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"
27 #include "util/u_math.h"
28 #include "util/bitscan.h"
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 nir_emit_impl(nir_shader_get_entrypoint((nir_shader
*)nir
));
46 fs_visitor::nir_setup_outputs()
48 if (stage
== MESA_SHADER_TESS_CTRL
|| stage
== MESA_SHADER_FRAGMENT
)
51 unsigned vec4s
[VARYING_SLOT_TESS_MAX
] = { 0, };
53 /* Calculate the size of output registers in a separate pass, before
54 * allocating them. With ARB_enhanced_layouts, multiple output variables
55 * may occupy the same slot, but have different type sizes.
57 nir_foreach_variable(var
, &nir
->outputs
) {
58 const int loc
= var
->data
.driver_location
;
59 const unsigned var_vec4s
=
60 var
->data
.compact
? DIV_ROUND_UP(glsl_get_length(var
->type
), 4)
61 : type_size_vec4(var
->type
, true);
62 vec4s
[loc
] = MAX2(vec4s
[loc
], var_vec4s
);
65 for (unsigned loc
= 0; loc
< ARRAY_SIZE(vec4s
);) {
66 if (vec4s
[loc
] == 0) {
71 unsigned reg_size
= vec4s
[loc
];
73 /* Check if there are any ranges that start within this range and extend
74 * past it. If so, include them in this allocation.
76 for (unsigned i
= 1; i
< reg_size
; i
++)
77 reg_size
= MAX2(vec4s
[i
+ loc
] + i
, reg_size
);
79 fs_reg reg
= bld
.vgrf(BRW_REGISTER_TYPE_F
, 4 * reg_size
);
80 for (unsigned i
= 0; i
< reg_size
; i
++)
81 outputs
[loc
+ i
] = offset(reg
, bld
, 4 * i
);
88 fs_visitor::nir_setup_uniforms()
90 /* Only the first compile gets to set up uniforms. */
91 if (push_constant_loc
) {
92 assert(pull_constant_loc
);
96 uniforms
= nir
->num_uniforms
/ 4;
98 if (stage
== MESA_SHADER_COMPUTE
) {
99 /* Add a uniform for the thread local id. It must be the last uniform
102 assert(uniforms
== prog_data
->nr_params
);
103 uint32_t *param
= brw_stage_prog_data_add_params(prog_data
, 1);
104 *param
= BRW_PARAM_BUILTIN_SUBGROUP_ID
;
105 subgroup_id
= fs_reg(UNIFORM
, uniforms
++, BRW_REGISTER_TYPE_UD
);
110 emit_system_values_block(nir_block
*block
, fs_visitor
*v
)
114 nir_foreach_instr(instr
, block
) {
115 if (instr
->type
!= nir_instr_type_intrinsic
)
118 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
119 switch (intrin
->intrinsic
) {
120 case nir_intrinsic_load_vertex_id
:
121 case nir_intrinsic_load_base_vertex
:
122 unreachable("should be lowered by nir_lower_system_values().");
124 case nir_intrinsic_load_vertex_id_zero_base
:
125 case nir_intrinsic_load_is_indexed_draw
:
126 case nir_intrinsic_load_first_vertex
:
127 case nir_intrinsic_load_instance_id
:
128 case nir_intrinsic_load_base_instance
:
129 case nir_intrinsic_load_draw_id
:
130 unreachable("should be lowered by brw_nir_lower_vs_inputs().");
132 case nir_intrinsic_load_invocation_id
:
133 if (v
->stage
== MESA_SHADER_TESS_CTRL
)
135 assert(v
->stage
== MESA_SHADER_GEOMETRY
);
136 reg
= &v
->nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
137 if (reg
->file
== BAD_FILE
) {
138 const fs_builder abld
= v
->bld
.annotate("gl_InvocationID", NULL
);
139 fs_reg
g1(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
140 fs_reg iid
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
141 abld
.SHR(iid
, g1
, brw_imm_ud(27u));
146 case nir_intrinsic_load_sample_pos
:
147 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
148 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
149 if (reg
->file
== BAD_FILE
)
150 *reg
= *v
->emit_samplepos_setup();
153 case nir_intrinsic_load_sample_id
:
154 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
155 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
156 if (reg
->file
== BAD_FILE
)
157 *reg
= *v
->emit_sampleid_setup();
160 case nir_intrinsic_load_sample_mask_in
:
161 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
162 assert(v
->devinfo
->gen
>= 7);
163 reg
= &v
->nir_system_values
[SYSTEM_VALUE_SAMPLE_MASK_IN
];
164 if (reg
->file
== BAD_FILE
)
165 *reg
= *v
->emit_samplemaskin_setup();
168 case nir_intrinsic_load_work_group_id
:
169 assert(v
->stage
== MESA_SHADER_COMPUTE
);
170 reg
= &v
->nir_system_values
[SYSTEM_VALUE_WORK_GROUP_ID
];
171 if (reg
->file
== BAD_FILE
)
172 *reg
= *v
->emit_cs_work_group_id_setup();
175 case nir_intrinsic_load_helper_invocation
:
176 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
177 reg
= &v
->nir_system_values
[SYSTEM_VALUE_HELPER_INVOCATION
];
178 if (reg
->file
== BAD_FILE
) {
179 const fs_builder abld
=
180 v
->bld
.annotate("gl_HelperInvocation", NULL
);
182 /* On Gen6+ (gl_HelperInvocation is only exposed on Gen7+) the
183 * pixel mask is in g1.7 of the thread payload.
185 * We move the per-channel pixel enable bit to the low bit of each
186 * channel by shifting the byte containing the pixel mask by the
187 * vector immediate 0x76543210UV.
189 * The region of <1,8,0> reads only 1 byte (the pixel masks for
190 * subspans 0 and 1) in SIMD8 and an additional byte (the pixel
191 * masks for 2 and 3) in SIMD16.
193 fs_reg shifted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
195 for (unsigned i
= 0; i
< DIV_ROUND_UP(v
->dispatch_width
, 16); i
++) {
196 const fs_builder hbld
= abld
.group(MIN2(16, v
->dispatch_width
), i
);
197 hbld
.SHR(offset(shifted
, hbld
, i
),
198 stride(retype(brw_vec1_grf(1 + i
, 7),
199 BRW_REGISTER_TYPE_UB
),
201 brw_imm_v(0x76543210));
204 /* A set bit in the pixel mask means the channel is enabled, but
205 * that is the opposite of gl_HelperInvocation so we need to invert
208 * The negate source-modifier bit of logical instructions on Gen8+
209 * performs 1's complement negation, so we can use that instead of
212 fs_reg inverted
= negate(shifted
);
213 if (v
->devinfo
->gen
< 8) {
214 inverted
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
215 abld
.NOT(inverted
, shifted
);
218 /* We then resolve the 0/1 result to 0/~0 boolean values by ANDing
219 * with 1 and negating.
221 fs_reg anded
= abld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
222 abld
.AND(anded
, inverted
, brw_imm_uw(1));
224 fs_reg dst
= abld
.vgrf(BRW_REGISTER_TYPE_D
, 1);
225 abld
.MOV(dst
, negate(retype(anded
, BRW_REGISTER_TYPE_D
)));
239 fs_visitor::nir_emit_system_values()
241 nir_system_values
= ralloc_array(mem_ctx
, fs_reg
, SYSTEM_VALUE_MAX
);
242 for (unsigned i
= 0; i
< SYSTEM_VALUE_MAX
; i
++) {
243 nir_system_values
[i
] = fs_reg();
246 /* Always emit SUBGROUP_INVOCATION. Dead code will clean it up if we
247 * never end up using it.
250 const fs_builder abld
= bld
.annotate("gl_SubgroupInvocation", NULL
);
251 fs_reg
®
= nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
];
252 reg
= abld
.vgrf(BRW_REGISTER_TYPE_UW
);
254 const fs_builder allbld8
= abld
.group(8, 0).exec_all();
255 allbld8
.MOV(reg
, brw_imm_v(0x76543210));
256 if (dispatch_width
> 8)
257 allbld8
.ADD(byte_offset(reg
, 16), reg
, brw_imm_uw(8u));
258 if (dispatch_width
> 16) {
259 const fs_builder allbld16
= abld
.group(16, 0).exec_all();
260 allbld16
.ADD(byte_offset(reg
, 32), reg
, brw_imm_uw(16u));
264 nir_function_impl
*impl
= nir_shader_get_entrypoint((nir_shader
*)nir
);
265 nir_foreach_block(block
, impl
)
266 emit_system_values_block(block
, this);
270 * Returns a type based on a reference_type (word, float, half-float) and a
273 * Reference BRW_REGISTER_TYPE are HF,F,DF,W,D,UW,UD.
275 * @FIXME: 64-bit return types are always DF on integer types to maintain
276 * compability with uses of DF previously to the introduction of int64
280 brw_reg_type_from_bit_size(const unsigned bit_size
,
281 const brw_reg_type reference_type
)
283 switch(reference_type
) {
284 case BRW_REGISTER_TYPE_HF
:
285 case BRW_REGISTER_TYPE_F
:
286 case BRW_REGISTER_TYPE_DF
:
289 return BRW_REGISTER_TYPE_HF
;
291 return BRW_REGISTER_TYPE_F
;
293 return BRW_REGISTER_TYPE_DF
;
295 unreachable("Invalid bit size");
297 case BRW_REGISTER_TYPE_B
:
298 case BRW_REGISTER_TYPE_W
:
299 case BRW_REGISTER_TYPE_D
:
300 case BRW_REGISTER_TYPE_Q
:
303 return BRW_REGISTER_TYPE_B
;
305 return BRW_REGISTER_TYPE_W
;
307 return BRW_REGISTER_TYPE_D
;
309 return BRW_REGISTER_TYPE_Q
;
311 unreachable("Invalid bit size");
313 case BRW_REGISTER_TYPE_UB
:
314 case BRW_REGISTER_TYPE_UW
:
315 case BRW_REGISTER_TYPE_UD
:
316 case BRW_REGISTER_TYPE_UQ
:
319 return BRW_REGISTER_TYPE_UB
;
321 return BRW_REGISTER_TYPE_UW
;
323 return BRW_REGISTER_TYPE_UD
;
325 return BRW_REGISTER_TYPE_UQ
;
327 unreachable("Invalid bit size");
330 unreachable("Unknown type");
335 fs_visitor::nir_emit_impl(nir_function_impl
*impl
)
337 nir_locals
= ralloc_array(mem_ctx
, fs_reg
, impl
->reg_alloc
);
338 for (unsigned i
= 0; i
< impl
->reg_alloc
; i
++) {
339 nir_locals
[i
] = fs_reg();
342 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
343 unsigned array_elems
=
344 reg
->num_array_elems
== 0 ? 1 : reg
->num_array_elems
;
345 unsigned size
= array_elems
* reg
->num_components
;
346 const brw_reg_type reg_type
=
347 brw_reg_type_from_bit_size(reg
->bit_size
, BRW_REGISTER_TYPE_F
);
348 nir_locals
[reg
->index
] = bld
.vgrf(reg_type
, size
);
351 nir_ssa_values
= reralloc(mem_ctx
, nir_ssa_values
, fs_reg
,
354 nir_emit_cf_list(&impl
->body
);
358 fs_visitor::nir_emit_cf_list(exec_list
*list
)
360 exec_list_validate(list
);
361 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
362 switch (node
->type
) {
364 nir_emit_if(nir_cf_node_as_if(node
));
367 case nir_cf_node_loop
:
368 nir_emit_loop(nir_cf_node_as_loop(node
));
371 case nir_cf_node_block
:
372 nir_emit_block(nir_cf_node_as_block(node
));
376 unreachable("Invalid CFG node block");
382 fs_visitor::nir_emit_if(nir_if
*if_stmt
)
387 /* If the condition has the form !other_condition, use other_condition as
388 * the source, but invert the predicate on the if instruction.
390 nir_alu_instr
*const cond
= nir_src_as_alu_instr(&if_stmt
->condition
);
391 if (cond
!= NULL
&& cond
->op
== nir_op_inot
) {
392 assert(!cond
->src
[0].negate
);
393 assert(!cond
->src
[0].abs
);
396 cond_reg
= get_nir_src(cond
->src
[0].src
);
399 cond_reg
= get_nir_src(if_stmt
->condition
);
402 /* first, put the condition into f0 */
403 fs_inst
*inst
= bld
.MOV(bld
.null_reg_d(),
404 retype(cond_reg
, BRW_REGISTER_TYPE_D
));
405 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
407 bld
.IF(BRW_PREDICATE_NORMAL
)->predicate_inverse
= invert
;
409 nir_emit_cf_list(&if_stmt
->then_list
);
411 /* note: if the else is empty, dead CF elimination will remove it */
412 bld
.emit(BRW_OPCODE_ELSE
);
414 nir_emit_cf_list(&if_stmt
->else_list
);
416 bld
.emit(BRW_OPCODE_ENDIF
);
418 if (devinfo
->gen
< 7)
419 limit_dispatch_width(16, "Non-uniform control flow unsupported "
424 fs_visitor::nir_emit_loop(nir_loop
*loop
)
426 bld
.emit(BRW_OPCODE_DO
);
428 nir_emit_cf_list(&loop
->body
);
430 bld
.emit(BRW_OPCODE_WHILE
);
432 if (devinfo
->gen
< 7)
433 limit_dispatch_width(16, "Non-uniform control flow unsupported "
438 fs_visitor::nir_emit_block(nir_block
*block
)
440 nir_foreach_instr(instr
, block
) {
441 nir_emit_instr(instr
);
446 fs_visitor::nir_emit_instr(nir_instr
*instr
)
448 const fs_builder abld
= bld
.annotate(NULL
, instr
);
450 switch (instr
->type
) {
451 case nir_instr_type_alu
:
452 nir_emit_alu(abld
, nir_instr_as_alu(instr
));
455 case nir_instr_type_deref
:
456 /* Derefs can exist for images but they do nothing */
459 case nir_instr_type_intrinsic
:
461 case MESA_SHADER_VERTEX
:
462 nir_emit_vs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
464 case MESA_SHADER_TESS_CTRL
:
465 nir_emit_tcs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
467 case MESA_SHADER_TESS_EVAL
:
468 nir_emit_tes_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
470 case MESA_SHADER_GEOMETRY
:
471 nir_emit_gs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
473 case MESA_SHADER_FRAGMENT
:
474 nir_emit_fs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
476 case MESA_SHADER_COMPUTE
:
477 nir_emit_cs_intrinsic(abld
, nir_instr_as_intrinsic(instr
));
480 unreachable("unsupported shader stage");
484 case nir_instr_type_tex
:
485 nir_emit_texture(abld
, nir_instr_as_tex(instr
));
488 case nir_instr_type_load_const
:
489 nir_emit_load_const(abld
, nir_instr_as_load_const(instr
));
492 case nir_instr_type_ssa_undef
:
493 /* We create a new VGRF for undefs on every use (by handling
494 * them in get_nir_src()), rather than for each definition.
495 * This helps register coalescing eliminate MOVs from undef.
499 case nir_instr_type_jump
:
500 nir_emit_jump(abld
, nir_instr_as_jump(instr
));
504 unreachable("unknown instruction type");
509 * Recognizes a parent instruction of nir_op_extract_* and changes the type to
513 fs_visitor::optimize_extract_to_float(nir_alu_instr
*instr
,
514 const fs_reg
&result
)
516 if (!instr
->src
[0].src
.is_ssa
||
517 !instr
->src
[0].src
.ssa
->parent_instr
)
520 if (instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
523 nir_alu_instr
*src0
=
524 nir_instr_as_alu(instr
->src
[0].src
.ssa
->parent_instr
);
526 if (src0
->op
!= nir_op_extract_u8
&& src0
->op
!= nir_op_extract_u16
&&
527 src0
->op
!= nir_op_extract_i8
&& src0
->op
!= nir_op_extract_i16
)
530 /* If either opcode has source modifiers, bail.
532 * TODO: We can potentially handle source modifiers if both of the opcodes
533 * we're combining are signed integers.
535 if (instr
->src
[0].abs
|| instr
->src
[0].negate
||
536 src0
->src
[0].abs
|| src0
->src
[0].negate
)
539 unsigned element
= nir_src_as_uint(src0
->src
[1].src
);
541 /* Element type to extract.*/
542 const brw_reg_type type
= brw_int_type(
543 src0
->op
== nir_op_extract_u16
|| src0
->op
== nir_op_extract_i16
? 2 : 1,
544 src0
->op
== nir_op_extract_i16
|| src0
->op
== nir_op_extract_i8
);
546 fs_reg op0
= get_nir_src(src0
->src
[0].src
);
547 op0
.type
= brw_type_for_nir_type(devinfo
,
548 (nir_alu_type
)(nir_op_infos
[src0
->op
].input_types
[0] |
549 nir_src_bit_size(src0
->src
[0].src
)));
550 op0
= offset(op0
, bld
, src0
->src
[0].swizzle
[0]);
552 set_saturate(instr
->dest
.saturate
,
553 bld
.MOV(result
, subscript(op0
, type
, element
)));
558 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr
*instr
,
559 const fs_reg
&result
)
561 if (!instr
->src
[0].src
.is_ssa
||
562 instr
->src
[0].src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
565 nir_intrinsic_instr
*src0
=
566 nir_instr_as_intrinsic(instr
->src
[0].src
.ssa
->parent_instr
);
568 if (src0
->intrinsic
!= nir_intrinsic_load_front_face
)
571 if (!nir_src_is_const(instr
->src
[1].src
) ||
572 !nir_src_is_const(instr
->src
[2].src
))
575 const float value1
= nir_src_as_float(instr
->src
[1].src
);
576 const float value2
= nir_src_as_float(instr
->src
[2].src
);
577 if (fabsf(value1
) != 1.0f
|| fabsf(value2
) != 1.0f
)
580 /* nir_opt_algebraic should have gotten rid of bcsel(b, a, a) */
581 assert(value1
== -value2
);
583 fs_reg tmp
= vgrf(glsl_type::int_type
);
585 if (devinfo
->gen
>= 6) {
586 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
587 fs_reg g0
= fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W
));
589 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
591 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
592 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
594 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
596 * This negation looks like it's safe in practice, because bits 0:4 will
597 * surely be TRIANGLES
600 if (value1
== -1.0f
) {
604 bld
.OR(subscript(tmp
, BRW_REGISTER_TYPE_W
, 1),
605 g0
, brw_imm_uw(0x3f80));
607 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
608 fs_reg g1_6
= fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D
));
610 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
612 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
613 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
615 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
617 * This negation looks like it's safe in practice, because bits 0:4 will
618 * surely be TRIANGLES
621 if (value1
== -1.0f
) {
625 bld
.OR(tmp
, g1_6
, brw_imm_d(0x3f800000));
627 bld
.AND(retype(result
, BRW_REGISTER_TYPE_D
), tmp
, brw_imm_d(0xbf800000));
633 emit_find_msb_using_lzd(const fs_builder
&bld
,
634 const fs_reg
&result
,
642 /* LZD of an absolute value source almost always does the right
643 * thing. There are two problem values:
645 * * 0x80000000. Since abs(0x80000000) == 0x80000000, LZD returns
646 * 0. However, findMSB(int(0x80000000)) == 30.
648 * * 0xffffffff. Since abs(0xffffffff) == 1, LZD returns
649 * 31. Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
651 * For a value of zero or negative one, -1 will be returned.
653 * * Negative powers of two. LZD(abs(-(1<<x))) returns x, but
654 * findMSB(-(1<<x)) should return x-1.
656 * For all negative number cases, including 0x80000000 and
657 * 0xffffffff, the correct value is obtained from LZD if instead of
658 * negating the (already negative) value the logical-not is used. A
659 * conditonal logical-not can be achieved in two instructions.
661 temp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
663 bld
.ASR(temp
, src
, brw_imm_d(31));
664 bld
.XOR(temp
, temp
, src
);
667 bld
.LZD(retype(result
, BRW_REGISTER_TYPE_UD
),
668 retype(temp
, BRW_REGISTER_TYPE_UD
));
670 /* LZD counts from the MSB side, while GLSL's findMSB() wants the count
671 * from the LSB side. Subtract the result from 31 to convert the MSB
672 * count into an LSB count. If no bits are set, LZD will return 32.
673 * 31-32 = -1, which is exactly what findMSB() is supposed to return.
675 inst
= bld
.ADD(result
, retype(result
, BRW_REGISTER_TYPE_D
), brw_imm_d(31));
676 inst
->src
[0].negate
= true;
680 brw_rnd_mode_from_nir_op (const nir_op op
) {
682 case nir_op_f2f16_rtz
:
683 return BRW_RND_MODE_RTZ
;
684 case nir_op_f2f16_rtne
:
685 return BRW_RND_MODE_RTNE
;
687 unreachable("Operation doesn't support rounding mode");
692 fs_visitor::prepare_alu_destination_and_sources(const fs_builder
&bld
,
693 nir_alu_instr
*instr
,
698 need_dest
? get_nir_dest(instr
->dest
.dest
) : bld
.null_reg_ud();
700 result
.type
= brw_type_for_nir_type(devinfo
,
701 (nir_alu_type
)(nir_op_infos
[instr
->op
].output_type
|
702 nir_dest_bit_size(instr
->dest
.dest
)));
704 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
705 op
[i
] = get_nir_src(instr
->src
[i
].src
);
706 op
[i
].type
= brw_type_for_nir_type(devinfo
,
707 (nir_alu_type
)(nir_op_infos
[instr
->op
].input_types
[i
] |
708 nir_src_bit_size(instr
->src
[i
].src
)));
709 op
[i
].abs
= instr
->src
[i
].abs
;
710 op
[i
].negate
= instr
->src
[i
].negate
;
713 /* Move and vecN instrutions may still be vectored. Return the raw,
714 * vectored source and destination so that fs_visitor::nir_emit_alu can
715 * handle it. Other callers should not have to handle these kinds of
729 /* At this point, we have dealt with any instruction that operates on
730 * more than a single channel. Therefore, we can just adjust the source
731 * and destination registers for that channel and emit the instruction.
733 unsigned channel
= 0;
734 if (nir_op_infos
[instr
->op
].output_size
== 0) {
735 /* Since NIR is doing the scalarizing for us, we should only ever see
736 * vectorized operations with a single channel.
738 assert(util_bitcount(instr
->dest
.write_mask
) == 1);
739 channel
= ffs(instr
->dest
.write_mask
) - 1;
741 result
= offset(result
, bld
, channel
);
744 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
745 assert(nir_op_infos
[instr
->op
].input_sizes
[i
] < 2);
746 op
[i
] = offset(op
[i
], bld
, instr
->src
[i
].swizzle
[channel
]);
753 fs_visitor::resolve_inot_sources(const fs_builder
&bld
, nir_alu_instr
*instr
,
756 for (unsigned i
= 0; i
< 2; i
++) {
757 nir_alu_instr
*const inot_instr
=
758 nir_src_as_alu_instr(&instr
->src
[i
].src
);
760 if (inot_instr
!= NULL
&& inot_instr
->op
== nir_op_inot
&&
761 !inot_instr
->src
[0].abs
&& !inot_instr
->src
[0].negate
) {
762 /* The source of the inot is now the source of instr. */
763 prepare_alu_destination_and_sources(bld
, inot_instr
, &op
[i
], false);
765 assert(!op
[i
].negate
);
768 op
[i
] = resolve_source_modifiers(op
[i
]);
774 fs_visitor::try_emit_b2fi_of_inot(const fs_builder
&bld
,
776 nir_alu_instr
*instr
)
778 if (devinfo
->gen
< 6 || devinfo
->gen
>= 12)
781 nir_alu_instr
*const inot_instr
= nir_src_as_alu_instr(&instr
->src
[0].src
);
783 if (inot_instr
== NULL
|| inot_instr
->op
!= nir_op_inot
)
786 /* HF is also possible as a destination on BDW+. For nir_op_b2i, the set
787 * of valid size-changing combinations is a bit more complex.
789 * The source restriction is just because I was lazy about generating the
792 if (nir_dest_bit_size(instr
->dest
.dest
) != 32 ||
793 nir_src_bit_size(inot_instr
->src
[0].src
) != 32)
796 /* b2[fi](inot(a)) maps a=0 => 1, a=-1 => 0. Since a can only be 0 or -1,
797 * this is float(1 + a).
801 prepare_alu_destination_and_sources(bld
, inot_instr
, &op
, false);
803 /* Ignore the saturate modifier, if there is one. The result of the
804 * arithmetic can only be 0 or 1, so the clamping will do nothing anyway.
