#include "main/mtypes.h"
#include "compiler/glsl/glsl_to_nir.h"
#include "compiler/nir_types.h"
-#include "main/imports.h"
+#include "util/imports.h"
#include "compiler/nir/nir_builder.h"
#include "disassemble.h"
static bi_block *emit_cf_list(bi_context *ctx, struct exec_list *list);
static bi_instruction *bi_emit_branch(bi_context *ctx);
-static void bi_block_add_successor(bi_block *block, bi_block *successor);
static void bi_schedule_barrier(bi_context *ctx);
static void
unreachable("Unhandled jump type");
}
- bi_block_add_successor(ctx->current_block, branch->branch.target);
+ pan_block_add_successor(&ctx->current_block->base, &branch->branch.target->base);
}
/* Gets a bytemask for a complete vecN write */
ins.load_vary.interp_mode = BIFROST_INTERP_DEFAULT; /* TODO */
ins.load_vary.reuse = false; /* TODO */
ins.load_vary.flat = instr->intrinsic != nir_intrinsic_load_interpolated_input;
- ins.dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
+ ins.dest_type = nir_type_float | nir_dest_bit_size(instr->dest);
+
+ if (nir_src_is_const(*nir_get_io_offset_src(instr))) {
+ /* Zero it out for direct */
+ ins.src[1] = BIR_INDEX_ZERO;
+ } else {
+ /* R61 contains sample mask stuff, TODO RA XXX */
+ ins.src[1] = BIR_INDEX_REGISTER | 61;
+ }
+
bi_emit(ctx, ins);
}
{
if (!ctx->emitted_atest) {
bi_instruction ins = {
- .type = BI_ATEST
+ .type = BI_ATEST,
+ .src = {
+ BIR_INDEX_REGISTER | 60 /* TODO: RA */,
+ bir_src_index(&instr->src[0])
+ },
+ .src_types = {
+ nir_type_uint32,
+ nir_type_float32
+ },
+ .swizzle = {
+ { 0 },
+ { 3, 0 } /* swizzle out the alpha */
+ },
+ .dest = BIR_INDEX_REGISTER | 60 /* TODO: RA */,
+ .dest_type = nir_type_uint32,
+ .writemask = 0xF
};
bi_emit(ctx, ins);
.type = BI_BLEND,
.blend_location = nir_intrinsic_base(instr),
.src = {
+ BIR_INDEX_REGISTER | 60 /* Can this be arbitrary? */,
bir_src_index(&instr->src[0])
},
+ .src_types = {
+ nir_type_uint32,
+ nir_type_float32,
+ },
.swizzle = {
+ { 0 },
{ 0, 1, 2, 3 }
- }
+ },
+ .dest = BIR_INDEX_REGISTER | 48 /* Looks like magic */,
+ .dest_type = nir_type_uint32,
+ .writemask = 0xF
};
bi_emit(ctx, blend);
bi_schedule_barrier(ctx);
}
+static bi_instruction
+bi_load_with_r61(enum bi_class T, nir_intrinsic_instr *instr)
+{
+ bi_instruction ld = bi_load(T, instr);
+ ld.src[1] = BIR_INDEX_REGISTER | 61; /* TODO: RA */
+ ld.src[2] = BIR_INDEX_REGISTER | 62;
+ ld.src[3] = 0;
+ ld.src_types[1] = nir_type_uint32;
+ ld.src_types[2] = nir_type_uint32;
+ ld.src_types[3] = nir_intrinsic_type(instr);
+ return ld;
+}
+
static void
bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr)
{
- bi_instruction address = bi_load(BI_LOAD_VAR_ADDRESS, instr);
+ bi_instruction address = bi_load_with_r61(BI_LOAD_VAR_ADDRESS, instr);
address.dest = bi_make_temp(ctx);
- address.dest_type = nir_type_uint64;
- address.writemask = (1 << 8) - 1;
+ address.dest_type = nir_type_uint32;
