertex_count - 1)) % 32).
1875 fs_reg mask
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1876 abld
.SHL(mask
, sid
, shift_count
);
1877 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1881 fs_visitor::emit_gs_vertex(const nir_src
&vertex_count_nir_src
,
1884 assert(stage
== MESA_SHADER_GEOMETRY
);
1886 struct brw_gs_prog_data
*gs_prog_data
=
1887 (struct brw_gs_prog_data
*) prog_data
;
1889 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1890 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1892 /* Haswell and later hardware ignores the "Render Stream Select" bits
1893 * from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
1894 * and instead sends all primitives down the pipeline for rasterization.
1895 * If the SOL stage is enabled, "Render Stream Select" is honored and
1896 * primitives bound to non-zero streams are discarded after stream output.
1898 * Since the only purpose of primives sent to non-zero streams is to
1899 * be recorded by transform feedback, we can simply discard all geometry
1900 * bound to these streams when transform feedback is disabled.
1902 if (stream_id
> 0 && !nir
->info
.has_transform_feedback_varyings
)
1905 /* If we're outputting 32 control data bits or less, then we can wait
1906 * until the shader is over to output them all. Otherwise we need to
1907 * output them as we go. Now is the time to do it, since we're about to
1908 * output the vertex_count'th vertex, so it's guaranteed that the
1909 * control data bits associated with the (vertex_count - 1)th vertex are
1912 if (gs_compile
->control_data_header_size_bits
> 32) {
1913 const fs_builder abld
=
1914 bld
.annotate("emit vertex: emit control data bits");
1916 /* Only emit control data bits if we've finished accumulating a batch
1917 * of 32 bits. This is the case when:
1919 * (vertex_count * bits_per_vertex) % 32 == 0
1921 * (in other words, when the last 5 bits of vertex_count *
1922 * bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
1923 * integer n (which is always the case, since bits_per_vertex is
1924 * always 1 or 2), this is equivalent to requiring that the last 5-n
1925 * bits of vertex_count are 0:
1927 * vertex_count & (2^(5-n) - 1) == 0
1929 * 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
1932 * vertex_count & (32 / bits_per_vertex - 1) == 0
1934 * TODO: If vertex_count is an immediate, we could do some of this math
1935 * at compile time...
1938 abld
.AND(bld
.null_reg_d(), vertex_count
,
1939 brw_imm_ud(32u / gs_compile
->control_data_bits_per_vertex
- 1u));
1940 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
1942 abld
.IF(BRW_PREDICATE_NORMAL
);
1943 /* If vertex_count is 0, then no control data bits have been
1944 * accumulated yet, so we can skip emitting them.
1946 abld
.CMP(bld
.null_reg_d(), vertex_count
, brw_imm_ud(0u),
1947 BRW_CONDITIONAL_NEQ
);
1948 abld
.IF(BRW_PREDICATE_NORMAL
);
1949 emit_gs_control_data_bits(vertex_count
);
1950 abld
.emit(BRW_OPCODE_ENDIF
);
1952 /* Reset control_data_bits to 0 so we can start accumulating a new
1955 * Note: in the case where vertex_count == 0, this neutralizes the
1956 * effect of any call to EndPrimitive() that the shader may have
1957 * made before outputting its first vertex.
1959 inst
= abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
1960 inst
->force_writemask_all
= true;
1961 abld
.emit(BRW_OPCODE_ENDIF
);
1964 emit_urb_writes(vertex_count
);
1966 /* In stream mode we have to set control data bits for all vertices
1967 * unless we have disabled control data bits completely (which we do
1968 * do for GL_POINTS outputs that don't use streams).
1970 if (gs_compile
->control_data_header_size_bits
> 0 &&
1971 gs_prog_data
->control_data_format
==
1972 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
) {
1973 set_gs_stream_control_data_bits(vertex_count
, stream_id
);
1978 fs_visitor::emit_gs_input_load(const fs_reg
&dst
,
1979 const nir_src
&vertex_src
,
1980 unsigned base_offset
,
1981 const nir_src
&offset_src
,
1982 unsigned num_components
)
1984 struct brw_gs_prog_data
*gs_prog_data
= (struct brw_gs_prog_data
*) prog_data
;
1986 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
1987 nir_const_value
*offset_const
= nir_src_as_const_value(offset_src
);
1988 const unsigned push_reg_count
= gs_prog_data
->base
.urb_read_length
* 8;
1990 /* Offset 0 is the VUE header, which contains VARYING_SLOT_LAYER [.y],
1991 * VARYING_SLOT_VIEWPORT [.z], and VARYING_SLOT_PSIZ [.w]. Only
1992 * gl_PointSize is available as a GS input, however, so it must be that.
1994 const bool is_point_size
= (base_offset
== 0);
1996 /* TODO: figure out push input layout for invocations == 1 */
1997 if (gs_prog_data
->invocations
== 1 &&
1998 offset_const
!= NULL
&& vertex_const
!= NULL
&&
1999 4 * (base_offset
+ offset_const
->u32
[0]) < push_reg_count
) {
2000 int imm_offset
= (base_offset
+ offset_const
->u32
[0]) * 4 +
2001 vertex_const
->u32
[0] * push_reg_count
;
2002 /* This input was pushed into registers. */
2003 if (is_point_size
) {
2004 /* gl_PointSize comes in .w */
2005 bld
.MOV(dst
, fs_reg(ATTR
, imm_offset
+ 3, dst
.type
));
2007 for (unsigned i
= 0; i
< num_components
; i
++) {
2008 bld
.MOV(offset(dst
, bld
, i
),
2009 fs_reg(ATTR
, imm_offset
+ i
, dst
.type
));
2015 /* Resort to the pull model. Ensure the VUE handles are provided. */
2016 gs_prog_data
->base
.include_vue_handles
= true;
2018 unsigned first_icp_handle
= gs_prog_data
->include_primitive_id
? 3 : 2;
2019 fs_reg icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2021 if (gs_prog_data
->invocations
== 1) {
2023 /* The vertex index is constant; just select the proper URB handle. */
2025 retype(brw_vec8_grf(first_icp_handle
+ vertex_const
->i32
[0], 0),
2026 BRW_REGISTER_TYPE_UD
);
2028 /* The vertex index is non-constant. We need to use indirect
2029 * addressing to fetch the proper URB handle.
2031 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
2032 * indicating that channel <n> should read the handle from
2033 * DWord <n>. We convert that to bytes by multiplying by 4.
2035 * Next, we convert the vertex index to bytes by multiplying
2036 * by 32 (shifting by 5), and add the two together. This is
2037 * the final indirect byte offset.
2039 fs_reg sequence
= bld
.vgrf(BRW_REGISTER_TYPE_W
, 1);
2040 fs_reg channel_offsets
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2041 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2042 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2044 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
2045 bld
.MOV(sequence
, fs_reg(brw_imm_v(0x76543210)));
2046 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
2047 bld
.SHL(channel_offsets
, sequence
, brw_imm_ud(2u));
2048 /* Convert vertex_index to bytes (multiply by 32) */
2049 bld
.SHL(vertex_offset_bytes
,
2050 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2052 bld
.ADD(icp_offset_bytes
, vertex_offset_bytes
, channel_offsets
);
2054 /* Use first_icp_handle as the base offset. There is one register
2055 * of URB handles per vertex, so inform the register allocator that
2056 * we might read up to nir->info.gs.vertices_in registers.
2058 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2059 fs_reg(brw_vec8_grf(first_icp_handle
, 0)),
2060 fs_reg(icp_offset_bytes
),
2061 brw_imm_ud(nir
->info
.gs
.vertices_in
* REG_SIZE
));
2064 assert(gs_prog_data
->invocations
> 1);
2067 assert(devinfo
->gen
>= 9 || vertex_const
->i32
[0] <= 5);
2069 retype(brw_vec1_grf(first_icp_handle
+
2070 vertex_const
->i32
[0] / 8,
2071 vertex_const
->i32
[0] % 8),
2072 BRW_REGISTER_TYPE_UD
));
2074 /* The vertex index is non-constant. We need to use indirect
2075 * addressing to fetch the proper URB handle.
2078 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2080 /* Convert vertex_index to bytes (multiply by 4) */
2081 bld
.SHL(icp_offset_bytes
,
2082 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2085 /* Use first_icp_handle as the base offset. There is one DWord
2086 * of URB handles per vertex, so inform the register allocator that
2087 * we might read up to ceil(nir->info.gs.vertices_in / 8) registers.
2089 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2090 fs_reg(brw_vec8_grf(first_icp_handle
, 0)),
2091 fs_reg(icp_offset_bytes
),
2092 brw_imm_ud(DIV_ROUND_UP(nir
->info
.gs
.vertices_in
, 8) *
2099 /* Constant indexing - use global offset. */
2100 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2101 inst
->offset
= base_offset
+ offset_const
->u32
[0];
2102 inst
->base_mrf
= -1;
2104 inst
->regs_written
= num_components
;
2106 /* Indirect indexing - use per-slot offsets as well. */
2107 const fs_reg srcs
[] = { icp_handle
, get_nir_src(offset_src
) };
2108 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2109 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2111 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2112 inst
->offset
= base_offset
;
2113 inst
->base_mrf
= -1;
2115 inst
->regs_written
= num_components
;
2118 if (is_point_size
) {
2119 /* Read the whole VUE header (because of alignment) and read .w. */
2120 fs_reg tmp
= bld
.vgrf(dst
.type
, 4);
2122 inst
->regs_written
= 4;
2123 bld
.MOV(dst
, offset(tmp
, bld
, 3));
2128 fs_visitor::get_indirect_offset(nir_intrinsic_instr
*instr
)
2130 nir_src
*offset_src
= nir_get_io_offset_src(instr
);
2131 nir_const_value
*const_value
= nir_src_as_const_value(*offset_src
);
2134 /* The only constant offset we should find is 0. brw_nir.c's
2135 * add_const_offset_to_base() will fold other constant offsets
2136 * into instr->const_index[0].
2138 assert(const_value
->u32
[0] == 0);
2142 return get_nir_src(*offset_src
);
2146 do_untyped_vector_read(const fs_builder
&bld
,
2148 const fs_reg surf_index
,
2149 const fs_reg offset_reg
,
2150 unsigned num_components
)
2152 if (type_sz(dest
.type
) == 4) {
2153 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, offset_reg
,
2156 BRW_PREDICATE_NONE
);
2157 read_result
.type
= dest
.type
;
2158 for (unsigned i
= 0; i
< num_components
; i
++)
2159 bld
.MOV(offset(dest
, bld
, i
), offset(read_result
, bld
, i
));
2160 } else if (type_sz(dest
.type
) == 8) {
2161 /* Reading a dvec, so we need to:
2163 * 1. Multiply num_components by 2, to account for the fact that we
2164 * need to read 64-bit components.
2165 * 2. Shuffle the result of the load to form valid 64-bit elements
2166 * 3. Emit a second load (for components z/w) if needed.
