#include "brw_program.h"
#include "brw_dead_control_flow.h"
#include "compiler/glsl_types.h"
+#include "compiler/nir/nir_builder.h"
#include "program/prog_parameter.h"
using namespace brw;
+static unsigned get_lowered_simd_width(const struct gen_device_info *devinfo,
+ const fs_inst *inst);
+
void
fs_inst::init(enum opcode opcode, uint8_t exec_size, const fs_reg &dst,
const fs_reg *src, unsigned sources)
this->dst = dst;
this->sources = sources;
this->exec_size = exec_size;
+ this->base_mrf = -1;
assert(dst.file != IMM && dst.file != UNIFORM);
case FIXED_GRF:
case MRF:
case ATTR:
- this->regs_written = DIV_ROUND_UP(dst.component_size(exec_size),
- REG_SIZE);
+ this->size_written = dst.component_size(exec_size);
break;
case BAD_FILE:
- this->regs_written = 0;
+ this->size_written = 0;
break;
case IMM:
case UNIFORM:
* be any component of a vector, and then we load 4 contiguous
* components starting from that.
*
- * We break down the const_offset to a portion added to the variable
- * offset and a portion done using reg_offset, which means that if you
- * have GLSL using something like "uniform vec4 a[20]; gl_FragColor =
- * a[i]", we'll temporarily generate 4 vec4 loads from offset i * 4, and
- * CSE can later notice that those loads are all the same and eliminate
- * the redundant ones.
+ * We break down the const_offset to a portion added to the variable offset
+ * and a portion done using fs_reg::offset, which means that if you have
+ * GLSL using something like "uniform vec4 a[20]; gl_FragColor = a[i]",
+ * we'll temporarily generate 4 vec4 loads from offset i * 4, and CSE can
+ * later notice that those loads are all the same and eliminate the
+ * redundant ones.
*/
fs_reg vec4_offset = vgrf(glsl_type::uint_type);
bld.ADD(vec4_offset, varying_offset, brw_imm_ud(const_offset & ~0xf));
fs_reg vec4_result = bld.vgrf(BRW_REGISTER_TYPE_F, 4);
fs_inst *inst = bld.emit(FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_LOGICAL,
vec4_result, surf_index, vec4_offset);
- inst->regs_written = 4 * bld.dispatch_width() / 8;
+ inst->size_written = 4 * vec4_result.component_size(inst->exec_size);
if (type_sz(dst.type) == 8) {
shuffle_32bit_load_result_to_64bit_data(
offset == inst->offset);
}
-bool
-fs_inst::overwrites_reg(const fs_reg ®) const
-{
- return reg.in_range(dst, regs_written);
-}
-
bool
fs_inst::is_send_from_grf() const
{
switch (opcode) {
case FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7:
case SHADER_OPCODE_SHADER_TIME_ADD:
- case FS_OPCODE_INTERPOLATE_AT_CENTROID:
case FS_OPCODE_INTERPOLATE_AT_SAMPLE:
case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET:
case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
return src[1].file == VGRF;
case FS_OPCODE_FB_WRITE:
+ case FS_OPCODE_FB_READ:
return src[0].file == VGRF;
default:
if (is_tex())
return false;
fs_reg reg = this->src[0];
- if (reg.file != VGRF || reg.reg_offset != 0 || reg.stride == 0)
+ if (reg.file != VGRF || reg.offset != 0 || reg.stride != 1)
return false;
- if (grf_alloc.sizes[reg.nr] != this->regs_written)
+ if (grf_alloc.sizes[reg.nr] * REG_SIZE != this->size_written)
return false;
for (int i = 0; i < this->sources; i++) {
return false;
if (i < this->header_size) {
- reg.reg_offset += 1;
+ reg.offset += REG_SIZE;
} else {
reg = horiz_offset(reg, this->exec_size);
}
}
bool
-fs_inst::can_do_source_mods(const struct brw_device_info *devinfo)
+fs_inst::can_do_source_mods(const struct gen_device_info *devinfo)
{
if (devinfo->gen == 6 && is_math())
return false;
fs_reg::fs_reg(struct ::brw_reg reg) :
backend_reg(reg)
{
- this->reg_offset = 0;
- this->subreg_offset = 0;
+ this->offset = 0;
this->stride = 1;
if (this->file == IMM &&
(this->type != BRW_REGISTER_TYPE_V &&
fs_reg::equals(const fs_reg &r) const
{
return (this->backend_reg::equals(r) &&
- subreg_offset == r.subreg_offset &&
stride == r.stride);
}
-fs_reg &
-fs_reg::set_smear(unsigned subreg)
-{
- assert(file != ARF && file != FIXED_GRF && file != IMM);
- subreg_offset = subreg * type_sz(type);
- stride = 0;
- return *this;
-}
-
bool
fs_reg::is_contiguous() const
{
return 0;
}
-/**
- * Returns the number of scalar components needed to store type, assuming
- * that vectors are padded out to vec4.
- *
- * This has the packing rules of type_size_vec4(), but counts components
- * similar to type_size_scalar().
- */
-extern "C" int
-type_size_vec4_times_4(const struct glsl_type *type)
-{
- return 4 * type_size_vec4(type);
-}
-
/* Attribute arrays are loaded as one vec4 per element (or matrix column),
* except for double-precision types, which are loaded as one dvec4.
*/
void
fs_visitor::emit_shader_time_begin()
{
- shader_start_time = get_timestamp(bld.annotate("shader time start"));
-
/* We want only the low 32 bits of the timestamp. Since it's running
* at the GPU clock rate of ~1.2ghz, it will roll over every ~3 seconds,
* which is plenty of time for our purposes. It is identical across the
* EUs, but since it's tracking GPU core speed it will increment at a
* varying rate as render P-states change.
*/
- shader_start_time.set_smear(0);
+ shader_start_time = component(
+ get_timestamp(bld.annotate("shader time start")), 0);
}
void
assert(end && ((fs_inst *) end)->eot);
const fs_builder ibld = bld.annotate("shader time end")
.exec_all().at(NULL, end);
-
- fs_reg shader_end_time = get_timestamp(ibld);
+ const fs_reg timestamp = get_timestamp(ibld);
/* We only use the low 32 bits of the timestamp - see
* emit_shader_time_begin()).
* else that might disrupt timing) by setting smear to 2 and checking if
* that field is != 0.
*/
- shader_end_time.set_smear(0);
+ const fs_reg shader_end_time = component(timestamp, 0);
/* Check that there weren't any timestamp reset events (assuming these
* were the only two timestamp reads that happened).
*/
- fs_reg reset = shader_end_time;
- reset.set_smear(2);
+ const fs_reg reset = component(timestamp, 2);
set_condmod(BRW_CONDITIONAL_Z,
ibld.AND(ibld.null_reg_ud(), reset, brw_imm_ud(1u)));
ibld.IF(BRW_PREDICATE_NORMAL);
fs_reg start = shader_start_time;
start.negate = true;
- fs_reg diff = fs_reg(VGRF, alloc.allocate(1), BRW_REGISTER_TYPE_UD);
- diff.set_smear(0);
-
+ const fs_reg diff = component(fs_reg(VGRF, alloc.allocate(1),
+ BRW_REGISTER_TYPE_UD),
+ 0);
const fs_builder cbld = ibld.group(1, 0);
cbld.group(1, 0).ADD(diff, start, shader_end_time);
return ((this->predicate && this->opcode != BRW_OPCODE_SEL) ||
(this->exec_size * type_sz(this->dst.type)) < 32 ||
!this->dst.is_contiguous() ||
- this->dst.subreg_offset > 0);
+ this->dst.offset % REG_SIZE != 0);
}
unsigned
fs_inst::components_read(unsigned i) const
{
+ /* Return zero if the source is not present. */
+ if (src[i].file == BAD_FILE)
+ return 0;
+
switch (opcode) {
case FS_OPCODE_LINTERP:
if (i == 0)
}
}
-int
-fs_inst::regs_read(int arg) const
+unsigned
+fs_inst::size_read(int arg) const
{
switch (opcode) {
case FS_OPCODE_FB_WRITE:
+ case FS_OPCODE_FB_READ:
case SHADER_OPCODE_URB_WRITE_SIMD8:
case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT:
case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED:
case SHADER_OPCODE_TYPED_SURFACE_WRITE:
case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
if (arg == 0)
- return mlen;
+ return mlen * REG_SIZE;
break;
case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7:
/* The payload is actually stored in src1 */
if (arg == 1)
- return mlen;
+ return mlen * REG_SIZE;
break;
case FS_OPCODE_LINTERP:
if (arg == 1)
- return 1;
+ return 16;
break;
case SHADER_OPCODE_LOAD_PAYLOAD:
if (arg < this->header_size)
- return 1;
+ return REG_SIZE;
break;
case CS_OPCODE_CS_TERMINATE:
case SHADER_OPCODE_BARRIER:
- return 1;
+ return REG_SIZE;
case SHADER_OPCODE_MOV_INDIRECT:
if (arg == 0) {
assert(src[2].file == IMM);
- unsigned region_length = src[2].ud;
-
- if (src[0].file == UNIFORM) {
- assert(region_length % 4 == 0);
- return region_length / 4;
- } else if (src[0].file == FIXED_GRF) {
- /* If the start of the region is not register aligned, then
- * there's some portion of the register that's technically
- * unread at the beginning.
- *
- * However, the register allocator works in terms of whole
- * registers, and does not use subnr. It assumes that the
- * read starts at the beginning of the register, and extends
- * regs_read() whole registers beyond that.
- *
- * To compensate, we extend the region length to include this
- * unread portion at the beginning.
- */
- if (src[0].subnr)
- region_length += src[0].subnr;
-
- return DIV_ROUND_UP(region_length, REG_SIZE);
- } else {
- assert(!"Invalid register file");
- }
+ return src[2].ud;
}
break;
default:
if (is_tex() && arg == 0 && src[0].file == VGRF)
- return mlen;
+ return mlen * REG_SIZE;
break;
}
switch (src[arg].file) {
- case BAD_FILE:
- return 0;
case UNIFORM:
case IMM:
- return 1;
+ return components_read(arg) * type_sz(src[arg].type);
+ case BAD_FILE:
case ARF:
case FIXED_GRF:
case VGRF:
case ATTR:
- return DIV_ROUND_UP(components_read(arg) *
- src[arg].component_size(exec_size),
- REG_SIZE);
+ return components_read(arg) * src[arg].component_size(exec_size);
case MRF:
unreachable("MRF registers are not allowed as sources");
}
}
unsigned
-fs_inst::flags_read(const brw_device_info *devinfo) const
+fs_inst::flags_read(const gen_device_info *devinfo) const
{
/* XXX - This doesn't consider explicit uses of the flag register as source
* region.
