From 897f6d3c7d06316b0535971cc2de318157c23692 Mon Sep 17 00:00:00 2001 From: Kenneth Graunke Date: Thu, 14 Oct 2010 11:40:19 -0700 Subject: i965: Correctly emit the RNDZ instruction. Simply using RNDU, RNDZ, or RNDE does not produce the desired result. Rather, the RND* instructions place a value in the destination register that may be 1 less than the correct answer. They can also set per-channel "increment bits" in a flag register, which, if set, mean dest needs to be incremented by 1. A second instruction - a predicated add - completes the job. Notably, RNDD always produces the correct answer in a single instruction. Fixes piglit test glsl-fs-trunc. --- src/mesa/drivers/dri/i965/brw_eu.h | 9 +++++++-- src/mesa/drivers/dri/i965/brw_eu_emit.c | 25 ++++++++++++++++++++++++- 2 files changed, 31 insertions(+), 3 deletions(-) (limited to 'src') diff --git a/src/mesa/drivers/dri/i965/brw_eu.h b/src/mesa/drivers/dri/i965/brw_eu.h index c0deb238c2..c16613e3df 100644 --- a/src/mesa/drivers/dri/i965/brw_eu.h +++ b/src/mesa/drivers/dri/i965/brw_eu.h @@ -789,6 +789,10 @@ struct brw_instruction *brw_##OP(struct brw_compile *p, \ struct brw_reg src0, \ struct brw_reg src1); +#define ROUND(OP) \ +void brw_##OP(struct brw_compile *p, struct brw_reg dest, struct brw_reg src0); + + ALU1(MOV) ALU2(SEL) ALU1(NOT) @@ -805,7 +809,6 @@ ALU2(ADD) ALU2(MUL) ALU1(FRC) ALU1(RNDD) -ALU1(RNDZ) ALU2(MAC) ALU2(MACH) ALU1(LZD) @@ -816,9 +819,11 @@ ALU2(DP2) ALU2(LINE) ALU2(PLN) +ROUND(RNDZ) + #undef ALU1 #undef ALU2 - +#undef ROUND /* Helpers for SEND instruction: diff --git a/src/mesa/drivers/dri/i965/brw_eu_emit.c b/src/mesa/drivers/dri/i965/brw_eu_emit.c index 41286382d0..68c6c47763 100644 --- a/src/mesa/drivers/dri/i965/brw_eu_emit.c +++ b/src/mesa/drivers/dri/i965/brw_eu_emit.c @@ -654,6 +654,26 @@ struct brw_instruction *brw_##OP(struct brw_compile *p, \ return brw_alu2(p, BRW_OPCODE_##OP, dest, src0, src1); \ } +/* Rounding operations (other than RNDD) require two instructions - the first + * stores a rounded value (possibly the wrong way) in the dest register, but + * also sets a per-channel "increment bit" in the flag register. A predicated + * add of 1.0 fixes dest to contain the desired result. + */ +#define ROUND(OP) \ +void brw_##OP(struct brw_compile *p, \ + struct brw_reg dest, \ + struct brw_reg src) \ +{ \ + struct brw_instruction *rnd, *add; \ + rnd = next_insn(p, BRW_OPCODE_##OP); \ + brw_set_dest(rnd, dest); \ + brw_set_src0(rnd, src); \ + rnd->header.destreg__conditionalmod = 0x7; /* turn on round-increments */ \ + \ + add = brw_ADD(p, dest, dest, brw_imm_f(1.0f)); \ + add->header.predicate_control = BRW_PREDICATE_NORMAL; \ +} + ALU1(MOV) ALU2(SEL) @@ -668,7 +688,6 @@ ALU2(RSL) ALU2(ASR) ALU1(FRC) ALU1(RNDD) -ALU1(RNDZ) ALU2(MAC) ALU2(MACH) ALU1(LZD) @@ -679,6 +698,10 @@ ALU2(DP2) ALU2(LINE) ALU2(PLN) + +ROUND(RNDZ) + + struct brw_instruction *brw_ADD(struct brw_compile *p, struct brw_reg dest, struct brw_reg src0, -- cgit v1.2.3