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_minmax(uint32_t conditionalmod, fs_reg dst,
203 fs_reg src0, fs_reg src1)
207 if (intel->gen >= 6) {
208 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
209 inst->conditional_mod = conditionalmod;
211 emit(CMP(reg_null_d, src0, src1, conditionalmod));
213 inst = emit(BRW_OPCODE_SEL, dst, src0, src1);
214 inst->predicate = BRW_PREDICATE_NORMAL;
218 /* Instruction selection: Produce a MOV.sat instead of
219 * MIN(MAX(val, 0), 1) when possible.
222 fs_visitor::try_emit_saturate(ir_expression *ir)
224 ir_rvalue *sat_val = ir->as_rvalue_to_saturate();
229 fs_inst *pre_inst = (fs_inst *) this->instructions.get_tail();
231 sat_val->accept(this);
232 fs_reg src = this->result;
234 fs_inst *last_inst = (fs_inst *) this->instructions.get_tail();
236 /* If the last instruction from our accept() didn't generate our
237 * src, generate a saturated MOV
239 fs_inst *modify = get_instruction_generating_reg(pre_inst, last_inst, src);
240 if (!modify || modify->regs_written() != 1) {
241 this->result = fs_reg(this, ir->type);
242 fs_inst *inst = emit(MOV(this->result, src));
243 inst->saturate = true;
245 modify->saturate = true;
254 fs_visitor::try_emit_mad(ir_expression *ir, int mul_arg)
256 /* 3-src instructions were introduced in gen6. */
260 /* MAD can only handle floating-point data. */
261 if (ir->type != glsl_type::float_type)
264 ir_rvalue *nonmul = ir->operands[1 - mul_arg];
265 ir_expression *mul = ir->operands[mul_arg]->as_expression();
267 if (!mul || mul->operation != ir_binop_mul)
270 if (nonmul->as_constant() ||
271 mul->operands[0]->as_constant() ||
272 mul->operands[1]->as_constant())
275 nonmul->accept(this);
276 fs_reg src0 = this->result;
278 mul->operands[0]->accept(this);
279 fs_reg src1 = this->result;
281 mul->operands[1]->accept(this);
282 fs_reg src2 = this->result;
284 this->result = fs_reg(this, ir->type);
285 emit(BRW_OPCODE_MAD, this->result, src0, src1, src2);
291 fs_visitor::visit(ir_expression *ir)
293 unsigned int operand;
297 assert(ir->get_num_operands() <= 2);
299 if (try_emit_saturate(ir))
301 if (ir->operation == ir_binop_add) {
302 if (try_emit_mad(ir, 0) || try_emit_mad(ir, 1))
306 for (operand = 0; operand < ir->get_num_operands(); operand++) {
307 ir->operands[operand]->accept(this);
308 if (this->result.file == BAD_FILE) {
310 fail("Failed to get tree for expression operand:\n");
311 ir->operands[operand]->accept(&v);
313 op[operand] = this->result;
315 /* Matrix expression operands should have been broken down to vector
316 * operations already.
318 assert(!ir->operands[operand]->type->is_matrix());
319 /* And then those vector operands should have been broken down to scalar.
321 assert(!ir->operands[operand]->type->is_vector());
324 /* Storage for our result. If our result goes into an assignment, it will
325 * just get copy-propagated out, so no worries.
327 this->result = fs_reg(this, ir->type);
329 switch (ir->operation) {
330 case ir_unop_logic_not:
331 /* Note that BRW_OPCODE_NOT is not appropriate here, since it is
332 * ones complement of the whole register, not just bit 0.
334 emit(XOR(this->result, op[0], fs_reg(1)));
337 op[0].negate = !op[0].negate;
338 this->result = op[0];
342 op[0].negate = false;
343 this->result = op[0];
346 temp = fs_reg(this, ir->type);
348 emit(MOV(this->result, fs_reg(0.0f)));
350 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_G));
351 inst = emit(MOV(this->result, fs_reg(1.0f)));
352 inst->predicate = BRW_PREDICATE_NORMAL;
354 emit(CMP(reg_null_f, op[0], fs_reg(0.0f), BRW_CONDITIONAL_L));
355 inst = emit(MOV(this->result, fs_reg(-1.0f)));
356 inst->predicate = BRW_PREDICATE_NORMAL;
360 emit_math(SHADER_OPCODE_RCP, this->result, op[0]);
364 emit_math(SHADER_OPCODE_EXP2, this->result, op[0]);
367 emit_math(SHADER_OPCODE_LOG2, this->result, op[0]);
371 assert(!"not reached: should be handled by ir_explog_to_explog2");
374 case ir_unop_sin_reduced:
375 emit_math(SHADER_OPCODE_SIN, this->result, op[0]);
378 case ir_unop_cos_reduced:
379 emit_math(SHADER_OPCODE_COS, this->result, op[0]);
383 emit(FS_OPCODE_DDX, this->result, op[0]);
386 emit(FS_OPCODE_DDY, this->result, op[0]);
390 emit(ADD(this->result, op[0], op[1]));
393 assert(!"not reached: should be handled by ir_sub_to_add_neg");
397 if (ir->type->is_integer()) {
398 /* For integer multiplication, the MUL uses the low 16 bits
399 * of one of the operands (src0 on gen6, src1 on gen7). The
400 * MACH accumulates in the contribution of the upper 16 bits
403 * FINISHME: Emit just the MUL if we know an operand is small
406 if (intel->gen >= 7 && dispatch_width == 16)
407 fail("16-wide explicit accumulator operands unsupported\n");
409 struct brw_reg acc = retype(brw_acc_reg(), BRW_REGISTER_TYPE_D);
411 emit(MUL(acc, op[0], op[1]));
412 emit(MACH(reg_null_d, op[0], op[1]));
413 emit(MOV(this->result, fs_reg(acc)));
415 emit(MUL(this->result, op[0], op[1]));
419 /* Floating point should be lowered by DIV_TO_MUL_RCP in the compiler. */
420 assert(ir->type->is_integer());
421 emit_math(SHADER_OPCODE_INT_QUOTIENT, this->result, op[0], op[1]);
424 /* Floating point should be lowered by MOD_TO_FRACT in the compiler. */
425 assert(ir->type->is_integer());
426 emit_math(SHADER_OPCODE_INT_REMAINDER, this->result, op[0], op[1]);
430 case ir_binop_greater:
431 case ir_binop_lequal:
432 case ir_binop_gequal:
434 case ir_binop_all_equal:
435 case ir_binop_nequal:
436 case ir_binop_any_nequal:
437 resolve_bool_comparison(ir->operands[0], &op[0]);
438 resolve_bool_comparison(ir->operands[1], &op[1]);
440 emit(CMP(this->result, op[0], op[1],
441 brw_conditional_for_comparison(ir->operation)));
444 case ir_binop_logic_xor:
445 emit(XOR(this->result, op[0], op[1]));
448 case ir_binop_logic_or:
449 emit(OR(this->result, op[0], op[1]));
452 case ir_binop_logic_and:
453 emit(AND(this->result, op[0], op[1]));
458 assert(!"not reached: should be handled by brw_fs_channel_expressions");
462 assert(!"not reached: should be handled by lower_noise");
465 case ir_quadop_vector:
466 assert(!"not reached: should be handled by lower_quadop_vector");
470 emit_math(SHADER_OPCODE_SQRT, this->result, op[0]);
474 emit_math(SHADER_OPCODE_RSQ, this->result, op[0]);
477 case ir_unop_bitcast_i2f:
478 case ir_unop_bitcast_u2f:
479 op[0].type = BRW_REGISTER_TYPE_F;
480 this->result = op[0];
483 case ir_unop_bitcast_f2u:
484 op[0].type = BRW_REGISTER_TYPE_UD;
485 this->result = op[0];
488 case ir_unop_bitcast_f2i:
489 op[0].type = BRW_REGISTER_TYPE_D;
490 this->result = op[0];
496 emit(MOV(this->result, op[0]));
500 inst = emit(AND(this->result, op[0], fs_reg(1)));
503 temp = fs_reg(this, glsl_type::int_type);
504 emit(AND(temp, op[0], fs_reg(1)));
505 emit(MOV(this->result, temp));
509 emit(CMP(this->result, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
512 emit(CMP(this->result, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
516 emit(RNDZ(this->result, op[0]));
519 op[0].negate = !op[0].negate;
520 inst = emit(RNDD(this->result, op[0]));
521 this->result.negate = true;
524 inst = emit(RNDD(this->result, op[0]));
527 inst = emit(FRC(this->result, op[0]));
529 case ir_unop_round_even:
530 emit(RNDE(this->result, op[0]));
535 resolve_ud_negate(&op[0]);
536 resolve_ud_negate(&op[1]);
537 emit_minmax(ir->operation == ir_binop_min ?
