diff options
| -rw-r--r-- | src/panfrost/bifrost/bi_pack.c | 25 | ||||
| -rw-r--r-- | src/panfrost/bifrost/compiler.h | 6 |
2 files changed, 26 insertions, 5 deletions
diff --git a/src/panfrost/bifrost/bi_pack.c b/src/panfrost/bifrost/bi_pack.c index 65218b23010..2f961650d3e 100644 --- a/src/panfrost/bifrost/bi_pack.c +++ b/src/panfrost/bifrost/bi_pack.c @@ -34,11 +34,19 @@ * bits on the wire (as well as fixup branches) */ static uint64_t -bi_pack_header(bi_clause *clause, bi_clause *next) +bi_pack_header(bi_clause *clause, bi_clause *next, bool is_fragment) { struct bifrost_header header = { - /* stub */ + .back_to_back = clause->back_to_back, .no_end_of_shader = (next != NULL), + .elide_writes = is_fragment, + .branch_cond = clause->branch_conditional, + .datareg_writebarrier = clause->data_register_write_barrier, + .datareg = clause->data_register, + .scoreboard_deps = clause->dependencies, + .scoreboard_index = clause->scoreboard_id, + .clause_type = clause->clause_type, + .next_clause_type = next ? next->clause_type : 0, }; uint64_t u = 0; @@ -384,7 +392,7 @@ bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs) } static unsigned -bi_pack_add_ld_vary(bi_instruction *ins, struct bi_registers *regs) +bi_pack_add_ld_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs) { unsigned size = nir_alu_type_get_type_size(ins->dest_type); assert(size == 32 || size == 16); @@ -408,6 +416,10 @@ bi_pack_add_ld_vary(bi_instruction *ins, struct bi_registers *regs) packed_addr = bi_get_src(ins, regs, 0, false) | 0b11000; } + /* The destination is thrown in the data register */ + assert(ins->dest & BIR_INDEX_REGISTER); + clause->data_register = ins->dest & ~BIR_INDEX_REGISTER; + assert(channels >= 1 && channels <= 4); struct bifrost_ld_var pack = { @@ -445,7 +457,7 @@ bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs) case BI_LOAD_ATTR: return BIFROST_ADD_NOP; case BI_LOAD_VAR: - return bi_pack_add_ld_vary(bundle.add, regs); + return bi_pack_add_ld_vary(clause, bundle.add, regs); case BI_LOAD_VAR_ADDRESS: case BI_MINMAX: case BI_MOV: @@ -492,9 +504,12 @@ bi_pack_clause(bi_context *ctx, bi_clause *clause, bi_clause *next, struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true); assert(clause->bundle_count == 1); + /* Used to decide if we elide writes */ + bool is_fragment = ctx->stage == MESA_SHADER_FRAGMENT; + struct bifrost_fmt1 quad_1 = { .tag = BIFROST_FMT1_FINAL, - .header = bi_pack_header(clause, next), + .header = bi_pack_header(clause, next, is_fragment), .ins_1 = ins_1.lo, .ins_2 = ins_1.hi & ((1 << 11) - 1), .ins_0 = (ins_1.hi >> 11) & 0b111, diff --git a/src/panfrost/bifrost/compiler.h b/src/panfrost/bifrost/compiler.h index 2d3633502ec..3c84c2671d5 100644 --- a/src/panfrost/bifrost/compiler.h +++ b/src/panfrost/bifrost/compiler.h @@ -307,12 +307,18 @@ typedef struct { bool back_to_back; bool branch_conditional; + /* Assigned data register */ + unsigned data_register; + /* Corresponds to the usual bit but shifted by a clause */ bool data_register_write_barrier; /* Constants read by this clause. ISA limit. */ uint64_t constants[8]; unsigned constant_count; + + /* What type of high latency instruction is here, basically */ + unsigned clause_type; } bi_clause; typedef struct bi_block { |