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 /** @file brw_fs_visitor.cpp
26 * This file supports generating the FS LIR from the GLSL IR. The LIR
27 * makes it easier to do backend-specific optimizations than doing so
28 * in the GLSL IR or in the native code.
32 #include <sys/types.h>
34 #include "main/macros.h"
35 #include "main/shaderobj.h"
36 #include "main/uniforms.h"
37 #include "program/prog_parameter.h"
38 #include "program/prog_print.h"
39 #include "program/prog_optimize.h"
40 #include "program/register_allocate.h"
41 #include "program/sampler.h"
42 #include "program/hash_table.h"
43 #include "brw_context.h"
48 #include "glsl/glsl_types.h"
49 #include "glsl/ir_optimization.h"
50 #include "glsl/ir_print_visitor.h"
53 fs_visitor::visit(ir_variable *ir)
57 if (variable_storage(ir))
60 if (ir->mode == ir_var_shader_in) {
61 if (!strcmp(ir->name, "gl_FragCoord")) {
62 reg = emit_fragcoord_interpolation(ir);
63 } else if (!strcmp(ir->name, "gl_FrontFacing")) {
64 reg = emit_frontfacing_interpolation(ir);
66 reg = emit_general_interpolation(ir);
69 hash_table_insert(this->variable_ht, reg, ir);
71 } else if (ir->mode == ir_var_shader_out) {
72 reg = new(this->mem_ctx) fs_reg(this, ir->type);
75 assert(ir->location == FRAG_RESULT_DATA0);
76 assert(ir->index == 1);
77 this->dual_src_output = *reg;
78 } else if (ir->location == FRAG_RESULT_COLOR) {
79 /* Writing gl_FragColor outputs to all color regions. */
80 for (unsigned int i = 0; i < MAX2(c->key.nr_color_regions, 1); i++) {
81 this->outputs[i] = *reg;
82 this->output_components[i] = 4;
84 } else if (ir->location == FRAG_RESULT_DEPTH) {
85 this->frag_depth = *reg;
87 /* gl_FragData or a user-defined FS output */
88 assert(ir->location >= FRAG_RESULT_DATA0 &&
89 ir->location < FRAG_RESULT_DATA0 + BRW_MAX_DRAW_BUFFERS);
92 ir->type->is_array() ? ir->type->fields.array->vector_elements
93 : ir->type->vector_elements;
95 /* General color output. */
96 for (unsigned int i = 0; i < MAX2(1, ir->type->length); i++) {
97 int output = ir->location - FRAG_RESULT_DATA0 + i;
98 this->outputs[output] = *reg;
99 this->outputs[output].reg_offset += vector_elements * i;
100 this->output_components[output] = vector_elements;
103 } else if (ir->mode == ir_var_uniform) {
104 int param_index = c->prog_data.nr_params;
106 /* Thanks to the lower_ubo_reference pass, we will see only
107 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
108 * variables, so no need for them to be in variable_ht.
110 if (ir->is_in_uniform_block())
113 if (dispatch_width == 16) {
114 if (!variable_storage(ir)) {
115 fail("Failed to find uniform '%s' in 16-wide\n", ir->name);
120 param_size[param_index] = type_size(ir->type);
121 if (!strncmp(ir->name, "gl_", 3)) {
122 setup_builtin_uniform_values(ir);
124 setup_uniform_values(ir);
127 reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
128 reg->type = brw_type_for_base_type(ir->type);
132 reg = new(this->mem_ctx) fs_reg(this, ir->type);
134 hash_table_insert(this->variable_ht, reg, ir);
138 fs_visitor::visit(ir_dereference_variable *ir)
140 fs_reg *reg = variable_storage(ir->var);
145 fs_visitor::visit(ir_dereference_record *ir)
147 const glsl_type *struct_type = ir->record->type;
149 ir->record->accept(this);
151 unsigned int offset = 0;
152 for (unsigned int i = 0; i < struct_type->length; i++) {
153 if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
155 offset += type_size(struct_type->fields.structure[i].type);
157 this->result.reg_offset += offset;
158 this->result.type = brw_type_for_base_type(ir->type);
162 fs_visitor::visit(ir_dereference_array *ir)
164 ir_constant *constant_index;
166 int element_size = type_size(ir->type);
168 constant_index = ir->array_index->as_constant();
170 ir->array->accept(this);
172 src.type = brw_type_for_base_type(ir->type);
174 if (constant_index) {
175 assert(src.file == UNIFORM || src.file == GRF);
176 src.reg_offset += constant_index->value.i[0] * element_size;
178 /* Variable index array dereference. We attach the variable index
179 * component to the reg as a pointer to a register containing the
180 * offset. Currently only uniform arrays are supported in this patch,
181 * and that reladdr pointer is resolved by
182 * move_uniform_array_access_to_pull_constants(). All other array types
183 * are lowered by lower_variable_index_to_cond_assign().
185 ir->array_index->accept(this);
188 index_reg = fs_reg(this, glsl_type::int_type);
189 emit(BRW_OPCODE_MUL, index_reg, this->result, fs_reg(element_size));
192 emit(BRW_OPCODE_ADD, index_reg, *src.reladdr, index_reg);
195 src.reladdr = ralloc(mem_ctx, fs_reg);
196 memcpy(src.reladdr, &index_reg, sizeof(index_reg));
202 fs_visitor::emit_lrp(fs_reg dst, fs_reg x, fs_reg y, fs_reg a)
204 if (intel->gen < 6 || x.file != GRF || y.file != GRF || a.file != GRF) {
205 /* We can't use the LRP instruction. Emit x*(1-a) + y*a. */
206 fs_reg y_times_a = fs_reg(this, glsl_type::float_type);
207 fs_reg one_minus_a = fs_reg(this, glsl_type::float_type);
208 fs_reg x_times_one_minus_a = fs_reg(this, glsl_type::float_type);
210 emit(MUL(y_times_a, y, a));
212 a.negate = !a.negate;
213 emit(ADD(one_minus_a, a, fs_reg(1.0f)));
214 emit(MUL(x_times_one_minus_a, x, one_minus_a));
216 emit(ADD(dst, x_times_one_minus_a, y_times_a));
218 /* The LRP instruction actually does op1 * op0 + op2 * (1 - op0), so
219 * we need to reorder the operands.
221 emit(LRP(dst, a, y, x));
226 fs_visitor::emit_minmax(uint32_t conditionalmod, fs_reg dst,
227 fs_reg src0, fs_reg src1)
231 if (intel->gen >= 6) {
232 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
233 inst->conditional_mod = conditionalmod;
235 emit(CMP(reg_null_d, src0, src1, conditionalmod));
237 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
238 inst->predicate = BRW_PREDICATE_NORMAL;
242 /* Instruction selection: Produce a MOV.sat instead of
243 * MIN(MAX(val, 0), 1) when possible.
