#include "brw_vec4.h"
#include "brw_vec4_builder.h"
#include "brw_vec4_surface_builder.h"
-#include "glsl/ir_uniform.h"
+#include "brw_program.h"
using namespace brw;
using namespace brw::surface_access;
void
vec4_visitor::emit_nir_code()
{
- if (nir->num_inputs > 0)
- nir_setup_inputs();
-
if (nir->num_uniforms > 0)
nir_setup_uniforms();
nir_setup_system_values();
/* get the main function and emit it */
- nir_foreach_overload(nir, overload) {
- assert(strcmp(overload->function->name, "main") == 0);
- assert(overload->impl);
- nir_emit_impl(overload->impl);
+ nir_foreach_function(nir, function) {
+ assert(strcmp(function->name, "main") == 0);
+ assert(function->impl);
+ nir_emit_impl(function->impl);
}
}
glsl_type::int_type);
break;
+ case nir_intrinsic_load_base_instance:
+ reg = &nir_system_values[SYSTEM_VALUE_BASE_INSTANCE];
+ if (reg->file == BAD_FILE)
+ *reg = *make_reg_for_system_value(SYSTEM_VALUE_BASE_INSTANCE,
+ glsl_type::int_type);
+ break;
+
+ case nir_intrinsic_load_draw_id:
+ reg = &nir_system_values[SYSTEM_VALUE_DRAW_ID];
+ if (reg->file == BAD_FILE)
+ *reg = *make_reg_for_system_value(SYSTEM_VALUE_DRAW_ID,
+ glsl_type::int_type);
+ break;
+
default:
break;
}
vec4_visitor::nir_setup_system_values()
{
nir_system_values = ralloc_array(mem_ctx, dst_reg, SYSTEM_VALUE_MAX);
-
- nir_foreach_overload(nir, overload) {
- assert(strcmp(overload->function->name, "main") == 0);
- assert(overload->impl);
- nir_foreach_block(overload->impl, setup_system_values_block, this);
+ for (unsigned i = 0; i < SYSTEM_VALUE_MAX; i++) {
+ nir_system_values[i] = dst_reg();
}
-}
-void
-vec4_visitor::nir_setup_inputs()
-{
- nir_inputs = ralloc_array(mem_ctx, src_reg, nir->num_inputs);
-
- nir_foreach_variable(var, &nir->inputs) {
- int offset = var->data.driver_location;
- unsigned size = type_size_vec4(var->type);
- for (unsigned i = 0; i < size; i++) {
- src_reg src = src_reg(ATTR, var->data.location + i, var->type);
- nir_inputs[offset + i] = src;
- }
+ nir_foreach_function(nir, function) {
+ assert(strcmp(function->name, "main") == 0);
+ assert(function->impl);
+ nir_foreach_block(function->impl, setup_system_values_block, this);
}
}
void
vec4_visitor::nir_setup_uniforms()
{
- uniforms = nir->num_uniforms;
-
- nir_foreach_variable(var, &nir->uniforms) {
- /* UBO's and atomics don't take up space in the uniform file */
- if (var->interface_type != NULL || var->type->contains_atomic())
- continue;
-
- if (type_size_vec4(var->type) > 0)
- uniform_size[var->data.driver_location] = type_size_vec4(var->type);
- }
+ uniforms = nir->num_uniforms / 16;
}
void
vec4_visitor::nir_emit_impl(nir_function_impl *impl)
{
nir_locals = ralloc_array(mem_ctx, dst_reg, impl->reg_alloc);
+ for (unsigned i = 0; i < impl->reg_alloc; i++) {
+ nir_locals[i] = dst_reg();
+ }
foreach_list_typed(nir_register, reg, node, &impl->registers) {
unsigned array_elems =
reg->num_array_elems == 0 ? 1 : reg->num_array_elems;
- nir_locals[reg->index] = dst_reg(GRF, alloc.allocate(array_elems));
+ nir_locals[reg->index] = dst_reg(VGRF, alloc.allocate(array_elems));
}
nir_ssa_values = ralloc_array(mem_ctx, dst_reg, impl->ssa_alloc);
vec4_visitor::get_nir_dest(nir_dest dest)
{
if (dest.is_ssa) {
- dst_reg dst = dst_reg(GRF, alloc.allocate(1));
+ dst_reg dst = dst_reg(VGRF, alloc.allocate(1));
nir_ssa_values[dest.ssa.index] = dst;
return dst;
} else {
return get_nir_src(src, nir_type_int, num_components);
}
+src_reg
+vec4_visitor::get_indirect_offset(nir_intrinsic_instr *instr)
+{
+ nir_src *offset_src = nir_get_io_offset_src(instr);
+ nir_const_value *const_value = nir_src_as_const_value(*offset_src);
+
+ if (const_value) {
+ /* The only constant offset we should find is 0. brw_nir.c's
+ * add_const_offset_to_base() will fold other constant offsets
+ * into instr->const_index[0].
