2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "compiler/glsl/ir.h"
26 #include "brw_fs_surface_builder.h"
30 using namespace brw::surface_access
;
33 fs_visitor::emit_nir_code()
35 /* emit the arrays used for inputs and outputs - load/store intrinsics will
36 * be converted to reads/writes of these arrays
40 nir_emit_system_values();
42 /* get the main function and emit it */
43 nir_foreach_function(function
, nir
) {
44 assert(strcmp(function
->name
, "main") == 0);
45 assert(function
->impl
);
46 nir_emit_impl(function
->impl
);
51 fs_visitor::nir_setup_outputs()
53 if (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_FRAGMENT
)
56 unsigned vec4s
[VARYING_SLOT_TESS_MAX
] = { 0, };
58 /* Calculate the size of output registers in a separate pass, before
59 * allocating them. With ARB_enhanced_layouts, multiple output variables
60 * may occupy the same slot, but have different type sizes.
62 nir_foreach_variable(var
, &nir
->outputs
) {
63 const int loc
= var
->data
.driver_location
;
64 const unsigned var_vec4s
=
65 var
->data
.compact
? DIV_ROUND_UP(glsl_get_length(var
->type
), 4)
66 : type_size_vec4(var
->type
);
67 vec4s
[loc
] = MAX2(vec4s
[loc
], var_vec4s
);
70 nir_foreach_variable(var
, &nir
->outputs
) {
71 const int loc
= var
->data
.driver_location
;
72 if (outputs
[loc
].file
== BAD_FILE
) {
73 fs_reg reg
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4 * vec4s
[loc
]);
74 for (unsigned i
= 0; i
< vec4s
[loc
]; i
++) {
75 outputs
[loc
+ i
] = offset(reg
, bld
, 4 * i
);
82 fs_visitor::nir_setup_uniforms()
84 /* Only the first compile gets to set up uniforms. */
85 if (push_constant_loc
) {
86 assert(pull_constant_loc
);
90 uniforms
= nir
->num_uniforms
/ 4;
92 if (stage
== MESA_SHADER_COMPUTE
) {
93 /* Add a uniform for the thread local id. It must be the last uniform
96 assert(uniforms
== prog_data
->nr_params
);
97 uint32_t *param
= brw_stage_prog_data_add_params(prog_data
, 1);
98 *param
= BRW_PARAM_BUILTIN_SUBGROUP_ID
;
99 subgroup_id
= fs_reg(UNIFORM
, uniforms
++, BRW_REGISTER_TYPE_UD
);
104 emit_system_values_block(nir_block
*block
, fs_visitor
*v
)
108 nir_foreach_instr(instr
, block
) {
109 if (instr
->type
!= nir_instr_type_intrinsic
)
112 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
113 switch (intrin
->intrinsic
) {
114 case nir_intrinsic_load_vertex_id
:
115 unreachable("should be lowered by lower_vertex_id().");
117 case nir_intrinsic_load_vertex_id_zero_base
:
118 case nir_intrinsic_load_base_vertex
:
119 case nir_intrinsic_load_is_indexed_draw
:
120 case nir_intrinsic_load_first_vertex
:
121 case nir_intrinsic_load_instance_id
:
122 case nir_intrinsic_load_base_instance
:
123 case nir_intrinsic_load_draw_id
:
124 unreachable("should be lowered by brw_nir_lower_vs_inputs().");
126 case nir_intrinsic_load_invocation_id
:
127 if (v
->stage
== MESA_SHADER_TESS_CTRL
)
129 assert(v
->stage
== MESA_SHADER_GEOMETRY
);
130 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
131 if (reg
->file
== BAD_FILE
) {
132 const fs_builder abld
= v
->bld
.annotate("gl_InvocationID", NULL
);
133 fs_reg
g1(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
134 fs_reg iid
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
135 abld
.SHR(iid
, g1
, brw_imm_ud(27u));
140 case nir_intrinsic_load_sample_pos
:
141 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
142 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
143 if (reg
->file
== BAD_FILE
)
144 *reg
= *v
->emit_samplepos_setup();
147 case nir_intrinsic_load_sample_id
:
148 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
149 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
150 if (reg
->file
== BAD_FILE
)
151 *reg
= *v
->emit_sampleid_setup();
154 case nir_intrinsic_load_sample_mask_in
:
155 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
156 assert(v
->devinfo
->gen
>= 7);
157 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
158 if (reg
->file
== BAD_FILE
)
159 *reg
= *v
->emit_samplemaskin_setup();
162 case nir_intrinsic_load_work_group_id
:
163 assert(v
->stage
== MESA_SHADER_COMPUTE
);
164 reg
= &v
->nir_system_values
[SYSTEM_VALUE_WORK_GROUP_ID
];
165 if (reg
->file
== BAD_FILE
)
166 *reg
= *v
->emit_cs_work_group_id_setup();
169 case nir_intrinsic_load_helper_invocation
:
170 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
171 reg
= &v
->nir_system_values
[SYSTEM_VALUE_HELPER_INVOCATION
];
172 if (reg
->file
== BAD_FILE
) {
173 const fs_builder abld
=
174 v
->bld
.annotate("gl_HelperInvocation", NULL
);
176 /* On Gen6+ (gl_HelperInvocation is only exposed on Gen7+) the
177 * pixel mask is in g1.7 of the thread payload.
179 * We move the per-channel pixel enable bit to the low bit of each
180 * channel by shifting the byte containing the pixel mask by the
181 * vector immediate 0x76543210UV.
183 * The region of <1,8,0> reads only 1 byte (the pixel masks for
184 * subspans 0 and 1) in SIMD8 and an additional byte (the pixel
185 * masks for 2 and 3) in SIMD16.
187 fs_reg shifted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
189 stride(byte_offset(retype(brw_vec1_grf(1, 0),
190 BRW_REGISTER_TYPE_UB
), 28),
192 brw_imm_v(0x76543210));
194 /* A set bit in the pixel mask means the channel is enabled, but
195 * that is the opposite of gl_HelperInvocation so we need to invert
198 * The negate source-modifier bit of logical instructions on Gen8+
199 * performs 1's complement negation, so we can use that instead of
202 fs_reg inverted
= negate(shifted
);
203 if (v
->devinfo
->gen
< 8) {
204 inverted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
205 abld
.NOT(inverted
, shifted
);
208 /* We then resolve the 0/1 result to 0/~0 boolean values by ANDing
209 * with 1 and negating.
211 fs_reg anded
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
212 abld
.AND(anded
, inverted
, brw_imm_uw(1));
214 fs_reg dst
= abld
.vgrf(BRW_REGISTER_TYPE_D
, 1);
215 abld
.MOV(dst
, negate(retype(anded
, BRW_REGISTER_TYPE_D
)));
229 fs_visitor::nir_emit_system_values()
231 nir_system_values
= ralloc_array(mem_ctx
, fs_reg
, SYSTEM_VALUE_MAX
);
232 for (unsigned i
= 0; i
< SYSTEM_VALUE_MAX
; i
++) {
233 nir_system_values
[i
] = fs_reg();
236 /* Always emit SUBGROUP_INVOCATION. Dead code will clean it up if we
237 * never end up using it.
240 const fs_builder abld
= bld
.annotate("gl_SubgroupInvocation", NULL
);
241 fs_reg
®
= nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
];
242 reg
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
244 const fs_builder allbld8
= abld
.group(8, 0).exec_all();
245 allbld8
.MOV(reg
, brw_imm_v(0x76543210));
246 if (dispatch_width
> 8)
247 allbld8
.ADD(byte_offset(reg
, 16), reg
, brw_imm_uw(8u));
248 if (dispatch_width
> 16) {
249 const fs_builder allbld16
= abld
.group(16, 0).exec_all();
250 allbld16
.ADD(byte_offset(reg
, 32), reg
, brw_imm_uw(16u));
254 nir_foreach_function(function
, nir
) {
255 assert(strcmp(function
->name
, "main") == 0);
256 assert(function
->impl
);
257 nir_foreach_block(block
, function
->impl
) {
258 emit_system_values_block(block
, this);
264 * Returns a type based on a reference_type (word, float, half-float) and a
267 * Reference BRW_REGISTER_TYPE are HF,F,DF,W,D,UW,UD.
269 * @FIXME: 64-bit return types are always DF on integer types to maintain
270 * compability with uses of DF previously to the introduction of int64
274 brw_reg_type_from_bit_size(const unsigned bit_size
,
275 const brw_reg_type reference_type
)
277 switch(reference_type
) {
278 case BRW_REGISTER_TYPE_HF
:
279 case BRW_REGISTER_TYPE_F
:
280 case BRW_REGISTER_TYPE_DF
:
283 return BRW_REGISTER_TYPE_HF
;
285 return BRW_REGISTER_TYPE_F
;
287 return BRW_REGISTER_TYPE_DF
;
289 unreachable("Invalid bit size");
291 case BRW_REGISTER_TYPE_W
:
292 case BRW_REGISTER_TYPE_D
:
293 case BRW_REGISTER_TYPE_Q
:
296 return BRW_REGISTER_TYPE_W
;
298 return BRW_REGISTER_TYPE_D
;
300 return BRW_REGISTER_TYPE_Q
;
302 unreachable("Invalid bit size");
304 case BRW_REGISTER_TYPE_UW
:
305 case BRW_REGISTER_TYPE_UD
:
306 case BRW_REGISTER_TYPE_UQ
:
309 return BRW_REGISTER_TYPE_UW
;
311 return BRW_REGISTER_TYPE_UD
;
313 return BRW_REGISTER_TYPE_UQ
;
315 unreachable("Invalid bit size");
318 unreachable("Unknown type");
323 fs_visitor::nir_emit_impl(nir_function_impl
*impl
)
325 nir_locals
= ralloc_array(mem_ctx
, fs_reg
, impl
->reg_alloc
);
326 for (unsigned i
= 0; i
< impl
->reg_alloc
; i
++) {
327 nir_locals
[i
] = fs_reg();
330 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
331 unsigned array_elems
=
332 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
333 unsigned size
= array_elems
* reg
->num_components
;
334 const brw_reg_type reg_type
=
335 brw_reg_type_from_bit_size(reg
->bit_size
, BRW_REGISTER_TYPE_F
);
336 nir_locals
[reg
->index
] = bld
.vgrf(reg_type
, size
);
339 nir_ssa_values
= reralloc(mem_ctx
, nir_ssa_values
, fs_reg
,
342 nir_emit_cf_list(&impl
->body
);
346 fs_visitor::nir_emit_cf_list(exec_list
*list
)
348 exec_list_validate(list
);
349 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
350 switch (node
->type
) {
352 nir_emit_if(nir_cf_node_as_if(node
));
355 case nir_cf_node_loop
:
356 nir_emit_loop(nir_cf_node_as_loop(node
));
359 case nir_cf_node_block
:
360 nir_emit_block(nir_cf_node_as_block(node
));
364 unreachable("Invalid CFG node block");
370 fs_visitor::nir_emit_if(nir_if
*if_stmt
)
372 /* first, put the condition into f0 */
373 fs_inst
*inst
= bld
.MOV(bld
.null_reg_d(),
374 retype(get_nir_src(if_stmt
->condition
),
375 BRW_REGISTER_TYPE_D
));
376 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
378 bld
.IF(BRW_PREDICATE_NORMAL
);
380 nir_emit_cf_list(&if_stmt
->then_list
);
382 /* note: if the else is empty, dead CF elimination will remove it */
383 bld
.emit(BRW_OPCODE_ELSE
);
385 nir_emit_cf_list(&if_stmt
->else_list
);
387 bld
.emit(BRW_OPCODE_ENDIF
);
391 fs_visitor::nir_emit_loop(nir_loop
*loop
)
393 bld
.emit(BRW_OPCODE_DO
);
395 nir_emit_cf_list(&loop
->body
);
397 bld
.emit(BRW_OPCODE_WHILE
);
401 fs_visitor::nir_emit_block(nir_block
*block
)
403 nir_foreach_instr(instr
, block
) {
404 nir_emit_instr(instr
);
409 fs_visitor::nir_emit_instr(nir_instr
*instr
)
411 const fs_builder abld
= bld
.annotate(NULL
, instr
);
413 switch (instr
->type
) {
414 case nir_instr_type_alu
:
415 nir_emit_alu(abld
, nir_instr_as_alu(instr
));
418 case nir_instr_type_intrinsic
:
420 case MESA_SHADER_VERTEX
:
421 nir_emit_vs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
423 case MESA_SHADER_TESS_CTRL
:
424 nir_emit_tcs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
426 case MESA_SHADER_TESS_EVAL
:
427 nir_emit_tes_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
429 case MESA_SHADER_GEOMETRY
:
430 nir_emit_gs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
432 case MESA_SHADER_FRAGMENT
:
433 nir_emit_fs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
435 case MESA_SHADER_COMPUTE
:
436 nir_emit_cs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
439 unreachable("unsupported shader stage");
443 case nir_instr_type_tex
:
444 nir_emit_texture(abld
, nir_instr_as_tex(instr
));
447 case nir_instr_type_load_const
:
448 nir_emit_load_const(abld
, nir_instr_as_load_const(instr
));
451 case nir_instr_type_ssa_undef
:
452 /* We create a new VGRF for undefs on every use (by handling
453 * them in get_nir_src()), rather than for each definition.
454 * This helps register coalescing eliminate MOVs from undef.
458 case nir_instr_type_jump
:
459 nir_emit_jump(abld
, nir_instr_as_jump(instr
));
463 unreachable("unknown instruction type");
468 * Recognizes a parent instruction of nir_op_extract_* and changes the type to
472 fs_visitor::optimize_extract_to_float(nir_alu_instr
*instr
,
473 const fs_reg
&result
)
475 if (!instr
->src
[0].src
.is_ssa
||
476 !instr
->src
[0].src
.ssa
->parent_instr
)
479 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
482 nir_alu_instr
*src0
=
483 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
485 if (src0
->op
!= nir_op_extract_u8
&& src0
->op
!= nir_op_extract_u16
&&
486 src0
->op
!= nir_op_extract_i8
&& src0
->op
!= nir_op_extract_i16
)
489 nir_const_value
*element
= nir_src_as_const_value(src0
->src
[1].src
);
490 assert(element
!= NULL
);
492 /* Element type to extract.*/
493 const brw_reg_type type
= brw_int_type(
494 src0
->op
== nir_op_extract_u16
|| src0
->op
== nir_op_extract_i16
? 2 : 1,
495 src0
->op
== nir_op_extract_i16
|| src0
->op
== nir_op_extract_i8
);
497 fs_reg op0
= get_nir_src(src0
->src
[0].src
);
498 op0
.type
= brw_type_for_nir_type(devinfo
,
499 (nir_alu_type
)(nir_op_infos
[src0
->op
].input_types
[0] |
500 nir_src_bit_size(src0
->src
[0].src
)));
501 op0
= offset(op0
, bld
, src0
->src
[0].swizzle
[0]);
503 set_saturate(instr
->dest
.saturate
,
504 bld
.MOV(result
, subscript(op0
, type
, element
->u32
[0])));
509 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr
*instr
,
510 const fs_reg
&result
)
512 if (!instr
->src
[0].src
.is_ssa
||
513 instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
516 nir_intrinsic_instr
*src0
=
517 nir_instr_as_intrinsic(instr
->src
[0].src
.ssa
->parent_instr
);
519 if (src0
->intrinsic
!= nir_intrinsic_load_front_face
)
522 nir_const_value
*value1
= nir_src_as_const_value(instr
->src
[1].src
);
523 if (!value1
|| fabsf(value1
->f32
[0]) != 1.0f
)
526 nir_const_value
*value2
= nir_src_as_const_value(instr
->src
[2].src
);
527 if (!value2
|| fabsf(value2
->f32
[0]) != 1.0f
)
530 fs_reg tmp
= vgrf(glsl_type::int_type
);
532 if (devinfo
->gen
>= 6) {
533 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
534 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
536 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
538 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
539 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
541 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
543 * This negation looks like it's safe in practice, because bits 0:4 will
544 * surely be TRIANGLES
547 if (value1
->f32
[0] == -1.0f
) {
551 bld
.OR(subscript(tmp
, BRW_REGISTER_TYPE_W
, 1),
552 g0
, brw_imm_uw(0x3f80));
554 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
555 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
557 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
559 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
560 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
562 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
564 * This negation looks like it's safe in practice, because bits 0:4 will
565 * surely be TRIANGLES
568 if (value1
->f32
[0] == -1.0f
) {
572 bld
.OR(tmp
, g1_6
, brw_imm_d(0x3f800000));
574 bld
.AND(retype(result
, BRW_REGISTER_TYPE_D
), tmp
, brw_imm_d(0xbf800000));
580 emit_find_msb_using_lzd(const fs_builder
&bld
,
581 const fs_reg
&result
,
589 /* LZD of an absolute value source almost always does the right
590 * thing. There are two problem values:
592 * * 0x80000000. Since abs(0x80000000) == 0x80000000, LZD returns
593 * 0. However, findMSB(int(0x80000000)) == 30.
595 * * 0xffffffff. Since abs(0xffffffff) == 1, LZD returns
596 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
598 * For a value of zero or negative one, -1 will be returned.
600 * * Negative powers of two. LZD(abs(-(1<<x))) returns x, but
601 * findMSB(-(1<<x)) should return x-1.
603 * For all negative number cases, including 0x80000000 and
604 * 0xffffffff, the correct value is obtained from LZD if instead of
605 * negating the (already negative) value the logical-not is used. A
606 * conditonal logical-not can be achieved in two instructions.
608 temp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
610 bld
.ASR(temp
, src
, brw_imm_d(31));
611 bld
.XOR(temp
, temp
, src
);
614 bld
.LZD(retype(result
, BRW_REGISTER_TYPE_UD
),
615 retype(temp
, BRW_REGISTER_TYPE_UD
));
617 /* LZD counts from the MSB side, while GLSL's findMSB() wants the count
618 * from the LSB side. Subtract the result from 31 to convert the MSB
619 * count into an LSB count. If no bits are set, LZD will return 32.
620 * 31-32 = -1, which is exactly what findMSB() is supposed to return.
622 inst
= bld
.ADD(result
, retype(result
, BRW_REGISTER_TYPE_D
), brw_imm_d(31));
623 inst
->src
[0].negate
= true;
627 brw_rnd_mode_from_nir_op (const nir_op op
) {
629 case nir_op_f2f16_rtz
:
630 return BRW_RND_MODE_RTZ
;
631 case nir_op_f2f16_rtne
:
632 return BRW_RND_MODE_RTNE
;
634 unreachable("Operation doesn't support rounding mode");
639 fs_visitor::nir_emit_alu(const fs_builder
&bld
, nir_alu_instr
*instr
)
641 struct brw_wm_prog_key
*fs_key
= (struct brw_wm_prog_key
*) this->key
;
644 fs_reg result
= get_nir_dest(instr
->dest
.dest
);
645 result
.type
= brw_type_for_nir_type(devinfo
,
646 (nir_alu_type
)(nir_op_infos
[instr
->op
].output_type
|
647 nir_dest_bit_size(instr
->dest
.dest
)));
650 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
651 op
[i
] = get_nir_src(instr
->src
[i
].src
);
652 op
[i
].type
= brw_type_for_nir_type(devinfo
,
653 (nir_alu_type
)(nir_op_infos
[instr
->op
].input_types
[i
] |
654 nir_src_bit_size(instr
->src
[i
].src
)));
655 op
[i
].abs
= instr
->src
[i
].abs
;
656 op
[i
].negate
= instr
->src
[i
].negate
;
659 /* We get a bunch of mov's out of the from_ssa pass and they may still
660 * be vectorized. We'll handle them as a special-case. We'll also
661 * handle vecN here because it's basically the same thing.
669 fs_reg temp
= result
;
670 bool need_extra_copy
= false;
671 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
672 if (!instr
->src
[i
].src
.is_ssa
&&
673 instr
->dest
.dest
.reg
.reg
== instr
->src
[i
].src
.reg
.reg
) {
674 need_extra_copy
= true;
675 temp
= bld
.vgrf(result
.type
, 4);
680 for (unsigned i
= 0; i
< 4; i
++) {
681 if (!(instr
->dest
.write_mask
& (1 << i
)))
684 if (instr
->op
== nir_op_imov
|| instr
->op
== nir_op_fmov
) {
685 inst
= bld
.MOV(offset(temp
, bld
, i
),
686 offset(op
[0], bld
, instr
->src
[0].swizzle
[i
]));
688 inst
= bld
.MOV(offset(temp
, bld
, i
),
689 offset(op
[i
], bld
, instr
->src
[i
].swizzle
[0]));
691 inst
->saturate
= instr
->dest
.saturate
;
694 /* In this case the source and destination registers were the same,
695 * so we need to insert an extra set of moves in order to deal with
698 if (need_extra_copy
) {
699 for (unsigned i
= 0; i
< 4; i
++) {
700 if (!(instr
->dest
.write_mask
& (1 << i
)))
703 bld
.MOV(offset(result
, bld
, i
), offset(temp
, bld
, i
));
712 /* At this point, we have dealt with any instruction that operates on
713 * more than a single channel. Therefore, we can just adjust the source
714 * and destination registers for that channel and emit the instruction.
716 unsigned channel
= 0;
717 if (nir_op_infos
[instr
->op
].output_size
== 0) {
718 /* Since NIR is doing the scalarizing for us, we should only ever see
719 * vectorized operations with a single channel.
721 assert(_mesa_bitcount(instr
->dest
.write_mask
) == 1);
722 channel
= ffs(instr
->dest
.write_mask
) - 1;
724 result
= offset(result
, bld
, channel
);
727 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
728 assert(nir_op_infos
[instr
->op
].input_sizes
[i
] < 2);
729 op
[i
] = offset(op
[i
], bld
, instr
->src
[i
].swizzle
[channel
]);
735 if (optimize_extract_to_float(instr
, result
))
737 inst
= bld
.MOV(result
, op
[0]);
738 inst
->saturate
= instr
->dest
.saturate
;
741 case nir_op_f2f16_rtne
:
742 case nir_op_f2f16_rtz
:
743 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(),
744 brw_imm_d(brw_rnd_mode_from_nir_op(instr
->op
)));
747 /* In theory, it would be better to use BRW_OPCODE_F32TO16. Depending
748 * on the HW gen, it is a special hw opcode or just a MOV, and
749 * brw_F32TO16 (at brw_eu_emit) would do the work to chose.
751 * But if we want to use that opcode, we need to provide support on
752 * different optimizations and lowerings. As right now HF support is
753 * only for gen8+, it will be better to use directly the MOV, and use
754 * BRW_OPCODE_F32TO16 when/if we work for HF support on gen7.