2168 fs_reg read_offset
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
2169 bld
.MOV(read_offset
, offset_reg
);
2171 int iters
= num_components
<= 2 ? 1 : 2;
2173 /* Load the dvec, the first iteration loads components x/y, the second
2174 * iteration, if needed, loads components z/w
2176 for (int it
= 0; it
< iters
; it
++) {
2177 /* Compute number of components to read in this iteration */
2178 int iter_components
= MIN2(2, num_components
);
2179 num_components
-= iter_components
;
2181 /* Read. Since this message reads 32-bit components, we need to
2182 * read twice as many components.
2184 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, read_offset
,
2186 iter_components
* 2,
2187 BRW_PREDICATE_NONE
);
2189 /* Shuffle the 32-bit load result into valid 64-bit data */
2190 const fs_reg packed_result
= bld
.vgrf(dest
.type
, iter_components
);
2191 shuffle_32bit_load_result_to_64bit_data(
2192 bld
, packed_result
, read_result
, iter_components
);
2194 /* Move each component to its destination */
2195 read_result
= retype(read_result
, BRW_REGISTER_TYPE_DF
);
2196 for (int c
= 0; c
< iter_components
; c
++) {
2197 bld
.MOV(offset(dest
, bld
, it
* 2 + c
),
2198 offset(packed_result
, bld
, c
));
2201 bld
.ADD(read_offset
, read_offset
, brw_imm_ud(16));
2204 unreachable("Unsupported type");
2209 fs_visitor::nir_emit_vs_intrinsic(const fs_builder
&bld
,
2210 nir_intrinsic_instr
*instr
)
2212 assert(stage
== MESA_SHADER_VERTEX
);
2215 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2216 dest
= get_nir_dest(instr
->dest
);
2218 switch (instr
->intrinsic
) {
2219 case nir_intrinsic_load_vertex_id
:
2220 unreachable("should be lowered by lower_vertex_id()");
2222 case nir_intrinsic_load_vertex_id_zero_base
:
2223 case nir_intrinsic_load_base_vertex
:
2224 case nir_intrinsic_load_instance_id
:
2225 case nir_intrinsic_load_base_instance
:
2226 case nir_intrinsic_load_draw_id
: {
2227 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
2228 fs_reg val
= nir_system_values
[sv
];
2229 assert(val
.file
!= BAD_FILE
);
2230 dest
.type
= val
.type
;
2236 nir_emit_intrinsic(bld
, instr
);
2242 fs_visitor::nir_emit_tcs_intrinsic(const fs_builder
&bld
,
2243 nir_intrinsic_instr
*instr
)
2245 assert(stage
== MESA_SHADER_TESS_CTRL
);
2246 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
2247 struct brw_tcs_prog_data
*tcs_prog_data
=
2248 (struct brw_tcs_prog_data
*) prog_data
;
2251 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2252 dst
= get_nir_dest(instr
->dest
);
2254 switch (instr
->intrinsic
) {
2255 case nir_intrinsic_load_primitive_id
:
2256 bld
.MOV(dst
, fs_reg(brw_vec1_grf(0, 1)));
2258 case nir_intrinsic_load_invocation_id
:
2259 bld
.MOV(retype(dst
, invocation_id
.type
), invocation_id
);
2261 case nir_intrinsic_load_patch_vertices_in
:
2262 bld
.MOV(retype(dst
, BRW_REGISTER_TYPE_D
),
2263 brw_imm_d(tcs_key
->input_vertices
));
2266 case nir_intrinsic_barrier
: {
2267 if (tcs_prog_data
->instances
== 1)
2270 fs_reg m0
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2271 fs_reg m0_2
= byte_offset(m0
, 2 * sizeof(uint32_t));
2273 const fs_builder fwa_bld
= bld
.exec_all();
2275 /* Zero the message header */
2276 fwa_bld
.MOV(m0
, brw_imm_ud(0u));
2278 /* Copy "Barrier ID" from r0.2, bits 16:13 */
2279 fwa_bld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
2280 brw_imm_ud(INTEL_MASK(16, 13)));
2282 /* Shift it up to bits 27:24. */
2283 fwa_bld
.SHL(m0_2
, m0_2
, brw_imm_ud(11));
2285 /* Set the Barrier Count and the enable bit */
2286 fwa_bld
.OR(m0_2
, m0_2
,
2287 brw_imm_ud(tcs_prog_data
->instances
<< 8 | (1 << 15)));
2289 bld
.emit(SHADER_OPCODE_BARRIER
, bld
.null_reg_ud(), m0
);
2293 case nir_intrinsic_load_input
:
2294 unreachable("nir_lower_io should never give us these.");
2297 case nir_intrinsic_load_per_vertex_input
: {
2298 fs_reg indirect_offset
= get_indirect_offset(instr
);
2299 unsigned imm_offset
= instr
->const_index
[0];
2301 const nir_src
&vertex_src
= instr
->src
[0];
2302 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
2309 /* Emit a MOV to resolve <0,1,0> regioning. */
2310 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2312 retype(brw_vec1_grf(1 + (vertex_const
->i32
[0] >> 3),
2313 vertex_const
->i32
[0] & 7),
2314 BRW_REGISTER_TYPE_UD
));
2315 } else if (tcs_prog_data
->instances
== 1 &&
2316 vertex_src
.is_ssa
&&
2317 vertex_src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
&&
2318 nir_instr_as_intrinsic(vertex_src
.ssa
->parent_instr
)->intrinsic
== nir_intrinsic_load_invocation_id
) {
2319 /* For the common case of only 1 instance, an array index of
2320 * gl_InvocationID means reading g1. Skip all the indirect work.
2322 icp_handle
= retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
);
2324 /* The vertex index is non-constant. We need to use indirect
2325 * addressing to fetch the proper URB handle.
2327 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2329 /* Each ICP handle is a single DWord (4 bytes) */
2330 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2331 bld
.SHL(vertex_offset_bytes
,
2332 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2335 /* Start at g1. We might read up to 4 registers. */
2336 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2337 fs_reg(brw_vec8_grf(1, 0)), vertex_offset_bytes
,
2338 brw_imm_ud(4 * REG_SIZE
));
2341 /* We can only read two double components with each URB read, so
2342 * we send two read messages in that case, each one loading up to
2343 * two double components.
2345 unsigned num_iterations
= 1;
2346 unsigned num_components
= instr
->num_components
;
2347 fs_reg orig_dst
= dst
;
2348 if (type_sz(dst
.type
) == 8) {
2349 if (instr
->num_components
> 2) {
2354 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(4), dst
.type
);
2358 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2359 if (indirect_offset
.file
== BAD_FILE
) {
2360 /* Constant indexing - use global offset. */
2361 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2362 inst
->offset
= imm_offset
;
2364 inst
->base_mrf
= -1;
2366 /* Indirect indexing - use per-slot offsets as well. */
2367 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2368 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2369 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2371 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2372 inst
->offset
= imm_offset
;
2373 inst
->base_mrf
= -1;
2376 inst
->regs_written
= num_components
* type_sz(dst
.type
) / 4;
2378 /* If we are reading 64-bit data using 32-bit read messages we need
2379 * build proper 64-bit data elements by shuffling the low and high
2380 * 32-bit components around like we do for other things like UBOs
2383 if (type_sz(dst
.type
) == 8) {
2384 shuffle_32bit_load_result_to_64bit_data(
2385 bld
, dst
, retype(dst
, BRW_REGISTER_TYPE_F
), num_components
);
2387 for (unsigned c
= 0; c
< num_components
; c
++) {
2388 bld
.MOV(offset(orig_dst
, bld
, iter
* 2 + c
),
2389 offset(dst
, bld
, c
));
2393 /* Copy the temporary to the destination to deal with writemasking.
2395 * Also attempt to deal with gl_PointSize being in the .w component.
2397 if (inst
->offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
2398 assert(type_sz(dst
.type
) < 8);
2399 inst
->dst
= bld
.vgrf(dst
.type
, 4);
2400 inst
->regs_written
= 4;
2401 bld
.MOV(dst
, offset(inst
->dst
, bld
, 3));
2404 /* If we are loading double data and we need a second read message
2405 * adjust the write offset
2407 if (num_iterations
> 1) {
2408 num_components
= instr
->num_components
- 2;
2409 if (indirect_offset
.file
== BAD_FILE
) {
2412 fs_reg new_indirect
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2413 bld
.ADD(new_indirect
, indirect_offset
, brw_imm_ud(1u));
2414 indirect_offset
= new_indirect
;
2421 case nir_intrinsic_load_output
:
2422 case nir_intrinsic_load_per_vertex_output
: {
2423 fs_reg indirect_offset
= get_indirect_offset(instr
);
2424 unsigned imm_offset
= instr
->const_index
[0];
2427 if (indirect_offset
.file
== BAD_FILE
) {
2428 /* Replicate the patch handle to all enabled channels */
2429 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2430 bld
.MOV(patch_handle
,
2431 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2433 if (imm_offset
== 0) {
2434 /* This is a read of gl_TessLevelInner[], which lives in the
2435 * Patch URB header. The layout depends on the domain.
2437 dst
.type
= BRW_REGISTER_TYPE_F
;
2438 switch (tcs_key
->tes_primitive_mode
) {
2440 /* DWords 3-2 (reversed) */
2441 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
2443 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, patch_handle
);
2446 inst
->base_mrf
= -1;
2447 inst
->regs_written
= 4;
2449 /* dst.xy = tmp.wz */
2450 bld
.MOV(dst
, offset(tmp
, bld
, 3));
2451 bld
.MOV(offset(dst
, bld
, 1), offset(tmp
, bld
, 2));
2455 /* DWord 4; hardcode offset = 1 and regs_written = 1 */
2456 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, patch_handle
);
2459 inst
->base_mrf
= -1;
2460 inst
->regs_written
= 1;
2463 /* All channels are undefined. */
2466 unreachable("Bogus tessellation domain");
2468 } else if (imm_offset
== 1) {
2469 /* This is a read of gl_TessLevelOuter[], which lives in the
2470 * Patch URB header. The layout depends on the domain.
2472 dst
.type
= BRW_REGISTER_TYPE_F
;
2474 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4);
2475 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, patch_handle
);
2478 inst
->base_mrf
= -1;
2479 inst
->regs_written
= 4;
2481 /* Reswizzle: WZYX */
2483 offset(tmp
, bld
, 3),
2484 offset(tmp
, bld
, 2),
2485 offset(tmp
, bld
, 1),
2486 offset(tmp
, bld
, 0),
2489 unsigned num_components
;
2490 switch (tcs_key
->tes_primitive_mode
) {
2498 /* Isolines are not reversed; swizzle .zw -> .xy */
2499 srcs
[0] = offset(tmp
, bld
, 2);
2500 srcs
[1] = offset(tmp
, bld
, 3);
2504 unreachable("Bogus tessellation domain");
2506 bld
.LOAD_PAYLOAD(dst
, srcs
, num_components
, 0);
2508 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, patch_handle
);
2509 inst
->offset
= imm_offset
;
2511 inst
->base_mrf
= -1;
2512 inst
->regs_written
= instr
->num_components
;
2515 /* Indirect indexing - use per-slot offsets as well. */
2516 const fs_reg srcs
[] = {
2517 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2520 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2521 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2523 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2524 inst
->offset
= imm_offset
;
2526 inst
->base_mrf
= -1;
2527 inst
->regs_written
= instr
->num_components
;
2532 case nir_intrinsic_store_output
:
2533 case nir_intrinsic_store_per_vertex_output
: {
2534 fs_reg value
= get_nir_src(instr
->src
[0]);
2535 bool is_64bit
= (instr
->src
[0].is_ssa
?