+ address.writemask = (1 << 12) - 1;
bi_instruction st = {
.type = BI_STORE_VAR,
.src = {
- address.dest,
- bir_src_index(&instr->src[0])
+ bir_src_index(&instr->src[0]),
+ address.dest, address.dest, address.dest,
+ },
+ .src_types = {
+ nir_type_uint32,
+ nir_type_uint32, nir_type_uint32, nir_type_uint32,
},
.swizzle = {
- { 0, 1, 2, 3 }
- }
+ { 0, 1, 2, 3 },
+ { 0 }, { 1 }, { 2}
+ },
+ .store_channels = 4, /* TODO: WRITEMASK */
};
bi_emit(ctx, address);
{
bi_instruction ld = bi_load(BI_LOAD_UNIFORM, instr);
ld.src[1] = BIR_INDEX_ZERO; /* TODO: UBO index */
+
+ /* TODO: Indirect access, since we need to multiply by the element
+ * size. I believe we can get this lowering automatically via
+ * nir_lower_io (as mul instructions) with the proper options, but this
+ * is TODO */
+ assert(ld.src[0] & BIR_INDEX_CONSTANT);
+ ld.constant.u64 += ctx->sysvals.sysval_count;
+ ld.constant.u64 *= 16;
+
bi_emit(ctx, ld);
}
+static void
+bi_emit_sysval(bi_context *ctx, nir_instr *instr,
+ unsigned nr_components, unsigned offset)
+{
+ nir_dest nir_dest;
+
+ /* Figure out which uniform this is */
+ int sysval = panfrost_sysval_for_instr(instr, &nir_dest);
+ void *val = _mesa_hash_table_u64_search(ctx->sysvals.sysval_to_id, sysval);
+
+ /* Sysvals are prefix uniforms */
+ unsigned uniform = ((uintptr_t) val) - 1;
+
+ /* Emit the read itself -- this is never indirect */
+
+ bi_instruction load = {
+ .type = BI_LOAD_UNIFORM,
+ .writemask = (1 << (nr_components * 4)) - 1,
+ .src = { BIR_INDEX_CONSTANT, BIR_INDEX_ZERO },
+ .constant = { (uniform * 16) + offset },
+ .dest = bir_dest_index(&nir_dest),
+ .dest_type = nir_type_uint32, /* TODO */
+ };
+
+ bi_emit(ctx, load);
+}
+
static void
emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
{
if (ctx->stage == MESA_SHADER_FRAGMENT)
bi_emit_ld_vary(ctx, instr);
else if (ctx->stage == MESA_SHADER_VERTEX)
- bi_emit(ctx, bi_load(BI_LOAD_ATTR, instr));
+ bi_emit(ctx, bi_load_with_r61(BI_LOAD_ATTR, instr));
else {
unreachable("Unsupported shader stage");
}
bi_emit_ld_uniform(ctx, instr);
break;
+ case nir_intrinsic_load_ssbo_address:
+ bi_emit_sysval(ctx, &instr->instr, 1, 0);
+ break;
+
+ case nir_intrinsic_get_buffer_size:
+ bi_emit_sysval(ctx, &instr->instr, 1, 8);
+ break;
+
+ case nir_intrinsic_load_viewport_scale:
+ case nir_intrinsic_load_viewport_offset:
+ case nir_intrinsic_load_num_work_groups:
+ case nir_intrinsic_load_sampler_lod_parameters_pan:
+ bi_emit_sysval(ctx, &instr->instr, 3, 0);
+ break;
+
default:
/* todo */
break;
.src = {
BIR_INDEX_CONSTANT
},
+ .src_types = {
+ instr->def.bit_size | nir_type_uint,
+ },
.constant = {
.u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size)
}
bi_emit(ctx, move);
}
+#define BI_CASE_CMP(op) \
+ case op##8: \
+ case op##16: \
+ case op##32: \
+
static enum bi_class
bi_class_for_nir_alu(nir_op op)
{
case nir_op_isub:
return BI_ISUB;
- case nir_op_flt:
- case nir_op_fge:
- case nir_op_feq:
- case nir_op_fne:
- case nir_op_ilt:
- case nir_op_ige:
- case nir_op_ieq:
- case nir_op_ine:
+ BI_CASE_CMP(nir_op_flt)
+ BI_CASE_CMP(nir_op_fge)
+ BI_CASE_CMP(nir_op_feq)
+ BI_CASE_CMP(nir_op_fne)
+ BI_CASE_CMP(nir_op_ilt)
+ BI_CASE_CMP(nir_op_ige)
+ BI_CASE_CMP(nir_op_ieq)
+ BI_CASE_CMP(nir_op_ine)
return BI_CMP;
- case nir_op_bcsel:
+ case nir_op_b8csel:
+ case nir_op_b16csel:
+ case nir_op_b32csel:
return BI_CSEL;
case nir_op_i2i8:
case nir_op_u2f16:
case nir_op_u2f32:
case nir_op_u2f64:
+ case nir_op_f2f16:
+ case nir_op_f2f32:
+ case nir_op_f2f64:
+ case nir_op_f2fmp:
return BI_CONVERT;
+ case nir_op_vec2:
+ case nir_op_vec3:
+ case nir_op_vec4:
+ return BI_COMBINE;
+
+ case nir_op_vec8:
+ case nir_op_vec16:
+ unreachable("should've been lowered");
+
+ case nir_op_ffma:
case nir_op_fmul:
return BI_FMA;
case nir_op_fsat:
case nir_op_fneg:
case nir_op_fabs:
+ return BI_FMOV;
case nir_op_mov:
return BI_MOV;
+ case nir_op_fround_even:
+ case nir_op_fceil:
+ case nir_op_ffloor:
+ case nir_op_ftrunc:
+ return BI_ROUND;
+
case nir_op_frcp:
case nir_op_frsq:
case nir_op_fsin:
}
}
+/* Gets a bi_cond for a given NIR comparison opcode. In soft mode, it will
+ * return BI_COND_ALWAYS as a sentinel if it fails to do so (when used for
+ * optimizations). Otherwise it will bail (when used for primary code
+ * generation). */
+
static enum bi_cond
-bi_cond_for_nir(nir_op op)
+bi_cond_for_nir(nir_op op, bool soft)
{
switch (op) {
- case nir_op_flt:
- case nir_op_ilt:
+ BI_CASE_CMP(nir_op_flt)
+ BI_CASE_CMP(nir_op_ilt)
return BI_COND_LT;
- case nir_op_fge:
- case nir_op_ige:
+
+ BI_CASE_CMP(nir_op_fge)
+ BI_CASE_CMP(nir_op_ige)
return BI_COND_GE;
- case nir_op_feq:
- case nir_op_ieq:
+
+ BI_CASE_CMP(nir_op_feq)
+ BI_CASE_CMP(nir_op_ieq)
return BI_COND_EQ;
- case nir_op_fne:
- case nir_op_ine:
+
+ BI_CASE_CMP(nir_op_fne)
+ BI_CASE_CMP(nir_op_ine)
return BI_COND_NE;
default:
- unreachable("Invalid compare");
+ if (soft)
+ return BI_COND_ALWAYS;
+ else
+ unreachable("Invalid compare");
+ }
+}
+
+static void
+bi_copy_src(bi_instruction *alu, nir_alu_instr *instr, unsigned i, unsigned to,
+ unsigned *constants_left, unsigned *constant_shift)
+{
+ unsigned bits = nir_src_bit_size(instr->src[i].src);
+ unsigned dest_bits = nir_dest_bit_size(instr->dest.dest);
+
+ alu->src_types[to] = nir_op_infos[instr->op].input_types[i]
+ | bits;
+
+ /* Try to inline a constant */
+ if (nir_src_is_const(instr->src[i].src) && *constants_left && (dest_bits == bits)) {
+ alu->constant.u64 |=
+ (nir_src_as_uint(instr->src[i].src)) << *constant_shift;
+
+ alu->src[to] = BIR_INDEX_CONSTANT | (*constant_shift);
+ --(*constants_left);
+ (*constant_shift) += dest_bits;
+ return;
}
+
+ alu->src[to] = bir_src_index(&instr->src[i].src);
+
+ /* We assert scalarization above */
+ alu->swizzle[to][0] = instr->src[i].swizzle[0];
+}
+
+static void