757 case nir_op_f2f16_undef
:
760 /* TODO: Fixing aligment rules for conversions from 32-bits to
761 * 16-bit types should be moved to lower_conversions
763 fs_reg tmp
= bld
.vgrf(op
[0].type
, 1);
764 tmp
= subscript(tmp
, result
.type
, 0);
765 inst
= bld
.MOV(tmp
, op
[0]);
766 inst
->saturate
= instr
->dest
.saturate
;
767 inst
= bld
.MOV(result
, tmp
);
768 inst
->saturate
= instr
->dest
.saturate
;
779 /* CHV PRM, vol07, 3D Media GPGPU Engine, Register Region Restrictions:
781 * "When source or destination is 64b (...), regioning in Align1
782 * must follow these rules:
784 * 1. Source and destination horizontal stride must be aligned to
788 * This means that 32-bit to 64-bit conversions need to have the 32-bit
789 * data elements aligned to 64-bit. This restriction does not apply to
792 if (nir_dest_bit_size(instr
->dest
.dest
) == 64 &&
793 nir_src_bit_size(instr
->src
[0].src
) == 32 &&
794 (devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
))) {
795 fs_reg tmp
= bld
.vgrf(result
.type
, 1);
796 tmp
= subscript(tmp
, op
[0].type
, 0);
797 inst
= bld
.MOV(tmp
, op
[0]);
798 inst
= bld
.MOV(result
, tmp
);
799 inst
->saturate
= instr
->dest
.saturate
;
808 inst
= bld
.MOV(result
, op
[0]);
809 inst
->saturate
= instr
->dest
.saturate
;
814 /* Straightforward since the source can be assumed to be
817 set_condmod(BRW_CONDITIONAL_NZ
, bld
.MOV(result
, op
[0]));
818 set_predicate(BRW_PREDICATE_NORMAL
, bld
.MOV(result
, brw_imm_f(1.0f
)));
820 } else if (type_sz(op
[0].type
) < 8) {
821 /* AND(val, 0x80000000) gives the sign bit.
823 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
826 bld
.CMP(bld
.null_reg_f(), op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
828 fs_reg result_int
= retype(result
, BRW_REGISTER_TYPE_UD
);
829 op
[0].type
= BRW_REGISTER_TYPE_UD
;
830 result
.type
= BRW_REGISTER_TYPE_UD
;
831 bld
.AND(result_int
, op
[0], brw_imm_ud(0x80000000u
));
833 inst
= bld
.OR(result_int
, result_int
, brw_imm_ud(0x3f800000u
));
834 inst
->predicate
= BRW_PREDICATE_NORMAL
;
835 if (instr
->dest
.saturate
) {
836 inst
= bld
.MOV(result
, result
);
837 inst
->saturate
= true;
840 /* For doubles we do the same but we need to consider:
842 * - 2-src instructions can't operate with 64-bit immediates
843 * - The sign is encoded in the high 32-bit of each DF
844 * - We need to produce a DF result.
847 fs_reg zero
= vgrf(glsl_type::double_type
);
848 bld
.MOV(zero
, setup_imm_df(bld
, 0.0));
849 bld
.CMP(bld
.null_reg_df(), op
[0], zero
, BRW_CONDITIONAL_NZ
);
851 bld
.MOV(result
, zero
);
853 fs_reg r
= subscript(result
, BRW_REGISTER_TYPE_UD
, 1);
854 bld
.AND(r
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1),
855 brw_imm_ud(0x80000000u
));
857 set_predicate(BRW_PREDICATE_NORMAL
,
858 bld
.OR(r
, r
, brw_imm_ud(0x3ff00000u
)));
860 if (instr
->dest
.saturate
) {
861 inst
= bld
.MOV(result
, result
);
862 inst
->saturate
= true;
869 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
870 * -> non-negative val generates 0x00000000.
871 * Predicated OR sets 1 if val is positive.
873 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
874 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_G
);
875 bld
.ASR(result
, op
[0], brw_imm_d(31));
876 inst
= bld
.OR(result
, result
, brw_imm_d(1));
877 inst
->predicate
= BRW_PREDICATE_NORMAL
;
881 inst
= bld
.emit(SHADER_OPCODE_RCP
, result
, op
[0]);
882 inst
->saturate
= instr
->dest
.saturate
;
886 inst
= bld
.emit(SHADER_OPCODE_EXP2
, result
, op
[0]);
887 inst
->saturate
= instr
->dest
.saturate
;
891 inst
= bld
.emit(SHADER_OPCODE_LOG2
, result
, op
[0]);
892 inst
->saturate
= instr
->dest
.saturate
;
896 inst
= bld
.emit(SHADER_OPCODE_SIN
, result
, op
[0]);
897 inst
->saturate
= instr
->dest
.saturate
;
901 inst
= bld
.emit(SHADER_OPCODE_COS
, result
, op
[0]);
902 inst
->saturate
= instr
->dest
.saturate
;
906 if (fs_key
->high_quality_derivatives
) {
907 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
909 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
911 inst
->saturate
= instr
->dest
.saturate
;
913 case nir_op_fddx_fine
:
914 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
915 inst
->saturate
= instr
->dest
.saturate
;
917 case nir_op_fddx_coarse
:
918 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
919 inst
->saturate
= instr
->dest
.saturate
;
922 if (fs_key
->high_quality_derivatives
) {
923 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
925 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
927 inst
->saturate
= instr
->dest
.saturate
;
929 case nir_op_fddy_fine
:
930 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
931 inst
->saturate
= instr
->dest
.saturate
;
933 case nir_op_fddy_coarse
:
934 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
935 inst
->saturate
= instr
->dest
.saturate
;
940 inst
= bld
.ADD(result
, op
[0], op
[1]);
941 inst
->saturate
= instr
->dest
.saturate
;
945 inst
= bld
.MUL(result
, op
[0], op
[1]);
946 inst
->saturate
= instr
->dest
.saturate
;
950 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
951 bld
.MUL(result
, op
[0], op
[1]);
954 case nir_op_imul_high
:
955 case nir_op_umul_high
:
956 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
957 bld
.emit(SHADER_OPCODE_MULH
, result
, op
[0], op
[1]);
962 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
963 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, result
, op
[0], op
[1]);
966 case nir_op_uadd_carry
:
967 unreachable("Should have been lowered by carry_to_arith().");
969 case nir_op_usub_borrow
:
970 unreachable("Should have been lowered by borrow_to_arith().");
974 /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
975 * appears that our hardware just does the right thing for signed
978 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
979 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
983 /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
984 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
986 /* Math instructions don't support conditional mod */
987 inst
= bld
.MOV(bld
.null_reg_d(), result
);
988 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
990 /* Now, we need to determine if signs of the sources are different.
991 * When we XOR the sources, the top bit is 0 if they are the same and 1
992 * if they are different. We can then use a conditional modifier to
993 * turn that into a predicate. This leads us to an XOR.l instruction.
995 * Technically, according to the PRM, you're not allowed to use .l on a
996 * XOR instruction. However, emperical experiments and Curro's reading
997 * of the simulator source both indicate that it's safe.
999 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1000 inst
= bld
.XOR(tmp
, op
[0], op
[1]);
1001 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1002 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1004 /* If the result of the initial remainder operation is non-zero and the
1005 * two sources have different signs, add in a copy of op[1] to get the
1006 * final integer modulus value.
1008 inst
= bld
.ADD(result
, result
, op
[1]);
1009 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1017 fs_reg dest
= result
;
1018 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
1019 dest
= bld
.vgrf(BRW_REGISTER_TYPE_DF
, 1);
1021 brw_conditional_mod cond
;
1022 switch (instr
->op
) {
1024 cond
= BRW_CONDITIONAL_L
;
1027 cond
= BRW_CONDITIONAL_GE
;
1030 cond
= BRW_CONDITIONAL_Z
;
1033 cond
= BRW_CONDITIONAL_NZ
;
1036 unreachable("bad opcode");
1038 bld
.CMP(dest
, op
[0], op
[1], cond
);
1039 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
1040 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1051 fs_reg dest
= result
;
1052 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
1053 dest
= bld
.vgrf(BRW_REGISTER_TYPE_UQ
, 1);
1056 brw_conditional_mod cond
;
1057 switch (instr
->op
) {
1060 cond
= BRW_CONDITIONAL_L
;
1064 cond
= BRW_CONDITIONAL_GE
;
1067 cond
= BRW_CONDITIONAL_Z
;
1070 cond
= BRW_CONDITIONAL_NZ
;
1073 unreachable("bad opcode");
1075 bld
.CMP(dest
, op
[0], op
[1], cond
);
1076 if (nir_src_bit_size(instr
->src
[0].src
) > 32) {
1077 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1083 if (devinfo
->gen
>= 8) {
1084 op
[0] = resolve_source_modifiers(op
[0]);
1086 bld
.NOT(result
, op
[0]);
1089 if (devinfo
->gen
>= 8) {
1090 op
[0] = resolve_source_modifiers(op
[0]);
1091 op
[1] = resolve_source_modifiers(op
[1]);
1093 bld
.XOR(result
, op
[0], op
[1]);
1096 if (devinfo
->gen
>= 8) {
1097 op
[0] = resolve_source_modifiers(op
[0]);
1098 op
[1] = resolve_source_modifiers(op
[1]);
1100 bld
.OR(result
, op
[0], op
[1]);
1103 if (devinfo
->gen
>= 8) {
1104 op
[0] = resolve_source_modifiers(op
[0]);
1105 op
[1] = resolve_source_modifiers(op
[1]);
1107 bld
.AND(result
, op
[0], op
[1]);
1113 case nir_op_ball_fequal2
:
1114 case nir_op_ball_iequal2
:
1115 case nir_op_ball_fequal3
:
1116 case nir_op_ball_iequal3
:
1117 case nir_op_ball_fequal4
:
1118 case nir_op_ball_iequal4
:
1119 case nir_op_bany_fnequal2
:
1120 case nir_op_bany_inequal2
:
1121 case nir_op_bany_fnequal3
:
1122 case nir_op_bany_inequal3
:
1123 case nir_op_bany_fnequal4
:
1124 case nir_op_bany_inequal4
:
1125 unreachable("Lowered by nir_lower_alu_reductions");
1127 case nir_op_fnoise1_1
:
1128 case nir_op_fnoise1_2
:
1129 case nir_op_fnoise1_3
:
1130 case nir_op_fnoise1_4
:
1131 case nir_op_fnoise2_1
:
1132 case nir_op_fnoise2_2
:
1133 case nir_op_fnoise2_3
:
1134 case nir_op_fnoise2_4
:
1135 case nir_op_fnoise3_1
:
1136 case nir_op_fnoise3_2
:
1137 case nir_op_fnoise3_3
:
1138 case nir_op_fnoise3_4
:
1139 case nir_op_fnoise4_1
:
1140 case nir_op_fnoise4_2
:
1141 case nir_op_fnoise4_3
:
1142 case nir_op_fnoise4_4
:
1143 unreachable("not reached: should be handled by lower_noise");
1146 unreachable("not reached: should be handled by ldexp_to_arith()");
1149 inst
= bld
.emit(SHADER_OPCODE_SQRT
, result
, op
[0]);
1150 inst
->saturate
= instr
->dest
.saturate
;
1154 inst
= bld
.emit(SHADER_OPCODE_RSQ
, result
, op
[0]);
1155 inst
->saturate
= instr
->dest
.saturate
;
1160 bld
.MOV(result
, negate(op
[0]));
1165 if (nir_src_bit_size(instr
->src
[0].src
) == 64) {
1166 /* two-argument instructions can't take 64-bit immediates */
1170 if (instr
->op
== nir_op_f2b
) {
1171 zero
= vgrf(glsl_type::double_type
);
1172 tmp
= vgrf(glsl_type::double_type
);
1173 bld
.MOV(zero
, setup_imm_df(bld
, 0.0));
1175 zero
= vgrf(glsl_type::int64_t_type
);
1176 tmp
= vgrf(glsl_type::int64_t_type
);
1177 bld
.MOV(zero
, brw_imm_q(0));
1180 /* A SIMD16 execution needs to be split in two instructions, so use
1181 * a vgrf instead of the flag register as dst so instruction splitting
1184 bld
.CMP(tmp
, op
[0], zero
, BRW_CONDITIONAL_NZ
);
1185 bld
.MOV(result
, subscript(tmp
, BRW_REGISTER_TYPE_UD
, 0));
1187 if (instr
->op
== nir_op_f2b
) {
1188 bld
.CMP(result
, op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
1190 bld
.CMP(result
, op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1196 inst
= bld
.RNDZ(result
, op
[0]);
1197 inst
->saturate
= instr
->dest
.saturate
;
1200 case nir_op_fceil
: {
1201 op
[0].negate
= !op
[0].negate
;
1202 fs_reg temp
= vgrf(glsl_type::float_type
);
1203 bld
.RNDD(temp
, op
[0]);
1205 inst
= bld
.MOV(result
, temp
);
1206 inst
->saturate
= instr
->dest
.saturate
;
1210 inst
= bld
.RNDD(result
, op
[0]);
1211 inst
->saturate
= instr
->dest
.saturate
;
1214 inst
= bld
.FRC(result
, op
[0]);
1215 inst
->saturate
= instr
->dest
.saturate
;
1217 case nir_op_fround_even
:
1218 inst
= bld
.RNDE(result
, op
[0]);
1219 inst
->saturate
= instr
->dest
.saturate
;
1222 case nir_op_fquantize2f16
: {
1223 fs_reg tmp16
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1224 fs_reg tmp32
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1225 fs_reg zero
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1227 /* The destination stride must be at least as big as the source stride. */
1228 tmp16
.type
= BRW_REGISTER_TYPE_W
;
1231 /* Check for denormal */
1232 fs_reg abs_src0
= op
[0];
1233 abs_src0
.abs
= true;
1234 bld
.CMP(bld
.null_reg_f(), abs_src0
, brw_imm_f(ldexpf(1.0, -14)),
1236 /* Get the appropriately signed zero */
1237 bld
.AND(retype(zero
, BRW_REGISTER_TYPE_UD
),
1238 retype(op
[0], BRW_REGISTER_TYPE_UD
),
1239 brw_imm_ud(0x80000000));
1240 /* Do the actual F32 -> F16 -> F32 conversion */
1241 bld
.emit(BRW_OPCODE_F32TO16
, tmp16
, op
[0]);
1242 bld
.emit(BRW_OPCODE_F16TO32
, tmp32
, tmp16
);
1243 /* Select that or zero based on normal status */
1244 inst
= bld
.SEL(result
, zero
, tmp32
);
1245 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1246 inst
->saturate
= instr
->dest
.saturate
;
1253 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1254 inst
->saturate
= instr
->dest
.saturate
;
1260 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1261 inst
->saturate
= instr
->dest
.saturate
;
1264 case nir_op_pack_snorm_2x16
:
1265 case nir_op_pack_snorm_4x8
:
1266 case nir_op_pack_unorm_2x16
:
1267 case nir_op_pack_unorm_4x8
:
1268 case nir_op_unpack_snorm_2x16
:
1269 case nir_op_unpack_snorm_4x8
:
1270 case nir_op_unpack_unorm_2x16
:
1271 case nir_op_unpack_unorm_4x8
:
1272 case nir_op_unpack_half_2x16
:
1273 case nir_op_pack_half_2x16
:
1274 unreachable("not reached: should be handled by lower_packing_builtins");
1276 case nir_op_unpack_half_2x16_split_x
:
1277 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X
, result
, op
[0]);
1278 inst
->saturate
= instr
->dest
.saturate
;
1280 case nir_op_unpack_half_2x16_split_y
:
1281 inst
= bld
.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y
, result
, op
[0]);
1282 inst
->saturate
= instr
->dest
.saturate
;
1285 case nir_op_pack_64_2x32_split
:
1286 bld
.emit(FS_OPCODE_PACK
, result
, op
[0], op
[1]);
1289 case nir_op_unpack_64_2x32_split_x
:
1290 case nir_op_unpack_64_2x32_split_y
: {
1291 if (instr
->op
== nir_op_unpack_64_2x32_split_x
)
1292 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 0));
1294 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1));
1299 inst
= bld
.emit(SHADER_OPCODE_POW
, result
, op
[0], op
[1]);
1300 inst
->saturate
= instr
->dest
.saturate
;
1303 case nir_op_bitfield_reverse
:
1304 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1305 bld
.BFREV(result
, op
[0]);
1308 case nir_op_bit_count
:
1309 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1310 bld
.CBIT(result
, op
[0]);
1313 case nir_op_ufind_msb
: {
1314 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1315 emit_find_msb_using_lzd(bld
, result
, op
[0], false);
1319 case nir_op_ifind_msb
: {
1320 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1322 if (devinfo
->gen
< 7) {
1323 emit_find_msb_using_lzd(bld
, result
, op
[0], true);
1325 bld
.FBH(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0]);
1327 /* FBH counts from the MSB side, while GLSL's findMSB() wants the
1328 * count from the LSB side. If FBH didn't return an error
1329 * (0xFFFFFFFF), then subtract the result from 31 to convert the MSB
1330 * count into an LSB count.
1332 bld
.CMP(bld
.null_reg_d(), result
, brw_imm_d(-1), BRW_CONDITIONAL_NZ
);
1334 inst
= bld
.ADD(result
, result
, brw_imm_d(31));
1335 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1336 inst
->src
[0].negate
= true;
1341 case nir_op_find_lsb
:
1342 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1344 if (devinfo
->gen
< 7) {
1345 fs_reg temp
= vgrf(glsl_type::int_type
);
1347 /* (x & -x) generates a value that consists of only the LSB of x.
1348 * For all powers of 2, findMSB(y) == findLSB(y).
1350 fs_reg src
= retype(op
[0], BRW_REGISTER_TYPE_D
);
1351 fs_reg negated_src
= src
;
1353 /* One must be negated, and the other must be non-negated. It
1354 * doesn't matter which is which.
1356 negated_src
.negate
= true;
1359 bld
.AND(temp
, src
, negated_src
);
1360 emit_find_msb_using_lzd(bld
, result
, temp
, false);
1362 bld
.FBL(result
, op
[0]);
1366 case nir_op_ubitfield_extract
:
1367 case nir_op_ibitfield_extract
:
1368 unreachable("should have been lowered");
1371 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1372 bld
.BFE(result
, op
[2], op
[1], op
[0]);
1375 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1376 bld
.BFI1(result
, op
[0], op
[1]);
1379 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1380 bld
.BFI2(result
, op
[0], op
[1], op
[2]);
1383 case nir_op_bitfield_insert
:
1384 unreachable("not reached: should have been lowered");
1389 fs_reg shift_count
= op
[1];
1391 if (devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
)) {
1392 if (op
[1].file
== VGRF
&&
1393 (result
.type
== BRW_REGISTER_TYPE_Q
||
1394 result
.type
== BRW_REGISTER_TYPE_UQ
)) {
1395 shift_count
= fs_reg(VGRF
, alloc
.allocate(dispatch_width
/ 4),
1396 BRW_REGISTER_TYPE_UD
);
1397 shift_count
.stride
= 2;
1398 bld
.MOV(shift_count
, op
[1]);
1402 switch (instr
->op
) {
1404 bld
.SHL(result
, op
[0], shift_count
);
1407 bld
.ASR(result
, op
[0], shift_count
);
1410 bld
.SHR(result
, op
[0], shift_count
);
1413 unreachable("not reached");
1418 case nir_op_pack_half_2x16_split
:
1419 bld
.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, result
, op
[0], op
[1]);
1423 inst
= bld
.MAD(result
, op
[2], op
[1], op
[0]);
1424 inst
->saturate
= instr
->dest
.saturate
;
1428 inst
= bld
.LRP(result
, op
[0], op
[1], op
[2]);
1429 inst
->saturate
= instr
->dest
.saturate
;
1433 if (optimize_frontfacing_ternary(instr
, result
))
1436 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1437 inst
= bld
.SEL(result
, op
[1], op
[2]);
1438 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1441 case nir_op_extract_u8
:
1442 case nir_op_extract_i8
: {
1443 nir_const_value
*byte
= nir_src_as_const_value(instr
->src
[1].src
);
1444 assert(byte
!= NULL
);
1449 * There is no direct conversion from B/UB to Q/UQ or Q/UQ to B/UB.
1450 * Use two instructions and a word or DWord intermediate integer type.
1452 if (nir_dest_bit_size(instr
->dest
.dest
) == 64) {
1453 const brw_reg_type type
= brw_int_type(2, instr
->op
== nir_op_extract_i8
);
1455 if (instr
->op
== nir_op_extract_i8
) {
1456 /* If we need to sign extend, extract to a word first */
1457 fs_reg w_temp
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
1458 bld
.MOV(w_temp
, subscript(op
[0], type
, byte
->u32
[0]));
1459 bld
.MOV(result
, w_temp
);
1461 /* Otherwise use an AND with 0xff and a word type */
1462 bld
.AND(result
, subscript(op
[0], type
, byte
->u32
[0] / 2), brw_imm_uw(0xff));
1465 const brw_reg_type type
= brw_int_type(1, instr
->op
== nir_op_extract_i8
);
1466 bld
.MOV(result
, subscript(op
[0], type
, byte
->u32
[0]));
1471 case nir_op_extract_u16
:
1472 case nir_op_extract_i16
: {
1473 const brw_reg_type type
= brw_int_type(2, instr
->op
== nir_op_extract_i16
);
1474 nir_const_value
*word
= nir_src_as_const_value(instr
->src
[1].src
);
1475 assert(word
!= NULL
);
1476 bld
.MOV(result
, subscript(op
[0], type
, word
->u32
[0]));
1481 unreachable("unhandled instruction");
1484 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1485 * to sign extend the low bit to 0/~0
1487 if (devinfo
->gen
<= 5 &&
1488 (instr
->instr
.pass_flags
& BRW_NIR_BOOLEAN_MASK
) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE
) {
1489 fs_reg masked
= vgrf(glsl_type::int_type
);
1490 bld
.AND(masked
, result
, brw_imm_d(1));
1491 masked
.negate
= true;
1492 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), masked
);
1497 fs_visitor::nir_emit_load_const(const fs_builder
&bld
,
1498 nir_load_const_instr
*instr
)
1500 const brw_reg_type reg_type
=
1501 brw_reg_type_from_bit_size(instr
->def
.bit_size
, BRW_REGISTER_TYPE_D
);
1502 fs_reg reg
= bld
.vgrf(reg_type
, instr
->def
.num_components
);
1504 switch (instr
->def
.bit_size
) {
1506 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1507 bld
.MOV(offset(reg
, bld
, i
), brw_imm_d(instr
->value
.i32
[i
]));
1511 assert(devinfo
->gen
>= 7);
1512 if (devinfo
->gen
== 7) {
1513 /* We don't get 64-bit integer types until gen8 */
1514 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++) {
1515 bld
.MOV(retype(offset(reg
, bld
, i
), BRW_REGISTER_TYPE_DF
),
1516 setup_imm_df(bld
, instr
->value
.f64
[i
]));
1519 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1520 bld
.MOV(offset(reg
, bld
, i
), brw_imm_q(instr
->value
.i64
[i
]));
1525 unreachable("Invalid bit size");
1528 nir_ssa_values
[instr
->def
.index
] = reg
;
1532 fs_visitor::get_nir_src(const nir_src
&src
)
1536 if (src
.ssa
->parent_instr
->type
== nir_instr_type_ssa_undef
) {
1537 const brw_reg_type reg_type
=
1538 brw_reg_type_from_bit_size(src
.ssa
->bit_size
, BRW_REGISTER_TYPE_D
);
1539 reg
= bld
.vgrf(reg_type
, src
.ssa
->num_components
);
1541 reg
= nir_ssa_values
[src
.ssa
->index
];
1544 /* We don't handle indirects on locals */
1545 assert(src
.reg
.indirect
== NULL
);
1546 reg
= offset(nir_locals
[src
.reg
.reg
->index
], bld
,
1547 src
.reg
.base_offset
* src
.reg
.reg
->num_components
);
1550 if (nir_src_bit_size(src
) == 64 && devinfo
->gen
== 7) {
1551 /* The only 64-bit type available on gen7 is DF, so use that. */
1552 reg
.type
= BRW_REGISTER_TYPE_DF
;
1554 /* To avoid floating-point denorm flushing problems, set the type by
1555 * default to an integer type - instructions that need floating point
1556 * semantics will set this to F if they need to
1558 reg
.type
= brw_reg_type_from_bit_size(nir_src_bit_size(src
),
1559 BRW_REGISTER_TYPE_D
);
1566 * Return an IMM for constants; otherwise call get_nir_src() as normal.