246 fs_visitor::try_emit_saturate(ir_expression *ir)
248 ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
253 fs_inst *pre_inst = (fs_inst *) this->instructions.get_tail();
255 sat_val->accept(this);
256 fs_reg src = this->result;
258 fs_inst *last_inst = (fs_inst *) this->instructions.get_tail();
260 /* If the last instruction from our accept() didn't generate our
261 * src, generate a saturated MOV
263 fs_inst *modify = get_instruction_generating_reg(pre_inst, last_inst, src);
264 if (!modify || modify->regs_written() != 1) {
265 this->result = fs_reg(this, ir->type);
266 fs_inst *inst = emit(MOV(this->result, src));
267 inst->saturate = true;
269 modify->saturate = true;
278 fs_visitor::try_emit_mad(ir_expression *ir, int mul_arg)
280 /* 3-src instructions were introduced in gen6. */
284 /* MAD can only handle floating-point data. */
285 if (ir->type != glsl_type::float_type)
288 ir_rvalue *nonmul = ir->operands[1 - mul_arg];
289 ir_expression *mul = ir->operands[mul_arg]->as_expression();
291 if (!mul || mul->operation != ir_binop_mul)
294 if (nonmul->as_constant() ||
295 mul->operands[0]->as_constant() ||
296 mul->operands[1]->as_constant())
299 nonmul->accept(this);
300 fs_reg src0 = this->result;
302 mul->operands[0]->accept(this);
303 fs_reg src1 = this->result;
305 mul->operands[1]->accept(this);
306 fs_reg src2 = this->result;
308 this->result = fs_reg(this, ir->type);
309 emit(BRW_OPCODE_MAD, this->result, src0, src1, src2);
315 fs_visitor::visit(ir_expression *ir)
317 unsigned int operand;
321 assert(ir->get_num_operands() <= 3);
323 if (try_emit_saturate(ir))
325 if (ir->operation == ir_binop_add) {
326 if (try_emit_mad(ir, 0) || try_emit_mad(ir, 1))
330 for (operand = 0; operand < ir->get_num_operands(); operand++) {
331 ir->operands[operand]->accept(this);
332 if (this->result.file == BAD_FILE) {
334 fail("Failed to get tree for expression operand:\n");
335 ir->operands[operand]->accept(&v);
337 op[operand] = this->result;
339 /* Matrix expression operands should have been broken down to vector
340 * operations already.
342 assert(!ir->operands[operand]->type->is_matrix());
343 /* And then those vector operands should have been broken down to scalar.
345 assert(!ir->operands[operand]->type->is_vector());
348 /* Storage for our result. If our result goes into an assignment, it will
349 * just get copy-propagated out, so no worries.
351 this->result = fs_reg(this, ir->type);
353 switch (ir->operation) {
354 case ir_unop_logic_not:
355 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
356 * ones complement of the whole register, not just bit 0.
358 emit(XOR(this->result, op[0], fs_reg(1)));
361 op[0].negate = !op[0].negate;
362 this->result = op[0];
366 op[0].negate = false;
367 this->result = op[0];
370 temp = fs_reg(this, ir->type);
372 emit(MOV(this->result, fs_reg(0.0f)));
374 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_G));
375 inst = emit(MOV(this->result, fs_reg(1.0f)));
376 inst->predicate = BRW_PREDICATE_NORMAL;
378 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_L));
379 inst = emit(MOV(this->result, fs_reg(-1.0f)));
380 inst->predicate = BRW_PREDICATE_NORMAL;
384 emit_math(SHADER_OPCODE_RCP, this->result, op[0]);
388 emit_math(SHADER_OPCODE_EXP2, this->result, op[0]);
391 emit_math(SHADER_OPCODE_LOG2, this->result, op[0]);
395 assert(!"not reached: should be handled by ir_explog_to_explog2");
398 case ir_unop_sin_reduced:
399 emit_math(SHADER_OPCODE_SIN, this->result, op[0]);
402 case ir_unop_cos_reduced:
403 emit_math(SHADER_OPCODE_COS, this->result, op[0]);
407 emit(FS_OPCODE_DDX, this->result, op[0]);
410 emit(FS_OPCODE_DDY, this->result, op[0]);
414 emit(ADD(this->result, op[0], op[1]));
417 assert(!"not reached: should be handled by ir_sub_to_add_neg");
421 if (ir->type->is_integer()) {
422 /* For integer multiplication, the MUL uses the low 16 bits
423 * of one of the operands (src0 on gen6, src1 on gen7). The
424 * MACH accumulates in the contribution of the upper 16 bits
427 * FINISHME: Emit just the MUL if we know an operand is small
430 if (intel->gen >= 7 && dispatch_width == 16)
431 fail("16-wide explicit accumulator operands unsupported\n");
433 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_D);
435 emit(MUL(acc, op[0], op[1]));
436 emit(MACH(reg_null_d, op[0], op[1]));
437 emit(MOV(this->result, fs_reg(acc)));
439 emit(MUL(this->result, op[0], op[1]));
443 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
444 assert(ir->type->is_integer());
445 emit_math(SHADER_OPCODE_INT_QUOTIENT, this->result, op[0], op[1]);
448 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
449 assert(ir->type->is_integer());
450 emit_math(SHADER_OPCODE_INT_REMAINDER, this->result, op[0], op[1]);
454 case ir_binop_greater:
455 case ir_binop_lequal:
456 case ir_binop_gequal:
458 case ir_binop_all_equal:
459 case ir_binop_nequal:
460 case ir_binop_any_nequal:
461 resolve_bool_comparison(ir->operands[0], &op[0]);
462 resolve_bool_comparison(ir->operands[1], &op[1]);
464 emit(CMP(this->result, op[0], op[1],
465 brw_conditional_for_comparison(ir->operation)));
468 case ir_binop_logic_xor:
469 emit(XOR(this->result, op[0], op[1]));
472 case ir_binop_logic_or:
473 emit(OR(this->result, op[0], op[1]));
476 case ir_binop_logic_and:
477 emit(AND(this->result, op[0], op[1]));
482 assert(!"not reached: should be handled by brw_fs_channel_expressions");
486 assert(!"not reached: should be handled by lower_noise");
489 case ir_quadop_vector:
490 assert(!"not reached: should be handled by lower_quadop_vector");
494 emit_math(SHADER_OPCODE_SQRT, this->result, op[0]);
498 emit_math(SHADER_OPCODE_RSQ, this->result, op[0]);
501 case ir_unop_bitcast_i2f:
502 case ir_unop_bitcast_u2f:
503 op[0].type = BRW_REGISTER_TYPE_F;
504 this->result = op[0];
507 case ir_unop_bitcast_f2u:
508 op[0].type = BRW_REGISTER_TYPE_UD;
509 this->result = op[0];
512 case ir_unop_bitcast_f2i:
513 op[0].type = BRW_REGISTER_TYPE_D;
514 this->result = op[0];
520 emit(MOV(this->result, op[0]));
524 inst = emit(AND(this->result, op[0], fs_reg(1)));
527 temp = fs_reg(this, glsl_type::int_type);
528 emit(AND(temp, op[0], fs_reg(1)));
529 emit(MOV(this->result, temp));
533 emit(CMP(this->result, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
536 emit(CMP(this->result, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
540 emit(RNDZ(this->result, op[0]));
543 op[0].negate = !op[0].negate;
544 inst = emit(RNDD(this->result, op[0]));
545 this->result.negate = true;
548 inst = emit(RNDD(this->result, op[0]));
551 inst = emit(FRC(this->result, op[0]));
553 case ir_unop_round_even:
554 emit(RNDE(this->result, op[0]));
559 resolve_ud_negate(&op[0]);
560 resolve_ud_negate(&op[1]);
561 emit_minmax(ir->operation == ir_binop_min ?
562 BRW_CONDITIONAL_L : BRW_CONDITIONAL_GE,
563 this->result, op[0], op[1]);
565 case ir_unop_pack_snorm_2x16:
566 case ir_unop_pack_snorm_4x8:
567 case ir_unop_pack_unorm_2x16:
568 case ir_unop_pack_unorm_4x8:
569 case ir_unop_unpack_snorm_2x16:
570 case ir_unop_unpack_snorm_4x8:
571 case ir_unop_unpack_unorm_2x16:
572 case ir_unop_unpack_unorm_4x8:
573 case ir_unop_unpack_half_2x16:
574 case ir_unop_pack_half_2x16:
575 assert(!"not reached: should be handled by lower_packing_builtins");
577 case ir_unop_unpack_half_2x16_split_x:
578 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X, this->result, op[0]);
580 case ir_unop_unpack_half_2x16_split_y:
581 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y, this->result, op[0]);
584 emit_math(SHADER_OPCODE_POW, this->result, op[0], op[1]);
587 case ir_unop_bit_not:
588 inst = emit(NOT(this->result, op[0]));
590 case ir_binop_bit_and:
591 inst = emit(AND(this->result, op[0], op[1]));
593 case ir_binop_bit_xor:
594 inst = emit(XOR(this->result, op[0], op[1]));
596 case ir_binop_bit_or:
597 inst = emit(OR(this->result, op[0], op[1]));
600 case ir_binop_lshift:
601 inst = emit(SHL(this->result, op[0], op[1]));
604 case ir_binop_rshift:
605 if (ir->type->base_type == GLSL_TYPE_INT)
606 inst = emit(ASR(this->result, op[0], op[1]));
608 inst = emit(SHR(this->result, op[0], op[1]));
610 case ir_binop_pack_half_2x16_split:
611 emit(FS_OPCODE_PACK_HALF_2x16_SPLIT, this->result, op[0], op[1]);
613 case ir_binop_ubo_load: {
614 /* This IR node takes a constant uniform block and a constant or
615 * variable byte offset within the block and loads a vector from that.
