#include "vc4_context.h"
#include "vc4_qir.h"
#include "vc4_qpu.h"
+#include "util/ralloc.h"
static void
-vc4_dump_program(struct qcompile *c)
+vc4_dump_program(struct vc4_compile *c)
{
- fprintf(stderr, "%s:\n", qir_get_stage_name(c->stage));
+ fprintf(stderr, "%s prog %d/%d QPU:\n",
+ qir_get_stage_name(c->stage),
+ c->program_id, c->variant_id);
for (int i = 0; i < c->qpu_inst_count; i++) {
fprintf(stderr, "0x%016"PRIx64" ", c->qpu_insts[i]);
vc4_qpu_disasm(&c->qpu_insts[i], 1);
fprintf(stderr, "\n");
}
+ fprintf(stderr, "\n");
}
-struct queued_qpu_inst {
- struct simple_node link;
- uint64_t inst;
-};
-
static void
-queue(struct qcompile *c, uint64_t inst)
+queue(struct vc4_compile *c, uint64_t inst)
{
- struct queued_qpu_inst *q = calloc(1, sizeof(*q));
+ struct queued_qpu_inst *q = rzalloc(c, struct queued_qpu_inst);
q->inst = inst;
- insert_at_tail(&c->qpu_inst_list, &q->link);
+ list_addtail(&q->link, &c->qpu_inst_list);
}
static uint64_t *
-last_inst(struct qcompile *c)
+last_inst(struct vc4_compile *c)
{
struct queued_qpu_inst *q =
- (struct queued_qpu_inst *)last_elem(&c->qpu_inst_list);
+ (struct queued_qpu_inst *)c->qpu_inst_list.prev;
return &q->inst;
}
-/**
- * This is used to resolve the fact that we might register-allocate two
- * different operands of an instruction to the same physical register file
- * even though instructions have only one field for the register file source
- * address.
- *
- * In that case, we need to move one to a temporary that can be used in the
- * instruction, instead.
- */
static void
-fixup_raddr_conflict(struct qcompile *c,
- struct qpu_reg src0, struct qpu_reg *src1)
+set_last_cond_add(struct vc4_compile *c, uint32_t cond)
{
- if ((src0.mux == QPU_MUX_A || src0.mux == QPU_MUX_B) &&
- (src1->mux == QPU_MUX_A || src1->mux == QPU_MUX_B) &&
- src0.addr != src1->addr) {
- queue(c, qpu_inst(qpu_a_MOV(qpu_r3(), *src1),
- qpu_m_NOP()));
- *src1 = qpu_r3();
- }
+ *last_inst(c) = qpu_set_cond_add(*last_inst(c), cond);
}
static void
-serialize_one_inst(struct qcompile *c, uint64_t inst)
+set_last_cond_mul(struct vc4_compile *c, uint32_t cond)
{
- if (c->qpu_inst_count >= c->qpu_inst_size) {
- c->qpu_inst_size = MAX2(16, c->qpu_inst_size * 2);
- c->qpu_insts = realloc(c->qpu_insts,
- c->qpu_inst_size * sizeof(uint64_t));
+ *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 bool
+swap_file(struct qpu_reg *src)
+{
+ switch (src->addr) {
+ case QPU_R_UNIF:
+ case QPU_R_VARY:
+ if (src->mux == QPU_MUX_SMALL_IMM) {
+ return false;
+ } else {
+ if (src->mux == QPU_MUX_A)
+ src->mux = QPU_MUX_B;
+ else
+ src->mux = QPU_MUX_A;
+ return true;
+ }
+
+ default:
+ return false;
}
- c->qpu_insts[c->qpu_inst_count++] = inst;
}
+/**
+ * This is used to resolve the fact that we might register-allocate two
+ * different operands of an instruction to the same physical register file
+ * even though instructions have only one field for the register file source
+ * address.
+ *
+ * In that case, we need to move one to a temporary that can be used in the
+ * instruction, instead. We reserve ra31/rb31 for this purpose.
