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"
28 #include "nir_search_helpers.h"
29 #include "util/u_math.h"
30 #include "util/bitscan.h"
35 fs_visitor::emit_nir_code()
37 emit_shader_float_controls_execution_mode();
39 /* emit the arrays used for inputs and outputs - load/store intrinsics will
40 * be converted to reads/writes of these arrays
44 nir_emit_system_values();
46 nir_emit_impl(nir_shader_get_entrypoint((nir_shader
*)nir
));
50 fs_visitor::nir_setup_outputs()
52 if (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_FRAGMENT
)
55 unsigned vec4s
[VARYING_SLOT_TESS_MAX
] = { 0, };
57 /* Calculate the size of output registers in a separate pass, before
58 * allocating them. With ARB_enhanced_layouts, multiple output variables
59 * may occupy the same slot, but have different type sizes.
61 nir_foreach_variable(var
, &nir
->outputs
) {
62 const int loc
= var
->data
.driver_location
;
63 const unsigned var_vec4s
=
64 var
->data
.compact
? DIV_ROUND_UP(glsl_get_length(var
->type
), 4)
65 : type_size_vec4(var
->type
, true);
66 vec4s
[loc
] = MAX2(vec4s
[loc
], var_vec4s
);
69 for (unsigned loc
= 0; loc
< ARRAY_SIZE(vec4s
);) {
70 if (vec4s
[loc
] == 0) {
75 unsigned reg_size
= vec4s
[loc
];
77 /* Check if there are any ranges that start within this range and extend
78 * past it. If so, include them in this allocation.
80 for (unsigned i
= 1; i
< reg_size
; i
++)
81 reg_size
= MAX2(vec4s
[i
+ loc
] + i
, reg_size
);
83 fs_reg reg
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4 * reg_size
);
84 for (unsigned i
= 0; i
< reg_size
; i
++)
85 outputs
[loc
+ i
] = offset(reg
, bld
, 4 * i
);
92 fs_visitor::nir_setup_uniforms()
94 /* Only the first compile gets to set up uniforms. */
95 if (push_constant_loc
) {
96 assert(pull_constant_loc
);
100 uniforms
= nir
->num_uniforms
/ 4;
102 if (stage
== MESA_SHADER_COMPUTE
) {
103 /* Add a uniform for the thread local id. It must be the last uniform
106 assert(uniforms
== prog_data
->nr_params
);
107 uint32_t *param
= brw_stage_prog_data_add_params(prog_data
, 1);
108 *param
= BRW_PARAM_BUILTIN_SUBGROUP_ID
;
109 subgroup_id
= fs_reg(UNIFORM
, uniforms
++, BRW_REGISTER_TYPE_UD
);
114 emit_system_values_block(nir_block
*block
, fs_visitor
*v
)
118 nir_foreach_instr(instr
, block
) {
119 if (instr
->type
!= nir_instr_type_intrinsic
)
122 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
123 switch (intrin
->intrinsic
) {
124 case nir_intrinsic_load_vertex_id
:
125 case nir_intrinsic_load_base_vertex
:
126 unreachable("should be lowered by nir_lower_system_values().");
128 case nir_intrinsic_load_vertex_id_zero_base
:
129 case nir_intrinsic_load_is_indexed_draw
:
130 case nir_intrinsic_load_first_vertex
:
131 case nir_intrinsic_load_instance_id
:
132 case nir_intrinsic_load_base_instance
:
133 case nir_intrinsic_load_draw_id
:
134 unreachable("should be lowered by brw_nir_lower_vs_inputs().");
136 case nir_intrinsic_load_invocation_id
:
137 if (v
->stage
== MESA_SHADER_TESS_CTRL
)
139 assert(v
->stage
== MESA_SHADER_GEOMETRY
);
140 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
141 if (reg
->file
== BAD_FILE
) {
142 const fs_builder abld
= v
->bld
.annotate("gl_InvocationID", NULL
);
143 fs_reg
g1(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
144 fs_reg iid
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
145 abld
.SHR(iid
, g1
, brw_imm_ud(27u));
150 case nir_intrinsic_load_sample_pos
:
151 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
152 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
153 if (reg
->file
== BAD_FILE
)
154 *reg
= *v
->emit_samplepos_setup();
157 case nir_intrinsic_load_sample_id
:
158 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
159 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
160 if (reg
->file
== BAD_FILE
)
161 *reg
= *v
->emit_sampleid_setup();
164 case nir_intrinsic_load_sample_mask_in
:
165 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
166 assert(v
->devinfo
->gen
>= 7);
167 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
168 if (reg
->file
== BAD_FILE
)
169 *reg
= *v
->emit_samplemaskin_setup();
172 case nir_intrinsic_load_work_group_id
:
173 assert(v
->stage
== MESA_SHADER_COMPUTE
);
174 reg
= &v
->nir_system_values
[SYSTEM_VALUE_WORK_GROUP_ID
];
175 if (reg
->file
== BAD_FILE
)
176 *reg
= *v
->emit_cs_work_group_id_setup();
179 case nir_intrinsic_load_helper_invocation
:
180 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
181 reg
= &v
->nir_system_values
[SYSTEM_VALUE_HELPER_INVOCATION
];
182 if (reg
->file
== BAD_FILE
) {
183 const fs_builder abld
=
184 v
->bld
.annotate("gl_HelperInvocation", NULL
);
186 /* On Gen6+ (gl_HelperInvocation is only exposed on Gen7+) the
187 * pixel mask is in g1.7 of the thread payload.
189 * We move the per-channel pixel enable bit to the low bit of each
190 * channel by shifting the byte containing the pixel mask by the
191 * vector immediate 0x76543210UV.
193 * The region of <1,8,0> reads only 1 byte (the pixel masks for
194 * subspans 0 and 1) in SIMD8 and an additional byte (the pixel
195 * masks for 2 and 3) in SIMD16.
197 fs_reg shifted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
199 for (unsigned i
= 0; i
< DIV_ROUND_UP(v
->dispatch_width
, 16); i
++) {
200 const fs_builder hbld
= abld
.group(MIN2(16, v
->dispatch_width
), i
);
201 hbld
.SHR(offset(shifted
, hbld
, i
),
202 stride(retype(brw_vec1_grf(1 + i
, 7),
203 BRW_REGISTER_TYPE_UB
),
205 brw_imm_v(0x76543210));
208 /* A set bit in the pixel mask means the channel is enabled, but
209 * that is the opposite of gl_HelperInvocation so we need to invert
212 * The negate source-modifier bit of logical instructions on Gen8+
213 * performs 1's complement negation, so we can use that instead of
216 fs_reg inverted
= negate(shifted
);
217 if (v
->devinfo
->gen
< 8) {
218 inverted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
219 abld
.NOT(inverted
, shifted
);
222 /* We then resolve the 0/1 result to 0/~0 boolean values by ANDing
223 * with 1 and negating.
225 fs_reg anded
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
226 abld
.AND(anded
, inverted
, brw_imm_uw(1));
228 fs_reg dst
= abld
.vgrf(BRW_REGISTER_TYPE_D
, 1);
229 abld
.MOV(dst
, negate(retype(anded
, BRW_REGISTER_TYPE_D
)));
243 fs_visitor::nir_emit_system_values()
245 nir_system_values
= ralloc_array(mem_ctx
, fs_reg
, SYSTEM_VALUE_MAX
);
246 for (unsigned i
= 0; i
< SYSTEM_VALUE_MAX
; i
++) {
247 nir_system_values
[i
] = fs_reg();
250 /* Always emit SUBGROUP_INVOCATION. Dead code will clean it up if we
251 * never end up using it.
254 const fs_builder abld
= bld
.annotate("gl_SubgroupInvocation", NULL
);
255 fs_reg
®
= nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
];
256 reg
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
258 const fs_builder allbld8
= abld
.group(8, 0).exec_all();
259 allbld8
.MOV(reg
, brw_imm_v(0x76543210));
260 if (dispatch_width
> 8)
261 allbld8
.ADD(byte_offset(reg
, 16), reg
, brw_imm_uw(8u));
262 if (dispatch_width
> 16) {
263 const fs_builder allbld16
= abld
.group(16, 0).exec_all();
264 allbld16
.ADD(byte_offset(reg
, 32), reg
, brw_imm_uw(16u));
268 nir_function_impl
*impl
= nir_shader_get_entrypoint((nir_shader
*)nir
);
269 nir_foreach_block(block
, impl
)
270 emit_system_values_block(block
, this);
274 * Returns a type based on a reference_type (word, float, half-float) and a
277 * Reference BRW_REGISTER_TYPE are HF,F,DF,W,D,UW,UD.
279 * @FIXME: 64-bit return types are always DF on integer types to maintain
280 * compability with uses of DF previously to the introduction of int64
284 brw_reg_type_from_bit_size(const unsigned bit_size
,
285 const brw_reg_type reference_type
)
287 switch(reference_type
) {
288 case BRW_REGISTER_TYPE_HF
:
289 case BRW_REGISTER_TYPE_F
:
290 case BRW_REGISTER_TYPE_DF
:
293 return BRW_REGISTER_TYPE_HF
;
295 return BRW_REGISTER_TYPE_F
;
297 return BRW_REGISTER_TYPE_DF
;
299 unreachable("Invalid bit size");
301 case BRW_REGISTER_TYPE_B
:
302 case BRW_REGISTER_TYPE_W
:
303 case BRW_REGISTER_TYPE_D
:
304 case BRW_REGISTER_TYPE_Q
:
307 return BRW_REGISTER_TYPE_B
;
309 return BRW_REGISTER_TYPE_W
;
311 return BRW_REGISTER_TYPE_D
;
313 return BRW_REGISTER_TYPE_Q
;
315 unreachable("Invalid bit size");
317 case BRW_REGISTER_TYPE_UB
:
318 case BRW_REGISTER_TYPE_UW
:
319 case BRW_REGISTER_TYPE_UD
:
320 case BRW_REGISTER_TYPE_UQ
:
323 return BRW_REGISTER_TYPE_UB
;
325 return BRW_REGISTER_TYPE_UW
;
327 return BRW_REGISTER_TYPE_UD
;
329 return BRW_REGISTER_TYPE_UQ
;
331 unreachable("Invalid bit size");
334 unreachable("Unknown type");
339 fs_visitor::nir_emit_impl(nir_function_impl
*impl
)
341 nir_locals
= ralloc_array(mem_ctx
, fs_reg
, impl
->reg_alloc
);
342 for (unsigned i
= 0; i
< impl
->reg_alloc
; i
++) {
343 nir_locals
[i
] = fs_reg();
346 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
347 unsigned array_elems
=
348 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
349 unsigned size
= array_elems
* reg
->num_components
;
350 const brw_reg_type reg_type
= reg
->bit_size
== 8 ? BRW_REGISTER_TYPE_B
:
351 brw_reg_type_from_bit_size(reg
->bit_size
, BRW_REGISTER_TYPE_F
);
352 nir_locals
[reg
->index
] = bld
.vgrf(reg_type
, size
);
355 nir_ssa_values
= reralloc(mem_ctx
, nir_ssa_values
, fs_reg
,
358 nir_emit_cf_list(&impl
->body
);
362 fs_visitor::nir_emit_cf_list(exec_list
*list
)
364 exec_list_validate(list
);
365 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
366 switch (node
->type
) {
368 nir_emit_if(nir_cf_node_as_if(node
));
371 case nir_cf_node_loop
:
372 nir_emit_loop(nir_cf_node_as_loop(node
));
375 case nir_cf_node_block
:
376 nir_emit_block(nir_cf_node_as_block(node
));
380 unreachable("Invalid CFG node block");
386 fs_visitor::nir_emit_if(nir_if
*if_stmt
)
391 /* If the condition has the form !other_condition, use other_condition as
392 * the source, but invert the predicate on the if instruction.
394 nir_alu_instr
*cond
= nir_src_as_alu_instr(if_stmt
->condition
);
395 if (cond
!= NULL
&& cond
->op
== nir_op_inot
) {
396 assert(!cond
->src
[0].negate
);
397 assert(!cond
->src
[0].abs
);
400 cond_reg
= get_nir_src(cond
->src
[0].src
);
403 cond_reg
= get_nir_src(if_stmt
->condition
);
406 /* first, put the condition into f0 */
407 fs_inst
*inst
= bld
.MOV(bld
.null_reg_d(),
408 retype(cond_reg
, BRW_REGISTER_TYPE_D
));
409 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
411 bld
.IF(BRW_PREDICATE_NORMAL
)->predicate_inverse
= invert
;
413 nir_emit_cf_list(&if_stmt
->then_list
);
415 if (!nir_cf_list_is_empty_block(&if_stmt
->else_list
)) {
416 bld
.emit(BRW_OPCODE_ELSE
);
417 nir_emit_cf_list(&if_stmt
->else_list
);
420 bld
.emit(BRW_OPCODE_ENDIF
);
422 if (devinfo
->gen
< 7)
423 limit_dispatch_width(16, "Non-uniform control flow unsupported "
428 fs_visitor::nir_emit_loop(nir_loop
*loop
)
430 bld
.emit(BRW_OPCODE_DO
);
432 nir_emit_cf_list(&loop
->body
);
434 bld
.emit(BRW_OPCODE_WHILE
);
436 if (devinfo
->gen
< 7)
437 limit_dispatch_width(16, "Non-uniform control flow unsupported "
442 fs_visitor::nir_emit_block(nir_block
*block
)
444 nir_foreach_instr(instr
, block
) {
445 nir_emit_instr(instr
);
450 fs_visitor::nir_emit_instr(nir_instr
*instr
)
452 const fs_builder abld
= bld
.annotate(NULL
, instr
);
454 switch (instr
->type
) {
455 case nir_instr_type_alu
:
456 nir_emit_alu(abld
, nir_instr_as_alu(instr
), true);
459 case nir_instr_type_deref
:
460 unreachable("All derefs should've been lowered");
463 case nir_instr_type_intrinsic
:
465 case MESA_SHADER_VERTEX
:
466 nir_emit_vs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
468 case MESA_SHADER_TESS_CTRL
:
469 nir_emit_tcs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
471 case MESA_SHADER_TESS_EVAL
:
472 nir_emit_tes_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
474 case MESA_SHADER_GEOMETRY
:
475 nir_emit_gs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
477 case MESA_SHADER_FRAGMENT
:
478 nir_emit_fs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
480 case MESA_SHADER_COMPUTE
:
481 nir_emit_cs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
484 unreachable("unsupported shader stage");
488 case nir_instr_type_tex
:
489 nir_emit_texture(abld
, nir_instr_as_tex(instr
));
492 case nir_instr_type_load_const
:
493 nir_emit_load_const(abld
, nir_instr_as_load_const(instr
));
496 case nir_instr_type_ssa_undef
:
497 /* We create a new VGRF for undefs on every use (by handling
498 * them in get_nir_src()), rather than for each definition.
499 * This helps register coalescing eliminate MOVs from undef.
503 case nir_instr_type_jump
:
504 nir_emit_jump(abld
, nir_instr_as_jump(instr
));
508 unreachable("unknown instruction type");
513 * Recognizes a parent instruction of nir_op_extract_* and changes the type to
517 fs_visitor::optimize_extract_to_float(nir_alu_instr
*instr
,
518 const fs_reg
&result
)
520 if (!instr
->src
[0].src
.is_ssa
||
521 !instr
->src
[0].src
.ssa
->parent_instr
)
524 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
527 nir_alu_instr
*src0
=
528 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
530 if (src0
->op
!= nir_op_extract_u8
&& src0
->op
!= nir_op_extract_u16
&&
531 src0
->op
!= nir_op_extract_i8
&& src0
->op
!= nir_op_extract_i16
)
534 /* If either opcode has source modifiers, bail.
536 * TODO: We can potentially handle source modifiers if both of the opcodes
537 * we're combining are signed integers.
539 if (instr
->src
[0].abs
|| instr
->src
[0].negate
||
540 src0
->src
[0].abs
|| src0
->src
[0].negate
)
543 unsigned element
= nir_src_as_uint(src0
->src
[1].src
);
545 /* Element type to extract.*/
546 const brw_reg_type type
= brw_int_type(
547 src0
->op
== nir_op_extract_u16
|| src0
->op
== nir_op_extract_i16
? 2 : 1,
548 src0
->op
== nir_op_extract_i16
|| src0
->op
== nir_op_extract_i8
);
550 fs_reg op0
= get_nir_src(src0
->src
[0].src
);
551 op0
.type
= brw_type_for_nir_type(devinfo
,
552 (nir_alu_type
)(nir_op_infos
[src0
->op
].input_types
[0] |
553 nir_src_bit_size(src0
->src
[0].src
)));
554 op0
= offset(op0
, bld
, src0
->src
[0].swizzle
[0]);
556 set_saturate(instr
->dest
.saturate
,
557 bld
.MOV(result
, subscript(op0
, type
, element
)));
562 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr
*instr
,
563 const fs_reg
&result
)
565 nir_intrinsic_instr
*src0
= nir_src_as_intrinsic(instr
->src
[0].src
);
566 if (src0
== NULL
|| src0
->intrinsic
!= nir_intrinsic_load_front_face
)
569 if (!nir_src_is_const(instr
->src
[1].src
) ||
570 !nir_src_is_const(instr
->src
[2].src
))
573 const float value1
= nir_src_as_float(instr
->src
[1].src
);
574 const float value2
= nir_src_as_float(instr
->src
[2].src
);
575 if (fabsf(value1
) != 1.0f
|| fabsf(value2
) != 1.0f
)
578 /* nir_opt_algebraic should have gotten rid of bcsel(b, a, a) */
579 assert(value1
== -value2
);
581 fs_reg tmp
= vgrf(glsl_type::int_type
);
583 if (devinfo
->gen
>= 6) {
584 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
585 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
587 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
589 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
590 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
592 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
594 * This negation looks like it's safe in practice, because bits 0:4 will
595 * surely be TRIANGLES
598 if (value1
== -1.0f
) {
602 bld
.OR(subscript(tmp
, BRW_REGISTER_TYPE_W
, 1),
603 g0
, brw_imm_uw(0x3f80));
605 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
606 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
608 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
610 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
611 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
613 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
615 * This negation looks like it's safe in practice, because bits 0:4 will
616 * surely be TRIANGLES
619 if (value1
== -1.0f
) {
623 bld
.OR(tmp
, g1_6
, brw_imm_d(0x3f800000));
625 bld
.AND(retype(result
, BRW_REGISTER_TYPE_D
), tmp
, brw_imm_d(0xbf800000));
631 emit_find_msb_using_lzd(const fs_builder
&bld
,
632 const fs_reg
&result
,
640 /* LZD of an absolute value source almost always does the right
641 * thing. There are two problem values:
643 * * 0x80000000. Since abs(0x80000000) == 0x80000000, LZD returns
644 * 0. However, findMSB(int(0x80000000)) == 30.
646 * * 0xffffffff. Since abs(0xffffffff) == 1, LZD returns
647 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
649 * For a value of zero or negative one, -1 will be returned.
651 * * Negative powers of two. LZD(abs(-(1<<x))) returns x, but
652 * findMSB(-(1<<x)) should return x-1.
654 * For all negative number cases, including 0x80000000 and
655 * 0xffffffff, the correct value is obtained from LZD if instead of
656 * negating the (already negative) value the logical-not is used. A
657 * conditonal logical-not can be achieved in two instructions.
659 temp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
661 bld
.ASR(temp
, src
, brw_imm_d(31));
662 bld
.XOR(temp
, temp
, src
);
665 bld
.LZD(retype(result
, BRW_REGISTER_TYPE_UD
),
666 retype(temp
, BRW_REGISTER_TYPE_UD
));
668 /* LZD counts from the MSB side, while GLSL's findMSB() wants the count
669 * from the LSB side. Subtract the result from 31 to convert the MSB
670 * count into an LSB count. If no bits are set, LZD will return 32.
671 * 31-32 = -1, which is exactly what findMSB() is supposed to return.
673 inst
= bld
.ADD(result
, retype(result
, BRW_REGISTER_TYPE_D
), brw_imm_d(31));
674 inst
->src
[0].negate
= true;
678 brw_rnd_mode_from_nir_op (const nir_op op
) {
680 case nir_op_f2f16_rtz
:
681 return BRW_RND_MODE_RTZ
;
682 case nir_op_f2f16_rtne
:
683 return BRW_RND_MODE_RTNE
;
685 unreachable("Operation doesn't support rounding mode");
690 brw_rnd_mode_from_execution_mode(unsigned execution_mode
)
692 if (nir_has_any_rounding_mode_rtne(execution_mode
))
693 return BRW_RND_MODE_RTNE
;
694 if (nir_has_any_rounding_mode_rtz(execution_mode
))
695 return BRW_RND_MODE_RTZ
;
696 return BRW_RND_MODE_UNSPECIFIED
;
700 fs_visitor::prepare_alu_destination_and_sources(const fs_builder
&bld
,
701 nir_alu_instr
*instr
,
706 need_dest
? get_nir_dest(instr
->dest
.dest
) : bld
.null_reg_ud();
708 result
.type
= brw_type_for_nir_type(devinfo
,
709 (nir_alu_type
)(nir_op_infos
[instr
->op
].output_type
|
710 nir_dest_bit_size(instr
->dest
.dest
)));
712 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
713 op
[i
] = get_nir_src(instr
->src
[i
].src
);
714 op
[i
].type
= brw_type_for_nir_type(devinfo
,
715 (nir_alu_type
)(nir_op_infos
[instr
->op
].input_types
[i
] |
716 nir_src_bit_size(instr
->src
[i
].src
)));
717 op
[i
].abs
= instr
->src
[i
].abs
;
718 op
[i
].negate
= instr
->src
[i
].negate
;
721 /* Move and vecN instrutions may still be vectored. Return the raw,
722 * vectored source and destination so that fs_visitor::nir_emit_alu can
723 * handle it. Other callers should not have to handle these kinds of
736 /* At this point, we have dealt with any instruction that operates on
737 * more than a single channel. Therefore, we can just adjust the source
738 * and destination registers for that channel and emit the instruction.
740 unsigned channel
= 0;
741 if (nir_op_infos
[instr
->op
].output_size
== 0) {
742 /* Since NIR is doing the scalarizing for us, we should only ever see
743 * vectorized operations with a single channel.
745 assert(util_bitcount(instr
->dest
.write_mask
) == 1);
746 channel
= ffs(instr
->dest
.write_mask
) - 1;
748 result
= offset(result
, bld
, channel
);
751 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
752 assert(nir_op_infos
[instr
->op
].input_sizes
[i
] < 2);
753 op
[i
] = offset(op
[i
], bld
, instr
->src
[i
].swizzle
[channel
]);
760 fs_visitor::resolve_inot_sources(const fs_builder
&bld
, nir_alu_instr
*instr
,
763 for (unsigned i
= 0; i
< 2; i
++) {
764 nir_alu_instr
*inot_instr
= nir_src_as_alu_instr(instr
->src
[i
].src
);
766 if (inot_instr
!= NULL
&& inot_instr
->op
== nir_op_inot
&&
767 !inot_instr
->src
[0].abs
&& !inot_instr
->src
[0].negate
) {
768 /* The source of the inot is now the source of instr. */
769 prepare_alu_destination_and_sources(bld
, inot_instr
, &op
[i
], false);
771 assert(!op
[i
].negate
);
774 op
[i
] = resolve_source_modifiers(op
[i
]);
780 fs_visitor::try_emit_b2fi_of_inot(const fs_builder
&bld
,
782 nir_alu_instr
*instr
)
784 if (devinfo
->gen
< 6 || devinfo
->gen
>= 12)
787 nir_alu_instr
*inot_instr
= nir_src_as_alu_instr(instr
->src
[0].src
);
789 if (inot_instr
== NULL
|| inot_instr
->op
!= nir_op_inot
)
792 /* HF is also possible as a destination on BDW+. For nir_op_b2i, the set
793 * of valid size-changing combinations is a bit more complex.
795 * The source restriction is just because I was lazy about generating the
798 if (nir_dest_bit_size(instr
->dest
.dest
) != 32 ||
799 nir_src_bit_size(inot_instr
->src
[0].src
) != 32)
802 /* b2[fi](inot(a)) maps a=0 => 1, a=-1 => 0. Since a can only be 0 or -1,
803 * this is float(1 + a).
807 prepare_alu_destination_and_sources(bld
, inot_instr
, &op
, false);
809 /* Ignore the saturate modifier, if there is one. The result of the
810 * arithmetic can only be 0 or 1, so the clamping will do nothing anyway.
812 bld
.ADD(result
, op
, brw_imm_d(1));
818 * Emit code for nir_op_fsign possibly fused with a nir_op_fmul
820 * If \c instr is not the \c nir_op_fsign, then \c fsign_src is the index of
821 * the source of \c instr that is a \c nir_op_fsign.