617 ir_constant *uniform_block = ir->operands[0]->as_constant();
618 ir_constant *const_offset = ir->operands[1]->as_constant();
619 fs_reg surf_index = fs_reg((unsigned)SURF_INDEX_WM_UBO(uniform_block->value.u[0]));
621 fs_reg packed_consts = fs_reg(this, glsl_type::float_type);
622 packed_consts.type = result.type;
624 fs_reg const_offset_reg = fs_reg(const_offset->value.u[0] & ~15);
625 emit(fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
626 packed_consts, surf_index, const_offset_reg));
628 packed_consts.smear = const_offset->value.u[0] % 16 / 4;
629 for (int i = 0; i < ir->type->vector_elements; i++) {
630 /* UBO bools are any nonzero value. We consider bools to be
631 * values with the low bit set to 1. Convert them using CMP.
633 if (ir->type->base_type == GLSL_TYPE_BOOL) {
634 emit(CMP(result, packed_consts, fs_reg(0u), BRW_CONDITIONAL_NZ));
636 emit(MOV(result, packed_consts));
639 packed_consts.smear++;
642 /* The std140 packing rules don't allow vectors to cross 16-byte
643 * boundaries, and a reg is 32 bytes.
645 assert(packed_consts.smear < 8);
648 /* Turn the byte offset into a dword offset. */
649 fs_reg base_offset = fs_reg(this, glsl_type::int_type);
650 emit(SHR(base_offset, op[1], fs_reg(2)));
652 for (int i = 0; i < ir->type->vector_elements; i++) {
653 fs_reg offset = fs_reg(this, glsl_type::int_type);
654 emit(ADD(offset, base_offset, fs_reg(i)));
655 emit(VARYING_PULL_CONSTANT_LOAD(result, surf_index, offset));
657 if (ir->type->base_type == GLSL_TYPE_BOOL)
658 emit(CMP(result, result, fs_reg(0), BRW_CONDITIONAL_NZ));
664 result.reg_offset = 0;
669 emit_lrp(this->result, op[0], op[1], op[2]);
675 fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
676 const glsl_type *type, bool predicated)
678 switch (type->base_type) {
679 case GLSL_TYPE_FLOAT:
683 for (unsigned int i = 0; i < type->components(); i++) {
684 l.type = brw_type_for_base_type(type);
685 r.type = brw_type_for_base_type(type);
687 if (predicated || !l.equals(r)) {
688 fs_inst *inst = emit(MOV(l, r));
689 inst->predicate = predicated ? BRW_PREDICATE_NORMAL : BRW_PREDICATE_NONE;
696 case GLSL_TYPE_ARRAY:
697 for (unsigned int i = 0; i < type->length; i++) {
698 emit_assignment_writes(l, r, type->fields.array, predicated);
702 case GLSL_TYPE_STRUCT:
703 for (unsigned int i = 0; i < type->length; i++) {
704 emit_assignment_writes(l, r, type->fields.structure[i].type,
709 case GLSL_TYPE_SAMPLER:
713 case GLSL_TYPE_ERROR:
714 case GLSL_TYPE_INTERFACE:
715 assert(!"not reached");
720 /* If the RHS processing resulted in an instruction generating a
721 * temporary value, and it would be easy to rewrite the instruction to
722 * generate its result right into the LHS instead, do so. This ends
723 * up reliably removing instructions where it can be tricky to do so
724 * later without real UD chain information.
727 fs_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir,
730 fs_inst *pre_rhs_inst,
731 fs_inst *last_rhs_inst)
733 /* Only attempt if we're doing a direct assignment. */
735 !(ir->lhs->type->is_scalar() ||
736 (ir->lhs->type->is_vector() &&
737 ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1)))
740 /* Make sure the last instruction generated our source reg. */
741 fs_inst *modify = get_instruction_generating_reg(pre_rhs_inst,
747 /* If last_rhs_inst wrote a different number of components than our LHS,
748 * we can't safely rewrite it.
750 if (virtual_grf_sizes[dst.reg] != modify->regs_written())
753 /* Success! Rewrite the instruction. */
760 fs_visitor::visit(ir_assignment *ir)
765 /* FINISHME: arrays on the lhs */
766 ir->lhs->accept(this);
769 fs_inst *pre_rhs_inst = (fs_inst *) this->instructions.get_tail();
771 ir->rhs->accept(this);
774 fs_inst *last_rhs_inst = (fs_inst *) this->instructions.get_tail();
776 assert(l.file != BAD_FILE);
777 assert(r.file != BAD_FILE);
779 if (try_rewrite_rhs_to_dst(ir, l, r, pre_rhs_inst, last_rhs_inst))
783 emit_bool_to_cond_code(ir->condition);
786 if (ir->lhs->type->is_scalar() ||
787 ir->lhs->type->is_vector()) {
788 for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
789 if (ir->write_mask & (1 << i)) {
790 inst = emit(MOV(l, r));
792 inst->predicate = BRW_PREDICATE_NORMAL;
798 emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
803 fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
804 fs_reg shadow_c, fs_reg lod, fs_reg dPdy)
814 if (ir->shadow_comparitor) {
815 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
816 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
817 coordinate.reg_offset++;
819 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
822 if (ir->op == ir_tex) {
823 /* There's no plain shadow compare message, so we use shadow
824 * compare with a bias of 0.0.
826 emit(MOV(fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f)));
828 } else if (ir->op == ir_txb || ir->op == ir_txl) {
829 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
832 assert(!"Should not get here.");
835 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
837 } else if (ir->op == ir_tex) {
838 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
839 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
840 coordinate.reg_offset++;
842 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
844 } else if (ir->op == ir_txd) {
847 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
848 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
849 coordinate.reg_offset++;
851 /* the slots for u and v are always present, but r is optional */
852 mlen += MAX2(ir->coordinate->type->vector_elements, 2);
855 * dPdx = dudx, dvdx, drdx
856 * dPdy = dudy, dvdy, drdy
858 * 1-arg: Does not exist.
860 * 2-arg: dudx dvdx dudy dvdy
861 * dPdx.x dPdx.y dPdy.x dPdy.y
864 * 3-arg: dudx dvdx drdx dudy dvdy drdy
865 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
868 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
869 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdx));
872 mlen += MAX2(ir->lod_info.grad.dPdx->type->vector_elements, 2);
874 for (int i = 0; i < ir->lod_info.grad.dPdy->type->vector_elements; i++) {
875 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdy));
878 mlen += MAX2(ir->lod_info.grad.dPdy->type->vector_elements, 2);
879 } else if (ir->op == ir_txs) {
880 /* There's no SIMD8 resinfo message on Gen4. Use SIMD16 instead. */
882 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
885 /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
886 * instructions. We'll need to do SIMD16 here.
889 assert(ir->op == ir_txb || ir->op == ir_txl || ir->op == ir_txf);
891 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
892 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2, coordinate.type),
894 coordinate.reg_offset++;
897 /* Initialize the rest of u/v/r with 0.0. Empirically, this seems to
898 * be necessary for TXF (ld), but seems wise to do for all messages.