538 BRW_CONDITIONAL_L : BRW_CONDITIONAL_GE,
539 this->result, op[0], op[1]);
541 case ir_unop_pack_snorm_2x16:
542 case ir_unop_pack_snorm_4x8:
543 case ir_unop_pack_unorm_2x16:
544 case ir_unop_pack_unorm_4x8:
545 case ir_unop_unpack_snorm_2x16:
546 case ir_unop_unpack_snorm_4x8:
547 case ir_unop_unpack_unorm_2x16:
548 case ir_unop_unpack_unorm_4x8:
549 case ir_unop_unpack_half_2x16:
550 case ir_unop_pack_half_2x16:
551 assert(!"not reached: should be handled by lower_packing_builtins");
553 case ir_unop_unpack_half_2x16_split_x:
554 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X, this->result, op[0]);
556 case ir_unop_unpack_half_2x16_split_y:
557 emit(FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y, this->result, op[0]);
560 emit_math(SHADER_OPCODE_POW, this->result, op[0], op[1]);
563 case ir_unop_bit_not:
564 inst = emit(NOT(this->result, op[0]));
566 case ir_binop_bit_and:
567 inst = emit(AND(this->result, op[0], op[1]));
569 case ir_binop_bit_xor:
570 inst = emit(XOR(this->result, op[0], op[1]));
572 case ir_binop_bit_or:
573 inst = emit(OR(this->result, op[0], op[1]));
576 case ir_binop_lshift:
577 inst = emit(SHL(this->result, op[0], op[1]));
580 case ir_binop_rshift:
581 if (ir->type->base_type == GLSL_TYPE_INT)
582 inst = emit(ASR(this->result, op[0], op[1]));
584 inst = emit(SHR(this->result, op[0], op[1]));
586 case ir_binop_pack_half_2x16_split:
587 emit(FS_OPCODE_PACK_HALF_2x16_SPLIT, this->result, op[0], op[1]);
589 case ir_binop_ubo_load:
590 /* This IR node takes a constant uniform block and a constant or
591 * variable byte offset within the block and loads a vector from that.
593 ir_constant *uniform_block = ir->operands[0]->as_constant();
594 ir_constant *const_offset = ir->operands[1]->as_constant();
595 fs_reg surf_index = fs_reg((unsigned)SURF_INDEX_WM_UBO(uniform_block->value.u[0]));
597 fs_reg packed_consts = fs_reg(this, glsl_type::float_type);
598 packed_consts.type = result.type;
600 fs_reg const_offset_reg = fs_reg(const_offset->value.u[0] & ~15);
601 emit(fs_inst(FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD,
602 packed_consts, surf_index, const_offset_reg));
604 packed_consts.smear = const_offset->value.u[0] % 16 / 4;
605 for (int i = 0; i < ir->type->vector_elements; i++) {
606 /* UBO bools are any nonzero value. We consider bools to be
607 * values with the low bit set to 1. Convert them using CMP.
609 if (ir->type->base_type == GLSL_TYPE_BOOL) {
610 emit(CMP(result, packed_consts, fs_reg(0u), BRW_CONDITIONAL_NZ));
612 emit(MOV(result, packed_consts));
615 packed_consts.smear++;
618 /* The std140 packing rules don't allow vectors to cross 16-byte
619 * boundaries, and a reg is 32 bytes.
621 assert(packed_consts.smear < 8);
624 /* Turn the byte offset into a dword offset. */
625 fs_reg base_offset = fs_reg(this, glsl_type::int_type);
626 emit(SHR(base_offset, op[1], fs_reg(2)));
628 for (int i = 0; i < ir->type->vector_elements; i++) {
629 fs_reg offset = fs_reg(this, glsl_type::int_type);
630 emit(ADD(offset, base_offset, fs_reg(i)));
631 emit(VARYING_PULL_CONSTANT_LOAD(result, surf_index, offset));
633 if (ir->type->base_type == GLSL_TYPE_BOOL)
634 emit(CMP(result, result, fs_reg(0), BRW_CONDITIONAL_NZ));
640 result.reg_offset = 0;
646 fs_visitor::emit_assignment_writes(fs_reg &l, fs_reg &r,
647 const glsl_type *type, bool predicated)
649 switch (type->base_type) {
650 case GLSL_TYPE_FLOAT:
654 for (unsigned int i = 0; i < type->components(); i++) {
655 l.type = brw_type_for_base_type(type);
656 r.type = brw_type_for_base_type(type);
658 if (predicated || !l.equals(r)) {
659 fs_inst *inst = emit(MOV(l, r));
660 inst->predicate = predicated ? BRW_PREDICATE_NORMAL : BRW_PREDICATE_NONE;
667 case GLSL_TYPE_ARRAY:
668 for (unsigned int i = 0; i < type->length; i++) {
669 emit_assignment_writes(l, r, type->fields.array, predicated);
673 case GLSL_TYPE_STRUCT:
674 for (unsigned int i = 0; i < type->length; i++) {
675 emit_assignment_writes(l, r, type->fields.structure[i].type,
680 case GLSL_TYPE_SAMPLER:
684 case GLSL_TYPE_ERROR:
685 case GLSL_TYPE_INTERFACE:
686 assert(!"not reached");
691 /* If the RHS processing resulted in an instruction generating a
692 * temporary value, and it would be easy to rewrite the instruction to
693 * generate its result right into the LHS instead, do so. This ends
694 * up reliably removing instructions where it can be tricky to do so
695 * later without real UD chain information.
698 fs_visitor::try_rewrite_rhs_to_dst(ir_assignment *ir,
701 fs_inst *pre_rhs_inst,
702 fs_inst *last_rhs_inst)
704 /* Only attempt if we're doing a direct assignment. */
706 !(ir->lhs->type->is_scalar() ||
707 (ir->lhs->type->is_vector() &&
708 ir->write_mask == (1 << ir->lhs->type->vector_elements) - 1)))
711 /* Make sure the last instruction generated our source reg. */
712 fs_inst *modify = get_instruction_generating_reg(pre_rhs_inst,
718 /* If last_rhs_inst wrote a different number of components than our LHS,
719 * we can't safely rewrite it.
