return (dpp_ctrl)(((unsigned) _dpp_row_sr) | amount);
}
+inline dpp_ctrl
+dpp_row_rr(unsigned amount)
+{
+ assert(amount > 0 && amount < 16);
+ return (dpp_ctrl)(((unsigned) _dpp_row_rr) | amount);
+}
+
inline unsigned
ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
{
aco_ptr<Instruction> create_s_mov(Definition dst, Operand src);
+extern uint8_t int8_mul_table[512];
+
enum sendmsg {
sendmsg_none = 0,
_sendmsg_gs = 2,
std::vector<aco_ptr<Instruction>> *instructions;
std::vector<aco_ptr<Instruction>>::iterator it;
+ bool is_precise = false;
+ bool is_nuw = false;
- Builder(Program *pgm) : program(pgm), use_iterator(false), start(false), lm(pgm->lane_mask), instructions(NULL) {}
+ Builder(Program *pgm) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), instructions(NULL) {}
Builder(Program *pgm, Block *block) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), instructions(&block->instructions) {}
Builder(Program *pgm, std::vector<aco_ptr<Instruction>> *instrs) : program(pgm), use_iterator(false), start(false), lm(pgm ? pgm->lane_mask : s2), instructions(instrs) {}
+ Builder precise() const {
+ Builder res = *this;
+ res.is_precise = true;
+ return res;
+ };
+
+ Builder nuw() const {
+ Builder res = *this;
+ res.is_nuw = true;
+ return res;
+ }
+
void moveEnd(Block *block) {
instructions = &block->instructions;
}
% for fixed in ['m0', 'vcc', 'exec', 'scc']:
Operand ${fixed}(Temp tmp) {
% if fixed == 'vcc' or fixed == 'exec':
- assert(tmp.regClass() == lm);
+ //vcc_hi and exec_hi can still be used in wave32
+ assert(tmp.type() == RegType::sgpr && tmp.bytes() <= 8);
% endif
Operand op(tmp);
op.setFixed(aco::${fixed});
Definition ${fixed}(Definition def) {
% if fixed == 'vcc' or fixed == 'exec':
- assert(def.regClass() == lm);
+ //vcc_hi and exec_hi can still be used in wave32
+ assert(def.regClass().type() == RegType::sgpr && def.bytes() <= 8);
% endif
def.setFixed(aco::${fixed});
return def;
Definition hint_${fixed}(Definition def) {
% if fixed == 'vcc' or fixed == 'exec':
- assert(def.regClass() == lm);
+ //vcc_hi and exec_hi can still be used in wave32
+ assert(def.regClass().type() == RegType::sgpr && def.bytes() <= 8);
% endif
def.setHint(aco::${fixed});
return def;
Result copy(Definition dst, Op op_) {
Operand op = op_.op;
+ assert(op.bytes() == dst.bytes());
if (dst.regClass() == s1 && op.size() == 1 && op.isLiteral()) {
uint32_t imm = op.constantValue();
if (imm == 0x3e22f983) {
}
}
- if (dst.regClass() == s2) {
+ if (dst.regClass() == s1) {
+ return sop1(aco_opcode::s_mov_b32, dst, op);
+ } else if (dst.regClass() == s2) {
return sop1(aco_opcode::s_mov_b64, dst, op);
- } else if (op.size() > 1) {
- return pseudo(aco_opcode::p_create_vector, dst, op);
} else if (dst.regClass() == v1 || dst.regClass() == v1.as_linear()) {
return vop1(aco_opcode::v_mov_b32, dst, op);
+ } else if (op.bytes() > 2) {
+ return pseudo(aco_opcode::p_create_vector, dst, op);
+ } else if (op.bytes() == 1 && op.isConstant()) {
+ uint8_t val = op.constantValue();
+ Operand op32((uint32_t)val | (val & 0x80u ? 0xffffff00u : 0u));
+ aco_ptr<SDWA_instruction> sdwa;
+ if (op32.isLiteral()) {
+ sdwa.reset(create_instruction<SDWA_instruction>(aco_opcode::v_mul_u32_u24, asSDWA(Format::VOP2), 2, 1));
+ uint32_t a = (uint32_t)int8_mul_table[val * 2];
+ uint32_t b = (uint32_t)int8_mul_table[val * 2 + 1];
+ sdwa->operands[0] = Operand(a | (a & 0x80u ? 0xffffff00u : 0x0u));
+ sdwa->operands[1] = Operand(b | (b & 0x80u ? 0xffffff00u : 0x0u));
+ } else {
+ sdwa.reset(create_instruction<SDWA_instruction>(aco_opcode::v_mov_b32, asSDWA(Format::VOP1), 1, 1));
+ sdwa->operands[0] = op32;
+ }
+ sdwa->definitions[0] = dst;
+ sdwa->sel[0] = sdwa_udword;
+ sdwa->sel[1] = sdwa_udword;
+ sdwa->dst_sel = sdwa_ubyte;
+ sdwa->dst_preserve = true;
+ return insert(std::move(sdwa));
+ } else if (op.bytes() == 2 && op.isConstant() && !op.isLiteral()) {
+ aco_ptr<SDWA_instruction> sdwa{create_instruction<SDWA_instruction>(aco_opcode::v_add_f16, asSDWA(Format::VOP2), 2, 1)};
+ sdwa->operands[0] = op;