case FS_OPCODE_TXB:
case SHADER_OPCODE_TXD:
case SHADER_OPCODE_TXF:
- case SHADER_OPCODE_TXF_LZ:
case SHADER_OPCODE_TXF_CMS:
- case SHADER_OPCODE_TXF_CMS_W:
case SHADER_OPCODE_TXF_MCS:
case SHADER_OPCODE_TG4:
case SHADER_OPCODE_TG4_OFFSET:
case SHADER_OPCODE_TXL:
- case SHADER_OPCODE_TXL_LZ:
case SHADER_OPCODE_TXS:
case SHADER_OPCODE_LOD:
case SHADER_OPCODE_SAMPLEINFO:
return 1;
case FS_OPCODE_FB_WRITE:
return 2;
- case FS_OPCODE_GET_BUFFER_SIZE:
case FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD:
case SHADER_OPCODE_GEN4_SCRATCH_READ:
return 1;
return inst->mlen;
case SHADER_OPCODE_GEN4_SCRATCH_WRITE:
return inst->mlen;
- case SHADER_OPCODE_UNTYPED_ATOMIC:
- case SHADER_OPCODE_UNTYPED_SURFACE_READ:
- case SHADER_OPCODE_UNTYPED_SURFACE_WRITE:
- case SHADER_OPCODE_TYPED_ATOMIC:
- case SHADER_OPCODE_TYPED_SURFACE_READ:
- case SHADER_OPCODE_TYPED_SURFACE_WRITE:
- case SHADER_OPCODE_URB_WRITE_SIMD8:
- case SHADER_OPCODE_URB_WRITE_SIMD8_PER_SLOT:
- case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED:
- case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT:
- case FS_OPCODE_INTERPOLATE_AT_CENTROID:
- case FS_OPCODE_INTERPOLATE_AT_SAMPLE:
- case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET:
- case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
- return 0;
default:
unreachable("not reached");
}
this->uniforms = v->uniforms;
}
-fs_reg *
-fs_visitor::emit_fragcoord_interpolation()
+void
+fs_visitor::emit_fragcoord_interpolation(fs_reg wpos)
{
assert(stage == MESA_SHADER_FRAGMENT);
- fs_reg *reg = new(this->mem_ctx) fs_reg(vgrf(glsl_type::vec4_type));
- fs_reg wpos = *reg;
/* gl_FragCoord.x */
bld.MOV(wpos, this->pixel_x);
bld.MOV(wpos, fs_reg(brw_vec8_grf(payload.source_depth_reg, 0)));
} else {
bld.emit(FS_OPCODE_LINTERP, wpos,
- this->delta_xy[BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC],
+ this->delta_xy[BRW_BARYCENTRIC_PERSPECTIVE_PIXEL],
interp_reg(VARYING_SLOT_POS, 2));
}
wpos = offset(wpos, bld, 1);
/* gl_FragCoord.w: Already set up in emit_interpolation */
bld.MOV(wpos, this->wpos_w);
-
- return reg;
}
-fs_inst *
-fs_visitor::emit_linterp(const fs_reg &attr, const fs_reg &interp,
- glsl_interp_qualifier interpolation_mode,
- bool is_centroid, bool is_sample)
+enum brw_barycentric_mode
+brw_barycentric_mode(enum glsl_interp_mode mode, nir_intrinsic_op op)
{
- brw_wm_barycentric_interp_mode barycoord_mode;
- if (devinfo->gen >= 6) {
- if (is_centroid) {
- if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
- barycoord_mode = BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC;
- else
- barycoord_mode = BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC;
- } else if (is_sample) {
- if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
- barycoord_mode = BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC;
- else
- barycoord_mode = BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC;
- } else {
- if (interpolation_mode == INTERP_QUALIFIER_SMOOTH)
- barycoord_mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
- else
- barycoord_mode = BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
- }
- } else {
- /* On Ironlake and below, there is only one interpolation mode.
- * Centroid interpolation doesn't mean anything on this hardware --
- * there is no multisampling.
- */
- barycoord_mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
- }
- return bld.emit(FS_OPCODE_LINTERP, attr,
- this->delta_xy[barycoord_mode], interp);
-}
+ /* Barycentric modes don't make sense for flat inputs. */
+ assert(mode != INTERP_MODE_FLAT);
-void
-fs_visitor::emit_general_interpolation(fs_reg *attr, const char *name,
- const glsl_type *type,
- glsl_interp_qualifier interpolation_mode,
- int *location, bool mod_centroid,
- bool mod_sample)
-{
- assert(stage == MESA_SHADER_FRAGMENT);
- brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
- brw_wm_prog_key *key = (brw_wm_prog_key*) this->key;
-
- if (interpolation_mode == INTERP_QUALIFIER_NONE) {
- bool is_gl_Color =
- *location == VARYING_SLOT_COL0 || *location == VARYING_SLOT_COL1;
- if (key->flat_shade && is_gl_Color) {
- interpolation_mode = INTERP_QUALIFIER_FLAT;
- } else {
- interpolation_mode = INTERP_QUALIFIER_SMOOTH;
- }
+ unsigned bary;
+ switch (op) {
+ case nir_intrinsic_load_barycentric_pixel:
+ case nir_intrinsic_load_barycentric_at_offset:
+ bary = BRW_BARYCENTRIC_PERSPECTIVE_PIXEL;
+ break;
+ case nir_intrinsic_load_barycentric_centroid:
+ bary = BRW_BARYCENTRIC_PERSPECTIVE_CENTROID;
+ break;
+ case nir_intrinsic_load_barycentric_sample:
+ case nir_intrinsic_load_barycentric_at_sample:
+ bary = BRW_BARYCENTRIC_PERSPECTIVE_SAMPLE;
+ break;
+ default:
+ unreachable("invalid intrinsic");
}
- if (type->is_array() || type->is_matrix()) {
- const glsl_type *elem_type = glsl_get_array_element(type);
- const unsigned length = glsl_get_length(type);
-
- for (unsigned i = 0; i < length; i++) {
- emit_general_interpolation(attr, name, elem_type, interpolation_mode,
- location, mod_centroid, mod_sample);
- }
- } else if (type->is_record()) {
- for (unsigned i = 0; i < type->length; i++) {
- const glsl_type *field_type = type->fields.structure[i].type;
- emit_general_interpolation(attr, name, field_type, interpolation_mode,
- location, mod_centroid, mod_sample);
- }
- } else {
- assert(type->is_scalar() || type->is_vector());
+ if (mode == INTERP_MODE_NOPERSPECTIVE)
+ bary += 3;
- if (prog_data->urb_setup[*location] == -1) {
- /* If there's no incoming setup data for this slot, don't
- * emit interpolation for it.
- */
- *attr = offset(*attr, bld, type->vector_elements);
- (*location)++;
- return;
- }
-
- attr->type = brw_type_for_base_type(type->get_scalar_type());
-
- if (interpolation_mode == INTERP_QUALIFIER_FLAT) {
- /* Constant interpolation (flat shading) case. The SF has
- * handed us defined values in only the constant offset
- * field of the setup reg.
- */
- for (unsigned int i = 0; i < type->vector_elements; i++) {
- struct brw_reg interp = interp_reg(*location, i);
- interp = suboffset(interp, 3);
- interp.type = attr->type;
- bld.emit(FS_OPCODE_CINTERP, *attr, fs_reg(interp));
- *attr = offset(*attr, bld, 1);
- }
- } else {
- /* Smooth/noperspective interpolation case. */
- for (unsigned int i = 0; i < type->vector_elements; i++) {
- struct brw_reg interp = interp_reg(*location, i);
- if (devinfo->needs_unlit_centroid_workaround && mod_centroid) {
- /* Get the pixel/sample mask into f0 so that we know
- * which pixels are lit. Then, for each channel that is
- * unlit, replace the centroid data with non-centroid
- * data.
- */
- bld.emit(FS_OPCODE_MOV_DISPATCH_TO_FLAGS);
-
- fs_inst *inst;
- inst = emit_linterp(*attr, fs_reg(interp), interpolation_mode,
- false, false);
- inst->predicate = BRW_PREDICATE_NORMAL;
- inst->predicate_inverse = true;
- if (devinfo->has_pln)
- inst->no_dd_clear = true;
-
- inst = emit_linterp(*attr, fs_reg(interp), interpolation_mode,
- mod_centroid && !key->persample_interp,
- mod_sample || key->persample_interp);
- inst->predicate = BRW_PREDICATE_NORMAL;
- inst->predicate_inverse = false;
- if (devinfo->has_pln)
- inst->no_dd_check = true;
+ return (enum brw_barycentric_mode) bary;
+}
- } else {
- emit_linterp(*attr, fs_reg(interp), interpolation_mode,
- mod_centroid && !key->persample_interp,
- mod_sample || key->persample_interp);
- }
- if (devinfo->gen < 6 && interpolation_mode == INTERP_QUALIFIER_SMOOTH) {
- bld.MUL(*attr, *attr, this->pixel_w);
- }
- *attr = offset(*attr, bld, 1);
- }
- }
- (*location)++;
- }
+/**
+ * Turn one of the two CENTROID barycentric modes into PIXEL mode.
+ */
+static enum brw_barycentric_mode
+centroid_to_pixel(enum brw_barycentric_mode bary)
+{
+ assert(bary == BRW_BARYCENTRIC_PERSPECTIVE_CENTROID ||
+ bary == BRW_BARYCENTRIC_NONPERSPECTIVE_CENTROID);
+ return (enum brw_barycentric_mode) ((unsigned) bary - 1);
}
fs_reg *
fs_visitor::compute_sample_position(fs_reg dst, fs_reg int_sample_pos)
{
assert(stage == MESA_SHADER_FRAGMENT);
- brw_wm_prog_data *wm_prog_data = (brw_wm_prog_data *) this->prog_data;
+ struct brw_wm_prog_data *wm_prog_data = brw_wm_prog_data(this->prog_data);
assert(dst.type == BRW_REGISTER_TYPE_F);
if (wm_prog_data->persample_dispatch) {
brw_imm_v(0x44440000));
abld.AND(*reg, tmp, brw_imm_w(0xf));
} else {
- fs_reg t1(VGRF, alloc.allocate(1), BRW_REGISTER_TYPE_D);
- t1.set_smear(0);
- fs_reg t2(VGRF, alloc.allocate(1), BRW_REGISTER_TYPE_W);
+ const fs_reg t1 = component(fs_reg(VGRF, alloc.allocate(1),
+ BRW_REGISTER_TYPE_D), 0);
+ const fs_reg t2(VGRF, alloc.allocate(1), BRW_REGISTER_TYPE_W);
/* The PS will be run in MSDISPMODE_PERSAMPLE. For example with
* 8x multisampling, subspan 0 will represent sample N (where N
fs_visitor::emit_samplemaskin_setup()
{
assert(stage == MESA_SHADER_FRAGMENT);
- brw_wm_prog_data *wm_prog_data = (brw_wm_prog_data *) this->prog_data;
+ struct brw_wm_prog_data *wm_prog_data = brw_wm_prog_data(this->prog_data);
assert(devinfo->gen >= 6);
fs_reg *reg = new(this->mem_ctx) fs_reg(vgrf(glsl_type::int_type));
void
fs_visitor::emit_discard_jump()
{
- assert(((brw_wm_prog_data*) this->prog_data)->uses_kill);
+ assert(brw_wm_prog_data(this->prog_data)->uses_kill);
/* For performance, after a discard, jump to the end of the
* shader if all relevant channels have been discarded.
fs_inst *discard_jump = bld.emit(FS_OPCODE_DISCARD_JUMP);
discard_jump->flag_subreg = 1;
- discard_jump->predicate = (dispatch_width == 8)
- ? BRW_PREDICATE_ALIGN1_ANY8H
- : BRW_PREDICATE_ALIGN1_ANY16H;
+ discard_jump->predicate = BRW_PREDICATE_ALIGN1_ANY4H;
discard_jump->predicate_inverse = true;
}
{
assert(stage == MESA_SHADER_GEOMETRY);
- struct brw_gs_prog_data *gs_prog_data =
- (struct brw_gs_prog_data *) prog_data;
+ struct brw_gs_prog_data *gs_prog_data = brw_gs_prog_data(prog_data);
if (gs_compile->control_data_header_size_bits > 0) {
emit_gs_control_data_bits(this->final_gs_vertex_count);
foreach_block_and_inst(block, fs_inst, inst, cfg) {
for (unsigned int i = 0; i < inst->sources; i++) {
if (inst->src[i].file == UNIFORM) {
- int uniform_nr = inst->src[i].nr + inst->src[i].reg_offset;
+ int uniform_nr = inst->src[i].nr + inst->src[i].offset / 4;
int constant_nr;
if (uniform_nr >= 0 && uniform_nr < (int) uniforms) {
constant_nr = push_constant_loc[uniform_nr];
assert(inst->src[i].stride == 0);
inst->src[i] = byte_offset(
retype(brw_reg, inst->src[i].type),
- inst->src[i].subreg_offset);
+ inst->src[i].offset % 4);
}
}
}
fs_visitor::calculate_urb_setup()
{
assert(stage == MESA_SHADER_FRAGMENT);
- brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
+ struct brw_wm_prog_data *prog_data = brw_wm_prog_data(this->prog_data);
brw_wm_prog_key *key = (brw_wm_prog_key*) this->key;
memset(prog_data->urb_setup, -1,
int urb_next = 0;
/* Figure out where each of the incoming setup attributes lands. */
if (devinfo->gen >= 6) {
- if (_mesa_bitcount_64(nir->info.inputs_read &
+ if (_mesa_bitcount_64(nir->info->inputs_read &
BRW_FS_VARYING_INPUT_MASK) <= 16) {
/* The SF/SBE pipeline stage can do arbitrary rearrangement of the
* first 16 varying inputs, so we can put them wherever we want.
