return false;
}
+static unsigned get_number_of_slots(struct lower_io_state *state,
+ const nir_variable *var)
+{
+ const struct glsl_type *type = var->type;
+
+ if (nir_is_per_vertex_io(var, state->builder.shader->info.stage)) {
+ assert(glsl_type_is_array(type));
+ type = glsl_get_array_element(type);
+ }
+
+ return state->type_size(type, var->data.bindless);
+}
+
static nir_ssa_def *
get_io_offset(nir_builder *b, nir_deref_instr *deref,
nir_ssa_def **vertex_index,
load->intrinsic == nir_intrinsic_load_uniform)
nir_intrinsic_set_type(load, type);
+ if (load->intrinsic != nir_intrinsic_load_uniform) {
+ nir_io_semantics semantics = {0};
+ semantics.location = var->data.location;
+ semantics.num_slots = get_number_of_slots(state, var);
+ semantics.fb_fetch_output = var->data.fb_fetch_output;
+ nir_intrinsic_set_io_semantics(load, semantics);
+ }
+
if (vertex_index) {
load->src[0] = nir_src_for_ssa(vertex_index);
load->src[1] = nir_src_for_ssa(offset);
store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
+ unsigned gs_streams = 0;
+ if (state->builder.shader->info.stage == MESA_SHADER_GEOMETRY) {
+ if (var->data.stream & NIR_STREAM_PACKED) {
+ gs_streams = var->data.stream & ~NIR_STREAM_PACKED;
+ } else {
+ assert(var->data.stream < 4);
+ gs_streams = 0;
+ for (unsigned i = 0; i < num_components; ++i)
+ gs_streams |= var->data.stream << (2 * i);
+ }
+ }
+
+ nir_io_semantics semantics = {0};
+ semantics.location = var->data.location;
+ semantics.num_slots = get_number_of_slots(state, var);
+ semantics.dual_source_blend_index = var->data.index;
+ semantics.gs_streams = gs_streams;
+ nir_intrinsic_set_io_semantics(store, semantics);
+
nir_builder_instr_insert(b, &store->instr);
}
nir_intrinsic_set_base(load, var->data.driver_location);
nir_intrinsic_set_component(load, component);
+ nir_io_semantics semantics = {0};
+ semantics.location = var->data.location;
+ semantics.num_slots = get_number_of_slots(state, var);
+ nir_intrinsic_set_io_semantics(load, semantics);
+
load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
load->src[1] = nir_src_for_ssa(offset);
assert(addr->bit_size == offset->bit_size);
return nir_vec2(b, nir_channel(b, addr, 0),
nir_iadd(b, nir_channel(b, addr, 1), offset));
+
+ case nir_address_format_32bit_index_offset_pack64:
+ assert(addr->num_components == 1);
+ assert(offset->bit_size == 32);
+ return nir_pack_64_2x32_split(b,
+ nir_iadd(b, nir_unpack_64_2x32_split_x(b, addr), offset),
+ nir_unpack_64_2x32_split_y(b, addr));
+
case nir_address_format_vec2_index_32bit_offset:
assert(addr->num_components == 3);
assert(offset->bit_size == 32);
return nir_vec3(b, nir_channel(b, addr, 0), nir_channel(b, addr, 1),
nir_iadd(b, nir_channel(b, addr, 2), offset));
+
case nir_address_format_logical:
unreachable("Unsupported address format");
}
static unsigned
addr_get_offset_bit_size(nir_ssa_def *addr, nir_address_format addr_format)
{
- if (addr_format == nir_address_format_32bit_offset_as_64bit)
+ if (addr_format == nir_address_format_32bit_offset_as_64bit ||
+ addr_format == nir_address_format_32bit_index_offset_pack64)
return 32;
return addr->bit_size;
}
addr_get_offset_bit_size(addr, addr_format)));
}
+static nir_ssa_def *
+build_addr_for_var(nir_builder *b, nir_variable *var,
+ nir_address_format addr_format)
+{
+ assert(var->data.mode & (nir_var_uniform | nir_var_mem_shared |
+ nir_var_shader_temp | nir_var_function_temp |
+ nir_var_mem_constant));
+
+ const unsigned num_comps = nir_address_format_num_components(addr_format);
+ const unsigned bit_size = nir_address_format_bit_size(addr_format);
+
+ switch (addr_format) {
+ case nir_address_format_32bit_global:
+ case nir_address_format_64bit_global: {
+ nir_ssa_def *base_addr;