2536 instr
->src
[0].ssa
->bit_size
: instr
->src
[0].reg
.reg
->bit_size
) == 64;
2537 fs_reg indirect_offset
= get_indirect_offset(instr
);
2538 unsigned imm_offset
= instr
->const_index
[0];
2539 unsigned swiz
= BRW_SWIZZLE_XYZW
;
2540 unsigned mask
= instr
->const_index
[1];
2541 unsigned header_regs
= 0;
2543 srcs
[header_regs
++] = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
);
2545 if (indirect_offset
.file
!= BAD_FILE
) {
2546 srcs
[header_regs
++] = indirect_offset
;
2547 } else if (!is_passthrough_shader
) {
2548 if (imm_offset
== 0) {
2549 value
.type
= BRW_REGISTER_TYPE_F
;
2551 mask
&= (1 << tesslevel_inner_components(tcs_key
->tes_primitive_mode
)) - 1;
2553 /* This is a write to gl_TessLevelInner[], which lives in the
2554 * Patch URB header. The layout depends on the domain.
2556 switch (tcs_key
->tes_primitive_mode
) {
2558 /* gl_TessLevelInner[].xy lives at DWords 3-2 (reversed).
2559 * We use an XXYX swizzle to reverse put .xy in the .wz
2560 * channels, and use a .zw writemask.
2562 mask
= writemask_for_backwards_vector(mask
);
2563 swiz
= BRW_SWIZZLE4(0, 0, 1, 0);
2566 /* gl_TessLevelInner[].x lives at DWord 4, so we set the
2567 * writemask to X and bump the URB offset by 1.
2572 /* Skip; gl_TessLevelInner[] doesn't exist for isolines. */
2575 unreachable("Bogus tessellation domain");
2577 } else if (imm_offset
== 1) {
2578 /* This is a write to gl_TessLevelOuter[] which lives in the
2579 * Patch URB Header at DWords 4-7. However, it's reversed, so
2580 * instead of .xyzw we have .wzyx.
2582 value
.type
= BRW_REGISTER_TYPE_F
;
2584 mask
&= (1 << tesslevel_outer_components(tcs_key
->tes_primitive_mode
)) - 1;
2586 if (tcs_key
->tes_primitive_mode
== GL_ISOLINES
) {
2587 /* Isolines .xy should be stored in .zw, in order. */
2588 swiz
= BRW_SWIZZLE4(0, 0, 0, 1);
2591 /* Other domains are reversed; store .wzyx instead of .xyzw */
2592 swiz
= BRW_SWIZZLE_WZYX
;
2593 mask
= writemask_for_backwards_vector(mask
);
2601 unsigned num_components
= _mesa_fls(mask
);
2604 /* We can only pack two 64-bit components in a single message, so send
2605 * 2 messages if we have more components
2607 unsigned num_iterations
= 1;
2608 unsigned iter_components
= num_components
;
2609 if (is_64bit
&& instr
->num_components
> 2) {
2611 iter_components
= 2;
2614 /* 64-bit data needs to me shuffled before we can write it to the URB.
2615 * We will use this temporary to shuffle the components in each
2619 fs_reg(VGRF
, alloc
.allocate(2 * iter_components
), value
.type
);
2621 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2622 if (!is_64bit
&& mask
!= WRITEMASK_XYZW
) {
2623 srcs
[header_regs
++] = brw_imm_ud(mask
<< 16);
2624 opcode
= indirect_offset
.file
!= BAD_FILE
?
2625 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2626 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2627 } else if (is_64bit
&& ((mask
& WRITEMASK_XY
) != WRITEMASK_XY
)) {
2628 /* Expand the 64-bit mask to 32-bit channels. We only handle
2629 * two channels in each iteration, so we only care about X/Y.
2631 unsigned mask32
= 0;
2632 if (mask
& WRITEMASK_X
)
2633 mask32
|= WRITEMASK_XY
;
2634 if (mask
& WRITEMASK_Y
)
2635 mask32
|= WRITEMASK_ZW
;
2637 /* If the mask does not include any of the channels X or Y there
2638 * is nothing to do in this iteration. Move on to the next couple
2639 * of 64-bit channels.
2647 srcs
[header_regs
++] = brw_imm_ud(mask32
<< 16);
2648 opcode
= indirect_offset
.file
!= BAD_FILE
?
2649 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2650 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2652 opcode
= indirect_offset
.file
!= BAD_FILE
?
2653 SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2654 SHADER_OPCODE_URB_WRITE_SIMD8
;
2657 for (unsigned i
= 0; i
< iter_components
; i
++) {
2658 if (!(mask
& (1 << i
)))
2662 srcs
[header_regs
+ i
] = offset(value
, bld
, BRW_GET_SWZ(swiz
, i
));
2664 /* We need to shuffle the 64-bit data to match the layout
2665 * expected by our 32-bit URB write messages. We use a temporary
2668 unsigned channel
= BRW_GET_SWZ(swiz
, iter
* 2 + i
);
2669 shuffle_64bit_data_for_32bit_write(bld
,
2670 retype(offset(tmp
, bld
, 2 * i
), BRW_REGISTER_TYPE_F
),
2671 retype(offset(value
, bld
, 2 * channel
), BRW_REGISTER_TYPE_DF
),
2674 /* Now copy the data to the destination */
2675 fs_reg dest
= fs_reg(VGRF
, alloc
.allocate(2), value
.type
);
2676 unsigned idx
= 2 * i
;
2677 bld
.MOV(dest
, offset(tmp
, bld
, idx
));
2678 bld
.MOV(offset(dest
, bld
, 1), offset(tmp
, bld
, idx
+ 1));
2679 srcs
[header_regs
+ idx
] = dest
;
2680 srcs
[header_regs
+ idx
+ 1] = offset(dest
, bld
, 1);
2685 header_regs
+ (is_64bit
? 2 * iter_components
: iter_components
);
2687 bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2688 bld
.LOAD_PAYLOAD(payload
, srcs
, mlen
, header_regs
);
2690 fs_inst
*inst
= bld
.emit(opcode
, bld
.null_reg_ud(), payload
);
2691 inst
->offset
= imm_offset
;
2693 inst
->base_mrf
= -1;
2695 /* If this is a 64-bit attribute, select the next two 64-bit channels
2696 * to be handled in the next iteration.
2707 nir_emit_intrinsic(bld
, instr
);
2713 fs_visitor::nir_emit_tes_intrinsic(const fs_builder
&bld
,
2714 nir_intrinsic_instr
*instr
)
2716 assert(stage
== MESA_SHADER_TESS_EVAL
);
2717 struct brw_tes_prog_data
*tes_prog_data
= (struct brw_tes_prog_data
*) prog_data
;
2720 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2721 dest
= get_nir_dest(instr
->dest
);
2723 switch (instr
->intrinsic
) {
2724 case nir_intrinsic_load_primitive_id
:
2725 bld
.MOV(dest
, fs_reg(brw_vec1_grf(0, 1)));
2727 case nir_intrinsic_load_tess_coord
:
2728 /* gl_TessCoord is part of the payload in g1-3 */
2729 for (unsigned i
= 0; i
< 3; i
++) {
2730 bld
.MOV(offset(dest
, bld
, i
), fs_reg(brw_vec8_grf(1 + i
, 0)));
2734 case nir_intrinsic_load_tess_level_outer
:
2735 /* When the TES reads gl_TessLevelOuter, we ensure that the patch header
2736 * appears as a push-model input. So, we can simply use the ATTR file
2737 * rather than issuing URB read messages. The data is stored in the
2738 * high DWords in reverse order - DWord 7 contains .x, DWord 6 contains
2741 switch (tes_prog_data
->domain
) {
2742 case BRW_TESS_DOMAIN_QUAD
:
2743 for (unsigned i
= 0; i
< 4; i
++)
2744 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2746 case BRW_TESS_DOMAIN_TRI
:
2747 for (unsigned i
= 0; i
< 3; i
++)
2748 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2750 case BRW_TESS_DOMAIN_ISOLINE
:
2751 for (unsigned i
= 0; i
< 2; i
++)
2752 bld
.MOV(offset(dest
, bld
, i
), component(fs_reg(ATTR
, 0), 7 - i
));
2757 case nir_intrinsic_load_tess_level_inner
:
2758 /* When the TES reads gl_TessLevelInner, we ensure that the patch header
2759 * appears as a push-model input. So, we can simply use the ATTR file
2760 * rather than issuing URB read messages.
2762 switch (tes_prog_data
->domain
) {
2763 case BRW_TESS_DOMAIN_QUAD
:
2764 bld
.MOV(dest
, component(fs_reg(ATTR
, 0), 3));
2765 bld
.MOV(offset(dest
, bld
, 1), component(fs_reg(ATTR
, 0), 2));
2767 case BRW_TESS_DOMAIN_TRI
:
2768 bld
.MOV(dest
, component(fs_reg(ATTR
, 0), 4));
2770 case BRW_TESS_DOMAIN_ISOLINE
:
2771 /* ignore - value is undefined */
2776 case nir_intrinsic_load_input
:
2777 case nir_intrinsic_load_per_vertex_input
: {
2778 fs_reg indirect_offset
= get_indirect_offset(instr
);
2779 unsigned imm_offset
= instr
->const_index
[0];
2782 if (indirect_offset
.file
== BAD_FILE
) {
2783 /* Arbitrarily only push up to 32 vec4 slots worth of data,
2784 * which is 16 registers (since each holds 2 vec4 slots).