* a different vertex (or geometry) shader.
*/
for (unsigned int i = 0; i < VARYING_SLOT_MAX; i++) {
- if (nir->info.inputs_read & BRW_FS_VARYING_INPUT_MASK &
+ if (nir->info->inputs_read & BRW_FS_VARYING_INPUT_MASK &
BITFIELD64_BIT(i)) {
prog_data->urb_setup[i] = urb_next++;
}
}
} else {
bool include_vue_header =
- nir->info.inputs_read & (VARYING_BIT_LAYER | VARYING_BIT_VIEWPORT);
+ nir->info->inputs_read & (VARYING_BIT_LAYER | VARYING_BIT_VIEWPORT);
/* We have enough input varyings that the SF/SBE pipeline stage can't
* arbitrarily rearrange them to suit our whim; we have to put them
struct brw_vue_map prev_stage_vue_map;
brw_compute_vue_map(devinfo, &prev_stage_vue_map,
key->input_slots_valid,
- nir->info.separate_shader);
+ nir->info->separate_shader);
int first_slot =
include_vue_header ? 0 : 2 * BRW_SF_URB_ENTRY_READ_OFFSET;
slot++) {
int varying = prev_stage_vue_map.slot_to_varying[slot];
if (varying != BRW_VARYING_SLOT_PAD &&
- (nir->info.inputs_read & BRW_FS_VARYING_INPUT_MASK &
+ (nir->info->inputs_read & BRW_FS_VARYING_INPUT_MASK &
BITFIELD64_BIT(varying))) {
prog_data->urb_setup[varying] = slot - first_slot;
}
*
* See compile_sf_prog() for more info.
*/
- if (nir->info.inputs_read & BITFIELD64_BIT(VARYING_SLOT_PNTC))
+ if (nir->info->inputs_read & BITFIELD64_BIT(VARYING_SLOT_PNTC))
prog_data->urb_setup[VARYING_SLOT_PNTC] = urb_next++;
}
fs_visitor::assign_urb_setup()
{
assert(stage == MESA_SHADER_FRAGMENT);
- brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
+ struct brw_wm_prog_data *prog_data = brw_wm_prog_data(this->prog_data);
int urb_start = payload.num_regs + prog_data->base.curb_read_length;
int grf = payload.num_regs +
prog_data->curb_read_length +
inst->src[i].nr +
- inst->src[i].reg_offset;
+ inst->src[i].offset / REG_SIZE;
/* As explained at brw_reg_from_fs_reg, From the Haswell PRM:
*
unsigned width = inst->src[i].stride == 0 ? 1 : exec_size;
struct brw_reg reg =
stride(byte_offset(retype(brw_vec8_grf(grf, 0), inst->src[i].type),
- inst->src[i].subreg_offset),
+ inst->src[i].offset % REG_SIZE),
exec_size * inst->src[i].stride,
width, inst->src[i].stride);
reg.abs = inst->src[i].abs;
void
fs_visitor::assign_vs_urb_setup()
{
- brw_vs_prog_data *vs_prog_data = (brw_vs_prog_data *) prog_data;
+ struct brw_vs_prog_data *vs_prog_data = brw_vs_prog_data(prog_data);
assert(stage == MESA_SHADER_VERTEX);
{
assert(stage == MESA_SHADER_TESS_EVAL);
- brw_vue_prog_data *vue_prog_data = (brw_vue_prog_data *) prog_data;
+ struct brw_vue_prog_data *vue_prog_data = brw_vue_prog_data(prog_data);
first_non_payload_grf += 8 * vue_prog_data->urb_read_length;
{
assert(stage == MESA_SHADER_GEOMETRY);
- brw_vue_prog_data *vue_prog_data = (brw_vue_prog_data *) prog_data;
+ struct brw_vue_prog_data *vue_prog_data = brw_vue_prog_data(prog_data);
first_non_payload_grf +=
- 8 * vue_prog_data->urb_read_length * nir->info.gs.vertices_in;
+ 8 * vue_prog_data->urb_read_length * nir->info->gs.vertices_in;
foreach_block_and_inst(block, fs_inst, inst, cfg) {
/* Rewrite all ATTR file references to GRFs. */
void
fs_visitor::split_virtual_grfs()
{
+ /* Compact the register file so we eliminate dead vgrfs. This
+ * only defines split points for live registers, so if we have
+ * too large dead registers they will hit assertions later.
+ */
+ compact_virtual_grfs();
+
int num_vars = this->alloc.count;
/* Count the total number of registers */
foreach_block_and_inst(block, fs_inst, inst, cfg) {
if (inst->dst.file == VGRF) {
- int reg = vgrf_to_reg[inst->dst.nr] + inst->dst.reg_offset;
- for (int j = 1; j < inst->regs_written; j++)
+ int reg = vgrf_to_reg[inst->dst.nr] + inst->dst.offset / REG_SIZE;
+ for (unsigned j = 1; j < regs_written(inst); j++)
split_points[reg + j] = false;
}
for (int i = 0; i < inst->sources; i++) {
if (inst->src[i].file == VGRF) {
- int reg = vgrf_to_reg[inst->src[i].nr] + inst->src[i].reg_offset;
- for (int j = 1; j < inst->regs_read(i); j++)
+ int reg = vgrf_to_reg[inst->src[i].nr] + inst->src[i].offset / REG_SIZE;
+ for (unsigned j = 1; j < regs_read(inst, i); j++)
split_points[reg + j] = false;
}
}
foreach_block_and_inst(block, fs_inst, inst, cfg) {
if (inst->dst.file == VGRF) {
- reg = vgrf_to_reg[inst->dst.nr] + inst->dst.reg_offset;
+ reg = vgrf_to_reg[inst->dst.nr] + inst->dst.offset / REG_SIZE;
inst->dst.nr = new_virtual_grf[reg];
- inst->dst.reg_offset = new_reg_offset[reg];
+ inst->dst.offset = new_reg_offset[reg] * REG_SIZE +
+ inst->dst.offset % REG_SIZE;
assert((unsigned)new_reg_offset[reg] < alloc.sizes[new_virtual_grf[reg]]);
}
for (int i = 0; i < inst->sources; i++) {
if (inst->src[i].file == VGRF) {
- reg = vgrf_to_reg[inst->src[i].nr] + inst->src[i].reg_offset;
+ reg = vgrf_to_reg[inst->src[i].nr] + inst->src[i].offset / REG_SIZE;
inst->src[i].nr = new_virtual_grf[reg];
- inst->src[i].reg_offset = new_reg_offset[reg];
+ inst->src[i].offset = new_reg_offset[reg] * REG_SIZE +
+ inst->src[i].offset % REG_SIZE;
assert((unsigned)new_reg_offset[reg] < alloc.sizes[new_virtual_grf[reg]]);
}
}
bool contiguous[uniforms];
memset(contiguous, 0, sizeof(contiguous));
+ int thread_local_id_index =
+ (stage == MESA_SHADER_COMPUTE) ?
+ brw_cs_prog_data(stage_prog_data)->thread_local_id_index : -1;
+
/* First, we walk through the instructions and do two things:
*
* 1) Figure out which uniforms are live.
if (inst->src[i].file != UNIFORM)
continue;
- int constant_nr = inst->src[i].nr + inst->src[i].reg_offset;
+ int constant_nr = inst->src[i].nr + inst->src[i].offset / 4;
if (inst->opcode == SHADER_OPCODE_MOV_INDIRECT && i == 0) {
assert(inst->src[2].ud % 4 == 0);
}
}
+ if (thread_local_id_index >= 0 && !is_live[thread_local_id_index])
+ thread_local_id_index = -1;
+
/* Only allow 16 registers (128 uniform components) as push constants.
*
* Just demote the end of the list. We could probably do better
* If changing this value, note the limitation about total_regs in
* brw_curbe.c.
*/
- const unsigned int max_push_components = 16 * 8;
+ unsigned int max_push_components = 16 * 8;
+ if (thread_local_id_index >= 0)
+ max_push_components--; /* Save a slot for the thread ID */
/* We push small arrays, but no bigger than 16 floats. This is big enough
* for a vec4 but hopefully not large enough to push out other stuff. We
if (!is_live[u] || is_live_64bit[u])
continue;
+ /* Skip thread_local_id_index to put it in the last push register. */
+ if (thread_local_id_index == (int)u)
+ continue;
+
set_push_pull_constant_loc(u, &chunk_start, contiguous[u],
push_constant_loc, pull_constant_loc,
&num_push_constants, &num_pull_constants,
stage_prog_data);
}
+ /* Add the CS local thread ID uniform at the end of the push constants */
+ if (thread_local_id_index >= 0)
+ push_constant_loc[thread_local_id_index] = num_push_constants++;
+
/* As the uniforms are going to be reordered, take the data from a temporary
* copy of the original param[].
*/
stage_prog_data->nr_params = num_push_constants;
stage_prog_data->nr_pull_params = num_pull_constants;
- /* Up until now, the param[] array has been indexed by reg + reg_offset
+ /* Up until now, the param[] array has been indexed by reg + offset
* of UNIFORM registers. Move pull constants into pull_param[] and
* condense param[] to only contain the uniforms we chose to push.
*
* push_constant_loc[i] <= i and we can do it in one smooth loop without
* having to make a copy.
*/
+ int new_thread_local_id_index = -1;
for (unsigned int i = 0; i < uniforms; i++) {
const gl_constant_value *value = param[i];
stage_prog_data->pull_param[pull_constant_loc[i]] = value;
} else if (push_constant_loc[i] != -1) {
stage_prog_data->param[push_constant_loc[i]] = value;
+ if (thread_local_id_index == (int)i)
+ new_thread_local_id_index = push_constant_loc[i];
}
}
ralloc_free(param);
+
+ if (stage == MESA_SHADER_COMPUTE)
+ brw_cs_prog_data(stage_prog_data)->thread_local_id_index =
+ new_thread_local_id_index;
}
/**
if (inst->opcode == SHADER_OPCODE_MOV_INDIRECT && i == 0)
continue;
- unsigned location = inst->src[i].nr + inst->src[i].reg_offset;
+ unsigned location = inst->src[i].nr + inst->src[i].offset / 4;
if (location >= uniforms)
continue; /* Out of bounds access */
/* Rewrite the instruction to use the temporary VGRF. */
inst->src[i].file = VGRF;
inst->src[i].nr = dst.nr;
- inst->src[i].reg_offset = 0;
- inst->src[i].set_smear((pull_index & 3) * 4 /
- type_sz(inst->src[i].type));
+ inst->src[i].offset = (pull_index & 3) * 4 + inst->src[i].offset % 4;
brw_mark_surface_used(prog_data, index);
}
if (inst->opcode == SHADER_OPCODE_MOV_INDIRECT &&
inst->src[0].file == UNIFORM) {
- unsigned location = inst->src[0].nr + inst->src[0].reg_offset;
+ unsigned location = inst->src[0].nr + inst->src[0].offset / 4;
if (location >= uniforms)
continue; /* Out of bounds access */
for (unsigned i = 0; i < FB_WRITE_LOGICAL_NUM_SRCS; i++) {
if (i == FB_WRITE_LOGICAL_SRC_COLOR0) {
if (!fb_write->src[i].equals(tex_inst->dst) ||
- fb_write->regs_read(i) != tex_inst->regs_written)
+ fb_write->size_read(i) != tex_inst->size_written)
return false;
} else if (i != FB_WRITE_LOGICAL_SRC_COMPONENTS) {
if (fb_write->src[i].file != BAD_FILE)
tex_inst->offset |= fb_write->target << 24;
tex_inst->eot = true;
tex_inst->dst = ibld.null_reg_ud();
- tex_inst->regs_written = 0;
+ tex_inst->size_written = 0;
fb_write->remove(cfg->blocks[cfg->num_blocks - 1]);
/* Marking EOT is sufficient, lower_logical_sends() will notice the EOT
if (depth == 0 &&
inst->dst.file == VGRF &&
- alloc.sizes[inst->dst.nr] == inst->regs_written &&
+ alloc.sizes[inst->dst.nr] * REG_SIZE == inst->size_written &&
!inst->is_partial_write()) {
if (remap[dst] == -1) {
remap[dst] = dst;
} else {
- remap[dst] = alloc.allocate(inst->regs_written);
+ remap[dst] = alloc.allocate(regs_written(inst));
inst->dst.nr = remap[dst];
progress = true;
}
return progress;
}
+/**
+ * Compute a bitmask with GRF granularity with a bit set for each GRF starting
+ * from \p r.offset which overlaps the region starting at \p s.offset and
+ * spanning \p ds bytes.
