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 "program/prog_parameter.h"
37 #include "program/prog_print.h"
38 #include "program/prog_optimize.h"
39 #include "program/register_allocate.h"
40 #include "program/sampler.h"
41 #include "program/hash_table.h"
42 #include "brw_context.h"
47 #include "main/uniforms.h"
48 #include "glsl/glsl_types.h"
49 #include "glsl/ir_optimization.h"
52 fs_visitor::visit(ir_variable *ir)
56 if (variable_storage(ir))
59 if (ir->data.mode == ir_var_shader_in) {
60 if (!strcmp(ir->name, "gl_FragCoord")) {
61 reg = emit_fragcoord_interpolation(ir);
62 } else if (!strcmp(ir->name, "gl_FrontFacing")) {
63 reg = emit_frontfacing_interpolation(ir);
65 reg = emit_general_interpolation(ir);
68 hash_table_insert(this->variable_ht, reg, ir);
70 } else if (ir->data.mode == ir_var_shader_out) {
71 reg = new(this->mem_ctx) fs_reg(this, ir->type);
73 if (ir->data.index > 0) {
74 assert(ir->data.location == FRAG_RESULT_DATA0);
75 assert(ir->data.index == 1);
76 this->dual_src_output = *reg;
77 } else if (ir->data.location == FRAG_RESULT_COLOR) {
78 /* Writing gl_FragColor outputs to all color regions. */
79 for (unsigned int i = 0; i < MAX2(c->key.nr_color_regions, 1); i++) {
80 this->outputs[i] = *reg;
81 this->output_components[i] = 4;
83 } else if (ir->data.location == FRAG_RESULT_DEPTH) {
84 this->frag_depth = *reg;
85 } else if (ir->data.location == FRAG_RESULT_SAMPLE_MASK) {
86 this->sample_mask = *reg;
88 /* gl_FragData or a user-defined FS output */
89 assert(ir->data.location >= FRAG_RESULT_DATA0 &&
90 ir->data.location < FRAG_RESULT_DATA0 + BRW_MAX_DRAW_BUFFERS);
93 ir->type->is_array() ? ir->type->fields.array->vector_elements
94 : ir->type->vector_elements;
96 /* General color output. */
97 for (unsigned int i = 0; i < MAX2(1, ir->type->length); i++) {
98 int output = ir->data.location - FRAG_RESULT_DATA0 + i;
99 this->outputs[output] = *reg;
100 this->outputs[output].reg_offset += vector_elements * i;
101 this->output_components[output] = vector_elements;
104 } else if (ir->data.mode == ir_var_uniform) {
105 int param_index = c->prog_data.nr_params;
107 /* Thanks to the lower_ubo_reference pass, we will see only
108 * ir_binop_ubo_load expressions and not ir_dereference_variable for UBO
109 * variables, so no need for them to be in variable_ht.
111 * Atomic counters take no uniform storage, no need to do
114 if (ir->is_in_uniform_block() || ir->type->contains_atomic())
117 if (dispatch_width == 16) {
118 if (!variable_storage(ir)) {
119 fail("Failed to find uniform '%s' in SIMD16\n", ir->name);
124 param_size[param_index] = type_size(ir->type);
125 if (!strncmp(ir->name, "gl_", 3)) {
126 setup_builtin_uniform_values(ir);
128 setup_uniform_values(ir);
131 reg = new(this->mem_ctx) fs_reg(UNIFORM, param_index);
132 reg->type = brw_type_for_base_type(ir->type);
134 } else if (ir->data.mode == ir_var_system_value) {
135 if (ir->data.location == SYSTEM_VALUE_SAMPLE_POS) {
136 reg = emit_samplepos_setup(ir);
137 } else if (ir->data.location == SYSTEM_VALUE_SAMPLE_ID) {
138 reg = emit_sampleid_setup(ir);
139 } else if (ir->data.location == SYSTEM_VALUE_SAMPLE_MASK_IN) {
140 reg = emit_samplemaskin_setup(ir);
145 reg = new(this->mem_ctx) fs_reg(this, ir->type);
147 hash_table_insert(this->variable_ht, reg, ir);
151 fs_visitor::visit(ir_dereference_variable *ir)
153 fs_reg *reg = variable_storage(ir->var);
158 fs_visitor::visit(ir_dereference_record *ir)
160 const glsl_type *struct_type = ir->record->type;
162 ir->record->accept(this);
164 unsigned int offset = 0;
165 for (unsigned int i = 0; i < struct_type->length; i++) {
166 if (strcmp(struct_type->fields.structure[i].name, ir->field) == 0)
168 offset += type_size(struct_type->fields.structure[i].type);
170 this->result.reg_offset += offset;
171 this->result.type = brw_type_for_base_type(ir->type);
175 fs_visitor::visit(ir_dereference_array *ir)
177 ir_constant *constant_index;
179 int element_size = type_size(ir->type);
181 constant_index = ir->array_index->as_constant();
183 ir->array->accept(this);
185 src.type = brw_type_for_base_type(ir->type);
187 if (constant_index) {
188 assert(src.file == UNIFORM || src.file == GRF);
189 src.reg_offset += constant_index->value.i[0] * element_size;
191 /* Variable index array dereference. We attach the variable index
192 * component to the reg as a pointer to a register containing the
193 * offset. Currently only uniform arrays are supported in this patch,
194 * and that reladdr pointer is resolved by
195 * move_uniform_array_access_to_pull_constants(). All other array types
196 * are lowered by lower_variable_index_to_cond_assign().
198 ir->array_index->accept(this);
201 index_reg = fs_reg(this, glsl_type::int_type);
202 emit(BRW_OPCODE_MUL, index_reg, this->result, fs_reg(element_size));
205 emit(BRW_OPCODE_ADD, index_reg, *src.reladdr, index_reg);
208 src.reladdr = ralloc(mem_ctx, fs_reg);
209 memcpy(src.reladdr, &index_reg, sizeof(index_reg));
215 fs_visitor::emit_lrp(fs_reg dst, fs_reg x, fs_reg y, fs_reg a)
218 !x.is_valid_3src() ||
219 !y.is_valid_3src() ||
220 !a.is_valid_3src()) {
221 /* We can't use the LRP instruction. Emit x*(1-a) + y*a. */
222 fs_reg y_times_a = fs_reg(this, glsl_type::float_type);
223 fs_reg one_minus_a = fs_reg(this, glsl_type::float_type);
224 fs_reg x_times_one_minus_a = fs_reg(this, glsl_type::float_type);
226 emit(MUL(y_times_a, y, a));
228 a.negate = !a.negate;
229 emit(ADD(one_minus_a, a, fs_reg(1.0f)));
230 emit(MUL(x_times_one_minus_a, x, one_minus_a));
232 emit(ADD(dst, x_times_one_minus_a, y_times_a));
234 /* The LRP instruction actually does op1 * op0 + op2 * (1 - op0), so
235 * we need to reorder the operands.
237 emit(LRP(dst, a, y, x));
242 fs_visitor::emit_minmax(uint32_t conditionalmod, fs_reg dst,
243 fs_reg src0, fs_reg src1)
248 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
249 inst->conditional_mod = conditionalmod;
251 emit(CMP(reg_null_d, src0, src1, conditionalmod));
253 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
254 inst->predicate = BRW_PREDICATE_NORMAL;
258 /* Instruction selection: Produce a MOV.sat instead of
259 * MIN(MAX(val, 0), 1) when possible.
262 fs_visitor::try_emit_saturate(ir_expression *ir)
264 ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
269 fs_inst *pre_inst = (fs_inst *) this->instructions.get_tail();
271 sat_val->accept(this);
272 fs_reg src = this->result;
274 fs_inst *last_inst = (fs_inst *) this->instructions.get_tail();
276 /* If the last instruction from our accept() didn't generate our
277 * src, generate a saturated MOV
279 fs_inst *modify = get_instruction_generating_reg(pre_inst, last_inst, src);
280 if (!modify || modify->regs_written != 1) {
281 this->result = fs_reg(this, ir->type);
282 fs_inst *inst = emit(MOV(this->result, src));
283 inst->saturate = true;
285 modify->saturate = true;
294 fs_visitor::try_emit_mad(ir_expression *ir, int mul_arg)
296 /* 3-src instructions were introduced in gen6. */
300 /* MAD can only handle floating-point data. */
301 if (ir->type != glsl_type::float_type)
304 ir_rvalue *nonmul = ir->operands[1 - mul_arg];
305 ir_expression *mul = ir->operands[mul_arg]->as_expression();
307 if (!mul || mul->operation != ir_binop_mul)
310 if (nonmul->as_constant() ||
311 mul->operands[0]->as_constant() ||
312 mul->operands[1]->as_constant())
315 nonmul->accept(this);
316 fs_reg src0 = this->result;
318 mul->operands[0]->accept(this);
319 fs_reg src1 = this->result;
321 mul->operands[1]->accept(this);
322 fs_reg src2 = this->result;
324 this->result = fs_reg(this, ir->type);
325 emit(BRW_OPCODE_MAD, this->result, src0, src1, src2);
331 fs_visitor::visit(ir_expression *ir)
333 unsigned int operand;
337 assert(ir->get_num_operands() <= 3);
339 if (try_emit_saturate(ir))
341 if (ir->operation == ir_binop_add) {
342 if (try_emit_mad(ir, 0) || try_emit_mad(ir, 1))
346 for (operand = 0; operand < ir->get_num_operands(); operand++) {
347 ir->operands[operand]->accept(this);
348 if (this->result.file == BAD_FILE) {
349 fail("Failed to get tree for expression operand:\n");
350 ir->operands[operand]->print();
353 assert(this->result.is_valid_3src());
354 op[operand] = this->result;
356 /* Matrix expression operands should have been broken down to vector
357 * operations already.
359 assert(!ir->operands[operand]->type->is_matrix());
360 /* And then those vector operands should have been broken down to scalar.
362 assert(!ir->operands[operand]->type->is_vector());
365 /* Storage for our result. If our result goes into an assignment, it will
366 * just get copy-propagated out, so no worries.
368 this->result = fs_reg(this, ir->type);
370 switch (ir->operation) {
371 case ir_unop_logic_not:
372 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
373 * ones complement of the whole register, not just bit 0.