+ */
+ assert(const_value->u32[0] == 0);
+ return src_reg();
+ }
+
+ return get_nir_src(*offset_src, BRW_REGISTER_TYPE_UD, 1);
+}
+
void
vec4_visitor::nir_emit_load_const(nir_load_const_instr *instr)
{
- dst_reg reg = dst_reg(GRF, alloc.allocate(1));
+ dst_reg reg = dst_reg(VGRF, alloc.allocate(1));
reg.type = BRW_REGISTER_TYPE_D;
unsigned remaining = brw_writemask_for_size(instr->def.num_components);
continue;
for (unsigned j = i; j < instr->def.num_components; j++) {
- if (instr->value.u[i] == instr->value.u[j]) {
+ if (instr->value.u32[i] == instr->value.u32[j]) {
writemask |= 1 << j;
}
}
reg.writemask = writemask;
- emit(MOV(reg, src_reg(instr->value.i[i])));
+ emit(MOV(reg, brw_imm_d(instr->value.i32[i])));
remaining &= ~writemask;
}
dst_reg dest;
src_reg src;
- bool has_indirect = false;
-
switch (instr->intrinsic) {
- case nir_intrinsic_load_input_indirect:
- has_indirect = true;
- /* fallthrough */
case nir_intrinsic_load_input: {
- int offset = instr->const_index[0];
- src = nir_inputs[offset];
+ nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
+
+ /* We set EmitNoIndirectInput for VS */
+ assert(const_offset);
+
+ src = src_reg(ATTR, instr->const_index[0] + const_offset->u32[0],
+ glsl_type::uvec4_type);
- if (has_indirect) {
- dest.reladdr = new(mem_ctx) src_reg(get_nir_src(instr->src[0],
- BRW_REGISTER_TYPE_D,
- 1));
- }
dest = get_nir_dest(instr->dest, src.type);
dest.writemask = brw_writemask_for_size(instr->num_components);
break;
}
- case nir_intrinsic_store_output_indirect:
- has_indirect = true;
- /* fallthrough */
case nir_intrinsic_store_output: {
- int varying = instr->const_index[0];
+ nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
+ assert(const_offset);
+
+ int varying = instr->const_index[0] + const_offset->u32[0];
src = get_nir_src(instr->src[0], BRW_REGISTER_TYPE_F,
instr->num_components);
- dest = dst_reg(src);
- if (has_indirect) {
- dest.reladdr = new(mem_ctx) src_reg(get_nir_src(instr->src[1],
- BRW_REGISTER_TYPE_D,
- 1));
- }
- output_reg[varying] = dest;
+ output_reg[varying] = dst_reg(src);
break;
}
case nir_intrinsic_get_buffer_size: {
nir_const_value *const_uniform_block = nir_src_as_const_value(instr->src[0]);
- unsigned ssbo_index = const_uniform_block ? const_uniform_block->u[0] : 0;
+ unsigned ssbo_index = const_uniform_block ? const_uniform_block->u32[0] : 0;
const unsigned index =
prog_data->base.binding_table.ssbo_start + ssbo_index;
inst->base_mrf = 2;
inst->mlen = 1; /* always at least one */
- inst->src[1] = src_reg(index);
+ inst->src[1] = brw_imm_ud(index);
/* MRF for the first parameter */
- src_reg lod = src_reg(0);
+ src_reg lod = brw_imm_d(0);
int param_base = inst->base_mrf;
int writemask = WRITEMASK_X;
emit(MOV(dst_reg(MRF, param_base, glsl_type::int_type, writemask), lod));
break;
}
- case nir_intrinsic_store_ssbo_indirect:
- has_indirect = true;
- /* fallthrough */
case nir_intrinsic_store_ssbo: {
assert(devinfo->gen >= 7);
nir_src_as_const_value(instr->src[1]);
if (const_uniform_block) {
unsigned index = prog_data->base.binding_table.ssbo_start +
- const_uniform_block->u[0];
- surf_index = src_reg(index);
+ const_uniform_block->u32[0];
+ surf_index = brw_imm_ud(index);
brw_mark_surface_used(&prog_data->base, index);
} else {
surf_index = src_reg(this, glsl_type::uint_type);
emit(ADD(dst_reg(surf_index), get_nir_src(instr->src[1], 1),
- src_reg(prog_data->base.binding_table.ssbo_start)));
+ brw_imm_ud(prog_data->base.binding_table.ssbo_start)));
surf_index = emit_uniformize(surf_index);
brw_mark_surface_used(&prog_data->base,
}
/* Offset */
- src_reg offset_reg = src_reg(this, glsl_type::uint_type);
- unsigned const_offset_bytes = 0;
- if (has_indirect) {
- emit(MOV(dst_reg(offset_reg), get_nir_src(instr->src[2], 1)));
+ src_reg offset_reg;
+ nir_const_value *const_offset = nir_src_as_const_value(instr->src[2]);
+ if (const_offset) {
+ offset_reg = brw_imm_ud(const_offset->u32[0]);
} else {
- const_offset_bytes = instr->const_index[0];
- emit(MOV(dst_reg(offset_reg), src_reg(const_offset_bytes)));
+ offset_reg = get_nir_src(instr->src[2], 1);
}
/* Value */
src_reg val_reg = get_nir_src(instr->src[0], 4);
/* Writemask */
- unsigned write_mask = instr->const_index[1];
+ unsigned write_mask = instr->const_index[0];
/* IvyBridge does not have a native SIMD4x2 untyped write message so untyped
* writes will use SIMD8 mode. In order to hide this and keep symmetry across
* write at to skip the channels we skipped, if any.
