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"
25 #include "main/shaderimage.h"
27 #include "brw_fs_surface_builder.h"
29 #include "brw_program.h"
32 using namespace brw::surface_access;
35 fs_visitor::emit_nir_code()
37 /* emit the arrays used for inputs and outputs - load/store intrinsics will
38 * be converted to reads/writes of these arrays
43 nir_emit_system_values();
45 /* get the main function and emit it */
46 nir_foreach_function(nir, function) {
47 assert(strcmp(function->name, "main") == 0);
48 assert(function->impl);
49 nir_emit_impl(function->impl);
54 fs_visitor::nir_setup_inputs()
56 if (stage != MESA_SHADER_FRAGMENT)
59 nir_inputs = bld.vgrf(BRW_REGISTER_TYPE_F, nir->num_inputs);
61 nir_foreach_variable(var, &nir->inputs) {
62 fs_reg input = offset(nir_inputs, bld, var->data.driver_location);
65 if (var->data.location == VARYING_SLOT_POS) {
66 reg = *emit_fragcoord_interpolation(var->data.pixel_center_integer,
67 var->data.origin_upper_left);
68 emit_percomp(bld, fs_inst(BRW_OPCODE_MOV, bld.dispatch_width(),
70 } else if (var->data.location == VARYING_SLOT_LAYER) {
71 struct brw_reg reg = suboffset(interp_reg(VARYING_SLOT_LAYER, 1), 3);
72 reg.type = BRW_REGISTER_TYPE_D;
73 bld.emit(FS_OPCODE_CINTERP, retype(input, BRW_REGISTER_TYPE_D), reg);
74 } else if (var->data.location == VARYING_SLOT_VIEWPORT) {
75 struct brw_reg reg = suboffset(interp_reg(VARYING_SLOT_VIEWPORT, 2), 3);
76 reg.type = BRW_REGISTER_TYPE_D;
77 bld.emit(FS_OPCODE_CINTERP, retype(input, BRW_REGISTER_TYPE_D), reg);
79 int location = var->data.location;
80 emit_general_interpolation(&input, var->name, var->type,
81 (glsl_interp_qualifier) var->data.interpolation,
82 &location, var->data.centroid,
89 fs_visitor::nir_setup_single_output_varying(fs_reg *reg,
90 const glsl_type *type,
93 if (type->is_array() || type->is_matrix()) {
94 const struct glsl_type *elem_type = glsl_get_array_element(type);
95 const unsigned length = glsl_get_length(type);
97 for (unsigned i = 0; i < length; i++) {
98 nir_setup_single_output_varying(reg, elem_type, location);
100 } else if (type->is_record()) {
101 for (unsigned i = 0; i < type->length; i++) {
102 const struct glsl_type *field_type = type->fields.structure[i].type;
103 nir_setup_single_output_varying(reg, field_type, location);
106 assert(type->is_scalar() || type->is_vector());
107 this->outputs[*location] = *reg;
108 this->output_components[*location] = type->vector_elements;
109 *reg = offset(*reg, bld, 4);
115 fs_visitor::nir_setup_outputs()
117 brw_wm_prog_key *key = (brw_wm_prog_key*) this->key;
119 nir_outputs = bld.vgrf(BRW_REGISTER_TYPE_F, nir->num_outputs);
121 nir_foreach_variable(var, &nir->outputs) {
122 fs_reg reg = offset(nir_outputs, bld, var->data.driver_location);
125 case MESA_SHADER_VERTEX:
126 case MESA_SHADER_TESS_EVAL:
127 case MESA_SHADER_GEOMETRY: {
128 unsigned location = var->data.location;
129 nir_setup_single_output_varying(®, var->type, &location);
132 case MESA_SHADER_FRAGMENT:
133 if (key->force_dual_color_blend &&
134 var->data.location == FRAG_RESULT_DATA1) {
135 this->dual_src_output = reg;
136 this->do_dual_src = true;
137 } else if (var->data.index > 0) {
138 assert(var->data.location == FRAG_RESULT_DATA0);
139 assert(var->data.index == 1);
140 this->dual_src_output = reg;
141 this->do_dual_src = true;
142 } else if (var->data.location == FRAG_RESULT_COLOR) {
143 /* Writing gl_FragColor outputs to all color regions. */
144 for (unsigned int i = 0; i < MAX2(key->nr_color_regions, 1); i++) {
145 this->outputs[i] = reg;
146 this->output_components[i] = 4;
148 } else if (var->data.location == FRAG_RESULT_DEPTH) {
149 this->frag_depth = reg;
150 } else if (var->data.location == FRAG_RESULT_STENCIL) {
151 this->frag_stencil = reg;
152 } else if (var->data.location == FRAG_RESULT_SAMPLE_MASK) {
153 this->sample_mask = reg;
155 int vector_elements = var->type->without_array()->vector_elements;
157 /* gl_FragData or a user-defined FS output */
158 assert(var->data.location >= FRAG_RESULT_DATA0 &&
159 var->data.location < FRAG_RESULT_DATA0+BRW_MAX_DRAW_BUFFERS);
161 /* General color output. */
162 for (unsigned int i = 0; i < MAX2(1, var->type->length); i++) {
163 int output = var->data.location - FRAG_RESULT_DATA0 + i;
164 this->outputs[output] = offset(reg, bld, vector_elements * i);
165 this->output_components[output] = vector_elements;
170 unreachable("unhandled shader stage");
176 fs_visitor::nir_setup_uniforms()
178 if (dispatch_width != 8)
181 uniforms = nir->num_uniforms / 4;
185 emit_system_values_block(nir_block *block, void *void_visitor)
187 fs_visitor *v = (fs_visitor *)void_visitor;
190 nir_foreach_instr(block, instr) {
191 if (instr->type != nir_instr_type_intrinsic)
194 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
195 switch (intrin->intrinsic) {
196 case nir_intrinsic_load_vertex_id:
197 unreachable("should be lowered by lower_vertex_id().");
199 case nir_intrinsic_load_vertex_id_zero_base:
200 assert(v->stage == MESA_SHADER_VERTEX);
201 reg = &v->nir_system_values[SYSTEM_VALUE_VERTEX_ID_ZERO_BASE];
202 if (reg->file == BAD_FILE)
203 *reg = *v->emit_vs_system_value(SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
206 case nir_intrinsic_load_base_vertex:
207 assert(v->stage == MESA_SHADER_VERTEX);
208 reg = &v->nir_system_values[SYSTEM_VALUE_BASE_VERTEX];
209 if (reg->file == BAD_FILE)
210 *reg = *v->emit_vs_system_value(SYSTEM_VALUE_BASE_VERTEX);
213 case nir_intrinsic_load_instance_id:
214 assert(v->stage == MESA_SHADER_VERTEX);
215 reg = &v->nir_system_values[SYSTEM_VALUE_INSTANCE_ID];
216 if (reg->file == BAD_FILE)
217 *reg = *v->emit_vs_system_value(SYSTEM_VALUE_INSTANCE_ID);
220 case nir_intrinsic_load_base_instance:
221 assert(v->stage == MESA_SHADER_VERTEX);
222 reg = &v->nir_system_values[SYSTEM_VALUE_BASE_INSTANCE];
223 if (reg->file == BAD_FILE)
224 *reg = *v->emit_vs_system_value(SYSTEM_VALUE_BASE_INSTANCE);
227 case nir_intrinsic_load_draw_id:
228 assert(v->stage == MESA_SHADER_VERTEX);
229 reg = &v->nir_system_values[SYSTEM_VALUE_DRAW_ID];
230 if (reg->file == BAD_FILE)
231 *reg = *v->emit_vs_system_value(SYSTEM_VALUE_DRAW_ID);
234 case nir_intrinsic_load_invocation_id:
235 assert(v->stage == MESA_SHADER_GEOMETRY);
236 reg = &v->nir_system_values[SYSTEM_VALUE_INVOCATION_ID];
237 if (reg->file == BAD_FILE) {
238 const fs_builder abld = v->bld.annotate("gl_InvocationID", NULL);
239 fs_reg g1(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD));
240 fs_reg iid = abld.vgrf(BRW_REGISTER_TYPE_UD, 1);
241 abld.SHR(iid, g1, brw_imm_ud(27u));
246 case nir_intrinsic_load_sample_pos:
247 assert(v->stage == MESA_SHADER_FRAGMENT);
248 reg = &v->nir_system_values[SYSTEM_VALUE_SAMPLE_POS];
249 if (reg->file == BAD_FILE)
250 *reg = *v->emit_samplepos_setup();
253 case nir_intrinsic_load_sample_id:
254 assert(v->stage == MESA_SHADER_FRAGMENT);
255 reg = &v->nir_system_values[SYSTEM_VALUE_SAMPLE_ID];
256 if (reg->file == BAD_FILE)
257 *reg = *v->emit_sampleid_setup();
260 case nir_intrinsic_load_sample_mask_in:
261 assert(v->stage == MESA_SHADER_FRAGMENT);
262 assert(v->devinfo->gen >= 7);
263 reg = &v->nir_system_values[SYSTEM_VALUE_SAMPLE_MASK_IN];
264 if (reg->file == BAD_FILE)
265 *reg = fs_reg(retype(brw_vec8_grf(v->payload.sample_mask_in_reg, 0),
266 BRW_REGISTER_TYPE_D));
269 case nir_intrinsic_load_local_invocation_id:
270 assert(v->stage == MESA_SHADER_COMPUTE);
271 reg = &v->nir_system_values[SYSTEM_VALUE_LOCAL_INVOCATION_ID];
272 if (reg->file == BAD_FILE)
273 *reg = *v->emit_cs_local_invocation_id_setup();
276 case nir_intrinsic_load_work_group_id:
277 assert(v->stage == MESA_SHADER_COMPUTE);
278 reg = &v->nir_system_values[SYSTEM_VALUE_WORK_GROUP_ID];
279 if (reg->file == BAD_FILE)
280 *reg = *v->emit_cs_work_group_id_setup();
283 case nir_intrinsic_load_helper_invocation:
284 assert(v->stage == MESA_SHADER_FRAGMENT);
285 reg = &v->nir_system_values[SYSTEM_VALUE_HELPER_INVOCATION];
286 if (reg->file == BAD_FILE) {
287 const fs_builder abld =
288 v->bld.annotate("gl_HelperInvocation", NULL);
290 /* On Gen6+ (gl_HelperInvocation is only exposed on Gen7+) the
291 * pixel mask is in g1.7 of the thread payload.
293 * We move the per-channel pixel enable bit to the low bit of each
294 * channel by shifting the byte containing the pixel mask by the
295 * vector immediate 0x76543210UV.
297 * The region of <1,8,0> reads only 1 byte (the pixel masks for
298 * subspans 0 and 1) in SIMD8 and an additional byte (the pixel
299 * masks for 2 and 3) in SIMD16.
301 fs_reg shifted = abld.vgrf(BRW_REGISTER_TYPE_UW, 1);
303 stride(byte_offset(retype(brw_vec1_grf(1, 0),
304 BRW_REGISTER_TYPE_UB), 28),
306 brw_imm_uv(0x76543210));
308 /* A set bit in the pixel mask means the channel is enabled, but
309 * that is the opposite of gl_HelperInvocation so we need to invert
312 * The negate source-modifier bit of logical instructions on Gen8+
313 * performs 1's complement negation, so we can use that instead of
316 fs_reg inverted = negate(shifted);
317 if (v->devinfo->gen < 8) {
318 inverted = abld.vgrf(BRW_REGISTER_TYPE_UW);
319 abld.NOT(inverted, shifted);
322 /* We then resolve the 0/1 result to 0/~0 boolean values by ANDing
323 * with 1 and negating.