375 emit(XOR(this->result, op[0], fs_reg(1)));
378 op[0].negate = !op[0].negate;
379 emit(MOV(this->result, op[0]));
383 op[0].negate = false;
384 emit(MOV(this->result, op[0]));
387 if (ir->type->is_float()) {
388 /* AND(val, 0x80000000) gives the sign bit.
390 * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
393 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
395 op[0].type = BRW_REGISTER_TYPE_UD;
396 this->result.type = BRW_REGISTER_TYPE_UD;
397 emit(AND(this->result, op[0], fs_reg(0x80000000u)));
399 inst = emit(OR(this->result, this->result, fs_reg(0x3f800000u)));
400 inst->predicate = BRW_PREDICATE_NORMAL;
402 this->result.type = BRW_REGISTER_TYPE_F;
404 /* ASR(val, 31) -> negative val generates 0xffffffff (signed -1).
405 * -> non-negative val generates 0x00000000.
406 * Predicated OR sets 1 if val is positive.
408 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_G));
410 emit(ASR(this->result, op[0], fs_reg(31)));
412 inst = emit(OR(this->result, this->result, fs_reg(1)));
413 inst->predicate = BRW_PREDICATE_NORMAL;
417 emit_math(SHADER_OPCODE_RCP, this->result, op[0]);
421 emit_math(SHADER_OPCODE_EXP2, this->result, op[0]);
424 emit_math(SHADER_OPCODE_LOG2, this->result, op[0]);
428 assert(!"not reached: should be handled by ir_explog_to_explog2");
431 case ir_unop_sin_reduced:
432 emit_math(SHADER_OPCODE_SIN, this->result, op[0]);
435 case ir_unop_cos_reduced:
436 emit_math(SHADER_OPCODE_COS, this->result, op[0]);
440 emit(FS_OPCODE_DDX, this->result, op[0]);
443 emit(FS_OPCODE_DDY, this->result, op[0]);
447 emit(ADD(this->result, op[0], op[1]));
450 assert(!"not reached: should be handled by ir_sub_to_add_neg");
454 if (brw->gen < 8 && ir->type->is_integer()) {
455 /* For integer multiplication, the MUL uses the low 16 bits
456 * of one of the operands (src0 on gen6, src1 on gen7). The
457 * MACH accumulates in the contribution of the upper 16 bits
460 * FINISHME: Emit just the MUL if we know an operand is small
463 if (brw->gen >= 7 && dispatch_width == 16)
464 fail("SIMD16 explicit accumulator operands unsupported\n");
466 struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
468 emit(MUL(acc, op[0], op[1]));
469 emit(MACH(reg_null_d, op[0], op[1]));
470 emit(MOV(this->result, fs_reg(acc)));
472 emit(MUL(this->result, op[0], op[1]));
475 case ir_binop_imul_high: {
476 if (brw->gen >= 7 && dispatch_width == 16)
477 fail("SIMD16 explicit accumulator operands unsupported\n");
479 struct brw_reg acc = retype(brw_acc_reg(), this->result.type);
481 emit(MUL(acc, op[0], op[1]));
482 emit(MACH(this->result, op[0], op[1]));
486 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
487 assert(ir->type->is_integer());
488 emit_math(SHADER_OPCODE_INT_QUOTIENT, this->result, op[0], op[1]);
490 case ir_binop_carry: {
491 if (brw->gen >= 7 && dispatch_width == 16)
492 fail("SIMD16 explicit accumulator operands unsupported\n");
494 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
496 emit(ADDC(reg_null_ud, op[0], op[1]));
497 emit(MOV(this->result, fs_reg(acc)));
500 case ir_binop_borrow: {
501 if (brw->gen >= 7 && dispatch_width == 16)
502 fail("SIMD16 explicit accumulator operands unsupported\n");
504 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_UD);
506 emit(SUBB(reg_null_ud, op[0], op[1]));
507 emit(MOV(this->result, fs_reg(acc)));
511 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
512 assert(ir->type->is_integer());
513 emit_math(SHADER_OPCODE_INT_REMAINDER, this->result, op[0], op[1]);
517 case ir_binop_greater:
518 case ir_binop_lequal:
519 case ir_binop_gequal:
521 case ir_binop_all_equal:
522 case ir_binop_nequal:
523 case ir_binop_any_nequal:
524 resolve_bool_comparison(ir->operands[0], &op[0]);
525 resolve_bool_comparison(ir->operands[1], &op[1]);
527 emit(CMP(this->result, op[0], op[1],
528 brw_conditional_for_comparison(ir->operation)));
531 case ir_binop_logic_xor:
532 emit(XOR(this->result, op[0], op[1]));
535 case ir_binop_logic_or:
536 emit(OR(this->result, op[0], op[1]));
539 case ir_binop_logic_and:
540 emit(AND(this->result, op[0], op[1]));
545 assert(!"not reached: should be handled by brw_fs_channel_expressions");
549 assert(!"not reached: should be handled by lower_noise");
552 case ir_quadop_vector:
553 assert(!"not reached: should be handled by lower_quadop_vector");
556 case ir_binop_vector_extract:
557 assert(!"not reached: should be handled by lower_vec_index_to_cond_assign()");
560 case ir_triop_vector_insert:
561 assert(!"not reached: should be handled by lower_vector_insert()");
565 assert(!"not reached: should be handled by ldexp_to_arith()");
569 emit_math(SHADER_OPCODE_SQRT, this->result, op[0]);
573 emit_math(SHADER_OPCODE_RSQ, this->result, op[0]);
576 case ir_unop_bitcast_i2f:
577 case ir_unop_bitcast_u2f:
578 op[0].type = BRW_REGISTER_TYPE_F;
579 this->result = op[0];
582 case ir_unop_bitcast_f2u:
583 op[0].type = BRW_REGISTER_TYPE_UD;
584 this->result = op[0];
587 case ir_unop_bitcast_f2i:
588 op[0].type = BRW_REGISTER_TYPE_D;
589 this->result = op[0];
595 emit(MOV(this->result, op[0]));
599 emit(AND(this->result, op[0], fs_reg(1)));
602 temp = fs_reg(this, glsl_type::int_type);
603 emit(AND(temp, op[0], fs_reg(1)));
604 emit(MOV(this->result, temp));
608 emit(CMP(this->result, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
611 emit(CMP(this->result, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
615 emit(RNDZ(this->result, op[0]));
618 op[0].negate = !op[0].negate;
619 emit(RNDD(this->result, op[0]));
620 this->result.negate = true;
623 emit(RNDD(this->result, op[0]));
626 emit(FRC(this->result, op[0]));
628 case ir_unop_round_even:
629 emit(RNDE(this->result, op[0]));
634 resolve_ud_negate(&op[0]);
635 resolve_ud_negate(&op[1]);
636 emit_minmax(ir->operation == ir_binop_min ?
637 BRW_CONDITIONAL_L : BRW_CONDITIONAL_GE,
638 this->result, op[0], op[1]);
640 case ir_unop_pack_snorm_2x16:
641 case ir_unop_pack_snorm_4x8:
642 case ir_unop_pack_unorm_2x16:
643 case ir_unop_pack_unorm_4x8:
644 case ir_unop_unpack_snorm_2x16:
645 case ir_unop_unpack_snorm_4x8:
646 case ir_unop_unpack_unorm_2x16:
647 case ir_unop_unpack_unorm_4x8:
648 case ir_unop_unpack_half_2x16:
649 case ir_unop_pack_half_2x16:
650 assert(!"not reached: should be handled by lower_packing_builtins");
652 case ir_unop_unpack_half_2x16_split_x:
653 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X, this->result, op[0]);
655 case ir_unop_unpack_half_2x16_split_y:
656 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y, this->result, op[0]);
659 emit_math(SHADER_OPCODE_POW, this->result, op[0], op[1]);
662 case ir_unop_bitfield_reverse:
663 emit(BFREV(this->result, op[0]));
665 case ir_unop_bit_count:
666 emit(CBIT(this->result, op[0]));
668 case ir_unop_find_msb:
669 temp = fs_reg(this, glsl_type::uint_type);
670 emit(FBH(temp, op[0]));
672 /* FBH counts from the MSB side, while GLSL's findMSB() wants the count
673 * from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
674 * subtract the result from 31 to convert the MSB count into an LSB count.
677 /* FBH only supports UD type for dst, so use a MOV to convert UD to D. */
678 emit(MOV(this->result, temp));
679 emit(CMP(reg_null_d, this->result, fs_reg(-1), BRW_CONDITIONAL_NZ));
682 inst = emit(ADD(this->result, temp, fs_reg(31)));
683 inst->predicate = BRW_PREDICATE_NORMAL;
685 case ir_unop_find_lsb:
686 emit(FBL(this->result, op[0]));
688 case ir_triop_bitfield_extract:
689 /* Note that the instruction's argument order is reversed from GLSL
692 emit(BFE(this->result, op[2], op[1], op[0]));
695 emit(BFI1(this->result, op[0], op[1]));
698 emit(BFI2(this->result, op[0], op[1], op[2]));
700 case ir_quadop_bitfield_insert:
701 assert(!"not reached: should be handled by "
702 "lower_instructions::bitfield_insert_to_bfm_bfi");
705 case ir_unop_bit_not:
706 emit(NOT(this->result, op[0]));
708 case ir_binop_bit_and:
709 emit(AND(this->result, op[0], op[1]));
711 case ir_binop_bit_xor:
712 emit(XOR(this->result, op[0], op[1]));
714 case ir_binop_bit_or:
715 emit(OR(this->result, op[0], op[1]));
718 case ir_binop_lshift:
719 emit(SHL(this->result, op[0], op[1]));
722 case ir_binop_rshift:
723 if (ir->type->base_type == GLSL_TYPE_INT)
724 emit(ASR(this->result, op[0], op[1]));
726 emit(SHR(this->result, op[0], op[1]));
728 case ir_binop_pack_half_2x16_split:
729 emit(FS_OPCODE_PACK_HALF_2x16_SPLIT, this->result, op[0], op[1]);
731 case ir_binop_ubo_load: {
732 /* This IR node takes a constant uniform block and a constant or
733 * variable byte offset within the block and loads a vector from that.
