static void
replace_with_mov(struct vc4_compile *c, struct qinst *inst, struct qreg arg)
{
dump_from(c, inst);
if (qir_is_mul(inst))
inst->op = QOP_MMOV;
else if (qir_is_float_input(inst))
inst->op = QOP_FMOV;
else
inst->op = QOP_MOV;
inst->src[0] = arg;
inst->src[1] = c->undef;
dump_to(c, inst);
}
void
vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c)
{
struct qpu_reg *temp_registers = vc4_register_allocate(vc4, c);
bool discard = false;
uint32_t inputs_remaining = c->num_inputs;
uint32_t vpm_read_fifo_count = 0;
uint32_t vpm_read_offset = 0;
int last_vpm_read_index = -1;
/* Map from the QIR ops enum order to QPU unpack bits. */
static const uint32_t unpack_map[] = {
QPU_UNPACK_8A,
QPU_UNPACK_8B,
QPU_UNPACK_8C,
QPU_UNPACK_8D,
QPU_UNPACK_16A_TO_F32,
QPU_UNPACK_16B_TO_F32,
};
list_inithead(&c->qpu_inst_list);
switch (c->stage) {
case QSTAGE_VERT:
case QSTAGE_COORD:
/* There's a 4-entry FIFO for VPMVCD reads, each of which can
* load up to 16 dwords (4 vec4s) per vertex.
*/
while (inputs_remaining) {
uint32_t num_entries = MIN2(inputs_remaining, 16);
queue(c, qpu_load_imm_ui(qpu_vrsetup(),
vpm_read_offset |
0x00001a00 |
((num_entries & 0xf) << 20)));
inputs_remaining -= num_entries;
vpm_read_offset += num_entries;
vpm_read_fifo_count++;
}
assert(vpm_read_fifo_count <= 4);
queue(c, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00));
break;
case QSTAGE_FRAG:
break;
}
list_for_each_entry(struct qinst, qinst, &c->instructions, link) {
#if 0
fprintf(stderr, "translating qinst to qpu: ");
qir_dump_inst(qinst);
fprintf(stderr, "\n");
#endif
static const struct {
uint32_t op;
} translate[] = {
#define A(name) [QOP_##name] = {QPU_A_##name}
#define M(name) [QOP_##name] = {QPU_M_##name}
A(FADD),
A(FSUB),
A(FMIN),
A(FMAX),
A(FMINABS),
A(FMAXABS),
A(FTOI),
A(ITOF),
A(ADD),
A(SUB),
A(SHL),
A(SHR),
A(ASR),
A(MIN),
A(MAX),
A(AND),
A(OR),
A(XOR),
A(NOT),
M(FMUL),
M(MUL24),
};
struct qpu_reg src[4];
for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
int index = qinst->src[i].index;
switch (qinst->src[i].file) {
case QFILE_NULL:
src[i] = qpu_rn(0);
break;
case QFILE_TEMP:
src[i] = temp_registers[index];
break;
case QFILE_UNIF:
src[i] = qpu_unif();
break;
case QFILE_VARY:
src[i] = qpu_vary();
break;
case QFILE_SMALL_IMM:
src[i].mux = QPU_MUX_SMALL_IMM;
src[i].addr = qpu_encode_small_immediate(qinst->src[i].index);
/* This should only have returned a valid
* small immediate field, not ~0 for failure.
