X-Git-Url: https://git.libre-soc.org/?p=mesa.git;a=blobdiff_plain;f=src%2Fbroadcom%2Fcompiler%2Fnir_to_vir.c;h=b05c892bba617100b83af8645317602d472bbab1;hp=4cbdb764eab7b96e4bde518ebbff58e71b2d5f24;hb=456557a837ea005763283b6cbd172fe3b9c7e8f4;hpb=e130ada24310f0e3adc0349c0fb2098ad20d65d9 diff --git a/src/broadcom/compiler/nir_to_vir.c b/src/broadcom/compiler/nir_to_vir.c index 4cbdb764eab..b05c892bba6 100644 --- a/src/broadcom/compiler/nir_to_vir.c +++ b/src/broadcom/compiler/nir_to_vir.c @@ -32,6 +32,57 @@ #include "common/v3d_device_info.h" #include "v3d_compiler.h" +#define GENERAL_TMU_LOOKUP_PER_QUAD (0 << 7) +#define GENERAL_TMU_LOOKUP_PER_PIXEL (1 << 7) +#define GENERAL_TMU_READ_OP_PREFETCH (0 << 3) +#define GENERAL_TMU_READ_OP_CACHE_CLEAR (1 << 3) +#define GENERAL_TMU_READ_OP_CACHE_FLUSH (3 << 3) +#define GENERAL_TMU_READ_OP_CACHE_CLEAN (3 << 3) +#define GENERAL_TMU_READ_OP_CACHE_L1T_CLEAR (4 << 3) +#define GENERAL_TMU_READ_OP_CACHE_L1T_FLUSH_AGGREGATION (5 << 3) +#define GENERAL_TMU_READ_OP_ATOMIC_INC (8 << 3) +#define GENERAL_TMU_READ_OP_ATOMIC_DEC (9 << 3) +#define GENERAL_TMU_READ_OP_ATOMIC_NOT (10 << 3) +#define GENERAL_TMU_READ_OP_READ (15 << 3) +#define GENERAL_TMU_LOOKUP_TYPE_8BIT_I (0 << 0) +#define GENERAL_TMU_LOOKUP_TYPE_16BIT_I (1 << 0) +#define GENERAL_TMU_LOOKUP_TYPE_VEC2 (2 << 0) +#define GENERAL_TMU_LOOKUP_TYPE_VEC3 (3 << 0) +#define GENERAL_TMU_LOOKUP_TYPE_VEC4 (4 << 0) +#define GENERAL_TMU_LOOKUP_TYPE_8BIT_UI (5 << 0) +#define GENERAL_TMU_LOOKUP_TYPE_16BIT_UI (6 << 0) +#define GENERAL_TMU_LOOKUP_TYPE_32BIT_UI (7 << 0) + +#define GENERAL_TMU_WRITE_OP_ATOMIC_ADD_WRAP (0 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_SUB_WRAP (1 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_XCHG (2 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG (3 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_UMIN (4 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_UMAX (5 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_SMIN (6 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_SMAX (7 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_AND (8 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_OR (9 << 3) +#define GENERAL_TMU_WRITE_OP_ATOMIC_XOR (10 << 3) +#define GENERAL_TMU_WRITE_OP_WRITE (15 << 3) + +#define V3D_TSY_SET_QUORUM 0 +#define V3D_TSY_INC_WAITERS 1 +#define V3D_TSY_DEC_WAITERS 2 +#define V3D_TSY_INC_QUORUM 3 +#define V3D_TSY_DEC_QUORUM 4 +#define V3D_TSY_FREE_ALL 5 +#define V3D_TSY_RELEASE 6 +#define V3D_TSY_ACQUIRE 7 +#define V3D_TSY_WAIT 8 +#define V3D_TSY_WAIT_INC 9 +#define V3D_TSY_WAIT_CHECK 10 +#define V3D_TSY_WAIT_INC_CHECK 11 +#define V3D_TSY_WAIT_CV 12 +#define V3D_TSY_INC_SEMAPHORE 13 +#define V3D_TSY_DEC_SEMAPHORE 14 +#define V3D_TSY_SET_QUORUM_FREE_ALL 15 + static void ntq_emit_cf_list(struct v3d_compile *c, struct exec_list *list); @@ -70,57 +121,215 @@ vir_emit_thrsw(struct v3d_compile *c) */ c->last_thrsw = vir_NOP(c); c->last_thrsw->qpu.sig.thrsw = true; - c->last_thrsw_at_top_level = (c->execute.file == QFILE_NULL); + c->last_thrsw_at_top_level = !c->in_control_flow; } -static struct qreg -vir_SFU(struct v3d_compile *c, int waddr, struct qreg src) +static uint32_t +v3d_general_tmu_op(nir_intrinsic_instr *instr) { - vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, waddr), src); - return vir_FMOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); + switch (instr->intrinsic) { + case nir_intrinsic_load_ssbo: + case nir_intrinsic_load_ubo: + case nir_intrinsic_load_uniform: + case nir_intrinsic_load_shared: + case nir_intrinsic_load_scratch: + return GENERAL_TMU_READ_OP_READ; + case nir_intrinsic_store_ssbo: + case nir_intrinsic_store_shared: + case nir_intrinsic_store_scratch: + return GENERAL_TMU_WRITE_OP_WRITE; + case nir_intrinsic_ssbo_atomic_add: + case nir_intrinsic_shared_atomic_add: + return GENERAL_TMU_WRITE_OP_ATOMIC_ADD_WRAP; + case nir_intrinsic_ssbo_atomic_imin: + case nir_intrinsic_shared_atomic_imin: + return GENERAL_TMU_WRITE_OP_ATOMIC_SMIN; + case nir_intrinsic_ssbo_atomic_umin: + case nir_intrinsic_shared_atomic_umin: + return GENERAL_TMU_WRITE_OP_ATOMIC_UMIN; + case nir_intrinsic_ssbo_atomic_imax: + case nir_intrinsic_shared_atomic_imax: + return GENERAL_TMU_WRITE_OP_ATOMIC_SMAX; + case nir_intrinsic_ssbo_atomic_umax: + case nir_intrinsic_shared_atomic_umax: + return GENERAL_TMU_WRITE_OP_ATOMIC_UMAX; + case nir_intrinsic_ssbo_atomic_and: + case nir_intrinsic_shared_atomic_and: + return GENERAL_TMU_WRITE_OP_ATOMIC_AND; + case nir_intrinsic_ssbo_atomic_or: + case nir_intrinsic_shared_atomic_or: + return GENERAL_TMU_WRITE_OP_ATOMIC_OR; + case nir_intrinsic_ssbo_atomic_xor: + case nir_intrinsic_shared_atomic_xor: + return GENERAL_TMU_WRITE_OP_ATOMIC_XOR; + case nir_intrinsic_ssbo_atomic_exchange: + case nir_intrinsic_shared_atomic_exchange: + return GENERAL_TMU_WRITE_OP_ATOMIC_XCHG; + case nir_intrinsic_ssbo_atomic_comp_swap: + case nir_intrinsic_shared_atomic_comp_swap: + return GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG; + default: + unreachable("unknown intrinsic op"); + } } -static struct qreg -indirect_uniform_load(struct v3d_compile *c, nir_intrinsic_instr *intr) +/** + * Implements indirect uniform loads and SSBO accesses through the TMU general + * memory access interface. + */ +static void +ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr, + bool is_shared_or_scratch) { - struct qreg indirect_offset = ntq_get_src(c, intr->src[0], 0); - uint32_t offset = nir_intrinsic_base(intr); - struct v3d_ubo_range *range = NULL; - unsigned i; - - for (i = 0; i < c->num_ubo_ranges; i++) { - range = &c->ubo_ranges[i]; - if (offset >= range->src_offset && - offset < range->src_offset + range->size) { - break; + /* XXX perf: We should turn add/sub of 1 to inc/dec. Perhaps NIR + * wants to have support for inc/dec? + */ + + uint32_t tmu_op = v3d_general_tmu_op(instr); + bool is_store = (instr->intrinsic == nir_intrinsic_store_ssbo || + instr->intrinsic == nir_intrinsic_store_scratch || + instr->intrinsic == nir_intrinsic_store_shared); + bool has_index = !is_shared_or_scratch; + + int offset_src; + int tmu_writes = 1; /* address */ + if (instr->intrinsic == nir_intrinsic_load_uniform) { + offset_src = 0; + } else if (instr->intrinsic == nir_intrinsic_load_ssbo || + instr->intrinsic == nir_intrinsic_load_ubo || + instr->intrinsic == nir_intrinsic_load_scratch || + instr->intrinsic == nir_intrinsic_load_shared) { + offset_src = 0 + has_index; + } else if (is_store) { + offset_src = 1 + has_index; + for (int i = 0; i < instr->num_components; i++) { + vir_MOV_dest(c, + vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD), + ntq_get_src(c, instr->src[0], i)); + tmu_writes++; + } + } else { + offset_src = 0 + has_index; + vir_MOV_dest(c, + vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD), + ntq_get_src(c, instr->src[1 + has_index], 0)); + tmu_writes++; + if (tmu_op == GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG) { + vir_MOV_dest(c, + vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD), + ntq_get_src(c, instr->src[2 + has_index], + 0)); + tmu_writes++; + } + } + + bool dynamic_src = !nir_src_is_const(instr->src[offset_src]); + uint32_t const_offset = 0; + if (!dynamic_src) + const_offset = nir_src_as_uint(instr->src[offset_src]); + + /* Make sure we won't exceed the 16-entry TMU fifo if each thread is + * storing at the same time. + */ + while (tmu_writes > 16 / c->threads) + c->threads /= 2; + + struct qreg offset; + if (instr->intrinsic == nir_intrinsic_load_uniform) { + const_offset += nir_intrinsic_base(instr); + offset = vir_uniform(c, QUNIFORM_UBO_ADDR, + v3d_unit_data_create(0, const_offset)); + const_offset = 0; + } else if (instr->intrinsic == nir_intrinsic_load_ubo) { + uint32_t index = nir_src_as_uint(instr->src[0]) + 1; + /* Note that QUNIFORM_UBO_ADDR takes a UBO index shifted up by + * 1 (0 is gallium's constant buffer 0). + */ + offset = vir_uniform(c, QUNIFORM_UBO_ADDR, + v3d_unit_data_create(index, const_offset)); + const_offset = 0; + } else if (is_shared_or_scratch) { + /* Shared and scratch variables have no buffer index, and all + * start from a common base that we set up at the start of + * dispatch. + */ + if (instr->intrinsic == nir_intrinsic_load_scratch || + instr->intrinsic == nir_intrinsic_store_scratch) { + offset = c->spill_base; + } else { + offset = c->cs_shared_offset; + const_offset += nir_intrinsic_base(instr); } + } else { + offset = vir_uniform(c, QUNIFORM_SSBO_OFFSET, + nir_src_as_uint(instr->src[is_store ? + 1 : 0])); } - /* The driver-location-based offset always has to be within a declared - * uniform range. + + /* The spec says that for atomics, the TYPE field is ignored, but that + * doesn't seem to be the case for CMPXCHG. Just use the number of + * tmud writes we did to decide the type (or choose "32bit" for atomic + * reads, which has been fine). */ - assert(i != c->num_ubo_ranges); - if (!c->ubo_range_used[i]) { - c->ubo_range_used[i] = true; - range->dst_offset = c->next_ubo_dst_offset; - c->next_ubo_dst_offset += range->size; + int num_components; + if (tmu_op == GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG) + num_components = 2; + else + num_components = instr->num_components; + + uint32_t config = (0xffffff00 | + tmu_op | + GENERAL_TMU_LOOKUP_PER_PIXEL); + if (num_components == 1) { + config |= GENERAL_TMU_LOOKUP_TYPE_32BIT_UI; + } else { + config |= GENERAL_TMU_LOOKUP_TYPE_VEC2 + num_components - 2; + } + + if (vir_in_nonuniform_control_flow(c)) { + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), + V3D_QPU_PF_PUSHZ); } - offset -= range->src_offset; + struct qreg tmua; + if (config == ~0) + tmua = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUA); + else + tmua = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUAU); + + struct qinst *tmu; + if (dynamic_src) { + if (const_offset != 0) { + offset = vir_ADD(c, offset, + vir_uniform_ui(c, const_offset)); + } + tmu = vir_ADD_dest(c, tmua, offset, + ntq_get_src(c, instr->src[offset_src], 0)); + } else { + if (const_offset != 0) { + tmu = vir_ADD_dest(c, tmua, offset, + vir_uniform_ui(c, const_offset)); + } else { + tmu = vir_MOV_dest(c, tmua, offset); + } + } - if (range->dst_offset + offset != 0) { - indirect_offset = vir_ADD(c, indirect_offset, - vir_uniform_ui(c, range->dst_offset + - offset)); + if (config != ~0) { + tmu->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT, + config); } - /* Adjust for where we stored the TGSI register base. */ - vir_ADD_dest(c, - vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUA), - vir_uniform(c, QUNIFORM_UBO_ADDR, 0), - indirect_offset); + if (vir_in_nonuniform_control_flow(c)) + vir_set_cond(tmu, V3D_QPU_COND_IFA); vir_emit_thrsw(c); - return vir_LDTMU(c); + + /* Read the result, or wait for the TMU op to complete. */ + for (int i = 0; i < nir_intrinsic_dest_components(instr); i++) + ntq_store_dest(c, &instr->dest, i, vir_MOV(c, vir_LDTMU(c))); + + if (nir_intrinsic_dest_components(instr) == 0) + vir_TMUWT(c); } static struct qreg * @@ -155,8 +364,7 @@ ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan, if (!list_empty(&c->cur_block->instructions)) last_inst = (struct qinst *)c->cur_block->instructions.prev; - assert(result.file == QFILE_UNIF || - (result.file == QFILE_TEMP && + assert((result.file == QFILE_TEMP && last_inst && last_inst == c->defs[result.index])); if (dest->is_ssa) { @@ -183,7 +391,8 @@ ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan, /* Insert a MOV if the source wasn't an SSA def in the * previous instruction. */ - if (result.file == QFILE_UNIF) { + if ((vir_in_nonuniform_control_flow(c) && + c->defs[last_inst->dst.index]->qpu.sig.ldunif)) { result = vir_MOV(c, result); last_inst = c->defs[result.index]; } @@ -195,17 +404,17 @@ ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan, /* If we're in control flow, then make this update of the reg * conditional on the execution mask. */ - if (c->execute.file != QFILE_NULL) { + if (vir_in_nonuniform_control_flow(c)) { last_inst->dst.index = qregs[chan].index; /* Set the flags to the current exec mask. */ c->cursor = vir_before_inst(last_inst); - vir_PF(c, c->execute, V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), + V3D_QPU_PF_PUSHZ); c->cursor = vir_after_inst(last_inst); vir_set_cond(last_inst, V3D_QPU_COND_IFA); - last_inst->cond_is_exec_mask = true; } } } @@ -244,14 +453,6 @@ ntq_get_alu_src(struct v3d_compile *c, nir_alu_instr *instr, return r; }; -static inline struct qreg -vir_SAT(struct v3d_compile *c, struct qreg val) -{ - return vir_FMAX(c, - vir_FMIN(c, val, vir_uniform_f(c, 1.0)), - vir_uniform_f(c, 0.0)); -} - static struct qreg ntq_minify(struct v3d_compile *c, struct qreg size, struct qreg level) { @@ -283,6 +484,7 @@ ntq_emit_txs(struct v3d_compile *c, nir_tex_instr *instr) switch (instr->sampler_dim) { case GLSL_SAMPLER_DIM_1D: case GLSL_SAMPLER_DIM_2D: + case GLSL_SAMPLER_DIM_MS: case GLSL_SAMPLER_DIM_3D: case GLSL_SAMPLER_DIM_CUBE: /* Don't minify the array size. */ @@ -338,8 +540,7 @@ ntq_fsincos(struct v3d_compile *c, struct qreg src, bool is_cos) input = vir_FADD(c, input, vir_uniform_f(c, 0.5)); struct qreg periods = vir_FROUND(c, input); - struct qreg sin_output = vir_SFU(c, V3D_QPU_WADDR_SIN, - vir_FSUB(c, input, periods)); + struct qreg sin_output = vir_SIN(c, vir_FSUB(c, input, periods)); return vir_XOR(c, sin_output, vir_SHL(c, vir_FTOIN(c, periods), vir_uniform_ui(c, -1))); @@ -351,39 +552,25 @@ ntq_fsign(struct v3d_compile *c, struct qreg src) struct qreg t = vir_get_temp(c); vir_MOV_dest(c, t, vir_uniform_f(c, 0.0)); - vir_PF(c, vir_FMOV(c, src), V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_FMOV_dest(c, vir_nop_reg(), src), V3D_QPU_PF_PUSHZ); vir_MOV_cond(c, V3D_QPU_COND_IFNA, t, vir_uniform_f(c, 1.0)); - vir_PF(c, vir_FMOV(c, src), V3D_QPU_PF_PUSHN); + vir_set_pf(vir_FMOV_dest(c, vir_nop_reg(), src), V3D_QPU_PF_PUSHN); vir_MOV_cond(c, V3D_QPU_COND_IFA, t, vir_uniform_f(c, -1.0)); return vir_MOV(c, t); } -static struct qreg -ntq_isign(struct v3d_compile *c, struct qreg src) -{ - struct qreg t = vir_get_temp(c); - - vir_MOV_dest(c, t, vir_uniform_ui(c, 0)); - vir_PF(c, vir_MOV(c, src), V3D_QPU_PF_PUSHZ); - vir_MOV_cond(c, V3D_QPU_COND_IFNA, t, vir_uniform_ui(c, 1)); - vir_PF(c, vir_MOV(c, src), V3D_QPU_PF_PUSHN); - vir_MOV_cond(c, V3D_QPU_COND_IFA, t, vir_uniform_ui(c, -1)); - return vir_MOV(c, t); -} - static void emit_fragcoord_input(struct v3d_compile *c, int attr) { c->inputs[attr * 4 + 0] = vir_FXCD(c); c->inputs[attr * 4 + 1] = vir_FYCD(c); c->inputs[attr * 4 + 2] = c->payload_z; - c->inputs[attr * 4 + 3] = vir_SFU(c, V3D_QPU_WADDR_RECIP, - c->payload_w); + c->inputs[attr * 4 + 3] = vir_RECIP(c, c->payload_w); } static struct qreg emit_fragment_varying(struct v3d_compile *c, nir_variable *var, - uint8_t swizzle) + uint8_t swizzle, int array_index) { struct qreg r3 = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R3); struct qreg r5 = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R5); @@ -408,8 +595,9 @@ emit_fragment_varying(struct v3d_compile *c, nir_variable *var, } int i = c->num_inputs++; - c->input_slots[i] = v3d_slot_from_slot_and_component(var->data.location, - swizzle); + c->input_slots[i] = + v3d_slot_from_slot_and_component(var->data.location + + array_index, swizzle); switch (var->data.interpolation) { case INTERP_MODE_NONE: @@ -417,7 +605,7 @@ emit_fragment_varying(struct v3d_compile *c, nir_variable *var, * qualifier, then if we're using glShadeModel(GL_FLAT) it * needs to be flat shaded. */ - switch (var->data.location) { + switch (var->data.location + array_index) { case VARYING_SLOT_COL0: case VARYING_SLOT_COL1: case VARYING_SLOT_BFC0: @@ -443,9 +631,7 @@ emit_fragment_varying(struct v3d_compile *c, nir_variable *var, return vir_FADD(c, vir_FMUL(c, vary, c->payload_w), r5); } case INTERP_MODE_NOPERSPECTIVE: - /* C appears after the mov from the varying. - XXX: improve ldvary setup. - */ + BITSET_SET(c->noperspective_flags, i); return vir_FADD(c, vir_MOV(c, vary), r5); case INTERP_MODE_FLAT: BITSET_SET(c->flat_shade_flags, i); @@ -457,12 +643,13 @@ emit_fragment_varying(struct v3d_compile *c, nir_variable *var, } static void -emit_fragment_input(struct v3d_compile *c, int attr, nir_variable *var) +emit_fragment_input(struct v3d_compile *c, int attr, nir_variable *var, + int array_index) { for (int i = 0; i < glsl_get_vector_elements(var->type); i++) { int chan = var->data.location_frac + i; c->inputs[attr * 4 + chan] = - emit_fragment_varying(c, var, chan); + emit_fragment_varying(c, var, chan, array_index); } } @@ -487,148 +674,127 @@ add_output(struct v3d_compile *c, v3d_slot_from_slot_and_component(slot, swizzle); } -static void -declare_uniform_range(struct v3d_compile *c, uint32_t start, uint32_t size) -{ - unsigned array_id = c->num_ubo_ranges++; - if (array_id >= c->ubo_ranges_array_size) { - c->ubo_ranges_array_size = MAX2(c->ubo_ranges_array_size * 2, - array_id + 1); - c->ubo_ranges = reralloc(c, c->ubo_ranges, - struct v3d_ubo_range, - c->ubo_ranges_array_size); - c->ubo_range_used = reralloc(c, c->ubo_range_used, - bool, - c->ubo_ranges_array_size); - } - - c->ubo_ranges[array_id].dst_offset = 0; - c->ubo_ranges[array_id].src_offset = start; - c->ubo_ranges[array_id].size = size; - c->ubo_range_used[array_id] = false; -} - /** * If compare_instr is a valid comparison instruction, emits the * compare_instr's comparison and returns the sel_instr's return value based * on the compare_instr's result. */ static bool -ntq_emit_comparison(struct v3d_compile *c, struct qreg *dest, +ntq_emit_comparison(struct v3d_compile *c, nir_alu_instr *compare_instr, - nir_alu_instr *sel_instr) + enum v3d_qpu_cond *out_cond) { struct qreg src0 = ntq_get_alu_src(c, compare_instr, 0); struct qreg src1; if (nir_op_infos[compare_instr->op].num_inputs > 1) src1 = ntq_get_alu_src(c, compare_instr, 1); bool cond_invert = false; + struct qreg nop = vir_nop_reg(); switch (compare_instr->op) { - case nir_op_feq: + case nir_op_feq32: case nir_op_seq: - vir_PF(c, vir_FCMP(c, src0, src1), V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_FCMP_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ); break; - case