806 bld
.ADD(result
, op
, brw_imm_d(1));
812 fs_visitor::nir_emit_alu(const fs_builder
&bld
, nir_alu_instr
*instr
)
814 struct brw_wm_prog_key
*fs_key
= (struct brw_wm_prog_key
*) this->key
;
818 fs_reg result
= prepare_alu_destination_and_sources(bld
, instr
, op
, true);
826 fs_reg temp
= result
;
827 bool need_extra_copy
= false;
828 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
829 if (!instr
->src
[i
].src
.is_ssa
&&
830 instr
->dest
.dest
.reg
.reg
== instr
->src
[i
].src
.reg
.reg
) {
831 need_extra_copy
= true;
832 temp
= bld
.vgrf(result
.type
, 4);
837 for (unsigned i
= 0; i
< 4; i
++) {
838 if (!(instr
->dest
.write_mask
& (1 << i
)))
841 if (instr
->op
== nir_op_imov
|| instr
->op
== nir_op_fmov
) {
842 inst
= bld
.MOV(offset(temp
, bld
, i
),
843 offset(op
[0], bld
, instr
->src
[0].swizzle
[i
]));
845 inst
= bld
.MOV(offset(temp
, bld
, i
),
846 offset(op
[i
], bld
, instr
->src
[i
].swizzle
[0]));
848 inst
->saturate
= instr
->dest
.saturate
;
851 /* In this case the source and destination registers were the same,
852 * so we need to insert an extra set of moves in order to deal with
855 if (need_extra_copy
) {
856 for (unsigned i
= 0; i
< 4; i
++) {
857 if (!(instr
->dest
.write_mask
& (1 << i
)))
860 bld
.MOV(offset(result
, bld
, i
), offset(temp
, bld
, i
));
868 if (optimize_extract_to_float(instr
, result
))
870 inst
= bld
.MOV(result
, op
[0]);
871 inst
->saturate
= instr
->dest
.saturate
;
874 case nir_op_f2f16_rtne
:
875 case nir_op_f2f16_rtz
:
876 bld
.emit(SHADER_OPCODE_RND_MODE
, bld
.null_reg_ud(),
877 brw_imm_d(brw_rnd_mode_from_nir_op(instr
->op
)));
880 /* In theory, it would be better to use BRW_OPCODE_F32TO16. Depending
881 * on the HW gen, it is a special hw opcode or just a MOV, and
882 * brw_F32TO16 (at brw_eu_emit) would do the work to chose.
884 * But if we want to use that opcode, we need to provide support on
885 * different optimizations and lowerings. As right now HF support is
886 * only for gen8+, it will be better to use directly the MOV, and use
887 * BRW_OPCODE_F32TO16 when/if we work for HF support on gen7.
893 assert(type_sz(op
[0].type
) < 8); /* brw_nir_lower_conversions */
894 inst
= bld
.MOV(result
, op
[0]);
895 inst
->saturate
= instr
->dest
.saturate
;
901 assert(type_sz(op
[0].type
) > 2); /* brw_nir_lower_conversions */
902 inst
= bld
.MOV(result
, op
[0]);
903 inst
->saturate
= instr
->dest
.saturate
;
913 if (try_emit_b2fi_of_inot(bld
, result
, instr
))
915 op
[0].type
= BRW_REGISTER_TYPE_D
;
916 op
[0].negate
= !op
[0].negate
;
922 assert(type_sz(op
[0].type
) > 1); /* brw_nir_lower_conversions */
935 inst
= bld
.MOV(result
, op
[0]);
936 inst
->saturate
= instr
->dest
.saturate
;
940 assert(!instr
->dest
.saturate
);
942 /* Straightforward since the source can be assumed to be either
943 * strictly >= 0 or strictly <= 0 depending on the setting of the
946 set_condmod(BRW_CONDITIONAL_NZ
, bld
.MOV(result
, op
[0]));
948 inst
= (op
[0].negate
)
949 ? bld
.MOV(result
, brw_imm_f(-1.0f
))
950 : bld
.MOV(result
, brw_imm_f(1.0f
));
952 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
953 } else if (type_sz(op
[0].type
) == 2) {
954 /* AND(val, 0x8000) gives the sign bit.
956 * Predicated OR ORs 1.0 (0x3c00) with the sign bit if val is not zero.
958 fs_reg zero
= retype(brw_imm_uw(0), BRW_REGISTER_TYPE_HF
);
959 bld
.CMP(bld
.null_reg_f(), op
[0], zero
, BRW_CONDITIONAL_NZ
);
961 op
[0].type
= BRW_REGISTER_TYPE_UW
;
962 result
.type
= BRW_REGISTER_TYPE_UW
;
963 bld
.AND(result
, op
[0], brw_imm_uw(0x8000u
));
965 inst
= bld
.OR(result
, result
, brw_imm_uw(0x3c00u
));
966 inst
->predicate
= BRW_PREDICATE_NORMAL
;
967 } else if (type_sz(op
[0].type
) == 4) {
968 /* AND(val, 0x80000000) gives the sign bit.
970 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
973 bld
.CMP(bld
.null_reg_f(), op
[0], brw_imm_f(0.0f
), BRW_CONDITIONAL_NZ
);
975 op
[0].type
= BRW_REGISTER_TYPE_UD
;
976 result
.type
= BRW_REGISTER_TYPE_UD
;
977 bld
.AND(result
, op
[0], brw_imm_ud(0x80000000u
));
979 inst
= bld
.OR(result
, result
, brw_imm_ud(0x3f800000u
));
980 inst
->predicate
= BRW_PREDICATE_NORMAL
;
982 /* For doubles we do the same but we need to consider:
984 * - 2-src instructions can't operate with 64-bit immediates
985 * - The sign is encoded in the high 32-bit of each DF
986 * - We need to produce a DF result.
988 assert(type_sz(op
[0].type
) == 8);
990 fs_reg zero
= vgrf(glsl_type::double_type
);
991 bld
.MOV(zero
, setup_imm_df(bld
, 0.0));
992 bld
.CMP(bld
.null_reg_df(), op
[0], zero
, BRW_CONDITIONAL_NZ
);
994 bld
.MOV(result
, zero
);
996 fs_reg r
= subscript(result
, BRW_REGISTER_TYPE_UD
, 1);
997 bld
.AND(r
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1),
998 brw_imm_ud(0x80000000u
));
1000 set_predicate(BRW_PREDICATE_NORMAL
,
1001 bld
.OR(r
, r
, brw_imm_ud(0x3ff00000u
)));
1007 inst
= bld
.emit(SHADER_OPCODE_RCP
, result
, op
[0]);
1008 inst
->saturate
= instr
->dest
.saturate
;
1012 inst
= bld
.emit(SHADER_OPCODE_EXP2
, result
, op
[0]);
1013 inst
->saturate
= instr
->dest
.saturate
;
1017 inst
= bld
.emit(SHADER_OPCODE_LOG2
, result
, op
[0]);
1018 inst
->saturate
= instr
->dest
.saturate
;
1022 inst
= bld
.emit(SHADER_OPCODE_SIN
, result
, op
[0]);
1023 inst
->saturate
= instr
->dest
.saturate
;
1027 inst
= bld
.emit(SHADER_OPCODE_COS
, result
, op
[0]);
1028 inst
->saturate
= instr
->dest
.saturate
;
1032 if (fs_key
->high_quality_derivatives
) {
1033 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
1035 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
1037 inst
->saturate
= instr
->dest
.saturate
;
1039 case nir_op_fddx_fine
:
1040 inst
= bld
.emit(FS_OPCODE_DDX_FINE
, result
, op
[0]);
1041 inst
->saturate
= instr
->dest
.saturate
;
1043 case nir_op_fddx_coarse
:
1044 inst
= bld
.emit(FS_OPCODE_DDX_COARSE
, result
, op
[0]);
1045 inst
->saturate
= instr
->dest
.saturate
;
1048 if (fs_key
->high_quality_derivatives
) {
1049 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
1051 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
1053 inst
->saturate
= instr
->dest
.saturate
;
1055 case nir_op_fddy_fine
:
1056 inst
= bld
.emit(FS_OPCODE_DDY_FINE
, result
, op
[0]);
1057 inst
->saturate
= instr
->dest
.saturate
;
1059 case nir_op_fddy_coarse
:
1060 inst
= bld
.emit(FS_OPCODE_DDY_COARSE
, result
, op
[0]);
1061 inst
->saturate
= instr
->dest
.saturate
;
1066 inst
= bld
.ADD(result
, op
[0], op
[1]);
1067 inst
->saturate
= instr
->dest
.saturate
;
1070 case nir_op_uadd_sat
:
1071 inst
= bld
.ADD(result
, op
[0], op
[1]);
1072 inst
->saturate
= true;
1076 inst
= bld
.MUL(result
, op
[0], op
[1]);
1077 inst
->saturate
= instr
->dest
.saturate
;
1080 case nir_op_imul_2x32_64
:
1081 case nir_op_umul_2x32_64
:
1082 bld
.MUL(result
, op
[0], op
[1]);
1086 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1087 bld
.MUL(result
, op
[0], op
[1]);
1090 case nir_op_imul_high
:
1091 case nir_op_umul_high
:
1092 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1093 bld
.emit(SHADER_OPCODE_MULH
, result
, op
[0], op
[1]);
1098 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1099 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
, result
, op
[0], op
[1]);
1102 case nir_op_uadd_carry
:
1103 unreachable("Should have been lowered by carry_to_arith().");
1105 case nir_op_usub_borrow
:
1106 unreachable("Should have been lowered by borrow_to_arith().");
1110 /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
1111 * appears that our hardware just does the right thing for signed
1114 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1115 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
1119 /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
1120 bld
.emit(SHADER_OPCODE_INT_REMAINDER
, result
, op
[0], op
[1]);
1122 /* Math instructions don't support conditional mod */
1123 inst
= bld
.MOV(bld
.null_reg_d(), result
);
1124 inst
->conditional_mod
= BRW_CONDITIONAL_NZ
;
1126 /* Now, we need to determine if signs of the sources are different.
1127 * When we XOR the sources, the top bit is 0 if they are the same and 1
1128 * if they are different. We can then use a conditional modifier to
1129 * turn that into a predicate. This leads us to an XOR.l instruction.
1131 * Technically, according to the PRM, you're not allowed to use .l on a
1132 * XOR instruction. However, emperical experiments and Curro's reading
1133 * of the simulator source both indicate that it's safe.
1135 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1136 inst
= bld
.XOR(tmp
, op
[0], op
[1]);
1137 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1138 inst
->conditional_mod
= BRW_CONDITIONAL_L
;
1140 /* If the result of the initial remainder operation is non-zero and the
1141 * two sources have different signs, add in a copy of op[1] to get the
1142 * final integer modulus value.
1144 inst
= bld
.ADD(result
, result
, op
[1]);
1145 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1152 case nir_op_fne32
: {
1153 fs_reg dest
= result
;
1155 const uint32_t bit_size
= nir_src_bit_size(instr
->src
[0].src
);
1157 dest
= bld
.vgrf(op
[0].type
, 1);
1159 brw_conditional_mod cond
;
1160 switch (instr
->op
) {
1162 cond
= BRW_CONDITIONAL_L
;
1165 cond
= BRW_CONDITIONAL_GE
;
1168 cond
= BRW_CONDITIONAL_Z
;
1171 cond
= BRW_CONDITIONAL_NZ
;
1174 unreachable("bad opcode");
1177 bld
.CMP(dest
, op
[0], op
[1], cond
);
1179 if (bit_size
> 32) {
1180 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1181 } else if(bit_size
< 32) {
1182 /* When we convert the result to 32-bit we need to be careful and do
1183 * it as a signed conversion to get sign extension (for 32-bit true)
1185 const brw_reg_type src_type
=
1186 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_D
);
1188 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), retype(dest
, src_type
));
1198 case nir_op_ine32
: {
1199 fs_reg dest
= result
;
1201 const uint32_t bit_size
= nir_src_bit_size(instr
->src
[0].src
);
1203 dest
= bld
.vgrf(op
[0].type
, 1);
1205 brw_conditional_mod cond
;
1206 switch (instr
->op
) {
1209 cond
= BRW_CONDITIONAL_L
;
1213 cond
= BRW_CONDITIONAL_GE
;
1216 cond
= BRW_CONDITIONAL_Z
;
1219 cond
= BRW_CONDITIONAL_NZ
;
1222 unreachable("bad opcode");
1224 bld
.CMP(dest
, op
[0], op
[1], cond
);
1226 if (bit_size
> 32) {
1227 bld
.MOV(result
, subscript(dest
, BRW_REGISTER_TYPE_UD
, 0));
1228 } else if (bit_size
< 32) {
1229 /* When we convert the result to 32-bit we need to be careful and do
1230 * it as a signed conversion to get sign extension (for 32-bit true)
1232 const brw_reg_type src_type
=
1233 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_D
);
1235 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), retype(dest
, src_type
));
1241 if (devinfo
->gen
>= 8) {
1242 nir_alu_instr
*const inot_src_instr
=
1243 nir_src_as_alu_instr(&instr
->src
[0].src
);
1245 if (inot_src_instr
!= NULL
&&
1246 (inot_src_instr
->op
== nir_op_ior
||
1247 inot_src_instr
->op
== nir_op_ixor
||
1248 inot_src_instr
->op
== nir_op_iand
) &&
1249 !inot_src_instr
->src
[0].abs
&&
1250 !inot_src_instr
->src
[0].negate
&&
1251 !inot_src_instr
->src
[1].abs
&&
1252 !inot_src_instr
->src
[1].negate
) {
1253 /* The sources of the source logical instruction are now the
1254 * sources of the instruction that will be generated.
1256 prepare_alu_destination_and_sources(bld
, inot_src_instr
, op
, false);
1257 resolve_inot_sources(bld
, inot_src_instr
, op
);
1259 /* Smash all of the sources and destination to be signed. This
1260 * doesn't matter for the operation of the instruction, but cmod
1261 * propagation fails on unsigned sources with negation (due to
1262 * fs_inst::can_do_cmod returning false).
1265 brw_type_for_nir_type(devinfo
,
1266 (nir_alu_type
)(nir_type_int
|
1267 nir_dest_bit_size(instr
->dest
.dest
)));
1269 brw_type_for_nir_type(devinfo
,
1270 (nir_alu_type
)(nir_type_int
|
1271 nir_src_bit_size(inot_src_instr
->src
[0].src
)));
1273 brw_type_for_nir_type(devinfo
,
1274 (nir_alu_type
)(nir_type_int
|
1275 nir_src_bit_size(inot_src_instr
->src
[1].src
)));
1277 /* For XOR, only invert one of the sources. Arbitrarily choose
1280 op
[0].negate
= !op
[0].negate
;
1281 if (inot_src_instr
->op
!= nir_op_ixor
)
1282 op
[1].negate
= !op
[1].negate
;
1284 switch (inot_src_instr
->op
) {
1286 bld
.AND(result
, op
[0], op
[1]);
1290 bld
.OR(result
, op
[0], op
[1]);
1294 bld
.XOR(result
, op
[0], op
[1]);
1298 unreachable("impossible opcode");
1301 op
[0] = resolve_source_modifiers(op
[0]);
1303 bld
.NOT(result
, op
[0]);
1306 if (devinfo
->gen
>= 8) {
1307 resolve_inot_sources(bld
, instr
, op
);
1309 bld
.XOR(result
, op
[0], op
[1]);
1312 if (devinfo
->gen
>= 8) {
1313 resolve_inot_sources(bld
, instr
, op
);
1315 bld
.OR(result
, op
[0], op
[1]);
1318 if (devinfo
->gen
>= 8) {
1319 resolve_inot_sources(bld
, instr
, op
);
1321 bld
.AND(result
, op
[0], op
[1]);
1327 case nir_op_b32all_fequal2
:
1328 case nir_op_b32all_iequal2
:
1329 case nir_op_b32all_fequal3
:
1330 case nir_op_b32all_iequal3
:
1331 case nir_op_b32all_fequal4
:
1332 case nir_op_b32all_iequal4
:
1333 case nir_op_b32any_fnequal2
:
1334 case nir_op_b32any_inequal2
:
1335 case nir_op_b32any_fnequal3
:
1336 case nir_op_b32any_inequal3
:
1337 case nir_op_b32any_fnequal4
:
1338 case nir_op_b32any_inequal4
:
1339 unreachable("Lowered by nir_lower_alu_reductions");
1341 case nir_op_fnoise1_1
:
1342 case nir_op_fnoise1_2
:
1343 case nir_op_fnoise1_3
:
1344 case nir_op_fnoise1_4
:
1345 case nir_op_fnoise2_1
:
1346 case nir_op_fnoise2_2
:
1347 case nir_op_fnoise2_3
:
1348 case nir_op_fnoise2_4
:
1349 case nir_op_fnoise3_1
:
1350 case nir_op_fnoise3_2
:
1351 case nir_op_fnoise3_3
:
1352 case nir_op_fnoise3_4
:
1353 case nir_op_fnoise4_1
:
1354 case nir_op_fnoise4_2
:
1355 case nir_op_fnoise4_3
:
1356 case nir_op_fnoise4_4
:
1357 unreachable("not reached: should be handled by lower_noise");
1360 unreachable("not reached: should be handled by ldexp_to_arith()");
1363 inst
= bld
.emit(SHADER_OPCODE_SQRT
, result
, op
[0]);
1364 inst
->saturate
= instr
->dest
.saturate
;
1368 inst
= bld
.emit(SHADER_OPCODE_RSQ
, result
, op
[0]);
1369 inst
->saturate
= instr
->dest
.saturate
;
1373 case nir_op_f2b32
: {
1374 uint32_t bit_size
= nir_src_bit_size(instr
->src
[0].src
);
1375 if (bit_size
== 64) {
1376 /* two-argument instructions can't take 64-bit immediates */
1380 if (instr
->op
== nir_op_f2b32
) {
1381 zero
= vgrf(glsl_type::double_type
);
1382 tmp
= vgrf(glsl_type::double_type
);
1383 bld
.MOV(zero
, setup_imm_df(bld
, 0.0));
1385 zero
= vgrf(glsl_type::int64_t_type
);
1386 tmp
= vgrf(glsl_type::int64_t_type
);
1387 bld
.MOV(zero
, brw_imm_q(0));
1390 /* A SIMD16 execution needs to be split in two instructions, so use
1391 * a vgrf instead of the flag register as dst so instruction splitting
1394 bld
.CMP(tmp
, op
[0], zero
, BRW_CONDITIONAL_NZ
);
1395 bld
.MOV(result
, subscript(tmp
, BRW_REGISTER_TYPE_UD
, 0));
1398 if (bit_size
== 32) {
1399 zero
= instr
->op
== nir_op_f2b32
? brw_imm_f(0.0f
) : brw_imm_d(0);
1401 assert(bit_size
== 16);
1402 zero
= instr
->op
== nir_op_f2b32
?
1403 retype(brw_imm_w(0), BRW_REGISTER_TYPE_HF
) : brw_imm_w(0);
1405 bld
.CMP(result
, op
[0], zero
, BRW_CONDITIONAL_NZ
);
1411 inst
= bld
.RNDZ(result
, op
[0]);
1412 inst
->saturate
= instr
->dest
.saturate
;
1415 case nir_op_fceil
: {
1416 op
[0].negate
= !op
[0].negate
;
1417 fs_reg temp
= vgrf(glsl_type::float_type
);
1418 bld
.RNDD(temp
, op
[0]);
1420 inst
= bld
.MOV(result
, temp
);
1421 inst
->saturate
= instr
->dest
.saturate
;
1425 inst
= bld
.RNDD(result
, op
[0]);
1426 inst
->saturate
= instr
->dest
.saturate
;
1429 inst
= bld
.FRC(result
, op
[0]);
1430 inst
->saturate
= instr
->dest
.saturate
;
1432 case nir_op_fround_even
:
1433 inst
= bld
.RNDE(result
, op
[0]);
1434 inst
->saturate
= instr
->dest
.saturate
;
1437 case nir_op_fquantize2f16
: {
1438 fs_reg tmp16
= bld
.vgrf(BRW_REGISTER_TYPE_D
);
1439 fs_reg tmp32
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1440 fs_reg zero
= bld
.vgrf(BRW_REGISTER_TYPE_F
);
1442 /* The destination stride must be at least as big as the source stride. */
1443 tmp16
.type
= BRW_REGISTER_TYPE_W
;
1446 /* Check for denormal */
1447 fs_reg abs_src0
= op
[0];
1448 abs_src0
.abs
= true;
1449 bld
.CMP(bld
.null_reg_f(), abs_src0
, brw_imm_f(ldexpf(1.0, -14)),
1451 /* Get the appropriately signed zero */
1452 bld
.AND(retype(zero
, BRW_REGISTER_TYPE_UD
),
1453 retype(op
[0], BRW_REGISTER_TYPE_UD
),
1454 brw_imm_ud(0x80000000));
1455 /* Do the actual F32 -> F16 -> F32 conversion */
1456 bld
.emit(BRW_OPCODE_F32TO16
, tmp16
, op
[0]);
1457 bld
.emit(BRW_OPCODE_F16TO32
, tmp32
, tmp16
);
1458 /* Select that or zero based on normal status */
1459 inst
= bld
.SEL(result
, zero
, tmp32
);
1460 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1461 inst
->saturate
= instr
->dest
.saturate
;
1468 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_L
);
1469 inst
->saturate
= instr
->dest
.saturate
;
1475 inst
= bld
.emit_minmax(result
, op
[0], op
[1], BRW_CONDITIONAL_GE
);
1476 inst
->saturate
= instr
->dest
.saturate
;
1479 case nir_op_pack_snorm_2x16
:
1480 case nir_op_pack_snorm_4x8
:
1481 case nir_op_pack_unorm_2x16
:
1482 case nir_op_pack_unorm_4x8
:
1483 case nir_op_unpack_snorm_2x16
:
1484 case nir_op_unpack_snorm_4x8
:
1485 case nir_op_unpack_unorm_2x16
:
1486 case nir_op_unpack_unorm_4x8
:
1487 case nir_op_unpack_half_2x16
:
1488 case nir_op_pack_half_2x16
:
1489 unreachable("not reached: should be handled by lower_packing_builtins");
1491 case nir_op_unpack_half_2x16_split_x
:
1492 inst
= bld
.emit(BRW_OPCODE_F16TO32
, result
,
1493 subscript(op
[0], BRW_REGISTER_TYPE_UW
, 0));
1494 inst
->saturate
= instr
->dest
.saturate
;
1496 case nir_op_unpack_half_2x16_split_y
:
1497 inst
= bld
.emit(BRW_OPCODE_F16TO32
, result
,
1498 subscript(op
[0], BRW_REGISTER_TYPE_UW
, 1));
1499 inst
->saturate
= instr
->dest
.saturate
;
1502 case nir_op_pack_64_2x32_split
:
1503 case nir_op_pack_32_2x16_split
:
1504 bld
.emit(FS_OPCODE_PACK
, result
, op
[0], op
[1]);
1507 case nir_op_unpack_64_2x32_split_x
:
1508 case nir_op_unpack_64_2x32_split_y
: {
1509 if (instr
->op
== nir_op_unpack_64_2x32_split_x
)
1510 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 0));
1512 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UD
, 1));
1516 case nir_op_unpack_32_2x16_split_x
:
1517 case nir_op_unpack_32_2x16_split_y
: {
1518 if (instr
->op
== nir_op_unpack_32_2x16_split_x
)
1519 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UW
, 0));
1521 bld
.MOV(result
, subscript(op
[0], BRW_REGISTER_TYPE_UW
, 1));
1526 inst
= bld
.emit(SHADER_OPCODE_POW
, result
, op
[0], op
[1]);
1527 inst
->saturate
= instr
->dest
.saturate
;
1530 case nir_op_bitfield_reverse
:
1531 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1532 bld
.BFREV(result
, op
[0]);
1535 case nir_op_bit_count
:
1536 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1537 bld
.CBIT(result
, op
[0]);
1540 case nir_op_ufind_msb
: {
1541 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1542 emit_find_msb_using_lzd(bld
, result
, op
[0], false);
1546 case nir_op_ifind_msb
: {
1547 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1549 if (devinfo
->gen
< 7) {
1550 emit_find_msb_using_lzd(bld
, result
, op
[0], true);
1552 bld
.FBH(retype(result
, BRW_REGISTER_TYPE_UD
), op
[0]);
1554 /* FBH counts from the MSB side, while GLSL's findMSB() wants the
1555 * count from the LSB side. If FBH didn't return an error
1556 * (0xFFFFFFFF), then subtract the result from 31 to convert the MSB
1557 * count into an LSB count.
