#include "program.h"
#include "compiler/nir/nir_control_flow.h"
#include "compiler/nir/nir_builder.h"
+#include "compiler/nir/nir_builtin_builder.h"
#include "compiler/nir/nir_deref.h"
#include "main/errors.h"
-#include "main/imports.h"
#include "main/mtypes.h"
#include "main/shaderobj.h"
#include "util/u_math.h"
virtual void visit(ir_loop *);
virtual void visit(ir_if *);
virtual void visit(ir_discard *);
+ virtual void visit(ir_demote *);
virtual void visit(ir_loop_jump *);
virtual void visit(ir_return *);
virtual void visit(ir_call *);
* inline functions. That way they get properly initialized at the top
* of the function and not at the top of its caller.
*/
- nir_lower_constant_initializers(shader, (nir_variable_mode)~0);
+ nir_lower_variable_initializers(shader, (nir_variable_mode)~0);
nir_lower_returns(shader);
nir_inline_functions(shader);
nir_opt_deref(shader);
}
}
- /* Remap the locations to slots so those requiring two slots will occupy
- * two locations. For instance, if we have in the IR code a dvec3 attr0 in
- * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
- * locations/slots 0 and 1, and attr1 will use location/slot 2 */
- if (shader->info.stage == MESA_SHADER_VERTEX)
- nir_remap_dual_slot_attributes(shader, &sh->Program->DualSlotInputs);
-
shader->info.name = ralloc_asprintf(shader, "GLSL%d", shader_prog->Name);
if (shader_prog->Label)
shader->info.label = ralloc_strdup(shader, shader_prog->Label);
break;
case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_FLOAT16:
case GLSL_TYPE_DOUBLE:
if (cols > 1) {
ret->elements = ralloc_array(mem_ctx, nir_constant *, cols);
col_const->values[r].f32 = ir->value.f[c * rows + r];
break;
+ case GLSL_TYPE_FLOAT16:
+ for (unsigned r = 0; r < rows; r++)
+ col_const->values[r].u16 = ir->value.f16[c * rows + r];
+ break;
+
case GLSL_TYPE_DOUBLE:
for (unsigned r = 0; r < rows; r++)
col_const->values[r].f64 = ir->value.d[c * rows + r];
ret->values[r].f32 = ir->value.f[r];
break;
+ case GLSL_TYPE_FLOAT16:
+ for (unsigned r = 0; r < rows; r++)
+ ret->values[r].u16 = ir->value.f16[r];
+ break;
+
case GLSL_TYPE_DOUBLE:
for (unsigned r = 0; r < rows; r++)
ret->values[r].f64 = ir->value.d[r];
return glsl_type::get_array_instance(elem_type, array_type->length);
}
+static unsigned
+get_nir_how_declared(unsigned how_declared)
+{
+ if (how_declared == ir_var_hidden)
+ return nir_var_hidden;
+
+ return nir_var_declared_normally;
+}
+
void
nir_visitor::visit(ir_variable *ir)
{
var->data.centroid = ir->data.centroid;
var->data.sample = ir->data.sample;
var->data.patch = ir->data.patch;
+ var->data.how_declared = get_nir_how_declared(ir->data.how_declared);
var->data.invariant = ir->data.invariant;
var->data.location = ir->data.location;
var->data.stream = ir->data.stream;
+ if (ir->data.stream & (1u << 31))
+ var->data.stream |= NIR_STREAM_PACKED;
+
+ var->data.precision = ir->data.precision;
+ var->data.explicit_location = ir->data.explicit_location;
+ var->data.matrix_layout = ir->data.matrix_layout;
+ var->data.from_named_ifc_block = ir->data.from_named_ifc_block;
var->data.compact = false;
switch(ir->data.mode) {
break;
case ir_var_shader_in:
- if (shader->info.stage == MESA_SHADER_FRAGMENT &&
- ir->data.location == VARYING_SLOT_FACE) {
- /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
- var->data.location = SYSTEM_VALUE_FRONT_FACE;
- var->data.mode = nir_var_system_value;