nir_op_ieq: - vir_PF(c, vir_XOR(c, src0, src1), V3D_QPU_PF_PUSHZ); + case nir_op_ieq32: + vir_set_pf(vir_XOR_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ); break; - case nir_op_fne: + case nir_op_fne32: case nir_op_sne: - vir_PF(c, vir_FCMP(c, src0, src1), V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_FCMP_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ); cond_invert = true; break; - case nir_op_ine: - vir_PF(c, vir_XOR(c, src0, src1), V3D_QPU_PF_PUSHZ); + case nir_op_ine32: + vir_set_pf(vir_XOR_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ); cond_invert = true; break; - case nir_op_fge: + case nir_op_fge32: case nir_op_sge: - vir_PF(c, vir_FCMP(c, src1, src0), V3D_QPU_PF_PUSHC); + vir_set_pf(vir_FCMP_dest(c, nop, src1, src0), V3D_QPU_PF_PUSHC); break; - case nir_op_ige: - vir_PF(c, vir_MIN(c, src1, src0), V3D_QPU_PF_PUSHC); + case nir_op_ige32: + vir_set_pf(vir_MIN_dest(c, nop, src1, src0), V3D_QPU_PF_PUSHC); cond_invert = true; break; - case nir_op_uge: - vir_PF(c, vir_SUB(c, src0, src1), V3D_QPU_PF_PUSHC); + case nir_op_uge32: + vir_set_pf(vir_SUB_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHC); cond_invert = true; break; case nir_op_slt: - case nir_op_flt: - vir_PF(c, vir_FCMP(c, src0, src1), V3D_QPU_PF_PUSHN); + case nir_op_flt32: + vir_set_pf(vir_FCMP_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHN); break; - case nir_op_ilt: - vir_PF(c, vir_MIN(c, src1, src0), V3D_QPU_PF_PUSHC); + case nir_op_ilt32: + vir_set_pf(vir_MIN_dest(c, nop, src1, src0), V3D_QPU_PF_PUSHC); break; - case nir_op_ult: - vir_PF(c, vir_SUB(c, src0, src1), V3D_QPU_PF_PUSHC); + case nir_op_ult32: + vir_set_pf(vir_SUB_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHC); break; - default: - return false; - } - - enum v3d_qpu_cond cond = (cond_invert ? - V3D_QPU_COND_IFNA : - V3D_QPU_COND_IFA); - - switch (sel_instr->op) { - case nir_op_seq: - case nir_op_sne: - case nir_op_sge: - case nir_op_slt: - *dest = vir_SEL(c, cond, - vir_uniform_f(c, 1.0), vir_uniform_f(c, 0.0)); + case nir_op_i2b32: + vir_set_pf(vir_MOV_dest(c, nop, src0), V3D_QPU_PF_PUSHZ); + cond_invert = true; break; - case nir_op_bcsel: - *dest = vir_SEL(c, cond, - ntq_get_alu_src(c, sel_instr, 1), - ntq_get_alu_src(c, sel_instr, 2)); + case nir_op_f2b32: + vir_set_pf(vir_FMOV_dest(c, nop, src0), V3D_QPU_PF_PUSHZ); + cond_invert = true; break; default: - *dest = vir_SEL(c, cond, - vir_uniform_ui(c, ~0), vir_uniform_ui(c, 0)); - break; + return false; } - /* Make the temporary for nir_store_dest(). */ - *dest = vir_MOV(c, *dest); + *out_cond = cond_invert ? V3D_QPU_COND_IFNA : V3D_QPU_COND_IFA; return true; } -/** - * Attempts to fold a comparison generating a boolean result into the - * condition code for selecting between two values, instead of comparing the - * boolean result against 0 to generate the condition code. +/* Finds an ALU instruction that generates our src value that could + * (potentially) be greedily emitted in the consuming instruction. */ -static struct qreg ntq_emit_bcsel(struct v3d_compile *c, nir_alu_instr *instr, - struct qreg *src) +static struct nir_alu_instr * +ntq_get_alu_parent(nir_src src) { - if (!instr->src[0].src.is_ssa) - goto out; - if (instr->src[0].src.ssa->parent_instr->type != nir_instr_type_alu) - goto out; - nir_alu_instr *compare = - nir_instr_as_alu(instr->src[0].src.ssa->parent_instr); + if (!src.is_ssa || src.ssa->parent_instr->type != nir_instr_type_alu) + return NULL; + nir_alu_instr *instr = nir_instr_as_alu(src.ssa->parent_instr); + if (!instr) + return NULL; + + /* If the ALU instr's srcs are non-SSA, then we would have to avoid + * moving emission of the ALU instr down past another write of the + * src. + */ + for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (!instr->src[i].src.is_ssa) + return NULL; + } + + return instr; +} + +/* Turns a NIR bool into a condition code to predicate on. */ +static enum v3d_qpu_cond +ntq_emit_bool_to_cond(struct v3d_compile *c, nir_src src) +{ + nir_alu_instr *compare = ntq_get_alu_parent(src); if (!compare) goto out; - struct qreg dest; - if (ntq_emit_comparison(c, &dest, compare, instr)) - return dest; + enum v3d_qpu_cond cond; + if (ntq_emit_comparison(c, compare, &cond)) + return cond; out: - vir_PF(c, src[0], V3D_QPU_PF_PUSHZ); - return vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFNA, src[1], src[2])); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), ntq_get_src(c, src, 0)), + V3D_QPU_PF_PUSHZ); + return V3D_QPU_COND_IFNA; } - static void ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) { @@ -661,8 +827,7 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) struct qreg result; switch (instr->op) { - case nir_op_fmov: - case nir_op_imov: + case nir_op_mov: result = vir_MOV(c, src[0]); break; @@ -689,9 +854,16 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) result = vir_FMAX(c, src[0], src[1]); break; - case nir_op_f2i32: - result = vir_FTOIZ(c, src[0]); + case nir_op_f2i32: { + nir_alu_instr *src0_alu = ntq_get_alu_parent(instr->src[0].src); + if (src0_alu && src0_alu->op == nir_op_fround_even) { + result = vir_FTOIN(c, ntq_get_alu_src(c, src0_alu, 0)); + } else { + result = vir_FTOIZ(c, src[0]); + } break; + } + case nir_op_f2u32: result = vir_FTOUZ(c, src[0]); break; @@ -701,19 +873,12 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) case nir_op_u2f32: result = vir_UTOF(c, src[0]); break; - case nir_op_b2f: + case nir_op_b2f32: result = vir_AND(c, src[0], vir_uniform_f(c, 1.0)); break; - case nir_op_b2i: + case nir_op_b2i32: result = vir_AND(c, src[0], vir_uniform_ui(c, 1)); break; - case nir_op_i2b: - case nir_op_f2b: - vir_PF(c, src[0], V3D_QPU_PF_PUSHZ); - result = vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFNA, - vir_uniform_ui(c, ~0), - vir_uniform_ui(c, 0))); - break; case nir_op_iadd: result = vir_ADD(c, src[0], src[1]); @@ -766,42 +931,62 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) case nir_op_seq: case nir_op_sne: case nir_op_sge: - case nir_op_slt: - case nir_op_feq: - case nir_op_fne: - case nir_op_fge: - case nir_op_flt: - case nir_op_ieq: - case nir_op_ine: - case nir_op_ige: - case nir_op_uge: - case nir_op_ilt: - case nir_op_ult: - if (!ntq_emit_comparison(c, &result, instr, instr)) { - fprintf(stderr, "Bad comparison instruction\n"); - } + case nir_op_slt: { + enum v3d_qpu_cond cond; + MAYBE_UNUSED bool ok = ntq_emit_comparison(c, instr, &cond); + assert(ok); + result = vir_MOV(c, vir_SEL(c, cond, + vir_uniform_f(c, 1.0), + vir_uniform_f(c, 0.0))); break; + } - case nir_op_bcsel: - result = ntq_emit_bcsel(c, instr, src); + case nir_op_i2b32: + case nir_op_f2b32: + case nir_op_feq32: + case nir_op_fne32: + case nir_op_fge32: + case nir_op_flt32: + case nir_op_ieq32: + case nir_op_ine32: + case nir_op_ige32: + case nir_op_uge32: + case nir_op_ilt32: + case nir_op_ult32: { + enum v3d_qpu_cond cond; + MAYBE_UNUSED bool ok = ntq_emit_comparison(c, instr, &cond); + assert(ok); + result = vir_MOV(c, vir_SEL(c, cond, + vir_uniform_ui(c, ~0), + vir_uniform_ui(c, 0))); break; + } + + case nir_op_b32csel: + result = vir_MOV(c, + vir_SEL(c, + ntq_emit_bool_to_cond(c, instr->src[0].src), + src[1], src[2])); + break; + case nir_op_fcsel: - vir_PF(c, src[0], V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), src[0]), + V3D_QPU_PF_PUSHZ); result = vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFNA, src[1], src[2])); break; case nir_op_frcp: - result = vir_SFU(c, V3D_QPU_WADDR_RECIP, src[0]); + result = vir_RECIP(c, src[0]); break; case nir_op_frsq: - result = vir_SFU(c, V3D_QPU_WADDR_RSQRT, src[0]); + result = vir_RSQRT(c, src[0]); break; case nir_op_fexp2: - result = vir_SFU(c, V3D_QPU_WADDR_EXP, src[0]); + result = vir_EXP(c, src[0]); break; case nir_op_flog2: - result = vir_SFU(c, V3D_QPU_WADDR_LOG, src[0]); + result = vir_LOG(c, src[0]); break; case nir_op_fceil: @@ -816,9 +1001,6 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) case nir_op_ftrunc: result = vir_FTRUNC(c, src[0]); break; - case nir_op_ffract: - result = vir_FSUB(c, src[0], vir_FFLOOR(c, src[0])); - break; case nir_op_fsin: result = ntq_fsincos(c, src[0], false); @@ -830,9 +1012,6 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) case nir_op_fsign: result = ntq_fsign(c, src[0]); break; - case nir_op_isign: - result = ntq_isign(c, src[0]); - break; case nir_op_fabs: { result = vir_FMOV(c, src[0]); @@ -841,8 +1020,7 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) } case nir_op_iabs: - result = vir_MAX(c, src[0], - vir_SUB(c, vir_uniform_ui(c, 0), src[0])); + result = vir_MAX(c, src[0], vir_NEG(c, src[0])); break; case nir_op_fddx: @@ -858,12 +1036,27 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) break; case nir_op_uadd_carry: - vir_PF(c, vir_ADD(c, src[0], src[1]), V3D_QPU_PF_PUSHC); + vir_set_pf(vir_ADD_dest(c, vir_nop_reg(), src[0], src[1]), + V3D_QPU_PF_PUSHC); result = vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFA, vir_uniform_ui(c, ~0), vir_uniform_ui(c, 0))); break; + case nir_op_pack_half_2x16_split: + result = vir_VFPACK(c, src[0], src[1]); + break; + + case nir_op_unpack_half_2x16_split_x: + result = vir_FMOV(c, src[0]); + vir_set_unpack(c->defs[result.index], 0, V3D_QPU_UNPACK_L); + break; + + case nir_op_unpack_half_2x16_split_y: + result = vir_FMOV(c, src[0]); + vir_set_unpack(c->defs[result.index], 0, V3D_QPU_UNPACK_H); + break; + default: fprintf(stderr, "unknown NIR ALU inst: "); nir_print_instr(&instr->instr, stderr); @@ -901,6 +1094,8 @@ ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr) #define TLB_TYPE_DEPTH ((2 << 6) | (0 << 4)) #define TLB_DEPTH_TYPE_INVARIANT (0 << 2) /* Unmodified sideband input used */ #define TLB_DEPTH_TYPE_PER_PIXEL (1 << 2) /* QPU result used */ +#define TLB_V42_DEPTH_TYPE_INVARIANT (0 << 3) /* Unmodified sideband input used */ +#define TLB_V42_DEPTH_TYPE_PER_PIXEL (1 << 3) /* QPU result used */ /* Stencil is a single 32-bit write. */ #define TLB_TYPE_STENCIL_ALPHA ((2 << 6) | (1 << 4)) @@ -915,22 +1110,42 @@ emit_frag_end(struct v3d_compile *c) */ bool has_any_tlb_color_write = false; - for (int rt = 0; rt < c->fs_key->nr_cbufs; rt++) { - if (c->output_color_var[rt]) + for (int rt = 0; rt < V3D_MAX_DRAW_BUFFERS; rt++) { + if (c->fs_key->cbufs & (1 << rt) && c->output_color_var[rt]) has_any_tlb_color_write = true; } + if (c->fs_key->sample_alpha_to_coverage && c->output_color_var[0]) { + struct nir_variable *var = c->output_color_var[0]; + struct qreg *color = &c->outputs[var->data.driver_location * 4]; + + vir_SETMSF_dest(c, vir_nop_reg(), + vir_AND(c, + vir_MSF(c), + vir_FTOC(c, color[3]))); + } + + struct qreg tlb_reg = vir_magic_reg(V3D_QPU_WADDR_TLB); + struct qreg tlbu_reg = vir_magic_reg(V3D_QPU_WADDR_TLBU); if (c->output_position_index != -1) { - struct qinst *inst = vir_MOV_dest(c, - vir_reg(QFILE_TLBU, 0), + struct qinst *inst = vir_MOV_dest(c, tlbu_reg, c->outputs[c->output_position_index]); - - inst->src[vir_get_implicit_uniform_src(inst)] = - vir_uniform_ui(c, - TLB_TYPE_DEPTH | - TLB_DEPTH_TYPE_PER_PIXEL | - 0xffffff00); - } else if (c->s->info.fs.uses_discard || !has_any_tlb_color_write) { + uint8_t tlb_specifier = TLB_TYPE_DEPTH; + + if (c->devinfo->ver >= 42) { + tlb_specifier |= (TLB_V42_DEPTH_TYPE_PER_PIXEL | + TLB_SAMPLE_MODE_PER_PIXEL); + } else + tlb_specifier |= TLB_DEPTH_TYPE_PER_PIXEL; + + inst->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT, + tlb_specifier | + 0xffffff00); + c->writes_z = true; + } else if (c->s->info.fs.uses_discard || + !c->s->info.fs.early_fragment_tests || + c->fs_key->sample_alpha_to_coverage || + !has_any_tlb_color_write) { /* Emit passthrough Z if it needed to be delayed until shader * end due to potential discards. * @@ -941,23 +1156,33 @@ emit_frag_end(struct v3d_compile *c) */ c->s->info.fs.uses_discard = true; - struct qinst *inst = vir_MOV_dest(c, - vir_reg(QFILE_TLBU, 0), - vir_reg(QFILE_NULL, 0)); + struct qinst *inst = vir_MOV_dest(c, tlbu_reg, + vir_nop_reg()); + uint8_t tlb_specifier = TLB_TYPE_DEPTH; + + if (c->devinfo->ver >= 42) { + /* The spec says the PER_PIXEL flag is ignored for + * invariant writes, but the simulator demands it. + */ + tlb_specifier |= (TLB_V42_DEPTH_TYPE_INVARIANT | + TLB_SAMPLE_MODE_PER_PIXEL); + } else { + tlb_specifier |= TLB_DEPTH_TYPE_INVARIANT; + } - inst->src[vir_get_implicit_uniform_src(inst)] = - vir_uniform_ui(c, - TLB_TYPE_DEPTH | - TLB_DEPTH_TYPE_INVARIANT | - 0xffffff00); + inst->uniform = vir_get_uniform_index(c, + QUNIFORM_CONSTANT, + tlb_specifier | + 0xffffff00); + c->writes_z = true; } /* XXX: Performance improvement: Merge Z write and color writes TLB * uniform setup */ - for (int rt = 0; rt < c->fs_key->nr_cbufs; rt++) { - if (!c->output_color_var[rt]) + for (int rt = 0; rt < V3D_MAX_DRAW_BUFFERS; rt++) { + if (!(c->fs_key->cbufs & (1 << rt)) || !c->output_color_var[rt]) continue; nir_variable *var = c->output_color_var[rt]; @@ -984,13 +1209,13 @@ emit_frag_end(struct v3d_compile *c) conf |= ((num_components - 1) << TLB_VEC_SIZE_MINUS_1_SHIFT); - inst = vir_MOV_dest(c, vir_reg(QFILE_TLBU, 0), color[0]); - inst->src[vir_get_implicit_uniform_src(inst)] = - vir_uniform_ui(c, conf); + inst = vir_MOV_dest(c, tlbu_reg, color[0]); + inst->uniform = vir_get_uniform_index(c, + QUNIFORM_CONSTANT, + conf); for (int i = 1; i < num_components; i++) { - inst = vir_MOV_dest(c, vir_reg(QFILE_TLB, 0), - color[i]); + inst = vir_MOV_dest(c, tlb_reg, color[i]); } break; @@ -1018,25 +1243,32 @@ emit_frag_end(struct v3d_compile *c) b = color[0]; } + if (c->fs_key->sample_alpha_to_one) + a = vir_uniform_f(c, 1.0); + if (c->fs_key->f32_color_rb & (1 << rt)) { - inst = vir_MOV_dest(c, vir_reg(QFILE_TLBU, 0), color[0]); - inst->src[vir_get_implicit_uniform_src(inst)] = - vir_uniform_ui(c, conf); + inst = vir_MOV_dest(c, tlbu_reg, r); + inst->uniform = vir_get_uniform_index(c, + QUNIFORM_CONSTANT, + conf); - for (int i = 1; i < num_components; i++) { - inst = vir_MOV_dest(c, vir_reg(QFILE_TLB, 0), - color[i]); - } + if (num_components >= 2) + vir_MOV_dest(c, tlb_reg, g); + if (num_components >= 3) + vir_MOV_dest(c, tlb_reg, b); + if (num_components >= 4) + vir_MOV_dest(c, tlb_reg, a); } else { - inst = vir_VFPACK_dest(c, vir_reg(QFILE_TLB, 0), r, g); + inst = vir_VFPACK_dest(c, tlb_reg, r, g); if (conf != ~0) { - inst->dst.file = QFILE_TLBU; - inst->src[vir_get_implicit_uniform_src(inst)] = - vir_uniform_ui(c, conf); + inst->dst = tlbu_reg; + inst->uniform = vir_get_uniform_index(c, + QUNIFORM_CONSTANT, + conf); } if (num_components >= 3) - inst = vir_VFPACK_dest(c, vir_reg(QFILE_TLB, 0), b, a); + inst = vir_VFPACK_dest(c, tlb_reg, b, a); } break; } @@ -1045,151 +1277,19 @@ emit_frag_end(struct v3d_compile *c) } static void -vir_VPM_WRITE(struct v3d_compile *c, struct qreg val, uint32_t *vpm_index) +vir_VPM_WRITE(struct v3d_compile *c, struct qreg val, uint32_t vpm_index) { if (c->devinfo->ver >= 40) { - vir_STVPMV(c, vir_uniform_ui(c, *vpm_index), val); - *vpm_index = *vpm_index + 1; + vir_STVPMV(c, vir_uniform_ui(c, vpm_index), val); } else { + /* XXX: v3d33_vir_vpm_write_setup(c); */ vir_MOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_VPM), val); } - - c->num_vpm_writes++; -} - -static void -emit_scaled_viewport_write(struct v3d_compile *c, struct qreg rcp_w, - uint32_t *vpm_index) -{ - for (int i = 0; i < 2; i++) { - struct qreg coord = c->outputs[c->output_position_index + i]; - coord = vir_FMUL(c, coord, - vir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, - 0)); - coord = vir_FMUL(c, coord, rcp_w); - vir_VPM_WRITE(c, vir_FTOIN(c, coord), vpm_index); - } - -} - -static void -emit_zs_write(struct v3d_compile *c, struct qreg rcp_w, uint32_t *vpm_index) -{ - struct qreg zscale = vir_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0); - struct qreg zoffset = vir_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0); - - struct qreg z = c->outputs[c->output_position_index + 2]; - z = vir_FMUL(c, z, zscale); - z = vir_FMUL(c, z, rcp_w); - z = vir_FADD(c, z, zoffset); - vir_VPM_WRITE(c, z, vpm_index); -} - -static void -emit_rcp_wc_write(struct v3d_compile *c, struct qreg rcp_w, uint32_t *vpm_index) -{ - vir_VPM_WRITE(c, rcp_w, vpm_index); -} - -static void -emit_point_size_write(struct v3d_compile *c, uint32_t *vpm_index) -{ - struct qreg point_size; - - if (c->output_point_size_index != -1) - point_size = c->outputs[c->output_point_size_index]; - else - point_size = vir_uniform_f(c, 1.0); - - /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835, - * BCM21553). - */ - point_size = vir_FMAX(c, point_size, vir_uniform_f(c, .125)); - - vir_VPM_WRITE(c, point_size, vpm_index); -} - -static void -emit_vpm_write_setup(struct v3d_compile *c) -{ - if (c->devinfo->ver >= 40) - return; - - v3d33_vir_vpm_write_setup(c); -} - -/** - * Sets up c->outputs[c->output_position_index] for the vertex shader - * epilogue, if an output vertex position wasn't specified in the user's - * shader. This may be the case for transform feedback with rasterizer - * discard enabled. - */ -static void -setup_default_position(struct v3d_compile *c) -{ - if (c->output_position_index != -1) - return; - - c->output_position_index = c->outputs_array_size; - for (int i = 0; i < 4; i++) { - add_output(c, - c->output_position_index + i, - VARYING_SLOT_POS, i); - } } static void emit_vert_end(struct v3d_compile *c) { - setup_default_position(c); - - uint32_t vpm_index = 0; - struct qreg rcp_w = vir_SFU(c, V3D_QPU_WADDR_RECIP, - c->outputs[c->output_position_index + 3]); - - emit_vpm_write_setup(c); - - if (c->vs_key->is_coord) { - for (int i = 0; i < 4; i++) - vir_VPM_WRITE(c, c->outputs[c->output_position_index + i], - &vpm_index); - emit_scaled_viewport_write(c, rcp_w, &vpm_index); - if (c->vs_key->per_vertex_point_size) { - emit_point_size_write(c, &vpm_index); - /* emit_rcp_wc_write(c, rcp_w); */ - } - /* XXX: Z-only rendering */ - if (0) - emit_zs_write(c, rcp_w, &vpm_index); - } else { - emit_scaled_viewport_write(c, rcp_w, &vpm_index); - emit_zs_write(c, rcp_w, &vpm_index); - emit_rcp_wc_write(c, rcp_w, &vpm_index); - if (c->vs_key->per_vertex_point_size) - emit_point_size_write(c, &vpm_index); - } - - for (int i = 0; i < c->vs_key->num_fs_inputs; i++) { - struct v3d_varying_slot input = c->vs_key->fs_inputs[i]; - int j; - - for (j = 0; j < c->num_outputs; j++) { - struct v3d_varying_slot output = c->output_slots[j]; - - if (!memcmp(&input, &output, sizeof(input))) { - vir_VPM_WRITE(c, c->outputs[j], - &vpm_index); - break; - } - } - /* Emit padding if we didn't find a declared VS output for - * this FS input. - */ - if (j == c->num_outputs) - vir_VPM_WRITE(c, vir_uniform_f(c, 0.0), - &vpm_index); - } - /* GFXH-1684: VPM writes need to be complete by the end of the shader. */ if (c->devinfo->ver >= 40 && c->devinfo->ver <= 42) @@ -1200,21 +1300,44 @@ void v3d_optimize_nir(struct nir_shader *s) { bool progress; + unsigned lower_flrp = + (s->options->lower_flrp16 ? 