1568 * This function should not be called on any value which may be 64 bits.
1569 * We could theoretically support 64-bit on gen8+ but we choose not to
1570 * because it wouldn't work in general (no gen7 support) and there are
1571 * enough restrictions in 64-bit immediates that you can't take the return
1572 * value and treat it the same as the result of get_nir_src().
1575 fs_visitor::get_nir_src_imm(const nir_src
&src
)
1577 nir_const_value
*val
= nir_src_as_const_value(src
);
1578 assert(nir_src_bit_size(src
) == 32);
1579 return val
? fs_reg(brw_imm_d(val
->i32
[0])) : get_nir_src(src
);
1583 fs_visitor::get_nir_dest(const nir_dest
&dest
)
1586 const brw_reg_type reg_type
=
1587 brw_reg_type_from_bit_size(dest
.ssa
.bit_size
, BRW_REGISTER_TYPE_F
);
1588 nir_ssa_values
[dest
.ssa
.index
] =
1589 bld
.vgrf(reg_type
, dest
.ssa
.num_components
);
1590 return nir_ssa_values
[dest
.ssa
.index
];
1592 /* We don't handle indirects on locals */
1593 assert(dest
.reg
.indirect
== NULL
);
1594 return offset(nir_locals
[dest
.reg
.reg
->index
], bld
,
1595 dest
.reg
.base_offset
* dest
.reg
.reg
->num_components
);
1600 fs_visitor::get_nir_image_deref(const nir_deref_var
*deref
)
1602 fs_reg
image(UNIFORM
, deref
->var
->data
.driver_location
/ 4,
1603 BRW_REGISTER_TYPE_UD
);
1605 unsigned indirect_max
= 0;
1607 for (const nir_deref
*tail
= &deref
->deref
; tail
->child
;
1608 tail
= tail
->child
) {
1609 const nir_deref_array
*deref_array
= nir_deref_as_array(tail
->child
);
1610 assert(tail
->child
->deref_type
== nir_deref_type_array
);
1611 const unsigned size
= glsl_get_length(tail
->type
);
1612 const unsigned element_size
= type_size_scalar(deref_array
->deref
.type
);
1613 const unsigned base
= MIN2(deref_array
->base_offset
, size
- 1);
1614 image
= offset(image
, bld
, base
* element_size
);
1616 if (deref_array
->deref_array_type
== nir_deref_array_type_indirect
) {
1617 fs_reg tmp
= vgrf(glsl_type::uint_type
);
1619 /* Accessing an invalid surface index with the dataport can result
1620 * in a hang. According to the spec "if the index used to
1621 * select an individual element is negative or greater than or
1622 * equal to the size of the array, the results of the operation
1623 * are undefined but may not lead to termination" -- which is one
1624 * of the possible outcomes of the hang. Clamp the index to
1625 * prevent access outside of the array bounds.
1627 bld
.emit_minmax(tmp
, retype(get_nir_src(deref_array
->indirect
),
1628 BRW_REGISTER_TYPE_UD
),
1629 brw_imm_ud(size
- base
- 1), BRW_CONDITIONAL_L
);
1631 indirect_max
+= element_size
* (tail
->type
->length
- 1);
1633 bld
.MUL(tmp
, tmp
, brw_imm_ud(element_size
* 4));
1634 if (indirect
.file
== BAD_FILE
) {
1637 bld
.ADD(indirect
, indirect
, tmp
);
1642 if (indirect
.file
== BAD_FILE
) {
1645 /* Emit a pile of MOVs to load the uniform into a temporary. The
1646 * dead-code elimination pass will get rid of what we don't use.
1648 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, BRW_IMAGE_PARAM_SIZE
);
1649 for (unsigned j
= 0; j
< BRW_IMAGE_PARAM_SIZE
; j
++) {
1650 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
1651 offset(tmp
, bld
, j
), offset(image
, bld
, j
),
1652 indirect
, brw_imm_ud((indirect_max
+ 1) * 4));
1659 fs_visitor::emit_percomp(const fs_builder
&bld
, const fs_inst
&inst
,
1662 for (unsigned i
= 0; i
< 4; i
++) {
1663 if (!((wr_mask
>> i
) & 1))
1666 fs_inst
*new_inst
= new(mem_ctx
) fs_inst(inst
);
1667 new_inst
->dst
= offset(new_inst
->dst
, bld
, i
);
1668 for (unsigned j
= 0; j
< new_inst
->sources
; j
++)
1669 if (new_inst
->src
[j
].file
== VGRF
)
1670 new_inst
->src
[j
] = offset(new_inst
->src
[j
], bld
, i
);
1677 * Get the matching channel register datatype for an image intrinsic of the
1678 * specified GLSL image type.
1681 get_image_base_type(const glsl_type
*type
)
1683 switch ((glsl_base_type
)type
->sampled_type
) {
1684 case GLSL_TYPE_UINT
:
1685 return BRW_REGISTER_TYPE_UD
;
1687 return BRW_REGISTER_TYPE_D
;
1688 case GLSL_TYPE_FLOAT
:
1689 return BRW_REGISTER_TYPE_F
;
1691 unreachable("Not reached.");
1696 * Get the appropriate atomic op for an image atomic intrinsic.
1699 get_image_atomic_op(nir_intrinsic_op op
, const glsl_type
*type
)
1702 case nir_intrinsic_image_var_atomic_add
:
1704 case nir_intrinsic_image_var_atomic_min
:
1705 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1706 BRW_AOP_IMIN
: BRW_AOP_UMIN
);
1707 case nir_intrinsic_image_var_atomic_max
:
1708 return (get_image_base_type(type
) == BRW_REGISTER_TYPE_D
?
1709 BRW_AOP_IMAX
: BRW_AOP_UMAX
);
1710 case nir_intrinsic_image_var_atomic_and
:
1712 case nir_intrinsic_image_var_atomic_or
:
1714 case nir_intrinsic_image_var_atomic_xor
:
1716 case nir_intrinsic_image_var_atomic_exchange
:
1718 case nir_intrinsic_image_var_atomic_comp_swap
:
1719 return BRW_AOP_CMPWR
;
1721 unreachable("Not reachable.");
1726 emit_pixel_interpolater_send(const fs_builder
&bld
,
1731 glsl_interp_mode interpolation
)
1733 struct brw_wm_prog_data
*wm_prog_data
=
1734 brw_wm_prog_data(bld
.shader
->stage_prog_data
);
1739 if (src
.file
== BAD_FILE
) {
1741 payload
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 1);
1745 mlen
= 2 * bld
.dispatch_width() / 8;
1748 inst
= bld
.emit(opcode
, dst
, payload
, desc
);
1750 /* 2 floats per slot returned */
1751 inst
->size_written
= 2 * dst
.component_size(inst
->exec_size
);
1752 inst
->pi_noperspective
= interpolation
== INTERP_MODE_NOPERSPECTIVE
;
1754 wm_prog_data
->pulls_bary
= true;
1760 * Computes 1 << x, given a D/UD register containing some value x.
1763 intexp2(const fs_builder
&bld
, const fs_reg
&x
)
1765 assert(x
.type
== BRW_REGISTER_TYPE_UD
|| x
.type
== BRW_REGISTER_TYPE_D
);
1767 fs_reg result
= bld
.vgrf(x
.type
, 1);
1768 fs_reg one
= bld
.vgrf(x
.type
, 1);
1770 bld
.MOV(one
, retype(brw_imm_d(1), one
.type
));
1771 bld
.SHL(result
, one
, x
);
1776 fs_visitor::emit_gs_end_primitive(const nir_src
&vertex_count_nir_src
)
1778 assert(stage
== MESA_SHADER_GEOMETRY
);
1780 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1782 if (gs_compile
->control_data_header_size_bits
== 0)
1785 /* We can only do EndPrimitive() functionality when the control data
1786 * consists of cut bits. Fortunately, the only time it isn't is when the
1787 * output type is points, in which case EndPrimitive() is a no-op.
1789 if (gs_prog_data
->control_data_format
!=
1790 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
) {
1794 /* Cut bits use one bit per vertex. */
1795 assert(gs_compile
->control_data_bits_per_vertex
== 1);
1797 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1798 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1800 /* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
1801 * vertex n, 0 otherwise. So all we need to do here is mark bit
1802 * (vertex_count - 1) % 32 in the cut_bits register to indicate that
1803 * EndPrimitive() was called after emitting vertex (vertex_count - 1);
1804 * vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
1806 * Note that if EndPrimitive() is called before emitting any vertices, this
1807 * will cause us to set bit 31 of the control_data_bits register to 1.
1808 * That's fine because:
1810 * - If max_vertices < 32, then vertex number 31 (zero-based) will never be
1811 * output, so the hardware will ignore cut bit 31.
1813 * - If max_vertices == 32, then vertex number 31 is guaranteed to be the
1814 * last vertex, so setting cut bit 31 has no effect (since the primitive
1815 * is automatically ended when the GS terminates).
1817 * - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
1818 * control_data_bits register to 0 when the first vertex is emitted.
1821 const fs_builder abld
= bld
.annotate("end primitive");
1823 /* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
1824 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1825 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1826 fs_reg mask
= intexp2(abld
, prev_count
);
1827 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1828 * attention to the lower 5 bits of its second source argument, so on this
1829 * architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
1830 * ((vertex_count - 1) % 32).
1832 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1836 fs_visitor::emit_gs_control_data_bits(const fs_reg
&vertex_count
)
1838 assert(stage
== MESA_SHADER_GEOMETRY
);
1839 assert(gs_compile
->control_data_bits_per_vertex
!= 0);
1841 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1843 const fs_builder abld
= bld
.annotate("emit control data bits");
1844 const fs_builder fwa_bld
= bld
.exec_all();
1846 /* We use a single UD register to accumulate control data bits (32 bits
1847 * for each of the SIMD8 channels). So we need to write a DWord (32 bits)
1850 * Unfortunately, the URB_WRITE_SIMD8 message uses 128-bit (OWord) offsets.
1851 * We have select a 128-bit group via the Global and Per-Slot Offsets, then
1852 * use the Channel Mask phase to enable/disable which DWord within that
1853 * group to write. (Remember, different SIMD8 channels may have emitted
1854 * different numbers of vertices, so we may need per-slot offsets.)
1856 * Channel masking presents an annoying problem: we may have to replicate
1857 * the data up to 4 times:
1859 * Msg = Handles, Per-Slot Offsets, Channel Masks, Data, Data, Data, Data.
1861 * To avoid penalizing shaders that emit a small number of vertices, we
1862 * can avoid these sometimes: if the size of the control data header is
1863 * <= 128 bits, then there is only 1 OWord. All SIMD8 channels will land
1864 * land in the same 128-bit group, so we can skip per-slot offsets.
1866 * Similarly, if the control data header is <= 32 bits, there is only one
1867 * DWord, so we can skip channel masks.
1869 enum opcode opcode
= SHADER_OPCODE_URB_WRITE_SIMD8
;
1871 fs_reg channel_mask
, per_slot_offset
;
1873 if (gs_compile
->control_data_header_size_bits
> 32) {
1874 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
1875 channel_mask
= vgrf(glsl_type::uint_type
);
1878 if (gs_compile
->control_data_header_size_bits
> 128) {
1879 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
;
1880 per_slot_offset
= vgrf(glsl_type::uint_type
);
1883 /* Figure out which DWord we're trying to write to using the formula:
1885 * dword_index = (vertex_count - 1) * bits_per_vertex / 32
1887 * Since bits_per_vertex is a power of two, and is known at compile
1888 * time, this can be optimized to:
1890 * dword_index = (vertex_count - 1) >> (6 - log2(bits_per_vertex))
1892 if (opcode
!= SHADER_OPCODE_URB_WRITE_SIMD8
) {
1893 fs_reg dword_index
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1894 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1895 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1896 unsigned log2_bits_per_vertex
=
1897 util_last_bit(gs_compile
->control_data_bits_per_vertex
);
1898 abld
.SHR(dword_index
, prev_count
, brw_imm_ud(6u - log2_bits_per_vertex
));
1900 if (per_slot_offset
.file
!= BAD_FILE
) {
1901 /* Set the per-slot offset to dword_index / 4, so that we'll write to
1902 * the appropriate OWord within the control data header.
1904 abld
.SHR(per_slot_offset
, dword_index
, brw_imm_ud(2u));
1907 /* Set the channel masks to 1 << (dword_index % 4), so that we'll
1908 * write to the appropriate DWORD within the OWORD.
1910 fs_reg channel
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1911 fwa_bld
.AND(channel
, dword_index
, brw_imm_ud(3u));
1912 channel_mask
= intexp2(fwa_bld
, channel
);
1913 /* Then the channel masks need to be in bits 23:16. */
1914 fwa_bld
.SHL(channel_mask
, channel_mask
, brw_imm_ud(16u));
1917 /* Store the control data bits in the message payload and send it. */
1919 if (channel_mask
.file
!= BAD_FILE
)
1920 mlen
+= 4; /* channel masks, plus 3 extra copies of the data */
1921 if (per_slot_offset
.file
!= BAD_FILE
)
1924 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
1925 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, mlen
);
1927 sources
[i
++] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
1928 if (per_slot_offset
.file
!= BAD_FILE
)
1929 sources
[i
++] = per_slot_offset
;
1930 if (channel_mask
.file
!= BAD_FILE
)
1931 sources
[i
++] = channel_mask
;
1933 sources
[i
++] = this->control_data_bits
;
1936 abld
.LOAD_PAYLOAD(payload
, sources
, mlen
, mlen
);
1937 fs_inst
*inst
= abld
.emit(opcode
, reg_undef
, payload
);
1939 /* We need to increment Global Offset by 256-bits to make room for
1940 * Broadwell's extra "Vertex Count" payload at the beginning of the
1941 * URB entry. Since this is an OWord message, Global Offset is counted
1942 * in 128-bit units, so we must set it to 2.
1944 if (gs_prog_data
->static_vertex_count
== -1)
1949 fs_visitor::set_gs_stream_control_data_bits(const fs_reg
&vertex_count
,
1952 /* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
1954 /* Note: we are calling this *before* increasing vertex_count, so
1955 * this->vertex_count == vertex_count - 1 in the formula above.
1958 /* Stream mode uses 2 bits per vertex */
1959 assert(gs_compile
->control_data_bits_per_vertex
== 2);
1961 /* Must be a valid stream */
1962 assert(stream_id
< MAX_VERTEX_STREAMS
);
1964 /* Control data bits are initialized to 0 so we don't have to set any
1965 * bits when sending vertices to stream 0.
1970 const fs_builder abld
= bld
.annotate("set stream control data bits", NULL
);
1972 /* reg::sid = stream_id */
1973 fs_reg sid
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1974 abld
.MOV(sid
, brw_imm_ud(stream_id
));
1976 /* reg:shift_count = 2 * (vertex_count - 1) */
1977 fs_reg shift_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1978 abld
.SHL(shift_count
, vertex_count
, brw_imm_ud(1u));
1980 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1981 * attention to the lower 5 bits of its second source argument, so on this
1982 * architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
1983 * stream_id << ((2 * (vertex_count - 1)) % 32).
1985 fs_reg mask
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1986 abld
.SHL(mask
, sid
, shift_count
);
1987 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1991 fs_visitor::emit_gs_vertex(const nir_src
&vertex_count_nir_src
,
1994 assert(stage
== MESA_SHADER_GEOMETRY
);
1996 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1998 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1999 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
2001 /* Haswell and later hardware ignores the "Render Stream Select" bits
2002 * from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
2003 * and instead sends all primitives down the pipeline for rasterization.
2004 * If the SOL stage is enabled, "Render Stream Select" is honored and
2005 * primitives bound to non-zero streams are discarded after stream output.
2007 * Since the only purpose of primives sent to non-zero streams is to
2008 * be recorded by transform feedback, we can simply discard all geometry
2009 * bound to these streams when transform feedback is disabled.
2011 if (stream_id
> 0 && !nir
->info
.has_transform_feedback_varyings
)
2014 /* If we're outputting 32 control data bits or less, then we can wait
2015 * until the shader is over to output them all. Otherwise we need to
2016 * output them as we go. Now is the time to do it, since we're about to
2017 * output the vertex_count'th vertex, so it's guaranteed that the
2018 * control data bits associated with the (vertex_count - 1)th vertex are
2021 if (gs_compile
->control_data_header_size_bits
> 32) {
2022 const fs_builder abld
=
2023 bld
.annotate("emit vertex: emit control data bits");
2025 /* Only emit control data bits if we've finished accumulating a batch
2026 * of 32 bits. This is the case when:
2028 * (vertex_count * bits_per_vertex) % 32 == 0
2030 * (in other words, when the last 5 bits of vertex_count *
2031 * bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
2032 * integer n (which is always the case, since bits_per_vertex is
2033 * always 1 or 2), this is equivalent to requiring that the last 5-n
2034 * bits of vertex_count are 0:
2036 * vertex_count & (2^(5-n) - 1) == 0
2038 * 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
2041 * vertex_count & (32 / bits_per_vertex - 1) == 0
2043 * TODO: If vertex_count is an immediate, we could do some of this math
2044 * at compile time...
2047 abld
.AND(bld
.null_reg_d(), vertex_count
,
2048 brw_imm_ud(32u / gs_compile
->control_data_bits_per_vertex
- 1u));
2049 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2051 abld
.IF(BRW_PREDICATE_NORMAL
);
2052 /* If vertex_count is 0, then no control data bits have been
2053 * accumulated yet, so we can skip emitting them.
2055 abld
.CMP(bld
.null_reg_d(), vertex_count
, brw_imm_ud(0u),
2056 BRW_CONDITIONAL_NEQ
);
2057 abld
.IF(BRW_PREDICATE_NORMAL
);
2058 emit_gs_control_data_bits(vertex_count
);
2059 abld
.emit(BRW_OPCODE_ENDIF
);
2061 /* Reset control_data_bits to 0 so we can start accumulating a new
2064 * Note: in the case where vertex_count == 0, this neutralizes the
2065 * effect of any call to EndPrimitive() that the shader may have
2066 * made before outputting its first vertex.
2068 inst
= abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
2069 inst
->force_writemask_all
= true;
2070 abld
.emit(BRW_OPCODE_ENDIF
);
2073 emit_urb_writes(vertex_count
);
2075 /* In stream mode we have to set control data bits for all vertices
2076 * unless we have disabled control data bits completely (which we do
2077 * do for GL_POINTS outputs that don't use streams).
2079 if (gs_compile
->control_data_header_size_bits
> 0 &&
2080 gs_prog_data
->control_data_format
==
2081 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
) {
2082 set_gs_stream_control_data_bits(vertex_count
, stream_id
);
2087 fs_visitor::emit_gs_input_load(const fs_reg
&dst
,
2088 const nir_src
&vertex_src
,
2089 unsigned base_offset
,
2090 const nir_src
&offset_src
,
2091 unsigned num_components
,
2092 unsigned first_component
)
2094 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
2096 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
2097 nir_const_value
*offset_const
= nir_src_as_const_value(offset_src
);
2098 const unsigned push_reg_count
= gs_prog_data
->base
.urb_read_length
* 8;
2100 /* TODO: figure out push input layout for invocations == 1 */
2101 /* TODO: make this work with 64-bit inputs */
2102 if (gs_prog_data
->invocations
== 1 &&
2103 type_sz(dst
.type
) <= 4 &&
2104 offset_const
!= NULL
&& vertex_const
!= NULL
&&
2105 4 * (base_offset
+ offset_const
->u32
[0]) < push_reg_count
) {
2106 int imm_offset
= (base_offset
+ offset_const
->u32
[0]) * 4 +
2107 vertex_const
->u32
[0] * push_reg_count
;
2108 for (unsigned i
= 0; i
< num_components
; i
++) {
2109 bld
.MOV(offset(dst
, bld
, i
),
2110 fs_reg(ATTR
, imm_offset
+ i
+ first_component
, dst
.type
));
2115 /* Resort to the pull model. Ensure the VUE handles are provided. */
2116 assert(gs_prog_data
->base
.include_vue_handles
);
2118 unsigned first_icp_handle
= gs_prog_data
->include_primitive_id
? 3 : 2;
2119 fs_reg icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2121 if (gs_prog_data
->invocations
== 1) {
2123 /* The vertex index is constant; just select the proper URB handle. */
2125 retype(brw_vec8_grf(first_icp_handle
+ vertex_const
->i32
[0], 0),
2126 BRW_REGISTER_TYPE_UD
);
2128 /* The vertex index is non-constant. We need to use indirect
2129 * addressing to fetch the proper URB handle.
2131 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
2132 * indicating that channel <n> should read the handle from
2133 * DWord <n>. We convert that to bytes by multiplying by 4.
2135 * Next, we convert the vertex index to bytes by multiplying
2136 * by 32 (shifting by 5), and add the two together. This is
2137 * the final indirect byte offset.
2139 fs_reg sequence
= bld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
2140 fs_reg channel_offsets
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2141 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2142 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2144 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
2145 bld
.MOV(sequence
, fs_reg(brw_imm_v(0x76543210)));
2146 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
2147 bld
.SHL(channel_offsets
, sequence
, brw_imm_ud(2u));
2148 /* Convert vertex_index to bytes (multiply by 32) */
2149 bld
.SHL(vertex_offset_bytes
,
2150 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2152 bld
.ADD(icp_offset_bytes
, vertex_offset_bytes
, channel_offsets
);
2154 /* Use first_icp_handle as the base offset. There is one register
2155 * of URB handles per vertex, so inform the register allocator that
2156 * we might read up to nir->info.gs.vertices_in registers.