721 if (virtual_grf_sizes[dst.reg] != modify->regs_written())
724 /* Success! Rewrite the instruction. */
731 fs_visitor::visit(ir_assignment *ir)
736 /* FINISHME: arrays on the lhs */
737 ir->lhs->accept(this);
740 fs_inst *pre_rhs_inst = (fs_inst *) this->instructions.get_tail();
742 ir->rhs->accept(this);
745 fs_inst *last_rhs_inst = (fs_inst *) this->instructions.get_tail();
747 assert(l.file != BAD_FILE);
748 assert(r.file != BAD_FILE);
750 if (try_rewrite_rhs_to_dst(ir, l, r, pre_rhs_inst, last_rhs_inst))
754 emit_bool_to_cond_code(ir->condition);
757 if (ir->lhs->type->is_scalar() ||
758 ir->lhs->type->is_vector()) {
759 for (int i = 0; i < ir->lhs->type->vector_elements; i++) {
760 if (ir->write_mask & (1 << i)) {
761 inst = emit(MOV(l, r));
763 inst->predicate = BRW_PREDICATE_NORMAL;
769 emit_assignment_writes(l, r, ir->lhs->type, ir->condition != NULL);
774 fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
775 fs_reg shadow_c, fs_reg lod, fs_reg dPdy)
785 if (ir->shadow_comparitor) {
786 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
787 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
788 coordinate.reg_offset++;
790 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
793 if (ir->op == ir_tex) {
794 /* There's no plain shadow compare message, so we use shadow
795 * compare with a bias of 0.0.
797 emit(MOV(fs_reg(MRF, base_mrf + mlen), fs_reg(0.0f)));
799 } else if (ir->op == ir_txb || ir->op == ir_txl) {
800 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
803 assert(!"Should not get here.");
806 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
808 } else if (ir->op == ir_tex) {
809 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
810 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
811 coordinate.reg_offset++;
813 /* gen4's SIMD8 sampler always has the slots for u,v,r present. */
815 } else if (ir->op == ir_txd) {
818 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
819 emit(MOV(fs_reg(MRF, base_mrf + mlen + i), coordinate));
820 coordinate.reg_offset++;
822 /* the slots for u and v are always present, but r is optional */
823 mlen += MAX2(ir->coordinate->type->vector_elements, 2);
826 * dPdx = dudx, dvdx, drdx
827 * dPdy = dudy, dvdy, drdy
829 * 1-arg: Does not exist.
831 * 2-arg: dudx dvdx dudy dvdy
832 * dPdx.x dPdx.y dPdy.x dPdy.y
835 * 3-arg: dudx dvdx drdx dudy dvdy drdy
836 * dPdx.x dPdx.y dPdx.z dPdy.x dPdy.y dPdy.z
839 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
840 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdx));
843 mlen += MAX2(ir->lod_info.grad.dPdx->type->vector_elements, 2);
845 for (int i = 0; i < ir->lod_info.grad.dPdy->type->vector_elements; i++) {
846 emit(MOV(fs_reg(MRF, base_mrf + mlen), dPdy));
849 mlen += MAX2(ir->lod_info.grad.dPdy->type->vector_elements, 2);
850 } else if (ir->op == ir_txs) {
851 /* There's no SIMD8 resinfo message on Gen4. Use SIMD16 instead. */
853 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
856 /* Oh joy. gen4 doesn't have SIMD8 non-shadow-compare bias/lod
857 * instructions. We'll need to do SIMD16 here.
860 assert(ir->op == ir_txb || ir->op == ir_txl || ir->op == ir_txf);
862 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
863 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2, coordinate.type),
865 coordinate.reg_offset++;
868 /* Initialize the rest of u/v/r with 0.0. Empirically, this seems to
869 * be necessary for TXF (ld), but seems wise to do for all messages.
871 for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
872 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * 2), fs_reg(0.0f)));
875 /* lod/bias appears after u/v/r. */
878 emit(MOV(fs_reg(MRF, base_mrf + mlen, lod.type), lod));
881 /* The unused upper half. */
886 /* Now, since we're doing simd16, the return is 2 interleaved
887 * vec4s where the odd-indexed ones are junk. We'll need to move
888 * this weirdness around to the expected layout.
891 const glsl_type *vec_type =
892 glsl_type::get_instance(ir->type->base_type, 4, 1);
893 dst = fs_reg(this, glsl_type::get_array_instance(vec_type, 2));
894 dst.type = intel->is_g4x ? brw_type_for_base_type(ir->type)
895 : BRW_REGISTER_TYPE_F;
898 fs_inst *inst = NULL;
901 inst = emit(SHADER_OPCODE_TEX, dst);
904 inst = emit(FS_OPCODE_TXB, dst);
907 inst = emit(SHADER_OPCODE_TXL, dst);
910 inst = emit(SHADER_OPCODE_TXD, dst);
913 inst = emit(SHADER_OPCODE_TXS, dst);
916 inst = emit(SHADER_OPCODE_TXF, dst);
919 inst->base_mrf = base_mrf;
921 inst->header_present = true;
924 for (int i = 0; i < 4; i++) {
925 emit(MOV(orig_dst, dst));
926 orig_dst.reg_offset++;
934 /* gen5's sampler has slots for u, v, r, array index, then optional
935 * parameters like shadow comparitor or LOD bias. If optional
936 * parameters aren't present, those base slots are optional and don't
937 * need to be included in the message.
939 * We don't fill in the unnecessary slots regardless, which may look
940 * surprising in the disassembly.
943 fs_visitor::emit_texture_gen5(ir_texture *ir, fs_reg dst, fs_reg coordinate,
944 fs_reg shadow_c, fs_reg lod, fs_reg lod2)
948 int reg_width = dispatch_width / 8;
949 bool header_present = false;
950 const int vector_elements =
951 ir->coordinate ? ir->coordinate->type->vector_elements : 0;
953 if (ir->offset != NULL && ir->op == ir_txf) {
954 /* It appears that the ld instruction used for txf does its
955 * address bounds check before adding in the offset. To work
956 * around this, just add the integer offset to the integer texel
957 * coordinate, and don't put the offset in the header.
959 ir_constant *offset = ir->offset->as_constant();
960 for (int i = 0; i < vector_elements; i++) {
961 emit(ADD(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
963 offset->value.i[i]));
964 coordinate.reg_offset++;
968 /* The offsets set up by the ir_texture visitor are in the
969 * m1 header, so we can't go headerless.
971 header_present = true;
976 for (int i = 0; i < vector_elements; i++) {
977 emit(MOV(fs_reg(MRF, base_mrf + mlen + i * reg_width, coordinate.type),
979 coordinate.reg_offset++;
982 mlen += vector_elements * reg_width;
984 if (ir->shadow_comparitor) {
985 mlen = MAX2(mlen, header_present + 4 * reg_width);
987 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
991 fs_inst *inst = NULL;
994 inst = emit(SHADER_OPCODE_TEX, dst);
997 mlen = MAX2(mlen, header_present + 4 * reg_width);
998 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1001 inst = emit(FS_OPCODE_TXB, dst);
1004 mlen = MAX2(mlen, header_present + 4 * reg_width);
1005 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1008 inst = emit(SHADER_OPCODE_TXL, dst);
1011 mlen = MAX2(mlen, header_present + 4 * reg_width); /* skip over 'ai' */
1015 * dPdx = dudx, dvdx, drdx
1016 * dPdy = dudy, dvdy, drdy
1018 * Load up these values:
1019 * - dudx dudy dvdx dvdy drdx drdy
1020 * - dPdx.x dPdy.x dPdx.y dPdy.y dPdx.z dPdy.z
1022 for (int i = 0; i < ir->lod_info.grad.dPdx->type->vector_elements; i++) {
1023 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1027 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
1032 inst = emit(SHADER_OPCODE_TXD, dst);
1036 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1038 inst = emit(SHADER_OPCODE_TXS, dst);
1041 mlen = header_present + 4 * reg_width;
1043 emit(MOV(fs_reg(MRF, base_mrf + mlen - reg_width, BRW_REGISTER_TYPE_UD),
1045 inst = emit(SHADER_OPCODE_TXF, dst);
1048 inst->base_mrf = base_mrf;
1050 inst->header_present = header_present;
1053 fail("Message length >11 disallowed by hardware\n");
1060 fs_visitor::emit_texture_gen7(ir_texture *ir, fs_reg dst, fs_reg coordinate,
1061 fs_reg shadow_c, fs_reg lod, fs_reg lod2)
1065 int reg_width = dispatch_width / 8;
1066 bool header_present = false;
1069 if (ir->offset && ir->op != ir_txf) {
1070 /* The offsets set up by the ir_texture visitor are in the
1071 * m1 header, so we can't go headerless.