1559 bld
.CMP(bld
.null_reg_d(), result
, brw_imm_d(-1), BRW_CONDITIONAL_NZ
);
1561 inst
= bld
.ADD(result
, result
, brw_imm_d(31));
1562 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1563 inst
->src
[0].negate
= true;
1568 case nir_op_find_lsb
:
1569 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1571 if (devinfo
->gen
< 7) {
1572 fs_reg temp
= vgrf(glsl_type::int_type
);
1574 /* (x & -x) generates a value that consists of only the LSB of x.
1575 * For all powers of 2, findMSB(y) == findLSB(y).
1577 fs_reg src
= retype(op
[0], BRW_REGISTER_TYPE_D
);
1578 fs_reg negated_src
= src
;
1580 /* One must be negated, and the other must be non-negated. It
1581 * doesn't matter which is which.
1583 negated_src
.negate
= true;
1586 bld
.AND(temp
, src
, negated_src
);
1587 emit_find_msb_using_lzd(bld
, result
, temp
, false);
1589 bld
.FBL(result
, op
[0]);
1593 case nir_op_ubitfield_extract
:
1594 case nir_op_ibitfield_extract
:
1595 unreachable("should have been lowered");
1598 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1599 bld
.BFE(result
, op
[2], op
[1], op
[0]);
1602 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1603 bld
.BFI1(result
, op
[0], op
[1]);
1606 assert(nir_dest_bit_size(instr
->dest
.dest
) < 64);
1607 bld
.BFI2(result
, op
[0], op
[1], op
[2]);
1610 case nir_op_bitfield_insert
:
1611 unreachable("not reached: should have been lowered");
1614 bld
.SHL(result
, op
[0], op
[1]);
1617 bld
.ASR(result
, op
[0], op
[1]);
1620 bld
.SHR(result
, op
[0], op
[1]);
1623 case nir_op_pack_half_2x16_split
:
1624 bld
.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT
, result
, op
[0], op
[1]);
1628 inst
= bld
.MAD(result
, op
[2], op
[1], op
[0]);
1629 inst
->saturate
= instr
->dest
.saturate
;
1633 inst
= bld
.LRP(result
, op
[0], op
[1], op
[2]);
1634 inst
->saturate
= instr
->dest
.saturate
;
1637 case nir_op_b32csel
:
1638 if (optimize_frontfacing_ternary(instr
, result
))
1641 bld
.CMP(bld
.null_reg_d(), op
[0], brw_imm_d(0), BRW_CONDITIONAL_NZ
);
1642 inst
= bld
.SEL(result
, op
[1], op
[2]);
1643 inst
->predicate
= BRW_PREDICATE_NORMAL
;
1646 case nir_op_extract_u8
:
1647 case nir_op_extract_i8
: {
1648 unsigned byte
= nir_src_as_uint(instr
->src
[1].src
);
1653 * There is no direct conversion from B/UB to Q/UQ or Q/UQ to B/UB.
1654 * Use two instructions and a word or DWord intermediate integer type.
1656 if (nir_dest_bit_size(instr
->dest
.dest
) == 64) {
1657 const brw_reg_type type
= brw_int_type(1, instr
->op
== nir_op_extract_i8
);
1659 if (instr
->op
== nir_op_extract_i8
) {
1660 /* If we need to sign extend, extract to a word first */
1661 fs_reg w_temp
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
1662 bld
.MOV(w_temp
, subscript(op
[0], type
, byte
));
1663 bld
.MOV(result
, w_temp
);
1664 } else if (byte
& 1) {
1665 /* Extract the high byte from the word containing the desired byte
1669 subscript(op
[0], BRW_REGISTER_TYPE_UW
, byte
/ 2),
1672 /* Otherwise use an AND with 0xff and a word type */
1674 subscript(op
[0], BRW_REGISTER_TYPE_UW
, byte
/ 2),
1678 const brw_reg_type type
= brw_int_type(1, instr
->op
== nir_op_extract_i8
);
1679 bld
.MOV(result
, subscript(op
[0], type
, byte
));
1684 case nir_op_extract_u16
:
1685 case nir_op_extract_i16
: {
1686 const brw_reg_type type
= brw_int_type(2, instr
->op
== nir_op_extract_i16
);
1687 unsigned word
= nir_src_as_uint(instr
->src
[1].src
);
1688 bld
.MOV(result
, subscript(op
[0], type
, word
));
1693 unreachable("unhandled instruction");
1696 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1697 * to sign extend the low bit to 0/~0
1699 if (devinfo
->gen
<= 5 &&
1700 (instr
->instr
.pass_flags
& BRW_NIR_BOOLEAN_MASK
) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE
) {
1701 fs_reg masked
= vgrf(glsl_type::int_type
);
1702 bld
.AND(masked
, result
, brw_imm_d(1));
1703 masked
.negate
= true;
1704 bld
.MOV(retype(result
, BRW_REGISTER_TYPE_D
), masked
);
1709 fs_visitor::nir_emit_load_const(const fs_builder
&bld
,
1710 nir_load_const_instr
*instr
)
1712 const brw_reg_type reg_type
=
1713 brw_reg_type_from_bit_size(instr
->def
.bit_size
, BRW_REGISTER_TYPE_D
);
1714 fs_reg reg
= bld
.vgrf(reg_type
, instr
->def
.num_components
);
1716 switch (instr
->def
.bit_size
) {
1718 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1719 bld
.MOV(offset(reg
, bld
, i
), setup_imm_b(bld
, instr
->value
[i
].i8
));
1723 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1724 bld
.MOV(offset(reg
, bld
, i
), brw_imm_w(instr
->value
[i
].i16
));
1728 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1729 bld
.MOV(offset(reg
, bld
, i
), brw_imm_d(instr
->value
[i
].i32
));
1733 assert(devinfo
->gen
>= 7);
1734 if (devinfo
->gen
== 7) {
1735 /* We don't get 64-bit integer types until gen8 */
1736 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++) {
1737 bld
.MOV(retype(offset(reg
, bld
, i
), BRW_REGISTER_TYPE_DF
),
1738 setup_imm_df(bld
, instr
->value
[i
].f64
));
1741 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++)
1742 bld
.MOV(offset(reg
, bld
, i
), brw_imm_q(instr
->value
[i
].i64
));
1747 unreachable("Invalid bit size");
1750 nir_ssa_values
[instr
->def
.index
] = reg
;
1754 fs_visitor::get_nir_src(const nir_src
&src
)
1758 if (src
.ssa
->parent_instr
->type
== nir_instr_type_ssa_undef
) {
1759 const brw_reg_type reg_type
=
1760 brw_reg_type_from_bit_size(src
.ssa
->bit_size
, BRW_REGISTER_TYPE_D
);
1761 reg
= bld
.vgrf(reg_type
, src
.ssa
->num_components
);
1763 reg
= nir_ssa_values
[src
.ssa
->index
];
1766 /* We don't handle indirects on locals */
1767 assert(src
.reg
.indirect
== NULL
);
1768 reg
= offset(nir_locals
[src
.reg
.reg
->index
], bld
,
1769 src
.reg
.base_offset
* src
.reg
.reg
->num_components
);
1772 if (nir_src_bit_size(src
) == 64 && devinfo
->gen
== 7) {
1773 /* The only 64-bit type available on gen7 is DF, so use that. */
1774 reg
.type
= BRW_REGISTER_TYPE_DF
;
1776 /* To avoid floating-point denorm flushing problems, set the type by
1777 * default to an integer type - instructions that need floating point
1778 * semantics will set this to F if they need to
1780 reg
.type
= brw_reg_type_from_bit_size(nir_src_bit_size(src
),
1781 BRW_REGISTER_TYPE_D
);
1788 * Return an IMM for constants; otherwise call get_nir_src() as normal.
1790 * This function should not be called on any value which may be 64 bits.
1791 * We could theoretically support 64-bit on gen8+ but we choose not to
1792 * because it wouldn't work in general (no gen7 support) and there are
1793 * enough restrictions in 64-bit immediates that you can't take the return
1794 * value and treat it the same as the result of get_nir_src().
1797 fs_visitor::get_nir_src_imm(const nir_src
&src
)
1799 assert(nir_src_bit_size(src
) == 32);
1800 return nir_src_is_const(src
) ?
1801 fs_reg(brw_imm_d(nir_src_as_int(src
))) : get_nir_src(src
);
1805 fs_visitor::get_nir_dest(const nir_dest
&dest
)
1808 const brw_reg_type reg_type
=
1809 brw_reg_type_from_bit_size(dest
.ssa
.bit_size
,
1810 dest
.ssa
.bit_size
== 8 ?
1811 BRW_REGISTER_TYPE_D
:
1812 BRW_REGISTER_TYPE_F
);
1813 nir_ssa_values
[dest
.ssa
.index
] =
1814 bld
.vgrf(reg_type
, dest
.ssa
.num_components
);
1815 return nir_ssa_values
[dest
.ssa
.index
];
1817 /* We don't handle indirects on locals */
1818 assert(dest
.reg
.indirect
== NULL
);
1819 return offset(nir_locals
[dest
.reg
.reg
->index
], bld
,
1820 dest
.reg
.base_offset
* dest
.reg
.reg
->num_components
);
1825 fs_visitor::emit_percomp(const fs_builder
&bld
, const fs_inst
&inst
,
1828 for (unsigned i
= 0; i
< 4; i
++) {
1829 if (!((wr_mask
>> i
) & 1))
1832 fs_inst
*new_inst
= new(mem_ctx
) fs_inst(inst
);
1833 new_inst
->dst
= offset(new_inst
->dst
, bld
, i
);
1834 for (unsigned j
= 0; j
< new_inst
->sources
; j
++)
1835 if (new_inst
->src
[j
].file
== VGRF
)
1836 new_inst
->src
[j
] = offset(new_inst
->src
[j
], bld
, i
);
1843 emit_pixel_interpolater_send(const fs_builder
&bld
,
1848 glsl_interp_mode interpolation
)
1850 struct brw_wm_prog_data
*wm_prog_data
=
1851 brw_wm_prog_data(bld
.shader
->stage_prog_data
);
1853 fs_inst
*inst
= bld
.emit(opcode
, dst
, src
, desc
);
1854 /* 2 floats per slot returned */
1855 inst
->size_written
= 2 * dst
.component_size(inst
->exec_size
);
1856 inst
->pi_noperspective
= interpolation
== INTERP_MODE_NOPERSPECTIVE
;
1858 wm_prog_data
->pulls_bary
= true;
1864 * Computes 1 << x, given a D/UD register containing some value x.
1867 intexp2(const fs_builder
&bld
, const fs_reg
&x
)
1869 assert(x
.type
== BRW_REGISTER_TYPE_UD
|| x
.type
== BRW_REGISTER_TYPE_D
);
1871 fs_reg result
= bld
.vgrf(x
.type
, 1);
1872 fs_reg one
= bld
.vgrf(x
.type
, 1);
1874 bld
.MOV(one
, retype(brw_imm_d(1), one
.type
));
1875 bld
.SHL(result
, one
, x
);
1880 fs_visitor::emit_gs_end_primitive(const nir_src
&vertex_count_nir_src
)
1882 assert(stage
== MESA_SHADER_GEOMETRY
);
1884 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1886 if (gs_compile
->control_data_header_size_bits
== 0)
1889 /* We can only do EndPrimitive() functionality when the control data
1890 * consists of cut bits. Fortunately, the only time it isn't is when the
1891 * output type is points, in which case EndPrimitive() is a no-op.
1893 if (gs_prog_data
->control_data_format
!=
1894 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT
) {
1898 /* Cut bits use one bit per vertex. */
1899 assert(gs_compile
->control_data_bits_per_vertex
== 1);
1901 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
1902 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
1904 /* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
1905 * vertex n, 0 otherwise. So all we need to do here is mark bit
1906 * (vertex_count - 1) % 32 in the cut_bits register to indicate that
1907 * EndPrimitive() was called after emitting vertex (vertex_count - 1);
1908 * vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
1910 * Note that if EndPrimitive() is called before emitting any vertices, this
1911 * will cause us to set bit 31 of the control_data_bits register to 1.
1912 * That's fine because:
1914 * - If max_vertices < 32, then vertex number 31 (zero-based) will never be
1915 * output, so the hardware will ignore cut bit 31.
1917 * - If max_vertices == 32, then vertex number 31 is guaranteed to be the
1918 * last vertex, so setting cut bit 31 has no effect (since the primitive
1919 * is automatically ended when the GS terminates).
1921 * - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
1922 * control_data_bits register to 0 when the first vertex is emitted.
1925 const fs_builder abld
= bld
.annotate("end primitive");
1927 /* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
1928 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1929 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
1930 fs_reg mask
= intexp2(abld
, prev_count
);
1931 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1932 * attention to the lower 5 bits of its second source argument, so on this
1933 * architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
1934 * ((vertex_count - 1) % 32).
1936 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
1940 fs_visitor::emit_gs_control_data_bits(const fs_reg
&vertex_count
)
1942 assert(stage
== MESA_SHADER_GEOMETRY
);
1943 assert(gs_compile
->control_data_bits_per_vertex
!= 0);
1945 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
1947 const fs_builder abld
= bld
.annotate("emit control data bits");
1948 const fs_builder fwa_bld
= bld
.exec_all();
1950 /* We use a single UD register to accumulate control data bits (32 bits
1951 * for each of the SIMD8 channels). So we need to write a DWord (32 bits)
1954 * Unfortunately, the URB_WRITE_SIMD8 message uses 128-bit (OWord) offsets.
1955 * We have select a 128-bit group via the Global and Per-Slot Offsets, then
1956 * use the Channel Mask phase to enable/disable which DWord within that
1957 * group to write. (Remember, different SIMD8 channels may have emitted
1958 * different numbers of vertices, so we may need per-slot offsets.)
1960 * Channel masking presents an annoying problem: we may have to replicate
1961 * the data up to 4 times:
1963 * Msg = Handles, Per-Slot Offsets, Channel Masks, Data, Data, Data, Data.
1965 * To avoid penalizing shaders that emit a small number of vertices, we
1966 * can avoid these sometimes: if the size of the control data header is
1967 * <= 128 bits, then there is only 1 OWord. All SIMD8 channels will land
1968 * land in the same 128-bit group, so we can skip per-slot offsets.
1970 * Similarly, if the control data header is <= 32 bits, there is only one
1971 * DWord, so we can skip channel masks.
1973 enum opcode opcode
= SHADER_OPCODE_URB_WRITE_SIMD8
;
1975 fs_reg channel_mask
, per_slot_offset
;
1977 if (gs_compile
->control_data_header_size_bits
> 32) {
1978 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
1979 channel_mask
= vgrf(glsl_type::uint_type
);
1982 if (gs_compile
->control_data_header_size_bits
> 128) {
1983 opcode
= SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
;
1984 per_slot_offset
= vgrf(glsl_type::uint_type
);
1987 /* Figure out which DWord we're trying to write to using the formula:
1989 * dword_index = (vertex_count - 1) * bits_per_vertex / 32
1991 * Since bits_per_vertex is a power of two, and is known at compile
1992 * time, this can be optimized to:
1994 * dword_index = (vertex_count - 1) >> (6 - log2(bits_per_vertex))
1996 if (opcode
!= SHADER_OPCODE_URB_WRITE_SIMD8
) {
1997 fs_reg dword_index
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1998 fs_reg prev_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
1999 abld
.ADD(prev_count
, vertex_count
, brw_imm_ud(0xffffffffu
));
2000 unsigned log2_bits_per_vertex
=
2001 util_last_bit(gs_compile
->control_data_bits_per_vertex
);
2002 abld
.SHR(dword_index
, prev_count
, brw_imm_ud(6u - log2_bits_per_vertex
));
2004 if (per_slot_offset
.file
!= BAD_FILE
) {
2005 /* Set the per-slot offset to dword_index / 4, so that we'll write to
2006 * the appropriate OWord within the control data header.
2008 abld
.SHR(per_slot_offset
, dword_index
, brw_imm_ud(2u));
2011 /* Set the channel masks to 1 << (dword_index % 4), so that we'll
2012 * write to the appropriate DWORD within the OWORD.
2014 fs_reg channel
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2015 fwa_bld
.AND(channel
, dword_index
, brw_imm_ud(3u));
2016 channel_mask
= intexp2(fwa_bld
, channel
);
2017 /* Then the channel masks need to be in bits 23:16. */
2018 fwa_bld
.SHL(channel_mask
, channel_mask
, brw_imm_ud(16u));
2021 /* Store the control data bits in the message payload and send it. */
2023 if (channel_mask
.file
!= BAD_FILE
)
2024 mlen
+= 4; /* channel masks, plus 3 extra copies of the data */
2025 if (per_slot_offset
.file
!= BAD_FILE
)
2028 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2029 fs_reg
*sources
= ralloc_array(mem_ctx
, fs_reg
, mlen
);
2031 sources
[i
++] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
));
2032 if (per_slot_offset
.file
!= BAD_FILE
)
2033 sources
[i
++] = per_slot_offset
;
2034 if (channel_mask
.file
!= BAD_FILE
)
2035 sources
[i
++] = channel_mask
;
2037 sources
[i
++] = this->control_data_bits
;
2040 abld
.LOAD_PAYLOAD(payload
, sources
, mlen
, mlen
);
2041 fs_inst
*inst
= abld
.emit(opcode
, reg_undef
, payload
);
2043 /* We need to increment Global Offset by 256-bits to make room for
2044 * Broadwell's extra "Vertex Count" payload at the beginning of the
2045 * URB entry. Since this is an OWord message, Global Offset is counted
2046 * in 128-bit units, so we must set it to 2.
2048 if (gs_prog_data
->static_vertex_count
== -1)
2053 fs_visitor::set_gs_stream_control_data_bits(const fs_reg
&vertex_count
,
2056 /* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
2058 /* Note: we are calling this *before* increasing vertex_count, so
2059 * this->vertex_count == vertex_count - 1 in the formula above.
2062 /* Stream mode uses 2 bits per vertex */
2063 assert(gs_compile
->control_data_bits_per_vertex
== 2);
2065 /* Must be a valid stream */
2066 assert(stream_id
< MAX_VERTEX_STREAMS
);
2068 /* Control data bits are initialized to 0 so we don't have to set any
2069 * bits when sending vertices to stream 0.
2074 const fs_builder abld
= bld
.annotate("set stream control data bits", NULL
);
2076 /* reg::sid = stream_id */
2077 fs_reg sid
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2078 abld
.MOV(sid
, brw_imm_ud(stream_id
));
2080 /* reg:shift_count = 2 * (vertex_count - 1) */
2081 fs_reg shift_count
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2082 abld
.SHL(shift_count
, vertex_count
, brw_imm_ud(1u));
2084 /* Note: we're relying on the fact that the GEN SHL instruction only pays
2085 * attention to the lower 5 bits of its second source argument, so on this
2086 * architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
2087 * stream_id << ((2 * (vertex_count - 1)) % 32).
2089 fs_reg mask
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2090 abld
.SHL(mask
, sid
, shift_count
);
2091 abld
.OR(this->control_data_bits
, this->control_data_bits
, mask
);
2095 fs_visitor::emit_gs_vertex(const nir_src
&vertex_count_nir_src
,
2098 assert(stage
== MESA_SHADER_GEOMETRY
);
2100 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
2102 fs_reg vertex_count
= get_nir_src(vertex_count_nir_src
);
2103 vertex_count
.type
= BRW_REGISTER_TYPE_UD
;
2105 /* Haswell and later hardware ignores the "Render Stream Select" bits
2106 * from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
2107 * and instead sends all primitives down the pipeline for rasterization.
2108 * If the SOL stage is enabled, "Render Stream Select" is honored and
2109 * primitives bound to non-zero streams are discarded after stream output.
2111 * Since the only purpose of primives sent to non-zero streams is to
2112 * be recorded by transform feedback, we can simply discard all geometry
2113 * bound to these streams when transform feedback is disabled.
2115 if (stream_id
> 0 && !nir
->info
.has_transform_feedback_varyings
)
2118 /* If we're outputting 32 control data bits or less, then we can wait
2119 * until the shader is over to output them all. Otherwise we need to
2120 * output them as we go. Now is the time to do it, since we're about to
2121 * output the vertex_count'th vertex, so it's guaranteed that the
2122 * control data bits associated with the (vertex_count - 1)th vertex are
2125 if (gs_compile
->control_data_header_size_bits
> 32) {
2126 const fs_builder abld
=
2127 bld
.annotate("emit vertex: emit control data bits");
2129 /* Only emit control data bits if we've finished accumulating a batch
2130 * of 32 bits. This is the case when:
2132 * (vertex_count * bits_per_vertex) % 32 == 0
2134 * (in other words, when the last 5 bits of vertex_count *
2135 * bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
2136 * integer n (which is always the case, since bits_per_vertex is
2137 * always 1 or 2), this is equivalent to requiring that the last 5-n
2138 * bits of vertex_count are 0:
2140 * vertex_count & (2^(5-n) - 1) == 0
2142 * 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
2145 * vertex_count & (32 / bits_per_vertex - 1) == 0
2147 * TODO: If vertex_count is an immediate, we could do some of this math
2148 * at compile time...
2151 abld
.AND(bld
.null_reg_d(), vertex_count
,
2152 brw_imm_ud(32u / gs_compile
->control_data_bits_per_vertex
- 1u));
2153 inst
->conditional_mod
= BRW_CONDITIONAL_Z
;
2155 abld
.IF(BRW_PREDICATE_NORMAL
);
2156 /* If vertex_count is 0, then no control data bits have been
2157 * accumulated yet, so we can skip emitting them.
2159 abld
.CMP(bld
.null_reg_d(), vertex_count
, brw_imm_ud(0u),
2160 BRW_CONDITIONAL_NEQ
);
2161 abld
.IF(BRW_PREDICATE_NORMAL
);
2162 emit_gs_control_data_bits(vertex_count
);
2163 abld
.emit(BRW_OPCODE_ENDIF
);
2165 /* Reset control_data_bits to 0 so we can start accumulating a new
2168 * Note: in the case where vertex_count == 0, this neutralizes the
2169 * effect of any call to EndPrimitive() that the shader may have
2170 * made before outputting its first vertex.
2172 inst
= abld
.MOV(this->control_data_bits
, brw_imm_ud(0u));
2173 inst
->force_writemask_all
= true;
2174 abld
.emit(BRW_OPCODE_ENDIF
);
2177 emit_urb_writes(vertex_count
);
2179 /* In stream mode we have to set control data bits for all vertices
2180 * unless we have disabled control data bits completely (which we do
2181 * do for GL_POINTS outputs that don't use streams).
2183 if (gs_compile
->control_data_header_size_bits
> 0 &&
2184 gs_prog_data
->control_data_format
==
2185 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID
) {
2186 set_gs_stream_control_data_bits(vertex_count
, stream_id
);
2191 fs_visitor::emit_gs_input_load(const fs_reg
&dst
,
2192 const nir_src
&vertex_src
,
2193 unsigned base_offset
,
2194 const nir_src
&offset_src
,
2195 unsigned num_components
,
2196 unsigned first_component
)
2198 struct brw_gs_prog_data
*gs_prog_data
= brw_gs_prog_data(prog_data
);
2199 const unsigned push_reg_count
= gs_prog_data
->base
.urb_read_length
* 8;
2201 /* TODO: figure out push input layout for invocations == 1 */
2202 /* TODO: make this work with 64-bit inputs */
2203 if (gs_prog_data
->invocations
== 1 &&
2204 type_sz(dst
.type
) <= 4 &&
2205 nir_src_is_const(offset_src
) && nir_src_is_const(vertex_src
) &&
2206 4 * (base_offset
+ nir_src_as_uint(offset_src
)) < push_reg_count
) {
2207 int imm_offset
= (base_offset
+ nir_src_as_uint(offset_src
)) * 4 +
2208 nir_src_as_uint(vertex_src
) * push_reg_count
;
2209 for (unsigned i
= 0; i
< num_components
; i
++) {
2210 bld
.MOV(offset(dst
, bld
, i
),
2211 fs_reg(ATTR
, imm_offset
+ i
+ first_component
, dst
.type
));
2216 /* Resort to the pull model. Ensure the VUE handles are provided. */
2217 assert(gs_prog_data
->base
.include_vue_handles
);
2219 unsigned first_icp_handle
= gs_prog_data
->include_primitive_id
? 3 : 2;
2220 fs_reg icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2222 if (gs_prog_data
->invocations
== 1) {
2223 if (nir_src_is_const(vertex_src
)) {
2224 /* The vertex index is constant; just select the proper URB handle. */
2226 retype(brw_vec8_grf(first_icp_handle
+ nir_src_as_uint(vertex_src
), 0),
2227 BRW_REGISTER_TYPE_UD
);
2229 /* The vertex index is non-constant. We need to use indirect
2230 * addressing to fetch the proper URB handle.