824 fs_visitor::emit_fsign(const fs_builder
&bld
, const nir_alu_instr
*instr
,
825 fs_reg result
, fs_reg
*op
, unsigned fsign_src
)
829 assert(instr
->op
== nir_op_fsign
|| instr
->op
== nir_op_fmul
);
830 assert(fsign_src
< nir_op_infos
[instr
->op
].num_inputs
);
832 if (instr
->op
!= nir_op_fsign
) {
833 const nir_alu_instr
*const fsign_instr
=
834 nir_src_as_alu_instr(instr
->src
[fsign_src
].src
);
836 assert(!fsign_instr
->dest
.saturate
);
838 /* op[fsign_src] has the nominal result of the fsign, and op[1 -
839 * fsign_src] has the other multiply source. This must be rearranged so
840 * that op[0] is the source of the fsign op[1] is the other multiply
846 op
[0] = get_nir_src(fsign_instr
->src
[0].src
);
848 const nir_alu_type t
=
849 (nir_alu_type
)(nir_op_infos
[instr
->op
].input_types
[0] |
850 nir_src_bit_size(fsign_instr
->src
[0].src
));
852 op
[0].type
= brw_type_for_nir_type(devinfo
, t
);
853 op
[0].abs
= fsign_instr
->src
[0].abs
;
854 op
[0].negate
= fsign_instr
->src
[0].negate
;
856 unsigned channel
= 0;
857 if (nir_op_infos
[instr
->op
].output_size
== 0) {
858 /* Since NIR is doing the scalarizing for us, we should only ever see
859 * vectorized operations with a single channel.
861 assert(util_bitcount(instr
->dest
.write_mask
) == 1);
862 channel
= ffs(instr
->dest
.write_mask
) - 1;
865 op
[0] = offset(op
[0], bld
, fsign_instr
->src
[0].swizzle
[channel
]);
867 assert(!instr
->dest
.saturate
);
871 /* Straightforward since the source can be assumed to be either strictly
872 * >= 0 or strictly <= 0 depending on the setting of the negate flag.
874 set_condmod(BRW_CONDITIONAL_NZ
, bld
.MOV(result
, op
[0]));
876 if (instr
->op
== nir_op_fsign
) {
877 inst
= (op
[0].negate
)
878 ? bld
.MOV(result
, brw_imm_f(-1.0f
))
879 : bld
.MOV(result
, brw_imm_f(1.0f
));
881 op
[1].negate
= (op
[0].negate
!= op
[1].negate
);
882 inst
= bld
.MOV(result
, op
[1]);
885 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
886 } else if (type_sz(op
[0].type
) == 2) {
887 /* AND(val, 0x8000) gives the sign bit.
889 * Predicated OR ORs 1.0 (0x3c00) with the sign bit if val is not zero.
891 fs_reg zero
= retype(brw_imm_uw(0), BRW_REGISTER_TYPE_HF
);
892 bld
.CMP(bld
.null_reg_f(), op
[0], zero
, BRW_CONDITIONAL_NZ
);
894 op
[0].type
= BRW_REGISTER_TYPE_UW
;
895 result
.type
= BRW_REGISTER_TYPE_UW
;
896 bld
.AND(result
, op
[0], brw_imm_uw(0x8000u
));
898 if (instr
->op
== nir_op_fsign
)
899 inst
= bld
.OR(result
, result
, brw_imm_uw(0x3c00u
));
901 /* Use XOR here to get the result sign correct. */
902 inst
= bld
.XOR(result
, result
, retype(op
[1], BRW_REGISTER_TYPE_UW
));
905 inst
->predicate
= BRW_PREDICATE_NORMAL
;
906 } else if (type_sz(op
[0].type
) == 4) {
907 /* AND(val, 0x80000000) gives the sign bit.
909 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
912 bld
.CMP(bld
.null_reg_f(), op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
914 op
[0].type
= BRW_REGISTER_TYPE_UD
;
915 result
.type
= BRW_REGISTER_TYPE_UD
;
916 bld
.AND(result
, op
[0], brw_imm_ud(0x80000000u
));
918 if (instr
->op
== nir_op_fsign
)
919 inst
= bld
.OR(result
, result
, brw_imm_ud(0x3f800000u
));
921 /* Use XOR here to get the result sign correct. */
922 inst
= bld
.XOR(result
, result
, retype(op
[1], BRW_REGISTER_TYPE_UD
));
925 inst
->predicate
= BRW_PREDICATE_NORMAL
;
927 /* For doubles we do the same but we need to consider:
929 * - 2-src instructions can't operate with 64-bit immediates
930 * - The sign is encoded in the high 32-bit of each DF
931 * - We need to produce a DF result.
934 fs_reg zero
= vgrf(glsl_type::double_type
);
935 bld
.MOV(zero
, setup_imm_df(bld
, 0.0));
936 bld
.CMP(bld
.null_reg_df(), op
[0], zero
, BRW_CONDITIONAL_NZ
);
938 bld
.MOV(result
, zero
);
940 fs_reg r
= subscript(result
, BRW_REGISTER_TYPE_UD
, 1);
941 bld
.AND(r
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1),
942 brw_imm_ud(0x80000000u
));
944 if (instr
->op
== nir_op_fsign
) {
945 set_predicate(BRW_PREDICATE_NORMAL
,
946 bld
.OR(r
, r
, brw_imm_ud(0x3ff00000u
)));
948 /* This could be done better in some cases. If the scale is an
949 * immediate with the low 32-bits all 0, emitting a separate XOR and
950 * OR would allow an algebraic optimization to remove the OR. There
951 * are currently zero instances of fsign(double(x))*IMM in shader-db
952 * or any test suite, so it is hard to care at this time.
954 fs_reg result_int64
= retype(result
, BRW_REGISTER_TYPE_UQ
);
955 inst
= bld
.XOR(result_int64
, result_int64
,
956 retype(op
[1], BRW_REGISTER_TYPE_UQ
));
962 * Deteremine whether sources of a nir_op_fmul can be fused with a nir_op_fsign
964 * Checks the operands of a \c nir_op_fmul to determine whether or not
965 * \c emit_fsign could fuse the multiplication with the \c sign() calculation.
967 * \param instr The multiplication instruction
969 * \param fsign_src The source of \c instr that may or may not be a
973 can_fuse_fmul_fsign(nir_alu_instr
*instr
, unsigned fsign_src
)
975 assert(instr
->op
== nir_op_fmul
);
977 nir_alu_instr
*const fsign_instr
=
978 nir_src_as_alu_instr(instr
->src
[fsign_src
].src
);
982 * 1. instr->src[fsign_src] must be a nir_op_fsign.
983 * 2. The nir_op_fsign can only be used by this multiplication.
984 * 3. The source that is the nir_op_fsign does not have source modifiers.
985 * \c emit_fsign only examines the source modifiers of the source of the
988 * The nir_op_fsign must also not have the saturate modifier, but steps
989 * have already been taken (in nir_opt_algebraic) to ensure that.
991 return fsign_instr
!= NULL
&& fsign_instr
->op
== nir_op_fsign
&&
992 is_used_once(fsign_instr
) &&
993 !instr
->src
[fsign_src
].abs
&& !instr
->src
[fsign_src
].negate
;
997 fs_visitor::nir_emit_alu(const fs_builder
&bld
, nir_alu_instr
*instr
,
1000 struct brw_wm_prog_key
*fs_key
= (struct brw_wm_prog_key
*) this->key
;
1002 unsigned execution_mode
=
1003 bld
.shader
->nir
->info
.float_controls_execution_mode
;
1006 fs_reg result
= prepare_alu_destination_and_sources(bld
, instr
, op
, need_dest
);
1008 switch (instr
->op
) {
1013 fs_reg temp
= result
;
1014 bool need_extra_copy
= false;
1015 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
1016 if (!instr
->src
[i
].src
.is_ssa
&&
1017 instr
->dest
.dest
.reg
.reg
== instr
->src
[i
].src
.reg
.reg
) {
1018 need_extra_copy
= true;
1019 temp
= bld
.vgrf(result
.type
, 4);
1024 for (unsigned i
= 0; i
< 4; i
++) {
1025 if (!(instr
->dest
.write_mask
& (1 << i
)))
1028 if (instr
->op
== nir_op_mov
) {
1029 inst
= bld
.MOV(offset(temp
, bld
, i
),
1030 offset(op
[0], bld
, instr
->src
[0].swizzle
[i
]));
1032 inst
= bld
.MOV(offset(temp
, bld
, i
),
1033 offset(op
[i
], bld
, instr
->src
[i
].swizzle
[0]));
1035 inst
->saturate
= instr
->dest
.saturate
;
1038 /* In this case the source and destination registers were the same,
1039 * so we need to insert an extra set of moves in order to deal with
1042 if (need_extra_copy
) {
1043 for (unsigned i
= 0; i
< 4; i
++) {
1044 if (!(instr
->dest
.write_mask
& (1 << i
)))
1047 bld
.MOV(offset(result
, bld
, i
), offset(temp
, bld
, i
));
1055 if (optimize_extract_to_float(instr
, result
))
1057 inst
= bld
.MOV(result
, op
[0]);
1058 inst
->saturate
= instr
->dest
.saturate
;
1061 case nir_op_f2f16_rtne
:
1062 case nir_op_f2f16_rtz
:
1063 case nir_op_f2f16
: {
1064 brw_rnd_mode rnd
= BRW_RND_MODE_UNSPECIFIED
;
1066 if (nir_op_f2f16
== instr
->op
)
1067 rnd
= brw_rnd_mode_from_execution_mode(execution_mode
);
1069 rnd
= brw_rnd_mode_from_nir_op(instr
->op
);
1071 if (BRW_RND_MODE_UNSPECIFIED
!= rnd
)
1072 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(), brw_imm_d(rnd
));
1074 /* In theory, it would be better to use BRW_OPCODE_F32TO16. Depending
1075 * on the HW gen, it is a special hw opcode or just a MOV, and
1076 * brw_F32TO16 (at brw_eu_emit) would do the work to chose.
1078 * But if we want to use that opcode, we need to provide support on
1079 * different optimizations and lowerings. As right now HF support is
1080 * only for gen8+, it will be better to use directly the MOV, and use
1081 * BRW_OPCODE_F32TO16 when/if we work for HF support on gen7.
1083 assert(type_sz(op
[0].type
) < 8); /* brw_nir_lower_conversions */
1084 inst
= bld
.MOV(result
, op
[0]);
1085 inst
->saturate
= instr
->dest
.saturate
;
1096 if (try_emit_b2fi_of_inot(bld
, result
, instr
))
1098 op
[0].type
= BRW_REGISTER_TYPE_D
;
1099 op
[0].negate
= !op
[0].negate
;
1122 if (result
.type
== BRW_REGISTER_TYPE_B
||
1123 result
.type
== BRW_REGISTER_TYPE_UB
||
1124 result
.type
== BRW_REGISTER_TYPE_HF
)
1125 assert(type_sz(op
[0].type
) < 8); /* brw_nir_lower_conversions */
1127 if (op
[0].type
== BRW_REGISTER_TYPE_B
||
1128 op
[0].type
== BRW_REGISTER_TYPE_UB
||
1129 op
[0].type
== BRW_REGISTER_TYPE_HF
)
1130 assert(type_sz(result
.type
) < 8); /* brw_nir_lower_conversions */
1132 inst
= bld
.MOV(result
, op
[0]);
1133 inst
->saturate
= instr
->dest
.saturate
;
1137 inst
= bld
.MOV(result
, op
[0]);
1138 inst
->saturate
= true;
1143 op
[0].negate
= true;
1144 inst
= bld
.MOV(result
, op
[0]);
1145 if (instr
->op
== nir_op_fneg
)
1146 inst
->saturate
= instr
->dest
.saturate
;
1151 op
[0].negate
= false;
1153 inst
= bld
.MOV(result
, op
[0]);
1154 if (instr
->op
== nir_op_fabs
)
1155 inst
->saturate
= instr
->dest
.saturate
;
1159 if (nir_has_any_rounding_mode_enabled(execution_mode
)) {
1161 brw_rnd_mode_from_execution_mode(execution_mode
);
1162 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(),
1166 if (op
[0].type
== BRW_REGISTER_TYPE_HF
)
1167 assert(type_sz(result
.type
) < 8); /* brw_nir_lower_conversions */
1169 inst
= bld
.MOV(result
, op
[0]);
1170 inst
->saturate
= instr
->dest
.saturate
;
1174 emit_fsign(bld
, instr
, result
, op
, 0);
1178 inst
= bld
.emit(SHADER_OPCODE_RCP
, result
, op
[0]);
1179 inst
->saturate
= instr
->dest
.saturate
;
1183 inst
= bld
.emit(SHADER_OPCODE_EXP2
, result
, op
[0]);
1184 inst
->saturate
= instr
->dest
.saturate
;
1188 inst
= bld
.emit(SHADER_OPCODE_LOG2
, result
, op
[0]);
1189 inst
->saturate
= instr
->dest
.saturate
;
1193 inst
= bld
.emit(SHADER_OPCODE_SIN
, result
, op
[0]);
1194 inst
->saturate
= instr
->dest
.saturate
;
1198 inst
= bld
.emit(SHADER_OPCODE_COS
, result
, op
[0]);
1199 inst
->saturate
= instr
->dest
.saturate
;
1203 if (fs_key
->high_quality_derivatives
) {
1204 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
1206 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
1208 inst
->saturate
= instr
->dest
.saturate
;
1210 case nir_op_fddx_fine
:
1211 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
1212 inst
->saturate
= instr
->dest
.saturate
;
1214 case nir_op_fddx_coarse
:
1215 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
1216 inst
->saturate
= instr
->dest
.saturate
;
1219 if (fs_key
->high_quality_derivatives
) {
1220 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
1222 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
1224 inst
->saturate
= instr
->dest
.saturate
;
1226 case nir_op_fddy_fine
:
1227 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
1228 inst
->saturate
= instr
->dest
.saturate
;
1230 case nir_op_fddy_coarse
:
1231 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
1232 inst
->saturate
= instr
->dest
.saturate
;
1236 if (nir_has_any_rounding_mode_enabled(execution_mode
)) {
1238 brw_rnd_mode_from_execution_mode(execution_mode
);
1239 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(),
1244 inst
= bld
.ADD(result
, op
[0], op
[1]);
1245 inst
->saturate
= instr
->dest
.saturate
;
1248 case nir_op_uadd_sat
:
1249 inst
= bld
.ADD(result
, op
[0], op
[1]);
1250 inst
->saturate
= true;
1254 for (unsigned i
= 0; i
< 2; i
++) {
1255 if (can_fuse_fmul_fsign(instr
, i
)) {
1256 emit_fsign(bld
, instr
, result
, op
, i
);
1261 /* We emit the rounding mode after the previous fsign optimization since
1262 * it won't result in a MUL, but will try to negate the value by other
1265 if (nir_has_any_rounding_mode_enabled(execution_mode
)) {
1267 brw_rnd_mode_from_execution_mode(execution_mode
);
1268 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(),
1272 inst
= bld
.MUL(result
, op
[0], op
[1]);
1273 inst
->saturate
= instr
->dest
.saturate
;
1276 case nir_op_imul_2x32_64
:
1277 case nir_op_umul_2x32_64
:
1278 bld
.MUL(result
, op
[0], op
[1]);
1282 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1283 bld
.MUL(result
, op
[0], op
[1]);
1286 case nir_op_imul_high
:
1287 case nir_op_umul_high
:
1288 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1289 bld
.emit(SHADER_OPCODE_MULH
, result
, op
[0], op
[1]);
1294 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1295 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, result
, op
[0], op
[1]);
1298 case nir_op_uadd_carry
:
1299 unreachable("Should have been lowered by carry_to_arith().");
1301 case nir_op_usub_borrow
:
1302 unreachable("Should have been lowered by borrow_to_arith().");
1306 /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
1307 * appears that our hardware just does the right thing for signed
1310 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1311 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
1315 /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
1316 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
1318 /* Math instructions don't support conditional mod */
1319 inst
= bld
.MOV(bld
.null_reg_d(), result
);
1320 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1322 /* Now, we need to determine if signs of the sources are different.
1323 * When we XOR the sources, the top bit is 0 if they are the same and 1
1324 * if they are different. We can then use a conditional modifier to
1325 * turn that into a predicate. This leads us to an XOR.l instruction.
1327 * Technically, according to the PRM, you're not allowed to use .l on a
1328 * XOR instruction. However, emperical experiments and Curro's reading
1329 * of the simulator source both indicate that it's safe.
1331 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1332 inst
= bld
.XOR(tmp
, op
[0], op
[1]);
1333 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1334 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1336 /* If the result of the initial remainder operation is non-zero and the
1337 * two sources have different signs, add in a copy of op[1] to get the
1338 * final integer modulus value.
1340 inst
= bld
.ADD(result
, result
, op
[1]);
1341 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1348 case nir_op_fne32
: {
1349 fs_reg dest
= result
;
1351 const uint32_t bit_size
= nir_src_bit_size(instr
->src
[0].src
);
1353 dest
= bld
.vgrf(op
[0].type
, 1);
1355 bld
.CMP(dest
, op
[0], op
[1], brw_cmod_for_nir_comparison(instr
->op
));
1357 if (bit_size
> 32) {
1358 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1359 } else if(bit_size
< 32) {
1360 /* When we convert the result to 32-bit we need to be careful and do
1361 * it as a signed conversion to get sign extension (for 32-bit true)
1363 const brw_reg_type src_type
=
1364 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_D
);
1366 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), retype(dest
, src_type
));
1376 case nir_op_ine32
: {
1377 fs_reg dest
= result
;
1379 /* On Gen11 we have an additional issue being that src1 cannot be a byte
1380 * type. So we convert both operands for the comparison.
1383 temp_op
[0] = bld
.fix_byte_src(op
[0]);
1384 temp_op
[1] = bld
.fix_byte_src(op
[1]);
1386 const uint32_t bit_size
= nir_src_bit_size(instr
->src
[0].src
);
1388 dest
= bld
.vgrf(temp_op
[0].type
, 1);
1390 bld
.CMP(dest
, temp_op
[0], temp_op
[1],
1391 brw_cmod_for_nir_comparison(instr
->op
));
1393 if (bit_size
> 32) {
1394 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1395 } else if (bit_size
< 32) {
1396 /* When we convert the result to 32-bit we need to be careful and do
1397 * it as a signed conversion to get sign extension (for 32-bit true)
1399 const brw_reg_type src_type
=
1400 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_D
);
1402 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), retype(dest
, src_type
));
1408 if (devinfo
->gen
>= 8) {
1409 nir_alu_instr
*inot_src_instr
= nir_src_as_alu_instr(instr
->src
[0].src
);
1411 if (inot_src_instr
!= NULL
&&
1412 (inot_src_instr
->op
== nir_op_ior
||
1413 inot_src_instr
->op
== nir_op_ixor
||
1414 inot_src_instr
->op
== nir_op_iand
) &&
1415 !inot_src_instr
->src
[0].abs
&&
1416 !inot_src_instr
->src
[0].negate
&&
1417 !inot_src_instr
->src
[1].abs
&&
1418 !inot_src_instr
->src
[1].negate
) {
1419 /* The sources of the source logical instruction are now the
1420 * sources of the instruction that will be generated.
1422 prepare_alu_destination_and_sources(bld
, inot_src_instr
, op
, false);
1423 resolve_inot_sources(bld
, inot_src_instr
, op
);
1425 /* Smash all of the sources and destination to be signed. This
1426 * doesn't matter for the operation of the instruction, but cmod
1427 * propagation fails on unsigned sources with negation (due to
1428 * fs_inst::can_do_cmod returning false).
1431 brw_type_for_nir_type(devinfo
,
1432 (nir_alu_type
)(nir_type_int
|
1433 nir_dest_bit_size(instr
->dest
.dest
)));
1435 brw_type_for_nir_type(devinfo
,
1436 (nir_alu_type
)(nir_type_int
|
1437 nir_src_bit_size(inot_src_instr
->src
[0].src
)));
1439 brw_type_for_nir_type(devinfo
,
1440 (nir_alu_type
)(nir_type_int
|
1441 nir_src_bit_size(inot_src_instr
->src
[1].src
)));
1443 /* For XOR, only invert one of the sources. Arbitrarily choose
1446 op
[0].negate
= !op
[0].negate
;
1447 if (inot_src_instr
->op
!= nir_op_ixor
)
1448 op
[1].negate
= !op
[1].negate
;
1450 switch (inot_src_instr
->op
) {
1452 bld
.AND(result
, op
[0], op
[1]);
1456 bld
.OR(result
, op
[0], op
[1]);
1460 bld
.XOR(result
, op
[0], op
[1]);
1464 unreachable("impossible opcode");
1467 op
[0] = resolve_source_modifiers(op
[0]);
1469 bld
.NOT(result
, op
[0]);
1472 if (devinfo
->gen
>= 8) {
1473 resolve_inot_sources(bld
, instr
, op
);
1475 bld
.XOR(result
, op
[0], op
[1]);
1478 if (devinfo
->gen
>= 8) {
1479 resolve_inot_sources(bld
, instr
, op
);
1481 bld
.OR(result
, op
[0], op
[1]);
1484 if (devinfo
->gen
>= 8) {
1485 resolve_inot_sources(bld
, instr
, op
);
1487 bld
.AND(result
, op
[0], op
[1]);
1493 case nir_op_b32all_fequal2
:
1494 case nir_op_b32all_iequal2
:
1495 case nir_op_b32all_fequal3
:
1496 case nir_op_b32all_iequal3
:
1497 case nir_op_b32all_fequal4
:
1498 case nir_op_b32all_iequal4
:
1499 case nir_op_b32any_fnequal2
:
1500 case nir_op_b32any_inequal2
:
1501 case nir_op_b32any_fnequal3
:
1502 case nir_op_b32any_inequal3
:
1503 case nir_op_b32any_fnequal4
:
1504 case nir_op_b32any_inequal4
:
1505 unreachable("Lowered by nir_lower_alu_reductions");
1507 case nir_op_fnoise1_1
:
1508 case nir_op_fnoise1_2
:
1509 case nir_op_fnoise1_3
:
1510 case nir_op_fnoise1_4
:
1511 case nir_op_fnoise2_1
:
1512 case nir_op_fnoise2_2
:
1513 case nir_op_fnoise2_3
:
1514 case nir_op_fnoise2_4
:
1515 case nir_op_fnoise3_1
:
1516 case nir_op_fnoise3_2
:
1517 case nir_op_fnoise3_3
:
1518 case nir_op_fnoise3_4
:
1519 case nir_op_fnoise4_1
:
1520 case nir_op_fnoise4_2
:
1521 case nir_op_fnoise4_3
:
1522 case nir_op_fnoise4_4
:
1523 unreachable("not reached: should be handled by lower_noise");
1526 unreachable("not reached: should be handled by ldexp_to_arith()");
1529 inst
= bld
.emit(SHADER_OPCODE_SQRT
, result
, op
[0]);
1530 inst
->saturate
= instr
->dest
.saturate
;
1534 inst
= bld
.emit(SHADER_OPCODE_RSQ
, result
, op
[0]);
1535 inst
->saturate
= instr
->dest
.saturate
;
1539 case nir_op_f2b32
: {
1540 uint32_t bit_size
= nir_src_bit_size(instr
->src
[0].src
);
1541 if (bit_size
== 64) {
1542 /* two-argument instructions can't take 64-bit immediates */
1546 if (instr
->op
== nir_op_f2b32
) {
1547 zero
= vgrf(glsl_type::double_type
);
1548 tmp
= vgrf(glsl_type::double_type
);
1549 bld
.MOV(zero
, setup_imm_df(bld
, 0.0));
1551 zero
= vgrf(glsl_type::int64_t_type
);
1552 tmp
= vgrf(glsl_type::int64_t_type
);
1553 bld
.MOV(zero
, brw_imm_q(0));
1556 /* A SIMD16 execution needs to be split in two instructions, so use
1557 * a vgrf instead of the flag register as dst so instruction splitting
1560 bld
.CMP(tmp
, op
[0], zero
, BRW_CONDITIONAL_NZ
);
1561 bld
.MOV(result
, subscript(tmp
, BRW_REGISTER_TYPE_UD
, 0));
1564 if (bit_size
== 32) {
1565 zero
= instr
->op
== nir_op_f2b32
? brw_imm_f(0.0f
) : brw_imm_d(0);
1567 assert(bit_size
== 16);
1568 zero
= instr
->op
== nir_op_f2b32
?
