#include "nir_builder.h"
#include "nir_deref.h"
+#include "util/u_math.h"
+
struct lower_io_state {
void *dead_ctx;
nir_builder builder;
}
}
+static nir_intrinsic_op
+shared_atomic_for_deref(nir_intrinsic_op deref_op)
+{
+ switch (deref_op) {
+#define OP(O) case nir_intrinsic_deref_##O: return nir_intrinsic_shared_##O;
+ OP(atomic_exchange)
+ OP(atomic_comp_swap)
+ OP(atomic_add)
+ OP(atomic_imin)
+ OP(atomic_umin)
+ OP(atomic_imax)
+ OP(atomic_umax)
+ OP(atomic_and)
+ OP(atomic_or)
+ OP(atomic_xor)
+ OP(atomic_fadd)
+ OP(atomic_fmin)
+ OP(atomic_fmax)
+ OP(atomic_fcomp_swap)
+#undef OP
+ default:
+ unreachable("Invalid shared atomic");
+ }
+}
+
void
nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
int (*type_size)(const struct glsl_type *, bool))
return offset;
}
-static nir_intrinsic_instr *
-lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
- nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
- unsigned component, const struct glsl_type *type)
+static nir_ssa_def *
+emit_load(struct lower_io_state *state,
+ nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
+ unsigned component, unsigned num_components, unsigned bit_size,
+ nir_alu_type type)
{
- const nir_shader *nir = state->builder.shader;
+ nir_builder *b = &state->builder;
+ const nir_shader *nir = b->shader;
nir_variable_mode mode = var->data.mode;
nir_ssa_def *barycentric = NULL;
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(state->builder.shader, op);
- load->num_components = intrin->num_components;
+ load->num_components = num_components;
nir_intrinsic_set_base(load, var->data.driver_location);
if (mode == nir_var_shader_in || mode == nir_var_shader_out)
if (load->intrinsic == nir_intrinsic_load_input ||
load->intrinsic == nir_intrinsic_load_uniform)
- nir_intrinsic_set_type(load, nir_get_nir_type_for_glsl_type(type));
+ nir_intrinsic_set_type(load, type);
if (vertex_index) {
load->src[0] = nir_src_for_ssa(vertex_index);
load->src[0] = nir_src_for_ssa(offset);
}
- return load;
+ nir_ssa_dest_init(&load->instr, &load->dest,
+ num_components, bit_size, NULL);
+ nir_builder_instr_insert(b, &load->instr);
+
+ return &load->dest.ssa;
}
-static nir_intrinsic_instr *
-lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
- nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
- unsigned component, const struct glsl_type *type)
+static nir_ssa_def *
+lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+ nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
+ unsigned component, const struct glsl_type *type)
{
+ assert(intrin->dest.is_ssa);
+ if (intrin->dest.ssa.bit_size == 64 &&
+ (state->options & nir_lower_io_lower_64bit_to_32)) {
+ nir_builder *b = &state->builder;
+
+ const unsigned slot_size = state->type_size(glsl_dvec_type(2), false);
+
+ nir_ssa_def *comp64[4];
+ assert(component == 0 || component == 2);
+ unsigned dest_comp = 0;
+ while (dest_comp < intrin->dest.ssa.num_components) {
+ const unsigned num_comps =
+ MIN2(intrin->dest.ssa.num_components - dest_comp,
+ (4 - component) / 2);
+
+ nir_ssa_def *data32 =
+ emit_load(state, vertex_index, var, offset, component,
+ num_comps * 2, 32, nir_type_uint32);
+ for (unsigned i = 0; i < num_comps; i++) {
+ comp64[dest_comp + i] =
+ nir_pack_64_2x32(b, nir_channels(b, data32, 3 << (i * 2)));
+ }
+
+ /* Only the first store has a component offset */
+ component = 0;
+ dest_comp += num_comps;
+ offset = nir_iadd_imm(b, offset, slot_size);
+ }
+
+ return nir_vec(b, comp64, intrin->dest.ssa.num_components);
+ } else {
+ return emit_load(state, vertex_index, var, offset, component,
+ intrin->dest.ssa.num_components,
+ intrin->dest.ssa.bit_size,
+ nir_get_nir_type_for_glsl_type(type));
+ }
+}
+
+static void
+emit_store(struct lower_io_state *state, nir_ssa_def *data,
+ nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
+ unsigned component, unsigned num_components,
+ nir_component_mask_t write_mask, nir_alu_type type)
+{
+ nir_builder *b = &state->builder;
nir_variable_mode mode = var->data.mode;
nir_intrinsic_op op;
nir_intrinsic_instr *store =
nir_intrinsic_instr_create(state->builder.shader, op);