+ */
+static inline unsigned
+mask_relative_to(const fs_reg &r, const fs_reg &s, unsigned ds)
+{
+ const int rel_offset = reg_offset(s) - reg_offset(r);
+ const int shift = rel_offset / REG_SIZE;
+ const unsigned n = DIV_ROUND_UP(rel_offset % REG_SIZE + ds, REG_SIZE);
+ assert(reg_space(r) == reg_space(s) &&
+ shift >= 0 && shift < int(8 * sizeof(unsigned)));
+ return ((1 << n) - 1) << shift;
+}
+
bool
fs_visitor::compute_to_mrf()
{
inst->dst.type != inst->src[0].type ||
inst->src[0].abs || inst->src[0].negate ||
!inst->src[0].is_contiguous() ||
- inst->src[0].subreg_offset)
+ inst->src[0].offset % REG_SIZE != 0)
continue;
/* Can't compute-to-MRF this GRF if someone else was going to
if (this->virtual_grf_end[inst->src[0].nr] > ip)
continue;
- /* Found a move of a GRF to a MRF. Let's see if we can go
- * rewrite the thing that made this GRF to write into the MRF.
+ /* Found a move of a GRF to a MRF. Let's see if we can go rewrite the
+ * things that computed the value of all GRFs of the source region. The
+ * regs_left bitset keeps track of the registers we haven't yet found a
+ * generating instruction for.
*/
- bool found = false;
+ unsigned regs_left = (1 << regs_read(inst, 0)) - 1;
foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
- if (regions_overlap(scan_inst->dst, scan_inst->regs_written * REG_SIZE,
- inst->src[0], inst->regs_read(0) * REG_SIZE)) {
+ if (regions_overlap(scan_inst->dst, scan_inst->size_written,
+ inst->src[0], inst->size_read(0))) {
/* Found the last thing to write our reg we want to turn
* into a compute-to-MRF.
*/
* would need us to understand coalescing out more than one MOV at
* a time.
*/
- if (scan_inst->dst.reg_offset < inst->src[0].reg_offset ||
- scan_inst->dst.reg_offset + scan_inst->regs_written >
- inst->src[0].reg_offset + inst->regs_read(0))
+ if (!region_contained_in(scan_inst->dst, scan_inst->size_written,
+ inst->src[0], inst->size_read(0)))
break;
/* SEND instructions can't have MRF as a destination. */
}
}
- if (scan_inst->dst.reg_offset == inst->src[0].reg_offset)
- found = true;
-
- break;
+ /* Clear the bits for any registers this instruction overwrites. */
+ regs_left &= ~mask_relative_to(
+ inst->src[0], scan_inst->dst, scan_inst->size_written);
+ if (!regs_left)
+ break;
}
/* We don't handle control flow here. Most computation of
*/
bool interfered = false;
for (int i = 0; i < scan_inst->sources; i++) {
- if (regions_overlap(scan_inst->src[i], scan_inst->regs_read(i) * REG_SIZE,
- inst->src[0], inst->regs_read(0) * REG_SIZE)) {
+ if (regions_overlap(scan_inst->src[i], scan_inst->size_read(i),
+ inst->src[0], inst->size_read(0))) {
interfered = true;
}
}
if (interfered)
break;
- if (regions_overlap(scan_inst->dst, scan_inst->regs_written * REG_SIZE,
- inst->dst, inst->regs_written * REG_SIZE)) {
+ if (regions_overlap(scan_inst->dst, scan_inst->size_written,
+ inst->dst, inst->size_written)) {
/* If somebody else writes our MRF here, we can't
* compute-to-MRF before that.
*/
if (scan_inst->mlen > 0 && scan_inst->base_mrf != -1 &&
regions_overlap(fs_reg(MRF, scan_inst->base_mrf), scan_inst->mlen * REG_SIZE,
- inst->dst, inst->regs_written * REG_SIZE)) {
+ inst->dst, inst->size_written)) {
/* Found a SEND instruction, which means that there are
* live values in MRFs from base_mrf to base_mrf +
* scan_inst->mlen - 1. Don't go pushing our MRF write up
}
}
- if (!found)
+ if (regs_left)
continue;
- /* Found all generating instructions of our MRF's source value.
+ /* Found all generating instructions of our MRF's source value, so it
+ * should be safe to rewrite them to point to the MRF directly.
*/
+ regs_left = (1 << regs_read(inst, 0)) - 1;
+
foreach_inst_in_block_reverse_starting_from(fs_inst, scan_inst, inst) {
- if (regions_overlap(scan_inst->dst, scan_inst->regs_written * REG_SIZE,
- inst->src[0], inst->regs_read(0) * REG_SIZE)) {
+ if (regions_overlap(scan_inst->dst, scan_inst->size_written,
+ inst->src[0], inst->size_read(0))) {
+ /* Clear the bits for any registers this instruction overwrites. */
+ regs_left &= ~mask_relative_to(
+ inst->src[0], scan_inst->dst, scan_inst->size_written);
+
+ const unsigned rel_offset = reg_offset(scan_inst->dst) -
+ reg_offset(inst->src[0]);
+
+ if (inst->dst.nr & BRW_MRF_COMPR4) {
+ /* Apply the same address transformation done by the hardware
+ * for COMPR4 MRF writes.
+ */
+ assert(rel_offset < 2 * REG_SIZE);
+ scan_inst->dst.nr = inst->dst.nr + rel_offset / REG_SIZE * 4;
+
+ /* Clear the COMPR4 bit if the generating instruction is not
+ * compressed.
+ */
+ if (scan_inst->size_written < 2 * REG_SIZE)
+ scan_inst->dst.nr &= ~BRW_MRF_COMPR4;
+
+ } else {
+ /* Calculate the MRF number the result of this instruction is
+ * ultimately written to.
+ */
+ scan_inst->dst.nr = inst->dst.nr + rel_offset / REG_SIZE;
+ }
+
scan_inst->dst.file = MRF;
- scan_inst->dst.nr = inst->dst.nr;
- scan_inst->dst.reg_offset = 0;
+ scan_inst->dst.offset = inst->dst.offset + rel_offset % REG_SIZE;
scan_inst->saturate |= inst->saturate;
- break;
+ if (!regs_left)
+ break;
}
}
+ assert(!regs_left);
inst->remove(block);
progress = true;
}
bool progress = false;
unsigned depth = 0;
+ if (!brw_stage_has_packed_dispatch(devinfo, stage, stage_prog_data)) {
+ /* The optimization below assumes that channel zero is live on thread
+ * dispatch, which may not be the case if the fixed function dispatches
+ * threads sparsely.
+ */
+ return false;
+ }
+
foreach_block_and_inst_safe(block, fs_inst, inst, cfg) {
switch (inst->opcode) {
case BRW_OPCODE_IF:
}
/* Clear out any MRF move records whose sources got overwritten. */
- if (inst->dst.file == VGRF) {
- for (unsigned int i = 0; i < ARRAY_SIZE(last_mrf_move); i++) {
- if (last_mrf_move[i] &&
- last_mrf_move[i]->src[0].nr == inst->dst.nr) {
- last_mrf_move[i] = NULL;
- }
- }
+ for (unsigned i = 0; i < ARRAY_SIZE(last_mrf_move); i++) {
+ if (last_mrf_move[i] &&
+ regions_overlap(inst->dst, inst->size_written,
+ last_mrf_move[i]->src[0],
+ last_mrf_move[i]->size_read(0))) {
+ last_mrf_move[i] = NULL;
+ }
}
if (inst->opcode == BRW_OPCODE_MOV &&
inst->dst.file == MRF &&
- inst->src[0].file == VGRF &&
+ inst->src[0].file != ARF &&
!inst->is_partial_write()) {
last_mrf_move[inst->dst.nr] = inst;
}
fs_visitor::insert_gen4_pre_send_dependency_workarounds(bblock_t *block,
fs_inst *inst)
{
- int write_len = inst->regs_written;
+ int write_len = regs_written(inst);
int first_write_grf = inst->dst.nr;
bool needs_dep[BRW_MAX_MRF(devinfo->gen)];
assert(write_len < (int)sizeof(needs_dep) - 1);
* dependency has more latency than a MOV.
*/
if (scan_inst->dst.file == VGRF) {
- for (int i = 0; i < scan_inst->regs_written; i++) {
+ for (unsigned i = 0; i < regs_written(scan_inst); i++) {
int reg = scan_inst->dst.nr + i;
if (reg >= first_write_grf &&
void
fs_visitor::insert_gen4_post_send_dependency_workarounds(bblock_t *block, fs_inst *inst)
{
- int write_len = inst->regs_written;
+ int write_len = regs_written(inst);
int first_write_grf = inst->dst.nr;
bool needs_dep[BRW_MAX_MRF(devinfo->gen)];
assert(write_len < (int)sizeof(needs_dep) - 1);
bool progress = false;
- /* Note that we're done with register allocation, so GRF fs_regs always
- * have a .reg_offset of 0.
- */
-
foreach_block_and_inst(block, fs_inst, inst, cfg) {
if (inst->mlen != 0 && inst->dst.file == VGRF) {
insert_gen4_pre_send_dependency_workarounds(block, inst);
* mode. Reserve space for the register.
