Path: blob/21.2-virgl/src/gallium/drivers/r600/eg_asm.c
4570 views
/*1* Copyright 2010 Jerome Glisse <[email protected]>2*3* Permission is hereby granted, free of charge, to any person obtaining a4* copy of this software and associated documentation files (the "Software"),5* to deal in the Software without restriction, including without limitation6* on the rights to use, copy, modify, merge, publish, distribute, sub7* license, and/or sell copies of the Software, and to permit persons to whom8* the Software is furnished to do so, subject to the following conditions:9*10* The above copyright notice and this permission notice (including the next11* paragraph) shall be included in all copies or substantial portions of the12* Software.13*14* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR15* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,16* FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL17* THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,18* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR19* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE20* USE OR OTHER DEALINGS IN THE SOFTWARE.21*/22#include "r600_pipe.h"23#include "r600_opcodes.h"24#include "r600_shader.h"2526#include "util/u_memory.h"27#include "eg_sq.h"28#include <errno.h>2930int eg_bytecode_cf_build(struct r600_bytecode *bc, struct r600_bytecode_cf *cf)31{32unsigned id = cf->id;3334if (cf->op == CF_NATIVE) {35bc->bytecode[id++] = cf->isa[0];36bc->bytecode[id++] = cf->isa[1];37} else {38const struct cf_op_info *cfop = r600_isa_cf(cf->op);39unsigned opcode = r600_isa_cf_opcode(bc->isa->hw_class, cf->op);40if (cfop->flags & CF_ALU) { /* ALU clauses */4142/* prepend ALU_EXTENDED if we need more than 2 kcache sets */43if (cf->eg_alu_extended) {44bc->bytecode[id++] =45S_SQ_CF_ALU_WORD0_EXT_KCACHE_BANK_INDEX_MODE0(cf->kcache[0].index_mode) |46S_SQ_CF_ALU_WORD0_EXT_KCACHE_BANK_INDEX_MODE1(cf->kcache[1].index_mode) |47S_SQ_CF_ALU_WORD0_EXT_KCACHE_BANK_INDEX_MODE2(cf->kcache[2].index_mode) |48S_SQ_CF_ALU_WORD0_EXT_KCACHE_BANK_INDEX_MODE3(cf->kcache[3].index_mode) |49S_SQ_CF_ALU_WORD0_EXT_KCACHE_BANK2(cf->kcache[2].bank) |50S_SQ_CF_ALU_WORD0_EXT_KCACHE_BANK3(cf->kcache[3].bank) |51S_SQ_CF_ALU_WORD0_EXT_KCACHE_MODE2(cf->kcache[2].mode);52bc->bytecode[id++] =53S_SQ_CF_ALU_WORD1_EXT_CF_INST(54r600_isa_cf_opcode(bc->isa->hw_class, CF_OP_ALU_EXT)) |55S_SQ_CF_ALU_WORD1_EXT_KCACHE_MODE3(cf->kcache[3].mode) |56S_SQ_CF_ALU_WORD1_EXT_KCACHE_ADDR2(cf->kcache[2].addr) |57S_SQ_CF_ALU_WORD1_EXT_KCACHE_ADDR3(cf->kcache[3].addr) |58S_SQ_CF_ALU_WORD1_EXT_BARRIER(1);59}60bc->bytecode[id++] = S_SQ_CF_ALU_WORD0_ADDR(cf->addr >> 1) |61S_SQ_CF_ALU_WORD0_KCACHE_MODE0(cf->kcache[0].mode) |62S_SQ_CF_ALU_WORD0_KCACHE_BANK0(cf->kcache[0].bank) |63S_SQ_CF_ALU_WORD0_KCACHE_BANK1(cf->kcache[1].bank);64bc->bytecode[id++] = S_SQ_CF_ALU_WORD1_CF_INST(opcode) |65S_SQ_CF_ALU_WORD1_KCACHE_MODE1(cf->kcache[1].mode) |66S_SQ_CF_ALU_WORD1_KCACHE_ADDR0(cf->kcache[0].addr) |67S_SQ_CF_ALU_WORD1_KCACHE_ADDR1(cf->kcache[1].addr) |68S_SQ_CF_ALU_WORD1_BARRIER(1) |69S_SQ_CF_ALU_WORD1_COUNT((cf->ndw / 2) - 1);70} else if (cfop->flags & CF_CLAUSE) {71/* CF_TEX/VTX (CF_ALU already handled above) */72bc->bytecode[id++] = S_SQ_CF_WORD0_ADDR(cf->addr >> 1);73bc->bytecode[id] = S_SQ_CF_WORD1_CF_INST(opcode) |74S_SQ_CF_WORD1_BARRIER(1) |75S_SQ_CF_WORD1_VALID_PIXEL_MODE(cf->vpm) |76S_SQ_CF_WORD1_COUNT((cf->ndw / 4) - 1);77if (bc->chip_class == EVERGREEN) /* no EOP on cayman */78bc->bytecode[id] |= S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->end_of_program);79id++;80} else if (cfop->flags & CF_EXP) {81/* EXPORT instructions */82bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(cf->output.gpr) |83S_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(cf->output.elem_size) |84S_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(cf->output.array_base) |85S_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(cf->output.type) |86S_SQ_CF_ALLOC_EXPORT_WORD0_INDEX_GPR(cf->output.index_gpr);87bc->bytecode[id] =88S_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(cf->output.burst_count - 1) |89S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(cf->output.swizzle_x) |90S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Y(cf->output.swizzle_y) |91S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Z(cf->output.swizzle_z) |92S_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_W(cf->output.swizzle_w) |93S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->barrier) |94S_SQ_CF_ALLOC_EXPORT_WORD1_MARK(cf->mark) |95S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(opcode);9697if (bc->chip_class == EVERGREEN) /* no EOP on cayman */98bc->bytecode[id] |= S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->end_of_program);99id++;100} else if (cfop->flags & CF_RAT) {101bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD0_RAT_RAT_ID(cf->rat.id) |102S_SQ_CF_ALLOC_EXPORT_WORD0_RAT_RAT_INST(cf->rat.inst) |103S_SQ_CF_ALLOC_EXPORT_WORD0_RAT_RAT_INDEX_MODE(cf->rat.index_mode) |104S_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(cf->output.gpr) |105S_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(cf->output.elem_size) |106S_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(cf->output.type) |107S_SQ_CF_ALLOC_EXPORT_WORD0_INDEX_GPR(cf->output.index_gpr);108bc->bytecode[id] = S_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(cf->output.burst_count - 1) |109S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->barrier) |110S_SQ_CF_ALLOC_EXPORT_WORD1_MARK(cf->mark) |111S_SQ_CF_ALLOC_EXPORT_WORD1_VALID_PIXEL_MODE(cf->vpm) |112S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(opcode) |113S_SQ_CF_ALLOC_EXPORT_WORD1_BUF_COMP_MASK(cf->output.comp_mask) |114S_SQ_CF_ALLOC_EXPORT_WORD1_BUF_ARRAY_SIZE(cf->output.array_size) |115S_SQ_CF_ALLOC_EXPORT_WORD1_MARK(cf->output.mark);116if (bc->chip_class == EVERGREEN) /* no EOP on cayman */117bc->bytecode[id] |= S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->end_of_program);118id++;119120} else if (cfop->flags & CF_MEM) {121/* MEM_STREAM, MEM_RING instructions */122bc->bytecode[id++] = S_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(cf->output.gpr) |123S_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(cf->output.elem_size) |124S_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(cf->output.array_base) |125S_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(cf->output.type) |126S_SQ_CF_ALLOC_EXPORT_WORD0_INDEX_GPR(cf->output.index_gpr);127bc->bytecode[id] = S_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(cf->output.burst_count - 1) |128S_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(cf->barrier) |129S_SQ_CF_ALLOC_EXPORT_WORD1_MARK(cf->mark) |130S_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(opcode) |131S_SQ_CF_ALLOC_EXPORT_WORD1_BUF_COMP_MASK(cf->output.comp_mask) |132S_SQ_CF_ALLOC_EXPORT_WORD1_BUF_ARRAY_SIZE(cf->output.array_size);133if (bc->chip_class == EVERGREEN) /* no EOP on cayman */134bc->bytecode[id] |= S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->end_of_program);135id++;136} else {137/* other instructions */138bc->bytecode[id++] = S_SQ_CF_WORD0_ADDR(cf->cf_addr >> 1);139bc->bytecode[id] = S_SQ_CF_WORD1_CF_INST(opcode) |140S_SQ_CF_WORD1_VALID_PIXEL_MODE(cf->vpm) |141S_SQ_CF_WORD1_BARRIER(1) |142S_SQ_CF_WORD1_COND(cf->cond) |143S_SQ_CF_WORD1_POP_COUNT(cf->pop_count) |144S_SQ_CF_WORD1_COUNT(cf->count);145if (bc->chip_class == EVERGREEN) /* no EOP on cayman */146bc->bytecode[id] |= S_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(cf->end_of_program);147id++;148}149}150return 0;151}152153#if 0154void eg_bytecode_export_read(struct r600_bytecode *bc,155struct r600_bytecode_output *output, uint32_t word0, uint32_t word1)156{157output->array_base = G_SQ_CF_ALLOC_EXPORT_WORD0_ARRAY_BASE(word0);158output->type = G_SQ_CF_ALLOC_EXPORT_WORD0_TYPE(word0);159output->gpr = G_SQ_CF_ALLOC_EXPORT_WORD0_RW_GPR(word0);160output->elem_size = G_SQ_CF_ALLOC_EXPORT_WORD0_ELEM_SIZE(word0);161162output->swizzle_x = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_X(word1);163output->swizzle_y = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Y(word1);164output->swizzle_z = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_Z(word1);165output->swizzle_w = G_SQ_CF_ALLOC_EXPORT_WORD1_SWIZ_SEL_W(word1);166output->burst_count = G_SQ_CF_ALLOC_EXPORT_WORD1_BURST_COUNT(word1);167output->end_of_program = G_SQ_CF_ALLOC_EXPORT_WORD1_END_OF_PROGRAM(word1);168output->op = r600_isa_cf_by_opcode(bc->isa,169G_SQ_CF_ALLOC_EXPORT_WORD1_CF_INST(word1), /* is_cf_alu = */ 0 );170output->barrier = G_SQ_CF_ALLOC_EXPORT_WORD1_BARRIER(word1);171output->array_size = G_SQ_CF_ALLOC_EXPORT_WORD1_BUF_ARRAY_SIZE(word1);172output->comp_mask = G_SQ_CF_ALLOC_EXPORT_WORD1_BUF_COMP_MASK(word1);173}174#endif175176int egcm_load_index_reg(struct r600_bytecode *bc, unsigned id, bool inside_alu_clause)177{178struct r600_bytecode_alu alu;179int r;180unsigned type;181182assert(id < 2);183assert(bc->chip_class >= EVERGREEN);184185if (bc->index_loaded[id])186return 0;187188memset(&alu, 0, sizeof(alu));189alu.op = ALU_OP1_MOVA_INT;190alu.src[0].sel = bc->index_reg[id];191alu.src[0].chan = bc->index_reg_chan[id];192if (bc->chip_class == CAYMAN)193alu.dst.sel = id == 0 ? CM_V_SQ_MOVA_DST_CF_IDX0 : CM_V_SQ_MOVA_DST_CF_IDX1;194195alu.last = 1;196r = r600_bytecode_add_alu(bc, &alu);197if (r)198return r;199200bc->ar_loaded = 0; /* clobbered */201202if (bc->chip_class == EVERGREEN) {203memset(&alu, 0, sizeof(alu));204alu.op = id == 0 ? ALU_OP0_SET_CF_IDX0 : ALU_OP0_SET_CF_IDX1;205alu.last = 1;206r = r600_bytecode_add_alu(bc, &alu);207if (r)208return r;209}210211/* Must split ALU group as index only applies to following group */212if (inside_alu_clause) {213type = bc->cf_last->op;214if ((r = r600_bytecode_add_cf(bc))) {215return r;216}217bc->cf_last->op = type;218}219220bc->index_loaded[id] = 1;221222return 0;223}224225int eg_bytecode_gds_build(struct r600_bytecode *bc, struct r600_bytecode_gds *gds, unsigned id)226{227unsigned gds_op = (r600_isa_fetch_opcode(bc->isa->hw_class, gds->op) >> 8) & 0x3f;228unsigned opcode;229if (gds->op == FETCH_OP_TF_WRITE) {230opcode = 5;231gds_op = 0;232} else233opcode = 4;234bc->bytecode[id++] = S_SQ_MEM_GDS_WORD0_MEM_INST(2) |235S_SQ_MEM_GDS_WORD0_MEM_OP(opcode) |236S_SQ_MEM_GDS_WORD0_SRC_GPR(gds->src_gpr) |237S_SQ_MEM_GDS_WORD0_SRC_REL(gds->src_rel) |238S_SQ_MEM_GDS_WORD0_SRC_SEL_X(gds->src_sel_x) |239S_SQ_MEM_GDS_WORD0_SRC_SEL_Y(gds->src_sel_y) |240S_SQ_MEM_GDS_WORD0_SRC_SEL_Z(gds->src_sel_z);241242bc->bytecode[id++] = S_SQ_MEM_GDS_WORD1_DST_GPR(gds->dst_gpr) |243S_SQ_MEM_GDS_WORD1_DST_REL(gds->dst_rel) |244S_SQ_MEM_GDS_WORD1_GDS_OP(gds_op) |245S_SQ_MEM_GDS_WORD1_SRC_GPR(gds->src_gpr2) |246S_SQ_MEM_GDS_WORD1_UAV_INDEX_MODE(gds->uav_index_mode) |247S_SQ_MEM_GDS_WORD1_UAV_ID(gds->uav_id) |248S_SQ_MEM_GDS_WORD1_ALLOC_CONSUME(gds->alloc_consume) |249S_SQ_MEM_GDS_WORD1_BCAST_FIRST_REQ(gds->bcast_first_req);250251bc->bytecode[id++] = S_SQ_MEM_GDS_WORD2_DST_SEL_X(gds->dst_sel_x) |252S_SQ_MEM_GDS_WORD2_DST_SEL_Y(gds->dst_sel_y) |253S_SQ_MEM_GDS_WORD2_DST_SEL_Z(gds->dst_sel_z) |254S_SQ_MEM_GDS_WORD2_DST_SEL_W(gds->dst_sel_w);255return 0;256}257258int eg_bytecode_alu_build(struct r600_bytecode *bc, struct r600_bytecode_alu *alu, unsigned id)259{260if (alu->is_lds_idx_op) {261assert(!alu->src[0].abs && !alu->src[1].abs && !alu->src[2].abs);262assert(!alu->src[0].neg && !alu->src[1].neg && !alu->src[2].neg);263bc->bytecode[id++] = S_SQ_ALU_WORD0_SRC0_SEL(alu->src[0].sel) |264S_SQ_ALU_WORD0_SRC0_REL(alu->src[0].rel) |265S_SQ_ALU_WORD0_SRC0_CHAN(alu->src[0].chan) |266S_SQ_ALU_WORD0_LDS_IDX_OP_IDX_OFFSET_4(alu->lds_idx >> 4) |267S_SQ_ALU_WORD0_SRC1_SEL(alu->src[1].sel) |268S_SQ_ALU_WORD0_SRC1_REL(alu->src[1].rel) |269S_SQ_ALU_WORD0_SRC1_CHAN(alu->src[1].chan) |270S_SQ_ALU_WORD0_LDS_IDX_OP_IDX_OFFSET_5(alu->lds_idx >> 5) |271S_SQ_ALU_WORD0_INDEX_MODE(alu->index_mode) |272S_SQ_ALU_WORD0_PRED_SEL(alu->pred_sel) |273S_SQ_ALU_WORD0_LAST(alu->last);274} else {275bc->bytecode[id++] = S_SQ_ALU_WORD0_SRC0_SEL(alu->src[0].sel) |276S_SQ_ALU_WORD0_SRC0_REL(alu->src[0].rel) |277S_SQ_ALU_WORD0_SRC0_CHAN(alu->src[0].chan) |278S_SQ_ALU_WORD0_SRC0_NEG(alu->src[0].neg) |279S_SQ_ALU_WORD0_SRC1_SEL(alu->src[1].sel) |280S_SQ_ALU_WORD0_SRC1_REL(alu->src[1].rel) |281S_SQ_ALU_WORD0_SRC1_CHAN(alu->src[1].chan) |282S_SQ_ALU_WORD0_SRC1_NEG(alu->src[1].neg) |283S_SQ_ALU_WORD0_PRED_SEL(alu->pred_sel) |284S_SQ_ALU_WORD0_LAST(alu->last);285}286287/* don't replace gpr by pv or ps for destination register */288if (alu->is_lds_idx_op) {289unsigned lds_op = r600_isa_alu_opcode(bc->isa->hw_class, alu->op);290bc->bytecode[id++] =291S_SQ_ALU_WORD1_OP3_SRC2_SEL(alu->src[2].sel) |292S_SQ_ALU_WORD1_OP3_SRC2_REL(alu->src[2].rel) |293S_SQ_ALU_WORD1_OP3_SRC2_CHAN(alu->src[2].chan) |294S_SQ_ALU_WORD1_LDS_IDX_OP_IDX_OFFSET_1(alu->lds_idx >> 1) |295296S_SQ_ALU_WORD1_OP3_ALU_INST(lds_op & 0xff) |297S_SQ_ALU_WORD1_BANK_SWIZZLE(alu->bank_swizzle) |298S_SQ_ALU_WORD1_LDS_IDX_OP_LDS_OP((lds_op >> 8) & 0xff) |299S_SQ_ALU_WORD1_LDS_IDX_OP_IDX_OFFSET_0(alu->lds_idx) |300S_SQ_ALU_WORD1_LDS_IDX_OP_IDX_OFFSET_2(alu->lds_idx >> 2) |301S_SQ_ALU_WORD1_DST_CHAN(alu->dst.chan) |302S_SQ_ALU_WORD1_LDS_IDX_OP_IDX_OFFSET_3(alu->lds_idx >> 3);303304} else if (alu->is_op3) {305assert(!alu->src[0].abs && !alu->src[1].abs && !alu->src[2].abs);306bc->bytecode[id++] = S_SQ_ALU_WORD1_DST_GPR(alu->dst.sel) |307S_SQ_ALU_WORD1_DST_CHAN(alu->dst.chan) |308S_SQ_ALU_WORD1_DST_REL(alu->dst.rel) |309S_SQ_ALU_WORD1_CLAMP(alu->dst.clamp) |310S_SQ_ALU_WORD1_OP3_SRC2_SEL(alu->src[2].sel) |311S_SQ_ALU_WORD1_OP3_SRC2_REL(alu->src[2].rel) |312S_SQ_ALU_WORD1_OP3_SRC2_CHAN(alu->src[2].chan) |313S_SQ_ALU_WORD1_OP3_SRC2_NEG(alu->src[2].neg) |314S_SQ_ALU_WORD1_OP3_ALU_INST(r600_isa_alu_opcode(bc->isa->hw_class, alu->op)) |315S_SQ_ALU_WORD1_BANK_SWIZZLE(alu->bank_swizzle);316} else {317bc->bytecode[id++] = S_SQ_ALU_WORD1_DST_GPR(alu->dst.sel) |318S_SQ_ALU_WORD1_DST_CHAN(alu->dst.chan) |319S_SQ_ALU_WORD1_DST_REL(alu->dst.rel) |320S_SQ_ALU_WORD1_CLAMP(alu->dst.clamp) |321S_SQ_ALU_WORD1_OP2_SRC0_ABS(alu->src[0].abs) |322S_SQ_ALU_WORD1_OP2_SRC1_ABS(alu->src[1].abs) |323S_SQ_ALU_WORD1_OP2_WRITE_MASK(alu->dst.write) |324S_SQ_ALU_WORD1_OP2_OMOD(alu->omod) |325S_SQ_ALU_WORD1_OP2_ALU_INST(r600_isa_alu_opcode(bc->isa->hw_class, alu->op)) |326S_SQ_ALU_WORD1_BANK_SWIZZLE(alu->bank_swizzle) |327S_SQ_ALU_WORD1_OP2_UPDATE_EXECUTE_MASK(alu->execute_mask) |328S_SQ_ALU_WORD1_OP2_UPDATE_PRED(alu->update_pred);329}330return 0;331}332333334