- } else if (shader->info.stage == MESA_SHADER_GEOMETRY &&
- ir->data.location == VARYING_SLOT_PRIMITIVE_ID) {
+ if (shader->info.stage == MESA_SHADER_GEOMETRY &&
+ ir->data.location == VARYING_SLOT_PRIMITIVE_ID) {
/* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
var->data.location = SYSTEM_VALUE_PRIMITIVE_ID;
var->data.mode = nir_var_system_value;
unreachable("not reached");
}
- unsigned image_access = 0;
+ unsigned mem_access = 0;
if (ir->data.memory_read_only)
- image_access |= ACCESS_NON_WRITEABLE;
+ mem_access |= ACCESS_NON_WRITEABLE;
if (ir->data.memory_write_only)
- image_access |= ACCESS_NON_READABLE;
+ mem_access |= ACCESS_NON_READABLE;
if (ir->data.memory_coherent)
- image_access |= ACCESS_COHERENT;
+ mem_access |= ACCESS_COHERENT;
if (ir->data.memory_volatile)
- image_access |= ACCESS_VOLATILE;
+ mem_access |= ACCESS_VOLATILE;
if (ir->data.memory_restrict)
- image_access |= ACCESS_RESTRICT;
+ mem_access |= ACCESS_RESTRICT;
+
+ var->interface_type = ir->get_interface_type();
/* For UBO and SSBO variables, we need explicit types */
if (var->data.mode & (nir_var_mem_ubo | nir_var_mem_ssbo)) {
const glsl_type *explicit_ifc_type =
ir->get_interface_type()->get_explicit_interface_type(supports_std430);
+ var->interface_type = explicit_ifc_type;
+
if (ir->type->without_array()->is_interface()) {
/* If the type contains the interface, wrap the explicit type in the
* right number of arrays.
var->type = field->type;
if (field->memory_read_only)
- image_access |= ACCESS_NON_WRITEABLE;
+ mem_access |= ACCESS_NON_WRITEABLE;
if (field->memory_write_only)
- image_access |= ACCESS_NON_READABLE;
+ mem_access |= ACCESS_NON_READABLE;
if (field->memory_coherent)
- image_access |= ACCESS_COHERENT;
+ mem_access |= ACCESS_COHERENT;
if (field->memory_volatile)
- image_access |= ACCESS_VOLATILE;
+ mem_access |= ACCESS_VOLATILE;
if (field->memory_restrict)
- image_access |= ACCESS_RESTRICT;
+ mem_access |= ACCESS_RESTRICT;
found = true;
break;
var->data.explicit_binding = ir->data.explicit_binding;
var->data.bindless = ir->data.bindless;
var->data.offset = ir->data.offset;
+ var->data.access = (gl_access_qualifier)mem_access;
- var->data.image.access = (gl_access_qualifier)image_access;
- var->data.image.format = ir->data.image_format;
+ if (var->type->without_array()->is_image()) {
+ var->data.image.format = ir->data.image_format;
+ } else if (var->data.mode == nir_var_shader_out) {
+ var->data.xfb.buffer = ir->data.xfb_buffer;
+ var->data.xfb.stride = ir->data.xfb_stride;
+ }
var->data.fb_fetch_output = ir->data.fb_fetch_output;
var->data.explicit_xfb_buffer = ir->data.explicit_xfb_buffer;
var->data.explicit_xfb_stride = ir->data.explicit_xfb_stride;
- var->data.xfb_buffer = ir->data.xfb_buffer;
- var->data.xfb_stride = ir->data.xfb_stride;
var->num_state_slots = ir->get_num_state_slots();
if (var->num_state_slots > 0) {
var->constant_initializer = constant_copy(ir->constant_initializer, var);
- var->interface_type = ir->get_interface_type();
-
if (var->data.mode == nir_var_function_temp)
nir_function_impl_add_variable(impl, var);
else
nir_builder_instr_insert(&b, &discard->instr);
}
+void
+nir_visitor::visit(ir_demote *ir)
+{
+ nir_intrinsic_instr *demote =
+ nir_intrinsic_instr_create(this->shader, nir_intrinsic_demote);
+
+ nir_builder_instr_insert(&b, &demote->instr);
+}
+
void
nir_visitor::visit(ir_emit_vertex *ir)
