#include "nir.h"
#include "nir_builder.h"
+static nir_ssa_def*
+build_local_group_size(nir_builder *b, unsigned bit_size)
+{
+ nir_ssa_def *local_size;
+
+ /*
+ * If the local work group size is variable it can't be lowered at this
+ * point, but its intrinsic can still be used.
+ */
+ if (b->shader->info.cs.local_size_variable) {
+ local_size = nir_load_local_group_size(b);
+ } else {
+ /* using a 32 bit constant is safe here as no device/driver needs more
+ * than 32 bits for the local size */
+ nir_const_value local_size_const[3];
+ memset(local_size_const, 0, sizeof(local_size_const));
+ local_size_const[0].u32 = b->shader->info.cs.local_size[0];
+ local_size_const[1].u32 = b->shader->info.cs.local_size[1];
+ local_size_const[2].u32 = b->shader->info.cs.local_size[2];
+ local_size = nir_build_imm(b, 3, 32, local_size_const);
+ }
+
+ return nir_u2u(b, local_size, bit_size);
+}
+
+static nir_ssa_def *
+build_local_invocation_id(nir_builder *b, unsigned bit_size)
+{
+ /* If lower_cs_local_id_from_index is true, then we derive the local
+ * index from the local id.
+ */
+ if (b->shader->options->lower_cs_local_id_from_index) {
+ /* We lower gl_LocalInvocationID from gl_LocalInvocationIndex based
+ * on this formula:
+ *
+ * gl_LocalInvocationID.x =
+ * gl_LocalInvocationIndex % gl_WorkGroupSize.x;
+ * gl_LocalInvocationID.y =
+ * (gl_LocalInvocationIndex / gl_WorkGroupSize.x) %
+ * gl_WorkGroupSize.y;
+ * gl_LocalInvocationID.z =
+ * (gl_LocalInvocationIndex /
+ * (gl_WorkGroupSize.x * gl_WorkGroupSize.y)) %
+ * gl_WorkGroupSize.z;
+ *
+ * However, the final % gl_WorkGroupSize.z does nothing unless we
+ * accidentally end up with a gl_LocalInvocationIndex that is too
+ * large so it can safely be omitted.
+ */
+ nir_ssa_def *local_index = nir_load_local_invocation_index(b);
+ nir_ssa_def *local_size = build_local_group_size(b, 32);
+
+ /* Because no hardware supports a local workgroup size greater than
+ * about 1K, this calculation can be done in 32-bit and can save some
+ * 64-bit arithmetic.
+ */
+ nir_ssa_def *id_x, *id_y, *id_z;
+ id_x = nir_umod(b, local_index,
+ nir_channel(b, local_size, 0));
+ id_y = nir_umod(b, nir_udiv(b, local_index,
+ nir_channel(b, local_size, 0)),
+ nir_channel(b, local_size, 1));
+ id_z = nir_udiv(b, local_index,
+ nir_imul(b, nir_channel(b, local_size, 0),
+ nir_channel(b, local_size, 1)));
+ return nir_u2u(b, nir_vec3(b, id_x, id_y, id_z), bit_size);
+ } else {
+ return nir_u2u(b, nir_load_local_invocation_id(b), bit_size);
+ }
+}
+
+static nir_ssa_def*
+build_global_group_size(nir_builder *b, unsigned bit_size)
+{
+ nir_ssa_def *group_size = build_local_group_size(b, bit_size);
+ nir_ssa_def *num_work_groups = nir_u2u(b, nir_load_num_work_groups(b), bit_size);
+ return nir_imul(b, group_size, num_work_groups);
+}
+
+static nir_ssa_def*
+build_global_invocation_id(nir_builder *b, unsigned bit_size)
+{
+ /* From the GLSL man page for gl_GlobalInvocationID:
+ *
+ * "The value of gl_GlobalInvocationID is equal to
+ * gl_WorkGroupID * gl_WorkGroupSize + gl_LocalInvocationID"
+ */
+ nir_ssa_def *group_size = build_local_group_size(b, bit_size);
+ nir_ssa_def *group_id = nir_u2u(b, nir_load_work_group_id(b), bit_size);
+ nir_ssa_def *local_id = build_local_invocation_id(b, bit_size);
+
+ return nir_iadd(b, nir_imul(b, group_id, group_size), local_id);
+}
+
static bool
convert_block(nir_block *block, nir_builder *b)
{
if (instr->type != nir_instr_type_intrinsic)
continue;
- nir_intrinsic_instr *load_var = nir_instr_as_intrinsic(instr);
-
- if (load_var->intrinsic != nir_intrinsic_load_var)
+ nir_intrinsic_instr *load_deref = nir_instr_as_intrinsic(instr);
+ if (load_deref->intrinsic != nir_intrinsic_load_deref)
continue;
- nir_variable *var = load_var->variables[0]->var;
- if (var->data.mode != nir_var_system_value)
+ nir_deref_instr *deref = nir_src_as_deref(load_deref->src[0]);
+ if (deref->mode != nir_var_system_value)
continue;
- b->cursor = nir_after_instr(&load_var->instr);
+ if (deref->deref_type != nir_deref_type_var) {
+ /* The only one system value that is an array and that is
+ * gl_SampleMask which is always an array of one element.