+ switch (var->data.mode) {
+ case nir_var_shader_temp:
+ base_addr = nir_load_scratch_base_ptr(b, 0, num_comps, bit_size);
+ break;
+
+ case nir_var_function_temp:
+ base_addr = nir_load_scratch_base_ptr(b, 1, num_comps, bit_size);
+ break;
+
+ case nir_var_mem_constant:
+ base_addr = nir_load_constant_base_ptr(b, num_comps, bit_size);
+ break;
+
+ default:
+ unreachable("Unsupported variable mode");
+ }
+
+ return build_addr_iadd_imm(b, base_addr, addr_format,
+ var->data.driver_location);
+ }
+
+ case nir_address_format_32bit_offset:
+ assert(var->data.driver_location <= UINT32_MAX);
+ return nir_imm_int(b, var->data.driver_location);
+
+ case nir_address_format_32bit_offset_as_64bit:
+ assert(var->data.driver_location <= UINT32_MAX);
+ return nir_imm_int64(b, var->data.driver_location);
+
+ default:
+ unreachable("Unsupported address format");
+ }
+}
+
static nir_ssa_def *
addr_to_index(nir_builder *b, nir_ssa_def *addr,
nir_address_format addr_format)
{
- if (addr_format == nir_address_format_32bit_index_offset) {
+ switch (addr_format) {
+ case nir_address_format_32bit_index_offset:
assert(addr->num_components == 2);
return nir_channel(b, addr, 0);
- } else if (addr_format == nir_address_format_vec2_index_32bit_offset) {
+ case nir_address_format_32bit_index_offset_pack64:
+ return nir_unpack_64_2x32_split_y(b, addr);
+ case nir_address_format_vec2_index_32bit_offset:
assert(addr->num_components == 3);
return nir_channels(b, addr, 0x3);
- } else {
- unreachable("bad address format for index");
+ default: unreachable("Invalid address format");
}
}
case nir_address_format_32bit_index_offset:
assert(addr->num_components == 2);
return nir_channel(b, addr, 1);
+ case nir_address_format_32bit_index_offset_pack64:
+ return nir_unpack_64_2x32_split_x(b, addr);
case nir_address_format_vec2_index_32bit_offset:
assert(addr->num_components == 3);
return nir_channel(b, addr, 2);
nir_u2u64(b, nir_channel(b, addr, 3)));
case nir_address_format_32bit_index_offset:
+ case nir_address_format_32bit_index_offset_pack64:
case nir_address_format_vec2_index_32bit_offset:
case nir_address_format_32bit_offset:
case nir_address_format_32bit_offset_as_64bit:
assert(addr_format_is_global(addr_format));
op = nir_intrinsic_load_global;
break;
- case nir_var_shader_in:
+ case nir_var_uniform:
assert(addr_format_is_offset(addr_format));
+ assert(b->shader->info.stage == MESA_SHADER_KERNEL);
op = nir_intrinsic_load_kernel_input;
break;
case nir_var_mem_shared:
op = nir_intrinsic_load_global;
}
break;
+ case nir_var_mem_constant:
+ if (addr_format_is_offset(addr_format)) {
+ op = nir_intrinsic_load_constant;
+ } else {
+ assert(addr_format_is_global(addr_format));
+ op = nir_intrinsic_load_global_constant;
+ }
+ break;
default:
unreachable("Unsupported explicit IO variable mode");
}
load->src[1] = nir_src_for_ssa(addr_to_offset(b, addr, addr_format));
}
- if (nir_intrinsic_infos[op].index_map[NIR_INTRINSIC_ACCESS] > 0)
+ if (nir_intrinsic_has_access(load))
nir_intrinsic_set_access(load, nir_intrinsic_access(intrin));
+ if (op == nir_intrinsic_load_constant) {
+ nir_intrinsic_set_base(load, 0);
+ nir_intrinsic_set_range(load, b->shader->constant_data_size);
+ }
+
unsigned bit_size = intrin->dest.ssa.bit_size;
if (bit_size == 1) {
/* TODO: Make the native bool bit_size an option. */
nir_intrinsic_set_write_mask(store, write_mask);
- if (nir_intrinsic_infos[op].index_map[NIR_INTRINSIC_ACCESS] > 0)
+ if (nir_intrinsic_has_access(store))
nir_intrinsic_set_access(store, nir_intrinsic_access(intrin));
/* TODO: We should try and provide a better alignment. For OpenCL, we need
/* Global atomics don't have access flags because they assume that the
* address may be non-uniform.