2786 const unsigned max_push_slots
= 32;
2787 if (imm_offset
< max_push_slots
) {
2788 fs_reg src
= fs_reg(ATTR
, imm_offset
/ 2, dest
.type
);
2789 for (int i
= 0; i
< instr
->num_components
; i
++) {
2790 unsigned comp
= 16 / type_sz(dest
.type
) * (imm_offset
% 2) + i
;
2791 bld
.MOV(offset(dest
, bld
, i
), component(src
, comp
));
2793 tes_prog_data
->base
.urb_read_length
=
2794 MAX2(tes_prog_data
->base
.urb_read_length
,
2795 DIV_ROUND_UP(imm_offset
+ 1, 2));
2797 /* Replicate the patch handle to all enabled channels */
2798 const fs_reg srcs
[] = {
2799 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)
2801 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2802 bld
.LOAD_PAYLOAD(patch_handle
, srcs
, ARRAY_SIZE(srcs
), 0);
2804 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dest
, patch_handle
);
2806 inst
->offset
= imm_offset
;
2807 inst
->base_mrf
= -1;
2808 inst
->regs_written
= instr
->num_components
;
2811 /* Indirect indexing - use per-slot offsets as well. */
2812 const fs_reg srcs
[] = {
2813 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2816 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2817 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2819 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dest
, payload
);
2821 inst
->offset
= imm_offset
;
2822 inst
->base_mrf
= -1;
2823 inst
->regs_written
= instr
->num_components
;
2828 nir_emit_intrinsic(bld
, instr
);
2834 fs_visitor::nir_emit_gs_intrinsic(const fs_builder
&bld
,
2835 nir_intrinsic_instr
*instr
)
2837 assert(stage
== MESA_SHADER_GEOMETRY
);
2838 fs_reg indirect_offset
;
2841 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2842 dest
= get_nir_dest(instr
->dest
);
2844 switch (instr
->intrinsic
) {
2845 case nir_intrinsic_load_primitive_id
:
2846 assert(stage
== MESA_SHADER_GEOMETRY
);
2847 assert(((struct brw_gs_prog_data
*)prog_data
)->include_primitive_id
);
2848 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
2849 retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD
));
2852 case nir_intrinsic_load_input
:
2853 unreachable("load_input intrinsics are invalid for the GS stage");
2855 case nir_intrinsic_load_per_vertex_input
:
2856 emit_gs_input_load(dest
, instr
->src
[0], instr
->const_index
[0],
2857 instr
->src
[1], instr
->num_components
);
2860 case nir_intrinsic_emit_vertex_with_counter
:
2861 emit_gs_vertex(instr
->src
[0], instr
->const_index
[0]);
2864 case nir_intrinsic_end_primitive_with_counter
:
2865 emit_gs_end_primitive(instr
->src
[0]);
2868 case nir_intrinsic_set_vertex_count
:
2869 bld
.MOV(this->final_gs_vertex_count
, get_nir_src(instr
->src
[0]));
2872 case nir_intrinsic_load_invocation_id
: {
2873 fs_reg val
= nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
2874 assert(val
.file
!= BAD_FILE
);
2875 dest
.type
= val
.type
;
2881 nir_emit_intrinsic(bld
, instr
);
2887 fs_visitor::nir_emit_fs_intrinsic(const fs_builder
&bld
,
2888 nir_intrinsic_instr
*instr
)
2890 assert(stage
== MESA_SHADER_FRAGMENT
);
2891 struct brw_wm_prog_data
*wm_prog_data
=
2892 (struct brw_wm_prog_data
*) prog_data
;
2893 const struct brw_wm_prog_key
*wm_key
= (const struct brw_wm_prog_key
*) key
;
2896 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2897 dest
= get_nir_dest(instr
->dest
);
2899 switch (instr
->intrinsic
) {
2900 case nir_intrinsic_load_front_face
:
2901 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
2902 *emit_frontfacing_interpolation());
2905 case nir_intrinsic_load_sample_pos
: {
2906 fs_reg sample_pos
= nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
2907 assert(sample_pos
.file
!= BAD_FILE
);
2908 dest
.type
= sample_pos
.type
;
2909 bld
.MOV(dest
, sample_pos
);
2910 bld
.MOV(offset(dest
, bld
, 1), offset(sample_pos
, bld
, 1));
2914 case nir_intrinsic_load_helper_invocation
:
2915 case nir_intrinsic_load_sample_mask_in
:
2916 case nir_intrinsic_load_sample_id
: {
2917 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
2918 fs_reg val
= nir_system_values
[sv
];
2919 assert(val
.file
!= BAD_FILE
);
2920 dest
.type
= val
.type
;
2925 case nir_intrinsic_discard
:
2926 case nir_intrinsic_discard_if
: {
2927 /* We track our discarded pixels in f0.1. By predicating on it, we can
2928 * update just the flag bits that aren't yet discarded. If there's no
2929 * condition, we emit a CMP of g0 != g0, so all currently executing
2930 * channels will get turned off.
2933 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
2934 cmp
= bld
.CMP(bld
.null_reg_f(), get_nir_src(instr
->src
[0]),
2935 brw_imm_d(0), BRW_CONDITIONAL_Z
);
2937 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
2938 BRW_REGISTER_TYPE_UW
));
2939 cmp
= bld
.CMP(bld
.null_reg_f(), some_reg
, some_reg
, BRW_CONDITIONAL_NZ
);
2941 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
2942 cmp
->flag_subreg
= 1;
2944 if (devinfo
->gen
>= 6) {
2945 emit_discard_jump();
2950 case nir_intrinsic_interp_var_at_centroid
:
2951 case nir_intrinsic_interp_var_at_sample
:
2952 case nir_intrinsic_interp_var_at_offset
: {
2953 /* Handle ARB_gpu_shader5 interpolation intrinsics
2955 * It's worth a quick word of explanation as to why we handle the full
2956 * variable-based interpolation intrinsic rather than a lowered version
2957 * with like we do for other inputs. We have to do that because the way
2958 * we set up inputs doesn't allow us to use the already setup inputs for
2959 * interpolation. At the beginning of the shader, we go through all of
2960 * the input variables and do the initial interpolation and put it in
2961 * the nir_inputs array based on its location as determined in
2962 * nir_lower_io. If the input isn't used, dead code cleans up and
2963 * everything works fine. However, when we get to the ARB_gpu_shader5
2964 * interpolation intrinsics, we need to reinterpolate the input
2965 * differently. If we used an intrinsic that just had an index it would
2966 * only give us the offset into the nir_inputs array. However, this is
2967 * useless because that value is post-interpolation and we need
2968 * pre-interpolation. In order to get the actual location of the bits
2969 * we get from the vertex fetching hardware, we need the variable.
2971 wm_prog_data
->pulls_bary
= true;
2973 fs_reg dst_xy
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 2);
2974 const glsl_interp_qualifier interpolation
=
2975 (glsl_interp_qualifier
) instr
->variables
[0]->var
->data
.interpolation
;
2977 switch (instr
->intrinsic
) {
2978 case nir_intrinsic_interp_var_at_centroid
:
2979 emit_pixel_interpolater_send(bld
,
2980 FS_OPCODE_INTERPOLATE_AT_CENTROID
,
2987 case nir_intrinsic_interp_var_at_sample
: {
2988 if (!wm_key
->multisample_fbo
) {
2989 /* From the ARB_gpu_shader5 specification:
2990 * "If multisample buffers are not available, the input varying
2991 * will be evaluated at the center of the pixel."
2993 emit_pixel_interpolater_send(bld
,
2994 FS_OPCODE_INTERPOLATE_AT_CENTROID
,
3002 nir_const_value
*const_sample
= nir_src_as_const_value(instr
->src
[0]);
3005 unsigned msg_data
= const_sample
->i32
[0] << 4;
3007 emit_pixel_interpolater_send(bld
,
3008 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3011 brw_imm_ud(msg_data
),
3014 const fs_reg sample_src
= retype(get_nir_src(instr
->src
[0]),
3015 BRW_REGISTER_TYPE_UD
);
3017 if (nir_src_is_dynamically_uniform(instr
->src
[0])) {
3018 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3019 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3020 bld
.exec_all().group(1, 0)
3021 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3022 emit_pixel_interpolater_send(bld
,
3023 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3029 /* Make a loop that sends a message to the pixel interpolater
3030 * for the sample number in each live channel. If there are
3031 * multiple channels with the same sample number then these
3032 * will be handled simultaneously with a single interation of
3035 bld
.emit(BRW_OPCODE_DO
);
3037 /* Get the next live sample number into sample_id_reg */
3038 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3040 /* Set the flag register so that we can perform the send
3041 * message on all channels that have the same sample number
3043 bld
.CMP(bld
.null_reg_ud(),
3044 sample_src
, sample_id
,
3045 BRW_CONDITIONAL_EQ
);
3046 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3047 bld
.exec_all().group(1, 0)
3048 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3050 emit_pixel_interpolater_send(bld
,
3051 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3056 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
3058 /* Continue the loop if there are any live channels left */
3059 set_predicate_inv(BRW_PREDICATE_NORMAL
,
3061 bld
.emit(BRW_OPCODE_WHILE
));
3068 case nir_intrinsic_interp_var_at_offset
: {
3069 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3072 unsigned off_x
= MIN2((int)(const_offset
->f32
[0] * 16), 7) & 0xf;
3073 unsigned off_y
= MIN2((int)(const_offset
->f32
[1] * 16), 7) & 0xf;
3075 emit_pixel_interpolater_send(bld
,
3076 FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
,
3079 brw_imm_ud(off_x
| (off_y
<< 4)),
3082 fs_reg src
= vgrf(glsl_type::ivec2_type
);
3083 fs_reg offset_src
= retype(get_nir_src(instr
->src
[0]),
3084 BRW_REGISTER_TYPE_F
);
3085 for (int i
= 0; i
< 2; i
++) {
3086 fs_reg temp
= vgrf(glsl_type::float_type
);
3087 bld
.MUL(temp
, offset(offset_src
, bld
, i
), brw_imm_f(16.0f
));
3088 fs_reg itemp
= vgrf(glsl_type::int_type
);
3090 bld
.MOV(itemp
, temp
);
3092 /* Clamp the upper end of the range to +7/16.
3093 * ARB_gpu_shader5 requires that we support a maximum offset
3094 * of +0.5, which isn't representable in a S0.4 value -- if
3095 * we didn't clamp it, we'd end up with -8/16, which is the
3096 * opposite of what the shader author wanted.