900 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
901 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2), fs_reg(0.0f)));
904 /* lod/bias appears after u/v/r. */
907 emit(MOV(fs_reg(MRF, base_mrf + mlen, lod.type), lod));
910 /* The unused upper half. */
915 /* Now, since we're doing simd16, the return is 2 interleaved
916 * vec4s where the odd-indexed ones are junk. We'll need to move
917 * this weirdness around to the expected layout.
920 const glsl_type *vec_type =
921 glsl_type::get_instance(ir->type->base_type, 4, 1);
922 dst = fs_reg(this, glsl_type::get_array_instance(vec_type, 2));
923 dst.type = intel->is_g4x ? brw_type_for_base_type(ir->type)
924 : BRW_REGISTER_TYPE_F;
927 fs_inst *inst = NULL;
930 inst = emit(SHADER_OPCODE_TEX, dst);
933 inst = emit(FS_OPCODE_TXB, dst);
936 inst = emit(SHADER_OPCODE_TXL, dst);
939 inst = emit(SHADER_OPCODE_TXD, dst);
942 inst = emit(SHADER_OPCODE_TXS, dst);
945 inst = emit(SHADER_OPCODE_TXF, dst);
948 fail("unrecognized texture opcode");
950 inst->base_mrf = base_mrf;
952 inst->header_present = true;
955 for (int i = 0; i < 4; i++) {
956 emit(MOV(orig_dst, dst));
957 orig_dst.reg_offset++;
965 /* gen5's sampler has slots for u, v, r, array index, then optional
966 * parameters like shadow comparitor or LOD bias. If optional
967 * parameters aren't present, those base slots are optional and don't
968 * need to be included in the message.
970 * We don't fill in the unnecessary slots regardless, which may look
971 * surprising in the disassembly.
974 fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
975 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
980 int reg_width = dispatch_width / 8;
981 bool header_present = false;
982 const int vector_elements =
983 ir->coordinate ? ir->coordinate->type->vector_elements : 0;
985 if (ir->offset != NULL && ir->op == ir_txf) {
986 /* It appears that the ld instruction used for txf does its
987 * address bounds check before adding in the offset. To work
988 * around this, just add the integer offset to the integer texel
989 * coordinate, and don't put the offset in the header.
991 ir_constant *offset = ir->offset->as_constant();
992 for (int i = 0; i < vector_elements; i++) {
993 emit(ADD(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
995 offset->value.i[i]));
996 coordinate.reg_offset++;
1000 /* The offsets set up by the ir_texture visitor are in the
1001 * m1 header, so we can't go headerless.
1003 header_present = true;
1008 for (int i = 0; i < vector_elements; i++) {
1009 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
1011 coordinate.reg_offset++;
1014 mlen += vector_elements * reg_width;
1016 if (ir->shadow_comparitor) {
1017 mlen = MAX2(mlen, header_present + 4 * reg_width);
1019 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
1023 fs_inst *inst = NULL;
1026 inst = emit(SHADER_OPCODE_TEX, dst);
1029 mlen = MAX2(mlen, header_present + 4 * reg_width);
1030 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1033 inst = emit(FS_OPCODE_TXB, dst);
1036 mlen = MAX2(mlen, header_present + 4 * reg_width);
1037 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1040 inst = emit(SHADER_OPCODE_TXL, dst);
1043 mlen = MAX2(mlen, header_present + 4 * reg_width); /* skip over 'ai' */
1047 * dPdx = dudx, dvdx, drdx
1048 * dPdy = dudy, dvdy, drdy
1050 * Load up these values:
1051 * - dudx dudy dvdx dvdy drdx drdy
1052 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
1054 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
1055 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1059 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
1064 inst = emit(SHADER_OPCODE_TXD, dst);
1068 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1070 inst = emit(SHADER_OPCODE_TXS, dst);
1073 mlen = header_present + 4 * reg_width;
1074 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), lod));
1075 inst = emit(SHADER_OPCODE_TXF, dst);
1078 mlen = header_present + 4 * reg_width;
1081 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), fs_reg(0)));
1083 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), sample_index));
1085 inst = emit(SHADER_OPCODE_TXF_MS, dst);
1088 inst = emit(SHADER_OPCODE_LOD, dst);
1091 inst->base_mrf = base_mrf;
1093 inst->header_present = header_present;
1096 fail("Message length >11 disallowed by hardware\n");
1103 fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1104 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
1105 fs_reg sample_index)
1109 int reg_width = dispatch_width / 8;
1110 bool header_present = false;
1113 if (ir->offset && ir->op != ir_txf) {
1114 /* The offsets set up by the ir_texture visitor are in the
1115 * m1 header, so we can't go headerless.
1117 header_present = true;
1122 if (ir->shadow_comparitor) {
1123 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
1127 /* Set up the LOD info */
1133 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1137 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1141 if (dispatch_width == 16)
1142 fail("Gen7 does not support sample_d/sample_d_c in SIMD16 mode.");
1144 /* Load dPdx and the coordinate together:
1145 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
1147 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1148 emit(MOV(fs_reg(MRF, base_mrf + mlen), coordinate));
1149 coordinate.reg_offset++;
1152 /* For cube map array, the coordinate is (u,v,r,ai) but there are
1153 * only derivatives for (u, v, r).
1155 if (i < ir->lod_info.grad.dPdx->type->vector_elements) {
1156 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1160 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
1168 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1172 /* It appears that the ld instruction used for txf does its
1173 * address bounds check before adding in the offset. To work
1174 * around this, just add the integer offset to the integer texel
1175 * coordinate, and don't put the offset in the header.
1178 ir_constant *offset = ir->offset->as_constant();
1179 offsets[0] = offset->value.i[0];
1180 offsets[1] = offset->value.i[1];
1181 offsets[2] = offset->value.i[2];
1183 memset(offsets, 0, sizeof(offsets));
1186 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r. */
1187 emit(ADD(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
1188 coordinate, offsets[0]));
1189 coordinate.reg_offset++;
1192 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D), lod));
1195 for (int i = 1; i < ir->coordinate->type->vector_elements; i++) {
1196 emit(ADD(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
1197 coordinate, offsets[i]));
1198 coordinate.reg_offset++;
1203 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), sample_index));
1206 /* constant zero MCS; we arrange to never actually have a compressed
1207 * multisample surface here for now. TODO: issue ld_mcs to get this first,
1208 * if we ever support texturing from compressed multisample surfaces
1210 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1213 /* there is no offsetting for this message; just copy in the integer
1214 * texture coordinates
1216 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1217 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
1219 coordinate.reg_offset++;
1225 /* Set up the coordinate (except for cases where it was done above) */
1226 if (ir->op != ir_txd && ir->op != ir_txs && ir->op != ir_txf && ir->op != ir_txf_ms) {
1227 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1228 emit(MOV(fs_reg(MRF, base_mrf + mlen), coordinate));
1229 coordinate.reg_offset++;
1234 /* Generate the SEND */
1235 fs_inst *inst = NULL;
1237 case ir_tex: inst = emit(SHADER_OPCODE_TEX, dst); break;
1238 case ir_txb: inst = emit(FS_OPCODE_TXB, dst); break;
1239 case ir_txl: inst = emit(SHADER_OPCODE_TXL, dst); break;
1240 case ir_txd: inst = emit(SHADER_OPCODE_TXD, dst); break;
1241 case ir_txf: inst = emit(SHADER_OPCODE_TXF, dst); break;
1242 case ir_txf_ms: inst = emit(SHADER_OPCODE_TXF_MS, dst); break;
1243 case ir_txs: inst = emit(SHADER_OPCODE_TXS, dst); break;
1244 case ir_lod: inst = emit(SHADER_OPCODE_LOD, dst); break;
1246 inst->base_mrf = base_mrf;
1248 inst->header_present = header_present;
1251 fail("Message length >11 disallowed by hardware\n");
1258 fs_visitor::rescale_texcoord(ir_texture *ir, fs_reg coordinate,
1259 bool is_rect, int sampler, int texunit)
1261 fs_inst *inst = NULL;
1262 bool needs_gl_clamp = true;
1263 fs_reg scale_x, scale_y;
1265 /* The 965 requires the EU to do the normalization of GL rectangle
1266 * texture coordinates. We use the program parameter state
1267 * tracking to get the scaling factor.
