/* * Copyright © 2014 Broadcom * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS * IN THE SOFTWARE. */ #include #include "vc4_context.h" #include "vc4_qir.h" #include "vc4_qpu.h" #include "util/ralloc.h" static void vc4_dump_program(struct vc4_compile *c) { 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"); } static void queue(struct qblock *block, uint64_t inst) { struct queued_qpu_inst *q = rzalloc(block, struct queued_qpu_inst); q->inst = inst; list_addtail(&q->link, &block->qpu_inst_list); } static uint64_t * last_inst(struct qblock *block) { struct queued_qpu_inst *q = (struct queued_qpu_inst *)block->qpu_inst_list.prev; return &q->inst; } static void set_last_cond_add(struct qblock *block, uint32_t cond) { *last_inst(block) = qpu_set_cond_add(*last_inst(block), cond); } static void set_last_cond_mul(struct qblock *block, uint32_t cond) { *last_inst(block) = qpu_set_cond_mul(*last_inst(block), 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; } } /** * Sets up the VPM read FIFO before we do any VPM read. * * VPM reads (vertex attribute input) and VPM writes (varyings output) from * the QPU reuse the VRI (varying interpolation) block's FIFOs to talk to the * VPM block. In the VS/CS (unlike in the FS), the block starts out * uninitialized, and you need to emit setup to the block before any VPM * reads/writes. * * VRI has a FIFO in each direction, with each FIFO able to hold four * 32-bit-per-vertex values. VPM reads come through the read FIFO and VPM * writes go through the write FIFO. The read/write setup values from QPU go * through the write FIFO as well, with a sideband signal indicating that * they're setup values. Once a read setup reaches the other side of the * FIFO, the VPM block will start asynchronously reading vertex attributes and * filling the read FIFO -- that way hopefully the QPU doesn't have to block * on reads later. * * VPM read setup can configure 16 32-bit-per-vertex values to be read at a * time, which is 4 vec4s. If more than that is being read (since we support * 8 vec4 vertex attributes), then multiple read setup writes need to be done. * * The existence of the FIFO makes it seem like you should be able to emit * both setups for the 5-8 attribute cases and then do all the attribute * reads. However, once the setup value makes it to the other end of the * write FIFO, it will immediately update the VPM block's setup register. * That updated setup register would be used for read FIFO fills from then on, * breaking whatever remaining VPM values were supposed to be read into the * read FIFO from the previous attribute set. * * As a result, we need to emit the read setup, pull every VPM read value from * that setup, and only then emit the second setup if applicable. */ static void setup_for_vpm_read(struct vc4_compile *c, struct qblock *block) { if (c->num_inputs_in_fifo) { c->num_inputs_in_fifo--; return; } c->num_inputs_in_fifo = MIN2(c->num_inputs_remaining, 16); queue(block, qpu_load_imm_ui(qpu_vrsetup(), c->vpm_read_offset | 0x00001a00 | ((c->num_inputs_in_fifo & 0xf) << 20))); c->num_inputs_remaining -= c->num_inputs_in_fifo; c->vpm_read_offset += c->num_inputs_in_fifo; c->num_inputs_in_fifo--; } /** * 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 ra14/rb14 for this purpose. */ static void fixup_raddr_conflict(struct qblock *block, struct qpu_reg dst, struct qpu_reg *src0, struct qpu_reg *src1, struct qinst *inst, uint64_t *unpack) { 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; } if (swap_file(src0) || swap_file(src1)) return; if (mux0 == QPU_MUX_A) { /* Make sure we use the same type of MOV as the instruction, * in case of unpacks. */ if (qir_is_float_input(inst)) queue(block, qpu_a_FMAX(qpu_rb(14), *src0, *src0)); else queue(block, qpu_a_MOV(qpu_rb(14), *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(block) |= *unpack; *unpack = 0; } *src0 = qpu_rb(14); } else { queue(block, qpu_a_MOV(qpu_ra(14), *src0)); *src0 = qpu_ra(14); } } static void set_last_dst_pack(struct qblock *block, struct qinst *inst) { 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(block) |= QPU_SET_FIELD(inst->dst.pack, QPU_PACK); if (qir_is_mul(inst)) { assert(!unpack || had_pm); *last_inst(block) |= QPU_PM; } else { assert(!unpack || !had_pm); assert(!had_ws); /* dst must be a-file to pack. */ } } static void handle_r4_qpu_write(struct qblock *block, struct qinst *qinst, struct qpu_reg dst) { if (dst.mux != QPU_MUX_R4) queue(block, qpu_a_MOV(dst, qpu_r4())); else if (qinst->sf) queue(block, qpu_a_MOV(qpu_ra(QPU_W_NOP), qpu_r4())); } static void vc4_generate_code_block(struct vc4_compile *c, struct qblock *block, struct qpu_reg *temp_registers) { int last_vpm_read_index = -1; qir_for_each_inst(qinst, block) { #if 0 fprintf(stderr, "translating qinst to qpu: "); qir_dump_inst(qinst); fprintf(stderr, "\n"); #endif static const struct { uint32_t op; } translate[] = { #define A(name) [QOP_##name] = {QPU_A_##name} #define M(name) [QOP_##name] = {QPU_M_##name} A(FADD), A(FSUB), A(FMIN), A(FMAX), A(FMINABS), 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 }, }; uint64_t unpack = 0; struct qpu_reg src[ARRAY_SIZE(qinst->src)]; for (int i = 0; i < qir_get_nsrc(qinst); i++) { 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: 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: src[i] = qpu_unif(); break; 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: setup_for_vpm_read(c, block); 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_QPU_ELEMENT: src[i] = qpu_ra(QPU_R_ELEM_QPU); break; case QFILE_TLB_COLOR_WRITE: case QFILE_TLB_COLOR_WRITE_MS: case QFILE_TLB_Z_WRITE: case QFILE_TLB_STENCIL_SETUP: case QFILE_TEX_S: case QFILE_TEX_S_DIRECT: case QFILE_TEX_T: case QFILE_TEX_R: case QFILE_TEX_B: unreachable("bad qir src file"); } } struct qpu_reg dst; switch (qinst->dst.file) { case QFILE_NULL: dst = qpu_ra(QPU_W_NOP); break; case QFILE_TEMP: dst = temp_registers[qinst->dst.index]; break; 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_TEX_S: case QFILE_TEX_S_DIRECT: dst = qpu_rb(QPU_W_TMU0_S); break; case QFILE_TEX_T: dst = qpu_rb(QPU_W_TMU0_T); break; case QFILE_TEX_R: dst = qpu_rb(QPU_W_TMU0_R); break; case QFILE_TEX_B: dst = qpu_rb(QPU_W_TMU0_B); 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: case QFILE_QPU_ELEMENT: 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(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP), src[0]) | unpack); break; case QOP_RSQ: queue(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT), src[0]) | unpack); break; case QOP_EXP2: queue(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP), src[0]) | unpack); break; case QOP_LOG2: queue(block, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG), src[0]) | unpack); break; default: abort(); } handle_r4_qpu_write(block, qinst, dst); break; case QOP_LOAD_IMM: assert(qinst->src[0].file == QFILE_LOAD_IMM); queue(block, qpu_load_imm_ui(dst, qinst->src[0].index)); break; case QOP_LOAD_IMM_U2: queue(block, qpu_load_imm_u2(dst, qinst->src[0].index)); break; case QOP_LOAD_IMM_I2: queue(block, qpu_load_imm_i2(dst, qinst->src[0].index)); break; case QOP_ROT_MUL: /* Rotation at the hardware level occurs on the inputs * to the MUL unit, and they must be accumulators in * order to have the time necessary to move things. */ assert(src[0].mux <= QPU_MUX_R3); queue(block, qpu_m_rot(dst, src[0], qinst->src[1].index - QPU_SMALL_IMM_MUL_ROT) | unpack); set_last_cond_mul(block, qinst->cond); handled_qinst_cond = true; set_last_dst_pack(block, qinst); break; case QOP_MS_MASK: src[1] = qpu_ra(QPU_R_MS_REV_FLAGS); fixup_raddr_conflict(block, dst, &src[0], &src[1], qinst, &unpack); queue(block, qpu_a_AND(qpu_ra(QPU_W_MS_FLAGS), src[0], src[1]) | unpack); break; case QOP_FRAG_Z: 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(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(block, qpu_a_FADD(dst, src[0], qpu_r5()) | unpack); break; case QOP_TEX_RESULT: 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_THRSW: queue(block, qpu_NOP()); *last_inst(block) = qpu_set_sig(*last_inst(block), QPU_SIG_THREAD_SWITCH); c->last_thrsw = last_inst(block); break; case QOP_BRANCH: /* The branch target will be updated at QPU scheduling * time. */ queue(block, (qpu_branch(qinst->cond, 0) | QPU_BRANCH_REL)); handled_qinst_cond = true; break; case QOP_UNIFORMS_RESET: fixup_raddr_conflict(block, dst, &src[0], &src[1], qinst, &unpack); queue(block, qpu_a_ADD(qpu_ra(QPU_W_UNIFORMS_ADDRESS), src[0], src[1])); 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_non_sideband_nsrc(qinst) == 1) src[1] = src[0]; fixup_raddr_conflict(block, dst, &src[0], &src[1], qinst, &unpack); if (qir_is_mul(qinst)) { queue(block, qpu_m_alu2(translate[qinst->op].op, dst, src[0], src[1]) | unpack); set_last_cond_mul(block, qinst->cond); } else { 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(block, qinst); break; } assert(qinst->cond == QPU_COND_ALWAYS || handled_qinst_cond); if (qinst->sf) *last_inst(block) |= QPU_SF; } } void vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c) { struct qblock *start_block = list_first_entry(&c->blocks, struct qblock, link); struct qpu_reg *temp_registers = vc4_register_allocate(vc4, c); if (!temp_registers) return; switch (c->stage) { case QSTAGE_VERT: case QSTAGE_COORD: c->num_inputs_remaining = c->num_inputs; 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); /* Switch the last SIG_THRSW instruction to SIG_LAST_THRSW. * * LAST_THRSW is a new signal in BCM2708B0 (including Raspberry Pi) * that ensures that a later thread doesn't try to lock the scoreboard * and terminate before an earlier-spawned thread on the same QPU, by * delaying switching back to the later shader until earlier has * finished. Otherwise, if the earlier thread was hitting the same * quad, the scoreboard would deadlock. */ if (c->last_thrsw) { assert(QPU_GET_FIELD(*c->last_thrsw, QPU_SIG) == QPU_SIG_THREAD_SWITCH); *c->last_thrsw = ((*c->last_thrsw & ~QPU_SIG_MASK) | QPU_SET_FIELD(QPU_SIG_LAST_THREAD_SWITCH, QPU_SIG)); } 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], QPU_WADDR_ADD) == QPU_W_VPM || QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1], 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()); /* 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); qpu_serialize_one_inst(c, qpu_NOP()); qpu_serialize_one_inst(c, qpu_NOP()); switch (c->stage) { case QSTAGE_VERT: case QSTAGE_COORD: break; case QSTAGE_FRAG: c->qpu_insts[c->qpu_inst_count - 1] = qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1], QPU_SIG_SCOREBOARD_UNLOCK); 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); }