{
nir_deref_path path;
nir_deref_path_init(&path, deref, NULL);
- unsigned qualifiers = path.path[0]->var->data.image.access;
+ unsigned qualifiers = path.path[0]->var->data.access;
const glsl_type *parent_type = path.path[0]->type;
for (nir_deref_instr **cur_ptr = &path.path[1]; *cur_ptr; cur_ptr++) {
if (field->memory_restrict)
qualifiers |= ACCESS_RESTRICT;
}
-
+
parent_type = cur->type;
}
: nir_intrinsic_image_deref_atomic_fadd;
break;
case ir_intrinsic_image_atomic_min:
- op = nir_intrinsic_image_deref_atomic_min;
+ if (ir->return_deref->type == glsl_type::int_type)
+ op = nir_intrinsic_image_deref_atomic_imin;
+ else if (ir->return_deref->type == glsl_type::uint_type)
+ op = nir_intrinsic_image_deref_atomic_umin;
+ else
+ unreachable("Invalid type");
break;
case ir_intrinsic_image_atomic_max:
- op = nir_intrinsic_image_deref_atomic_max;
+ if (ir->return_deref->type == glsl_type::int_type)
+ op = nir_intrinsic_image_deref_atomic_imax;
+ else if (ir->return_deref->type == glsl_type::uint_type)
+ op = nir_intrinsic_image_deref_atomic_umax;
+ else
+ unreachable("Invalid type");
break;
case ir_intrinsic_image_atomic_and:
op = nir_intrinsic_image_deref_atomic_and;
case ir_intrinsic_image_atomic_comp_swap:
op = nir_intrinsic_image_deref_atomic_comp_swap;
break;
+ case ir_intrinsic_image_atomic_inc_wrap:
+ op = nir_intrinsic_image_deref_atomic_inc_wrap;
+ break;
+ case ir_intrinsic_image_atomic_dec_wrap:
+ op = nir_intrinsic_image_deref_atomic_dec_wrap;
+ break;
case ir_intrinsic_memory_barrier:
op = nir_intrinsic_memory_barrier;
break;
op = nir_intrinsic_image_deref_samples;
break;
case ir_intrinsic_ssbo_store:
- op = nir_intrinsic_store_ssbo;
- break;
case ir_intrinsic_ssbo_load:
- op = nir_intrinsic_load_ssbo;
- break;
case ir_intrinsic_ssbo_atomic_add:
- op = ir->return_deref->type->is_integer_32_64()
- ? nir_intrinsic_ssbo_atomic_add : nir_intrinsic_ssbo_atomic_fadd;
- break;
case ir_intrinsic_ssbo_atomic_and:
- op = nir_intrinsic_ssbo_atomic_and;
- break;
case ir_intrinsic_ssbo_atomic_or:
- op = nir_intrinsic_ssbo_atomic_or;
- break;
case ir_intrinsic_ssbo_atomic_xor:
- op = nir_intrinsic_ssbo_atomic_xor;
- break;
case ir_intrinsic_ssbo_atomic_min:
- assert(ir->return_deref);
- if (ir->return_deref->type == glsl_type::int_type)
- op = nir_intrinsic_ssbo_atomic_imin;
- else if (ir->return_deref->type == glsl_type::uint_type)
- op = nir_intrinsic_ssbo_atomic_umin;
- else if (ir->return_deref->type == glsl_type::float_type)
- op = nir_intrinsic_ssbo_atomic_fmin;
- else
- unreachable("Invalid type");
- break;
case ir_intrinsic_ssbo_atomic_max:
- assert(ir->return_deref);
- if (ir->return_deref->type == glsl_type::int_type)
- op = nir_intrinsic_ssbo_atomic_imax;
- else if (ir->return_deref->type == glsl_type::uint_type)
- op = nir_intrinsic_ssbo_atomic_umax;
- else if (ir->return_deref->type == glsl_type::float_type)
- op = nir_intrinsic_ssbo_atomic_fmax;
- else
- unreachable("Invalid type");
- break;
case ir_intrinsic_ssbo_atomic_exchange:
- op = nir_intrinsic_ssbo_atomic_exchange;
- break;
case ir_intrinsic_ssbo_atomic_comp_swap:
- op = ir->return_deref->type->is_integer_32_64()
- ? nir_intrinsic_ssbo_atomic_comp_swap
- : nir_intrinsic_ssbo_atomic_fcomp_swap;
- break;
+ /* SSBO store/loads should only have been lowered in GLSL IR for
+ * non-nir drivers, NIR drivers make use of gl_nir_lower_buffers()
+ * instead.