1271 (intel->gen >= 6 && (c->key.tex.gl_clamp_mask[0] & (1 << sampler) ||
1272 c->key.tex.gl_clamp_mask[1] & (1 << sampler))))) {
1273 struct gl_program_parameter_list *params = fp->Base.Parameters;
1274 int tokens[STATE_LENGTH] = {
1276 STATE_TEXRECT_SCALE,
1282 if (dispatch_width == 16) {
1283 fail("rectangle scale uniform setup not supported on 16-wide\n");
1287 scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
1288 scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
1290 GLuint index = _mesa_add_state_reference(params,
1291 (gl_state_index *)tokens);
1292 c->prog_data.param[c->prog_data.nr_params++] =
1293 &fp->Base.Parameters->ParameterValues[index][0].f;
1294 c->prog_data.param[c->prog_data.nr_params++] =
1295 &fp->Base.Parameters->ParameterValues[index][1].f;
1298 /* The 965 requires the EU to do the normalization of GL rectangle
1299 * texture coordinates. We use the program parameter state
1300 * tracking to get the scaling factor.
1302 if (intel->gen < 6 && is_rect) {
1303 fs_reg dst = fs_reg(this, ir->coordinate->type);
1304 fs_reg src = coordinate;
1307 emit(MUL(dst, src, scale_x));
1310 emit(MUL(dst, src, scale_y));
1311 } else if (is_rect) {
1312 /* On gen6+, the sampler handles the rectangle coordinates
1313 * natively, without needing rescaling. But that means we have
1314 * to do GL_CLAMP clamping at the [0, width], [0, height] scale,
1315 * not [0, 1] like the default case below.
1317 needs_gl_clamp = false;
1319 for (int i = 0; i < 2; i++) {
1320 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1321 fs_reg chan = coordinate;
1322 chan.reg_offset += i;
1324 inst = emit(BRW_OPCODE_SEL, chan, chan, brw_imm_f(0.0));
1325 inst->conditional_mod = BRW_CONDITIONAL_G;
1327 /* Our parameter comes in as 1.0/width or 1.0/height,
1328 * because that's what people normally want for doing
1329 * texture rectangle handling. We need width or height
1330 * for clamping, but we don't care enough to make a new
1331 * parameter type, so just invert back.
1333 fs_reg limit = fs_reg(this, glsl_type::float_type);
1334 emit(MOV(limit, i == 0 ? scale_x : scale_y));
1335 emit(SHADER_OPCODE_RCP, limit, limit);
1337 inst = emit(BRW_OPCODE_SEL, chan, chan, limit);
1338 inst->conditional_mod = BRW_CONDITIONAL_L;
1343 if (ir->coordinate && needs_gl_clamp) {
1344 for (unsigned int i = 0;
1345 i < MIN2(ir->coordinate->type->vector_elements, 3); i++) {
1346 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1347 fs_reg chan = coordinate;
1348 chan.reg_offset += i;
1350 fs_inst *inst = emit(MOV(chan, chan));
1351 inst->saturate = true;
1359 fs_visitor::visit(ir_texture *ir)
1361 fs_inst *inst = NULL;
1363 int sampler = _mesa_get_sampler_uniform_value(ir->sampler, prog, &fp->Base);
1364 /* FINISHME: We're failing to recompile our programs when the sampler is
1365 * updated. This only matters for the texture rectangle scale parameters
1366 * (pre-gen6, or gen6+ with GL_CLAMP).
1368 int texunit = fp->Base.SamplerUnits[sampler];
1370 /* Should be lowered by do_lower_texture_projection */
1371 assert(!ir->projector);
1373 /* Generate code to compute all the subexpression trees. This has to be
1374 * done before loading any values into MRFs for the sampler message since
1375 * generating these values may involve SEND messages that need the MRFs.
1378 if (ir->coordinate) {
1379 ir->coordinate->accept(this);
1381 coordinate = rescale_texcoord(ir, this->result,
1382 ir->sampler->type->sampler_dimensionality ==
1383 GLSL_SAMPLER_DIM_RECT,
1387 fs_reg shadow_comparitor;
1388 if (ir->shadow_comparitor) {
1389 ir->shadow_comparitor->accept(this);
1390 shadow_comparitor = this->result;
1393 fs_reg lod, lod2, sample_index;
1399 ir->lod_info.bias->accept(this);
1403 ir->lod_info.grad.dPdx->accept(this);
1406 ir->lod_info.grad.dPdy->accept(this);
1407 lod2 = this->result;
1412 ir->lod_info.lod->accept(this);
1416 ir->lod_info.sample_index->accept(this);
1417 sample_index = this->result;
1421 /* Writemasking doesn't eliminate channels on SIMD8 texture
1422 * samples, so don't worry about them.
1424 fs_reg dst = fs_reg(this, glsl_type::get_instance(ir->type->base_type, 4, 1));
1426 if (intel->gen >= 7) {
1427 inst = emit_texture_gen7(ir, dst, coordinate, shadow_comparitor,
1428 lod, lod2, sample_index);
1429 } else if (intel->gen >= 5) {
1430 inst = emit_texture_gen5(ir, dst, coordinate, shadow_comparitor,
1431 lod, lod2, sample_index);
1433 inst = emit_texture_gen4(ir, dst, coordinate, shadow_comparitor,
1437 /* The header is set up by generate_tex() when necessary. */
1438 inst->src[0] = reg_undef;
1440 if (ir->offset != NULL && ir->op != ir_txf)
1441 inst->texture_offset = brw_texture_offset(ir->offset->as_constant());
1443 inst->sampler = sampler;
1445 if (ir->shadow_comparitor)
1446 inst->shadow_compare = true;
1448 /* fixup #layers for cube map arrays */
1449 if (ir->op == ir_txs) {
1450 glsl_type const *type = ir->sampler->type;
1451 if (type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE &&
1452 type->sampler_array) {
1454 depth.reg_offset = 2;
1455 emit_math(SHADER_OPCODE_INT_QUOTIENT, depth, depth, fs_reg(6));
1459 swizzle_result(ir, dst, sampler);
1463 * Swizzle the result of a texture result. This is necessary for
1464 * EXT_texture_swizzle as well as DEPTH_TEXTURE_MODE for shadow comparisons.
1467 fs_visitor::swizzle_result(ir_texture *ir, fs_reg orig_val, int sampler)
1469 this->result = orig_val;
1471 if (ir->op == ir_txs)
1474 if (ir->type == glsl_type::float_type) {
1475 /* Ignore DEPTH_TEXTURE_MODE swizzling. */
1476 assert(ir->sampler->type->sampler_shadow);
1477 } else if (c->key.tex.swizzles[sampler] != SWIZZLE_NOOP) {
1478 fs_reg swizzled_result = fs_reg(this, glsl_type::vec4_type);
1480 for (int i = 0; i < 4; i++) {
1481 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], i);
1482 fs_reg l = swizzled_result;
1485 if (swiz == SWIZZLE_ZERO) {
1486 emit(MOV(l, fs_reg(0.0f)));
1487 } else if (swiz == SWIZZLE_ONE) {
1488 emit(MOV(l, fs_reg(1.0f)));
1490 fs_reg r = orig_val;
1491 r.reg_offset += GET_SWZ(c->key.tex.swizzles[sampler], i);
1495 this->result = swizzled_result;
1500 fs_visitor::visit(ir_swizzle *ir)
1502 ir->val->accept(this);
1503 fs_reg val = this->result;
1505 if (ir->type->vector_elements == 1) {
1506 this->result.reg_offset += ir->mask.x;
1510 fs_reg result = fs_reg(this, ir->type);
1511 this->result = result;
1513 for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
1514 fs_reg channel = val;
1532 channel.reg_offset += swiz;
1533 emit(MOV(result, channel));
1534 result.reg_offset++;
1539 fs_visitor::visit(ir_discard *ir)
1541 assert(ir->condition == NULL); /* FINISHME */
1543 /* We track our discarded pixels in f0.1. By predicating on it, we can
1544 * update just the flag bits that aren't yet discarded. By emitting a
1545 * CMP of g0 != g0, all our currently executing channels will get turned
1548 fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0),
1549 BRW_REGISTER_TYPE_UW));
1550 fs_inst *cmp = emit(CMP(reg_null_f, some_reg, some_reg,
1551 BRW_CONDITIONAL_NZ));
1552 cmp->predicate = BRW_PREDICATE_NORMAL;
1553 cmp->flag_subreg = 1;
1555 if (intel->gen >= 6) {
1556 /* For performance, after a discard, jump to the end of the shader.