+ sdwa->operands[1] = Operand(0u);
+ sdwa->definitions[0] = dst;
+ sdwa->sel[0] = sdwa_uword;
+ sdwa->sel[1] = sdwa_udword;
+ sdwa->dst_sel = dst.bytes() == 1 ? sdwa_ubyte : sdwa_uword;
+ sdwa->dst_preserve = true;
+ return insert(std::move(sdwa));
+ } else if (dst.regClass().is_subdword()) {
+ if (program->chip_class >= GFX8) {
+ aco_ptr<SDWA_instruction> sdwa{create_instruction<SDWA_instruction>(aco_opcode::v_mov_b32, asSDWA(Format::VOP1), 1, 1)};
+ sdwa->operands[0] = op;
+ sdwa->definitions[0] = dst;
+ sdwa->sel[0] = op.bytes() == 1 ? sdwa_ubyte : sdwa_uword;
+ sdwa->dst_sel = dst.bytes() == 1 ? sdwa_ubyte : sdwa_uword;
+ sdwa->dst_preserve = true;
+ return insert(std::move(sdwa));
+ } else {
+ return vop1(aco_opcode::v_mov_b32, dst, op);
+ }
} else {
- assert(dst.regClass() == s1);
- return sop1(aco_opcode::s_mov_b32, dst, op);
+ unreachable("Unhandled case in bld.copy()");
}
}
<%
import itertools
formats = [("pseudo", [Format.PSEUDO], 'Pseudo_instruction', list(itertools.product(range(5), range(5))) + [(8, 1), (1, 8)]),
- ("sop1", [Format.SOP1], 'SOP1_instruction', [(1, 1), (2, 1), (3, 2)]),
+ ("sop1", [Format.SOP1], 'SOP1_instruction', [(0, 1), (1, 0), (1, 1), (2, 1), (3, 2)]),
("sop2", [Format.SOP2], 'SOP2_instruction', itertools.product([1, 2], [2, 3])),
("sopk", [Format.SOPK], 'SOPK_instruction', itertools.product([0, 1, 2], [0, 1])),
- ("sopp", [Format.SOPP], 'SOPP_instruction', [(0, 0), (0, 1)]),
+ ("sopp", [Format.SOPP], 'SOPP_instruction', itertools.product([0, 1], [0, 1])),
("sopc", [Format.SOPC], 'SOPC_instruction', [(1, 2)]),
("smem", [Format.SMEM], 'SMEM_instruction', [(0, 4), (0, 3), (1, 0), (1, 3), (1, 2), (0, 0)]),
("ds", [Format.DS], 'DS_instruction', [(1, 1), (1, 2), (0, 3), (0, 4)]),
("mtbuf", [Format.MTBUF], 'MTBUF_instruction', [(0, 4), (1, 3)]),
("mimg", [Format.MIMG], 'MIMG_instruction', [(0, 3), (1, 3)]),
("exp", [Format.EXP], 'Export_instruction', [(0, 4)]),
- ("branch", [Format.PSEUDO_BRANCH], 'Pseudo_branch_instruction', itertools.product([0], [0, 1])),
+ ("branch", [Format.PSEUDO_BRANCH], 'Pseudo_branch_instruction', itertools.product([1], [0, 1])),
("barrier", [Format.PSEUDO_BARRIER], 'Pseudo_barrier_instruction', [(0, 0)]),
- ("reduction", [Format.PSEUDO_REDUCTION], 'Pseudo_reduction_instruction', [(3, 2), (3, 4)]),
- ("vop1", [Format.VOP1], 'VOP1_instruction', [(1, 1), (2, 2)]),
+ ("reduction", [Format.PSEUDO_REDUCTION], 'Pseudo_reduction_instruction', [(3, 2)]),
+ ("vop1", [Format.VOP1], 'VOP1_instruction', [(0, 0), (1, 1), (2, 2)]),
("vop2", [Format.VOP2], 'VOP2_instruction', itertools.product([1, 2], [2, 3])),
+ ("vop2_sdwa", [Format.VOP2, Format.SDWA], 'SDWA_instruction', itertools.product([1, 2], [2, 3])),
("vopc", [Format.VOPC], 'VOPC_instruction', itertools.product([1, 2], [2])),
("vop3", [Format.VOP3A], 'VOP3A_instruction', [(1, 3), (1, 2), (1, 1), (2, 2)]),
("vintrp", [Format.VINTRP], 'Interp_instruction', [(1, 2), (1, 3)]),
("vopc_e64", [Format.VOPC, Format.VOP3A], 'VOP3A_instruction', itertools.product([1, 2], [2])),
("flat", [Format.FLAT], 'FLAT_instruction', [(0, 3), (1, 2)]),
("global", [Format.GLOBAL], 'FLAT_instruction', [(0, 3), (1, 2)])]
+formats = [(f if len(f) == 5 else f + ('',)) for f in formats]
%>\\
-% for name, formats, struct, shapes in formats:
+% for name, formats, struct, shapes, extra_field_setup in formats:
% for num_definitions, num_operands in shapes:
<%
args = ['aco_opcode opcode']
${struct} *instr = create_instruction<${struct}>(opcode, (Format)(${'|'.join('(int)Format::%s' % f.name for f in formats)}), ${num_operands}, ${num_definitions});
% for i in range(num_definitions):
instr->definitions[${i}] = def${i};
+ instr->definitions[${i}].setPrecise(is_precise);
+ instr->definitions[${i}].setNUW(is_nuw);
% endfor
% for i in range(num_operands):
instr->operands[${i}] = op${i}.op;
% for dest, field_name in zip(f.get_builder_field_dests(), f.get_builder_field_names()):
instr->${dest} = ${field_name};
% endfor
+ ${f.get_builder_initialization(num_operands)}
% endfor
+ ${extra_field_setup}
return insert(instr);
}