+ */
static void
-serialize_insts(struct qcompile *c)
+fixup_raddr_conflict(struct vc4_compile *c,
+ struct qpu_reg dst,
+ struct qpu_reg *src0, struct qpu_reg *src1,
+ struct qinst *inst, uint64_t *unpack)
{
- int last_sfu_write = -10;
- bool scoreboard_wait_emitted = false;
-
- while (!is_empty_list(&c->qpu_inst_list)) {
- struct queued_qpu_inst *q =
- (struct queued_qpu_inst *)first_elem(&c->qpu_inst_list);
- uint32_t last_waddr_a = QPU_W_NOP, last_waddr_b = QPU_W_NOP;
- uint32_t raddr_a = QPU_GET_FIELD(q->inst, QPU_RADDR_A);
- uint32_t raddr_b = QPU_GET_FIELD(q->inst, QPU_RADDR_B);
-
- if (c->qpu_inst_count > 0) {
- uint64_t last_inst = c->qpu_insts[c->qpu_inst_count -
- 1];
- uint32_t last_waddr_add = QPU_GET_FIELD(last_inst,
- QPU_WADDR_ADD);
- uint32_t last_waddr_mul = QPU_GET_FIELD(last_inst,
- QPU_WADDR_MUL);
-
- if (last_inst & QPU_WS) {
- last_waddr_a = last_waddr_mul;
- last_waddr_b = last_waddr_add;
- } else {
- last_waddr_a = last_waddr_add;
- last_waddr_b = last_waddr_mul;
- }
- }
+ 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;
+
+ if (mux0 <= QPU_MUX_R5 ||
+ mux0 != mux1 ||
+ (src0->addr == src1->addr &&
+ src0->mux == src1->mux)) {
+ return;
+ }
- uint32_t src_muxes[] = {
- QPU_GET_FIELD(q->inst, QPU_ADD_A),
- QPU_GET_FIELD(q->inst, QPU_ADD_B),
- QPU_GET_FIELD(q->inst, QPU_MUL_A),
- QPU_GET_FIELD(q->inst, QPU_MUL_B),
- };
+ if (swap_file(src0) || swap_file(src1))
+ return;
- /* "An instruction must not read from a location in physical
- * regfile A or B that was written to by the previous
- * instruction."
+ if (mux0 == QPU_MUX_A) {
+ /* Make sure we use the same type of MOV as the instruction,
+ * in case of unpacks.
*/
- bool needs_raddr_vs_waddr_nop = false;
- bool reads_r4 = false;
- for (int i = 0; i < ARRAY_SIZE(src_muxes); i++) {
- if ((raddr_a < 32 &&
- src_muxes[i] == QPU_MUX_A &&
- last_waddr_a == raddr_a) ||
- (raddr_b < 32 &&
- src_muxes[i] == QPU_MUX_B &&
- last_waddr_b == raddr_b)) {
- needs_raddr_vs_waddr_nop = true;
- }
- if (src_muxes[i] == QPU_MUX_R4)
- reads_r4 = true;
- }
+ 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 (needs_raddr_vs_waddr_nop) {
- serialize_one_inst(c, qpu_inst(qpu_a_NOP(),
- qpu_m_NOP()));
- }
-
- /* "After an SFU lookup instruction, accumulator r4 must not
- * be read in the following two instructions. Any other
- * instruction that results in r4 being written (that is, TMU
- * read, TLB read, SFU lookup) cannot occur in the two
- * instructions following an SFU lookup."
+ /* 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 (reads_r4) {
- while (c->qpu_inst_count - last_sfu_write < 3) {
- serialize_one_inst(c, qpu_inst(qpu_a_NOP(),
- qpu_m_NOP()));
- }
+ 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));
+ *src0 = qpu_ra(31);
+ }
+}
- uint32_t waddr_a = QPU_GET_FIELD(q->inst, QPU_WADDR_ADD);
- uint32_t waddr_m = QPU_GET_FIELD(q->inst, QPU_WADDR_MUL);
- if ((waddr_a >= QPU_W_SFU_RECIP && waddr_a <= QPU_W_SFU_LOG) ||
- (waddr_m >= QPU_W_SFU_RECIP && waddr_m <= QPU_W_SFU_LOG)) {
- last_sfu_write = c->qpu_inst_count;
- }
+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);
- /* "A scoreboard wait must not occur in the first two
- * instructions of a fragment shader. This is either the
- * explicit Wait for Scoreboard signal or an implicit wait
- * with the first tile-buffer read or write instruction."