2232 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
2233 * indicating that channel <n> should read the handle from
2234 * DWord <n>. We convert that to bytes by multiplying by 4.
2236 * Next, we convert the vertex index to bytes by multiplying
2237 * by 32 (shifting by 5), and add the two together. This is
2238 * the final indirect byte offset.
2240 fs_reg sequence
= bld
.vgrf(BRW_REGISTER_TYPE_UW
, 1);
2241 fs_reg channel_offsets
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2242 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2243 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2245 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
2246 bld
.MOV(sequence
, fs_reg(brw_imm_v(0x76543210)));
2247 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
2248 bld
.SHL(channel_offsets
, sequence
, brw_imm_ud(2u));
2249 /* Convert vertex_index to bytes (multiply by 32) */
2250 bld
.SHL(vertex_offset_bytes
,
2251 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2253 bld
.ADD(icp_offset_bytes
, vertex_offset_bytes
, channel_offsets
);
2255 /* Use first_icp_handle as the base offset. There is one register
2256 * of URB handles per vertex, so inform the register allocator that
2257 * we might read up to nir->info.gs.vertices_in registers.
2259 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2260 retype(brw_vec8_grf(first_icp_handle
, 0), icp_handle
.type
),
2261 fs_reg(icp_offset_bytes
),
2262 brw_imm_ud(nir
->info
.gs
.vertices_in
* REG_SIZE
));
2265 assert(gs_prog_data
->invocations
> 1);
2267 if (nir_src_is_const(vertex_src
)) {
2268 unsigned vertex
= nir_src_as_uint(vertex_src
);
2269 assert(devinfo
->gen
>= 9 || vertex
<= 5);
2271 retype(brw_vec1_grf(first_icp_handle
+ vertex
/ 8, vertex
% 8),
2272 BRW_REGISTER_TYPE_UD
));
2274 /* The vertex index is non-constant. We need to use indirect
2275 * addressing to fetch the proper URB handle.
2278 fs_reg icp_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2280 /* Convert vertex_index to bytes (multiply by 4) */
2281 bld
.SHL(icp_offset_bytes
,
2282 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2285 /* Use first_icp_handle as the base offset. There is one DWord
2286 * of URB handles per vertex, so inform the register allocator that
2287 * we might read up to ceil(nir->info.gs.vertices_in / 8) registers.
2289 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2290 retype(brw_vec8_grf(first_icp_handle
, 0), icp_handle
.type
),
2291 fs_reg(icp_offset_bytes
),
2292 brw_imm_ud(DIV_ROUND_UP(nir
->info
.gs
.vertices_in
, 8) *
2299 fs_reg tmp_dst
= dst
;
2300 fs_reg indirect_offset
= get_nir_src(offset_src
);
2301 unsigned num_iterations
= 1;
2302 unsigned orig_num_components
= num_components
;
2304 if (type_sz(dst
.type
) == 8) {
2305 if (num_components
> 2) {
2309 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(4), dst
.type
);
2311 first_component
= first_component
/ 2;
2314 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2315 if (nir_src_is_const(offset_src
)) {
2316 /* Constant indexing - use global offset. */
2317 if (first_component
!= 0) {
2318 unsigned read_components
= num_components
+ first_component
;
2319 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2320 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, icp_handle
);
2321 inst
->size_written
= read_components
*
2322 tmp
.component_size(inst
->exec_size
);
2323 for (unsigned i
= 0; i
< num_components
; i
++) {
2324 bld
.MOV(offset(tmp_dst
, bld
, i
),
2325 offset(tmp
, bld
, i
+ first_component
));
2328 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp_dst
,
2330 inst
->size_written
= num_components
*
2331 tmp_dst
.component_size(inst
->exec_size
);
2333 inst
->offset
= base_offset
+ nir_src_as_uint(offset_src
);
2336 /* Indirect indexing - use per-slot offsets as well. */
2337 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2338 unsigned read_components
= num_components
+ first_component
;
2339 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2340 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2341 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2342 if (first_component
!= 0) {
2343 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2345 inst
->size_written
= read_components
*
2346 tmp
.component_size(inst
->exec_size
);
2347 for (unsigned i
= 0; i
< num_components
; i
++) {
2348 bld
.MOV(offset(tmp_dst
, bld
, i
),
2349 offset(tmp
, bld
, i
+ first_component
));
2352 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp_dst
,
2354 inst
->size_written
= num_components
*
2355 tmp_dst
.component_size(inst
->exec_size
);
2357 inst
->offset
= base_offset
;
2361 if (type_sz(dst
.type
) == 8) {
2362 shuffle_from_32bit_read(bld
,
2363 offset(dst
, bld
, iter
* 2),
2364 retype(tmp_dst
, BRW_REGISTER_TYPE_D
),
2369 if (num_iterations
> 1) {
2370 num_components
= orig_num_components
- 2;
2371 if(nir_src_is_const(offset_src
)) {
2374 fs_reg new_indirect
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2375 bld
.ADD(new_indirect
, indirect_offset
, brw_imm_ud(1u));
2376 indirect_offset
= new_indirect
;
2383 fs_visitor::get_indirect_offset(nir_intrinsic_instr
*instr
)
2385 nir_src
*offset_src
= nir_get_io_offset_src(instr
);
2387 if (nir_src_is_const(*offset_src
)) {
2388 /* The only constant offset we should find is 0. brw_nir.c's
2389 * add_const_offset_to_base() will fold other constant offsets
2390 * into instr->const_index[0].
2392 assert(nir_src_as_uint(*offset_src
) == 0);
2396 return get_nir_src(*offset_src
);
2400 fs_visitor::nir_emit_vs_intrinsic(const fs_builder
&bld
,
2401 nir_intrinsic_instr
*instr
)
2403 assert(stage
== MESA_SHADER_VERTEX
);
2406 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2407 dest
= get_nir_dest(instr
->dest
);
2409 switch (instr
->intrinsic
) {
2410 case nir_intrinsic_load_vertex_id
:
2411 case nir_intrinsic_load_base_vertex
:
2412 unreachable("should be lowered by nir_lower_system_values()");
2414 case nir_intrinsic_load_input
: {
2415 fs_reg src
= fs_reg(ATTR
, nir_intrinsic_base(instr
) * 4, dest
.type
);
2416 unsigned first_component
= nir_intrinsic_component(instr
);
2417 unsigned num_components
= instr
->num_components
;
2419 src
= offset(src
, bld
, nir_src_as_uint(instr
->src
[0]));
2421 if (type_sz(dest
.type
) == 8)
2422 first_component
/= 2;
2424 /* For 16-bit support maybe a temporary will be needed to copy from
2427 shuffle_from_32bit_read(bld
, dest
, retype(src
, BRW_REGISTER_TYPE_D
),
2428 first_component
, num_components
);
2432 case nir_intrinsic_load_vertex_id_zero_base
:
2433 case nir_intrinsic_load_instance_id
:
2434 case nir_intrinsic_load_base_instance
:
2435 case nir_intrinsic_load_draw_id
:
2436 case nir_intrinsic_load_first_vertex
:
2437 case nir_intrinsic_load_is_indexed_draw
:
2438 unreachable("lowered by brw_nir_lower_vs_inputs");
2441 nir_emit_intrinsic(bld
, instr
);
2447 fs_visitor::nir_emit_tcs_intrinsic(const fs_builder
&bld
,
2448 nir_intrinsic_instr
*instr
)
2450 assert(stage
== MESA_SHADER_TESS_CTRL
);
2451 struct brw_tcs_prog_key
*tcs_key
= (struct brw_tcs_prog_key
*) key
;
2452 struct brw_tcs_prog_data
*tcs_prog_data
= brw_tcs_prog_data(prog_data
);
2455 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2456 dst
= get_nir_dest(instr
->dest
);
2458 switch (instr
->intrinsic
) {
2459 case nir_intrinsic_load_primitive_id
:
2460 bld
.MOV(dst
, fs_reg(brw_vec1_grf(0, 1)));
2462 case nir_intrinsic_load_invocation_id
:
2463 bld
.MOV(retype(dst
, invocation_id
.type
), invocation_id
);
2465 case nir_intrinsic_load_patch_vertices_in
:
2466 bld
.MOV(retype(dst
, BRW_REGISTER_TYPE_D
),
2467 brw_imm_d(tcs_key
->input_vertices
));
2470 case nir_intrinsic_barrier
: {
2471 if (tcs_prog_data
->instances
== 1)
2474 fs_reg m0
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2475 fs_reg m0_2
= component(m0
, 2);
2477 const fs_builder chanbld
= bld
.exec_all().group(1, 0);
2479 /* Zero the message header */
2480 bld
.exec_all().MOV(m0
, brw_imm_ud(0u));
2482 if (devinfo
->gen
< 11) {
2483 /* Copy "Barrier ID" from r0.2, bits 16:13 */
2484 chanbld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
2485 brw_imm_ud(INTEL_MASK(16, 13)));
2487 /* Shift it up to bits 27:24. */
2488 chanbld
.SHL(m0_2
, m0_2
, brw_imm_ud(11));
2490 chanbld
.AND(m0_2
, retype(brw_vec1_grf(0, 2), BRW_REGISTER_TYPE_UD
),
2491 brw_imm_ud(INTEL_MASK(30, 24)));
2494 /* Set the Barrier Count and the enable bit */
2495 if (devinfo
->gen
< 11) {
2496 chanbld
.OR(m0_2
, m0_2
,
2497 brw_imm_ud(tcs_prog_data
->instances
<< 9 | (1 << 15)));
2499 chanbld
.OR(m0_2
, m0_2
,
2500 brw_imm_ud(tcs_prog_data
->instances
<< 8 | (1 << 15)));
2503 bld
.emit(SHADER_OPCODE_BARRIER
, bld
.null_reg_ud(), m0
);
2507 case nir_intrinsic_load_input
:
2508 unreachable("nir_lower_io should never give us these.");
2511 case nir_intrinsic_load_per_vertex_input
: {
2512 fs_reg indirect_offset
= get_indirect_offset(instr
);
2513 unsigned imm_offset
= instr
->const_index
[0];
2515 const nir_src
&vertex_src
= instr
->src
[0];
2521 if (nir_src_is_const(vertex_src
)) {
2522 /* Emit a MOV to resolve <0,1,0> regioning. */
2523 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2524 unsigned vertex
= nir_src_as_uint(vertex_src
);
2526 retype(brw_vec1_grf(1 + (vertex
>> 3), vertex
& 7),
2527 BRW_REGISTER_TYPE_UD
));
2528 } else if (tcs_prog_data
->instances
== 1 &&
2529 vertex_src
.is_ssa
&&
2530 vertex_src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
&&
2531 nir_instr_as_intrinsic(vertex_src
.ssa
->parent_instr
)->intrinsic
== nir_intrinsic_load_invocation_id
) {
2532 /* For the common case of only 1 instance, an array index of
2533 * gl_InvocationID means reading g1. Skip all the indirect work.
2535 icp_handle
= retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD
);
2537 /* The vertex index is non-constant. We need to use indirect
2538 * addressing to fetch the proper URB handle.
2540 icp_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2542 /* Each ICP handle is a single DWord (4 bytes) */
2543 fs_reg vertex_offset_bytes
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2544 bld
.SHL(vertex_offset_bytes
,
2545 retype(get_nir_src(vertex_src
), BRW_REGISTER_TYPE_UD
),
2548 /* Start at g1. We might read up to 4 registers. */
2549 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
, icp_handle
,
2550 retype(brw_vec8_grf(1, 0), icp_handle
.type
), vertex_offset_bytes
,
2551 brw_imm_ud(4 * REG_SIZE
));
2554 /* We can only read two double components with each URB read, so
2555 * we send two read messages in that case, each one loading up to
2556 * two double components.
2558 unsigned num_iterations
= 1;
2559 unsigned num_components
= instr
->num_components
;
2560 unsigned first_component
= nir_intrinsic_component(instr
);
2561 fs_reg orig_dst
= dst
;
2562 if (type_sz(dst
.type
) == 8) {
2563 first_component
= first_component
/ 2;
2564 if (instr
->num_components
> 2) {
2569 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(4), dst
.type
);
2573 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2574 if (indirect_offset
.file
== BAD_FILE
) {
2575 /* Constant indexing - use global offset. */
2576 if (first_component
!= 0) {
2577 unsigned read_components
= num_components
+ first_component
;
2578 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2579 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
, icp_handle
);
2580 for (unsigned i
= 0; i
< num_components
; i
++) {
2581 bld
.MOV(offset(dst
, bld
, i
),
2582 offset(tmp
, bld
, i
+ first_component
));
2585 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
, icp_handle
);
2587 inst
->offset
= imm_offset
;
2590 /* Indirect indexing - use per-slot offsets as well. */
2591 const fs_reg srcs
[] = { icp_handle
, indirect_offset
};
2592 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2593 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
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_PER_SLOT
, tmp
,
2599 for (unsigned i
= 0; i
< num_components
; i
++) {
2600 bld
.MOV(offset(dst
, bld
, i
),
2601 offset(tmp
, bld
, i
+ first_component
));
2604 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
,
2607 inst
->offset
= imm_offset
;
2610 inst
->size_written
= (num_components
+ first_component
) *
2611 inst
->dst
.component_size(inst
->exec_size
);
2613 /* If we are reading 64-bit data using 32-bit read messages we need
2614 * build proper 64-bit data elements by shuffling the low and high
2615 * 32-bit components around like we do for other things like UBOs
2618 if (type_sz(dst
.type
) == 8) {
2619 shuffle_from_32bit_read(bld
,
2620 offset(orig_dst
, bld
, iter
* 2),
2621 retype(dst
, BRW_REGISTER_TYPE_D
),
2625 /* Copy the temporary to the destination to deal with writemasking.
2627 * Also attempt to deal with gl_PointSize being in the .w component.
2629 if (inst
->offset
== 0 && indirect_offset
.file
== BAD_FILE
) {
2630 assert(type_sz(dst
.type
) < 8);
2631 inst
->dst
= bld
.vgrf(dst
.type
, 4);
2632 inst
->size_written
= 4 * REG_SIZE
;
2633 bld
.MOV(dst
, offset(inst
->dst
, bld
, 3));
2636 /* If we are loading double data and we need a second read message
2637 * adjust the write offset
2639 if (num_iterations
> 1) {
2640 num_components
= instr
->num_components
- 2;
2647 case nir_intrinsic_load_output
:
2648 case nir_intrinsic_load_per_vertex_output
: {
2649 fs_reg indirect_offset
= get_indirect_offset(instr
);
2650 unsigned imm_offset
= instr
->const_index
[0];
2651 unsigned first_component
= nir_intrinsic_component(instr
);
2654 if (indirect_offset
.file
== BAD_FILE
) {
2655 /* Replicate the patch handle to all enabled channels */
2656 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2657 bld
.MOV(patch_handle
,
2658 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
));
2661 if (first_component
!= 0) {
2662 unsigned read_components
=
2663 instr
->num_components
+ first_component
;
2664 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2665 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
,
2667 inst
->size_written
= read_components
* REG_SIZE
;
2668 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2669 bld
.MOV(offset(dst
, bld
, i
),
2670 offset(tmp
, bld
, i
+ first_component
));
2673 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dst
,
2675 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2677 inst
->offset
= imm_offset
;
2681 /* Indirect indexing - use per-slot offsets as well. */
2682 const fs_reg srcs
[] = {
2683 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2686 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2687 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2688 if (first_component
!= 0) {
2689 unsigned read_components
=
2690 instr
->num_components
+ first_component
;
2691 fs_reg tmp
= bld
.vgrf(dst
.type
, read_components
);
2692 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2694 inst
->size_written
= read_components
* REG_SIZE
;
2695 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2696 bld
.MOV(offset(dst
, bld
, i
),
2697 offset(tmp
, bld
, i
+ first_component
));
2700 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dst
,
2702 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2704 inst
->offset
= imm_offset
;
2710 case nir_intrinsic_store_output
:
2711 case nir_intrinsic_store_per_vertex_output
: {
2712 fs_reg value
= get_nir_src(instr
->src
[0]);
2713 bool is_64bit
= (instr
->src
[0].is_ssa
?
2714 instr
->src
[0].ssa
->bit_size
: instr
->src
[0].reg
.reg
->bit_size
) == 64;
2715 fs_reg indirect_offset
= get_indirect_offset(instr
);
2716 unsigned imm_offset
= instr
->const_index
[0];
2717 unsigned mask
= instr
->const_index
[1];
2718 unsigned header_regs
= 0;
2720 srcs
[header_regs
++] = retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
);
2722 if (indirect_offset
.file
!= BAD_FILE
) {
2723 srcs
[header_regs
++] = indirect_offset
;
2729 unsigned num_components
= util_last_bit(mask
);
2732 /* We can only pack two 64-bit components in a single message, so send
2733 * 2 messages if we have more components
2735 unsigned num_iterations
= 1;
2736 unsigned iter_components
= num_components
;
2737 unsigned first_component
= nir_intrinsic_component(instr
);
2739 first_component
= first_component
/ 2;
2740 if (instr
->num_components
> 2) {
2742 iter_components
= 2;
2746 mask
= mask
<< first_component
;
2748 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2749 if (!is_64bit
&& mask
!= WRITEMASK_XYZW
) {
2750 srcs
[header_regs
++] = brw_imm_ud(mask
<< 16);
2751 opcode
= indirect_offset
.file
!= BAD_FILE
?
2752 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2753 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2754 } else if (is_64bit
&& ((mask
& WRITEMASK_XY
) != WRITEMASK_XY
)) {
2755 /* Expand the 64-bit mask to 32-bit channels. We only handle
2756 * two channels in each iteration, so we only care about X/Y.
2758 unsigned mask32
= 0;
2759 if (mask
& WRITEMASK_X
)
2760 mask32
|= WRITEMASK_XY
;
2761 if (mask
& WRITEMASK_Y
)
2762 mask32
|= WRITEMASK_ZW
;
2764 /* If the mask does not include any of the channels X or Y there
2765 * is nothing to do in this iteration. Move on to the next couple
2766 * of 64-bit channels.
2774 srcs
[header_regs
++] = brw_imm_ud(mask32
<< 16);
2775 opcode
= indirect_offset
.file
!= BAD_FILE
?
2776 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT
:
2777 SHADER_OPCODE_URB_WRITE_SIMD8_MASKED
;
2779 opcode
= indirect_offset
.file
!= BAD_FILE
?
2780 SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT
:
2781 SHADER_OPCODE_URB_WRITE_SIMD8
;
2784 for (unsigned i
= 0; i
< iter_components
; i
++) {
2785 if (!(mask
& (1 << (i
+ first_component
))))
2789 srcs
[header_regs
+ i
+ first_component
] = offset(value
, bld
, i
);
2791 /* We need to shuffle the 64-bit data to match the layout
2792 * expected by our 32-bit URB write messages. We use a temporary
2795 unsigned channel
= iter
* 2 + i
;
2796 fs_reg dest
= shuffle_for_32bit_write(bld
, value
, channel
, 1);
2798 srcs
[header_regs
+ (i
+ first_component
) * 2] = dest
;
2799 srcs
[header_regs
+ (i
+ first_component
) * 2 + 1] =
2800 offset(dest
, bld
, 1);
2805 header_regs
+ (is_64bit
? 2 * iter_components
: iter_components
) +
2806 (is_64bit
? 2 * first_component
: first_component
);
2808 bld
.vgrf(BRW_REGISTER_TYPE_UD
, mlen
);
2809 bld
.LOAD_PAYLOAD(payload
, srcs
, mlen
, header_regs
);
2811 fs_inst
*inst
= bld
.emit(opcode
, bld
.null_reg_ud(), payload
);
2812 inst
->offset
= imm_offset
;
2815 /* If this is a 64-bit attribute, select the next two 64-bit channels
2816 * to be handled in the next iteration.
2827 nir_emit_intrinsic(bld
, instr
);
2833 fs_visitor::nir_emit_tes_intrinsic(const fs_builder
&bld
,
2834 nir_intrinsic_instr
*instr
)
2836 assert(stage
== MESA_SHADER_TESS_EVAL
);
2837 struct brw_tes_prog_data
*tes_prog_data
= brw_tes_prog_data(prog_data
);
2840 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2841 dest
= get_nir_dest(instr
->dest
);
2843 switch (instr
->intrinsic
) {
2844 case nir_intrinsic_load_primitive_id
:
2845 bld
.MOV(dest
, fs_reg(brw_vec1_grf(0, 1)));
2847 case nir_intrinsic_load_tess_coord
:
2848 /* gl_TessCoord is part of the payload in g1-3 */
2849 for (unsigned i
= 0; i
< 3; i
++) {
2850 bld
.MOV(offset(dest
, bld
, i
), fs_reg(brw_vec8_grf(1 + i
, 0)));
2854 case nir_intrinsic_load_input
:
2855 case nir_intrinsic_load_per_vertex_input
: {
2856 fs_reg indirect_offset
= get_indirect_offset(instr
);
2857 unsigned imm_offset
= instr
->const_index
[0];
2858 unsigned first_component
= nir_intrinsic_component(instr
);
2860 if (type_sz(dest
.type
) == 8) {
2861 first_component
= first_component
/ 2;
2865 if (indirect_offset
.file
== BAD_FILE
) {
2866 /* Arbitrarily only push up to 32 vec4 slots worth of data,
2867 * which is 16 registers (since each holds 2 vec4 slots).
2869 unsigned slot_count
= 1;
2870 if (type_sz(dest
.type
) == 8 && instr
->num_components
> 2)
2873 const unsigned max_push_slots
= 32;
2874 if (imm_offset
+ slot_count
<= max_push_slots
) {
2875 fs_reg src
= fs_reg(ATTR
, imm_offset
/ 2, dest
.type
);
2876 for (int i
= 0; i
< instr
->num_components
; i
++) {
2877 unsigned comp
= 16 / type_sz(dest
.type
) * (imm_offset
% 2) +
2878 i
+ first_component
;
2879 bld
.MOV(offset(dest
, bld
, i
), component(src
, comp
));
2882 tes_prog_data
->base
.urb_read_length
=
2883 MAX2(tes_prog_data
->base
.urb_read_length
,
2884 DIV_ROUND_UP(imm_offset
+ slot_count
, 2));
2886 /* Replicate the patch handle to all enabled channels */
2887 const fs_reg srcs
[] = {
2888 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
)
2890 fs_reg patch_handle
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 1);
2891 bld
.LOAD_PAYLOAD(patch_handle
, srcs
, ARRAY_SIZE(srcs
), 0);
2893 if (first_component
!= 0) {
2894 unsigned read_components
=
2895 instr
->num_components
+ first_component
;
2896 fs_reg tmp
= bld
.vgrf(dest
.type
, read_components
);
2897 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, tmp
,
2899 inst
->size_written
= read_components
* REG_SIZE
;
2900 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
2901 bld
.MOV(offset(dest
, bld
, i
),
2902 offset(tmp
, bld
, i
+ first_component
));
2905 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8
, dest
,
2907 inst
->size_written
= instr
->num_components
* REG_SIZE
;
2910 inst
->offset
= imm_offset
;
2913 /* Indirect indexing - use per-slot offsets as well. */
2915 /* We can only read two double components with each URB read, so
2916 * we send two read messages in that case, each one loading up to
2917 * two double components.