735 ir_constant *uniform_block = ir->operands[0]->as_constant();
736 ir_constant *const_offset = ir->operands[1]->as_constant();
737 fs_reg surf_index = fs_reg(c->prog_data.base.binding_table.ubo_start +
738 uniform_block->value.u[0]);
740 fs_reg packed_consts = fs_reg(this, glsl_type::float_type);
741 packed_consts.type = result.type;
743 fs_reg const_offset_reg = fs_reg(const_offset->value.u[0] & ~15);
744 emit(fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
745 packed_consts, surf_index, const_offset_reg));
747 packed_consts.smear = const_offset->value.u[0] % 16 / 4;
748 for (int i = 0; i < ir->type->vector_elements; i++) {
749 /* UBO bools are any nonzero value. We consider bools to be
750 * values with the low bit set to 1. Convert them using CMP.
752 if (ir->type->base_type == GLSL_TYPE_BOOL) {
753 emit(CMP(result, packed_consts, fs_reg(0u), BRW_CONDITIONAL_NZ));
755 emit(MOV(result, packed_consts));
758 packed_consts.smear++;
761 /* The std140 packing rules don't allow vectors to cross 16-byte
762 * boundaries, and a reg is 32 bytes.
764 assert(packed_consts.smear < 8);
767 /* Turn the byte offset into a dword offset. */
768 fs_reg base_offset = fs_reg(this, glsl_type::int_type);
769 emit(SHR(base_offset, op[1], fs_reg(2)));
771 for (int i = 0; i < ir->type->vector_elements; i++) {
772 emit(VARYING_PULL_CONSTANT_LOAD(result, surf_index,
775 if (ir->type->base_type == GLSL_TYPE_BOOL)
776 emit(CMP(result, result, fs_reg(0), BRW_CONDITIONAL_NZ));
782 result.reg_offset = 0;
787 /* Note that the instruction's argument order is reversed from GLSL
790 emit(MAD(this->result, op[2], op[1], op[0]));
794 emit_lrp(this->result, op[0], op[1], op[2]);
798 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
799 inst = emit(BRW_OPCODE_SEL, this->result, op[1], op[2]);
800 inst->predicate = BRW_PREDICATE_NORMAL;
806 fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
807 const glsl_type *type, bool predicated)
809 switch (type->base_type) {
810 case GLSL_TYPE_FLOAT:
814 for (unsigned int i = 0; i < type->components(); i++) {
815 l.type = brw_type_for_base_type(type);
816 r.type = brw_type_for_base_type(type);
818 if (predicated || !l.equals(r)) {
819 fs_inst *inst = emit(MOV(l, r));
820 inst->predicate = predicated ? BRW_PREDICATE_NORMAL : BRW_PREDICATE_NONE;
827 case GLSL_TYPE_ARRAY:
828 for (unsigned int i = 0; i < type->length; i++) {
829 emit_assignment_writes(l, r, type->fields.array, predicated);
833 case GLSL_TYPE_STRUCT:
834 for (unsigned int i = 0; i < type->length; i++) {
835 emit_assignment_writes(l, r, type->fields.structure[i].type,
840 case GLSL_TYPE_SAMPLER:
841 case GLSL_TYPE_ATOMIC_UINT:
845 case GLSL_TYPE_ERROR:
846 case GLSL_TYPE_INTERFACE:
847 assert(!"not reached");
852 /* If the RHS processing resulted in an instruction generating a
853 * temporary value, and it would be easy to rewrite the instruction to
854 * generate its result right into the LHS instead, do so. This ends
855 * up reliably removing instructions where it can be tricky to do so
856 * later without real UD chain information.
859 fs_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir,
862 fs_inst *pre_rhs_inst,
863 fs_inst *last_rhs_inst)
865 /* Only attempt if we're doing a direct assignment. */
867 !(ir->lhs->type->is_scalar() ||
868 (ir->lhs->type->is_vector() &&
869 ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1)))
872 /* Make sure the last instruction generated our source reg. */
873 fs_inst *modify = get_instruction_generating_reg(pre_rhs_inst,
879 /* If last_rhs_inst wrote a different number of components than our LHS,
880 * we can't safely rewrite it.
882 if (virtual_grf_sizes[dst.reg] != modify->regs_written)
885 /* Success! Rewrite the instruction. */
892 fs_visitor::visit(ir_assignment *ir)
897 /* FINISHME: arrays on the lhs */
898 ir->lhs->accept(this);
901 fs_inst *pre_rhs_inst = (fs_inst *) this->instructions.get_tail();
903 ir->rhs->accept(this);
906 fs_inst *last_rhs_inst = (fs_inst *) this->instructions.get_tail();
908 assert(l.file != BAD_FILE);
909 assert(r.file != BAD_FILE);
911 if (try_rewrite_rhs_to_dst(ir, l, r, pre_rhs_inst, last_rhs_inst))
915 emit_bool_to_cond_code(ir->condition);
918 if (ir->lhs->type->is_scalar() ||
919 ir->lhs->type->is_vector()) {
920 for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
921 if (ir->write_mask & (1 << i)) {
922 inst = emit(MOV(l, r));
924 inst->predicate = BRW_PREDICATE_NORMAL;
930 emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
935 fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
936 fs_reg shadow_c, fs_reg lod, fs_reg dPdy)
946 if (ir->shadow_comparitor) {
947 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
948 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
949 coordinate.reg_offset++;
952 /* gen4's SIMD8 sampler always has the slots for u,v,r present.
953 * the unused slots must be zeroed.
955 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
956 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), fs_reg(0.0f)));
960 if (ir->op == ir_tex) {
961 /* There's no plain shadow compare message, so we use shadow
962 * compare with a bias of 0.0.
964 emit(MOV(fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f)));
966 } else if (ir->op == ir_txb || ir->op == ir_txl) {
967 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
970 assert(!"Should not get here.");
973 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
975 } else if (ir->op == ir_tex) {
976 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
977 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
978 coordinate.reg_offset++;
980 /* zero the others. */
981 for (int i = ir->coordinate->type->vector_elements; i<3; i++) {
982 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), fs_reg(0.0f)));
984 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
986 } else if (ir->op == ir_txd) {
989 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
990 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
991 coordinate.reg_offset++;
993 /* the slots for u and v are always present, but r is optional */
994 mlen += MAX2(ir->coordinate->type->vector_elements, 2);
997 * dPdx = dudx, dvdx, drdx
998 * dPdy = dudy, dvdy, drdy
1000 * 1-arg: Does not exist.
1002 * 2-arg: dudx dvdx dudy dvdy
1003 * dPdx.x dPdx.y dPdy.x dPdy.y
1006 * 3-arg: dudx dvdx drdx dudy dvdy drdy
1007 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
1008 * m5 m6 m7 m8 m9 m10
1010 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
1011 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdx));
1014 mlen += MAX2(ir->lod_info.grad.dPdx->type->vector_elements, 2);
1016 for (int i = 0; i < ir->lod_info.grad.dPdy->type->vector_elements; i++) {
1017 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdy));
1020 mlen += MAX2(ir->lod_info.grad.dPdy->type->vector_elements, 2);
1021 } else if (ir->op == ir_txs) {
1022 /* There's no SIMD8 resinfo message on Gen4. Use SIMD16 instead. */
1024 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1027 /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
1028 * instructions. We'll need to do SIMD16 here.
1031 assert(ir->op == ir_txb || ir->op == ir_txl || ir->op == ir_txf);
1033 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1034 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2, coordinate.type),
1036 coordinate.reg_offset++;
1039 /* Initialize the rest of u/v/r with 0.0. Empirically, this seems to
1040 * be necessary for TXF (ld), but seems wise to do for all messages.
1042 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
1043 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2), fs_reg(0.0f)));
1046 /* lod/bias appears after u/v/r. */
1049 emit(MOV(fs_reg(MRF, base_mrf + mlen, lod.type), lod));
1052 /* The unused upper half. */
1057 /* Now, since we're doing simd16, the return is 2 interleaved
1058 * vec4s where the odd-indexed ones are junk. We'll need to move
1059 * this weirdness around to the expected layout.
1062 dst = fs_reg(GRF, virtual_grf_alloc(8),
1064 brw_type_for_base_type(ir->type) :
1065 BRW_REGISTER_TYPE_F));
1068 fs_inst *inst = NULL;
1071 inst = emit(SHADER_OPCODE_TEX, dst);
1074 inst = emit(FS_OPCODE_TXB, dst);
1077 inst = emit(SHADER_OPCODE_TXL, dst);
1080 inst = emit(SHADER_OPCODE_TXD, dst);
1083 inst = emit(SHADER_OPCODE_TXS, dst);
1086 inst = emit(SHADER_OPCODE_TXF, dst);
1089 fail("unrecognized texture opcode");
1091 inst->base_mrf = base_mrf;
1093 inst->header_present = true;
1094 inst->regs_written = simd16 ? 8 : 4;
1097 for (int i = 0; i < 4; i++) {
1098 emit(MOV(orig_dst, dst));
1099 orig_dst.reg_offset++;
1100 dst.reg_offset += 2;
1107 /* gen5's sampler has slots for u, v, r, array index, then optional
1108 * parameters like shadow comparitor or LOD bias. If optional
1109 * parameters aren't present, those base slots are optional and don't
1110 * need to be included in the message.
1112 * We don't fill in the unnecessary slots regardless, which may look
1113 * surprising in the disassembly.
1116 fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1117 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
1118 fs_reg sample_index)
1122 int reg_width = dispatch_width / 8;
1123 bool header_present = false;
1124 const int vector_elements =
1125 ir->coordinate ? ir->coordinate->type->vector_elements : 0;
1128 /* The offsets set up by the ir_texture visitor are in the
1129 * m1 header, so we can't go headerless.