*/
assert(src[i].addr <= 47);
break;
case QFILE_VPM:
assert((int)qinst->src[i].index >=
last_vpm_read_index);
(void)last_vpm_read_index;
last_vpm_read_index = qinst->src[i].index;
src[i] = qpu_ra(QPU_R_VPM);
break;
}
}
struct qpu_reg dst;
switch (qinst->dst.file) {
case QFILE_NULL:
dst = qpu_ra(QPU_W_NOP);
break;
case QFILE_TEMP:
dst = temp_registers[qinst->dst.index];
break;
case QFILE_VPM:
dst = qpu_ra(QPU_W_VPM);
break;
case QFILE_VARY:
case QFILE_UNIF:
case QFILE_SMALL_IMM:
assert(!"not reached");
break;
}
switch (qinst->op) {
case QOP_MOV:
/* Skip emitting the MOV if it's a no-op. */
if (dst.mux == QPU_MUX_A || dst.mux == QPU_MUX_B ||
dst.mux != src[0].mux || dst.addr != src[0].addr) {
queue(c, qpu_a_MOV(dst, src[0]));
}
break;
case QOP_SEL_X_0_ZS:
case QOP_SEL_X_0_ZC:
case QOP_SEL_X_0_NS:
case QOP_SEL_X_0_NC:
case QOP_SEL_X_0_CS:
case QOP_SEL_X_0_CC:
queue(c, qpu_a_MOV(dst, src[0]));
set_last_cond_add(c, qinst->op - QOP_SEL_X_0_ZS +
QPU_COND_ZS);
queue(c, qpu_a_XOR(dst, qpu_r0(), qpu_r0()));
set_last_cond_add(c, ((qinst->op - QOP_SEL_X_0_ZS) ^
1) + QPU_COND_ZS);
break;
case QOP_SEL_X_Y_ZS:
case QOP_SEL_X_Y_ZC:
case QOP_SEL_X_Y_NS:
case QOP_SEL_X_Y_NC:
case QOP_SEL_X_Y_CS:
case QOP_SEL_X_Y_CC:
queue(c, qpu_a_MOV(dst, src[0]));
set_last_cond_add(c, qinst->op - QOP_SEL_X_Y_ZS +
QPU_COND_ZS);
queue(c, qpu_a_MOV(dst, src[1]));
set_last_cond_add(c, ((qinst->op - QOP_SEL_X_Y_ZS) ^
1) + QPU_COND_ZS);
break;
case QOP_RCP:
case QOP_RSQ:
case QOP_EXP2:
case QOP_LOG2:
switch (qinst->op) {
case QOP_RCP:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP),
src[0]));
break;
case QOP_RSQ:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT),
src[0]));
break;
case QOP_EXP2:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP),
src[0]));
break;
case QOP_LOG2:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG),
src[0]));
break;
default:
abort();
}
if (dst.mux != QPU_MUX_R4)
queue(c, qpu_a_MOV(dst, qpu_r4()));
break;
case QOP_PACK_8888_F:
queue(c, qpu_m_MOV(dst, src[0]));
*last_inst(c) |= QPU_PM;
*last_inst(c) |= QPU_SET_FIELD(QPU_PACK_MUL_8888,
QPU_PACK);
break;
case QOP_PACK_8A_F:
case QOP_PACK_8B_F:
case QOP_PACK_8C_F:
case QOP_PACK_8D_F:
queue(c,
qpu_m_MOV(dst, src[0]) |
QPU_PM |
QPU_SET_FIELD(QPU_PACK_MUL_8A +
qinst->op - QOP_PACK_8A_F,
QPU_PACK));
break;
case QOP_FRAG_X:
queue(c, qpu_a_ITOF(dst,
qpu_ra(QPU_R_XY_PIXEL_COORD)));
break;
case QOP_FRAG_Y:
queue(c, qpu_a_ITOF(dst,
qpu_rb(QPU_R_XY_PIXEL_COORD)));
break;
case QOP_FRAG_REV_FLAG:
queue(c, qpu_a_ITOF(dst,
qpu_rb(QPU_R_MS_REV_FLAGS)));
break;
case QOP_FRAG_Z:
case QOP_FRAG_W:
/* QOP_FRAG_Z/W don't emit instructions, just allocate
* the register to the Z/W payload.
