void
gen8_vec4_generator::generate_scratch_write(vec4_instruction *ir,
struct brw_reg dst,
struct brw_reg src,
struct brw_reg index)
{
struct brw_reg header = brw_vec8_grf(GEN7_MRF_HACK_START + ir->base_mrf, 0);
MOV_RAW(header, brw_vec8_grf(0, 0));
generate_oword_dual_block_offsets(brw_message_reg(ir->base_mrf + 1), index);
MOV(retype(brw_message_reg(ir->base_mrf + 2), BRW_REGISTER_TYPE_D),
retype(src, BRW_REGISTER_TYPE_D));
/* Each of the 8 channel enables is considered for whether each
* dword is written.
*/
gen8_instruction *send = next_inst(BRW_OPCODE_SEND);
gen8_set_dst(brw, send, dst);
gen8_set_src0(brw, send, header);
gen8_set_pred_control(send, ir->predicate);
gen8_set_dp_message(brw, send, GEN7_SFID_DATAPORT_DATA_CACHE,
255, /* binding table index: stateless access */
GEN7_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE,
BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
3, /* mlen */
0, /* rlen */
true, /* header present */
false); /* EOT */
}
void
vec4_generator::generate_scratch_read(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg index)
{
struct brw_reg header = brw_vec8_grf(0, 0);
gen6_resolve_implied_move(p, &header, inst->base_mrf);
generate_oword_dual_block_offsets(brw_message_reg(inst->base_mrf + 1),
index);
uint32_t msg_type;
if (brw->gen >= 6)
msg_type = GEN6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
else if (brw->gen == 5 || brw->is_g4x)
msg_type = G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
else
msg_type = BRW_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ;
/* Each of the 8 channel enables is considered for whether each
* dword is written.
*/
struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, send, dst);
brw_set_src0(p, send, header);
if (brw->gen < 6)
send->header.destreg__conditionalmod = inst->base_mrf;
brw_set_dp_read_message(p, send,
255, /* binding table index: stateless access */
BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
msg_type,
BRW_DATAPORT_READ_TARGET_RENDER_CACHE,
2, /* mlen */
true, /* header_present */
1 /* rlen */);
}
void
vec4_generator::generate_scratch_write(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src,
struct brw_reg index)
{
struct brw_reg header = brw_vec8_grf(0, 0);
bool write_commit;
/* If the instruction is predicated, we'll predicate the send, not
* the header setup.
*/
brw_set_predicate_control(p, false);
gen6_resolve_implied_move(p, &header, inst->base_mrf);
generate_oword_dual_block_offsets(brw_message_reg(inst->base_mrf + 1),
index);
brw_MOV(p,
retype(brw_message_reg(inst->base_mrf + 2), BRW_REGISTER_TYPE_D),
retype(src, BRW_REGISTER_TYPE_D));
uint32_t msg_type;
if (brw->gen >= 7)
msg_type = GEN7_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE;
else if (brw->gen == 6)
msg_type = GEN6_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE;
else
msg_type = BRW_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE;
brw_set_predicate_control(p, inst->predicate);
/* Pre-gen6, we have to specify write commits to ensure ordering
* between reads and writes within a thread. Afterwards, that's
* guaranteed and write commits only matter for inter-thread
* synchronization.
*/
if (brw->gen >= 6) {
write_commit = false;
} else {
/* The visitor set up our destination register to be g0. This
* means that when the next read comes along, we will end up
* reading from g0 and causing a block on the write commit. For
* write-after-read, we are relying on the value of the previous
* read being used (and thus blocking on completion) before our
* write is executed. This means we have to be careful in
* instruction scheduling to not violate this assumption.
*/
write_commit = true;
}
/* Each of the 8 channel enables is considered for whether each
* dword is written.
*/
struct brw_instruction *send = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, send, dst);
brw_set_src0(p, send, header);
if (brw->gen < 6)
send->header.destreg__conditionalmod = inst->base_mrf;
brw_set_dp_write_message(p, send,
255, /* binding table index: stateless access */
BRW_DATAPORT_OWORD_DUAL_BLOCK_1OWORD,
msg_type,
3, /* mlen */
true, /* header present */
false, /* not a render target write */
write_commit, /* rlen */
false, /* eot */
write_commit);
}