*/
if (skipped_channels > 0) {
- if (!has_indirect) {
- const_offset_bytes += 4 * skipped_channels;
- offset_reg = src_reg(const_offset_bytes);
+ if (offset_reg.file == IMM) {
+ offset_reg.ud += 4 * skipped_channels;
} else {
emit(ADD(dst_reg(offset_reg), offset_reg,
brw_imm_ud(4 * skipped_channels)));
break;
}
- case nir_intrinsic_load_ssbo_indirect:
- has_indirect = true;
- /* fallthrough */
case nir_intrinsic_load_ssbo: {
assert(devinfo->gen >= 7);
src_reg surf_index;
if (const_uniform_block) {
unsigned index = prog_data->base.binding_table.ssbo_start +
- const_uniform_block->u[0];
- surf_index = src_reg(index);
+ const_uniform_block->u32[0];
+ surf_index = brw_imm_ud(index);
brw_mark_surface_used(&prog_data->base, index);
} else {
surf_index = src_reg(this, glsl_type::uint_type);
emit(ADD(dst_reg(surf_index), get_nir_src(instr->src[0], 1),
- src_reg(prog_data->base.binding_table.ssbo_start)));
+ brw_imm_ud(prog_data->base.binding_table.ssbo_start)));
surf_index = emit_uniformize(surf_index);
/* Assume this may touch any UBO. It would be nice to provide
nir->info.num_ssbos - 1);
}
- src_reg offset_reg = src_reg(this, glsl_type::uint_type);
- unsigned const_offset_bytes = 0;
- if (has_indirect) {
- emit(MOV(dst_reg(offset_reg), get_nir_src(instr->src[1], 1)));
+ src_reg offset_reg;
+ nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
+ if (const_offset) {
+ offset_reg = brw_imm_ud(const_offset->u32[0]);
} else {
- const_offset_bytes = instr->const_index[0];
- emit(MOV(dst_reg(offset_reg), src_reg(const_offset_bytes)));
+ offset_reg = get_nir_src(instr->src[1], 1);
}
/* Read the vector */
case nir_intrinsic_load_vertex_id_zero_base:
case nir_intrinsic_load_base_vertex:
- case nir_intrinsic_load_instance_id: {
+ case nir_intrinsic_load_instance_id:
+ case nir_intrinsic_load_base_instance:
+ case nir_intrinsic_load_draw_id:
+ case nir_intrinsic_load_invocation_id: {
gl_system_value sv = nir_system_value_from_intrinsic(instr->intrinsic);
src_reg val = src_reg(nir_system_values[sv]);
assert(val.file != BAD_FILE);
break;
}
- case nir_intrinsic_load_uniform_indirect:
- has_indirect = true;
- /* fallthrough */
case nir_intrinsic_load_uniform: {
+ /* Offsets are in bytes but they should always be multiples of 16 */
+ assert(instr->const_index[0] % 16 == 0);
+
dest = get_nir_dest(instr->dest);
- src = src_reg(dst_reg(UNIFORM, instr->const_index[0]));
- src.reg_offset = instr->const_index[1];
+ src = src_reg(dst_reg(UNIFORM, instr->const_index[0] / 16));
+ src.type = dest.type;
- if (has_indirect) {
- src_reg tmp = get_nir_src(instr->src[0], BRW_REGISTER_TYPE_D, 1);
- src.reladdr = new(mem_ctx) src_reg(tmp);
- }
+ nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
+ if (const_offset) {
+ /* Offsets are in bytes but they should always be multiples of 16 */
+ assert(const_offset->u32[0] % 16 == 0);
+ src.reg_offset = const_offset->u32[0] / 16;
- emit(MOV(dest, src));
+ emit(MOV(dest, src));
+ } else {
+ src_reg indirect = get_nir_src(instr->src[0], BRW_REGISTER_TYPE_UD, 1);
+
+ emit(SHADER_OPCODE_MOV_INDIRECT, dest, src,
+ indirect, brw_imm_ud(instr->const_index[1]));
+ }
break;
}
(unsigned) instr->const_index[0];
src_reg offset = get_nir_src(instr->src[0], nir_type_int,
instr->num_components);
+ const src_reg surface = brw_imm_ud(surf_index);
+ const vec4_builder bld =
+ vec4_builder(this).at_end().annotate(current_annotation, base_ir);
+ src_reg tmp;
+
dest = get_nir_dest(instr->dest);
switch (instr->intrinsic) {
- case nir_intrinsic_atomic_counter_inc:
- emit_untyped_atomic(BRW_AOP_INC, surf_index, dest, offset,
- src_reg(), src_reg());
- break;
- case nir_intrinsic_atomic_counter_dec:
- emit_untyped_atomic(BRW_AOP_PREDEC, surf_index, dest, offset,
- src_reg(), src_reg());
- break;
- case nir_intrinsic_atomic_counter_read:
- emit_untyped_surface_read(surf_index, dest, offset);
- break;
- default:
- unreachable("Unreachable");
+ case nir_intrinsic_atomic_counter_inc:
+ tmp = emit_untyped_atomic(bld, surface, offset,
+ src_reg(), src_reg(),
+ 1, 1,
+ BRW_AOP_INC);
+ break;
+ case nir_intrinsic_atomic_counter_dec:
+ tmp = emit_untyped_atomic(bld, surface, offset,
+ src_reg(), src_reg(),
+ 1, 1,
+ BRW_AOP_PREDEC);
+ break;
+ case nir_intrinsic_atomic_counter_read:
+ tmp = emit_untyped_read(bld, surface, offset, 1, 1);
+ break;
+ default:
+ unreachable("Unreachable");
}
+ bld.MOV(retype(dest, tmp.type), tmp);
brw_mark_surface_used(stage_prog_data, surf_index);
break;
}
- case nir_intrinsic_load_ubo_indirect:
- has_indirect = true;
- /* fallthrough */
case nir_intrinsic_load_ubo: {
nir_const_value *const_block_index = nir_src_as_const_value(instr->src[0]);
src_reg surf_index;
* as an immediate.