325 fs_reg anded = abld.vgrf(BRW_REGISTER_TYPE_UD, 1);
326 abld.AND(anded, inverted, brw_imm_uw(1));
328 fs_reg dst = abld.vgrf(BRW_REGISTER_TYPE_D, 1);
329 abld.MOV(dst, negate(retype(anded, BRW_REGISTER_TYPE_D)));
343 fs_visitor::nir_emit_system_values()
345 nir_system_values = ralloc_array(mem_ctx, fs_reg, SYSTEM_VALUE_MAX);
346 for (unsigned i = 0; i < SYSTEM_VALUE_MAX; i++) {
347 nir_system_values[i] = fs_reg();
350 nir_foreach_function(nir, function) {
351 assert(strcmp(function->name, "main") == 0);
352 assert(function->impl);
353 nir_foreach_block(function->impl, emit_system_values_block, this);
358 fs_visitor::nir_emit_impl(nir_function_impl *impl)
360 nir_locals = ralloc_array(mem_ctx, fs_reg, impl->reg_alloc);
361 for (unsigned i = 0; i < impl->reg_alloc; i++) {
362 nir_locals[i] = fs_reg();
365 foreach_list_typed(nir_register, reg, node, &impl->registers) {
366 unsigned array_elems =
367 reg->num_array_elems == 0 ? 1 : reg->num_array_elems;
368 unsigned size = array_elems * reg->num_components;
369 nir_locals[reg->index] = bld.vgrf(BRW_REGISTER_TYPE_F, size);
372 nir_ssa_values = reralloc(mem_ctx, nir_ssa_values, fs_reg,
375 nir_emit_cf_list(&impl->body);
379 fs_visitor::nir_emit_cf_list(exec_list *list)
381 exec_list_validate(list);
382 foreach_list_typed(nir_cf_node, node, node, list) {
383 switch (node->type) {
385 nir_emit_if(nir_cf_node_as_if(node));
388 case nir_cf_node_loop:
389 nir_emit_loop(nir_cf_node_as_loop(node));
392 case nir_cf_node_block:
393 nir_emit_block(nir_cf_node_as_block(node));
397 unreachable("Invalid CFG node block");
403 fs_visitor::nir_emit_if(nir_if *if_stmt)
405 /* first, put the condition into f0 */
406 fs_inst *inst = bld.MOV(bld.null_reg_d(),
407 retype(get_nir_src(if_stmt->condition),
408 BRW_REGISTER_TYPE_D));
409 inst->conditional_mod = BRW_CONDITIONAL_NZ;
411 bld.IF(BRW_PREDICATE_NORMAL);
413 nir_emit_cf_list(&if_stmt->then_list);
415 /* note: if the else is empty, dead CF elimination will remove it */
416 bld.emit(BRW_OPCODE_ELSE);
418 nir_emit_cf_list(&if_stmt->else_list);
420 bld.emit(BRW_OPCODE_ENDIF);
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);
434 fs_visitor::nir_emit_block(nir_block *block)
436 nir_foreach_instr(block, instr) {
437 nir_emit_instr(instr);
442 fs_visitor::nir_emit_instr(nir_instr *instr)
444 const fs_builder abld = bld.annotate(NULL, instr);
446 switch (instr->type) {
447 case nir_instr_type_alu:
448 nir_emit_alu(abld, nir_instr_as_alu(instr));
451 case nir_instr_type_intrinsic:
453 case MESA_SHADER_VERTEX:
454 nir_emit_vs_intrinsic(abld, nir_instr_as_intrinsic(instr));
456 case MESA_SHADER_TESS_EVAL:
457 nir_emit_tes_intrinsic(abld, nir_instr_as_intrinsic(instr));
459 case MESA_SHADER_GEOMETRY:
460 nir_emit_gs_intrinsic(abld, nir_instr_as_intrinsic(instr));
462 case MESA_SHADER_FRAGMENT:
463 nir_emit_fs_intrinsic(abld, nir_instr_as_intrinsic(instr));
465 case MESA_SHADER_COMPUTE:
466 nir_emit_cs_intrinsic(abld, nir_instr_as_intrinsic(instr));
469 unreachable("unsupported shader stage");
473 case nir_instr_type_tex:
474 nir_emit_texture(abld, nir_instr_as_tex(instr));
477 case nir_instr_type_load_const:
478 nir_emit_load_const(abld, nir_instr_as_load_const(instr));
481 case nir_instr_type_ssa_undef:
482 nir_emit_undef(abld, nir_instr_as_ssa_undef(instr));
485 case nir_instr_type_jump:
486 nir_emit_jump(abld, nir_instr_as_jump(instr));
490 unreachable("unknown instruction type");
495 * Recognizes a parent instruction of nir_op_extract_* and changes the type to
499 fs_visitor::optimize_extract_to_float(nir_alu_instr *instr,
500 const fs_reg &result)
502 if (!instr->src[0].src.is_ssa ||
503 !instr->src[0].src.ssa->parent_instr)
506 if (instr->src[0].src.ssa->parent_instr->type != nir_instr_type_alu)
509 nir_alu_instr *src0 =
510 nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
512 if (src0->op != nir_op_extract_u8 && src0->op != nir_op_extract_u16 &&
513 src0->op != nir_op_extract_i8 && src0->op != nir_op_extract_i16)
516 nir_const_value *element = nir_src_as_const_value(src0->src[1].src);
517 assert(element != NULL);
519 enum opcode extract_op;
520 if (src0->op == nir_op_extract_u16 || src0->op == nir_op_extract_i16) {
521 assert(element->u[0] <= 1);
522 extract_op = SHADER_OPCODE_EXTRACT_WORD;
524 assert(element->u[0] <= 3);
525 extract_op = SHADER_OPCODE_EXTRACT_BYTE;
528 fs_reg op0 = get_nir_src(src0->src[0].src);
529 op0.type = brw_type_for_nir_type(nir_op_infos[src0->op].input_types[0]);
530 op0 = offset(op0, bld, src0->src[0].swizzle[0]);
532 set_saturate(instr->dest.saturate,
533 bld.emit(extract_op, result, op0, brw_imm_ud(element->u[0])));
538 fs_visitor::optimize_frontfacing_ternary(nir_alu_instr *instr,
539 const fs_reg &result)
541 if (!instr->src[0].src.is_ssa ||
542 instr->src[0].src.ssa->parent_instr->type != nir_instr_type_intrinsic)
545 nir_intrinsic_instr *src0 =
546 nir_instr_as_intrinsic(instr->src[0].src.ssa->parent_instr);
548 if (src0->intrinsic != nir_intrinsic_load_front_face)
551 nir_const_value *value1 = nir_src_as_const_value(instr->src[1].src);
552 if (!value1 || fabsf(value1->f[0]) != 1.0f)
555 nir_const_value *value2 = nir_src_as_const_value(instr->src[2].src);
556 if (!value2 || fabsf(value2->f[0]) != 1.0f)
559 fs_reg tmp = vgrf(glsl_type::int_type);
561 if (devinfo->gen >= 6) {
562 /* Bit 15 of g0.0 is 0 if the polygon is front facing. */
563 fs_reg g0 = fs_reg(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_W));
565 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
567 * or(8) tmp.1<2>W g0.0<0,1,0>W 0x00003f80W
568 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
570 * and negate g0.0<0,1,0>W for (gl_FrontFacing ? -1.0 : 1.0).
572 * This negation looks like it's safe in practice, because bits 0:4 will
573 * surely be TRIANGLES
576 if (value1->f[0] == -1.0f) {
580 tmp.type = BRW_REGISTER_TYPE_W;
581 tmp.subreg_offset = 2;
584 bld.OR(tmp, g0, brw_imm_uw(0x3f80));
586 tmp.type = BRW_REGISTER_TYPE_D;
587 tmp.subreg_offset = 0;
590 /* Bit 31 of g1.6 is 0 if the polygon is front facing. */
591 fs_reg g1_6 = fs_reg(retype(brw_vec1_grf(1, 6), BRW_REGISTER_TYPE_D));
593 /* For (gl_FrontFacing ? 1.0 : -1.0), emit:
595 * or(8) tmp<1>D g1.6<0,1,0>D 0x3f800000D
596 * and(8) dst<1>D tmp<8,8,1>D 0xbf800000D
598 * and negate g1.6<0,1,0>D for (gl_FrontFacing ? -1.0 : 1.0).
600 * This negation looks like it's safe in practice, because bits 0:4 will
601 * surely be TRIANGLES
604 if (value1->f[0] == -1.0f) {
608 bld.OR(tmp, g1_6, brw_imm_d(0x3f800000));
610 bld.AND(retype(result, BRW_REGISTER_TYPE_D), tmp, brw_imm_d(0xbf800000));
616 fs_visitor::nir_emit_alu(const fs_builder &bld, nir_alu_instr *instr)
618 struct brw_wm_prog_key *fs_key = (struct brw_wm_prog_key *) this->key;
621 fs_reg result = get_nir_dest(instr->dest.dest);
622 result.type = brw_type_for_nir_type(nir_op_infos[instr->op].output_type);
625 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
626 op[i] = get_nir_src(instr->src[i].src);
627 op[i].type = brw_type_for_nir_type(nir_op_infos[instr->op].input_types[i]);
628 op[i].abs = instr->src[i].abs;
629 op[i].negate = instr->src[i].negate;
632 /* We get a bunch of mov's out of the from_ssa pass and they may still
633 * be vectorized. We'll handle them as a special-case. We'll also
634 * handle vecN here because it's basically the same thing.
642 fs_reg temp = result;
643 bool need_extra_copy = false;
644 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
645 if (!instr->src[i].src.is_ssa &&
646 instr->dest.dest.reg.reg == instr->src[i].src.reg.reg) {
647 need_extra_copy = true;
648 temp = bld.vgrf(result.type, 4);
653 for (unsigned i = 0; i < 4; i++) {
654 if (!(instr->dest.write_mask & (1 << i)))
657 if (instr->op == nir_op_imov || instr->op == nir_op_fmov) {
658 inst = bld.MOV(offset(temp, bld, i),
659 offset(op[0], bld, instr->src[0].swizzle[i]));
661 inst = bld.MOV(offset(temp, bld, i),
662 offset(op[i], bld, instr->src[i].swizzle[0]));
664 inst->saturate = instr->dest.saturate;
667 /* In this case the source and destination registers were the same,
668 * so we need to insert an extra set of moves in order to deal with
671 if (need_extra_copy) {
672 for (unsigned i = 0; i < 4; i++) {
673 if (!(instr->dest.write_mask & (1 << i)))
676 bld.MOV(offset(result, bld, i), offset(temp, bld, i));
685 /* At this point, we have dealt with any instruction that operates on
686 * more than a single channel. Therefore, we can just adjust the source
687 * and destination registers for that channel and emit the instruction.
689 unsigned channel = 0;
690 if (nir_op_infos[instr->op].output_size == 0) {
691 /* Since NIR is doing the scalarizing for us, we should only ever see
692 * vectorized operations with a single channel.
694 assert(_mesa_bitcount(instr->dest.write_mask) == 1);
695 channel = ffs(instr->dest.write_mask) - 1;
697 result = offset(result, bld, channel);
700 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
701 assert(nir_op_infos[instr->op].input_sizes[i] < 2);
702 op[i] = offset(op[i], bld, instr->src[i].swizzle[channel]);
708 if (optimize_extract_to_float(instr, result))
711 inst = bld.MOV(result, op[0]);
712 inst->saturate = instr->dest.saturate;
717 bld.MOV(result, op[0]);
721 /* AND(val, 0x80000000) gives the sign bit.
723 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
726 bld.CMP(bld.null_reg_f(), op[0], brw_imm_f(0.0f), BRW_CONDITIONAL_NZ);
728 fs_reg result_int = retype(result, BRW_REGISTER_TYPE_UD);
729 op[0].type = BRW_REGISTER_TYPE_UD;
730 result.type = BRW_REGISTER_TYPE_UD;
731 bld.AND(result_int, op[0], brw_imm_ud(0x80000000u));
733 inst = bld.OR(result_int, result_int, brw_imm_ud(0x3f800000u));
734 inst->predicate = BRW_PREDICATE_NORMAL;
735 if (instr->dest.saturate) {
736 inst = bld.MOV(result, result);
737 inst->saturate = true;
743 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
744 * -> non-negative val generates 0x00000000.
745 * Predicated OR sets 1 if val is positive.
747 bld.CMP(bld.null_reg_d(), op[0], brw_imm_d(0), BRW_CONDITIONAL_G);
748 bld.ASR(result, op[0], brw_imm_d(31));
749 inst = bld.OR(result, result, brw_imm_d(1));
750 inst->predicate = BRW_PREDICATE_NORMAL;
754 inst = bld.emit(SHADER_OPCODE_RCP, result, op[0]);
755 inst->saturate = instr->dest.saturate;
759 inst = bld.emit(SHADER_OPCODE_EXP2, result, op[0]);
760 inst->saturate = instr->dest.saturate;
764 inst = bld.emit(SHADER_OPCODE_LOG2, result, op[0]);
765 inst->saturate = instr->dest.saturate;
769 fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_F);
770 inst = bld.emit(SHADER_OPCODE_SIN, tmp, op[0]);
771 if (instr->dest.saturate) {
773 inst->saturate = true;
775 bld.MUL(result, tmp, brw_imm_f(0.99997));
781 fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_F);
782 inst = bld.emit(SHADER_OPCODE_COS, tmp, op[0]);
783 if (instr->dest.saturate) {
785 inst->saturate = true;
787 bld.MUL(result, tmp, brw_imm_f(0.99997));
793 if (fs_key->high_quality_derivatives) {
794 inst = bld.emit(FS_OPCODE_DDX_FINE, result, op[0]);
796 inst = bld.emit(FS_OPCODE_DDX_COARSE, result, op[0]);
798 inst->saturate = instr->dest.saturate;
800 case nir_op_fddx_fine:
801 inst = bld.emit(FS_OPCODE_DDX_FINE, result, op[0]);
802 inst->saturate = instr->dest.saturate;
804 case nir_op_fddx_coarse:
805 inst = bld.emit(FS_OPCODE_DDX_COARSE, result, op[0]);
806 inst->saturate = instr->dest.saturate;
809 if (fs_key->high_quality_derivatives) {
810 inst = bld.emit(FS_OPCODE_DDY_FINE, result, op[0],
811 brw_imm_d(fs_key->render_to_fbo));
813 inst = bld.emit(FS_OPCODE_DDY_COARSE, result, op[0],
814 brw_imm_d(fs_key->render_to_fbo));
816 inst->saturate = instr->dest.saturate;
818 case nir_op_fddy_fine:
819 inst = bld.emit(FS_OPCODE_DDY_FINE, result, op[0],
820 brw_imm_d(fs_key->render_to_fbo));
821 inst->saturate = instr->dest.saturate;
823 case nir_op_fddy_coarse:
824 inst = bld.emit(FS_OPCODE_DDY_COARSE, result, op[0],
825 brw_imm_d(fs_key->render_to_fbo));
826 inst->saturate = instr->dest.saturate;
831 inst = bld.ADD(result, op[0], op[1]);
832 inst->saturate = instr->dest.saturate;
836 inst = bld.MUL(result, op[0], op[1]);
837 inst->saturate = instr->dest.saturate;
841 bld.MUL(result, op[0], op[1]);
844 case nir_op_imul_high:
845 case nir_op_umul_high:
846 bld.emit(SHADER_OPCODE_MULH, result, op[0], op[1]);
851 bld.emit(SHADER_OPCODE_INT_QUOTIENT, result, op[0], op[1]);
854 case nir_op_uadd_carry:
855 unreachable("Should have been lowered by carry_to_arith().");
857 case nir_op_usub_borrow:
858 unreachable("Should have been lowered by borrow_to_arith().");
862 /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
863 * appears that our hardware just does the right thing for signed
866 bld.emit(SHADER_OPCODE_INT_REMAINDER, result, op[0], op[1]);
870 /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
871 bld.emit(SHADER_OPCODE_INT_REMAINDER, result, op[0], op[1]);
873 /* Math instructions don't support conditional mod */
874 inst = bld.MOV(bld.null_reg_d(), result);
875 inst->conditional_mod = BRW_CONDITIONAL_NZ;
877 /* Now, we need to determine if signs of the sources are different.
878 * When we XOR the sources, the top bit is 0 if they are the same and 1
879 * if they are different. We can then use a conditional modifier to
880 * turn that into a predicate. This leads us to an XOR.l instruction.
882 fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_D);
883 inst = bld.XOR(tmp, op[0], op[1]);
884 inst->predicate = BRW_PREDICATE_NORMAL;
885 inst->conditional_mod = BRW_CONDITIONAL_L;
887 /* If the result of the initial remainder operation is non-zero and the
888 * two sources have different signs, add in a copy of op[1] to get the
889 * final integer modulus value.