*/
- if (nir_intrinsic_infos[op].index_map[NIR_INTRINSIC_ACCESS] > 0)
+ if (nir_intrinsic_has_access(atomic))
nir_intrinsic_set_access(atomic, nir_intrinsic_access(intrin));
assert(intrin->dest.ssa.num_components == 1);
assert(deref->dest.is_ssa);
switch (deref->deref_type) {
case nir_deref_type_var:
- assert(deref->mode & (nir_var_shader_in | nir_var_mem_shared |
- nir_var_shader_temp | nir_var_function_temp));
- if (addr_format_is_global(addr_format)) {
- assert(nir_var_shader_temp | nir_var_function_temp);
- base_addr =
- nir_load_scratch_base_ptr(b, !(deref->mode & nir_var_shader_temp),
- nir_address_format_num_components(addr_format),
- nir_address_format_bit_size(addr_format));
- return build_addr_iadd_imm(b, base_addr, addr_format,
- deref->var->data.driver_location);
- } else {
- assert(deref->var->data.driver_location <= UINT32_MAX);
- return nir_imm_intN_t(b, deref->var->data.driver_location,
- deref->dest.ssa.bit_size);
- }
+ return build_addr_for_var(b, deref->var, addr_format);
case nir_deref_type_array: {
- nir_deref_instr *parent = nir_deref_instr_parent(deref);
-
- unsigned stride = glsl_get_explicit_stride(parent->type);
- if ((glsl_type_is_matrix(parent->type) &&
- glsl_matrix_type_is_row_major(parent->type)) ||
- (glsl_type_is_vector(parent->type) && stride == 0))
- stride = type_scalar_size_bytes(parent->type);
-
+ unsigned stride = nir_deref_instr_array_stride(deref);
assert(stride > 0);
nir_ssa_def *index = nir_ssa_for_src(b, deref->arr.index, 1);
case nir_deref_type_ptr_as_array: {
nir_ssa_def *index = nir_ssa_for_src(b, deref->arr.index, 1);
index = nir_i2i(b, index, addr_get_offset_bit_size(base_addr, addr_format));
- unsigned stride = nir_deref_instr_ptr_as_array_stride(deref);
+ unsigned stride = nir_deref_instr_array_stride(deref);
return build_addr_iadd(b, base_addr, addr_format,
nir_amul_imm(b, index, stride));
}
nir_instr_remove(&intrin->instr);
}
+bool
+nir_get_explicit_deref_align(nir_deref_instr *deref,
+ bool default_to_type_align,
+ uint32_t *align_mul,
+ uint32_t *align_offset)
+{
+ if (deref->deref_type == nir_deref_type_var) {
+ /* If we see a variable, align_mul is effectively infinite because we
+ * know the offset exactly (up to the offset of the base pointer for the
+ * given variable mode). We have to pick something so we choose 256B
+ * as an arbitrary alignment which seems high enough for any reasonable
+ * wide-load use-case. Back-ends should clamp alignments down if 256B
+ * is too large for some reason.