+ */
+ unreachable("Invalid operation nir doesn't want lowered ssbo "
+ "store/loads");
case ir_intrinsic_shader_clock:
op = nir_intrinsic_shader_clock;
break;
case ir_intrinsic_read_first_invocation:
op = nir_intrinsic_read_first_invocation;
break;
+ case ir_intrinsic_helper_invocation:
+ op = nir_intrinsic_is_helper_invocation;
+ break;
default:
unreachable("not reached");
}
case nir_intrinsic_image_deref_load:
case nir_intrinsic_image_deref_store:
case nir_intrinsic_image_deref_atomic_add:
- case nir_intrinsic_image_deref_atomic_min:
- case nir_intrinsic_image_deref_atomic_max:
+ case nir_intrinsic_image_deref_atomic_imin:
+ case nir_intrinsic_image_deref_atomic_umin:
+ case nir_intrinsic_image_deref_atomic_imax:
+ case nir_intrinsic_image_deref_atomic_umax:
case nir_intrinsic_image_deref_atomic_and:
case nir_intrinsic_image_deref_atomic_or:
case nir_intrinsic_image_deref_atomic_xor:
case nir_intrinsic_image_deref_atomic_comp_swap:
case nir_intrinsic_image_deref_atomic_fadd:
case nir_intrinsic_image_deref_samples:
- case nir_intrinsic_image_deref_size: {
+ case nir_intrinsic_image_deref_size:
+ case nir_intrinsic_image_deref_atomic_inc_wrap:
+ case nir_intrinsic_image_deref_atomic_dec_wrap: {
nir_ssa_undef_instr *instr_undef =
nir_ssa_undef_instr_create(shader, 1, 32);
nir_builder_instr_insert(&b, &instr_undef->instr);
instr->src[3] =
nir_src_for_ssa(evaluate_rvalue((ir_dereference *)param));
param = param->get_next();
+ } else if (op == nir_intrinsic_image_deref_load) {
+ instr->src[3] = nir_src_for_ssa(nir_imm_int(&b, 0)); /* LOD */
}
if (!param->is_tail_sentinel()) {
instr->src[4] =
nir_src_for_ssa(evaluate_rvalue((ir_dereference *)param));
param = param->get_next();
+ } else if (op == nir_intrinsic_image_deref_store) {
+ instr->src[4] = nir_src_for_ssa(nir_imm_int(&b, 0)); /* LOD */
}
+
nir_builder_instr_insert(&b, &instr->instr);
break;
}
case nir_intrinsic_shader_clock:
nir_ssa_dest_init(&instr->instr, &instr->dest, 2, 32, NULL);
instr->num_components = 2;
+ nir_intrinsic_set_memory_scope(instr, NIR_SCOPE_SUBGROUP);
nir_builder_instr_insert(&b, &instr->instr);
break;
case nir_intrinsic_begin_invocation_interlock:
nir_builder_instr_insert(&b, &instr->instr);
break;
}
- case nir_intrinsic_load_ssbo: {
- exec_node *param = ir->actual_parameters.get_head();
- ir_rvalue *block = ((ir_instruction *)param)->as_rvalue();
-
- param = param->get_next();
- ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue();
-
- instr->src[0] = nir_src_for_ssa(evaluate_rvalue(block));
- instr->src[1] = nir_src_for_ssa(evaluate_rvalue(offset));
-
- const glsl_type *type = ir->return_deref->var->type;
- instr->num_components = type->vector_elements;
- intrinsic_set_std430_align(instr, type);
-
- /* Setup destination register */
- unsigned bit_size = type->is_boolean() ? 32 : glsl_get_bit_size(type);
- nir_ssa_dest_init(&instr->instr, &instr->dest,
- type->vector_elements, bit_size, NULL);
-
- /* Insert the created nir instruction now since in the case of boolean
- * result we will need to emit another instruction after it
- */
- nir_builder_instr_insert(&b, &instr->instr);
-
- /*
- * In SSBO/UBO's, a true boolean value is any non-zero value, but we
- * consider a true boolean to be ~0. Fix this up with a != 0
- * comparison.