+bi_fuse_csel_cond(bi_instruction *csel, nir_alu_src cond,
+ unsigned *constants_left, unsigned *constant_shift)
+{
+ /* Bail for vector weirdness */
+ if (cond.swizzle[0] != 0)
+ return;
+
+ if (!cond.src.is_ssa)
+ return;
+
+ nir_ssa_def *def = cond.src.ssa;
+ nir_instr *parent = def->parent_instr;
+
+ if (parent->type != nir_instr_type_alu)
+ return;
+
+ nir_alu_instr *alu = nir_instr_as_alu(parent);
+
+ /* Try to match a condition */
+ enum bi_cond bcond = bi_cond_for_nir(alu->op, true);
+
+ if (bcond == BI_COND_ALWAYS)
+ return;
+
+ /* We found one, let's fuse it in */
+ csel->csel_cond = bcond;
+ bi_copy_src(csel, alu, 0, 0, constants_left, constant_shift);
+ bi_copy_src(csel, alu, 1, 3, constants_left, constant_shift);
}
static void
/* Construct a writemask */
unsigned bits_per_comp = instr->dest.dest.ssa.bit_size;
unsigned comps = instr->dest.dest.ssa.num_components;
- assert(comps == 1);
+
+ if (alu.type != BI_COMBINE)
+ assert(comps == 1);
+
unsigned bits = bits_per_comp * comps;
- unsigned bytes = MAX2(bits / 8, 1);
+ unsigned bytes = bits / 8;
alu.writemask = (1 << bytes) - 1;
} else {
unsigned comp_mask = instr->dest.write_mask;
unsigned constants_left = (64 / dest_bits);
unsigned constant_shift = 0;
+ if (alu.type == BI_COMBINE)
+ constants_left = 0;
+
/* Copy sources */
unsigned num_inputs = nir_op_infos[instr->op].num_inputs;
assert(num_inputs <= ARRAY_SIZE(alu.src));
- for (unsigned i = 0; i < num_inputs; ++i) {
- unsigned bits = nir_src_bit_size(instr->src[i].src);
- alu.src_types[i] = nir_op_infos[instr->op].input_types[i]
- | bits;
-
- /* Try to inline a constant */
- if (nir_src_is_const(instr->src[i].src) && constants_left && (dest_bits == bits)) {
- alu.constant.u64 |=
- (nir_src_as_uint(instr->src[i].src)) << constant_shift;
-
- alu.src[i] = BIR_INDEX_CONSTANT | constant_shift;
- --constants_left;
- constant_shift += dest_bits;
- continue;
- }
-
- alu.src[i] = bir_src_index(&instr->src[i].src);
-
- /* We assert scalarization above */
- alu.swizzle[i][0] = instr->src[i].swizzle[0];
- }
+ for (unsigned i = 0; i < num_inputs; ++i)
+ bi_copy_src(&alu, instr, i, i, &constants_left, &constant_shift);
/* Op-specific fixup */
switch (instr->op) {
case nir_op_fmul:
alu.src[2] = BIR_INDEX_ZERO; /* FMA */
+ alu.src_types[2] = alu.src_types[1];
break;
case nir_op_fsat:
- alu.outmod = BIFROST_SAT; /* MOV */
+ alu.outmod = BIFROST_SAT; /* FMOV */
break;
case nir_op_fneg:
- alu.src_neg[0] = true; /* MOV */
+ alu.src_neg[0] = true; /* FMOV */
break;
case nir_op_fabs:
- alu.src_abs[0] = true; /* MOV */
+ alu.src_abs[0] = true; /* FMOV */
break;
case nir_op_fsub:
- alu.src_neg[1] = true; /* ADD */
+ alu.src_neg[1] = true; /* FADD */
break;
case nir_op_fmax:
case nir_op_imax:
case nir_op_fcos:
alu.op.special = BI_SPECIAL_FCOS;
break;
- case nir_op_flt:
- case nir_op_ilt:
- case nir_op_fge:
- case nir_op_ige:
- case nir_op_feq:
- case nir_op_ieq:
- case nir_op_fne:
- case nir_op_ine:
- alu.op.compare = bi_cond_for_nir(instr->op);
+ BI_CASE_CMP(nir_op_flt)
+ BI_CASE_CMP(nir_op_ilt)
+ BI_CASE_CMP(nir_op_fge)
+ BI_CASE_CMP(nir_op_ige)
+ BI_CASE_CMP(nir_op_feq)
+ BI_CASE_CMP(nir_op_ieq)
+ BI_CASE_CMP(nir_op_fne)
+ BI_CASE_CMP(nir_op_ine)
+ alu.op.compare = bi_cond_for_nir(instr->op, false);
+ break;
+ case nir_op_fround_even:
+ alu.op.round = BI_ROUND_MODE;
+ alu.roundmode = BIFROST_RTE;
+ break;
+ case nir_op_fceil:
+ alu.op.round = BI_ROUND_MODE;
+ alu.roundmode = BIFROST_RTP;
+ break;
+ case nir_op_ffloor:
+ alu.op.round = BI_ROUND_MODE;
+ alu.roundmode = BIFROST_RTN;
+ break;
+ case nir_op_ftrunc:
+ alu.op.round = BI_ROUND_MODE;
+ alu.roundmode = BIFROST_RTZ;
break;
default:
break;
}
+ if (alu.type == BI_CSEL) {
+ /* Default to csel3 */
+ alu.csel_cond = BI_COND_NE;
+ alu.src[3] = BIR_INDEX_ZERO;
+ alu.src_types[3] = alu.src_types[0];
+
+ bi_fuse_csel_cond(&alu, instr->src[0],
+ &constants_left, &constant_shift);
+ }
+
bi_emit(ctx, alu);
}
{
bi_block *blk = rzalloc(ctx, bi_block);
- blk->predecessors = _mesa_set_create(blk,
+ blk->base.predecessors = _mesa_set_create(blk,
_mesa_hash_pointer,
_mesa_key_pointer_equal);
- blk->name = ctx->block_name_count++;
+ blk->base.name = ctx->block_name_count++;
return blk;
}
-static void
-bi_block_add_successor(bi_block *block, bi_block *successor)
-{
- assert(block);
- assert(successor);
-
- for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
- if (block->successors[i]) {
- if (block->successors[i] == successor)
- return;
- else
- continue;
- }
-
- block->successors[i] = successor;
- _mesa_set_add(successor->predecessors, block);
- return;
- }
-
- unreachable("Too many successors");
-}
-
static void
bi_schedule_barrier(bi_context *ctx)
{
bi_block *temp = ctx->after_block;
ctx->after_block = create_empty_block(ctx);
- list_addtail(&ctx->after_block->link, &ctx->blocks);
- list_inithead(&ctx->after_block->instructions);
- bi_block_add_successor(ctx->current_block, ctx->after_block);
+ list_addtail(&ctx->after_block->base.link, &ctx->blocks);
+ list_inithead(&ctx->after_block->base.instructions);
+ pan_block_add_successor(&ctx->current_block->base, &ctx->after_block->base);
ctx->current_block = ctx->after_block;
ctx->after_block = temp;
}
ctx->current_block = create_empty_block(ctx);
}
- list_addtail(&ctx->current_block->link, &ctx->blocks);
- list_inithead(&ctx->current_block->instructions);
+ list_addtail(&ctx->current_block->base.link, &ctx->blocks);
+ list_inithead(&ctx->current_block->base.instructions);
nir_foreach_instr(instr, block) {
emit_instr(ctx, instr);
} else {
then_branch->branch.target = else_block;
then_exit->branch.target = ctx->after_block;
- bi_block_add_successor(end_then_block, then_exit->branch.target);
+ pan_block_add_successor(&end_then_block->base, &then_exit->branch.target->base);
}