1569 retype(brw_imm_w(0), BRW_REGISTER_TYPE_HF
) : brw_imm_w(0);
1571 bld
.CMP(result
, op
[0], zero
, BRW_CONDITIONAL_NZ
);
1577 inst
= bld
.RNDZ(result
, op
[0]);
1578 inst
->saturate
= instr
->dest
.saturate
;
1581 case nir_op_fceil
: {
1582 op
[0].negate
= !op
[0].negate
;
1583 fs_reg temp
= vgrf(glsl_type::float_type
);
1584 bld
.RNDD(temp
, op
[0]);
1586 inst
= bld
.MOV(result
, temp
);
1587 inst
->saturate
= instr
->dest
.saturate
;
1591 inst
= bld
.RNDD(result
, op
[0]);
1592 inst
->saturate
= instr
->dest
.saturate
;
1595 inst
= bld
.FRC(result
, op
[0]);
1596 inst
->saturate
= instr
->dest
.saturate
;
1598 case nir_op_fround_even
:
1599 inst
= bld
.RNDE(result
, op
[0]);
1600 inst
->saturate
= instr
->dest
.saturate
;
1603 case nir_op_fquantize2f16
: {
1604 fs_reg tmp16
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1605 fs_reg tmp32
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1606 fs_reg zero
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1608 /* The destination stride must be at least as big as the source stride. */
1609 tmp16
.type
= BRW_REGISTER_TYPE_W
;
1612 /* Check for denormal */
1613 fs_reg abs_src0
= op
[0];
1614 abs_src0
.abs
= true;
1615 bld
.CMP(bld
.null_reg_f(), abs_src0
, brw_imm_f(ldexpf(1.0, -14)),
1617 /* Get the appropriately signed zero */
1618 bld
.AND(retype(zero
, BRW_REGISTER_TYPE_UD
),
1619 retype(op
[0], BRW_REGISTER_TYPE_UD
),
1620 brw_imm_ud(0x80000000));
1621 /* Do the actual F32 -> F16 -> F32 conversion */
1622 bld
.emit(BRW_OPCODE_F32TO16
, tmp16
, op
[0]);
1623 bld
.emit(BRW_OPCODE_F16TO32
, tmp32
, tmp16
);
1624 /* Select that or zero based on normal status */
1625 inst
= bld
.SEL(result
, zero
, tmp32
);
1626 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1627 inst
->saturate
= instr
->dest
.saturate
;
1634 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1635 inst
->saturate
= instr
->dest
.saturate
;
1641 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1642 inst
->saturate
= instr
->dest
.saturate
;
1645 case nir_op_pack_snorm_2x16
:
1646 case nir_op_pack_snorm_4x8
:
1647 case nir_op_pack_unorm_2x16
:
1648 case nir_op_pack_unorm_4x8
:
1649 case nir_op_unpack_snorm_2x16
:
1650 case nir_op_unpack_snorm_4x8
:
1651 case nir_op_unpack_unorm_2x16
:
1652 case nir_op_unpack_unorm_4x8
:
1653 case nir_op_unpack_half_2x16
:
1654 case nir_op_pack_half_2x16
:
1655 unreachable("not reached: should be handled by lower_packing_builtins");
1657 case nir_op_unpack_half_2x16_split_x_flush_to_zero
:
1658 assert(FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
& execution_mode
);
1660 case nir_op_unpack_half_2x16_split_x
:
1661 inst
= bld
.emit(BRW_OPCODE_F16TO32
, result
,
1662 subscript(op
[0], BRW_REGISTER_TYPE_UW
, 0));
1663 inst
->saturate
= instr
->dest
.saturate
;
1666 case nir_op_unpack_half_2x16_split_y_flush_to_zero
:
1667 assert(FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
& execution_mode
);
1669 case nir_op_unpack_half_2x16_split_y
:
1670 inst
= bld
.emit(BRW_OPCODE_F16TO32
, result
,
1671 subscript(op
[0], BRW_REGISTER_TYPE_UW
, 1));
1672 inst
->saturate
= instr
->dest
.saturate
;
1675 case nir_op_pack_64_2x32_split
:
1676 case nir_op_pack_32_2x16_split
:
1677 bld
.emit(FS_OPCODE_PACK
, result
, op
[0], op
[1]);
1680 case nir_op_unpack_64_2x32_split_x
:
1681 case nir_op_unpack_64_2x32_split_y
: {
1682 if (instr
->op
== nir_op_unpack_64_2x32_split_x
)
1683 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 0));
1685 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1));
1689 case nir_op_unpack_32_2x16_split_x
:
1690 case nir_op_unpack_32_2x16_split_y
: {
1691 if (instr
->op
== nir_op_unpack_32_2x16_split_x
)
1692 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UW
, 0));
1694 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UW
, 1));
1699 inst
= bld
.emit(SHADER_OPCODE_POW
, result
, op
[0], op
[1]);
1700 inst
->saturate
= instr
->dest
.saturate
;
1703 case nir_op_bitfield_reverse
:
1704 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1705 bld
.BFREV(result
, op
[0]);
1708 case nir_op_bit_count
:
1709 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1710 bld
.CBIT(result
, op
[0]);
1713 case nir_op_ufind_msb
: {
1714 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1715 emit_find_msb_using_lzd(bld
, result
, op
[0], false);
1719 case nir_op_ifind_msb
: {
1720 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1722 if (devinfo
->gen
< 7) {
1723 emit_find_msb_using_lzd(bld
, result
, op
[0], true);
1725 bld
.FBH(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0]);
1727 /* FBH counts from the MSB side, while GLSL's findMSB() wants the
1728 * count from the LSB side. If FBH didn't return an error
1729 * (0xFFFFFFFF), then subtract the result from 31 to convert the MSB
1730 * count into an LSB count.
1732 bld
.CMP(bld
.null_reg_d(), result
, brw_imm_d(-1), BRW_CONDITIONAL_NZ
);
1734 inst
= bld
.ADD(result
, result
, brw_imm_d(31));
1735 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1736 inst
->src
[0].negate
= true;
1741 case nir_op_find_lsb
:
1742 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1744 if (devinfo
->gen
< 7) {
1745 fs_reg temp
= vgrf(glsl_type::int_type
);
1747 /* (x & -x) generates a value that consists of only the LSB of x.
1748 * For all powers of 2, findMSB(y) == findLSB(y).
1750 fs_reg src
= retype(op
[0], BRW_REGISTER_TYPE_D
);
1751 fs_reg negated_src
= src
;
1753 /* One must be negated, and the other must be non-negated. It
1754 * doesn't matter which is which.
1756 negated_src
.negate
= true;
1759 bld
.AND(temp
, src
, negated_src
);
1760 emit_find_msb_using_lzd(bld
, result
, temp
, false);
1762 bld
.FBL(result
, op
[0]);
1766 case nir_op_ubitfield_extract
:
1767 case nir_op_ibitfield_extract
:
1768 unreachable("should have been lowered");
1771 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1772 bld
.BFE(result
, op
[2], op
[1], op
[0]);
1775 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1776 bld
.BFI1(result
, op
[0], op
[1]);
1779 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1780 bld
.BFI2(result
, op
[0], op
[1], op
[2]);
1783 case nir_op_bitfield_insert
:
1784 unreachable("not reached: should have been lowered");
1787 bld
.SHL(result
, op
[0], op
[1]);
1790 bld
.ASR(result
, op
[0], op
[1]);
1793 bld
.SHR(result
, op
[0], op
[1]);
1797 bld
.ROL(result
, op
[0], op
[1]);
1800 bld
.ROR(result
, op
[0], op
[1]);
1803 case nir_op_pack_half_2x16_split
:
1804 bld
.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, result
, op
[0], op
[1]);
1808 if (nir_has_any_rounding_mode_enabled(execution_mode
)) {
1810 brw_rnd_mode_from_execution_mode(execution_mode
);
1811 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(),
1815 inst
= bld
.MAD(result
, op
[2], op
[1], op
[0]);
1816 inst
->saturate
= instr
->dest
.saturate
;
1820 if (nir_has_any_rounding_mode_enabled(execution_mode
)) {
1822 brw_rnd_mode_from_execution_mode(execution_mode
);
1823 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(),
1827 inst
= bld
.LRP(result
, op
[0], op
[1], op
[2]);
1828 inst
->saturate
= instr
->dest
.saturate
;
1831 case nir_op_b32csel
:
1832 if (optimize_frontfacing_ternary(instr
, result
))
1835 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1836 inst
= bld
.SEL(result
, op
[1], op
[2]);
1837 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1840 case nir_op_extract_u8
:
1841 case nir_op_extract_i8
: {
1842 unsigned byte
= nir_src_as_uint(instr
->src
[1].src
);
1847 * There is no direct conversion from B/UB to Q/UQ or Q/UQ to B/UB.
1848 * Use two instructions and a word or DWord intermediate integer type.
1850 if (nir_dest_bit_size(instr
->dest
.dest
) == 64) {
1851 const brw_reg_type type
= brw_int_type(1, instr
->op
== nir_op_extract_i8
);
1853 if (instr
->op
== nir_op_extract_i8
) {
1854 /* If we need to sign extend, extract to a word first */
1855 fs_reg w_temp
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
1856 bld
.MOV(w_temp
, subscript(op
[0], type
, byte
));
1857 bld
.MOV(result
, w_temp
);
1858 } else if (byte
& 1) {
1859 /* Extract the high byte from the word containing the desired byte
1863 subscript(op
[0], BRW_REGISTER_TYPE_UW
, byte
/ 2),
1866 /* Otherwise use an AND with 0xff and a word type */
1868 subscript(op
[0], BRW_REGISTER_TYPE_UW
, byte
/ 2),
1872 const brw_reg_type type
= brw_int_type(1, instr
->op
== nir_op_extract_i8
);
1873 bld
.MOV(result
, subscript(op
[0], type
, byte
));
1878 case nir_op_extract_u16
:
1879 case nir_op_extract_i16
: {
1880 const brw_reg_type type
= brw_int_type(2, instr
->op
== nir_op_extract_i16
);
1881 unsigned word
= nir_src_as_uint(instr
->src
[1].src
);
1882 bld
.MOV(result
, subscript(op
[0], type
, word
));
1887 unreachable("unhandled instruction");
1890 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1891 * to sign extend the low bit to 0/~0
1893 if (devinfo
->gen
<= 5 &&
1894 !result
.is_null() &&
1895 (instr
->instr
.pass_flags
& BRW_NIR_BOOLEAN_MASK
) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE
) {
1896 fs_reg masked
= vgrf(glsl_type::int_type
);
1897 bld
.AND(masked
, result
, brw_imm_d(1));
1898 masked
.negate
= true;
1899 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), masked
);
1904 fs_visitor::nir_emit_load_const(const fs_builder
&bld
,
1905 nir_load_const_instr
*instr
)
1907 const brw_reg_type reg_type
=
1908 brw_reg_type_from_bit_size(instr
->def
.bit_size
, BRW_REGISTER_TYPE_D
);
1909 fs_reg reg
= bld
.vgrf(reg_type
, instr
->def
.num_components
);
1911 switch (instr
->def
.bit_size
) {
1913 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1914 bld
.MOV(offset(reg
, bld
, i
), setup_imm_b(bld
, instr
->value
[i
].i8
));
1918 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1919 bld
.MOV(offset(reg
, bld
, i
), brw_imm_w(instr
->value
[i
].i16
));
1923 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1924 bld
.MOV(offset(reg
, bld
, i
), brw_imm_d(instr
->value
[i
].i32
));
1928 assert(devinfo
->gen
>= 7);
1929 if (devinfo
->gen
== 7) {
1930 /* We don't get 64-bit integer types until gen8 */
1931 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++) {
1932 bld
.MOV(retype(offset(reg
, bld
, i
), BRW_REGISTER_TYPE_DF
),
1933 setup_imm_df(bld
, instr
->value
[i
].f64
));
1936 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1937 bld
.MOV(offset(reg
, bld
, i
), brw_imm_q(instr
->value
[i
].i64
));
1942 unreachable("Invalid bit size");
1945 nir_ssa_values
[instr
->def
.index
] = reg
;
1949 fs_visitor::get_nir_src(const nir_src
&src
)
1953 if (src
.ssa
->parent_instr
->type
== nir_instr_type_ssa_undef
) {
1954 const brw_reg_type reg_type
=
1955 brw_reg_type_from_bit_size(src
.ssa
->bit_size
, BRW_REGISTER_TYPE_D
);
1956 reg
= bld
.vgrf(reg_type
, src
.ssa
->num_components
);
1958 reg
= nir_ssa_values
[src
.ssa
->index
];
1961 /* We don't handle indirects on locals */
1962 assert(src
.reg
.indirect
== NULL
);
1963 reg
= offset(nir_locals
[src
.reg
.reg
->index
], bld
,
1964 src
.reg
.base_offset
* src
.reg
.reg
->num_components
);
1967 if (nir_src_bit_size(src
) == 64 && devinfo
->gen
== 7) {
1968 /* The only 64-bit type available on gen7 is DF, so use that. */
1969 reg
.type
= BRW_REGISTER_TYPE_DF
;
1971 /* To avoid floating-point denorm flushing problems, set the type by
1972 * default to an integer type - instructions that need floating point
1973 * semantics will set this to F if they need to
1975 reg
.type
= brw_reg_type_from_bit_size(nir_src_bit_size(src
),
1976 BRW_REGISTER_TYPE_D
);
1983 * Return an IMM for constants; otherwise call get_nir_src() as normal.
1985 * This function should not be called on any value which may be 64 bits.
1986 * We could theoretically support 64-bit on gen8+ but we choose not to
1987 * because it wouldn't work in general (no gen7 support) and there are
1988 * enough restrictions in 64-bit immediates that you can't take the return
1989 * value and treat it the same as the result of get_nir_src().
1992 fs_visitor::get_nir_src_imm(const nir_src
&src
)
1994 assert(nir_src_bit_size(src
) == 32);
1995 return nir_src_is_const(src
) ?
1996 fs_reg(brw_imm_d(nir_src_as_int(src
))) : get_nir_src(src
);
2000 fs_visitor::get_nir_dest(const nir_dest
&dest
)
2003 const brw_reg_type reg_type
=
2004 brw_reg_type_from_bit_size(dest
.ssa
.bit_size
,
2005 dest
.ssa
.bit_size
== 8 ?
2006 BRW_REGISTER_TYPE_D
:
2007 BRW_REGISTER_TYPE_F
);
2008 nir_ssa_values
[dest
.ssa
.index
] =
2009 bld
.vgrf(reg_type
, dest
.ssa
.num_components
);
2010 bld
.UNDEF(nir_ssa_values
[dest
.ssa
.index
]);
2011 return nir_ssa_values
[dest
.ssa
.index
];
2013 /* We don't handle indirects on locals */
2014 assert(dest
.reg
.indirect
== NULL
);
2015 return offset(nir_locals
[dest
.reg
.reg
->index
], bld
,
2016 dest
.reg
.base_offset
* dest
.reg
.reg
->num_components
);
2021 fs_visitor::emit_percomp(const fs_builder
&bld
, const fs_inst
&inst
,
2024 for (unsigned i
= 0; i
< 4; i
++) {
2025 if (!((wr_mask
>> i
) & 1))
2028 fs_inst
*new_inst
= new(mem_ctx
) fs_inst(inst
);
2029 new_inst
->dst
= offset(new_inst
->dst
, bld
, i
);
2030 for (unsigned j
= 0; j
< new_inst
->sources
; j
++)
2031 if (new_inst
->src
[j
].file
== VGRF
)
2032 new_inst
->src
[j
] = offset(new_inst
->src
[j
], bld
, i
);
2039 emit_pixel_interpolater_send(const fs_builder
&bld
,
2044 glsl_interp_mode interpolation
)
2046 struct brw_wm_prog_data
*wm_prog_data
=
2047 brw_wm_prog_data(bld
.shader
->stage_prog_data
);
2049 fs_inst
*inst
= bld
.emit(opcode
, dst
, src
, desc
);
2050 /* 2 floats per slot returned */
2051 inst
->size_written
= 2 * dst
.component_size(inst
->exec_size
);
2052 inst
->pi_noperspective
= interpolation
== INTERP_MODE_NOPERSPECTIVE
;
2054 wm_prog_data
->pulls_bary
= true;
2060 * Computes 1 << x, given a D/UD register containing some value x.
2063 intexp2(const fs_builder
&bld
, const fs_reg
&x
)
2065 assert(x
.type
== BRW_REGISTER_TYPE_UD
|| x
.type
== BRW_REGISTER_TYPE_D
);
2067 fs_reg result
= bld
.vgrf(x
.type
, 1);
2068 fs_reg one
= bld
.vgrf(x
.type
, 1);
2070 bld
.MOV(one
, retype(brw_imm_d(1), one
.type
));
2071 bld
.SHL(result
, one
, x
);
2076 fs_visitor::emit_gs_end_primitive(const nir_src
&vertex_count_nir_src
)
2078 assert(stage
== MESA_SHADER_GEOMETRY
);
2080 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
2082 if (gs_compile
->control_data_header_size_bits
== 0)
2085 /* We can only do EndPrimitive() functionality when the control data
2086 * consists of cut bits. Fortunately, the only time it isn't is when the
2087 * output type is points, in which case EndPrimitive() is a no-op.
2089 if (gs_prog_data
->control_data_format
!=
2090 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
) {
2094 /* Cut bits use one bit per vertex. */
2095 assert(gs_compile
->control_data_bits_per_vertex
== 1);
2097 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
2098 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
2100 /* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
2101 * vertex n, 0 otherwise. So all we need to do here is mark bit
2102 * (vertex_count - 1) % 32 in the cut_bits register to indicate that
2103 * EndPrimitive() was called after emitting vertex (vertex_count - 1);
2104 * vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
2106 * Note that if EndPrimitive() is called before emitting any vertices, this
2107 * will cause us to set bit 31 of the control_data_bits register to 1.
2108 * That's fine because:
2110 * - If max_vertices < 32, then vertex number 31 (zero-based) will never be
2111 * output, so the hardware will ignore cut bit 31.
2113 * - If max_vertices == 32, then vertex number 31 is guaranteed to be the
2114 * last vertex, so setting cut bit 31 has no effect (since the primitive
2115 * is automatically ended when the GS terminates).
2117 * - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
2118 * control_data_bits register to 0 when the first vertex is emitted.
2121 const fs_builder abld
= bld
.annotate("end primitive");
2123 /* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
2124 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2125 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
2126 fs_reg mask
= intexp2(abld
, prev_count
);
2127 /* Note: we're relying on the fact that the GEN SHL instruction only pays
2128 * attention to the lower 5 bits of its second source argument, so on this
2129 * architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
2130 * ((vertex_count - 1) % 32).
2132 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
2136 fs_visitor::emit_gs_control_data_bits(const fs_reg
&vertex_count
)
2138 assert(stage
== MESA_SHADER_GEOMETRY
);
2139 assert(gs_compile
->control_data_bits_per_vertex
!= 0);
2141 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
2143 const fs_builder abld
= bld
.annotate("emit control data bits");
2144 const fs_builder fwa_bld
= bld
.exec_all();
2146 /* We use a single UD register to accumulate control data bits (32 bits
2147 * for each of the SIMD8 channels). So we need to write a DWord (32 bits)
2150 * Unfortunately, the URB_WRITE_SIMD8 message uses 128-bit (OWord) offsets.
2151 * We have select a 128-bit group via the Global and Per-Slot Offsets, then
2152 * use the Channel Mask phase to enable/disable which DWord within that
2153 * group to write. (Remember, different SIMD8 channels may have emitted
2154 * different numbers of vertices, so we may need per-slot offsets.)
2156 * Channel masking presents an annoying problem: we may have to replicate
2157 * the data up to 4 times:
2159 * Msg = Handles, Per-Slot Offsets, Channel Masks, Data, Data, Data, Data.
2161 * To avoid penalizing shaders that emit a small number of vertices, we
2162 * can avoid these sometimes: if the size of the control data header is
2163 * <= 128 bits, then there is only 1 OWord. All SIMD8 channels will land
2164 * land in the same 128-bit group, so we can skip per-slot offsets.
2166 * Similarly, if the control data header is <= 32 bits, there is only one
2167 * DWord, so we can skip channel masks.
2169 enum opcode opcode
= SHADER_OPCODE_URB_WRITE_SIMD8
;
2171 fs_reg channel_mask
, per_slot_offset
;
2173 if (gs_compile
->control_data_header_size_bits
> 32) {
2174 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2175 channel_mask
= vgrf(glsl_type::uint_type
);
2178 if (gs_compile
->control_data_header_size_bits
> 128) {
2179 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
;
2180 per_slot_offset
= vgrf(glsl_type::uint_type
);
2183 /* Figure out which DWord we're trying to write to using the formula:
2185 * dword_index = (vertex_count - 1) * bits_per_vertex / 32
2187 * Since bits_per_vertex is a power of two, and is known at compile
2188 * time, this can be optimized to:
2190 * dword_index = (vertex_count - 1) >> (6 - log2(bits_per_vertex))
2192 if (opcode
!= SHADER_OPCODE_URB_WRITE_SIMD8
) {
2193 fs_reg dword_index
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2194 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2195 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
2196 unsigned log2_bits_per_vertex
=
2197 util_last_bit(gs_compile
->control_data_bits_per_vertex
);
2198 abld
.SHR(dword_index
, prev_count
, brw_imm_ud(6u - log2_bits_per_vertex
));
2200 if (per_slot_offset
.file
!= BAD_FILE
) {
2201 /* Set the per-slot offset to dword_index / 4, so that we'll write to
2202 * the appropriate OWord within the control data header.
2204 abld
.SHR(per_slot_offset
, dword_index
, brw_imm_ud(2u));
2207 /* Set the channel masks to 1 << (dword_index % 4), so that we'll
2208 * write to the appropriate DWORD within the OWORD.
2210 fs_reg channel
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2211 fwa_bld
.AND(channel
, dword_index
, brw_imm_ud(3u));
2212 channel_mask
= intexp2(fwa_bld
, channel
);
2213 /* Then the channel masks need to be in bits 23:16. */
2214 fwa_bld
.SHL(channel_mask
, channel_mask
, brw_imm_ud(16u));
2217 /* Store the control data bits in the message payload and send it. */
2219 if (channel_mask
.file
!= BAD_FILE
)
2220 mlen
+= 4; /* channel masks, plus 3 extra copies of the data */
2221 if (per_slot_offset
.file
!= BAD_FILE
)
2224 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2225 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, mlen
);
2227 sources
[i
++] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
2228 if (per_slot_offset
.file
!= BAD_FILE
)
2229 sources
[i
++] = per_slot_offset
;
2230 if (channel_mask
.file
!= BAD_FILE
)
2231 sources
[i
++] = channel_mask
;
2233 sources
[i
++] = this->control_data_bits
;
2236 abld
.LOAD_PAYLOAD(payload
, sources
, mlen
, mlen
);
2237 fs_inst
*inst
= abld
.emit(opcode
, reg_undef
, payload
);
2239 /* We need to increment Global Offset by 256-bits to make room for
2240 * Broadwell's extra "Vertex Count" payload at the beginning of the
2241 * URB entry. Since this is an OWord message, Global Offset is counted
2242 * in 128-bit units, so we must set it to 2.
2244 if (gs_prog_data
->static_vertex_count
== -1)
2249 fs_visitor::set_gs_stream_control_data_bits(const fs_reg
&vertex_count
,
2252 /* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
2254 /* Note: we are calling this *before* increasing vertex_count, so
2255 * this->vertex_count == vertex_count - 1 in the formula above.
2258 /* Stream mode uses 2 bits per vertex */
2259 assert(gs_compile
->control_data_bits_per_vertex
== 2);
2261 /* Must be a valid stream */
2262 assert(stream_id
< MAX_VERTEX_STREAMS
);
2264 /* Control data bits are initialized to 0 so we don't have to set any
2265 * bits when sending vertices to stream 0.
2270 const fs_builder abld
= bld
.annotate("set stream control data bits", NULL
);
2272 /* reg::sid = stream_id */
2273 fs_reg sid
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2274 abld
.MOV(sid
, brw_imm_ud(stream_id
));
2276 /* reg:shift_count = 2 * (vertex_count - 1) */
2277 fs_reg shift_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2278 abld
.SHL(shift_count
, vertex_count
, brw_imm_ud(1u));
2280 /* Note: we're relying on the fact that the GEN SHL instruction only pays
2281 * attention to the lower 5 bits of its second source argument, so on this
2282 * architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
2283 * stream_id << ((2 * (vertex_count - 1)) % 32).
2285 fs_reg mask
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2286 abld
.SHL(mask
, sid
, shift_count
);
2287 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
2291 fs_visitor::emit_gs_vertex(const nir_src
&vertex_count_nir_src
,
2294 assert(stage
== MESA_SHADER_GEOMETRY
);
2296 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
2298 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
2299 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
2301 /* Haswell and later hardware ignores the "Render Stream Select" bits
2302 * from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
2303 * and instead sends all primitives down the pipeline for rasterization.
2304 * If the SOL stage is enabled, "Render Stream Select" is honored and
2305 * primitives bound to non-zero streams are discarded after stream output.
2307 * Since the only purpose of primives sent to non-zero streams is to
2308 * be recorded by transform feedback, we can simply discard all geometry
2309 * bound to these streams when transform feedback is disabled.
2311 if (stream_id
> 0 && !nir
->info
.has_transform_feedback_varyings
)
2314 /* If we're outputting 32 control data bits or less, then we can wait
2315 * until the shader is over to output them all. Otherwise we need to
2316 * output them as we go. Now is the time to do it, since we're about to
2317 * output the vertex_count'th vertex, so it's guaranteed that the
2318 * control data bits associated with the (vertex_count - 1)th vertex are
2321 if (gs_compile
->control_data_header_size_bits
> 32) {
2322 const fs_builder abld
=
2323 bld
.annotate("emit vertex: emit control data bits");
2325 /* Only emit control data bits if we've finished accumulating a batch
2326 * of 32 bits. This is the case when:
2328 * (vertex_count * bits_per_vertex) % 32 == 0
2330 * (in other words, when the last 5 bits of vertex_count *
2331 * bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
2332 * integer n (which is always the case, since bits_per_vertex is
2333 * always 1 or 2), this is equivalent to requiring that the last 5-n
2334 * bits of vertex_count are 0:
2336 * vertex_count & (2^(5-n) - 1) == 0
2338 * 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
2341 * vertex_count & (32 / bits_per_vertex - 1) == 0
2343 * TODO: If vertex_count is an immediate, we could do some of this math
2344 * at compile time...
