From a59da513d3229c883809ac2088c9612abcec1470 Mon Sep 17 00:00:00 2001 From: Eric Anholt Date: Tue, 15 Mar 2016 17:53:36 -0700 Subject: [PATCH] vc4: Move the QPU instructions to schedule into each block. We'll want to schedule them individually, to handle delay slots. --- src/gallium/drivers/vc4/vc4_qir.c | 1 + src/gallium/drivers/vc4/vc4_qir.h | 2 + src/gallium/drivers/vc4/vc4_qpu_emit.c | 205 +++++++++++---------- src/gallium/drivers/vc4/vc4_qpu_schedule.c | 113 +++++++----- 4 files changed, 180 insertions(+), 141 deletions(-) diff --git a/src/gallium/drivers/vc4/vc4_qir.c b/src/gallium/drivers/vc4/vc4_qir.c index 982e8298ae9..e1d663dd3a7 100644 --- a/src/gallium/drivers/vc4/vc4_qir.c +++ b/src/gallium/drivers/vc4/vc4_qir.c @@ -536,6 +536,7 @@ qir_new_block(struct vc4_compile *c) struct qblock *block = rzalloc(c, struct qblock); list_inithead(&block->instructions); + list_inithead(&block->qpu_inst_list); block->predecessors = _mesa_set_create(block, _mesa_hash_pointer, diff --git a/src/gallium/drivers/vc4/vc4_qir.h b/src/gallium/drivers/vc4/vc4_qir.h index fe85b769775..fabdf10e24d 100644 --- a/src/gallium/drivers/vc4/vc4_qir.h +++ b/src/gallium/drivers/vc4/vc4_qir.h @@ -355,6 +355,7 @@ struct qblock { struct list_head link; struct list_head instructions; + struct list_head qpu_inst_list; struct set *predecessors; struct qblock *successors[2]; @@ -467,6 +468,7 @@ struct vc4_compile { struct qblock *loop_break_block; struct list_head qpu_inst_list; + uint64_t *qpu_insts; uint32_t qpu_inst_count; uint32_t qpu_inst_size; diff --git a/src/gallium/drivers/vc4/vc4_qpu_emit.c b/src/gallium/drivers/vc4/vc4_qpu_emit.c index 2257dcce83b..9001643507e 100644 --- a/src/gallium/drivers/vc4/vc4_qpu_emit.c +++ b/src/gallium/drivers/vc4/vc4_qpu_emit.c @@ -44,31 +44,31 @@ vc4_dump_program(struct vc4_compile *c) } static void -queue(struct vc4_compile *c, uint64_t inst) +queue(struct qblock *block, uint64_t inst) { - struct queued_qpu_inst *q = rzalloc(c, struct queued_qpu_inst); + struct queued_qpu_inst *q = rzalloc(block, struct queued_qpu_inst); q->inst = inst; - list_addtail(&q->link, &c->qpu_inst_list); + list_addtail(&q->link, &block->qpu_inst_list); } static uint64_t * -last_inst(struct vc4_compile *c) +last_inst(struct qblock *block) { struct queued_qpu_inst *q = - (struct queued_qpu_inst *)c->qpu_inst_list.prev; + (struct queued_qpu_inst *)block->qpu_inst_list.prev; return &q->inst; } static void -set_last_cond_add(struct vc4_compile *c, uint32_t cond) +set_last_cond_add(struct qblock *block, uint32_t cond) { - *last_inst(c) = qpu_set_cond_add(*last_inst(c), cond); + *last_inst(block) = qpu_set_cond_add(*last_inst(block), cond); } static void -set_last_cond_mul(struct vc4_compile *c, uint32_t cond) +set_last_cond_mul(struct qblock *block, uint32_t cond) { - *last_inst(c) = qpu_set_cond_mul(*last_inst(c), cond); + *last_inst(block) = qpu_set_cond_mul(*last_inst(block), cond); } /** @@ -106,7 +106,7 @@ swap_file(struct qpu_reg *src) * instruction, instead. We reserve ra31/rb31 for this purpose. */ static void -fixup_raddr_conflict(struct vc4_compile *c, +fixup_raddr_conflict(struct qblock *block, struct qpu_reg dst, struct qpu_reg *src0, struct qpu_reg *src1, struct qinst *inst, uint64_t *unpack) @@ -129,39 +129,39 @@ fixup_raddr_conflict(struct vc4_compile *c, * in case of unpacks. */ if (qir_is_float_input(inst)) - queue(c, qpu_a_FMAX(qpu_rb(31), *src0, *src0)); + queue(block, qpu_a_FMAX(qpu_rb(31), *src0, *src0)); else - queue(c, qpu_a_MOV(qpu_rb(31), *src0)); + queue(block, 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; + *last_inst(block) |= *unpack; *unpack = 0; } *src0 = qpu_rb(31); } else { - queue(c, qpu_a_MOV(qpu_ra(31), *src0)); + queue(block, qpu_a_MOV(qpu_ra(31), *src0)); *src0 = qpu_ra(31); } } static void -set_last_dst_pack(struct