/* Wire up the successors */
- bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
+ pan_block_add_successor(&before_block->base, &then_branch->branch.target->base); /* then_branch */
- bi_block_add_successor(before_block, then_block); /* fallthrough */
- bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
+ pan_block_add_successor(&before_block->base, &then_block->base); /* fallthrough */
+ pan_block_add_successor(&end_else_block->base, &ctx->after_block->base); /* fallthrough */
}
static void
/* Branch back to loop back */
bi_instruction *br_back = bi_emit_branch(ctx);
br_back->branch.target = ctx->continue_block;
- bi_block_add_successor(start_block, ctx->continue_block);
- bi_block_add_successor(ctx->current_block, ctx->continue_block);
+ pan_block_add_successor(&start_block->base, &ctx->continue_block->base);
+ pan_block_add_successor(&ctx->current_block->base, &ctx->continue_block->base);
ctx->after_block = ctx->break_block;
} while (progress);
NIR_PASS(progress, nir, nir_opt_algebraic_late);
+ NIR_PASS(progress, nir, nir_lower_bool_to_int32);
NIR_PASS(progress, nir, bifrost_nir_lower_algebraic_late);
NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
/* Take us out of SSA */
NIR_PASS(progress, nir, nir_lower_locals_to_regs);
- NIR_PASS(progress, nir, nir_convert_from_ssa, true);
-
- /* We're a primary scalar architecture but there's enough vector that
- * we use a vector IR so let's not also deal with scalar hacks on top
- * of the vector hacks */
-
NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest);
- NIR_PASS(progress, nir, nir_lower_vec_to_movs);
- NIR_PASS(progress, nir, nir_opt_dce);
+ NIR_PASS(progress, nir, nir_convert_from_ssa, true);
}
void
-bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
+bifrost_compile_shader_nir(nir_shader *nir, panfrost_program *program, unsigned product_id)
{
bi_context *ctx = rzalloc(NULL, bi_context);
ctx->nir = nir;
bi_optimize_nir(nir);
nir_print_shader(nir, stdout);
+ panfrost_nir_assign_sysvals(&ctx->sysvals, nir);
+ program->sysval_count = ctx->sysvals.sysval_count;
+ memcpy(program->sysvals, ctx->sysvals.sysvals, sizeof(ctx->sysvals.sysvals[0]) * ctx->sysvals.sysval_count);
+
nir_foreach_function(func, nir) {
if (!func->impl)
continue;
break; /* TODO: Multi-function shaders */
}
+ bi_foreach_block(ctx, _block) {
+ bi_block *block = (bi_block *) _block;
+ bi_lower_combine(ctx, block);
+ }
+
+ bool progress = false;
+
+ do {
+ progress = false;
+
+ bi_foreach_block(ctx, _block) {
+ bi_block *block = (bi_block *) _block;
+ progress |= bi_opt_dead_code_eliminate(ctx, block);
+ }
+ } while(progress);
+
bi_print_shader(ctx, stdout);
bi_schedule(ctx);
+ bi_register_allocate(ctx);
+ bi_print_shader(ctx, stdout);
+ bi_pack(ctx, &program->compiled);
+ disassemble_bifrost(stdout, program->compiled.data, program->compiled.size, true);
ralloc_free(ctx);
}
projects / mesa.git / blobdiff