2347 abld
.AND(bld
.null_reg_d(), vertex_count
,
2348 brw_imm_ud(32u / gs_compile
->control_data_bits_per_vertex
- 1u));
2349 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2351 abld
.IF(BRW_PREDICATE_NORMAL
);
2352 /* If vertex_count is 0, then no control data bits have been
2353 * accumulated yet, so we can skip emitting them.
2355 abld
.CMP(bld
.null_reg_d(), vertex_count
, brw_imm_ud(0u),
2356 BRW_CONDITIONAL_NEQ
);
2357 abld
.IF(BRW_PREDICATE_NORMAL
);
2358 emit_gs_control_data_bits(vertex_count
);
2359 abld
.emit(BRW_OPCODE_ENDIF
);
2361 /* Reset control_data_bits to 0 so we can start accumulating a new
2364 * Note: in the case where vertex_count == 0, this neutralizes the
2365 * effect of any call to EndPrimitive() that the shader may have
2366 * made before outputting its first vertex.
2368 inst
= abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
2369 inst
->force_writemask_all
= true;
2370 abld
.emit(BRW_OPCODE_ENDIF
);
2373 emit_urb_writes(vertex_count
);
2375 /* In stream mode we have to set control data bits for all vertices
2376 * unless we have disabled control data bits completely (which we do
2377 * do for GL_POINTS outputs that don't use streams).
2379 if (gs_compile
->control_data_header_size_bits
> 0 &&
2380 gs_prog_data
->control_data_format
==
2381 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
) {
2382 set_gs_stream_control_data_bits(vertex_count
, stream_id
);
2387 fs_visitor::emit_gs_input_load(const fs_reg
&dst
,
2388 const nir_src
&vertex_src
,
2389 unsigned base_offset
,
2390 const nir_src
&offset_src
,
2391 unsigned num_components
,
2392 unsigned first_component
)
2394 assert(type_sz(dst
.type
) == 4);
2395 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
2396 const unsigned push_reg_count
= gs_prog_data
->base
.urb_read_length
* 8;
2398 /* TODO: figure out push input layout for invocations == 1 */
2399 if (gs_prog_data
->invocations
== 1 &&
2400 nir_src_is_const(offset_src
) && nir_src_is_const(vertex_src
) &&
2401 4 * (base_offset
+ nir_src_as_uint(offset_src
)) < push_reg_count
) {
2402 int imm_offset
= (base_offset
+ nir_src_as_uint(offset_src
)) * 4 +
2403 nir_src_as_uint(vertex_src
) * push_reg_count
;
2404 for (unsigned i
= 0; i
< num_components
; i
++) {
2405 bld
.MOV(offset(dst
, bld
, i
),
2406 fs_reg(ATTR
, imm_offset
+ i
+ first_component
, dst
.type
));
2411 /* Resort to the pull model. Ensure the VUE handles are provided. */
2412 assert(gs_prog_data
->base
.include_vue_handles
);
2414 unsigned first_icp_handle
= gs_prog_data
->include_primitive_id
? 3 : 2;
2415 fs_reg icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2417 if (gs_prog_data
->invocations
== 1) {
2418 if (nir_src_is_const(vertex_src
)) {
2419 /* The vertex index is constant; just select the proper URB handle. */
2421 retype(brw_vec8_grf(first_icp_handle
+ nir_src_as_uint(vertex_src
), 0),
2422 BRW_REGISTER_TYPE_UD
);
2424 /* The vertex index is non-constant. We need to use indirect
2425 * addressing to fetch the proper URB handle.
2427 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
2428 * indicating that channel <n> should read the handle from
2429 * DWord <n>. We convert that to bytes by multiplying by 4.
2431 * Next, we convert the vertex index to bytes by multiplying
2432 * by 32 (shifting by 5), and add the two together. This is
2433 * the final indirect byte offset.
2435 fs_reg sequence
= bld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
2436 fs_reg channel_offsets
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2437 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2438 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2440 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
2441 bld
.MOV(sequence
, fs_reg(brw_imm_v(0x76543210)));
2442 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
2443 bld
.SHL(channel_offsets
, sequence
, brw_imm_ud(2u));
2444 /* Convert vertex_index to bytes (multiply by 32) */
2445 bld
.SHL(vertex_offset_bytes
,
2446 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2448 bld
.ADD(icp_offset_bytes
, vertex_offset_bytes
, channel_offsets
);
2450 /* Use first_icp_handle as the base offset. There is one register
2451 * of URB handles per vertex, so inform the register allocator that
2452 * we might read up to nir->info.gs.vertices_in registers.
2454 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2455 retype(brw_vec8_grf(first_icp_handle
, 0), icp_handle
.type
),
2456 fs_reg(icp_offset_bytes
),
2457 brw_imm_ud(nir
->info
.gs
.vertices_in
* REG_SIZE
));
2460 assert(gs_prog_data
->invocations
> 1);
2462 if (nir_src_is_const(vertex_src
)) {
2463 unsigned vertex
= nir_src_as_uint(vertex_src
);
2464 assert(devinfo
->gen
>= 9 || vertex
<= 5);
2466 retype(brw_vec1_grf(first_icp_handle
+ vertex
/ 8, vertex
% 8),
2467 BRW_REGISTER_TYPE_UD
));
2469 /* The vertex index is non-constant. We need to use indirect
2470 * addressing to fetch the proper URB handle.
2473 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2475 /* Convert vertex_index to bytes (multiply by 4) */
2476 bld
.SHL(icp_offset_bytes
,
2477 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2480 /* Use first_icp_handle as the base offset. There is one DWord
2481 * of URB handles per vertex, so inform the register allocator that
2482 * we might read up to ceil(nir->info.gs.vertices_in / 8) registers.
2484 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2485 retype(brw_vec8_grf(first_icp_handle
, 0), icp_handle
.type
),
2486 fs_reg(icp_offset_bytes
),
2487 brw_imm_ud(DIV_ROUND_UP(nir
->info
.gs
.vertices_in
, 8) *
2493 fs_reg indirect_offset
= get_nir_src(offset_src
);
2495 if (nir_src_is_const(offset_src
)) {
2496 /* Constant indexing - use global offset. */
2497 if (first_component
!= 0) {
2498 unsigned read_components
= num_components
+ first_component
;
2499 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2500 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, icp_handle
);
2501 inst
->size_written
= read_components
*
2502 tmp
.component_size(inst
->exec_size
);
2503 for (unsigned i
= 0; i
< num_components
; i
++) {
2504 bld
.MOV(offset(dst
, bld
, i
),
2505 offset(tmp
, bld
, i
+ first_component
));
2508 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2509 inst
->size_written
= num_components
*
2510 dst
.component_size(inst
->exec_size
);
2512 inst
->offset
= base_offset
+ nir_src_as_uint(offset_src
);
2515 /* Indirect indexing - use per-slot offsets as well. */
2516 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2517 unsigned read_components
= num_components
+ first_component
;
2518 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2519 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2520 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2521 if (first_component
!= 0) {
2522 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2524 inst
->size_written
= read_components
*
2525 tmp
.component_size(inst
->exec_size
);
2526 for (unsigned i
= 0; i
< num_components
; i
++) {
2527 bld
.MOV(offset(dst
, bld
, i
),
2528 offset(tmp
, bld
, i
+ first_component
));
2531 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
, payload
);
2532 inst
->size_written
= num_components
*
2533 dst
.component_size(inst
->exec_size
);
2535 inst
->offset
= base_offset
;
2541 fs_visitor::get_indirect_offset(nir_intrinsic_instr
*instr
)
2543 nir_src
*offset_src
= nir_get_io_offset_src(instr
);
2545 if (nir_src_is_const(*offset_src
)) {
2546 /* The only constant offset we should find is 0. brw_nir.c's
2547 * add_const_offset_to_base() will fold other constant offsets
2548 * into instr->const_index[0].
2550 assert(nir_src_as_uint(*offset_src
) == 0);
2554 return get_nir_src(*offset_src
);
2558 fs_visitor::nir_emit_vs_intrinsic(const fs_builder
&bld
,
2559 nir_intrinsic_instr
*instr
)
2561 assert(stage
== MESA_SHADER_VERTEX
);
2564 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2565 dest
= get_nir_dest(instr
->dest
);
2567 switch (instr
->intrinsic
) {
2568 case nir_intrinsic_load_vertex_id
:
2569 case nir_intrinsic_load_base_vertex
:
2570 unreachable("should be lowered by nir_lower_system_values()");
2572 case nir_intrinsic_load_input
: {
2573 assert(nir_dest_bit_size(instr
->dest
) == 32);
2574 fs_reg src
= fs_reg(ATTR
, nir_intrinsic_base(instr
) * 4, dest
.type
);
2575 src
= offset(src
, bld
, nir_intrinsic_component(instr
));
2576 src
= offset(src
, bld
, nir_src_as_uint(instr
->src
[0]));
2578 for (unsigned i
= 0; i
< instr
->num_components
; i
++)
2579 bld
.MOV(offset(dest
, bld
, i
), offset(src
, bld
, i
));
2583 case nir_intrinsic_load_vertex_id_zero_base
:
2584 case nir_intrinsic_load_instance_id
:
2585 case nir_intrinsic_load_base_instance
:
2586 case nir_intrinsic_load_draw_id
:
2587 case nir_intrinsic_load_first_vertex
:
2588 case nir_intrinsic_load_is_indexed_draw
:
2589 unreachable("lowered by brw_nir_lower_vs_inputs");
2592 nir_emit_intrinsic(bld
, instr
);
2598 fs_visitor::get_tcs_single_patch_icp_handle(const fs_builder
&bld
,
2599 nir_intrinsic_instr
*instr
)
2601 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
2602 const nir_src
&vertex_src
= instr
->src
[0];
2603 nir_intrinsic_instr
*vertex_intrin
= nir_src_as_intrinsic(vertex_src
);
2606 if (nir_src_is_const(vertex_src
)) {
2607 /* Emit a MOV to resolve <0,1,0> regioning. */
2608 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2609 unsigned vertex
= nir_src_as_uint(vertex_src
);
2611 retype(brw_vec1_grf(1 + (vertex
>> 3), vertex
& 7),
2612 BRW_REGISTER_TYPE_UD
));
2613 } else if (tcs_prog_data
->instances
== 1 && vertex_intrin
&&
2614 vertex_intrin
->intrinsic
== nir_intrinsic_load_invocation_id
) {
2615 /* For the common case of only 1 instance, an array index of
2616 * gl_InvocationID means reading g1. Skip all the indirect work.
2618 icp_handle
= retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
);
2620 /* The vertex index is non-constant. We need to use indirect
2621 * addressing to fetch the proper URB handle.
2623 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2625 /* Each ICP handle is a single DWord (4 bytes) */
2626 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2627 bld
.SHL(vertex_offset_bytes
,
2628 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2631 /* Start at g1. We might read up to 4 registers. */
2632 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2633 retype(brw_vec8_grf(1, 0), icp_handle
.type
), vertex_offset_bytes
,
2634 brw_imm_ud(4 * REG_SIZE
));
2641 fs_visitor::get_tcs_eight_patch_icp_handle(const fs_builder
&bld
,
2642 nir_intrinsic_instr
*instr
)
2644 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
2645 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
2646 const nir_src
&vertex_src
= instr
->src
[0];
2648 unsigned first_icp_handle
= tcs_prog_data
->include_primitive_id
? 3 : 2;
2650 if (nir_src_is_const(vertex_src
)) {
2651 return fs_reg(retype(brw_vec8_grf(first_icp_handle
+
2652 nir_src_as_uint(vertex_src
), 0),
2653 BRW_REGISTER_TYPE_UD
));
2656 /* The vertex index is non-constant. We need to use indirect
2657 * addressing to fetch the proper URB handle.
2659 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
2660 * indicating that channel <n> should read the handle from
2661 * DWord <n>. We convert that to bytes by multiplying by 4.
2663 * Next, we convert the vertex index to bytes by multiplying
2664 * by 32 (shifting by 5), and add the two together. This is
2665 * the final indirect byte offset.
2667 fs_reg icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2668 fs_reg sequence
= bld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
2669 fs_reg channel_offsets
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2670 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2671 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2673 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
2674 bld
.MOV(sequence
, fs_reg(brw_imm_v(0x76543210)));
2675 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
2676 bld
.SHL(channel_offsets
, sequence
, brw_imm_ud(2u));
2677 /* Convert vertex_index to bytes (multiply by 32) */
2678 bld
.SHL(vertex_offset_bytes
,
2679 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2681 bld
.ADD(icp_offset_bytes
, vertex_offset_bytes
, channel_offsets
);
2683 /* Use first_icp_handle as the base offset. There is one register
2684 * of URB handles per vertex, so inform the register allocator that
2685 * we might read up to nir->info.gs.vertices_in registers.
2687 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2688 retype(brw_vec8_grf(first_icp_handle
, 0), icp_handle
.type
),
2689 icp_offset_bytes
, brw_imm_ud(tcs_key
->input_vertices
* REG_SIZE
));
2695 fs_visitor::get_tcs_output_urb_handle()
2697 struct brw_vue_prog_data
*vue_prog_data
= brw_vue_prog_data(prog_data
);
2699 if (vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_SINGLE_PATCH
) {
2700 return retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
);
2702 assert(vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
);
2703 return retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
);
2708 fs_visitor::nir_emit_tcs_intrinsic(const fs_builder
&bld
,
2709 nir_intrinsic_instr
*instr
)
2711 assert(stage
== MESA_SHADER_TESS_CTRL
);
2712 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
2713 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
2714 struct brw_vue_prog_data
*vue_prog_data
= &tcs_prog_data
->base
;
2717 vue_prog_data
->dispatch_mode
== DISPATCH_MODE_TCS_8_PATCH
;
2720 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2721 dst
= get_nir_dest(instr
->dest
);
2723 switch (instr
->intrinsic
) {
2724 case nir_intrinsic_load_primitive_id
:
2725 bld
.MOV(dst
, fs_reg(eight_patch
? brw_vec8_grf(2, 0)
2726 : brw_vec1_grf(0, 1)));
2728 case nir_intrinsic_load_invocation_id
:
2729 bld
.MOV(retype(dst
, invocation_id
.type
), invocation_id
);
2731 case nir_intrinsic_load_patch_vertices_in
:
2732 bld
.MOV(retype(dst
, BRW_REGISTER_TYPE_D
),
2733 brw_imm_d(tcs_key
->input_vertices
));
2736 case nir_intrinsic_barrier
: {
2737 if (tcs_prog_data
->instances
== 1)
2740 fs_reg m0
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2741 fs_reg m0_2
= component(m0
, 2);
2743 const fs_builder chanbld
= bld
.exec_all().group(1, 0);
2745 /* Zero the message header */
2746 bld
.exec_all().MOV(m0
, brw_imm_ud(0u));
2748 if (devinfo
->gen
< 11) {
2749 /* Copy "Barrier ID" from r0.2, bits 16:13 */
2750 chanbld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
2751 brw_imm_ud(INTEL_MASK(16, 13)));
2753 /* Shift it up to bits 27:24. */
2754 chanbld
.SHL(m0_2
, m0_2
, brw_imm_ud(11));
2756 chanbld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
2757 brw_imm_ud(INTEL_MASK(30, 24)));
2760 /* Set the Barrier Count and the enable bit */
2761 if (devinfo
->gen
< 11) {
2762 chanbld
.OR(m0_2
, m0_2
,
2763 brw_imm_ud(tcs_prog_data
->instances
<< 9 | (1 << 15)));
2765 chanbld
.OR(m0_2
, m0_2
,
2766 brw_imm_ud(tcs_prog_data
->instances
<< 8 | (1 << 15)));
2769 bld
.emit(SHADER_OPCODE_BARRIER
, bld
.null_reg_ud(), m0
);
2773 case nir_intrinsic_load_input
:
2774 unreachable("nir_lower_io should never give us these.");
2777 case nir_intrinsic_load_per_vertex_input
: {
2778 assert(nir_dest_bit_size(instr
->dest
) == 32);
2779 fs_reg indirect_offset
= get_indirect_offset(instr
);
2780 unsigned imm_offset
= instr
->const_index
[0];
2784 eight_patch
? get_tcs_eight_patch_icp_handle(bld
, instr
)
2785 : get_tcs_single_patch_icp_handle(bld
, instr
);
2787 /* We can only read two double components with each URB read, so
2788 * we send two read messages in that case, each one loading up to
2789 * two double components.
2791 unsigned num_components
= instr
->num_components
;
2792 unsigned first_component
= nir_intrinsic_component(instr
);
2794 if (indirect_offset
.file
== BAD_FILE
) {
2795 /* Constant indexing - use global offset. */
2796 if (first_component
!= 0) {
2797 unsigned read_components
= num_components
+ first_component
;
2798 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2799 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, icp_handle
);
2800 for (unsigned i
= 0; i
< num_components
; i
++) {
2801 bld
.MOV(offset(dst
, bld
, i
),
2802 offset(tmp
, bld
, i
+ first_component
));
2805 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2807 inst
->offset
= imm_offset
;
2810 /* Indirect indexing - use per-slot offsets as well. */
2811 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2812 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2813 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2814 if (first_component
!= 0) {
2815 unsigned read_components
= num_components
+ first_component
;
2816 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2817 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2819 for (unsigned i
= 0; i
< num_components
; i
++) {
2820 bld
.MOV(offset(dst
, bld
, i
),
2821 offset(tmp
, bld
, i
+ first_component
));
2824 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
,
2827 inst
->offset
= imm_offset
;
2830 inst
->size_written
= (num_components
+ first_component
) *
2831 inst
->dst
.component_size(inst
->exec_size
);
2833 /* Copy the temporary to the destination to deal with writemasking.
2835 * Also attempt to deal with gl_PointSize being in the .w component.
2837 if (inst
->offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
2838 assert(type_sz(dst
.type
) == 4);
2839 inst
->dst
= bld
.vgrf(dst
.type
, 4);
2840 inst
->size_written
= 4 * REG_SIZE
;
2841 bld
.MOV(dst
, offset(inst
->dst
, bld
, 3));
2846 case nir_intrinsic_load_output
:
2847 case nir_intrinsic_load_per_vertex_output
: {
2848 assert(nir_dest_bit_size(instr
->dest
) == 32);
2849 fs_reg indirect_offset
= get_indirect_offset(instr
);
2850 unsigned imm_offset
= instr
->const_index
[0];
2851 unsigned first_component
= nir_intrinsic_component(instr
);
2853 struct brw_reg output_handles
= get_tcs_output_urb_handle();
2856 if (indirect_offset
.file
== BAD_FILE
) {
2857 /* This MOV replicates the output handle to all enabled channels
2858 * is SINGLE_PATCH mode.
2860 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2861 bld
.MOV(patch_handle
, output_handles
);
2864 if (first_component
!= 0) {
2865 unsigned read_components
=
2866 instr
->num_components
+ first_component
;
2867 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2868 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
,
2870 inst
->size_written
= read_components
* REG_SIZE
;
2871 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2872 bld
.MOV(offset(dst
, bld
, i
),
2873 offset(tmp
, bld
, i
+ first_component
));
2876 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
,
2878 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2880 inst
->offset
= imm_offset
;
2884 /* Indirect indexing - use per-slot offsets as well. */
2885 const fs_reg srcs
[] = { output_handles
, indirect_offset
};
2886 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2887 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2888 if (first_component
!= 0) {
2889 unsigned read_components
=
2890 instr
->num_components
+ first_component
;
2891 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2892 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2894 inst
->size_written
= read_components
* REG_SIZE
;
2895 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2896 bld
.MOV(offset(dst
, bld
, i
),
2897 offset(tmp
, bld
, i
+ first_component
));
2900 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
,
2902 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2904 inst
->offset
= imm_offset
;
2910 case nir_intrinsic_store_output
:
2911 case nir_intrinsic_store_per_vertex_output
: {
2912 assert(nir_src_bit_size(instr
->src
[0]) == 32);
2913 fs_reg value
= get_nir_src(instr
->src
[0]);
2914 fs_reg indirect_offset
= get_indirect_offset(instr
);
2915 unsigned imm_offset
= instr
->const_index
[0];
2916 unsigned mask
= instr
->const_index
[1];
2917 unsigned header_regs
= 0;
2918 struct brw_reg output_handles
= get_tcs_output_urb_handle();
2921 srcs
[header_regs
++] = output_handles
;
2923 if (indirect_offset
.file
!= BAD_FILE
) {
2924 srcs
[header_regs
++] = indirect_offset
;
2930 unsigned num_components
= util_last_bit(mask
);
2933 /* We can only pack two 64-bit components in a single message, so send
2934 * 2 messages if we have more components
2936 unsigned first_component
= nir_intrinsic_component(instr
);
2937 mask
= mask
<< first_component
;
2939 if (mask
!= WRITEMASK_XYZW
) {
2940 srcs
[header_regs
++] = brw_imm_ud(mask
<< 16);
2941 opcode
= indirect_offset
.file
!= BAD_FILE
?
2942 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2943 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2945 opcode
= indirect_offset
.file
!= BAD_FILE
?
2946 SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2947 SHADER_OPCODE_URB_WRITE_SIMD8
;
2950 for (unsigned i
= 0; i
< num_components
; i
++) {
2951 if (!(mask
& (1 << (i
+ first_component
))))
2954 srcs
[header_regs
+ i
+ first_component
] = offset(value
, bld
, i
);
2957 unsigned mlen
= header_regs
+ num_components
+ first_component
;
2959 bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2960 bld
.LOAD_PAYLOAD(payload
, srcs
, mlen
, header_regs
);
2962 fs_inst
*inst
= bld
.emit(opcode
, bld
.null_reg_ud(), payload
);
2963 inst
->offset
= imm_offset
;
2969 nir_emit_intrinsic(bld
, instr
);
2975 fs_visitor::nir_emit_tes_intrinsic(const fs_builder
&bld
,
2976 nir_intrinsic_instr
*instr
)
2978 assert(stage
== MESA_SHADER_TESS_EVAL
);
2979 struct brw_tes_prog_data
*tes_prog_data
= brw_tes_prog_data(prog_data
);
2982 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2983 dest
= get_nir_dest(instr
->dest
);
2985 switch (instr
->intrinsic
) {
2986 case nir_intrinsic_load_primitive_id
:
2987 bld
.MOV(dest
, fs_reg(brw_vec1_grf(0, 1)));
2989 case nir_intrinsic_load_tess_coord
:
2990 /* gl_TessCoord is part of the payload in g1-3 */
2991 for (unsigned i
= 0; i
< 3; i
++) {
2992 bld
.MOV(offset(dest
, bld
, i
), fs_reg(brw_vec8_grf(1 + i
, 0)));
2996 case nir_intrinsic_load_input
:
2997 case nir_intrinsic_load_per_vertex_input
: {
2998 assert(nir_dest_bit_size(instr
->dest
) == 32);
2999 fs_reg indirect_offset
= get_indirect_offset(instr
);
3000 unsigned imm_offset
= instr
->const_index
[0];
3001 unsigned first_component
= nir_intrinsic_component(instr
);
3004 if (indirect_offset
.file
== BAD_FILE
) {
3005 /* Arbitrarily only push up to 32 vec4 slots worth of data,
3006 * which is 16 registers (since each holds 2 vec4 slots).
3008 const unsigned max_push_slots
= 32;
3009 if (imm_offset
< max_push_slots
) {
3010 fs_reg src
= fs_reg(ATTR
, imm_offset
/ 2, dest
.type
);
3011 for (int i
= 0; i
< instr
->num_components
; i
++) {
3012 unsigned comp
= 4 * (imm_offset
% 2) + i
+ first_component
;
3013 bld
.MOV(offset(dest
, bld
, i
), component(src
, comp
));
3016 tes_prog_data
->base
.urb_read_length
=
3017 MAX2(tes_prog_data
->base
.urb_read_length
,
3018 (imm_offset
/ 2) + 1);
3020 /* Replicate the patch handle to all enabled channels */
3021 const fs_reg srcs
[] = {
3022 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)
3024 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
3025 bld
.LOAD_PAYLOAD(patch_handle
, srcs
, ARRAY_SIZE(srcs
), 0);
3027 if (first_component
!= 0) {
3028 unsigned read_components
=
3029 instr
->num_components
+ first_component
;
3030 fs_reg tmp
= bld
.vgrf(dest
.type
, read_components
);
3031 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
,
3033 inst
->size_written
= read_components
* REG_SIZE
;
3034 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
3035 bld
.MOV(offset(dest
, bld
, i
),
3036 offset(tmp
, bld
, i
+ first_component
));
3039 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dest
,
3041 inst
->size_written
= instr
->num_components
* REG_SIZE
;
3044 inst
->offset
= imm_offset
;
3047 /* Indirect indexing - use per-slot offsets as well. */
3049 /* We can only read two double components with each URB read, so
3050 * we send two read messages in that case, each one loading up to
3051 * two double components.