1131 header_present = true;
1136 for (int i = 0; i < vector_elements; i++) {
1137 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
1139 coordinate.reg_offset++;
1141 mlen += vector_elements * reg_width;
1143 if (ir->shadow_comparitor) {
1144 mlen = MAX2(mlen, header_present + 4 * reg_width);
1146 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
1150 fs_inst *inst = NULL;
1153 inst = emit(SHADER_OPCODE_TEX, dst);
1156 mlen = MAX2(mlen, header_present + 4 * reg_width);
1157 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1160 inst = emit(FS_OPCODE_TXB, dst);
1163 mlen = MAX2(mlen, header_present + 4 * reg_width);
1164 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1167 inst = emit(SHADER_OPCODE_TXL, dst);
1170 mlen = MAX2(mlen, header_present + 4 * reg_width); /* skip over 'ai' */
1174 * dPdx = dudx, dvdx, drdx
1175 * dPdy = dudy, dvdy, drdy
1177 * Load up these values:
1178 * - dudx dudy dvdx dvdy drdx drdy
1179 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
1181 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
1182 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1186 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
1191 inst = emit(SHADER_OPCODE_TXD, dst);
1195 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1197 inst = emit(SHADER_OPCODE_TXS, dst);
1199 case ir_query_levels:
1200 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1202 inst = emit(SHADER_OPCODE_TXS, dst);
1205 mlen = header_present + 4 * reg_width;
1206 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), lod));
1207 inst = emit(SHADER_OPCODE_TXF, dst);
1210 mlen = header_present + 4 * reg_width;
1213 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD), fs_reg(0)));
1215 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), sample_index));
1217 inst = emit(SHADER_OPCODE_TXF_CMS, dst);
1220 inst = emit(SHADER_OPCODE_LOD, dst);
1223 inst = emit(SHADER_OPCODE_TG4, dst);
1226 fail("unrecognized texture opcode");
1229 inst->base_mrf = base_mrf;
1231 inst->header_present = header_present;
1232 inst->regs_written = 4;
1234 if (mlen > MAX_SAMPLER_MESSAGE_SIZE) {
1235 fail("Message length >" STRINGIFY(MAX_SAMPLER_MESSAGE_SIZE)
1236 " disallowed by hardware\n");
1243 fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1244 fs_reg shadow_c, fs_reg lod, fs_reg lod2,
1245 fs_reg sample_index, fs_reg mcs, int sampler)
1247 int reg_width = dispatch_width / 8;
1248 bool header_present = false;
1250 fs_reg payload = fs_reg(this, glsl_type::float_type);
1251 fs_reg next = payload;
1253 if (ir->op == ir_tg4 || (ir->offset && ir->op != ir_txf) || sampler >= 16) {
1254 /* For general texture offsets (no txf workaround), we need a header to
1255 * put them in. Note that for SIMD16 we're making space for two actual
1256 * hardware registers here, so the emit will have to fix up for this.
1258 * * ir4_tg4 needs to place its channel select in the header,
1259 * for interaction with ARB_texture_swizzle
1261 * The sampler index is only 4-bits, so for larger sampler numbers we
1262 * need to offset the Sampler State Pointer in the header.
1264 header_present = true;
1268 if (ir->shadow_comparitor) {
1269 emit(MOV(next, shadow_c));
1273 bool has_nonconstant_offset = ir->offset && !ir->offset->as_constant();
1274 bool coordinate_done = false;
1276 /* Set up the LOD info */
1282 emit(MOV(next, lod));
1286 emit(MOV(next, lod));
1290 if (dispatch_width == 16)
1291 fail("Gen7 does not support sample_d/sample_d_c in SIMD16 mode.");
1293 /* Load dPdx and the coordinate together:
1294 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
1296 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1297 emit(MOV(next, coordinate));
1298 coordinate.reg_offset++;
1301 /* For cube map array, the coordinate is (u,v,r,ai) but there are
1302 * only derivatives for (u, v, r).
1304 if (i < ir->lod_info.grad.dPdx->type->vector_elements) {
1305 emit(MOV(next, lod));
1309 emit(MOV(next, lod2));
1315 coordinate_done = true;
1319 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), lod));
1322 case ir_query_levels:
1323 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), fs_reg(0u)));
1327 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r. */
1328 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1329 coordinate.reg_offset++;
1332 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), lod));
1335 for (int i = 1; i < ir->coordinate->type->vector_elements; i++) {
1336 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1337 coordinate.reg_offset++;
1341 coordinate_done = true;
1344 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), sample_index));
1347 /* data from the multisample control surface */
1348 emit(MOV(next.retype(BRW_REGISTER_TYPE_UD), mcs));
1351 /* there is no offsetting for this message; just copy in the integer
1352 * texture coordinates
1354 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1355 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1356 coordinate.reg_offset++;
1360 coordinate_done = true;
1363 if (has_nonconstant_offset) {
1364 if (ir->shadow_comparitor && dispatch_width == 16)
1365 fail("Gen7 does not support gather4_po_c in SIMD16 mode.");
1367 /* More crazy intermixing */
1368 ir->offset->accept(this);
1369 fs_reg offset_value = this->result;
1371 for (int i = 0; i < 2; i++) { /* u, v */
1372 emit(MOV(next, coordinate));
1373 coordinate.reg_offset++;
1377 for (int i = 0; i < 2; i++) { /* offu, offv */
1378 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), offset_value));
1379 offset_value.reg_offset++;
1383 if (ir->coordinate->type->vector_elements == 3) { /* r if present */
1384 emit(MOV(next, coordinate));
1385 coordinate.reg_offset++;
1389 coordinate_done = true;
1394 /* Set up the coordinate (except for cases where it was done above) */
1395 if (ir->coordinate && !coordinate_done) {
1396 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1397 emit(MOV(next, coordinate));
1398 coordinate.reg_offset++;
1403 /* Generate the SEND */
1404 fs_inst *inst = NULL;
1406 case ir_tex: inst = emit(SHADER_OPCODE_TEX, dst, payload); break;
1407 case ir_txb: inst = emit(FS_OPCODE_TXB, dst, payload); break;
1408 case ir_txl: inst = emit(SHADER_OPCODE_TXL, dst, payload); break;
1409 case ir_txd: inst = emit(SHADER_OPCODE_TXD, dst, payload); break;
1410 case ir_txf: inst = emit(SHADER_OPCODE_TXF, dst, payload); break;
1411 case ir_txf_ms: inst = emit(SHADER_OPCODE_TXF_CMS, dst, payload); break;
1412 case ir_txs: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
1413 case ir_query_levels: inst = emit(SHADER_OPCODE_TXS, dst, payload); break;
1414 case ir_lod: inst = emit(SHADER_OPCODE_LOD, dst, payload); break;
1416 if (has_nonconstant_offset)
1417 inst = emit(SHADER_OPCODE_TG4_OFFSET, dst, payload);
1419 inst = emit(SHADER_OPCODE_TG4, dst, payload);
1422 inst->base_mrf = -1;
1424 inst->mlen = next.reg_offset * reg_width - header_present;
1426 inst->mlen = next.reg_offset * reg_width;
1427 inst->header_present = header_present;
1428 inst->regs_written = 4;
1430 virtual_grf_sizes[payload.reg] = next.reg_offset;
1431 if (inst->mlen > MAX_SAMPLER_MESSAGE_SIZE) {
1432 fail("Message length >" STRINGIFY(MAX_SAMPLER_MESSAGE_SIZE)
1433 " disallowed by hardware\n");
1440 fs_visitor::rescale_texcoord(ir_texture *ir, fs_reg coordinate,
1441 bool is_rect, int sampler, int texunit)
1443 fs_inst *inst = NULL;
1444 bool needs_gl_clamp = true;
1445 fs_reg scale_x, scale_y;
1447 /* The 965 requires the EU to do the normalization of GL rectangle
1448 * texture coordinates. We use the program parameter state
1449 * tracking to get the scaling factor.
1453 (brw->gen >= 6 && (c->key.tex.gl_clamp_mask[0] & (1 << sampler) ||
1454 c->key.tex.gl_clamp_mask[1] & (1 << sampler))))) {
1455 struct gl_program_parameter_list *params = prog->Parameters;
1456 int tokens[STATE_LENGTH] = {
1458 STATE_TEXRECT_SCALE,
1464 if (dispatch_width == 16) {
1465 fail("rectangle scale uniform setup not supported on SIMD16\n");
1469 scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
1470 scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
1472 GLuint index = _mesa_add_state_reference(params,
1473 (gl_state_index *)tokens);
1474 c->prog_data.param[c->prog_data.nr_params++] =
1475 &prog->Parameters->ParameterValues[index][0].f;
1476 c->prog_data.param[c->prog_data.nr_params++] =
1477 &prog->Parameters->ParameterValues[index][1].f;
1480 /* The 965 requires the EU to do the normalization of GL rectangle
1481 * texture coordinates. We use the program parameter state
1482 * tracking to get the scaling factor.
1484 if (brw->gen < 6 && is_rect) {
1485 fs_reg dst = fs_reg(this, ir->coordinate->type);
1486 fs_reg src = coordinate;
1489 emit(MUL(dst, src, scale_x));
1492 emit(MUL(dst, src, scale_y));
1493 } else if (is_rect) {
1494 /* On gen6+, the sampler handles the rectangle coordinates
1495 * natively, without needing rescaling. But that means we have
1496 * to do GL_CLAMP clamping at the [0, width], [0, height] scale,
1497 * not [0, 1] like the default case below.
1499 needs_gl_clamp = false;
1501 for (int i = 0; i < 2; i++) {
1502 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1503 fs_reg chan = coordinate;
1504 chan.reg_offset += i;
1506 inst = emit(BRW_OPCODE_SEL, chan, chan, brw_imm_f(0.0));
1507 inst->conditional_mod = BRW_CONDITIONAL_G;
1509 /* Our parameter comes in as 1.0/width or 1.0/height,
1510 * because that's what people normally want for doing
1511 * texture rectangle handling. We need width or height
1512 * for clamping, but we don't care enough to make a new
1513 * parameter type, so just invert back.