*/
break;
case QOP_TLB_DISCARD_SETUP:
discard = true;
queue(c, qpu_a_MOV(src[0], src[0]));
*last_inst(c) |= QPU_SF;
break;
case QOP_TLB_STENCIL_SETUP:
queue(c, qpu_a_MOV(qpu_ra(QPU_W_TLB_STENCIL_SETUP), src[0]));
break;
case QOP_TLB_Z_WRITE:
queue(c, qpu_a_MOV(qpu_ra(QPU_W_TLB_Z), src[0]));
if (discard) {
set_last_cond_add(c, QPU_COND_ZS);
}
break;
case QOP_TLB_COLOR_READ:
queue(c, qpu_NOP());
*last_inst(c) = qpu_set_sig(*last_inst(c),
QPU_SIG_COLOR_LOAD);
if (dst.mux != QPU_MUX_R4)
queue(c, qpu_a_MOV(dst, qpu_r4()));
break;
case QOP_TLB_COLOR_WRITE:
queue(c, qpu_a_MOV(qpu_tlbc(), src[0]));
if (discard) {
set_last_cond_add(c, QPU_COND_ZS);
}
break;
case QOP_VARY_ADD_C:
queue(c, qpu_a_FADD(dst, src[0], qpu_r5()));
break;
case QOP_TEX_S:
case QOP_TEX_T:
case QOP_TEX_R:
case QOP_TEX_B:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_TMU0_S +
(qinst->op - QOP_TEX_S)),
src[0]));
break;
case QOP_TEX_DIRECT:
fixup_raddr_conflict(c, dst, &src[0], &src[1]);
queue(c, qpu_a_ADD(qpu_rb(QPU_W_TMU0_S), src[0], src[1]));
break;
case QOP_TEX_RESULT:
queue(c, qpu_NOP());
*last_inst(c) = qpu_set_sig(*last_inst(c),
QPU_SIG_LOAD_TMU0);
if (dst.mux != QPU_MUX_R4)
queue(c, qpu_a_MOV(dst, qpu_r4()));
break;
case QOP_UNPACK_8A_F:
case QOP_UNPACK_8B_F:
case QOP_UNPACK_8C_F:
case QOP_UNPACK_8D_F:
case QOP_UNPACK_16A_F:
case QOP_UNPACK_16B_F: {
if (src[0].mux == QPU_MUX_R4) {
queue(c, qpu_a_MOV(dst, src[0]));
*last_inst(c) |= QPU_PM;
*last_inst(c) |= QPU_SET_FIELD(QPU_UNPACK_8A +
(qinst->op -
QOP_UNPACK_8A_F),
QPU_UNPACK);
} else {
assert(src[0].mux == QPU_MUX_A);
/* Since we're setting the pack bits, if the
* destination is in A it would get re-packed.
*/
queue(c, qpu_a_FMAX((dst.mux == QPU_MUX_A ?
qpu_rb(31) : dst),
src[0], src[0]));
*last_inst(c) |=
QPU_SET_FIELD(unpack_map[qinst->op -
QOP_UNPACK_8A_F],
QPU_UNPACK);
if (dst.mux == QPU_MUX_A) {
queue(c, qpu_a_MOV(dst, qpu_rb(31)));
}
}
}
break;
case QOP_UNPACK_8A_I:
case QOP_UNPACK_8B_I:
case QOP_UNPACK_8C_I:
case QOP_UNPACK_8D_I:
case QOP_UNPACK_16A_I:
case QOP_UNPACK_16B_I: {
assert(src[0].mux == QPU_MUX_A);
/* Since we're setting the pack bits, if the
* destination is in A it would get re-packed.
*/
queue(c, qpu_a_MOV((dst.mux == QPU_MUX_A ?
qpu_rb(31) : dst), src[0]));
*last_inst(c) |= QPU_SET_FIELD(unpack_map[qinst->op -
QOP_UNPACK_8A_I],
QPU_UNPACK);
if (dst.mux == QPU_MUX_A) {
queue(c, qpu_a_MOV(dst, qpu_rb(31)));
}
}
break;
default:
assert(qinst->op < ARRAY_SIZE(translate));
assert(translate[qinst->op].op != 0); /* NOPs */
/* If we have only one source, put it in the second
* argument slot as well so that we don't take up
* another raddr just to get unused data.