2158 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2159 retype(brw_vec8_grf(first_icp_handle
, 0), icp_handle
.type
),
2160 fs_reg(icp_offset_bytes
),
2161 brw_imm_ud(nir
->info
.gs
.vertices_in
* REG_SIZE
));
2164 assert(gs_prog_data
->invocations
> 1);
2167 assert(devinfo
->gen
>= 9 || vertex_const
->i32
[0] <= 5);
2169 retype(brw_vec1_grf(first_icp_handle
+
2170 vertex_const
->i32
[0] / 8,
2171 vertex_const
->i32
[0] % 8),
2172 BRW_REGISTER_TYPE_UD
));
2174 /* The vertex index is non-constant. We need to use indirect
2175 * addressing to fetch the proper URB handle.
2178 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2180 /* Convert vertex_index to bytes (multiply by 4) */
2181 bld
.SHL(icp_offset_bytes
,
2182 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2185 /* Use first_icp_handle as the base offset. There is one DWord
2186 * of URB handles per vertex, so inform the register allocator that
2187 * we might read up to ceil(nir->info.gs.vertices_in / 8) registers.
2189 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2190 retype(brw_vec8_grf(first_icp_handle
, 0), icp_handle
.type
),
2191 fs_reg(icp_offset_bytes
),
2192 brw_imm_ud(DIV_ROUND_UP(nir
->info
.gs
.vertices_in
, 8) *
2199 fs_reg tmp_dst
= dst
;
2200 fs_reg indirect_offset
= get_nir_src(offset_src
);
2201 unsigned num_iterations
= 1;
2202 unsigned orig_num_components
= num_components
;
2204 if (type_sz(dst
.type
) == 8) {
2205 if (num_components
> 2) {
2209 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(4), dst
.type
);
2211 first_component
= first_component
/ 2;
2214 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2216 /* Constant indexing - use global offset. */
2217 if (first_component
!= 0) {
2218 unsigned read_components
= num_components
+ first_component
;
2219 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2220 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, icp_handle
);
2221 inst
->size_written
= read_components
*
2222 tmp
.component_size(inst
->exec_size
);
2223 for (unsigned i
= 0; i
< num_components
; i
++) {
2224 bld
.MOV(offset(tmp_dst
, bld
, i
),
2225 offset(tmp
, bld
, i
+ first_component
));
2228 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp_dst
,
2230 inst
->size_written
= num_components
*
2231 tmp_dst
.component_size(inst
->exec_size
);
2233 inst
->offset
= base_offset
+ offset_const
->u32
[0];
2236 /* Indirect indexing - use per-slot offsets as well. */
2237 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2238 unsigned read_components
= num_components
+ first_component
;
2239 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2240 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2241 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2242 if (first_component
!= 0) {
2243 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2245 inst
->size_written
= read_components
*
2246 tmp
.component_size(inst
->exec_size
);
2247 for (unsigned i
= 0; i
< num_components
; i
++) {
2248 bld
.MOV(offset(tmp_dst
, bld
, i
),
2249 offset(tmp
, bld
, i
+ first_component
));
2252 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp_dst
,
2254 inst
->size_written
= num_components
*
2255 tmp_dst
.component_size(inst
->exec_size
);
2257 inst
->offset
= base_offset
;
2261 if (type_sz(dst
.type
) == 8) {
2262 shuffle_32bit_load_result_to_64bit_data(
2263 bld
, tmp_dst
, retype(tmp_dst
, BRW_REGISTER_TYPE_F
), num_components
);
2265 for (unsigned c
= 0; c
< num_components
; c
++)
2266 bld
.MOV(offset(dst
, bld
, iter
* 2 + c
), offset(tmp_dst
, bld
, c
));
2269 if (num_iterations
> 1) {
2270 num_components
= orig_num_components
- 2;
2274 fs_reg new_indirect
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2275 bld
.ADD(new_indirect
, indirect_offset
, brw_imm_ud(1u));
2276 indirect_offset
= new_indirect
;
2283 fs_visitor::get_indirect_offset(nir_intrinsic_instr
*instr
)
2285 nir_src
*offset_src
= nir_get_io_offset_src(instr
);
2286 nir_const_value
*const_value
= nir_src_as_const_value(*offset_src
);
2289 /* The only constant offset we should find is 0. brw_nir.c's
2290 * add_const_offset_to_base() will fold other constant offsets
2291 * into instr->const_index[0].
2293 assert(const_value
->u32
[0] == 0);
2297 return get_nir_src(*offset_src
);
2301 do_untyped_vector_read(const fs_builder
&bld
,
2303 const fs_reg surf_index
,
2304 const fs_reg offset_reg
,
2305 unsigned num_components
)
2307 if (type_sz(dest
.type
) <= 2) {
2308 assert(dest
.stride
== 1);
2309 boolean is_const_offset
= offset_reg
.file
== BRW_IMMEDIATE_VALUE
;
2311 if (is_const_offset
) {
2312 uint32_t start
= offset_reg
.ud
& ~3;
2313 uint32_t end
= offset_reg
.ud
+ num_components
* type_sz(dest
.type
);
2314 end
= ALIGN(end
, 4);
2315 assert (end
- start
<= 16);
2317 /* At this point we have 16-bit component/s that have constant
2318 * offset aligned to 4-bytes that can be read with untyped_reads.
2319 * untyped_read message requires 32-bit aligned offsets.
2321 unsigned first_component
= (offset_reg
.ud
& 3) / type_sz(dest
.type
);
2322 unsigned num_components_32bit
= (end
- start
) / 4;
2324 fs_reg read_result
=
2325 emit_untyped_read(bld
, surf_index
, brw_imm_ud(start
),
2327 num_components_32bit
,
2328 BRW_PREDICATE_NONE
);
2329 shuffle_32bit_load_result_to_16bit_data(bld
,
2330 retype(dest
, BRW_REGISTER_TYPE_W
),
2331 retype(read_result
, BRW_REGISTER_TYPE_D
),
2332 first_component
, num_components
);
2334 fs_reg read_offset
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
2335 for (unsigned i
= 0; i
< num_components
; i
++) {
2337 bld
.MOV(read_offset
, offset_reg
);
2339 bld
.ADD(read_offset
, offset_reg
,
2340 brw_imm_ud(i
* type_sz(dest
.type
)));
2342 /* Non constant offsets are not guaranteed to be aligned 32-bits
2343 * so they are read using one byte_scattered_read message
2344 * for each component.
2346 fs_reg read_result
=
2347 emit_byte_scattered_read(bld
, surf_index
, read_offset
,
2349 type_sz(dest
.type
) * 8 /* bit_size */,
2350 BRW_PREDICATE_NONE
);
2351 bld
.MOV(offset(dest
, bld
, i
),
2352 subscript (read_result
, dest
.type
, 0));
2355 } else if (type_sz(dest
.type
) == 4) {
2356 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, offset_reg
,
2359 BRW_PREDICATE_NONE
);
2360 read_result
.type
= dest
.type
;
2361 for (unsigned i
= 0; i
< num_components
; i
++)
2362 bld
.MOV(offset(dest
, bld
, i
), offset(read_result
, bld
, i
));
2363 } else if (type_sz(dest
.type
) == 8) {
2364 /* Reading a dvec, so we need to:
2366 * 1. Multiply num_components by 2, to account for the fact that we
2367 * need to read 64-bit components.
2368 * 2. Shuffle the result of the load to form valid 64-bit elements
2369 * 3. Emit a second load (for components z/w) if needed.
2371 fs_reg read_offset
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
2372 bld
.MOV(read_offset
, offset_reg
);
2374 int iters
= num_components
<= 2 ? 1 : 2;
2376 /* Load the dvec, the first iteration loads components x/y, the second
2377 * iteration, if needed, loads components z/w
2379 for (int it
= 0; it
< iters
; it
++) {
2380 /* Compute number of components to read in this iteration */
2381 int iter_components
= MIN2(2, num_components
);
2382 num_components
-= iter_components
;
2384 /* Read. Since this message reads 32-bit components, we need to
2385 * read twice as many components.
2387 fs_reg read_result
= emit_untyped_read(bld
, surf_index
, read_offset
,
2389 iter_components
* 2,
2390 BRW_PREDICATE_NONE
);
2392 /* Shuffle the 32-bit load result into valid 64-bit data */
2393 const fs_reg packed_result
= bld
.vgrf(dest
.type
, iter_components
);
2394 shuffle_32bit_load_result_to_64bit_data(
2395 bld
, packed_result
, read_result
, iter_components
);
2397 /* Move each component to its destination */
2398 read_result
= retype(read_result
, BRW_REGISTER_TYPE_DF
);
2399 for (int c
= 0; c
< iter_components
; c
++) {
2400 bld
.MOV(offset(dest
, bld
, it
* 2 + c
),
2401 offset(packed_result
, bld
, c
));
2404 bld
.ADD(read_offset
, read_offset
, brw_imm_ud(16));
2407 unreachable("Unsupported type");
2412 fs_visitor::nir_emit_vs_intrinsic(const fs_builder
&bld
,
2413 nir_intrinsic_instr
*instr
)
2415 assert(stage
== MESA_SHADER_VERTEX
);
2418 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2419 dest
= get_nir_dest(instr
->dest
);
2421 switch (instr
->intrinsic
) {
2422 case nir_intrinsic_load_vertex_id
:
2423 unreachable("should be lowered by lower_vertex_id()");
2425 case nir_intrinsic_load_vertex_id_zero_base
:
2426 case nir_intrinsic_load_base_vertex
:
2427 case nir_intrinsic_load_instance_id
:
2428 case nir_intrinsic_load_base_instance
:
2429 case nir_intrinsic_load_draw_id
: {
2430 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
2431 fs_reg val
= nir_system_values
[sv
];
2432 assert(val
.file
!= BAD_FILE
);
2433 dest
.type
= val
.type
;
2438 case nir_intrinsic_load_input
: {
2439 fs_reg src
= fs_reg(ATTR
, nir_intrinsic_base(instr
) * 4, dest
.type
);
2440 unsigned first_component
= nir_intrinsic_component(instr
);
2441 unsigned num_components
= instr
->num_components
;
2443 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
2444 assert(const_offset
&& "Indirect input loads not allowed");
2445 src
= offset(src
, bld
, const_offset
->u32
[0]);
2447 if (type_sz(dest
.type
) == 8)
2448 first_component
/= 2;
2450 for (unsigned j
= 0; j
< num_components
; j
++) {
2451 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
+ first_component
));
2454 if (type_sz(dest
.type
) == 8) {
2455 shuffle_32bit_load_result_to_64bit_data(bld
,
2457 retype(dest
, BRW_REGISTER_TYPE_F
),
2458 instr
->num_components
);
2463 case nir_intrinsic_load_first_vertex
:
2464 case nir_intrinsic_load_is_indexed_draw
:
2465 unreachable("lowered by brw_nir_lower_vs_inputs");
2468 nir_emit_intrinsic(bld
, instr
);
2474 fs_visitor::nir_emit_tcs_intrinsic(const fs_builder
&bld
,
2475 nir_intrinsic_instr
*instr
)
2477 assert(stage
== MESA_SHADER_TESS_CTRL
);
2478 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
2479 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
2482 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2483 dst
= get_nir_dest(instr
->dest
);
2485 switch (instr
->intrinsic
) {
2486 case nir_intrinsic_load_primitive_id
:
2487 bld
.MOV(dst
, fs_reg(brw_vec1_grf(0, 1)));
2489 case nir_intrinsic_load_invocation_id
:
2490 bld
.MOV(retype(dst
, invocation_id
.type
), invocation_id
);
2492 case nir_intrinsic_load_patch_vertices_in
:
2493 bld
.MOV(retype(dst
, BRW_REGISTER_TYPE_D
),
2494 brw_imm_d(tcs_key
->input_vertices
));
2497 case nir_intrinsic_barrier
: {
2498 if (tcs_prog_data
->instances
== 1)
2501 fs_reg m0
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2502 fs_reg m0_2
= component(m0
, 2);
2504 const fs_builder chanbld
= bld
.exec_all().group(1, 0);
2506 /* Zero the message header */
2507 bld
.exec_all().MOV(m0
, brw_imm_ud(0u));
2509 /* Copy "Barrier ID" from r0.2, bits 16:13 */
2510 chanbld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
2511 brw_imm_ud(INTEL_MASK(16, 13)));
2513 /* Shift it up to bits 27:24. */
2514 chanbld
.SHL(m0_2
, m0_2
, brw_imm_ud(11));
2516 /* Set the Barrier Count and the enable bit */
2517 chanbld
.OR(m0_2
, m0_2
,
2518 brw_imm_ud(tcs_prog_data
->instances
<< 9 | (1 << 15)));
2520 bld
.emit(SHADER_OPCODE_BARRIER
, bld
.null_reg_ud(), m0
);
2524 case nir_intrinsic_load_input
:
2525 unreachable("nir_lower_io should never give us these.");
2528 case nir_intrinsic_load_per_vertex_input
: {
2529 fs_reg indirect_offset
= get_indirect_offset(instr
);
2530 unsigned imm_offset
= instr
->const_index
[0];
2532 const nir_src
&vertex_src
= instr
->src
[0];
2533 nir_const_value
*vertex_const
= nir_src_as_const_value(vertex_src
);
2540 /* Emit a MOV to resolve <0,1,0> regioning. */
2541 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2543 retype(brw_vec1_grf(1 + (vertex_const
->i32
[0] >> 3),
2544 vertex_const
->i32
[0] & 7),
2545 BRW_REGISTER_TYPE_UD
));
2546 } else if (tcs_prog_data
->instances
== 1 &&
2547 vertex_src
.is_ssa
&&
2548 vertex_src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
&&
2549 nir_instr_as_intrinsic(vertex_src
.ssa
->parent_instr
)->intrinsic
== nir_intrinsic_load_invocation_id
) {
2550 /* For the common case of only 1 instance, an array index of
2551 * gl_InvocationID means reading g1. Skip all the indirect work.
2553 icp_handle
= retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
);
2555 /* The vertex index is non-constant. We need to use indirect
2556 * addressing to fetch the proper URB handle.
2558 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2560 /* Each ICP handle is a single DWord (4 bytes) */
2561 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2562 bld
.SHL(vertex_offset_bytes
,
2563 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2566 /* Start at g1. We might read up to 4 registers. */
2567 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2568 retype(brw_vec8_grf(1, 0), icp_handle
.type
), vertex_offset_bytes
,
2569 brw_imm_ud(4 * REG_SIZE
));
2572 /* We can only read two double components with each URB read, so
2573 * we send two read messages in that case, each one loading up to
2574 * two double components.
2576 unsigned num_iterations
= 1;
2577 unsigned num_components
= instr
->num_components
;
2578 unsigned first_component
= nir_intrinsic_component(instr
);
2579 fs_reg orig_dst
= dst
;
2580 if (type_sz(dst
.type
) == 8) {
2581 first_component
= first_component
/ 2;
2582 if (instr
->num_components
> 2) {
2587 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(4), dst
.type
);
2591 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2592 if (indirect_offset
.file
== BAD_FILE
) {
2593 /* Constant indexing - use global offset. */
2594 if (first_component
!= 0) {
2595 unsigned read_components
= num_components
+ first_component
;
2596 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2597 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, icp_handle
);
2598 for (unsigned i
= 0; i
< num_components
; i
++) {
2599 bld
.MOV(offset(dst
, bld
, i
),
2600 offset(tmp
, bld
, i
+ first_component
));
2603 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2605 inst
->offset
= imm_offset
;
2608 /* Indirect indexing - use per-slot offsets as well. */
2609 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2610 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2611 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2612 if (first_component
!= 0) {
2613 unsigned read_components
= num_components
+ first_component
;
2614 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2615 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2617 for (unsigned i
= 0; i
< num_components
; i
++) {
2618 bld
.MOV(offset(dst
, bld
, i
),
2619 offset(tmp
, bld
, i
+ first_component
));
2622 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
,
2625 inst
->offset
= imm_offset
;
2628 inst
->size_written
= (num_components
+ first_component
) *
2629 inst
->dst
.component_size(inst
->exec_size
);
2631 /* If we are reading 64-bit data using 32-bit read messages we need
2632 * build proper 64-bit data elements by shuffling the low and high
2633 * 32-bit components around like we do for other things like UBOs
2636 if (type_sz(dst
.type
) == 8) {
2637 shuffle_32bit_load_result_to_64bit_data(
2638 bld
, dst
, retype(dst
, BRW_REGISTER_TYPE_F
), num_components
);
2640 for (unsigned c
= 0; c
< num_components
; c
++) {
2641 bld
.MOV(offset(orig_dst
, bld
, iter
* 2 + c
),
2642 offset(dst
, bld
, c
));
2646 /* Copy the temporary to the destination to deal with writemasking.
2648 * Also attempt to deal with gl_PointSize being in the .w component.
2650 if (inst
->offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
2651 assert(type_sz(dst
.type
) < 8);
2652 inst
->dst
= bld
.vgrf(dst
.type
, 4);
2653 inst
->size_written
= 4 * REG_SIZE
;
2654 bld
.MOV(dst
, offset(inst
->dst
, bld
, 3));
2657 /* If we are loading double data and we need a second read message
2658 * adjust the write offset
2660 if (num_iterations
> 1) {
2661 num_components
= instr
->num_components
- 2;
2668 case nir_intrinsic_load_output
:
2669 case nir_intrinsic_load_per_vertex_output
: {
2670 fs_reg indirect_offset
= get_indirect_offset(instr
);
2671 unsigned imm_offset
= instr
->const_index
[0];
2672 unsigned first_component
= nir_intrinsic_component(instr
);
2675 if (indirect_offset
.file
== BAD_FILE
) {
2676 /* Replicate the patch handle to all enabled channels */
2677 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2678 bld
.MOV(patch_handle
,
2679 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2682 if (first_component
!= 0) {
2683 unsigned read_components
=
2684 instr
->num_components
+ first_component
;
2685 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2686 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
,
2688 inst
->size_written
= read_components
* REG_SIZE
;
2689 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2690 bld
.MOV(offset(dst
, bld
, i
),
2691 offset(tmp
, bld
, i
+ first_component
));
2694 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
,
2696 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2698 inst
->offset
= imm_offset
;
2702 /* Indirect indexing - use per-slot offsets as well. */
2703 const fs_reg srcs
[] = {
2704 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2707 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2708 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2709 if (first_component
!= 0) {
2710 unsigned read_components
=
2711 instr
->num_components
+ first_component
;
2712 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2713 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2715 inst
->size_written
= read_components
* REG_SIZE
;
2716 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2717 bld
.MOV(offset(dst
, bld
, i
),
2718 offset(tmp
, bld
, i
+ first_component
));
2721 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
,
2723 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2725 inst
->offset
= imm_offset
;
2731 case nir_intrinsic_store_output
:
2732 case nir_intrinsic_store_per_vertex_output
: {
2733 fs_reg value
= get_nir_src(instr
->src
[0]);
2734 bool is_64bit
= (instr
->src
[0].is_ssa
?
2735 instr
->src
[0].ssa
->bit_size
: instr
->src
[0].reg
.reg
->bit_size
) == 64;
2736 fs_reg indirect_offset
= get_indirect_offset(instr
);
2737 unsigned imm_offset
= instr
->const_index
[0];
2738 unsigned mask
= instr
->const_index
[1];
2739 unsigned header_regs
= 0;
2741 srcs
[header_regs
++] = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
);
2743 if (indirect_offset
.file
!= BAD_FILE
) {
2744 srcs
[header_regs
++] = indirect_offset
;
2750 unsigned num_components
= util_last_bit(mask
);
2753 /* We can only pack two 64-bit components in a single message, so send
2754 * 2 messages if we have more components
2756 unsigned num_iterations
= 1;
2757 unsigned iter_components
= num_components
;
2758 unsigned first_component
= nir_intrinsic_component(instr
);
2760 first_component
= first_component
/ 2;
2761 if (instr
->num_components
> 2) {
2763 iter_components
= 2;
2767 mask
= mask
<< first_component
;
2769 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2770 if (!is_64bit
&& mask
!= WRITEMASK_XYZW
) {
2771 srcs
[header_regs
++] = brw_imm_ud(mask
<< 16);
2772 opcode
= indirect_offset
.file
!= BAD_FILE
?
2773 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2774 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2775 } else if (is_64bit
&& ((mask
& WRITEMASK_XY
) != WRITEMASK_XY
)) {
2776 /* Expand the 64-bit mask to 32-bit channels. We only handle
2777 * two channels in each iteration, so we only care about X/Y.
2779 unsigned mask32
= 0;
2780 if (mask
& WRITEMASK_X
)
2781 mask32
|= WRITEMASK_XY
;
2782 if (mask
& WRITEMASK_Y
)
2783 mask32
|= WRITEMASK_ZW
;
2785 /* If the mask does not include any of the channels X or Y there
2786 * is nothing to do in this iteration. Move on to the next couple
2787 * of 64-bit channels.
2795 srcs
[header_regs
++] = brw_imm_ud(mask32
<< 16);
2796 opcode
= indirect_offset
.file
!= BAD_FILE
?
2797 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2798 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2800 opcode
= indirect_offset
.file
!= BAD_FILE
?
2801 SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2802 SHADER_OPCODE_URB_WRITE_SIMD8
;
2805 for (unsigned i
= 0; i
< iter_components
; i
++) {
2806 if (!(mask
& (1 << (i
+ first_component
))))
2810 srcs
[header_regs
+ i
+ first_component
] = offset(value
, bld
, i
);
2812 /* We need to shuffle the 64-bit data to match the layout
2813 * expected by our 32-bit URB write messages. We use a temporary
2816 unsigned channel
= iter
* 2 + i
;
2817 fs_reg dest
= shuffle_64bit_data_for_32bit_write(bld
,
2818 offset(value
, bld
, channel
), 1);
2820 srcs
[header_regs
+ (i
+ first_component
) * 2] = dest
;
2821 srcs
[header_regs
+ (i
+ first_component
) * 2 + 1] =
2822 offset(dest
, bld
, 1);
2827 header_regs
+ (is_64bit
? 2 * iter_components
: iter_components
) +
2828 (is_64bit
? 2 * first_component
: first_component
);
2830 bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2831 bld
.LOAD_PAYLOAD(payload
, srcs
, mlen
, header_regs
);
2833 fs_inst
*inst
= bld
.emit(opcode
, bld
.null_reg_ud(), payload
);
2834 inst
->offset
= imm_offset
;
2837 /* If this is a 64-bit attribute, select the next two 64-bit channels
2838 * to be handled in the next iteration.