1515 fs_reg limit = fs_reg(this, glsl_type::float_type);
1516 emit(MOV(limit, i == 0 ? scale_x : scale_y));
1517 emit(SHADER_OPCODE_RCP, limit, limit);
1519 inst = emit(BRW_OPCODE_SEL, chan, chan, limit);
1520 inst->conditional_mod = BRW_CONDITIONAL_L;
1525 if (ir->coordinate && needs_gl_clamp) {
1526 for (unsigned int i = 0;
1527 i < MIN2(ir->coordinate->type->vector_elements, 3); i++) {
1528 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1529 fs_reg chan = coordinate;
1530 chan.reg_offset += i;
1532 fs_inst *inst = emit(MOV(chan, chan));
1533 inst->saturate = true;
1540 /* Sample from the MCS surface attached to this multisample texture. */
1542 fs_visitor::emit_mcs_fetch(ir_texture *ir, fs_reg coordinate, int sampler)
1544 int reg_width = dispatch_width / 8;
1545 fs_reg payload = fs_reg(this, glsl_type::float_type);
1546 fs_reg dest = fs_reg(this, glsl_type::uvec4_type);
1547 fs_reg next = payload;
1549 /* parameters are: u, v, r, lod; missing parameters are treated as zero */
1550 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1551 emit(MOV(next.retype(BRW_REGISTER_TYPE_D), coordinate));
1552 coordinate.reg_offset++;
1556 fs_inst *inst = emit(SHADER_OPCODE_TXF_MCS, dest, payload);
1557 virtual_grf_sizes[payload.reg] = next.reg_offset;
1558 inst->base_mrf = -1;
1559 inst->mlen = next.reg_offset * reg_width;
1560 inst->header_present = false;
1561 inst->regs_written = 4 * reg_width; /* we only care about one reg of response,
1562 * but the sampler always writes 4/8
1564 inst->sampler = sampler;
1570 fs_visitor::visit(ir_texture *ir)
1572 fs_inst *inst = NULL;
1575 _mesa_get_sampler_uniform_value(ir->sampler, shader_prog, prog);
1576 /* FINISHME: We're failing to recompile our programs when the sampler is
1577 * updated. This only matters for the texture rectangle scale parameters
1578 * (pre-gen6, or gen6+ with GL_CLAMP).
1580 int texunit = prog->SamplerUnits[sampler];
1582 if (ir->op == ir_tg4) {
1583 /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
1584 * emitting anything other than setting up the constant result.
1586 ir_constant *chan = ir->lod_info.component->as_constant();
1587 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
1588 if (swiz == SWIZZLE_ZERO || swiz == SWIZZLE_ONE) {
1590 fs_reg res = fs_reg(this, glsl_type::vec4_type);
1593 for (int i=0; i<4; i++) {
1594 emit(MOV(res, fs_reg(swiz == SWIZZLE_ZERO ? 0.0f : 1.0f)));
1601 /* Should be lowered by do_lower_texture_projection */
1602 assert(!ir->projector);
1604 /* Should be lowered */
1605 assert(!ir->offset || !ir->offset->type->is_array());
1607 /* Generate code to compute all the subexpression trees. This has to be
1608 * done before loading any values into MRFs for the sampler message since
1609 * generating these values may involve SEND messages that need the MRFs.
1612 if (ir->coordinate) {
1613 ir->coordinate->accept(this);
1615 coordinate = rescale_texcoord(ir, this->result,
1616 ir->sampler->type->sampler_dimensionality ==
1617 GLSL_SAMPLER_DIM_RECT,
1621 fs_reg shadow_comparitor;
1622 if (ir->shadow_comparitor) {
1623 ir->shadow_comparitor->accept(this);
1624 shadow_comparitor = this->result;
1627 fs_reg lod, lod2, sample_index, mcs;
1632 case ir_query_levels:
1635 ir->lod_info.bias->accept(this);
1639 ir->lod_info.grad.dPdx->accept(this);
1642 ir->lod_info.grad.dPdy->accept(this);
1643 lod2 = this->result;
1648 ir->lod_info.lod->accept(this);
1652 ir->lod_info.sample_index->accept(this);
1653 sample_index = this->result;
1655 if (brw->gen >= 7 && c->key.tex.compressed_multisample_layout_mask & (1<<sampler))
1656 mcs = emit_mcs_fetch(ir, coordinate, sampler);
1661 assert(!"Unrecognized texture opcode");
1664 /* Writemasking doesn't eliminate channels on SIMD8 texture
1665 * samples, so don't worry about them.
1667 fs_reg dst = fs_reg(this, glsl_type::get_instance(ir->type->base_type, 4, 1));
1669 if (brw->gen >= 7) {
1670 inst = emit_texture_gen7(ir, dst, coordinate, shadow_comparitor,
1671 lod, lod2, sample_index, mcs, sampler);
1672 } else if (brw->gen >= 5) {
1673 inst = emit_texture_gen5(ir, dst, coordinate, shadow_comparitor,
1674 lod, lod2, sample_index);
1676 inst = emit_texture_gen4(ir, dst, coordinate, shadow_comparitor,
1680 if (ir->offset != NULL && ir->op != ir_txf)
1681 inst->texture_offset = brw_texture_offset(ctx, ir->offset->as_constant());
1683 if (ir->op == ir_tg4)
1684 inst->texture_offset |= gather_channel(ir, sampler) << 16; // M0.2:16-17
1686 inst->sampler = sampler;
1688 if (ir->shadow_comparitor)
1689 inst->shadow_compare = true;
1691 /* fixup #layers for cube map arrays */
1692 if (ir->op == ir_txs) {
1693 glsl_type const *type = ir->sampler->type;
1694 if (type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE &&
1695 type->sampler_array) {
1697 depth.reg_offset = 2;
1698 emit_math(SHADER_OPCODE_INT_QUOTIENT, depth, depth, fs_reg(6));
1702 swizzle_result(ir, dst, sampler);
1706 * Set up the gather channel based on the swizzle, for gather4.
1709 fs_visitor::gather_channel(ir_texture *ir, int sampler)
1711 ir_constant *chan = ir->lod_info.component->as_constant();
1712 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], chan->value.i[0]);
1714 case SWIZZLE_X: return 0;
1716 /* gather4 sampler is broken for green channel on RG32F --
1717 * we must ask for blue instead.
1719 if (c->key.tex.gather_channel_quirk_mask & (1<<sampler))
1722 case SWIZZLE_Z: return 2;
1723 case SWIZZLE_W: return 3;
1725 assert(!"Not reached"); /* zero, one swizzles handled already */
1731 * Swizzle the result of a texture result. This is necessary for
1732 * EXT_texture_swizzle as well as DEPTH_TEXTURE_MODE for shadow comparisons.
1735 fs_visitor::swizzle_result(ir_texture *ir, fs_reg orig_val, int sampler)
1737 if (ir->op == ir_query_levels) {
1738 /* # levels is in .w */
1739 orig_val.reg_offset += 3;
1740 this->result = orig_val;
1744 this->result = orig_val;
1746 /* txs,lod don't actually sample the texture, so swizzling the result
1749 if (ir->op == ir_txs || ir->op == ir_lod || ir->op == ir_tg4)
1752 if (ir->type == glsl_type::float_type) {
1753 /* Ignore DEPTH_TEXTURE_MODE swizzling. */
1754 assert(ir->sampler->type->sampler_shadow);
1755 } else if (c->key.tex.swizzles[sampler] != SWIZZLE_NOOP) {
1756 fs_reg swizzled_result = fs_reg(this, glsl_type::vec4_type);
1758 for (int i = 0; i < 4; i++) {
1759 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], i);
1760 fs_reg l = swizzled_result;
1763 if (swiz == SWIZZLE_ZERO) {
1764 emit(MOV(l, fs_reg(0.0f)));
1765 } else if (swiz == SWIZZLE_ONE) {
1766 emit(MOV(l, fs_reg(1.0f)));
1768 fs_reg r = orig_val;
1769 r.reg_offset += GET_SWZ(c->key.tex.swizzles[sampler], i);
1773 this->result = swizzled_result;
1778 fs_visitor::visit(ir_swizzle *ir)
1780 ir->val->accept(this);
1781 fs_reg val = this->result;
1783 if (ir->type->vector_elements == 1) {
1784 this->result.reg_offset += ir->mask.x;
1788 fs_reg result = fs_reg(this, ir->type);
1789 this->result = result;
1791 for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
1792 fs_reg channel = val;
1810 channel.reg_offset += swiz;
1811 emit(MOV(result, channel));
1812 result.reg_offset++;
1817 fs_visitor::visit(ir_discard *ir)
1819 assert(ir->condition == NULL); /* FINISHME */
1821 /* We track our discarded pixels in f0.1. By predicating on it, we can
1822 * update just the flag bits that aren't yet discarded. By emitting a
1823 * CMP of g0 != g0, all our currently executing channels will get turned
1826 fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0),
1827 BRW_REGISTER_TYPE_UW));
1828 fs_inst *cmp = emit(CMP(reg_null_f, some_reg, some_reg,
1829 BRW_CONDITIONAL_NZ));
1830 cmp->predicate = BRW_PREDICATE_NORMAL;
1831 cmp->flag_subreg = 1;
1833 if (brw->gen >= 6) {
1834 /* For performance, after a discard, jump to the end of the shader.
1835 * However, many people will do foliage by discarding based on a
1836 * texture's alpha mask, and then continue on to texture with the
1837 * remaining pixels. To avoid trashing the derivatives for those
1838 * texture samples, we'll only jump if all of the pixels in the subspan
1839 * have been discarded.
1841 fs_inst *discard_jump = emit(FS_OPCODE_DISCARD_JUMP);
1842 discard_jump->flag_subreg = 1;
1843 discard_jump->predicate = BRW_PREDICATE_ALIGN1_ANY4H;
1844 discard_jump->predicate_inverse = true;
1849 fs_visitor::visit(ir_constant *ir)
1851 /* Set this->result to reg at the bottom of the function because some code
1852 * paths will cause this visitor to be applied to other fields. This will
1853 * cause the value stored in this->result to be modified.