1073 header_present = true;
1078 if (ir->shadow_comparitor) {
1079 emit(MOV(fs_reg(MRF, base_mrf + mlen), shadow_c));
1083 /* Set up the LOD info */
1088 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1092 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1096 if (dispatch_width == 16)
1097 fail("Gen7 does not support sample_d/sample_d_c in SIMD16 mode.");
1099 /* Load dPdx and the coordinate together:
1100 * [hdr], [ref], x, dPdx.x, dPdy.x, y, dPdx.y, dPdy.y, z, dPdx.z, dPdy.z
1102 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1103 emit(MOV(fs_reg(MRF, base_mrf + mlen), coordinate));
1104 coordinate.reg_offset++;
1107 /* For cube map array, the coordinate is (u,v,r,ai) but there are
1108 * only derivatives for (u, v, r).
1110 if (i < ir->lod_info.grad.dPdx->type->vector_elements) {
1111 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod));
1115 emit(MOV(fs_reg(MRF, base_mrf + mlen), lod2));
1123 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_UD), lod));
1127 /* It appears that the ld instruction used for txf does its
1128 * address bounds check before adding in the offset. To work
1129 * around this, just add the integer offset to the integer texel
1130 * coordinate, and don't put the offset in the header.
1133 ir_constant *offset = ir->offset->as_constant();
1134 offsets[0] = offset->value.i[0];
1135 offsets[1] = offset->value.i[1];
1136 offsets[2] = offset->value.i[2];
1138 memset(offsets, 0, sizeof(offsets));
1141 /* Unfortunately, the parameters for LD are intermixed: u, lod, v, r. */
1142 emit(ADD(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
1143 coordinate, offsets[0]));
1144 coordinate.reg_offset++;
1147 emit(MOV(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D), lod));
1150 for (int i = 1; i < ir->coordinate->type->vector_elements; i++) {
1151 emit(ADD(fs_reg(MRF, base_mrf + mlen, BRW_REGISTER_TYPE_D),
1152 coordinate, offsets[i]));
1153 coordinate.reg_offset++;
1159 /* Set up the coordinate (except for cases where it was done above) */
1160 if (ir->op != ir_txd && ir->op != ir_txs && ir->op != ir_txf) {
1161 for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
1162 emit(MOV(fs_reg(MRF, base_mrf + mlen), coordinate));
1163 coordinate.reg_offset++;
1168 /* Generate the SEND */
1169 fs_inst *inst = NULL;
1171 case ir_tex: inst = emit(SHADER_OPCODE_TEX, dst); break;
1172 case ir_txb: inst = emit(FS_OPCODE_TXB, dst); break;
1173 case ir_txl: inst = emit(SHADER_OPCODE_TXL, dst); break;
1174 case ir_txd: inst = emit(SHADER_OPCODE_TXD, dst); break;
1175 case ir_txf: inst = emit(SHADER_OPCODE_TXF, dst); break;
1176 case ir_txs: inst = emit(SHADER_OPCODE_TXS, dst); break;
1178 inst->base_mrf = base_mrf;
1180 inst->header_present = header_present;
1183 fail("Message length >11 disallowed by hardware\n");
1190 fs_visitor::rescale_texcoord(ir_texture *ir, fs_reg coordinate,
1191 bool is_rect, int sampler, int texunit)
1193 fs_inst *inst = NULL;
1194 bool needs_gl_clamp = true;
1195 fs_reg scale_x, scale_y;
1197 /* The 965 requires the EU to do the normalization of GL rectangle
1198 * texture coordinates. We use the program parameter state
1199 * tracking to get the scaling factor.
1203 (intel->gen >= 6 && (c->key.tex.gl_clamp_mask[0] & (1 << sampler) ||
1204 c->key.tex.gl_clamp_mask[1] & (1 << sampler))))) {
1205 struct gl_program_parameter_list *params = fp->Base.Parameters;
1206 int tokens[STATE_LENGTH] = {
1208 STATE_TEXRECT_SCALE,
1214 if (dispatch_width == 16) {
1215 fail("rectangle scale uniform setup not supported on 16-wide\n");
1219 scale_x = fs_reg(UNIFORM, c->prog_data.nr_params);
1220 scale_y = fs_reg(UNIFORM, c->prog_data.nr_params + 1);
1222 GLuint index = _mesa_add_state_reference(params,
1223 (gl_state_index *)tokens);
1224 c->prog_data.param[c->prog_data.nr_params++] =
1225 &fp->Base.Parameters->ParameterValues[index][0].f;
1226 c->prog_data.param[c->prog_data.nr_params++] =
1227 &fp->Base.Parameters->ParameterValues[index][1].f;
1230 /* The 965 requires the EU to do the normalization of GL rectangle
1231 * texture coordinates. We use the program parameter state
1232 * tracking to get the scaling factor.
1234 if (intel->gen < 6 && is_rect) {
1235 fs_reg dst = fs_reg(this, ir->coordinate->type);
1236 fs_reg src = coordinate;
1239 emit(MUL(dst, src, scale_x));
1242 emit(MUL(dst, src, scale_y));
1243 } else if (is_rect) {
1244 /* On gen6+, the sampler handles the rectangle coordinates
1245 * natively, without needing rescaling. But that means we have
1246 * to do GL_CLAMP clamping at the [0, width], [0, height] scale,
1247 * not [0, 1] like the default case below.
1249 needs_gl_clamp = false;
1251 for (int i = 0; i < 2; i++) {
1252 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1253 fs_reg chan = coordinate;
1254 chan.reg_offset += i;
1256 inst = emit(BRW_OPCODE_SEL, chan, chan, brw_imm_f(0.0));
1257 inst->conditional_mod = BRW_CONDITIONAL_G;
1259 /* Our parameter comes in as 1.0/width or 1.0/height,
1260 * because that's what people normally want for doing
1261 * texture rectangle handling. We need width or height
1262 * for clamping, but we don't care enough to make a new
1263 * parameter type, so just invert back.