2849 nir_emit_intrinsic(bld
, instr
);
2855 fs_visitor::nir_emit_tes_intrinsic(const fs_builder
&bld
,
2856 nir_intrinsic_instr
*instr
)
2858 assert(stage
== MESA_SHADER_TESS_EVAL
);
2859 struct brw_tes_prog_data
*tes_prog_data
= brw_tes_prog_data(prog_data
);
2862 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2863 dest
= get_nir_dest(instr
->dest
);
2865 switch (instr
->intrinsic
) {
2866 case nir_intrinsic_load_primitive_id
:
2867 bld
.MOV(dest
, fs_reg(brw_vec1_grf(0, 1)));
2869 case nir_intrinsic_load_tess_coord
:
2870 /* gl_TessCoord is part of the payload in g1-3 */
2871 for (unsigned i
= 0; i
< 3; i
++) {
2872 bld
.MOV(offset(dest
, bld
, i
), fs_reg(brw_vec8_grf(1 + i
, 0)));
2876 case nir_intrinsic_load_input
:
2877 case nir_intrinsic_load_per_vertex_input
: {
2878 fs_reg indirect_offset
= get_indirect_offset(instr
);
2879 unsigned imm_offset
= instr
->const_index
[0];
2880 unsigned first_component
= nir_intrinsic_component(instr
);
2882 if (type_sz(dest
.type
) == 8) {
2883 first_component
= first_component
/ 2;
2887 if (indirect_offset
.file
== BAD_FILE
) {
2888 /* Arbitrarily only push up to 32 vec4 slots worth of data,
2889 * which is 16 registers (since each holds 2 vec4 slots).
2891 unsigned slot_count
= 1;
2892 if (type_sz(dest
.type
) == 8 && instr
->num_components
> 2)
2895 const unsigned max_push_slots
= 32;
2896 if (imm_offset
+ slot_count
<= max_push_slots
) {
2897 fs_reg src
= fs_reg(ATTR
, imm_offset
/ 2, dest
.type
);
2898 for (int i
= 0; i
< instr
->num_components
; i
++) {
2899 unsigned comp
= 16 / type_sz(dest
.type
) * (imm_offset
% 2) +
2900 i
+ first_component
;
2901 bld
.MOV(offset(dest
, bld
, i
), component(src
, comp
));
2904 tes_prog_data
->base
.urb_read_length
=
2905 MAX2(tes_prog_data
->base
.urb_read_length
,
2906 DIV_ROUND_UP(imm_offset
+ slot_count
, 2));
2908 /* Replicate the patch handle to all enabled channels */
2909 const fs_reg srcs
[] = {
2910 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)
2912 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2913 bld
.LOAD_PAYLOAD(patch_handle
, srcs
, ARRAY_SIZE(srcs
), 0);
2915 if (first_component
!= 0) {
2916 unsigned read_components
=
2917 instr
->num_components
+ first_component
;
2918 fs_reg tmp
= bld
.vgrf(dest
.type
, read_components
);
2919 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
,
2921 inst
->size_written
= read_components
* REG_SIZE
;
2922 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2923 bld
.MOV(offset(dest
, bld
, i
),
2924 offset(tmp
, bld
, i
+ first_component
));
2927 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dest
,
2929 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2932 inst
->offset
= imm_offset
;
2935 /* Indirect indexing - use per-slot offsets as well. */
2937 /* We can only read two double components with each URB read, so
2938 * we send two read messages in that case, each one loading up to
2939 * two double components.
2941 unsigned num_iterations
= 1;
2942 unsigned num_components
= instr
->num_components
;
2943 fs_reg orig_dest
= dest
;
2944 if (type_sz(dest
.type
) == 8) {
2945 if (instr
->num_components
> 2) {
2949 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(4), dest
.type
);
2953 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2954 const fs_reg srcs
[] = {
2955 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2958 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2959 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2961 if (first_component
!= 0) {
2962 unsigned read_components
=
2963 num_components
+ first_component
;
2964 fs_reg tmp
= bld
.vgrf(dest
.type
, read_components
);
2965 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2967 for (unsigned i
= 0; i
< num_components
; i
++) {
2968 bld
.MOV(offset(dest
, bld
, i
),
2969 offset(tmp
, bld
, i
+ first_component
));
2972 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dest
,
2976 inst
->offset
= imm_offset
;
2977 inst
->size_written
= (num_components
+ first_component
) *
2978 inst
->dst
.component_size(inst
->exec_size
);
2980 /* If we are reading 64-bit data using 32-bit read messages we need
2981 * build proper 64-bit data elements by shuffling the low and high
2982 * 32-bit components around like we do for other things like UBOs
2985 if (type_sz(dest
.type
) == 8) {
2986 shuffle_32bit_load_result_to_64bit_data(
2987 bld
, dest
, retype(dest
, BRW_REGISTER_TYPE_F
), num_components
);
2989 for (unsigned c
= 0; c
< num_components
; c
++) {
2990 bld
.MOV(offset(orig_dest
, bld
, iter
* 2 + c
),
2991 offset(dest
, bld
, c
));
2995 /* If we are loading double data and we need a second read message
2998 if (num_iterations
> 1) {
2999 num_components
= instr
->num_components
- 2;
3007 nir_emit_intrinsic(bld
, instr
);
3013 fs_visitor::nir_emit_gs_intrinsic(const fs_builder
&bld
,
3014 nir_intrinsic_instr
*instr
)
3016 assert(stage
== MESA_SHADER_GEOMETRY
);
3017 fs_reg indirect_offset
;
3020 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3021 dest
= get_nir_dest(instr
->dest
);
3023 switch (instr
->intrinsic
) {
3024 case nir_intrinsic_load_primitive_id
:
3025 assert(stage
== MESA_SHADER_GEOMETRY
);
3026 assert(brw_gs_prog_data(prog_data
)->include_primitive_id
);
3027 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
3028 retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD
));
3031 case nir_intrinsic_load_input
:
3032 unreachable("load_input intrinsics are invalid for the GS stage");
3034 case nir_intrinsic_load_per_vertex_input
:
3035 emit_gs_input_load(dest
, instr
->src
[0], instr
->const_index
[0],
3036 instr
->src
[1], instr
->num_components
,
3037 nir_intrinsic_component(instr
));
3040 case nir_intrinsic_emit_vertex_with_counter
:
3041 emit_gs_vertex(instr
->src
[0], instr
->const_index
[0]);
3044 case nir_intrinsic_end_primitive_with_counter
:
3045 emit_gs_end_primitive(instr
->src
[0]);
3048 case nir_intrinsic_set_vertex_count
:
3049 bld
.MOV(this->final_gs_vertex_count
, get_nir_src(instr
->src
[0]));
3052 case nir_intrinsic_load_invocation_id
: {
3053 fs_reg val
= nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
3054 assert(val
.file
!= BAD_FILE
);
3055 dest
.type
= val
.type
;
3061 nir_emit_intrinsic(bld
, instr
);
3067 * Fetch the current render target layer index.
3070 fetch_render_target_array_index(const fs_builder
&bld
)
3072 if (bld
.shader
->devinfo
->gen
>= 6) {
3073 /* The render target array index is provided in the thread payload as
3074 * bits 26:16 of r0.0.
3076 const fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3077 bld
.AND(idx
, brw_uw1_reg(BRW_GENERAL_REGISTER_FILE
, 0, 1),
3081 /* Pre-SNB we only ever render into the first layer of the framebuffer
3082 * since layered rendering is not implemented.
3084 return brw_imm_ud(0);
3089 * Fake non-coherent framebuffer read implemented using TXF to fetch from the
3090 * framebuffer at the current fragment coordinates and sample index.
3093 fs_visitor::emit_non_coherent_fb_read(const fs_builder
&bld
, const fs_reg
&dst
,
3096 const struct gen_device_info
*devinfo
= bld
.shader
->devinfo
;
3098 assert(bld
.shader
->stage
== MESA_SHADER_FRAGMENT
);
3099 const brw_wm_prog_key
*wm_key
=
3100 reinterpret_cast<const brw_wm_prog_key
*>(key
);
3101 assert(!wm_key
->coherent_fb_fetch
);
3102 const struct brw_wm_prog_data
*wm_prog_data
=
3103 brw_wm_prog_data(stage_prog_data
);
3105 /* Calculate the surface index relative to the start of the texture binding
3106 * table block, since that's what the texturing messages expect.
3108 const unsigned surface
= target
+
3109 wm_prog_data
->binding_table
.render_target_read_start
-
3110 wm_prog_data
->base
.binding_table
.texture_start
;
3112 brw_mark_surface_used(
3113 bld
.shader
->stage_prog_data
,
3114 wm_prog_data
->binding_table
.render_target_read_start
+ target
);
3116 /* Calculate the fragment coordinates. */
3117 const fs_reg coords
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 3);
3118 bld
.MOV(offset(coords
, bld
, 0), pixel_x
);
3119 bld
.MOV(offset(coords
, bld
, 1), pixel_y
);
3120 bld
.MOV(offset(coords
, bld
, 2), fetch_render_target_array_index(bld
));
3122 /* Calculate the sample index and MCS payload when multisampling. Luckily
3123 * the MCS fetch message behaves deterministically for UMS surfaces, so it
3124 * shouldn't be necessary to recompile based on whether the framebuffer is
3127 if (wm_key
->multisample_fbo
&&
3128 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
].file
== BAD_FILE
)
3129 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
] = *emit_sampleid_setup();
3131 const fs_reg sample
= nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
3132 const fs_reg mcs
= wm_key
->multisample_fbo
?
3133 emit_mcs_fetch(coords
, 3, brw_imm_ud(surface
)) : fs_reg();
3135 /* Use either a normal or a CMS texel fetch message depending on whether
3136 * the framebuffer is single or multisample. On SKL+ use the wide CMS
3137 * message just in case the framebuffer uses 16x multisampling, it should
3138 * be equivalent to the normal CMS fetch for lower multisampling modes.
3140 const opcode op
= !wm_key
->multisample_fbo
? SHADER_OPCODE_TXF_LOGICAL
:
3141 devinfo
->gen
>= 9 ? SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
3142 SHADER_OPCODE_TXF_CMS_LOGICAL
;
3144 /* Emit the instruction. */
3145 const fs_reg srcs
[] = { coords
, fs_reg(), brw_imm_ud(0), fs_reg(),
3147 brw_imm_ud(surface
), brw_imm_ud(0),
3148 fs_reg(), brw_imm_ud(3), brw_imm_ud(0) };
3149 STATIC_ASSERT(ARRAY_SIZE(srcs
) == TEX_LOGICAL_NUM_SRCS
);
3151 fs_inst
*inst
= bld
.emit(op
, dst
, srcs
, ARRAY_SIZE(srcs
));
3152 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
3158 * Actual coherent framebuffer read implemented using the native render target
3159 * read message. Requires SKL+.
3162 emit_coherent_fb_read(const fs_builder
&bld
, const fs_reg
&dst
, unsigned target
)
3164 assert(bld
.shader
->devinfo
->gen
>= 9);
3165 fs_inst
*inst
= bld
.emit(FS_OPCODE_FB_READ_LOGICAL
, dst
);
3166 inst
->target
= target
;
3167 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
3173 alloc_temporary(const fs_builder
&bld
, unsigned size
, fs_reg
*regs
, unsigned n
)
3175 if (n
&& regs
[0].file
!= BAD_FILE
) {
3179 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, size
);
3181 for (unsigned i
= 0; i
< n
; i
++)
3189 alloc_frag_output(fs_visitor
*v
, unsigned location
)
3191 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
3192 const brw_wm_prog_key
*const key
=
3193 reinterpret_cast<const brw_wm_prog_key
*>(v
->key
);
3194 const unsigned l
= GET_FIELD(location
, BRW_NIR_FRAG_OUTPUT_LOCATION
);
3195 const unsigned i
= GET_FIELD(location
, BRW_NIR_FRAG_OUTPUT_INDEX
);
3197 if (i
> 0 || (key
->force_dual_color_blend
&& l
== FRAG_RESULT_DATA1
))
3198 return alloc_temporary(v
->bld
, 4, &v
->dual_src_output
, 1);
3200 else if (l
== FRAG_RESULT_COLOR
)
3201 return alloc_temporary(v
->bld
, 4, v
->outputs
,
3202 MAX2(key
->nr_color_regions
, 1));
3204 else if (l
== FRAG_RESULT_DEPTH
)
3205 return alloc_temporary(v
->bld
, 1, &v
->frag_depth
, 1);
3207 else if (l
== FRAG_RESULT_STENCIL
)
3208 return alloc_temporary(v
->bld
, 1, &v
->frag_stencil
, 1);
3210 else if (l
== FRAG_RESULT_SAMPLE_MASK
)
3211 return alloc_temporary(v
->bld
, 1, &v
->sample_mask
, 1);
3213 else if (l
>= FRAG_RESULT_DATA0
&&
3214 l
< FRAG_RESULT_DATA0
+ BRW_MAX_DRAW_BUFFERS
)
3215 return alloc_temporary(v
->bld
, 4,
3216 &v
->outputs
[l
- FRAG_RESULT_DATA0
], 1);
3219 unreachable("Invalid location");
3223 fs_visitor::nir_emit_fs_intrinsic(const fs_builder
&bld
,
3224 nir_intrinsic_instr
*instr
)
3226 assert(stage
== MESA_SHADER_FRAGMENT
);
3229 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3230 dest
= get_nir_dest(instr
->dest
);
3232 switch (instr
->intrinsic
) {
3233 case nir_intrinsic_load_front_face
:
3234 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
3235 *emit_frontfacing_interpolation());
3238 case nir_intrinsic_load_sample_pos
: {
3239 fs_reg sample_pos
= nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
3240 assert(sample_pos
.file
!= BAD_FILE
);
3241 dest
.type
= sample_pos
.type
;
3242 bld
.MOV(dest
, sample_pos
);
3243 bld
.MOV(offset(dest
, bld
, 1), offset(sample_pos
, bld
, 1));
3247 case nir_intrinsic_load_layer_id
:
3248 dest
.type
= BRW_REGISTER_TYPE_UD
;
3249 bld
.MOV(dest
, fetch_render_target_array_index(bld
));
3252 case nir_intrinsic_load_helper_invocation
:
3253 case nir_intrinsic_load_sample_mask_in
:
3254 case nir_intrinsic_load_sample_id
: {
3255 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3256 fs_reg val
= nir_system_values
[sv
];
3257 assert(val
.file
!= BAD_FILE
);
3258 dest
.type
= val
.type
;
3263 case nir_intrinsic_store_output
: {
3264 const fs_reg src
= get_nir_src(instr
->src
[0]);
3265 const nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3266 assert(const_offset
&& "Indirect output stores not allowed");
3267 const unsigned location
= nir_intrinsic_base(instr
) +
3268 SET_FIELD(const_offset
->u32
[0], BRW_NIR_FRAG_OUTPUT_LOCATION
);
3269 const fs_reg new_dest
= retype(alloc_frag_output(this, location
),
3272 for (unsigned j
= 0; j
< instr
->num_components
; j
++)
3273 bld
.MOV(offset(new_dest
, bld
, nir_intrinsic_component(instr
) + j
),
3274 offset(src
, bld
, j
));
3279 case nir_intrinsic_load_output
: {
3280 const unsigned l
= GET_FIELD(nir_intrinsic_base(instr
),
3281 BRW_NIR_FRAG_OUTPUT_LOCATION
);
3282 assert(l
>= FRAG_RESULT_DATA0
);
3283 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3284 assert(const_offset
&& "Indirect output loads not allowed");
3285 const unsigned target
= l
- FRAG_RESULT_DATA0
+ const_offset
->u32
[0];
3286 const fs_reg tmp
= bld
.vgrf(dest
.type
, 4);
3288 if (reinterpret_cast<const brw_wm_prog_key
*>(key
)->coherent_fb_fetch
)
3289 emit_coherent_fb_read(bld
, tmp
, target
);
3291 emit_non_coherent_fb_read(bld
, tmp
, target
);
3293 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3294 bld
.MOV(offset(dest
, bld
, j
),
3295 offset(tmp
, bld
, nir_intrinsic_component(instr
) + j
));
3301 case nir_intrinsic_discard
:
3302 case nir_intrinsic_discard_if
: {
3303 /* We track our discarded pixels in f0.1. By predicating on it, we can
3304 * update just the flag bits that aren't yet discarded. If there's no
3305 * condition, we emit a CMP of g0 != g0, so all currently executing
3306 * channels will get turned off.
3309 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
3310 cmp
= bld
.CMP(bld
.null_reg_f(), get_nir_src(instr
->src
[0]),
3311 brw_imm_d(0), BRW_CONDITIONAL_Z
);
3313 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
3314 BRW_REGISTER_TYPE_UW
));
3315 cmp
= bld
.CMP(bld
.null_reg_f(), some_reg
, some_reg
, BRW_CONDITIONAL_NZ
);
3317 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
3318 cmp
->flag_subreg
= 1;
3320 if (devinfo
->gen
>= 6) {
3321 emit_discard_jump();
3326 case nir_intrinsic_load_input
: {
3327 /* load_input is only used for flat inputs */
3328 unsigned base
= nir_intrinsic_base(instr
);
3329 unsigned component
= nir_intrinsic_component(instr
);
3330 unsigned num_components
= instr
->num_components
;
3331 enum brw_reg_type type
= dest
.type
;
3333 /* Special case fields in the VUE header */
3334 if (base
== VARYING_SLOT_LAYER
)
3336 else if (base
== VARYING_SLOT_VIEWPORT
)
3339 if (nir_dest_bit_size(instr
->dest
) == 64) {
3340 /* const_index is in 32-bit type size units that could not be aligned
3341 * with DF. We need to read the double vector as if it was a float
3342 * vector of twice the number of components to fetch the right data.
3344 type
= BRW_REGISTER_TYPE_F
;
3345 num_components
*= 2;
3348 for (unsigned int i
= 0; i
< num_components
; i
++) {
3349 struct brw_reg interp
= interp_reg(base
, component
+ i
);
3350 interp
= suboffset(interp
, 3);
3351 bld
.emit(FS_OPCODE_CINTERP
, offset(retype(dest
, type
), bld
, i
),
3352 retype(fs_reg(interp
), type
));
3355 if (nir_dest_bit_size(instr
->dest
) == 64) {
3356 shuffle_32bit_load_result_to_64bit_data(bld
,
3359 instr
->num_components
);
3364 case nir_intrinsic_load_barycentric_pixel
:
3365 case nir_intrinsic_load_barycentric_centroid
:
3366 case nir_intrinsic_load_barycentric_sample
:
3367 /* Do nothing - load_interpolated_input handling will handle it later. */
3370 case nir_intrinsic_load_barycentric_at_sample
: {
3371 const glsl_interp_mode interpolation
=
3372 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(instr
);
3374 nir_const_value
*const_sample
= nir_src_as_const_value(instr
->src
[0]);
3377 unsigned msg_data
= const_sample
->i32
[0] << 4;
3379 emit_pixel_interpolater_send(bld
,
3380 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3383 brw_imm_ud(msg_data
),
3386 const fs_reg sample_src
= retype(get_nir_src(instr
->src
[0]),
3387 BRW_REGISTER_TYPE_UD
);
3389 if (nir_src_is_dynamically_uniform(instr
->src
[0])) {
3390 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3391 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3392 bld
.exec_all().group(1, 0)
3393 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3394 emit_pixel_interpolater_send(bld
,
3395 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3401 /* Make a loop that sends a message to the pixel interpolater
3402 * for the sample number in each live channel. If there are
3403 * multiple channels with the same sample number then these
3404 * will be handled simultaneously with a single interation of
3407 bld
.emit(BRW_OPCODE_DO
);
3409 /* Get the next live sample number into sample_id_reg */
3410 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3412 /* Set the flag register so that we can perform the send
3413 * message on all channels that have the same sample number
3415 bld
.CMP(bld
.null_reg_ud(),
3416 sample_src
, sample_id
,
3417 BRW_CONDITIONAL_EQ
);
3418 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3419 bld
.exec_all().group(1, 0)
3420 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3422 emit_pixel_interpolater_send(bld
,
3423 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3428 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
3430 /* Continue the loop if there are any live channels left */
3431 set_predicate_inv(BRW_PREDICATE_NORMAL
,
3433 bld
.emit(BRW_OPCODE_WHILE
));
3439 case nir_intrinsic_load_barycentric_at_offset
: {
3440 const glsl_interp_mode interpolation
=
3441 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(instr
);
3443 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3446 unsigned off_x
= MIN2((int)(const_offset
->f32
[0] * 16), 7) & 0xf;
3447 unsigned off_y
= MIN2((int)(const_offset
->f32
[1] * 16), 7) & 0xf;
3449 emit_pixel_interpolater_send(bld
,
3450 FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
,
3453 brw_imm_ud(off_x
| (off_y
<< 4)),
3456 fs_reg src
= vgrf(glsl_type::ivec2_type
);
3457 fs_reg offset_src
= retype(get_nir_src(instr
->src
[0]),
3458 BRW_REGISTER_TYPE_F
);
3459 for (int i
= 0; i
< 2; i
++) {
3460 fs_reg temp
= vgrf(glsl_type::float_type
);
3461 bld
.MUL(temp
, offset(offset_src
, bld
, i
), brw_imm_f(16.0f
));
3462 fs_reg itemp
= vgrf(glsl_type::int_type
);
3464 bld
.MOV(itemp
, temp
);
3466 /* Clamp the upper end of the range to +7/16.
3467 * ARB_gpu_shader5 requires that we support a maximum offset
3468 * of +0.5, which isn't representable in a S0.4 value -- if
3469 * we didn't clamp it, we'd end up with -8/16, which is the
3470 * opposite of what the shader author wanted.