2919 unsigned num_iterations
= 1;
2920 unsigned num_components
= instr
->num_components
;
2921 fs_reg orig_dest
= dest
;
2922 if (type_sz(dest
.type
) == 8) {
2923 if (instr
->num_components
> 2) {
2927 fs_reg tmp
= fs_reg(VGRF
, alloc
.allocate(4), dest
.type
);
2931 for (unsigned iter
= 0; iter
< num_iterations
; iter
++) {
2932 const fs_reg srcs
[] = {
2933 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD
),
2936 fs_reg payload
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
2937 bld
.LOAD_PAYLOAD(payload
, srcs
, ARRAY_SIZE(srcs
), 0);
2939 if (first_component
!= 0) {
2940 unsigned read_components
=
2941 num_components
+ first_component
;
2942 fs_reg tmp
= bld
.vgrf(dest
.type
, read_components
);
2943 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, tmp
,
2945 for (unsigned i
= 0; i
< num_components
; i
++) {
2946 bld
.MOV(offset(dest
, bld
, i
),
2947 offset(tmp
, bld
, i
+ first_component
));
2950 inst
= bld
.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT
, dest
,
2954 inst
->offset
= imm_offset
;
2955 inst
->size_written
= (num_components
+ first_component
) *
2956 inst
->dst
.component_size(inst
->exec_size
);
2958 /* If we are reading 64-bit data using 32-bit read messages we need
2959 * build proper 64-bit data elements by shuffling the low and high
2960 * 32-bit components around like we do for other things like UBOs
2963 if (type_sz(dest
.type
) == 8) {
2964 shuffle_from_32bit_read(bld
,
2965 offset(orig_dest
, bld
, iter
* 2),
2966 retype(dest
, BRW_REGISTER_TYPE_D
),
2970 /* If we are loading double data and we need a second read message
2973 if (num_iterations
> 1) {
2974 num_components
= instr
->num_components
- 2;
2982 nir_emit_intrinsic(bld
, instr
);
2988 fs_visitor::nir_emit_gs_intrinsic(const fs_builder
&bld
,
2989 nir_intrinsic_instr
*instr
)
2991 assert(stage
== MESA_SHADER_GEOMETRY
);
2992 fs_reg indirect_offset
;
2995 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
2996 dest
= get_nir_dest(instr
->dest
);
2998 switch (instr
->intrinsic
) {
2999 case nir_intrinsic_load_primitive_id
:
3000 assert(stage
== MESA_SHADER_GEOMETRY
);
3001 assert(brw_gs_prog_data(prog_data
)->include_primitive_id
);
3002 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
3003 retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD
));
3006 case nir_intrinsic_load_input
:
3007 unreachable("load_input intrinsics are invalid for the GS stage");
3009 case nir_intrinsic_load_per_vertex_input
:
3010 emit_gs_input_load(dest
, instr
->src
[0], instr
->const_index
[0],
3011 instr
->src
[1], instr
->num_components
,
3012 nir_intrinsic_component(instr
));
3015 case nir_intrinsic_emit_vertex_with_counter
:
3016 emit_gs_vertex(instr
->src
[0], instr
->const_index
[0]);
3019 case nir_intrinsic_end_primitive_with_counter
:
3020 emit_gs_end_primitive(instr
->src
[0]);
3023 case nir_intrinsic_set_vertex_count
:
3024 bld
.MOV(this->final_gs_vertex_count
, get_nir_src(instr
->src
[0]));
3027 case nir_intrinsic_load_invocation_id
: {
3028 fs_reg val
= nir_system_values
[SYSTEM_VALUE_INVOCATION_ID
];
3029 assert(val
.file
!= BAD_FILE
);
3030 dest
.type
= val
.type
;
3036 nir_emit_intrinsic(bld
, instr
);
3042 * Fetch the current render target layer index.
3045 fetch_render_target_array_index(const fs_builder
&bld
)
3047 if (bld
.shader
->devinfo
->gen
>= 6) {
3048 /* The render target array index is provided in the thread payload as
3049 * bits 26:16 of r0.0.
3051 const fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3052 bld
.AND(idx
, brw_uw1_reg(BRW_GENERAL_REGISTER_FILE
, 0, 1),
3056 /* Pre-SNB we only ever render into the first layer of the framebuffer
3057 * since layered rendering is not implemented.
3059 return brw_imm_ud(0);
3064 * Fake non-coherent framebuffer read implemented using TXF to fetch from the
3065 * framebuffer at the current fragment coordinates and sample index.
3068 fs_visitor::emit_non_coherent_fb_read(const fs_builder
&bld
, const fs_reg
&dst
,
3071 const struct gen_device_info
*devinfo
= bld
.shader
->devinfo
;
3073 assert(bld
.shader
->stage
== MESA_SHADER_FRAGMENT
);
3074 const brw_wm_prog_key
*wm_key
=
3075 reinterpret_cast<const brw_wm_prog_key
*>(key
);
3076 assert(!wm_key
->coherent_fb_fetch
);
3077 const struct brw_wm_prog_data
*wm_prog_data
=
3078 brw_wm_prog_data(stage_prog_data
);
3080 /* Calculate the surface index relative to the start of the texture binding
3081 * table block, since that's what the texturing messages expect.
3083 const unsigned surface
= target
+
3084 wm_prog_data
->binding_table
.render_target_read_start
-
3085 wm_prog_data
->base
.binding_table
.texture_start
;
3087 /* Calculate the fragment coordinates. */
3088 const fs_reg coords
= bld
.vgrf(BRW_REGISTER_TYPE_UD
, 3);
3089 bld
.MOV(offset(coords
, bld
, 0), pixel_x
);
3090 bld
.MOV(offset(coords
, bld
, 1), pixel_y
);
3091 bld
.MOV(offset(coords
, bld
, 2), fetch_render_target_array_index(bld
));
3093 /* Calculate the sample index and MCS payload when multisampling. Luckily
3094 * the MCS fetch message behaves deterministically for UMS surfaces, so it
3095 * shouldn't be necessary to recompile based on whether the framebuffer is
3098 if (wm_key
->multisample_fbo
&&
3099 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
].file
== BAD_FILE
)
3100 nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
] = *emit_sampleid_setup();
3102 const fs_reg sample
= nir_system_values
[SYSTEM_VALUE_SAMPLE_ID
];
3103 const fs_reg mcs
= wm_key
->multisample_fbo
?
3104 emit_mcs_fetch(coords
, 3, brw_imm_ud(surface
)) : fs_reg();
3106 /* Use either a normal or a CMS texel fetch message depending on whether
3107 * the framebuffer is single or multisample. On SKL+ use the wide CMS
3108 * message just in case the framebuffer uses 16x multisampling, it should
3109 * be equivalent to the normal CMS fetch for lower multisampling modes.
3111 const opcode op
= !wm_key
->multisample_fbo
? SHADER_OPCODE_TXF_LOGICAL
:
3112 devinfo
->gen
>= 9 ? SHADER_OPCODE_TXF_CMS_W_LOGICAL
:
3113 SHADER_OPCODE_TXF_CMS_LOGICAL
;
3115 /* Emit the instruction. */
3116 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
3117 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = coords
;
3118 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_ud(0);
3119 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = sample
;
3120 srcs
[TEX_LOGICAL_SRC_MCS
] = mcs
;
3121 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(surface
);
3122 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(0);
3123 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_ud(3);
3124 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_ud(0);
3126 fs_inst
*inst
= bld
.emit(op
, dst
, srcs
, ARRAY_SIZE(srcs
));
3127 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
3133 * Actual coherent framebuffer read implemented using the native render target
3134 * read message. Requires SKL+.
3137 emit_coherent_fb_read(const fs_builder
&bld
, const fs_reg
&dst
, unsigned target
)
3139 assert(bld
.shader
->devinfo
->gen
>= 9);
3140 fs_inst
*inst
= bld
.emit(FS_OPCODE_FB_READ_LOGICAL
, dst
);
3141 inst
->target
= target
;
3142 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
3148 alloc_temporary(const fs_builder
&bld
, unsigned size
, fs_reg
*regs
, unsigned n
)
3150 if (n
&& regs
[0].file
!= BAD_FILE
) {
3154 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_F
, size
);
3156 for (unsigned i
= 0; i
< n
; i
++)
3164 alloc_frag_output(fs_visitor
*v
, unsigned location
)
3166 assert(v
->stage
== MESA_SHADER_FRAGMENT
);
3167 const brw_wm_prog_key
*const key
=
3168 reinterpret_cast<const brw_wm_prog_key
*>(v
->key
);
3169 const unsigned l
= GET_FIELD(location
, BRW_NIR_FRAG_OUTPUT_LOCATION
);
3170 const unsigned i
= GET_FIELD(location
, BRW_NIR_FRAG_OUTPUT_INDEX
);
3172 if (i
> 0 || (key
->force_dual_color_blend
&& l
== FRAG_RESULT_DATA1
))
3173 return alloc_temporary(v
->bld
, 4, &v
->dual_src_output
, 1);
3175 else if (l
== FRAG_RESULT_COLOR
)
3176 return alloc_temporary(v
->bld
, 4, v
->outputs
,
3177 MAX2(key
->nr_color_regions
, 1));
3179 else if (l
== FRAG_RESULT_DEPTH
)
3180 return alloc_temporary(v
->bld
, 1, &v
->frag_depth
, 1);
3182 else if (l
== FRAG_RESULT_STENCIL
)
3183 return alloc_temporary(v
->bld
, 1, &v
->frag_stencil
, 1);
3185 else if (l
== FRAG_RESULT_SAMPLE_MASK
)
3186 return alloc_temporary(v
->bld
, 1, &v
->sample_mask
, 1);
3188 else if (l
>= FRAG_RESULT_DATA0
&&
3189 l
< FRAG_RESULT_DATA0
+ BRW_MAX_DRAW_BUFFERS
)
3190 return alloc_temporary(v
->bld
, 4,
3191 &v
->outputs
[l
- FRAG_RESULT_DATA0
], 1);
3194 unreachable("Invalid location");
3198 fs_visitor::nir_emit_fs_intrinsic(const fs_builder
&bld
,
3199 nir_intrinsic_instr
*instr
)
3201 assert(stage
== MESA_SHADER_FRAGMENT
);
3204 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3205 dest
= get_nir_dest(instr
->dest
);
3207 switch (instr
->intrinsic
) {
3208 case nir_intrinsic_load_front_face
:
3209 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
3210 *emit_frontfacing_interpolation());
3213 case nir_intrinsic_load_sample_pos
: {
3214 fs_reg sample_pos
= nir_system_values
[SYSTEM_VALUE_SAMPLE_POS
];
3215 assert(sample_pos
.file
!= BAD_FILE
);
3216 dest
.type
= sample_pos
.type
;
3217 bld
.MOV(dest
, sample_pos
);
3218 bld
.MOV(offset(dest
, bld
, 1), offset(sample_pos
, bld
, 1));
3222 case nir_intrinsic_load_layer_id
:
3223 dest
.type
= BRW_REGISTER_TYPE_UD
;
3224 bld
.MOV(dest
, fetch_render_target_array_index(bld
));
3227 case nir_intrinsic_load_helper_invocation
:
3228 case nir_intrinsic_load_sample_mask_in
:
3229 case nir_intrinsic_load_sample_id
: {
3230 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3231 fs_reg val
= nir_system_values
[sv
];
3232 assert(val
.file
!= BAD_FILE
);
3233 dest
.type
= val
.type
;
3238 case nir_intrinsic_store_output
: {
3239 const fs_reg src
= get_nir_src(instr
->src
[0]);
3240 const unsigned store_offset
= nir_src_as_uint(instr
->src
[1]);
3241 const unsigned location
= nir_intrinsic_base(instr
) +
3242 SET_FIELD(store_offset
, BRW_NIR_FRAG_OUTPUT_LOCATION
);
3243 const fs_reg new_dest
= retype(alloc_frag_output(this, location
),
3246 for (unsigned j
= 0; j
< instr
->num_components
; j
++)
3247 bld
.MOV(offset(new_dest
, bld
, nir_intrinsic_component(instr
) + j
),
3248 offset(src
, bld
, j
));
3253 case nir_intrinsic_load_output
: {
3254 const unsigned l
= GET_FIELD(nir_intrinsic_base(instr
),
3255 BRW_NIR_FRAG_OUTPUT_LOCATION
);
3256 assert(l
>= FRAG_RESULT_DATA0
);
3257 const unsigned load_offset
= nir_src_as_uint(instr
->src
[0]);
3258 const unsigned target
= l
- FRAG_RESULT_DATA0
+ load_offset
;
3259 const fs_reg tmp
= bld
.vgrf(dest
.type
, 4);
3261 if (reinterpret_cast<const brw_wm_prog_key
*>(key
)->coherent_fb_fetch
)
3262 emit_coherent_fb_read(bld
, tmp
, target
);
3264 emit_non_coherent_fb_read(bld
, tmp
, target
);
3266 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
3267 bld
.MOV(offset(dest
, bld
, j
),
3268 offset(tmp
, bld
, nir_intrinsic_component(instr
) + j
));
3274 case nir_intrinsic_discard
:
3275 case nir_intrinsic_discard_if
: {
3276 /* We track our discarded pixels in f0.1. By predicating on it, we can
3277 * update just the flag bits that aren't yet discarded. If there's no
3278 * condition, we emit a CMP of g0 != g0, so all currently executing
3279 * channels will get turned off.
3282 if (instr
->intrinsic
== nir_intrinsic_discard_if
) {
3283 cmp
= bld
.CMP(bld
.null_reg_f(), get_nir_src(instr
->src
[0]),
3284 brw_imm_d(0), BRW_CONDITIONAL_Z
);
3286 fs_reg some_reg
= fs_reg(retype(brw_vec8_grf(0, 0),
3287 BRW_REGISTER_TYPE_UW
));
3288 cmp
= bld
.CMP(bld
.null_reg_f(), some_reg
, some_reg
, BRW_CONDITIONAL_NZ
);
3290 cmp
->predicate
= BRW_PREDICATE_NORMAL
;
3291 cmp
->flag_subreg
= 1;
3293 if (devinfo
->gen
>= 6) {
3294 emit_discard_jump();
3297 limit_dispatch_width(16, "Fragment discard not implemented in SIMD32 mode.");
3301 case nir_intrinsic_load_input
: {
3302 /* load_input is only used for flat inputs */
3303 unsigned base
= nir_intrinsic_base(instr
);
3304 unsigned comp
= nir_intrinsic_component(instr
);
3305 unsigned num_components
= instr
->num_components
;
3306 fs_reg orig_dest
= dest
;
3307 enum brw_reg_type type
= dest
.type
;
3309 /* Special case fields in the VUE header */
3310 if (base
== VARYING_SLOT_LAYER
)
3312 else if (base
== VARYING_SLOT_VIEWPORT
)
3315 if (nir_dest_bit_size(instr
->dest
) == 64) {
3316 /* const_index is in 32-bit type size units that could not be aligned
3317 * with DF. We need to read the double vector as if it was a float
3318 * vector of twice the number of components to fetch the right data.
3320 type
= BRW_REGISTER_TYPE_F
;
3321 num_components
*= 2;
3322 dest
= bld
.vgrf(type
, num_components
);
3325 for (unsigned int i
= 0; i
< num_components
; i
++) {
3326 bld
.MOV(offset(retype(dest
, type
), bld
, i
),
3327 retype(component(interp_reg(base
, comp
+ i
), 3), type
));
3330 if (nir_dest_bit_size(instr
->dest
) == 64) {
3331 shuffle_from_32bit_read(bld
, orig_dest
, dest
, 0,
3332 instr
->num_components
);
3337 case nir_intrinsic_load_barycentric_pixel
:
3338 case nir_intrinsic_load_barycentric_centroid
:
3339 case nir_intrinsic_load_barycentric_sample
:
3340 /* Do nothing - load_interpolated_input handling will handle it later. */
3343 case nir_intrinsic_load_barycentric_at_sample
: {
3344 const glsl_interp_mode interpolation
=
3345 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(instr
);
3347 if (nir_src_is_const(instr
->src
[0])) {
3348 unsigned msg_data
= nir_src_as_uint(instr
->src
[0]) << 4;
3350 emit_pixel_interpolater_send(bld
,
3351 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3354 brw_imm_ud(msg_data
),
3357 const fs_reg sample_src
= retype(get_nir_src(instr
->src
[0]),
3358 BRW_REGISTER_TYPE_UD
);
3360 if (nir_src_is_dynamically_uniform(instr
->src
[0])) {
3361 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3362 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3363 bld
.exec_all().group(1, 0)
3364 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3365 emit_pixel_interpolater_send(bld
,
3366 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3372 /* Make a loop that sends a message to the pixel interpolater
3373 * for the sample number in each live channel. If there are
3374 * multiple channels with the same sample number then these
3375 * will be handled simultaneously with a single interation of
3378 bld
.emit(BRW_OPCODE_DO
);
3380 /* Get the next live sample number into sample_id_reg */
3381 const fs_reg sample_id
= bld
.emit_uniformize(sample_src
);
3383 /* Set the flag register so that we can perform the send
3384 * message on all channels that have the same sample number
3386 bld
.CMP(bld
.null_reg_ud(),
3387 sample_src
, sample_id
,
3388 BRW_CONDITIONAL_EQ
);
3389 const fs_reg msg_data
= vgrf(glsl_type::uint_type
);
3390 bld
.exec_all().group(1, 0)
3391 .SHL(msg_data
, sample_id
, brw_imm_ud(4u));
3393 emit_pixel_interpolater_send(bld
,
3394 FS_OPCODE_INTERPOLATE_AT_SAMPLE
,
3397 component(msg_data
, 0),
3399 set_predicate(BRW_PREDICATE_NORMAL
, inst
);
3401 /* Continue the loop if there are any live channels left */
3402 set_predicate_inv(BRW_PREDICATE_NORMAL
,
3404 bld
.emit(BRW_OPCODE_WHILE
));
3410 case nir_intrinsic_load_barycentric_at_offset
: {
3411 const glsl_interp_mode interpolation
=
3412 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(instr
);
3414 nir_const_value
*const_offset
= nir_src_as_const_value(instr
->src
[0]);
3417 assert(nir_src_bit_size(instr
->src
[0]) == 32);
3418 unsigned off_x
= MIN2((int)(const_offset
[0].f32
* 16), 7) & 0xf;
3419 unsigned off_y
= MIN2((int)(const_offset
[1].f32
* 16), 7) & 0xf;
3421 emit_pixel_interpolater_send(bld
,
3422 FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET
,
3425 brw_imm_ud(off_x
| (off_y
<< 4)),
3428 fs_reg src
= vgrf(glsl_type::ivec2_type
);
3429 fs_reg offset_src
= retype(get_nir_src(instr
->src
[0]),
3430 BRW_REGISTER_TYPE_F
);
3431 for (int i
= 0; i
< 2; i
++) {
3432 fs_reg temp
= vgrf(glsl_type::float_type
);
3433 bld
.MUL(temp
, offset(offset_src
, bld
, i
), brw_imm_f(16.0f
));
3434 fs_reg itemp
= vgrf(glsl_type::int_type
);
3436 bld
.MOV(itemp
, temp
);
3438 /* Clamp the upper end of the range to +7/16.
3439 * ARB_gpu_shader5 requires that we support a maximum offset
3440 * of +0.5, which isn't representable in a S0.4 value -- if
3441 * we didn't clamp it, we'd end up with -8/16, which is the
3442 * opposite of what the shader author wanted.
