diff options
Diffstat (limited to 'src/mesa/drivers/dri/i965/brw_vs_emit.c')
-rw-r--r-- | src/mesa/drivers/dri/i965/brw_vs_emit.c | 666 |
1 files changed, 531 insertions, 135 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_vs_emit.c b/src/mesa/drivers/dri/i965/brw_vs_emit.c index 174331a765..1638ef8111 100644 --- a/src/mesa/drivers/dri/i965/brw_vs_emit.c +++ b/src/mesa/drivers/dri/i965/brw_vs_emit.c @@ -38,18 +38,55 @@ #include "brw_vs.h" +static struct brw_reg get_tmp( struct brw_vs_compile *c ) +{ + struct brw_reg tmp = brw_vec8_grf(c->last_tmp, 0); + + if (++c->last_tmp > c->prog_data.total_grf) + c->prog_data.total_grf = c->last_tmp; + + return tmp; +} + +static void release_tmp( struct brw_vs_compile *c, struct brw_reg tmp ) +{ + if (tmp.nr == c->last_tmp-1) + c->last_tmp--; +} + +static void release_tmps( struct brw_vs_compile *c ) +{ + c->last_tmp = c->first_tmp; +} + -/* Do things as simply as possible. Allocate and populate all regs +/** + * Preallocate GRF register before code emit. + * Do things as simply as possible. Allocate and populate all regs * ahead of time. */ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) { GLuint i, reg = 0, mrf; - GLuint nr_params; + int attributes_in_vue; + + /* Determine whether to use a real constant buffer or use a block + * of GRF registers for constants. The later is faster but only + * works if everything fits in the GRF. + * XXX this heuristic/check may need some fine tuning... + */ + if (c->vp->program.Base.Parameters->NumParameters + + c->vp->program.Base.NumTemporaries + 20 > BRW_MAX_GRF) + c->vp->use_const_buffer = GL_TRUE; + else + c->vp->use_const_buffer = GL_FALSE; + + /*printf("use_const_buffer = %d\n", c->vp->use_const_buffer);*/ /* r0 -- reserved as usual */ - c->r0 = brw_vec8_grf(reg, 0); reg++; + c->r0 = brw_vec8_grf(reg, 0); + reg++; /* User clip planes from curbe: */ @@ -60,39 +97,59 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) /* Deal with curbe alignment: */ - reg += ((6+c->key.nr_userclip+3)/4)*2; + reg += ((6 + c->key.nr_userclip + 3) / 4) * 2; } /* Vertex program parameters from curbe: */ - nr_params = c->vp->program.Base.Parameters->NumParameters; - for (i = 0; i < nr_params; i++) { - c->regs[PROGRAM_STATE_VAR][i] = stride( brw_vec4_grf(reg+i/2, (i%2) * 4), 0, 4, 1); - } - reg += (nr_params+1)/2; + if (c->vp->use_const_buffer) { + /* get constants from a real constant buffer */ + c->prog_data.curb_read_length = 0; + c->prog_data.nr_params = 4; /* XXX 0 causes a bug elsewhere... */ + } + else { + /* use a section of the GRF for constants */ + GLuint nr_params = c->vp->program.Base.Parameters->NumParameters; + for (i = 0; i < nr_params; i++) { + c->regs[PROGRAM_STATE_VAR][i] = stride( brw_vec4_grf(reg+i/2, (i%2) * 4), 0, 4, 1); + } + reg += (nr_params + 1) / 2; + c->prog_data.curb_read_length = reg - 1; - c->prog_data.curb_read_length = reg - 1; + c->prog_data.nr_params = nr_params * 4; + } /* Allocate input regs: */ c->nr_inputs = 0; for (i = 0; i < VERT_ATTRIB_MAX; i++) { - if (c->prog_data.inputs_read & (1<<i)) { + if (c->prog_data.inputs_read & (1 << i)) { c->nr_inputs++; c->regs[PROGRAM_INPUT][i] = brw_vec8_grf(reg, 0); reg++; } } + /* If there are no inputs, we'll still be reading one attribute's worth + * because it's required -- see urb_read_length setting. + */ + if (c->nr_inputs == 0) + reg++; - /* Allocate outputs: TODO: could organize the non-position outputs - * to go straight into message regs. + /* Allocate outputs. The non-position outputs go straight into message regs. */ c->nr_outputs = 0; c->first_output = reg; - mrf = 4; + c->first_overflow_output = 0; + + if (BRW_IS_IGDNG(c->func.brw)) + mrf = 8; + else + mrf = 4; + for (i = 0; i < VERT_RESULT_MAX; i++) { - if (c->prog_data.outputs_written & (1<<i)) { + if (c->prog_data.outputs_written & (1 << i)) { c->nr_outputs++; + assert(i < Elements(c->regs[PROGRAM_OUTPUT])); if (i == VERT_RESULT_HPOS) { c->regs[PROGRAM_OUTPUT][i] = brw_vec8_grf(reg, 0); reg++; @@ -103,8 +160,17 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) mrf++; /* just a placeholder? XXX fix later stages & remove this */ } else { - c->regs[PROGRAM_OUTPUT][i] = brw_message_reg(mrf); - mrf++; + if (mrf < 16) { + c->regs[PROGRAM_OUTPUT][i] = brw_message_reg(mrf); + mrf++; + } + else { + /* too many vertex results to fit in MRF, use GRF for overflow */ + if (!c->first_overflow_output) + c->first_overflow_output = i; + c->regs[PROGRAM_OUTPUT][i] = brw_vec8_grf(reg, 0); + reg++; + } } } } @@ -132,17 +198,24 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) reg++; } + if (c->vp->use_const_buffer) { + for (i = 0; i < 3; i++) { + c->current_const[i].index = -1; + c->current_const[i].reg = brw_vec8_grf(reg, 0); + reg++; + } + } + for (i = 0; i < 128; i++) { - if (c->output_regs[i].used_in_src) { - c->output_regs[i].reg = brw_vec8_grf(reg, 0); - reg++; - } + if (c->output_regs[i].used_in_src) { + c->output_regs[i].reg = brw_vec8_grf(reg, 0); + reg++; + } } c->stack = brw_uw16_reg(BRW_GENERAL_REGISTER_FILE, reg, 0); reg += 2; - - + /* Some opcodes need an internal temporary: */ c->first_tmp = reg; @@ -152,35 +225,38 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) * urb_read_length is the number of registers read from *each* * vertex urb, so is half the amount: */ - c->prog_data.urb_read_length = (c->nr_inputs+1)/2; - - c->prog_data.urb_entry_size = (c->nr_outputs+2+3)/4; - c->prog_data.total_grf = reg; -} - + c->prog_data.urb_read_length = (c->nr_inputs + 1) / 2; + /* Setting this field to 0 leads to undefined behavior according to the + * the VS_STATE docs. Our VUEs will always have at least one attribute + * sitting in them, even if it's padding. + */ + if (c->prog_data.urb_read_length == 0) + c->prog_data.urb_read_length = 1; -static struct brw_reg get_tmp( struct brw_vs_compile *c ) -{ - struct brw_reg tmp = brw_vec8_grf(c->last_tmp, 0); + /* The VS VUEs are shared by VF (outputting our inputs) and VS, so size + * them to fit the biggest thing they need to. + */ + attributes_in_vue = MAX2(c->nr_outputs, c->nr_inputs); - if (++c->last_tmp > c->prog_data.total_grf) - c->prog_data.total_grf = c->last_tmp; + if (BRW_IS_IGDNG(c->func.brw)) + c->prog_data.urb_entry_size = (attributes_in_vue + 6 + 3) / 4; + else + c->prog_data.urb_entry_size = (attributes_in_vue + 2 + 3) / 4; - return tmp; -} + c->prog_data.total_grf = reg; -static void release_tmp( struct brw_vs_compile *c, struct brw_reg tmp ) -{ - if (tmp.nr == c->last_tmp-1) - c->last_tmp--; -} - -static void release_tmps( struct brw_vs_compile *c ) -{ - c->last_tmp = c->first_tmp; + if (INTEL_DEBUG & DEBUG_VS) { + _mesa_printf("%s NumAddrRegs %d\n", __FUNCTION__, c->vp->program.Base.NumAddressRegs); + _mesa_printf("%s NumTemps %d\n", __FUNCTION__, c->vp->program.Base.NumTemporaries); + _mesa_printf("%s reg = %d\n", __FUNCTION__, reg); + } } +/** + * If an instruction uses a temp reg both as a src and the dest, we + * sometimes need to allocate an intermediate temporary. + */ static void unalias1( struct brw_vs_compile *c, struct brw_reg dst, struct brw_reg arg0, @@ -200,6 +276,10 @@ static void unalias1( struct brw_vs_compile *c, } } +/** + * \sa unalias2 + * Checkes if 2-operand instruction needs an intermediate temporary. + */ static void unalias2( struct brw_vs_compile *c, struct brw_reg dst, struct brw_reg arg0, @@ -222,6 +302,10 @@ static void unalias2( struct brw_vs_compile *c, } } +/** + * \sa unalias2 + * Checkes if 3-operand instruction needs an intermediate temporary. + */ static void unalias3( struct brw_vs_compile *c, struct brw_reg dst, struct brw_reg arg0, @@ -615,6 +699,8 @@ static void emit_lit_noalias( struct brw_vs_compile *c, } brw_ENDIF(p, if_insn); + + release_tmp(c, tmp); } static void emit_lrp_noalias(struct brw_vs_compile *c, @@ -655,13 +741,84 @@ static void emit_nrm( struct brw_vs_compile *c, } +static struct brw_reg +get_constant(struct brw_vs_compile *c, + const struct prog_instruction *inst, + GLuint argIndex) +{ + const struct prog_src_register *src = &inst->SrcReg[argIndex]; + struct brw_compile *p = &c->func; + struct brw_reg const_reg; + struct brw_reg const2_reg; + const GLboolean relAddr = src->RelAddr; + + assert(argIndex < 3); + + if (c->current_const[argIndex].index != src->Index || relAddr) { + struct brw_reg addrReg = c->regs[PROGRAM_ADDRESS][0]; + + c->current_const[argIndex].index = src->Index; + +#if 0 + printf(" fetch const[%d] for arg %d into reg %d\n", + src->Index, argIndex, c->current_const[argIndex].reg.nr); +#endif + /* need to fetch the constant now */ + brw_dp_READ_4_vs(p, + c->current_const[argIndex].reg,/* writeback dest */ + 0, /* oword */ + relAddr, /* relative indexing? */ + addrReg, /* address register */ + 16 * src->Index, /* byte offset */ + SURF_INDEX_VERT_CONST_BUFFER /* binding table index */ + ); + + if (relAddr) { + /* second read */ + const2_reg = get_tmp(c); + + /* use upper half of address reg for second read */ + addrReg = stride(addrReg, 0, 4, 0); + addrReg.subnr = 16; + + brw_dp_READ_4_vs(p, + const2_reg, /* writeback dest */ + 1, /* oword */ + relAddr, /* relative indexing? */ + addrReg, /* address register */ + 16 * src->Index, /* byte offset */ + SURF_INDEX_VERT_CONST_BUFFER + ); + } + } + + const_reg = c->current_const[argIndex].reg; + + if (relAddr) { + /* merge the two Owords into the constant register */ + /* const_reg[7..4] = const2_reg[7..4] */ + brw_MOV(p, + suboffset(stride(const_reg, 0, 4, 1), 4), + suboffset(stride(const2_reg, 0, 4, 1), 4)); + release_tmp(c, const2_reg); + } + else { + /* replicate lower four floats into upper half (to get XYZWXYZW) */ + const_reg = stride(const_reg, 0, 4, 0); + const_reg.subnr = 0; + } + + return const_reg; +} + + + /* TODO: relative addressing! */ static struct brw_reg get_reg( struct brw_vs_compile *c, - GLuint file, + gl_register_file file, GLuint index ) { - switch (file) { case PROGRAM_TEMPORARY: case PROGRAM_INPUT: @@ -690,13 +847,17 @@ static struct brw_reg get_reg( struct brw_vs_compile *c, } +/** + * Indirect addressing: get reg[[arg] + offset]. + */ static struct brw_reg deref( struct brw_vs_compile *c, struct brw_reg arg, GLint offset) { struct brw_compile *p = &c->func; struct brw_reg tmp = vec4(get_tmp(c)); - struct brw_reg vp_address = retype(vec1(get_reg(c, PROGRAM_ADDRESS, 0)), BRW_REGISTER_TYPE_UW); + struct brw_reg addr_reg = c->regs[PROGRAM_ADDRESS][0]; + struct brw_reg vp_address = retype(vec1(addr_reg), BRW_REGISTER_TYPE_UW); GLuint byte_offset = arg.nr * 32 + arg.subnr + offset * 16; struct brw_reg indirect = brw_vec4_indirect(0,0); @@ -717,10 +878,67 @@ static struct brw_reg deref( struct brw_vs_compile *c, brw_pop_insn_state(p); } + /* NOTE: tmp not released */ return vec8(tmp); } +/** + * Get brw