vc4_qpu_disasm(&c->qpu_insts[i], 1);
fprintf(stderr, "\n");
}
+ fprintf(stderr, "\n");
}
static void
*last_inst(c) = qpu_set_cond_add(*last_inst(c), cond);
}
+static void
+set_last_cond_mul(struct vc4_compile *c, uint32_t cond)
+{
+ *last_inst(c) = qpu_set_cond_mul(*last_inst(c), cond);
+}
+
/**
* Some special registers can be read from either file, which lets us resolve
* raddr conflicts without extra MOVs.
static void
fixup_raddr_conflict(struct vc4_compile *c,
struct qpu_reg dst,
- struct qpu_reg *src0, struct qpu_reg *src1)
+ struct qpu_reg *src0, struct qpu_reg *src1,
+ struct qinst *inst, uint64_t *unpack)
{
uint32_t mux0 = src0->mux == QPU_MUX_SMALL_IMM ? QPU_MUX_B : src0->mux;
uint32_t mux1 = src1->mux == QPU_MUX_SMALL_IMM ? QPU_MUX_B : src1->mux;
return;
if (mux0 == QPU_MUX_A) {
- queue(c, qpu_a_MOV(qpu_rb(31), *src0));
+ /* Make sure we use the same type of MOV as the instruction,
+ * in case of unpacks.
+ */
+ if (qir_is_float_input(inst))
+ queue(c, qpu_a_FMAX(qpu_rb(31), *src0, *src0));
+ else
+ queue(c, qpu_a_MOV(qpu_rb(31), *src0));
+
+ /* If we had an unpack on this A-file source, we need to put
+ * it into this MOV, not into the later move from regfile B.
+ */
+ if (inst->src[0].pack) {
+ *last_inst(c) |= *unpack;
+ *unpack = 0;
+ }
*src0 = qpu_rb(31);
} else {
queue(c, qpu_a_MOV(qpu_ra(31), *src0));
}
}
+static void
+set_last_dst_pack(struct vc4_compile *c, struct qinst *inst)
+{
+ bool had_pm = *last_inst(c) & QPU_PM;
+ bool had_ws = *last_inst(c) & QPU_WS;
+ uint32_t unpack = QPU_GET_FIELD(*last_inst(c), QPU_UNPACK);
+
+ if (!inst->dst.pack)
+ return;
+
+ *last_inst(c) |= QPU_SET_FIELD(inst->dst.pack, QPU_PACK);
+
+ if (qir_is_mul(inst)) {
+ assert(!unpack || had_pm);
+ *last_inst(c) |= QPU_PM;
+ } else {
+ assert(!unpack || !had_pm);
+ assert(!had_ws); /* dst must be a-file to pack. */
+ }
+}
+
+static void
+handle_r4_qpu_write(struct vc4_compile *c, struct qinst *qinst,
+ struct qpu_reg dst)
+{
+ if (dst.mux != QPU_MUX_R4)
+ queue(c, qpu_a_MOV(dst, qpu_r4()));
+ else if (qinst->sf)
+ queue(c, qpu_a_MOV(qpu_ra(QPU_W_NOP), qpu_r4()));
+}
+
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,
- QPU_UNPACK_16B,
- };
list_inithead(&c->qpu_inst_list);
break;
}
- list_for_each_entry(struct qinst, qinst, &c->instructions, link) {
+ qir_for_each_inst_inorder(qinst, c) {
#if 0
fprintf(stderr, "translating qinst to qpu: ");
qir_dump_inst(qinst);
* out the same as a MOV.