3444 * This is legal due to ARB_gpu_shader5's quantization
3447 * "Not all values of <offset> may be supported; x and y
3448 * offsets may be rounded to fixed-point values with the
3449 * number of fraction bits given by the
3450 * implementation-dependent constant
3451 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
3453 set_condmod(BRW_CONDITIONAL_L
,
3454 bld
.SEL(offset(src
, bld
, i
), itemp
, brw_imm_d(7)));
3457 const enum opcode opcode
= FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET
;
3458 emit_pixel_interpolater_send(bld
,
3468 case nir_intrinsic_load_interpolated_input
: {
3469 if (nir_intrinsic_base(instr
) == VARYING_SLOT_POS
) {
3470 emit_fragcoord_interpolation(dest
);
3474 assert(instr
->src
[0].ssa
&&
3475 instr
->src
[0].ssa
->parent_instr
->type
== nir_instr_type_intrinsic
);
3476 nir_intrinsic_instr
*bary_intrinsic
=
3477 nir_instr_as_intrinsic(instr
->src
[0].ssa
->parent_instr
);
3478 nir_intrinsic_op bary_intrin
= bary_intrinsic
->intrinsic
;
3479 enum glsl_interp_mode interp_mode
=
3480 (enum glsl_interp_mode
) nir_intrinsic_interp_mode(bary_intrinsic
);
3483 if (bary_intrin
== nir_intrinsic_load_barycentric_at_offset
||
3484 bary_intrin
== nir_intrinsic_load_barycentric_at_sample
) {
3485 /* Use the result of the PI message */
3486 dst_xy
= retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_F
);
3488 /* Use the delta_xy values computed from the payload */
3489 enum brw_barycentric_mode bary
=
3490 brw_barycentric_mode(interp_mode
, bary_intrin
);
3492 dst_xy
= this->delta_xy
[bary
];
3495 for (unsigned int i
= 0; i
< instr
->num_components
; i
++) {
3497 component(interp_reg(nir_intrinsic_base(instr
),
3498 nir_intrinsic_component(instr
) + i
), 0);
3499 interp
.type
= BRW_REGISTER_TYPE_F
;
3500 dest
.type
= BRW_REGISTER_TYPE_F
;
3502 if (devinfo
->gen
< 6 && interp_mode
== INTERP_MODE_SMOOTH
) {
3503 fs_reg tmp
= vgrf(glsl_type::float_type
);
3504 bld
.emit(FS_OPCODE_LINTERP
, tmp
, dst_xy
, interp
);
3505 bld
.MUL(offset(dest
, bld
, i
), tmp
, this->pixel_w
);
3507 bld
.emit(FS_OPCODE_LINTERP
, offset(dest
, bld
, i
), dst_xy
, interp
);
3514 nir_emit_intrinsic(bld
, instr
);
3520 get_op_for_atomic_add(nir_intrinsic_instr
*instr
, unsigned src
)
3522 if (nir_src_is_const(instr
->src
[src
])) {
3523 int64_t add_val
= nir_src_as_int(instr
->src
[src
]);
3526 else if (add_val
== -1)
3534 fs_visitor::nir_emit_cs_intrinsic(const fs_builder
&bld
,
3535 nir_intrinsic_instr
*instr
)
3537 assert(stage
== MESA_SHADER_COMPUTE
);
3538 struct brw_cs_prog_data
*cs_prog_data
= brw_cs_prog_data(prog_data
);
3541 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3542 dest
= get_nir_dest(instr
->dest
);
3544 switch (instr
->intrinsic
) {
3545 case nir_intrinsic_barrier
:
3547 cs_prog_data
->uses_barrier
= true;
3550 case nir_intrinsic_load_subgroup_id
:
3551 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
), subgroup_id
);
3554 case nir_intrinsic_load_local_invocation_id
:
3555 case nir_intrinsic_load_work_group_id
: {
3556 gl_system_value sv
= nir_system_value_from_intrinsic(instr
->intrinsic
);
3557 fs_reg val
= nir_system_values
[sv
];
3558 assert(val
.file
!= BAD_FILE
);
3559 dest
.type
= val
.type
;
3560 for (unsigned i
= 0; i
< 3; i
++)
3561 bld
.MOV(offset(dest
, bld
, i
), offset(val
, bld
, i
));
3565 case nir_intrinsic_load_num_work_groups
: {
3566 const unsigned surface
=
3567 cs_prog_data
->binding_table
.work_groups_start
;
3569 cs_prog_data
->uses_num_work_groups
= true;
3571 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3572 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(surface
);
3573 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3574 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(1); /* num components */
3576 /* Read the 3 GLuint components of gl_NumWorkGroups */
3577 for (unsigned i
= 0; i
< 3; i
++) {
3578 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = brw_imm_ud(i
<< 2);
3579 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
3580 offset(dest
, bld
, i
), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3585 case nir_intrinsic_shared_atomic_add
:
3586 nir_emit_shared_atomic(bld
, get_op_for_atomic_add(instr
, 1), instr
);
3588 case nir_intrinsic_shared_atomic_imin
:
3589 nir_emit_shared_atomic(bld
, BRW_AOP_IMIN
, instr
);
3591 case nir_intrinsic_shared_atomic_umin
:
3592 nir_emit_shared_atomic(bld
, BRW_AOP_UMIN
, instr
);
3594 case nir_intrinsic_shared_atomic_imax
:
3595 nir_emit_shared_atomic(bld
, BRW_AOP_IMAX
, instr
);
3597 case nir_intrinsic_shared_atomic_umax
:
3598 nir_emit_shared_atomic(bld
, BRW_AOP_UMAX
, instr
);
3600 case nir_intrinsic_shared_atomic_and
:
3601 nir_emit_shared_atomic(bld
, BRW_AOP_AND
, instr
);
3603 case nir_intrinsic_shared_atomic_or
:
3604 nir_emit_shared_atomic(bld
, BRW_AOP_OR
, instr
);
3606 case nir_intrinsic_shared_atomic_xor
:
3607 nir_emit_shared_atomic(bld
, BRW_AOP_XOR
, instr
);
3609 case nir_intrinsic_shared_atomic_exchange
:
3610 nir_emit_shared_atomic(bld
, BRW_AOP_MOV
, instr
);
3612 case nir_intrinsic_shared_atomic_comp_swap
:
3613 nir_emit_shared_atomic(bld
, BRW_AOP_CMPWR
, instr
);
3615 case nir_intrinsic_shared_atomic_fmin
:
3616 nir_emit_shared_atomic_float(bld
, BRW_AOP_FMIN
, instr
);
3618 case nir_intrinsic_shared_atomic_fmax
:
3619 nir_emit_shared_atomic_float(bld
, BRW_AOP_FMAX
, instr
);
3621 case nir_intrinsic_shared_atomic_fcomp_swap
:
3622 nir_emit_shared_atomic_float(bld
, BRW_AOP_FCMPWR
, instr
);
3625 case nir_intrinsic_load_shared
: {
3626 assert(devinfo
->gen
>= 7);
3627 assert(stage
== MESA_SHADER_COMPUTE
);
3629 const unsigned bit_size
= nir_dest_bit_size(instr
->dest
);
3630 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3631 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
3632 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[0]);
3633 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3635 /* Make dest unsigned because that's what the temporary will be */
3636 dest
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
3638 /* Read the vector */
3639 if (nir_intrinsic_align(instr
) >= 4) {
3640 assert(nir_dest_bit_size(instr
->dest
) == 32);
3641 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3643 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
3644 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3645 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
3647 assert(nir_dest_bit_size(instr
->dest
) <= 32);
3648 assert(nir_dest_num_components(instr
->dest
) == 1);
3649 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
3651 fs_reg read_result
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3652 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
,
3653 read_result
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3654 bld
.MOV(dest
, read_result
);
3659 case nir_intrinsic_store_shared
: {
3660 assert(devinfo
->gen
>= 7);
3661 assert(stage
== MESA_SHADER_COMPUTE
);
3663 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
3664 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3665 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
3666 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
3667 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3669 fs_reg data
= get_nir_src(instr
->src
[0]);
3670 data
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
3672 assert(nir_intrinsic_write_mask(instr
) ==
3673 (1u << instr
->num_components
) - 1);
3674 if (nir_intrinsic_align(instr
) >= 4) {
3675 assert(nir_src_bit_size(instr
->src
[0]) == 32);
3676 assert(nir_src_num_components(instr
->src
[0]) <= 4);
3677 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
3678 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3679 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
,
3680 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3682 assert(nir_src_bit_size(instr
->src
[0]) <= 32);
3683 assert(nir_src_num_components(instr
->src
[0]) == 1);
3684 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
3686 srcs
[SURFACE_LOGICAL_SRC_DATA
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
3687 bld
.MOV(srcs
[SURFACE_LOGICAL_SRC_DATA
], data
);
3689 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
,
3690 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3696 nir_emit_intrinsic(bld
, instr
);
3702 brw_nir_reduction_op_identity(const fs_builder
&bld
,
3703 nir_op op
, brw_reg_type type
)
3705 nir_const_value value
= nir_alu_binop_identity(op
, type_sz(type
) * 8);
3706 switch (type_sz(type
)) {
3708 assert(type
!= BRW_REGISTER_TYPE_HF
);
3709 return retype(brw_imm_uw(value
.u16
), type
);
3711 return retype(brw_imm_ud(value
.u32
), type
);
3713 if (type
== BRW_REGISTER_TYPE_DF
)
3714 return setup_imm_df(bld
, value
.f64
);
3716 return retype(brw_imm_u64(value
.u64
), type
);
3718 unreachable("Invalid type size");
3723 brw_op_for_nir_reduction_op(nir_op op
)
3726 case nir_op_iadd
: return BRW_OPCODE_ADD
;
3727 case nir_op_fadd
: return BRW_OPCODE_ADD
;
3728 case nir_op_imul
: return BRW_OPCODE_MUL
;
3729 case nir_op_fmul
: return BRW_OPCODE_MUL
;
3730 case nir_op_imin
: return BRW_OPCODE_SEL
;
3731 case nir_op_umin
: return BRW_OPCODE_SEL
;
3732 case nir_op_fmin
: return BRW_OPCODE_SEL
;
3733 case nir_op_imax
: return BRW_OPCODE_SEL
;
3734 case nir_op_umax
: return BRW_OPCODE_SEL
;
3735 case nir_op_fmax
: return BRW_OPCODE_SEL
;
3736 case nir_op_iand
: return BRW_OPCODE_AND
;
3737 case nir_op_ior
: return BRW_OPCODE_OR
;
3738 case nir_op_ixor
: return BRW_OPCODE_XOR
;
3740 unreachable("Invalid reduction operation");
3744 static brw_conditional_mod
3745 brw_cond_mod_for_nir_reduction_op(nir_op op
)
3748 case nir_op_iadd
: return BRW_CONDITIONAL_NONE
;
3749 case nir_op_fadd
: return BRW_CONDITIONAL_NONE
;
3750 case nir_op_imul
: return BRW_CONDITIONAL_NONE
;
3751 case nir_op_fmul
: return BRW_CONDITIONAL_NONE
;
3752 case nir_op_imin
: return BRW_CONDITIONAL_L
;
3753 case nir_op_umin
: return BRW_CONDITIONAL_L
;
3754 case nir_op_fmin
: return BRW_CONDITIONAL_L
;
3755 case nir_op_imax
: return BRW_CONDITIONAL_GE
;
3756 case nir_op_umax
: return BRW_CONDITIONAL_GE
;
3757 case nir_op_fmax
: return BRW_CONDITIONAL_GE
;
3758 case nir_op_iand
: return BRW_CONDITIONAL_NONE
;
3759 case nir_op_ior
: return BRW_CONDITIONAL_NONE
;
3760 case nir_op_ixor
: return BRW_CONDITIONAL_NONE
;
3762 unreachable("Invalid reduction operation");
3767 fs_visitor::get_nir_image_intrinsic_image(const brw::fs_builder
&bld
,
3768 nir_intrinsic_instr
*instr
)
3770 fs_reg image
= retype(get_nir_src_imm(instr
->src
[0]), BRW_REGISTER_TYPE_UD
);
3772 if (stage_prog_data
->binding_table
.image_start
> 0) {
3773 if (image
.file
== BRW_IMMEDIATE_VALUE
) {
3774 image
.d
+= stage_prog_data
->binding_table
.image_start
;
3776 bld
.ADD(image
, image
,
3777 brw_imm_d(stage_prog_data
->binding_table
.image_start
));
3781 return bld
.emit_uniformize(image
);
3785 fs_visitor::get_nir_ssbo_intrinsic_index(const brw::fs_builder
&bld
,
3786 nir_intrinsic_instr
*instr
)
3788 /* SSBO stores are weird in that their index is in src[1] */
3789 const unsigned src
= instr
->intrinsic
== nir_intrinsic_store_ssbo
? 1 : 0;
3792 if (nir_src_is_const(instr
->src
[src
])) {
3793 unsigned index
= stage_prog_data
->binding_table
.ssbo_start
+
3794 nir_src_as_uint(instr
->src
[src
]);
3795 surf_index
= brw_imm_ud(index
);
3797 surf_index
= vgrf(glsl_type::uint_type
);
3798 bld
.ADD(surf_index
, get_nir_src(instr
->src
[src
]),
3799 brw_imm_ud(stage_prog_data
->binding_table
.ssbo_start
));
3802 return bld
.emit_uniformize(surf_index
);
3806 image_intrinsic_coord_components(nir_intrinsic_instr
*instr
)
3808 switch (nir_intrinsic_image_dim(instr
)) {
3809 case GLSL_SAMPLER_DIM_1D
:
3810 return 1 + nir_intrinsic_image_array(instr
);
3811 case GLSL_SAMPLER_DIM_2D
:
3812 case GLSL_SAMPLER_DIM_RECT
:
3813 return 2 + nir_intrinsic_image_array(instr
);
3814 case GLSL_SAMPLER_DIM_3D
:
3815 case GLSL_SAMPLER_DIM_CUBE
:
3817 case GLSL_SAMPLER_DIM_BUF
:
3819 case GLSL_SAMPLER_DIM_MS
:
3820 return 2 + nir_intrinsic_image_array(instr
);
3822 unreachable("Invalid image dimension");
3827 fs_visitor::nir_emit_intrinsic(const fs_builder
&bld
, nir_intrinsic_instr
*instr
)
3830 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
3831 dest
= get_nir_dest(instr
->dest
);
3833 switch (instr
->intrinsic
) {
3834 case nir_intrinsic_image_load
:
3835 case nir_intrinsic_image_store
:
3836 case nir_intrinsic_image_atomic_add
:
3837 case nir_intrinsic_image_atomic_min
:
3838 case nir_intrinsic_image_atomic_max
:
3839 case nir_intrinsic_image_atomic_and
:
3840 case nir_intrinsic_image_atomic_or
:
3841 case nir_intrinsic_image_atomic_xor
:
3842 case nir_intrinsic_image_atomic_exchange
:
3843 case nir_intrinsic_image_atomic_comp_swap
: {
3844 if (stage
== MESA_SHADER_FRAGMENT
&&
3845 instr
->intrinsic
!= nir_intrinsic_image_load
)
3846 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
3848 /* Get some metadata from the image intrinsic. */
3849 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
3850 const GLenum format
= nir_intrinsic_format(instr
);
3852 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3853 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
3854 get_nir_image_intrinsic_image(bld
, instr
);
3855 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
3856 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] =
3857 brw_imm_ud(image_intrinsic_coord_components(instr
));
3859 /* Emit an image load, store or atomic op. */
3860 if (instr
->intrinsic
== nir_intrinsic_image_load
) {
3861 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3863 bld
.emit(SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL
,
3864 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3865 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
3866 } else if (instr
->intrinsic
== nir_intrinsic_image_store
) {
3867 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3868 srcs
[SURFACE_LOGICAL_SRC_DATA
] = get_nir_src(instr
->src
[3]);
3869 bld
.emit(SHADER_OPCODE_TYPED_SURFACE_WRITE_LOGICAL
,
3870 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3873 unsigned num_srcs
= info
->num_srcs
;
3875 switch (instr
->intrinsic
) {
3876 case nir_intrinsic_image_atomic_add
:
3877 assert(num_srcs
== 4);
3879 op
= get_op_for_atomic_add(instr
, 3);
3881 if (op
!= BRW_AOP_ADD
)
3884 case nir_intrinsic_image_atomic_min
:
3885 assert(format
== GL_R32UI
|| format
== GL_R32I
);
3886 op
= (format
== GL_R32I
) ? BRW_AOP_IMIN
: BRW_AOP_UMIN
;
3888 case nir_intrinsic_image_atomic_max
:
3889 assert(format
== GL_R32UI
|| format
== GL_R32I
);
3890 op
= (format
== GL_R32I
) ? BRW_AOP_IMAX
: BRW_AOP_UMAX
;
3892 case nir_intrinsic_image_atomic_and
:
3895 case nir_intrinsic_image_atomic_or
:
3898 case nir_intrinsic_image_atomic_xor
:
3901 case nir_intrinsic_image_atomic_exchange
:
3904 case nir_intrinsic_image_atomic_comp_swap
:
3908 unreachable("Not reachable.");
3911 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
3915 data
= get_nir_src(instr
->src
[3]);
3916 if (num_srcs
>= 5) {
3917 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
3918 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[4]) };
3919 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
3922 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
3924 bld
.emit(SHADER_OPCODE_TYPED_ATOMIC_LOGICAL
,
3925 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3930 case nir_intrinsic_image_size
: {
3931 /* Unlike the [un]typed load and store opcodes, the TXS that this turns
3932 * into will handle the binding table index for us in the geneerator.
3934 fs_reg image
= retype(get_nir_src_imm(instr
->src
[0]),
3935 BRW_REGISTER_TYPE_UD
);
3936 image
= bld
.emit_uniformize(image
);
3938 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
3939 srcs
[TEX_LOGICAL_SRC_SURFACE
] = image
;
3940 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_d(0);
3941 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(0);
3942 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(0);
3944 /* Since the image size is always uniform, we can just emit a SIMD8
3945 * query instruction and splat the result out.
3947 const fs_builder ubld
= bld
.exec_all().group(8, 0);
3949 fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 4);
3950 fs_inst
*inst
= ubld
.emit(SHADER_OPCODE_IMAGE_SIZE_LOGICAL
,
3951 tmp
, srcs
, ARRAY_SIZE(srcs
));
3952 inst
->size_written
= 4 * REG_SIZE
;
3954 for (unsigned c
= 0; c
< instr
->dest
.ssa
.num_components
; ++c
) {
3955 if (c
== 2 && nir_intrinsic_image_dim(instr
) == GLSL_SAMPLER_DIM_CUBE
) {
3956 bld
.emit(SHADER_OPCODE_INT_QUOTIENT
,
3957 offset(retype(dest
, tmp
.type
), bld
, c
),
3958 component(offset(tmp
, ubld
, c
), 0), brw_imm_ud(6));
3960 bld
.MOV(offset(retype(dest
, tmp
.type
), bld
, c
),
3961 component(offset(tmp
, ubld
, c
), 0));
3967 case nir_intrinsic_image_load_raw_intel
: {
3968 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3969 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
3970 get_nir_image_intrinsic_image(bld
, instr
);
3971 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
3972 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3973 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3976 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
3977 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3978 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
3982 case nir_intrinsic_image_store_raw_intel
: {
3983 if (stage
== MESA_SHADER_FRAGMENT
)
3984 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
3986 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
3987 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
3988 get_nir_image_intrinsic_image(bld
, instr
);
3989 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
3990 srcs
[SURFACE_LOGICAL_SRC_DATA
] = get_nir_src(instr
->src
[2]);
3991 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
3992 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
3994 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
,
3995 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
3999 case nir_intrinsic_group_memory_barrier
:
4000 case nir_intrinsic_memory_barrier_shared
:
4001 case nir_intrinsic_memory_barrier_atomic_counter
:
4002 case nir_intrinsic_memory_barrier_buffer
:
4003 case nir_intrinsic_memory_barrier_image
:
4004 case nir_intrinsic_memory_barrier
: {
4005 const fs_builder ubld
= bld
.group(8, 0);
4006 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
4007 ubld
.emit(SHADER_OPCODE_MEMORY_FENCE
, tmp
)
4008 ->size_written
= 2 * REG_SIZE
;
4012 case nir_intrinsic_shader_clock
: {
4013 /* We cannot do anything if there is an event, so ignore it for now */
4014 const fs_reg shader_clock
= get_timestamp(bld
);
4015 const fs_reg srcs
[] = { component(shader_clock
, 0),
4016 component(shader_clock
, 1) };
4017 bld
.LOAD_PAYLOAD(dest
, srcs
, ARRAY_SIZE(srcs
), 0);
4021 case nir_intrinsic_image_samples
:
4022 /* The driver does not support multi-sampled images. */
4023 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), brw_imm_d(1));
4026 case nir_intrinsic_load_uniform
: {
4027 /* Offsets are in bytes but they should always aligned to
4030 assert(instr
->const_index
[0] % 4 == 0 ||
4031 instr
->const_index
[0] % type_sz(dest
.type
) == 0);
4033 fs_reg
src(UNIFORM
, instr
->const_index
[0] / 4, dest
.type
);
4035 if (nir_src_is_const(instr
->src
[0])) {
4036 unsigned load_offset
= nir_src_as_uint(instr
->src
[0]);
4037 assert(load_offset
% type_sz(dest
.type
) == 0);
4038 /* For 16-bit types we add the module of the const_index[0]
4039 * offset to access to not 32-bit aligned element
4041 src
.offset
= load_offset
+ instr
->const_index
[0] % 4;
4043 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
4044 bld
.MOV(offset(dest
, bld
, j
), offset(src
, bld
, j
));
4047 fs_reg indirect
= retype(get_nir_src(instr
->src
[0]),
4048 BRW_REGISTER_TYPE_UD
);
4050 /* We need to pass a size to the MOV_INDIRECT but we don't want it to
4051 * go past the end of the uniform. In order to keep the n'th
4052 * component from running past, we subtract off the size of all but
4053 * one component of the vector.
4055 assert(instr
->const_index
[1] >=
4056 instr
->num_components
* (int) type_sz(dest
.type
));
4057 unsigned read_size
= instr
->const_index
[1] -
4058 (instr
->num_components
- 1) * type_sz(dest
.type
);
4060 bool supports_64bit_indirects
=
4061 !devinfo
->is_cherryview
&& !gen_device_info_is_9lp(devinfo
);
4063 if (type_sz(dest
.type
) != 8 || supports_64bit_indirects
) {
4064 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
4065 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
4066 offset(dest
, bld
, j
), offset(src
, bld
, j
),
4067 indirect
, brw_imm_ud(read_size
));
4070 const unsigned num_mov_indirects
=
4071 type_sz(dest
.type
) / type_sz(BRW_REGISTER_TYPE_UD
);
4072 /* We read a little bit less per MOV INDIRECT, as they are now
4073 * 32-bits ones instead of 64-bit. Fix read_size then.
4075 const unsigned read_size_32bit
= read_size
-
4076 (num_mov_indirects
- 1) * type_sz(BRW_REGISTER_TYPE_UD
);
4077 for (unsigned j
= 0; j
< instr
->num_components
; j
++) {
4078 for (unsigned i
= 0; i
< num_mov_indirects
; i
++) {
4079 bld
.emit(SHADER_OPCODE_MOV_INDIRECT
,
4080 subscript(offset(dest
, bld
, j
), BRW_REGISTER_TYPE_UD
, i
),
4081 subscript(offset(src
, bld
, j
), BRW_REGISTER_TYPE_UD
, i
),
4082 indirect
, brw_imm_ud(read_size_32bit
));
4090 case nir_intrinsic_load_ubo
: {
4092 if (nir_src_is_const(instr
->src
[0])) {
4093 const unsigned index
= stage_prog_data
->binding_table
.ubo_start
+
4094 nir_src_as_uint(instr
->src
[0]);
4095 surf_index
= brw_imm_ud(index
);
4097 /* The block index is not a constant. Evaluate the index expression
4098 * per-channel and add the base UBO index; we have to select a value
4099 * from any live channel.
4101 surf_index
= vgrf(glsl_type::uint_type
);
4102 bld
.ADD(surf_index
, get_nir_src(instr
->src
[0]),
4103 brw_imm_ud(stage_prog_data
->binding_table
.ubo_start
));
4104 surf_index
= bld
.emit_uniformize(surf_index
);
4107 if (!nir_src_is_const(instr
->src
[1])) {
4108 fs_reg base_offset
= retype(get_nir_src(instr
->src
[1]),
4109 BRW_REGISTER_TYPE_UD
);
4111 for (int i
= 0; i
< instr
->num_components
; i
++)
4112 VARYING_PULL_CONSTANT_LOAD(bld
, offset(dest
, bld
, i
), surf_index
,
4113 base_offset
, i
* type_sz(dest
.type
));
4115 /* Even if we are loading doubles, a pull constant load will load
4116 * a 32-bit vec4, so should only reserve vgrf space for that. If we
4117 * need to load a full dvec4 we will have to emit 2 loads. This is
4118 * similar to demote_pull_constants(), except that in that case we
4119 * see individual accesses to each component of the vector and then
4120 * we let CSE deal with duplicate loads. Here we see a vector access
4121 * and we have to split it if necessary.