+ */
+ *align_mul = 256;
+ *align_offset = deref->var->data.driver_location % 256;
+ return true;
+ }
+
+ /* If we're a cast deref that has an alignment, use that. */
+ if (deref->deref_type == nir_deref_type_cast && deref->cast.align_mul > 0) {
+ *align_mul = deref->cast.align_mul;
+ *align_offset = deref->cast.align_offset;
+ return true;
+ }
+
+ /* Otherwise, we need to compute the alignment based on the parent */
+ nir_deref_instr *parent = nir_deref_instr_parent(deref);
+ if (parent == NULL) {
+ assert(deref->deref_type == nir_deref_type_cast);
+ if (default_to_type_align) {
+ /* If we don't have a parent, assume the type's alignment, if any. */
+ unsigned type_align = glsl_get_explicit_alignment(deref->type);
+ if (type_align == 0)
+ return false;
+
+ *align_mul = type_align;
+ *align_offset = 0;
+ return true;
+ } else {
+ return false;
+ }
+ }
+
+ uint32_t parent_mul, parent_offset;
+ if (!nir_get_explicit_deref_align(parent, default_to_type_align,
+ &parent_mul, &parent_offset))
+ return false;
+
+ switch (deref->deref_type) {
+ case nir_deref_type_var:
+ unreachable("Handled above");
+
+ case nir_deref_type_array:
+ case nir_deref_type_array_wildcard:
+ case nir_deref_type_ptr_as_array: {
+ const unsigned stride = nir_deref_instr_array_stride(deref);
+ if (stride == 0)
+ return false;
+
+ if (deref->deref_type != nir_deref_type_array_wildcard &&
+ nir_src_is_const(deref->arr.index)) {
+ unsigned offset = nir_src_as_uint(deref->arr.index) * stride;
+ *align_mul = parent_mul;
+ *align_offset = (parent_offset + offset) % parent_mul;
+ } else {
+ /* If this is a wildcard or an indirect deref, we have to go with the
+ * power-of-two gcd.
+ */
+ *align_mul = MIN3(parent_mul,
+ 1 << (ffs(parent_offset) - 1),
+ 1 << (ffs(stride) - 1));
+ *align_offset = 0;
+ }
+ return true;
+ }
+
+ case nir_deref_type_struct: {
+ const int offset = glsl_get_struct_field_offset(parent->type,
+ deref->strct.index);
+ if (offset < 0)
+ return false;
+
+ *align_mul = parent_mul;
+ *align_offset = (parent_offset + offset) % parent_mul;
+ return true;
+ }
+
+ case nir_deref_type_cast:
+ /* We handled the explicit alignment case above. */
+ assert(deref->cast.align_mul == 0);
+ *align_mul = parent_mul;
+ *align_offset = parent_offset;
+ return true;
+ }
+
+ unreachable("Invalid deref_instr_type");
+}
+
static void
lower_explicit_io_deref(nir_builder *b, nir_deref_instr *deref,
nir_address_format addr_format)
nir_ssa_def *addr = nir_explicit_io_address_from_deref(b, deref, base_addr,
addr_format);
+ assert(addr->bit_size == deref->dest.ssa.bit_size);
+ assert(addr->num_components == deref->dest.ssa.num_components);
nir_instr_remove(&deref->instr);
nir_ssa_def_rewrite_uses(&deref->dest.ssa, nir_src_for_ssa(addr));
unsigned stride = glsl_get_explicit_stride(deref->type);
assert(stride > 0);
- assert(addr_format == nir_address_format_32bit_index_offset ||
- addr_format == nir_address_format_vec2_index_32bit_offset);
nir_ssa_def *addr = &deref->dest.ssa;
nir_ssa_def *index = addr_to_index(b, addr, addr_format);
nir_ssa_def *offset = addr_to_offset(b, addr, addr_format);
case nir_var_mem_shared:
offset = 0;
break;
+ case nir_var_mem_constant:
+ offset = shader->constant_data_size;
+ break;
default:
unreachable("Unsupported mode");
}
const struct glsl_type *explicit_type =
glsl_get_explicit_type_for_size_align(var->type, type_info, &size, &align);
- if (explicit_type != var->type) {
- progress = true;
+ if (explicit_type != var->type)
var->type = explicit_type;
- }
var->data.driver_location = ALIGN_POT(offset, align);
offset = var->data.driver_location + size;
+ progress = true;
}
switch (mode) {
break;
case nir_var_mem_shared:
shader->info.cs.shared_size = offset;
- shader->num_shared = offset;
+ shader->shared_size = offset;
+ break;
+ case nir_var_mem_constant:
+ shader->constant_data_size = offset;
break;
default:
unreachable("Unsupported mode");
return progress;
}
+static void
+write_constant(void *dst, const nir_constant *c, const struct glsl_type *type)
+{
+ if (glsl_type_is_vector_or_scalar(type)) {
+ const unsigned num_components = glsl_get_vector_elements(type);
+ const unsigned bit_size = glsl_get_bit_size(type);
+ if (bit_size == 1) {
+ /* Booleans are special-cased to be 32-bit
+ *
+ * TODO: Make the native bool bit_size an option.