3053 unsigned num_components
= instr
->num_components
;
3054 const fs_reg srcs
[] = {
3055 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
3058 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
3059 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
3061 if (first_component
!= 0) {
3062 unsigned read_components
=
3063 num_components
+ first_component
;
3064 fs_reg tmp
= bld
.vgrf(dest
.type
, read_components
);
3065 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
3067 for (unsigned i
= 0; i
< num_components
; i
++) {
3068 bld
.MOV(offset(dest
, bld
, i
),
3069 offset(tmp
, bld
, i
+ first_component
));
3072 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dest
,
3076 inst
->offset
= imm_offset
;
3077 inst
->size_written
= (num_components
+ first_component
) *
3078 inst
->dst
.component_size(inst
->exec_size
);
3083 nir_emit_intrinsic(bld
, instr
);
3089 fs_visitor::nir_emit_gs_intrinsic(const fs_builder
&bld
,
3090 nir_intrinsic_instr
*instr
)
3092 assert(stage
== MESA_SHADER_GEOMETRY
);
3093 fs_reg indirect_offset
;
3096 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3097 dest
= get_nir_dest(instr
->dest
);
3099 switch (instr
->intrinsic
) {
3100 case nir_intrinsic_load_primitive_id
:
3101 assert(stage
== MESA_SHADER_GEOMETRY
);
3102 assert(brw_gs_prog_data(prog_data
)->include_primitive_id
);
3103 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
3104 retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD
));
3107 case nir_intrinsic_load_input
:
3108 unreachable("load_input intrinsics are invalid for the GS stage");
3110 case nir_intrinsic_load_per_vertex_input
:
3111 emit_gs_input_load(dest
, instr
->src
[0], instr
->const_index
[0],
3112 instr
->src
[1], instr
->num_components
,
3113 nir_intrinsic_component(instr
));
3116 case nir_intrinsic_emit_vertex_with_counter
:
3117 emit_gs_vertex(instr
->src
[0], instr
->const_index
[0]);
3120 case nir_intrinsic_end_primitive_with_counter
:
3121 emit_gs_end_primitive(instr
->src
[0]);
3124 case nir_intrinsic_set_vertex_count
:
3125 bld
.MOV(this->final_gs_vertex_count
, get_nir_src(instr
->src
[0]));
3128 case nir_intrinsic_load_invocation_id
: {
3129 fs_reg val
= nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
3130 assert(val
.file
!= BAD_FILE
);
3131 dest
.type
= val
.type
;
3137 nir_emit_intrinsic(bld
, instr
);
3143 * Fetch the current render target layer index.
3146 fetch_render_target_array_index(const fs_builder
&bld
)
3148 if (bld
.shader
->devinfo
->gen
>= 6) {
3149 /* The render target array index is provided in the thread payload as
3150 * bits 26:16 of r0.0.
3152 const fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3153 bld
.AND(idx
, brw_uw1_reg(BRW_GENERAL_REGISTER_FILE
, 0, 1),
3157 /* Pre-SNB we only ever render into the first layer of the framebuffer
3158 * since layered rendering is not implemented.
3160 return brw_imm_ud(0);
3165 * Fake non-coherent framebuffer read implemented using TXF to fetch from the
3166 * framebuffer at the current fragment coordinates and sample index.
3169 fs_visitor::emit_non_coherent_fb_read(const fs_builder
&bld
, const fs_reg
&dst
,
3172 const struct gen_device_info
*devinfo
= bld
.shader
->devinfo
;
3174 assert(bld
.shader
->stage
== MESA_SHADER_FRAGMENT
);
3175 const brw_wm_prog_key
*wm_key
=
3176 reinterpret_cast<const brw_wm_prog_key
*>(key
);
3177 assert(!wm_key
->coherent_fb_fetch
);
3178 const struct brw_wm_prog_data
*wm_prog_data
=
3179 brw_wm_prog_data(stage_prog_data
);
3181 /* Calculate the surface index relative to the start of the texture binding
3182 * table block, since that's what the texturing messages expect.
3184 const unsigned surface
= target
+
3185 wm_prog_data
->binding_table
.render_target_read_start
-
3186 wm_prog_data
->base
.binding_table
.texture_start
;
3188 /* Calculate the fragment coordinates. */
3189 const fs_reg coords
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 3);
3190 bld
.MOV(offset(coords
, bld
, 0), pixel_x
);
3191 bld
.MOV(offset(coords
, bld
, 1), pixel_y
);
3192 bld
.MOV(offset(coords
, bld
, 2), fetch_render_target_array_index(bld
));
3194 /* Calculate the sample index and MCS payload when multisampling. Luckily
3195 * the MCS fetch message behaves deterministically for UMS surfaces, so it
3196 * shouldn't be necessary to recompile based on whether the framebuffer is
3199 if (wm_key
->multisample_fbo
&&
3200 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
].file
== BAD_FILE
)
3201 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
] = *emit_sampleid_setup();
3203 const fs_reg sample
= nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
3204 const fs_reg mcs
= wm_key
->multisample_fbo
?
3205 emit_mcs_fetch(coords
, 3, brw_imm_ud(surface
), fs_reg()) : fs_reg();
3207 /* Use either a normal or a CMS texel fetch message depending on whether
3208 * the framebuffer is single or multisample. On SKL+ use the wide CMS
3209 * message just in case the framebuffer uses 16x multisampling, it should
3210 * be equivalent to the normal CMS fetch for lower multisampling modes.
3212 const opcode op
= !wm_key
->multisample_fbo
? SHADER_OPCODE_TXF_LOGICAL
:
3213 devinfo
->gen
>= 9 ? SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
3214 SHADER_OPCODE_TXF_CMS_LOGICAL
;
3216 /* Emit the instruction. */
3217 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
3218 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = coords
;
3219 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_ud(0);
3220 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = sample
;
3221 srcs
[TEX_LOGICAL_SRC_MCS
] = mcs
;
3222 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(surface
);
3223 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(0);
3224 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_ud(3);
3225 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_ud(0);
3227 fs_inst
*inst
= bld
.emit(op
, dst
, srcs
, ARRAY_SIZE(srcs
));
3228 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
3234 * Actual coherent framebuffer read implemented using the native render target
3235 * read message. Requires SKL+.
3238 emit_coherent_fb_read(const fs_builder
&bld
, const fs_reg
&dst
, unsigned target
)
3240 assert(bld
.shader
->devinfo
->gen
>= 9);
3241 fs_inst
*inst
= bld
.emit(FS_OPCODE_FB_READ_LOGICAL
, dst
);
3242 inst
->target
= target
;
3243 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
3249 alloc_temporary(const fs_builder
&bld
, unsigned size
, fs_reg
*regs
, unsigned n
)
3251 if (n
&& regs
[0].file
!= BAD_FILE
) {
3255 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, size
);
3257 for (unsigned i
= 0; i
< n
; i
++)
3265 alloc_frag_output(fs_visitor
*v
, unsigned location
)
3267 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
3268 const brw_wm_prog_key
*const key
=
3269 reinterpret_cast<const brw_wm_prog_key
*>(v
->key
);
3270 const unsigned l
= GET_FIELD(location
, BRW_NIR_FRAG_OUTPUT_LOCATION
);
3271 const unsigned i
= GET_FIELD(location
, BRW_NIR_FRAG_OUTPUT_INDEX
);
3273 if (i
> 0 || (key
->force_dual_color_blend
&& l
== FRAG_RESULT_DATA1
))
3274 return alloc_temporary(v
->bld
, 4, &v
->dual_src_output
, 1);
3276 else if (l
== FRAG_RESULT_COLOR
)
3277 return alloc_temporary(v
->bld
, 4, v
->outputs
,
3278 MAX2(key
->nr_color_regions
, 1));
3280 else if (l
== FRAG_RESULT_DEPTH
)
3281 return alloc_temporary(v
->bld
, 1, &v
->frag_depth
, 1);
3283 else if (l
== FRAG_RESULT_STENCIL
)
3284 return alloc_temporary(v
->bld
, 1, &v
->frag_stencil
, 1);
3286 else if (l
== FRAG_RESULT_SAMPLE_MASK
)
3287 return alloc_temporary(v
->bld
, 1, &v
->sample_mask
, 1);
3289 else if (l
>= FRAG_RESULT_DATA0
&&
3290 l
< FRAG_RESULT_DATA0
+ BRW_MAX_DRAW_BUFFERS
)
3291 return alloc_temporary(v
->bld
, 4,
3292 &v
->outputs
[l
- FRAG_RESULT_DATA0
], 1);
3295 unreachable("Invalid location");
3298 /* Annoyingly, we get the barycentrics into the shader in a layout that's
3299 * optimized for PLN but it doesn't work nearly as well as one would like for
3300 * manual interpolation.
3303 shuffle_from_pln_layout(const fs_builder
&bld
, fs_reg dest
, fs_reg pln_data
)
3305 dest
.type
= BRW_REGISTER_TYPE_F
;
3306 pln_data
.type
= BRW_REGISTER_TYPE_F
;
3307 const fs_reg dest_u
= offset(dest
, bld
, 0);
3308 const fs_reg dest_v
= offset(dest
, bld
, 1);
3310 for (unsigned g
= 0; g
< bld
.dispatch_width() / 8; g
++) {
3311 const fs_builder gbld
= bld
.group(8, g
);
3312 gbld
.MOV(horiz_offset(dest_u
, g
* 8),
3313 byte_offset(pln_data
, (g
* 2 + 0) * REG_SIZE
));
3314 gbld
.MOV(horiz_offset(dest_v
, g
* 8),
3315 byte_offset(pln_data
, (g
* 2 + 1) * REG_SIZE
));
3320 shuffle_to_pln_layout(const fs_builder
&bld
, fs_reg pln_data
, fs_reg src
)
3322 pln_data
.type
= BRW_REGISTER_TYPE_F
;
3323 src
.type
= BRW_REGISTER_TYPE_F
;
3324 const fs_reg src_u
= offset(src
, bld
, 0);
3325 const fs_reg src_v
= offset(src
, bld
, 1);
3327 for (unsigned g
= 0; g
< bld
.dispatch_width() / 8; g
++) {
3328 const fs_builder gbld
= bld
.group(8, g
);
3329 gbld
.MOV(byte_offset(pln_data
, (g
* 2 + 0) * REG_SIZE
),
3330 horiz_offset(src_u
, g
* 8));
3331 gbld
.MOV(byte_offset(pln_data
, (g
* 2 + 1) * REG_SIZE
),
3332 horiz_offset(src_v
, g
* 8));
3337 fs_visitor::nir_emit_fs_intrinsic(const fs_builder
&bld
,
3338 nir_intrinsic_instr
*instr
)
3340 assert(stage
== MESA_SHADER_FRAGMENT
);
3343 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3344 dest
= get_nir_dest(instr
->dest
);
3346 switch (instr
->intrinsic
) {
3347 case nir_intrinsic_load_front_face
:
3348 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
3349 *emit_frontfacing_interpolation());
3352 case nir_intrinsic_load_sample_pos
: {
3353 fs_reg sample_pos
= nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
3354 assert(sample_pos
.file
!= BAD_FILE
);
3355 dest
.type
= sample_pos
.type
;
3356 bld
.MOV(dest
, sample_pos
);
3357 bld
.MOV(offset(dest
, bld
, 1), offset(sample_pos
, bld
, 1));
3361 case nir_intrinsic_load_layer_id
:
3362 dest
.type
= BRW_REGISTER_TYPE_UD
;
3363 bld
.MOV(dest
, fetch_render_target_array_index(bld
));
3366 case nir_intrinsic_is_helper_invocation
: {
3367 /* Unlike the regular gl_HelperInvocation, that is defined at dispatch,
3368 * the helperInvocationEXT() (aka SpvOpIsHelperInvocationEXT) takes into
3369 * consideration demoted invocations. That information is stored in
3372 dest
.type
= BRW_REGISTER_TYPE_UD
;
3374 bld
.MOV(dest
, brw_imm_ud(0));
3376 fs_inst
*mov
= bld
.MOV(dest
, brw_imm_ud(~0));
3377 mov
->predicate
= BRW_PREDICATE_NORMAL
;
3378 mov
->predicate_inverse
= true;
3379 mov
->flag_subreg
= 1;
3383 case nir_intrinsic_load_helper_invocation
:
3384 case nir_intrinsic_load_sample_mask_in
:
3385 case nir_intrinsic_load_sample_id
: {
3386 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3387 fs_reg val
= nir_system_values
[sv
];
3388 assert(val
.file
!= BAD_FILE
);
3389 dest
.type
= val
.type
;
3394 case nir_intrinsic_store_output
: {
3395 const fs_reg src
= get_nir_src(instr
->src
[0]);
3396 const unsigned store_offset
= nir_src_as_uint(instr
->src
[1]);
3397 const unsigned location
= nir_intrinsic_base(instr
) +
3398 SET_FIELD(store_offset
, BRW_NIR_FRAG_OUTPUT_LOCATION
);
3399 const fs_reg new_dest
= retype(alloc_frag_output(this, location
),
3402 for (unsigned j
= 0; j
< instr
->num_components
; j
++)
3403 bld
.MOV(offset(new_dest
, bld
, nir_intrinsic_component(instr
) + j
),
3404 offset(src
, bld
, j
));
3409 case nir_intrinsic_load_output
: {
3410 const unsigned l
= GET_FIELD(nir_intrinsic_base(instr
),
3411 BRW_NIR_FRAG_OUTPUT_LOCATION
);
3412 assert(l
>= FRAG_RESULT_DATA0
);
3413 const unsigned load_offset
= nir_src_as_uint(instr
->src
[0]);
3414 const unsigned target
= l
- FRAG_RESULT_DATA0
+ load_offset
;
3415 const fs_reg tmp
= bld
.vgrf(dest
.type
, 4);
3417 if (reinterpret_cast<const brw_wm_prog_key
*>(key
)->coherent_fb_fetch
)
3418 emit_coherent_fb_read(bld
, tmp
, target
);
3420 emit_non_coherent_fb_read(bld
, tmp
, target
);
3422 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3423 bld
.MOV(offset(dest
, bld
, j
),
3424 offset(tmp
, bld
, nir_intrinsic_component(instr
) + j
));
3430 case nir_intrinsic_demote
:
3431 case nir_intrinsic_discard
:
3432 case nir_intrinsic_demote_if
:
3433 case nir_intrinsic_discard_if
: {
3434 /* We track our discarded pixels in f0.1. By predicating on it, we can
3435 * update just the flag bits that aren't yet discarded. If there's no
3436 * condition, we emit a CMP of g0 != g0, so all currently executing
3437 * channels will get turned off.
3439 fs_inst
*cmp
= NULL
;
3440 if (instr
->intrinsic
== nir_intrinsic_demote_if
||
3441 instr
->intrinsic
== nir_intrinsic_discard_if
) {
3442 nir_alu_instr
*alu
= nir_src_as_alu_instr(instr
->src
[0]);
3445 alu
->op
!= nir_op_bcsel
&&
3446 alu
->op
!= nir_op_inot
) {
3447 /* Re-emit the instruction that generated the Boolean value, but
3448 * do not store it. Since this instruction will be conditional,
3449 * other instructions that want to use the real Boolean value may
3450 * get garbage. This was a problem for piglit's fs-discard-exit-2
3453 * Ideally we'd detect that the instruction cannot have a
3454 * conditional modifier before emitting the instructions. Alas,
3455 * that is nigh impossible. Instead, we're going to assume the
3456 * instruction (or last instruction) generated can have a
3457 * conditional modifier. If it cannot, fallback to the old-style
3458 * compare, and hope dead code elimination will clean up the
3459 * extra instructions generated.
3461 nir_emit_alu(bld
, alu
, false);
3463 cmp
= (fs_inst
*) instructions
.get_tail();
3464 if (cmp
->conditional_mod
== BRW_CONDITIONAL_NONE
) {
3465 if (cmp
->can_do_cmod())
3466 cmp
->conditional_mod
= BRW_CONDITIONAL_Z
;
3470 /* The old sequence that would have been generated is,
3471 * basically, bool_result == false. This is equivalent to
3472 * !bool_result, so negate the old modifier.
3474 cmp
->conditional_mod
= brw_negate_cmod(cmp
->conditional_mod
);
3479 cmp
= bld
.CMP(bld
.null_reg_f(), get_nir_src(instr
->src
[0]),
3480 brw_imm_d(0), BRW_CONDITIONAL_Z
);
3483 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
3484 BRW_REGISTER_TYPE_UW
));
3485 cmp
= bld
.CMP(bld
.null_reg_f(), some_reg
, some_reg
, BRW_CONDITIONAL_NZ
);
3488 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
3489 cmp
->flag_subreg
= 1;
3491 if (devinfo
->gen
>= 6) {
3492 /* Due to the way we implement discard, the jump will only happen
3493 * when the whole quad is discarded. So we can do this even for
3494 * demote as it won't break its uniformity promises.
3496 emit_discard_jump();
3499 limit_dispatch_width(16, "Fragment discard/demote not implemented in SIMD32 mode.");
3503 case nir_intrinsic_load_input
: {
3504 /* load_input is only used for flat inputs */
3505 assert(nir_dest_bit_size(instr
->dest
) == 32);
3506 unsigned base
= nir_intrinsic_base(instr
);
3507 unsigned comp
= nir_intrinsic_component(instr
);
3508 unsigned num_components
= instr
->num_components
;
3510 /* Special case fields in the VUE header */
3511 if (base
== VARYING_SLOT_LAYER
)
3513 else if (base
== VARYING_SLOT_VIEWPORT
)
3516 for (unsigned int i
= 0; i
< num_components
; i
++) {
3517 bld
.MOV(offset(dest
, bld
, i
),
3518 retype(component(interp_reg(base
, comp
+ i
), 3), dest
.type
));
3523 case nir_intrinsic_load_fs_input_interp_deltas
: {
3524 assert(stage
== MESA_SHADER_FRAGMENT
);
3525 assert(nir_src_as_uint(instr
->src
[0]) == 0);
3526 fs_reg interp
= interp_reg(nir_intrinsic_base(instr
),
3527 nir_intrinsic_component(instr
));
3528 dest
.type
= BRW_REGISTER_TYPE_F
;
3529 bld
.MOV(offset(dest
, bld
, 0), component(interp
, 3));
3530 bld
.MOV(offset(dest
, bld
, 1), component(interp
, 1));
3531 bld
.MOV(offset(dest
, bld
, 2), component(interp
, 0));
3535 case nir_intrinsic_load_barycentric_pixel
:
3536 case nir_intrinsic_load_barycentric_centroid
:
3537 case nir_intrinsic_load_barycentric_sample
: {
3538 /* Use the delta_xy values computed from the payload */
3539 const glsl_interp_mode interp_mode
=
3540 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(instr
);
3541 enum brw_barycentric_mode bary
=
3542 brw_barycentric_mode(interp_mode
, instr
->intrinsic
);
3544 shuffle_from_pln_layout(bld
, dest
, this->delta_xy
[bary
]);
3548 case nir_intrinsic_load_barycentric_at_sample
: {
3549 const glsl_interp_mode interpolation
=
3550 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(instr
);
3552 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 2);
3553 if (nir_src_is_const(instr
->src
[0])) {
3554 unsigned msg_data
= nir_src_as_uint(instr
->src
[0]) << 4;
3556 emit_pixel_interpolater_send(bld
,
3557 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3560 brw_imm_ud(msg_data
),
3563 const fs_reg sample_src
= retype(get_nir_src(instr
->src
[0]),
3564 BRW_REGISTER_TYPE_UD
);
3566 if (nir_src_is_dynamically_uniform(instr
->src
[0])) {
3567 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3568 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3569 bld
.exec_all().group(1, 0)
3570 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3571 emit_pixel_interpolater_send(bld
,
3572 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3578 /* Make a loop that sends a message to the pixel interpolater
3579 * for the sample number in each live channel. If there are
3580 * multiple channels with the same sample number then these
3581 * will be handled simultaneously with a single interation of
3584 bld
.emit(BRW_OPCODE_DO
);
3586 /* Get the next live sample number into sample_id_reg */
3587 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3589 /* Set the flag register so that we can perform the send
3590 * message on all channels that have the same sample number
3592 bld
.CMP(bld
.null_reg_ud(),
3593 sample_src
, sample_id
,
3594 BRW_CONDITIONAL_EQ
);
3595 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3596 bld
.exec_all().group(1, 0)
3597 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3599 emit_pixel_interpolater_send(bld
,
3600 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3603 component(msg_data
, 0),
3605 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
3607 /* Continue the loop if there are any live channels left */
3608 set_predicate_inv(BRW_PREDICATE_NORMAL
,
3610 bld
.emit(BRW_OPCODE_WHILE
));
3613 shuffle_from_pln_layout(bld
, dest
, tmp
);
3617 case nir_intrinsic_load_barycentric_at_offset
: {
3618 const glsl_interp_mode interpolation
=
3619 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(instr
);
3621 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3623 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 2);
3625 assert(nir_src_bit_size(instr
->src
[0]) == 32);
3626 unsigned off_x
= MIN2((int)(const_offset
[0].f32
* 16), 7) & 0xf;
3627 unsigned off_y
= MIN2((int)(const_offset
[1].f32
* 16), 7) & 0xf;
3629 emit_pixel_interpolater_send(bld
,
3630 FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
,
3633 brw_imm_ud(off_x
| (off_y
<< 4)),
3636 fs_reg src
= vgrf(glsl_type::ivec2_type
);
3637 fs_reg offset_src
= retype(get_nir_src(instr
->src
[0]),
3638 BRW_REGISTER_TYPE_F
);
3639 for (int i
= 0; i
< 2; i
++) {
3640 fs_reg temp
= vgrf(glsl_type::float_type
);
3641 bld
.MUL(temp
, offset(offset_src
, bld
, i
), brw_imm_f(16.0f
));
3642 fs_reg itemp
= vgrf(glsl_type::int_type
);
3644 bld
.MOV(itemp
, temp
);
3646 /* Clamp the upper end of the range to +7/16.
3647 * ARB_gpu_shader5 requires that we support a maximum offset
3648 * of +0.5, which isn't representable in a S0.4 value -- if
3649 * we didn't clamp it, we'd end up with -8/16, which is the
3650 * opposite of what the shader author wanted.
