bi_block_add_successor(ctx->current_block, branch->branch.target);
}
-static void
-bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
+/* Gets a bytemask for a complete vecN write */
+static unsigned
+bi_mask_for_channels_32(unsigned i)
{
- bi_instruction ins = {
- .type = BI_LOAD_VAR,
- .load_vary = {
- .load = {
- .location = nir_intrinsic_base(instr),
- .channels = instr->num_components,
- },
- .interp_mode = BIFROST_INTERP_DEFAULT, /* TODO */
- .reuse = false, /* TODO */
- .flat = instr->intrinsic != nir_intrinsic_load_interpolated_input
- },
- .dest = bir_dest_index(&instr->dest),
- .dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
+ return (1 << (4 * i)) - 1;
+}
+
+static bi_instruction
+bi_load(enum bi_class T, nir_intrinsic_instr *instr)
+{
+ bi_instruction load = {
+ .type = T,
+ .writemask = bi_mask_for_channels_32(instr->num_components),
+ .src = { BIR_INDEX_CONSTANT },
+ .constant = { .u64 = nir_intrinsic_base(instr) },
};
+ const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
+
+ if (info->has_dest)
+ load.dest = bir_dest_index(&instr->dest);
+
+ if (info->has_dest && info->index_map[NIR_INTRINSIC_TYPE] > 0)
+ load.dest_type = nir_intrinsic_type(instr);
+
nir_src *offset = nir_get_io_offset_src(instr);
if (nir_src_is_const(*offset))
- ins.load_vary.load.location += nir_src_as_uint(*offset);
+ load.constant.u64 += nir_src_as_uint(*offset);
else
- ins.src[0] = bir_src_index(offset);
+ load.src[0] = bir_src_index(offset);
+
+ return load;
+}
+static void
+bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
+{
+ bi_instruction ins = bi_load(BI_LOAD_VAR, instr);
+ 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),
bi_emit(ctx, ins);
}
.blend_location = nir_intrinsic_base(instr),
.src = {
bir_src_index(&instr->src[0])
+ },
+ .swizzle = {
+ { 0, 1, 2, 3 }
}
};
bi_schedule_barrier(ctx);
}
-static struct bi_load
-bi_direct_load_for_instr(nir_intrinsic_instr *instr)
+static void
+bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr)
{
- nir_src *offset = nir_get_io_offset_src(instr);
- assert(nir_src_is_const(*offset)); /* no indirects */
+ bi_instruction address = bi_load(BI_LOAD_VAR_ADDRESS, instr);
+ address.dest = bi_make_temp(ctx);
+ address.dest_type = nir_type_uint64;
+ address.writemask = (1 << 8) - 1;
- struct bi_load load = {
- .location = nir_intrinsic_base(instr) + nir_src_as_uint(*offset),
- .channels = instr->num_components
+ bi_instruction st = {
+ .type = BI_STORE_VAR,
+ .src = {
+ address.dest,
+ bir_src_index(&instr->src[0])
+ },
+ .swizzle = {
+ { 0, 1, 2, 3 }
+ }
};
- return load;
+ bi_emit(ctx, address);
+ bi_emit(ctx, st);
}
static void
-bi_emit_ld_attr(bi_context *ctx, nir_intrinsic_instr *instr)
+bi_emit_ld_uniform(bi_context *ctx, nir_intrinsic_instr *instr)
{
- bi_instruction load = {
- .type = BI_LOAD_ATTR,
- .load = bi_direct_load_for_instr(instr),
- .dest = bir_dest_index(&instr->dest),
- .dest_type = nir_intrinsic_type(instr)
- };
-
- bi_emit(ctx, load);
+ bi_instruction ld = bi_load(BI_LOAD_UNIFORM, instr);
+ ld.src[1] = BIR_INDEX_ZERO; /* TODO: UBO index */
+ bi_emit(ctx, ld);
}
static void
if (ctx->stage == MESA_SHADER_FRAGMENT)
bi_emit_ld_vary(ctx, instr);
else if (ctx->stage == MESA_SHADER_VERTEX)
- bi_emit_ld_attr(ctx, instr);
+ bi_emit(ctx, bi_load(BI_LOAD_ATTR, instr));
else {
unreachable("Unsupported shader stage");
}
case nir_intrinsic_store_output:
if (ctx->stage == MESA_SHADER_FRAGMENT)
bi_emit_frag_out(ctx, instr);
- else {
- /* TODO */
- }
+ else if (ctx->stage == MESA_SHADER_VERTEX)
+ bi_emit_st_vary(ctx, instr);
+ else
+ unreachable("Unsupported shader stage");
+ break;
+
+ case nir_intrinsic_load_uniform:
+ bi_emit_ld_uniform(ctx, instr);
break;
+
default:
/* todo */
break;
}
}
+static void
+emit_load_const(bi_context *ctx, nir_load_const_instr *instr)
+{
+ /* Make sure we've been lowered */
+ assert(instr->def.num_components == 1);