- store->num_components = intrin->num_components;
+ store->num_components = num_components;
- nir_src_copy(&store->src[0], &intrin->src[1], store);
+ store->src[0] = nir_src_for_ssa(data);
nir_intrinsic_set_base(store, var->data.driver_location);
nir_intrinsic_set_component(store, component);
if (store->intrinsic == nir_intrinsic_store_output)
- nir_intrinsic_set_type(store, nir_get_nir_type_for_glsl_type(type));
+ nir_intrinsic_set_type(store, type);
- nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
+ nir_intrinsic_set_write_mask(store, write_mask);
if (vertex_index)
store->src[1] = nir_src_for_ssa(vertex_index);
store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
- return store;
+ nir_builder_instr_insert(b, &store->instr);
+}
+
+static void
+lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
+ nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
+ unsigned component, const struct glsl_type *type)
+{
+ assert(intrin->src[1].is_ssa);
+ if (intrin->src[1].ssa->bit_size == 64 &&
+ (state->options & nir_lower_io_lower_64bit_to_32)) {
+ nir_builder *b = &state->builder;
+
+ const unsigned slot_size = state->type_size(glsl_dvec_type(2), false);
+
+ assert(component == 0 || component == 2);
+ unsigned src_comp = 0;
+ nir_component_mask_t write_mask = nir_intrinsic_write_mask(intrin);
+ while (src_comp < intrin->num_components) {
+ const unsigned num_comps =
+ MIN2(intrin->num_components - src_comp,
+ (4 - component) / 2);
+
+ if (write_mask & BITFIELD_MASK(num_comps)) {
+ nir_ssa_def *data =
+ nir_channels(b, intrin->src[1].ssa,
+ BITFIELD_RANGE(src_comp, num_comps));
+ nir_ssa_def *data32 = nir_bitcast_vector(b, data, 32);
+
+ nir_component_mask_t write_mask32 = 0;
+ for (unsigned i = 0; i < num_comps; i++) {
+ if (write_mask & BITFIELD_MASK(num_comps) & (1 << i))
+ write_mask32 |= 3 << (i * 2);
+ }
+
+ emit_store(state, data32, vertex_index, var, offset,
+ component, data32->num_components, write_mask32,
+ nir_type_uint32);
+ }
+
+ /* Only the first store has a component offset */
+ component = 0;
+ src_comp += num_comps;
+ write_mask >>= num_comps;
+ offset = nir_iadd_imm(b, offset, slot_size);
+ }
+ } else {
+ emit_store(state, intrin->src[1].ssa, vertex_index, var, offset,
+ component, intrin->num_components,
+ nir_intrinsic_write_mask(intrin),
+ nir_get_nir_type_for_glsl_type(type));
+ }
}
-static nir_intrinsic_instr *
+static nir_ssa_def *
lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
nir_variable *var, nir_ssa_def *offset)
{
+ nir_builder *b = &state->builder;
assert(var->data.mode == nir_var_mem_shared);
- nir_intrinsic_op op;
- switch (intrin->intrinsic) {
-#define OP(O) case nir_intrinsic_deref_##O: op = nir_intrinsic_shared_##O; break;
- OP(atomic_exchange)
- OP(atomic_comp_swap)
- OP(atomic_add)
- OP(atomic_imin)
- OP(atomic_umin)
- OP(atomic_imax)
- OP(atomic_umax)
- OP(atomic_and)
- OP(atomic_or)
- OP(atomic_xor)
- OP(atomic_fadd)
- OP(atomic_fmin)
- OP(atomic_fmax)
- OP(atomic_fcomp_swap)
-#undef OP
- default:
- unreachable("Invalid atomic");
- }
+ nir_intrinsic_op op = shared_atomic_for_deref(intrin->intrinsic);
nir_intrinsic_instr *atomic =
nir_intrinsic_instr_create(state->builder.shader, op);
nir_src_copy(&atomic->src[i], &intrin->src[i], atomic);
}
- return atomic;
+ if (nir_intrinsic_infos[op].has_dest) {
+ assert(intrin->dest.is_ssa);
+ assert(nir_intrinsic_infos[intrin->intrinsic].has_dest);
+ nir_ssa_dest_init(&atomic->instr, &atomic->dest,
+ intrin->dest.ssa.num_components,
+ intrin->dest.ssa.bit_size, NULL);
+ }
+
+ nir_builder_instr_insert(b, &atomic->instr);
+
+ return nir_intrinsic_infos[op].has_dest ? &atomic->dest.ssa : NULL;
}
-static nir_intrinsic_instr *
+static nir_ssa_def *
lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
nir_variable *var, nir_ssa_def *offset, unsigned component,
const struct glsl_type *type)
{
+ nir_builder *b = &state->builder;
assert(var->data.mode == nir_var_shader_in);
/* Ignore interpolateAt() for flat variables - flat is flat. */
if (var->data.interpolation == INTERP_MODE_FLAT)