reg corresponding to the instruction's [argIndex] src reg. + * TODO: relative addressing! + */ +static struct brw_reg +get_src_reg( struct brw_vs_compile *c, + const struct prog_instruction *inst, + GLuint argIndex ) +{ + const GLuint file = inst->SrcReg[argIndex].File; + const GLint index = inst->SrcReg[argIndex].Index; + const GLboolean relAddr = inst->SrcReg[argIndex].RelAddr; + + switch (file) { + case PROGRAM_TEMPORARY: + case PROGRAM_INPUT: + case PROGRAM_OUTPUT: + if (relAddr) { + return deref(c, c->regs[file][0], index); + } + else { + assert(c->regs[file][index].nr != 0); + return c->regs[file][index]; + } + + case PROGRAM_STATE_VAR: + case PROGRAM_CONSTANT: + case PROGRAM_UNIFORM: + case PROGRAM_ENV_PARAM: + if (c->vp->use_const_buffer) { + return get_constant(c, inst, argIndex); + } + else if (relAddr) { + return deref(c, c->regs[PROGRAM_STATE_VAR][0], index); + } + else { + assert(c->regs[PROGRAM_STATE_VAR][index].nr != 0); + return c->regs[PROGRAM_STATE_VAR][index]; + } + case PROGRAM_ADDRESS: + assert(index == 0); + return c->regs[file][index]; + + case PROGRAM_UNDEFINED: + /* this is a normal case since we loop over all three src args */ + return brw_null_reg(); + + case PROGRAM_LOCAL_PARAM: + case PROGRAM_WRITE_ONLY: + default: + assert(0); + return brw_null_reg(); + } +} + + static void emit_arl( struct brw_vs_compile *c, struct brw_reg dst, struct brw_reg arg0 ) @@ -732,30 +950,31 @@ static void emit_arl( struct brw_vs_compile *c, if (need_tmp) tmp = get_tmp(c); - brw_RNDD(p, tmp, arg0); - brw_MUL(p, dst, tmp, brw_imm_d(16)); + brw_RNDD(p, tmp, arg0); /* tmp = round(arg0) */ + brw_MUL(p, dst, tmp, brw_imm_d(16)); /* dst = tmp * 16 */ if (need_tmp) release_tmp(c, tmp); } -/* Will return mangled results for SWZ op. The emit_swz() function +/** + * Return the brw reg for the given instruction's src argument. + * Will return mangled results for SWZ op. The emit_swz() function * ignores this result and recalculates taking extended swizzles into * account. */ static struct brw_reg get_arg( struct brw_vs_compile *c, - struct prog_src_register *src ) + const struct prog_instruction *inst, + GLuint argIndex ) { + const struct prog_src_register *src = &inst->SrcReg[argIndex]; struct brw_reg reg; if (src->File == PROGRAM_UNDEFINED) return brw_null_reg(); - if (src->RelAddr) - reg = deref(c, c->regs[PROGRAM_STATE_VAR][0], src->Index); - else - reg = get_reg(c, src->File, src->Index); + reg = get_src_reg(c, inst, argIndex); /* Convert 3-bit swizzle to 2-bit. */ @@ -766,16 +985,38 @@ static struct brw_reg get_arg( struct brw_vs_compile *c, /* Note this is ok for non-swizzle instructions: */ - reg.negate = src->NegateBase ? 1 : 0; + reg.negate = src->Negate ? 1 : 0; return reg; } +/** + * Get brw register for the given program dest register. + */ static struct brw_reg get_dst( struct brw_vs_compile *c, struct prog_dst_register dst ) { - struct brw_reg reg = get_reg(c, dst.File, dst.Index); + struct brw_reg reg; + + switch (dst.File) { + case PROGRAM_TEMPORARY: + case PROGRAM_OUTPUT: + assert(c->regs[dst.File][dst.Index].nr != 0); + reg = c->regs[dst.File][dst.Index]; + break; + case PROGRAM_ADDRESS: + assert(dst.Index == 0); + reg = c->regs[dst.File][dst.Index]; + break; + case PROGRAM_UNDEFINED: + /* we may hit this for OPCODE_END, OPCODE_KIL, etc */ + reg = brw_null_reg(); + break; + default: + assert(0); + reg = brw_null_reg(); + } reg.dw1.bits.writemask = dst.WriteMask; @@ -785,14 +1026,16 @@ static struct brw_reg get_dst( struct brw_vs_compile *c, static void emit_swz( struct brw_vs_compile *c, struct