- */
- if (!scoreboard_wait_emitted &&
- (waddr_a == QPU_W_TLB_Z || waddr_m == QPU_W_TLB_Z ||
- waddr_a == QPU_W_TLB_COLOR_MS ||
- waddr_m == QPU_W_TLB_COLOR_MS ||
- waddr_a == QPU_W_TLB_COLOR_ALL ||
- waddr_m == QPU_W_TLB_COLOR_ALL ||
- QPU_GET_FIELD(q->inst, QPU_SIG) == QPU_SIG_COLOR_LOAD)) {
- while (c->qpu_inst_count < 3 ||
- QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
- QPU_SIG) != QPU_SIG_NONE) {
- serialize_one_inst(c, qpu_inst(qpu_a_NOP(),
- qpu_m_NOP()));
- }
- c->qpu_insts[c->qpu_inst_count - 1] =
- qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1],
- QPU_SIG_WAIT_FOR_SCOREBOARD);
- scoreboard_wait_emitted = true;
- }
+ if (!inst->dst.pack)
+ return;
- serialize_one_inst(c, q->inst);
+ *last_inst(c) |= QPU_SET_FIELD(inst->dst.pack, QPU_PACK);
- remove_from_list(&q->link);
- free(q);
+ 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 qcompile *c)
+vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c)
{
- struct qpu_reg allocate_to_qpu_reg[3 + 32 + 32];
- bool reg_in_use[ARRAY_SIZE(allocate_to_qpu_reg)];
- int *reg_allocated = calloc(c->num_temps, sizeof(*reg_allocated));
- int *reg_uses_remaining =
- calloc(c->num_temps, sizeof(*reg_uses_remaining));
- bool discard = false;
-
- for (int i = 0; i < ARRAY_SIZE(reg_in_use); i++)
- reg_in_use[i] = false;
- for (int i = 0; i < c->num_temps; i++)
- reg_allocated[i] = -1;
- for (int i = 0; i < 3; i++)
- allocate_to_qpu_reg[i] = qpu_rn(i);
- for (int i = 0; i < 32; i++)
- allocate_to_qpu_reg[i + 3] = qpu_ra(i);
- for (int i = 0; i < 32; i++)
- allocate_to_qpu_reg[i + 3 + 32] = qpu_rb(i);
-
- make_empty_list(&c->qpu_inst_list);
-
- struct simple_node *node;
- foreach(node, &c->instructions) {
- struct qinst *qinst = (struct qinst *)node;
-
- if (qinst->dst.file == QFILE_TEMP)
- reg_uses_remaining[qinst->dst.index]++;
- for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
- if (qinst->src[i].file == QFILE_TEMP)
- reg_uses_remaining[qinst->src[i].index]++;
- }
- if (qinst->op == QOP_TLB_PASSTHROUGH_Z_WRITE ||
- qinst->op == QOP_FRAG_Z)
- reg_in_use[3 + 32 + QPU_R_FRAG_PAYLOAD_ZW] = true;
- }
+ struct qpu_reg *temp_registers = vc4_register_allocate(vc4, c);
+ 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;
+
+ list_inithead(&c->qpu_inst_list);
switch (c->stage) {
case QSTAGE_VERT:
case QSTAGE_COORD:
- queue(c, qpu_load_imm_ui(qpu_vrsetup(),
- (0x00001a00 +
- 0x00100000 * c->num_inputs)));
+ /* 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;
}
- foreach(node, &c->instructions) {
- struct qinst *qinst = (struct qinst *)node;
-
+ list_for_each_entry(struct qinst, qinst, &c->instructions, link) {
#if 0
fprintf(stderr, "translating qinst to qpu: ");
qir_dump_inst(qinst);
static const struct {
uint32_t op;
- bool is_mul;
} translate[] = {
-#define A(name) [QOP_##name] = {QPU_A_##name, false}
-#define M(name) [QOP_##name] = {QPU_M_##name, true}
+#define A(name) [QOP_##name] = {QPU_A_##name}
+#define M(name) [QOP_##name] = {QPU_M_##name}
A(FADD),
A(FSUB),
A(FMIN),
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(V8MULD),
+ M(V8MIN),
+ M(V8MAX),
+ M(V8ADDS),
+ M(V8SUBS),
+ M(MUL24),
+
+ /* If we replicate src[0] out to src[1], this works
+ * out the same as a MOV.