16 : 0) | + (s->options->lower_flrp32 ? 32 : 0) | + (s->options->lower_flrp64 ? 64 : 0); do { progress = false; NIR_PASS_V(s, nir_lower_vars_to_ssa); - NIR_PASS(progress, s, nir_lower_alu_to_scalar); + NIR_PASS(progress, s, nir_lower_alu_to_scalar, NULL); NIR_PASS(progress, s, nir_lower_phis_to_scalar); NIR_PASS(progress, s, nir_copy_prop); NIR_PASS(progress, s, nir_opt_remove_phis); NIR_PASS(progress, s, nir_opt_dce); NIR_PASS(progress, s, nir_opt_dead_cf); NIR_PASS(progress, s, nir_opt_cse); - NIR_PASS(progress, s, nir_opt_peephole_select, 8); + NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true); NIR_PASS(progress, s, nir_opt_algebraic); NIR_PASS(progress, s, nir_opt_constant_folding); + + if (lower_flrp != 0) { + bool lower_flrp_progress = false; + + NIR_PASS(lower_flrp_progress, s, nir_lower_flrp, + lower_flrp, + false /* always_precise */, + s->options->lower_ffma); + if (lower_flrp_progress) { + NIR_PASS(progress, s, nir_opt_constant_folding); + progress = true; + } + + /* Nothing should rematerialize any flrps, so we only + * need to do this lowering once. + */ + lower_flrp = 0; + } + NIR_PASS(progress, s, nir_opt_undef); } while (progress); @@ -1246,7 +1369,6 @@ ntq_emit_vpm_read(struct v3d_compile *c, if (*num_components_queued != 0) { (*num_components_queued)--; - c->num_inputs++; return vir_MOV(c, vpm); } @@ -1256,13 +1378,103 @@ ntq_emit_vpm_read(struct v3d_compile *c, *num_components_queued = num_components - 1; *remaining -= num_components; - c->num_inputs++; return vir_MOV(c, vpm); } static void -ntq_setup_inputs(struct v3d_compile *c) +ntq_setup_vpm_inputs(struct v3d_compile *c) +{ + /* Figure out how many components of each vertex attribute the shader + * uses. Each variable should have been split to individual + * components and unused ones DCEed. The vertex fetcher will load + * from the start of the attribute to the number of components we + * declare we need in c->vattr_sizes[]. + */ + nir_foreach_variable(var, &c->s->inputs) { + /* No VS attribute array support. */ + assert(MAX2(glsl_get_length(var->type), 1) == 1); + + unsigned loc = var->data.driver_location; + int start_component = var->data.location_frac; + int num_components = glsl_get_components(var->type); + + c->vattr_sizes[loc] = MAX2(c->vattr_sizes[loc], + start_component + num_components); + } + + unsigned num_components = 0; + uint32_t vpm_components_queued = 0; + bool uses_iid = c->s->info.system_values_read & + (1ull << SYSTEM_VALUE_INSTANCE_ID); + bool uses_vid = c->s->info.system_values_read & + (1ull << SYSTEM_VALUE_VERTEX_ID); + num_components += uses_iid; + num_components += uses_vid; + + for (int i = 0; i < ARRAY_SIZE(c->vattr_sizes); i++) + num_components += c->vattr_sizes[i]; + + if (uses_iid) { + c->iid = ntq_emit_vpm_read(c, &vpm_components_queued, + &num_components, ~0); + } + + if (uses_vid) { + c->vid = ntq_emit_vpm_read(c, &vpm_components_queued, + &num_components, ~0); + } + + /* The actual loads will happen directly in nir_intrinsic_load_input + * on newer versions. + */ + if (c->devinfo->ver >= 40) + return; + + for (int loc = 0; loc < ARRAY_SIZE(c->vattr_sizes); loc++) { + resize_qreg_array(c, &c->inputs, &c->inputs_array_size, + (loc + 1) * 4); + + for (int i = 0; i < c->vattr_sizes[loc]; i++) { + c->inputs[loc * 4 + i] = + ntq_emit_vpm_read(c, + &vpm_components_queued, + &num_components, + loc * 4 + i); + + } + } + + if (c->devinfo->ver >= 40) { + assert(vpm_components_queued == num_components); + } else { + assert(vpm_components_queued == 0); + assert(num_components == 0); + } +} + +static bool +var_needs_point_coord(struct v3d_compile *c, nir_variable *var) +{ + return (var->data.location == VARYING_SLOT_PNTC || + (var->data.location >= VARYING_SLOT_VAR0 && + (c->fs_key->point_sprite_mask & + (1 << (var->data.location - VARYING_SLOT_VAR0))))); +} + +static bool +program_reads_point_coord(struct v3d_compile *c) +{ + nir_foreach_variable(var, &c->s->inputs) { + if (var_needs_point_coord(c, var)) + return true; + } + + return false; +} + +static void +ntq_setup_fs_inputs(struct v3d_compile *c) { unsigned num_entries = 0; unsigned num_components = 0; @@ -1284,71 +1496,22 @@ ntq_setup_inputs(struct v3d_compile *c) */ qsort(&vars, num_entries, sizeof(*vars), driver_location_compare); - uint32_t vpm_components_queued = 0; - if (c->s->info.stage == MESA_SHADER_VERTEX) { - bool uses_iid = c->s->info.system_values_read & - (1ull << SYSTEM_VALUE_INSTANCE_ID); - bool uses_vid = c->s->info.system_values_read & - (1ull << SYSTEM_VALUE_VERTEX_ID); - - num_components += uses_iid; - num_components += uses_vid; - - if (uses_iid) { - c->iid = ntq_emit_vpm_read(c, &vpm_components_queued, - &num_components, ~0); - } - - if (uses_vid) { - c->vid = ntq_emit_vpm_read(c, &vpm_components_queued, - &num_components, ~0); - } - } - for (unsigned i = 0; i < num_entries; i++) { nir_variable *var = vars[i]; unsigned array_len = MAX2(glsl_get_length(var->type), 1); unsigned loc = var->data.driver_location; - assert(array_len == 1); - (void)array_len; resize_qreg_array(c, &c->inputs, &c->inputs_array_size, - (loc + 1) * 4); + (loc + array_len) * 4); - if (c->s->info.stage == MESA_SHADER_FRAGMENT) { - if (var->data.location == VARYING_SLOT_POS) { - emit_fragcoord_input(c, loc); - } else if (var->data.location == VARYING_SLOT_PNTC || - (var->data.location >= VARYING_SLOT_VAR0 && - (c->fs_key->point_sprite_mask & - (1 << (var->data.location - - VARYING_SLOT_VAR0))))) { - c->inputs[loc * 4 + 0] = c->point_x; - c->inputs[loc * 4 + 1] = c->point_y; - } else { - emit_fragment_input(c, loc, var); - } + if (var->data.location == VARYING_SLOT_POS) { + emit_fragcoord_input(c, loc); + } else if (var_needs_point_coord(c, var)) { + c->inputs[loc * 4 + 0] = c->point_x; + c->inputs[loc * 4 + 1] = c->point_y; } else { - int var_components = glsl_get_components(var->type); - - for (int i = 0; i < var_components; i++) { - c->inputs[loc * 4 + i] = - ntq_emit_vpm_read(c, - &vpm_components_queued, - &num_components, - loc * 4 + i); - - } - c->vattr_sizes[loc] = var_components; - } - } - - if (c->s->info.stage == MESA_SHADER_VERTEX) { - if (c->devinfo->ver >= 40) { - assert(vpm_components_queued == num_components); - } else { - assert(vpm_components_queued == 0); - assert(num_components == 0); + for (int j = 0; j < array_len; j++) + emit_fragment_input(c, loc + j, var, j); } } } @@ -1356,6 +1519,9 @@ ntq_setup_inputs(struct v3d_compile *c) static void ntq_setup_outputs(struct v3d_compile *c) { + if (c->s->info.stage != MESA_SHADER_FRAGMENT) + return; + nir_foreach_variable(var, &c->s->outputs) { unsigned array_len = MAX2(glsl_get_length(var->type), 1); unsigned loc = var->data.driver_location * 4; @@ -1363,61 +1529,36 @@ ntq_setup_outputs(struct v3d_compile *c) assert(array_len == 1); (void)array_len; - for (int i = 0; i < 4; i++) { + for (int i = 0; i < 4 - var->data.location_frac; i++) { add_output(c, loc + var->data.location_frac + i, var->data.location, var->data.location_frac + i); } - if (c->s->info.stage == MESA_SHADER_FRAGMENT) { - switch (var->data.location) { - case FRAG_RESULT_COLOR: - c->output_color_var[0] = var; - c->output_color_var[1] = var; - c->output_color_var[2] = var; - c->output_color_var[3] = var; - break; - case FRAG_RESULT_DATA0: - case FRAG_RESULT_DATA1: - case FRAG_RESULT_DATA2: - case FRAG_RESULT_DATA3: - c->output_color_var[var->data.location - - FRAG_RESULT_DATA0] = var; - break; - case FRAG_RESULT_DEPTH: - c->output_position_index = loc; - break; - case FRAG_RESULT_SAMPLE_MASK: - c->output_sample_mask_index = loc; - break; - } - } else { - switch (var->data.location) { - case VARYING_SLOT_POS: - c->output_position_index = loc; - break; - case VARYING_SLOT_PSIZ: - c->output_point_size_index = loc; - break; - } + switch (var->data.location) { + case FRAG_RESULT_COLOR: + c->output_color_var[0] = var; + c->output_color_var[1] = var; + c->output_color_var[2] = var; + c->output_color_var[3] = var; + break; + case FRAG_RESULT_DATA0: + case FRAG_RESULT_DATA1: + case FRAG_RESULT_DATA2: + case FRAG_RESULT_DATA3: + c->output_color_var[var->data.location - + FRAG_RESULT_DATA0] = var; + break; + case FRAG_RESULT_DEPTH: + c->output_position_index = loc; + break; + case FRAG_RESULT_SAMPLE_MASK: + c->output_sample_mask_index = loc; + break; } } } -static void -ntq_setup_uniforms(struct v3d_compile *c) -{ - nir_foreach_variable(var, &c->s->uniforms) { - uint32_t vec4_count = glsl_count_attribute_slots(var->type, - false); - unsigned vec4_size = 4 * sizeof(float); - - declare_uniform_range(c, var->data.driver_location * vec4_size, - vec4_count * vec4_size); - - } -} - /** * Sets up the mapping from nir_register to struct qreg *. * @@ -1442,9 +1583,13 @@ ntq_setup_registers(struct v3d_compile *c, struct exec_list *list) static void ntq_emit_load_const(struct v3d_compile *c, nir_load_const_instr *instr) { + /* XXX perf: Experiment with using immediate loads to avoid having + * these end up in the uniform stream. Watch out for breaking the + * small immediates optimization in the process! + */ struct qreg *qregs = ntq_init_ssa_def(c, &instr->def); for (int i = 0; i < instr->def.num_components; i++) - qregs[i] = vir_uniform_ui(c, instr->value.u32[i]); + qregs[i] = vir_uniform_ui(c, instr->value[i].u32); _mesa_hash_table_insert(c->def_ht, &instr->def, qregs); } @@ -1461,61 +1606,108 @@ ntq_emit_ssa_undef(struct v3d_compile *c, nir_ssa_undef_instr *instr) qregs[i] = vir_uniform_ui(c, 0); } +static void +ntq_emit_image_size(struct v3d_compile *c, nir_intrinsic_instr *instr) +{ + assert(instr->intrinsic == nir_intrinsic_image_deref_size); + nir_variable *var = nir_intrinsic_get_var(instr, 0); + unsigned image_index = var->data.driver_location; + const struct glsl_type *sampler_type = glsl_without_array(var->type); + bool is_array = glsl_sampler_type_is_array(sampler_type); + + ntq_store_dest(c, &instr->dest, 0, + vir_uniform(c, QUNIFORM_IMAGE_WIDTH, image_index)); + if (instr->num_components > 1) { + ntq_store_dest(c, &instr->dest, 1, + vir_uniform(c, QUNIFORM_IMAGE_HEIGHT, + image_index)); + } + if (instr->num_components > 2) { + ntq_store_dest(c, &instr->dest, 2, + vir_uniform(c, + is_array ? + QUNIFORM_IMAGE_ARRAY_SIZE : + QUNIFORM_IMAGE_DEPTH, + image_index)); + } +} + static void ntq_emit_intrinsic(struct v3d_compile *c, nir_intrinsic_instr *instr) { - nir_const_value *const_offset; unsigned offset; switch (instr->intrinsic) { case nir_intrinsic_load_uniform: - assert(instr->num_components == 1); - const_offset = nir_src_as_const_value(instr->src[0]); - if (const_offset) { - offset = nir_intrinsic_base(instr) + const_offset->u32[0]; + if (nir_src_is_const(instr->src[0])) { + int offset = (nir_intrinsic_base(instr) + + nir_src_as_uint(instr->src[0])); assert(offset % 4 == 0); /* We need dwords */ offset = offset / 4; - ntq_store_dest(c, &instr->dest, 0, - vir_uniform(c, QUNIFORM_UNIFORM, - offset)); + for (int i = 0; i < instr->num_components; i++) { + ntq_store_dest(c, &instr->dest, i, + vir_uniform(c, QUNIFORM_UNIFORM, + offset + i)); + } } else { - ntq_store_dest(c, &instr->dest, 0, - indirect_uniform_load(c, instr)); + ntq_emit_tmu_general(c, instr, false); } break; case nir_intrinsic_load_ubo: - for (int i = 0; i < instr->num_components; i++) { - int ubo = nir_src_as_const_value(instr->src[0])->u32[0]; - - /* Adjust for where we stored the TGSI register base. */ - vir_ADD_dest(c, - vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUA), - vir_uniform(c, QUNIFORM_UBO_ADDR, 1 + ubo), - vir_ADD(c, - ntq_get_src(c, instr->src[1], 0), - vir_uniform_ui(c, i * 4))); - - vir_emit_thrsw(c); - - ntq_store_dest(c, &instr->dest, i, vir_LDTMU(c)); - } - break; - - const_offset = nir_src_as_const_value(instr->src[0]); - if (const_offset) { - offset = nir_intrinsic_base(instr) + const_offset->u32[0]; - assert(offset % 4 == 0); - /* We need dwords */ - offset = offset / 4; - ntq_store_dest(c, &instr->dest, 0, - vir_uniform(c, QUNIFORM_UNIFORM, - offset)); - } else { - ntq_store_dest(c, &instr->dest, 0, - indirect_uniform_load(c, instr)); - } + ntq_emit_tmu_general(c, instr, false); + break; + + case nir_intrinsic_ssbo_atomic_add: + case nir_intrinsic_ssbo_atomic_imin: + case nir_intrinsic_ssbo_atomic_umin: + case nir_intrinsic_ssbo_atomic_imax: + case nir_intrinsic_ssbo_atomic_umax: + case nir_intrinsic_ssbo_atomic_and: + case nir_intrinsic_ssbo_atomic_or: + case nir_intrinsic_ssbo_atomic_xor: + case nir_intrinsic_ssbo_atomic_exchange: + case nir_intrinsic_ssbo_atomic_comp_swap: + case nir_intrinsic_load_ssbo: + case nir_intrinsic_store_ssbo: + ntq_emit_tmu_general(c, instr, false); + break; + + case nir_intrinsic_shared_atomic_add: + case nir_intrinsic_shared_atomic_imin: + case nir_intrinsic_shared_atomic_umin: + case nir_intrinsic_shared_atomic_imax: + case nir_intrinsic_shared_atomic_umax: + case nir_intrinsic_shared_atomic_and: + case nir_intrinsic_shared_atomic_or: + case nir_intrinsic_shared_atomic_xor: + case nir_intrinsic_shared_atomic_exchange: + case nir_intrinsic_shared_atomic_comp_swap: + case nir_intrinsic_load_shared: + case nir_intrinsic_store_shared: + case nir_intrinsic_load_scratch: + case nir_intrinsic_store_scratch: + ntq_emit_tmu_general(c, instr, true); + break; + + case nir_intrinsic_image_deref_load: + case nir_intrinsic_image_deref_store: + case nir_intrinsic_image_deref_atomic_add: + case nir_intrinsic_image_deref_atomic_min: + case nir_intrinsic_image_deref_atomic_max: + case nir_intrinsic_image_deref_atomic_and: + case nir_intrinsic_image_deref_atomic_or: + case nir_intrinsic_image_deref_atomic_xor: + case nir_intrinsic_image_deref_atomic_exchange: + case nir_intrinsic_image_deref_atomic_comp_swap: + v3d40_vir_emit_image_load_store(c, instr); + break; + + case nir_intrinsic_get_buffer_size: + ntq_store_dest(c, &instr->dest, 0, + vir_uniform(c, QUNIFORM_GET_BUFFER_SIZE, + nir_src_as_uint(instr->src[0]))); break; case nir_intrinsic_load_user_clip_plane: @@ -1527,14 +1719,41 @@ ntq_emit_intrinsic(struct v3d_compile *c, nir_intrinsic_instr *instr) } break; + case nir_intrinsic_load_viewport_x_scale: + ntq_store_dest(c, &instr->dest, 0, + vir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE, 0)); + break; + + case nir_intrinsic_load_viewport_y_scale: + ntq_store_dest(c, &instr->dest, 0, + vir_uniform(c, QUNIFORM_VIEWPORT_Y_SCALE, 0)); + break; + + case nir_intrinsic_load_viewport_z_scale: + ntq_store_dest(c, &instr->dest, 0, + vir_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0)); + break; + + case nir_intrinsic_load_viewport_z_offset: + ntq_store_dest(c, &instr->dest, 0, + vir_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0)); + break; + case nir_intrinsic_load_alpha_ref_float: ntq_store_dest(c, &instr->dest, 0, vir_uniform(c, QUNIFORM_ALPHA_REF, 0)); break; case nir_intrinsic_load_sample_mask_in: + ntq_store_dest(c, &instr->dest, 0, vir_MSF(c)); + break; + + case nir_intrinsic_load_helper_invocation: + vir_set_pf(vir_MSF_dest(c, vir_nop_reg()), V3D_QPU_PF_PUSHZ); ntq_store_dest(c, &instr->dest, 0, - vir_uniform(c, QUNIFORM_SAMPLE_MASK, 0)); + vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFA, + vir_uniform_ui(c, ~0), + vir_uniform_ui(c, 0)))); break; case nir_intrinsic_load_front_face: @@ -1556,67 +1775,182 @@ ntq_emit_intrinsic(struct v3d_compile *c, nir_intrinsic_instr *instr) break; case nir_intrinsic_load_input: - const_offset = nir_src_as_const_value(instr->src[0]); - assert(const_offset && "v3d doesn't support indirect inputs"); - for (int i = 0; i < instr->num_components; i++) { - offset = nir_intrinsic_base(instr) + const_offset->u32[0]; - int comp = nir_intrinsic_component(instr) + i; - ntq_store_dest(c, &instr->dest, i, - vir_MOV(c, c->inputs[offset * 4 + comp])); + /* Use ldvpmv (uniform offset) or ldvpmd (non-uniform offset) + * and enable PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR. + */ + offset = (nir_intrinsic_base(instr) + + nir_src_as_uint(instr->src[0])); + if (c->s->info.stage != MESA_SHADER_FRAGMENT && + c->devinfo->ver >= 40) { + /* Emit the LDVPM directly now, rather than at the top + * of the shader like we did for V3D 3.x (which needs + * vpmsetup when not just taking the next offset). + * + * Note that delaying like this may introduce stalls, + * as LDVPMV takes a minimum of 1 instruction but may + * be slower if the VPM unit is busy with another QPU. + */ + int index = 0; + if (c->s->info.system_values_read & + (1ull << SYSTEM_VALUE_INSTANCE_ID)) { + index++; + } + if (c->s->info.system_values_read & + (1ull << SYSTEM_VALUE_VERTEX_ID)) { + index++; + } + for (int i = 0; i < offset; i++) + index += c->vattr_sizes[i]; + index += nir_intrinsic_component(instr); + for (int i = 0; i < instr->num_components; i++) { + struct qreg vpm_offset = + vir_uniform_ui(c, index++); + ntq_store_dest(c, &instr->dest, i, + vir_LDVPMV_IN(c, vpm_offset)); + } + } else { + for (int i = 0; i < instr->num_components; i++) { + int comp = nir_intrinsic_component(instr) + i; + ntq_store_dest(c, &instr->dest, i, + vir_MOV(c, c->inputs[offset * 4 + + comp])); + } } break; case nir_intrinsic_store_output: - const_offset = nir_src_as_const_value(instr->src[1]); - assert(const_offset && "v3d doesn't support indirect outputs"); - offset = ((nir_intrinsic_base(instr) + - const_offset->u32[0]) * 4 + - nir_intrinsic_component(instr)); + /* XXX perf: Use stvpmv with uniform non-constant offsets and + * stvpmd with non-uniform