1557 * However, many people will do foliage by discarding based on a
1558 * texture's alpha mask, and then continue on to texture with the
1559 * remaining pixels. To avoid trashing the derivatives for those
1560 * texture samples, we'll only jump if all of the pixels in the subspan
1561 * have been discarded.
1563 fs_inst *discard_jump = emit(FS_OPCODE_DISCARD_JUMP);
1564 discard_jump->flag_subreg = 1;
1565 discard_jump->predicate = BRW_PREDICATE_ALIGN1_ANY4H;
1566 discard_jump->predicate_inverse = true;
1571 fs_visitor::visit(ir_constant *ir)
1573 /* Set this->result to reg at the bottom of the function because some code
1574 * paths will cause this visitor to be applied to other fields. This will
1575 * cause the value stored in this->result to be modified.
1577 * Make reg constant so that it doesn't get accidentally modified along the
1578 * way. Yes, I actually had this problem. :(
1580 const fs_reg reg(this, ir->type);
1581 fs_reg dst_reg = reg;
1583 if (ir->type->is_array()) {
1584 const unsigned size = type_size(ir->type->fields.array);
1586 for (unsigned i = 0; i < ir->type->length; i++) {
1587 ir->array_elements[i]->accept(this);
1588 fs_reg src_reg = this->result;
1590 dst_reg.type = src_reg.type;
1591 for (unsigned j = 0; j < size; j++) {
1592 emit(MOV(dst_reg, src_reg));
1593 src_reg.reg_offset++;
1594 dst_reg.reg_offset++;
1597 } else if (ir->type->is_record()) {
1598 foreach_list(node, &ir->components) {
1599 ir_constant *const field = (ir_constant *) node;
1600 const unsigned size = type_size(field->type);
1602 field->accept(this);
1603 fs_reg src_reg = this->result;
1605 dst_reg.type = src_reg.type;
1606 for (unsigned j = 0; j < size; j++) {
1607 emit(MOV(dst_reg, src_reg));
1608 src_reg.reg_offset++;
1609 dst_reg.reg_offset++;
1613 const unsigned size = type_size(ir->type);
1615 for (unsigned i = 0; i < size; i++) {
1616 switch (ir->type->base_type) {
1617 case GLSL_TYPE_FLOAT:
1618 emit(MOV(dst_reg, fs_reg(ir->value.f[i])));
1620 case GLSL_TYPE_UINT:
1621 emit(MOV(dst_reg, fs_reg(ir->value.u[i])));
1624 emit(MOV(dst_reg, fs_reg(ir->value.i[i])));
1626 case GLSL_TYPE_BOOL:
1627 emit(MOV(dst_reg, fs_reg((int)ir->value.b[i])));
1630 assert(!"Non-float/uint/int/bool constant");
1632 dst_reg.reg_offset++;
1640 fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
1642 ir_expression *expr = ir->as_expression();
1648 assert(expr->get_num_operands() <= 2);
1649 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1650 assert(expr->operands[i]->type->is_scalar());
1652 expr->operands[i]->accept(this);
1653 op[i] = this->result;
1655 resolve_ud_negate(&op[i]);
1658 switch (expr->operation) {
1659 case ir_unop_logic_not:
1660 inst = emit(AND(reg_null_d, op[0], fs_reg(1)));
1661 inst->conditional_mod = BRW_CONDITIONAL_Z;
1664 case ir_binop_logic_xor:
1665 case ir_binop_logic_or:
1666 case ir_binop_logic_and:
1670 if (intel->gen >= 6) {
1671 emit(CMP(reg_null_d, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
1673 inst = emit(MOV(reg_null_f, op[0]));
1674 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1679 if (intel->gen >= 6) {
1680 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1682 inst = emit(MOV(reg_null_d, op[0]));
1683 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1687 case ir_binop_greater:
1688 case ir_binop_gequal:
1690 case ir_binop_lequal:
1691 case ir_binop_equal:
1692 case ir_binop_all_equal:
1693 case ir_binop_nequal:
1694 case ir_binop_any_nequal:
1695 resolve_bool_comparison(expr->operands[0], &op[0]);
1696 resolve_bool_comparison(expr->operands[1], &op[1]);
1698 emit(CMP(reg_null_d, op[0], op[1],
1699 brw_conditional_for_comparison(expr->operation)));
1703 assert(!"not reached");
1704 fail("bad cond code\n");
1713 fs_inst *inst = emit(AND(reg_null_d, this->result, fs_reg(1)));
1714 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1718 * Emit a gen6 IF statement with the comparison folded into the IF
1722 fs_visitor::emit_if_gen6(ir_if *ir)
1724 ir_expression *expr = ir->condition->as_expression();
1731 assert(expr->get_num_operands() <= 2);
1732 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1733 assert(expr->operands[i]->type->is_scalar());
1735 expr->operands[i]->accept(this);
1736 op[i] = this->result;
1739 switch (expr->operation) {
1740 case ir_unop_logic_not:
1741 case ir_binop_logic_xor:
1742 case ir_binop_logic_or:
1743 case ir_binop_logic_and:
1744 /* For operations on bool arguments, only the low bit of the bool is
1745 * valid, and the others are undefined. Fall back to the condition
1751 inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
1752 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1756 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1759 case ir_binop_greater:
1760 case ir_binop_gequal:
1762 case ir_binop_lequal:
1763 case ir_binop_equal:
1764 case ir_binop_all_equal:
1765 case ir_binop_nequal:
1766 case ir_binop_any_nequal:
1767 resolve_bool_comparison(expr->operands[0], &op[0]);
1768 resolve_bool_comparison(expr->operands[1], &op[1]);
1770 emit(IF(op[0], op[1],
1771 brw_conditional_for_comparison(expr->operation)));
1774 assert(!"not reached");
1775 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1776 fail("bad condition\n");
1781 emit_bool_to_cond_code(ir->condition);
1782 fs_inst *inst = emit(BRW_OPCODE_IF);
1783 inst->predicate = BRW_PREDICATE_NORMAL;
1787 fs_visitor::visit(ir_if *ir)
1789 if (intel->gen < 6 && dispatch_width == 16) {
1790 fail("Can't support (non-uniform) control flow on 16-wide\n");
1793 /* Don't point the annotation at the if statement, because then it plus
1794 * the then and else blocks get printed.