*/
const unsigned index = prog_data->base.binding_table.ubo_start +
- const_block_index->u[0];
- surf_index = src_reg(index);
+ const_block_index->u32[0];
+ surf_index = brw_imm_ud(index);
brw_mark_surface_used(&prog_data->base, index);
} else {
/* The block index is not a constant. Evaluate the index expression
surf_index = src_reg(this, glsl_type::uint_type);
emit(ADD(dst_reg(surf_index), get_nir_src(instr->src[0], nir_type_int,
instr->num_components),
- src_reg(prog_data->base.binding_table.ubo_start)));
+ brw_imm_ud(prog_data->base.binding_table.ubo_start)));
surf_index = emit_uniformize(surf_index);
/* Assume this may touch any UBO. It would be nice to provide
nir->info.num_ubos - 1);
}
- unsigned const_offset = instr->const_index[0];
src_reg offset;
-
- if (!has_indirect) {
- offset = src_reg(const_offset / 16);
+ nir_const_value *const_offset = nir_src_as_const_value(instr->src[1]);
+ if (const_offset) {
+ offset = brw_imm_ud(const_offset->u32[0] & ~15);
} else {
- offset = src_reg(this, glsl_type::uint_type);
- emit(SHR(dst_reg(offset), get_nir_src(instr->src[1], nir_type_int, 1),
- src_reg(4u)));
+ offset = get_nir_src(instr->src[1], nir_type_int, 1);
}
src_reg packed_consts = src_reg(this, glsl_type::vec4_type);
NULL, NULL /* before_block/inst */);
packed_consts.swizzle = brw_swizzle_for_size(instr->num_components);
- packed_consts.swizzle += BRW_SWIZZLE4(const_offset % 16 / 4,
- const_offset % 16 / 4,
- const_offset % 16 / 4,
- const_offset % 16 / 4);
+ if (const_offset) {
+ packed_consts.swizzle += BRW_SWIZZLE4(const_offset->u32[0] % 16 / 4,
+ const_offset->u32[0] % 16 / 4,
+ const_offset->u32[0] % 16 / 4,
+ const_offset->u32[0] % 16 / 4);
+ }
emit(MOV(dest, packed_consts));
break;
nir_const_value *const_surface = nir_src_as_const_value(instr->src[0]);
if (const_surface) {
unsigned surf_index = prog_data->base.binding_table.ssbo_start +
- const_surface->u[0];
- surface = src_reg(surf_index);
+ const_surface->u32[0];
+ surface = brw_imm_ud(surf_index);
brw_mark_surface_used(&prog_data->base, surf_index);
} else {
surface = src_reg(this, glsl_type::uint_type);
emit(ADD(dst_reg(surface), get_nir_src(instr->src[0]),
- src_reg(prog_data->base.binding_table.ssbo_start)));
+ brw_imm_ud(prog_data->base.binding_table.ssbo_start)));
/* Assume this may touch any UBO. This is the same we do for other
* UBO/SSBO accesses with non-constant surface.