1265 fs_reg limit = fs_reg(this, glsl_type::float_type);
1266 emit(MOV(limit, i == 0 ? scale_x : scale_y));
1267 emit(SHADER_OPCODE_RCP, limit, limit);
1269 inst = emit(BRW_OPCODE_SEL, chan, chan, limit);
1270 inst->conditional_mod = BRW_CONDITIONAL_L;
1275 if (ir->coordinate && needs_gl_clamp) {
1276 for (unsigned int i = 0;
1277 i < MIN2(ir->coordinate->type->vector_elements, 3); i++) {
1278 if (c->key.tex.gl_clamp_mask[i] & (1 << sampler)) {
1279 fs_reg chan = coordinate;
1280 chan.reg_offset += i;
1282 fs_inst *inst = emit(MOV(chan, chan));
1283 inst->saturate = true;
1291 fs_visitor::visit(ir_texture *ir)
1293 fs_inst *inst = NULL;
1295 int sampler = _mesa_get_sampler_uniform_value(ir->sampler, prog, &fp->Base);
1296 /* FINISHME: We're failing to recompile our programs when the sampler is
1297 * updated. This only matters for the texture rectangle scale parameters
1298 * (pre-gen6, or gen6+ with GL_CLAMP).
1300 int texunit = fp->Base.SamplerUnits[sampler];
1302 /* Should be lowered by do_lower_texture_projection */
1303 assert(!ir->projector);
1305 /* Generate code to compute all the subexpression trees. This has to be
1306 * done before loading any values into MRFs for the sampler message since
1307 * generating these values may involve SEND messages that need the MRFs.
1310 if (ir->coordinate) {
1311 ir->coordinate->accept(this);
1313 coordinate = rescale_texcoord(ir, this->result,
1314 ir->sampler->type->sampler_dimensionality ==
1315 GLSL_SAMPLER_DIM_RECT,
1319 fs_reg shadow_comparitor;
1320 if (ir->shadow_comparitor) {
1321 ir->shadow_comparitor->accept(this);
1322 shadow_comparitor = this->result;
1330 ir->lod_info.bias->accept(this);
1334 ir->lod_info.grad.dPdx->accept(this);
1337 ir->lod_info.grad.dPdy->accept(this);
1338 lod2 = this->result;
1343 ir->lod_info.lod->accept(this);
1348 /* Writemasking doesn't eliminate channels on SIMD8 texture
1349 * samples, so don't worry about them.
1351 fs_reg dst = fs_reg(this, glsl_type::get_instance(ir->type->base_type, 4, 1));
1353 if (intel->gen >= 7) {
1354 inst = emit_texture_gen7(ir, dst, coordinate, shadow_comparitor,
1356 } else if (intel->gen >= 5) {
1357 inst = emit_texture_gen5(ir, dst, coordinate, shadow_comparitor,
1360 inst = emit_texture_gen4(ir, dst, coordinate, shadow_comparitor,
1364 /* The header is set up by generate_tex() when necessary. */
1365 inst->src[0] = reg_undef;
1367 if (ir->offset != NULL && ir->op != ir_txf)
1368 inst->texture_offset = brw_texture_offset(ir->offset->as_constant());
1370 inst->sampler = sampler;
1372 if (ir->shadow_comparitor)
1373 inst->shadow_compare = true;
1375 /* fixup #layers for cube map arrays */
1376 if (ir->op == ir_txs) {
1377 glsl_type const *type = ir->sampler->type;
1378 if (type->sampler_dimensionality == GLSL_SAMPLER_DIM_CUBE &&
1379 type->sampler_array) {
1381 depth.reg_offset = 2;
1382 emit_math(SHADER_OPCODE_INT_QUOTIENT, depth, depth, fs_reg(6));
1386 swizzle_result(ir, dst, sampler);
1390 * Swizzle the result of a texture result. This is necessary for
1391 * EXT_texture_swizzle as well as DEPTH_TEXTURE_MODE for shadow comparisons.
1394 fs_visitor::swizzle_result(ir_texture *ir, fs_reg orig_val, int sampler)
1396 this->result = orig_val;
1398 if (ir->op == ir_txs)
1401 if (ir->type == glsl_type::float_type) {
1402 /* Ignore DEPTH_TEXTURE_MODE swizzling. */
1403 assert(ir->sampler->type->sampler_shadow);
1404 } else if (c->key.tex.swizzles[sampler] != SWIZZLE_NOOP) {
1405 fs_reg swizzled_result = fs_reg(this, glsl_type::vec4_type);
1407 for (int i = 0; i < 4; i++) {
1408 int swiz = GET_SWZ(c->key.tex.swizzles[sampler], i);
1409 fs_reg l = swizzled_result;
1412 if (swiz == SWIZZLE_ZERO) {
1413 emit(MOV(l, fs_reg(0.0f)));
1414 } else if (swiz == SWIZZLE_ONE) {
1415 emit(MOV(l, fs_reg(1.0f)));
1417 fs_reg r = orig_val;
1418 r.reg_offset += GET_SWZ(c->key.tex.swizzles[sampler], i);
1422 this->result = swizzled_result;
1427 fs_visitor::visit(ir_swizzle *ir)
1429 ir->val->accept(this);
1430 fs_reg val = this->result;
1432 if (ir->type->vector_elements == 1) {
1433 this->result.reg_offset += ir->mask.x;
1437 fs_reg result = fs_reg(this, ir->type);
1438 this->result = result;
1440 for (unsigned int i = 0; i < ir->type->vector_elements; i++) {
1441 fs_reg channel = val;
1459 channel.reg_offset += swiz;
1460 emit(MOV(result, channel));
1461 result.reg_offset++;
1466 fs_visitor::visit(ir_discard *ir)
1468 assert(ir->condition == NULL); /* FINISHME */
1470 /* We track our discarded pixels in f0.1. By predicating on it, we can
1471 * update just the flag bits that aren't yet discarded. By emitting a
1472 * CMP of g0 != g0, all our currently executing channels will get turned
1475 fs_reg some_reg = fs_reg(retype(brw_vec8_grf(0, 0),
1476 BRW_REGISTER_TYPE_UW));
1477 fs_inst *cmp = emit(CMP(reg_null_f, some_reg, some_reg,
1478 BRW_CONDITIONAL_NZ));
1479 cmp->predicate = BRW_PREDICATE_NORMAL;
1480 cmp->flag_subreg = 1;
1482 if (intel->gen >= 6) {
1483 /* For performance, after a discard, jump to the end of the shader.
1484 * However, many people will do foliage by discarding based on a
1485 * texture's alpha mask, and then continue on to texture with the
1486 * remaining pixels. To avoid trashing the derivatives for those
1487 * texture samples, we'll only jump if all of the pixels in the subspan
1488 * have been discarded.
1490 fs_inst *discard_jump = emit(FS_OPCODE_DISCARD_JUMP);
1491 discard_jump->flag_subreg = 1;
1492 discard_jump->predicate = BRW_PREDICATE_ALIGN1_ANY4H;
1493 discard_jump->predicate_inverse = true;
1498 fs_visitor::visit(ir_constant *ir)
1500 /* Set this->result to reg at the bottom of the function because some code
1501 * paths will cause this visitor to be applied to other fields. This will
1502 * cause the value stored in this->result to be modified.