1855 * Make reg constant so that it doesn't get accidentally modified along the
1856 * way. Yes, I actually had this problem. :(
1858 const fs_reg reg(this, ir->type);
1859 fs_reg dst_reg = reg;
1861 if (ir->type->is_array()) {
1862 const unsigned size = type_size(ir->type->fields.array);
1864 for (unsigned i = 0; i < ir->type->length; i++) {
1865 ir->array_elements[i]->accept(this);
1866 fs_reg src_reg = this->result;
1868 dst_reg.type = src_reg.type;
1869 for (unsigned j = 0; j < size; j++) {
1870 emit(MOV(dst_reg, src_reg));
1871 src_reg.reg_offset++;
1872 dst_reg.reg_offset++;
1875 } else if (ir->type->is_record()) {
1876 foreach_list(node, &ir->components) {
1877 ir_constant *const field = (ir_constant *) node;
1878 const unsigned size = type_size(field->type);
1880 field->accept(this);
1881 fs_reg src_reg = this->result;
1883 dst_reg.type = src_reg.type;
1884 for (unsigned j = 0; j < size; j++) {
1885 emit(MOV(dst_reg, src_reg));
1886 src_reg.reg_offset++;
1887 dst_reg.reg_offset++;
1891 const unsigned size = type_size(ir->type);
1893 for (unsigned i = 0; i < size; i++) {
1894 switch (ir->type->base_type) {
1895 case GLSL_TYPE_FLOAT:
1896 emit(MOV(dst_reg, fs_reg(ir->value.f[i])));
1898 case GLSL_TYPE_UINT:
1899 emit(MOV(dst_reg, fs_reg(ir->value.u[i])));
1902 emit(MOV(dst_reg, fs_reg(ir->value.i[i])));
1904 case GLSL_TYPE_BOOL:
1905 emit(MOV(dst_reg, fs_reg((int)ir->value.b[i])));
1908 assert(!"Non-float/uint/int/bool constant");
1910 dst_reg.reg_offset++;
1918 fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
1920 ir_expression *expr = ir->as_expression();
1923 expr->operation != ir_binop_logic_and &&
1924 expr->operation != ir_binop_logic_or &&
1925 expr->operation != ir_binop_logic_xor) {
1929 assert(expr->get_num_operands() <= 2);
1930 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1931 assert(expr->operands[i]->type->is_scalar());
1933 expr->operands[i]->accept(this);
1934 op[i] = this->result;
1936 resolve_ud_negate(&op[i]);
1939 switch (expr->operation) {
1940 case ir_unop_logic_not:
1941 inst = emit(AND(reg_null_d, op[0], fs_reg(1)));
1942 inst->conditional_mod = BRW_CONDITIONAL_Z;
1946 if (brw->gen >= 6) {
1947 emit(CMP(reg_null_d, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
1949 inst = emit(MOV(reg_null_f, op[0]));
1950 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1955 if (brw->gen >= 6) {
1956 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1958 inst = emit(MOV(reg_null_d, op[0]));
1959 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1963 case ir_binop_greater:
1964 case ir_binop_gequal:
1966 case ir_binop_lequal:
1967 case ir_binop_equal:
1968 case ir_binop_all_equal:
1969 case ir_binop_nequal:
1970 case ir_binop_any_nequal:
1971 resolve_bool_comparison(expr->operands[0], &op[0]);
1972 resolve_bool_comparison(expr->operands[1], &op[1]);
1974 emit(CMP(reg_null_d, op[0], op[1],
1975 brw_conditional_for_comparison(expr->operation)));
1979 assert(!"not reached");
1980 fail("bad cond code\n");
1988 fs_inst *inst = emit(AND(reg_null_d, this->result, fs_reg(1)));
1989 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1993 * Emit a gen6 IF statement with the comparison folded into the IF
1997 fs_visitor::emit_if_gen6(ir_if *ir)
1999 ir_expression *expr = ir->condition->as_expression();
2006 assert(expr->get_num_operands() <= 2);
2007 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
2008 assert(expr->operands[i]->type->is_scalar());
2010 expr->operands[i]->accept(this);
2011 op[i] = this->result;
2014 switch (expr->operation) {
2015 case ir_unop_logic_not:
2016 case ir_binop_logic_xor:
2017 case ir_binop_logic_or:
2018 case ir_binop_logic_and:
2019 /* For operations on bool arguments, only the low bit of the bool is
2020 * valid, and the others are undefined. Fall back to the condition
2026 inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
2027 inst->conditional_mod = BRW_CONDITIONAL_NZ;
2031 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
2034 case ir_binop_greater:
2035 case ir_binop_gequal:
2037 case ir_binop_lequal:
2038 case ir_binop_equal:
2039 case ir_binop_all_equal:
2040 case ir_binop_nequal:
2041 case ir_binop_any_nequal:
2042 resolve_bool_comparison(expr->operands[0], &op[0]);
2043 resolve_bool_comparison(expr->operands[1], &op[1]);
2045 emit(IF(op[0], op[1],
2046 brw_conditional_for_comparison(expr->operation)));
2049 assert(!"not reached");
2050 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
2051 fail("bad condition\n");
2056 emit_bool_to_cond_code(ir->condition);
2057 fs_inst *inst = emit(BRW_OPCODE_IF);
2058 inst->predicate = BRW_PREDICATE_NORMAL;
2062 * Try to replace IF/MOV/ELSE/MOV/ENDIF with SEL.
2064 * Many GLSL shaders contain the following pattern:
2066 * x = condition ? foo : bar
2068 * The compiler emits an ir_if tree for this, since each subexpression might be
2069 * a complex tree that could have side-effects or short-circuit logic.
2071 * However, the common case is to simply select one of two constants or
2072 * variable values---which is exactly what SEL is for. In this case, the
2073 * assembly looks like:
2081 * which can be easily translated into:
2083 * (+f0) SEL dst src0 src1
2085 * If src0 is an immediate value, we promote it to a temporary GRF.
2088 fs_visitor::try_replace_with_sel()
2090 fs_inst *endif_inst = (fs_inst *) instructions.get_tail();
2091 assert(endif_inst->opcode == BRW_OPCODE_ENDIF);
2093 /* Pattern match in reverse: IF, MOV, ELSE, MOV, ENDIF. */
2095 BRW_OPCODE_IF, BRW_OPCODE_MOV, BRW_OPCODE_ELSE, BRW_OPCODE_MOV,
2098 fs_inst *match = (fs_inst *) endif_inst->prev;
2099 for (int i = 0; i < 4; i++) {
2100 if (match->is_head_sentinel() || match->opcode != opcodes[4-i-1])
2102 match = (fs_inst *) match->prev;
2105 /* The opcodes match; it looks like the right sequence of instructions. */
2106 fs_inst *else_mov = (fs_inst *) endif_inst->prev;
2107 fs_inst *then_mov = (fs_inst *) else_mov->prev->prev;
2108 fs_inst *if_inst = (fs_inst *) then_mov->prev;
2110 /* Check that the MOVs are the right form. */
2111 if (then_mov->dst.equals(else_mov->dst) &&
2112 !then_mov->is_partial_write() &&
2113 !else_mov->is_partial_write()) {
2115 /* Remove the matched instructions; we'll emit a SEL to replace them. */
2116 while (!if_inst->next->is_tail_sentinel())
2117 if_inst->next->remove();
2120 /* Only the last source register can be a constant, so if the MOV in
2121 * the "then" clause uses a constant, we need to put it in a temporary.
2123 fs_reg src0(then_mov->src[0]);
2124 if (src0.file == IMM) {
2125 src0 = fs_reg(this, glsl_type::float_type);
2126 src0.type = then_mov->src[0].type;
2127 emit(MOV(src0, then_mov->src[0]));
2131 if (if_inst->conditional_mod) {
2132 /* Sandybridge-specific IF with embedded comparison */
2133 emit(CMP(reg_null_d, if_inst->src[0], if_inst->src[1],
2134 if_inst->conditional_mod));
2135 sel = emit(BRW_OPCODE_SEL, then_mov->dst, src0, else_mov->src[0]);
2136 sel->predicate = BRW_PREDICATE_NORMAL;
2138 /* Separate CMP and IF instructions */
2139 sel = emit(BRW_OPCODE_SEL, then_mov->dst, src0, else_mov->src[0]);
2140 sel->predicate = if_inst->predicate;
2141 sel->predicate_inverse = if_inst->predicate_inverse;
2147 fs_visitor::visit(ir_if *ir)
2149 if (brw->gen < 6 && dispatch_width == 16) {
2150 fail("Can't support (non-uniform) control flow on SIMD16\n");
2153 /* Don't point the annotation at the if statement, because then it plus
2154 * the then and else blocks get printed.
2156 this->base_ir = ir->condition;
2158 if (brw->gen == 6) {
2161 emit_bool_to_cond_code(ir->condition);
2163 emit(IF(BRW_PREDICATE_NORMAL));
2166 foreach_list(node, &ir->then_instructions) {
2167 ir_instruction *ir = (ir_instruction *)node;
2173 if (!ir->else_instructions.is_empty()) {
2174 emit(BRW_OPCODE_ELSE);
2176 foreach_list(node, &ir->else_instructions) {
2177 ir_instruction *ir = (ir_instruction *)node;
2184 emit(BRW_OPCODE_ENDIF);
2186 try_replace_with_sel();
2190 fs_visitor::visit(ir_loop *ir)
2192 if (brw->gen < 6 && dispatch_width == 16) {
2193 fail("Can't support (non-uniform) control flow on SIMD16\n");
2196 this->base_ir = NULL;
2197 emit(BRW_OPCODE_DO);
2199 foreach_list(node, &ir->body_instructions) {
2200 ir_instruction *ir = (ir_instruction *)node;
2206 this->base_ir = NULL;
2207 emit(BRW_OPCODE_WHILE);
2211 fs_visitor::visit(ir_loop_jump *ir)
2214 case ir_loop_jump::jump_break:
2215 emit(BRW_OPCODE_BREAK);
2217 case ir_loop_jump::jump_continue:
2218 emit(BRW_OPCODE_CONTINUE);
2224 fs_visitor::visit_atomic_counter_intrinsic(ir_call *ir)
2226 ir_dereference *deref = static_cast<ir_dereference *>(
2227 ir->actual_parameters.get_head());
2228 ir_variable *location = deref->variable_referenced();
2229 unsigned surf_index = (c->prog_data.base.binding_table.abo_start +
2230 location->data.atomic.buffer_index);
2232 /* Calculate the surface offset */
2233 fs_reg offset(this, glsl_type::uint_type);
2234 ir_dereference_array *deref_array = deref->as_dereference_array();
2237 deref_array->array_index->accept(this);
2239 fs_reg tmp(this, glsl_type::uint_type);
2240 emit(MUL(tmp, this->result, ATOMIC_COUNTER_SIZE));
2241 emit(ADD(offset, tmp, location->data.atomic.offset));
2243 offset = location->data.atomic.offset;
2246 /* Emit the appropriate machine instruction */
2247 const char *callee = ir->callee->function_name();
2248 ir->return_deref->accept(this);
2249 fs_reg dst = this->result;
2251 if (!strcmp("__intrinsic_atomic_read", callee)) {
2252 emit_untyped_surface_read(surf_index, dst, offset);
2254 } else if (!strcmp("__intrinsic_atomic_increment", callee)) {
2255 emit_untyped_atomic(BRW_AOP_INC, surf_index, dst, offset,
2256 fs_reg(), fs_reg());
2258 } else if (!strcmp("__intrinsic_atomic_predecrement", callee)) {
2259 emit_untyped_atomic(BRW_AOP_PREDEC, surf_index, dst, offset,
2260 fs_reg(), fs_reg());
2265 fs_visitor::visit(ir_call *ir)
2267 const char *callee = ir->callee->function_name();
2269 if (!strcmp("__intrinsic_atomic_read", callee) ||
2270 !strcmp("__intrinsic_atomic_increment", callee) ||
2271 !strcmp("__intrinsic_atomic_predecrement", callee)) {
2272 visit_atomic_counter_intrinsic(ir);
2274 assert(!"Unsupported intrinsic.");
2279 fs_visitor::visit(ir_return *ir)
2281 assert(!"FINISHME");
2285 fs_visitor::visit(ir_function *ir)
2287 /* Ignore function bodies other than main() -- we shouldn't see calls to
2288 * them since they should all be inlined before we get to ir_to_mesa.