4123 const unsigned type_size
= type_sz(dest
.type
);
4124 const unsigned load_offset
= nir_src_as_uint(instr
->src
[1]);
4126 /* See if we've selected this as a push constant candidate */
4127 if (nir_src_is_const(instr
->src
[0])) {
4128 const unsigned ubo_block
= nir_src_as_uint(instr
->src
[0]);
4129 const unsigned offset_256b
= load_offset
/ 32;
4132 for (int i
= 0; i
< 4; i
++) {
4133 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
4134 if (range
->block
== ubo_block
&&
4135 offset_256b
>= range
->start
&&
4136 offset_256b
< range
->start
+ range
->length
) {
4138 push_reg
= fs_reg(UNIFORM
, UBO_START
+ i
, dest
.type
);
4139 push_reg
.offset
= load_offset
- 32 * range
->start
;
4144 if (push_reg
.file
!= BAD_FILE
) {
4145 for (unsigned i
= 0; i
< instr
->num_components
; i
++) {
4146 bld
.MOV(offset(dest
, bld
, i
),
4147 byte_offset(push_reg
, i
* type_size
));
4153 const unsigned block_sz
= 64; /* Fetch one cacheline at a time. */
4154 const fs_builder ubld
= bld
.exec_all().group(block_sz
/ 4, 0);
4155 const fs_reg packed_consts
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4157 for (unsigned c
= 0; c
< instr
->num_components
;) {
4158 const unsigned base
= load_offset
+ c
* type_size
;
4159 /* Number of usable components in the next block-aligned load. */
4160 const unsigned count
= MIN2(instr
->num_components
- c
,
4161 (block_sz
- base
% block_sz
) / type_size
);
4163 ubld
.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD
,
4164 packed_consts
, surf_index
,
4165 brw_imm_ud(base
& ~(block_sz
- 1)));
4167 const fs_reg consts
=
4168 retype(byte_offset(packed_consts
, base
& (block_sz
- 1)),
4171 for (unsigned d
= 0; d
< count
; d
++)
4172 bld
.MOV(offset(dest
, bld
, c
+ d
), component(consts
, d
));
4180 case nir_intrinsic_load_global
: {
4181 assert(devinfo
->gen
>= 8);
4183 if (nir_intrinsic_align(instr
) >= 4) {
4184 assert(nir_dest_bit_size(instr
->dest
) == 32);
4185 fs_inst
*inst
= bld
.emit(SHADER_OPCODE_A64_UNTYPED_READ_LOGICAL
,
4187 get_nir_src(instr
->src
[0]), /* Address */
4188 fs_reg(), /* No source data */
4189 brw_imm_ud(instr
->num_components
));
4190 inst
->size_written
= instr
->num_components
*
4191 inst
->dst
.component_size(inst
->exec_size
);
4193 const unsigned bit_size
= nir_dest_bit_size(instr
->dest
);
4194 assert(bit_size
<= 32);
4195 assert(nir_dest_num_components(instr
->dest
) == 1);
4196 brw_reg_type data_type
=
4197 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
4198 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4199 bld
.emit(SHADER_OPCODE_A64_BYTE_SCATTERED_READ_LOGICAL
,
4201 get_nir_src(instr
->src
[0]), /* Address */
4202 fs_reg(), /* No source data */
4203 brw_imm_ud(bit_size
));
4204 bld
.MOV(retype(dest
, data_type
), tmp
);
4209 case nir_intrinsic_store_global
:
4210 assert(devinfo
->gen
>= 8);
4212 if (stage
== MESA_SHADER_FRAGMENT
)
4213 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4215 if (nir_intrinsic_align(instr
) >= 4) {
4216 assert(nir_src_bit_size(instr
->src
[0]) == 32);
4217 bld
.emit(SHADER_OPCODE_A64_UNTYPED_WRITE_LOGICAL
,
4219 get_nir_src(instr
->src
[1]), /* Address */
4220 get_nir_src(instr
->src
[0]), /* Data */
4221 brw_imm_ud(instr
->num_components
));
4223 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4224 assert(bit_size
<= 32);
4225 assert(nir_src_num_components(instr
->src
[0]) == 1);
4226 brw_reg_type data_type
=
4227 brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
4228 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4229 bld
.MOV(tmp
, retype(get_nir_src(instr
->src
[0]), data_type
));
4230 bld
.emit(SHADER_OPCODE_A64_BYTE_SCATTERED_WRITE_LOGICAL
,
4232 get_nir_src(instr
->src
[1]), /* Address */
4234 brw_imm_ud(nir_src_bit_size(instr
->src
[0])));
4238 case nir_intrinsic_global_atomic_add
:
4239 nir_emit_global_atomic(bld
, get_op_for_atomic_add(instr
, 1), instr
);
4241 case nir_intrinsic_global_atomic_imin
:
4242 nir_emit_global_atomic(bld
, BRW_AOP_IMIN
, instr
);
4244 case nir_intrinsic_global_atomic_umin
:
4245 nir_emit_global_atomic(bld
, BRW_AOP_UMIN
, instr
);
4247 case nir_intrinsic_global_atomic_imax
:
4248 nir_emit_global_atomic(bld
, BRW_AOP_IMAX
, instr
);
4250 case nir_intrinsic_global_atomic_umax
:
4251 nir_emit_global_atomic(bld
, BRW_AOP_UMAX
, instr
);
4253 case nir_intrinsic_global_atomic_and
:
4254 nir_emit_global_atomic(bld
, BRW_AOP_AND
, instr
);
4256 case nir_intrinsic_global_atomic_or
:
4257 nir_emit_global_atomic(bld
, BRW_AOP_OR
, instr
);
4259 case nir_intrinsic_global_atomic_xor
:
4260 nir_emit_global_atomic(bld
, BRW_AOP_XOR
, instr
);
4262 case nir_intrinsic_global_atomic_exchange
:
4263 nir_emit_global_atomic(bld
, BRW_AOP_MOV
, instr
);
4265 case nir_intrinsic_global_atomic_comp_swap
:
4266 nir_emit_global_atomic(bld
, BRW_AOP_CMPWR
, instr
);
4268 case nir_intrinsic_global_atomic_fmin
:
4269 nir_emit_global_atomic_float(bld
, BRW_AOP_FMIN
, instr
);
4271 case nir_intrinsic_global_atomic_fmax
:
4272 nir_emit_global_atomic_float(bld
, BRW_AOP_FMAX
, instr
);
4274 case nir_intrinsic_global_atomic_fcomp_swap
:
4275 nir_emit_global_atomic_float(bld
, BRW_AOP_FCMPWR
, instr
);
4278 case nir_intrinsic_load_ssbo
: {
4279 assert(devinfo
->gen
>= 7);
4281 const unsigned bit_size
= nir_dest_bit_size(instr
->dest
);
4282 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4283 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
4284 get_nir_ssbo_intrinsic_index(bld
, instr
);
4285 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
4286 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4288 /* Make dest unsigned because that's what the temporary will be */
4289 dest
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
4291 /* Read the vector */
4292 if (nir_intrinsic_align(instr
) >= 4) {
4293 assert(nir_dest_bit_size(instr
->dest
) == 32);
4294 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4296 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_READ_LOGICAL
,
4297 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4298 inst
->size_written
= instr
->num_components
* dispatch_width
* 4;
4300 assert(nir_dest_bit_size(instr
->dest
) <= 32);
4301 assert(nir_dest_num_components(instr
->dest
) == 1);
4302 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
4304 fs_reg read_result
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4305 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_READ_LOGICAL
,
4306 read_result
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4307 bld
.MOV(dest
, read_result
);
4312 case nir_intrinsic_store_ssbo
: {
4313 assert(devinfo
->gen
>= 7);
4315 if (stage
== MESA_SHADER_FRAGMENT
)
4316 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4318 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4319 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4320 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] =
4321 get_nir_ssbo_intrinsic_index(bld
, instr
);
4322 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[2]);
4323 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4325 fs_reg data
= get_nir_src(instr
->src
[0]);
4326 data
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_UD
);
4328 assert(nir_intrinsic_write_mask(instr
) ==
4329 (1u << instr
->num_components
) - 1);
4330 if (nir_intrinsic_align(instr
) >= 4) {
4331 assert(nir_src_bit_size(instr
->src
[0]) == 32);
4332 assert(nir_src_num_components(instr
->src
[0]) <= 4);
4333 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
4334 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(instr
->num_components
);
4335 bld
.emit(SHADER_OPCODE_UNTYPED_SURFACE_WRITE_LOGICAL
,
4336 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4338 assert(nir_src_bit_size(instr
->src
[0]) <= 32);
4339 assert(nir_src_num_components(instr
->src
[0]) == 1);
4340 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(bit_size
);
4342 srcs
[SURFACE_LOGICAL_SRC_DATA
] = bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4343 bld
.MOV(srcs
[SURFACE_LOGICAL_SRC_DATA
], data
);
4345 bld
.emit(SHADER_OPCODE_BYTE_SCATTERED_WRITE_LOGICAL
,
4346 fs_reg(), srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4351 case nir_intrinsic_store_output
: {
4352 fs_reg src
= get_nir_src(instr
->src
[0]);
4354 unsigned store_offset
= nir_src_as_uint(instr
->src
[1]);
4355 unsigned num_components
= instr
->num_components
;
4356 unsigned first_component
= nir_intrinsic_component(instr
);
4357 if (nir_src_bit_size(instr
->src
[0]) == 64) {
4358 src
= shuffle_for_32bit_write(bld
, src
, 0, num_components
);
4359 num_components
*= 2;
4362 fs_reg new_dest
= retype(offset(outputs
[instr
->const_index
[0]], bld
,
4363 4 * store_offset
), src
.type
);
4364 for (unsigned j
= 0; j
< num_components
; j
++) {
4365 bld
.MOV(offset(new_dest
, bld
, j
+ first_component
),
4366 offset(src
, bld
, j
));
4371 case nir_intrinsic_ssbo_atomic_add
:
4372 nir_emit_ssbo_atomic(bld
, get_op_for_atomic_add(instr
, 2), instr
);
4374 case nir_intrinsic_ssbo_atomic_imin
:
4375 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMIN
, instr
);
4377 case nir_intrinsic_ssbo_atomic_umin
:
4378 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMIN
, instr
);
4380 case nir_intrinsic_ssbo_atomic_imax
:
4381 nir_emit_ssbo_atomic(bld
, BRW_AOP_IMAX
, instr
);
4383 case nir_intrinsic_ssbo_atomic_umax
:
4384 nir_emit_ssbo_atomic(bld
, BRW_AOP_UMAX
, instr
);
4386 case nir_intrinsic_ssbo_atomic_and
:
4387 nir_emit_ssbo_atomic(bld
, BRW_AOP_AND
, instr
);
4389 case nir_intrinsic_ssbo_atomic_or
:
4390 nir_emit_ssbo_atomic(bld
, BRW_AOP_OR
, instr
);
4392 case nir_intrinsic_ssbo_atomic_xor
:
4393 nir_emit_ssbo_atomic(bld
, BRW_AOP_XOR
, instr
);
4395 case nir_intrinsic_ssbo_atomic_exchange
:
4396 nir_emit_ssbo_atomic(bld
, BRW_AOP_MOV
, instr
);
4398 case nir_intrinsic_ssbo_atomic_comp_swap
:
4399 nir_emit_ssbo_atomic(bld
, BRW_AOP_CMPWR
, instr
);
4401 case nir_intrinsic_ssbo_atomic_fmin
:
4402 nir_emit_ssbo_atomic_float(bld
, BRW_AOP_FMIN
, instr
);
4404 case nir_intrinsic_ssbo_atomic_fmax
:
4405 nir_emit_ssbo_atomic_float(bld
, BRW_AOP_FMAX
, instr
);
4407 case nir_intrinsic_ssbo_atomic_fcomp_swap
:
4408 nir_emit_ssbo_atomic_float(bld
, BRW_AOP_FCMPWR
, instr
);
4411 case nir_intrinsic_get_buffer_size
: {
4412 assert(nir_src_num_components(instr
->src
[0]) == 1);
4413 unsigned ssbo_index
= nir_src_is_const(instr
->src
[0]) ?
4414 nir_src_as_uint(instr
->src
[0]) : 0;
4416 /* A resinfo's sampler message is used to get the buffer size. The
4417 * SIMD8's writeback message consists of four registers and SIMD16's
4418 * writeback message consists of 8 destination registers (two per each
4419 * component). Because we are only interested on the first channel of
4420 * the first returned component, where resinfo returns the buffer size
4421 * for SURFTYPE_BUFFER, we can just use the SIMD8 variant regardless of
4422 * the dispatch width.
4424 const fs_builder ubld
= bld
.exec_all().group(8, 0);
4425 fs_reg src_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4426 fs_reg ret_payload
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 4);
4429 ubld
.MOV(src_payload
, brw_imm_d(0));
4431 const unsigned index
= prog_data
->binding_table
.ssbo_start
+ ssbo_index
;
4432 fs_inst
*inst
= ubld
.emit(SHADER_OPCODE_GET_BUFFER_SIZE
, ret_payload
,
4433 src_payload
, brw_imm_ud(index
));
4434 inst
->header_size
= 0;
4436 inst
->size_written
= 4 * REG_SIZE
;
4438 /* SKL PRM, vol07, 3D Media GPGPU Engine, Bounds Checking and Faulting:
4440 * "Out-of-bounds checking is always performed at a DWord granularity. If
4441 * any part of the DWord is out-of-bounds then the whole DWord is
4442 * considered out-of-bounds."
4444 * This implies that types with size smaller than 4-bytes need to be
4445 * padded if they don't complete the last dword of the buffer. But as we
4446 * need to maintain the original size we need to reverse the padding
4447 * calculation to return the correct size to know the number of elements
4448 * of an unsized array. As we stored in the last two bits of the surface
4449 * size the needed padding for the buffer, we calculate here the
4450 * original buffer_size reversing the surface_size calculation:
4452 * surface_size = isl_align(buffer_size, 4) +
4453 * (isl_align(buffer_size) - buffer_size)
4455 * buffer_size = surface_size & ~3 - surface_size & 3
4458 fs_reg size_aligned4
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4459 fs_reg size_padding
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4460 fs_reg buffer_size
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
);
4462 ubld
.AND(size_padding
, ret_payload
, brw_imm_ud(3));
4463 ubld
.AND(size_aligned4
, ret_payload
, brw_imm_ud(~3));
4464 ubld
.ADD(buffer_size
, size_aligned4
, negate(size_padding
));
4466 bld
.MOV(retype(dest
, ret_payload
.type
), component(buffer_size
, 0));
4470 case nir_intrinsic_load_subgroup_invocation
:
4471 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
),
4472 nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
]);
4475 case nir_intrinsic_load_subgroup_eq_mask
:
4476 case nir_intrinsic_load_subgroup_ge_mask
:
4477 case nir_intrinsic_load_subgroup_gt_mask
:
4478 case nir_intrinsic_load_subgroup_le_mask
:
4479 case nir_intrinsic_load_subgroup_lt_mask
:
4480 unreachable("not reached");
4482 case nir_intrinsic_vote_any
: {
4483 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4485 /* The any/all predicates do not consider channel enables. To prevent
4486 * dead channels from affecting the result, we initialize the flag with
4487 * with the identity value for the logical operation.
4489 if (dispatch_width
== 32) {
4490 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4491 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4494 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0));
4496 bld
.CMP(bld
.null_reg_d(), get_nir_src(instr
->src
[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ
);
4498 /* For some reason, the any/all predicates don't work properly with
4499 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4500 * doesn't read the correct subset of the flag register and you end up
4501 * getting garbage in the second half. Work around this by using a pair
4502 * of 1-wide MOVs and scattering the result.
4504 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4505 ubld
.MOV(res1
, brw_imm_d(0));
4506 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ANY8H
:
4507 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ANY16H
:
4508 BRW_PREDICATE_ALIGN1_ANY32H
,
4509 ubld
.MOV(res1
, brw_imm_d(-1)));
4511 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4514 case nir_intrinsic_vote_all
: {
4515 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4517 /* The any/all predicates do not consider channel enables. To prevent
4518 * dead channels from affecting the result, we initialize the flag with
4519 * with the identity value for the logical operation.
4521 if (dispatch_width
== 32) {
4522 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4523 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4524 brw_imm_ud(0xffffffff));
4526 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0xffff));
4528 bld
.CMP(bld
.null_reg_d(), get_nir_src(instr
->src
[0]), brw_imm_d(0), BRW_CONDITIONAL_NZ
);
4530 /* For some reason, the any/all predicates don't work properly with
4531 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4532 * doesn't read the correct subset of the flag register and you end up
4533 * getting garbage in the second half. Work around this by using a pair
4534 * of 1-wide MOVs and scattering the result.
4536 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4537 ubld
.MOV(res1
, brw_imm_d(0));
4538 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ALL8H
:
4539 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ALL16H
:
4540 BRW_PREDICATE_ALIGN1_ALL32H
,
4541 ubld
.MOV(res1
, brw_imm_d(-1)));
4543 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4546 case nir_intrinsic_vote_feq
:
4547 case nir_intrinsic_vote_ieq
: {
4548 fs_reg value
= get_nir_src(instr
->src
[0]);
4549 if (instr
->intrinsic
== nir_intrinsic_vote_feq
) {
4550 const unsigned bit_size
= nir_src_bit_size(instr
->src
[0]);
4551 value
.type
= brw_reg_type_from_bit_size(bit_size
, BRW_REGISTER_TYPE_F
);
4554 fs_reg uniformized
= bld
.emit_uniformize(value
);
4555 const fs_builder ubld
= bld
.exec_all().group(1, 0);
4557 /* The any/all predicates do not consider channel enables. To prevent
4558 * dead channels from affecting the result, we initialize the flag with
4559 * with the identity value for the logical operation.
4561 if (dispatch_width
== 32) {
4562 /* For SIMD32, we use a UD type so we fill both f0.0 and f0.1. */
4563 ubld
.MOV(retype(brw_flag_reg(0, 0), BRW_REGISTER_TYPE_UD
),
4564 brw_imm_ud(0xffffffff));
4566 ubld
.MOV(brw_flag_reg(0, 0), brw_imm_uw(0xffff));
4568 bld
.CMP(bld
.null_reg_d(), value
, uniformized
, BRW_CONDITIONAL_Z
);
4570 /* For some reason, the any/all predicates don't work properly with
4571 * SIMD32. In particular, it appears that a SEL with a QtrCtrl of 2H
4572 * doesn't read the correct subset of the flag register and you end up
4573 * getting garbage in the second half. Work around this by using a pair
4574 * of 1-wide MOVs and scattering the result.
4576 fs_reg res1
= ubld
.vgrf(BRW_REGISTER_TYPE_D
);
4577 ubld
.MOV(res1
, brw_imm_d(0));
4578 set_predicate(dispatch_width
== 8 ? BRW_PREDICATE_ALIGN1_ALL8H
:
4579 dispatch_width
== 16 ? BRW_PREDICATE_ALIGN1_ALL16H
:
4580 BRW_PREDICATE_ALIGN1_ALL32H
,
4581 ubld
.MOV(res1
, brw_imm_d(-1)));
4583 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_D
), component(res1
, 0));
4587 case nir_intrinsic_ballot
: {
4588 const fs_reg value
= retype(get_nir_src(instr
->src
[0]),
4589 BRW_REGISTER_TYPE_UD
);
4590 struct brw_reg flag
= brw_flag_reg(0, 0);
4591 /* FIXME: For SIMD32 programs, this causes us to stomp on f0.1 as well
4592 * as f0.0. This is a problem for fragment programs as we currently use
4593 * f0.1 for discards. Fortunately, we don't support SIMD32 fragment
4594 * programs yet so this isn't a problem. When we do, something will
4597 if (dispatch_width
== 32)
4598 flag
.type
= BRW_REGISTER_TYPE_UD
;
4600 bld
.exec_all().group(1, 0).MOV(flag
, brw_imm_ud(0u));
4601 bld
.CMP(bld
.null_reg_ud(), value
, brw_imm_ud(0u), BRW_CONDITIONAL_NZ
);
4603 if (instr
->dest
.ssa
.bit_size
> 32) {
4604 dest
.type
= BRW_REGISTER_TYPE_UQ
;
4606 dest
.type
= BRW_REGISTER_TYPE_UD
;
4608 bld
.MOV(dest
, flag
);
4612 case nir_intrinsic_read_invocation
: {
4613 const fs_reg value
= get_nir_src(instr
->src
[0]);
4614 const fs_reg invocation
= get_nir_src(instr
->src
[1]);
4615 fs_reg tmp
= bld
.vgrf(value
.type
);
4617 bld
.exec_all().emit(SHADER_OPCODE_BROADCAST
, tmp
, value
,
4618 bld
.emit_uniformize(invocation
));
4620 bld
.MOV(retype(dest
, value
.type
), fs_reg(component(tmp
, 0)));
4624 case nir_intrinsic_read_first_invocation
: {
4625 const fs_reg value
= get_nir_src(instr
->src
[0]);
4626 bld
.MOV(retype(dest
, value
.type
), bld
.emit_uniformize(value
));
4630 case nir_intrinsic_shuffle
: {
4631 const fs_reg value
= get_nir_src(instr
->src
[0]);
4632 const fs_reg index
= get_nir_src(instr
->src
[1]);
4634 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, index
);
4638 case nir_intrinsic_first_invocation
: {
4639 fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UD
);
4640 bld
.exec_all().emit(SHADER_OPCODE_FIND_LIVE_CHANNEL
, tmp
);
4641 bld
.MOV(retype(dest
, BRW_REGISTER_TYPE_UD
),
4642 fs_reg(component(tmp
, 0)));
4646 case nir_intrinsic_quad_broadcast
: {
4647 const fs_reg value
= get_nir_src(instr
->src
[0]);
4648 const unsigned index
= nir_src_as_uint(instr
->src
[1]);
4650 bld
.emit(SHADER_OPCODE_CLUSTER_BROADCAST
, retype(dest
, value
.type
),
4651 value
, brw_imm_ud(index
), brw_imm_ud(4));
4655 case nir_intrinsic_quad_swap_horizontal
: {
4656 const fs_reg value
= get_nir_src(instr
->src
[0]);
4657 const fs_reg tmp
= bld
.vgrf(value
.type
);
4658 const fs_builder ubld
= bld
.exec_all().group(dispatch_width
/ 2, 0);
4660 const fs_reg src_left
= horiz_stride(value
, 2);
4661 const fs_reg src_right
= horiz_stride(horiz_offset(value
, 1), 2);
4662 const fs_reg tmp_left
= horiz_stride(tmp
, 2);
4663 const fs_reg tmp_right
= horiz_stride(horiz_offset(tmp
, 1), 2);
4665 ubld
.MOV(tmp_left
, src_right
);
4666 ubld
.MOV(tmp_right
, src_left
);
4668 bld
.MOV(retype(dest
, value
.type
), tmp
);
4672 case nir_intrinsic_quad_swap_vertical
: {
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(2,3,0,1)));
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_quad_swap_diagonal
: {
4694 const fs_reg value
= get_nir_src(instr
->src
[0]);
4695 if (nir_src_bit_size(instr
->src
[0]) == 32) {
4696 /* For 32-bit, we can use a SIMD4x2 instruction to do this easily */
4697 const fs_reg tmp
= bld
.vgrf(value
.type
);
4698 const fs_builder ubld
= bld
.exec_all();
4699 ubld
.emit(SHADER_OPCODE_QUAD_SWIZZLE
, tmp
, value
,
4700 brw_imm_ud(BRW_SWIZZLE4(3,2,1,0)));
4701 bld
.MOV(retype(dest
, value
.type
), tmp
);
4703 /* For larger data types, we have to either emit dispatch_width many
4704 * MOVs or else fall back to doing indirects.
4706 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
4707 bld
.XOR(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
4709 bld
.emit(SHADER_OPCODE_SHUFFLE
, retype(dest
, value
.type
), value
, idx
);
4714 case nir_intrinsic_reduce
: {
4715 fs_reg src
= get_nir_src(instr
->src
[0]);
4716 nir_op redop
= (nir_op
)nir_intrinsic_reduction_op(instr
);
4717 unsigned cluster_size
= nir_intrinsic_cluster_size(instr
);
4718 if (cluster_size
== 0 || cluster_size
> dispatch_width
)
4719 cluster_size
= dispatch_width
;
4721 /* Figure out the source type */
4722 src
.type
= brw_type_for_nir_type(devinfo
,
4723 (nir_alu_type
)(nir_op_infos
[redop
].input_types
[0] |
4724 nir_src_bit_size(instr
->src
[0])));
4726 fs_reg identity
= brw_nir_reduction_op_identity(bld
, redop
, src
.type
);
4727 opcode brw_op
= brw_op_for_nir_reduction_op(redop
);
4728 brw_conditional_mod cond_mod
= brw_cond_mod_for_nir_reduction_op(redop
);
4730 /* Set up a register for all of our scratching around and initialize it
4731 * to reduction operation's identity value.
4733 fs_reg scan
= bld
.vgrf(src
.type
);
4734 bld
.exec_all().emit(SHADER_OPCODE_SEL_EXEC
, scan
, src
, identity
);
4736 bld
.emit_scan(brw_op
, scan
, cluster_size
, cond_mod
);
4738 dest
.type
= src
.type
;
4739 if (cluster_size
* type_sz(src
.type
) >= REG_SIZE
* 2) {
4740 /* In this case, CLUSTER_BROADCAST instruction isn't needed because
4741 * the distance between clusters is at least 2 GRFs. In this case,
4742 * we don't need the weird striding of the CLUSTER_BROADCAST
4743 * instruction and can just do regular MOVs.
4745 assert((cluster_size
* type_sz(src
.type
)) % (REG_SIZE
* 2) == 0);
4746 const unsigned groups
=
4747 (dispatch_width
* type_sz(src
.type
)) / (REG_SIZE
* 2);
4748 const unsigned group_size
= dispatch_width
/ groups
;
4749 for (unsigned i
= 0; i
< groups
; i
++) {
4750 const unsigned cluster
= (i
* group_size
) / cluster_size
;
4751 const unsigned comp
= cluster
* cluster_size
+ (cluster_size
- 1);
4752 bld
.group(group_size
, i
).MOV(horiz_offset(dest
, i
* group_size
),
4753 component(scan
, comp
));
4756 bld
.emit(SHADER_OPCODE_CLUSTER_BROADCAST
, dest
, scan
,
4757 brw_imm_ud(cluster_size
- 1), brw_imm_ud(cluster_size
));
4762 case nir_intrinsic_inclusive_scan
:
4763 case nir_intrinsic_exclusive_scan
: {
4764 fs_reg src
= get_nir_src(instr
->src
[0]);
4765 nir_op redop
= (nir_op
)nir_intrinsic_reduction_op(instr
);
4767 /* Figure out the source type */
4768 src
.type
= brw_type_for_nir_type(devinfo
,
4769 (nir_alu_type
)(nir_op_infos
[redop
].input_types
[0] |
4770 nir_src_bit_size(instr
->src
[0])));
4772 fs_reg identity
= brw_nir_reduction_op_identity(bld
, redop
, src
.type
);
4773 opcode brw_op
= brw_op_for_nir_reduction_op(redop
);
4774 brw_conditional_mod cond_mod
= brw_cond_mod_for_nir_reduction_op(redop
);
4776 /* Set up a register for all of our scratching around and initialize it
4777 * to reduction operation's identity value.
4779 fs_reg scan
= bld
.vgrf(src
.type
);
4780 const fs_builder allbld
= bld
.exec_all();
4781 allbld
.emit(SHADER_OPCODE_SEL_EXEC
, scan
, src
, identity
);
4783 if (instr
->intrinsic
== nir_intrinsic_exclusive_scan
) {
4784 /* Exclusive scan is a bit harder because we have to do an annoying
4785 * shift of the contents before we can begin. To make things worse,
4786 * we can't do this with a normal stride; we have to use indirects.