1504 * Make reg constant so that it doesn't get accidentally modified along the
1505 * way. Yes, I actually had this problem. :(
1507 const fs_reg reg(this, ir->type);
1508 fs_reg dst_reg = reg;
1510 if (ir->type->is_array()) {
1511 const unsigned size = type_size(ir->type->fields.array);
1513 for (unsigned i = 0; i < ir->type->length; i++) {
1514 ir->array_elements[i]->accept(this);
1515 fs_reg src_reg = this->result;
1517 dst_reg.type = src_reg.type;
1518 for (unsigned j = 0; j < size; j++) {
1519 emit(MOV(dst_reg, src_reg));
1520 src_reg.reg_offset++;
1521 dst_reg.reg_offset++;
1524 } else if (ir->type->is_record()) {
1525 foreach_list(node, &ir->components) {
1526 ir_constant *const field = (ir_constant *) node;
1527 const unsigned size = type_size(field->type);
1529 field->accept(this);
1530 fs_reg src_reg = this->result;
1532 dst_reg.type = src_reg.type;
1533 for (unsigned j = 0; j < size; j++) {
1534 emit(MOV(dst_reg, src_reg));
1535 src_reg.reg_offset++;
1536 dst_reg.reg_offset++;
1540 const unsigned size = type_size(ir->type);
1542 for (unsigned i = 0; i < size; i++) {
1543 switch (ir->type->base_type) {
1544 case GLSL_TYPE_FLOAT:
1545 emit(MOV(dst_reg, fs_reg(ir->value.f[i])));
1547 case GLSL_TYPE_UINT:
1548 emit(MOV(dst_reg, fs_reg(ir->value.u[i])));
1551 emit(MOV(dst_reg, fs_reg(ir->value.i[i])));
1553 case GLSL_TYPE_BOOL:
1554 emit(MOV(dst_reg, fs_reg((int)ir->value.b[i])));
1557 assert(!"Non-float/uint/int/bool constant");
1559 dst_reg.reg_offset++;
1567 fs_visitor::emit_bool_to_cond_code(ir_rvalue *ir)
1569 ir_expression *expr = ir->as_expression();
1575 assert(expr->get_num_operands() <= 2);
1576 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1577 assert(expr->operands[i]->type->is_scalar());
1579 expr->operands[i]->accept(this);
1580 op[i] = this->result;
1582 resolve_ud_negate(&op[i]);
1585 switch (expr->operation) {
1586 case ir_unop_logic_not:
1587 inst = emit(AND(reg_null_d, op[0], fs_reg(1)));
1588 inst->conditional_mod = BRW_CONDITIONAL_Z;
1591 case ir_binop_logic_xor:
1592 case ir_binop_logic_or:
1593 case ir_binop_logic_and:
1597 if (intel->gen >= 6) {
1598 emit(CMP(reg_null_d, op[0], fs_reg(0.0f), BRW_CONDITIONAL_NZ));
1600 inst = emit(MOV(reg_null_f, op[0]));
1601 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1606 if (intel->gen >= 6) {
1607 emit(CMP(reg_null_d, op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1609 inst = emit(MOV(reg_null_d, op[0]));
1610 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1614 case ir_binop_greater:
1615 case ir_binop_gequal:
1617 case ir_binop_lequal:
1618 case ir_binop_equal:
1619 case ir_binop_all_equal:
1620 case ir_binop_nequal:
1621 case ir_binop_any_nequal:
1622 resolve_bool_comparison(expr->operands[0], &op[0]);
1623 resolve_bool_comparison(expr->operands[1], &op[1]);
1625 emit(CMP(reg_null_d, op[0], op[1],
1626 brw_conditional_for_comparison(expr->operation)));
1630 assert(!"not reached");
1631 fail("bad cond code\n");
1640 fs_inst *inst = emit(AND(reg_null_d, this->result, fs_reg(1)));
1641 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1645 * Emit a gen6 IF statement with the comparison folded into the IF
1649 fs_visitor::emit_if_gen6(ir_if *ir)
1651 ir_expression *expr = ir->condition->as_expression();
1658 assert(expr->get_num_operands() <= 2);
1659 for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
1660 assert(expr->operands[i]->type->is_scalar());
1662 expr->operands[i]->accept(this);
1663 op[i] = this->result;
1666 switch (expr->operation) {
1667 case ir_unop_logic_not:
1668 case ir_binop_logic_xor:
1669 case ir_binop_logic_or:
1670 case ir_binop_logic_and:
1671 /* For operations on bool arguments, only the low bit of the bool is
1672 * valid, and the others are undefined. Fall back to the condition
1678 inst = emit(BRW_OPCODE_IF, reg_null_f, op[0], fs_reg(0));
1679 inst->conditional_mod = BRW_CONDITIONAL_NZ;
1683 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1686 case ir_binop_greater:
1687 case ir_binop_gequal:
1689 case ir_binop_lequal:
1690 case ir_binop_equal:
1691 case ir_binop_all_equal:
1692 case ir_binop_nequal:
1693 case ir_binop_any_nequal:
1694 resolve_bool_comparison(expr->operands[0], &op[0]);
1695 resolve_bool_comparison(expr->operands[1], &op[1]);
1697 emit(IF(op[0], op[1],
1698 brw_conditional_for_comparison(expr->operation)));
1701 assert(!"not reached");
1702 emit(IF(op[0], fs_reg(0), BRW_CONDITIONAL_NZ));
1703 fail("bad condition\n");
1708 emit_bool_to_cond_code(ir->condition);
1709 fs_inst *inst = emit(BRW_OPCODE_IF);
1710 inst->predicate = BRW_PREDICATE_NORMAL;
1714 fs_visitor::visit(ir_if *ir)
1716 if (intel->gen < 6 && dispatch_width == 16) {
1717 fail("Can't support (non-uniform) control flow on 16-wide\n");
1720 /* Don't point the annotation at the if statement, because then it plus
1721 * the then and else blocks get printed.
1723 this->base_ir = ir->condition;
1725 if (intel->gen == 6) {
1728 emit_bool_to_cond_code(ir->condition);
1730 emit(IF(BRW_PREDICATE_NORMAL));
1733 foreach_list(node, &ir->then_instructions) {
1734 ir_instruction *ir = (ir_instruction *)node;
1740 if (!ir->else_instructions.is_empty()) {
1741 emit(BRW_OPCODE_ELSE);
1743 foreach_list(node, &ir->else_instructions) {
1744 ir_instruction *ir = (ir_instruction *)node;
1751 emit(BRW_OPCODE_ENDIF);
1755 fs_visitor::visit(ir_loop *ir)
1757 fs_reg counter = reg_undef;
1759 if (intel->gen < 6 && dispatch_width == 16) {
1760 fail("Can't support (non-uniform) control flow on 16-wide\n");
1764 this->base_ir = ir->counter;
1765 ir->counter->accept(this);
1766 counter = *(variable_storage(ir->counter));
1769 this->base_ir = ir->from;
1770 ir->from->accept(this);
1772 emit(MOV(counter, this->result));
1776 this->base_ir = NULL;
1777 emit(BRW_OPCODE_DO);
1780 this->base_ir = ir->to;
1781 ir->to->accept(this);
1783 emit(CMP(reg_null_d, counter, this->result,
1784 brw_conditional_for_comparison(ir->cmp)));
1786 fs_inst *inst = emit(BRW_OPCODE_BREAK);
1787 inst->predicate = BRW_PREDICATE_NORMAL;
1790 foreach_list(node, &ir->body_instructions) {
1791 ir_instruction *ir = (ir_instruction *)node;
1797 if (ir->increment) {
1798 this->base_ir = ir->increment;
1799 ir->increment->accept(this);
1800 emit(ADD(counter, counter, this->result));
1803 this->base_ir = NULL;
1804 emit(BRW_OPCODE_WHILE);
1808 fs_visitor::visit(ir_loop_jump *ir)
1811 case ir_loop_jump::jump_break:
1812 emit(BRW_OPCODE_BREAK);
1814 case ir_loop_jump::jump_continue:
1815 emit(BRW_OPCODE_CONTINUE);
1821 fs_visitor::visit(ir_call *ir)
1823 assert(!"FINISHME");
1827 fs_visitor::visit(ir_return *ir)
1829 assert(!"FINISHME");
1833 fs_visitor::visit(ir_function *ir)
1835 /* Ignore function bodies other than main() -- we shouldn't see calls to
1836 * them since they should all be inlined before we get to ir_to_mesa.
