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 | 180 |
1 files changed, 152 insertions, 28 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_vs_emit.c b/src/mesa/drivers/dri/i965/brw_vs_emit.c index 72cd36a603..108e19cdbc 100644 --- a/src/mesa/drivers/dri/i965/brw_vs_emit.c +++ b/src/mesa/drivers/dri/i965/brw_vs_emit.c @@ -70,13 +70,18 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) GLuint i, reg = 0, mrf; int attributes_in_vue; -#if 0 - if (c->vp->program.Base.Parameters->NumParameters >= 6) - c->vp->use_const_buffer = 1; + /* 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 -#endif c->vp->use_const_buffer = GL_FALSE; - /*printf("use_const_buffer = %d\n", c->use_const_buffer);*/ + + /*printf("use_const_buffer = %d\n", c->vp->use_const_buffer);*/ /* r0 -- reserved as usual */ @@ -130,15 +135,21 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) 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)) { 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++; @@ -149,8 +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++; + } } } } @@ -218,7 +238,10 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c ) */ attributes_in_vue = MAX2(c->nr_outputs, c->nr_inputs); - c->prog_data.urb_entry_size = (attributes_in_vue + 2 + 3) / 4; + 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; c->prog_data.total_grf = reg; @@ -727,10 +750,11 @@ get_constant(struct brw_vs_compile *c, 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 || src->RelAddr) { + 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; @@ -743,13 +767,13 @@ get_constant(struct brw_vs_compile *c, brw_dp_READ_4_vs(p, c->current_const[argIndex].reg,/* writeback dest */ 0, /* oword */ - src->RelAddr, /* relative indexing? */ + relAddr, /* relative indexing? */ addrReg, /* address register */ 16 * src->Index, /* byte offset */ SURF_INDEX_VERT_CONST_BUFFER /* binding table index */ ); - if (src->RelAddr) { + if (relAddr) { /* second read */ const2_reg = get_tmp(c); @@ -760,7 +784,7 @@ get_constant(struct brw_vs_compile *c, brw_dp_READ_4_vs(p, const2_reg, /* writeback dest */ 1, /* oword */ - src->RelAddr, /* relative indexing? */ + relAddr, /* relative indexing? */ addrReg, /* address register */ 16 * src->Index, /* byte offset */ SURF_INDEX_VERT_CONST_BUFFER @@ -770,7 +794,7 @@ get_constant(struct brw_vs_compile *c, const_reg = c->current_const[argIndex].reg; - if (src->RelAddr) { + if (relAddr) { /* merge the two Owords into the constant register */ /* const_reg[7..4] = const2_reg[7..4] */ brw_MOV(p, @@ -887,6 +911,7 @@ get_src_reg( struct brw_vs_compile *c, 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); } @@ -906,7 +931,6 @@ get_src_reg( struct brw_vs_compile *c, return brw_null_reg(); case PROGRAM_LOCAL_PARAM: - case PROGRAM_ENV_PARAM: case PROGRAM_WRITE_ONLY: default: assert(0); @@ -1079,6 +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, @@ -1088,14 +1114,16 @@ static void emit_vertex_write( struct brw_vs_compile *c) /* Build ndc coords */ 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.nr_userclip || BRW_IS_965(p->brw)) { struct brw_reg header1 = retype(get_tmp(c), BRW_REGISTER_TYPE_UD); GLuint i; @@ -1126,7 +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)) { + if (BRW_IS_965(p->brw)) { brw_CMP(p, vec8(brw_null_reg()), BRW_CONDITIONAL_L, @@ -1153,7 +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 */ @@ -1161,12 +1205,43 @@ 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 */ + eot, /* eot */ 1, /* 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); + } } @@ -1187,7 +1262,11 @@ post_vs_emit( struct brw_vs_compile *c, /* patch up the END code to jump past subroutines, etc */ offset = last_inst - end_inst; - brw_set_src1(end_inst, brw_imm_d(offset * 16)); + if (offset > 1) { + brw_set_src1(end_inst, brw_imm_d(offset * 16)); + } else { + end_inst->header.opcode = BRW_OPCODE_NOP; + } } static uint32_t @@ -1213,18 +1292,19 @@ get_predicate(uint32_t swizzle) void brw_vs_emit(struct brw_vs_compile *c ) { #define MAX_IF_DEPTH 32 +#define MAX_LOOP_DEPTH 32 struct brw_compile *p = &c->func; const GLuint nr_insns = c->vp->program.Base.NumInstructions; - GLuint insn, if_depth = 0; + 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]; + 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"); } @@ -1253,7 +1333,7 @@ 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; @@ -1266,7 +1346,7 @@ void brw_vs_emit(struct brw_vs_compile *c ) */ 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) @@ -1423,6 +1503,41 @@ void brw_vs_emit(struct brw_vs_compile *c ) assert(if_depth > 0); brw_ENDIF(p, if_inst[--if_depth]); break; +#if 0 + case OPCODE_BGNLOOP: + loop_inst[loop_depth++] = brw_DO(p, BRW_EXECUTE_8); + break; + case OPCODE_BRK: + brw_BREAK(p); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + break; + case OPCODE_CONT: + brw_CONT(p); + brw_set_predicate_control(p, BRW_PREDICATE_NONE); + break; + case OPCODE_ENDLOOP: + { + struct brw_instruction *inst0, *inst1; + loop_depth--; + 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 = inst1 - inst0 + 1; + inst0->bits3.if_else.pop_count = 0; + } + else if (inst0->header.opcode == BRW_OPCODE_CONTINUE) { + inst0->bits3.if_else.jump_count = inst1 - inst0; + inst0->bits3.if_else.pop_count = 0; + } + } + } + break; +#else + (void) loop_inst; + (void) loop_depth; +#endif case OPCODE_BRA: brw_set_predicate_control(p, BRW_PREDICATE_NORMAL); brw_ADD(p, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16)); @@ -1517,4 +1632,13 @@ void brw_vs_emit(struct brw_vs_compile *c ) emit_vertex_write(c); 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"); + } } |