3098 * This is legal due to ARB_gpu_shader5's quantization
3101 * "Not all values of <offset> may be supported; x and y
3102 * offsets may be rounded to fixed-point values with the
3103 * number of fraction bits given by the
3104 * implementation-dependent constant
3105 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
3107 set_condmod(BRW_CONDITIONAL_L
,
3108 bld
.SEL(offset(src
, bld
, i
), itemp
, brw_imm_d(7)));
3111 const enum opcode opcode
= FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
;
3112 emit_pixel_interpolater_send(bld
,
3123 unreachable("Invalid intrinsic");
3126 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3127 fs_reg src
= interp_reg(instr
->variables
[0]->var
->data
.location
, j
);
3128 src
.type
= dest
.type
;
3130 bld
.emit(FS_OPCODE_LINTERP
, dest
, dst_xy
, src
);
3131 dest
= offset(dest
, bld
, 1);
3136 nir_emit_intrinsic(bld
, instr
);
3142 fs_visitor::nir_emit_cs_intrinsic(const fs_builder
&bld
,
3143 nir_intrinsic_instr
*instr
)
3145 assert(stage
== MESA_SHADER_COMPUTE
);
3146 struct brw_cs_prog_data
*cs_prog_data
=
3147 (struct brw_cs_prog_data
*) prog_data
;
3150 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3151 dest
= get_nir_dest(instr
->dest
);
3153 switch (instr
->intrinsic
) {
3154 case nir_intrinsic_barrier
:
3156 cs_prog_data
->uses_barrier
= true;
3159 case nir_intrinsic_load_local_invocation_id
:
3160 case nir_intrinsic_load_work_group_id
: {
3161 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3162 fs_reg val
= nir_system_values
[sv
];
3163 assert(val
.file
!= BAD_FILE
);
3164 dest
.type
= val
.type
;
3165 for (unsigned i
= 0; i
< 3; i
++)
3166 bld
.MOV(offset(dest
, bld
, i
), offset(val
, bld
, i
));
3170 case nir_intrinsic_load_num_work_groups
: {
3171 const unsigned surface
=
3172 cs_prog_data
->binding_table
.work_groups_start
;
3174 cs_prog_data
->uses_num_work_groups
= true;
3176 fs_reg surf_index
= brw_imm_ud(surface
);
3177 brw_mark_surface_used(prog_data
, surface
);
3179 /* Read the 3 GLuint components of gl_NumWorkGroups */
3180 for (unsigned i
= 0; i
< 3; i
++) {
3181 fs_reg read_result
=
3182 emit_untyped_read(bld
, surf_index
,
3184 1 /* dims */, 1 /* size */,
3185 BRW_PREDICATE_NONE
);
3186 read_result
.type
= dest
.type
;
3187 bld
.MOV(dest
, read_result
);
3188 dest
= offset(dest
, bld
, 1);
3193 case nir_intrinsic_shared_atomic_add
:
3194 nir_emit_shared_atomic(bld
, BRW_AOP_ADD
, instr
);
3196 case nir_intrinsic_shared_atomic_imin
:
3197 nir_emit_shared_atomic(bld
, BRW_AOP_IMIN
, instr
);
3199 case nir_intrinsic_shared_atomic_umin
:
3200 nir_emit_shared_atomic(bld
, BRW_AOP_UMIN
, instr
);
3202 case nir_intrinsic_shared_atomic_imax
:
3203 nir_emit_shared_atomic(bld
, BRW_AOP_IMAX
, instr
);
3205 case nir_intrinsic_shared_atomic_umax
:
3206 nir_emit_shared_atomic(bld
, BRW_AOP_UMAX
, instr
);
3208 case nir_intrinsic_shared_atomic_and
:
3209 nir_emit_shared_atomic(bld
, BRW_AOP_AND
, instr
);
3211 case nir_intrinsic_shared_atomic_or
:
3212 nir_emit_shared_atomic(bld
, BRW_AOP_OR
, instr
);
3214 case nir_intrinsic_shared_atomic_xor
:
3215 nir_emit_shared_atomic(bld
, BRW_AOP_XOR
, instr
);
3217 case nir_intrinsic_shared_atomic_exchange
:
3218 nir_emit_shared_atomic(bld
, BRW_AOP_MOV
, instr
);
3220 case nir_intrinsic_shared_atomic_comp_swap
:
3221 nir_emit_shared_atomic(bld
, BRW_AOP_CMPWR
, instr
);
3224 case nir_intrinsic_load_shared
: {
3225 assert(devinfo
->gen
>= 7);
3227 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
3229 /* Get the offset to read from */
3231 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3233 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0]);
3235 offset_reg
= vgrf(glsl_type::uint_type
);
3237 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
3238 brw_imm_ud(instr
->const_index
[0]));
3241 /* Read the vector */
3242 do_untyped_vector_read(bld
, dest
, surf_index
, offset_reg
,
3243 instr
->num_components
);
3247 case nir_intrinsic_store_shared
: {
3248 assert(devinfo
->gen
>= 7);
3251 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
3254 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
3257 unsigned writemask
= instr
->const_index
[1];
3259 /* get_nir_src() retypes to integer. Be wary of 64-bit types though
3260 * since the untyped writes below operate in units of 32-bits, which
3261 * means that we need to write twice as many components each time.
3262 * Also, we have to suffle 64-bit data to be in the appropriate layout
3263 * expected by our 32-bit write messages.
3265 unsigned type_size
= 4;
3266 unsigned bit_size
= instr
->src
[0].is_ssa
?
3267 instr
->src
[0].ssa
->bit_size
: instr
->src
[0].reg
.reg
->bit_size
;
3268 if (bit_size
== 64) {
3271 fs_reg(VGRF
, alloc
.allocate(alloc
.sizes
[val_reg
.nr
]), val_reg
.type
);
3272 shuffle_64bit_data_for_32bit_write(
3274 retype(tmp
, BRW_REGISTER_TYPE_F
),
3275 retype(val_reg
, BRW_REGISTER_TYPE_DF
),
3276 instr
->num_components
);
3280 unsigned type_slots
= type_size
/ 4;
3282 /* Combine groups of consecutive enabled channels in one write
3283 * message. We use ffs to find the first enabled channel and then ffs on
3284 * the bit-inverse, down-shifted writemask to determine the length of
3285 * the block of enabled bits.
3288 unsigned first_component
= ffs(writemask
) - 1;
3289 unsigned length
= ffs(~(writemask
>> first_component
)) - 1;
3291 /* We can't write more than 2 64-bit components at once. Limit the
3292 * length of the write to what we can do and let the next iteration
3296 length
= MIN2(2, length
);
3299 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3301 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0] +
3302 type_size
* first_component
);
3304 offset_reg
= vgrf(glsl_type::uint_type
);
3306 retype(get_nir_src(instr
->src
[1]), BRW_REGISTER_TYPE_UD
),
3307 brw_imm_ud(instr
->const_index
[0] + type_size
* first_component
));
3310 emit_untyped_write(bld
, surf_index
, offset_reg
,
3311 offset(val_reg
, bld
, first_component
* type_slots
),
3312 1 /* dims */, length
* type_slots
,
3313 BRW_PREDICATE_NONE
);
3315 /* Clear the bits in the writemask that we just wrote, then try
3316 * again to see if more channels are left.
3318 writemask
&= (15 << (first_component
+ length
));
3325 nir_emit_intrinsic(bld
, instr
);
3331 fs_visitor::nir_emit_intrinsic(const fs_builder
&bld
, nir_intrinsic_instr
*instr
)
3334 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3335 dest
= get_nir_dest(instr
->dest
);
3337 switch (instr
->intrinsic
) {
3338 case nir_intrinsic_atomic_counter_inc
:
3339 case nir_intrinsic_atomic_counter_dec
:
3340 case nir_intrinsic_atomic_counter_read
: {
3341 /* Get the arguments of the atomic intrinsic. */
3342 const fs_reg offset
= get_nir_src(instr
->src
[0]);
3343 const unsigned surface
= (stage_prog_data
->binding_table
.abo_start
+
3344 instr
->const_index
[0]);
3347 /* Emit a surface read or atomic op. */
3348 switch (instr
->intrinsic
) {
3349 case nir_intrinsic_atomic_counter_read
:
3350 tmp
= emit_untyped_read(bld
, brw_imm_ud(surface
), offset
, 1, 1);
3353 case nir_intrinsic_atomic_counter_inc
:
3354 tmp
= emit_untyped_atomic(bld
, brw_imm_ud(surface
), offset
, fs_reg(),
3355 fs_reg(), 1, 1, BRW_AOP_INC
);
3358 case nir_intrinsic_atomic_counter_dec
:
3359 tmp
= emit_untyped_atomic(bld
, brw_imm_ud(surface
), offset
, fs_reg(),
3360 fs_reg(), 1, 1, BRW_AOP_PREDEC
);
3364 unreachable("Unreachable");
3367 /* Assign the result. */
3368 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
), tmp
);
3370 /* Mark the surface as used. */
3371 brw_mark_surface_used(stage_prog_data
, surface
);
3375 case nir_intrinsic_image_load
:
3376 case nir_intrinsic_image_store
:
3377 case nir_intrinsic_image_atomic_add
:
3378 case nir_intrinsic_image_atomic_min
:
3379 case nir_intrinsic_image_atomic_max
:
3380 case nir_intrinsic_image_atomic_and
:
3381 case nir_intrinsic_image_atomic_or
:
3382 case nir_intrinsic_image_atomic_xor
:
3383 case nir_intrinsic_image_atomic_exchange
:
3384 case nir_intrinsic_image_atomic_comp_swap
: {
3385 using namespace image_access
;
3387 /* Get the referenced image variable and type. */
3388 const nir_variable
*var
= instr
->variables
[0]->var
;
3389 const glsl_type
*type
= var
->type
->without_array();
3390 const brw_reg_type base_type
= get_image_base_type(type
);
3392 /* Get some metadata from the image intrinsic. */
3393 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
3394 const unsigned arr_dims
= type
->sampler_array
? 1 : 0;
3395 const unsigned surf_dims
= type
->coordinate_components() - arr_dims
;
3396 const unsigned format
= var
->data
.image
.format
;
3398 /* Get the arguments of the image intrinsic. */
3399 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
3400 const fs_reg addr
= retype(get_nir_src(instr
->src
[0]),
3401 BRW_REGISTER_TYPE_UD
);
3402 const fs_reg src0
= (info
->num_srcs
>= 3 ?
3403 retype(get_nir_src(instr
->src
[2]), base_type
) :
3405 const fs_reg src1
= (info
->num_srcs
>= 4 ?
3406 retype(get_nir_src(instr
->src
[3]), base_type
) :
3410 /* Emit an image load, store or atomic op. */
3411 if (instr
->intrinsic
== nir_intrinsic_image_load
)
3412 tmp
= emit_image_load(bld
, image
, addr
, surf_dims
, arr_dims
, format
);
3414 else if (instr
->intrinsic
== nir_intrinsic_image_store
)
3415 emit_image_store(bld
, image
, addr
, src0
, surf_dims
, arr_dims
,
3416 var
->data
.image
.write_only
? GL_NONE
: format
);
3419 tmp
= emit_image_atomic(bld
, image
, addr
, src0
, src1
,
3420 surf_dims
, arr_dims
, info
->dest_components
,
3421 get_image_atomic_op(instr
->intrinsic
, type
));
3423 /* Assign the result. */
3424 for (unsigned c
= 0; c
< info
->dest_components
; ++c
)
3425 bld
.MOV(offset(retype(dest
, base_type
), bld
, c
),
3426 offset(tmp
, bld
, c
));
3430 case nir_intrinsic_memory_barrier_atomic_counter
:
3431 case nir_intrinsic_memory_barrier_buffer
:
3432 case nir_intrinsic_memory_barrier_image
:
3433 case nir_intrinsic_memory_barrier
: {
3434 const fs_builder ubld
= bld
.group(8, 0);
3435 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
3436 ubld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
)
3441 case nir_intrinsic_group_memory_barrier
:
3442 case nir_intrinsic_memory_barrier_shared
:
3443 /* We treat these workgroup-level barriers as no-ops. This should be
3444 * safe at present and as long as:
3446 * - Memory access instructions are not subsequently reordered by the
3447 * compiler back-end.