brw_reg dst, - struct prog_src_register src ) + const struct prog_instruction *inst) { + const GLuint argIndex = 0; + const struct prog_src_register src = inst->SrcReg[argIndex]; struct brw_compile *p = &c->func; GLuint zeros_mask = 0; GLuint ones_mask = 0; GLuint src_mask = 0; GLubyte src_swz[4]; - GLboolean need_tmp = (src.NegateBase && + GLboolean need_tmp = (src.Negate && dst.file != BRW_GENERAL_REGISTER_FILE); struct brw_reg tmp = dst; GLuint i; @@ -826,10 +1069,7 @@ static void emit_swz( struct brw_vs_compile *c, if (src_mask) { struct brw_reg arg0; - if (src.RelAddr) - arg0 = deref(c, c->regs[PROGRAM_STATE_VAR][0], src.Index); - else - arg0 = get_reg(c, src.File, src.Index); + arg0 = get_src_reg(c, inst, argIndex); arg0 = brw_swizzle(arg0, src_swz[0], src_swz[1], @@ -844,8 +1084,8 @@ static void emit_swz( struct brw_vs_compile *c, if (ones_mask) brw_MOV(p, brw_writemask(tmp, ones_mask), brw_imm_f(1)); - if (src.NegateBase) - brw_MOV(p, brw_writemask(tmp, src.NegateBase), negate(tmp)); + if (src.Negate) + brw_MOV(p, brw_writemask(tmp, src.Negate), negate(tmp)); if (need_tmp) { brw_MOV(p, dst, tmp); @@ -863,6 +1103,8 @@ static void emit_vertex_write( struct brw_vs_compile *c) struct brw_reg m0 = brw_message_reg(0); struct brw_reg pos = c->regs[PROGRAM_OUTPUT][VERT_RESULT_HPOS]; struct brw_reg ndc; + int eot; + GLuint len_vertext_header = 2; if (c->key.copy_edgeflag) { brw_MOV(p, @@ -871,21 +1113,17 @@ static void emit_vertex_write( struct brw_vs_compile *c) } /* Build ndc coords */ - if (!c->key.know_w_is_one) { - ndc = get_tmp(c); - emit_math1(c, BRW_MATH_FUNCTION_INV, ndc, brw_swizzle1(pos, 3), BRW_MATH_PRECISION_FULL); - brw_MUL(p, brw_writemask(ndc, WRITEMASK_XYZ), pos, ndc); - } - else { - ndc = pos; - } + ndc = get_tmp(c); + /* ndc = 1.0 / pos.w */ + emit_math1(c, BRW_MATH_FUNCTION_INV, ndc, brw_swizzle1(pos, 3), BRW_MATH_PRECISION_FULL); + /* ndc.xyz = pos * ndc */ + brw_MUL(p, brw_writemask(ndc, WRITEMASK_XYZ), pos, ndc); /* Update the header for point size, user clipping flags, and -ve rhw * workaround. */ if ((c->prog_data.outputs_written & (1<<VERT_RESULT_PSIZ)) || - c->key.nr_userclip || - (!BRW_IS_G4X(p->brw) && !c->key.know_w_is_one)) + c->key.nr_userclip || BRW_IS_965(p->brw)) { struct brw_reg header1 = retype(get_tmp(c), BRW_REGISTER_TYPE_UD); GLuint i; @@ -916,7 +1154,7 @@ static void emit_vertex_write( struct brw_vs_compile *c) * Later, clipping will detect ucp[6] and ensure the primitive is * clipped against all fixed planes. */ - if (!BRW_IS_G4X(p->brw) && !c->key.know_w_is_one) { + if (BRW_IS_965(p->brw)) { brw_CMP(p, vec8(brw_null_reg()), BRW_CONDITIONAL_L, @@ -943,7 +1181,23 @@ static void emit_vertex_write( struct brw_vs_compile *c) */ brw_set_access_mode(p, BRW_ALIGN_1); brw_MOV(p, offset(m0, 2), ndc); - brw_MOV(p, offset(m0, 3), pos); + + if (BRW_IS_IGDNG(p->brw)) { + /* There are 20 DWs (D0-D19) in VUE vertex header on IGDNG */ + brw_MOV(p, offset(m0, 3), pos); /* a portion of vertex header */ + /* m4, m5 contain the distances from vertex to the user clip planeXXX. + * Seems it is useless for us. + * m6 is used for aligning, so that the remainder of vertex element is + * reg-aligned. + */ + brw_MOV(p, offset(m0, 7), pos); /* the remainder of vertex element */ + len_vertext_header = 6; + } else { + brw_MOV(p, offset(m0, 3), pos); + len_vertext_header = 2; + } + + eot = (c->first_overflow_output == 0); brw_urb_WRITE(p, brw_null_reg(), /* dest */ @@ -951,62 +1205,126 @@ static void emit_vertex_write( struct brw_vs_compile *c) c->r0, /* src */ 0, /* allocate */ 1, /* used */ - c->nr_outputs + 3, /* msg len */ + MIN2(c->nr_outputs + 1 + len_vertext_header, (BRW_MAX_MRF-1)), /* msg len */ 0, /* response len */ - 1, /* eot */ - 1, /* writes complete */ + eot, /* eot */ + eot, /* writes complete */ 0, /* urb destination offset */ BRW_URB_SWIZZLE_INTERLEAVE); + + if (c->first_overflow_output > 0) { + /* Not all of the vertex outputs/results fit into the MRF. + * Move the overflowed attributes from the GRF to the MRF and + * issue another brw_urb_WRITE(). + */ + /* XXX I'm not 100% sure about which MRF regs to use here. Starting + * at mrf[4] atm... + */ + GLuint i, mrf = 0; + for (i = c->first_overflow_output; i < VERT_RESULT_MAX; i++) { + if (c->prog_data.outputs_written & (1 << i)) { + /* move from GRF to MRF */ + brw_MOV(p, brw_message_reg(4+mrf), c->regs[PROGRAM_OUTPUT][i]); + mrf++; + } + } + + brw_urb_WRITE(p, + brw_null_reg(), /* dest */ + 4, /* starting mrf reg nr */ + c->r0, /* src */ + 0, /* allocate */ + 1, /* used */ + mrf+1, /* msg len */ + 0, /* response len */ + 1, /* eot */ + 1, /* writes complete */ + BRW_MAX_MRF-1, /* urb destination offset */ + BRW_URB_SWIZZLE_INTERLEAVE); + } } +/** + * Called after code generation to resolve subroutine calls and the + * END instruction. + * \param end_inst points to brw code for END instruction + * \param last_inst points to last instruction emitted before vertex write + */ static void -post_vs_emit( struct brw_vs_compile *c, struct brw_instruction *end_inst ) +post_vs_emit( struct brw_vs_compile *c, + struct brw_instruction *end_inst, + struct brw_instruction *last_inst ) { - GLuint nr_insns = c->vp->program.Base.NumInstructions; - GLuint insn, target_insn; - struct prog_instruction *inst1, *inst2; - struct brw_instruction *brw_inst1, *brw_inst2; - int offset; - for (insn = 0; insn < nr_insns; insn++) { - inst1 = &c->vp->program.Base.Instructions[insn]; - brw_inst1 = inst1->Data; - switch (inst1->Opcode) { - case OPCODE_CAL: - case OPCODE_BRA: - target_insn = inst1->BranchTarget; - inst2 = &c->vp->program.Base.Instructions[target_insn]; - brw_inst2 = inst2->Data; - offset = brw_inst2 - brw_inst1; - brw_set_src1(brw_inst1, brw_imm_d(offset*16)); - break; - case OPCODE_END: - offset = end_inst - brw_inst1; - brw_set_src1(brw_inst1, brw_imm_d(offset*16)); - break; - default: - break; - } + GLint offset; + + brw_resolve_cals(&c->func); + + /* patch up the END code to jump past subroutines, etc */ + offset = last_inst - end_inst; + if (offset > 1) { + brw_set_src1(end_inst, brw_imm_d(offset * 16)); + } else { + end_inst->header.opcode = BRW_OPCODE_NOP; + } +} + +static uint32_t +get_predicate(const struct prog_instruction *inst) +{ + if (inst->DstReg.CondMask == COND_TR) + return BRW_PREDICATE_NONE; + + /* All of GLSL only produces predicates for COND_NE and one channel per + * vector. Fail badly if someone starts doing something else, as it might + * mean infinite looping or something. + * + * We'd like to support all the condition codes, but our hardware doesn't + * quite match the Mesa IR, which is modeled after the NV extensions. For + * those, the instruction may update the condition codes or not, then any + * later instruction may use one of those condition codes. For gen4, the + * instruction may update the flags register based on one of the condition + * codes output by the instruction, and then further instructions may + * predicate on that. We can probably support this, but it won't + * necessarily be easy. + */ + assert(inst->DstReg.CondMask == COND_NE); + + switch (inst->DstReg.CondSwizzle) { + case SWIZZLE_XXXX: + return