/** * Generate exp(x) */ LLVMValueRef lp_build_exp(struct lp_build_context *bld, LLVMValueRef x) { /* log2(e) = 1/log(2) */ LLVMValueRef log2e = lp_build_const_scalar(bld->type, 1.4426950408889634); return lp_build_mul(bld, log2e, lp_build_exp2(bld, x)); }
/** * Generate log(x) */ LLVMValueRef lp_build_log(struct lp_build_context *bld, LLVMValueRef x) { /* log(2) */ LLVMValueRef log2 = lp_build_const_scalar(bld->type, 0.69314718055994529); return lp_build_mul(bld, log2, lp_build_exp2(bld, x)); }
/** * Generate pow(x, y) */ LLVMValueRef lp_build_pow(struct lp_build_context *bld, LLVMValueRef x, LLVMValueRef y) { /* TODO: optimize the constant case */ if(LLVMIsConstant(x) && LLVMIsConstant(y)) debug_printf("%s: inefficient/imprecise constant arithmetic\n", __FUNCTION__); return lp_build_exp2(bld, lp_build_mul(bld, lp_build_log2(bld, x), y)); }
/** * Emit LLVM for one TGSI instruction. * \param return TRUE for success, FALSE otherwise */ boolean lp_emit_instruction_aos( struct lp_build_tgsi_aos_context *bld, const struct tgsi_full_instruction *inst, const struct tgsi_opcode_info *info, int *pc) { LLVMValueRef src0, src1, src2; LLVMValueRef tmp0, tmp1; LLVMValueRef dst0 = NULL; /* * Stores and write masks are handled in a general fashion after the long * instruction opcode switch statement. * * Although not stricitly necessary, we avoid generating instructions for * channels which won't be stored, in cases where's that easy. For some * complex instructions, like texture sampling, it is more convenient to * assume a full writemask and then let LLVM optimization passes eliminate * redundant code. */ (*pc)++; assert(info->num_dst <= 1); if (info->num_dst) { dst0 = bld->bld_base.base.undef; } switch (inst->Instruction.Opcode) { case TGSI_OPCODE_ARL: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_floor(&bld->bld_base.base, src0); break; case TGSI_OPCODE_MOV: dst0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); break; case TGSI_OPCODE_LIT: return FALSE; case TGSI_OPCODE_RCP: /* TGSI_OPCODE_RECIP */ src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_rcp(&bld->bld_base.base, src0); break; case TGSI_OPCODE_RSQ: /* TGSI_OPCODE_RECIPSQRT */ src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); tmp0 = lp_build_emit_llvm_unary(&bld->bld_base, TGSI_OPCODE_ABS, src0); dst0 = lp_build_rsqrt(&bld->bld_base.base, tmp0); break; case TGSI_OPCODE_EXP: return FALSE; case TGSI_OPCODE_LOG: return FALSE; case TGSI_OPCODE_MUL: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); dst0 = lp_build_mul(&bld->bld_base.base, src0, src1); break; case TGSI_OPCODE_ADD: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); dst0 = lp_build_add(&bld->bld_base.base, src0, src1); break; case TGSI_OPCODE_DP3: /* TGSI_OPCODE_DOT3 */ return FALSE; case TGSI_OPCODE_DP4: /* TGSI_OPCODE_DOT4 */ return FALSE; case TGSI_OPCODE_DST: return FALSE; case TGSI_OPCODE_MIN: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); dst0 = lp_build_max(&bld->bld_base.base, src0, src1); break; case TGSI_OPCODE_MAX: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); dst0 = lp_build_max(&bld->bld_base.base, src0, src1); break; case TGSI_OPCODE_SLT: /* TGSI_OPCODE_SETLT */ src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, src1); dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero); break; case TGSI_OPCODE_SGE: /* TGSI_OPCODE_SETGE */ src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GEQUAL, src0, src1); dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero); break; case TGSI_OPCODE_MAD: /* TGSI_OPCODE_MADD */ src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL); tmp0 = lp_build_mul(&bld->bld_base.base, src0, src1); dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2); break; case TGSI_OPCODE_SUB: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); dst0 = lp_build_sub(&bld->bld_base.base, src0, src1); break; case TGSI_OPCODE_LRP: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL); tmp0 = lp_build_sub(&bld->bld_base.base, src1, src2); tmp0 = lp_build_mul(&bld->bld_base.base, src0, tmp0); dst0 = lp_build_add(&bld->bld_base.base, tmp0, src2); break; case TGSI_OPCODE_CND: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL); tmp1 = lp_build_const_vec(bld->bld_base.base.gallivm, bld->bld_base.base.type, 0.5); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src2, tmp1); dst0 = lp_build_select(&bld->bld_base.base, tmp0, src0, src1); break; case TGSI_OPCODE_DP2A: return FALSE; case TGSI_OPCODE_FRC: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); tmp0 = lp_build_floor(&bld->bld_base.base, src0); dst0 = lp_build_sub(&bld->bld_base.base, src0, tmp0); break; case TGSI_OPCODE_CLAMP: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL); tmp0 = lp_build_max(&bld->bld_base.base, src0, src1); dst0 = lp_build_min(&bld->bld_base.base, tmp0, src2); break; case TGSI_OPCODE_FLR: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_floor(&bld->bld_base.base, src0); break; case TGSI_OPCODE_ROUND: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_round(&bld->bld_base.base, src0); break; case TGSI_OPCODE_EX2: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); tmp0 = lp_build_swizzle_scalar_aos(&bld->bld_base.base, src0, TGSI_SWIZZLE_X, TGSI_NUM_CHANNELS); dst0 = lp_build_exp2(&bld->bld_base.base, tmp0); break; case TGSI_OPCODE_LG2: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X); dst0 = lp_build_log2(&bld->bld_base.base, tmp0); break; case TGSI_OPCODE_POW: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); src1 = swizzle_scalar_aos(bld, src1, TGSI_SWIZZLE_X); dst0 = lp_build_pow(&bld->bld_base.base, src0, src1); break; case TGSI_OPCODE_XPD: return FALSE; case TGSI_OPCODE_RCC: /* deprecated? */ assert(0); return FALSE; case TGSI_OPCODE_DPH: return FALSE; case TGSI_OPCODE_COS: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X); dst0 = lp_build_cos(&bld->bld_base.base, tmp0); break; case TGSI_OPCODE_DDX: return FALSE; case TGSI_OPCODE_DDY: return FALSE; case TGSI_OPCODE_KILP: /* predicated kill */ return FALSE; case TGSI_OPCODE_KIL: /* conditional kill */ return FALSE; case TGSI_OPCODE_PK2H: return FALSE; break; case TGSI_OPCODE_PK2US: return FALSE; break; case TGSI_OPCODE_PK4B: return FALSE; break; case TGSI_OPCODE_PK4UB: return FALSE; case TGSI_OPCODE_RFL: return FALSE; case TGSI_OPCODE_SEQ: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_EQUAL, src0, src1); dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero); break; case TGSI_OPCODE_SFL: dst0 = bld->bld_base.base.zero; break; case TGSI_OPCODE_SGT: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_GREATER, src0, src1); dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero); break; case TGSI_OPCODE_SIN: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); tmp0 = swizzle_scalar_aos(bld, src0, TGSI_SWIZZLE_X); dst0 = lp_build_sin(&bld->bld_base.base, tmp0); break; case TGSI_OPCODE_SLE: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LEQUAL, src0, src1); dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero); break; case TGSI_OPCODE_SNE: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_NOTEQUAL, src0, src1); dst0 = lp_build_select(&bld->bld_base.base, tmp0, bld->bld_base.base.one, bld->bld_base.base.zero); break; case TGSI_OPCODE_STR: dst0 = bld->bld_base.base.one; break; case TGSI_OPCODE_TEX: dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_NONE); break; case TGSI_OPCODE_TXD: dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV); break; case TGSI_OPCODE_UP2H: /* deprecated */ assert (0); return FALSE; break; case TGSI_OPCODE_UP2US: /* deprecated */ assert(0); return FALSE; break; case TGSI_OPCODE_UP4B: /* deprecated */ assert(0); return FALSE; break; case TGSI_OPCODE_UP4UB: /* deprecated */ assert(0); return FALSE; break; case TGSI_OPCODE_X2D: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_ARA: /* deprecated */ assert(0); return FALSE; break; case TGSI_OPCODE_ARR: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_round(&bld->bld_base.base, src0); break; case TGSI_OPCODE_BRA: /* deprecated */ assert(0); return FALSE; break; case TGSI_OPCODE_CAL: return FALSE; case TGSI_OPCODE_RET: return FALSE; case TGSI_OPCODE_END: *pc = -1; break; case TGSI_OPCODE_SSG: /* TGSI_OPCODE_SGN */ tmp0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_sgn(&bld->bld_base.base, tmp0); break; case TGSI_OPCODE_CMP: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); src1 = lp_build_emit_fetch(&bld->bld_base, inst, 1, LP_CHAN_ALL); src2 = lp_build_emit_fetch(&bld->bld_base, inst, 2, LP_CHAN_ALL); tmp0 = lp_build_cmp(&bld->bld_base.base, PIPE_FUNC_LESS, src0, bld->bld_base.base.zero); dst0 = lp_build_select(&bld->bld_base.base, tmp0, src1, src2); break; case TGSI_OPCODE_SCS: return FALSE; case TGSI_OPCODE_TXB: dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_LOD_BIAS); break; case TGSI_OPCODE_NRM: /* fall-through */ case TGSI_OPCODE_NRM4: return FALSE; case TGSI_OPCODE_DIV: /* deprecated */ assert(0); return FALSE; break; case TGSI_OPCODE_DP2: return FALSE; case TGSI_OPCODE_TXL: dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_EXPLICIT_LOD); break; case TGSI_OPCODE_TXP: dst0 = emit_tex(bld, inst, LP_BLD_TEX_MODIFIER_PROJECTED); break; case TGSI_OPCODE_BRK: return FALSE; case TGSI_OPCODE_IF: return FALSE; case TGSI_OPCODE_BGNLOOP: return FALSE; case TGSI_OPCODE_BGNSUB: return FALSE; case TGSI_OPCODE_ELSE: return FALSE; case TGSI_OPCODE_ENDIF: return FALSE; case TGSI_OPCODE_ENDLOOP: return FALSE; case TGSI_OPCODE_ENDSUB: return FALSE; case TGSI_OPCODE_PUSHA: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_POPA: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_CEIL: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_ceil(&bld->bld_base.base, src0); break; case TGSI_OPCODE_I2F: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_NOT: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_TRUNC: src0 = lp_build_emit_fetch(&bld->bld_base, inst, 0, LP_CHAN_ALL); dst0 = lp_build_trunc(&bld->bld_base.base, src0); break; case TGSI_OPCODE_SHL: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_ISHR: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_AND: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_OR: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_MOD: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_XOR: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_SAD: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_TXF: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_TXQ: /* deprecated? */ assert(0); return FALSE; break; case TGSI_OPCODE_CONT: return FALSE; case TGSI_OPCODE_EMIT: return FALSE; break; case TGSI_OPCODE_ENDPRIM: return FALSE; break; case TGSI_OPCODE_NOP: break; default: return FALSE; } if (info->num_dst) { lp_emit_store_aos(bld, inst, 0, dst0); } return TRUE; }