vc4_compile *c, struct qinst *inst) +set_last_dst_pack(struct qblock *block, 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); + bool had_pm = *last_inst(block) & QPU_PM; + bool had_ws = *last_inst(block) & QPU_WS; + uint32_t unpack = QPU_GET_FIELD(*last_inst(block), QPU_UNPACK); if (!inst->dst.pack) return; - *last_inst(c) |= QPU_SET_FIELD(inst->dst.pack, QPU_PACK); + *last_inst(block) |= QPU_SET_FIELD(inst->dst.pack, QPU_PACK); if (qir_is_mul(inst)) { assert(!unpack || had_pm); - *last_inst(c) |= QPU_PM; + *last_inst(block) |= QPU_PM; } else { assert(!unpack || !had_pm); assert(!had_ws); /* dst must be a-file to pack. */ @@ -169,51 +169,23 @@ set_last_dst_pack(struct vc4_compile *c, struct qinst *inst) } static void -handle_r4_qpu_write(struct vc4_compile *c, struct qinst *qinst, +handle_r4_qpu_write(struct qblock *block, struct qinst *qinst, struct qpu_reg dst) { if (dst.mux != QPU_MUX_R4) - queue(c, qpu_a_MOV(dst, qpu_r4())); + queue(block, qpu_a_MOV(dst, qpu_r4())); else if (qinst->sf) - queue(c, qpu_a_MOV(qpu_ra(QPU_W_NOP), qpu_r4())); + queue(block, qpu_a_MOV(qpu_ra(QPU_W_NOP), qpu_r4())); } -void -vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) +static void +vc4_generate_code_block(struct vc4_compile *c, + struct qblock *block, + struct qpu_reg *temp_registers) { - 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: - /* 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; - } - - qir_for_each_inst_inorder(qinst, c) { + qir_for_each_inst(qinst, block) { #if 0 fprintf(stderr, "translating qinst to qpu: "); qir_dump_inst(qinst); @@ -369,40 +341,40 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) case QOP_LOG2: switch (qinst->op) { case QOP_RCP: - queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP), - src[0]) | unpack); + queue(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP), + src[0]) | unpack); break; case QOP_RSQ: - queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT), - src[0]) | unpack); + queue(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT), + src[0]) | unpack); break; case QOP_EXP2: - queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP), - src[0]) | unpack); + queue(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP), + src[0]) | unpack); break; case QOP_LOG2: - queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG), - src[0]) | unpack); + queue(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG), + src[0]) | unpack); break; default: abort(); } - handle_r4_qpu_write(c, qinst, dst); + handle_r4_qpu_write(block, qinst, dst); break; case QOP_LOAD_IMM: assert(qinst->src[0].file == QFILE_LOAD_IMM); - queue(c, qpu_load_imm_ui(dst, qinst->src[0].index)); + queue(block, qpu_load_imm_ui(dst, qinst->src[0].index)); break; case QOP_MS_MASK: src[1] = qpu_ra(QPU_R_MS_REV_FLAGS); - fixup_raddr_conflict(c, dst, &src[0], &src[1], + fixup_raddr_conflict(block, dst, &src[0], &src[1], qinst, &unpack); - queue(c, qpu_a_AND(qpu_ra(QPU_W_MS_FLAGS), - src[0], src[1]) | unpack); + queue(block, qpu_a_AND(qpu_ra(QPU_W_MS_FLAGS), + src[0], src[1]) | unpack); break; case QOP_FRAG_Z: @@ -413,45 +385,45 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) break; case QOP_TLB_COLOR_READ: - queue(c, qpu_NOP()); - *last_inst(c) = qpu_set_sig(*last_inst(c), - QPU_SIG_COLOR_LOAD); - handle_r4_qpu_write(c, qinst, dst); + queue(block, qpu_NOP()); + *last_inst(block) = qpu_set_sig(*last_inst(block), + QPU_SIG_COLOR_LOAD); + handle_r4_qpu_write(block, qinst, dst); break; case QOP_VARY_ADD_C: - queue(c, qpu_a_FADD(dst, src[0], qpu_r5()) | unpack); + queue(block, 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_a_MOV(qpu_rb(QPU_W_TMU0_S + - (qinst->op - QOP_TEX_S)), - src[0]) | unpack); + queue(block, 