3652 * This is legal due to ARB_gpu_shader5's quantization
3655 * "Not all values of <offset> may be supported; x and y
3656 * offsets may be rounded to fixed-point values with the
3657 * number of fraction bits given by the
3658 * implementation-dependent constant
3659 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
3661 set_condmod(BRW_CONDITIONAL_L
,
3662 bld
.SEL(offset(src
, bld
, i
), itemp
, brw_imm_d(7)));
3665 const enum opcode opcode
= FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
;
3666 emit_pixel_interpolater_send(bld
,
3673 shuffle_from_pln_layout(bld
, dest
, tmp
);
3677 case nir_intrinsic_load_frag_coord
:
3678 emit_fragcoord_interpolation(dest
);
3681 case nir_intrinsic_load_interpolated_input
: {
3682 assert(instr
->src
[0].ssa
&&
3683 instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_intrinsic
);
3684 nir_intrinsic_instr
*bary_intrinsic
=
3685 nir_instr_as_intrinsic(instr
->src
[0].ssa
->parent_instr
);
3686 nir_intrinsic_op bary_intrin
= bary_intrinsic
->intrinsic
;
3687 enum glsl_interp_mode interp_mode
=
3688 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(bary_intrinsic
);
3691 if (bary_intrin
== nir_intrinsic_load_barycentric_at_offset
||
3692 bary_intrin
== nir_intrinsic_load_barycentric_at_sample
) {
3693 /* Use the result of the PI message. Because the load_barycentric
3694 * intrinsics return a regular vec2 and we need it in PLN layout, we
3695 * have to do a translation. Fortunately, copy-prop cleans this up
3698 dst_xy
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 2);
3699 shuffle_to_pln_layout(bld
, dst_xy
, get_nir_src(instr
->src
[0]));
3701 /* Use the delta_xy values computed from the payload */
3702 enum brw_barycentric_mode bary
=
3703 brw_barycentric_mode(interp_mode
, bary_intrin
);
3705 dst_xy
= this->delta_xy
[bary
];
3708 for (unsigned int i
= 0; i
< instr
->num_components
; i
++) {
3710 component(interp_reg(nir_intrinsic_base(instr
),
3711 nir_intrinsic_component(instr
) + i
), 0);
3712 interp
.type
= BRW_REGISTER_TYPE_F
;
3713 dest
.type
= BRW_REGISTER_TYPE_F
;
3715 if (devinfo
->gen
< 6 && interp_mode
== INTERP_MODE_SMOOTH
) {
3716 fs_reg tmp
= vgrf(glsl_type::float_type
);
3717 bld
.emit(FS_OPCODE_LINTERP
, tmp
, dst_xy
, interp
);
3718 bld
.MUL(offset(dest
, bld
, i
), tmp
, this->pixel_w
);
3720 bld
.emit(FS_OPCODE_LINTERP
, offset(dest
, bld
, i
), dst_xy
, interp
);
3727 nir_emit_intrinsic(bld
, instr
);
3733 fs_visitor::nir_emit_cs_intrinsic(const fs_builder
&bld
,
3734 nir_intrinsic_instr
*instr
)
3736 assert(stage
== MESA_SHADER_COMPUTE
);
3737 struct brw_cs_prog_data
*cs_prog_data
= brw_cs_prog_data(prog_data
);
3740 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3741 dest
= get_nir_dest(instr
->dest
);
3743 switch (instr
->intrinsic
) {
3744 case nir_intrinsic_barrier
:
3746 cs_prog_data
->uses_barrier
= true;
3749 case nir_intrinsic_load_subgroup_id
:
3750 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
), subgroup_id
);
3753 case nir_intrinsic_load_local_invocation_id
:
3754 case nir_intrinsic_load_work_group_id
: {
3755 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3756 fs_reg val
= nir_system_values
[sv
];
3757 assert(val
.file
!= BAD_FILE
);
3758 dest
.type
= val
.type
;
3759 for (unsigned i
= 0; i
< 3; i
++)
3760 bld
.MOV(offset(dest
, bld
, i
), offset(val
, bld
, i
));
3764 case nir_intrinsic_load_num_work_groups
: {
3765 const unsigned surface
=
3766 cs_prog_data
->binding_table
.work_groups_start
;
3768 cs_prog_data
->uses_num_work_groups
= true;
3770 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3771 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(surface
);
3772 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3773 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(1); /* num components */
3775 /* Read the 3 GLuint components of gl_NumWorkGroups */
3776 for (unsigned i
= 0; i
< 3; i
++) {
3777 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = brw_imm_ud(i
<< 2);
3778 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
3779 offset(dest
, bld
, i
), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3784 case nir_intrinsic_shared_atomic_add
:
3785 case nir_intrinsic_shared_atomic_imin
:
3786 case nir_intrinsic_shared_atomic_umin
:
3787 case nir_intrinsic_shared_atomic_imax
:
3788 case nir_intrinsic_shared_atomic_umax
:
3789 case nir_intrinsic_shared_atomic_and
:
3790 case nir_intrinsic_shared_atomic_or
:
3791 case nir_intrinsic_shared_atomic_xor
:
3792 case nir_intrinsic_shared_atomic_exchange
:
3793 case nir_intrinsic_shared_atomic_comp_swap
:
3794 nir_emit_shared_atomic(bld
, brw_aop_for_nir_intrinsic(instr
), instr
);
3796 case nir_intrinsic_shared_atomic_fmin
:
3797 case nir_intrinsic_shared_atomic_fmax
:
3798 case nir_intrinsic_shared_atomic_fcomp_swap
:
3799 nir_emit_shared_atomic_float(bld
, brw_aop_for_nir_intrinsic(instr
), instr
);
3802 case nir_intrinsic_load_shared
: {
3803 assert(devinfo
->gen
>= 7);
3804 assert(stage
== MESA_SHADER_COMPUTE
);
3806 const unsigned bit_size
= nir_dest_bit_size(instr
->dest
);
3807 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3808 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
3809 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[0]);
3810 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3812 /* Make dest unsigned because that's what the temporary will be */
3813 dest
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
3815 /* Read the vector */
3816 if (nir_intrinsic_align(instr
) >= 4) {
3817 assert(nir_dest_bit_size(instr
->dest
) == 32);
3818 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3820 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
3821 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3822 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
3824 assert(nir_dest_bit_size(instr
->dest
) <= 32);
3825 assert(nir_dest_num_components(instr
->dest
) == 1);
3826 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
3828 fs_reg read_result
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3829 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
,
3830 read_result
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3831 bld
.MOV(dest
, subscript(read_result
, dest
.type
, 0));
3836 case nir_intrinsic_store_shared
: {
3837 assert(devinfo
->gen
>= 7);
3838 assert(stage
== MESA_SHADER_COMPUTE
);
3840 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
3841 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3842 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
3843 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
3844 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3846 fs_reg data
= get_nir_src(instr
->src
[0]);
3847 data
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
3849 assert(nir_intrinsic_write_mask(instr
) ==
3850 (1u << instr
->num_components
) - 1);
3851 if (nir_intrinsic_align(instr
) >= 4) {
3852 assert(nir_src_bit_size(instr
->src
[0]) == 32);
3853 assert(nir_src_num_components(instr
->src
[0]) <= 4);
3854 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
3855 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3856 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
,
3857 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3859 assert(nir_src_bit_size(instr
->src
[0]) <= 32);
3860 assert(nir_src_num_components(instr
->src
[0]) == 1);
3861 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
3863 srcs
[SURFACE_LOGICAL_SRC_DATA
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3864 bld
.MOV(srcs
[SURFACE_LOGICAL_SRC_DATA
], data
);
3866 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
,
3867 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3873 nir_emit_intrinsic(bld
, instr
);
3879 brw_nir_reduction_op_identity(const fs_builder
&bld
,
3880 nir_op op
, brw_reg_type type
)
3882 nir_const_value value
= nir_alu_binop_identity(op
, type_sz(type
) * 8);
3883 switch (type_sz(type
)) {
3885 if (type
== BRW_REGISTER_TYPE_UB
) {
3886 return brw_imm_uw(value
.u8
);
3888 assert(type
== BRW_REGISTER_TYPE_B
);
3889 return brw_imm_w(value
.i8
);
3892 return retype(brw_imm_uw(value
.u16
), type
);
3894 return retype(brw_imm_ud(value
.u32
), type
);
3896 if (type
== BRW_REGISTER_TYPE_DF
)
3897 return setup_imm_df(bld
, value
.f64
);
3899 return retype(brw_imm_u64(value
.u64
), type
);
3901 unreachable("Invalid type size");
3906 brw_op_for_nir_reduction_op(nir_op op
)
3909 case nir_op_iadd
: return BRW_OPCODE_ADD
;
3910 case nir_op_fadd
: return BRW_OPCODE_ADD
;
3911 case nir_op_imul
: return BRW_OPCODE_MUL
;
3912 case nir_op_fmul
: return BRW_OPCODE_MUL
;
3913 case nir_op_imin
: return BRW_OPCODE_SEL
;
3914 case nir_op_umin
: return BRW_OPCODE_SEL
;
3915 case nir_op_fmin
: return BRW_OPCODE_SEL
;
3916 case nir_op_imax
: return BRW_OPCODE_SEL
;
3917 case nir_op_umax
: return BRW_OPCODE_SEL
;
3918 case nir_op_fmax
: return BRW_OPCODE_SEL
;
3919 case nir_op_iand
: return BRW_OPCODE_AND
;
3920 case nir_op_ior
: return BRW_OPCODE_OR
;
3921 case nir_op_ixor
: return BRW_OPCODE_XOR
;
3923 unreachable("Invalid reduction operation");
3927 static brw_conditional_mod
3928 brw_cond_mod_for_nir_reduction_op(nir_op op
)
3931 case nir_op_iadd
: return BRW_CONDITIONAL_NONE
;
3932 case nir_op_fadd
: return BRW_CONDITIONAL_NONE
;
3933 case nir_op_imul
: return BRW_CONDITIONAL_NONE
;
3934 case nir_op_fmul
: return BRW_CONDITIONAL_NONE
;
3935 case nir_op_imin
: return BRW_CONDITIONAL_L
;
3936 case nir_op_umin
: return BRW_CONDITIONAL_L
;
3937 case nir_op_fmin
: return BRW_CONDITIONAL_L
;
3938 case nir_op_imax
: return BRW_CONDITIONAL_GE
;
3939 case nir_op_umax
: return BRW_CONDITIONAL_GE
;
3940 case nir_op_fmax
: return BRW_CONDITIONAL_GE
;
3941 case nir_op_iand
: return BRW_CONDITIONAL_NONE
;
3942 case nir_op_ior
: return BRW_CONDITIONAL_NONE
;
3943 case nir_op_ixor
: return BRW_CONDITIONAL_NONE
;
3945 unreachable("Invalid reduction operation");
3950 fs_visitor::get_nir_image_intrinsic_image(const brw::fs_builder
&bld
,
3951 nir_intrinsic_instr
*instr
)
3953 fs_reg image
= retype(get_nir_src_imm(instr
->src
[0]), BRW_REGISTER_TYPE_UD
);
3955 if (stage_prog_data
->binding_table
.image_start
> 0) {
3956 if (image
.file
== BRW_IMMEDIATE_VALUE
) {
3957 image
.d
+= stage_prog_data
->binding_table
.image_start
;
3959 bld
.ADD(image
, image
,
3960 brw_imm_d(stage_prog_data
->binding_table
.image_start
));
3964 return bld
.emit_uniformize(image
);
3968 fs_visitor::get_nir_ssbo_intrinsic_index(const brw::fs_builder
&bld
,
3969 nir_intrinsic_instr
*instr
)
3971 /* SSBO stores are weird in that their index is in src[1] */
3972 const unsigned src
= instr
->intrinsic
== nir_intrinsic_store_ssbo
? 1 : 0;
3975 if (nir_src_is_const(instr
->src
[src
])) {
3976 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3977 nir_src_as_uint(instr
->src
[src
]);
3978 surf_index
= brw_imm_ud(index
);
3980 surf_index
= vgrf(glsl_type::uint_type
);
3981 bld
.ADD(surf_index
, get_nir_src(instr
->src
[src
]),
3982 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3985 return bld
.emit_uniformize(surf_index
);
3989 image_intrinsic_coord_components(nir_intrinsic_instr
*instr
)
3991 switch (nir_intrinsic_image_dim(instr
)) {
3992 case GLSL_SAMPLER_DIM_1D
:
3993 return 1 + nir_intrinsic_image_array(instr
);
3994 case GLSL_SAMPLER_DIM_2D
:
3995 case GLSL_SAMPLER_DIM_RECT
:
3996 return 2 + nir_intrinsic_image_array(instr
);
3997 case GLSL_SAMPLER_DIM_3D
:
3998 case GLSL_SAMPLER_DIM_CUBE
:
4000 case GLSL_SAMPLER_DIM_BUF
:
4002 case GLSL_SAMPLER_DIM_MS
:
4003 return 2 + nir_intrinsic_image_array(instr
);
4005 unreachable("Invalid image dimension");
4010 fs_visitor::nir_emit_intrinsic(const fs_builder
&bld
, nir_intrinsic_instr
*instr
)
4013 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4014 dest
= get_nir_dest(instr
->dest
);
4016 switch (instr
->intrinsic
) {
4017 case nir_intrinsic_image_load
:
4018 case nir_intrinsic_image_store
:
4019 case nir_intrinsic_image_atomic_add
:
4020 case nir_intrinsic_image_atomic_imin
:
4021 case nir_intrinsic_image_atomic_umin
:
4022 case nir_intrinsic_image_atomic_imax
:
4023 case nir_intrinsic_image_atomic_umax
:
4024 case nir_intrinsic_image_atomic_and
:
4025 case nir_intrinsic_image_atomic_or
:
4026 case nir_intrinsic_image_atomic_xor
:
4027 case nir_intrinsic_image_atomic_exchange
:
4028 case nir_intrinsic_image_atomic_comp_swap
:
4029 case nir_intrinsic_bindless_image_load
:
4030 case nir_intrinsic_bindless_image_store
:
4031 case nir_intrinsic_bindless_image_atomic_add
:
4032 case nir_intrinsic_bindless_image_atomic_imin
:
4033 case nir_intrinsic_bindless_image_atomic_umin
:
4034 case nir_intrinsic_bindless_image_atomic_imax
:
4035 case nir_intrinsic_bindless_image_atomic_umax
:
4036 case nir_intrinsic_bindless_image_atomic_and
:
4037 case nir_intrinsic_bindless_image_atomic_or
:
4038 case nir_intrinsic_bindless_image_atomic_xor
:
4039 case nir_intrinsic_bindless_image_atomic_exchange
:
4040 case nir_intrinsic_bindless_image_atomic_comp_swap
: {
4041 if (stage
== MESA_SHADER_FRAGMENT
&&
4042 instr
->intrinsic
!= nir_intrinsic_image_load
)
4043 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4045 /* Get some metadata from the image intrinsic. */
4046 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
4048 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4050 switch (instr
->intrinsic
) {
4051 case nir_intrinsic_image_load
:
4052 case nir_intrinsic_image_store
:
4053 case nir_intrinsic_image_atomic_add
:
4054 case nir_intrinsic_image_atomic_imin
:
4055 case nir_intrinsic_image_atomic_umin
:
4056 case nir_intrinsic_image_atomic_imax
:
4057 case nir_intrinsic_image_atomic_umax
:
4058 case nir_intrinsic_image_atomic_and
:
4059 case nir_intrinsic_image_atomic_or
:
4060 case nir_intrinsic_image_atomic_xor
:
4061 case nir_intrinsic_image_atomic_exchange
:
4062 case nir_intrinsic_image_atomic_comp_swap
:
4063 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
4064 get_nir_image_intrinsic_image(bld
, instr
);
4069 srcs
[SURFACE_LOGICAL_SRC_SURFACE_HANDLE
] =
4070 bld
.emit_uniformize(get_nir_src(instr
->src
[0]));
4074 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
4075 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] =
4076 brw_imm_ud(image_intrinsic_coord_components(instr
));
4078 /* Emit an image load, store or atomic op. */
4079 if (instr
->intrinsic
== nir_intrinsic_image_load
||
4080 instr
->intrinsic
== nir_intrinsic_bindless_image_load
) {
4081 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4083 bld
.emit(SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
,
4084 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4085 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
4086 } else if (instr
->intrinsic
== nir_intrinsic_image_store
||
4087 instr
->intrinsic
== nir_intrinsic_bindless_image_store
) {
4088 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4089 srcs
[SURFACE_LOGICAL_SRC_DATA
] = get_nir_src(instr
->src
[3]);
4090 bld
.emit(SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
,
4091 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4093 unsigned num_srcs
= info
->num_srcs
;
4094 int op
= brw_aop_for_nir_intrinsic(instr
);
4095 if (op
== BRW_AOP_INC
|| op
== BRW_AOP_DEC
) {
4096 assert(num_srcs
== 4);
4100 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
4104 data
= get_nir_src(instr
->src
[3]);
4105 if (num_srcs
>= 5) {
4106 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
4107 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[4]) };
4108 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
4111 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
4113 bld
.emit(SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
,
4114 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4119 case nir_intrinsic_image_size
:
4120 case nir_intrinsic_bindless_image_size
: {
4121 /* Unlike the [un]typed load and store opcodes, the TXS that this turns
4122 * into will handle the binding table index for us in the geneerator.
4123 * Incidentally, this means that we can handle bindless with exactly the
4126 fs_reg image
= retype(get_nir_src_imm(instr
->src
[0]),
4127 BRW_REGISTER_TYPE_UD
);
4128 image
= bld
.emit_uniformize(image
);
4130 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
4131 if (instr
->intrinsic
== nir_intrinsic_image_size
)
4132 srcs
[TEX_LOGICAL_SRC_SURFACE
] = image
;
4134 srcs
[TEX_LOGICAL_SRC_SURFACE_HANDLE
] = image
;
4135 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_d(0);
4136 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(0);
4137 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(0);
4139 /* Since the image size is always uniform, we can just emit a SIMD8
4140 * query instruction and splat the result out.
4142 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4144 fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 4);
4145 fs_inst
*inst
= ubld
.emit(SHADER_OPCODE_IMAGE_SIZE_LOGICAL
,
4146 tmp
, srcs
, ARRAY_SIZE(srcs
));
4147 inst
->size_written
= 4 * REG_SIZE
;
4149 for (unsigned c
= 0; c
< instr
->dest
.ssa
.num_components
; ++c
) {
4150 if (c
== 2 && nir_intrinsic_image_dim(instr
) == GLSL_SAMPLER_DIM_CUBE
) {
4151 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
,
4152 offset(retype(dest
, tmp
.type
), bld
, c
),
4153 component(offset(tmp
, ubld
, c
), 0), brw_imm_ud(6));
4155 bld
.MOV(offset(retype(dest
, tmp
.type
), bld
, c
),
4156 component(offset(tmp
, ubld
, c
), 0));
4162 case nir_intrinsic_image_load_raw_intel
: {
4163 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4164 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
4165 get_nir_image_intrinsic_image(bld
, instr
);
4166 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
4167 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4168 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4171 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
4172 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4173 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
4177 case nir_intrinsic_image_store_raw_intel
: {
4178 if (stage
== MESA_SHADER_FRAGMENT
)
4179 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4181 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4182 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
4183 get_nir_image_intrinsic_image(bld
, instr
);
4184 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
4185 srcs
[SURFACE_LOGICAL_SRC_DATA
] = get_nir_src(instr
->src
[2]);
4186 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4187 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4189 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
,
4190 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4194 case nir_intrinsic_group_memory_barrier
:
4195 case nir_intrinsic_memory_barrier_shared
:
4196 case nir_intrinsic_memory_barrier_atomic_counter
:
4197 case nir_intrinsic_memory_barrier_buffer
:
4198 case nir_intrinsic_memory_barrier_image
:
4199 case nir_intrinsic_memory_barrier
: {
4200 bool l3_fence
, slm_fence
;
4201 if (devinfo
->gen
>= 11) {
4202 l3_fence
= instr
->intrinsic
!= nir_intrinsic_memory_barrier_shared
;
4203 slm_fence
= instr
->intrinsic
== nir_intrinsic_group_memory_barrier
||
4204 instr
->intrinsic
== nir_intrinsic_memory_barrier
||
4205 instr
->intrinsic
== nir_intrinsic_memory_barrier_shared
;
4207 /* Prior to gen11, we only have one kind of fence. */
4212 /* Be conservative in Gen11+ and always stall in a fence. Since there
4213 * are two different fences, and shader might want to synchronize
4216 * TODO: Improve NIR so that scope and visibility information for the
4217 * barriers is available here to make a better decision.
4219 * TODO: When emitting more than one fence, it might help emit all
4220 * the fences first and then generate the stall moves.
4222 const bool stall
= devinfo
->gen
>= 11;
4224 const fs_builder ubld
= bld
.group(8, 0);
4225 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
4228 ubld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
,
4229 brw_vec8_grf(0, 0), brw_imm_ud(stall
),
4230 /* bti */ brw_imm_ud(0))
4231 ->size_written
= 2 * REG_SIZE
;
4235 ubld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
,
4236 brw_vec8_grf(0, 0), brw_imm_ud(stall
),
4237 brw_imm_ud(GEN7_BTI_SLM
))
4238 ->size_written
= 2 * REG_SIZE
;
4244 case nir_intrinsic_shader_clock
: {
4245 /* We cannot do anything if there is an event, so ignore it for now */
4246 const fs_reg shader_clock
= get_timestamp(bld
);
4247 const fs_reg srcs
[] = { component(shader_clock
, 0),
4248 component(shader_clock
, 1) };
4249 bld
.LOAD_PAYLOAD(dest
, srcs
, ARRAY_SIZE(srcs
), 0);
4253 case nir_intrinsic_image_samples
:
4254 /* The driver does not support multi-sampled images. */
4255 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), brw_imm_d(1));
4258 case nir_intrinsic_load_uniform
: {
4259 /* Offsets are in bytes but they should always aligned to
4262 assert(instr
->const_index
[0] % 4 == 0 ||
4263 instr
->const_index
[0] % type_sz(dest
.type
) == 0);
4265 fs_reg
src(UNIFORM
, instr
->const_index
[0] / 4, dest
.type
);
4267 if (nir_src_is_const(instr
->src
[0])) {
4268 unsigned load_offset
= nir_src_as_uint(instr
->src
[0]);
4269 assert(load_offset
% type_sz(dest
.type
) == 0);
4270 /* For 16-bit types we add the module of the const_index[0]
4271 * offset to access to not 32-bit aligned element
4273 src
.offset
= load_offset
+ instr
->const_index
[0] % 4;
4275 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
4276 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
4279 fs_reg indirect
= retype(get_nir_src(instr
->src
[0]),
4280 BRW_REGISTER_TYPE_UD
);
4282 /* We need to pass a size to the MOV_INDIRECT but we don't want it to
4283 * go past the end of the uniform. In order to keep the n'th
4284 * component from running past, we subtract off the size of all but
4285 * one component of the vector.
4287 assert(instr
->const_index
[1] >=
4288 instr
->num_components
* (int) type_sz(dest
.type
));
4289 unsigned read_size
= instr
->const_index
[1] -
4290 (instr
->num_components
- 1) * type_sz(dest
.type
);
4292 bool supports_64bit_indirects
=
4293 !devinfo
->is_cherryview
&& !gen_device_info_is_9lp(devinfo
);
4295 if (type_sz(dest
.type
) != 8 || supports_64bit_indirects
) {
4296 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
4297 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
4298 offset(dest
, bld
, j
), offset(src
, bld
, j
),
4299 indirect
, brw_imm_ud(read_size
));
4302 const unsigned num_mov_indirects
=
4303 type_sz(dest
.type
) / type_sz(BRW_REGISTER_TYPE_UD
);
4304 /* We read a little bit less per MOV INDIRECT, as they are now
4305 * 32-bits ones instead of 64-bit. Fix read_size then.
4307 const unsigned read_size_32bit
= read_size
-
4308 (num_mov_indirects
- 1) * type_sz(BRW_REGISTER_TYPE_UD
);
4309 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
4310 for (unsigned i
= 0; i
< num_mov_indirects
; i
++) {
4311 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
4312 subscript(offset(dest
, bld
, j
), BRW_REGISTER_TYPE_UD
, i
),
4313 subscript(offset(src
, bld
, j
), BRW_REGISTER_TYPE_UD
, i
),
4314 indirect
, brw_imm_ud(read_size_32bit
));
4322 case nir_intrinsic_load_ubo
: {
4324 if (nir_src_is_const(instr
->src
[0])) {
4325 const unsigned index
= stage_prog_data
->binding_table
.ubo_start
+
4326 nir_src_as_uint(instr
->src
[0]);
4327 surf_index
= brw_imm_ud(index
);
4329 /* The block index is not a constant. Evaluate the index expression
4330 * per-channel and add the base UBO index; we have to select a value
4331 * from any live channel.
4333 surf_index
= vgrf(glsl_type::uint_type
);
4334 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
4335 brw_imm_ud(stage_prog_data
->binding_table
.ubo_start
));
4336 surf_index
= bld
.emit_uniformize(surf_index
);
4339 if (!nir_src_is_const(instr
->src
[1])) {
4340 fs_reg base_offset
= retype(get_nir_src(instr
->src
[1]),
4341 BRW_REGISTER_TYPE_UD
);
4343 for (int i
= 0; i
< instr
->num_components
; i
++)
4344 VARYING_PULL_CONSTANT_LOAD(bld
, offset(dest
, bld
, i
), surf_index
,
4345 base_offset
, i
* type_sz(dest
.type
));
4347 prog_data
->has_ubo_pull
= true;
4349 /* Even if we are loading doubles, a pull constant load will load
4350 * a 32-bit vec4, so should only reserve vgrf space for that. If we
4351 * need to load a full dvec4 we will have to emit 2 loads. This is
4352 * similar to demote_pull_constants(), except that in that case we
4353 * see individual accesses to each component of the vector and then
4354 * we let CSE deal with duplicate loads. Here we see a vector access
4355 * and we have to split it if necessary.