const vec4_builder bld =
vec4_builder(this).at_end().annotate(current_annotation, base_ir);
- src_reg atomic_result =
- surface_access::emit_untyped_atomic(bld, surface, offset,
- data1, data2,
- 1 /* dims */, 1 /* rsize */,
- op,
- BRW_PREDICATE_NONE);
+ src_reg atomic_result = emit_untyped_atomic(bld, surface, offset,
+ data1, data2,
+ 1 /* dims */, 1 /* rsize */,
+ op,
+ BRW_PREDICATE_NONE);
dest.type = atomic_result.type;
bld.MOV(dest, atomic_result);
}
}
}
+bool
+vec4_visitor::optimize_predicate(nir_alu_instr *instr,
+ enum brw_predicate *predicate)
+{
+ if (!instr->src[0].src.is_ssa ||
+ instr->src[0].src.ssa->parent_instr->type != nir_instr_type_alu)
+ return false;
+
+ nir_alu_instr *cmp_instr =
+ nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
+
+ switch (cmp_instr->op) {
+ case nir_op_bany_fnequal2:
+ case nir_op_bany_inequal2:
+ case nir_op_bany_fnequal3:
+ case nir_op_bany_inequal3:
+ case nir_op_bany_fnequal4:
+ case nir_op_bany_inequal4:
+ *predicate = BRW_PREDICATE_ALIGN16_ANY4H;
+ break;
+ case nir_op_ball_fequal2:
+ case nir_op_ball_iequal2:
+ case nir_op_ball_fequal3:
+ case nir_op_ball_iequal3:
+ case nir_op_ball_fequal4:
+ case nir_op_ball_iequal4:
+ *predicate = BRW_PREDICATE_ALIGN16_ALL4H;
+ break;
+ default:
+ return false;
+ }
+
+ unsigned size_swizzle =
+ brw_swizzle_for_size(nir_op_infos[cmp_instr->op].input_sizes[0]);
+
+ src_reg op[2];
+ assert(nir_op_infos[cmp_instr->op].num_inputs == 2);
+ for (unsigned i = 0; i < 2; i++) {
+ op[i] = get_nir_src(cmp_instr->src[i].src,
+ nir_op_infos[cmp_instr->op].input_types[i], 4);
+ unsigned base_swizzle =
+ brw_swizzle_for_nir_swizzle(cmp_instr->src[i].swizzle);
+ op[i].swizzle = brw_compose_swizzle(size_swizzle, base_swizzle);
+ op[i].abs = cmp_instr->src[i].abs;
+ op[i].negate = cmp_instr->src[i].negate;
+ }
+
+ emit(CMP(dst_null_d(), op[0], op[1],
+ brw_conditional_for_nir_comparison(cmp_instr->op)));
+
+ return true;
+}
+
void
vec4_visitor::nir_emit_alu(nir_alu_instr *instr)
{
* operand. If we can determine that one of the args is in the low
* 16 bits, though, we can just emit a single MUL.
*/
- if (value0 && value0->u[0] < (1 << 16)) {
+ if (value0 && value0->u32[0] < (1 << 16)) {
if (devinfo->gen < 7)
emit(MUL(dst, op[0], op[1]));
else
emit(MUL(dst, op[1], op[0]));
- } else if (value1 && value1->u[0] < (1 << 16)) {
+ } else if (value1 && value1->u32[0] < (1 << 16)) {
if (devinfo->gen < 7)
emit(MUL(dst, op[1], op[0]));
else
case nir_op_umul_high: {
struct brw_reg acc = retype(brw_acc_reg(8), dst.type);
- emit(MUL(acc, op[0], op[1]));
+ if (devinfo->gen >= 8)
+ emit(MUL(acc, op[0], retype(op[1], BRW_REGISTER_TYPE_UW)));
+ else
+ emit(MUL(acc, op[0], op[1]));
+
emit(MACH(dst, op[0], op[1]));
break;
}
break;
case nir_op_umod:
+ case nir_op_irem:
+ /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
+ * appears that our hardware just does the right thing for signed
+ * remainder.
+ */
emit_math(SHADER_OPCODE_INT_REMAINDER, dst, op[0], op[1]);
break;
+ case nir_op_imod: {
+ /* Get a regular C-style remainder. If a % b == 0, set the predicate. */
+ inst = emit_math(SHADER_OPCODE_INT_REMAINDER, dst, op[0], op[1]);
+
+ /* Math instructions don't support conditional mod */
+ inst = emit(MOV(dst_null_d(), src_reg(dst)));
+ inst->conditional_mod = BRW_CONDITIONAL_NZ;
+
+ /* Now, we need to determine if signs of the sources are different.
+ * When we XOR the sources, the top bit is 0 if they are the same and 1
+ * if they are different. We can then use a conditional modifier to
+ * turn that into a predicate. This leads us to an XOR.l instruction.
+ *
+ * Technically, according to the PRM, you're not allowed to use .l on a
+ * XOR instruction. However, emperical experiments and Curro's reading
+ * of the simulator source both indicate that it's safe.
+ */
+ src_reg tmp = src_reg(this, glsl_type::ivec4_type);
+ inst = emit(XOR(dst_reg(tmp), op[0], op[1]));
+ inst->predicate = BRW_PREDICATE_NORMAL;
+ inst->conditional_mod = BRW_CONDITIONAL_L;
+
+ /* If the result of the initial remainder operation is non-zero and the
+ * two sources have different signs, add in a copy of op[1] to get the
+ * final integer modulus value.
