16-bit abs/neg creates v_xor_b32/v_and_b32 with v2b definitions. These
instructions never do partial writes without SDWA.
No shader-db changes.
Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5040>
else if (instr->format == Format::PSEUDO)
return std::make_pair(4, rc.size() * 4u);
else if (instr->format == Format::PSEUDO)
return std::make_pair(4, rc.size() * 4u);
- bool can_do_partial = chip >= GFX10;
+ unsigned bytes_written = chip >= GFX10 ? rc.bytes() : 4u;
switch (instr->opcode) {
case aco_opcode::v_mad_f16:
case aco_opcode::v_mad_u16:
switch (instr->opcode) {
case aco_opcode::v_mad_f16:
case aco_opcode::v_mad_u16:
case aco_opcode::v_fma_f16:
case aco_opcode::v_div_fixup_f16:
case aco_opcode::v_interp_p2_f16:
case aco_opcode::v_fma_f16:
case aco_opcode::v_div_fixup_f16:
case aco_opcode::v_interp_p2_f16:
- can_do_partial = chip >= GFX9;
+ bytes_written = chip >= GFX9 ? rc.bytes() : 4u;
+ bytes_written = MAX2(bytes_written, instr_info.definition_size[(int)instr->opcode] / 8u);
if (can_use_SDWA(chip, instr)) {
return std::make_pair(rc.bytes(), rc.bytes());
if (can_use_SDWA(chip, instr)) {
return std::make_pair(rc.bytes(), rc.bytes());
- return std::make_pair(4u, can_do_partial ? rc.bytes() : 4u);
+ return std::make_pair(4u, bytes_written);
}
void add_subdword_definition(Program *program, aco_ptr<Instruction>& instr, unsigned idx, PhysReg reg, bool is_partial)
}
void add_subdword_definition(Program *program, aco_ptr<Instruction>& instr, unsigned idx, PhysReg reg, bool is_partial)
- return chip >= GFX10 ? def.bytes() : 4;
+ return MAX2(chip >= GFX10 ? def.bytes() : 4, instr_info.definition_size[(int)instr->opcode] / 8u);