*/
[QOP_MOV] = { QPU_A_OR },
+ [QOP_FMOV] = { QPU_A_FMAX },
+ [QOP_MMOV] = { QPU_M_V8MIN },
};
+ uint64_t unpack = 0;
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:
+ case QFILE_LOAD_IMM:
src[i] = qpu_rn(0);
break;
case QFILE_TEMP:
src[i] = temp_registers[index];
+ if (qinst->src[i].pack) {
+ assert(!unpack ||
+ unpack == qinst->src[i].pack);
+ unpack = QPU_SET_FIELD(qinst->src[i].pack,
+ QPU_UNPACK);
+ if (src[i].mux == QPU_MUX_R4)
+ unpack |= QPU_PM;
+ }
break;
case QFILE_UNIF:
src[i] = qpu_unif();
last_vpm_read_index = qinst->src[i].index;
src[i] = qpu_ra(QPU_R_VPM);
break;
+
+ case QFILE_FRAG_X:
+ src[i] = qpu_ra(QPU_R_XY_PIXEL_COORD);
+ break;
+ case QFILE_FRAG_Y:
+ src[i] = qpu_rb(QPU_R_XY_PIXEL_COORD);
+ break;
+ case QFILE_FRAG_REV_FLAG:
+ src[i] = qpu_rb(QPU_R_MS_REV_FLAGS);
+ break;
+
+ case QFILE_TLB_COLOR_WRITE:
+ case QFILE_TLB_COLOR_WRITE_MS:
+ case QFILE_TLB_Z_WRITE:
+ case QFILE_TLB_STENCIL_SETUP:
+ unreachable("bad qir src file");
}
}
case QFILE_VPM:
dst = qpu_ra(QPU_W_VPM);
break;
- case QFILE_VARY:
- case QFILE_UNIF:
- case QFILE_SMALL_IMM:
- assert(!"not reached");
+
+ case QFILE_TLB_COLOR_WRITE:
+ dst = qpu_tlbc();
break;
- }
- switch (qinst->op) {
- case QOP_SEL_X_0_ZS:
- case QOP_SEL_X_0_ZC:
- case QOP_SEL_X_0_NS:
- case QOP_SEL_X_0_NC:
- 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);
+ case QFILE_TLB_COLOR_WRITE_MS:
+ dst = qpu_tlbc_ms();
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:
- queue(c, qpu_a_MOV(dst, src[0]));
- set_last_cond_add(c, qinst->op - QOP_SEL_X_Y_ZS +
- QPU_COND_ZS);
+ case QFILE_TLB_Z_WRITE:
+ dst = qpu_ra(QPU_W_TLB_Z);
+ break;
- queue(c, qpu_a_MOV(dst, src[1]));
- set_last_cond_add(c, ((qinst->op - QOP_SEL_X_Y_ZS) ^
- 1) + QPU_COND_ZS);
+ case QFILE_TLB_STENCIL_SETUP:
+ dst = qpu_ra(QPU_W_TLB_STENCIL_SETUP);
+ break;
+ case QFILE_VARY:
+ case QFILE_UNIF:
+ case QFILE_SMALL_IMM:
+ case QFILE_LOAD_IMM:
+ case QFILE_FRAG_X:
+ case QFILE_FRAG_Y:
+ case QFILE_FRAG_REV_FLAG:
+ assert(!"not reached");
break;
+ }
+
+ bool handled_qinst_cond = false;
+ switch (qinst->op) {
case QOP_RCP:
case QOP_RSQ:
case QOP_EXP2:
switch (qinst->op) {
case QOP_RCP:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP),
- src[0]));
+ src[0]) | unpack);
break;
case QOP_RSQ:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT),
- src[0]));
+ src[0]) | unpack);
break;
case QOP_EXP2:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP),
- src[0]));
+ src[0]) | unpack);
break;
case QOP_LOG2:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG),
- src[0]));
+ src[0]) | unpack);
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;
+ handle_r4_qpu_write(c, qinst, dst);
- 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)));
+ case QOP_LOAD_IMM:
+ assert(qinst->src[0].file == QFILE_LOAD_IMM);
+ queue(c, qpu_load_imm_ui(dst, qinst->src[0].index));
break;
- case QOP_FRAG_REV_FLAG:
- queue(c, qpu_a_ITOF(dst,
- qpu_rb(QPU_R_MS_REV_FLAGS)));
+ case QOP_MS_MASK:
+ src[1] = qpu_ra(QPU_R_MS_REV_FLAGS);
+ fixup_raddr_conflict(c, dst, &src[0], &src[1],
+ qinst, &unpack);
+ queue(c, qpu_a_AND(qpu_ra(QPU_W_MS_FLAGS),
+ src[0], src[1]) | unpack);
break;
case QOP_FRAG_Z:
*/
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);
- }
+ handle_r4_qpu_write(c, qinst, dst);
break;
case QOP_VARY_ADD_C:
- queue(c, qpu_a_FADD(dst, src[0], qpu_r5()));
+ queue(c, qpu_a_FADD(dst, src[0], qpu_r5()) | unpack);
break;
case QOP_TEX_S:
case QOP_TEX_B:
queue(c, qpu_a_MOV(qpu_rb(QPU_W_TMU0_S +
(qinst->op - QOP_TEX_S)),
- src[0]));
+ src[0]) | unpack);
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]));
+ fixup_raddr_conflict(c, dst, &src[0], &src[1],
+ qinst, &unpack);
+ queue(c, qpu_a_ADD(qpu_rb(QPU_W_TMU0_S),
+ src[0], src[1]) | unpack);
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);
-
- queue(c, qpu_a_FMAX(dst, src[0], src[0]));
- *last_inst(c) |=
- QPU_SET_FIELD(unpack_map[qinst->op -
- QOP_UNPACK_8A_F],
- QPU_UNPACK);
- }
+ handle_r4_qpu_write(c, qinst, dst);
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);
-
- queue(c, qpu_a_MOV(dst, src[0]));
- *last_inst(c) |= QPU_SET_FIELD(unpack_map[qinst->op -
- QOP_UNPACK_8A_I],
- QPU_UNPACK);
+ case QOP_BRANCH:
+ /* The branch target will be updated at QPU scheduling
+ * time.
+ */
+ queue(c, (qpu_branch(qinst->cond, 0) |
+ QPU_BRANCH_REL));
+ handled_qinst_cond = true;
break;
default:
if (qir_get_op_nsrc(qinst->op) == 1)
src[1] = src[0];
- fixup_raddr_conflict(c, dst, &src[0], &src[1]);
+ fixup_raddr_conflict(c, dst, &src[0], &src[1],
+ qinst, &unpack);
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);
- }
+ src[0], src[1]) | unpack);
+ set_last_cond_mul(c, qinst->cond);
} 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);
- }
+ src[0], src[1]) | unpack);
+ set_last_cond_add(c, qinst->cond);
}
+ handled_qinst_cond = true;
+ set_last_dst_pack(c, qinst);
break;
}
- if (qinst->sf) {
- assert(!qir_is_multi_instruction(qinst));
+ assert(qinst->cond == QPU_COND_ALWAYS ||
+ handled_qinst_cond);
+
+ if (qinst->sf)
*last_inst(c) |= QPU_SF;
- }
}
- qpu_schedule_instructions(c);
+ uint32_t cycles = qpu_schedule_instructions(c);
+ uint32_t inst_count_at_schedule_time = c->qpu_inst_count;
/* thread end can't have VPM write or read */
if (QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
if (qpu_inst_is_tlb(c->qpu_insts[c->qpu_inst_count - 1]))
qpu_serialize_one_inst(c, qpu_NOP());
+ /* Make sure there's no existing signal set (like for a small
+ * immediate)
+ */
+ if (QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
+ QPU_SIG) != QPU_SIG_NONE) {
+ 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);
break;
}
+ cycles += c->qpu_inst_count - inst_count_at_schedule_time;
+
+ if (vc4_debug & VC4_DEBUG_SHADERDB) {
+ fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d estimated cycles\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id,
+ cycles);
+ }
+
if (vc4_debug & VC4_DEBUG_QPU)
vc4_dump_program(c);