891 inst = bld.ADD(result, result, op[1]);
892 inst->predicate = BRW_PREDICATE_NORMAL;
899 bld.CMP(result, op[0], op[1], BRW_CONDITIONAL_L);
905 bld.CMP(result, op[0], op[1], BRW_CONDITIONAL_GE);
910 bld.CMP(result, op[0], op[1], BRW_CONDITIONAL_Z);
915 bld.CMP(result, op[0], op[1], BRW_CONDITIONAL_NZ);
919 if (devinfo->gen >= 8) {
920 op[0] = resolve_source_modifiers(op[0]);
922 bld.NOT(result, op[0]);
925 if (devinfo->gen >= 8) {
926 op[0] = resolve_source_modifiers(op[0]);
927 op[1] = resolve_source_modifiers(op[1]);
929 bld.XOR(result, op[0], op[1]);
932 if (devinfo->gen >= 8) {
933 op[0] = resolve_source_modifiers(op[0]);
934 op[1] = resolve_source_modifiers(op[1]);
936 bld.OR(result, op[0], op[1]);
939 if (devinfo->gen >= 8) {
940 op[0] = resolve_source_modifiers(op[0]);
941 op[1] = resolve_source_modifiers(op[1]);
943 bld.AND(result, op[0], op[1]);
949 case nir_op_ball_fequal2:
950 case nir_op_ball_iequal2:
951 case nir_op_ball_fequal3:
952 case nir_op_ball_iequal3:
953 case nir_op_ball_fequal4:
954 case nir_op_ball_iequal4:
955 case nir_op_bany_fnequal2:
956 case nir_op_bany_inequal2:
957 case nir_op_bany_fnequal3:
958 case nir_op_bany_inequal3:
959 case nir_op_bany_fnequal4:
960 case nir_op_bany_inequal4:
961 unreachable("Lowered by nir_lower_alu_reductions");
963 case nir_op_fnoise1_1:
964 case nir_op_fnoise1_2:
965 case nir_op_fnoise1_3:
966 case nir_op_fnoise1_4:
967 case nir_op_fnoise2_1:
968 case nir_op_fnoise2_2:
969 case nir_op_fnoise2_3:
970 case nir_op_fnoise2_4:
971 case nir_op_fnoise3_1:
972 case nir_op_fnoise3_2:
973 case nir_op_fnoise3_3:
974 case nir_op_fnoise3_4:
975 case nir_op_fnoise4_1:
976 case nir_op_fnoise4_2:
977 case nir_op_fnoise4_3:
978 case nir_op_fnoise4_4:
979 unreachable("not reached: should be handled by lower_noise");
982 unreachable("not reached: should be handled by ldexp_to_arith()");
985 inst = bld.emit(SHADER_OPCODE_SQRT, result, op[0]);
986 inst->saturate = instr->dest.saturate;
990 inst = bld.emit(SHADER_OPCODE_RSQ, result, op[0]);
991 inst->saturate = instr->dest.saturate;
996 bld.MOV(result, negate(op[0]));
1000 bld.CMP(result, op[0], brw_imm_f(0.0f), BRW_CONDITIONAL_NZ);
1003 bld.CMP(result, op[0], brw_imm_d(0), BRW_CONDITIONAL_NZ);
1007 inst = bld.RNDZ(result, op[0]);
1008 inst->saturate = instr->dest.saturate;
1011 case nir_op_fceil: {
1012 op[0].negate = !op[0].negate;
1013 fs_reg temp = vgrf(glsl_type::float_type);
1014 bld.RNDD(temp, op[0]);
1016 inst = bld.MOV(result, temp);
1017 inst->saturate = instr->dest.saturate;
1021 inst = bld.RNDD(result, op[0]);
1022 inst->saturate = instr->dest.saturate;
1025 inst = bld.FRC(result, op[0]);
1026 inst->saturate = instr->dest.saturate;
1028 case nir_op_fround_even:
1029 inst = bld.RNDE(result, op[0]);
1030 inst->saturate = instr->dest.saturate;
1033 case nir_op_fquantize2f16: {
1034 fs_reg tmp16 = bld.vgrf(BRW_REGISTER_TYPE_D);
1035 fs_reg tmp32 = bld.vgrf(BRW_REGISTER_TYPE_F);
1036 fs_reg zero = bld.vgrf(BRW_REGISTER_TYPE_F);
1038 /* The destination stride must be at least as big as the source stride. */
1039 tmp16.type = BRW_REGISTER_TYPE_W;
1042 /* Check for denormal */
1043 fs_reg abs_src0 = op[0];
1044 abs_src0.abs = true;
1045 bld.CMP(bld.null_reg_f(), abs_src0, brw_imm_f(ldexpf(1.0, -14)),
1047 /* Get the appropriately signed zero */
1048 bld.AND(retype(zero, BRW_REGISTER_TYPE_UD),
1049 retype(op[0], BRW_REGISTER_TYPE_UD),
1050 brw_imm_ud(0x80000000));
1051 /* Do the actual F32 -> F16 -> F32 conversion */
1052 bld.emit(BRW_OPCODE_F32TO16, tmp16, op[0]);
1053 bld.emit(BRW_OPCODE_F16TO32, tmp32, tmp16);
1054 /* Select that or zero based on normal status */
1055 inst = bld.SEL(result, zero, tmp32);
1056 inst->predicate = BRW_PREDICATE_NORMAL;
1057 inst->saturate = instr->dest.saturate;
1064 inst = bld.emit_minmax(result, op[0], op[1], BRW_CONDITIONAL_L);
1065 inst->saturate = instr->dest.saturate;
1071 inst = bld.emit_minmax(result, op[0], op[1], BRW_CONDITIONAL_GE);
1072 inst->saturate = instr->dest.saturate;
1075 case nir_op_pack_snorm_2x16:
1076 case nir_op_pack_snorm_4x8:
1077 case nir_op_pack_unorm_2x16:
1078 case nir_op_pack_unorm_4x8:
1079 case nir_op_unpack_snorm_2x16:
1080 case nir_op_unpack_snorm_4x8:
1081 case nir_op_unpack_unorm_2x16:
1082 case nir_op_unpack_unorm_4x8:
1083 case nir_op_unpack_half_2x16:
1084 case nir_op_pack_half_2x16:
1085 unreachable("not reached: should be handled by lower_packing_builtins");
1087 case nir_op_unpack_half_2x16_split_x:
1088 inst = bld.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X, result, op[0]);
1089 inst->saturate = instr->dest.saturate;
1091 case nir_op_unpack_half_2x16_split_y:
1092 inst = bld.emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y, result, op[0]);
1093 inst->saturate = instr->dest.saturate;
1097 inst = bld.emit(SHADER_OPCODE_POW, result, op[0], op[1]);
1098 inst->saturate = instr->dest.saturate;
1101 case nir_op_bitfield_reverse:
1102 bld.BFREV(result, op[0]);
1105 case nir_op_bit_count:
1106 bld.CBIT(result, op[0]);
1109 case nir_op_ufind_msb:
1110 case nir_op_ifind_msb: {
1111 bld.FBH(retype(result, BRW_REGISTER_TYPE_UD), op[0]);
1113 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
1114 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
1115 * subtract the result from 31 to convert the MSB count into an LSB count.
1117 bld.CMP(bld.null_reg_d(), result, brw_imm_d(-1), BRW_CONDITIONAL_NZ);
1119 inst = bld.ADD(result, result, brw_imm_d(31));
1120 inst->predicate = BRW_PREDICATE_NORMAL;
1121 inst->src[0].negate = true;
1125 case nir_op_find_lsb:
1126 bld.FBL(result, op[0]);
1129 case nir_op_ubitfield_extract:
1130 case nir_op_ibitfield_extract:
1131 unreachable("should have been lowered");
1134 bld.BFE(result, op[2], op[1], op[0]);
1137 bld.BFI1(result, op[0], op[1]);
1140 bld.BFI2(result, op[0], op[1], op[2]);
1143 case nir_op_bitfield_insert:
1144 unreachable("not reached: should have been lowered");
1147 bld.SHL(result, op[0], op[1]);
1150 bld.ASR(result, op[0], op[1]);
1153 bld.SHR(result, op[0], op[1]);
1156 case nir_op_pack_half_2x16_split:
1157 bld.emit(FS_OPCODE_PACK_HALF_2x16_SPLIT, result, op[0], op[1]);
1161 inst = bld.MAD(result, op[2], op[1], op[0]);
1162 inst->saturate = instr->dest.saturate;
1166 inst = bld.LRP(result, op[0], op[1], op[2]);
1167 inst->saturate = instr->dest.saturate;
1171 if (optimize_frontfacing_ternary(instr, result))
1174 bld.CMP(bld.null_reg_d(), op[0], brw_imm_d(0), BRW_CONDITIONAL_NZ);
1175 inst = bld.SEL(result, op[1], op[2]);
1176 inst->predicate = BRW_PREDICATE_NORMAL;
1179 case nir_op_extract_u8:
1180 case nir_op_extract_i8: {
1181 nir_const_value *byte = nir_src_as_const_value(instr->src[1].src);
1182 bld.emit(SHADER_OPCODE_EXTRACT_BYTE,
1183 result, op[0], brw_imm_ud(byte->u[0]));
1187 case nir_op_extract_u16:
1188 case nir_op_extract_i16: {
1189 nir_const_value *word = nir_src_as_const_value(instr->src[1].src);
1190 bld.emit(SHADER_OPCODE_EXTRACT_WORD,
1191 result, op[0], brw_imm_ud(word->u[0]));
1196 unreachable("unhandled instruction");
1199 /* If we need to do a boolean resolve, replace the result with -(x & 1)
1200 * to sign extend the low bit to 0/~0
1202 if (devinfo->gen <= 5 &&
1203 (instr->instr.pass_flags & BRW_NIR_BOOLEAN_MASK) == BRW_NIR_BOOLEAN_NEEDS_RESOLVE) {
1204 fs_reg masked = vgrf(glsl_type::int_type);
1205 bld.AND(masked, result, brw_imm_d(1));
1206 masked.negate = true;
1207 bld.MOV(retype(result, BRW_REGISTER_TYPE_D), masked);
1212 fs_visitor::nir_emit_load_const(const fs_builder &bld,
1213 nir_load_const_instr *instr)
1215 fs_reg reg = bld.vgrf(BRW_REGISTER_TYPE_D, instr->def.num_components);
1217 for (unsigned i = 0; i < instr->def.num_components; i++)
1218 bld.MOV(offset(reg, bld, i), brw_imm_d(instr->value.i[i]));
1220 nir_ssa_values[instr->def.index] = reg;
1224 fs_visitor::nir_emit_undef(const fs_builder &bld, nir_ssa_undef_instr *instr)
1226 nir_ssa_values[instr->def.index] = bld.vgrf(BRW_REGISTER_TYPE_D,
1227 instr->def.num_components);
1231 fs_visitor::get_nir_src(nir_src src)
1235 reg = nir_ssa_values[src.ssa->index];
1237 /* We don't handle indirects on locals */
1238 assert(src.reg.indirect == NULL);
1239 reg = offset(nir_locals[src.reg.reg->index], bld,
1240 src.reg.base_offset * src.reg.reg->num_components);
1243 /* to avoid floating-point denorm flushing problems, set the type by
1244 * default to D - instructions that need floating point semantics will set
1245 * this to F if they need to
1247 return retype(reg, BRW_REGISTER_TYPE_D);
1251 fs_visitor::get_nir_dest(nir_dest dest)
1254 nir_ssa_values[dest.ssa.index] = bld.vgrf(BRW_REGISTER_TYPE_F,
1255 dest.ssa.num_components);
1256 return nir_ssa_values[dest.ssa.index];
1258 /* We don't handle indirects on locals */
1259 assert(dest.reg.indirect == NULL);
1260 return offset(nir_locals[dest.reg.reg->index], bld,
1261 dest.reg.base_offset * dest.reg.reg->num_components);
1266 fs_visitor::get_nir_image_deref(const nir_deref_var *deref)
1268 fs_reg image(UNIFORM, deref->var->data.driver_location / 4,
1269 BRW_REGISTER_TYPE_UD);
1271 unsigned indirect_max = 0;
1273 for (const nir_deref *tail = &deref->deref; tail->child;
1274 tail = tail->child) {
1275 const nir_deref_array *deref_array = nir_deref_as_array(tail->child);
1276 assert(tail->child->deref_type == nir_deref_type_array);
1277 const unsigned size = glsl_get_length(tail->type);
1278 const unsigned element_size = type_size_scalar(deref_array->deref.type);
1279 const unsigned base = MIN2(deref_array->base_offset, size - 1);
1280 image = offset(image, bld, base * element_size);
1282 if (deref_array->deref_array_type == nir_deref_array_type_indirect) {
1283 fs_reg tmp = vgrf(glsl_type::uint_type);
1285 if (devinfo->gen == 7 && !devinfo->is_haswell) {
1286 /* IVB hangs when trying to access an invalid surface index with
1287 * the dataport. According to the spec "if the index used to
1288 * select an individual element is negative or greater than or
1289 * equal to the size of the array, the results of the operation
1290 * are undefined but may not lead to termination" -- which is one
1291 * of the possible outcomes of the hang. Clamp the index to
1292 * prevent access outside of the array bounds.
1294 bld.emit_minmax(tmp, retype(get_nir_src(deref_array->indirect),
1295 BRW_REGISTER_TYPE_UD),
1296 brw_imm_ud(size - base - 1), BRW_CONDITIONAL_L);
1298 bld.MOV(tmp, get_nir_src(deref_array->indirect));
1301 indirect_max += element_size * (tail->type->length - 1);
1303 bld.MUL(tmp, tmp, brw_imm_ud(element_size * 4));
1304 if (indirect.file == BAD_FILE) {
1307 bld.ADD(indirect, indirect, tmp);
1312 if (indirect.file == BAD_FILE) {
1315 /* Emit a pile of MOVs to load the uniform into a temporary. The
1316 * dead-code elimination pass will get rid of what we don't use.