3472 * This is legal due to ARB_gpu_shader5's quantization
3475 * "Not all values of <offset> may be supported; x and y
3476 * offsets may be rounded to fixed-point values with the
3477 * number of fraction bits given by the
3478 * implementation-dependent constant
3479 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
3481 set_condmod(BRW_CONDITIONAL_L
,
3482 bld
.SEL(offset(src
, bld
, i
), itemp
, brw_imm_d(7)));
3485 const enum opcode opcode
= FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
;
3486 emit_pixel_interpolater_send(bld
,
3496 case nir_intrinsic_load_interpolated_input
: {
3497 if (nir_intrinsic_base(instr
) == VARYING_SLOT_POS
) {
3498 emit_fragcoord_interpolation(dest
);
3502 assert(instr
->src
[0].ssa
&&
3503 instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_intrinsic
);
3504 nir_intrinsic_instr
*bary_intrinsic
=
3505 nir_instr_as_intrinsic(instr
->src
[0].ssa
->parent_instr
);
3506 nir_intrinsic_op bary_intrin
= bary_intrinsic
->intrinsic
;
3507 enum glsl_interp_mode interp_mode
=
3508 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(bary_intrinsic
);
3511 if (bary_intrin
== nir_intrinsic_load_barycentric_at_offset
||
3512 bary_intrin
== nir_intrinsic_load_barycentric_at_sample
) {
3513 /* Use the result of the PI message */
3514 dst_xy
= retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_F
);
3516 /* Use the delta_xy values computed from the payload */
3517 enum brw_barycentric_mode bary
=
3518 brw_barycentric_mode(interp_mode
, bary_intrin
);
3520 dst_xy
= this->delta_xy
[bary
];
3523 for (unsigned int i
= 0; i
< instr
->num_components
; i
++) {
3525 fs_reg(interp_reg(nir_intrinsic_base(instr
),
3526 nir_intrinsic_component(instr
) + i
));
3527 interp
.type
= BRW_REGISTER_TYPE_F
;
3528 dest
.type
= BRW_REGISTER_TYPE_F
;
3530 if (devinfo
->gen
< 6 && interp_mode
== INTERP_MODE_SMOOTH
) {
3531 fs_reg tmp
= vgrf(glsl_type::float_type
);
3532 bld
.emit(FS_OPCODE_LINTERP
, tmp
, dst_xy
, interp
);
3533 bld
.MUL(offset(dest
, bld
, i
), tmp
, this->pixel_w
);
3535 bld
.emit(FS_OPCODE_LINTERP
, offset(dest
, bld
, i
), dst_xy
, interp
);
3542 nir_emit_intrinsic(bld
, instr
);
3548 fs_visitor::nir_emit_cs_intrinsic(const fs_builder
&bld
,
3549 nir_intrinsic_instr
*instr
)
3551 assert(stage
== MESA_SHADER_COMPUTE
);
3552 struct brw_cs_prog_data
*cs_prog_data
= brw_cs_prog_data(prog_data
);
3555 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3556 dest
= get_nir_dest(instr
->dest
);
3558 switch (instr
->intrinsic
) {
3559 case nir_intrinsic_barrier
:
3561 cs_prog_data
->uses_barrier
= true;
3564 case nir_intrinsic_load_subgroup_id
:
3565 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
), subgroup_id
);
3568 case nir_intrinsic_load_local_invocation_id
:
3569 case nir_intrinsic_load_work_group_id
: {
3570 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3571 fs_reg val
= nir_system_values
[sv
];
3572 assert(val
.file
!= BAD_FILE
);
3573 dest
.type
= val
.type
;
3574 for (unsigned i
= 0; i
< 3; i
++)
3575 bld
.MOV(offset(dest
, bld
, i
), offset(val
, bld
, i
));
3579 case nir_intrinsic_load_num_work_groups
: {
3580 const unsigned surface
=
3581 cs_prog_data
->binding_table
.work_groups_start
;
3583 cs_prog_data
->uses_num_work_groups
= true;
3585 fs_reg surf_index
= brw_imm_ud(surface
);
3586 brw_mark_surface_used(prog_data
, surface
);
3588 /* Read the 3 GLuint components of gl_NumWorkGroups */
3589 for (unsigned i
= 0; i
< 3; i
++) {
3590 fs_reg read_result
=
3591 emit_untyped_read(bld
, surf_index
,
3593 1 /* dims */, 1 /* size */,
3594 BRW_PREDICATE_NONE
);
3595 read_result
.type
= dest
.type
;
3596 bld
.MOV(dest
, read_result
);
3597 dest
= offset(dest
, bld
, 1);
3602 case nir_intrinsic_shared_atomic_add
:
3603 nir_emit_shared_atomic(bld
, BRW_AOP_ADD
, instr
);
3605 case nir_intrinsic_shared_atomic_imin
:
3606 nir_emit_shared_atomic(bld
, BRW_AOP_IMIN
, instr
);
3608 case nir_intrinsic_shared_atomic_umin
:
3609 nir_emit_shared_atomic(bld
, BRW_AOP_UMIN
, instr
);
3611 case nir_intrinsic_shared_atomic_imax
:
3612 nir_emit_shared_atomic(bld
, BRW_AOP_IMAX
, instr
);
3614 case nir_intrinsic_shared_atomic_umax
:
3615 nir_emit_shared_atomic(bld
, BRW_AOP_UMAX
, instr
);
3617 case nir_intrinsic_shared_atomic_and
:
3618 nir_emit_shared_atomic(bld
, BRW_AOP_AND
, instr
);
3620 case nir_intrinsic_shared_atomic_or
:
3621 nir_emit_shared_atomic(bld
, BRW_AOP_OR
, instr
);
3623 case nir_intrinsic_shared_atomic_xor
:
3624 nir_emit_shared_atomic(bld
, BRW_AOP_XOR
, instr
);
3626 case nir_intrinsic_shared_atomic_exchange
:
3627 nir_emit_shared_atomic(bld
, BRW_AOP_MOV
, instr
);
3629 case nir_intrinsic_shared_atomic_comp_swap
:
3630 nir_emit_shared_atomic(bld
, BRW_AOP_CMPWR
, instr
);
3633 case nir_intrinsic_load_shared
: {
3634 assert(devinfo
->gen
>= 7);
3636 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
3638 /* Get the offset to read from */
3640 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3642 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0]);
3644 offset_reg
= vgrf(glsl_type::uint_type
);
3646 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
3647 brw_imm_ud(instr
->const_index
[0]));
3650 /* Read the vector */
3651 do_untyped_vector_read(bld
, dest
, surf_index
, offset_reg
,
3652 instr
->num_components
);
3656 case nir_intrinsic_store_shared
: {
3657 assert(devinfo
->gen
>= 7);
3660 fs_reg surf_index
= brw_imm_ud(GEN7_BTI_SLM
);
3663 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
3666 unsigned writemask
= instr
->const_index
[1];
3668 /* get_nir_src() retypes to integer. Be wary of 64-bit types though
3669 * since the untyped writes below operate in units of 32-bits, which
3670 * means that we need to write twice as many components each time.
3671 * Also, we have to suffle 64-bit data to be in the appropriate layout
3672 * expected by our 32-bit write messages.
3674 unsigned type_size
= 4;
3675 if (nir_src_bit_size(instr
->src
[0]) == 64) {
3677 val_reg
= shuffle_64bit_data_for_32bit_write(bld
,
3678 val_reg
, instr
->num_components
);
3681 unsigned type_slots
= type_size
/ 4;
3683 /* Combine groups of consecutive enabled channels in one write
3684 * message. We use ffs to find the first enabled channel and then ffs on
3685 * the bit-inverse, down-shifted writemask to determine the length of
3686 * the block of enabled bits.
3689 unsigned first_component
= ffs(writemask
) - 1;
3690 unsigned length
= ffs(~(writemask
>> first_component
)) - 1;
3692 /* We can't write more than 2 64-bit components at once. Limit the
3693 * length of the write to what we can do and let the next iteration
3697 length
= MIN2(2, length
);
3700 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
3702 offset_reg
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0] +
3703 type_size
* first_component
);
3705 offset_reg
= vgrf(glsl_type::uint_type
);
3707 retype(get_nir_src(instr
->src
[1]), BRW_REGISTER_TYPE_UD
),
3708 brw_imm_ud(instr
->const_index
[0] + type_size
* first_component
));
3711 emit_untyped_write(bld
, surf_index
, offset_reg
,
3712 offset(val_reg
, bld
, first_component
* type_slots
),
3713 1 /* dims */, length
* type_slots
,
3714 BRW_PREDICATE_NONE
);
3716 /* Clear the bits in the writemask that we just wrote, then try
3717 * again to see if more channels are left.
3719 writemask
&= (15 << (first_component
+ length
));
3726 nir_emit_intrinsic(bld
, instr
);
3732 brw_nir_reduction_op_identity(const fs_builder
&bld
,
3733 nir_op op
, brw_reg_type type
)
3735 nir_const_value value
= nir_alu_binop_identity(op
, type_sz(type
) * 8);
3736 switch (type_sz(type
)) {
3738 assert(type
!= BRW_REGISTER_TYPE_HF
);
3739 return retype(brw_imm_uw(value
.u16
[0]), type
);
3741 return retype(brw_imm_ud(value
.u32
[0]), type
);
3743 if (type
== BRW_REGISTER_TYPE_DF
)
3744 return setup_imm_df(bld
, value
.f64
[0]);
3746 return retype(brw_imm_u64(value
.u64
[0]), type
);
3748 unreachable("Invalid type size");
3753 brw_op_for_nir_reduction_op(nir_op op
)
3756 case nir_op_iadd
: return BRW_OPCODE_ADD
;
3757 case nir_op_fadd
: return BRW_OPCODE_ADD
;
3758 case nir_op_imul
: return BRW_OPCODE_MUL
;
3759 case nir_op_fmul
: return BRW_OPCODE_MUL
;
3760 case nir_op_imin
: return BRW_OPCODE_SEL
;
3761 case nir_op_umin
: return BRW_OPCODE_SEL
;
3762 case nir_op_fmin
: return BRW_OPCODE_SEL
;
3763 case nir_op_imax
: return BRW_OPCODE_SEL
;
3764 case nir_op_umax
: return BRW_OPCODE_SEL
;
3765 case nir_op_fmax
: return BRW_OPCODE_SEL
;
3766 case nir_op_iand
: return BRW_OPCODE_AND
;
3767 case nir_op_ior
: return BRW_OPCODE_OR
;
3768 case nir_op_ixor
: return BRW_OPCODE_XOR
;
3770 unreachable("Invalid reduction operation");
3774 static brw_conditional_mod
3775 brw_cond_mod_for_nir_reduction_op(nir_op op
)
3778 case nir_op_iadd
: return BRW_CONDITIONAL_NONE
;
3779 case nir_op_fadd
: return BRW_CONDITIONAL_NONE
;
3780 case nir_op_imul
: return BRW_CONDITIONAL_NONE
;
3781 case nir_op_fmul
: return BRW_CONDITIONAL_NONE
;
3782 case nir_op_imin
: return BRW_CONDITIONAL_L
;
3783 case nir_op_umin
: return BRW_CONDITIONAL_L
;
3784 case nir_op_fmin
: return BRW_CONDITIONAL_L
;
3785 case nir_op_imax
: return BRW_CONDITIONAL_GE
;
3786 case nir_op_umax
: return BRW_CONDITIONAL_GE
;
3787 case nir_op_fmax
: return BRW_CONDITIONAL_GE
;
3788 case nir_op_iand
: return BRW_CONDITIONAL_NONE
;
3789 case nir_op_ior
: return BRW_CONDITIONAL_NONE
;
3790 case nir_op_ixor
: return BRW_CONDITIONAL_NONE
;
3792 unreachable("Invalid reduction operation");
3797 fs_visitor::nir_emit_intrinsic(const fs_builder
&bld
, nir_intrinsic_instr
*instr
)
3800 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3801 dest
= get_nir_dest(instr
->dest
);
3803 switch (instr
->intrinsic
) {
3804 case nir_intrinsic_image_var_load
:
3805 case nir_intrinsic_image_var_store
:
3806 case nir_intrinsic_image_var_atomic_add
:
3807 case nir_intrinsic_image_var_atomic_min
:
3808 case nir_intrinsic_image_var_atomic_max
:
3809 case nir_intrinsic_image_var_atomic_and
:
3810 case nir_intrinsic_image_var_atomic_or
:
3811 case nir_intrinsic_image_var_atomic_xor
:
3812 case nir_intrinsic_image_var_atomic_exchange
:
3813 case nir_intrinsic_image_var_atomic_comp_swap
: {
3814 using namespace image_access
;
3816 if (stage
== MESA_SHADER_FRAGMENT
&&
3817 instr
->intrinsic
!= nir_intrinsic_image_var_load
)
3818 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
3820 /* Get the referenced image variable and type. */
3821 const nir_variable
*var
= instr
->variables
[0]->var
;
3822 const glsl_type
*type
= var
->type
->without_array();
3823 const brw_reg_type base_type
= get_image_base_type(type
);
3825 /* Get some metadata from the image intrinsic. */
3826 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
3827 const unsigned arr_dims
= type
->sampler_array
? 1 : 0;
3828 const unsigned surf_dims
= type
->coordinate_components() - arr_dims
;
3829 const unsigned format
= var
->data
.image
.format
;
3830 const unsigned dest_components
= nir_intrinsic_dest_components(instr
);
3832 /* Get the arguments of the image intrinsic. */
3833 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
3834 const fs_reg addr
= retype(get_nir_src(instr
->src
[0]),
3835 BRW_REGISTER_TYPE_UD
);
3836 const fs_reg src0
= (info
->num_srcs
>= 3 ?
3837 retype(get_nir_src(instr
->src
[2]), base_type
) :
3839 const fs_reg src1
= (info
->num_srcs
>= 4 ?
3840 retype(get_nir_src(instr
->src
[3]), base_type
) :
3844 /* Emit an image load, store or atomic op. */
3845 if (instr
->intrinsic
== nir_intrinsic_image_var_load
)
3846 tmp
= emit_image_load(bld
, image
, addr
, surf_dims
, arr_dims
, format
);
3848 else if (instr
->intrinsic
== nir_intrinsic_image_var_store
)
3849 emit_image_store(bld
, image
, addr
, src0
, surf_dims
, arr_dims
,
3850 var
->data
.image
.write_only
? GL_NONE
: format
);
3853 tmp
= emit_image_atomic(bld
, image
, addr
, src0
, src1
,
3854 surf_dims
, arr_dims
, dest_components
,
3855 get_image_atomic_op(instr
->intrinsic
, type
));
3857 /* Assign the result. */
3858 for (unsigned c
= 0; c
< dest_components
; ++c
) {
3859 bld
.MOV(offset(retype(dest
, base_type
), bld
, c
),
3860 offset(tmp
, bld
, c
));
3865 case nir_intrinsic_memory_barrier_atomic_counter
:
3866 case nir_intrinsic_memory_barrier_buffer
:
3867 case nir_intrinsic_memory_barrier_image
:
3868 case nir_intrinsic_memory_barrier
: {
3869 const fs_builder ubld
= bld
.group(8, 0);
3870 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
3871 ubld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
)
3872 ->size_written
= 2 * REG_SIZE
;
3876 case nir_intrinsic_group_memory_barrier
:
3877 case nir_intrinsic_memory_barrier_shared
:
3878 /* We treat these workgroup-level barriers as no-ops. This should be
3879 * safe at present and as long as:
3881 * - Memory access instructions are not subsequently reordered by the
3882 * compiler back-end.
3884 * - All threads from a given compute shader workgroup fit within a
3885 * single subslice and therefore talk to the same HDC shared unit
3886 * what supposedly guarantees ordering and coherency between threads
3887 * from the same workgroup. This may change in the future when we
3888 * start splitting workgroups across multiple subslices.
3890 * - The context is not in fault-and-stream mode, which could cause
3891 * memory transactions (including to SLM) prior to the barrier to be
3892 * replayed after the barrier if a pagefault occurs. This shouldn't
3893 * be a problem up to and including SKL because fault-and-stream is
3894 * not usable due to hardware issues, but that's likely to change in
3899 case nir_intrinsic_shader_clock
: {
3900 /* We cannot do anything if there is an event, so ignore it for now */
3901 const fs_reg shader_clock
= get_timestamp(bld
);
3902 const fs_reg srcs
[] = { component(shader_clock
, 0),
3903 component(shader_clock
, 1) };
3904 bld
.LOAD_PAYLOAD(dest
, srcs
, ARRAY_SIZE(srcs
), 0);
3908 case nir_intrinsic_image_var_size
: {
3909 /* Get the referenced image variable and type. */
3910 const nir_variable
*var
= instr
->variables
[0]->var
;
3911 const glsl_type
*type
= var
->type
->without_array();
3913 /* Get the size of the image. */
3914 const fs_reg image
= get_nir_image_deref(instr
->variables
[0]);
3915 const fs_reg size
= offset(image
, bld
, BRW_IMAGE_PARAM_SIZE_OFFSET
);
3917 /* For 1DArray image types, the array index is stored in the Z component.
3918 * Fix this by swizzling the Z component to the Y component.
3920 const bool is_1d_array_image
=
3921 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_1D
&&
3922 type
->sampler_array
;
3924 /* For CubeArray images, we should count the number of cubes instead
3925 * of the number of faces. Fix it by dividing the (Z component) by 6.
3927 const bool is_cube_array_image
=
3928 type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_CUBE
&&
3929 type
->sampler_array
;
3931 /* Copy all the components. */
3932 for (unsigned c
= 0; c
< instr
->dest
.ssa
.num_components
; ++c
) {
3933 if ((int)c
>= type
->coordinate_components()) {
3934 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3936 } else if (c
== 1 && is_1d_array_image
) {
3937 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3938 offset(size
, bld
, 2));
3939 } else if (c
== 2 && is_cube_array_image
) {
3940 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
,
3941 offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3942 offset(size
, bld
, c
), brw_imm_d(6));
3944 bld
.MOV(offset(retype(dest
, BRW_REGISTER_TYPE_D
), bld
, c
),
3945 offset(size
, bld
, c
));
3952 case nir_intrinsic_image_var_samples
:
3953 /* The driver does not support multi-sampled images. */
3954 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), brw_imm_d(1));
3957 case nir_intrinsic_load_uniform
: {
3958 /* Offsets are in bytes but they should always aligned to
3961 assert(instr
->const_index
[0] % 4 == 0 ||
3962 instr
->const_index
[0] % type_sz(dest
.type
) == 0);
3964 fs_reg
src(UNIFORM
, instr
->const_index
[0] / 4, dest
.type
);
3966 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3968 assert(const_offset
->u32
[0] % type_sz(dest
.type
) == 0);
3969 /* For 16-bit types we add the module of the const_index[0]
3970 * offset to access to not 32-bit aligned element
3972 src
.offset
= const_offset
->u32
[0] + instr
->const_index
[0] % 4;
3974 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3975 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
3978 fs_reg indirect
= retype(get_nir_src(instr
->src
[0]),
3979 BRW_REGISTER_TYPE_UD
);
3981 /* We need to pass a size to the MOV_INDIRECT but we don't want it to
3982 * go past the end of the uniform. In order to keep the n'th
3983 * component from running past, we subtract off the size of all but
3984 * one component of the vector.
3986 assert(instr
->const_index
[1] >=
3987 instr
->num_components
* (int) type_sz(dest
.type
));
3988 unsigned read_size
= instr
->const_index
[1] -
3989 (instr
->num_components
- 1) * type_sz(dest
.type
);
3991 bool supports_64bit_indirects
=
3992 !devinfo
->is_cherryview
&& !gen_device_info_is_9lp(devinfo
);
3994 if (type_sz(dest
.type
) != 8 || supports_64bit_indirects
) {
3995 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3996 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
3997 offset(dest
, bld
, j
), offset(src
, bld
, j
),
3998 indirect
, brw_imm_ud(read_size
));
4001 const unsigned num_mov_indirects
=
4002 type_sz(dest
.type
) / type_sz(BRW_REGISTER_TYPE_UD
);
4003 /* We read a little bit less per MOV INDIRECT, as they are now
4004 * 32-bits ones instead of 64-bit. Fix read_size then.
4006 const unsigned read_size_32bit
= read_size
-
4007 (num_mov_indirects
- 1) * type_sz(BRW_REGISTER_TYPE_UD
);
4008 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
4009 for (unsigned i
= 0; i
< num_mov_indirects
; i
++) {
4010 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
4011 subscript(offset(dest
, bld
, j
), BRW_REGISTER_TYPE_UD
, i
),
4012 subscript(offset(src
, bld
, j
), BRW_REGISTER_TYPE_UD
, i
),
4013 indirect
, brw_imm_ud(read_size_32bit
));
4021 case nir_intrinsic_load_ubo
: {
4022 nir_const_value
*const_index
= nir_src_as_const_value(instr
->src
[0]);
4026 const unsigned index
= stage_prog_data
->binding_table
.ubo_start
+
4027 const_index
->u32
[0];
4028 surf_index
= brw_imm_ud(index
);
4029 brw_mark_surface_used(prog_data
, index
);
4031 /* The block index is not a constant. Evaluate the index expression
4032 * per-channel and add the base UBO index; we have to select a value
4033 * from any live channel.
4035 surf_index
= vgrf(glsl_type::uint_type
);
4036 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
4037 brw_imm_ud(stage_prog_data
->binding_table
.ubo_start
));
4038 surf_index
= bld
.emit_uniformize(surf_index
);
4040 /* Assume this may touch any UBO. It would be nice to provide
4041 * a tighter bound, but the array information is already lowered away.
4043 brw_mark_surface_used(prog_data
,
4044 stage_prog_data
->binding_table
.ubo_start
+
4045 nir
->info
.num_ubos
- 1);
4048 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
4049 if (const_offset
== NULL
) {
4050 fs_reg base_offset
= retype(get_nir_src(instr
->src
[1]),
4051 BRW_REGISTER_TYPE_UD
);
4053 for (int i
= 0; i
< instr
->num_components
; i
++)
4054 VARYING_PULL_CONSTANT_LOAD(bld
, offset(dest
, bld
, i
), surf_index
,
4055 base_offset
, i
* type_sz(dest
.type
));
4057 /* Even if we are loading doubles, a pull constant load will load
4058 * a 32-bit vec4, so should only reserve vgrf space for that. If we
4059 * need to load a full dvec4 we will have to emit 2 loads. This is
4060 * similar to demote_pull_constants(), except that in that case we
4061 * see individual accesses to each component of the vector and then
4062 * we let CSE deal with duplicate loads. Here we see a vector access
4063 * and we have to split it if necessary.
4065 const unsigned type_size
= type_sz(dest
.type
);
4067 /* See if we've selected this as a push constant candidate */
4069 const unsigned ubo_block
= const_index
->u32
[0];
4070 const unsigned offset_256b
= const_offset
->u32
[0] / 32;
4073 for (int i
= 0; i
< 4; i
++) {
4074 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4075 if (range
->block
== ubo_block
&&
4076 offset_256b
>= range
->start
&&
4077 offset_256b
< range
->start
+ range
->length
) {
4079 push_reg
= fs_reg(UNIFORM
, UBO_START
+ i
, dest
.type
);
4080 push_reg
.offset
= const_offset
->u32
[0] - 32 * range
->start
;
4085 if (push_reg
.file
!= BAD_FILE
) {
4086 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
4087 bld
.MOV(offset(dest
, bld
, i
),
4088 byte_offset(push_reg
, i
* type_size
));
4094 const unsigned block_sz
= 64; /* Fetch one cacheline at a time. */
4095 const fs_builder ubld
= bld
.exec_all().group(block_sz
/ 4, 0);
4096 const fs_reg packed_consts
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4098 for (unsigned c
= 0; c
< instr
->num_components
;) {
4099 const unsigned base
= const_offset
->u32
[0] + c
* type_size
;
4100 /* Number of usable components in the next block-aligned load. */
4101 const unsigned count
= MIN2(instr
->num_components
- c
,
4102 (block_sz
- base
% block_sz
) / type_size
);
4104 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
4105 packed_consts
, surf_index
,
4106 brw_imm_ud(base
& ~(block_sz
- 1)));
4108 const fs_reg consts
=
4109 retype(byte_offset(packed_consts
, base
& (block_sz
- 1)),
4112 for (unsigned d
= 0; d
< count
; d
++)
4113 bld
.MOV(offset(dest
, bld
, c
+ d
), component(consts
, d
));
4121 case nir_intrinsic_load_ssbo
: {
4122 assert(devinfo
->gen
>= 7);
4124 nir_const_value
*const_uniform_block
=
4125 nir_src_as_const_value(instr
->src
[0]);
4128 if (const_uniform_block
) {
4129 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
4130 const_uniform_block
->u32
[0];
4131 surf_index
= brw_imm_ud(index
);
4132 brw_mark_surface_used(prog_data
, index
);
4134 surf_index
= vgrf(glsl_type::uint_type
);
4135 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
4136 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
4138 /* Assume this may touch any UBO. It would be nice to provide
4139 * a tighter bound, but the array information is already lowered away.