return lower_load(intrin, state, NULL, var, offset, component, type);
+ /* None of the supported APIs allow interpolation on 64-bit things */
+ assert(intrin->dest.is_ssa && intrin->dest.ssa.bit_size <= 32);
+
nir_intrinsic_op bary_op;
switch (intrin->intrinsic) {
case nir_intrinsic_interp_deref_at_centroid:
intrin->intrinsic == nir_intrinsic_interp_deref_at_offset)
nir_src_copy(&bary_setup->src[0], &intrin->src[1], bary_setup);
- nir_builder_instr_insert(&state->builder, &bary_setup->instr);
+ nir_builder_instr_insert(b, &bary_setup->instr);
nir_intrinsic_instr *load =
nir_intrinsic_instr_create(state->builder.shader,
load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
load->src[1] = nir_src_for_ssa(offset);
- return load;
+ assert(intrin->dest.is_ssa);
+ nir_ssa_dest_init(&load->instr, &load->dest,
+ intrin->dest.ssa.num_components,
+ intrin->dest.ssa.bit_size, NULL);
+ nir_builder_instr_insert(b, &load->instr);
+
+ return &load->dest.ssa;
}
static bool
state->type_size, &component_offset,
bindless_type_size);
- nir_intrinsic_instr *replacement;
+ nir_ssa_def *replacement = NULL;
switch (intrin->intrinsic) {
case nir_intrinsic_load_deref:
break;
case nir_intrinsic_store_deref:
- replacement = lower_store(intrin, state, vertex_index, var, offset,
- component_offset, deref->type);
+ lower_store(intrin, state, vertex_index, var, offset,
+ component_offset, deref->type);
break;
case nir_intrinsic_deref_atomic_add:
continue;
}
- if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
- if (intrin->dest.is_ssa) {
- nir_ssa_dest_init(&replacement->instr, &replacement->dest,
- intrin->dest.ssa.num_components,
- intrin->dest.ssa.bit_size, NULL);
- nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
- nir_src_for_ssa(&replacement->dest.ssa));
- } else {
- nir_dest_copy(&replacement->dest, &intrin->dest, &intrin->instr);
- }
+ if (replacement) {
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
+ nir_src_for_ssa(replacement));
}
-
- nir_instr_insert_before(&intrin->instr, &replacement->instr);
nir_instr_remove(&intrin->instr);
progress = true;
}
assert(addr_format_is_global(addr_format));
op = nir_intrinsic_load_kernel_input;
break;
+ case nir_var_mem_shared:
+ assert(addr_format == nir_address_format_32bit_offset);
+ op = nir_intrinsic_load_shared;
+ break;
default:
unreachable("Unsupported explicit IO variable mode");
}
if (addr_format_is_global(addr_format)) {
load->src[0] = nir_src_for_ssa(addr_to_global(b, addr, addr_format));
+ } else if (addr_format == nir_address_format_32bit_offset) {
+ assert(addr->num_components == 1);
+ load->src[0] = nir_src_for_ssa(addr);
} else {
load->src[0] = nir_src_for_ssa(addr_to_index(b, addr, addr_format));
load->src[1] = nir_src_for_ssa(addr_to_offset(b, addr, addr_format));
}
- if (mode != nir_var_mem_ubo && mode != nir_var_shader_in)
+ if (mode != nir_var_mem_ubo && mode != nir_var_shader_in && mode != nir_var_mem_shared)
nir_intrinsic_set_access(load, nir_intrinsic_access(intrin));
/* TODO: We should try and provide a better alignment. For OpenCL, we need
assert(addr_format_is_global(addr_format));
op = nir_intrinsic_store_global;
break;
+ case nir_var_mem_shared:
+ assert(addr_format == nir_address_format_32bit_offset);
+ op = nir_intrinsic_store_shared;
+ break;
default:
unreachable("Unsupported explicit IO variable mode");
}
store->src[0] = nir_src_for_ssa(value);
if (addr_format_is_global(addr_format)) {
store->src[1] = nir_src_for_ssa(addr_to_global(b, addr, addr_format));
+ } else if (addr_format == nir_address_format_32bit_offset) {
+ assert(addr->num_components == 1);
+ store->src[1] = nir_src_for_ssa(addr);
} else {
store->src[1] = nir_src_for_ssa(addr_to_index(b, addr, addr_format));
store->src[2] = nir_src_for_ssa(addr_to_offset(b, addr, addr_format));
nir_intrinsic_set_write_mask(store, write_mask);
- nir_intrinsic_set_access(store, nir_intrinsic_access(intrin));
+ if (mode != nir_var_mem_shared)
+ nir_intrinsic_set_access(store, nir_intrinsic_access(intrin));
/* TODO: We should try and provide a better alignment. For OpenCL, we need
* to plumb the alignment through from SPIR-V when we have one.