BRW_PREDICATE_ALIGN16_REPLICATE_X; + case SWIZZLE_YYYY: + return BRW_PREDICATE_ALIGN16_REPLICATE_Y; + case SWIZZLE_ZZZZ: + return BRW_PREDICATE_ALIGN16_REPLICATE_Z; + case SWIZZLE_WWWW: + return BRW_PREDICATE_ALIGN16_REPLICATE_W; + default: + _mesa_problem(NULL, "Unexpected predicate: 0x%08x\n", + inst->DstReg.CondMask); + return BRW_PREDICATE_NORMAL; } } -/* Emit the fragment program instructions here. +/* Emit the vertex program instructions here. */ void brw_vs_emit(struct brw_vs_compile *c ) { -#define MAX_IFSN 32 +#define MAX_IF_DEPTH 32 +#define MAX_LOOP_DEPTH 32 struct brw_compile *p = &c->func; - GLuint nr_insns = c->vp->program.Base.NumInstructions; - GLuint insn, if_insn = 0; - struct brw_instruction *end_inst; - struct brw_instruction *if_inst[MAX_IFSN]; - struct brw_indirect stack_index = brw_indirect(0, 0); - + struct brw_context *brw = p->brw; + const GLuint nr_insns = c->vp->program.Base.NumInstructions; + GLuint insn, if_depth = 0, loop_depth = 0; + GLuint end_offset = 0; + struct brw_instruction *end_inst, *last_inst; + struct brw_instruction *if_inst[MAX_IF_DEPTH], *loop_inst[MAX_LOOP_DEPTH]; + const struct brw_indirect stack_index = brw_indirect(0, 0); GLuint index; GLuint file; if (INTEL_DEBUG & DEBUG_VS) { - _mesa_printf("vs-emit:\n"); + _mesa_printf("vs-mesa:\n"); _mesa_print_program(&c->vp->program.Base); _mesa_printf("\n"); } @@ -1035,22 +1353,26 @@ void brw_vs_emit(struct brw_vs_compile *c ) for (insn = 0; insn < nr_insns; insn++) { - struct prog_instruction *inst = &c->vp->program.Base.Instructions[insn]; + const struct prog_instruction *inst = &c->vp->program.Base.Instructions[insn]; struct brw_reg args[3], dst; GLuint i; +#if 0 + printf("%d: ", insn); + _mesa_print_instruction(inst); +#endif + /* Get argument regs. SWZ is special and does this itself. */ - inst->Data = &p->store[p->nr_insn]; if (inst->Opcode != OPCODE_SWZ) for (i = 0; i < 3; i++) { - struct prog_src_register *src = &inst->SrcReg[i]; + const struct prog_src_register *src = &inst->SrcReg[i]; index = src->Index; file = src->File; - if (file == PROGRAM_OUTPUT&&c->output_regs[index].used_in_src) + if (file == PROGRAM_OUTPUT && c->output_regs[index].used_in_src) args[i] = c->output_regs[index].reg; else - args[i] = get_arg(c, src); + args[i] = get_arg(c, inst, i); } /* Get dest regs. Note that it is possible for a reg to be both @@ -1178,7 +1500,7 @@ void brw_vs_emit(struct brw_vs_compile *c ) /* The args[0] value can't be used here as it won't have * correctly encoded the full swizzle: */ - emit_swz(c, dst, inst->SrcReg[0] ); + emit_swz(c, dst, inst); break; case OPCODE_TRUNC: /* round toward zero */ @@ -1188,20 +1510,61 @@ void brw_vs_emit(struct brw_vs_compile *c ) emit_xpd(p, dst, args[0], args[1]); break; case OPCODE_IF: - assert(if_insn < MAX_IFSN); - if_inst[if_insn++] = brw_IF(p, BRW_EXECUTE_8); + assert(if_depth < MAX_IF_DEPTH); + if_inst[if_depth] = brw_IF(p, BRW_EXECUTE_8); + /* Note that brw_IF smashes the predicate_control field. */ + if_inst[if_depth]->header.predicate_control = get_predicate(inst); + if_depth++; break; case OPCODE_ELSE: - if_inst[if_insn-1] = brw_ELSE(p, if_inst[if_insn-1]); + if_inst[if_depth-1] = brw_ELSE(p, if_inst[if_depth-1]); break; case OPCODE_ENDIF: - assert(if_insn > 0); - brw_ENDIF(p, if_inst[--if_insn]); + assert(if_depth > 0); + brw_ENDIF(p, if_inst[--if_depth]); break; + case OPCODE_BGNLOOP: + loop_inst[loop_depth++] = brw_DO(p, BRW_EXECUTE_8); + break; + case OPCODE_BRK: + brw_set_predicate_control(p, get_predicate(inst)); + brw_BREAK(p); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + break; + case OPCODE_CONT: + brw_set_predicate_control(p, get_predicate(inst)); + brw_CONT(p); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + break; + case OPCODE_ENDLOOP: + { + struct brw_instruction *inst0, *inst1; + GLuint br = 1; + + loop_depth--; + + if (BRW_IS_IGDNG(brw)) + br = 2; + + inst0 = inst1 = brw_WHILE(p, loop_inst[loop_depth]); + /* patch all the BREAK/CONT instructions from last BEGINLOOP */ + while (inst0 > loop_inst[loop_depth]) { + inst0--; + if (inst0->header.opcode == BRW_OPCODE_BREAK) { + inst0->bits3.if_else.jump_count = br * (inst1 - inst0 + 1); + inst0->bits3.if_else.pop_count = 0; + } + else if (inst0->header.opcode == BRW_OPCODE_CONTINUE) { + inst0->bits3.if_else.jump_count = br * (inst1 - inst0); + inst0->bits3.if_else.pop_count = 0; + } + } + } + break; case OPCODE_BRA: - brw_set_predicate_control(p, BRW_PREDICATE_NORMAL); + brw_set_predicate_control(p, get_predicate(inst)); brw_ADD(p, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16)); - brw_set_predicate_control_flag_value(p, 0xff); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); break; case OPCODE_CAL: brw_set_access_mode(p, BRW_ALIGN_1); @@ -1209,7 +1572,7 @@ void brw_vs_emit(struct brw_vs_compile *c ) brw_set_access_mode(p, BRW_ALIGN_16); brw_ADD(p, get_addr_reg(stack_index), get_addr_reg(stack_index), brw_imm_d(4)); - inst->Data = &p->store[p->nr_insn]; + brw_save_call(p, inst->Comment, p->nr_insn); brw_ADD(p, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16)); break; case OPCODE_RET: @@ -1218,14 +1581,23 @@ void brw_vs_emit(struct brw_vs_compile *c ) brw_set_access_mode(p, BRW_ALIGN_1); brw_MOV(p, brw_ip_reg(), deref_1d(stack_index, 0)); brw_set_access_mode(p, BRW_ALIGN_16); + break; case OPCODE_END: + end_offset = p->nr_insn; + /* this instruction will get patched later to jump past subroutine + * code, etc. + */ brw_ADD(p, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16)); break; case OPCODE_PRINT: + /* no-op */ + break; case OPCODE_BGNSUB: + brw_save_label(p, inst->Comment, p->nr_insn); + break; case OPCODE_ENDSUB: - /* no-op instructions */ - break; + /* no-op */ + break; default: _mesa_problem(NULL, "Unsupported opcode %i (%s) in vertex shader", inst->Opcode, inst->Opcode < MAX_OPCODE ? @@ -1233,6 +1605,19 @@ void brw_vs_emit(struct brw_vs_compile *c ) "unknown"); } + /* Set the predication update on the last instruction of the native + * instruction sequence. + * + * This would be problematic if it was set on a math instruction, + * but that shouldn't be the case with the current GLSL compiler. + */ + if (inst->CondUpdate) { + struct brw_instruction *hw_insn = &p->store[p->nr_insn - 1]; + + assert(hw_insn->header.destreg__conditionalmod == 0); + hw_insn->header.destreg__conditionalmod = BRW_CONDITIONAL_NZ; + } + if ((inst->DstReg.File == PROGRAM_OUTPUT) && (inst->DstReg.Index != VERT_RESULT_HPOS) && c->output_regs[inst->DstReg.Index].used_in_src) { @@ -1263,9 +1648,20 @@ void brw_vs_emit(struct brw_vs_compile *c ) release_tmps(c); } - end_inst = &p->store[p->nr_insn]; + end_inst = &p->store[end_offset]; + last_inst = &p->store[p->nr_insn]; + + /* The END instruction will be patched to jump to this code */ emit_vertex_write(c); - post_vs_emit(c, end_inst); - for (insn = 0; insn < nr_insns; insn++) - c->vp->program.Base.Instructions[insn].Data = NULL; + + post_vs_emit(c, end_inst, last_inst); + + if (INTEL_DEBUG & DEBUG_VS) { + int i; + + _mesa_printf("vs-native:\n"); + for (i = 0; i < p->nr_insn; i++) + brw_disasm(stderr, &p->store[i]); + _mesa_printf("\n"); + } } |