X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fvc4%2Fvc4_qpu_emit.c;h=2257dcce83be3f70c180c612abbe815d75f6fb2f;hb=ee69cfd11d34e6570c579c42f9cd8b5c8ea36bcf;hp=133e138517839c4addd17586341812fc3b14fd25;hpb=3359ad6cda49fb977d837eb00e8ae4d781d95c2a;p=mesa.git diff --git a/src/gallium/drivers/vc4/vc4_qpu_emit.c b/src/gallium/drivers/vc4/vc4_qpu_emit.c index 133e1385178..2257dcce83b 100644 --- a/src/gallium/drivers/vc4/vc4_qpu_emit.c +++ b/src/gallium/drivers/vc4/vc4_qpu_emit.c @@ -40,6 +40,7 @@ vc4_dump_program(struct vc4_compile *c) vc4_qpu_disasm(&c->qpu_insts[i], 1); fprintf(stderr, "\n"); } + fprintf(stderr, "\n"); } static void @@ -64,6 +65,12 @@ set_last_cond_add(struct vc4_compile *c, uint32_t cond) *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. @@ -161,11 +168,20 @@ set_last_dst_pack(struct vc4_compile *c, struct qinst *inst) } } +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; @@ -197,7 +213,7 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) 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); @@ -251,6 +267,7 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) 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: @@ -285,6 +302,22 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) 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"); } } @@ -299,45 +332,37 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) case QFILE_VPM: dst = qpu_ra(QPU_W_VPM); break; + + case QFILE_TLB_COLOR_WRITE: + dst = qpu_tlbc(); + break; + + case QFILE_TLB_COLOR_WRITE_MS: + dst = qpu_tlbc_ms(); + break; + + case QFILE_TLB_Z_WRITE: + dst = qpu_ra(QPU_W_TLB_Z); + break; + + 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; } - 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]) | unpack); - 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: - queue(c, qpu_a_MOV(dst, src[0])); - if (qinst->src[0].pack) - *(last_inst(c)) |= unpack; - set_last_cond_add(c, qinst->op - QOP_SEL_X_Y_ZS + - QPU_COND_ZS); - - queue(c, qpu_a_MOV(dst, src[1])); - if (qinst->src[1].pack) - *(last_inst(c)) |= unpack; - set_last_cond_add(c, ((qinst->op - QOP_SEL_X_Y_ZS) ^ - 1) + QPU_COND_ZS); - - break; + bool handled_qinst_cond = false; + switch (qinst->op) { case QOP_RCP: case QOP_RSQ: case QOP_EXP2: @@ -363,24 +388,21 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) abort(); } - if (dst.mux != QPU_MUX_R4) - queue(c, qpu_a_MOV(dst, qpu_r4())); - - 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: @@ -390,40 +412,11 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) */ break; - case QOP_TLB_DISCARD_SETUP: - discard = true; - queue(c, qpu_a_MOV(src[0], src[0]) | unpack); - *last_inst(c) |= QPU_SF; - break; - - case QOP_TLB_STENCIL_SETUP: - assert(!unpack); - queue(c, qpu_a_MOV(qpu_ra(QPU_W_TLB_STENCIL_SETUP), - src[0]) | unpack); - break; - - case QOP_TLB_Z_WRITE: - queue(c, qpu_a_MOV(qpu_ra(QPU_W_TLB_Z), - src[0]) | unpack); - 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]) | unpack); - if (discard) { - set_last_cond_add(c, QPU_COND_ZS); - } + handle_r4_qpu_write(c, qinst, dst); break; case QOP_VARY_ADD_C: @@ -450,8 +443,16 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) 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())); + handle_r4_qpu_write(c, 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)); + handled_qinst_cond = true; break; default: @@ -478,23 +479,28 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) 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_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; } - 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], @@ -520,6 +526,14 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) 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); @@ -537,6 +551,15 @@ vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) 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);