assert(addr_format_is_global(addr_format));
op = global_atomic_for_deref(intrin->intrinsic);
break;
+ case nir_var_mem_shared:
+ assert(addr_format == nir_address_format_32bit_offset);
+ op = shared_atomic_for_deref(intrin->intrinsic);
+ break;
default:
unreachable("Unsupported explicit IO variable mode");
}
unsigned src = 0;
if (addr_format_is_global(addr_format)) {
atomic->src[src++] = nir_src_for_ssa(addr_to_global(b, addr, addr_format));
+ } else if (addr_format == nir_address_format_32bit_offset) {
+ assert(addr->num_components == 1);
+ atomic->src[src++] = nir_src_for_ssa(addr);
} else {
atomic->src[src++] = nir_src_for_ssa(addr_to_index(b, addr, addr_format));
atomic->src[src++] = nir_src_for_ssa(addr_to_offset(b, addr, addr_format));
/* Global atomics don't have access flags because they assume that the
* address may be non-uniform.
*/
- if (!addr_format_is_global(addr_format))
+ if (!addr_format_is_global(addr_format) && mode != nir_var_mem_shared)
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);
+ assert(deref->mode & (nir_var_shader_in | nir_var_mem_shared));
return nir_imm_intN_t(b, deref->var->data.driver_location,
deref->dest.ssa.bit_size);
return progress;
}
+static bool
+nir_lower_vars_to_explicit_types_impl(nir_function_impl *impl,
+ nir_variable_mode modes,
+ glsl_type_size_align_func type_info)
+{
+ bool progress = false;
+
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type != nir_instr_type_deref)
+ continue;
+
+ nir_deref_instr *deref = nir_instr_as_deref(instr);
+ if (!(deref->mode & modes))
+ continue;
+
+ unsigned size, alignment;
+ const struct glsl_type *new_type =
+ glsl_get_explicit_type_for_size_align(deref->type, type_info, &size, &alignment);
+ if (new_type != deref->type) {
+ progress = true;
+ deref->type = new_type;
+ }
+ if (deref->deref_type == nir_deref_type_cast) {
+ /* See also glsl_type::get_explicit_type_for_size_align() */
+ unsigned new_stride = align(size, alignment);
+ if (new_stride != deref->cast.ptr_stride) {
+ deref->cast.ptr_stride = new_stride;
+ progress = true;
+ }
+ }
+ }
+ }
+
+ if (progress) {
+ nir_metadata_preserve(impl, nir_metadata_block_index |
+ nir_metadata_dominance |
+ nir_metadata_live_ssa_defs |
+ nir_metadata_loop_analysis);
+ }
+
+ return progress;
+}
+
+static bool
+lower_vars_to_explicit(nir_shader *shader,
+ struct exec_list *vars, nir_variable_mode mode,
+ glsl_type_size_align_func type_info)
+{
+ bool progress = false;
+ unsigned offset = 0;
+ nir_foreach_variable(var, vars) {
+ unsigned size, align;
+ 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;
+ var->type = explicit_type;
+ }
+
+ var->data.driver_location = ALIGN_POT(offset, align);
+ offset = var->data.driver_location + size;
+ }
+
+ if (mode == nir_var_mem_shared) {
+ shader->info.cs.shared_size = offset;
+ shader->num_shared = offset;
+ }
+
+ return progress;
+}
+
+bool
+nir_lower_vars_to_explicit_types(nir_shader *shader,
+ nir_variable_mode modes,
+ glsl_type_size_align_func type_info)
+{
+ /* TODO: Situations which need to be handled to support more modes:
+ * - row-major matrices
+ * - compact shader inputs/outputs
+ * - interface types
+ */
+ nir_variable_mode supported = nir_var_mem_shared | nir_var_shader_temp | nir_var_function_temp;
+ assert(!(modes & ~supported) && "unsupported");
+
+ bool progress = false;
+
+ if (modes & nir_var_mem_shared)
+ progress |= lower_vars_to_explicit(shader, &shader->shared, nir_var_mem_shared, type_info);
+ if (modes & nir_var_shader_temp)
+ progress |= lower_vars_to_explicit(shader, &shader->globals, nir_var_shader_temp, type_info);
+
+ nir_foreach_function(function, shader) {
+ if (function->impl) {
+ if (modes & nir_var_function_temp)
+ progress |= lower_vars_to_explicit(shader, &function->impl->locals, nir_var_function_temp, type_info);
+
+ progress |= nir_lower_vars_to_explicit_types_impl(function->impl, modes, type_info);
+ }
+ }
+
+ return progress;
+}
+
/**
* Return the offset source for a load/store intrinsic.
*/
projects / mesa.git / blobdiff