diff options
Diffstat (limited to 'src/gallium/drivers/freedreno')
-rw-r--r-- | src/gallium/drivers/freedreno/a3xx/fd3_compiler.c | 21 | ||||
-rw-r--r-- | src/gallium/drivers/freedreno/a3xx/ir3.h | 3 | ||||
-rw-r--r-- | src/gallium/drivers/freedreno/a3xx/ir3_ra.c | 2 | ||||
-rw-r--r-- | src/gallium/drivers/freedreno/a3xx/ir3_sched.c | 32 |
4 files changed, 19 insertions, 39 deletions
diff --git a/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c b/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c index 1138ec9be34..3a388f8f3d7 100644 --- a/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c +++ b/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c @@ -43,31 +43,10 @@ #include "fd3_compiler.h" #include "fd3_program.h" -#include "fd3_util.h" #include "instr-a3xx.h" #include "ir3.h" -/* NOTE on half/full precision: - * Currently, the front end (ie. basically this file) does everything in - * full precision (with the exception of trans_arl() which doesn't work - * currently.. we reject anything with relative addressing and fallback - * to old compiler). - * - * In the RA step, if half_precision, it will assign the output to hr0.x - * but use full precision everywhere else. - * - * Eventually we'll need a better way to communicate type information - * to RA so that it can more properly assign both half and full precision - * registers. (And presumably double precision pairs for a4xx?) This - * would let us make more use of half precision registers, while still - * keeping things like tex coords in full precision registers. - * - * Since the RA is dealing with patching instruction types for half - * precision output, we can ignore that in the front end and just always - * create full precision instructions. - */ - struct fd3_compile_context { const struct tgsi_token *tokens; bool free_tokens; diff --git a/src/gallium/drivers/freedreno/a3xx/ir3.h b/src/gallium/drivers/freedreno/a3xx/ir3.h index e6f9a401366..0a8e53831c6 100644 --- a/src/gallium/drivers/freedreno/a3xx/ir3.h +++ b/src/gallium/drivers/freedreno/a3xx/ir3.h @@ -354,7 +354,7 @@ static inline bool is_meta(struct ir3_instruction *instr) return (instr->category == -1); } -static inline bool is_deref(struct ir3_instruction *instr) +static inline bool is_addr(struct ir3_instruction *instr) { return is_meta(instr) && (instr->opc == OPC_META_DEREF); } @@ -368,7 +368,6 @@ static inline bool writes_addr(struct ir3_instruction *instr) return false; } -/* TODO combine is_gpr()/reg_gpr().. */ static inline bool reg_gpr(struct ir3_register *r) { if (r->flags & (IR3_REG_CONST | IR3_REG_IMMED | IR3_REG_RELATIV | IR3_REG_SSA | IR3_REG_ADDR)) diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_ra.c b/src/gallium/drivers/freedreno/a3xx/ir3_ra.c index 67f92e670e9..ce9e50019a2 100644 --- a/src/gallium/drivers/freedreno/a3xx/ir3_ra.c +++ b/src/gallium/drivers/freedreno/a3xx/ir3_ra.c @@ -566,7 +566,7 @@ static void ir3_instr_ra(struct ir3_ra_ctx *ctx, return; /* allocate register(s): */ - if (is_deref(instr)) { + if (is_addr(instr)) { num = instr->regs[2]->num; } else if (reg_gpr(dst)) { struct ir3_ra_assignment a; diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_sched.c b/src/gallium/drivers/freedreno/a3xx/ir3_sched.c index 4fd3da58b46..ef84c54a601 100644 --- a/src/gallium/drivers/freedreno/a3xx/ir3_sched.c +++ b/src/gallium/drivers/freedreno/a3xx/ir3_sched.c @@ -52,7 +52,7 @@ enum { struct ir3_sched_ctx { struct ir3_instruction *scheduled; /* last scheduled instr */ - struct ir3_instruction *deref; /* current deref, if any */ + struct ir3_instruction *addr; /* current a0.x user, if any */ unsigned cnt; }; @@ -130,8 +130,8 @@ static void schedule(struct ir3_sched_ctx *ctx, } if (writes_addr(instr)) { - assert(ctx->deref == NULL); - ctx->deref = instr; + assert(ctx->addr == NULL); + ctx->addr = instr; } instr->flags |= IR3_INSTR_MARK; @@ -227,8 +227,8 @@ static int trysched(struct ir3_sched_ctx *ctx, /* if this is a write to address register, and addr register * is currently in use, we need to defer until it is free: */ - if (writes_addr(instr) && ctx->deref) { - assert(ctx->deref != instr); + if (writes_addr(instr) && ctx->addr) { + assert(ctx->addr != instr); return DELAYED; } @@ -248,17 +248,17 @@ static struct ir3_instruction * reverse(struct ir3_instruction *instr) return reversed; } -static bool uses_current_deref(struct ir3_sched_ctx *ctx, +static bool uses_current_addr(struct ir3_sched_ctx *ctx, struct ir3_instruction *instr) { unsigned i; for (i = 1; i < instr->regs_count; i++) { struct ir3_register *reg = instr->regs[i]; if (reg->flags & IR3_REG_SSA) { - if (is_deref(reg->instr)) { - struct ir3_instruction *deref; - deref = reg->instr->regs[1]->instr; /* the mova */ - if (ctx->deref == deref) + if (is_addr(reg->instr)) { + struct ir3_instruction *addr; + addr = reg->instr->regs[1]->instr; /* the mova */ + if (ctx->addr == addr) return true; } } @@ -274,26 +274,28 @@ static int block_sched_undelayed(struct ir3_sched_ctx *ctx, struct ir3_block *block) { struct ir3_instruction *instr = block->head; - bool in_use = false; + bool addr_in_use = false; unsigned cnt = ~0; while (instr) { struct ir3_instruction *next = instr->next; + bool addr = uses_current_addr(ctx, instr); - if (uses_current_deref(ctx, instr)) { + if (addr) { int ret = trysched(ctx, instr); if (ret == SCHEDULED) cnt = 0; else if (ret > 0) cnt = MIN2(cnt, ret); - in_use = true; + if (addr) + addr_in_use = true; } instr = next; } - if (!in_use) - ctx->deref = NULL; + if (!addr_in_use) + ctx->addr = NULL; return cnt; } |