diff options
Diffstat (limited to 'src/gallium')
| -rw-r--r-- | src/gallium/drivers/nv50/nv50_program.c | 77 |
1 files changed, 75 insertions, 2 deletions
diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c index 2ab2ac35c2a..8e66fdca49b 100644 --- a/src/gallium/drivers/nv50/nv50_program.c +++ b/src/gallium/drivers/nv50/nv50_program.c @@ -90,8 +90,9 @@ struct nv50_reg { int acc; /* instruction where this reg is last read (first insn == 1) */ }; -/* arbitrary limit */ +/* arbitrary limits */ #define MAX_IF_DEPTH 4 +#define MAX_LOOP_DEPTH 4 struct nv50_pc { struct nv50_program *p; @@ -127,7 +128,9 @@ struct nv50_pc { struct nv50_program_exec *if_cond; struct nv50_program_exec *if_insn[MAX_IF_DEPTH]; struct nv50_program_exec *br_join[MAX_IF_DEPTH]; - int if_lvl; + struct nv50_program_exec *br_loop[MAX_LOOP_DEPTH]; /* for BRK branch */ + int if_lvl, loop_lvl; + unsigned loop_pos[MAX_LOOP_DEPTH]; /* current instruction and total number of insns */ unsigned insn_cur; @@ -204,6 +207,10 @@ alloc_reg(struct nv50_pc *pc, struct nv50_reg *reg) assert(0); } +/* XXX: For shaders that aren't executed linearly (e.g. shaders that + * contain loops), we need to assign all hw regs to TGSI TEMPs early, + * lest we risk temp_temps overwriting regs alloc'd "later". + */ static struct nv50_reg * alloc_temp(struct nv50_pc *pc, struct nv50_reg *dst) { @@ -1485,6 +1492,55 @@ nv50_tgsi_dst_revdep(unsigned op, int s, int c) } } +static INLINE boolean +has_pred(struct nv50_program_exec *e, unsigned cc) +{ + if (!is_long(e) || is_immd(e)) + return FALSE; + return ((e->inst[1] & 0x780) == (cc << 7)); +} + +/* on ENDIF see if we can do "@p0.neu single_op" instead of: + * join_at ENDIF + * @p0.eq bra ENDIF + * single_op + * ENDIF: nop.join + */ +static boolean +nv50_kill_branch(struct nv50_pc *pc) +{ + int lvl = pc->if_lvl; + + if (pc->if_insn[lvl]->next != pc->p->exec_tail) + return FALSE; + + /* if ccode == 'true', the BRA is from an ELSE and the predicate + * reg may no longer be valid, since we currently always use $p0 + */ + if (has_pred(pc->if_insn[lvl], 0xf)) + return FALSE; + assert(pc->if_insn[lvl] && pc->br_join[lvl]); + + /* We'll use the exec allocated for JOIN_AT (as we can't easily + * update prev's next); if exec_tail is BRK, update the pointer. + */ + if (pc->loop_lvl && pc->br_loop[pc->loop_lvl - 1] == pc->p->exec_tail) + pc->br_loop[pc->loop_lvl - 1] = pc->br_join[lvl]; + + pc->p->exec_size -= 4; /* remove JOIN_AT and BRA */ + + *pc->br_join[lvl] = *pc->p->exec_tail; + + FREE(pc->if_insn[lvl]); + FREE(pc->p->exec_tail); + + pc->p->exec_tail = pc->br_join[lvl]; + pc->p->exec_tail->next = NULL; + set_pred(pc, 0xd, 0, pc->p->exec_tail); + + return TRUE; +} + static boolean nv50_program_tx_insn(struct nv50_pc *pc, const struct tgsi_full_instruction *inst) @@ -1554,6 +1610,14 @@ nv50_program_tx_insn(struct nv50_pc *pc, emit_add(pc, dst[c], src[0][c], src[1][c]); } break; + case TGSI_OPCODE_BGNLOOP: + pc->loop_pos[pc->loop_lvl++] = pc->p->exec_size; + break; + case TGSI_OPCODE_BRK: + emit_branch(pc, -1, 0, NULL); + assert(pc->loop_lvl > 0); + pc->br_loop[pc->loop_lvl - 1] = pc->p->exec_tail; + break; case TGSI_OPCODE_CEIL: for (c = 0; c < 4; c++) { if (!(mask & (1 << c))) @@ -1609,6 +1673,10 @@ nv50_program_tx_insn(struct nv50_pc *pc, case TGSI_OPCODE_ENDIF: pc->if_insn[--pc->if_lvl]->param.index = pc->p->exec_size; + /* try to replace branch over 1 insn with a predicated insn */ + if (nv50_kill_branch(pc) == TRUE) + break; + if (pc->br_join[pc->if_lvl]) { pc->br_join[pc->if_lvl]->param.index = pc->p->exec_size; pc->br_join[pc->if_lvl] = NULL; @@ -1619,6 +1687,11 @@ nv50_program_tx_insn(struct nv50_pc *pc, emit_nop(pc); pc->p->exec_tail->inst[1] |= 2; break; + case TGSI_OPCODE_ENDLOOP: + emit_branch(pc, -1, 0, NULL); + pc->p->exec_tail->param.index = pc->loop_pos[--pc->loop_lvl]; + pc->br_loop[pc->loop_lvl]->param.index = pc->p->exec_size; + break; case TGSI_OPCODE_EX2: emit_preex2(pc, temp, src[0][0]); emit_flop(pc, 6, brdc, temp); |