+ */
+ inst = emit(ADD(dst, src_reg(dst), op[1]));
+ inst->predicate = BRW_PREDICATE_NORMAL;
+ break;
+ }
+
case nir_op_ldexp:
unreachable("not reached: should be handled by ldexp_to_arith()");
inst->saturate = instr->dest.saturate;
break;
+ case nir_op_fquantize2f16: {
+ /* See also vec4_visitor::emit_pack_half_2x16() */
+ src_reg tmp16 = src_reg(this, glsl_type::uvec4_type);
+ src_reg tmp32 = src_reg(this, glsl_type::vec4_type);
+ src_reg zero = src_reg(this, glsl_type::vec4_type);
+
+ /* Check for denormal */
+ src_reg abs_src0 = op[0];
+ abs_src0.abs = true;
+ emit(CMP(dst_null_f(), abs_src0, brw_imm_f(ldexpf(1.0, -14)),
+ BRW_CONDITIONAL_L));
+ /* Get the appropriately signed zero */
+ emit(AND(retype(dst_reg(zero), BRW_REGISTER_TYPE_UD),
+ retype(op[0], BRW_REGISTER_TYPE_UD),
+ brw_imm_ud(0x80000000)));
+ /* Do the actual F32 -> F16 -> F32 conversion */
+ emit(F32TO16(dst_reg(tmp16), op[0]));
+ emit(F16TO32(dst_reg(tmp32), tmp16));
+ /* Select that or zero based on normal status */
+ inst = emit(BRW_OPCODE_SEL, dst, zero, tmp32);
+ inst->predicate = BRW_PREDICATE_NORMAL;
+ inst->saturate = instr->dest.saturate;
+ break;
+ }
+
case nir_op_fmin:
case nir_op_imin:
case nir_op_umin:
emit(CMP(dst_null_d(), swizzle(op[0], swiz), swizzle(op[1], swiz),
brw_conditional_for_nir_comparison(instr->op)));
- emit(MOV(dst, src_reg(0)));
- inst = emit(MOV(dst, src_reg(~0)));
+ emit(MOV(dst, brw_imm_d(0)));
+ inst = emit(MOV(dst, brw_imm_d(~0)));
inst->predicate = BRW_PREDICATE_ALIGN16_ALL4H;
break;
}
emit(CMP(dst_null_d(), swizzle(op[0], swiz), swizzle(op[1], swiz),
brw_conditional_for_nir_comparison(instr->op)));
- emit(MOV(dst, src_reg(0)));
- inst = emit(MOV(dst, src_reg(~0)));
+ emit(MOV(dst, brw_imm_d(0)));
+ inst = emit(MOV(dst, brw_imm_d(~0)));
inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H;
break;
}
break;
case nir_op_f2b:
- emit(CMP(dst, op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ));
+ emit(CMP(dst, op[0], brw_imm_f(0.0f), BRW_CONDITIONAL_NZ));
break;
case nir_op_i2b:
- emit(CMP(dst, op[0], src_reg(0), BRW_CONDITIONAL_NZ));
+ emit(CMP(dst, op[0], brw_imm_d(0), BRW_CONDITIONAL_NZ));
break;
case nir_op_fnoise1_1:
case nir_op_pack_unorm_2x16:
unreachable("not reached: should be handled by lower_packing_builtins");
+ case nir_op_pack_uvec4_to_uint:
+ unreachable("not reached");
+
+ case nir_op_pack_uvec2_to_uint: {
+ dst_reg tmp1 = dst_reg(this, glsl_type::uint_type);
+ tmp1.writemask = WRITEMASK_X;
+ op[0].swizzle = BRW_SWIZZLE_YYYY;
+ emit(SHL(tmp1, op[0], src_reg(brw_imm_ud(16u))));
+
+ dst_reg tmp2 = dst_reg(this, glsl_type::uint_type);
+ tmp2.writemask = WRITEMASK_X;
+ op[0].swizzle = BRW_SWIZZLE_XXXX;
+ emit(AND(tmp2, op[0], src_reg(brw_imm_ud(0xffffu))));
+
+ emit(OR(dst, src_reg(tmp1), src_reg(tmp2)));
+ break;
+ }
+
case nir_op_unpack_half_2x16:
/* As NIR does not guarantee that we have a correct swizzle outside the
* boundaries of a vector, and the implementation of emit_unpack_half_2x16
* subtract the result from 31 to convert the MSB count into an LSB count.
*/
src_reg src(dst);
- emit(CMP(dst_null_d(), src, src_reg(-1), BRW_CONDITIONAL_NZ));
+ emit(CMP(dst_null_d(), src, brw_imm_d(-1), BRW_CONDITIONAL_NZ));
- inst = emit(ADD(dst, src, src_reg(31)));
+ inst = emit(ADD(dst, src, brw_imm_d(31)));
inst->predicate = BRW_PREDICATE_NORMAL;
inst->src[0].negate = true;
break;
case nir_op_ubitfield_extract:
case nir_op_ibitfield_extract:
+ unreachable("should have been lowered");
+ case nir_op_ubfe:
+ case nir_op_ibfe:
op[0] = fix_3src_operand(op[0]);
op[1] = fix_3src_operand(op[1]);
op[2] = fix_3src_operand(op[2]);
break;
case nir_op_bitfield_insert:
- unreachable("not reached: should be handled by "
- "lower_instructions::bitfield_insert_to_bfm_bfi");
+ unreachable("not reached: should have been lowered");
case nir_op_fsign:
/* AND(val, 0x80000000) gives the sign bit.
* Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
* zero.
*/
- emit(CMP(dst_null_f(), op[0], src_reg(0.0f), BRW_CONDITIONAL_NZ));
+ emit(CMP(dst_null_f(), op[0], brw_imm_f(0.0f), BRW_CONDITIONAL_NZ));
op[0].type = BRW_REGISTER_TYPE_UD;
dst.type = BRW_REGISTER_TYPE_UD;
- emit(AND(dst, op[0], src_reg(0x80000000u)));
+ emit(AND(dst, op[0], brw_imm_ud(0x80000000u)));
- inst = emit(OR(dst, src_reg(dst), src_reg(0x3f800000u)));
+ inst = emit(OR(dst, src_reg(dst), brw_imm_ud(0x3f800000u)));
inst->predicate = BRW_PREDICATE_NORMAL;
dst.type = BRW_REGISTER_TYPE_F;
* -> non-negative val generates 0x00000000.