2290 if (strcmp(ir->name, "main") == 0) {
2291 const ir_function_signature *sig;
2294 sig = ir->matching_signature(NULL, &empty);
2298 foreach_list(node, &sig->body) {
2299 ir_instruction *ir = (ir_instruction *)node;
2308 fs_visitor::visit(ir_function_signature *ir)
2310 assert(!"not reached");
2315 fs_visitor::visit(ir_emit_vertex *)
2317 assert(!"not reached");
2321 fs_visitor::visit(ir_end_primitive *)
2323 assert(!"not reached");
2327 fs_visitor::emit_untyped_atomic(unsigned atomic_op, unsigned surf_index,
2328 fs_reg dst, fs_reg offset, fs_reg src0,
2331 const unsigned operand_len = dispatch_width / 8;
2334 /* Initialize the sample mask in the message header. */
2335 emit(MOV(brw_uvec_mrf(8, mlen, 0), brw_imm_ud(0)))
2336 ->force_writemask_all = true;
2339 emit(MOV(brw_uvec_mrf(1, mlen, 7), brw_flag_reg(0, 1)))
2340 ->force_writemask_all = true;
2342 emit(MOV(brw_uvec_mrf(1, mlen, 7),
2343 retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UD)))
2344 ->force_writemask_all = true;
2349 /* Set the atomic operation offset. */
2350 emit(MOV(brw_uvec_mrf(dispatch_width, mlen, 0), offset));
2351 mlen += operand_len;
2353 /* Set the atomic operation arguments. */
2354 if (src0.file != BAD_FILE) {
2355 emit(MOV(brw_uvec_mrf(dispatch_width, mlen, 0), src0));
2356 mlen += operand_len;
2359 if (src1.file != BAD_FILE) {
2360 emit(MOV(brw_uvec_mrf(dispatch_width, mlen, 0), src1));
2361 mlen += operand_len;
2364 /* Emit the instruction. */
2365 fs_inst inst(SHADER_OPCODE_UNTYPED_ATOMIC, dst, atomic_op, surf_index);
2372 fs_visitor::emit_untyped_surface_read(unsigned surf_index, fs_reg dst,
2375 const unsigned operand_len = dispatch_width / 8;
2378 /* Initialize the sample mask in the message header. */
2379 emit(MOV(brw_uvec_mrf(8, mlen, 0), brw_imm_ud(0)))
2380 ->force_writemask_all = true;
2383 emit(MOV(brw_uvec_mrf(1, mlen, 7), brw_flag_reg(0, 1)))
2384 ->force_writemask_all = true;
2386 emit(MOV(brw_uvec_mrf(1, mlen, 7),
2387 retype(brw_vec1_grf(1, 7), BRW_REGISTER_TYPE_UD)))
2388 ->force_writemask_all = true;
2393 /* Set the surface read offset. */
2394 emit(MOV(brw_uvec_mrf(dispatch_width, mlen, 0), offset));
2395 mlen += operand_len;
2397 /* Emit the instruction. */
2398 fs_inst inst(SHADER_OPCODE_UNTYPED_SURFACE_READ, dst, surf_index);
2405 fs_visitor::emit(fs_inst inst)
2407 fs_inst *list_inst = new(mem_ctx) fs_inst;
2414 fs_visitor::emit(fs_inst *inst)
2416 if (force_uncompressed_stack > 0)
2417 inst->force_uncompressed = true;
2419 inst->annotation = this->current_annotation;
2420 inst->ir = this->base_ir;
2422 this->instructions.push_tail(inst);
2428 fs_visitor::emit(exec_list list)
2430 foreach_list_safe(node, &list) {
2431 fs_inst *inst = (fs_inst *)node;
2437 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
2439 fs_visitor::emit_dummy_fs()
2441 int reg_width = dispatch_width / 8;
2443 /* Everyone's favorite color. */
2444 emit(MOV(fs_reg(MRF, 2 + 0 * reg_width), fs_reg(1.0f)));
2445 emit(MOV(fs_reg(MRF, 2 + 1 * reg_width), fs_reg(0.0f)));
2446 emit(MOV(fs_reg(MRF, 2 + 2 * reg_width), fs_reg(1.0f)));
2447 emit(MOV(fs_reg(MRF, 2 + 3 * reg_width), fs_reg(0.0f)));
2450 write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
2451 write->base_mrf = 2;
2452 write->mlen = 4 * reg_width;
2456 /* The register location here is relative to the start of the URB
2457 * data. It will get adjusted to be a real location before
2458 * generate_code() time.
2461 fs_visitor::interp_reg(int location, int channel)
2463 int regnr = c->prog_data.urb_setup[location] * 2 + channel / 2;
2464 int stride = (channel & 1) * 4;
2466 assert(c->prog_data.urb_setup[location] != -1);
2468 return brw_vec1_grf(regnr, stride);
2471 /** Emits the interpolation for the varying inputs. */
2473 fs_visitor::emit_interpolation_setup_gen4()
2475 this->current_annotation = "compute pixel centers";
2476 this->pixel_x = fs_reg(this, glsl_type::uint_type);
2477 this->pixel_y = fs_reg(this, glsl_type::uint_type);
2478 this->pixel_x.type = BRW_REGISTER_TYPE_UW;
2479 this->pixel_y.type = BRW_REGISTER_TYPE_UW;
2481 emit(FS_OPCODE_PIXEL_X, this->pixel_x);
2482 emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
2484 this->current_annotation = "compute pixel deltas from v0";
2486 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2487 fs_reg(this, glsl_type::vec2_type);
2488 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2489 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC];
2490 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg_offset++;
2492 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2493 fs_reg(this, glsl_type::float_type);
2494 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
2495 fs_reg(this, glsl_type::float_type);
2497 emit(ADD(this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2498 this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0)))));
2499 emit(ADD(this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2500 this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1)))));
2502 this->current_annotation = "compute pos.w and 1/pos.w";
2503 /* Compute wpos.w. It's always in our setup, since it's needed to
2504 * interpolate the other attributes.
2506 this->wpos_w = fs_reg(this, glsl_type::float_type);
2507 emit(FS_OPCODE_LINTERP, wpos_w,
2508 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2509 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
2510 interp_reg(VARYING_SLOT_POS, 3));
2511 /* Compute the pixel 1/W value from wpos.w. */
2512 this->pixel_w = fs_reg(this, glsl_type::float_type);
2513 emit_math(SHADER_OPCODE_RCP, this->pixel_w, wpos_w);
2514 this->current_annotation = NULL;
2517 /** Emits the interpolation for the varying inputs. */
2519 fs_visitor::emit_interpolation_setup_gen6()
2521 struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
2523 /* If the pixel centers end up used, the setup is the same as for gen4. */
2524 this->current_annotation = "compute pixel centers";
2525 fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
2526 fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
2527 int_pixel_x.type = BRW_REGISTER_TYPE_UW;
2528 int_pixel_y.type = BRW_REGISTER_TYPE_UW;
2529 emit(ADD(int_pixel_x,
2530 fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
2531 fs_reg(brw_imm_v(0x10101010))));
2532 emit(ADD(int_pixel_y,
2533 fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
2534 fs_reg(brw_imm_v(0x11001100))));
2536 /* As of gen6, we can no longer mix float and int sources. We have
2537 * to turn the integer pixel centers into floats for their actual
2540 this->pixel_x = fs_reg(this, glsl_type::float_type);
2541 this->pixel_y = fs_reg(this, glsl_type::float_type);
2542 emit(MOV(this->pixel_x, int_pixel_x));
2543 emit(MOV(this->pixel_y, int_pixel_y));
2545 this->current_annotation = "compute pos.w";
2546 this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
2547 this->wpos_w = fs_reg(this, glsl_type::float_type);
2548 emit_math(SHADER_OPCODE_RCP, this->wpos_w, this->pixel_w);
2550 for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
2551 uint8_t reg = c->barycentric_coord_reg[i];
2552 this->delta_x[i] = fs_reg(brw_vec8_grf(reg, 0));
2553 this->delta_y[i] = fs_reg(brw_vec8_grf(reg + 1, 0));
2556 this->current_annotation = NULL;
2560 fs_visitor::emit_color_write(int target, int index, int first_color_mrf)
2562 int reg_width = dispatch_width / 8;
2564 fs_reg color = outputs[target];
2567 /* If there's no color data to be written, skip it. */
2568 if (color.file == BAD_FILE)
2571 color.reg_offset += index;
2573 if (dispatch_width == 8 || brw->gen >= 6) {
2574 /* SIMD8 write looks like:
2580 * gen6 SIMD16 DP write looks like:
2590 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index * reg_width,
2593 inst->saturate = c->key.clamp_fragment_color;
2595 /* pre-gen6 SIMD16 single source DP write looks like:
2605 if (brw->has_compr4) {
2606 /* By setting the high bit of the MRF register number, we
2607 * indicate that we want COMPR4 mode - instead of doing the
2608 * usual destination + 1 for the second half we get
2611 inst = emit(MOV(fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index,
2614 inst->saturate = c->key.clamp_fragment_color;
2616 push_force_uncompressed();
2617 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index, color.type),
2619 inst->saturate = c->key.clamp_fragment_color;
2620 pop_force_uncompressed();
2622 color.sechalf = true;
2623 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index + 4, color.type),
2625 inst->force_sechalf = true;
2626 inst->saturate = c->key.clamp_fragment_color;
2627 color.sechalf = false;
2633 cond_for_alpha_func(GLenum func)
2637 return BRW_CONDITIONAL_G;
2639 return BRW_CONDITIONAL_GE;
2641 return BRW_CONDITIONAL_L;
2643 return BRW_CONDITIONAL_LE;
2645 return BRW_CONDITIONAL_EQ;
2647 return BRW_CONDITIONAL_NEQ;
2649 assert(!"Not reached");
2655 * Alpha test support for when we compile it into the shader instead
2656 * of using the normal fixed-function alpha test.