1318 fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_UD, BRW_IMAGE_PARAM_SIZE);
1319 for (unsigned j = 0; j < BRW_IMAGE_PARAM_SIZE; j++) {
1320 bld.emit(SHADER_OPCODE_MOV_INDIRECT,
1321 offset(tmp, bld, j), offset(image, bld, j),
1322 indirect, brw_imm_ud((indirect_max + 1) * 4));
1329 fs_visitor::emit_percomp(const fs_builder &bld, const fs_inst &inst,
1332 for (unsigned i = 0; i < 4; i++) {
1333 if (!((wr_mask >> i) & 1))
1336 fs_inst *new_inst = new(mem_ctx) fs_inst(inst);
1337 new_inst->dst = offset(new_inst->dst, bld, i);
1338 for (unsigned j = 0; j < new_inst->sources; j++)
1339 if (new_inst->src[j].file == VGRF)
1340 new_inst->src[j] = offset(new_inst->src[j], bld, i);
1347 * Get the matching channel register datatype for an image intrinsic of the
1348 * specified GLSL image type.
1351 get_image_base_type(const glsl_type *type)
1353 switch ((glsl_base_type)type->sampled_type) {
1354 case GLSL_TYPE_UINT:
1355 return BRW_REGISTER_TYPE_UD;
1357 return BRW_REGISTER_TYPE_D;
1358 case GLSL_TYPE_FLOAT:
1359 return BRW_REGISTER_TYPE_F;
1361 unreachable("Not reached.");
1366 * Get the appropriate atomic op for an image atomic intrinsic.
1369 get_image_atomic_op(nir_intrinsic_op op, const glsl_type *type)
1372 case nir_intrinsic_image_atomic_add:
1374 case nir_intrinsic_image_atomic_min:
1375 return (get_image_base_type(type) == BRW_REGISTER_TYPE_D ?
1376 BRW_AOP_IMIN : BRW_AOP_UMIN);
1377 case nir_intrinsic_image_atomic_max:
1378 return (get_image_base_type(type) == BRW_REGISTER_TYPE_D ?
1379 BRW_AOP_IMAX : BRW_AOP_UMAX);
1380 case nir_intrinsic_image_atomic_and:
1382 case nir_intrinsic_image_atomic_or:
1384 case nir_intrinsic_image_atomic_xor:
1386 case nir_intrinsic_image_atomic_exchange:
1388 case nir_intrinsic_image_atomic_comp_swap:
1389 return BRW_AOP_CMPWR;
1391 unreachable("Not reachable.");
1396 emit_pixel_interpolater_send(const fs_builder &bld,
1401 glsl_interp_qualifier interpolation)
1407 if (src.file == BAD_FILE) {
1409 payload = bld.vgrf(BRW_REGISTER_TYPE_F, 1);
1413 mlen = 2 * bld.dispatch_width() / 8;
1416 inst = bld.emit(opcode, dst, payload, desc);
1418 /* 2 floats per slot returned */
1419 inst->regs_written = 2 * bld.dispatch_width() / 8;
1420 inst->pi_noperspective = interpolation == INTERP_QUALIFIER_NOPERSPECTIVE;
1426 * Computes 1 << x, given a D/UD register containing some value x.
1429 intexp2(const fs_builder &bld, const fs_reg &x)
1431 assert(x.type == BRW_REGISTER_TYPE_UD || x.type == BRW_REGISTER_TYPE_D);
1433 fs_reg result = bld.vgrf(x.type, 1);
1434 fs_reg one = bld.vgrf(x.type, 1);
1436 bld.MOV(one, retype(brw_imm_d(1), one.type));
1437 bld.SHL(result, one, x);
1442 fs_visitor::emit_gs_end_primitive(const nir_src &vertex_count_nir_src)
1444 assert(stage == MESA_SHADER_GEOMETRY);
1446 struct brw_gs_prog_data *gs_prog_data =
1447 (struct brw_gs_prog_data *) prog_data;
1449 /* We can only do EndPrimitive() functionality when the control data
1450 * consists of cut bits. Fortunately, the only time it isn't is when the
1451 * output type is points, in which case EndPrimitive() is a no-op.
1453 if (gs_prog_data->control_data_format !=
1454 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_CUT) {
1458 /* Cut bits use one bit per vertex. */
1459 assert(gs_compile->control_data_bits_per_vertex == 1);
1461 fs_reg vertex_count = get_nir_src(vertex_count_nir_src);
1462 vertex_count.type = BRW_REGISTER_TYPE_UD;
1464 /* Cut bit n should be set to 1 if EndPrimitive() was called after emitting
1465 * vertex n, 0 otherwise. So all we need to do here is mark bit
1466 * (vertex_count - 1) % 32 in the cut_bits register to indicate that
1467 * EndPrimitive() was called after emitting vertex (vertex_count - 1);
1468 * vec4_gs_visitor::emit_control_data_bits() will take care of the rest.
1470 * Note that if EndPrimitive() is called before emitting any vertices, this
1471 * will cause us to set bit 31 of the control_data_bits register to 1.
1472 * That's fine because:
1474 * - If max_vertices < 32, then vertex number 31 (zero-based) will never be
1475 * output, so the hardware will ignore cut bit 31.
1477 * - If max_vertices == 32, then vertex number 31 is guaranteed to be the
1478 * last vertex, so setting cut bit 31 has no effect (since the primitive
1479 * is automatically ended when the GS terminates).
1481 * - If max_vertices > 32, then the ir_emit_vertex visitor will reset the
1482 * control_data_bits register to 0 when the first vertex is emitted.
1485 const fs_builder abld = bld.annotate("end primitive");
1487 /* control_data_bits |= 1 << ((vertex_count - 1) % 32) */
1488 fs_reg prev_count = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1489 abld.ADD(prev_count, vertex_count, brw_imm_ud(0xffffffffu));
1490 fs_reg mask = intexp2(abld, prev_count);
1491 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1492 * attention to the lower 5 bits of its second source argument, so on this
1493 * architecture, 1 << (vertex_count - 1) is equivalent to 1 <<
1494 * ((vertex_count - 1) % 32).
1496 abld.OR(this->control_data_bits, this->control_data_bits, mask);
1500 fs_visitor::emit_gs_control_data_bits(const fs_reg &vertex_count)
1502 assert(stage == MESA_SHADER_GEOMETRY);
1503 assert(gs_compile->control_data_bits_per_vertex != 0);
1505 struct brw_gs_prog_data *gs_prog_data =
1506 (struct brw_gs_prog_data *) prog_data;
1508 const fs_builder abld = bld.annotate("emit control data bits");
1509 const fs_builder fwa_bld = bld.exec_all();
1511 /* We use a single UD register to accumulate control data bits (32 bits
1512 * for each of the SIMD8 channels). So we need to write a DWord (32 bits)
1515 * Unfortunately, the URB_WRITE_SIMD8 message uses 128-bit (OWord) offsets.
1516 * We have select a 128-bit group via the Global and Per-Slot Offsets, then
1517 * use the Channel Mask phase to enable/disable which DWord within that
1518 * group to write. (Remember, different SIMD8 channels may have emitted
1519 * different numbers of vertices, so we may need per-slot offsets.)
1521 * Channel masking presents an annoying problem: we may have to replicate
1522 * the data up to 4 times:
1524 * Msg = Handles, Per-Slot Offsets, Channel Masks, Data, Data, Data, Data.
1526 * To avoid penalizing shaders that emit a small number of vertices, we
1527 * can avoid these sometimes: if the size of the control data header is
1528 * <= 128 bits, then there is only 1 OWord. All SIMD8 channels will land
1529 * land in the same 128-bit group, so we can skip per-slot offsets.
1531 * Similarly, if the control data header is <= 32 bits, there is only one
1532 * DWord, so we can skip channel masks.
1534 enum opcode opcode = SHADER_OPCODE_URB_WRITE_SIMD8;
1536 fs_reg channel_mask, per_slot_offset;
1538 if (gs_compile->control_data_header_size_bits > 32) {
1539 opcode = SHADER_OPCODE_URB_WRITE_SIMD8_MASKED;
1540 channel_mask = vgrf(glsl_type::uint_type);
1543 if (gs_compile->control_data_header_size_bits > 128) {
1544 opcode = SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT;
1545 per_slot_offset = vgrf(glsl_type::uint_type);
1548 /* Figure out which DWord we're trying to write to using the formula:
1550 * dword_index = (vertex_count - 1) * bits_per_vertex / 32
1552 * Since bits_per_vertex is a power of two, and is known at compile
1553 * time, this can be optimized to:
1555 * dword_index = (vertex_count - 1) >> (6 - log2(bits_per_vertex))
1557 if (opcode != SHADER_OPCODE_URB_WRITE_SIMD8) {
1558 fs_reg dword_index = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1559 fs_reg prev_count = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1560 abld.ADD(prev_count, vertex_count, brw_imm_ud(0xffffffffu));
1561 unsigned log2_bits_per_vertex =
1562 _mesa_fls(gs_compile->control_data_bits_per_vertex);
1563 abld.SHR(dword_index, prev_count, brw_imm_ud(6u - log2_bits_per_vertex));
1565 if (per_slot_offset.file != BAD_FILE) {
1566 /* Set the per-slot offset to dword_index / 4, so that we'll write to
1567 * the appropriate OWord within the control data header.
1569 abld.SHR(per_slot_offset, dword_index, brw_imm_ud(2u));
1572 /* Set the channel masks to 1 << (dword_index % 4), so that we'll
1573 * write to the appropriate DWORD within the OWORD.
1575 fs_reg channel = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1576 fwa_bld.AND(channel, dword_index, brw_imm_ud(3u));
1577 channel_mask = intexp2(fwa_bld, channel);
1578 /* Then the channel masks need to be in bits 23:16. */
1579 fwa_bld.SHL(channel_mask, channel_mask, brw_imm_ud(16u));
1582 /* Store the control data bits in the message payload and send it. */
1584 if (channel_mask.file != BAD_FILE)
1585 mlen += 4; /* channel masks, plus 3 extra copies of the data */
1586 if (per_slot_offset.file != BAD_FILE)
1589 fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, mlen);
1590 fs_reg *sources = ralloc_array(mem_ctx, fs_reg, mlen);
1592 sources[i++] = fs_reg(retype(brw_vec8_grf(1, 0), BRW_REGISTER_TYPE_UD));
1593 if (per_slot_offset.file != BAD_FILE)
1594 sources[i++] = per_slot_offset;
1595 if (channel_mask.file != BAD_FILE)
1596 sources[i++] = channel_mask;
1598 sources[i++] = this->control_data_bits;
1601 abld.LOAD_PAYLOAD(payload, sources, mlen, mlen);
1602 fs_inst *inst = abld.emit(opcode, reg_undef, payload);
1604 /* We need to increment Global Offset by 256-bits to make room for
1605 * Broadwell's extra "Vertex Count" payload at the beginning of the
1606 * URB entry. Since this is an OWord message, Global Offset is counted
1607 * in 128-bit units, so we must set it to 2.
1609 if (gs_prog_data->static_vertex_count == -1)
1614 fs_visitor::set_gs_stream_control_data_bits(const fs_reg &vertex_count,
1617 /* control_data_bits |= stream_id << ((2 * (vertex_count - 1)) % 32) */
1619 /* Note: we are calling this *before* increasing vertex_count, so
1620 * this->vertex_count == vertex_count - 1 in the formula above.
1623 /* Stream mode uses 2 bits per vertex */
1624 assert(gs_compile->control_data_bits_per_vertex == 2);
1626 /* Must be a valid stream */
1627 assert(stream_id >= 0 && stream_id < MAX_VERTEX_STREAMS);
1629 /* Control data bits are initialized to 0 so we don't have to set any
1630 * bits when sending vertices to stream 0.
1635 const fs_builder abld = bld.annotate("set stream control data bits", NULL);
1637 /* reg::sid = stream_id */
1638 fs_reg sid = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1639 abld.MOV(sid, brw_imm_ud(stream_id));
1641 /* reg:shift_count = 2 * (vertex_count - 1) */
1642 fs_reg shift_count = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1643 abld.SHL(shift_count, vertex_count, brw_imm_ud(1u));
1645 /* Note: we're relying on the fact that the GEN SHL instruction only pays
1646 * attention to the lower 5 bits of its second source argument, so on this
1647 * architecture, stream_id << 2 * (vertex_count - 1) is equivalent to
1648 * stream_id << ((2 * (vertex_count - 1)) % 32).
1650 fs_reg mask = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1651 abld.SHL(mask, sid, shift_count);
1652 abld.OR(this->control_data_bits, this->control_data_bits, mask);
1656 fs_visitor::emit_gs_vertex(const nir_src &vertex_count_nir_src,
1659 assert(stage == MESA_SHADER_GEOMETRY);
1661 struct brw_gs_prog_data *gs_prog_data =
1662 (struct brw_gs_prog_data *) prog_data;
1664 fs_reg vertex_count = get_nir_src(vertex_count_nir_src);
1665 vertex_count.type = BRW_REGISTER_TYPE_UD;
1667 /* Haswell and later hardware ignores the "Render Stream Select" bits
1668 * from the 3DSTATE_STREAMOUT packet when the SOL stage is disabled,
1669 * and instead sends all primitives down the pipeline for rasterization.
1670 * If the SOL stage is enabled, "Render Stream Select" is honored and
1671 * primitives bound to non-zero streams are discarded after stream output.
1673 * Since the only purpose of primives sent to non-zero streams is to
1674 * be recorded by transform feedback, we can simply discard all geometry
1675 * bound to these streams when transform feedback is disabled.
1677 if (stream_id > 0 && !nir->info.has_transform_feedback_varyings)
1680 /* If we're outputting 32 control data bits or less, then we can wait
1681 * until the shader is over to output them all. Otherwise we need to
1682 * output them as we go. Now is the time to do it, since we're about to
1683 * output the vertex_count'th vertex, so it's guaranteed that the
1684 * control data bits associated with the (vertex_count - 1)th vertex are
1687 if (gs_compile->control_data_header_size_bits > 32) {
1688 const fs_builder abld =
1689 bld.annotate("emit vertex: emit control data bits");
1691 /* Only emit control data bits if we've finished accumulating a batch
1692 * of 32 bits. This is the case when:
1694 * (vertex_count * bits_per_vertex) % 32 == 0
1696 * (in other words, when the last 5 bits of vertex_count *
1697 * bits_per_vertex are 0). Assuming bits_per_vertex == 2^n for some
1698 * integer n (which is always the case, since bits_per_vertex is
1699 * always 1 or 2), this is equivalent to requiring that the last 5-n
1700 * bits of vertex_count are 0:
1702 * vertex_count & (2^(5-n) - 1) == 0
1704 * 2^(5-n) == 2^5 / 2^n == 32 / bits_per_vertex, so this is
1707 * vertex_count & (32 / bits_per_vertex - 1) == 0
1709 * TODO: If vertex_count is an immediate, we could do some of this math
1710 * at compile time...