* Predicated OR sets 1 if val is positive.
*/
- emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_G));
- emit(ASR(dst, op[0], src_reg(31)));
- inst = emit(OR(dst, src_reg(dst), src_reg(1)));
+ emit(CMP(dst_null_d(), op[0], brw_imm_d(0), BRW_CONDITIONAL_G));
+ emit(ASR(dst, op[0], brw_imm_d(31)));
+ inst = emit(OR(dst, src_reg(dst), brw_imm_d(1)));
inst->predicate = BRW_PREDICATE_NORMAL;
break;
break;
case nir_op_bcsel:
- emit(CMP(dst_null_d(), op[0], src_reg(0), BRW_CONDITIONAL_NZ));
- inst = emit(BRW_OPCODE_SEL, dst, op[1], op[2]);
- switch (dst.writemask) {
- case WRITEMASK_X:
- inst->predicate = BRW_PREDICATE_ALIGN16_REPLICATE_X;
- break;
- case WRITEMASK_Y:
- inst->predicate = BRW_PREDICATE_ALIGN16_REPLICATE_Y;
- break;
- case WRITEMASK_Z:
- inst->predicate = BRW_PREDICATE_ALIGN16_REPLICATE_Z;
- break;
- case WRITEMASK_W:
- inst->predicate = BRW_PREDICATE_ALIGN16_REPLICATE_W;
- break;
- default:
- inst->predicate = BRW_PREDICATE_NORMAL;
- break;
+ enum brw_predicate predicate;
+ if (!optimize_predicate(instr, &predicate)) {
+ emit(CMP(dst_null_d(), op[0], brw_imm_d(0), BRW_CONDITIONAL_NZ));
+ switch (dst.writemask) {
+ case WRITEMASK_X:
+ predicate = BRW_PREDICATE_ALIGN16_REPLICATE_X;
+ break;
+ case WRITEMASK_Y:
+ predicate = BRW_PREDICATE_ALIGN16_REPLICATE_Y;
+ break;
+ case WRITEMASK_Z:
+ predicate = BRW_PREDICATE_ALIGN16_REPLICATE_Z;
+ break;
+ case WRITEMASK_W:
+ predicate = BRW_PREDICATE_ALIGN16_REPLICATE_W;
+ break;
+ default:
+ predicate = BRW_PREDICATE_NORMAL;
+ break;
+ }
}
+ inst = emit(BRW_OPCODE_SEL, dst, op[1], op[2]);
+ inst->predicate = predicate;
break;
case nir_op_fdot_replicated2:
inst->saturate = instr->dest.saturate;
break;
- case nir_op_bany2:
- case nir_op_bany3:
- case nir_op_bany4: {
- unsigned swiz =
- brw_swizzle_for_size(nir_op_infos[instr->op].input_sizes[0]);
-
- emit(CMP(dst_null_d(), swizzle(op[0], swiz), src_reg(0),
- BRW_CONDITIONAL_NZ));
- emit(MOV(dst, src_reg(0)));
- inst = emit(MOV(dst, src_reg(~0)));
- inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H;
- break;
- }
-
case nir_op_fabs:
case nir_op_iabs:
case nir_op_fneg:
BRW_NIR_BOOLEAN_NEEDS_RESOLVE) {
dst_reg masked = dst_reg(this, glsl_type::int_type);
masked.writemask = dst.writemask;
- emit(AND(masked, src_reg(dst), src_reg(1)));
+ emit(AND(masked, src_reg(dst), brw_imm_d(1)));
src_reg masked_neg = src_reg(masked);
masked_neg.negate = true;
emit(MOV(retype(dst, BRW_REGISTER_TYPE_D), masked_neg));
case nir_texop_txf_ms: op = ir_txf_ms; break;
case nir_texop_txl: op = ir_txl; break;
case nir_texop_txs: op = ir_txs; break;
+ case nir_texop_samples_identical: op = ir_samples_identical; break;
default:
unreachable("unknown texture opcode");
}
return glsl_type::vec(components);
case nir_type_int:
return glsl_type::ivec(components);
- case nir_type_unsigned:
+ case nir_type_uint:
return glsl_type::uvec(components);
case nir_type_bool:
return glsl_type::bvec(components);
void
vec4_visitor::nir_emit_texture(nir_tex_instr *instr)
{
+ unsigned texture = instr->texture_index;
unsigned sampler = instr->sampler_index;
- src_reg sampler_reg = src_reg(sampler);
+ src_reg texture_reg = brw_imm_ud(texture);
+ src_reg sampler_reg = brw_imm_ud(sampler);
src_reg coordinate;
const glsl_type *coord_type = NULL;
src_reg shadow_comparitor;
nir_tex_instr_dest_size(instr));
dst_reg dest = get_nir_dest(instr->dest, instr->dest_type);
- /* When tg4 is used with the degenerate ZERO/ONE swizzles, don't bother
- * emitting anything other than setting up the constant result.