+ */
+ assert(deref->deref_type == nir_deref_type_array);
+ deref = nir_deref_instr_parent(deref);
+ assert(deref->deref_type == nir_deref_type_var);
+ assert(deref->var->data.location == SYSTEM_VALUE_SAMPLE_MASK_IN);
+ }
+ nir_variable *var = deref->var;
+ b->cursor = nir_after_instr(&load_deref->instr);
+
+ unsigned bit_size = nir_dest_bit_size(load_deref->dest);
nir_ssa_def *sysval = NULL;
switch (var->data.location) {
case SYSTEM_VALUE_GLOBAL_INVOCATION_ID: {
- /* From the GLSL man page for gl_GlobalInvocationID:
- *
- * "The value of gl_GlobalInvocationID is equal to
- * gl_WorkGroupID * gl_WorkGroupSize + gl_LocalInvocationID"
- */
-
- nir_const_value local_size;
- memset(&local_size, 0, sizeof(local_size));
- local_size.u32[0] = b->shader->info.cs.local_size[0];
- local_size.u32[1] = b->shader->info.cs.local_size[1];
- local_size.u32[2] = b->shader->info.cs.local_size[2];
+ sysval = build_global_invocation_id(b, bit_size);
+ break;
+ }
- nir_ssa_def *group_id = nir_load_work_group_id(b);
- nir_ssa_def *local_id = nir_load_local_invocation_id(b);
+ case SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX: {
+ nir_ssa_def *global_id = build_global_invocation_id(b, bit_size);
+ nir_ssa_def *global_size = build_global_group_size(b, bit_size);
- sysval = nir_iadd(b, nir_imul(b, group_id,
- nir_build_imm(b, 3, 32, local_size)),
- local_id);
+ /* index = id.x + ((id.y + (id.z * size.y)) * size.x) */
+ sysval = nir_imul(b, nir_channel(b, global_id, 2),
+ nir_channel(b, global_size, 1));
+ sysval = nir_iadd(b, nir_channel(b, global_id, 1), sysval);
+ sysval = nir_imul(b, nir_channel(b, global_size, 0), sysval);
+ sysval = nir_iadd(b, nir_channel(b, global_id, 0), sysval);
break;
}
nir_ssa_def *size_y =
nir_imm_int(b, b->shader->info.cs.local_size[1]);
+ /* Because no hardware supports a local workgroup size greater than
+ * about 1K, this calculation can be done in 32-bit and can save some
+ * 64-bit arithmetic.
+ */
sysval = nir_imul(b, nir_channel(b, local_id, 2),
nir_imul(b, size_x, size_y));
sysval = nir_iadd(b, sysval,
nir_imul(b, nir_channel(b, local_id, 1), size_x));
sysval = nir_iadd(b, sysval, nir_channel(b, local_id, 0));
+ sysval = nir_u2u(b, sysval, bit_size);
break;
}
+ case SYSTEM_VALUE_LOCAL_INVOCATION_ID:
+ sysval = build_local_invocation_id(b, bit_size);
+ break;
+
case SYSTEM_VALUE_LOCAL_GROUP_SIZE: {
- nir_const_value local_size;
- memset(&local_size, 0, sizeof(local_size));
- local_size.u32[0] = b->shader->info.cs.local_size[0];
- local_size.u32[1] = b->shader->info.cs.local_size[1];
- local_size.u32[2] = b->shader->info.cs.local_size[2];
- sysval = nir_build_imm(b, 3, 32, local_size);
+ sysval = build_local_group_size(b, bit_size);
break;
}
nir_load_first_vertex(b));
break;
+ case SYSTEM_VALUE_HELPER_INVOCATION:
+ if (b->shader->options->lower_helper_invocation) {
+ nir_ssa_def *tmp;
+
+ tmp = nir_ishl(b,
+ nir_imm_int(b, 1),
+ nir_load_sample_id_no_per_sample(b));
+
+ tmp = nir_iand(b,
+ nir_load_sample_mask_in(b),
+ tmp);
+
+ sysval = nir_inot(b, nir_i2b(b, tmp));
+ }
+
+ break;
+
case SYSTEM_VALUE_INSTANCE_INDEX:
sysval = nir_iadd(b,
nir_load_instance_id(b),
sysval = nir_imm_int(b, 0);
break;
+ case SYSTEM_VALUE_GLOBAL_GROUP_SIZE: {
+ sysval = build_global_group_size(b, bit_size);
+ break;
+ }
+
default:
break;
}
if (sysval == NULL) {
nir_intrinsic_op sysval_op =
nir_intrinsic_from_system_value(var->data.location);
- sysval = nir_load_system_value(b, sysval_op, 0);
+ sysval = nir_load_system_value(b, sysval_op, 0,
+ load_deref->dest.ssa.bit_size);
}
- nir_ssa_def_rewrite_uses(&load_var->dest.ssa, nir_src_for_ssa(sysval));
- nir_instr_remove(&load_var->instr);
+ nir_ssa_def_rewrite_uses(&load_deref->dest.ssa, nir_src_for_ssa(sysval));
+ nir_instr_remove(&load_deref->instr);
progress = true;
}
progress = convert_impl(function->impl) || progress;
}
+ /* We're going to delete the variables so we need to clean up all those
+ * derefs we left lying around.
+ */
+ nir_remove_dead_derefs(shader);
+
exec_list_make_empty(&shader->system_values);
return progress;