1713 abld.AND(bld.null_reg_d(), vertex_count,
1714 brw_imm_ud(32u / gs_compile->control_data_bits_per_vertex - 1u));
1715 inst->conditional_mod = BRW_CONDITIONAL_Z;
1717 abld.IF(BRW_PREDICATE_NORMAL);
1718 /* If vertex_count is 0, then no control data bits have been
1719 * accumulated yet, so we can skip emitting them.
1721 abld.CMP(bld.null_reg_d(), vertex_count, brw_imm_ud(0u),
1722 BRW_CONDITIONAL_NEQ);
1723 abld.IF(BRW_PREDICATE_NORMAL);
1724 emit_gs_control_data_bits(vertex_count);
1725 abld.emit(BRW_OPCODE_ENDIF);
1727 /* Reset control_data_bits to 0 so we can start accumulating a new
1730 * Note: in the case where vertex_count == 0, this neutralizes the
1731 * effect of any call to EndPrimitive() that the shader may have
1732 * made before outputting its first vertex.
1734 inst = abld.MOV(this->control_data_bits, brw_imm_ud(0u));
1735 inst->force_writemask_all = true;
1736 abld.emit(BRW_OPCODE_ENDIF);
1739 emit_urb_writes(vertex_count);
1741 /* In stream mode we have to set control data bits for all vertices
1742 * unless we have disabled control data bits completely (which we do
1743 * do for GL_POINTS outputs that don't use streams).
1745 if (gs_compile->control_data_header_size_bits > 0 &&
1746 gs_prog_data->control_data_format ==
1747 GEN7_GS_CONTROL_DATA_FORMAT_GSCTL_SID) {
1748 set_gs_stream_control_data_bits(vertex_count, stream_id);
1753 fs_visitor::emit_gs_input_load(const fs_reg &dst,
1754 const nir_src &vertex_src,
1755 unsigned base_offset,
1756 const nir_src &offset_src,
1757 unsigned num_components)
1759 struct brw_gs_prog_data *gs_prog_data = (struct brw_gs_prog_data *) prog_data;
1761 nir_const_value *vertex_const = nir_src_as_const_value(vertex_src);
1762 nir_const_value *offset_const = nir_src_as_const_value(offset_src);
1763 const unsigned push_reg_count = gs_prog_data->base.urb_read_length * 8;
1765 /* Offset 0 is the VUE header, which contains VARYING_SLOT_LAYER [.y],
1766 * VARYING_SLOT_VIEWPORT [.z], and VARYING_SLOT_PSIZ [.w]. Only
1767 * gl_PointSize is available as a GS input, however, so it must be that.
1769 const bool is_point_size = (base_offset == 0);
1771 if (offset_const != NULL && vertex_const != NULL &&
1772 4 * (base_offset + offset_const->u[0]) < push_reg_count) {
1773 int imm_offset = (base_offset + offset_const->u[0]) * 4 +
1774 vertex_const->u[0] * push_reg_count;
1775 /* This input was pushed into registers. */
1776 if (is_point_size) {
1777 /* gl_PointSize comes in .w */
1778 assert(imm_offset == 0);
1779 bld.MOV(dst, fs_reg(ATTR, imm_offset + 3, dst.type));
1781 for (unsigned i = 0; i < num_components; i++) {
1782 bld.MOV(offset(dst, bld, i),
1783 fs_reg(ATTR, imm_offset + i, dst.type));
1787 /* Resort to the pull model. Ensure the VUE handles are provided. */
1788 gs_prog_data->base.include_vue_handles = true;
1790 unsigned first_icp_handle = gs_prog_data->include_primitive_id ? 3 : 2;
1794 /* The vertex index is constant; just select the proper URB handle. */
1796 retype(brw_vec8_grf(first_icp_handle + vertex_const->i[0], 0),
1797 BRW_REGISTER_TYPE_UD);
1799 /* The vertex index is non-constant. We need to use indirect
1800 * addressing to fetch the proper URB handle.
1802 * First, we start with the sequence <7, 6, 5, 4, 3, 2, 1, 0>
1803 * indicating that channel <n> should read the handle from
1804 * DWord <n>. We convert that to bytes by multiplying by 4.
1806 * Next, we convert the vertex index to bytes by multiplying
1807 * by 32 (shifting by 5), and add the two together. This is
1808 * the final indirect byte offset.
1810 fs_reg sequence = bld.vgrf(BRW_REGISTER_TYPE_W, 1);
1811 fs_reg channel_offsets = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1812 fs_reg vertex_offset_bytes = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1813 fs_reg icp_offset_bytes = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1814 icp_handle = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
1816 /* sequence = <7, 6, 5, 4, 3, 2, 1, 0> */
1817 bld.MOV(sequence, fs_reg(brw_imm_v(0x76543210)));
1818 /* channel_offsets = 4 * sequence = <28, 24, 20, 16, 12, 8, 4, 0> */
1819 bld.SHL(channel_offsets, sequence, brw_imm_ud(2u));
1820 /* Convert vertex_index to bytes (multiply by 32) */
1821 bld.SHL(vertex_offset_bytes,
1822 retype(get_nir_src(vertex_src), BRW_REGISTER_TYPE_UD),
1824 bld.ADD(icp_offset_bytes, vertex_offset_bytes, channel_offsets);
1826 /* Use first_icp_handle as the base offset. There is one register
1827 * of URB handles per vertex, so inform the register allocator that
1828 * we might read up to nir->info.gs.vertices_in registers.
1830 bld.emit(SHADER_OPCODE_MOV_INDIRECT, icp_handle,
1831 fs_reg(brw_vec8_grf(first_icp_handle, 0)),
1832 fs_reg(icp_offset_bytes),
1833 brw_imm_ud(nir->info.gs.vertices_in * REG_SIZE));
1838 /* Constant indexing - use global offset. */
1839 inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dst, icp_handle);
1840 inst->offset = base_offset + offset_const->u[0];
1841 inst->base_mrf = -1;
1843 inst->regs_written = num_components;
1845 /* Indirect indexing - use per-slot offsets as well. */
1846 const fs_reg srcs[] = { icp_handle, get_nir_src(offset_src) };
1847 fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
1848 bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
1850 inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dst, payload);
1851 inst->offset = base_offset;
1852 inst->base_mrf = -1;
1854 inst->regs_written = num_components;
1857 if (is_point_size) {
1858 /* Read the whole VUE header (because of alignment) and read .w. */
1859 fs_reg tmp = bld.vgrf(dst.type, 4);
1861 inst->regs_written = 4;
1862 bld.MOV(dst, offset(tmp, bld, 3));
1868 fs_visitor::get_indirect_offset(nir_intrinsic_instr *instr)
1870 nir_src *offset_src = nir_get_io_offset_src(instr);
1871 nir_const_value *const_value = nir_src_as_const_value(*offset_src);
1874 /* The only constant offset we should find is 0. brw_nir.c's
1875 * add_const_offset_to_base() will fold other constant offsets
1876 * into instr->const_index[0].
1878 assert(const_value->u[0] == 0);
1882 return get_nir_src(*offset_src);
1886 fs_visitor::nir_emit_vs_intrinsic(const fs_builder &bld,
1887 nir_intrinsic_instr *instr)
1889 assert(stage == MESA_SHADER_VERTEX);
1892 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
1893 dest = get_nir_dest(instr->dest);
1895 switch (instr->intrinsic) {
1896 case nir_intrinsic_load_vertex_id:
1897 unreachable("should be lowered by lower_vertex_id()");
1899 case nir_intrinsic_load_vertex_id_zero_base:
1900 case nir_intrinsic_load_base_vertex:
1901 case nir_intrinsic_load_instance_id:
1902 case nir_intrinsic_load_base_instance:
1903 case nir_intrinsic_load_draw_id: {
1904 gl_system_value sv = nir_system_value_from_intrinsic(instr->intrinsic);
1905 fs_reg val = nir_system_values[sv];
1906 assert(val.file != BAD_FILE);
1907 dest.type = val.type;
1913 nir_emit_intrinsic(bld, instr);
1919 fs_visitor::nir_emit_tes_intrinsic(const fs_builder &bld,
1920 nir_intrinsic_instr *instr)
1922 assert(stage == MESA_SHADER_TESS_EVAL);
1923 struct brw_tes_prog_data *tes_prog_data = (struct brw_tes_prog_data *) prog_data;
1926 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
1927 dest = get_nir_dest(instr->dest);
1929 switch (instr->intrinsic) {
1930 case nir_intrinsic_load_primitive_id:
1931 bld.MOV(dest, fs_reg(brw_vec1_grf(0, 1)));
1933 case nir_intrinsic_load_tess_coord:
1934 /* gl_TessCoord is part of the payload in g1-3 */
1935 for (unsigned i = 0; i < 3; i++) {
1936 bld.MOV(offset(dest, bld, i), fs_reg(brw_vec8_grf(1 + i, 0)));
1940 case nir_intrinsic_load_tess_level_outer:
1941 /* When the TES reads gl_TessLevelOuter, we ensure that the patch header
1942 * appears as a push-model input. So, we can simply use the ATTR file
1943 * rather than issuing URB read messages. The data is stored in the
1944 * high DWords in reverse order - DWord 7 contains .x, DWord 6 contains
1947 switch (tes_prog_data->domain) {
1948 case BRW_TESS_DOMAIN_QUAD:
1949 for (unsigned i = 0; i < 4; i++)
1950 bld.MOV(offset(dest, bld, i), component(fs_reg(ATTR, 0), 7 - i));
1952 case BRW_TESS_DOMAIN_TRI:
1953 for (unsigned i = 0; i < 3; i++)
1954 bld.MOV(offset(dest, bld, i), component(fs_reg(ATTR, 0), 7 - i));
1956 case BRW_TESS_DOMAIN_ISOLINE:
1957 for (unsigned i = 0; i < 2; i++)
1958 bld.MOV(offset(dest, bld, i), component(fs_reg(ATTR, 0), 7 - i));
1963 case nir_intrinsic_load_tess_level_inner:
1964 /* When the TES reads gl_TessLevelInner, we ensure that the patch header
1965 * appears as a push-model input. So, we can simply use the ATTR file
1966 * rather than issuing URB read messages.
1968 switch (tes_prog_data->domain) {
1969 case BRW_TESS_DOMAIN_QUAD:
1970 bld.MOV(dest, component(fs_reg(ATTR, 0), 3));
1971 bld.MOV(offset(dest, bld, 1), component(fs_reg(ATTR, 0), 2));
1973 case BRW_TESS_DOMAIN_TRI:
1974 bld.MOV(dest, component(fs_reg(ATTR, 0), 4));
1976 case BRW_TESS_DOMAIN_ISOLINE:
1977 /* ignore - value is undefined */
1982 case nir_intrinsic_load_input:
1983 case nir_intrinsic_load_per_vertex_input: {
1984 fs_reg indirect_offset = get_indirect_offset(instr);
1985 unsigned imm_offset = instr->const_index[0];
1988 if (indirect_offset.file == BAD_FILE) {
1989 /* Arbitrarily only push up to 32 vec4 slots worth of data,
1990 * which is 16 registers (since each holds 2 vec4 slots).
1992 const unsigned max_push_slots = 32;
1993 if (imm_offset < max_push_slots) {
1994 fs_reg src = fs_reg(ATTR, imm_offset / 2, dest.type);
1995 for (int i = 0; i < instr->num_components; i++) {
1996 bld.MOV(offset(dest, bld, i),
1997 component(src, 4 * (imm_offset % 2) + i));
1999 tes_prog_data->base.urb_read_length =
2000 MAX2(tes_prog_data->base.urb_read_length,
2001 DIV_ROUND_UP(imm_offset + 1, 2));
2003 /* Replicate the patch handle to all enabled channels */
2004 const fs_reg srcs[] = {
2005 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD)
2007 fs_reg patch_handle = bld.vgrf(BRW_REGISTER_TYPE_UD, 1);
2008 bld.LOAD_PAYLOAD(patch_handle, srcs, ARRAY_SIZE(srcs), 0);
2010 inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8, dest, patch_handle);
2012 inst->offset = imm_offset;
2013 inst->base_mrf = -1;
2014 inst->regs_written = instr->num_components;
2017 /* Indirect indexing - use per-slot offsets as well. */
2018 const fs_reg srcs[] = {
2019 retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD),
2022 fs_reg payload = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
2023 bld.LOAD_PAYLOAD(payload, srcs, ARRAY_SIZE(srcs), 0);
2025 inst = bld.emit(SHADER_OPCODE_URB_READ_SIMD8_PER_SLOT, dest, payload);
2027 inst->offset = imm_offset;
2028 inst->base_mrf = -1;
2029 inst->regs_written = instr->num_components;
2034 nir_emit_intrinsic(bld, instr);
2040 fs_visitor::nir_emit_gs_intrinsic(const fs_builder &bld,
2041 nir_intrinsic_instr *instr)
2043 assert(stage == MESA_SHADER_GEOMETRY);
2044 fs_reg indirect_offset;
2047 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
2048 dest = get_nir_dest(instr->dest);
2050 switch (instr->intrinsic) {
2051 case nir_intrinsic_load_primitive_id:
2052 assert(stage == MESA_SHADER_GEOMETRY);
2053 assert(((struct brw_gs_prog_data *)prog_data)->include_primitive_id);
2054 bld.MOV(retype(dest, BRW_REGISTER_TYPE_UD),
2055 retype(fs_reg(brw_vec8_grf(2, 0)), BRW_REGISTER_TYPE_UD));
2058 case nir_intrinsic_load_input:
2059 unreachable("load_input intrinsics are invalid for the GS stage");
2061 case nir_intrinsic_load_per_vertex_input:
2062 emit_gs_input_load(dest, instr->src[0], instr->const_index[0],
2063 instr->src[1], instr->num_components);
2066 case nir_intrinsic_emit_vertex_with_counter:
2067 emit_gs_vertex(instr->src[0], instr->const_index[0]);
2070 case nir_intrinsic_end_primitive_with_counter:
2071 emit_gs_end_primitive(instr->src[0]);
2074 case nir_intrinsic_set_vertex_count:
2075 bld.MOV(this->final_gs_vertex_count, get_nir_src(instr->src[0]));
2078 case nir_intrinsic_load_invocation_id: {
2079 fs_reg val = nir_system_values[SYSTEM_VALUE_INVOCATION_ID];
2080 assert(val.file != BAD_FILE);
2081 dest.type = val.type;
2087 nir_emit_intrinsic(bld, instr);
2093 fs_visitor::nir_emit_fs_intrinsic(const fs_builder &bld,
2094 nir_intrinsic_instr *instr)
2096 assert(stage == MESA_SHADER_FRAGMENT);
2097 struct brw_wm_prog_data *wm_prog_data =
2098 (struct brw_wm_prog_data *) prog_data;
2101 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
2102 dest = get_nir_dest(instr->dest);
2104 switch (instr->intrinsic) {
2105 case nir_intrinsic_load_front_face:
2106 bld.MOV(retype(dest, BRW_REGISTER_TYPE_D),
2107 *emit_frontfacing_interpolation());
2110 case nir_intrinsic_load_sample_pos: {
2111 fs_reg sample_pos = nir_system_values[SYSTEM_VALUE_SAMPLE_POS];
2112 assert(sample_pos.file != BAD_FILE);
2113 dest.type = sample_pos.type;
2114 bld.MOV(dest, sample_pos);
2115 bld.MOV(offset(dest, bld, 1), offset(sample_pos, bld, 1));
2119 case nir_intrinsic_load_helper_invocation:
2120 case nir_intrinsic_load_sample_mask_in:
2121 case nir_intrinsic_load_sample_id: {
2122 gl_system_value sv = nir_system_value_from_intrinsic(instr->intrinsic);
2123 fs_reg val = nir_system_values[sv];
2124 assert(val.file != BAD_FILE);
2125 dest.type = val.type;
2130 case nir_intrinsic_discard:
2131 case nir_intrinsic_discard_if: {
2132 /* We track our discarded pixels in f0.1. By predicating on it, we can
2133 * update just the flag bits that aren't yet discarded. If there's no
2134 * condition, we emit a CMP of g0 != g0, so all currently executing
2135 * channels will get turned off.
