From eea1d36915cb97ee1a6eb6aeaf15dd5689f03148 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Fri, 4 Jul 2014 11:51:31 -0700 Subject: [PATCH] vc4: Make scheduling of NOPs a separate step from QIR -> QPU translation. This should also be used as a way to pair QIR instructions into QPU instructions later. --- src/gallium/drivers/vc4/vc4_program.c | 8 +- src/gallium/drivers/vc4/vc4_qir.h | 4 +- src/gallium/drivers/vc4/vc4_qpu_emit.c | 290 +++++++++++++++++-------- 3 files changed, 212 insertions(+), 90 deletions(-) diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c index 12d188185cd..0e415a8baf6 100644 --- a/src/gallium/drivers/vc4/vc4_program.c +++ b/src/gallium/drivers/vc4/vc4_program.c @@ -600,7 +600,7 @@ vc4_shader_tgsi_to_qir(struct vc4_compiled_shader *shader, enum qstage stage, if (vc4_debug & VC4_DEBUG_SHADERDB) { fprintf(stderr, "SHADER-DB: %s: %d instructions\n", - qir_get_stage_name(c->stage), c->num_qpu_insts); + qir_get_stage_name(c->stage), c->qpu_inst_count); fprintf(stderr, "SHADER-DB: %s: %d uniforms\n", qir_get_stage_name(c->stage), trans->num_uniforms); } @@ -647,7 +647,7 @@ vc4_fs_compile(struct vc4_context *vc4, struct vc4_compiled_shader *shader, shader->num_inputs = trans->num_inputs; copy_uniform_state_to_shader(shader, 0, trans); shader->bo = vc4_bo_alloc_mem(vc4->screen, trans->c->qpu_insts, - trans->c->num_qpu_insts * sizeof(uint64_t), + trans->c->qpu_inst_count * sizeof(uint64_t), "fs_code"); qir_compile_destroy(trans->c); @@ -668,8 +668,8 @@ vc4_vs_compile(struct vc4_context *vc4, struct vc4_compiled_shader *shader, &key->base); copy_uniform_state_to_shader(shader, 1, cs_trans); - uint32_t vs_size = vs_trans->c->num_qpu_insts * sizeof(uint64_t); - uint32_t cs_size = cs_trans->c->num_qpu_insts * sizeof(uint64_t); + uint32_t vs_size = vs_trans->c->qpu_inst_count * sizeof(uint64_t); + uint32_t cs_size = cs_trans->c->qpu_inst_count * sizeof(uint64_t); shader->coord_shader_offset = vs_size; /* XXX: alignment? */ shader->bo = vc4_bo_alloc(vc4->screen, shader->coord_shader_offset + cs_size, diff --git a/src/gallium/drivers/vc4/vc4_qir.h b/src/gallium/drivers/vc4/vc4_qir.h index 75b1933e6a6..cd7133cdc88 100644 --- a/src/gallium/drivers/vc4/vc4_qir.h +++ b/src/gallium/drivers/vc4/vc4_qir.h @@ -126,8 +126,10 @@ struct qcompile { struct simple_node instructions; uint32_t immediates[1024]; + struct simple_node qpu_inst_list; uint64_t *qpu_insts; - uint32_t num_qpu_insts; + uint32_t qpu_inst_count; + uint32_t qpu_inst_size; }; struct qcompile *qir_compile_init(void); diff --git a/src/gallium/drivers/vc4/vc4_qpu_emit.c b/src/gallium/drivers/vc4/vc4_qpu_emit.c index f02fe822b07..b8a506eb174 100644 --- a/src/gallium/drivers/vc4/vc4_qpu_emit.c +++ b/src/gallium/drivers/vc4/vc4_qpu_emit.c @@ -33,13 +33,34 @@ vc4_dump_program(struct qcompile *c) { fprintf(stderr, "%s:\n", qir_get_stage_name(c->stage)); - for (int i = 0; i < c->num_qpu_insts; i++) { + 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"); } } +struct queued_qpu_inst { + struct simple_node link; + uint64_t inst; +}; + +static void +queue(struct qcompile *c, uint64_t inst) +{ + struct queued_qpu_inst *q = calloc(1, sizeof(*q)); + q->inst = inst; + insert_at_tail(&c->qpu_inst_list, &q->link); +} + +static uint64_t * +last_inst(struct qcompile *c) +{ + struct queued_qpu_inst *q = + (struct queued_qpu_inst *)last_elem(&c->qpu_inst_list); + 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 @@ -50,23 +71,119 @@ vc4_dump_program(struct qcompile *c) * instruction, instead. */ static void -fixup_raddr_conflict(uint64_t *insts, uint32_t *ni, +fixup_raddr_conflict(struct qcompile *c, struct qpu_reg src0, struct qpu_reg *src1) { 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) { - insts[(*ni)++] = qpu_inst(qpu_a_MOV(qpu_r3(), *src1), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(qpu_r3(), *src1), + qpu_m_NOP())); *src1 = qpu_r3(); } } +static void +serialize_one_inst(struct qcompile *c, uint64_t inst) +{ + 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)); + } + c->qpu_insts[c->qpu_inst_count++] = inst; +} + +static void +serialize_insts(struct qcompile *c) +{ + int last_sfu_write = -10; + + 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 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), + }; + + /* "An instruction must not read from a location in physical + * regfile A or B that was written to by the previous + * instruction." + */ + 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 (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 (reads_r4) { + while (c->qpu_inst_count - last_sfu_write < 3) { + serialize_one_inst(c, qpu_inst(qpu_a_NOP(), + qpu_m_NOP())); + } + } + + 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; + } + + serialize_one_inst(c, q->inst); + + remove_from_list(&q->link); + free(q); + } +} + void vc4_generate_code(struct qcompile *c) { - uint64_t *insts = malloc(sizeof(uint64_t) * 1024); /* XXX: sizing */ - uint32_t ni = 0; 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)); @@ -84,6 +201,8 @@ vc4_generate_code(struct qcompile *c) 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; @@ -99,8 +218,8 @@ vc4_generate_code(struct qcompile *c) switch (c->stage) { case QSTAGE_VERT: case QSTAGE_COORD: - insts[ni++] = qpu_load_imm_ui(qpu_vrsetup(), 0x00401a00); - insts[ni++] = qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00); + queue(c, qpu_load_imm_ui(qpu_vrsetup(), 0x00401a00)); + queue(c, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00)); break; case QSTAGE_FRAG: break; @@ -208,8 +327,8 @@ vc4_generate_code(struct qcompile *c) /* 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) { - insts[ni++] = qpu_inst(qpu_a_MOV(dst, src[0]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(dst, src[0]), + qpu_m_NOP())); } break; @@ -217,30 +336,28 @@ vc4_generate_code(struct qcompile *c) case QOP_SNE: case QOP_SGE: case QOP_SLT: - fixup_raddr_conflict(insts, &ni, src[0], &src[1]); - insts[ni++] = qpu_inst(qpu_a_SUB(qpu_ra(QPU_W_NOP), - src[0], src[1]), - qpu_m_NOP()); - insts[ni - 1] |= QPU_SF; + fixup_raddr_conflict(c, src[0], &src[1]); + queue(c, qpu_inst(qpu_a_SUB(qpu_ra(QPU_W_NOP), + src[0], src[1]), + qpu_m_NOP())); + *last_inst(c) |= QPU_SF; - insts[ni++] = qpu_load_imm_f(dst, 0.0); - insts[ni++] = qpu_load_imm_f(dst, 1.0); - insts[ni - 1] = ((insts[ni - 1] & ~QPU_COND_ADD_MASK) + queue(c, qpu_load_imm_f(dst, 0.0)); + queue(c, qpu_load_imm_f(dst, 1.0)); + *last_inst(c) = ((*last_inst(c) & ~QPU_COND_ADD_MASK) | QPU_SET_FIELD(compareflags[qinst->op - QOP_SEQ], QPU_COND_ADD)); break; case QOP_VPM_WRITE: - insts[ni++] = qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_VPM), - src[0]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(qpu_ra(QPU_W_VPM), src[0]), + qpu_m_NOP())); break; case QOP_VPM_READ: - insts[ni++] = qpu_inst(qpu_a_MOV(dst, - qpu_ra(QPU_R_VPM)), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(dst, qpu_ra(QPU_R_VPM)), + qpu_m_NOP())); break; case QOP_RCP: @@ -249,70 +366,70 @@ vc4_generate_code(struct qcompile *c) case QOP_LOG2: switch (qinst->op) { case QOP_RCP: - insts[ni++] = qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP), - src[0]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP), + src[0]), + qpu_m_NOP())); break; case QOP_RSQ: - insts[ni++] = qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT), - src[0]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT), + src[0]), + qpu_m_NOP())); break; case QOP_EXP2: - insts[ni++] = qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP), - src[0]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP), + src[0]), + qpu_m_NOP())); break; case