3449 * - All threads from a given compute shader workgroup fit within a
3450 * single subslice and therefore talk to the same HDC shared unit
3451 * what supposedly guarantees ordering and coherency between threads
3452 * from the same workgroup. This may change in the future when we
3453 * start splitting workgroups across multiple subslices.
3455 * - The context is not in fault-and-stream mode, which could cause
3456 * memory transactions (including to SLM) prior to the barrier to be
3457 * replayed after the barrier if a pagefault occurs. This shouldn't
3458 * be a problem up to and including SKL because fault-and-stream is
3459 * not usable due to hardware issues, but that's likely to change in
3464 case nir_intrinsic_shader_clock
: {
3465 /* We cannot do anything if there is an event, so ignore it for now */
3466 fs_reg shader_clock
= get_timestamp(bld
);
3467 const fs_reg srcs
[] = { shader_clock
.set_smear(0), shader_clock
.set_smear(1) };
3469 bld
.LOAD_PAYLOAD(dest
, srcs
, ARRAY_SIZE(srcs
), 0);
3473 case nir_intrinsic_image_size
: {
3474 /* Get the referenced image variable and type. */
3475 const nir_variable
*var
= instr
->variables
[0]->var
;
3476 const glsl_type
*type
= var
->type
->without_array();
3478 /* Get the size of the image. */
3479 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
3480 const fs_reg size
= offset(image
, bld
, BRW_IMAGE_PARAM_SIZE_OFFSET
);
3482 /* For 1DArray image types, the array index is stored in the Z component.
3483 * Fix this by swizzling the Z component to the Y component.
3485 const bool is_1d_array_image
=
3486 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_1D
&&
3487 type
->sampler_array
;
3489 /* For CubeArray images, we should count the number of cubes instead
3490 * of the number of faces. Fix it by dividing the (Z component) by 6.
3492 const bool is_cube_array_image
=
3493 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
3494 type
->sampler_array
;
3496 /* Copy all the components. */
3497 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
3498 for (unsigned c
= 0; c
< info
->dest_components
; ++c
) {
3499 if ((int)c
>= type
->coordinate_components()) {
3500 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3502 } else if (c
== 1 && is_1d_array_image
) {
3503 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3504 offset(size
, bld
, 2));
3505 } else if (c
== 2 && is_cube_array_image
) {
3506 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
,
3507 offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3508 offset(size
, bld
, c
), brw_imm_d(6));
3510 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3511 offset(size
, bld
, c
));
3518 case nir_intrinsic_image_samples
:
3519 /* The driver does not support multi-sampled images. */
3520 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), brw_imm_d(1));
3523 case nir_intrinsic_load_uniform
: {
3524 /* Offsets are in bytes but they should always be multiples of 4 */
3525 assert(instr
->const_index
[0] % 4 == 0);
3527 fs_reg
src(UNIFORM
, instr
->const_index
[0] / 4, dest
.type
);
3529 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3531 /* Offsets are in bytes but they should always be multiples of 4 */
3532 assert(const_offset
->u32
[0] % 4 == 0);
3533 src
.reg_offset
= const_offset
->u32
[0] / 4;
3535 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3536 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
3539 fs_reg indirect
= retype(get_nir_src(instr
->src
[0]),
3540 BRW_REGISTER_TYPE_UD
);
3542 /* We need to pass a size to the MOV_INDIRECT but we don't want it to
3543 * go past the end of the uniform. In order to keep the n'th
3544 * component from running past, we subtract off the size of all but
3545 * one component of the vector.
3547 assert(instr
->const_index
[1] >=
3548 instr
->num_components
* (int) type_sz(dest
.type
));
3549 unsigned read_size
= instr
->const_index
[1] -
3550 (instr
->num_components
- 1) * type_sz(dest
.type
);
3552 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3553 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
3554 offset(dest
, bld
, j
), offset(src
, bld
, j
),
3555 indirect
, brw_imm_ud(read_size
));
3561 case nir_intrinsic_load_ubo
: {
3562 nir_const_value
*const_index
= nir_src_as_const_value(instr
->src
[0]);
3566 const unsigned index
= stage_prog_data
->binding_table
.ubo_start
+
3567 const_index
->u32
[0];
3568 surf_index
= brw_imm_ud(index
);
3569 brw_mark_surface_used(prog_data
, index
);
3571 /* The block index is not a constant. Evaluate the index expression
3572 * per-channel and add the base UBO index; we have to select a value
3573 * from any live channel.
3575 surf_index
= vgrf(glsl_type::uint_type
);
3576 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
3577 brw_imm_ud(stage_prog_data
->binding_table
.ubo_start
));
3578 surf_index
= bld
.emit_uniformize(surf_index
);
3580 /* Assume this may touch any UBO. It would be nice to provide
3581 * a tighter bound, but the array information is already lowered away.
3583 brw_mark_surface_used(prog_data
,
3584 stage_prog_data
->binding_table
.ubo_start
+
3585 nir
->info
.num_ubos
- 1);
3588 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3589 if (const_offset
== NULL
) {
3590 fs_reg base_offset
= retype(get_nir_src(instr
->src
[1]),
3591 BRW_REGISTER_TYPE_UD
);
3593 for (int i
= 0; i
< instr
->num_components
; i
++)
3594 VARYING_PULL_CONSTANT_LOAD(bld
, offset(dest
, bld
, i
), surf_index
,
3595 base_offset
, i
* type_sz(dest
.type
));
3597 /* Even if we are loading doubles, a pull constant load will load
3598 * a 32-bit vec4, so should only reserve vgrf space for that. If we
3599 * need to load a full dvec4 we will have to emit 2 loads. This is
3600 * similar to demote_pull_constants(), except that in that case we
3601 * see individual accesses to each component of the vector and then
3602 * we let CSE deal with duplicate loads. Here we see a vector access
3603 * and we have to split it if necessary.
3605 const unsigned type_size
= type_sz(dest
.type
);
3606 const fs_reg packed_consts
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
3607 for (unsigned c
= 0; c
< instr
->num_components
;) {
3608 const unsigned base
= const_offset
->u32
[0] + c
* type_size
;
3610 /* Number of usable components in the next 16B-aligned load */
3611 const unsigned count
= MIN2(instr
->num_components
- c
,
3612 (16 - base
% 16) / type_size
);
3615 .emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
3616 packed_consts
, surf_index
, brw_imm_ud(base
& ~15));
3618 const fs_reg consts
=
3619 retype(byte_offset(packed_consts
, base
& 15), dest
.type
);
3621 for (unsigned d
= 0; d
< count
; d
++)
3622 bld
.MOV(offset(dest
, bld
, c
+ d
), component(consts
, d
));
3630 case nir_intrinsic_load_ssbo
: {
3631 assert(devinfo
->gen
>= 7);
3633 nir_const_value
*const_uniform_block
=
3634 nir_src_as_const_value(instr
->src
[0]);
3637 if (const_uniform_block
) {
3638 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3639 const_uniform_block
->u32
[0];
3640 surf_index
= brw_imm_ud(index
);
3641 brw_mark_surface_used(prog_data
, index
);
3643 surf_index
= vgrf(glsl_type::uint_type
);
3644 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
3645 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3647 /* Assume this may touch any UBO. It would be nice to provide
3648 * a tighter bound, but the array information is already lowered away.
3650 brw_mark_surface_used(prog_data
,
3651 stage_prog_data
->binding_table
.ssbo_start
+
3652 nir
->info
.num_ssbos
- 1);
3656 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3658 offset_reg
= brw_imm_ud(const_offset
->u32
[0]);
3660 offset_reg
= get_nir_src(instr
->src
[1]);
3663 /* Read the vector */
3664 do_untyped_vector_read(bld
, dest
, surf_index
, offset_reg
,
3665 instr
->num_components
);
3670 case nir_intrinsic_load_input
: {
3672 if (stage
== MESA_SHADER_VERTEX
) {
3673 src
= fs_reg(ATTR
, instr
->const_index
[0], dest
.type
);
3675 src
= offset(retype(nir_inputs
, dest
.type
), bld
,
3676 instr
->const_index
[0]);
3679 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3680 assert(const_offset
&& "Indirect input loads not allowed");
3681 src
= offset(src
, bld
, const_offset
->u32
[0]);
3683 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3684 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
3687 if (type_sz(src
.type
) == 8) {
3688 shuffle_32bit_load_result_to_64bit_data(bld
,
3690 retype(dest
, BRW_REGISTER_TYPE_F
),
3691 instr
->num_components
);
3697 case nir_intrinsic_store_ssbo
: {
3698 assert(devinfo
->gen
>= 7);
3702 nir_const_value
*const_uniform_block
=
3703 nir_src_as_const_value(instr
->src
[1]);
3704 if (const_uniform_block
) {
3705 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3706 const_uniform_block
->u32
[0];
3707 surf_index
= brw_imm_ud(index
);
3708 brw_mark_surface_used(prog_data
, index
);
3710 surf_index
= vgrf(glsl_type::uint_type
);
3711 bld
.ADD(surf_index
, get_nir_src(instr
->src
[1]),
3712 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3714 brw_mark_surface_used(prog_data
,
3715 stage_prog_data
->binding_table
.ssbo_start
+
3716 nir
->info
.num_ssbos
- 1);
3720 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
3723 unsigned writemask
= instr
->const_index
[0];
3725 /* get_nir_src() retypes to integer. Be wary of 64-bit types though
3726 * since the untyped writes below operate in units of 32-bits, which
3727 * means that we need to write twice as many components each time.
3728 * Also, we have to suffle 64-bit data to be in the appropriate layout
3729 * expected by our 32-bit write messages.
3731 unsigned type_size
= 4;
3732 unsigned bit_size
= instr
->src
[0].is_ssa
?
3733 instr
->src
[0].ssa
->bit_size
: instr
->src
[0].reg
.reg
->bit_size
;
3734 if (bit_size
== 64) {
3737 fs_reg(VGRF
, alloc
.allocate(alloc
.sizes
[val_reg
.nr
]), val_reg
.type
);
3738 shuffle_64bit_data_for_32bit_write(bld
,
3739 retype(tmp
, BRW_REGISTER_TYPE_F
),
3740 retype(val_reg
, BRW_REGISTER_TYPE_DF
),
3741 instr
->num_components
);
3745 unsigned type_slots
= type_size
/ 4;
3747 /* Combine groups of consecutive enabled channels in one write
3748 * message. We use ffs to find the first enabled channel and then ffs on
3749 * the bit-inverse, down-shifted writemask to determine the length of
3750 * the block of enabled bits.