- */
- if (type->is_boolean())
- ret = nir_i2b(&b, &instr->dest.ssa);
- break;
- }
- case nir_intrinsic_ssbo_atomic_add:
- case nir_intrinsic_ssbo_atomic_imin:
- case nir_intrinsic_ssbo_atomic_umin:
- case nir_intrinsic_ssbo_atomic_imax:
- case nir_intrinsic_ssbo_atomic_umax:
- case nir_intrinsic_ssbo_atomic_and:
- case nir_intrinsic_ssbo_atomic_or:
- case nir_intrinsic_ssbo_atomic_xor:
- case nir_intrinsic_ssbo_atomic_exchange:
- case nir_intrinsic_ssbo_atomic_comp_swap:
- case nir_intrinsic_ssbo_atomic_fadd:
- case nir_intrinsic_ssbo_atomic_fmin:
- case nir_intrinsic_ssbo_atomic_fmax:
- case nir_intrinsic_ssbo_atomic_fcomp_swap: {
- int param_count = ir->actual_parameters.length();
- assert(param_count == 3 || param_count == 4);
-
- /* Block index */
- exec_node *param = ir->actual_parameters.get_head();
- ir_instruction *inst = (ir_instruction *) param;
- instr->src[0] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue()));
-
- /* Offset */
- param = param->get_next();
- inst = (ir_instruction *) param;
- instr->src[1] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue()));
-
- /* data1 parameter (this is always present) */
- param = param->get_next();
- inst = (ir_instruction *) param;
- instr->src[2] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue()));
-
- /* data2 parameter (only with atomic_comp_swap) */
- if (param_count == 4) {
- assert(op == nir_intrinsic_ssbo_atomic_comp_swap ||
- op == nir_intrinsic_ssbo_atomic_fcomp_swap);
- param = param->get_next();
- inst = (ir_instruction *) param;
- instr->src[3] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue()));
- }
-
- /* Atomic result */
- assert(ir->return_deref);
- nir_ssa_dest_init(&instr->instr, &instr->dest,
- ir->return_deref->type->vector_elements, 32, NULL);
- nir_builder_instr_insert(&b, &instr->instr);
- break;
- }
case nir_intrinsic_load_shared: {
exec_node *param = ir->actual_parameters.get_head();
ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue();
/* The value in shared memory is a 32-bit value */
if (type->is_boolean())
- ret = nir_i2b(&b, &instr->dest.ssa);
+ ret = nir_b2b1(&b, &instr->dest.ssa);
break;
}
case nir_intrinsic_store_shared: {
nir_ssa_def *nir_val = evaluate_rvalue(val);
/* The value in shared memory is a 32-bit value */
if (val->type->is_boolean())
- nir_val = nir_b2i32(&b, nir_val);
+ nir_val = nir_b2b32(&b, nir_val);
instr->src[0] = nir_src_for_ssa(nir_val);
instr->num_components = val->type->vector_elements;
nir_builder_instr_insert(&b, &instr->instr);
break;
}
+ case nir_intrinsic_is_helper_invocation: {
+ nir_ssa_dest_init(&instr->instr, &instr->dest, 1, 1, NULL);
+ instr->num_components = 1;
+ nir_builder_instr_insert(&b, &instr->instr);
+ break;
+ }
default:
unreachable("not reached");
}
{
/* Some special cases */
switch (ir->operation) {
- case ir_binop_ubo_load: {
- nir_intrinsic_instr *load =
- nir_intrinsic_instr_create(this->shader, nir_intrinsic_load_ubo);
- unsigned bit_size = ir->type->is_boolean() ? 32 :
- glsl_get_bit_size(ir->type);
- load->num_components = ir->type->vector_elements;
- load->src[0] = nir_src_for_ssa(evaluate_rvalue(ir->operands[0]));
- load->src[1] = nir_src_for_ssa(evaluate_rvalue(ir->operands[1]));
- intrinsic_set_std430_align(load, ir->type);
- add_instr(&load->instr, ir->type->vector_elements, bit_size);
-
- /*
- * In UBO's, a true boolean value is any non-zero value, but we consider
- * a true boolean to be ~0. Fix this up with a != 0 comparison.