4141 brw_mark_surface_used(prog_data
,
4142 stage_prog_data
->binding_table
.ssbo_start
+
4143 nir
->info
.num_ssbos
- 1);
4147 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
4149 offset_reg
= brw_imm_ud(const_offset
->u32
[0]);
4151 offset_reg
= retype(get_nir_src(instr
->src
[1]), BRW_REGISTER_TYPE_UD
);
4154 /* Read the vector */
4155 do_untyped_vector_read(bld
, dest
, surf_index
, offset_reg
,
4156 instr
->num_components
);
4161 case nir_intrinsic_store_ssbo
: {
4162 assert(devinfo
->gen
>= 7);
4164 if (stage
== MESA_SHADER_FRAGMENT
)
4165 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4169 nir_const_value
*const_uniform_block
=
4170 nir_src_as_const_value(instr
->src
[1]);
4171 if (const_uniform_block
) {
4172 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
4173 const_uniform_block
->u32
[0];
4174 surf_index
= brw_imm_ud(index
);
4175 brw_mark_surface_used(prog_data
, index
);
4177 surf_index
= vgrf(glsl_type::uint_type
);
4178 bld
.ADD(surf_index
, get_nir_src(instr
->src
[1]),
4179 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
4181 brw_mark_surface_used(prog_data
,
4182 stage_prog_data
->binding_table
.ssbo_start
+
4183 nir
->info
.num_ssbos
- 1);
4187 fs_reg val_reg
= get_nir_src(instr
->src
[0]);
4190 unsigned writemask
= instr
->const_index
[0];
4192 /* get_nir_src() retypes to integer. Be wary of 64-bit types though
4193 * since the untyped writes below operate in units of 32-bits, which
4194 * means that we need to write twice as many components each time.
4195 * Also, we have to suffle 64-bit data to be in the appropriate layout
4196 * expected by our 32-bit write messages.
4198 unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4199 unsigned type_size
= bit_size
/ 8;
4201 /* Combine groups of consecutive enabled channels in one write
4202 * message. We use ffs to find the first enabled channel and then ffs on
4203 * the bit-inverse, down-shifted writemask to determine the num_components
4204 * of the block of enabled bits.
4207 unsigned first_component
= ffs(writemask
) - 1;
4208 unsigned num_components
= ffs(~(writemask
>> first_component
)) - 1;
4209 fs_reg write_src
= offset(val_reg
, bld
, first_component
);
4211 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[2]);
4213 if (type_size
> 4) {
4214 /* We can't write more than 2 64-bit components at once. Limit
4215 * the num_components of the write to what we can do and let the next
4216 * iteration handle the rest.
4218 num_components
= MIN2(2, num_components
);
4219 write_src
= shuffle_64bit_data_for_32bit_write(bld
, write_src
,
4221 } else if (type_size
< 4) {
4222 assert(type_size
== 2);
4223 /* For 16-bit types we pack two consecutive values into a 32-bit
4224 * word and use an untyped write message. For single values or not
4225 * 32-bit-aligned we need to use byte-scattered writes because
4226 * untyped writes works with 32-bit components with 32-bit
4227 * alignment. byte_scattered_write messages only support one
4228 * 16-bit component at a time. As VK_KHR_relaxed_block_layout
4229 * could be enabled we can not guarantee that not constant offsets
4230 * to be 32-bit aligned for 16-bit types. For example an array, of
4231 * 16-bit vec3 with array element stride of 6.
4233 * In the case of 32-bit aligned constant offsets if there is
4234 * a 3-components vector we submit one untyped-write message
4235 * of 32-bit (first two components), and one byte-scattered
4236 * write message (the last component).
4239 if ( !const_offset
|| ((const_offset
->u32
[0] +
4240 type_size
* first_component
) % 4)) {
4241 /* If we use a .yz writemask we also need to emit 2
4242 * byte-scattered write messages because of y-component not
4243 * being aligned to 32-bit.
4246 } else if (num_components
> 2 && (num_components
% 2)) {
4247 /* If there is an odd number of consecutive components we left
4248 * the not paired component for a following emit of length == 1
4249 * with byte_scattered_write.
4253 /* For num_components == 1 we are also shuffling the component
4254 * because byte scattered writes of 16-bit need values to be dword
4255 * aligned. Shuffling only one component would be the same as
4258 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
,
4259 DIV_ROUND_UP(num_components
, 2));
4260 shuffle_16bit_data_for_32bit_write(bld
, tmp
, write_src
,
4268 offset_reg
= brw_imm_ud(const_offset
->u32
[0] +
4269 type_size
* first_component
);
4271 offset_reg
= vgrf(glsl_type::uint_type
);
4273 retype(get_nir_src(instr
->src
[2]), BRW_REGISTER_TYPE_UD
),
4274 brw_imm_ud(type_size
* first_component
));
4277 if (type_size
< 4 && num_components
== 1) {
4278 assert(type_size
== 2);
4279 /* Untyped Surface messages have a fixed 32-bit size, so we need
4280 * to rely on byte scattered in order to write 16-bit elements.
4281 * The byte_scattered_write message needs that every written 16-bit
4282 * type to be aligned 32-bits (stride=2).
4284 emit_byte_scattered_write(bld
, surf_index
, offset_reg
,
4288 BRW_PREDICATE_NONE
);
4290 assert(num_components
* type_size
<= 16);
4291 assert((num_components
* type_size
) % 4 == 0);
4292 assert(offset_reg
.file
!= BRW_IMMEDIATE_VALUE
||
4293 offset_reg
.ud
% 4 == 0);
4294 unsigned num_slots
= (num_components
* type_size
) / 4;
4296 emit_untyped_write(bld
, surf_index
, offset_reg
,
4298 1 /* dims */, num_slots
,
4299 BRW_PREDICATE_NONE
);
4302 /* Clear the bits in the writemask that we just wrote, then try
4303 * again to see if more channels are left.
4305 writemask
&= (15 << (first_component
+ num_components
));
4310 case nir_intrinsic_store_output
: {
4311 fs_reg src
= get_nir_src(instr
->src
[0]);
4313 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[1]);
4314 assert(const_offset
&& "Indirect output stores not allowed");
4316 unsigned num_components
= instr
->num_components
;
4317 unsigned first_component
= nir_intrinsic_component(instr
);
4318 if (nir_src_bit_size(instr
->src
[0]) == 64) {
4319 src
= shuffle_64bit_data_for_32bit_write(bld
, src
, num_components
);
4320 num_components
*= 2;
4323 fs_reg new_dest
= retype(offset(outputs
[instr
->const_index
[0]], bld
,
4324 4 * const_offset
->u32
[0]), src
.type
);
4325 for (unsigned j
= 0; j
< num_components
; j
++) {
4326 bld
.MOV(offset(new_dest
, bld
, j
+ first_component
),
4327 offset(src
, bld
, j
));
4332 case nir_intrinsic_ssbo_atomic_add
:
4333 nir_emit_ssbo_atomic(bld
, BRW_AOP_ADD
, instr
);
4335 case nir_intrinsic_ssbo_atomic_imin
:
4336 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMIN
, instr
);
4338 case nir_intrinsic_ssbo_atomic_umin
:
4339 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMIN
, instr
);
4341 case nir_intrinsic_ssbo_atomic_imax
:
4342 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMAX
, instr
);
4344 case nir_intrinsic_ssbo_atomic_umax
:
4345 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMAX
, instr
);
4347 case nir_intrinsic_ssbo_atomic_and
:
4348 nir_emit_ssbo_atomic(bld
, BRW_AOP_AND
, instr
);
4350 case nir_intrinsic_ssbo_atomic_or
:
4351 nir_emit_ssbo_atomic(bld
, BRW_AOP_OR
, instr
);
4353 case nir_intrinsic_ssbo_atomic_xor
:
4354 nir_emit_ssbo_atomic(bld
, BRW_AOP_XOR
, instr
);
4356 case nir_intrinsic_ssbo_atomic_exchange
:
4357 nir_emit_ssbo_atomic(bld
, BRW_AOP_MOV
, instr
);
4359 case nir_intrinsic_ssbo_atomic_comp_swap
:
4360 nir_emit_ssbo_atomic(bld
, BRW_AOP_CMPWR
, instr
);
4363 case nir_intrinsic_get_buffer_size
: {
4364 nir_const_value
*const_uniform_block
= nir_src_as_const_value(instr
->src
[0]);
4365 unsigned ssbo_index
= const_uniform_block
? const_uniform_block
->u32
[0] : 0;
4367 /* A resinfo's sampler message is used to get the buffer size. The
4368 * SIMD8's writeback message consists of four registers and SIMD16's
4369 * writeback message consists of 8 destination registers (two per each
4370 * component). Because we are only interested on the first channel of
4371 * the first returned component, where resinfo returns the buffer size
4372 * for SURFTYPE_BUFFER, we can just use the SIMD8 variant regardless of
4373 * the dispatch width.
4375 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4376 fs_reg src_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4377 fs_reg ret_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 4);
4380 ubld
.MOV(src_payload
, brw_imm_d(0));
4382 const unsigned index
= prog_data
->binding_table
.ssbo_start
+ ssbo_index
;
4383 fs_inst
*inst
= ubld
.emit(SHADER_OPCODE_GET_BUFFER_SIZE
, ret_payload
,
4384 src_payload
, brw_imm_ud(index
));
4385 inst
->header_size
= 0;
4387 inst
->size_written
= 4 * REG_SIZE
;
4389 /* SKL PRM, vol07, 3D Media GPGPU Engine, Bounds Checking and Faulting:
4391 * "Out-of-bounds checking is always performed at a DWord granularity. If
4392 * any part of the DWord is out-of-bounds then the whole DWord is
4393 * considered out-of-bounds."
4395 * This implies that types with size smaller than 4-bytes need to be
4396 * padded if they don't complete the last dword of the buffer. But as we
4397 * need to maintain the original size we need to reverse the padding
4398 * calculation to return the correct size to know the number of elements
4399 * of an unsized array. As we stored in the last two bits of the surface
4400 * size the needed padding for the buffer, we calculate here the
4401 * original buffer_size reversing the surface_size calculation:
4403 * surface_size = isl_align(buffer_size, 4) +
4404 * (isl_align(buffer_size) - buffer_size)
4406 * buffer_size = surface_size & ~3 - surface_size & 3
4409 fs_reg size_aligned4
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4410 fs_reg size_padding
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4411 fs_reg buffer_size
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4413 ubld
.AND(size_padding
, ret_payload
, brw_imm_ud(3));
4414 ubld
.AND(size_aligned4
, ret_payload
, brw_imm_ud(~3));
4415 ubld
.ADD(buffer_size
, size_aligned4
, negate(size_padding
));
4417 bld
.MOV(retype(dest
, ret_payload
.type
), component(buffer_size
, 0));
4419 brw_mark_surface_used(prog_data
, index
);
4423 case nir_intrinsic_load_subgroup_invocation
:
4424 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
4425 nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
]);
4428 case nir_intrinsic_load_subgroup_eq_mask
:
4429 case nir_intrinsic_load_subgroup_ge_mask
:
4430 case nir_intrinsic_load_subgroup_gt_mask
:
4431 case nir_intrinsic_load_subgroup_le_mask
:
4432 case nir_intrinsic_load_subgroup_lt_mask
:
4433 unreachable("not reached");
4435 case nir_intrinsic_vote_any
: {
4436 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4438 /* The any/all predicates do not consider channel enables. To prevent
4439 * dead channels from affecting the result, we initialize the flag with
4440 * with the identity value for the logical operation.
4442 if (dispatch_width
== 32) {
4443 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4444 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4447 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0));
4449 bld
.CMP(bld
.null_reg_d(), get_nir_src(instr
->src
[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ
);
4451 /* For some reason, the any/all predicates don't work properly with
4452 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4453 * doesn't read the correct subset of the flag register and you end up
4454 * getting garbage in the second half. Work around this by using a pair
4455 * of 1-wide MOVs and scattering the result.
4457 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4458 ubld
.MOV(res1
, brw_imm_d(0));
4459 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ANY8H
:
4460 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ANY16H
:
4461 BRW_PREDICATE_ALIGN1_ANY32H
,
4462 ubld
.MOV(res1
, brw_imm_d(-1)));
4464 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4467 case nir_intrinsic_vote_all
: {
4468 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4470 /* The any/all predicates do not consider channel enables. To prevent
4471 * dead channels from affecting the result, we initialize the flag with
4472 * with the identity value for the logical operation.
4474 if (dispatch_width
== 32) {
4475 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4476 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4477 brw_imm_ud(0xffffffff));
4479 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0xffff));
4481 bld
.CMP(bld
.null_reg_d(), get_nir_src(instr
->src
[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ
);
4483 /* For some reason, the any/all predicates don't work properly with
4484 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4485 * doesn't read the correct subset of the flag register and you end up
4486 * getting garbage in the second half. Work around this by using a pair
4487 * of 1-wide MOVs and scattering the result.
4489 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4490 ubld
.MOV(res1
, brw_imm_d(0));
4491 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ALL8H
:
4492 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ALL16H
:
4493 BRW_PREDICATE_ALIGN1_ALL32H
,
4494 ubld
.MOV(res1
, brw_imm_d(-1)));
4496 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4499 case nir_intrinsic_vote_feq
:
4500 case nir_intrinsic_vote_ieq
: {
4501 fs_reg value
= get_nir_src(instr
->src
[0]);
4502 if (instr
->intrinsic
== nir_intrinsic_vote_feq
) {
4503 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4504 value
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_F
);
4507 fs_reg uniformized
= bld
.emit_uniformize(value
);
4508 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4510 /* The any/all predicates do not consider channel enables. To prevent
4511 * dead channels from affecting the result, we initialize the flag with
4512 * with the identity value for the logical operation.
4514 if (dispatch_width
== 32) {
4515 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4516 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4517 brw_imm_ud(0xffffffff));
4519 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0xffff));
4521 bld
.CMP(bld
.null_reg_d(), value
, uniformized
, BRW_CONDITIONAL_Z
);
4523 /* For some reason, the any/all predicates don't work properly with
4524 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4525 * doesn't read the correct subset of the flag register and you end up
4526 * getting garbage in the second half. Work around this by using a pair
4527 * of 1-wide MOVs and scattering the result.
4529 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4530 ubld
.MOV(res1
, brw_imm_d(0));
4531 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ALL8H
:
4532 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ALL16H
:
4533 BRW_PREDICATE_ALIGN1_ALL32H
,
4534 ubld
.MOV(res1
, brw_imm_d(-1)));
4536 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4540 case nir_intrinsic_ballot
: {
4541 const fs_reg value
= retype(get_nir_src(instr
->src
[0]),
4542 BRW_REGISTER_TYPE_UD
);
4543 struct brw_reg flag
= brw_flag_reg(0, 0);
4544 /* FIXME: For SIMD32 programs, this causes us to stomp on f0.1 as well
4545 * as f0.0. This is a problem for fragment programs as we currently use
4546 * f0.1 for discards. Fortunately, we don't support SIMD32 fragment
4547 * programs yet so this isn't a problem. When we do, something will
4550 if (dispatch_width
== 32)
4551 flag
.type
= BRW_REGISTER_TYPE_UD
;
4553 bld
.exec_all().group(1, 0).MOV(flag
, brw_imm_ud(0u));
4554 bld
.CMP(bld
.null_reg_ud(), value
, brw_imm_ud(0u), BRW_CONDITIONAL_NZ
);
4556 if (instr
->dest
.ssa
.bit_size
> 32) {
4557 dest
.type
= BRW_REGISTER_TYPE_UQ
;
4559 dest
.type
= BRW_REGISTER_TYPE_UD
;
4561 bld
.MOV(dest
, flag
);
4565 case nir_intrinsic_read_invocation
: {
4566 const fs_reg value
= get_nir_src(instr
->src
[0]);
4567 const fs_reg invocation
= get_nir_src(instr
->src
[1]);
4568 fs_reg tmp
= bld
.vgrf(value
.type
);
4570 bld
.exec_all().emit(SHADER_OPCODE_BROADCAST
, tmp
, value
,
4571 bld
.emit_uniformize(invocation
));
4573 bld
.MOV(retype(dest
, value
.type
), fs_reg(component(tmp
, 0)));
4577 case nir_intrinsic_read_first_invocation
: {
4578 const fs_reg value
= get_nir_src(instr
->src
[0]);
4579 bld
.MOV(retype(dest
, value
.type
), bld
.emit_uniformize(value
));
4583 case nir_intrinsic_shuffle
: {
4584 const fs_reg value
= get_nir_src(instr
->src
[0]);
4585 const fs_reg index
= get_nir_src(instr
->src
[1]);
4587 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, index
);
4591 case nir_intrinsic_first_invocation
: {
4592 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4593 bld
.exec_all().emit(SHADER_OPCODE_FIND_LIVE_CHANNEL
, tmp
);
4594 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
4595 fs_reg(component(tmp
, 0)));
4599 case nir_intrinsic_quad_broadcast
: {
4600 const fs_reg value
= get_nir_src(instr
->src
[0]);
4601 nir_const_value
*index
= nir_src_as_const_value(instr
->src
[1]);
4602 assert(nir_src_bit_size(instr
->src
[1]) == 32);
4604 bld
.emit(SHADER_OPCODE_CLUSTER_BROADCAST
, retype(dest
, value
.type
),
4605 value
, brw_imm_ud(index
->u32
[0]), brw_imm_ud(4));
4609 case nir_intrinsic_quad_swap_horizontal
: {
4610 const fs_reg value
= get_nir_src(instr
->src
[0]);
4611 const fs_reg tmp
= bld
.vgrf(value
.type
);
4612 const fs_builder ubld
= bld
.exec_all().group(dispatch_width
/ 2, 0);
4614 const fs_reg src_left
= horiz_stride(value
, 2);
4615 const fs_reg src_right
= horiz_stride(horiz_offset(value
, 1), 2);
4616 const fs_reg tmp_left
= horiz_stride(tmp
, 2);
4617 const fs_reg tmp_right
= horiz_stride(horiz_offset(tmp
, 1), 2);
4619 /* From the Cherryview PRM Vol. 7, "Register Region Restrictiosn":
4621 * "When source or destination datatype is 64b or operation is
4622 * integer DWord multiply, regioning in Align1 must follow
4627 * 3. Source and Destination offset must be the same, except
4628 * the case of scalar source."
4630 * In order to work around this, we have to emit two 32-bit MOVs instead
4631 * of a single 64-bit MOV to do the shuffle.
4633 if (type_sz(value
.type
) > 4 &&
4634 (devinfo
->is_cherryview
|| gen_device_info_is_9lp(devinfo
))) {
4635 ubld
.MOV(subscript(tmp_left
, BRW_REGISTER_TYPE_D
, 0),
4636 subscript(src_right
, BRW_REGISTER_TYPE_D
, 0));
4637 ubld
.MOV(subscript(tmp_left
, BRW_REGISTER_TYPE_D
, 1),
4638 subscript(src_right
, BRW_REGISTER_TYPE_D
, 1));
4639 ubld
.MOV(subscript(tmp_right
, BRW_REGISTER_TYPE_D
, 0),
4640 subscript(src_left
, BRW_REGISTER_TYPE_D
, 0));
4641 ubld
.MOV(subscript(tmp_right
, BRW_REGISTER_TYPE_D
, 1),
4642 subscript(src_left
, BRW_REGISTER_TYPE_D
, 1));
4644 ubld
.MOV(tmp_left
, src_right
);
4645 ubld
.MOV(tmp_right
, src_left
);
4647 bld
.MOV(retype(dest
, value
.type
), tmp
);
4651 case nir_intrinsic_quad_swap_vertical
: {
4652 const fs_reg value
= get_nir_src(instr
->src
[0]);
4653 if (nir_src_bit_size(instr
->src
[0]) == 32) {
4654 /* For 32-bit, we can use a SIMD4x2 instruction to do this easily */
4655 const fs_reg tmp
= bld
.vgrf(value
.type
);
4656 const fs_builder ubld
= bld
.exec_all();
4657 ubld
.emit(SHADER_OPCODE_QUAD_SWIZZLE
, tmp
, value
,
4658 brw_imm_ud(BRW_SWIZZLE4(2,3,0,1)));
4659 bld
.MOV(retype(dest
, value
.type
), tmp
);
4661 /* For larger data types, we have to either emit dispatch_width many
4662 * MOVs or else fall back to doing indirects.
4664 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
4665 bld
.XOR(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
4667 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, idx
);
4672 case nir_intrinsic_quad_swap_diagonal
: {
4673 const fs_reg value
= get_nir_src(instr
->src
[0]);
4674 if (nir_src_bit_size(instr
->src
[0]) == 32) {
4675 /* For 32-bit, we can use a SIMD4x2 instruction to do this easily */
4676 const fs_reg tmp
= bld
.vgrf(value
.type
);
4677 const fs_builder ubld
= bld
.exec_all();
4678 ubld
.emit(SHADER_OPCODE_QUAD_SWIZZLE
, tmp
, value
,
4679 brw_imm_ud(BRW_SWIZZLE4(3,2,1,0)));
4680 bld
.MOV(retype(dest
, value
.type
), tmp
);
4682 /* For larger data types, we have to either emit dispatch_width many
4683 * MOVs or else fall back to doing indirects.
4685 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
4686 bld
.XOR(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
4688 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, idx
);
4693 case nir_intrinsic_reduce
: {
4694 fs_reg src
= get_nir_src(instr
->src
[0]);
4695 nir_op redop
= (nir_op
)nir_intrinsic_reduction_op(instr
);
4696 unsigned cluster_size
= nir_intrinsic_cluster_size(instr
);
4697 if (cluster_size
== 0 || cluster_size
> dispatch_width
)
4698 cluster_size
= dispatch_width
;
4700 /* Figure out the source type */
4701 src
.type
= brw_type_for_nir_type(devinfo
,
4702 (nir_alu_type
)(nir_op_infos
[redop
].input_types
[0] |
4703 nir_src_bit_size(instr
->src
[0])));
4705 fs_reg identity
= brw_nir_reduction_op_identity(bld
, redop
, src
.type
);
4706 opcode brw_op
= brw_op_for_nir_reduction_op(redop
);
4707 brw_conditional_mod cond_mod
= brw_cond_mod_for_nir_reduction_op(redop
);
4709 /* Set up a register for all of our scratching around and initialize it
4710 * to reduction operation's identity value.
4712 fs_reg scan
= bld
.vgrf(src
.type
);
4713 bld
.exec_all().emit(SHADER_OPCODE_SEL_EXEC
, scan
, src
, identity
);
4715 bld
.emit_scan(brw_op
, scan
, cluster_size
, cond_mod
);
4717 dest
.type
= src
.type
;
4718 if (cluster_size
* type_sz(src
.type
) >= REG_SIZE
* 2) {
4719 /* In this case, CLUSTER_BROADCAST instruction isn't needed because
4720 * the distance between clusters is at least 2 GRFs. In this case,
4721 * we don't need the weird striding of the CLUSTER_BROADCAST
4722 * instruction and can just do regular MOVs.