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], + fixup_raddr_conflict(block, dst, &src[0], &src[1], qinst, &unpack); - queue(c, qpu_a_ADD(qpu_rb(QPU_W_TMU0_S), - src[0], src[1]) | unpack); + queue(block, 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); - handle_r4_qpu_write(c, qinst, dst); + queue(block, qpu_NOP()); + *last_inst(block) = qpu_set_sig(*last_inst(block), + QPU_SIG_LOAD_TMU0); + handle_r4_qpu_write(block, qinst, dst); break; case QOP_BRANCH: /* The branch target will be updated at QPU scheduling * time. */ - queue(c, (qpu_branch(qinst->cond, 0) | - QPU_BRANCH_REL)); + queue(block, (qpu_branch(qinst->cond, 0) | + QPU_BRANCH_REL)); handled_qinst_cond = true; break; @@ -472,22 +444,22 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) if (qir_get_op_nsrc(qinst->op) == 1) src[1] = src[0]; - fixup_raddr_conflict(c, dst, &src[0], &src[1], + fixup_raddr_conflict(block, 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]) | unpack); - set_last_cond_mul(c, qinst->cond); + queue(block, qpu_m_alu2(translate[qinst->op].op, + dst, + src[0], src[1]) | unpack); + set_last_cond_mul(block, qinst->cond); } else { - queue(c, qpu_a_alu2(translate[qinst->op].op, - dst, - src[0], src[1]) | unpack); - set_last_cond_add(c, qinst->cond); + queue(block, qpu_a_alu2(translate[qinst->op].op, + dst, + src[0], src[1]) | unpack); + set_last_cond_add(block, qinst->cond); } handled_qinst_cond = true; - set_last_dst_pack(c, qinst); + set_last_dst_pack(block, qinst); break; } @@ -496,9 +468,48 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) handled_qinst_cond); if (qinst->sf) - *last_inst(c) |= QPU_SF; + *last_inst(block) |= QPU_SF; + } +} + +void +vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) +{ + 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; + struct qblock *start_block = list_first_entry(&c->blocks, + struct qblock, link); + + 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(start_block, + 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(start_block, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00)); + break; + case QSTAGE_FRAG: + break; } + qir_for_each_block(block, c) + vc4_generate_code_block(c, block, temp_registers); + uint32_t cycles = qpu_schedule_instructions(c); uint32_t inst_count_at_schedule_time = c->qpu_inst_count; diff --git a/src/gallium/drivers/vc4/vc4_qpu_schedule.c b/src/gallium/drivers/vc4/vc4_qpu_schedule.c index 4af605c109c..fad10e509e2 100644 --- a/src/gallium/drivers/vc4/vc4_qpu_schedule.c +++ b/src/gallium/drivers/vc4/vc4_qpu_schedule.c @@ -722,22 +722,14 @@ mark_instruction_scheduled(struct list_head *schedule_list, } static uint32_t -schedule_instructions(struct vc4_compile *c, struct list_head *schedule_list) +schedule_instructions(struct vc4_compile *c, struct list_head *schedule_list, + enum quniform_contents *orig_uniform_contents, + uint32_t *orig_uniform_data, + uint32_t *next_uniform) { struct choose_scoreboard scoreboard; uint32_t time = 0; - /* We reorder the uniforms as we schedule instructions, so save the - * old data off and replace it. - */ - uint32_t *uniform_data = c->uniform_data; - enum quniform_contents *uniform_contents = c->uniform_contents; - c->uniform_contents = ralloc_array(c, enum quniform_contents, - c->num_uniforms); - c->uniform_data = ralloc_array(c, uint32_t, c->num_uniforms); - c->uniform_array_size = c->num_uniforms; - uint32_t next_uniform = 0; - memset(&scoreboard, 0, sizeof(scoreboard)); scoreboard.last_waddr_a = ~0; scoreboard.last_waddr_b = ~0; @@ -785,11 +777,11 @@ schedule_instructions(struct vc4_compile *c, struct list_head *schedule_list) mark_instruction_scheduled(schedule_list, time, chosen, true); if (chosen->uniform != -1) { - c->uniform_data[next_uniform] = - uniform_data[chosen->uniform]; - c->uniform_contents[next_uniform] = - uniform_contents[chosen->uniform]; - next_uniform++; + c->uniform_data[*next_uniform] = + orig_uniform_data[chosen->uniform]; + c->uniform_contents[*next_uniform] = + orig_uniform_contents[chosen->uniform]; + (*next_uniform)++; } merge = choose_instruction_to_schedule(&scoreboard, @@ -801,11 +793,11 @@ schedule_instructions(struct vc4_compile *c, struct list_head *schedule_list) inst = qpu_merge_inst(inst, merge->inst->inst); assert(inst != 0); if (merge->uniform != -1) { - c->uniform_data[next_uniform] = - uniform_data[merge->uniform]; - c->uniform_contents[next_uniform] = - uniform_contents[merge->uniform]; - next_uniform++; + c->uniform_data[*next_uniform] = + orig_uniform_data[merge->uniform]; + c->uniform_contents[*next_uniform] = + orig_uniform_contents[merge->uniform]; + (*next_uniform)++; } if (debug) { @@ -840,47 +832,37 @@ schedule_instructions(struct vc4_compile *c, struct list_head *schedule_list) time++; } - assert(next_uniform == c->num_uniforms); - return time; } -uint32_t -qpu_schedule_instructions(struct vc4_compile *c) +static uint32_t +qpu_schedule_instructions_block(struct vc4_compile *c, struct qblock *block, + enum quniform_contents *orig_uniform_contents, + uint32_t *orig_uniform_data, + uint32_t *next_uniform) { void *mem_ctx = ralloc_context(NULL); struct list_head schedule_list; list_inithead(&schedule_list); - if (debug) { - fprintf(stderr, "Pre-schedule instructions\n"); - list_for_each_entry(struct queued_qpu_inst, q, - &c->qpu_inst_list, link) { - vc4_qpu_disasm(&q->inst, 1); - fprintf(stderr, "\n"); - } - fprintf(stderr, "\n"); - } - /* Wrap each instruction in a scheduler structure. */ - uint32_t next_uniform = 0; - while (!list_empty(&c->qpu_inst_list)) { + uint32_t next_sched_uniform = *next_uniform; + while (!list_empty(&block->qpu_inst_list)) { struct queued_qpu_inst *inst = - (struct queued_qpu_inst *)c->qpu_inst_list.next; + (struct queued_qpu_inst *)block->qpu_inst_list.next; struct schedule_node *n = rzalloc(mem_ctx, struct schedule_node); n->inst = inst; if (reads_uniform(inst->inst)) { - n->uniform = next_uniform++; + n->uniform = next_sched_uniform++; } else { n->uniform = -1; } list_del(&inst->link); list_addtail(&n->link, &schedule_list); } - assert(next_uniform == c->num_uniforms); calculate_forward_deps(c, &schedule_list); calculate_reverse_deps(c, &schedule_list); @@ -889,7 +871,52 @@ qpu_schedule_instructions(struct vc4_compile *c) compute_delay(n); } - uint32_t cycles = schedule_instructions(c, &schedule_list); + uint32_t cycles = schedule_instructions(c, &schedule_list, + orig_uniform_contents, + orig_uniform_data, + next_uniform); + + ralloc_free(mem_ctx); + + return cycles; +} + +uint32_t +qpu_schedule_instructions(struct vc4_compile *c) +{ + /* We reorder the uniforms as we schedule instructions, so save the + * old data off and replace it. + */ + uint32_t *uniform_data = c->uniform_data; + enum quniform_contents *uniform_contents = c->uniform_contents; + c->uniform_contents = ralloc_array(c, enum quniform_contents, + c->num_uniforms); + c->uniform_data = ralloc_array(c, uint32_t, c->num_uniforms); + c->uniform_array_size = c->num_uniforms; + uint32_t next_uniform = 0; + + if (debug) { + fprintf(stderr, "Pre-schedule instructions\n"); + qir_for_each_block(block, c) { + fprintf(stderr, "BLOCK %d\n", block->index); + list_for_each_entry(struct queued_qpu_inst, q, + &block->qpu_inst_list, link) { + vc4_qpu_disasm(&q->inst, 1); + fprintf(stderr, "\n"); + } + } + fprintf(stderr, "\n"); + } + + uint32_t cycles = 0; + qir_for_each_block(block, c) { + cycles += qpu_schedule_instructions_block(c, block, + uniform_contents, + uniform_data, + &next_uniform); + } + + assert(next_uniform == c->num_uniforms); if (debug) { fprintf(stderr, "Post-schedule instructions\n"); @@ -897,7 +924,5 @@ qpu_schedule_instructions(struct vc4_compile *c) fprintf(stderr, "\n"); } - ralloc_free(mem_ctx); - return cycles; } -- 2.30.2