*/
offset = payload = fs_reg(VGRF, alloc.allocate(2));
- offset.reg_offset++;
+ offset.offset += REG_SIZE;
inst->mlen = 2;
} else {
offset = payload = fs_reg(VGRF, alloc.allocate(1));
*/
inst->opcode = FS_OPCODE_UNIFORM_PULL_CONSTANT_LOAD_GEN7;
inst->src[1] = payload;
- inst->base_mrf = -1;
invalidate_live_intervals();
} else {
ibld.MOV(imm, inst->src[1]);
ibld.MUL(inst->dst, imm, inst->src[0]);
} else {
- ibld.MUL(inst->dst, inst->src[0], inst->src[1]);
+ const bool ud = (inst->src[1].type == BRW_REGISTER_TYPE_UD);
+ ibld.MUL(inst->dst, inst->src[0],
+ ud ? brw_imm_uw(inst->src[1].ud)
+ : brw_imm_w(inst->src[1].d));
}
} else {
/* Gen < 8 (and some Gen8+ low-power parts like Cherryview) cannot
inst->dst.type);
if (devinfo->gen >= 7) {
- fs_reg src1_0_w = inst->src[1];
- fs_reg src1_1_w = inst->src[1];
-
if (inst->src[1].file == IMM) {
- src1_0_w.ud &= 0xffff;
- src1_1_w.ud >>= 16;
+ ibld.MUL(low, inst->src[0],
+ brw_imm_uw(inst->src[1].ud & 0xffff));
+ ibld.MUL(high, inst->src[0],
+ brw_imm_uw(inst->src[1].ud >> 16));
} else {
- src1_0_w.type = BRW_REGISTER_TYPE_UW;
- if (src1_0_w.stride != 0) {
- assert(src1_0_w.stride == 1);
- src1_0_w.stride = 2;
- }
-
- src1_1_w.type = BRW_REGISTER_TYPE_UW;
- if (src1_1_w.stride != 0) {
- assert(src1_1_w.stride == 1);
- src1_1_w.stride = 2;
- }
- src1_1_w.subreg_offset += type_sz(BRW_REGISTER_TYPE_UW);
+ ibld.MUL(low, inst->src[0],
+ subscript(inst->src[1], BRW_REGISTER_TYPE_UW, 0));
+ ibld.MUL(high, inst->src[0],
+ subscript(inst->src[1], BRW_REGISTER_TYPE_UW, 1));
}
- ibld.MUL(low, inst->src[0], src1_0_w);
- ibld.MUL(high, inst->src[0], src1_1_w);
} else {
- fs_reg src0_0_w = inst->src[0];
- fs_reg src0_1_w = inst->src[0];
-
- src0_0_w.type = BRW_REGISTER_TYPE_UW;
- if (src0_0_w.stride != 0) {
- assert(src0_0_w.stride == 1);
- src0_0_w.stride = 2;
- }
-
- src0_1_w.type = BRW_REGISTER_TYPE_UW;
- if (src0_1_w.stride != 0) {
- assert(src0_1_w.stride == 1);
- src0_1_w.stride = 2;
- }
- src0_1_w.subreg_offset += type_sz(BRW_REGISTER_TYPE_UW);
-
- ibld.MUL(low, src0_0_w, inst->src[1]);
- ibld.MUL(high, src0_1_w, inst->src[1]);
+ ibld.MUL(low, subscript(inst->src[0], BRW_REGISTER_TYPE_UW, 0),
+ inst->src[1]);
+ ibld.MUL(high, subscript(inst->src[0], BRW_REGISTER_TYPE_UW, 1),
+ inst->src[1]);
}
- fs_reg dst = inst->dst;
- dst.type = BRW_REGISTER_TYPE_UW;
- dst.subreg_offset = 2;
- dst.stride = 2;
-
- high.type = BRW_REGISTER_TYPE_UW;
- high.stride = 2;
-
- low.type = BRW_REGISTER_TYPE_UW;
- low.subreg_offset = 2;
- low.stride = 2;
-
- ibld.ADD(dst, low, high);
+ ibld.ADD(subscript(inst->dst, BRW_REGISTER_TYPE_UW, 1),
+ subscript(low, BRW_REGISTER_TYPE_UW, 1),
+ subscript(high, BRW_REGISTER_TYPE_UW, 0));
if (inst->conditional_mod || orig_dst.file == MRF) {
set_condmod(inst->conditional_mod,
} else if (inst->opcode == SHADER_OPCODE_MULH) {
/* Should have been lowered to 8-wide. */
- assert(inst->exec_size <= 8);
+ assert(inst->exec_size <= get_lowered_simd_width(devinfo, inst));
const fs_reg acc = retype(brw_acc_reg(inst->exec_size),
inst->dst.type);
fs_inst *mul = ibld.MUL(acc, inst->src[0], inst->src[1]);
static void
lower_fb_write_logical_send(const fs_builder &bld, fs_inst *inst,
- const brw_wm_prog_data *prog_data,
+ const struct brw_wm_prog_data *prog_data,
const brw_wm_prog_key *key,
const fs_visitor::thread_payload &payload)
{
assert(inst->src[FB_WRITE_LOGICAL_SRC_COMPONENTS].file == IMM);
- const brw_device_info *devinfo = bld.shader->devinfo;
+ const gen_device_info *devinfo = bld.shader->devinfo;
const fs_reg &color0 = inst->src[FB_WRITE_LOGICAL_SRC_COLOR0];
const fs_reg &color1 = inst->src[FB_WRITE_LOGICAL_SRC_COLOR1];
const fs_reg &src0_alpha = inst->src[FB_WRITE_LOGICAL_SRC_SRC0_ALPHA];
/* Send from the GRF */
fs_reg payload = fs_reg(VGRF, -1, BRW_REGISTER_TYPE_F);
load = bld.LOAD_PAYLOAD(payload, sources, length, payload_header_size);
- payload.nr = bld.shader->alloc.allocate(load->regs_written);
+ payload.nr = bld.shader->alloc.allocate(regs_written(load));
load->dst = payload;
inst->src[0] = payload;
inst->resize_sources(1);
- inst->base_mrf = -1;
} else {
/* Send from the MRF */
load = bld.LOAD_PAYLOAD(fs_reg(MRF, 1, BRW_REGISTER_TYPE_F),
}
inst->opcode = FS_OPCODE_FB_WRITE;
- inst->mlen = load->regs_written;
+ inst->mlen = regs_written(load);
inst->header_size = header_size;
}
+static void
+lower_fb_read_logical_send(const fs_builder &bld, fs_inst *inst)
+{
+ const fs_builder &ubld = bld.exec_all();
+ const unsigned length = 2;
+ const fs_reg header = ubld.group(8, 0).vgrf(BRW_REGISTER_TYPE_UD, length);
+
+ ubld.group(16, 0)
+ .MOV(header, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
+
+ inst->resize_sources(1);
+ inst->src[0] = header;
+ inst->opcode = FS_OPCODE_FB_READ;
+ inst->mlen = length;
+ inst->header_size = length;
+}
+
static void
lower_sampler_logical_send_gen4(const fs_builder &bld, fs_inst *inst, opcode op,
const fs_reg &coordinate,
}
static bool
-is_high_sampler(const struct brw_device_info *devinfo, const fs_reg &sampler)
+is_high_sampler(const struct gen_device_info *devinfo, const fs_reg &sampler)
{
if (devinfo->gen < 8 && !devinfo->is_haswell)
return false;
unsigned coord_components,
unsigned grad_components)
{
- const brw_device_info *devinfo = bld.shader->devinfo;
- int reg_width = bld.dispatch_width() / 8;
+ const gen_device_info *devinfo = bld.shader->devinfo;
+ unsigned reg_width = bld.dispatch_width() / 8;
unsigned header_size = 0, length = 0;
fs_reg sources[MAX_SAMPLER_MESSAGE_SIZE];
for (unsigned i = 0; i < ARRAY_SIZE(sources); i++)
* and we have an explicit header, we need to set up the sampler
* writemask. It's reversed from normal: 1 means "don't write".
*/
- if (!inst->eot && inst->regs_written != 4 * reg_width) {
- assert((inst->regs_written % reg_width) == 0);
- unsigned mask = ~((1 << (inst->regs_written / reg_width)) - 1) & 0xf;
+ if (!inst->eot && regs_written(inst) != 4 * reg_width) {
+ assert(regs_written(inst) % reg_width == 0);
+ unsigned mask = ~((1 << (regs_written(inst) / reg_width)) - 1) & 0xf;
inst->offset |= mask << 12;
}
}
bool coordinate_done = false;
- /* The sampler can only meaningfully compute LOD for fragment shader
- * messages. For all other stages, we change the opcode to TXL and
- * hardcode the LOD to 0.
- */
- if (bld.shader->stage != MESA_SHADER_FRAGMENT &&
- op == SHADER_OPCODE_TEX) {
- op = SHADER_OPCODE_TXL;
- lod = brw_imm_f(0.0f);
- }
-
/* Set up the LOD info */
switch (op) {
case FS_OPCODE_TXB:
if (coord_components >= 2) {
bld.MOV(retype(sources[length], BRW_REGISTER_TYPE_D),
offset(coordinate, bld, 1));
+ } else {
+ sources[length] = brw_imm_d(0);
}
length++;
}
coordinate_done = true;
break;
case SHADER_OPCODE_TG4_OFFSET:
- /* gather4_po_c should have been lowered in SIMD16 mode. */
- assert(bld.dispatch_width() == 8 || shadow_c.file == BAD_FILE);
-
/* More crazy intermixing */
for (unsigned i = 0; i < 2; i++) /* u, v */
bld.MOV(sources[length++], offset(coordinate, bld, i));
inst->src[1] = surface;
inst->src[2] = sampler;
inst->resize_sources(3);
- inst->base_mrf = -1;
inst->mlen = mlen;
inst->header_size = header_size;
static void
lower_sampler_logical_send(const fs_builder &bld, fs_inst *inst, opcode op)
{
- const brw_device_info *devinfo = bld.shader->devinfo;
+ const gen_device_info *devinfo = bld.shader->devinfo;
const fs_reg &coordinate = inst->src[TEX_LOGICAL_SRC_COORDINATE];
const fs_reg &shadow_c = inst->src[TEX_LOGICAL_SRC_SHADOW_C];
const fs_reg &lod = inst->src[TEX_LOGICAL_SRC_LOD];
static void
lower_varying_pull_constant_logical_send(const fs_builder &bld, fs_inst *inst)
{
- const brw_device_info *devinfo = bld.shader->devinfo;
+ const gen_device_info *devinfo = bld.shader->devinfo;
if (devinfo->gen >= 7) {
+ /* We are switching the instruction from an ALU-like instruction to a
+ * send-from-grf instruction. Since sends can't handle strides or
+ * source modifiers, we have to make a copy of the offset source.
+ */
+ fs_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_UD);
+ bld.MOV(tmp, inst->src[1]);
+ inst->src[1] = tmp;
+
inst->opcode = FS_OPCODE_VARYING_PULL_CONSTANT_LOAD_GEN7;
} else {
case FS_OPCODE_FB_WRITE_LOGICAL:
assert(stage == MESA_SHADER_FRAGMENT);
lower_fb_write_logical_send(ibld, inst,
- (const brw_wm_prog_data *)prog_data,
+ brw_wm_prog_data(prog_data),
(const brw_wm_prog_key *)key,
payload);
break;
+ case FS_OPCODE_FB_READ_LOGICAL:
+ lower_fb_read_logical_send(ibld, inst);
+ break;
+
case SHADER_OPCODE_TEX_LOGICAL:
lower_sampler_logical_send(ibld, inst, SHADER_OPCODE_TEX);
break;
* excessively restrictive.
*/
static unsigned
-get_fpu_lowered_simd_width(const struct brw_device_info *devinfo,
+get_fpu_lowered_simd_width(const struct gen_device_info *devinfo,
const fs_inst *inst)
{
/* Maximum execution size representable in the instruction controls. */
* which is the one that is going to limit the overall execution size of
* the instruction due to this rule.
*/
- unsigned reg_count = inst->regs_written;
+ unsigned reg_count = DIV_ROUND_UP(inst->size_written, REG_SIZE);
for (unsigned i = 0; i < inst->sources; i++)
- reg_count = MAX2(reg_count, (unsigned)inst->regs_read(i));
+ reg_count = MAX2(reg_count, DIV_ROUND_UP(inst->size_read(i), REG_SIZE));
/* Calculate the maximum execution size of the instruction based on the
* factor by which it goes over the hardware limit of 2 GRFs.
*/
if (devinfo->gen < 8) {
for (unsigned i = 0; i < inst->sources; i++) {
- if (inst->regs_written == 2 &&
- inst->regs_read(i) != 0 && inst->regs_read(i) != 2 &&
+ if (inst->size_written > REG_SIZE &&
+ inst->size_read(i) != 0 && inst->size_read(i) <= REG_SIZE &&
!is_uniform(inst->src[i]) &&
!(type_sz(inst->dst.type) == 4 && inst->dst.stride == 1 &&
- type_sz(inst->src[i].type) == 2 && inst->src[i].stride == 1))
- max_width = MIN2(max_width, inst->exec_size /
- inst->regs_written);
+ type_sz(inst->src[i].type) == 2 && inst->src[i].stride == 1)) {
+ const unsigned reg_count = DIV_ROUND_UP(inst->size_written, REG_SIZE);
+ max_width = MIN2(max_width, inst->exec_size / reg_count);
+ }
}
}
max_width = MIN2(max_width, 16);
/* From the IVB PRMs (applies to other devices that don't have the
- * brw_device_info::supports_simd16_3src flag set):
+ * gen_device_info::supports_simd16_3src flag set):
* "In Align16 access mode, SIMD16 is not allowed for DW operations and
* SIMD8 is not allowed for DF operations."
*/
if (inst->is_3src(devinfo) && !devinfo->supports_simd16_3src)
max_width = MIN2(max_width, inst->exec_size / reg_count);
+ /* Pre-Gen8 EUs are hardwired to use the QtrCtrl+1 (where QtrCtrl is
+ * the 8-bit quarter of the execution mask signals specified in the
+ * instruction control fields) for the second compressed half of any
+ * single-precision instruction (for double-precision instructions
+ * it's hardwired to use NibCtrl+1, at least on HSW), which means that
+ * the EU will apply the wrong execution controls for the second
+ * sequential GRF write if the number of channels per GRF is not exactly
+ * eight in single-precision mode (or four in double-float mode).