+ */
+ for (unsigned i = 0; i < num_components; i++) {
+ int32_t b32 = -(int)c->values[i].b;
+ memcpy((char *)dst + i * 4, &b32, 4);
+ }
+ } else {
+ assert(bit_size >= 8 && bit_size % 8 == 0);
+ const unsigned byte_size = bit_size / 8;
+ for (unsigned i = 0; i < num_components; i++) {
+ /* Annoyingly, thanks to packed structs, we can't make any
+ * assumptions about the alignment of dst. To avoid any strange
+ * issues with unaligned writes, we always use memcpy.
+ */
+ memcpy((char *)dst + i * byte_size, &c->values[i], byte_size);
+ }
+ }
+ } else if (glsl_type_is_array_or_matrix(type)) {
+ const unsigned array_len = glsl_get_length(type);
+ const unsigned stride = glsl_get_explicit_stride(type);
+ assert(stride > 0);
+ const struct glsl_type *elem_type = glsl_get_array_element(type);
+ for (unsigned i = 0; i < array_len; i++)
+ write_constant((char *)dst + i * stride, c->elements[i], elem_type);
+ } else {
+ assert(glsl_type_is_struct_or_ifc(type));
+ const unsigned num_fields = glsl_get_length(type);
+ for (unsigned i = 0; i < num_fields; i++) {
+ const int field_offset = glsl_get_struct_field_offset(type, i);
+ assert(field_offset >= 0);
+ const struct glsl_type *field_type = glsl_get_struct_field(type, i);
+ write_constant((char *)dst + field_offset, c->elements[i], field_type);
+ }
+ }
+}
+
+bool
+nir_lower_mem_constant_vars(nir_shader *shader,
+ glsl_type_size_align_func type_info)
+{
+ bool progress = false;
+
+ unsigned old_constant_data_size = shader->constant_data_size;
+ if (lower_vars_to_explicit(shader, &shader->variables,
+ nir_var_mem_constant, type_info)) {
+ assert(shader->constant_data_size > old_constant_data_size);
+ shader->constant_data = rerzalloc_size(shader, shader->constant_data,
+ old_constant_data_size,
+ shader->constant_data_size);
+
+ nir_foreach_variable_with_modes(var, shader, nir_var_mem_constant) {
+ write_constant((char *)shader->constant_data +
+ var->data.driver_location,
+ var->constant_initializer, var->type);
+ }
+ progress = true;
+ }
+
+ nir_foreach_function(function, shader) {
+ if (!function->impl)
+ continue;
+
+ if (nir_lower_vars_to_explicit_types_impl(function->impl,
+ nir_var_mem_constant,
+ type_info))
+ progress = true;
+ }
+
+ return progress;
+}
+
/**
* Return the offset source for a load/store intrinsic.