- */
-
- if (ir->type->is_boolean())
- this->result = nir_i2b(&b, &load->dest.ssa);
-
- return;
- }
-
case ir_unop_interpolate_at_centroid:
case ir_binop_interpolate_at_offset:
case ir_binop_interpolate_at_sample: {
return;
}
+ case ir_binop_ubo_load:
+ /* UBO loads should only have been lowered in GLSL IR for non-nir drivers,
+ * NIR drivers make use of gl_nir_lower_buffers() instead.
+ */
+ unreachable("Invalid operation nir doesn't want lowered ubo loads");
default:
break;
}
result = type_is_float(types[0]) ? nir_fabs(&b, srcs[0])
: nir_iabs(&b, srcs[0]);
break;
+ case ir_unop_clz:
+ result = nir_uclz(&b, srcs[0]);
+ break;
case ir_unop_saturate:
assert(type_is_float(types[0]));
result = nir_fsat(&b, srcs[0]);
case ir_unop_b2i64:
case ir_unop_d2f:
case ir_unop_f2d:
+ case ir_unop_f162f:
+ case ir_unop_f2f16:
+ case ir_unop_f162b:
+ case ir_unop_b2f16:
case ir_unop_d2i:
case ir_unop_d2u:
case ir_unop_d2b:
break;
}
+ case ir_unop_f2fmp: {
+ result = nir_build_alu(&b, nir_op_f2fmp, srcs[0], NULL, NULL, NULL);
+ break;
+ }
+
case ir_unop_bitcast_i2f:
case ir_unop_bitcast_f2i:
case ir_unop_bitcast_u2f:
result = nir_find_lsb(&b, srcs[0]);
break;
- case ir_unop_noise:
- switch (ir->type->vector_elements) {
- case 1:
- switch (ir->operands[0]->type->vector_elements) {
- case 1: result = nir_fnoise1_1(&b, srcs[0]); break;
- case 2: result = nir_fnoise1_2(&b, srcs[0]); break;
- case 3: result = nir_fnoise1_3(&b, srcs[0]); break;
- case 4: result = nir_fnoise1_4(&b, srcs[0]); break;
- default: unreachable("not reached");
- }
- break;
- case 2:
- switch (ir->operands[0]->type->vector_elements) {
- case 1: result = nir_fnoise2_1(&b, srcs[0]); break;
- case 2: result = nir_fnoise2_2(&b, srcs[0]); break;
- case 3: result = nir_fnoise2_3(&b, srcs[0]); break;
- case 4: result = nir_fnoise2_4(&b, srcs[0]); break;
- default: unreachable("not reached");
- }
- break;
- case 3:
- switch (ir->operands[0]->type->vector_elements) {
- case 1: result = nir_fnoise3_1(&b, srcs[0]); break;
- case 2: result = nir_fnoise3_2(&b, srcs[0]); break;
- case 3: result = nir_fnoise3_3(&b, srcs[0]); break;
- case 4: result = nir_fnoise3_4(&b, srcs[0]); break;
- default: unreachable("not reached");
- }
- break;
- case 4:
- switch (ir->operands[0]->type->vector_elements) {
- case 1: result = nir_fnoise4_1(&b, srcs[0]); break;
- case 2: result = nir_fnoise4_2(&b, srcs[0]); break;
- case 3: result = nir_fnoise4_3(&b, srcs[0]); break;
- case 4: result = nir_fnoise4_4(&b, srcs[0]); break;
- default: unreachable("not reached");
- }
- break;
- default:
- unreachable("not reached");
- }
- break;
case ir_unop_get_buffer_size: {
nir_intrinsic_instr *load = nir_intrinsic_instr_create(
this->shader,
return;
}
+ case ir_unop_atan:
+ result = nir_atan(&b, srcs[0]);
+ break;
+
case ir_binop_add:
result = type_is_float(out_type) ? nir_fadd(&b, srcs[0], srcs[1])
: nir_iadd(&b, srcs[0], srcs[1]);
break;
+ case ir_binop_add_sat:
+ result = type_is_signed(out_type) ? nir_iadd_sat(&b, srcs[0], srcs[1])
+ : nir_uadd_sat(&b, srcs[0], srcs[1]);
+ break;
case ir_binop_sub:
result = type_is_float(out_type) ? nir_fsub(&b, srcs[0], srcs[1])
: nir_isub(&b, srcs[0], srcs[1]);
break;
+ case ir_binop_sub_sat:
+ result = type_is_signed(out_type) ? nir_isub_sat(&b, srcs[0], srcs[1])
+ : nir_usub_sat(&b, srcs[0], srcs[1]);
+ break;
+ case ir_binop_abs_sub:
+ /* out_type is always unsigned for ir_binop_abs_sub, so we have to key
+ * on the type of the sources.