1838 if (strcmp(ir->name, "main") == 0) {
1839 const ir_function_signature *sig;
1842 sig = ir->matching_signature(&empty);
1846 foreach_list(node, &sig->body) {
1847 ir_instruction *ir = (ir_instruction *)node;
1856 fs_visitor::visit(ir_function_signature *ir)
1858 assert(!"not reached");
1863 fs_visitor::emit(fs_inst inst)
1865 fs_inst *list_inst = new(mem_ctx) fs_inst;
1872 fs_visitor::emit(fs_inst *inst)
1874 if (force_uncompressed_stack > 0)
1875 inst->force_uncompressed = true;
1876 else if (force_sechalf_stack > 0)
1877 inst->force_sechalf = true;
1879 inst->annotation = this->current_annotation;
1880 inst->ir = this->base_ir;
1882 this->instructions.push_tail(inst);
1888 fs_visitor::emit(exec_list list)
1890 foreach_list_safe(node, &list) {
1891 fs_inst *inst = (fs_inst *)node;
1897 /** Emits a dummy fragment shader consisting of magenta for bringup purposes. */
1899 fs_visitor::emit_dummy_fs()
1901 int reg_width = dispatch_width / 8;
1903 /* Everyone's favorite color. */
1904 emit(MOV(fs_reg(MRF, 2 + 0 * reg_width), fs_reg(1.0f)));
1905 emit(MOV(fs_reg(MRF, 2 + 1 * reg_width), fs_reg(0.0f)));
1906 emit(MOV(fs_reg(MRF, 2 + 2 * reg_width), fs_reg(1.0f)));
1907 emit(MOV(fs_reg(MRF, 2 + 3 * reg_width), fs_reg(0.0f)));
1910 write = emit(FS_OPCODE_FB_WRITE, fs_reg(0), fs_reg(0));
1911 write->base_mrf = 2;
1912 write->mlen = 4 * reg_width;
1916 /* The register location here is relative to the start of the URB
1917 * data. It will get adjusted to be a real location before
1918 * generate_code() time.
1921 fs_visitor::interp_reg(int location, int channel)
1923 int regnr = urb_setup[location] * 2 + channel / 2;
1924 int stride = (channel & 1) * 4;
1926 assert(urb_setup[location] != -1);
1928 return brw_vec1_grf(regnr, stride);
1931 /** Emits the interpolation for the varying inputs. */
1933 fs_visitor::emit_interpolation_setup_gen4()
1935 this->current_annotation = "compute pixel centers";
1936 this->pixel_x = fs_reg(this, glsl_type::uint_type);
1937 this->pixel_y = fs_reg(this, glsl_type::uint_type);
1938 this->pixel_x.type = BRW_REGISTER_TYPE_UW;
1939 this->pixel_y.type = BRW_REGISTER_TYPE_UW;
1941 emit(FS_OPCODE_PIXEL_X, this->pixel_x);
1942 emit(FS_OPCODE_PIXEL_Y, this->pixel_y);
1944 this->current_annotation = "compute pixel deltas from v0";
1946 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
1947 fs_reg(this, glsl_type::vec2_type);
1948 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
1949 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC];
1950 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC].reg_offset++;
1952 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
1953 fs_reg(this, glsl_type::float_type);
1954 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC] =
1955 fs_reg(this, glsl_type::float_type);
1957 emit(ADD(this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
1958 this->pixel_x, fs_reg(negate(brw_vec1_grf(1, 0)))));
1959 emit(ADD(this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
1960 this->pixel_y, fs_reg(negate(brw_vec1_grf(1, 1)))));
1962 this->current_annotation = "compute pos.w and 1/pos.w";
1963 /* Compute wpos.w. It's always in our setup, since it's needed to
1964 * interpolate the other attributes.
1966 this->wpos_w = fs_reg(this, glsl_type::float_type);
1967 emit(FS_OPCODE_LINTERP, wpos_w,
1968 this->delta_x[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
1969 this->delta_y[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
1970 interp_reg(FRAG_ATTRIB_WPOS, 3));
1971 /* Compute the pixel 1/W value from wpos.w. */
1972 this->pixel_w = fs_reg(this, glsl_type::float_type);
1973 emit_math(SHADER_OPCODE_RCP, this->pixel_w, wpos_w);
1974 this->current_annotation = NULL;
1977 /** Emits the interpolation for the varying inputs. */
1979 fs_visitor::emit_interpolation_setup_gen6()
1981 struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
1983 /* If the pixel centers end up used, the setup is the same as for gen4. */
1984 this->current_annotation = "compute pixel centers";
1985 fs_reg int_pixel_x = fs_reg(this, glsl_type::uint_type);
1986 fs_reg int_pixel_y = fs_reg(this, glsl_type::uint_type);
1987 int_pixel_x.type = BRW_REGISTER_TYPE_UW;
1988 int_pixel_y.type = BRW_REGISTER_TYPE_UW;
1989 emit(ADD(int_pixel_x,
1990 fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
1991 fs_reg(brw_imm_v(0x10101010))));
1992 emit(ADD(int_pixel_y,
1993 fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
1994 fs_reg(brw_imm_v(0x11001100))));
1996 /* As of gen6, we can no longer mix float and int sources. We have
1997 * to turn the integer pixel centers into floats for their actual
2000 this->pixel_x = fs_reg(this, glsl_type::float_type);
2001 this->pixel_y = fs_reg(this, glsl_type::float_type);
2002 emit(MOV(this->pixel_x, int_pixel_x));
2003 emit(MOV(this->pixel_y, int_pixel_y));
2005 this->current_annotation = "compute pos.w";
2006 this->pixel_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
2007 this->wpos_w = fs_reg(this, glsl_type::float_type);
2008 emit_math(SHADER_OPCODE_RCP, this->wpos_w, this->pixel_w);
2010 for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
2011 uint8_t reg = c->barycentric_coord_reg[i];
2012 this->delta_x[i] = fs_reg(brw_vec8_grf(reg, 0));
2013 this->delta_y[i] = fs_reg(brw_vec8_grf(reg + 1, 0));
2016 this->current_annotation = NULL;
2020 fs_visitor::emit_color_write(int target, int index, int first_color_mrf)
2022 int reg_width = dispatch_width / 8;
2024 fs_reg color = outputs[target];
2027 /* If there's no color data to be written, skip it. */
2028 if (color.file == BAD_FILE)
2031 color.reg_offset += index;
2033 if (dispatch_width == 8 || intel->gen >= 6) {
2034 /* SIMD8 write looks like:
2040 * gen6 SIMD16 DP write looks like:
2050 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index * reg_width,
2053 inst->saturate = c->key.clamp_fragment_color;
2055 /* pre-gen6 SIMD16 single source DP write looks like:
2065 if (brw->has_compr4) {
2066 /* By setting the high bit of the MRF register number, we
2067 * indicate that we want COMPR4 mode - instead of doing the
2068 * usual destination + 1 for the second half we get
2071 inst = emit(MOV(fs_reg(MRF, BRW_MRF_COMPR4 + first_color_mrf + index,
2074 inst->saturate = c->key.clamp_fragment_color;
2076 push_force_uncompressed();
2077 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index, color.type),
2079 inst->saturate = c->key.clamp_fragment_color;
2080 pop_force_uncompressed();
2082 push_force_sechalf();
2083 color.sechalf = true;
2084 inst = emit(MOV(fs_reg(MRF, first_color_mrf + index + 4, color.type),
2086 inst->saturate = c->key.clamp_fragment_color;
2087 pop_force_sechalf();
2088 color.sechalf = false;
2094 fs_visitor::emit_fb_writes()
2096 this->current_annotation = "FB write header";
2097 bool header_present = true;
2098 /* We can potentially have a message length of up to 15, so we have to set
2099 * base_mrf to either 0 or 1 in order to fit in m0..m15.