- */
- if (instr->op == nir_texop_tg4) {
- int swiz = GET_SWZ(key_tex->swizzles[sampler], instr->component);
- if (swiz == SWIZZLE_ZERO || swiz == SWIZZLE_ONE) {
- emit(MOV(dest, src_reg(swiz == SWIZZLE_ONE ? 1.0f : 0.0f)));
- return;
- }
- }
+ /* The hardware requires a LOD for buffer textures */
+ if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
+ lod = brw_imm_d(0);
/* Load the texture operation sources */
+ uint32_t constant_offset = 0;
for (unsigned i = 0; i < instr->num_srcs; i++) {
switch (instr->src[i].src_type) {
case nir_tex_src_comparitor:
switch (instr->op) {
case nir_texop_txf:
case nir_texop_txf_ms:
+ case nir_texop_samples_identical:
coordinate = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D,
src_size);
coord_type = glsl_type::ivec(src_size);
case nir_tex_src_ms_index: {
sample_index = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D, 1);
- assert(coord_type != NULL);
- if (devinfo->gen >= 7 &&
- key_tex->compressed_multisample_layout_mask & (1 << sampler)) {
- mcs = emit_mcs_fetch(coord_type, coordinate, sampler_reg);
- } else {
- mcs = src_reg(0u);
- }
- mcs = retype(mcs, BRW_REGISTER_TYPE_UD);
break;
}
- case nir_tex_src_offset:
- offset_value = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D, 2);
+ case nir_tex_src_offset: {
+ nir_const_value *const_offset =
+ nir_src_as_const_value(instr->src[i].src);
+ if (const_offset) {
+ constant_offset = brw_texture_offset(const_offset->i32, 3);
+ } else {
+ offset_value =
+ get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D, 2);
+ }
break;
+ }
- case nir_tex_src_sampler_offset: {
- /* The highest sampler which may be used by this operation is
+ case nir_tex_src_texture_offset: {
+ /* The highest texture which may be used by this operation is
* the last element of the array. Mark it here, because the generator
* doesn't have enough information to determine the bound.
*/
- uint32_t array_size = instr->sampler_array_size;
- uint32_t max_used = sampler + array_size - 1;
+ uint32_t array_size = instr->texture_array_size;
+ uint32_t max_used = texture + array_size - 1;
if (instr->op == nir_texop_tg4) {
max_used += prog_data->base.binding_table.gather_texture_start;
} else {
/* Emit code to evaluate the actual indexing expression */
src_reg src = get_nir_src(instr->src[i].src, 1);
src_reg temp(this, glsl_type::uint_type);
- emit(ADD(dst_reg(temp), src, src_reg(sampler)));
+ emit(ADD(dst_reg(temp), src, brw_imm_ud(texture)));
+ texture_reg = emit_uniformize(temp);
+ break;
+ }
+
+ case nir_tex_src_sampler_offset: {
+ /* Emit code to evaluate the actual indexing expression */
+ src_reg src = get_nir_src(instr->src[i].src, 1);
+ src_reg temp(this, glsl_type::uint_type);
+ emit(ADD(dst_reg(temp), src, brw_imm_ud(sampler)));
sampler_reg = emit_uniformize(temp);
break;
}
}
}
- uint32_t constant_offset = 0;
- for (unsigned i = 0; i < 3; i++) {
- if (instr->const_offset[i] != 0) {
- constant_offset = brw_texture_offset(instr->const_offset, 3);
- break;
+ if (instr->op == nir_texop_txf_ms ||
+ instr->op == nir_texop_samples_identical) {
+ assert(coord_type != NULL);
+ if (devinfo->gen >= 7 &&
+ key_tex->compressed_multisample_layout_mask & (1 << texture)) {
+ mcs = emit_mcs_fetch(coord_type, coordinate, texture_reg);
+ } else {
+ mcs = brw_imm_ud(0u);
}
}
/* Stuff the channel select bits in the top of the texture offset */
- if (instr->op == nir_texop_tg4)
- constant_offset |= gather_channel(instr->component, sampler) << 16;
+ if (instr->op == nir_texop_tg4) {
+ if (instr->component == 1 &&
+ (key_tex->gather_channel_quirk_mask & (1 << texture))) {
+ /* gather4 sampler is broken for green channel on RG32F --
+ * we must ask for blue instead.
+ */
+ constant_offset |= 2 << 16;
+ } else {
+ constant_offset |= instr->component << 16;
+ }
+ }
ir_texture_opcode op = ir_texture_opcode_for_nir_texop(instr->op);
shadow_comparitor,
lod, lod2, sample_index,
constant_offset, offset_value,
- mcs, is_cube_array, sampler, sampler_reg);
+ mcs, is_cube_array,
+ texture, texture_reg, sampler, sampler_reg);
}
void
vec4_visitor::nir_emit_undef(nir_ssa_undef_instr *instr)
{
- nir_ssa_values[instr->def.index] = dst_reg(GRF, alloc.allocate(1));
+ nir_ssa_values[instr->def.index] = dst_reg(VGRF, alloc.allocate(1));
}
}
projects / mesa.git / blobdiff