+ */
+ [QOP_MOV] = { QPU_A_OR },
+ [QOP_FMOV] = { QPU_A_FMAX },
+ [QOP_MMOV] = { QPU_M_V8MIN },
};
- static const uint32_t compareflags[] = {
- [QOP_SEQ - QOP_SEQ] = QPU_COND_ZS,
- [QOP_SNE - QOP_SEQ] = QPU_COND_ZC,
- [QOP_SLT - QOP_SEQ] = QPU_COND_NS,
- [QOP_SGE - QOP_SEQ] = QPU_COND_NC,
- };
-
+ 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;
src[i] = qpu_rn(0);
break;
case QFILE_TEMP:
- if (reg_allocated[index] == -1) {
- fprintf(stderr, "undefined reg use: ");
- qir_dump_inst(qinst);
- fprintf(stderr, "\n");
-
- src[i] = qpu_rn(0);
- } else {
- src[i] = allocate_to_qpu_reg[reg_allocated[index]];
- reg_uses_remaining[index]--;
- if (reg_uses_remaining[index] == 0)
- reg_in_use[reg_allocated[index]] = false;
+ 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:
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;
+
+ 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_NULL:
dst = qpu_ra(QPU_W_NOP);
break;
-
case QFILE_TEMP:
- if (reg_allocated[qinst->dst.index] == -1) {
- int alloc;
- for (alloc = 0;
- alloc < ARRAY_SIZE(reg_in_use);
- alloc++) {
- /* The pack flags require an A-file register. */
- if (qinst->op == QOP_PACK_SCALED &&
- allocate_to_qpu_reg[alloc].mux != QPU_MUX_A) {
- continue;
- }
-
- if (!reg_in_use[alloc])
- break;
- }
- assert(alloc != ARRAY_SIZE(reg_in_use) && "need better reg alloc");
- reg_in_use[alloc] = true;
- reg_allocated[qinst->dst.index] = alloc;
- }
-
- dst = allocate_to_qpu_reg[reg_allocated[qinst->dst.index]];
-
- reg_uses_remaining[qinst->dst.index]--;
- if (reg_uses_remaining[qinst->dst.index] == 0) {
- reg_in_use[reg_allocated[qinst->dst.index]] =
- false;
- }
+ dst = temp_registers[qinst->dst.index];
break;
-
- case QFILE_VARY:
- case QFILE_UNIF:
- assert(!"not reached");
+ case QFILE_VPM:
+ dst = qpu_ra(QPU_W_VPM);
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_inst(qpu_a_MOV(dst, src[0]),
- qpu_m_NOP()));
- }
+ case QFILE_TLB_COLOR_WRITE:
+ dst = qpu_tlbc();
break;
- case QOP_CMP:
- queue(c, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_NOP),
- src[0]),
- qpu_m_NOP()));
- *last_inst(c) |= QPU_SF;
-
- if (dst.mux <= QPU_MUX_R3) {
- fixup_raddr_conflict(c, src[1], &src[2]);
- queue(c, qpu_inst(qpu_a_MOV(dst, src[1]),
- qpu_m_MOV(dst, src[2])));
- *last_inst(c) = qpu_set_cond_add(*last_inst(c),
- QPU_COND_NS);
- *last_inst(c) = qpu_set_cond_mul(*last_inst(c),
- QPU_COND_NC);
- } else {
- if (dst.mux == src[1].mux &&
- dst.addr == src[1].addr) {
- queue(c, qpu_inst(qpu_a_MOV(dst, src[1]),
- qpu_m_NOP()));
-
- queue(c, qpu_inst(qpu_a_MOV(dst, src[2]),
- qpu_m_NOP()));
- *last_inst(c) = qpu_set_cond_add(*last_inst(c),
- QPU_COND_NC);
- } else {
- queue(c, qpu_inst(qpu_a_MOV(dst, src[2]),
- qpu_m_NOP()));
-
- queue(c, qpu_inst(qpu_a_MOV(dst, src[1]),
- qpu_m_NOP()));
- *last_inst(c) = qpu_set_cond_add(*last_inst(c),
- QPU_COND_NS);
- }
- }
+ case QFILE_TLB_COLOR_WRITE_MS:
+ dst = qpu_tlbc_ms();
break;
- case QOP_SEQ:
- case QOP_SNE:
- case QOP_SGE:
- case QOP_SLT:
- fixup_raddr_conflict(c, src[0], &src[1]);
- queue(c, qpu_inst(qpu_a_FSUB(qpu_ra(QPU_W_NOP),
- src[0], src[1]),
- qpu_m_NOP()));
- *last_inst(c) |= QPU_SF;
-
- queue(c, qpu_load_imm_f(dst, 0.0));
- queue(c, qpu_load_imm_f(dst, 1.0));