2138 if (instr->intrinsic == nir_intrinsic_discard_if) {
2139 cmp = bld.CMP(bld.null_reg_f(), get_nir_src(instr->src[0]),
2140 brw_imm_d(0), BRW_CONDITIONAL_Z);
2142 fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0),
2143 BRW_REGISTER_TYPE_UW));
2144 cmp = bld.CMP(bld.null_reg_f(), some_reg, some_reg, BRW_CONDITIONAL_NZ);
2146 cmp->predicate = BRW_PREDICATE_NORMAL;
2147 cmp->flag_subreg = 1;
2149 if (devinfo->gen >= 6) {
2150 emit_discard_jump();
2155 case nir_intrinsic_interp_var_at_centroid:
2156 case nir_intrinsic_interp_var_at_sample:
2157 case nir_intrinsic_interp_var_at_offset: {
2158 /* Handle ARB_gpu_shader5 interpolation intrinsics
2160 * It's worth a quick word of explanation as to why we handle the full
2161 * variable-based interpolation intrinsic rather than a lowered version
2162 * with like we do for other inputs. We have to do that because the way
2163 * we set up inputs doesn't allow us to use the already setup inputs for
2164 * interpolation. At the beginning of the shader, we go through all of
2165 * the input variables and do the initial interpolation and put it in
2166 * the nir_inputs array based on its location as determined in
2167 * nir_lower_io. If the input isn't used, dead code cleans up and
2168 * everything works fine. However, when we get to the ARB_gpu_shader5
2169 * interpolation intrinsics, we need to reinterpolate the input
2170 * differently. If we used an intrinsic that just had an index it would
2171 * only give us the offset into the nir_inputs array. However, this is
2172 * useless because that value is post-interpolation and we need
2173 * pre-interpolation. In order to get the actual location of the bits
2174 * we get from the vertex fetching hardware, we need the variable.
2176 wm_prog_data->pulls_bary = true;
2178 fs_reg dst_xy = bld.vgrf(BRW_REGISTER_TYPE_F, 2);
2179 const glsl_interp_qualifier interpolation =
2180 (glsl_interp_qualifier) instr->variables[0]->var->data.interpolation;
2182 switch (instr->intrinsic) {
2183 case nir_intrinsic_interp_var_at_centroid:
2184 emit_pixel_interpolater_send(bld,
2185 FS_OPCODE_INTERPOLATE_AT_CENTROID,
2192 case nir_intrinsic_interp_var_at_sample: {
2193 nir_const_value *const_sample = nir_src_as_const_value(instr->src[0]);
2196 unsigned msg_data = const_sample->i[0] << 4;
2198 emit_pixel_interpolater_send(bld,
2199 FS_OPCODE_INTERPOLATE_AT_SAMPLE,
2202 brw_imm_ud(msg_data),
2205 const fs_reg sample_src = retype(get_nir_src(instr->src[0]),
2206 BRW_REGISTER_TYPE_UD);
2208 if (nir_src_is_dynamically_uniform(instr->src[0])) {
2209 const fs_reg sample_id = bld.emit_uniformize(sample_src);
2210 const fs_reg msg_data = vgrf(glsl_type::uint_type);
2211 bld.exec_all().group(1, 0)
2212 .SHL(msg_data, sample_id, brw_imm_ud(4u));
2213 emit_pixel_interpolater_send(bld,
2214 FS_OPCODE_INTERPOLATE_AT_SAMPLE,
2220 /* Make a loop that sends a message to the pixel interpolater
2221 * for the sample number in each live channel. If there are
2222 * multiple channels with the same sample number then these
2223 * will be handled simultaneously with a single interation of
2226 bld.emit(BRW_OPCODE_DO);
2228 /* Get the next live sample number into sample_id_reg */
2229 const fs_reg sample_id = bld.emit_uniformize(sample_src);
2231 /* Set the flag register so that we can perform the send
2232 * message on all channels that have the same sample number
2234 bld.CMP(bld.null_reg_ud(),
2235 sample_src, sample_id,
2236 BRW_CONDITIONAL_EQ);
2237 const fs_reg msg_data = vgrf(glsl_type::uint_type);
2238 bld.exec_all().group(1, 0)
2239 .SHL(msg_data, sample_id, brw_imm_ud(4u));
2241 emit_pixel_interpolater_send(bld,
2242 FS_OPCODE_INTERPOLATE_AT_SAMPLE,
2247 set_predicate(BRW_PREDICATE_NORMAL, inst);
2249 /* Continue the loop if there are any live channels left */
2250 set_predicate_inv(BRW_PREDICATE_NORMAL,
2252 bld.emit(BRW_OPCODE_WHILE));
2259 case nir_intrinsic_interp_var_at_offset: {
2260 nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
2263 unsigned off_x = MIN2((int)(const_offset->f[0] * 16), 7) & 0xf;
2264 unsigned off_y = MIN2((int)(const_offset->f[1] * 16), 7) & 0xf;
2266 emit_pixel_interpolater_send(bld,
2267 FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET,
2270 brw_imm_ud(off_x | (off_y << 4)),
2273 fs_reg src = vgrf(glsl_type::ivec2_type);
2274 fs_reg offset_src = retype(get_nir_src(instr->src[0]),
2275 BRW_REGISTER_TYPE_F);
2276 for (int i = 0; i < 2; i++) {
2277 fs_reg temp = vgrf(glsl_type::float_type);
2278 bld.MUL(temp, offset(offset_src, bld, i), brw_imm_f(16.0f));
2279 fs_reg itemp = vgrf(glsl_type::int_type);
2280 bld.MOV(itemp, temp); /* float to int */
2282 /* Clamp the upper end of the range to +7/16.
2283 * ARB_gpu_shader5 requires that we support a maximum offset
2284 * of +0.5, which isn't representable in a S0.4 value -- if
2285 * we didn't clamp it, we'd end up with -8/16, which is the
2286 * opposite of what the shader author wanted.
2288 * This is legal due to ARB_gpu_shader5's quantization
2291 * "Not all values of <offset> may be supported; x and y
2292 * offsets may be rounded to fixed-point values with the
2293 * number of fraction bits given by the
2294 * implementation-dependent constant
2295 * FRAGMENT_INTERPOLATION_OFFSET_BITS"
2297 set_condmod(BRW_CONDITIONAL_L,
2298 bld.SEL(offset(src, bld, i), itemp, brw_imm_d(7)));
2301 const enum opcode opcode = FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET;
2302 emit_pixel_interpolater_send(bld,
2313 unreachable("Invalid intrinsic");
2316 for (unsigned j = 0; j < instr->num_components; j++) {
2317 fs_reg src = interp_reg(instr->variables[0]->var->data.location, j);
2318 src.type = dest.type;
2320 bld.emit(FS_OPCODE_LINTERP, dest, dst_xy, src);
2321 dest = offset(dest, bld, 1);
2326 nir_emit_intrinsic(bld, instr);
2332 fs_visitor::nir_emit_cs_intrinsic(const fs_builder &bld,
2333 nir_intrinsic_instr *instr)
2335 assert(stage == MESA_SHADER_COMPUTE);
2336 struct brw_cs_prog_data *cs_prog_data =
2337 (struct brw_cs_prog_data *) prog_data;
2340 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
2341 dest = get_nir_dest(instr->dest);
2343 switch (instr->intrinsic) {
2344 case nir_intrinsic_barrier:
2346 cs_prog_data->uses_barrier = true;
2349 case nir_intrinsic_load_local_invocation_id:
2350 case nir_intrinsic_load_work_group_id: {
2351 gl_system_value sv = nir_system_value_from_intrinsic(instr->intrinsic);
2352 fs_reg val = nir_system_values[sv];
2353 assert(val.file != BAD_FILE);
2354 dest.type = val.type;
2355 for (unsigned i = 0; i < 3; i++)
2356 bld.MOV(offset(dest, bld, i), offset(val, bld, i));
2360 case nir_intrinsic_load_num_work_groups: {
2361 const unsigned surface =
2362 cs_prog_data->binding_table.work_groups_start;
2364 cs_prog_data->uses_num_work_groups = true;
2366 fs_reg surf_index = brw_imm_ud(surface);
2367 brw_mark_surface_used(prog_data, surface);
2369 /* Read the 3 GLuint components of gl_NumWorkGroups */
2370 for (unsigned i = 0; i < 3; i++) {
2371 fs_reg read_result =
2372 emit_untyped_read(bld, surf_index,
2374 1 /* dims */, 1 /* size */,
2375 BRW_PREDICATE_NONE);
2376 read_result.type = dest.type;
2377 bld.MOV(dest, read_result);
2378 dest = offset(dest, bld, 1);
2383 case nir_intrinsic_shared_atomic_add:
2384 nir_emit_shared_atomic(bld, BRW_AOP_ADD, instr);
2386 case nir_intrinsic_shared_atomic_imin:
2387 nir_emit_shared_atomic(bld, BRW_AOP_IMIN, instr);
2389 case nir_intrinsic_shared_atomic_umin:
2390 nir_emit_shared_atomic(bld, BRW_AOP_UMIN, instr);
2392 case nir_intrinsic_shared_atomic_imax:
2393 nir_emit_shared_atomic(bld, BRW_AOP_IMAX, instr);
2395 case nir_intrinsic_shared_atomic_umax:
2396 nir_emit_shared_atomic(bld, BRW_AOP_UMAX, instr);
2398 case nir_intrinsic_shared_atomic_and:
2399 nir_emit_shared_atomic(bld, BRW_AOP_AND, instr);
2401 case nir_intrinsic_shared_atomic_or:
2402 nir_emit_shared_atomic(bld, BRW_AOP_OR, instr);
2404 case nir_intrinsic_shared_atomic_xor:
2405 nir_emit_shared_atomic(bld, BRW_AOP_XOR, instr);
2407 case nir_intrinsic_shared_atomic_exchange:
2408 nir_emit_shared_atomic(bld, BRW_AOP_MOV, instr);
2410 case nir_intrinsic_shared_atomic_comp_swap:
2411 nir_emit_shared_atomic(bld, BRW_AOP_CMPWR, instr);
2414 case nir_intrinsic_load_shared: {
2415 assert(devinfo->gen >= 7);
2417 fs_reg surf_index = brw_imm_ud(GEN7_BTI_SLM);
2419 /* Get the offset to read from */
2421 nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
2423 offset_reg = brw_imm_ud(instr->const_index[0] + const_offset->u[0]);
2425 offset_reg = vgrf(glsl_type::uint_type);
2427 retype(get_nir_src(instr->src[0]), BRW_REGISTER_TYPE_UD),
2428 brw_imm_ud(instr->const_index[0]));
2431 /* Read the vector */
2432 fs_reg read_result = emit_untyped_read(bld, surf_index, offset_reg,
2434 instr->num_components,
2435 BRW_PREDICATE_NONE);
2436 read_result.type = dest.type;
2437 for (int i = 0; i < instr->num_components; i++)
2438 bld.MOV(offset(dest, bld, i), offset(read_result, bld, i));
2443 case nir_intrinsic_store_shared: {
2444 assert(devinfo->gen >= 7);
2447 fs_reg surf_index = brw_imm_ud(GEN7_BTI_SLM);
2450 fs_reg val_reg = get_nir_src(instr->src[0]);
2453 unsigned writemask = instr->const_index[1];
2455 /* Combine groups of consecutive enabled channels in one write
2456 * message. We use ffs to find the first enabled channel and then ffs on
2457 * the bit-inverse, down-shifted writemask to determine the length of
2458 * the block of enabled bits.
2461 unsigned first_component = ffs(writemask) - 1;
2462 unsigned length = ffs(~(writemask >> first_component)) - 1;
2465 nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
2467 offset_reg = brw_imm_ud(instr->const_index[0] + const_offset->u[0] +
2468 4 * first_component);
2470 offset_reg = vgrf(glsl_type::uint_type);
2472 retype(get_nir_src(instr->src[1]), BRW_REGISTER_TYPE_UD),
2473 brw_imm_ud(instr->const_index[0] + 4 * first_component));
2476 emit_untyped_write(bld, surf_index, offset_reg,
2477 offset(val_reg, bld, first_component),
2478 1 /* dims */, length,
2479 BRW_PREDICATE_NONE);
2481 /* Clear the bits in the writemask that we just wrote, then try
2482 * again to see if more channels are left.