*/
if (qir_get_op_nsrc(qinst->op) == 1)
src[1] = src[0];
fixup_raddr_conflict(c, dst, &src[0], &src[1]);
if (qir_is_mul(qinst)) {
queue(c, qpu_m_alu2(translate[qinst->op].op,
dst,
src[0], src[1]));
if (qinst->dst.pack) {
*last_inst(c) |= QPU_PM;
*last_inst(c) |= QPU_SET_FIELD(qinst->dst.pack,
QPU_PACK);
}
} else {
queue(c, qpu_a_alu2(translate[qinst->op].op,
dst,
src[0], src[1]));
if (qinst->dst.pack) {
assert(dst.mux == QPU_MUX_A);
*last_inst(c) |= QPU_SET_FIELD(qinst->dst.pack,
QPU_PACK);
}
}
break;
}
if (qinst->sf) {
assert(!qir_is_multi_instruction(qinst));
*last_inst(c) |= QPU_SF;
}
}
qpu_schedule_instructions(c);
/* thread end can't have VPM write or read */
if (QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
QPU_WADDR_ADD) == QPU_W_VPM ||
QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
QPU_WADDR_MUL) == QPU_W_VPM ||
QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
QPU_RADDR_A) == QPU_R_VPM ||
QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
QPU_RADDR_B) == QPU_R_VPM) {
qpu_serialize_one_inst(c, qpu_NOP());
}
/* thread end can't have uniform read */
if (QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
QPU_RADDR_A) == QPU_R_UNIF ||
QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
QPU_RADDR_B) == QPU_R_UNIF) {
qpu_serialize_one_inst(c, qpu_NOP());
}
/* thread end can't have TLB operations */
if (qpu_inst_is_tlb(c->qpu_insts[c->qpu_inst_count - 1]))
qpu_serialize_one_inst(c, qpu_NOP());
c->qpu_insts[c->qpu_inst_count - 1] =
qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1],
QPU_SIG_PROG_END);
qpu_serialize_one_inst(c, qpu_NOP());
qpu_serialize_one_inst(c, qpu_NOP());
switch (c->stage) {
case QSTAGE_VERT:
case QSTAGE_COORD:
break;
case QSTAGE_FRAG:
c->qpu_insts[c->qpu_inst_count - 1] =
qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1],
QPU_SIG_SCOREBOARD_UNLOCK);
break;
}
if (vc4_debug & VC4_DEBUG_QPU)
vc4_dump_program(c);
vc4_qpu_validate(c->qpu_insts, c->qpu_inst_count);
free(temp_registers);
}
/**
* Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
*
* The return value should be freed by the caller.
*/
struct qpu_reg *
vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
{
struct node_to_temp_map map[c->num_temps];
uint32_t temp_to_node[c->num_temps];
uint8_t class_bits[c->num_temps];
struct qpu_reg *temp_registers = calloc(c->num_temps,
sizeof(*temp_registers));
/* If things aren't ever written (undefined values), just read from
* r0.
*/
for (uint32_t i = 0; i < c->num_temps; i++)
temp_registers[i] = qpu_rn(0);
vc4_alloc_reg_set(vc4);
struct ra_graph *g = ra_alloc_interference_graph(vc4->regs,
c->num_temps);
/* Compute the live ranges so we can figure out interference. */
qir_calculate_live_intervals(c);
for (uint32_t i = 0; i < c->num_temps; i++) {
map[i].temp = i;
map[i].priority = c->temp_end[i] - c->temp_start[i];
}
qsort(map, c->num_temps, sizeof(map[0]), node_to_temp_priority);
for (uint32_t i = 0; i < c->num_temps; i++) {
temp_to_node[map[i].temp] = i;
}
/* Figure out our register classes and preallocated registers. We
* start with any temp being able to be in any file, then instructions
* incrementally remove bits that the temp definitely can't be in.