1796 this->base_ir = ir->condition;
1798 if (intel->gen == 6) {
1801 emit_bool_to_cond_code(ir->condition);
1803 emit(IF(BRW_PREDICATE_NORMAL));
1806 foreach_list(node, &ir->then_instructions) {
1807 ir_instruction *ir = (ir_instruction *)node;
1813 if (!ir->else_instructions.is_empty()) {
1814 emit(BRW_OPCODE_ELSE);
1816 foreach_list(node, &ir->else_instructions) {
1817 ir_instruction *ir = (ir_instruction *)node;
1824 emit(BRW_OPCODE_ENDIF);
1828 fs_visitor::visit(ir_loop *ir)
1830 fs_reg counter = reg_undef;
1832 if (intel->gen < 6 && dispatch_width == 16) {
1833 fail("Can't support (non-uniform) control flow on 16-wide\n");
1837 this->base_ir = ir->counter;
1838 ir->counter->accept(this);
1839 counter = *(variable_storage(ir->counter));
1842 this->base_ir = ir->from;
1843 ir->from->accept(this);
1845 emit(MOV(counter, this->result));
1849 this->base_ir = NULL;
1850 emit(BRW_OPCODE_DO);
1853 this->base_ir = ir->to;
1854 ir->to->accept(this);
1856 emit(CMP(reg_null_d, counter, this->result,
1857 brw_conditional_for_comparison(ir->cmp)));
1859 fs_inst *inst = emit(BRW_OPCODE_BREAK);
1860 inst->predicate = BRW_PREDICATE_NORMAL;
1863 foreach_list(node, &ir->body_instructions) {
1864 ir_instruction *ir = (ir_instruction *)node;
1870 if (ir->increment) {
1871 this->base_ir = ir->increment;
1872 ir->increment->accept(this);
1873 emit(ADD(counter, counter, this->result));
1876 this->base_ir = NULL;
1877 emit(BRW_OPCODE_WHILE);
1881 fs_visitor::visit(ir_loop_jump *ir)
1884 case ir_loop_jump::jump_break:
1885 emit(BRW_OPCODE_BREAK);
1887 case ir_loop_jump::jump_continue:
1888 emit(BRW_OPCODE_CONTINUE);
1894 fs_visitor::visit(ir_call *ir)
1896 assert(!"FINISHME");
1900 fs_visitor::visit(ir_return *ir)
1902 assert(!"FINISHME");
1906 fs_visitor::visit(ir_function *ir)
1908 /* Ignore function bodies other than main() -- we shouldn't see calls to
1909 * them since they should all be inlined before we get to ir_to_mesa.
1911 if (strcmp(ir->name, "main") == 0) {
1912 const ir_function_signature *sig;
1915 sig = ir->matching_signature(&empty);
1919 foreach_list(node, &sig->body) {
1920 ir_instruction *ir = (ir_instruction *)node;
1929 fs_visitor::visit(ir_function_signature *ir)
1931 assert(!"not reached");
1936 fs_visitor::emit(fs_inst inst)
1938 fs_inst *list_inst = new(mem_ctx) fs_inst;
1945 fs_visitor::emit(fs_inst *inst)
1947 if (force_uncompressed_stack > 0)
1948 inst->force_uncompressed = true;
1949 else if (force_sechalf_stack > 0)
1950 inst->force_sechalf = true;
1952 inst->annotation = this->current_annotation;
1953 inst->ir = this->base_ir;
1955 this->instructions.push_tail(inst);
1961 fs_visitor::emit(exec_list list)
1963 foreach_list_safe(node, &list) {
1964 fs_inst *inst = (fs_inst *)node;
1970 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
1972 fs_visitor::emit_dummy_fs()
1974 int reg_width = dispatch_width / 8;
1976 /* Everyone's favorite color. */
1977 emit(MOV(fs_reg(MRF, 2 + 0 * reg_width), fs_reg(1.0f)));
1978 emit(MOV(fs_reg(MRF, 2 + 1 * reg_width), fs_reg(0.0f)));
1979 emit(MOV(fs_reg(MRF, 2 + 2 * reg_width), fs_reg(1.0f)));
1980 emit(MOV(fs_reg(MRF, 2 + 3 * reg_width), fs_reg(0.0f)));
1983 write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
1984 write->base_mrf = 2;
1985 write->mlen = 4 * reg_width;
1989 /* The register location here is relative to the start of the URB
1990 * data. It will get adjusted to be a real location before
1991 * generate_code() time.
1994 fs_visitor::interp_reg(int location, int channel)
1996 int regnr = urb_setup[location] * 2 + channel / 2;
1997 int stride = (channel & 1) * 4;
1999 assert(urb_setup[location] != -1);
2001 return brw_vec1_grf(regnr, stride);
2004 /** Emits the interpolation for the varying inputs. */
2006 fs_visitor::emit_interpolation_setup_gen4()
2008 this->current_annotation = "compute pixel centers";
2009 this->pixel_x = fs_reg(this, glsl_type::uint_type);
2010 this->pixel_y = fs_reg(this, glsl_type::uint_type);
2011 this->pixel_x.type = BRW_REGISTER_TYPE_UW;
2012 this->pixel_y.type = BRW_REGISTER_TYPE_UW;
2014 emit(FS_OPCODE_PIXEL_X, this->pixel_x);
2015 emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
2017 this->current_annotation = "compute pixel deltas from v0";
2019 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2020 fs_reg(this, glsl_type::vec2_type);
2021 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2022 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC];
2023 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg_offset++;
2025 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2026 fs_reg(this, glsl_type::float_type);
2027 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2028 fs_reg(this, glsl_type::float_type);
2030 emit(ADD(this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2031 this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0)))));
2032 emit(ADD(this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2033 this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1)))));
2035 this->current_annotation = "compute pos.w and 1/pos.w";
2036 /* Compute wpos.w. It's always in our setup, since it's needed to
2037 * interpolate the other attributes.
2039 this->wpos_w = fs_reg(this, glsl_type::float_type);
2040 emit(FS_OPCODE_LINTERP, wpos_w,
2041 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2042 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2043 interp_reg(VARYING_SLOT_POS, 3));
2044 /* Compute the pixel 1/W value from wpos.w. */
2045 this->pixel_w = fs_reg(this, glsl_type::float_type);
2046 emit_math(SHADER_OPCODE_RCP, this->pixel_w, wpos_w);
2047 this->current_annotation = NULL;
2050 /** Emits the interpolation for the varying inputs. */
2052 fs_visitor::emit_interpolation_setup_gen6()
2054 struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
2056 /* If the pixel centers end up used, the setup is the same as for gen4. */
2057 this->current_annotation = "compute pixel centers";
2058 fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
2059 fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
2060 int_pixel_x.type = BRW_REGISTER_TYPE_UW;
2061 int_pixel_y.type = BRW_REGISTER_TYPE_UW;
2062 emit(ADD(int_pixel_x,
2063 fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
2064 fs_reg(brw_imm_v(0x10101010))));
2065 emit(ADD(int_pixel_y,
2066 fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
2067 fs_reg(brw_imm_v(0x11001100))));
2069 /* As of gen6, we can no longer mix float and int sources. We have
2070 * to turn the integer pixel centers into floats for their actual
2073 this->pixel_x = fs_reg(this, glsl_type::float_type);
2074 this->pixel_y = fs_reg(this, glsl_type::float_type);
2075 emit(MOV(this->pixel_x, int_pixel_x));
2076 emit(MOV(this->pixel_y, int_pixel_y));
2078 this->current_annotation = "compute pos.w";
2079 this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
2080 this->wpos_w = fs_reg(this, glsl_type::float_type);
2081 emit_math(SHADER_OPCODE_RCP, this->wpos_w, this->pixel_w);
2083 for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
2084 uint8_t reg = c->barycentric_coord_reg[i];
2085 this->delta_x[i] = fs_reg(brw_vec8_grf(reg, 0));
2086 this->delta_y[i] = fs_reg(brw_vec8_grf(reg + 1, 0));
2089 this->current_annotation = NULL;
2093 fs_visitor::emit_color_write(int target, int index, int first_color_mrf)
2095 int reg_width = dispatch_width / 8;
2097 fs_reg color = outputs[target];
2100 /* If there's no color data to be written, skip it. */
2101 if (color.file == BAD_FILE)
2104 color.reg_offset += index;
2106 if (dispatch_width == 8 || intel->gen >= 6) {
2107 /* SIMD8 write looks like:
2113 * gen6 SIMD16 DP write looks like:
2123 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index * reg_width,
2126 inst->saturate = c->key.clamp_fragment_color;
2128 /* pre-gen6 SIMD16 single source DP write looks like:
2138 if (brw->has_compr4) {
2139 /* By setting the high bit of the MRF register number, we
2140 * indicate that we want COMPR4 mode - instead of doing the
2141 * usual destination + 1 for the second half we get
2144 inst = emit(MOV(fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index,
2147 inst->saturate = c->key.clamp_fragment_color;
2149 push_force_uncompressed();
2150 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index, color.type),
2152 inst->saturate = c->key.clamp_fragment_color;
2153 pop_force_uncompressed();
2155 push_force_sechalf();
2156 color.sechalf = true;
2157 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index + 4, color.type),
2159 inst->saturate = c->key.clamp_fragment_color;
2160 pop_force_sechalf();
2161 color.sechalf = false;
2167 fs_visitor::emit_fb_writes()
2169 this->current_annotation = "FB write header";
2170 bool header_present = true;
2171 /* We can potentially have a message length of up to 15, so we have to set
2172 * base_mrf to either 0 or 1 in order to fit in m0..m15.