+ *
+ * In this situation we calculate the maximum size of the split
+ * instructions so they only ever write to a single register.
+ */
+ if (devinfo->gen < 8 && inst->size_written > REG_SIZE &&
+ !inst->force_writemask_all) {
+ const unsigned channels_per_grf = inst->exec_size /
+ DIV_ROUND_UP(inst->size_written, REG_SIZE);
+ unsigned exec_type_size = 0;
+ for (int i = 0; i < inst->sources; i++) {
+ if (inst->src[i].file != BAD_FILE)
+ exec_type_size = MAX2(exec_type_size, type_sz(inst->src[i].type));
+ }
+ assert(exec_type_size);
+
+ /* The hardware shifts exactly 8 channels per compressed half of the
+ * instruction in single-precision mode and exactly 4 in double-precision.
+ */
+ if (channels_per_grf != (exec_type_size == 8 ? 4 : 8))
+ max_width = MIN2(max_width, channels_per_grf);
+ }
+
/* Only power-of-two execution sizes are representable in the instruction
* control fields.
*/
return 1 << _mesa_logbase2(max_width);
}
+/**
+ * Get the maximum allowed SIMD width for instruction \p inst accounting for
+ * various payload size restrictions that apply to sampler message
+ * instructions.
+ *
+ * This is only intended to provide a maximum theoretical bound for the
+ * execution size of the message based on the number of argument components
+ * alone, which in most cases will determine whether the SIMD8 or SIMD16
+ * variant of the message can be used, though some messages may have
+ * additional restrictions not accounted for here (e.g. pre-ILK hardware uses
+ * the message length to determine the exact SIMD width and argument count,
+ * which makes a number of sampler message combinations impossible to
+ * represent).
+ */
+static unsigned
+get_sampler_lowered_simd_width(const struct gen_device_info *devinfo,
+ const fs_inst *inst)
+{
+ /* Calculate the number of coordinate components that have to be present
+ * assuming that additional arguments follow the texel coordinates in the
+ * message payload. On IVB+ there is no need for padding, on ILK-SNB we
+ * need to pad to four or three components depending on the message,
+ * pre-ILK we need to pad to at most three components.
+ */
+ const unsigned req_coord_components =
+ (devinfo->gen >= 7 ||
+ !inst->components_read(TEX_LOGICAL_SRC_COORDINATE)) ? 0 :
+ (devinfo->gen >= 5 && inst->opcode != SHADER_OPCODE_TXF_LOGICAL &&
+ inst->opcode != SHADER_OPCODE_TXF_CMS_LOGICAL) ? 4 :
+ 3;
+
+ /* On Gen9+ the LOD argument is for free if we're able to use the LZ
+ * variant of the TXL or TXF message.
+ */
+ const bool implicit_lod = devinfo->gen >= 9 &&
+ (inst->opcode == SHADER_OPCODE_TXL ||
+ inst->opcode == SHADER_OPCODE_TXF) &&
+ inst->src[TEX_LOGICAL_SRC_LOD].is_zero();
+
+ /* Calculate the total number of argument components that need to be passed
+ * to the sampler unit.
+ */
+ const unsigned num_payload_components =
+ MAX2(inst->components_read(TEX_LOGICAL_SRC_COORDINATE),
+ req_coord_components) +
+ inst->components_read(TEX_LOGICAL_SRC_SHADOW_C) +
+ (implicit_lod ? 0 : inst->components_read(TEX_LOGICAL_SRC_LOD)) +
+ inst->components_read(TEX_LOGICAL_SRC_LOD2) +
+ inst->components_read(TEX_LOGICAL_SRC_SAMPLE_INDEX) +
+ (inst->opcode == SHADER_OPCODE_TG4_OFFSET_LOGICAL ?
+ inst->components_read(TEX_LOGICAL_SRC_OFFSET_VALUE) : 0) +
+ inst->components_read(TEX_LOGICAL_SRC_MCS);
+
+ /* SIMD16 messages with more than five arguments exceed the maximum message
+ * size supported by the sampler, regardless of whether a header is
+ * provided or not.
+ */
+ return MIN2(inst->exec_size,
+ num_payload_components > MAX_SAMPLER_MESSAGE_SIZE / 2 ? 8 : 16);
+}
+
/**
* Get the closest native SIMD width supported by the hardware for instruction
* \p inst. The instruction will be left untouched by
* original execution size.
*/
static unsigned
-get_lowered_simd_width(const struct brw_device_info *devinfo,
+get_lowered_simd_width(const struct gen_device_info *devinfo,
const fs_inst *inst)
{
switch (inst->opcode) {
case FS_OPCODE_PACK_HALF_2x16_SPLIT:
case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_X:
case FS_OPCODE_UNPACK_HALF_2x16_SPLIT_Y:
- case FS_OPCODE_INTERPOLATE_AT_CENTROID:
case FS_OPCODE_INTERPOLATE_AT_SAMPLE:
case FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET:
case FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET:
return (inst->src[FB_WRITE_LOGICAL_SRC_COLOR1].file != BAD_FILE ?
8 : MIN2(16, inst->exec_size));
+ case FS_OPCODE_FB_READ_LOGICAL:
+ return MIN2(16, inst->exec_size);
+
case SHADER_OPCODE_TEX_LOGICAL:
case SHADER_OPCODE_TXF_CMS_LOGICAL:
case SHADER_OPCODE_TXF_UMS_LOGICAL:
case SHADER_OPCODE_LOD_LOGICAL:
case SHADER_OPCODE_TG4_LOGICAL:
case SHADER_OPCODE_SAMPLEINFO_LOGICAL:
- return MIN2(16, inst->exec_size);
+ case SHADER_OPCODE_TXF_CMS_W_LOGICAL:
+ case SHADER_OPCODE_TG4_OFFSET_LOGICAL:
+ return get_sampler_lowered_simd_width(devinfo, inst);
case SHADER_OPCODE_TXD_LOGICAL:
/* TXD is unsupported in SIMD16 mode. */
return 8;
- case SHADER_OPCODE_TG4_OFFSET_LOGICAL: {
- /* gather4_po_c is unsupported in SIMD16 mode. */
- const fs_reg &shadow_c = inst->src[TEX_LOGICAL_SRC_SHADOW_C];
- return (shadow_c.file != BAD_FILE ? 8 : MIN2(16, inst->exec_size));
- }
case SHADER_OPCODE_TXL_LOGICAL:
- case FS_OPCODE_TXB_LOGICAL: {
- /* Gen4 doesn't have SIMD8 non-shadow-compare bias/LOD instructions, and
- * Gen4-6 can't support TXL and TXB with shadow comparison in SIMD16
- * mode because the message exceeds the maximum length of 11.
+ case FS_OPCODE_TXB_LOGICAL:
+ /* Only one execution size is representable pre-ILK depending on whether
+ * the shadow reference argument is present.
*/
- const fs_reg &shadow_c = inst->src[TEX_LOGICAL_SRC_SHADOW_C];
- if (devinfo->gen == 4 && shadow_c.file == BAD_FILE)
- return 16;
- else if (devinfo->gen < 7 && shadow_c.file != BAD_FILE)
- return 8;
+ if (devinfo->gen == 4)
+ return inst->src[TEX_LOGICAL_SRC_SHADOW_C].file == BAD_FILE ? 16 : 8;
else
- return MIN2(16, inst->exec_size);
- }
+ return get_sampler_lowered_simd_width(devinfo, inst);
+
case SHADER_OPCODE_TXF_LOGICAL:
case SHADER_OPCODE_TXS_LOGICAL:
/* Gen4 doesn't have SIMD8 variants for the RESINFO and LD-with-LOD
if (devinfo->gen == 4)
return 16;
else
- return MIN2(16, inst->exec_size);
-
- case SHADER_OPCODE_TXF_CMS_W_LOGICAL: {
- /* This opcode can take up to 6 arguments which means that in some
- * circumstances it can end up with a message that is too long in SIMD16
- * mode.
- */
- const unsigned coord_components =
- inst->src[TEX_LOGICAL_SRC_COORD_COMPONENTS].ud;
- /* First three arguments are the sample index and the two arguments for
- * the MCS data.
- */
- if ((coord_components + 3) * 2 > MAX_SAMPLER_MESSAGE_SIZE)
- return 8;
- else
- return MIN2(16, inst->exec_size);
- }
+ return get_sampler_lowered_simd_width(devinfo, inst);
case SHADER_OPCODE_TYPED_ATOMIC_LOGICAL:
case SHADER_OPCODE_TYPED_SURFACE_READ_LOGICAL:
}
}
+/**
+ * Return true if splitting out the group of channels of instruction \p inst
+ * given by lbld.group() requires allocating a temporary for the destination
+ * of the lowered instruction and copying the data back to the original
+ * destination region.
+ */
+static inline bool
+needs_dst_copy(const fs_builder &lbld, const fs_inst *inst)
+{
+ /* If the instruction writes more than one component we'll have to shuffle
+ * the results of multiple lowered instructions in order to make sure that
+ * they end up arranged correctly in the original destination region.
+ */
+ if (inst->size_written > inst->dst.component_size(inst->exec_size))
+ return true;
+
+ /* If the lowered execution size is larger than the original the result of
+ * the instruction won't fit in the original destination, so we'll have to
+ * allocate a temporary in any case.
+ */
+ if (lbld.dispatch_width() > inst->exec_size)
+ return true;
+
+ for (unsigned i = 0; i < inst->sources; i++) {
+ /* If we already made a copy of the source for other reasons there won't
+ * be any overlap with the destination.
+ */
+ if (needs_src_copy(lbld, inst, i))
+ continue;
+
+ /* In order to keep the logic simple we emit a copy whenever the
+ * destination region doesn't exactly match an overlapping source, which
+ * may point at the source and destination not being aligned group by
+ * group which could cause one of the lowered instructions to overwrite
+ * the data read from the same source by other lowered instructions.
+ */
+ if (regions_overlap(inst->dst, inst->size_written,
+ inst->src[i], inst->size_read(i)) &&
+ !inst->dst.equals(inst->src[i]))
+ return true;
+ }
+
+ return false;
+}
+
/**
* Insert data from a packed temporary into the channel group given by
* lbld.group() of the destination region of instruction \p inst and return
/* Specified channel group from the destination region. */
const fs_reg dst = horiz_offset(inst->dst, lbld.group());
- const unsigned dst_size = inst->regs_written * REG_SIZE /
- inst->dst.component_size(inst->exec_size);
- const fs_reg tmp = lbld.vgrf(inst->dst.type, dst_size);
+ const unsigned dst_size = inst->size_written /
+ inst->dst.component_size(inst->exec_size);
+
+ if (needs_dst_copy(lbld, inst)) {
+ const fs_reg tmp = lbld.vgrf(inst->dst.type, dst_size);
+
+ if (inst->predicate) {
+ /* Handle predication by copying the original contents of
+ * the destination into the temporary before emitting the
+ * lowered instruction.
+ */
+ for (unsigned k = 0; k < dst_size; ++k)
+ cbld.at(block, inst)
+ .MOV(offset(tmp, lbld, k), offset(dst, inst->exec_size, k));
+ }
- if (inst->predicate) {
- /* Handle predication by copying the original contents of the
- * destination into the temporary before emitting the lowered
- * instruction.
- */
for (unsigned k = 0; k < dst_size; ++k)
- cbld.at(block, inst)
- .MOV(offset(tmp, lbld, k), offset(dst, inst->exec_size, k));
- }
+ cbld.at(block, inst->next)
+ .MOV(offset(dst, inst->exec_size, k), offset(tmp, lbld, k));
- for (unsigned k = 0; k < dst_size; ++k)
- cbld.at(block, inst->next)
- .MOV(offset(dst, inst->exec_size, k), offset(tmp, lbld, k));
+ return tmp;
- return tmp;
+ } else {
+ /* No need to allocate a temporary for the lowered instruction, just
+ * take the right group of channels from the original region.