2103 int reg_width = dispatch_width / 8;
2104 bool do_dual_src = this->dual_src_output.file != BAD_FILE;
2105 bool src0_alpha_to_render_target = false;
2107 if (dispatch_width == 16 && do_dual_src) {
2108 fail("GL_ARB_blend_func_extended not yet supported in 16-wide.");
2109 do_dual_src = false;
2112 /* From the Sandy Bridge PRM, volume 4, page 198:
2114 * "Dispatched Pixel Enables. One bit per pixel indicating
2115 * which pixels were originally enabled when the thread was
2116 * dispatched. This field is only required for the end-of-
2117 * thread message and on all dual-source messages."
2119 if (intel->gen >= 6 &&
2120 !this->fp->UsesKill &&
2122 c->key.nr_color_regions == 1) {
2123 header_present = false;
2126 if (header_present) {
2127 src0_alpha_to_render_target = intel->gen >= 6 &&
2129 c->key.nr_color_regions > 1 &&
2130 c->key.sample_alpha_to_coverage;
2135 if (c->aa_dest_stencil_reg) {
2136 push_force_uncompressed();
2137 emit(MOV(fs_reg(MRF, nr++),
2138 fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0))));
2139 pop_force_uncompressed();
2142 /* Reserve space for color. It'll be filled in per MRT below. */
2144 nr += 4 * reg_width;
2147 if (src0_alpha_to_render_target)
2150 if (c->source_depth_to_render_target) {
2151 if (intel->gen == 6 && dispatch_width == 16) {
2152 /* For outputting oDepth on gen6, SIMD8 writes have to be
2153 * used. This would require 8-wide moves of each half to
2154 * message regs, kind of like pre-gen5 SIMD16 FB writes.
2155 * Just bail on doing so for now.
2157 fail("Missing support for simd16 depth writes on gen6\n");
2160 if (fp->Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
2161 /* Hand over gl_FragDepth. */
2162 assert(this->frag_depth.file != BAD_FILE);
2163 emit(MOV(fs_reg(MRF, nr), this->frag_depth));
2165 /* Pass through the payload depth. */
2166 emit(MOV(fs_reg(MRF, nr),
2167 fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
2172 if (c->dest_depth_reg) {
2173 emit(MOV(fs_reg(MRF, nr),
2174 fs_reg(brw_vec8_grf(c->dest_depth_reg, 0))));
2179 fs_reg src0 = this->outputs[0];
2180 fs_reg src1 = this->dual_src_output;
2182 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2184 for (int i = 0; i < 4; i++) {
2185 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + i, src0.type), src0));
2187 inst->saturate = c->key.clamp_fragment_color;
2190 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2192 for (int i = 0; i < 4; i++) {
2193 fs_inst *inst = emit(MOV(fs_reg(MRF, color_mrf + 4 + i, src1.type),
2196 inst->saturate = c->key.clamp_fragment_color;
2199 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2201 inst->base_mrf = base_mrf;
2202 inst->mlen = nr - base_mrf;
2204 inst->header_present = header_present;
2206 c->prog_data.dual_src_blend = true;
2207 this->current_annotation = NULL;
2211 for (int target = 0; target < c->key.nr_color_regions; target++) {
2212 this->current_annotation = ralloc_asprintf(this->mem_ctx,
2213 "FB write target %d",
2215 /* If src0_alpha_to_render_target is true, include source zero alpha
2216 * data in RenderTargetWrite message for targets > 0.
2218 int write_color_mrf = color_mrf;
2219 if (src0_alpha_to_render_target && target != 0) {
2221 fs_reg color = outputs[0];
2222 color.reg_offset += 3;
2224 inst = emit(MOV(fs_reg(MRF, write_color_mrf, color.type),
2226 inst->saturate = c->key.clamp_fragment_color;
2227 write_color_mrf = color_mrf + reg_width;
2230 for (unsigned i = 0; i < this->output_components[target]; i++)
2231 emit_color_write(target, i, write_color_mrf);
2233 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2234 inst->target = target;
2235 inst->base_mrf = base_mrf;
2236 if (src0_alpha_to_render_target && target == 0)
2237 inst->mlen = nr - base_mrf - reg_width;
2239 inst->mlen = nr - base_mrf;
2240 if (target == c->key.nr_color_regions - 1)
2242 inst->header_present = header_present;
2245 if (c->key.nr_color_regions == 0) {
2246 /* Even if there's no color buffers enabled, we still need to send
2247 * alpha out the pipeline to our null renderbuffer to support
2248 * alpha-testing, alpha-to-coverage, and so on.
2250 emit_color_write(0, 3, color_mrf);
2252 fs_inst *inst = emit(FS_OPCODE_FB_WRITE);
2253 inst->base_mrf = base_mrf;
2254 inst->mlen = nr - base_mrf;
2256 inst->header_present = header_present;
2259 this->current_annotation = NULL;
2263 fs_visitor::resolve_ud_negate(fs_reg *reg)
2265 if (reg->type != BRW_REGISTER_TYPE_UD ||
2269 fs_reg temp = fs_reg(this, glsl_type::uint_type);
2270 emit(MOV(temp, *reg));
2275 fs_visitor::resolve_bool_comparison(ir_rvalue *rvalue, fs_reg *reg)
2277 if (rvalue->type != glsl_type::bool_type)
2280 fs_reg temp = fs_reg(this, glsl_type::bool_type);
2281 emit(AND(temp, *reg, fs_reg(1)));
2285 fs_visitor::fs_visitor(struct brw_context *brw,
2286 struct brw_wm_compile *c,
2287 struct gl_shader_program *prog,
2288 struct gl_fragment_program *fp,
2289 unsigned dispatch_width)
2290 : dispatch_width(dispatch_width)
2296 this->intel = &brw->intel;
2297 this->ctx = &intel->ctx;
2298 this->mem_ctx = ralloc_context(NULL);
2300 shader = (struct brw_shader *) prog->_LinkedShaders[MESA_SHADER_FRAGMENT];
2303 this->failed = false;
2304 this->variable_ht = hash_table_ctor(0,
2305 hash_table_pointer_hash,
2306 hash_table_pointer_compare);
2308 memset(this->outputs, 0, sizeof(this->outputs));
2309 memset(this->output_components, 0, sizeof(this->output_components));
2310 this->first_non_payload_grf = 0;
2311 this->max_grf = intel->gen >= 7 ? GEN7_MRF_HACK_START : BRW_MAX_GRF;
2313 this->current_annotation = NULL;
2314 this->base_ir = NULL;
2316 this->virtual_grf_sizes = NULL;
2317 this->virtual_grf_count = 0;
2318 this->virtual_grf_array_size = 0;
2319 this->virtual_grf_def = NULL;
2320 this->virtual_grf_use = NULL;
2321 this->live_intervals_valid = false;
2323 this->force_uncompressed_stack = 0;
2324 this->force_sechalf_stack = 0;
2326 memset(&this->param_size, 0, sizeof(this->param_size));
2329 fs_visitor::~fs_visitor()
2331 ralloc_free(this->mem_ctx);
2332 hash_table_dtor(this->variable_ht);