+ */
+ result = type_is_signed(types[0]) ? nir_uabs_isub(&b, srcs[0], srcs[1])
+ : nir_uabs_usub(&b, srcs[0], srcs[1]);
+ break;
+ case ir_binop_avg:
+ result = type_is_signed(out_type) ? nir_ihadd(&b, srcs[0], srcs[1])
+ : nir_uhadd(&b, srcs[0], srcs[1]);
+ break;
+ case ir_binop_avg_round:
+ result = type_is_signed(out_type) ? nir_irhadd(&b, srcs[0], srcs[1])
+ : nir_urhadd(&b, srcs[0], srcs[1]);
+ break;
+ case ir_binop_mul_32x16:
+ result = type_is_signed(out_type) ? nir_imul_32x16(&b, srcs[0], srcs[1])
+ : nir_umul_32x16(&b, srcs[0], srcs[1]);
+ break;
case ir_binop_mul:
if (type_is_float(out_type))
result = nir_fmul(&b, srcs[0], srcs[1]);
break;
}
+ case ir_binop_atan2:
+ result = nir_atan2(&b, srcs[0], srcs[1]);
+ break;
+
case ir_binop_ldexp: result = nir_ldexp(&b, srcs[0], srcs[1]); break;
case ir_triop_fma:
result = nir_ffma(&b, srcs[0], srcs[1], srcs[2]);
case GLSL_TYPE_FLOAT:
instr->dest_type = nir_type_float;
break;
+ case GLSL_TYPE_FLOAT16:
+ instr->dest_type = nir_type_float16;
+ break;
case GLSL_TYPE_INT:
instr->dest_type = nir_type_int;
break;
void
nir_visitor::visit(ir_barrier *)
{
+ if (shader->info.stage == MESA_SHADER_COMPUTE) {
+ nir_intrinsic_instr *shared_barrier =
+ nir_intrinsic_instr_create(this->shader,
+ nir_intrinsic_memory_barrier_shared);
+ nir_builder_instr_insert(&b, &shared_barrier->instr);
+ } else if (shader->info.stage == MESA_SHADER_TESS_CTRL) {
+ nir_intrinsic_instr *patch_barrier =
+ nir_intrinsic_instr_create(this->shader,
+ nir_intrinsic_memory_barrier_tcs_patch);
+ nir_builder_instr_insert(&b, &patch_barrier->instr);
+ }
+
nir_intrinsic_instr *instr =
- nir_intrinsic_instr_create(this->shader, nir_intrinsic_barrier);
+ nir_intrinsic_instr_create(this->shader, nir_intrinsic_control_barrier);
nir_builder_instr_insert(&b, &instr->instr);
}
nir_validate_shader(nir, "float64_funcs_to_nir");
- NIR_PASS_V(nir, nir_lower_constant_initializers, nir_var_function_temp);
+ NIR_PASS_V(nir, nir_lower_variable_initializers, nir_var_function_temp);
NIR_PASS_V(nir, nir_lower_returns);
NIR_PASS_V(nir, nir_inline_functions);
NIR_PASS_V(nir, nir_opt_deref);
+ /* Do some optimizations to clean up the shader now. By optimizing the
+ * functions in the library, we avoid having to re-do that work every
+ * time we inline a copy of a function. Reducing basic blocks also helps
+ * with compile times.
+ */
+ NIR_PASS_V(nir, nir_lower_vars_to_ssa);
+ NIR_PASS_V(nir, nir_copy_prop);
+ NIR_PASS_V(nir, nir_opt_dce);
+ NIR_PASS_V(nir, nir_opt_cse);
+ NIR_PASS_V(nir, nir_opt_gcm, true);
+ NIR_PASS_V(nir, nir_opt_peephole_select, 1, false, false);
+ NIR_PASS_V(nir, nir_opt_dce);
+
return nir;
}