+ */
+ return dst;
+ }
}
bool
* original or the lowered instruction, whichever is lower.
*/
const unsigned n = DIV_ROUND_UP(inst->exec_size, lower_width);
- const unsigned dst_size = inst->regs_written * REG_SIZE /
+ const unsigned dst_size = inst->size_written /
inst->dst.component_size(inst->exec_size);
assert(!inst->writes_accumulator && !inst->mlen);
split_inst.src[j] = emit_unzip(lbld, block, inst, j);
split_inst.dst = emit_zip(lbld, block, inst);
- split_inst.regs_written =
- DIV_ROUND_UP(type_sz(inst->dst.type) * dst_size * lower_width,
- REG_SIZE);
+ split_inst.size_written =
+ split_inst.dst.component_size(lower_width) * dst_size;
lbld.emit(split_inst);
}
fprintf(file, "(mlen: %d) ", inst->mlen);
}
+ if (inst->eot) {
+ fprintf(file, "(EOT) ");
+ }
+
switch (inst->dst.file) {
case VGRF:
fprintf(file, "vgrf%d", inst->dst.nr);
- if (alloc.sizes[inst->dst.nr] != inst->regs_written ||
- inst->dst.subreg_offset)
- fprintf(file, "+%d.%d",
- inst->dst.reg_offset, inst->dst.subreg_offset);
break;
case FIXED_GRF:
fprintf(file, "g%d", inst->dst.nr);
fprintf(file, "(null)");
break;
case UNIFORM:
- fprintf(file, "***u%d***", inst->dst.nr + inst->dst.reg_offset);
+ fprintf(file, "***u%d***", inst->dst.nr);
break;
case ATTR:
- fprintf(file, "***attr%d***", inst->dst.nr + inst->dst.reg_offset);
+ fprintf(file, "***attr%d***", inst->dst.nr);
break;
case ARF:
switch (inst->dst.nr) {
fprintf(file, "arf%d.%d", inst->dst.nr & 0xf, inst->dst.subnr);
break;
}
- if (inst->dst.subnr)
- fprintf(file, "+%d", inst->dst.subnr);
break;
case IMM:
unreachable("not reached");
}
+
+ if (inst->dst.offset ||
+ (inst->dst.file == VGRF &&
+ alloc.sizes[inst->dst.nr] * REG_SIZE != inst->size_written)) {
+ const unsigned reg_size = (inst->dst.file == UNIFORM ? 4 : REG_SIZE);
+ fprintf(file, "+%d.%d", inst->dst.offset / reg_size,
+ inst->dst.offset % reg_size);
+ }
+
if (inst->dst.stride != 1)
fprintf(file, "<%u>", inst->dst.stride);
fprintf(file, ":%s, ", brw_reg_type_letters(inst->dst.type));
switch (inst->src[i].file) {
case VGRF:
fprintf(file, "vgrf%d", inst->src[i].nr);
- if (alloc.sizes[inst->src[i].nr] != (unsigned)inst->regs_read(i) ||
- inst->src[i].subreg_offset)
- fprintf(file, "+%d.%d", inst->src[i].reg_offset,
- inst->src[i].subreg_offset);
break;
case FIXED_GRF:
fprintf(file, "g%d", inst->src[i].nr);
fprintf(file, "***m%d***", inst->src[i].nr);
break;
case ATTR:
- fprintf(file, "attr%d+%d", inst->src[i].nr, inst->src[i].reg_offset);
+ fprintf(file, "attr%d", inst->src[i].nr);
break;
case UNIFORM:
- fprintf(file, "u%d", inst->src[i].nr + inst->src[i].reg_offset);
- if (inst->src[i].subreg_offset) {
- fprintf(file, "+%d.%d", inst->src[i].reg_offset,
- inst->src[i].subreg_offset);
- }
+ fprintf(file, "u%d", inst->src[i].nr);
break;
case BAD_FILE:
fprintf(file, "(null)");
fprintf(file, "arf%d.%d", inst->src[i].nr & 0xf, inst->src[i].subnr);
break;
}
- if (inst->src[i].subnr)
- fprintf(file, "+%d", inst->src[i].subnr);
break;
}
+
+ if (inst->src[i].offset ||
+ (inst->src[i].file == VGRF &&
+ alloc.sizes[inst->src[i].nr] * REG_SIZE != inst->size_read(i))) {
+ const unsigned reg_size = (inst->src[i].file == UNIFORM ? 4 : REG_SIZE);
+ fprintf(file, "+%d.%d", inst->src[i].offset / reg_size,
+ inst->src[i].offset % reg_size);
+ }
+
if (inst->src[i].abs)
fprintf(file, "|");
fs_visitor::setup_fs_payload_gen6()
{
assert(stage == MESA_SHADER_FRAGMENT);
- brw_wm_prog_data *prog_data = (brw_wm_prog_data*) this->prog_data;
-
- unsigned barycentric_interp_modes =
- (stage == MESA_SHADER_FRAGMENT) ?
- ((brw_wm_prog_data*) this->prog_data)->barycentric_interp_modes : 0;
+ struct brw_wm_prog_data *prog_data = brw_wm_prog_data(this->prog_data);
assert(devinfo->gen >= 6);
/* R2: only for 32-pixel dispatch.*/
/* R3-26: barycentric interpolation coordinates. These appear in the
- * same order that they appear in the brw_wm_barycentric_interp_mode
+ * same order that they appear in the brw_barycentric_mode
* enum. Each set of coordinates occupies 2 registers if dispatch width
* == 8 and 4 registers if dispatch width == 16. Coordinates only
* appear if they were enabled using the "Barycentric Interpolation
* Mode" bits in WM_STATE.
*/
- for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
- if (barycentric_interp_modes & (1 << i)) {
+ for (int i = 0; i < BRW_BARYCENTRIC_MODE_COUNT; ++i) {
+ if (prog_data->barycentric_interp_modes & (1 << i)) {
payload.barycentric_coord_reg[i] = payload.num_regs;
payload.num_regs += 2;
if (dispatch_width == 16) {
/* R27: interpolated depth if uses source depth */
prog_data->uses_src_depth =
- (nir->info.inputs_read & (1 << VARYING_SLOT_POS)) != 0;
+ (nir->info->inputs_read & (1 << VARYING_SLOT_POS)) != 0;
if (prog_data->uses_src_depth) {
payload.source_depth_reg = payload.num_regs;
payload.num_regs++;
/* R29: interpolated W set if GEN6_WM_USES_SOURCE_W. */
prog_data->uses_src_w =
- (nir->info.inputs_read & (1 << VARYING_SLOT_POS)) != 0;
+ (nir->info->inputs_read & (1 << VARYING_SLOT_POS)) != 0;
if (prog_data->uses_src_w) {
payload.source_w_reg = payload.num_regs;
payload.num_regs++;
/* R31: MSAA position offsets. */
if (prog_data->persample_dispatch &&
- (nir->info.system_values_read & SYSTEM_BIT_SAMPLE_POS)) {
+ (nir->info->system_values_read & SYSTEM_BIT_SAMPLE_POS)) {
/* From the Ivy Bridge PRM documentation for 3DSTATE_PS:
*
* "MSDISPMODE_PERSAMPLE is required in order to select
/* R32: MSAA input coverage mask */
prog_data->uses_sample_mask =
- (nir->info.system_values_read & SYSTEM_BIT_SAMPLE_MASK_IN) != 0;
+ (nir->info->system_values_read & SYSTEM_BIT_SAMPLE_MASK_IN) != 0;
if (prog_data->uses_sample_mask) {
assert(devinfo->gen >= 7);
payload.sample_mask_in_reg = payload.num_regs;
/* R34-: bary for 32-pixel. */
/* R58-59: interp W for 32-pixel. */
- if (nir->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
+ if (nir->info->outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
source_depth_to_render_target = true;
}
}
payload.num_regs = 2;
}
-/**
- * We are building the local ID push constant data using the simplest possible
- * method. We simply push the local IDs directly as they should appear in the
- * registers for the uvec3 gl_LocalInvocationID variable.
- *
- * Therefore, for SIMD8, we use 3 full registers, and for SIMD16 we use 6
- * registers worth of push constant space.
- *
- * Note: Any updates to brw_cs_prog_local_id_payload_dwords,
- * fill_local_id_payload or fs_visitor::emit_cs_local_invocation_id_setup need
- * to coordinated.
- *
- * FINISHME: There are a few easy optimizations to consider.
- *
- * 1. If gl_WorkGroupSize x, y or z is 1, we can just use zero, and there is
- * no need for using push constant space for that dimension.
- *
- * 2. Since GL_MAX_COMPUTE_WORK_GROUP_SIZE is currently 1024 or less, we can
- * easily use 16-bit words rather than 32-bit dwords in the push constant
- * data.
- *
- * 3. If gl_WorkGroupSize x, y or z is small, then we can use bytes for
- * conveying the data, and thereby reduce push constant usage.
- *
- */
void
fs_visitor::setup_gs_payload()
{
assert(stage == MESA_SHADER_GEOMETRY);
- struct brw_gs_prog_data *gs_prog_data =
- (struct brw_gs_prog_data *) prog_data;
- struct brw_vue_prog_data *vue_prog_data =
- (struct brw_vue_prog_data *) prog_data;
+ struct brw_gs_prog_data *gs_prog_data = brw_gs_prog_data(prog_data);
+ struct brw_vue_prog_data *vue_prog_data = brw_vue_prog_data(prog_data);
/* R0: thread header, R1: output URB handles */
payload.num_regs = 2;
* Note that the GS reads <URB Read Length> HWords for every vertex - so we
* have to multiply by VerticesIn to obtain the total storage requirement.
*/
- if (8 * vue_prog_data->urb_read_length * nir->info.gs.vertices_in >
- max_push_components) {
+ if (8 * vue_prog_data->urb_read_length * nir->info->gs.vertices_in >
+ max_push_components || gs_prog_data->invocations > 1) {
gs_prog_data->base.include_vue_handles = true;
/* R3..RN: ICP Handles for each incoming vertex (when using pull model) */
- payload.num_regs += nir->info.gs.vertices_in;
+ payload.num_regs += nir->info->gs.vertices_in;
vue_prog_data->urb_read_length =
- ROUND_DOWN_TO(max_push_components / nir->info.gs.vertices_in, 8) / 8;
+ ROUND_DOWN_TO(max_push_components / nir->info->gs.vertices_in, 8) / 8;
}
}
fs_visitor::setup_cs_payload()
{
assert(devinfo->gen >= 7);
- brw_cs_prog_data *prog_data = (brw_cs_prog_data*) this->prog_data;
-
payload.num_regs = 1;
-
- if (nir->info.system_values_read & SYSTEM_BIT_LOCAL_INVOCATION_ID) {
- prog_data->local_invocation_id_regs = dispatch_width * 3 / 8;
- payload.local_invocation_id_reg = payload.num_regs;
- payload.num_regs += prog_data->local_invocation_id_regs;
- }
}
void
if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER) && this_progress) { \
char filename[64]; \
snprintf(filename, 64, "%s%d-%s-%02d-%02d-" #pass, \
- stage_abbrev, dispatch_width, nir->info.name, iteration, pass_num); \
+ stage_abbrev, dispatch_width, nir->info->name, iteration, pass_num); \
\
backend_shader::dump_instructions(filename); \
} \
if (unlikely(INTEL_DEBUG & DEBUG_OPTIMIZER)) {
char filename[64];
snprintf(filename, 64, "%s%d-%s-00-00-start",
- stage_abbrev, dispatch_width, nir->info.name);
+ stage_abbrev, dispatch_width, nir->info->name);
backend_shader::dump_instructions(filename);
}
progress = false;
pass_num = 0;
+ if (OPT(lower_pack)) {
+ OPT(register_coalesce);
+ OPT(dead_code_eliminate);
+ }
+
+ if (OPT(lower_d2x)) {
+ OPT(opt_copy_propagate);
+ OPT(dead_code_eliminate);
+ }
+
OPT(lower_simd_width);
/* After SIMD lowering just in case we had to unroll the EOT send. */
OPT(dead_code_eliminate);
}
- if (OPT(lower_pack)) {
- OPT(register_coalesce);
- OPT(dead_code_eliminate);
- }
-
- if (OPT(lower_d2x)) {
- OPT(opt_copy_propagate);
- OPT(dead_code_eliminate);
- }
-
OPT(opt_combine_constants);
OPT(lower_integer_multiplication);
}
if (!allocated_without_spills) {
+ if (!allow_spilling)
+ fail("Failure to register allocate and spilling is not allowed.");
+
/* We assume that any spilling is worse than just dropping back to
* SIMD8. There's probably actually some intermediate point where
* SIMD16 with a couple of spills is still better.