*/
case nir_intrinsic_load_shared:
case nir_intrinsic_load_uniform:
case nir_intrinsic_load_global:
+ case nir_intrinsic_load_global_constant:
case nir_intrinsic_load_scratch:
case nir_intrinsic_load_fs_input_interp_deltas:
+ case nir_intrinsic_shared_atomic_add:
+ case nir_intrinsic_shared_atomic_and:
+ case nir_intrinsic_shared_atomic_comp_swap:
+ case nir_intrinsic_shared_atomic_exchange:
+ case nir_intrinsic_shared_atomic_fadd:
+ case nir_intrinsic_shared_atomic_fcomp_swap:
+ case nir_intrinsic_shared_atomic_fmax:
+ case nir_intrinsic_shared_atomic_fmin:
+ case nir_intrinsic_shared_atomic_imax:
+ case nir_intrinsic_shared_atomic_imin:
+ case nir_intrinsic_shared_atomic_or:
+ case nir_intrinsic_shared_atomic_umax:
+ case nir_intrinsic_shared_atomic_umin:
+ case nir_intrinsic_shared_atomic_xor:
+ case nir_intrinsic_global_atomic_add:
+ case nir_intrinsic_global_atomic_and:
+ case nir_intrinsic_global_atomic_comp_swap:
+ case nir_intrinsic_global_atomic_exchange:
+ case nir_intrinsic_global_atomic_fadd:
+ case nir_intrinsic_global_atomic_fcomp_swap:
+ case nir_intrinsic_global_atomic_fmax:
+ case nir_intrinsic_global_atomic_fmin:
+ case nir_intrinsic_global_atomic_imax:
+ case nir_intrinsic_global_atomic_imin:
+ case nir_intrinsic_global_atomic_or:
+ case nir_intrinsic_global_atomic_umax:
+ case nir_intrinsic_global_atomic_umin:
+ case nir_intrinsic_global_atomic_xor:
return &instr->src[0];
case nir_intrinsic_load_ubo:
case nir_intrinsic_load_ssbo:
[nir_address_format_64bit_global] = {{0}},
[nir_address_format_64bit_bounded_global] = {{0}},
[nir_address_format_32bit_index_offset] = {{.u32 = ~0}, {.u32 = ~0}},
+ [nir_address_format_32bit_index_offset_pack64] = {{.u64 = ~0ull}},
[nir_address_format_vec2_index_32bit_offset] = {{.u32 = ~0}, {.u32 = ~0}, {.u32 = ~0}},
[nir_address_format_32bit_offset] = {{.u32 = ~0}},
[nir_address_format_32bit_offset_as_64bit] = {{.u64 = ~0ull}},
assert(addr0->num_components == 1 && addr1->num_components == 1);
return nir_ieq(b, nir_u2u32(b, addr0), nir_u2u32(b, addr1));
+ case nir_address_format_32bit_index_offset_pack64:
+ assert(addr0->num_components == 1 && addr1->num_components == 1);
+ return nir_ball_iequal(b, nir_unpack_64_2x32(b, addr0), nir_unpack_64_2x32(b, addr1));
+
case nir_address_format_logical:
unreachable("Unsupported address format");
}
case nir_address_format_32bit_global:
case nir_address_format_64bit_global:
case nir_address_format_32bit_offset:
+ case nir_address_format_32bit_index_offset_pack64:
assert(addr0->num_components == 1);
assert(addr1->num_components == 1);
return nir_isub(b, addr0, addr1);
intrin->intrinsic == nir_intrinsic_store_per_vertex_output;
}
+static bool is_dual_slot(nir_intrinsic_instr *intrin)
+{
+ if (intrin->intrinsic == nir_intrinsic_store_output ||
+ intrin->intrinsic == nir_intrinsic_store_per_vertex_output) {
+ return nir_src_bit_size(intrin->src[0]) == 64 &&
+ nir_src_num_components(intrin->src[0]) >= 3;
+ }
+
+ return nir_dest_bit_size(intrin->dest) == 64 &&
+ nir_dest_num_components(intrin->dest) >= 3;
+}
/**
* This pass adds constant offsets to instr->const_index[0] for input/output
nir_src *offset = nir_get_io_offset_src(intrin);
if (nir_src_is_const(*offset)) {
- intrin->const_index[0] += nir_src_as_uint(*offset);
+ unsigned off = nir_src_as_uint(*offset);
+
+ nir_intrinsic_set_base(intrin, nir_intrinsic_base(intrin) + off);
+
+ nir_io_semantics sem = nir_intrinsic_io_semantics(intrin);
+ sem.location += off;
+ /* non-indirect indexing should reduce num_slots */
+ sem.num_slots = is_dual_slot(intrin) ? 2 : 1;
+ nir_intrinsic_set_io_semantics(intrin, sem);
+
b->cursor = nir_before_instr(&intrin->instr);
nir_instr_rewrite_src(&intrin->instr, offset,
nir_src_for_ssa(nir_imm_int(b, 0)));
projects / mesa.git / blobdiff