2659 fs_visitor::emit_alpha_test()
2661 this->current_annotation = "Alpha test";
2664 if (c->key.alpha_test_func == GL_ALWAYS)
2667 if (c->key.alpha_test_func == GL_NEVER) {
2669 fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0),
2670 BRW_REGISTER_TYPE_UW));
2671 cmp = emit(CMP(reg_null_f, some_reg, some_reg,
2672 BRW_CONDITIONAL_NEQ));
2675 fs_reg color = outputs[0];
2676 color.reg_offset += 3;
2678 /* f0.1 &= func(color, ref) */
2679 cmp = emit(CMP(reg_null_f, color, fs_reg(c->key.alpha_test_ref),
2680 cond_for_alpha_func(c->key.alpha_test_func)));
2682 cmp->predicate = BRW_PREDICATE_NORMAL;
2683 cmp->flag_subreg = 1;
2687 fs_visitor::emit_fb_writes()
2689 this->current_annotation = "FB write header";
2690 bool header_present = true;
2691 /* We can potentially have a message length of up to 15, so we have to set
2692 * base_mrf to either 0 or 1 in order to fit in m0..m15.
2696 int reg_width = dispatch_width / 8;
2697 bool do_dual_src = this->dual_src_output.file != BAD_FILE;
2698 bool src0_alpha_to_render_target = false;
2700 if (dispatch_width == 16 && do_dual_src) {
2701 fail("GL_ARB_blend_func_extended not yet supported in SIMD16.");
2702 do_dual_src = false;
2705 /* From the Sandy Bridge PRM, volume 4, page 198:
2707 * "Dispatched Pixel Enables. One bit per pixel indicating
2708 * which pixels were originally enabled when the thread was
2709 * dispatched. This field is only required for the end-of-
2710 * thread message and on all dual-source messages."
2712 if (brw->gen >= 6 &&
2713 !this->fp->UsesKill &&
2715 c->key.nr_color_regions == 1) {
2716 header_present = false;
2719 if (header_present) {
2720 src0_alpha_to_render_target = brw->gen >= 6 &&
2722 c->key.replicate_alpha;
2727 if (c->aa_dest_stencil_reg) {
2728 push_force_uncompressed();
2729 emit(MOV(fs_reg(MRF, nr++),
2730 fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0))));
2731 pop_force_uncompressed();
2734 c->prog_data.uses_omask =
2735 fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK);
2736 if(c->prog_data.uses_omask) {
2737 this->current_annotation = "FB write oMask";
2738 assert(this->sample_mask.file != BAD_FILE);
2739 /* Hand over gl_SampleMask. Only lower 16 bits are relevant. */
2740 emit(FS_OPCODE_SET_OMASK, fs_reg(MRF, nr, BRW_REGISTER_TYPE_UW), this->sample_mask);
2744 /* Reserve space for color. It'll be filled in per MRT below. */
2746 nr += 4 * reg_width;
2749 if (src0_alpha_to_render_target)
2752 if (c->source_depth_to_render_target) {
2753 if (brw->gen == 6 && dispatch_width == 16) {
2754 /* For outputting oDepth on gen6, SIMD8 writes have to be
2755 * used. This would require SIMD8 moves of each half to
2756 * message regs, kind of like pre-gen5 SIMD16 FB writes.
2757 * Just bail on doing so for now.
2759 fail("Missing support for simd16 depth writes on gen6\n");
2762 if (prog->OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
2763 /* Hand over gl_FragDepth. */
2764 assert(this->frag_depth.file != BAD_FILE);
2765 emit(MOV(fs_reg(MRF, nr), this->frag_depth));
2767 /* Pass through the payload depth. */
2768 emit(MOV(fs_reg(MRF, nr),
2769 fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
2774 if (c->dest_depth_reg) {
2775 emit(MOV(fs_reg(MRF, nr),
2776 fs_reg(brw_vec8_grf(c->dest_depth_reg, 0))));
2781 fs_reg src0 = this->outputs[0];
2782 fs_reg src1 = this->dual_src_output;
2784 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2786 for (int i = 0; i < 4; i++) {
2787 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + i, src0.type), src0));
2789 inst->saturate = c->key.clamp_fragment_color;
2792 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2794 for (int i = 0; i < 4; i++) {
2795 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + 4 + i, src1.type),
2798 inst->saturate = c->key.clamp_fragment_color;
2801 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2802 emit_shader_time_end();
2804 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2806 inst->base_mrf = base_mrf;
2807 inst->mlen = nr - base_mrf;
2809 inst->header_present = header_present;
2811 c->prog_data.dual_src_blend = true;
2812 this->current_annotation = NULL;
2816 for (int target = 0; target < c->key.nr_color_regions; target++) {
2817 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2818 "FB write target %d",
2820 /* If src0_alpha_to_render_target is true, include source zero alpha
2821 * data in RenderTargetWrite message for targets > 0.
2823 int write_color_mrf = color_mrf;
2824 if (src0_alpha_to_render_target && target != 0) {
2826 fs_reg color = outputs[0];
2827 color.reg_offset += 3;
2829 inst = emit(MOV(fs_reg(MRF, write_color_mrf, color.type),
2831 inst->saturate = c->key.clamp_fragment_color;
2832 write_color_mrf = color_mrf + reg_width;
2835 for (unsigned i = 0; i < this->output_components[target]; i++)
2836 emit_color_write(target, i, write_color_mrf);
2839 if (target == c->key.nr_color_regions - 1) {
2842 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2843 emit_shader_time_end();
2846 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2847 inst->target = target;
2848 inst->base_mrf = base_mrf;
2849 if (src0_alpha_to_render_target && target == 0)
2850 inst->mlen = nr - base_mrf - reg_width;
2852 inst->mlen = nr - base_mrf;
2854 inst->header_present = header_present;
2857 if (c->key.nr_color_regions == 0) {
2858 /* Even if there's no color buffers enabled, we still need to send
2859 * alpha out the pipeline to our null renderbuffer to support
2860 * alpha-testing, alpha-to-coverage, and so on.
2862 emit_color_write(0, 3, color_mrf);
2864 if (INTEL_DEBUG & DEBUG_SHADER_TIME)
2865 emit_shader_time_end();
2867 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2868 inst->base_mrf = base_mrf;
2869 inst->mlen = nr - base_mrf;
2871 inst->header_present = header_present;
2874 this->current_annotation = NULL;
2878 fs_visitor::resolve_ud_negate(fs_reg *reg)
2880 if (reg->type != BRW_REGISTER_TYPE_UD ||
2884 fs_reg temp = fs_reg(this, glsl_type::uint_type);
2885 emit(MOV(temp, *reg));
2890 fs_visitor::resolve_bool_comparison(ir_rvalue *rvalue, fs_reg *reg)
2892 if (rvalue->type != glsl_type::bool_type)
2895 fs_reg temp = fs_reg(this, glsl_type::bool_type);
2896 emit(AND(temp, *reg, fs_reg(1)));
2900 fs_visitor::fs_visitor(struct brw_context *brw,
2901 struct brw_wm_compile *c,
2902 struct gl_shader_program *shader_prog,
2903 struct gl_fragment_program *fp,
2904 unsigned dispatch_width)
2905 : dispatch_width(dispatch_width)
2910 this->prog = &fp->Base;
2911 this->shader_prog = shader_prog;
2912 this->prog = &fp->Base;
2913 this->stage_prog_data = &c->prog_data.base;
2914 this->ctx = &brw->ctx;
2915 this->mem_ctx = ralloc_context(NULL);
2917 shader = (struct brw_shader *)
2918 shader_prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2921 this->failed = false;
2922 this->variable_ht = hash_table_ctor(0,
2923 hash_table_pointer_hash,
2924 hash_table_pointer_compare);
2926 memset(this->outputs, 0, sizeof(this->outputs));
2927 memset(this->output_components, 0, sizeof(this->output_components));
2928 this->first_non_payload_grf = 0;
2929 this->max_grf = brw->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF;
2931 this->current_annotation = NULL;
2932 this->base_ir = NULL;
2934 this->virtual_grf_sizes = NULL;
2935 this->virtual_grf_count = 0;
2936 this->virtual_grf_array_size = 0;
2937 this->virtual_grf_start = NULL;
2938 this->virtual_grf_end = NULL;
2939 this->live_intervals = NULL;
2940 this->regs_live_at_ip = NULL;
2942 this->params_remap = NULL;
2943 this->nr_params_remap = 0;
2945 this->force_uncompressed_stack = 0;
2947 this->spilled_any_registers = false;
2949 memset(&this->param_size, 0, sizeof(this->param_size));
2952 fs_visitor::~fs_visitor()
2954 ralloc_free(this->mem_ctx);
2955 hash_table_dtor(this->variable_ht);