3753 unsigned first_component
= ffs(writemask
) - 1;
3754 unsigned length
= ffs(~(writemask
>> first_component
)) - 1;
3756 /* We can't write more than 2 64-bit components at once. Limit the
3757 * length of the write to what we can do and let the next iteration
3761 length
= MIN2(2, length
);
3764 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[2]);
3766 offset_reg
= brw_imm_ud(const_offset
->u32
[0] +
3767 type_size
* first_component
);
3769 offset_reg
= vgrf(glsl_type::uint_type
);
3771 retype(get_nir_src(instr
->src
[2]), BRW_REGISTER_TYPE_UD
),
3772 brw_imm_ud(type_size
* first_component
));
3776 emit_untyped_write(bld
, surf_index
, offset_reg
,
3777 offset(val_reg
, bld
, first_component
* type_slots
),
3778 1 /* dims */, length
* type_slots
,
3779 BRW_PREDICATE_NONE
);
3781 /* Clear the bits in the writemask that we just wrote, then try
3782 * again to see if more channels are left.
3784 writemask
&= (15 << (first_component
+ length
));
3789 case nir_intrinsic_store_output
: {
3790 fs_reg src
= get_nir_src(instr
->src
[0]);
3791 fs_reg new_dest
= offset(retype(nir_outputs
, src
.type
), bld
,
3792 instr
->const_index
[0]);
3794 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3795 assert(const_offset
&& "Indirect output stores not allowed");
3796 new_dest
= offset(new_dest
, bld
, const_offset
->u32
[0]);
3798 unsigned num_components
= instr
->num_components
;
3799 unsigned bit_size
= instr
->src
[0].is_ssa
?
3800 instr
->src
[0].ssa
->bit_size
: instr
->src
[0].reg
.reg
->bit_size
;
3801 if (bit_size
== 64) {
3803 fs_reg(VGRF
, alloc
.allocate(2 * num_components
),
3804 BRW_REGISTER_TYPE_F
);
3805 shuffle_64bit_data_for_32bit_write(
3806 bld
, tmp
, retype(src
, BRW_REGISTER_TYPE_DF
), num_components
);
3807 src
= retype(tmp
, src
.type
);
3808 num_components
*= 2;
3811 for (unsigned j
= 0; j
< num_components
; j
++) {
3812 bld
.MOV(offset(new_dest
, bld
, j
), offset(src
, bld
, j
));
3817 case nir_intrinsic_ssbo_atomic_add
:
3818 nir_emit_ssbo_atomic(bld
, BRW_AOP_ADD
, instr
);
3820 case nir_intrinsic_ssbo_atomic_imin
:
3821 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMIN
, instr
);
3823 case nir_intrinsic_ssbo_atomic_umin
:
3824 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMIN
, instr
);
3826 case nir_intrinsic_ssbo_atomic_imax
:
3827 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMAX
, instr
);
3829 case nir_intrinsic_ssbo_atomic_umax
:
3830 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMAX
, instr
);
3832 case nir_intrinsic_ssbo_atomic_and
:
3833 nir_emit_ssbo_atomic(bld
, BRW_AOP_AND
, instr
);
3835 case nir_intrinsic_ssbo_atomic_or
:
3836 nir_emit_ssbo_atomic(bld
, BRW_AOP_OR
, instr
);
3838 case nir_intrinsic_ssbo_atomic_xor
:
3839 nir_emit_ssbo_atomic(bld
, BRW_AOP_XOR
, instr
);
3841 case nir_intrinsic_ssbo_atomic_exchange
:
3842 nir_emit_ssbo_atomic(bld
, BRW_AOP_MOV
, instr
);
3844 case nir_intrinsic_ssbo_atomic_comp_swap
:
3845 nir_emit_ssbo_atomic(bld
, BRW_AOP_CMPWR
, instr
);
3848 case nir_intrinsic_get_buffer_size
: {
3849 nir_const_value
*const_uniform_block
= nir_src_as_const_value(instr
->src
[0]);
3850 unsigned ssbo_index
= const_uniform_block
? const_uniform_block
->u32
[0] : 0;
3852 /* A resinfo's sampler message is used to get the buffer size. The
3853 * SIMD8's writeback message consists of four registers and SIMD16's
3854 * writeback message consists of 8 destination registers (two per each
3855 * component). Because we are only interested on the first channel of
3856 * the first returned component, where resinfo returns the buffer size
3857 * for SURFTYPE_BUFFER, we can just use the SIMD8 variant regardless of
3858 * the dispatch width.
3860 const fs_builder ubld
= bld
.exec_all().group(8, 0);
3861 fs_reg src_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
3862 fs_reg ret_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 4);
3865 ubld
.MOV(src_payload
, brw_imm_d(0));
3867 const unsigned index
= prog_data
->binding_table
.ssbo_start
+ ssbo_index
;
3868 fs_inst
*inst
= ubld
.emit(FS_OPCODE_GET_BUFFER_SIZE
, ret_payload
,
3869 src_payload
, brw_imm_ud(index
));
3870 inst
->header_size
= 0;
3872 inst
->regs_written
= 4;
3874 bld
.MOV(retype(dest
, ret_payload
.type
), component(ret_payload
, 0));
3875 brw_mark_surface_used(prog_data
, index
);
3880 unreachable("unknown intrinsic");
3885 fs_visitor::nir_emit_ssbo_atomic(const fs_builder
&bld
,
3886 int op
, nir_intrinsic_instr
*instr
)
3889 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3890 dest
= get_nir_dest(instr
->dest
);
3893 nir_const_value
*const_surface
= nir_src_as_const_value(instr
->src
[0]);
3894 if (const_surface
) {
3895 unsigned surf_index
= stage_prog_data
->binding_table
.ssbo_start
+
3896 const_surface
->u32
[0];
3897 surface
= brw_imm_ud(surf_index
);
3898 brw_mark_surface_used(prog_data
, surf_index
);
3900 surface
= vgrf(glsl_type::uint_type
);
3901 bld
.ADD(surface
, get_nir_src(instr
->src
[0]),
3902 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3904 /* Assume this may touch any SSBO. This is the same we do for other
3905 * UBO/SSBO accesses with non-constant surface.
3907 brw_mark_surface_used(prog_data
,
3908 stage_prog_data
->binding_table
.ssbo_start
+
3909 nir
->info
.num_ssbos
- 1);
3912 fs_reg offset
= get_nir_src(instr
->src
[1]);
3913 fs_reg data1
= get_nir_src(instr
->src
[2]);
3915 if (op
== BRW_AOP_CMPWR
)
3916 data2
= get_nir_src(instr
->src
[3]);
3918 /* Emit the actual atomic operation operation */
3920 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
3922 1 /* dims */, 1 /* rsize */,
3924 BRW_PREDICATE_NONE
);
3925 dest
.type
= atomic_result
.type
;
3926 bld
.MOV(dest
, atomic_result
);
3930 fs_visitor::nir_emit_shared_atomic(const fs_builder
&bld
,
3931 int op
, nir_intrinsic_instr
*instr
)
3934 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3935 dest
= get_nir_dest(instr
->dest
);
3937 fs_reg surface
= brw_imm_ud(GEN7_BTI_SLM
);
3938 fs_reg offset
= get_nir_src(instr
->src
[0]);
3939 fs_reg data1
= get_nir_src(instr
->src
[1]);
3941 if (op
== BRW_AOP_CMPWR
)
3942 data2
= get_nir_src(instr
->src
[2]);
3944 /* Emit the actual atomic operation operation */
3946 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
3948 1 /* dims */, 1 /* rsize */,
3950 BRW_PREDICATE_NONE
);
3951 dest
.type
= atomic_result
.type
;
3952 bld
.MOV(dest
, atomic_result
);
3956 fs_visitor::nir_emit_texture(const fs_builder
&bld
, nir_tex_instr
*instr
)
3958 unsigned texture
= instr
->texture_index
;
3959 unsigned sampler
= instr
->sampler_index
;
3961 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
3963 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture
);
3964 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(sampler
);
3966 int lod_components
= 0;
3968 /* The hardware requires a LOD for buffer textures */
3969 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
3970 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_d(0);
3972 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
3973 fs_reg src
= get_nir_src(instr
->src
[i
].src
);
3974 switch (instr
->src
[i
].src_type
) {
3975 case nir_tex_src_bias
:
3976 srcs
[TEX_LOGICAL_SRC_LOD
] =
3977 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
3979 case nir_tex_src_comparitor
:
3980 srcs
[TEX_LOGICAL_SRC_SHADOW_C
] = retype(src
, BRW_REGISTER_TYPE_F
);
3982 case nir_tex_src_coord
:
3983 switch (instr
->op
) {
3985 case nir_texop_txf_ms
:
3986 case nir_texop_txf_ms_mcs
:
3987 case nir_texop_samples_identical
:
3988 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_D
);
3991 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_F
);
3995 case nir_tex_src_ddx
:
3996 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
3997 lod_components
= nir_tex_instr_src_size(instr
, i
);
3999 case nir_tex_src_ddy
:
4000 srcs
[TEX_LOGICAL_SRC_LOD2
] = retype(src
, BRW_REGISTER_TYPE_F
);
4002 case nir_tex_src_lod
:
4003 switch (instr
->op
) {
4005 srcs
[TEX_LOGICAL_SRC_LOD
] =
4006 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_UD
);
4009 srcs
[TEX_LOGICAL_SRC_LOD
] =
4010 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_D
);
4013 srcs
[TEX_LOGICAL_SRC_LOD
] =
4014 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
4018 case nir_tex_src_ms_index
:
4019 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = retype(src
, BRW_REGISTER_TYPE_UD
);
4022 case nir_tex_src_offset
: {
4023 nir_const_value
*const_offset
=
4024 nir_src_as_const_value(instr
->src
[i
].src
);
4026 unsigned header_bits
= brw_texture_offset(const_offset
->i32
, 3);
4027 if (header_bits
!= 0)
4028 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
] = brw_imm_ud(header_bits
);
4030 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
] =
4031 retype(src
, BRW_REGISTER_TYPE_D
);
4036 case nir_tex_src_projector
:
4037 unreachable("should be lowered");
4039 case nir_tex_src_texture_offset
: {
4040 /* Figure out the highest possible texture index and mark it as used */
4041 uint32_t max_used
= texture
+ instr
->texture_array_size
- 1;
4042 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
< 8) {
4043 max_used
+= stage_prog_data
->binding_table
.gather_texture_start
;
4045 max_used
+= stage_prog_data
->binding_table
.texture_start
;
4047 brw_mark_surface_used(prog_data
, max_used
);
4049 /* Emit code to evaluate the actual indexing expression */
4050 fs_reg tmp
= vgrf(glsl_type::uint_type
);
4051 bld
.ADD(tmp
, src
, brw_imm_ud(texture
));
4052 srcs
[TEX_LOGICAL_SRC_SURFACE
] = bld
.emit_uniformize(tmp
);
4056 case nir_tex_src_sampler_offset
: {
4057 /* Emit code to evaluate the actual indexing expression */
4058 fs_reg tmp
= vgrf(glsl_type::uint_type
);
4059 bld
.ADD(tmp
, src
, brw_imm_ud(sampler
));
4060 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = bld
.emit_uniformize(tmp
);
4064 case nir_tex_src_ms_mcs
:
4065 assert(instr
->op
== nir_texop_txf_ms
);
4066 srcs
[TEX_LOGICAL_SRC_MCS
] = retype(src
, BRW_REGISTER_TYPE_D
);
4069 case nir_tex_src_plane
: {
4070 nir_const_value
*const_plane
=
4071 nir_src_as_const_value(instr
->src
[i
].src
);
4072 const uint32_t plane
= const_plane
->u32
[0];
4073 const uint32_t texture_index
=
4074 instr
->texture_index
+
4075 stage_prog_data
->binding_table
.plane_start
[plane
] -
4076 stage_prog_data
->binding_table
.texture_start
;
4078 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture_index
);
4083 unreachable("unknown texture source");
4087 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BAD_FILE
&&
4088 (instr
->op
== nir_texop_txf_ms
||
4089 instr
->op
== nir_texop_samples_identical
)) {
4090 if (devinfo
->gen
>= 7 &&
4091 key_tex
->compressed_multisample_layout_mask
& (1 << texture
)) {
4092 srcs
[TEX_LOGICAL_SRC_MCS
] =
4093 emit_mcs_fetch(srcs
[TEX_LOGICAL_SRC_COORDINATE
],
4094 instr
->coord_components
,
4095 srcs
[TEX_LOGICAL_SRC_SURFACE
]);
4097 srcs
[TEX_LOGICAL_SRC_MCS
] = brw_imm_ud(0u);
4101 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(instr
->coord_components
);
4102 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(lod_components
);
4104 if (instr
->op
== nir_texop_query_levels
) {
4105 /* textureQueryLevels() is implemented in terms of TXS so we need to
4106 * pass a valid LOD argument.