4357 const unsigned type_size
= type_sz(dest
.type
);
4358 const unsigned load_offset
= nir_src_as_uint(instr
->src
[1]);
4360 /* See if we've selected this as a push constant candidate */
4361 if (nir_src_is_const(instr
->src
[0])) {
4362 const unsigned ubo_block
= nir_src_as_uint(instr
->src
[0]);
4363 const unsigned offset_256b
= load_offset
/ 32;
4366 for (int i
= 0; i
< 4; i
++) {
4367 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4368 if (range
->block
== ubo_block
&&
4369 offset_256b
>= range
->start
&&
4370 offset_256b
< range
->start
+ range
->length
) {
4372 push_reg
= fs_reg(UNIFORM
, UBO_START
+ i
, dest
.type
);
4373 push_reg
.offset
= load_offset
- 32 * range
->start
;
4378 if (push_reg
.file
!= BAD_FILE
) {
4379 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
4380 bld
.MOV(offset(dest
, bld
, i
),
4381 byte_offset(push_reg
, i
* type_size
));
4387 prog_data
->has_ubo_pull
= true;
4389 const unsigned block_sz
= 64; /* Fetch one cacheline at a time. */
4390 const fs_builder ubld
= bld
.exec_all().group(block_sz
/ 4, 0);
4391 const fs_reg packed_consts
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4393 for (unsigned c
= 0; c
< instr
->num_components
;) {
4394 const unsigned base
= load_offset
+ c
* type_size
;
4395 /* Number of usable components in the next block-aligned load. */
4396 const unsigned count
= MIN2(instr
->num_components
- c
,
4397 (block_sz
- base
% block_sz
) / type_size
);
4399 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
4400 packed_consts
, surf_index
,
4401 brw_imm_ud(base
& ~(block_sz
- 1)));
4403 const fs_reg consts
=
4404 retype(byte_offset(packed_consts
, base
& (block_sz
- 1)),
4407 for (unsigned d
= 0; d
< count
; d
++)
4408 bld
.MOV(offset(dest
, bld
, c
+ d
), component(consts
, d
));
4416 case nir_intrinsic_load_global
: {
4417 assert(devinfo
->gen
>= 8);
4419 if (nir_intrinsic_align(instr
) >= 4) {
4420 assert(nir_dest_bit_size(instr
->dest
) == 32);
4421 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
,
4423 get_nir_src(instr
->src
[0]), /* Address */
4424 fs_reg(), /* No source data */
4425 brw_imm_ud(instr
->num_components
));
4426 inst
->size_written
= instr
->num_components
*
4427 inst
->dst
.component_size(inst
->exec_size
);
4429 const unsigned bit_size
= nir_dest_bit_size(instr
->dest
);
4430 assert(bit_size
<= 32);
4431 assert(nir_dest_num_components(instr
->dest
) == 1);
4432 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4433 bld
.emit(SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
,
4435 get_nir_src(instr
->src
[0]), /* Address */
4436 fs_reg(), /* No source data */
4437 brw_imm_ud(bit_size
));
4438 bld
.MOV(dest
, subscript(tmp
, dest
.type
, 0));
4443 case nir_intrinsic_store_global
:
4444 assert(devinfo
->gen
>= 8);
4446 if (stage
== MESA_SHADER_FRAGMENT
)
4447 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4449 if (nir_intrinsic_align(instr
) >= 4) {
4450 assert(nir_src_bit_size(instr
->src
[0]) == 32);
4451 bld
.emit(SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
,
4453 get_nir_src(instr
->src
[1]), /* Address */
4454 get_nir_src(instr
->src
[0]), /* Data */
4455 brw_imm_ud(instr
->num_components
));
4457 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4458 assert(bit_size
<= 32);
4459 assert(nir_src_num_components(instr
->src
[0]) == 1);
4460 brw_reg_type data_type
=
4461 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
4462 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4463 bld
.MOV(tmp
, retype(get_nir_src(instr
->src
[0]), data_type
));
4464 bld
.emit(SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
,
4466 get_nir_src(instr
->src
[1]), /* Address */
4468 brw_imm_ud(nir_src_bit_size(instr
->src
[0])));
4472 case nir_intrinsic_global_atomic_add
:
4473 case nir_intrinsic_global_atomic_imin
:
4474 case nir_intrinsic_global_atomic_umin
:
4475 case nir_intrinsic_global_atomic_imax
:
4476 case nir_intrinsic_global_atomic_umax
:
4477 case nir_intrinsic_global_atomic_and
:
4478 case nir_intrinsic_global_atomic_or
:
4479 case nir_intrinsic_global_atomic_xor
:
4480 case nir_intrinsic_global_atomic_exchange
:
4481 case nir_intrinsic_global_atomic_comp_swap
:
4482 nir_emit_global_atomic(bld
, brw_aop_for_nir_intrinsic(instr
), instr
);
4484 case nir_intrinsic_global_atomic_fmin
:
4485 case nir_intrinsic_global_atomic_fmax
:
4486 case nir_intrinsic_global_atomic_fcomp_swap
:
4487 nir_emit_global_atomic_float(bld
, brw_aop_for_nir_intrinsic(instr
), instr
);
4490 case nir_intrinsic_load_ssbo
: {
4491 assert(devinfo
->gen
>= 7);
4493 const unsigned bit_size
= nir_dest_bit_size(instr
->dest
);
4494 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4495 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
4496 get_nir_ssbo_intrinsic_index(bld
, instr
);
4497 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
4498 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4500 /* Make dest unsigned because that's what the temporary will be */
4501 dest
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
4503 /* Read the vector */
4504 if (nir_intrinsic_align(instr
) >= 4) {
4505 assert(nir_dest_bit_size(instr
->dest
) == 32);
4506 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4508 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
4509 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4510 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
4512 assert(nir_dest_bit_size(instr
->dest
) <= 32);
4513 assert(nir_dest_num_components(instr
->dest
) == 1);
4514 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
4516 fs_reg read_result
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4517 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
,
4518 read_result
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4519 bld
.MOV(dest
, subscript(read_result
, dest
.type
, 0));
4524 case nir_intrinsic_store_ssbo
: {
4525 assert(devinfo
->gen
>= 7);
4527 if (stage
== MESA_SHADER_FRAGMENT
)
4528 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4530 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4531 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4532 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
4533 get_nir_ssbo_intrinsic_index(bld
, instr
);
4534 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[2]);
4535 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4537 fs_reg data
= get_nir_src(instr
->src
[0]);
4538 data
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
4540 assert(nir_intrinsic_write_mask(instr
) ==
4541 (1u << instr
->num_components
) - 1);
4542 if (nir_intrinsic_align(instr
) >= 4) {
4543 assert(nir_src_bit_size(instr
->src
[0]) == 32);
4544 assert(nir_src_num_components(instr
->src
[0]) <= 4);
4545 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
4546 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4547 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
,
4548 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4550 assert(nir_src_bit_size(instr
->src
[0]) <= 32);
4551 assert(nir_src_num_components(instr
->src
[0]) == 1);
4552 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
4554 srcs
[SURFACE_LOGICAL_SRC_DATA
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4555 bld
.MOV(srcs
[SURFACE_LOGICAL_SRC_DATA
], data
);
4557 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
,
4558 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4563 case nir_intrinsic_store_output
: {
4564 assert(nir_src_bit_size(instr
->src
[0]) == 32);
4565 fs_reg src
= get_nir_src(instr
->src
[0]);
4567 unsigned store_offset
= nir_src_as_uint(instr
->src
[1]);
4568 unsigned num_components
= instr
->num_components
;
4569 unsigned first_component
= nir_intrinsic_component(instr
);
4571 fs_reg new_dest
= retype(offset(outputs
[instr
->const_index
[0]], bld
,
4572 4 * store_offset
), src
.type
);
4573 for (unsigned j
= 0; j
< num_components
; j
++) {
4574 bld
.MOV(offset(new_dest
, bld
, j
+ first_component
),
4575 offset(src
, bld
, j
));
4580 case nir_intrinsic_ssbo_atomic_add
:
4581 case nir_intrinsic_ssbo_atomic_imin
:
4582 case nir_intrinsic_ssbo_atomic_umin
:
4583 case nir_intrinsic_ssbo_atomic_imax
:
4584 case nir_intrinsic_ssbo_atomic_umax
:
4585 case nir_intrinsic_ssbo_atomic_and
:
4586 case nir_intrinsic_ssbo_atomic_or
:
4587 case nir_intrinsic_ssbo_atomic_xor
:
4588 case nir_intrinsic_ssbo_atomic_exchange
:
4589 case nir_intrinsic_ssbo_atomic_comp_swap
:
4590 nir_emit_ssbo_atomic(bld
, brw_aop_for_nir_intrinsic(instr
), instr
);
4592 case nir_intrinsic_ssbo_atomic_fmin
:
4593 case nir_intrinsic_ssbo_atomic_fmax
:
4594 case nir_intrinsic_ssbo_atomic_fcomp_swap
:
4595 nir_emit_ssbo_atomic_float(bld
, brw_aop_for_nir_intrinsic(instr
), instr
);
4598 case nir_intrinsic_get_buffer_size
: {
4599 assert(nir_src_num_components(instr
->src
[0]) == 1);
4600 unsigned ssbo_index
= nir_src_is_const(instr
->src
[0]) ?
4601 nir_src_as_uint(instr
->src
[0]) : 0;
4603 /* A resinfo's sampler message is used to get the buffer size. The
4604 * SIMD8's writeback message consists of four registers and SIMD16's
4605 * writeback message consists of 8 destination registers (two per each
4606 * component). Because we are only interested on the first channel of
4607 * the first returned component, where resinfo returns the buffer size
4608 * for SURFTYPE_BUFFER, we can just use the SIMD8 variant regardless of
4609 * the dispatch width.
4611 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4612 fs_reg src_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4613 fs_reg ret_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 4);
4616 ubld
.MOV(src_payload
, brw_imm_d(0));
4618 const unsigned index
= prog_data
->binding_table
.ssbo_start
+ ssbo_index
;
4619 fs_inst
*inst
= ubld
.emit(SHADER_OPCODE_GET_BUFFER_SIZE
, ret_payload
,
4620 src_payload
, brw_imm_ud(index
));
4621 inst
->header_size
= 0;
4623 inst
->size_written
= 4 * REG_SIZE
;
4625 /* SKL PRM, vol07, 3D Media GPGPU Engine, Bounds Checking and Faulting:
4627 * "Out-of-bounds checking is always performed at a DWord granularity. If
4628 * any part of the DWord is out-of-bounds then the whole DWord is
4629 * considered out-of-bounds."
4631 * This implies that types with size smaller than 4-bytes need to be
4632 * padded if they don't complete the last dword of the buffer. But as we
4633 * need to maintain the original size we need to reverse the padding
4634 * calculation to return the correct size to know the number of elements
4635 * of an unsized array. As we stored in the last two bits of the surface
4636 * size the needed padding for the buffer, we calculate here the
4637 * original buffer_size reversing the surface_size calculation:
4639 * surface_size = isl_align(buffer_size, 4) +
4640 * (isl_align(buffer_size) - buffer_size)
4642 * buffer_size = surface_size & ~3 - surface_size & 3
4645 fs_reg size_aligned4
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4646 fs_reg size_padding
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4647 fs_reg buffer_size
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4649 ubld
.AND(size_padding
, ret_payload
, brw_imm_ud(3));
4650 ubld
.AND(size_aligned4
, ret_payload
, brw_imm_ud(~3));
4651 ubld
.ADD(buffer_size
, size_aligned4
, negate(size_padding
));
4653 bld
.MOV(retype(dest
, ret_payload
.type
), component(buffer_size
, 0));
4657 case nir_intrinsic_load_subgroup_size
:
4658 /* This should only happen for fragment shaders because every other case
4659 * is lowered in NIR so we can optimize on it.
4661 assert(stage
== MESA_SHADER_FRAGMENT
);
4662 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), brw_imm_d(dispatch_width
));
4665 case nir_intrinsic_load_subgroup_invocation
:
4666 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
4667 nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
]);
4670 case nir_intrinsic_load_subgroup_eq_mask
:
4671 case nir_intrinsic_load_subgroup_ge_mask
:
4672 case nir_intrinsic_load_subgroup_gt_mask
:
4673 case nir_intrinsic_load_subgroup_le_mask
:
4674 case nir_intrinsic_load_subgroup_lt_mask
:
4675 unreachable("not reached");
4677 case nir_intrinsic_vote_any
: {
4678 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4680 /* The any/all predicates do not consider channel enables. To prevent
4681 * dead channels from affecting the result, we initialize the flag with
4682 * with the identity value for the logical operation.
4684 if (dispatch_width
== 32) {
4685 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4686 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4689 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0));
4691 bld
.CMP(bld
.null_reg_d(), get_nir_src(instr
->src
[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ
);
4693 /* For some reason, the any/all predicates don't work properly with
4694 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4695 * doesn't read the correct subset of the flag register and you end up
4696 * getting garbage in the second half. Work around this by using a pair
4697 * of 1-wide MOVs and scattering the result.
4699 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4700 ubld
.MOV(res1
, brw_imm_d(0));
4701 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ANY8H
:
4702 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ANY16H
:
4703 BRW_PREDICATE_ALIGN1_ANY32H
,
4704 ubld
.MOV(res1
, brw_imm_d(-1)));
4706 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4709 case nir_intrinsic_vote_all
: {
4710 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4712 /* The any/all predicates do not consider channel enables. To prevent
4713 * dead channels from affecting the result, we initialize the flag with
4714 * with the identity value for the logical operation.
4716 if (dispatch_width
== 32) {
4717 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4718 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4719 brw_imm_ud(0xffffffff));
4721 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0xffff));
4723 bld
.CMP(bld
.null_reg_d(), get_nir_src(instr
->src
[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ
);
4725 /* For some reason, the any/all predicates don't work properly with
4726 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4727 * doesn't read the correct subset of the flag register and you end up
4728 * getting garbage in the second half. Work around this by using a pair
4729 * of 1-wide MOVs and scattering the result.
4731 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4732 ubld
.MOV(res1
, brw_imm_d(0));
4733 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ALL8H
:
4734 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ALL16H
:
4735 BRW_PREDICATE_ALIGN1_ALL32H
,
4736 ubld
.MOV(res1
, brw_imm_d(-1)));
4738 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4741 case nir_intrinsic_vote_feq
:
4742 case nir_intrinsic_vote_ieq
: {
4743 fs_reg value
= get_nir_src(instr
->src
[0]);
4744 if (instr
->intrinsic
== nir_intrinsic_vote_feq
) {
4745 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4746 value
.type
= bit_size
== 8 ? BRW_REGISTER_TYPE_B
:
4747 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_F
);
4750 fs_reg uniformized
= bld
.emit_uniformize(value
);
4751 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4753 /* The any/all predicates do not consider channel enables. To prevent
4754 * dead channels from affecting the result, we initialize the flag with
4755 * with the identity value for the logical operation.
4757 if (dispatch_width
== 32) {
4758 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4759 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4760 brw_imm_ud(0xffffffff));
4762 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0xffff));
4764 bld
.CMP(bld
.null_reg_d(), value
, uniformized
, BRW_CONDITIONAL_Z
);
4766 /* For some reason, the any/all predicates don't work properly with
4767 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4768 * doesn't read the correct subset of the flag register and you end up
4769 * getting garbage in the second half. Work around this by using a pair
4770 * of 1-wide MOVs and scattering the result.
4772 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4773 ubld
.MOV(res1
, brw_imm_d(0));
4774 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ALL8H
:
4775 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ALL16H
:
4776 BRW_PREDICATE_ALIGN1_ALL32H
,
4777 ubld
.MOV(res1
, brw_imm_d(-1)));
4779 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4783 case nir_intrinsic_ballot
: {
4784 const fs_reg value
= retype(get_nir_src(instr
->src
[0]),
4785 BRW_REGISTER_TYPE_UD
);
4786 struct brw_reg flag
= brw_flag_reg(0, 0);
4787 /* FIXME: For SIMD32 programs, this causes us to stomp on f0.1 as well
4788 * as f0.0. This is a problem for fragment programs as we currently use
4789 * f0.1 for discards. Fortunately, we don't support SIMD32 fragment
4790 * programs yet so this isn't a problem. When we do, something will
4793 if (dispatch_width
== 32)
4794 flag
.type
= BRW_REGISTER_TYPE_UD
;
4796 bld
.exec_all().group(1, 0).MOV(flag
, brw_imm_ud(0u));
4797 bld
.CMP(bld
.null_reg_ud(), value
, brw_imm_ud(0u), BRW_CONDITIONAL_NZ
);
4799 if (instr
->dest
.ssa
.bit_size
> 32) {
4800 dest
.type
= BRW_REGISTER_TYPE_UQ
;
4802 dest
.type
= BRW_REGISTER_TYPE_UD
;
4804 bld
.MOV(dest
, flag
);
4808 case nir_intrinsic_read_invocation
: {
4809 const fs_reg value
= get_nir_src(instr
->src
[0]);
4810 const fs_reg invocation
= get_nir_src(instr
->src
[1]);
4811 fs_reg tmp
= bld
.vgrf(value
.type
);
4813 bld
.exec_all().emit(SHADER_OPCODE_BROADCAST
, tmp
, value
,
4814 bld
.emit_uniformize(invocation
));
4816 bld
.MOV(retype(dest
, value
.type
), fs_reg(component(tmp
, 0)));
4820 case nir_intrinsic_read_first_invocation
: {
4821 const fs_reg value
= get_nir_src(instr
->src
[0]);
4822 bld
.MOV(retype(dest
, value
.type
), bld
.emit_uniformize(value
));
4826 case nir_intrinsic_shuffle
: {
4827 const fs_reg value
= get_nir_src(instr
->src
[0]);
4828 const fs_reg index
= get_nir_src(instr
->src
[1]);
4830 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, index
);
4834 case nir_intrinsic_first_invocation
: {
4835 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4836 bld
.exec_all().emit(SHADER_OPCODE_FIND_LIVE_CHANNEL
, tmp
);
4837 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
4838 fs_reg(component(tmp
, 0)));
4842 case nir_intrinsic_quad_broadcast
: {
4843 const fs_reg value
= get_nir_src(instr
->src
[0]);
4844 const unsigned index
= nir_src_as_uint(instr
->src
[1]);
4846 bld
.emit(SHADER_OPCODE_CLUSTER_BROADCAST
, retype(dest
, value
.type
),
4847 value
, brw_imm_ud(index
), brw_imm_ud(4));
4851 case nir_intrinsic_quad_swap_horizontal
: {
4852 const fs_reg value
= get_nir_src(instr
->src
[0]);
4853 const fs_reg tmp
= bld
.vgrf(value
.type
);
4854 if (devinfo
->gen
<= 7) {
4855 /* The hardware doesn't seem to support these crazy regions with
4856 * compressed instructions on gen7 and earlier so we fall back to
4857 * using quad swizzles. Fortunately, we don't support 64-bit
4858 * anything in Vulkan on gen7.
4860 assert(nir_src_bit_size(instr
->src
[0]) == 32);
4861 const fs_builder ubld
= bld
.exec_all();
4862 ubld
.emit(SHADER_OPCODE_QUAD_SWIZZLE
, tmp
, value
,
4863 brw_imm_ud(BRW_SWIZZLE4(1,0,3,2)));
4864 bld
.MOV(retype(dest
, value
.type
), tmp
);
4866 const fs_builder ubld
= bld
.exec_all().group(dispatch_width
/ 2, 0);
4868 const fs_reg src_left
= horiz_stride(value
, 2);
4869 const fs_reg src_right
= horiz_stride(horiz_offset(value
, 1), 2);
4870 const fs_reg tmp_left
= horiz_stride(tmp
, 2);
4871 const fs_reg tmp_right
= horiz_stride(horiz_offset(tmp
, 1), 2);
4873 ubld
.MOV(tmp_left
, src_right
);
4874 ubld
.MOV(tmp_right
, src_left
);
4877 bld
.MOV(retype(dest
, value
.type
), tmp
);
4881 case nir_intrinsic_quad_swap_vertical
: {
4882 const fs_reg value
= get_nir_src(instr
->src
[0]);
4883 if (nir_src_bit_size(instr
->src
[0]) == 32) {
4884 /* For 32-bit, we can use a SIMD4x2 instruction to do this easily */
4885 const fs_reg tmp
= bld
.vgrf(value
.type
);
4886 const fs_builder ubld
= bld
.exec_all();
4887 ubld
.emit(SHADER_OPCODE_QUAD_SWIZZLE
, tmp
, value
,
4888 brw_imm_ud(BRW_SWIZZLE4(2,3,0,1)));
4889 bld
.MOV(retype(dest
, value
.type
), tmp
);
4891 /* For larger data types, we have to either emit dispatch_width many
4892 * MOVs or else fall back to doing indirects.
4894 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
4895 bld
.XOR(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
4897 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, idx
);
4902 case nir_intrinsic_quad_swap_diagonal
: {
4903 const fs_reg value
= get_nir_src(instr
->src
[0]);
4904 if (nir_src_bit_size(instr
->src
[0]) == 32) {
4905 /* For 32-bit, we can use a SIMD4x2 instruction to do this easily */
4906 const fs_reg tmp
= bld
.vgrf(value
.type
);
4907 const fs_builder ubld
= bld
.exec_all();
4908 ubld
.emit(SHADER_OPCODE_QUAD_SWIZZLE
, tmp
, value
,
4909 brw_imm_ud(BRW_SWIZZLE4(3,2,1,0)));
4910 bld
.MOV(retype(dest
, value
.type
), tmp
);
4912 /* For larger data types, we have to either emit dispatch_width many
4913 * MOVs or else fall back to doing indirects.
4915 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
4916 bld
.XOR(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
4918 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, idx
);
4923 case nir_intrinsic_reduce
: {
4924 fs_reg src
= get_nir_src(instr
->src
[0]);
4925 nir_op redop
= (nir_op
)nir_intrinsic_reduction_op(instr
);
4926 unsigned cluster_size
= nir_intrinsic_cluster_size(instr
);
4927 if (cluster_size
== 0 || cluster_size
> dispatch_width
)
4928 cluster_size
= dispatch_width
;
4930 /* Figure out the source type */
4931 src
.type
= brw_type_for_nir_type(devinfo
,
4932 (nir_alu_type
)(nir_op_infos
[redop
].input_types
[0] |
4933 nir_src_bit_size(instr
->src
[0])));
4935 fs_reg identity
= brw_nir_reduction_op_identity(bld
, redop
, src
.type
);
4936 opcode brw_op
= brw_op_for_nir_reduction_op(redop
);
4937 brw_conditional_mod cond_mod
= brw_cond_mod_for_nir_reduction_op(redop
);
4939 /* There are a couple of register region issues that make things
4940 * complicated for 8-bit types:
4942 * 1. Only raw moves are allowed to write to a packed 8-bit
4944 * 2. If we use a strided destination, the efficient way to do scan
4945 * operations ends up using strides that are too big to encode in
4948 * To get around these issues, we just do all 8-bit scan operations in
4949 * 16 bits. It's actually fewer instructions than what we'd have to do
4950 * if we were trying to do it in native 8-bit types and the results are
4951 * the same once we truncate to 8 bits at the end.
4953 brw_reg_type scan_type
= src
.type
;
4954 if (type_sz(scan_type
) == 1)
4955 scan_type
= brw_reg_type_from_bit_size(16, src
.type
);
4957 /* Set up a register for all of our scratching around and initialize it
4958 * to reduction operation's identity value.
4960 fs_reg scan
= bld
.vgrf(scan_type
);
4961 bld
.exec_all().emit(SHADER_OPCODE_SEL_EXEC
, scan
, src
, identity
);
4963 bld
.emit_scan(brw_op
, scan
, cluster_size
, cond_mod
);
4965 dest
.type
= src
.type
;
4966 if (cluster_size
* type_sz(src
.type
) >= REG_SIZE
* 2) {
4967 /* In this case, CLUSTER_BROADCAST instruction isn't needed because
4968 * the distance between clusters is at least 2 GRFs. In this case,
4969 * we don't need the weird striding of the CLUSTER_BROADCAST
4970 * instruction and can just do regular MOVs.
4972 assert((cluster_size
* type_sz(src
.type
)) % (REG_SIZE
* 2) == 0);
4973 const unsigned groups
=
4974 (dispatch_width
* type_sz(src
.type
)) / (REG_SIZE
* 2);
4975 const unsigned group_size
= dispatch_width
/ groups
;
4976 for (unsigned i
= 0; i
< groups
; i
++) {
4977 const unsigned cluster
= (i
* group_size
) / cluster_size
;
4978 const unsigned comp
= cluster
* cluster_size
+ (cluster_size
- 1);
4979 bld
.group(group_size
, i
).MOV(horiz_offset(dest
, i
* group_size
),
4980 component(scan
, comp
));
4983 bld
.emit(SHADER_OPCODE_CLUSTER_BROADCAST
, dest
, scan
,
4984 brw_imm_ud(cluster_size
- 1), brw_imm_ud(cluster_size
));
4989 case nir_intrinsic_inclusive_scan
:
4990 case nir_intrinsic_exclusive_scan
: {
4991 fs_reg src
= get_nir_src(instr
->src
[0]);
4992 nir_op redop
= (nir_op
)nir_intrinsic_reduction_op(instr
);
4994 /* Figure out the source type */
4995 src
.type
= brw_type_for_nir_type(devinfo
,
4996 (nir_alu_type
)(nir_op_infos
[redop
].input_types
[0] |
4997 nir_src_bit_size(instr
->src
[0])));
4999 fs_reg identity
= brw_nir_reduction_op_identity(bld
, redop
, src
.type
);
5000 opcode brw_op
= brw_op_for_nir_reduction_op(redop
);
5001 brw_conditional_mod cond_mod
= brw_cond_mod_for_nir_reduction_op(redop
);
5003 /* There are a couple of register region issues that make things
5004 * complicated for 8-bit types:
5006 * 1. Only raw moves are allowed to write to a packed 8-bit
5008 * 2. If we use a strided destination, the efficient way to do scan
5009 * operations ends up using strides that are too big to encode in
5012 * To get around these issues, we just do all 8-bit scan operations in
5013 * 16 bits. It's actually fewer instructions than what we'd have to do
5014 * if we were trying to do it in native 8-bit types and the results are
5015 * the same once we truncate to 8 bits at the end.
5017 brw_reg_type scan_type
= src
.type
;
5018 if (type_sz(scan_type
) == 1)
5019 scan_type
= brw_reg_type_from_bit_size(16, src
.type
);
5021 /* Set up a register for all of our scratching around and initialize it
5022 * to reduction operation's identity value.
5024 fs_reg scan
= bld
.vgrf(scan_type
);
5025 const fs_builder allbld
= bld
.exec_all();
5026 allbld
.emit(SHADER_OPCODE_SEL_EXEC
, scan
, src
, identity
);
5028 if (instr
->intrinsic
== nir_intrinsic_exclusive_scan
) {
5029 /* Exclusive scan is a bit harder because we have to do an annoying
5030 * shift of the contents before we can begin. To make things worse,
5031 * we can't do this with a normal stride; we have to use indirects.