2484 writemask &= (15 << (first_component + length));
2491 nir_emit_intrinsic(bld, instr);
2497 fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr)
2500 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
2501 dest = get_nir_dest(instr->dest);
2503 switch (instr->intrinsic) {
2504 case nir_intrinsic_atomic_counter_inc:
2505 case nir_intrinsic_atomic_counter_dec:
2506 case nir_intrinsic_atomic_counter_read: {
2507 /* Get the arguments of the atomic intrinsic. */
2508 const fs_reg offset = get_nir_src(instr->src[0]);
2509 const unsigned surface = (stage_prog_data->binding_table.abo_start +
2510 instr->const_index[0]);
2513 /* Emit a surface read or atomic op. */
2514 switch (instr->intrinsic) {
2515 case nir_intrinsic_atomic_counter_read:
2516 tmp = emit_untyped_read(bld, brw_imm_ud(surface), offset, 1, 1);
2519 case nir_intrinsic_atomic_counter_inc:
2520 tmp = emit_untyped_atomic(bld, brw_imm_ud(surface), offset, fs_reg(),
2521 fs_reg(), 1, 1, BRW_AOP_INC);
2524 case nir_intrinsic_atomic_counter_dec:
2525 tmp = emit_untyped_atomic(bld, brw_imm_ud(surface), offset, fs_reg(),
2526 fs_reg(), 1, 1, BRW_AOP_PREDEC);
2530 unreachable("Unreachable");
2533 /* Assign the result. */
2534 bld.MOV(retype(dest, BRW_REGISTER_TYPE_UD), tmp);
2536 /* Mark the surface as used. */
2537 brw_mark_surface_used(stage_prog_data, surface);
2541 case nir_intrinsic_image_load:
2542 case nir_intrinsic_image_store:
2543 case nir_intrinsic_image_atomic_add:
2544 case nir_intrinsic_image_atomic_min:
2545 case nir_intrinsic_image_atomic_max:
2546 case nir_intrinsic_image_atomic_and:
2547 case nir_intrinsic_image_atomic_or:
2548 case nir_intrinsic_image_atomic_xor:
2549 case nir_intrinsic_image_atomic_exchange:
2550 case nir_intrinsic_image_atomic_comp_swap: {
2551 using namespace image_access;
2553 /* Get the referenced image variable and type. */
2554 const nir_variable *var = instr->variables[0]->var;
2555 const glsl_type *type = var->type->without_array();
2556 const brw_reg_type base_type = get_image_base_type(type);
2558 /* Get some metadata from the image intrinsic. */
2559 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
2560 const unsigned arr_dims = type->sampler_array ? 1 : 0;
2561 const unsigned surf_dims = type->coordinate_components() - arr_dims;
2562 const mesa_format format =
2563 (var->data.image.write_only ? MESA_FORMAT_NONE :
2564 _mesa_get_shader_image_format(var->data.image.format));
2566 /* Get the arguments of the image intrinsic. */
2567 const fs_reg image = get_nir_image_deref(instr->variables[0]);
2568 const fs_reg addr = retype(get_nir_src(instr->src[0]),
2569 BRW_REGISTER_TYPE_UD);
2570 const fs_reg src0 = (info->num_srcs >= 3 ?
2571 retype(get_nir_src(instr->src[2]), base_type) :
2573 const fs_reg src1 = (info->num_srcs >= 4 ?
2574 retype(get_nir_src(instr->src[3]), base_type) :
2578 /* Emit an image load, store or atomic op. */
2579 if (instr->intrinsic == nir_intrinsic_image_load)
2580 tmp = emit_image_load(bld, image, addr, surf_dims, arr_dims, format);
2582 else if (instr->intrinsic == nir_intrinsic_image_store)
2583 emit_image_store(bld, image, addr, src0, surf_dims, arr_dims, format);
2586 tmp = emit_image_atomic(bld, image, addr, src0, src1,
2587 surf_dims, arr_dims, info->dest_components,
2588 get_image_atomic_op(instr->intrinsic, type));
2590 /* Assign the result. */
2591 for (unsigned c = 0; c < info->dest_components; ++c)
2592 bld.MOV(offset(retype(dest, base_type), bld, c),
2593 offset(tmp, bld, c));
2597 case nir_intrinsic_memory_barrier_atomic_counter:
2598 case nir_intrinsic_memory_barrier_buffer:
2599 case nir_intrinsic_memory_barrier_image:
2600 case nir_intrinsic_memory_barrier: {
2601 const fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_UD, 16 / dispatch_width);
2602 bld.emit(SHADER_OPCODE_MEMORY_FENCE, tmp)
2607 case nir_intrinsic_group_memory_barrier:
2608 case nir_intrinsic_memory_barrier_shared:
2609 /* We treat these workgroup-level barriers as no-ops. This should be
2610 * safe at present and as long as:
2612 * - Memory access instructions are not subsequently reordered by the
2613 * compiler back-end.
2615 * - All threads from a given compute shader workgroup fit within a
2616 * single subslice and therefore talk to the same HDC shared unit
2617 * what supposedly guarantees ordering and coherency between threads
2618 * from the same workgroup. This may change in the future when we
2619 * start splitting workgroups across multiple subslices.
2621 * - The context is not in fault-and-stream mode, which could cause
2622 * memory transactions (including to SLM) prior to the barrier to be
2623 * replayed after the barrier if a pagefault occurs. This shouldn't
2624 * be a problem up to and including SKL because fault-and-stream is
2625 * not usable due to hardware issues, but that's likely to change in
2630 case nir_intrinsic_shader_clock: {
2631 /* We cannot do anything if there is an event, so ignore it for now */
2632 fs_reg shader_clock = get_timestamp(bld);
2633 const fs_reg srcs[] = { shader_clock.set_smear(0), shader_clock.set_smear(1) };
2635 bld.LOAD_PAYLOAD(dest, srcs, ARRAY_SIZE(srcs), 0);
2639 case nir_intrinsic_image_size: {
2640 /* Get the referenced image variable and type. */
2641 const nir_variable *var = instr->variables[0]->var;
2642 const glsl_type *type = var->type->without_array();
2644 /* Get the size of the image. */
2645 const fs_reg image = get_nir_image_deref(instr->variables[0]);
2646 const fs_reg size = offset(image, bld, BRW_IMAGE_PARAM_SIZE_OFFSET);
2648 /* For 1DArray image types, the array index is stored in the Z component.
2649 * Fix this by swizzling the Z component to the Y component.
2651 const bool is_1d_array_image =
2652 type->sampler_dimensionality == GLSL_SAMPLER_DIM_1D &&
2653 type->sampler_array;
2655 /* For CubeArray images, we should count the number of cubes instead
2656 * of the number of faces. Fix it by dividing the (Z component) by 6.
2658 const bool is_cube_array_image =
2659 type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE &&
2660 type->sampler_array;
2662 /* Copy all the components. */
2663 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
2664 for (unsigned c = 0; c < info->dest_components; ++c) {
2665 if ((int)c >= type->coordinate_components()) {
2666 bld.MOV(offset(retype(dest, BRW_REGISTER_TYPE_D), bld, c),
2668 } else if (c == 1 && is_1d_array_image) {
2669 bld.MOV(offset(retype(dest, BRW_REGISTER_TYPE_D), bld, c),
2670 offset(size, bld, 2));
2671 } else if (c == 2 && is_cube_array_image) {
2672 bld.emit(SHADER_OPCODE_INT_QUOTIENT,
2673 offset(retype(dest, BRW_REGISTER_TYPE_D), bld, c),
2674 offset(size, bld, c), brw_imm_d(6));
2676 bld.MOV(offset(retype(dest, BRW_REGISTER_TYPE_D), bld, c),
2677 offset(size, bld, c));
2684 case nir_intrinsic_image_samples:
2685 /* The driver does not support multi-sampled images. */
2686 bld.MOV(retype(dest, BRW_REGISTER_TYPE_D), brw_imm_d(1));
2689 case nir_intrinsic_load_uniform: {
2690 /* Offsets are in bytes but they should always be multiples of 4 */
2691 assert(instr->const_index[0] % 4 == 0);
2693 fs_reg src(UNIFORM, instr->const_index[0] / 4, dest.type);
2695 nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
2697 /* Offsets are in bytes but they should always be multiples of 4 */
2698 assert(const_offset->u[0] % 4 == 0);
2699 src.reg_offset = const_offset->u[0] / 4;
2701 for (unsigned j = 0; j < instr->num_components; j++) {
2702 bld.MOV(offset(dest, bld, j), offset(src, bld, j));
2705 fs_reg indirect = retype(get_nir_src(instr->src[0]),
2706 BRW_REGISTER_TYPE_UD);
2708 /* We need to pass a size to the MOV_INDIRECT but we don't want it to
2709 * go past the end of the uniform. In order to keep the n'th
2710 * component from running past, we subtract off the size of all but
2711 * one component of the vector.
2713 assert(instr->const_index[1] >= instr->num_components * 4);
2714 unsigned read_size = instr->const_index[1] -
2715 (instr->num_components - 1) * 4;
2717 for (unsigned j = 0; j < instr->num_components; j++) {
2718 bld.emit(SHADER_OPCODE_MOV_INDIRECT,
2719 offset(dest, bld, j), offset(src, bld, j),
2720 indirect, brw_imm_ud(read_size));
2726 case nir_intrinsic_load_ubo: {
2727 nir_const_value *const_index = nir_src_as_const_value(instr->src[0]);
2731 const unsigned index = stage_prog_data->binding_table.ubo_start +
2733 surf_index = brw_imm_ud(index);
2734 brw_mark_surface_used(prog_data, index);
2736 /* The block index is not a constant. Evaluate the index expression
2737 * per-channel and add the base UBO index; we have to select a value
2738 * from any live channel.
2740 surf_index = vgrf(glsl_type::uint_type);
2741 bld.ADD(surf_index, get_nir_src(instr->src[0]),
2742 brw_imm_ud(stage_prog_data->binding_table.ubo_start));
2743 surf_index = bld.emit_uniformize(surf_index);
2745 /* Assume this may touch any UBO. It would be nice to provide
2746 * a tighter bound, but the array information is already lowered away.
2748 brw_mark_surface_used(prog_data,
2749 stage_prog_data->binding_table.ubo_start +
2750 nir->info.num_ubos - 1);
2753 nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
2754 if (const_offset == NULL) {
2755 fs_reg base_offset = retype(get_nir_src(instr->src[1]),
2756 BRW_REGISTER_TYPE_D);
2758 for (int i = 0; i < instr->num_components; i++)
2759 VARYING_PULL_CONSTANT_LOAD(bld, offset(dest, bld, i), surf_index,
2760 base_offset, i * 4);
2762 fs_reg packed_consts = vgrf(glsl_type::float_type);
2763 packed_consts.type = dest.type;
2765 struct brw_reg const_offset_reg = brw_imm_ud(const_offset->u[0] & ~15);
2766 bld.emit(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD, packed_consts,
2767 surf_index, const_offset_reg);
2769 for (unsigned i = 0; i < instr->num_components; i++) {
2770 packed_consts.set_smear(const_offset->u[0] % 16 / 4 + i);
2772 /* The std140 packing rules don't allow vectors to cross 16-byte
2773 * boundaries, and a reg is 32 bytes.
2775 assert(packed_consts.subreg_offset < 32);
2777 bld.MOV(dest, packed_consts);
2778 dest = offset(dest, bld, 1);
2784 case nir_intrinsic_load_ssbo: {
2785 assert(devinfo->gen >= 7);
2787 nir_const_value *const_uniform_block =
2788 nir_src_as_const_value(instr->src[0]);
2791 if (const_uniform_block) {
2792 unsigned index = stage_prog_data->binding_table.ssbo_start +
2793 const_uniform_block->u[0];
2794 surf_index = brw_imm_ud(index);
2795 brw_mark_surface_used(prog_data, index);
2797 surf_index = vgrf(glsl_type::uint_type);
2798 bld.ADD(surf_index, get_nir_src(instr->src[0]),
2799 brw_imm_ud(stage_prog_data->binding_table.ssbo_start));
2801 /* Assume this may touch any UBO. It would be nice to provide
2802 * a tighter bound, but the array information is already lowered away.
2804 brw_mark_surface_used(prog_data,
2805 stage_prog_data->binding_table.ssbo_start +
2806 nir->info.num_ssbos - 1);
2810 nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
2812 offset_reg = brw_imm_ud(const_offset->u[0]);
2814 offset_reg = get_nir_src(instr->src[1]);
2817 /* Read the vector */
2818 fs_reg read_result = emit_untyped_read(bld, surf_index, offset_reg,
2820 instr->num_components,
2821 BRW_PREDICATE_NONE);
2822 read_result.type = dest.type;
2823 for (int i = 0; i < instr->num_components; i++)
2824 bld.MOV(offset(dest, bld, i), offset(read_result, bld, i));
2829 case nir_intrinsic_load_input: {
2831 if (stage == MESA_SHADER_VERTEX) {
2832 src = fs_reg(ATTR, instr->const_index[0], dest.type);
2834 src = offset(retype(nir_inputs, dest.type), bld,
2835 instr->const_index[0]);
2838 nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
2839 assert(const_offset && "Indirect input loads not allowed");
2840 src = offset(src, bld, const_offset->u[0]);
2842 for (unsigned j = 0; j < instr->num_components; j++) {
2843 bld.MOV(offset(dest, bld, j), offset(src, bld, j));
2848 case nir_intrinsic_store_ssbo: {
2849 assert(devinfo->gen >= 7);
2853 nir_const_value *const_uniform_block =
2854 nir_src_as_const_value(instr->src[1]);
2855 if (const_uniform_block) {
2856 unsigned index = stage_prog_data->binding_table.ssbo_start +
2857 const_uniform_block->u[0];
2858 surf_index = brw_imm_ud(index);
2859 brw_mark_surface_used(prog_data, index);
2861 surf_index = vgrf(glsl_type::uint_type);
2862 bld.ADD(surf_index, get_nir_src(instr->src[1]),
2863 brw_imm_ud(stage_prog_data->binding_table.ssbo_start));
2865 brw_mark_surface_used(prog_data,
2866 stage_prog_data->binding_table.ssbo_start +
2867 nir->info.num_ssbos - 1);
2871 fs_reg val_reg = get_nir_src(instr->src[0]);
2874 unsigned writemask = instr->const_index[0];
2876 /* Combine groups of consecutive enabled channels in one write
2877 * message. We use ffs to find the first enabled channel and then ffs on
2878 * the bit-inverse, down-shifted writemask to determine the length of
2879 * the block of enabled bits.