+
+ bi_instruction move = {
+ .type = BI_MOV,
+ .dest = bir_ssa_index(&instr->def),
+ .dest_type = instr->def.bit_size | nir_type_uint,
+ .writemask = (1 << (instr->def.bit_size / 8)) - 1,
+ .src = {
+ BIR_INDEX_CONSTANT
+ },
+ .constant = {
+ .u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size)
+ }
+ };
+
+ bi_emit(ctx, move);
+}
+
+static enum bi_class
+bi_class_for_nir_alu(nir_op op)
+{
+ switch (op) {
+ case nir_op_iadd:
+ case nir_op_fadd:
+ return BI_ADD;
+
+ case nir_op_i2i8:
+ case nir_op_i2i16:
+ case nir_op_i2i32:
+ case nir_op_i2i64:
+ case nir_op_u2u8:
+ case nir_op_u2u16:
+ case nir_op_u2u32:
+ case nir_op_u2u64:
+ case nir_op_f2i16:
+ case nir_op_f2i32:
+ case nir_op_f2i64:
+ case nir_op_f2u16:
+ case nir_op_f2u32:
+ case nir_op_f2u64:
+ case nir_op_i2f16:
+ case nir_op_i2f32:
+ case nir_op_i2f64:
+ case nir_op_u2f16:
+ case nir_op_u2f32:
+ case nir_op_u2f64:
+ return BI_CONVERT;
+
+ case nir_op_fmul:
+ return BI_FMA;
+
+ case nir_op_imin:
+ case nir_op_imax:
+ case nir_op_umin:
+ case nir_op_umax:
+ case nir_op_fmin:
+ case nir_op_fmax:
+ return BI_MINMAX;
+
+ case nir_op_fsat:
+ case nir_op_mov:
+ return BI_MOV;
+
+ default:
+ unreachable("Unknown ALU op");
+ }
+}
+
+static void
+emit_alu(bi_context *ctx, nir_alu_instr *instr)
+{
+ /* Assume it's something we can handle normally */
+ bi_instruction alu = {
+ .type = bi_class_for_nir_alu(instr->op),
+ .dest = bir_dest_index(&instr->dest.dest),
+ .dest_type = nir_op_infos[instr->op].output_type
+ | nir_dest_bit_size(instr->dest.dest),
+ };
+
+ if (instr->dest.dest.is_ssa) {
+ /* Construct a writemask */
+ unsigned bits_per_comp = instr->dest.dest.ssa.bit_size;
+ unsigned comps = instr->dest.dest.ssa.num_components;
+ assert(comps == 1);
+ unsigned bits = bits_per_comp * comps;
+ unsigned bytes = MAX2(bits / 8, 1);
+ alu.writemask = (1 << bytes) - 1;
+ } else {
+ unsigned comp_mask = instr->dest.write_mask;
+
+ alu.writemask = pan_to_bytemask(nir_dest_bit_size(instr->dest.dest),
+ comp_mask);
+ }
+
+ /* We inline constants as we go. This tracks how many constants have
+ * been inlined, since we're limited to 64-bits of constants per
+ * instruction */
+
+ unsigned dest_bits = nir_dest_bit_size(instr->dest.dest);
+ unsigned constants_left = (64 / dest_bits);
+ unsigned constant_shift = 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];
+ }
+
+ /* Op-specific fixup */
+ switch (instr->op) {
+ case nir_op_fmul:
+ alu.src[2] = BIR_INDEX_ZERO; /* FMA */
+ break;
+ case nir_op_fsat:
+ alu.outmod = BIFROST_SAT; /* MOV */
+ break;
+ case nir_op_fmax:
+ case nir_op_imax:
+ case nir_op_umax:
+ alu.op.minmax = BI_MINMAX_MAX; /* MINMAX */
+ break;
+ default:
+ break;
+ }
+
+ bi_emit(ctx, alu);
+}
+
static void
emit_instr(bi_context *ctx, struct nir_instr *instr)
{
switch (instr->type) {
-#if 0
case nir_instr_type_load_const:
emit_load_const(ctx, nir_instr_as_load_const(instr));
break;
-#endif
case nir_instr_type_intrinsic:
emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
break;
-#if 0
case nir_instr_type_alu:
emit_alu(ctx, nir_instr_as_alu(instr));
break;
+#if 0
case nir_instr_type_tex:
emit_tex(ctx, nir_instr_as_tex(instr));
break;
};
NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
+ NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
+ NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
do {
progress = false;
} while (progress);
NIR_PASS(progress, nir, nir_opt_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);
}
void
NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
NIR_PASS_V(nir, nir_lower_ssbo);
- /* We have to lower ALU to scalar ourselves since viewport
- * transformations produce vector ops */
- NIR_PASS_V(nir, nir_lower_alu_to_scalar, NULL, NULL);
-
bi_optimize_nir(nir);
nir_print_shader(nir, stdout);
}
bi_print_shader(ctx, stdout);
+ bi_schedule(ctx);
ralloc_free(ctx);
}