4108 assert(srcs
[TEX_LOGICAL_SRC_LOD
].file
== BAD_FILE
);
4109 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_ud(0u);
4113 switch (instr
->op
) {
4115 opcode
= SHADER_OPCODE_TEX_LOGICAL
;
4118 opcode
= FS_OPCODE_TXB_LOGICAL
;
4121 opcode
= SHADER_OPCODE_TXL_LOGICAL
;
4124 opcode
= SHADER_OPCODE_TXD_LOGICAL
;
4127 opcode
= SHADER_OPCODE_TXF_LOGICAL
;
4129 case nir_texop_txf_ms
:
4130 if ((key_tex
->msaa_16
& (1 << sampler
)))
4131 opcode
= SHADER_OPCODE_TXF_CMS_W_LOGICAL
;
4133 opcode
= SHADER_OPCODE_TXF_CMS_LOGICAL
;
4135 case nir_texop_txf_ms_mcs
:
4136 opcode
= SHADER_OPCODE_TXF_MCS_LOGICAL
;
4138 case nir_texop_query_levels
:
4140 opcode
= SHADER_OPCODE_TXS_LOGICAL
;
4143 opcode
= SHADER_OPCODE_LOD_LOGICAL
;
4146 if (srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
!= BAD_FILE
&&
4147 srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
!= IMM
)
4148 opcode
= SHADER_OPCODE_TG4_OFFSET_LOGICAL
;
4150 opcode
= SHADER_OPCODE_TG4_LOGICAL
;
4152 case nir_texop_texture_samples
:
4153 opcode
= SHADER_OPCODE_SAMPLEINFO_LOGICAL
;
4155 case nir_texop_samples_identical
: {
4156 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
4158 /* If mcs is an immediate value, it means there is no MCS. In that case
4159 * just return false.
4161 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BRW_IMMEDIATE_VALUE
) {
4162 bld
.MOV(dst
, brw_imm_ud(0u));
4163 } else if ((key_tex
->msaa_16
& (1 << sampler
))) {
4164 fs_reg tmp
= vgrf(glsl_type::uint_type
);
4165 bld
.OR(tmp
, srcs
[TEX_LOGICAL_SRC_MCS
],
4166 offset(srcs
[TEX_LOGICAL_SRC_MCS
], bld
, 1));
4167 bld
.CMP(dst
, tmp
, brw_imm_ud(0u), BRW_CONDITIONAL_EQ
);
4169 bld
.CMP(dst
, srcs
[TEX_LOGICAL_SRC_MCS
], brw_imm_ud(0u),
4170 BRW_CONDITIONAL_EQ
);
4175 unreachable("unknown texture opcode");
4178 fs_reg dst
= bld
.vgrf(brw_type_for_nir_type(instr
->dest_type
), 4);
4179 fs_inst
*inst
= bld
.emit(opcode
, dst
, srcs
, ARRAY_SIZE(srcs
));
4181 const unsigned dest_size
= nir_tex_instr_dest_size(instr
);
4182 if (devinfo
->gen
>= 9 &&
4183 instr
->op
!= nir_texop_tg4
&& instr
->op
!= nir_texop_query_levels
) {
4184 unsigned write_mask
= instr
->dest
.is_ssa
?
4185 nir_ssa_def_components_read(&instr
->dest
.ssa
):
4186 (1 << dest_size
) - 1;
4187 assert(write_mask
!= 0); /* dead code should have been eliminated */
4188 inst
->regs_written
= _mesa_fls(write_mask
) * dispatch_width
/ 8;
4190 inst
->regs_written
= 4 * dispatch_width
/ 8;
4193 if (srcs
[TEX_LOGICAL_SRC_SHADOW_C
].file
!= BAD_FILE
)
4194 inst
->shadow_compare
= true;
4196 if (srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].file
== IMM
)
4197 inst
->offset
= srcs
[TEX_LOGICAL_SRC_OFFSET_VALUE
].ud
;
4199 if (instr
->op
== nir_texop_tg4
) {
4200 if (instr
->component
== 1 &&
4201 key_tex
->gather_channel_quirk_mask
& (1 << texture
)) {
4202 /* gather4 sampler is broken for green channel on RG32F --
4203 * we must ask for blue instead.
4205 inst
->offset
|= 2 << 16;
4207 inst
->offset
|= instr
->component
<< 16;
4210 if (devinfo
->gen
== 6)
4211 emit_gen6_gather_wa(key_tex
->gen6_gather_wa
[texture
], dst
);
4215 for (unsigned i
= 0; i
< dest_size
; i
++)
4216 nir_dest
[i
] = offset(dst
, bld
, i
);
4218 bool is_cube_array
= instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
&&
4221 if (instr
->op
== nir_texop_query_levels
) {
4222 /* # levels is in .w */
4223 nir_dest
[0] = offset(dst
, bld
, 3);
4224 } else if (instr
->op
== nir_texop_txs
&& dest_size
>= 3 &&
4225 (devinfo
->gen
< 7 || is_cube_array
)) {
4226 fs_reg depth
= offset(dst
, bld
, 2);
4227 fs_reg fixed_depth
= vgrf(glsl_type::int_type
);
4229 if (is_cube_array
) {
4230 /* fixup #layers for cube map arrays */
4231 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, fixed_depth
, depth
, brw_imm_d(6));
4232 } else if (devinfo
->gen
< 7) {
4233 /* Gen4-6 return 0 instead of 1 for single layer surfaces. */
4234 bld
.emit_minmax(fixed_depth
, depth
, brw_imm_d(1), BRW_CONDITIONAL_GE
);
4237 nir_dest
[2] = fixed_depth
;
4240 bld
.LOAD_PAYLOAD(get_nir_dest(instr
->dest
), nir_dest
, dest_size
, 0);
4244 fs_visitor::nir_emit_jump(const fs_builder
&bld
, nir_jump_instr
*instr
)
4246 switch (instr
->type
) {
4247 case nir_jump_break
:
4248 bld
.emit(BRW_OPCODE_BREAK
);
4250 case nir_jump_continue
:
4251 bld
.emit(BRW_OPCODE_CONTINUE
);
4253 case nir_jump_return
:
4255 unreachable("unknown jump");
4260 * This helper takes the result of a load operation that reads 32-bit elements
4268 * and shuffles the data to get this:
4275 * Which is exactly what we want if the load is reading 64-bit components
4276 * like doubles, where x represents the low 32-bit of the x double component
4277 * and y represents the high 32-bit of the x double component (likewise with
4278 * z and w for double component y). The parameter @components represents
4279 * the number of 64-bit components present in @src. This would typically be
4280 * 2 at most, since we can only fit 2 double elements in the result of a
4283 * Notice that @dst and @src can be the same register.
4286 shuffle_32bit_load_result_to_64bit_data(const fs_builder
&bld
,
4289 uint32_t components
)
4291 assert(type_sz(src
.type
) == 4);
4292 assert(type_sz(dst
.type
) == 8);
4294 /* A temporary that we will use to shuffle the 32-bit data of each
4295 * component in the vector into valid 64-bit data. We can't write directly
4296 * to dst because dst can be (and would usually be) the same as src
4297 * and in that case the first MOV in the loop below would overwrite the
4298 * data read in the second MOV.
4300 fs_reg tmp
= bld
.vgrf(dst
.type
);
4302 for (unsigned i
= 0; i
< components
; i
++) {
4303 const fs_reg component_i
= offset(src
, bld
, 2 * i
);
4305 bld
.MOV(subscript(tmp
, src
.type
, 0), component_i
);
4306 bld
.MOV(subscript(tmp
, src
.type
, 1), offset(component_i
, bld
, 1));
4308 bld
.MOV(offset(dst
, bld
, i
), tmp
);
4313 * This helper does the inverse operation of
4314 * SHUFFLE_32BIT_LOAD_RESULT_TO_64BIT_DATA.
4316 * We need to do this when we are going to use untyped write messsages that
4317 * operate with 32-bit components in order to arrange our 64-bit data to be
4318 * in the expected layout.
4320 * Notice that callers of this function, unlike in the case of the inverse
4321 * operation, would typically need to call this with dst and src being
4322 * different registers, since they would otherwise corrupt the original
4323 * 64-bit data they are about to write. Because of this the function checks
4324 * that the src and dst regions involved in the operation do not overlap.
4327 shuffle_64bit_data_for_32bit_write(const fs_builder
&bld
,
4330 uint32_t components
)
4332 assert(type_sz(src
.type
) == 8);
4333 assert(type_sz(dst
.type
) == 4);
4335 assert(!src
.in_range(dst
, 2 * components
* bld
.dispatch_width() / 8));
4337 for (unsigned i
= 0; i
< components
; i
++) {
4338 const fs_reg component_i
= offset(src
, bld
, i
);
4339 bld
.MOV(offset(dst
, bld
, 2 * i
), subscript(component_i
, dst
.type
, 0));
4340 bld
.MOV(offset(dst
, bld
, 2 * i
+ 1), subscript(component_i
, dst
.type
, 1));