- *last_inst(c) = qpu_set_cond_add(*last_inst(c),
- compareflags[qinst->op - QOP_SEQ]);
-
-
+ case QFILE_TLB_Z_WRITE:
+ dst = qpu_ra(QPU_W_TLB_Z);
break;
- case QOP_VPM_WRITE:
- queue(c, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_VPM), src[0]),
- qpu_m_NOP()));
+ case QFILE_TLB_STENCIL_SETUP:
+ dst = qpu_ra(QPU_W_TLB_STENCIL_SETUP);
break;
- case QOP_VPM_READ:
- queue(c, qpu_inst(qpu_a_MOV(dst, qpu_ra(QPU_R_VPM)),
- qpu_m_NOP()));
+ case QFILE_VARY:
+ case QFILE_UNIF:
+ case QFILE_SMALL_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:
case QOP_LOG2:
switch (qinst->op) {
case QOP_RCP:
- queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP),
- src[0]),
- qpu_m_NOP()));
+ queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP),
+ src[0]) | unpack);
break;
case QOP_RSQ:
- queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT),
- src[0]),
- qpu_m_NOP()));
+ queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT),
+ src[0]) | unpack);
break;
case QOP_EXP2:
- queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP),
- src[0]),
- qpu_m_NOP()));
+ queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP),
+ src[0]) | unpack);
break;
case QOP_LOG2:
- queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG),
- src[0]),
- qpu_m_NOP()));
+ queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG),
+ src[0]) | unpack);
break;
default:
abort();
}
- queue(c, qpu_inst(qpu_a_MOV(dst, qpu_r4()),
- qpu_m_NOP()));
-
- break;
-
- case QOP_PACK_COLORS:
- for (int i = 0; i < 4; i++) {
- queue(c, qpu_inst(qpu_a_NOP(),
- qpu_m_MOV(qpu_r3(), src[i])));
- *last_inst(c) |= QPU_PM;
- *last_inst(c) |= QPU_SET_FIELD(QPU_PACK_MUL_8A + i,
- QPU_PACK);
- }
-
- queue(c, qpu_inst(qpu_a_MOV(dst, qpu_r3()),
- qpu_m_NOP()));
-
- break;
+ handle_r4_qpu_write(c, qinst, dst);
- case QOP_FRAG_X:
- queue(c, qpu_inst(qpu_a_ITOF(dst,
- qpu_ra(QPU_R_XY_PIXEL_COORD)),
- qpu_m_NOP()));
break;
- case QOP_FRAG_Y:
- queue(c, qpu_inst(qpu_a_ITOF(dst,
- qpu_rb(QPU_R_XY_PIXEL_COORD)),
- qpu_m_NOP()));
+ 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:
- queue(c, qpu_inst(qpu_a_ITOF(dst,
- qpu_rb(QPU_R_FRAG_PAYLOAD_ZW)),
- qpu_m_NOP()));
- break;
-
- case QOP_FRAG_RCP_W:
- queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP),
- qpu_ra(QPU_R_FRAG_PAYLOAD_ZW)),
- qpu_m_NOP()));
-
- queue(c, qpu_inst(qpu_a_MOV(dst, qpu_r4()),
- qpu_m_NOP()));
- break;
-
- case QOP_TLB_DISCARD_SETUP:
- discard = true;
- queue(c, qpu_inst(qpu_a_MOV(src[0], src[0]),
- qpu_m_NOP()));
- *last_inst(c) |= QPU_SF;
- break;
-
- case QOP_TLB_PASSTHROUGH_Z_WRITE:
- queue(c, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_TLB_Z),
- qpu_rb(QPU_R_FRAG_PAYLOAD_ZW)),
- qpu_m_NOP()));
- if (discard) {
- *last_inst(c) = qpu_set_cond_add(*last_inst(c),
- QPU_COND_ZS);
- }
+ 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_COLOR_READ:
- queue(c, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
+ queue(c, qpu_NOP());
*last_inst(c) = qpu_set_sig(*last_inst(c),
QPU_SIG_COLOR_LOAD);
-
- break;
-
- case QOP_TLB_COLOR_WRITE:
- queue(c, qpu_inst(qpu_a_MOV(qpu_tlbc(),
- src[0]),
- qpu_m_NOP()));
- if (discard) {
- *last_inst(c) = qpu_set_cond_add(*last_inst(c),
- QPU_COND_ZS);