2176 int reg_width = dispatch_width / 8;
2177 bool do_dual_src = this->dual_src_output.file != BAD_FILE;
2178 bool src0_alpha_to_render_target = false;
2180 if (dispatch_width == 16 && do_dual_src) {
2181 fail("GL_ARB_blend_func_extended not yet supported in 16-wide.");
2182 do_dual_src = false;
2185 /* From the Sandy Bridge PRM, volume 4, page 198:
2187 * "Dispatched Pixel Enables. One bit per pixel indicating
2188 * which pixels were originally enabled when the thread was
2189 * dispatched. This field is only required for the end-of-
2190 * thread message and on all dual-source messages."
2192 if (intel->gen >= 6 &&
2193 !this->fp->UsesKill &&
2195 c->key.nr_color_regions == 1) {
2196 header_present = false;
2199 if (header_present) {
2200 src0_alpha_to_render_target = intel->gen >= 6 &&
2202 c->key.nr_color_regions > 1 &&
2203 c->key.sample_alpha_to_coverage;
2208 if (c->aa_dest_stencil_reg) {
2209 push_force_uncompressed();
2210 emit(MOV(fs_reg(MRF, nr++),
2211 fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0))));
2212 pop_force_uncompressed();
2215 /* Reserve space for color. It'll be filled in per MRT below. */
2217 nr += 4 * reg_width;
2220 if (src0_alpha_to_render_target)
2223 if (c->source_depth_to_render_target) {
2224 if (intel->gen == 6 && dispatch_width == 16) {
2225 /* For outputting oDepth on gen6, SIMD8 writes have to be
2226 * used. This would require 8-wide moves of each half to
2227 * message regs, kind of like pre-gen5 SIMD16 FB writes.
2228 * Just bail on doing so for now.
2230 fail("Missing support for simd16 depth writes on gen6\n");
2233 if (fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
2234 /* Hand over gl_FragDepth. */
2235 assert(this->frag_depth.file != BAD_FILE);
2236 emit(MOV(fs_reg(MRF, nr), this->frag_depth));
2238 /* Pass through the payload depth. */
2239 emit(MOV(fs_reg(MRF, nr),
2240 fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
2245 if (c->dest_depth_reg) {
2246 emit(MOV(fs_reg(MRF, nr),
2247 fs_reg(brw_vec8_grf(c->dest_depth_reg, 0))));
2252 fs_reg src0 = this->outputs[0];
2253 fs_reg src1 = this->dual_src_output;
2255 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2257 for (int i = 0; i < 4; i++) {
2258 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + i, src0.type), src0));
2260 inst->saturate = c->key.clamp_fragment_color;
2263 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2265 for (int i = 0; i < 4; i++) {
2266 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + 4 + i, src1.type),
2269 inst->saturate = c->key.clamp_fragment_color;
2272 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2273 emit_shader_time_end();
2275 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2277 inst->base_mrf = base_mrf;
2278 inst->mlen = nr - base_mrf;
2280 inst->header_present = header_present;
2282 c->prog_data.dual_src_blend = true;
2283 this->current_annotation = NULL;
2287 for (int target = 0; target < c->key.nr_color_regions; target++) {
2288 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2289 "FB write target %d",
2291 /* If src0_alpha_to_render_target is true, include source zero alpha
2292 * data in RenderTargetWrite message for targets > 0.
2294 int write_color_mrf = color_mrf;
2295 if (src0_alpha_to_render_target && target != 0) {
2297 fs_reg color = outputs[0];
2298 color.reg_offset += 3;
2300 inst = emit(MOV(fs_reg(MRF, write_color_mrf, color.type),
2302 inst->saturate = c->key.clamp_fragment_color;
2303 write_color_mrf = color_mrf + reg_width;
2306 for (unsigned i = 0; i < this->output_components[target]; i++)
2307 emit_color_write(target, i, write_color_mrf);
2310 if (target == c->key.nr_color_regions - 1) {
2313 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2314 emit_shader_time_end();
2317 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2318 inst->target = target;
2319 inst->base_mrf = base_mrf;
2320 if (src0_alpha_to_render_target && target == 0)
2321 inst->mlen = nr - base_mrf - reg_width;
2323 inst->mlen = nr - base_mrf;
2325 inst->header_present = header_present;
2328 if (c->key.nr_color_regions == 0) {
2329 /* Even if there's no color buffers enabled, we still need to send
2330 * alpha out the pipeline to our null renderbuffer to support
2331 * alpha-testing, alpha-to-coverage, and so on.
2333 emit_color_write(0, 3, color_mrf);
2335 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2336 emit_shader_time_end();
2338 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2339 inst->base_mrf = base_mrf;
2340 inst->mlen = nr - base_mrf;
2342 inst->header_present = header_present;
2345 this->current_annotation = NULL;
2349 fs_visitor::resolve_ud_negate(fs_reg *reg)
2351 if (reg->type != BRW_REGISTER_TYPE_UD ||
2355 fs_reg temp = fs_reg(this, glsl_type::uint_type);
2356 emit(MOV(temp, *reg));
2361 fs_visitor::resolve_bool_comparison(ir_rvalue *rvalue, fs_reg *reg)
2363 if (rvalue->type != glsl_type::bool_type)
2366 fs_reg temp = fs_reg(this, glsl_type::bool_type);
2367 emit(AND(temp, *reg, fs_reg(1)));
2371 fs_visitor::fs_visitor(struct brw_context *brw,
2372 struct brw_wm_compile *c,
2373 struct gl_shader_program *prog,
2374 struct gl_fragment_program *fp,
2375 unsigned dispatch_width)
2376 : dispatch_width(dispatch_width)
2382 this->intel = &brw->intel;
2383 this->ctx = &intel->ctx;
2384 this->mem_ctx = ralloc_context(NULL);
2386 shader = (struct brw_shader *) prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2389 this->failed = false;
2390 this->variable_ht = hash_table_ctor(0,
2391 hash_table_pointer_hash,
2392 hash_table_pointer_compare);
2394 memset(this->outputs, 0, sizeof(this->outputs));
2395 memset(this->output_components, 0, sizeof(this->output_components));
2396 this->first_non_payload_grf = 0;
2397 this->max_grf = intel->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF;
2399 this->current_annotation = NULL;
2400 this->base_ir = NULL;
2402 this->virtual_grf_sizes = NULL;
2403 this->virtual_grf_count = 0;
2404 this->virtual_grf_array_size = 0;
2405 this->virtual_grf_def = NULL;
2406 this->virtual_grf_use = NULL;
2407 this->live_intervals_valid = false;
2409 this->force_uncompressed_stack = 0;
2410 this->force_sechalf_stack = 0;
2412 memset(&this->param_size, 0, sizeof(this->param_size));
2415 fs_visitor::~fs_visitor()
2417 ralloc_free(this->mem_ctx);
2418 hash_table_dtor(this->variable_ht);