offsets and enable + * PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR. + */ + if (c->s->info.stage == MESA_SHADER_FRAGMENT) { + offset = ((nir_intrinsic_base(instr) + + nir_src_as_uint(instr->src[1])) * 4 + + nir_intrinsic_component(instr)); + for (int i = 0; i < instr->num_components; i++) { + c->outputs[offset + i] = + vir_MOV(c, + ntq_get_src(c, + instr->src[0], i)); + } + } else { + assert(instr->num_components == 1); - for (int i = 0; i < instr->num_components; i++) { - c->outputs[offset + i] = - vir_MOV(c, ntq_get_src(c, instr->src[0], i)); + vir_VPM_WRITE(c, + ntq_get_src(c, instr->src[0], 0), + nir_intrinsic_base(instr)); } - c->num_outputs = MAX2(c->num_outputs, - offset + instr->num_components); + break; + + case nir_intrinsic_image_deref_size: + ntq_emit_image_size(c, instr); break; case nir_intrinsic_discard: - if (c->execute.file != QFILE_NULL) { - vir_PF(c, c->execute, V3D_QPU_PF_PUSHZ); - vir_set_cond(vir_SETMSF_dest(c, vir_reg(QFILE_NULL, 0), + if (vir_in_nonuniform_control_flow(c)) { + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), + V3D_QPU_PF_PUSHZ); + vir_set_cond(vir_SETMSF_dest(c, vir_nop_reg(), vir_uniform_ui(c, 0)), V3D_QPU_COND_IFA); } else { - vir_SETMSF_dest(c, vir_reg(QFILE_NULL, 0), + vir_SETMSF_dest(c, vir_nop_reg(), vir_uniform_ui(c, 0)); } break; case nir_intrinsic_discard_if: { - /* true (~0) if we're discarding */ - struct qreg cond = ntq_get_src(c, instr->src[0], 0); + enum v3d_qpu_cond cond = ntq_emit_bool_to_cond(c, instr->src[0]); - if (c->execute.file != QFILE_NULL) { - /* execute == 0 means the channel is active. Invert - * the condition so that we can use zero as "executing - * and discarding." - */ - vir_PF(c, vir_OR(c, c->execute, vir_NOT(c, cond)), - V3D_QPU_PF_PUSHZ); - vir_set_cond(vir_SETMSF_dest(c, vir_reg(QFILE_NULL, 0), - vir_uniform_ui(c, 0)), - V3D_QPU_COND_IFA); - } else { - vir_PF(c, cond, V3D_QPU_PF_PUSHZ); - vir_set_cond(vir_SETMSF_dest(c, vir_reg(QFILE_NULL, 0), - vir_uniform_ui(c, 0)), - V3D_QPU_COND_IFNA); + if (vir_in_nonuniform_control_flow(c)) { + struct qinst *exec_flag = vir_MOV_dest(c, vir_nop_reg(), + c->execute); + if (cond == V3D_QPU_COND_IFA) { + vir_set_uf(exec_flag, V3D_QPU_UF_ANDZ); + } else { + vir_set_uf(exec_flag, V3D_QPU_UF_NORNZ); + cond = V3D_QPU_COND_IFA; + } } + vir_set_cond(vir_SETMSF_dest(c, vir_nop_reg(), + vir_uniform_ui(c, 0)), cond); + break; } + case nir_intrinsic_memory_barrier: + case nir_intrinsic_memory_barrier_atomic_counter: + case nir_intrinsic_memory_barrier_buffer: + case nir_intrinsic_memory_barrier_image: + case nir_intrinsic_memory_barrier_shared: + case nir_intrinsic_group_memory_barrier: + /* We don't do any instruction scheduling of these NIR + * instructions between each other, so we just need to make + * sure that the TMU operations before the barrier are flushed + * before the ones after the barrier. That is currently + * handled by having a THRSW in each of them and a LDTMU + * series or a TMUWT after. + */ + break; + + case nir_intrinsic_barrier: + /* Emit a TSY op to get all invocations in the workgroup + * (actually supergroup) to block until the last invocation + * reaches the TSY op. + */ + if (c->devinfo->ver >= 42) { + vir_BARRIERID_dest(c, vir_reg(QFILE_MAGIC, + V3D_QPU_WADDR_SYNCB)); + } else { + struct qinst *sync = + vir_BARRIERID_dest(c, + vir_reg(QFILE_MAGIC, + V3D_QPU_WADDR_SYNCU)); + sync->uniform = + vir_get_uniform_index(c, QUNIFORM_CONSTANT, + 0xffffff00 | + V3D_TSY_WAIT_INC_CHECK); + + } + + /* The blocking of a TSY op only happens at the next thread + * switch. No texturing may be outstanding at the time of a + * TSY blocking operation. + */ + vir_emit_thrsw(c); + break; + + case nir_intrinsic_load_num_work_groups: + for (int i = 0; i < 3; i++) { + ntq_store_dest(c, &instr->dest, i, + vir_uniform(c, QUNIFORM_NUM_WORK_GROUPS, + i)); + } + break; + + case nir_intrinsic_load_local_invocation_index: + ntq_store_dest(c, &instr->dest, 0, + vir_SHR(c, c->cs_payload[1], + vir_uniform_ui(c, 32 - c->local_invocation_index_bits))); + break; + + case nir_intrinsic_load_work_group_id: + ntq_store_dest(c, &instr->dest, 0, + vir_AND(c, c->cs_payload[0], + vir_uniform_ui(c, 0xffff))); + ntq_store_dest(c, &instr->dest, 1, + vir_SHR(c, c->cs_payload[0], + vir_uniform_ui(c, 16))); + ntq_store_dest(c, &instr->dest, 2, + vir_AND(c, c->cs_payload[1], + vir_uniform_ui(c, 0xffff))); + break; + + case nir_intrinsic_load_subgroup_id: + ntq_store_dest(c, &instr->dest, 0, vir_EIDX(c)); + break; + default: fprintf(stderr, "Unknown intrinsic: "); nir_print_instr(&instr->instr, stderr); @@ -1627,18 +1961,25 @@ ntq_emit_intrinsic(struct v3d_compile *c, nir_intrinsic_instr *instr) /* Clears (activates) the execute flags for any channels whose jump target * matches this block. + * + * XXX perf: Could we be using flpush/flpop somehow for our execution channel + * enabling? + * + * XXX perf: For uniform control flow, we should be able to skip c->execute + * handling entirely. */ static void ntq_activate_execute_for_block(struct v3d_compile *c) { - vir_PF(c, vir_XOR(c, c->execute, vir_uniform_ui(c, c->cur_block->index)), - V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_XOR_dest(c, vir_nop_reg(), + c->execute, vir_uniform_ui(c, c->cur_block->index)), + V3D_QPU_PF_PUSHZ); vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute, vir_uniform_ui(c, 0)); } static void -ntq_emit_if(struct v3d_compile *c, nir_if *if_stmt) +ntq_emit_uniform_if(struct v3d_compile *c, nir_if *if_stmt) { nir_block *nir_else_block = nir_if_first_else_block(if_stmt); bool empty_else_block = @@ -1653,28 +1994,83 @@ ntq_emit_if(struct v3d_compile *c, nir_if *if_stmt) else else_block = vir_new_block(c); - bool was_top_level = false; - if (c->execute.file == QFILE_NULL) { + /* Set up the flags for the IF condition (taking the THEN branch). */ + enum v3d_qpu_cond cond = ntq_emit_bool_to_cond(c, if_stmt->condition); + + /* Jump to ELSE. */ + vir_BRANCH(c, cond == V3D_QPU_COND_IFA ? + V3D_QPU_BRANCH_COND_ALLNA : + V3D_QPU_BRANCH_COND_ALLA); + vir_link_blocks(c->cur_block, else_block); + vir_link_blocks(c->cur_block, then_block); + + /* Process the THEN block. */ + vir_set_emit_block(c, then_block); + ntq_emit_cf_list(c, &if_stmt->then_list); + + if (!empty_else_block) { + /* At the end of the THEN block, jump to ENDIF */ + vir_BRANCH(c, V3D_QPU_BRANCH_COND_ALWAYS); + vir_link_blocks(c->cur_block, after_block); + + /* Emit the else block. */ + vir_set_emit_block(c, else_block); + ntq_emit_cf_list(c, &if_stmt->else_list); + } + + vir_link_blocks(c->cur_block, after_block); + + vir_set_emit_block(c, after_block); +} + +static void +ntq_emit_nonuniform_if(struct v3d_compile *c, nir_if *if_stmt) +{ + nir_block *nir_else_block = nir_if_first_else_block(if_stmt); + bool empty_else_block = + (nir_else_block == nir_if_last_else_block(if_stmt) && + exec_list_is_empty(&nir_else_block->instr_list)); + + struct qblock *then_block = vir_new_block(c); + struct qblock *after_block = vir_new_block(c); + struct qblock *else_block; + if (empty_else_block) + else_block = after_block; + else + else_block = vir_new_block(c); + + bool was_uniform_control_flow = false; + if (!vir_in_nonuniform_control_flow(c)) { c->execute = vir_MOV(c, vir_uniform_ui(c, 0)); - was_top_level = true; + was_uniform_control_flow = true; } - /* Set A for executing (execute == 0) and jumping (if->condition == - * 0) channels, and then update execute flags for those to point to - * the ELSE block. + /* Set up the flags for the IF condition (taking the THEN branch). */ + enum v3d_qpu_cond cond = ntq_emit_bool_to_cond(c, if_stmt->condition); + + /* Update the flags+cond to mean "Taking the ELSE branch (!cond) and + * was previously active (execute Z) for updating the exec flags. */ - vir_PF(c, vir_OR(c, - c->execute, - ntq_get_src(c, if_stmt->condition, 0)), - V3D_QPU_PF_PUSHZ); - vir_MOV_cond(c, V3D_QPU_COND_IFA, + if (was_uniform_control_flow) { + cond = v3d_qpu_cond_invert(cond); + } else { + struct qinst *inst = vir_MOV_dest(c, vir_nop_reg(), c->execute); + if (cond == V3D_QPU_COND_IFA) { + vir_set_uf(inst, V3D_QPU_UF_NORNZ); + } else { + vir_set_uf(inst, V3D_QPU_UF_ANDZ); + cond = V3D_QPU_COND_IFA; + } + } + + vir_MOV_cond(c, cond, c->execute, vir_uniform_ui(c, else_block->index)); /* Jump to ELSE if nothing is active for THEN, otherwise fall * through. */ - vir_PF(c, c->execute, V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), V3D_QPU_PF_PUSHZ); vir_BRANCH(c, V3D_QPU_BRANCH_COND_ALLNA); vir_link_blocks(c->cur_block, else_block); vir_link_blocks(c->cur_block, then_block); @@ -1688,14 +2084,16 @@ ntq_emit_if(struct v3d_compile *c, nir_if *if_stmt) * active channels update their execute flags to point to * ENDIF */ - vir_PF(c, c->execute, V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), + V3D_QPU_PF_PUSHZ); vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute, vir_uniform_ui(c, after_block->index)); /* If everything points at ENDIF, then jump there immediately. */ - vir_PF(c, vir_XOR(c, c->execute, - vir_uniform_ui(c, after_block->index)), - V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_XOR_dest(c, vir_nop_reg(), + c->execute, + vir_uniform_ui(c, after_block->index)), + V3D_QPU_PF_PUSHZ); vir_BRANCH(c, V3D_QPU_BRANCH_COND_ALLA); vir_link_blocks(c->cur_block, after_block); vir_link_blocks(c->cur_block, else_block); @@ -1708,24 +2106,40 @@ ntq_emit_if(struct v3d_compile *c, nir_if *if_stmt) vir_link_blocks(c->cur_block, after_block); vir_set_emit_block(c, after_block); - if (was_top_level) + if (was_uniform_control_flow) c->execute = c->undef; else ntq_activate_execute_for_block(c); } +static void +ntq_emit_if(struct v3d_compile *c, nir_if *nif) +{ + bool was_in_control_flow = c->in_control_flow; + c->in_control_flow = true; + if (!vir_in_nonuniform_control_flow(c) && + nir_src_is_dynamically_uniform(nif->condition)) { + ntq_emit_uniform_if(c, nif); + } else { + ntq_emit_nonuniform_if(c, nif); + } + c->in_control_flow = was_in_control_flow; +} + static void ntq_emit_jump(struct v3d_compile *c, nir_jump_instr *jump) { switch (jump->type) { case nir_jump_break: - vir_PF(c, c->execute, V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), + V3D_QPU_PF_PUSHZ); vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute, vir_uniform_ui(c, c->loop_break_block->index)); break; case nir_jump_continue: - vir_PF(c, c->execute, V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), + V3D_QPU_PF_PUSHZ); vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute, vir_uniform_ui(c, c->loop_cont_block->index)); break; @@ -1739,6 +2153,10 @@ static void ntq_emit_instr(struct v3d_compile *c, nir_instr *instr) { switch (instr->type) { + case nir_instr_type_deref: + /* ignored, will be walked by the intrinsic using it. */ + break; + case nir_instr_type_alu: ntq_emit_alu(c, nir_instr_as_alu(instr)); break; @@ -1784,10 +2202,13 @@ static void ntq_emit_cf_list(struct v3d_compile *c, struct exec_list *list); static void ntq_emit_loop(struct v3d_compile *c, nir_loop *loop) { - bool was_top_level = false; - if (c->execute.file == QFILE_NULL) { + bool was_in_control_flow = c->in_control_flow; + c->in_control_flow = true; + + bool was_uniform_control_flow = false; + if (!vir_in_nonuniform_control_flow(c)) { c->execute = vir_MOV(c, vir_uniform_ui(c, 0)); - was_top_level = true; + was_uniform_control_flow = true; } struct qblock *save_loop_cont_block = c->loop_cont_block; @@ -1807,13 +2228,14 @@ ntq_emit_loop(struct v3d_compile *c, nir_loop *loop) * * XXX: Use the .ORZ flags update, instead. */ - vir_PF(c, vir_XOR(c, - c->execute, - vir_uniform_ui(c, c->loop_cont_block->index)), - V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_XOR_dest(c, + vir_nop_reg(), + c->execute, + vir_uniform_ui(c, c->loop_cont_block->index)), + V3D_QPU_PF_PUSHZ); vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute, vir_uniform_ui(c, 0)); - vir_PF(c, c->execute, V3D_QPU_PF_PUSHZ); + vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), V3D_QPU_PF_PUSHZ); struct qinst *branch = vir_BRANCH(c, V3D_QPU_BRANCH_COND_ANYA); /* Pixels that were not dispatched or have been discarded should not @@ -1824,13 +2246,17 @@ ntq_emit_loop(struct v3d_compile *c, nir_loop *loop) vir_link_blocks(c->cur_block, c->loop_break_block); vir_set_emit_block(c, c->loop_break_block); - if (was_top_level) + if (was_uniform_control_flow) c->execute = c->undef; else ntq_activate_execute_for_block(c); c->loop_break_block = save_loop_break_block; c->loop_cont_block = save_loop_cont_block; + + c->loops++; + + c->in_control_flow = was_in_control_flow; } static void @@ -1878,23 +2304,81 @@ ntq_emit_impl(struct v3d_compile *c, nir_function_impl *impl) static void nir_to_vir(struct v3d_compile *c) { - if (c->s->info.stage == MESA_SHADER_FRAGMENT) { + switch (c->s->info.stage) { + case MESA_SHADER_FRAGMENT: c->payload_w = vir_MOV(c, vir_reg(QFILE_REG, 0)); c->payload_w_centroid = vir_MOV(c, vir_reg(QFILE_REG, 1)); c->payload_z = vir_MOV(c, vir_reg(QFILE_REG, 2)); - if (c->fs_key->is_points) { - c->point_x = emit_fragment_varying(c, NULL, 0); - c->point_y = emit_fragment_varying(c, NULL, 0); - } else if (c->fs_key->is_lines) { - c->line_x = emit_fragment_varying(c, NULL, 0); + /* V3D 4.x can disable implicit point coordinate varyings if + * they are not used. + */ + if (c->fs_key->is_points && + (c->devinfo->ver < 40 || program_reads_point_coord(c))) { + c->point_x = emit_fragment_varying(c, NULL, 0, 0); + c->point_y = emit_fragment_varying(c, NULL, 0, 0); + c->uses_implicit_point_line_varyings = true; + } else if (c->fs_key->is_lines && c->devinfo->ver < 40) { + c->line_x = emit_fragment_varying(c, NULL, 0, 0); + c->uses_implicit_point_line_varyings = true; + } + break; + case MESA_SHADER_COMPUTE: + /* Set up the TSO for barriers, assuming we do some. */ + if (c->devinfo->ver < 42) { + vir_BARRIERID_dest(c, vir_reg(QFILE_MAGIC, + V3D_QPU_WADDR_SYNC)); } + + c->cs_payload[0] = vir_MOV(c, vir_reg(QFILE_REG, 0)); + c->cs_payload[1] = vir_MOV(c, vir_reg(QFILE_REG, 2)); + + /* Set up the division between gl_LocalInvocationIndex and + * wg_in_mem in the payload reg. + */ + int wg_size = (c->s->info.cs.local_size[0] * + c->s->info.cs.local_size[1] * + c->s->info.cs.local_size[2]); + c->local_invocation_index_bits = + ffs(util_next_power_of_two(MAX2(wg_size, 64))) - 1; + assert(c->local_invocation_index_bits <= 8); + + if (c->s->info.cs.shared_size) { + struct qreg wg_in_mem = vir_SHR(c, c->cs_payload[1], + vir_uniform_ui(c, 16)); + if (c->s->info.cs.local_size[0] != 1 || + c->s->info.cs.local_size[1] != 1 || + c->s->info.cs.local_size[2] != 1) { + int wg_bits = (16 - + c->local_invocation_index_bits); + int wg_mask = (1 << wg_bits) - 1; + wg_in_mem = vir_AND(c, wg_in_mem, + vir_uniform_ui(c, wg_mask)); + } + struct qreg shared_per_wg = + vir_uniform_ui(c, c->s->info.cs.shared_size); + + c->cs_shared_offset = + vir_ADD(c, + vir_uniform(c, QUNIFORM_SHARED_OFFSET,0), + vir_UMUL(c, wg_in_mem, shared_per_wg)); + } + break; + default: + break; } - ntq_setup_inputs(c); + if (c->s->scratch_size) { + v3d_setup_spill_base(c); + c->spill_size += V3D_CHANNELS * c->s->scratch_size; + } + + if (c->s->info.stage == MESA_SHADER_FRAGMENT) + ntq_setup_fs_inputs(c); + else + ntq_setup_vpm_inputs(c); + ntq_setup_outputs(c); - ntq_setup_uniforms(c); - ntq_setup_registers(c, &c->s->registers); /* Find the main function and emit the body. */ nir_foreach_function(function, c->s) { @@ -1908,17 +2392,21 @@ const nir_shader_compiler_options v3d_nir_options = { .lower_all_io_to_temps = true, .lower_extract_byte = true, .lower_extract_word = true, - .lower_bfm = true, .lower_bitfield_insert_to_shifts = true, .lower_bitfield_extract_to_shifts = true, .lower_bitfield_reverse = true, .lower_bit_count = true, + .lower_cs_local_id_from_index = true, + .lower_ffract = true, + .lower_fmod = true, .lower_pack_unorm_2x16 = true, .lower_pack_snorm_2x16 = true, .lower_pack_unorm_4x8 = true, .lower_pack_snorm_4x8 = true, .lower_unpack_unorm_4x8 = true, .lower_unpack_snorm_4x8 = true, + .lower_pack_half_2x16 = true, + .lower_unpack_half_2x16 = true, .lower_fdiv = true, .lower_find_lsb = true, .lower_ffma = true, @@ -1927,29 +2415,13 @@ const nir_shader_compiler_options v3d_nir_options = { .lower_fsat = true, .lower_fsqrt = true, .lower_ifind_msb = true, + .lower_isign = true, .lower_ldexp = true, .lower_mul_high = true, - .native_integers = true, + .lower_wpos_pntc = true, + .lower_rotate = true, }; - -#if 0 -static int -count_nir_instrs(nir_shader *nir) -{ - int count = 0; - nir_foreach_function(function, nir) { - if (!function->impl) - continue; - nir_foreach_block(block, function->impl) { - nir_foreach_instr(instr, block) - count++; - } - } - return count; -} -#endif - /** * When demoting a shader down to single-threaded, removes the THRSW * instructions (one will still be inserted at v3d_vir_to_qpu() for the @@ -2052,6 +2524,8 @@ v3d_nir_to_vir(struct v3d_compile *c) case MESA_SHADER_VERTEX: emit_vert_end(c); break; + case MESA_SHADER_COMPUTE: + break; default: unreachable("bad stage"); } @@ -2066,11 +2540,17 @@ v3d_nir_to_vir(struct v3d_compile *c) } vir_optimize(c); - vir_lower_uniforms(c); vir_check_payload_w(c); - /* XXX: vir_schedule_instructions(c); */ + /* XXX perf: On VC4, we do a VIR-level instruction scheduling here. + * We used that on that platform to pipeline TMU writes and reduce the + * number of thread switches, as well as try (mostly successfully) to + * reduce maximum register pressure to allow more threads. We should + * do something of that sort for V3D -- either instruction scheduling + * here, or delay the the THRSW and LDTMUs from our texture + * instructions until the results are needed. + */ if (V3D_DEBUG & (V3D_DEBUG_VIR | v3d_debug_flag_for_shader_stage(c->s->info.stage))) { @@ -2109,5 +2589,15 @@ v3d_nir_to_vir(struct v3d_compile *c) vir_remove_thrsw(c); } + if (c->spills && + (V3D_DEBUG & (V3D_DEBUG_VIR | + v3d_debug_flag_for_shader_stage(c->s->info.stage)))) { + fprintf(stderr, "%s prog %d/%d spilled VIR:\n", + vir_get_stage_name(c), + c->program_id, c->variant_id); + vir_dump(c); + fprintf(stderr, "\n"); + } + v3d_vir_to_qpu(c, temp_registers); }