2882 unsigned first_component = ffs(writemask) - 1;
2883 unsigned length = ffs(~(writemask >> first_component)) - 1;
2886 nir_const_value *const_offset = nir_src_as_const_value(instr->src[2]);
2888 offset_reg = brw_imm_ud(const_offset->u[0] + 4 * first_component);
2890 offset_reg = vgrf(glsl_type::uint_type);
2892 retype(get_nir_src(instr->src[2]), BRW_REGISTER_TYPE_UD),
2893 brw_imm_ud(4 * first_component));
2896 emit_untyped_write(bld, surf_index, offset_reg,
2897 offset(val_reg, bld, first_component),
2898 1 /* dims */, length,
2899 BRW_PREDICATE_NONE);
2901 /* Clear the bits in the writemask that we just wrote, then try
2902 * again to see if more channels are left.
2904 writemask &= (15 << (first_component + length));
2909 case nir_intrinsic_store_output: {
2910 fs_reg src = get_nir_src(instr->src[0]);
2911 fs_reg new_dest = offset(retype(nir_outputs, src.type), bld,
2912 instr->const_index[0]);
2914 nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
2915 assert(const_offset && "Indirect output stores not allowed");
2916 new_dest = offset(new_dest, bld, const_offset->u[0]);
2918 for (unsigned j = 0; j < instr->num_components; j++) {
2919 bld.MOV(offset(new_dest, bld, j), offset(src, bld, j));
2924 case nir_intrinsic_ssbo_atomic_add:
2925 nir_emit_ssbo_atomic(bld, BRW_AOP_ADD, instr);
2927 case nir_intrinsic_ssbo_atomic_imin:
2928 nir_emit_ssbo_atomic(bld, BRW_AOP_IMIN, instr);
2930 case nir_intrinsic_ssbo_atomic_umin:
2931 nir_emit_ssbo_atomic(bld, BRW_AOP_UMIN, instr);
2933 case nir_intrinsic_ssbo_atomic_imax:
2934 nir_emit_ssbo_atomic(bld, BRW_AOP_IMAX, instr);
2936 case nir_intrinsic_ssbo_atomic_umax:
2937 nir_emit_ssbo_atomic(bld, BRW_AOP_UMAX, instr);
2939 case nir_intrinsic_ssbo_atomic_and:
2940 nir_emit_ssbo_atomic(bld, BRW_AOP_AND, instr);
2942 case nir_intrinsic_ssbo_atomic_or:
2943 nir_emit_ssbo_atomic(bld, BRW_AOP_OR, instr);
2945 case nir_intrinsic_ssbo_atomic_xor:
2946 nir_emit_ssbo_atomic(bld, BRW_AOP_XOR, instr);
2948 case nir_intrinsic_ssbo_atomic_exchange:
2949 nir_emit_ssbo_atomic(bld, BRW_AOP_MOV, instr);
2951 case nir_intrinsic_ssbo_atomic_comp_swap:
2952 nir_emit_ssbo_atomic(bld, BRW_AOP_CMPWR, instr);
2955 case nir_intrinsic_get_buffer_size: {
2956 nir_const_value *const_uniform_block = nir_src_as_const_value(instr->src[0]);
2957 unsigned ssbo_index = const_uniform_block ? const_uniform_block->u[0] : 0;
2958 int reg_width = dispatch_width / 8;
2961 fs_reg source = brw_imm_d(0);
2963 int mlen = 1 * reg_width;
2965 /* A resinfo's sampler message is used to get the buffer size.
2966 * The SIMD8's writeback message consists of four registers and
2967 * SIMD16's writeback message consists of 8 destination registers
2968 * (two per each component), although we are only interested on the
2969 * first component, where resinfo returns the buffer size for
2972 int regs_written = 4 * mlen;
2973 fs_reg src_payload = fs_reg(VGRF, alloc.allocate(mlen),
2974 BRW_REGISTER_TYPE_UD);
2975 bld.LOAD_PAYLOAD(src_payload, &source, 1, 0);
2976 fs_reg buffer_size = fs_reg(VGRF, alloc.allocate(regs_written),
2977 BRW_REGISTER_TYPE_UD);
2978 const unsigned index = prog_data->binding_table.ssbo_start + ssbo_index;
2979 fs_inst *inst = bld.emit(FS_OPCODE_GET_BUFFER_SIZE, buffer_size,
2980 src_payload, brw_imm_ud(index));
2981 inst->header_size = 0;
2983 inst->regs_written = regs_written;
2985 bld.MOV(retype(dest, buffer_size.type), buffer_size);
2987 brw_mark_surface_used(prog_data, index);
2992 unreachable("unknown intrinsic");
2997 fs_visitor::nir_emit_ssbo_atomic(const fs_builder &bld,
2998 int op, nir_intrinsic_instr *instr)
3001 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
3002 dest = get_nir_dest(instr->dest);
3005 nir_const_value *const_surface = nir_src_as_const_value(instr->src[0]);
3006 if (const_surface) {
3007 unsigned surf_index = stage_prog_data->binding_table.ssbo_start +
3008 const_surface->u[0];
3009 surface = brw_imm_ud(surf_index);
3010 brw_mark_surface_used(prog_data, surf_index);
3012 surface = vgrf(glsl_type::uint_type);
3013 bld.ADD(surface, get_nir_src(instr->src[0]),
3014 brw_imm_ud(stage_prog_data->binding_table.ssbo_start));
3016 /* Assume this may touch any SSBO. This is the same we do for other
3017 * UBO/SSBO accesses with non-constant surface.
3019 brw_mark_surface_used(prog_data,
3020 stage_prog_data->binding_table.ssbo_start +
3021 nir->info.num_ssbos - 1);
3024 fs_reg offset = get_nir_src(instr->src[1]);
3025 fs_reg data1 = get_nir_src(instr->src[2]);
3027 if (op == BRW_AOP_CMPWR)
3028 data2 = get_nir_src(instr->src[3]);
3030 /* Emit the actual atomic operation operation */
3032 fs_reg atomic_result = emit_untyped_atomic(bld, surface, offset,
3034 1 /* dims */, 1 /* rsize */,
3036 BRW_PREDICATE_NONE);
3037 dest.type = atomic_result.type;
3038 bld.MOV(dest, atomic_result);
3042 fs_visitor::nir_emit_shared_atomic(const fs_builder &bld,
3043 int op, nir_intrinsic_instr *instr)
3046 if (nir_intrinsic_infos[instr->intrinsic].has_dest)
3047 dest = get_nir_dest(instr->dest);
3049 fs_reg surface = brw_imm_ud(GEN7_BTI_SLM);
3050 fs_reg offset = get_nir_src(instr->src[0]);
3051 fs_reg data1 = get_nir_src(instr->src[1]);
3053 if (op == BRW_AOP_CMPWR)
3054 data2 = get_nir_src(instr->src[2]);
3056 /* Emit the actual atomic operation operation */
3058 fs_reg atomic_result = emit_untyped_atomic(bld, surface, offset,
3060 1 /* dims */, 1 /* rsize */,
3062 BRW_PREDICATE_NONE);
3063 dest.type = atomic_result.type;
3064 bld.MOV(dest, atomic_result);
3068 fs_visitor::nir_emit_texture(const fs_builder &bld, nir_tex_instr *instr)
3070 unsigned texture = instr->texture_index;
3071 unsigned sampler = instr->sampler_index;
3072 fs_reg texture_reg(brw_imm_ud(texture));
3073 fs_reg sampler_reg(brw_imm_ud(sampler));
3075 int gather_component = instr->component;
3077 bool is_cube_array = instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
3080 int lod_components = 0;
3082 fs_reg coordinate, shadow_comparitor, lod, lod2, sample_index, mcs, tex_offset;
3084 /* Our hardware requires a LOD for buffer textures */
3085 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
3088 for (unsigned i = 0; i < instr->num_srcs; i++) {
3089 fs_reg src = get_nir_src(instr->src[i].src);
3090 switch (instr->src[i].src_type) {
3091 case nir_tex_src_bias:
3092 lod = retype(src, BRW_REGISTER_TYPE_F);
3094 case nir_tex_src_comparitor:
3095 shadow_comparitor = retype(src, BRW_REGISTER_TYPE_F);
3097 case nir_tex_src_coord:
3098 switch (instr->op) {
3100 case nir_texop_txf_ms:
3101 case nir_texop_samples_identical:
3102 coordinate = retype(src, BRW_REGISTER_TYPE_D);
3105 coordinate = retype(src, BRW_REGISTER_TYPE_F);
3109 case nir_tex_src_ddx:
3110 lod = retype(src, BRW_REGISTER_TYPE_F);
3111 lod_components = nir_tex_instr_src_size(instr, i);
3113 case nir_tex_src_ddy:
3114 lod2 = retype(src, BRW_REGISTER_TYPE_F);
3116 case nir_tex_src_lod:
3117 switch (instr->op) {
3119 lod = retype(src, BRW_REGISTER_TYPE_UD);
3122 lod = retype(src, BRW_REGISTER_TYPE_D);
3125 lod = retype(src, BRW_REGISTER_TYPE_F);
3129 case nir_tex_src_ms_index:
3130 sample_index = retype(src, BRW_REGISTER_TYPE_UD);
3133 case nir_tex_src_offset: {
3134 nir_const_value *const_offset =
3135 nir_src_as_const_value(instr->src[i].src);
3137 tex_offset = brw_imm_ud(brw_texture_offset(const_offset->i, 3));
3139 tex_offset = retype(src, BRW_REGISTER_TYPE_D);
3144 case nir_tex_src_projector:
3145 unreachable("should be lowered");
3147 case nir_tex_src_texture_offset: {
3148 /* Figure out the highest possible texture index and mark it as used */
3149 uint32_t max_used = texture + instr->texture_array_size - 1;
3150 if (instr->op == nir_texop_tg4 && devinfo->gen < 8) {
3151 max_used += stage_prog_data->binding_table.gather_texture_start;
3153 max_used += stage_prog_data->binding_table.texture_start;
3155 brw_mark_surface_used(prog_data, max_used);
3157 /* Emit code to evaluate the actual indexing expression */
3158 texture_reg = vgrf(glsl_type::uint_type);
3159 bld.ADD(texture_reg, src, brw_imm_ud(texture));
3160 texture_reg = bld.emit_uniformize(texture_reg);
3164 case nir_tex_src_sampler_offset: {
3165 /* Emit code to evaluate the actual indexing expression */
3166 sampler_reg = vgrf(glsl_type::uint_type);
3167 bld.ADD(sampler_reg, src, brw_imm_ud(sampler));
3168 sampler_reg = bld.emit_uniformize(sampler_reg);
3173 unreachable("unknown texture source");
3177 if (instr->op == nir_texop_txf_ms ||
3178 instr->op == nir_texop_samples_identical) {
3179 if (devinfo->gen >= 7 &&
3180 key_tex->compressed_multisample_layout_mask & (1 << texture)) {
3181 mcs = emit_mcs_fetch(coordinate, instr->coord_components, texture_reg);
3183 mcs = brw_imm_ud(0u);
3187 enum glsl_base_type dest_base_type =
3188 brw_glsl_base_type_for_nir_type (instr->dest_type);
3190 const glsl_type *dest_type =
3191 glsl_type::get_instance(dest_base_type, nir_tex_instr_dest_size(instr),
3194 ir_texture_opcode op;
3195 switch (instr->op) {
3196 case nir_texop_lod: op = ir_lod; break;
3197 case nir_texop_query_levels: op = ir_query_levels; break;
3198 case nir_texop_tex: op = ir_tex; break;
3199 case nir_texop_tg4: op = ir_tg4; break;
3200 case nir_texop_txb: op = ir_txb; break;
3201 case nir_texop_txd: op = ir_txd; break;
3202 case nir_texop_txf: op = ir_txf; break;
3203 case nir_texop_txf_ms: op = ir_txf_ms; break;
3204 case nir_texop_txl: op = ir_txl; break;
3205 case nir_texop_txs: op = ir_txs; break;
3206 case nir_texop_texture_samples: {
3207 fs_reg dst = retype(get_nir_dest(instr->dest), BRW_REGISTER_TYPE_D);
3208 fs_inst *inst = bld.emit(SHADER_OPCODE_SAMPLEINFO, dst,
3209 bld.vgrf(BRW_REGISTER_TYPE_D, 1),
3210 texture_reg, texture_reg);
3212 inst->header_size = 1;
3213 inst->base_mrf = -1;
3216 case nir_texop_samples_identical: op = ir_samples_identical; break;
3218 unreachable("unknown texture opcode");
3221 emit_texture(op, dest_type, coordinate, instr->coord_components,
3222 shadow_comparitor, lod, lod2, lod_components, sample_index,
3223 tex_offset, mcs, gather_component, is_cube_array,
3224 texture, texture_reg, sampler, sampler_reg);
3226 fs_reg dest = get_nir_dest(instr->dest);
3227 dest.type = this->result.type;
3228 unsigned num_components = nir_tex_instr_dest_size(instr);
3229 emit_percomp(bld, fs_inst(BRW_OPCODE_MOV, bld.dispatch_width(),
3230 dest, this->result),
3231 (1 << num_components) - 1);
3235 fs_visitor::nir_emit_jump(const fs_builder &bld, nir_jump_instr *instr)
3237 switch (instr->type) {
3238 case nir_jump_break:
3239 bld.emit(BRW_OPCODE_BREAK);
3241 case nir_jump_continue:
3242 bld.emit(BRW_OPCODE_CONTINUE);
3244 case nir_jump_return:
3246 unreachable("unknown jump");