- }
+ handle_r4_qpu_write(c, qinst, dst);
break;
case QOP_VARY_ADD_C:
- queue(c, qpu_inst(qpu_a_FADD(dst,
- src[0], qpu_r5()),
- qpu_m_NOP()));
- break;
-
- case QOP_PACK_SCALED: {
- uint64_t a = (qpu_inst(qpu_a_MOV(dst, src[0]),
- qpu_m_NOP()) |
- QPU_SET_FIELD(QPU_PACK_A_16A,
- QPU_PACK));
- uint64_t b = (qpu_inst(qpu_a_MOV(dst, src[1]),
- qpu_m_NOP()) |
- QPU_SET_FIELD(QPU_PACK_A_16B,
- QPU_PACK));
-
- if (dst.mux == src[1].mux && dst.addr == src[1].addr) {
- queue(c, b);
- queue(c, a);
- } else {
- queue(c, a);
- queue(c, b);
- }
+ queue(c, qpu_a_FADD(dst, src[0], qpu_r5()) | unpack);
break;
- }
case QOP_TEX_S:
case QOP_TEX_T:
case QOP_TEX_R:
case QOP_TEX_B:
- queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_TMU0_S +
- (qinst->op -
- QOP_TEX_S)),
- src[0]),
- qpu_m_NOP()));
+ queue(c, qpu_a_MOV(qpu_rb(QPU_W_TMU0_S +
+ (qinst->op - QOP_TEX_S)),
+ src[0]) | unpack);
+ break;
+
+ case QOP_TEX_DIRECT:
+ 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_inst(qpu_a_NOP(), qpu_m_NOP()));
+ queue(c, qpu_NOP());
*last_inst(c) = qpu_set_sig(*last_inst(c),
QPU_SIG_LOAD_TMU0);
-
- break;
-
- case QOP_R4_UNPACK_A:
- case QOP_R4_UNPACK_B:
- case QOP_R4_UNPACK_C:
- case QOP_R4_UNPACK_D:
- queue(c, qpu_inst(qpu_a_MOV(dst, qpu_r4()),
- qpu_m_NOP()));
- *last_inst(c) |= QPU_PM;
- *last_inst(c) |= QPU_SET_FIELD(QPU_UNPACK_R4_8A +
- (qinst->op -
- QOP_R4_UNPACK_A),
- QPU_UNPACK);
-
+ handle_r4_qpu_write(c, qinst, dst);
break;
default:
assert(qinst->op < ARRAY_SIZE(translate));
assert(translate[qinst->op].op != 0); /* NOPs */
+ /* Skip emitting the MOV if it's a no-op. */
+ if (qir_is_raw_mov(qinst) &&
+ dst.mux == src[0].mux && dst.addr == src[0].addr) {
+ break;
+ }
+
/* 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, src[0], &src[1]);
+ fixup_raddr_conflict(c, dst, &src[0], &src[1],
+ qinst, &unpack);
- if (translate[qinst->op].is_mul) {
- queue(c, qpu_inst(qpu_a_NOP(),
- qpu_m_alu2(translate[qinst->op].op,
- dst,
- src[0], src[1])));
+ if (qir_is_mul(qinst)) {
+ queue(c, qpu_m_alu2(translate[qinst->op].op,
+ dst,
+ src[0], src[1]) | unpack);
+ set_last_cond_mul(c, qinst->cond);
} else {
- queue(c, qpu_inst(qpu_a_alu2(translate[qinst->op].op,
- dst,
- src[0], src[1]),
- qpu_m_NOP()));
+ queue(c, qpu_a_alu2(translate[qinst->op].op,
+ dst,
+ src[0], src[1]) | unpack);
+ set_last_cond_add(c, qinst->cond);
}
+ handled_qinst_cond = true;
+ set_last_dst_pack(c, qinst);
+
break;
}
+
+ assert(qinst->cond == QPU_COND_ALWAYS ||
+ handled_qinst_cond);
+
+ if (qinst->sf)
+ *last_inst(c) |= QPU_SF;
}
- serialize_insts(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 */
+ /* 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) {
- serialize_one_inst(c, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
+ 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);
- serialize_one_inst(c, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
- serialize_one_inst(c, qpu_inst(qpu_a_NOP(), qpu_m_NOP()));
+ qpu_serialize_one_inst(c, qpu_NOP());
+ qpu_serialize_one_inst(c, qpu_NOP());
switch (c->stage) {
case QSTAGE_VERT:
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);
vc4_qpu_validate(c->qpu_insts, c->qpu_inst_count);
+
+ free(temp_registers);
}