5033 fs_reg shifted
= bld
.vgrf(scan_type
);
5034 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
5035 allbld
.ADD(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
5037 allbld
.emit(SHADER_OPCODE_SHUFFLE
, shifted
, scan
, idx
);
5038 allbld
.group(1, 0).MOV(component(shifted
, 0), identity
);
5042 bld
.emit_scan(brw_op
, scan
, dispatch_width
, cond_mod
);
5044 bld
.MOV(retype(dest
, src
.type
), scan
);
5048 case nir_intrinsic_begin_invocation_interlock
: {
5049 const fs_builder ubld
= bld
.group(8, 0);
5050 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
5052 ubld
.emit(SHADER_OPCODE_INTERLOCK
, tmp
, brw_vec8_grf(0, 0))
5053 ->size_written
= 2 * REG_SIZE
;
5057 case nir_intrinsic_end_invocation_interlock
: {
5058 /* For endInvocationInterlock(), we need to insert a memory fence which
5059 * stalls in the shader until the memory transactions prior to that
5060 * fence are complete. This ensures that the shader does not end before
5061 * any writes from its critical section have landed. Otherwise, you can
5062 * end up with a case where the next invocation on that pixel properly
5063 * stalls for previous FS invocation on its pixel to complete but
5064 * doesn't actually wait for the dataport memory transactions from that
5065 * thread to land before submitting its own.
5067 const fs_builder ubld
= bld
.group(8, 0);
5068 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
5069 ubld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
,
5070 brw_vec8_grf(0, 0), brw_imm_ud(1), brw_imm_ud(0))
5071 ->size_written
= 2 * REG_SIZE
;
5076 unreachable("unknown intrinsic");
5081 fs_visitor::nir_emit_ssbo_atomic(const fs_builder
&bld
,
5082 int op
, nir_intrinsic_instr
*instr
)
5084 if (stage
== MESA_SHADER_FRAGMENT
)
5085 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
5087 /* The BTI untyped atomic messages only support 32-bit atomics. If you
5088 * just look at the big table of messages in the Vol 7 of the SKL PRM, they
5089 * appear to exist. However, if you look at Vol 2a, there are no message
5090 * descriptors provided for Qword atomic ops except for A64 messages.
5092 assert(nir_dest_bit_size(instr
->dest
) == 32);
5095 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
5096 dest
= get_nir_dest(instr
->dest
);
5098 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
5099 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = get_nir_ssbo_intrinsic_index(bld
, instr
);
5100 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
5101 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
5102 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
5105 if (op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
)
5106 data
= get_nir_src(instr
->src
[2]);
5108 if (op
== BRW_AOP_CMPWR
) {
5109 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
5110 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[3]) };
5111 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
5114 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
5116 /* Emit the actual atomic operation */
5118 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
,
5119 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
5123 fs_visitor::nir_emit_ssbo_atomic_float(const fs_builder
&bld
,
5124 int op
, nir_intrinsic_instr
*instr
)
5126 if (stage
== MESA_SHADER_FRAGMENT
)
5127 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
5130 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
5131 dest
= get_nir_dest(instr
->dest
);
5133 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
5134 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = get_nir_ssbo_intrinsic_index(bld
, instr
);
5135 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
5136 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
5137 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
5139 fs_reg data
= get_nir_src(instr
->src
[2]);
5140 if (op
== BRW_AOP_FCMPWR
) {
5141 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
5142 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[3]) };
5143 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
5146 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
5148 /* Emit the actual atomic operation */
5150 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
,
5151 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
5155 fs_visitor::nir_emit_shared_atomic(const fs_builder
&bld
,
5156 int op
, nir_intrinsic_instr
*instr
)
5159 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
5160 dest
= get_nir_dest(instr
->dest
);
5162 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
5163 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
5164 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
5165 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
5168 if (op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
)
5169 data
= get_nir_src(instr
->src
[1]);
5170 if (op
== BRW_AOP_CMPWR
) {
5171 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
5172 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
5173 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
5176 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
5178 /* Get the offset */
5179 if (nir_src_is_const(instr
->src
[0])) {
5180 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] =
5181 brw_imm_ud(instr
->const_index
[0] + nir_src_as_uint(instr
->src
[0]));
5183 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = vgrf(glsl_type::uint_type
);
5184 bld
.ADD(srcs
[SURFACE_LOGICAL_SRC_ADDRESS
],
5185 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
5186 brw_imm_ud(instr
->const_index
[0]));
5189 /* Emit the actual atomic operation operation */
5191 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
,
5192 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
5196 fs_visitor::nir_emit_shared_atomic_float(const fs_builder
&bld
,
5197 int op
, nir_intrinsic_instr
*instr
)
5200 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
5201 dest
= get_nir_dest(instr
->dest
);
5203 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
5204 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
5205 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
5206 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
5208 fs_reg data
= get_nir_src(instr
->src
[1]);
5209 if (op
== BRW_AOP_FCMPWR
) {
5210 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
5211 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
5212 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
5215 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
5217 /* Get the offset */
5218 if (nir_src_is_const(instr
->src
[0])) {
5219 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] =
5220 brw_imm_ud(instr
->const_index
[0] + nir_src_as_uint(instr
->src
[0]));
5222 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = vgrf(glsl_type::uint_type
);
5223 bld
.ADD(srcs
[SURFACE_LOGICAL_SRC_ADDRESS
],
5224 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
5225 brw_imm_ud(instr
->const_index
[0]));
5228 /* Emit the actual atomic operation operation */
5230 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
,
5231 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
5235 fs_visitor::nir_emit_global_atomic(const fs_builder
&bld
,
5236 int op
, nir_intrinsic_instr
*instr
)
5238 if (stage
== MESA_SHADER_FRAGMENT
)
5239 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
5242 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
5243 dest
= get_nir_dest(instr
->dest
);
5245 fs_reg addr
= get_nir_src(instr
->src
[0]);
5248 if (op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
)
5249 data
= get_nir_src(instr
->src
[1]);
5251 if (op
== BRW_AOP_CMPWR
) {
5252 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
5253 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
5254 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
5258 if (nir_dest_bit_size(instr
->dest
) == 64) {
5259 bld
.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_INT64_LOGICAL
,
5260 dest
, addr
, data
, brw_imm_ud(op
));
5262 assert(nir_dest_bit_size(instr
->dest
) == 32);
5263 bld
.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
,
5264 dest
, addr
, data
, brw_imm_ud(op
));
5269 fs_visitor::nir_emit_global_atomic_float(const fs_builder
&bld
,
5270 int op
, nir_intrinsic_instr
*instr
)
5272 if (stage
== MESA_SHADER_FRAGMENT
)
5273 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
5275 assert(nir_intrinsic_infos
[instr
->intrinsic
].has_dest
);
5276 fs_reg dest
= get_nir_dest(instr
->dest
);
5278 fs_reg addr
= get_nir_src(instr
->src
[0]);
5280 assert(op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
);
5281 fs_reg data
= get_nir_src(instr
->src
[1]);
5283 if (op
== BRW_AOP_FCMPWR
) {
5284 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
5285 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
5286 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
5290 bld
.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
,
5291 dest
, addr
, data
, brw_imm_ud(op
));
5295 fs_visitor::nir_emit_texture(const fs_builder
&bld
, nir_tex_instr
*instr
)
5297 unsigned texture
= instr
->texture_index
;
5298 unsigned sampler
= instr
->sampler_index
;
5300 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
5302 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture
);
5303 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(sampler
);
5305 int lod_components
= 0;
5307 /* The hardware requires a LOD for buffer textures */
5308 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
5309 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_d(0);
5311 uint32_t header_bits
= 0;
5312 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
5313 fs_reg src
= get_nir_src(instr
->src
[i
].src
);
5314 switch (instr
->src
[i
].src_type
) {
5315 case nir_tex_src_bias
:
5316 srcs
[TEX_LOGICAL_SRC_LOD
] =
5317 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
5319 case nir_tex_src_comparator
:
5320 srcs
[TEX_LOGICAL_SRC_SHADOW_C
] = retype(src
, BRW_REGISTER_TYPE_F
);
5322 case nir_tex_src_coord
:
5323 switch (instr
->op
) {
5325 case nir_texop_txf_ms
:
5326 case nir_texop_txf_ms_mcs
:
5327 case nir_texop_samples_identical
:
5328 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_D
);
5331 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_F
);
5335 case nir_tex_src_ddx
:
5336 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
5337 lod_components
= nir_tex_instr_src_size(instr
, i
);
5339 case nir_tex_src_ddy
:
5340 srcs
[TEX_LOGICAL_SRC_LOD2
] = retype(src
, BRW_REGISTER_TYPE_F
);
5342 case nir_tex_src_lod
:
5343 switch (instr
->op
) {
5345 srcs
[TEX_LOGICAL_SRC_LOD
] =
5346 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_UD
);
5349 srcs
[TEX_LOGICAL_SRC_LOD
] =
5350 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_D
);
5353 srcs
[TEX_LOGICAL_SRC_LOD
] =
5354 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
5358 case nir_tex_src_min_lod
:
5359 srcs
[TEX_LOGICAL_SRC_MIN_LOD
] =
5360 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
5362 case nir_tex_src_ms_index
:
5363 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = retype(src
, BRW_REGISTER_TYPE_UD
);
5366 case nir_tex_src_offset
: {
5367 uint32_t offset_bits
= 0;
5368 if (brw_texture_offset(instr
, i
, &offset_bits
)) {
5369 header_bits
|= offset_bits
;
5371 srcs
[TEX_LOGICAL_SRC_TG4_OFFSET
] =
5372 retype(src
, BRW_REGISTER_TYPE_D
);
5377 case nir_tex_src_projector
:
5378 unreachable("should be lowered");
5380 case nir_tex_src_texture_offset
: {
5381 /* Emit code to evaluate the actual indexing expression */
5382 fs_reg tmp
= vgrf(glsl_type::uint_type
);
5383 bld
.ADD(tmp
, src
, brw_imm_ud(texture
));
5384 srcs
[TEX_LOGICAL_SRC_SURFACE
] = bld
.emit_uniformize(tmp
);
5388 case nir_tex_src_sampler_offset
: {
5389 /* Emit code to evaluate the actual indexing expression */
5390 fs_reg tmp
= vgrf(glsl_type::uint_type
);
5391 bld
.ADD(tmp
, src
, brw_imm_ud(sampler
));
5392 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = bld
.emit_uniformize(tmp
);
5396 case nir_tex_src_texture_handle
:
5397 assert(nir_tex_instr_src_index(instr
, nir_tex_src_texture_offset
) == -1);
5398 srcs
[TEX_LOGICAL_SRC_SURFACE
] = fs_reg();
5399 srcs
[TEX_LOGICAL_SRC_SURFACE_HANDLE
] = bld
.emit_uniformize(src
);
5402 case nir_tex_src_sampler_handle
:
5403 assert(nir_tex_instr_src_index(instr
, nir_tex_src_sampler_offset
) == -1);
5404 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = fs_reg();
5405 srcs
[TEX_LOGICAL_SRC_SAMPLER_HANDLE
] = bld
.emit_uniformize(src
);
5408 case nir_tex_src_ms_mcs
:
5409 assert(instr
->op
== nir_texop_txf_ms
);
5410 srcs
[TEX_LOGICAL_SRC_MCS
] = retype(src
, BRW_REGISTER_TYPE_D
);
5413 case nir_tex_src_plane
: {
5414 const uint32_t plane
= nir_src_as_uint(instr
->src
[i
].src
);
5415 const uint32_t texture_index
=
5416 instr
->texture_index
+
5417 stage_prog_data
->binding_table
.plane_start
[plane
] -
5418 stage_prog_data
->binding_table
.texture_start
;
5420 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture_index
);
5425 unreachable("unknown texture source");
5429 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BAD_FILE
&&
5430 (instr
->op
== nir_texop_txf_ms
||
5431 instr
->op
== nir_texop_samples_identical
)) {
5432 if (devinfo
->gen
>= 7 &&
5433 key_tex
->compressed_multisample_layout_mask
& (1 << texture
)) {
5434 srcs
[TEX_LOGICAL_SRC_MCS
] =
5435 emit_mcs_fetch(srcs
[TEX_LOGICAL_SRC_COORDINATE
],
5436 instr
->coord_components
,
5437 srcs
[TEX_LOGICAL_SRC_SURFACE
],
5438 srcs
[TEX_LOGICAL_SRC_SURFACE_HANDLE
]);
5440 srcs
[TEX_LOGICAL_SRC_MCS
] = brw_imm_ud(0u);
5444 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(instr
->coord_components
);
5445 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(lod_components
);
5448 switch (instr
->op
) {
5450 opcode
= SHADER_OPCODE_TEX_LOGICAL
;
5453 opcode
= FS_OPCODE_TXB_LOGICAL
;
5456 opcode
= SHADER_OPCODE_TXL_LOGICAL
;
5459 opcode
= SHADER_OPCODE_TXD_LOGICAL
;
5462 opcode
= SHADER_OPCODE_TXF_LOGICAL
;
5464 case nir_texop_txf_ms
:
5465 if ((key_tex
->msaa_16
& (1 << sampler
)))
5466 opcode
= SHADER_OPCODE_TXF_CMS_W_LOGICAL
;
5468 opcode
= SHADER_OPCODE_TXF_CMS_LOGICAL
;
5470 case nir_texop_txf_ms_mcs
:
5471 opcode
= SHADER_OPCODE_TXF_MCS_LOGICAL
;
5473 case nir_texop_query_levels
:
5475 opcode
= SHADER_OPCODE_TXS_LOGICAL
;
5478 opcode
= SHADER_OPCODE_LOD_LOGICAL
;
5481 if (srcs
[TEX_LOGICAL_SRC_TG4_OFFSET
].file
!= BAD_FILE
)
5482 opcode
= SHADER_OPCODE_TG4_OFFSET_LOGICAL
;
5484 opcode
= SHADER_OPCODE_TG4_LOGICAL
;
5486 case nir_texop_texture_samples
:
5487 opcode
= SHADER_OPCODE_SAMPLEINFO_LOGICAL
;
5489 case nir_texop_samples_identical
: {
5490 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
5492 /* If mcs is an immediate value, it means there is no MCS. In that case
5493 * just return false.
5495 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BRW_IMMEDIATE_VALUE
) {
5496 bld
.MOV(dst
, brw_imm_ud(0u));
5497 } else if ((key_tex
->msaa_16
& (1 << sampler
))) {
5498 fs_reg tmp
= vgrf(glsl_type::uint_type
);
5499 bld
.OR(tmp
, srcs
[TEX_LOGICAL_SRC_MCS
],
5500 offset(srcs
[TEX_LOGICAL_SRC_MCS
], bld
, 1));
5501 bld
.CMP(dst
, tmp
, brw_imm_ud(0u), BRW_CONDITIONAL_EQ
);
5503 bld
.CMP(dst
, srcs
[TEX_LOGICAL_SRC_MCS
], brw_imm_ud(0u),
5504 BRW_CONDITIONAL_EQ
);
5509 unreachable("unknown texture opcode");
5512 if (instr
->op
== nir_texop_tg4
) {
5513 if (instr
->component
== 1 &&
5514 key_tex
->gather_channel_quirk_mask
& (1 << texture
)) {
5515 /* gather4 sampler is broken for green channel on RG32F --
5516 * we must ask for blue instead.
5518 header_bits
|= 2 << 16;
5520 header_bits
|= instr
->component
<< 16;
5524 fs_reg dst
= bld
.vgrf(brw_type_for_nir_type(devinfo
, instr
->dest_type
), 4);
5525 fs_inst
*inst
= bld
.emit(opcode
, dst
, srcs
, ARRAY_SIZE(srcs
));
5526 inst
->offset
= header_bits
;
5528 const unsigned dest_size
= nir_tex_instr_dest_size(instr
);
5529 if (devinfo
->gen
>= 9 &&
5530 instr
->op
!= nir_texop_tg4
&& instr
->op
!= nir_texop_query_levels
) {
5531 unsigned write_mask
= instr
->dest
.is_ssa
?
5532 nir_ssa_def_components_read(&instr
->dest
.ssa
):
5533 (1 << dest_size
) - 1;
5534 assert(write_mask
!= 0); /* dead code should have been eliminated */
5535 inst
->size_written
= util_last_bit(write_mask
) *
5536 inst
->dst
.component_size(inst
->exec_size
);
5538 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
5541 if (srcs
[TEX_LOGICAL_SRC_SHADOW_C
].file
!= BAD_FILE
)
5542 inst
->shadow_compare
= true;
5544 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
== 6)
5545 emit_gen6_gather_wa(key_tex
->gen6_gather_wa
[texture
], dst
);
5548 for (unsigned i
= 0; i
< dest_size
; i
++)
5549 nir_dest
[i
] = offset(dst
, bld
, i
);
5551 if (instr
->op
== nir_texop_query_levels
) {
5552 /* # levels is in .w */
5553 nir_dest
[0] = offset(dst
, bld
, 3);
5554 } else if (instr
->op
== nir_texop_txs
&&
5555 dest_size
>= 3 && devinfo
->gen
< 7) {
5556 /* Gen4-6 return 0 instead of 1 for single layer surfaces. */
5557 fs_reg depth
= offset(dst
, bld
, 2);
5558 nir_dest
[2] = vgrf(glsl_type::int_type
);
5559 bld
.emit_minmax(nir_dest
[2], depth
, brw_imm_d(1), BRW_CONDITIONAL_GE
);
5562 bld
.LOAD_PAYLOAD(get_nir_dest(instr
->dest
), nir_dest
, dest_size
, 0);
5566 fs_visitor::nir_emit_jump(const fs_builder
&bld
, nir_jump_instr
*instr
)
5568 switch (instr
->type
) {
5569 case nir_jump_break
:
5570 bld
.emit(BRW_OPCODE_BREAK
);
5572 case nir_jump_continue
:
5573 bld
.emit(BRW_OPCODE_CONTINUE
);
5575 case nir_jump_return
:
5577 unreachable("unknown jump");
5582 * This helper takes a source register and un/shuffles it into the destination
5585 * If source type size is smaller than destination type size the operation
5586 * needed is a component shuffle. The opposite case would be an unshuffle. If
5587 * source/destination type size is equal a shuffle is done that would be
5588 * equivalent to a simple MOV.
5590 * For example, if source is a 16-bit type and destination is 32-bit. A 3
5591 * components .xyz 16-bit vector on SIMD8 would be.
5593 * |x1|x2|x3|x4|x5|x6|x7|x8|y1|y2|y3|y4|y5|y6|y7|y8|
5594 * |z1|z2|z3|z4|z5|z6|z7|z8| | | | | | | | |
5596 * This helper will return the following 2 32-bit components with the 16-bit
5599 * |x1 y1|x2 y2|x3 y3|x4 y4|x5 y5|x6 y6|x7 y7|x8 y8|
5600 * |z1 |z2 |z3 |z4 |z5 |z6 |z7 |z8 |
5602 * For unshuffle, the example would be the opposite, a 64-bit type source
5603 * and a 32-bit destination. A 2 component .xy 64-bit vector on SIMD8
5606 * | x1l x1h | x2l x2h | x3l x3h | x4l x4h |
5607 * | x5l x5h | x6l x6h | x7l x7h | x8l x8h |
5608 * | y1l y1h | y2l y2h | y3l y3h | y4l y4h |
5609 * | y5l y5h | y6l y6h | y7l y7h | y8l y8h |
5611 * The returned result would be the following 4 32-bit components unshuffled:
5613 * | x1l | x2l | x3l | x4l | x5l | x6l | x7l | x8l |
5614 * | x1h | x2h | x3h | x4h | x5h | x6h | x7h | x8h |
5615 * | y1l | y2l | y3l | y4l | y5l | y6l | y7l | y8l |
5616 * | y1h | y2h | y3h | y4h | y5h | y6h | y7h | y8h |
5618 * - Source and destination register must not be overlapped.
5619 * - components units are measured in terms of the smaller type between
5620 * source and destination because we are un/shuffling the smaller
5621 * components from/into the bigger ones.
5622 * - first_component parameter allows skipping source components.
5625 shuffle_src_to_dst(const fs_builder
&bld
,
5628 uint32_t first_component
,
5629 uint32_t components
)
5631 if (type_sz(src
.type
) == type_sz(dst
.type
)) {
5632 assert(!regions_overlap(dst
,
5633 type_sz(dst
.type
) * bld
.dispatch_width() * components
,
5634 offset(src
, bld
, first_component
),
5635 type_sz(src
.type
) * bld
.dispatch_width() * components
));
5636 for (unsigned i
= 0; i
< components
; i
++) {
5637 bld
.MOV(retype(offset(dst
, bld
, i
), src
.type
),
5638 offset(src
, bld
, i
+ first_component
));
5640 } else if (type_sz(src
.type
) < type_sz(dst
.type
)) {
5641 /* Source is shuffled into destination */
5642 unsigned size_ratio
= type_sz(dst
.type
) / type_sz(src
.type
);
5643 assert(!regions_overlap(dst
,
5644 type_sz(dst
.type
) * bld
.dispatch_width() *
5645 DIV_ROUND_UP(components
, size_ratio
),
5646 offset(src
, bld
, first_component
),
5647 type_sz(src
.type
) * bld
.dispatch_width() * components
));
5649 brw_reg_type shuffle_type
=
5650 brw_reg_type_from_bit_size(8 * type_sz(src
.type
),
5651 BRW_REGISTER_TYPE_D
);
5652 for (unsigned i
= 0; i
< components
; i
++) {
5653 fs_reg shuffle_component_i
=
5654 subscript(offset(dst
, bld
, i
/ size_ratio
),
5655 shuffle_type
, i
% size_ratio
);
5656 bld
.MOV(shuffle_component_i
,
5657 retype(offset(src
, bld
, i
+ first_component
), shuffle_type
));
5660 /* Source is unshuffled into destination */
5661 unsigned size_ratio
= type_sz(src
.type
) / type_sz(dst
.type
);
5662 assert(!regions_overlap(dst
,
5663 type_sz(dst
.type
) * bld
.dispatch_width() * components
,
5664 offset(src
, bld
, first_component
/ size_ratio
),
5665 type_sz(src
.type
) * bld
.dispatch_width() *
5666 DIV_ROUND_UP(components
+ (first_component
% size_ratio
),
5669 brw_reg_type shuffle_type
=
5670 brw_reg_type_from_bit_size(8 * type_sz(dst
.type
),
5671 BRW_REGISTER_TYPE_D
);
5672 for (unsigned i
= 0; i
< components
; i
++) {
5673 fs_reg shuffle_component_i
=
5674 subscript(offset(src
, bld
, (first_component
+ i
) / size_ratio
),
5675 shuffle_type
, (first_component
+ i
) % size_ratio
);
5676 bld
.MOV(retype(offset(dst
, bld
, i
), shuffle_type
),
5677 shuffle_component_i
);
5683 shuffle_from_32bit_read(const fs_builder
&bld
,
5686 uint32_t first_component
,
5687 uint32_t components
)
5689 assert(type_sz(src
.type
) == 4);
5691 /* This function takes components in units of the destination type while
5692 * shuffle_src_to_dst takes components in units of the smallest type
5694 if (type_sz(dst
.type
) > 4) {
5695 assert(type_sz(dst
.type
) == 8);
5696 first_component
*= 2;
5700 shuffle_src_to_dst(bld
, dst
, src
, first_component
, components
);
5704 setup_imm_df(const fs_builder
&bld
, double v
)
5706 const struct gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5707 assert(devinfo
->gen
>= 7);
5709 if (devinfo
->gen
>= 8)
5710 return brw_imm_df(v
);
5712 /* gen7.5 does not support DF immediates straighforward but the DIM
5713 * instruction allows to set the 64-bit immediate value.
5715 if (devinfo
->is_haswell
) {
5716 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5717 fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_DF
, 1);
5718 ubld
.DIM(dst
, brw_imm_df(v
));
5719 return component(dst
, 0);
5722 /* gen7 does not support DF immediates, so we generate a 64-bit constant by
5723 * writing the low 32-bit of the constant to suboffset 0 of a VGRF and
5724 * the high 32-bit to suboffset 4 and then applying a stride of 0.
5726 * Alternatively, we could also produce a normal VGRF (without stride 0)
5727 * by writing to all the channels in the VGRF, however, that would hit the
5728 * gen7 bug where we have to split writes that span more than 1 register
5729 * into instructions with a width of 4 (otherwise the write to the second
5730 * register written runs into an execmask hardware bug) which isn't very
5743 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5744 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
5745 ubld
.MOV(tmp
, brw_imm_ud(di
.i1
));
5746 ubld
.MOV(horiz_offset(tmp
, 1), brw_imm_ud(di
.i2
));
5748 return component(retype(tmp
, BRW_REGISTER_TYPE_DF
), 0);
5752 setup_imm_b(const fs_builder
&bld
, int8_t v
)
5754 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_B
);
5755 bld
.MOV(tmp
, brw_imm_w(v
));
5760 setup_imm_ub(const fs_builder
&bld
, uint8_t v
)
5762 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UB
);
5763 bld
.MOV(tmp
, brw_imm_uw(v
));