*/
memset(class_bits,
CLASS_BIT_A | CLASS_BIT_B_OR_ACC | CLASS_BIT_R4,
sizeof(class_bits));
int ip = 0;
qir_for_each_inst_inorder(inst, c) {
if (qir_writes_r4(inst)) {
/* This instruction writes r4 (and optionally moves
* its result to a temp), so nothing else can be
* stored in r4 across it.
*/
for (int i = 0; i < c->num_temps; i++) {
if (c->temp_start[i] < ip && c->temp_end[i] > ip)
class_bits[i] &= ~CLASS_BIT_R4;
}
} else {
/* R4 can't be written as a general purpose
* register. (it's TMU_NOSWAP as a write address).
*/
if (inst->dst.file == QFILE_TEMP)
class_bits[inst->dst.index] &= ~CLASS_BIT_R4;
}
switch (inst->op) {
case QOP_FRAG_Z:
ra_set_node_reg(g, temp_to_node[inst->dst.index],
AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2 + 1);
break;
case QOP_FRAG_W:
ra_set_node_reg(g, temp_to_node[inst->dst.index],
AB_INDEX + QPU_R_FRAG_PAYLOAD_ZW * 2);
break;
case QOP_ROT_MUL:
assert(inst->src[0].file == QFILE_TEMP);
class_bits[inst->src[0].index] &= ~CLASS_BIT_R0_R3;
break;
default:
break;
}
if (inst->dst.pack && !qir_is_mul(inst)) {
/* The non-MUL pack flags require an A-file dst
* register.
*/
class_bits[inst->dst.index] &= CLASS_BIT_A;
}
/* Apply restrictions for src unpacks. The integer unpacks
* can only be done from regfile A, while float unpacks can be
* either A or R4.
*/
for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
if (inst->src[i].file == QFILE_TEMP &&
inst->src[i].pack) {
if (qir_is_float_input(inst)) {
class_bits[inst->src[i].index] &=
CLASS_BIT_A | CLASS_BIT_R4;
} else {
class_bits[inst->src[i].index] &=
CLASS_BIT_A;
}
}
}
ip++;
}
for (uint32_t i = 0; i < c->num_temps; i++) {
int node = temp_to_node[i];
switch (class_bits[i]) {
case CLASS_BIT_A | CLASS_BIT_B_OR_ACC | CLASS_BIT_R4:
ra_set_node_class(g, node, vc4->reg_class_any);
break;
case CLASS_BIT_A | CLASS_BIT_B_OR_ACC:
ra_set_node_class(g, node, vc4->reg_class_a_or_b_or_acc);
break;
case CLASS_BIT_A | CLASS_BIT_R4:
ra_set_node_class(g, node, vc4->reg_class_r4_or_a);
break;
case CLASS_BIT_A:
ra_set_node_class(g, node, vc4->reg_class_a);
break;
case CLASS_BIT_R0_R3:
ra_set_node_class(g, node, vc4->reg_class_r0_r3);
break;
default:
fprintf(stderr, "temp %d: bad class bits: 0x%x\n",
i, class_bits[i]);
abort();
break;
}
}
for (uint32_t i = 0; i < c->num_temps; i++) {
for (uint32_t j = i + 1; j < c->num_temps; j++) {
if (!(c->temp_start[i] >= c->temp_end[j] ||
c->temp_start[j] >= c->temp_end[i])) {
ra_add_node_interference(g,
temp_to_node[i],
temp_to_node[j]);
}
}
}
bool ok = ra_allocate(g);
if (!ok) {
fprintf(stderr, "Failed to register allocate:\n");
qir_dump(c);
c->failed = true;
return NULL;
}
for (uint32_t i = 0; i < c->num_temps; i++) {
temp_registers[i] = vc4_regs[ra_get_node_reg(g, temp_to_node[i])];
/* If the value's never used, just write to the NOP register
* for clarity in debug output.
*/
if (c->temp_start[i] == c->temp_end[i])
temp_registers[i] = qpu_ra(QPU_W_NOP);
}
ralloc_free(g);
return temp_registers;
}