4788 fs_reg shifted
= bld
.vgrf(src
.type
);
4789 fs_reg idx
= bld
.vgrf(BRW_REGISTER_TYPE_W
);
4790 allbld
.ADD(idx
, nir_system_values
[SYSTEM_VALUE_SUBGROUP_INVOCATION
],
4792 allbld
.emit(SHADER_OPCODE_SHUFFLE
, shifted
, scan
, idx
);
4793 allbld
.group(1, 0).MOV(component(shifted
, 0), identity
);
4797 bld
.emit_scan(brw_op
, scan
, dispatch_width
, cond_mod
);
4799 bld
.MOV(retype(dest
, src
.type
), scan
);
4803 case nir_intrinsic_begin_invocation_interlock
: {
4804 const fs_builder ubld
= bld
.group(8, 0);
4805 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
4807 ubld
.emit(SHADER_OPCODE_INTERLOCK
, tmp
)->size_written
= 2 *
4813 case nir_intrinsic_end_invocation_interlock
: {
4814 /* We don't need to do anything here */
4819 unreachable("unknown intrinsic");
4824 fs_visitor::nir_emit_ssbo_atomic(const fs_builder
&bld
,
4825 int op
, nir_intrinsic_instr
*instr
)
4827 if (stage
== MESA_SHADER_FRAGMENT
)
4828 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4831 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4832 dest
= get_nir_dest(instr
->dest
);
4834 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4835 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = get_nir_ssbo_intrinsic_index(bld
, instr
);
4836 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
4837 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4838 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
4841 if (op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
)
4842 data
= get_nir_src(instr
->src
[2]);
4844 if (op
== BRW_AOP_CMPWR
) {
4845 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
4846 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[3]) };
4847 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
4850 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
4852 /* Emit the actual atomic operation */
4854 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
,
4855 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4859 fs_visitor::nir_emit_ssbo_atomic_float(const fs_builder
&bld
,
4860 int op
, nir_intrinsic_instr
*instr
)
4862 if (stage
== MESA_SHADER_FRAGMENT
)
4863 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4866 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4867 dest
= get_nir_dest(instr
->dest
);
4869 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4870 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = get_nir_ssbo_intrinsic_index(bld
, instr
);
4871 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = get_nir_src(instr
->src
[1]);
4872 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4873 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
4875 fs_reg data
= get_nir_src(instr
->src
[2]);
4876 if (op
== BRW_AOP_FCMPWR
) {
4877 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
4878 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[3]) };
4879 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
4882 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
4884 /* Emit the actual atomic operation */
4886 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
,
4887 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4891 fs_visitor::nir_emit_shared_atomic(const fs_builder
&bld
,
4892 int op
, nir_intrinsic_instr
*instr
)
4895 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4896 dest
= get_nir_dest(instr
->dest
);
4898 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4899 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
4900 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4901 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
4904 if (op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
)
4905 data
= get_nir_src(instr
->src
[1]);
4906 if (op
== BRW_AOP_CMPWR
) {
4907 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
4908 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
4909 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
4912 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
4914 /* Get the offset */
4915 if (nir_src_is_const(instr
->src
[0])) {
4916 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] =
4917 brw_imm_ud(instr
->const_index
[0] + nir_src_as_uint(instr
->src
[0]));
4919 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = vgrf(glsl_type::uint_type
);
4920 bld
.ADD(srcs
[SURFACE_LOGICAL_SRC_ADDRESS
],
4921 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
4922 brw_imm_ud(instr
->const_index
[0]));
4925 /* Emit the actual atomic operation operation */
4927 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_LOGICAL
,
4928 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4932 fs_visitor::nir_emit_shared_atomic_float(const fs_builder
&bld
,
4933 int op
, nir_intrinsic_instr
*instr
)
4936 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4937 dest
= get_nir_dest(instr
->dest
);
4939 fs_reg srcs
[SURFACE_LOGICAL_NUM_SRCS
];
4940 srcs
[SURFACE_LOGICAL_SRC_SURFACE
] = brw_imm_ud(GEN7_BTI_SLM
);
4941 srcs
[SURFACE_LOGICAL_SRC_IMM_DIMS
] = brw_imm_ud(1);
4942 srcs
[SURFACE_LOGICAL_SRC_IMM_ARG
] = brw_imm_ud(op
);
4944 fs_reg data
= get_nir_src(instr
->src
[1]);
4945 if (op
== BRW_AOP_FCMPWR
) {
4946 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
4947 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
4948 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
4951 srcs
[SURFACE_LOGICAL_SRC_DATA
] = data
;
4953 /* Get the offset */
4954 if (nir_src_is_const(instr
->src
[0])) {
4955 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] =
4956 brw_imm_ud(instr
->const_index
[0] + nir_src_as_uint(instr
->src
[0]));
4958 srcs
[SURFACE_LOGICAL_SRC_ADDRESS
] = vgrf(glsl_type::uint_type
);
4959 bld
.ADD(srcs
[SURFACE_LOGICAL_SRC_ADDRESS
],
4960 retype(get_nir_src(instr
->src
[0]), BRW_REGISTER_TYPE_UD
),
4961 brw_imm_ud(instr
->const_index
[0]));
4964 /* Emit the actual atomic operation operation */
4966 bld
.emit(SHADER_OPCODE_UNTYPED_ATOMIC_FLOAT_LOGICAL
,
4967 dest
, srcs
, SURFACE_LOGICAL_NUM_SRCS
);
4971 fs_visitor::nir_emit_global_atomic(const fs_builder
&bld
,
4972 int op
, nir_intrinsic_instr
*instr
)
4974 if (stage
== MESA_SHADER_FRAGMENT
)
4975 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
4978 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
4979 dest
= get_nir_dest(instr
->dest
);
4981 fs_reg addr
= get_nir_src(instr
->src
[0]);
4984 if (op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
)
4985 data
= get_nir_src(instr
->src
[1]);
4987 if (op
== BRW_AOP_CMPWR
) {
4988 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
4989 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
4990 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
4994 bld
.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
,
4995 dest
, addr
, data
, brw_imm_ud(op
));
4999 fs_visitor::nir_emit_global_atomic_float(const fs_builder
&bld
,
5000 int op
, nir_intrinsic_instr
*instr
)
5002 if (stage
== MESA_SHADER_FRAGMENT
)
5003 brw_wm_prog_data(prog_data
)->has_side_effects
= true;
5005 assert(nir_intrinsic_infos
[instr
->intrinsic
].has_dest
);
5006 fs_reg dest
= get_nir_dest(instr
->dest
);
5008 fs_reg addr
= get_nir_src(instr
->src
[0]);
5010 assert(op
!= BRW_AOP_INC
&& op
!= BRW_AOP_DEC
&& op
!= BRW_AOP_PREDEC
);
5011 fs_reg data
= get_nir_src(instr
->src
[1]);
5013 if (op
== BRW_AOP_FCMPWR
) {
5014 fs_reg tmp
= bld
.vgrf(data
.type
, 2);
5015 fs_reg sources
[2] = { data
, get_nir_src(instr
->src
[2]) };
5016 bld
.LOAD_PAYLOAD(tmp
, sources
, 2, 0);
5020 bld
.emit(SHADER_OPCODE_A64_UNTYPED_ATOMIC_LOGICAL
,
5021 dest
, addr
, data
, brw_imm_ud(op
));
5025 fs_visitor::nir_emit_texture(const fs_builder
&bld
, nir_tex_instr
*instr
)
5027 unsigned texture
= instr
->texture_index
;
5028 unsigned sampler
= instr
->sampler_index
;
5030 fs_reg srcs
[TEX_LOGICAL_NUM_SRCS
];
5032 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture
);
5033 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = brw_imm_ud(sampler
);
5035 int lod_components
= 0;
5037 /* The hardware requires a LOD for buffer textures */
5038 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_BUF
)
5039 srcs
[TEX_LOGICAL_SRC_LOD
] = brw_imm_d(0);
5041 uint32_t header_bits
= 0;
5042 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
5043 fs_reg src
= get_nir_src(instr
->src
[i
].src
);
5044 switch (instr
->src
[i
].src_type
) {
5045 case nir_tex_src_bias
:
5046 srcs
[TEX_LOGICAL_SRC_LOD
] =
5047 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
5049 case nir_tex_src_comparator
:
5050 srcs
[TEX_LOGICAL_SRC_SHADOW_C
] = retype(src
, BRW_REGISTER_TYPE_F
);
5052 case nir_tex_src_coord
:
5053 switch (instr
->op
) {
5055 case nir_texop_txf_ms
:
5056 case nir_texop_txf_ms_mcs
:
5057 case nir_texop_samples_identical
:
5058 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_D
);
5061 srcs
[TEX_LOGICAL_SRC_COORDINATE
] = retype(src
, BRW_REGISTER_TYPE_F
);
5065 case nir_tex_src_ddx
:
5066 srcs
[TEX_LOGICAL_SRC_LOD
] = retype(src
, BRW_REGISTER_TYPE_F
);
5067 lod_components
= nir_tex_instr_src_size(instr
, i
);
5069 case nir_tex_src_ddy
:
5070 srcs
[TEX_LOGICAL_SRC_LOD2
] = retype(src
, BRW_REGISTER_TYPE_F
);
5072 case nir_tex_src_lod
:
5073 switch (instr
->op
) {
5075 srcs
[TEX_LOGICAL_SRC_LOD
] =
5076 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_UD
);
5079 srcs
[TEX_LOGICAL_SRC_LOD
] =
5080 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_D
);
5083 srcs
[TEX_LOGICAL_SRC_LOD
] =
5084 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
5088 case nir_tex_src_min_lod
:
5089 srcs
[TEX_LOGICAL_SRC_MIN_LOD
] =
5090 retype(get_nir_src_imm(instr
->src
[i
].src
), BRW_REGISTER_TYPE_F
);
5092 case nir_tex_src_ms_index
:
5093 srcs
[TEX_LOGICAL_SRC_SAMPLE_INDEX
] = retype(src
, BRW_REGISTER_TYPE_UD
);
5096 case nir_tex_src_offset
: {
5097 uint32_t offset_bits
= 0;
5098 if (brw_texture_offset(instr
, i
, &offset_bits
)) {
5099 header_bits
|= offset_bits
;
5101 srcs
[TEX_LOGICAL_SRC_TG4_OFFSET
] =
5102 retype(src
, BRW_REGISTER_TYPE_D
);
5107 case nir_tex_src_projector
:
5108 unreachable("should be lowered");
5110 case nir_tex_src_texture_offset
: {
5111 /* Emit code to evaluate the actual indexing expression */
5112 fs_reg tmp
= vgrf(glsl_type::uint_type
);
5113 bld
.ADD(tmp
, src
, brw_imm_ud(texture
));
5114 srcs
[TEX_LOGICAL_SRC_SURFACE
] = bld
.emit_uniformize(tmp
);
5118 case nir_tex_src_sampler_offset
: {
5119 /* Emit code to evaluate the actual indexing expression */
5120 fs_reg tmp
= vgrf(glsl_type::uint_type
);
5121 bld
.ADD(tmp
, src
, brw_imm_ud(sampler
));
5122 srcs
[TEX_LOGICAL_SRC_SAMPLER
] = bld
.emit_uniformize(tmp
);
5126 case nir_tex_src_ms_mcs
:
5127 assert(instr
->op
== nir_texop_txf_ms
);
5128 srcs
[TEX_LOGICAL_SRC_MCS
] = retype(src
, BRW_REGISTER_TYPE_D
);
5131 case nir_tex_src_plane
: {
5132 const uint32_t plane
= nir_src_as_uint(instr
->src
[i
].src
);
5133 const uint32_t texture_index
=
5134 instr
->texture_index
+
5135 stage_prog_data
->binding_table
.plane_start
[plane
] -
5136 stage_prog_data
->binding_table
.texture_start
;
5138 srcs
[TEX_LOGICAL_SRC_SURFACE
] = brw_imm_ud(texture_index
);
5143 unreachable("unknown texture source");
5147 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BAD_FILE
&&
5148 (instr
->op
== nir_texop_txf_ms
||
5149 instr
->op
== nir_texop_samples_identical
)) {
5150 if (devinfo
->gen
>= 7 &&
5151 key_tex
->compressed_multisample_layout_mask
& (1 << texture
)) {
5152 srcs
[TEX_LOGICAL_SRC_MCS
] =
5153 emit_mcs_fetch(srcs
[TEX_LOGICAL_SRC_COORDINATE
],
5154 instr
->coord_components
,
5155 srcs
[TEX_LOGICAL_SRC_SURFACE
]);
5157 srcs
[TEX_LOGICAL_SRC_MCS
] = brw_imm_ud(0u);
5161 srcs
[TEX_LOGICAL_SRC_COORD_COMPONENTS
] = brw_imm_d(instr
->coord_components
);
5162 srcs
[TEX_LOGICAL_SRC_GRAD_COMPONENTS
] = brw_imm_d(lod_components
);
5164 bool shader_supports_implicit_lod
= stage
== MESA_SHADER_FRAGMENT
||
5165 (stage
== MESA_SHADER_COMPUTE
&&
5166 nir
->info
.cs
.derivative_group
!= DERIVATIVE_GROUP_NONE
);
5169 switch (instr
->op
) {
5171 opcode
= shader_supports_implicit_lod
?
5172 SHADER_OPCODE_TEX_LOGICAL
: SHADER_OPCODE_TXL_LOGICAL
;
5175 opcode
= FS_OPCODE_TXB_LOGICAL
;
5178 opcode
= SHADER_OPCODE_TXL_LOGICAL
;
5181 opcode
= SHADER_OPCODE_TXD_LOGICAL
;
5184 opcode
= SHADER_OPCODE_TXF_LOGICAL
;
5186 case nir_texop_txf_ms
:
5187 if ((key_tex
->msaa_16
& (1 << sampler
)))
5188 opcode
= SHADER_OPCODE_TXF_CMS_W_LOGICAL
;
5190 opcode
= SHADER_OPCODE_TXF_CMS_LOGICAL
;
5192 case nir_texop_txf_ms_mcs
:
5193 opcode
= SHADER_OPCODE_TXF_MCS_LOGICAL
;
5195 case nir_texop_query_levels
:
5197 opcode
= SHADER_OPCODE_TXS_LOGICAL
;
5200 opcode
= SHADER_OPCODE_LOD_LOGICAL
;
5203 if (srcs
[TEX_LOGICAL_SRC_TG4_OFFSET
].file
!= BAD_FILE
)
5204 opcode
= SHADER_OPCODE_TG4_OFFSET_LOGICAL
;
5206 opcode
= SHADER_OPCODE_TG4_LOGICAL
;
5208 case nir_texop_texture_samples
:
5209 opcode
= SHADER_OPCODE_SAMPLEINFO_LOGICAL
;
5211 case nir_texop_samples_identical
: {
5212 fs_reg dst
= retype(get_nir_dest(instr
->dest
), BRW_REGISTER_TYPE_D
);
5214 /* If mcs is an immediate value, it means there is no MCS. In that case
5215 * just return false.
5217 if (srcs
[TEX_LOGICAL_SRC_MCS
].file
== BRW_IMMEDIATE_VALUE
) {
5218 bld
.MOV(dst
, brw_imm_ud(0u));
5219 } else if ((key_tex
->msaa_16
& (1 << sampler
))) {
5220 fs_reg tmp
= vgrf(glsl_type::uint_type
);
5221 bld
.OR(tmp
, srcs
[TEX_LOGICAL_SRC_MCS
],
5222 offset(srcs
[TEX_LOGICAL_SRC_MCS
], bld
, 1));
5223 bld
.CMP(dst
, tmp
, brw_imm_ud(0u), BRW_CONDITIONAL_EQ
);
5225 bld
.CMP(dst
, srcs
[TEX_LOGICAL_SRC_MCS
], brw_imm_ud(0u),
5226 BRW_CONDITIONAL_EQ
);
5231 unreachable("unknown texture opcode");
5234 if (instr
->op
== nir_texop_tg4
) {
5235 if (instr
->component
== 1 &&
5236 key_tex
->gather_channel_quirk_mask
& (1 << texture
)) {
5237 /* gather4 sampler is broken for green channel on RG32F --
5238 * we must ask for blue instead.
5240 header_bits
|= 2 << 16;
5242 header_bits
|= instr
->component
<< 16;
5246 fs_reg dst
= bld
.vgrf(brw_type_for_nir_type(devinfo
, instr
->dest_type
), 4);
5247 fs_inst
*inst
= bld
.emit(opcode
, dst
, srcs
, ARRAY_SIZE(srcs
));
5248 inst
->offset
= header_bits
;
5250 const unsigned dest_size
= nir_tex_instr_dest_size(instr
);
5251 if (devinfo
->gen
>= 9 &&
5252 instr
->op
!= nir_texop_tg4
&& instr
->op
!= nir_texop_query_levels
) {
5253 unsigned write_mask
= instr
->dest
.is_ssa
?
5254 nir_ssa_def_components_read(&instr
->dest
.ssa
):
5255 (1 << dest_size
) - 1;
5256 assert(write_mask
!= 0); /* dead code should have been eliminated */
5257 inst
->size_written
= util_last_bit(write_mask
) *
5258 inst
->dst
.component_size(inst
->exec_size
);
5260 inst
->size_written
= 4 * inst
->dst
.component_size(inst
->exec_size
);
5263 if (srcs
[TEX_LOGICAL_SRC_SHADOW_C
].file
!= BAD_FILE
)
5264 inst
->shadow_compare
= true;
5266 if (instr
->op
== nir_texop_tg4
&& devinfo
->gen
== 6)
5267 emit_gen6_gather_wa(key_tex
->gen6_gather_wa
[texture
], dst
);
5270 for (unsigned i
= 0; i
< dest_size
; i
++)
5271 nir_dest
[i
] = offset(dst
, bld
, i
);
5273 if (instr
->op
== nir_texop_query_levels
) {
5274 /* # levels is in .w */
5275 nir_dest
[0] = offset(dst
, bld
, 3);
5276 } else if (instr
->op
== nir_texop_txs
&&
5277 dest_size
>= 3 && devinfo
->gen
< 7) {
5278 /* Gen4-6 return 0 instead of 1 for single layer surfaces. */
5279 fs_reg depth
= offset(dst
, bld
, 2);
5280 nir_dest
[2] = vgrf(glsl_type::int_type
);
5281 bld
.emit_minmax(nir_dest
[2], depth
, brw_imm_d(1), BRW_CONDITIONAL_GE
);
5284 bld
.LOAD_PAYLOAD(get_nir_dest(instr
->dest
), nir_dest
, dest_size
, 0);
5288 fs_visitor::nir_emit_jump(const fs_builder
&bld
, nir_jump_instr
*instr
)
5290 switch (instr
->type
) {
5291 case nir_jump_break
:
5292 bld
.emit(BRW_OPCODE_BREAK
);
5294 case nir_jump_continue
:
5295 bld
.emit(BRW_OPCODE_CONTINUE
);
5297 case nir_jump_return
:
5299 unreachable("unknown jump");
5304 * This helper takes a source register and un/shuffles it into the destination
5307 * If source type size is smaller than destination type size the operation
5308 * needed is a component shuffle. The opposite case would be an unshuffle. If
5309 * source/destination type size is equal a shuffle is done that would be
5310 * equivalent to a simple MOV.
5312 * For example, if source is a 16-bit type and destination is 32-bit. A 3
5313 * components .xyz 16-bit vector on SIMD8 would be.
5315 * |x1|x2|x3|x4|x5|x6|x7|x8|y1|y2|y3|y4|y5|y6|y7|y8|
5316 * |z1|z2|z3|z4|z5|z6|z7|z8| | | | | | | | |
5318 * This helper will return the following 2 32-bit components with the 16-bit
5321 * |x1 y1|x2 y2|x3 y3|x4 y4|x5 y5|x6 y6|x7 y7|x8 y8|
5322 * |z1 |z2 |z3 |z4 |z5 |z6 |z7 |z8 |
5324 * For unshuffle, the example would be the opposite, a 64-bit type source
5325 * and a 32-bit destination. A 2 component .xy 64-bit vector on SIMD8
5328 * | x1l x1h | x2l x2h | x3l x3h | x4l x4h |
5329 * | x5l x5h | x6l x6h | x7l x7h | x8l x8h |
5330 * | y1l y1h | y2l y2h | y3l y3h | y4l y4h |
5331 * | y5l y5h | y6l y6h | y7l y7h | y8l y8h |
5333 * The returned result would be the following 4 32-bit components unshuffled:
5335 * | x1l | x2l | x3l | x4l | x5l | x6l | x7l | x8l |
5336 * | x1h | x2h | x3h | x4h | x5h | x6h | x7h | x8h |
5337 * | y1l | y2l | y3l | y4l | y5l | y6l | y7l | y8l |
5338 * | y1h | y2h | y3h | y4h | y5h | y6h | y7h | y8h |
5340 * - Source and destination register must not be overlapped.
5341 * - components units are measured in terms of the smaller type between
5342 * source and destination because we are un/shuffling the smaller
5343 * components from/into the bigger ones.
5344 * - first_component parameter allows skipping source components.
5347 shuffle_src_to_dst(const fs_builder
&bld
,
5350 uint32_t first_component
,
5351 uint32_t components
)
5353 if (type_sz(src
.type
) == type_sz(dst
.type
)) {
5354 assert(!regions_overlap(dst
,
5355 type_sz(dst
.type
) * bld
.dispatch_width() * components
,
5356 offset(src
, bld
, first_component
),
5357 type_sz(src
.type
) * bld
.dispatch_width() * components
));
5358 for (unsigned i
= 0; i
< components
; i
++) {
5359 bld
.MOV(retype(offset(dst
, bld
, i
), src
.type
),
5360 offset(src
, bld
, i
+ first_component
));
5362 } else if (type_sz(src
.type
) < type_sz(dst
.type
)) {
5363 /* Source is shuffled into destination */
5364 unsigned size_ratio
= type_sz(dst
.type
) / type_sz(src
.type
);
5365 assert(!regions_overlap(dst
,
5366 type_sz(dst
.type
) * bld
.dispatch_width() *
5367 DIV_ROUND_UP(components
, size_ratio
),
5368 offset(src
, bld
, first_component
),
5369 type_sz(src
.type
) * bld
.dispatch_width() * components
));
5371 brw_reg_type shuffle_type
=
5372 brw_reg_type_from_bit_size(8 * type_sz(src
.type
),
5373 BRW_REGISTER_TYPE_D
);
5374 for (unsigned i
= 0; i
< components
; i
++) {
5375 fs_reg shuffle_component_i
=
5376 subscript(offset(dst
, bld
, i
/ size_ratio
),
5377 shuffle_type
, i
% size_ratio
);
5378 bld
.MOV(shuffle_component_i
,
5379 retype(offset(src
, bld
, i
+ first_component
), shuffle_type
));
5382 /* Source is unshuffled into destination */
5383 unsigned size_ratio
= type_sz(src
.type
) / type_sz(dst
.type
);
5384 assert(!regions_overlap(dst
,
5385 type_sz(dst
.type
) * bld
.dispatch_width() * components
,
5386 offset(src
, bld
, first_component
/ size_ratio
),
5387 type_sz(src
.type
) * bld
.dispatch_width() *
5388 DIV_ROUND_UP(components
+ (first_component
% size_ratio
),
5391 brw_reg_type shuffle_type
=
5392 brw_reg_type_from_bit_size(8 * type_sz(dst
.type
),
5393 BRW_REGISTER_TYPE_D
);
5394 for (unsigned i
= 0; i
< components
; i
++) {
5395 fs_reg shuffle_component_i
=
5396 subscript(offset(src
, bld
, (first_component
+ i
) / size_ratio
),
5397 shuffle_type
, (first_component
+ i
) % size_ratio
);
5398 bld
.MOV(retype(offset(dst
, bld
, i
), shuffle_type
),
5399 shuffle_component_i
);
5405 shuffle_from_32bit_read(const fs_builder
&bld
,
5408 uint32_t first_component
,
5409 uint32_t components
)
5411 assert(type_sz(src
.type
) == 4);
5413 /* This function takes components in units of the destination type while
5414 * shuffle_src_to_dst takes components in units of the smallest type
5416 if (type_sz(dst
.type
) > 4) {
5417 assert(type_sz(dst
.type
) == 8);
5418 first_component
*= 2;
5422 shuffle_src_to_dst(bld
, dst
, src
, first_component
, components
);
5426 shuffle_for_32bit_write(const fs_builder
&bld
,
5428 uint32_t first_component
,
5429 uint32_t components
)
5431 fs_reg dst
= bld
.vgrf(BRW_REGISTER_TYPE_D
,
5432 DIV_ROUND_UP (components
* type_sz(src
.type
), 4));
5433 /* This function takes components in units of the source type while
5434 * shuffle_src_to_dst takes components in units of the smallest type
5436 if (type_sz(src
.type
) > 4) {
5437 assert(type_sz(src
.type
) == 8);
5438 first_component
*= 2;
5442 shuffle_src_to_dst(bld
, dst
, src
, first_component
, components
);
5448 setup_imm_df(const fs_builder
&bld
, double v
)
5450 const struct gen_device_info
*devinfo
= bld
.shader
->devinfo
;
5451 assert(devinfo
->gen
>= 7);
5453 if (devinfo
->gen
>= 8)
5454 return brw_imm_df(v
);
5456 /* gen7.5 does not support DF immediates straighforward but the DIM
5457 * instruction allows to set the 64-bit immediate value.
5459 if (devinfo
->is_haswell
) {
5460 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5461 fs_reg dst
= ubld
.vgrf(BRW_REGISTER_TYPE_DF
, 1);
5462 ubld
.DIM(dst
, brw_imm_df(v
));
5463 return component(dst
, 0);
5466 /* gen7 does not support DF immediates, so we generate a 64-bit constant by
5467 * writing the low 32-bit of the constant to suboffset 0 of a VGRF and
5468 * the high 32-bit to suboffset 4 and then applying a stride of 0.
5470 * Alternatively, we could also produce a normal VGRF (without stride 0)
5471 * by writing to all the channels in the VGRF, however, that would hit the
5472 * gen7 bug where we have to split writes that span more than 1 register
5473 * into instructions with a width of 4 (otherwise the write to the second
5474 * register written runs into an execmask hardware bug) which isn't very
5487 const fs_builder ubld
= bld
.exec_all().group(1, 0);
5488 const fs_reg tmp
= ubld
.vgrf(BRW_REGISTER_TYPE_UD
, 2);
5489 ubld
.MOV(tmp
, brw_imm_ud(di
.i1
));
5490 ubld
.MOV(horiz_offset(tmp
, 1), brw_imm_ud(di
.i2
));
5492 return component(retype(tmp
, BRW_REGISTER_TYPE_DF
), 0);
5496 setup_imm_b(const fs_builder
&bld
, int8_t v
)
5498 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_B
);
5499 bld
.MOV(tmp
, brw_imm_w(v
));
5504 setup_imm_ub(const fs_builder
&bld
, uint8_t v
)
5506 const fs_reg tmp
= bld
.vgrf(BRW_REGISTER_TYPE_UB
);
5507 bld
.MOV(tmp
, brw_imm_uw(v
));