4724 assert((cluster_size
* type_sz(src
.type
)) % (REG_SIZE
* 2) == 0);
4725 const unsigned groups
=
4726 (dispatch_width
* type_sz(src
.type
)) / (REG_SIZE
* 2);
4727 const unsigned group_size
= dispatch_width
/ groups
;
4728 for (unsigned i
= 0; i
< groups
; i
++) {
4729 const unsigned cluster
= (i
* group_size
) / cluster_size
;
4730 const unsigned comp
= cluster
* cluster_size
+ (cluster_size
- 1);
4731 bld
.group(group_size
, i
).MOV(horiz_offset(dest
, i
* group_size
),
4732 component(scan
, comp
));
4735 bld
.emit(SHADER_OPCODE_CLUSTER_BROADCAST
, dest
, scan
,
4736 brw_imm_ud(cluster_size
- 1), brw_imm_ud(cluster_size
));
4741 case nir_intrinsic_inclusive_scan
:
4742 case nir_intrinsic_exclusive_scan
: {
4743 fs_reg src
= get_nir_src(instr
->src
[0]);
4744 nir_op redop
= (nir_op
)nir_intrinsic_reduction_op(instr
);
4746 /* Figure out the source type */
4747 src
.type
= brw_type_for_nir_type(devinfo
,
4748 (nir_alu_type
)(nir_op_infos
[redop
].input_types
[0] |
4749 nir_src_bit_size(instr
->src
[0])));
4751 fs_reg identity
= brw_nir_reduction_op_identity(bld
, redop
, src
.type
);
4752 opcode brw_op
= brw_op_for_nir_reduction_op(redop
);
4753 brw_conditional_mod cond_mod
= brw_cond_mod_for_nir_reduction_op(redop
);
4755 /* Set up a register for all of our scratching around and initialize it
4756 * to reduction operation's identity value.
4758 fs_reg scan
= bld
.vgrf(src
.type
);
4759 const fs_builder allbld
= bld
.exec_all();
4760 allbld
.emit(SHADER_OPCODE_SEL_EXEC
, scan
, src
, identity
);
4762 if (instr
->intrinsic
== nir_intrinsic_exclusive_scan
) {
4763 /* Exclusive scan is a bit harder because we have to do an annoying
4764 * shift of the contents before we can begin. To make things worse,
4765 * we can't do this with a normal stride; we have to use indirects.
4767 fs_reg shifted
= bld
.vgrf(src
.type
);
4768 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
4769 allbld
.ADD(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
4771 allbld
.emit(SHADER_OPCODE_SHUFFLE
, shifted
, scan
, idx
);
4772 allbld
.group(1, 0).MOV(component(shifted
, 0), identity
);
4776 bld
.emit_scan(brw_op
, scan
, dispatch_width
, cond_mod
);
4778 bld
.MOV(retype(dest
, src
.type
), scan
);
4783 unreachable("unknown intrinsic");
4788 fs_visitor::nir_emit_ssbo_atomic(const fs_builder
&bld
,
4789 int op
, nir_intrinsic_instr
*instr
)
4791 if (stage
== MESA_SHADER_FRAGMENT
)
4792 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4795 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4796 dest
= get_nir_dest(instr
->dest
);
4799 nir_const_value
*const_surface
= nir_src_as_const_value(instr
->src
[0]);
4800 if (const_surface
) {
4801 unsigned surf_index
= stage_prog_data
->binding_table
.ssbo_start
+
4802 const_surface
->u32
[0];
4803 surface
= brw_imm_ud(surf_index
);
4804 brw_mark_surface_used(prog_data
, surf_index
);
4806 surface
= vgrf(glsl_type::uint_type
);
4807 bld
.ADD(surface
, get_nir_src(instr
->src
[0]),
4808 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
4810 /* Assume this may touch any SSBO. This is the same we do for other
4811 * UBO/SSBO accesses with non-constant surface.
4813 brw_mark_surface_used(prog_data
,
4814 stage_prog_data
->binding_table
.ssbo_start
+
4815 nir
->info
.num_ssbos
- 1);
4818 fs_reg offset
= get_nir_src(instr
->src
[1]);
4819 fs_reg data1
= get_nir_src(instr
->src
[2]);
4821 if (op
== BRW_AOP_CMPWR
)
4822 data2
= get_nir_src(instr
->src
[3]);
4824 /* Emit the actual atomic operation */
4826 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
4828 1 /* dims */, 1 /* rsize */,
4830 BRW_PREDICATE_NONE
);
4831 dest
.type
= atomic_result
.type
;
4832 bld
.MOV(dest
, atomic_result
);
4836 fs_visitor::nir_emit_shared_atomic(const fs_builder
&bld
,
4837 int op
, nir_intrinsic_instr
*instr
)
4840 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4841 dest
= get_nir_dest(instr
->dest
);
4843 fs_reg surface
= brw_imm_ud(GEN7_BTI_SLM
);
4845 fs_reg data1
= get_nir_src(instr
->src
[1]);
4847 if (op
== BRW_AOP_CMPWR
)
4848 data2
= get_nir_src(instr
->src
[2]);
4850 /* Get the offset */
4851 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
4853 offset
= brw_imm_ud(instr
->const_index
[0] + const_offset
->u32
[0]);
4855 offset
= vgrf(glsl_type::uint_type
);
4857 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
4858 brw_imm_ud(instr
->const_index
[0]));
4861 /* Emit the actual atomic operation operation */
4863 fs_reg atomic_result
= emit_untyped_atomic(bld
, surface
, offset
,
4865 1 /* dims */, 1 /* rsize */,
4867 BRW_PREDICATE_NONE
);
4868 dest
.type
= atomic_result
.type
;
4869 bld
.MOV(dest
, atomic_result
);
4873 fs_visitor::nir_emit_texture(const fs_builder
&bld
, nir_tex_instr
*instr
)
4875 unsigned texture
= instr
->texture_index
;
4876 unsigned sampler
= instr
->sampler_index
;
4878 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
4880 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture
);
4881 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(sampler
);
4883 int lod_components
= 0;
4885 /* The hardware requires a LOD for buffer textures */
4886 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
4887 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_d(0);
4889 uint32_t header_bits
= 0;
4890 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
4891 fs_reg src
= get_nir_src(instr
->src
[i
].src
);
4892 switch (instr
->src
[i
].src_type
) {
4893 case nir_tex_src_bias
:
4894 srcs
[TEX_LOGICAL_SRC_LOD
] =
4895 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
4897 case nir_tex_src_comparator
:
4898 srcs
[TEX_LOGICAL_SRC_SHADOW_C
] = retype(src
, BRW_REGISTER_TYPE_F
);
4900 case nir_tex_src_coord
:
4901 switch (instr
->op
) {
4903 case nir_texop_txf_ms
:
4904 case nir_texop_txf_ms_mcs
:
4905 case nir_texop_samples_identical
:
4906 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_D
);
4909 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_F
);
4913 case nir_tex_src_ddx
:
4914 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
4915 lod_components
= nir_tex_instr_src_size(instr
, i
);
4917 case nir_tex_src_ddy
:
4918 srcs
[TEX_LOGICAL_SRC_LOD2
] = retype(src
, BRW_REGISTER_TYPE_F
);
4920 case nir_tex_src_lod
:
4921 switch (instr
->op
) {
4923 srcs
[TEX_LOGICAL_SRC_LOD
] =
4924 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_UD
);
4927 srcs
[TEX_LOGICAL_SRC_LOD
] =
4928 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_D
);
4931 srcs
[TEX_LOGICAL_SRC_LOD
] =
4932 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
4936 case nir_tex_src_ms_index
:
4937 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = retype(src
, BRW_REGISTER_TYPE_UD
);
4940 case nir_tex_src_offset
: {
4941 nir_const_value
*const_offset
=
4942 nir_src_as_const_value(instr
->src
[i
].src
);
4943 unsigned offset_bits
= 0;
4945 brw_texture_offset(const_offset
->i32
,
4946 nir_tex_instr_src_size(instr
, i
),
4948 header_bits
|= offset_bits
;
4950 srcs
[TEX_LOGICAL_SRC_TG4_OFFSET
] =
4951 retype(src
, BRW_REGISTER_TYPE_D
);
4956 case nir_tex_src_projector
:
4957 unreachable("should be lowered");
4959 case nir_tex_src_texture_offset
: {
4960 /* Figure out the highest possible texture index and mark it as used */
4961 uint32_t max_used
= texture
+ instr
->texture_array_size
- 1;
4962 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
< 8) {
4963 max_used
+= stage_prog_data
->binding_table
.gather_texture_start
;
4965 max_used
+= stage_prog_data
->binding_table
.texture_start
;
4967 brw_mark_surface_used(prog_data
, max_used
);
4969 /* Emit code to evaluate the actual indexing expression */
4970 fs_reg tmp
= vgrf(glsl_type::uint_type
);
4971 bld
.ADD(tmp
, src
, brw_imm_ud(texture
));
4972 srcs
[TEX_LOGICAL_SRC_SURFACE
] = bld
.emit_uniformize(tmp
);
4976 case nir_tex_src_sampler_offset
: {
4977 /* Emit code to evaluate the actual indexing expression */
4978 fs_reg tmp
= vgrf(glsl_type::uint_type
);
4979 bld
.ADD(tmp
, src
, brw_imm_ud(sampler
));
4980 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = bld
.emit_uniformize(tmp
);
4984 case nir_tex_src_ms_mcs
:
4985 assert(instr
->op
== nir_texop_txf_ms
);
4986 srcs
[TEX_LOGICAL_SRC_MCS
] = retype(src
, BRW_REGISTER_TYPE_D
);
4989 case nir_tex_src_plane
: {
4990 nir_const_value
*const_plane
=
4991 nir_src_as_const_value(instr
->src
[i
].src
);
4992 const uint32_t plane
= const_plane
->u32
[0];
4993 const uint32_t texture_index
=
4994 instr
->texture_index
+
4995 stage_prog_data
->binding_table
.plane_start
[plane
] -
4996 stage_prog_data
->binding_table
.texture_start
;
4998 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture_index
);
5003 unreachable("unknown texture source");
5007 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BAD_FILE
&&
5008 (instr
->op
== nir_texop_txf_ms
||
5009 instr
->op
== nir_texop_samples_identical
)) {
5010 if (devinfo
->gen
>= 7 &&
5011 key_tex
->compressed_multisample_layout_mask
& (1 << texture
)) {
5012 srcs
[TEX_LOGICAL_SRC_MCS
] =
5013 emit_mcs_fetch(srcs
[TEX_LOGICAL_SRC_COORDINATE
],
5014 instr
->coord_components
,
5015 srcs
[TEX_LOGICAL_SRC_SURFACE
]);
5017 srcs
[TEX_LOGICAL_SRC_MCS
] = brw_imm_ud(0u);
5021 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(instr
->coord_components
);
5022 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(lod_components
);
5025 switch (instr
->op
) {
5027 opcode
= (stage
== MESA_SHADER_FRAGMENT
? SHADER_OPCODE_TEX_LOGICAL
:
5028 SHADER_OPCODE_TXL_LOGICAL
);
5031 opcode
= FS_OPCODE_TXB_LOGICAL
;
5034 opcode
= SHADER_OPCODE_TXL_LOGICAL
;
5037 opcode
= SHADER_OPCODE_TXD_LOGICAL
;
5040 opcode
= SHADER_OPCODE_TXF_LOGICAL
;
5042 case nir_texop_txf_ms
:
5043 if ((key_tex
->msaa_16
& (1 << sampler
)))
5044 opcode
= SHADER_OPCODE_TXF_CMS_W_LOGICAL
;
5046 opcode
= SHADER_OPCODE_TXF_CMS_LOGICAL
;
5048 case nir_texop_txf_ms_mcs
:
5049 opcode
= SHADER_OPCODE_TXF_MCS_LOGICAL
;
5051 case nir_texop_query_levels
:
5053 opcode
= SHADER_OPCODE_TXS_LOGICAL
;
5056 opcode
= SHADER_OPCODE_LOD_LOGICAL
;
5059 if (srcs
[TEX_LOGICAL_SRC_TG4_OFFSET
].file
!= BAD_FILE
)
5060 opcode
= SHADER_OPCODE_TG4_OFFSET_LOGICAL
;
5062 opcode
= SHADER_OPCODE_TG4_LOGICAL
;
5064 case nir_texop_texture_samples
:
5065 opcode
= SHADER_OPCODE_SAMPLEINFO_LOGICAL
;
5067 case nir_texop_samples_identical
: {
5068 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
5070 /* If mcs is an immediate value, it means there is no MCS. In that case
5071 * just return false.
5073 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BRW_IMMEDIATE_VALUE
) {
5074 bld
.MOV(dst
, brw_imm_ud(0u));
5075 } else if ((key_tex
->msaa_16
& (1 << sampler
))) {
5076 fs_reg tmp
= vgrf(glsl_type::uint_type
);
5077 bld
.OR(tmp
, srcs
[TEX_LOGICAL_SRC_MCS
],
5078 offset(srcs
[TEX_LOGICAL_SRC_MCS
], bld
, 1));
5079 bld
.CMP(dst
, tmp
, brw_imm_ud(0u), BRW_CONDITIONAL_EQ
);
5081 bld
.CMP(dst
, srcs
[TEX_LOGICAL_SRC_MCS
], brw_imm_ud(0u),
5082 BRW_CONDITIONAL_EQ
);
5087 unreachable("unknown texture opcode");
5090 if (instr
->op
== nir_texop_tg4
) {
5091 if (instr
->component
== 1 &&
5092 key_tex
->gather_channel_quirk_mask
& (1 << texture
)) {
5093 /* gather4 sampler is broken for green channel on RG32F --
5094 * we must ask for blue instead.
5096 header_bits
|= 2 << 16;
5098 header_bits
|= instr
->component
<< 16;
5102 fs_reg dst
= bld
.vgrf(brw_type_for_nir_type(devinfo
, instr
->dest_type
), 4);
5103 fs_inst
*inst
= bld
.emit(opcode
, dst
, srcs
, ARRAY_SIZE(srcs
));
5104 inst
->offset
= header_bits
;
5106 const unsigned dest_size
= nir_tex_instr_dest_size(instr
);
5107 if (devinfo
->gen
>= 9 &&
5108 instr
->op
!= nir_texop_tg4
&& instr
->op
!= nir_texop_query_levels
) {
5109 unsigned write_mask
= instr
->dest
.is_ssa
?
5110 nir_ssa_def_components_read(&instr
->dest
.ssa
):
5111 (1 << dest_size
) - 1;
5112 assert(write_mask
!= 0); /* dead code should have been eliminated */
5113 inst
->size_written
= util_last_bit(write_mask
) *
5114 inst
->dst
.component_size(inst
->exec_size
);
5116 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
5119 if (srcs
[TEX_LOGICAL_SRC_SHADOW_C
].file
!= BAD_FILE
)
5120 inst
->shadow_compare
= true;
5122 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
== 6)
5123 emit_gen6_gather_wa(key_tex
->gen6_gather_wa
[texture
], dst
);
5126 for (unsigned i
= 0; i
< dest_size
; i
++)
5127 nir_dest
[i
] = offset(dst
, bld
, i
);
5129 if (instr
->op
== nir_texop_query_levels
) {
5130 /* # levels is in .w */
5131 nir_dest
[0] = offset(dst
, bld
, 3);
5132 } else if (instr
->op
== nir_texop_txs
&&
5133 dest_size
>= 3 && devinfo
->gen
< 7) {
5134 /* Gen4-6 return 0 instead of 1 for single layer surfaces. */
5135 fs_reg depth
= offset(dst
, bld
, 2);
5136 nir_dest
[2] = vgrf(glsl_type::int_type
);
5137 bld
.emit_minmax(nir_dest
[2], depth
, brw_imm_d(1), BRW_CONDITIONAL_GE
);
5140 bld
.LOAD_PAYLOAD(get_nir_dest(instr
->dest
), nir_dest
, dest_size
, 0);
5144 fs_visitor::nir_emit_jump(const fs_builder
&bld
, nir_jump_instr
*instr
)
5146 switch (instr
->type
) {
5147 case nir_jump_break
:
5148 bld
.emit(BRW_OPCODE_BREAK
);
5150 case nir_jump_continue
:
5151 bld
.emit(BRW_OPCODE_CONTINUE
);
5153 case nir_jump_return
:
5155 unreachable("unknown jump");
5160 * This helper takes the result of a load operation that reads 32-bit elements
5168 * and shuffles the data to get this:
5175 * Which is exactly what we want if the load is reading 64-bit components
5176 * like doubles, where x represents the low 32-bit of the x double component
5177 * and y represents the high 32-bit of the x double component (likewise with
5178 * z and w for double component y). The parameter @components represents
5179 * the number of 64-bit components present in @src. This would typically be
5180 * 2 at most, since we can only fit 2 double elements in the result of a
5183 * Notice that @dst and @src can be the same register.
5186 shuffle_32bit_load_result_to_64bit_data(const fs_builder
&bld
,
5189 uint32_t components
)
5191 assert(type_sz(src
.type
) == 4);
5192 assert(type_sz(dst
.type
) == 8);
5194 /* A temporary that we will use to shuffle the 32-bit data of each
5195 * component in the vector into valid 64-bit data. We can't write directly
5196 * to dst because dst can be (and would usually be) the same as src
5197 * and in that case the first MOV in the loop below would overwrite the
5198 * data read in the second MOV.
5200 fs_reg tmp
= bld
.vgrf(dst
.type
);
5202 for (unsigned i
= 0; i
< components
; i
++) {
5203 const fs_reg component_i
= offset(src
, bld
, 2 * i
);
5205 bld
.MOV(subscript(tmp
, src
.type
, 0), component_i
);
5206 bld
.MOV(subscript(tmp
, src
.type
, 1), offset(component_i
, bld
, 1));
5208 bld
.MOV(offset(dst
, bld
, i
), tmp
);
5213 shuffle_32bit_load_result_to_16bit_data(const fs_builder
&bld
,
5216 uint32_t first_component
,
5217 uint32_t components
)
5219 assert(type_sz(src
.type
) == 4);
5220 assert(type_sz(dst
.type
) == 2);
5222 /* A temporary is used to un-shuffle the 32-bit data of each component in
5223 * into a valid 16-bit vector. We can't write directly to dst because it
5224 * can be the same register as src and in that case the first MOV in the
5225 * loop below would overwrite the data read in the second MOV.
5227 fs_reg tmp
= retype(bld
.vgrf(src
.type
), dst
.type
);
5229 for (unsigned i
= 0; i
< components
; i
++) {
5230 const fs_reg component_i
=
5231 subscript(offset(src
, bld
, (first_component
+ i
) / 2), dst
.type
,
5232 (first_component
+ i
) % 2);
5234 bld
.MOV(offset(tmp
, bld
, i
% 2), component_i
);
5237 bld
.MOV(offset(dst
, bld
, i
-1), offset(tmp
, bld
, 0));
5238 bld
.MOV(offset(dst
, bld
, i
), offset(tmp
, bld
, 1));
5241 if (components
% 2) {
5242 bld
.MOV(offset(dst
, bld
, components
- 1), tmp
);
5247 * This helper does the inverse operation of
5248 * SHUFFLE_32BIT_LOAD_RESULT_TO_64BIT_DATA.
5250 * We need to do this when we are going to use untyped write messsages that
5251 * operate with 32-bit components in order to arrange our 64-bit data to be
5252 * in the expected layout.
5254 * Notice that callers of this function, unlike in the case of the inverse
5255 * operation, would typically need to call this with dst and src being
5256 * different registers, since they would otherwise corrupt the original
5257 * 64-bit data they are about to write. Because of this the function checks
5258 * that the src and dst regions involved in the operation do not overlap.
5261 shuffle_64bit_data_for_32bit_write(const fs_builder
&bld
,
5263 uint32_t components
)
5265 assert(type_sz(src
.type
) == 8);
5267 fs_reg dst
= bld
.vgrf(BRW_REGISTER_TYPE_D
, 2 * components
);
5269 for (unsigned i
= 0; i
< components
; i
++) {
5270 const fs_reg component_i
= offset(src
, bld
, i
);
5271 bld
.MOV(offset(dst
, bld
, 2 * i
), subscript(component_i
, dst
.type
, 0));
5272 bld
.MOV(offset(dst
, bld
, 2 * i
+ 1), subscript(component_i
, dst
.type
, 1));
5279 shuffle_16bit_data_for_32bit_write(const fs_builder
&bld
,
5282 uint32_t components
)
5284 assert(type_sz(src
.type
) == 2);
5285 assert(type_sz(dst
.type
) == 4);
5287 /* A temporary is used to shuffle the 16-bit data of each component in the
5288 * 32-bit data vector. We can't write directly to dst because it can be the
5289 * same register as src and in that case the first MOV in the loop below
5290 * would overwrite the data read in the second MOV.
5292 fs_reg tmp
= bld
.vgrf(dst
.type
);
5294 for (unsigned i
= 0; i
< components
; i
++) {
5295 const fs_reg component_i
= offset(src
, bld
, i
);
5296 bld
.MOV(subscript(tmp
, src
.type
, i
% 2), component_i
);
5298 bld
.MOV(offset(dst
, bld
, i
/ 2), tmp
);
5301 if (components
% 2) {
5302 bld
.MOV(offset(dst
, bld
, components
/ 2), tmp
);
5307 setup_imm_df(const fs_builder
&bld
, double v
)
5309 const struct gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5310 assert(devinfo
->gen
>= 7);
5312 if (devinfo
->gen
>= 8)
5313 return brw_imm_df(v
);
5315 /* gen7.5 does not support DF immediates straighforward but the DIM
5316 * instruction allows to set the 64-bit immediate value.
5318 if (devinfo
->is_haswell
) {
5319 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5320 fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_DF
, 1);
5321 ubld
.DIM(dst
, brw_imm_df(v
));
5322 return component(dst
, 0);
5325 /* gen7 does not support DF immediates, so we generate a 64-bit constant by
5326 * writing the low 32-bit of the constant to suboffset 0 of a VGRF and
5327 * the high 32-bit to suboffset 4 and then applying a stride of 0.
5329 * Alternatively, we could also produce a normal VGRF (without stride 0)
5330 * by writing to all the channels in the VGRF, however, that would hit the
5331 * gen7 bug where we have to split writes that span more than 1 register
5332 * into instructions with a width of 4 (otherwise the write to the second
5333 * register written runs into an execmask hardware bug) which isn't very
5346 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5347 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
5348 ubld
.MOV(tmp
, brw_imm_ud(di
.i1
));
5349 ubld
.MOV(horiz_offset(tmp
, 1), brw_imm_ud(di
.i2
));
5351 return component(retype(tmp
, BRW_REGISTER_TYPE_DF
), 0);