QOP_LOG2: - insts[ni++] = qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG), - src[0]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG), + src[0]), + qpu_m_NOP())); break; default: abort(); } - insts[ni++] = qpu_inst(qpu_a_NOP(), qpu_m_NOP()); - insts[ni++] = qpu_inst(qpu_a_NOP(), qpu_m_NOP()); - - insts[ni++] = qpu_inst(qpu_a_MOV(dst, qpu_r4()), - qpu_m_NOP()); + 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++) { - insts[ni++] = qpu_inst(qpu_a_NOP(), - qpu_m_MOV(qpu_r3(), src[i])); - insts[ni - 1] |= QPU_PM; - insts[ni - 1] |= QPU_SET_FIELD(QPU_PACK_MUL_8A + 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); } - insts[ni++] = qpu_inst(qpu_a_MOV(dst, qpu_r3()), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(dst, qpu_r3()), + qpu_m_NOP())); + break; case QOP_TLB_COLOR_WRITE: - insts[ni++] = qpu_inst(qpu_a_MOV(qpu_tlbc(), - src[0]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_MOV(qpu_tlbc(), + src[0]), + qpu_m_NOP())); break; case QOP_VARY_ADD_C: - insts[ni++] = qpu_inst(qpu_a_FADD(dst, - src[0], qpu_r5()), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_FADD(dst, + src[0], qpu_r5()), + qpu_m_NOP())); break; case QOP_PACK_SCALED: - insts[ni++] = qpu_inst(qpu_a_MOV(dst, src[0]), - qpu_m_NOP()); - insts[ni - 1] |= QPU_SET_FIELD(QPU_PACK_A_16A, QPU_PACK); + queue(c, qpu_inst(qpu_a_MOV(dst, src[0]), + qpu_m_NOP())); + *last_inst(c) |= QPU_SET_FIELD(QPU_PACK_A_16A, + QPU_PACK); - insts[ni++] = qpu_inst(qpu_a_MOV(dst, src[1]), - qpu_m_NOP()); - insts[ni - 1] |= QPU_SET_FIELD(QPU_PACK_A_16B, QPU_PACK); + queue(c, qpu_inst(qpu_a_MOV(dst, src[1]), + qpu_m_NOP())); + *last_inst(c) |= QPU_SET_FIELD(QPU_PACK_A_16B, + QPU_PACK); break; @@ -327,51 +444,54 @@ vc4_generate_code(struct qcompile *c) if (qir_get_op_nsrc(qinst->op) == 1) src[1] = src[0]; - fixup_raddr_conflict(insts, &ni, src[0], &src[1]); + fixup_raddr_conflict(c, src[0], &src[1]); if (translate[qinst->op].is_mul) { - insts[ni++] = qpu_inst(qpu_a_NOP(), - qpu_m_alu2(translate[qinst->op].op, - dst, src[0], src[1])); + queue(c, qpu_inst(qpu_a_NOP(), + qpu_m_alu2(translate[qinst->op].op, + dst, + src[0], src[1]))); } else { - insts[ni++] = qpu_inst(qpu_a_alu2(translate[qinst->op].op, - dst, src[0], src[1]), - qpu_m_NOP()); + queue(c, qpu_inst(qpu_a_alu2(translate[qinst->op].op, + dst, + src[0], src[1]), + qpu_m_NOP())); } break; } - - if ((dst.mux == QPU_MUX_A || dst.mux == QPU_MUX_B) && - dst.addr < 32) - insts[ni++] = qpu_inst(qpu_a_NOP(), qpu_m_NOP()); } - /* thread end can't have VPM write */ - if (QPU_GET_FIELD(insts[ni - 1], QPU_WADDR_ADD) == QPU_W_VPM || - QPU_GET_FIELD(insts[ni - 1], QPU_WADDR_MUL) == QPU_W_VPM) - insts[ni++] = qpu_inst(qpu_a_NOP(), qpu_m_NOP()); + serialize_insts(c); - insts[ni - 1] = qpu_set_sig(insts[ni - 1], QPU_SIG_PROG_END); - insts[ni++] = qpu_inst(qpu_a_NOP(), qpu_m_NOP()); - insts[ni++] = qpu_inst(qpu_a_NOP(), qpu_m_NOP()); + /* thread end can't have VPM write */ + 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())); + } + + 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())); switch (c->stage) { case QSTAGE_VERT: case QSTAGE_COORD: break; case QSTAGE_FRAG: - insts[2] = qpu_set_sig(insts[2], QPU_SIG_WAIT_FOR_SCOREBOARD); - insts[ni - 1] = qpu_set_sig(insts[ni - 1], - QPU_SIG_SCOREBOARD_UNLOCK); + c->qpu_insts[2] = qpu_set_sig(c->qpu_insts[2], + QPU_SIG_WAIT_FOR_SCOREBOARD); + c->qpu_insts[c->qpu_inst_count - 1] = + qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1], + QPU_SIG_SCOREBOARD_UNLOCK); break; } - c->qpu_insts = insts; - c->num_qpu_insts = ni; - if (vc4_debug & VC4_DEBUG_QPU) vc4_dump_program(c); - vc4_qpu_validate(insts, ni); + vc4_qpu_validate(c->qpu_insts, c->qpu_inst_count); } -