}
}
- assert(last_scratch == 0 || allow_spilling);
-
/* This must come after all optimization and register allocation, since
* it inserts dead code that happens to have side effects, and it does
* so based on the actual physical registers in use.
schedule_instructions(SCHEDULE_POST);
- if (last_scratch > 0)
+ if (last_scratch > 0) {
+ MAYBE_UNUSED unsigned max_scratch_size = 2 * 1024 * 1024;
+
prog_data->total_scratch = brw_get_scratch_size(last_scratch);
+
+ if (stage == MESA_SHADER_COMPUTE) {
+ if (devinfo->is_haswell) {
+ /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
+ * field documentation, Haswell supports a minimum of 2kB of
+ * scratch space for compute shaders, unlike every other stage
+ * and platform.
+ */
+ prog_data->total_scratch = MAX2(prog_data->total_scratch, 2048);
+ } else if (devinfo->gen <= 7) {
+ /* According to the MEDIA_VFE_STATE's "Per Thread Scratch Space"
+ * field documentation, platforms prior to Haswell measure scratch
+ * size linearly with a range of [1kB, 12kB] and 1kB granularity.
+ */
+ prog_data->total_scratch = ALIGN(last_scratch, 1024);
+ max_scratch_size = 12 * 1024;
+ }
+ }
+
+ /* We currently only support up to 2MB of scratch space. If we
+ * need to support more eventually, the documentation suggests
+ * that we could allocate a larger buffer, and partition it out
+ * ourselves. We'd just have to undo the hardware's address
+ * calculation by subtracting (FFTID * Per Thread Scratch Space)
+ * and then add FFTID * (Larger Per Thread Scratch Space).
+ *
+ * See 3D-Media-GPGPU Engine > Media GPGPU Pipeline >
+ * Thread Group Tracking > Local Memory/Scratch Space.
+ */
+ assert(prog_data->total_scratch < max_scratch_size);
+ }
}
bool
{
assert(stage == MESA_SHADER_TESS_CTRL);
- struct brw_tcs_prog_data *tcs_prog_data =
- (struct brw_tcs_prog_data *) prog_data;
+ struct brw_tcs_prog_data *tcs_prog_data = brw_tcs_prog_data(prog_data);
/* r1-r4 contain the ICP handles. */
payload.num_regs = 5;
}
/* Fix the disptach mask */
- if (nir->info.tcs.vertices_out % 8) {
+ if (nir->info->tcs.vertices_out % 8) {
bld.CMP(bld.null_reg_ud(), invocation_id,
- brw_imm_ud(nir->info.tcs.vertices_out), BRW_CONDITIONAL_L);
+ brw_imm_ud(nir->info->tcs.vertices_out), BRW_CONDITIONAL_L);
bld.IF(BRW_PREDICATE_NORMAL);
}
emit_nir_code();
- if (nir->info.tcs.vertices_out % 8) {
+ if (nir->info->tcs.vertices_out % 8) {
bld.emit(BRW_OPCODE_ENDIF);
}
fs_inst *inst = bld.emit(SHADER_OPCODE_URB_WRITE_SIMD8_MASKED,
bld.null_reg_ud(), payload);
inst->mlen = 3;
- inst->base_mrf = -1;
inst->eot = true;
if (shader_time_index >= 0)
bool
fs_visitor::run_fs(bool allow_spilling, bool do_rep_send)
{
- brw_wm_prog_data *wm_prog_data = (brw_wm_prog_data *) this->prog_data;
+ struct brw_wm_prog_data *wm_prog_data = brw_wm_prog_data(this->prog_data);
brw_wm_prog_key *wm_key = (brw_wm_prog_key *) this->key;
assert(stage == MESA_SHADER_FRAGMENT);
emit_shader_time_begin();
calculate_urb_setup();
- if (nir->info.inputs_read > 0) {
+ if (nir->info->inputs_read > 0 ||
+ (nir->info->outputs_read > 0 && !wm_key->coherent_fb_fetch)) {
if (devinfo->gen < 6)
emit_interpolation_setup_gen4();
else
if (devinfo->is_haswell && prog_data->total_shared > 0) {
/* Move SLM index from g0.0[27:24] to sr0.1[11:8] */
const fs_builder abld = bld.exec_all().group(1, 0);
- abld.MOV(retype(suboffset(brw_sr0_reg(), 1), BRW_REGISTER_TYPE_UW),
+ abld.MOV(retype(brw_sr0_reg(1), BRW_REGISTER_TYPE_UW),
suboffset(retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UW), 1));
}
/**
* Return a bitfield where bit n is set if barycentric interpolation mode n
- * (see enum brw_wm_barycentric_interp_mode) is needed by the fragment shader.
+ * (see enum brw_barycentric_mode) is needed by the fragment shader.
+ *
+ * We examine the load_barycentric intrinsics rather than looking at input
+ * variables so that we catch interpolateAtCentroid() messages too, which
+ * also need the BRW_BARYCENTRIC_[NON]PERSPECTIVE_CENTROID mode set up.
*/
static unsigned
-brw_compute_barycentric_interp_modes(const struct brw_device_info *devinfo,
- bool shade_model_flat,
- bool persample_shading,
+brw_compute_barycentric_interp_modes(const struct gen_device_info *devinfo,
const nir_shader *shader)
{
unsigned barycentric_interp_modes = 0;
- nir_foreach_variable(var, &shader->inputs) {
- enum glsl_interp_qualifier interp_qualifier =
- (enum glsl_interp_qualifier)var->data.interpolation;
- bool is_centroid = var->data.centroid && !persample_shading;
- bool is_sample = var->data.sample || persample_shading;
- bool is_gl_Color = (var->data.location == VARYING_SLOT_COL0) ||
- (var->data.location == VARYING_SLOT_COL1);
-
- /* Ignore WPOS and FACE, because they don't require interpolation. */
- if (var->data.location == VARYING_SLOT_POS ||
- var->data.location == VARYING_SLOT_FACE)
+ nir_foreach_function(f, shader) {
+ if (!f->impl)
continue;
- /* Determine the set (or sets) of barycentric coordinates needed to
- * interpolate this variable. Note that when
- * brw->needs_unlit_centroid_workaround is set, centroid interpolation
- * uses PIXEL interpolation for unlit pixels and CENTROID interpolation
- * for lit pixels, so we need both sets of barycentric coordinates.
- */
- if (interp_qualifier == INTERP_QUALIFIER_NOPERSPECTIVE) {
- if (is_centroid) {
- barycentric_interp_modes |=
- 1 << BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC;
- } else if (is_sample) {
- barycentric_interp_modes |=
- 1 << BRW_WM_NONPERSPECTIVE_SAMPLE_BARYCENTRIC;
- }
- if ((!is_centroid && !is_sample) ||
- devinfo->needs_unlit_centroid_workaround) {
- barycentric_interp_modes |=
- 1 << BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC;
- }
- } else if (interp_qualifier == INTERP_QUALIFIER_SMOOTH ||
- (!(shade_model_flat && is_gl_Color) &&
- interp_qualifier == INTERP_QUALIFIER_NONE)) {
- if (is_centroid) {
- barycentric_interp_modes |=
- 1 << BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC;
- } else if (is_sample) {
- barycentric_interp_modes |=
- 1 << BRW_WM_PERSPECTIVE_SAMPLE_BARYCENTRIC;
- }
- if ((!is_centroid && !is_sample) ||
- devinfo->needs_unlit_centroid_workaround) {
- barycentric_interp_modes |=
- 1 << BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC;
+ nir_foreach_block(block, f->impl) {
+ nir_foreach_instr(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ if (intrin->intrinsic != nir_intrinsic_load_interpolated_input)
+ continue;
+
+ /* Ignore WPOS; it doesn't require interpolation. */
+ if (nir_intrinsic_base(intrin) == VARYING_SLOT_POS)
+ continue;
+
+ intrin = nir_instr_as_intrinsic(intrin->src[0].ssa->parent_instr);
+ enum glsl_interp_mode interp = (enum glsl_interp_mode)
+ nir_intrinsic_interp_mode(intrin);
+ nir_intrinsic_op bary_op = intrin->intrinsic;
+ enum brw_barycentric_mode bary =
+ brw_barycentric_mode(interp, bary_op);
+
+ barycentric_interp_modes |= 1 << bary;
+
+ if (devinfo->needs_unlit_centroid_workaround &&
+ bary_op == nir_intrinsic_load_barycentric_centroid)
+ barycentric_interp_modes |= 1 << centroid_to_pixel(bary);
}
}
}
static void
brw_compute_flat_inputs(struct brw_wm_prog_data *prog_data,
- bool shade_model_flat, const nir_shader *shader)
+ const nir_shader *shader)
{
prog_data->flat_inputs = 0;
nir_foreach_variable(var, &shader->inputs) {
- enum glsl_interp_qualifier interp_qualifier =
- (enum glsl_interp_qualifier)var->data.interpolation;
- bool is_gl_Color = (var->data.location == VARYING_SLOT_COL0) ||
- (var->data.location == VARYING_SLOT_COL1);
-
int input_index = prog_data->urb_setup[var->data.location];
if (input_index < 0)
continue;
/* flat shading */
- if (interp_qualifier == INTERP_QUALIFIER_FLAT ||
- (shade_model_flat && is_gl_Color &&
- interp_qualifier == INTERP_QUALIFIER_NONE))
+ if (var->data.interpolation == INTERP_MODE_FLAT)
prog_data->flat_inputs |= (1 << input_index);
}
}
static uint8_t
computed_depth_mode(const nir_shader *shader)
{
- if (shader->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
- switch (shader->info.fs.depth_layout) {
+ if (shader->info->outputs_written & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
+ switch (shader->info->fs.depth_layout) {
case FRAG_DEPTH_LAYOUT_NONE:
case FRAG_DEPTH_LAYOUT_ANY:
return BRW_PSCDEPTH_ON;
return BRW_PSCDEPTH_OFF;
}
+/**
+ * Move load_interpolated_input with simple (payload-based) barycentric modes
+ * to the top of the program so we don't emit multiple PLNs for the same input.
+ *
+ * This works around CSE not being able to handle non-dominating cases
+ * such as:
+ *
+ * if (...) {
+ * interpolate input
+ * } else {
+ * interpolate the same exact input
+ * }
+ *
+ * This should be replaced by global value numbering someday.
+ */
+void
+move_interpolation_to_top(nir_shader *nir)
+{
+ nir_foreach_function(f, nir) {
+ if (!f->impl)
+ continue;
+
+ nir_block *top = nir_start_block(f->impl);
+ exec_node *cursor_node = NULL;
+
+ nir_foreach_block(block, f->impl) {
+ if (block == top)
+ continue;
+
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ if (intrin->intrinsic != nir_intrinsic_load_interpolated_input)
+ continue;
+ nir_intrinsic_instr *bary_intrinsic =
+ nir_instr_as_intrinsic(intrin->src[0].ssa->parent_instr);
+ nir_intrinsic_op op = bary_intrinsic->intrinsic;
+
+ /* Leave interpolateAtSample/Offset() where they are. */
+ if (op == nir_intrinsic_load_barycentric_at_sample ||
+ op == nir_intrinsic_load_barycentric_at_offset)
+ continue;
+
+ nir_instr *move[3] = {
+ &bary_intrinsic->instr,
+ intrin->src[1].ssa->parent_instr,
+ instr
+ };
+
+ for (unsigned i = 0; i < ARRAY_SIZE(move); i++) {
+ if (move[i]->block != top) {
+ move[i]->block = top;
+ exec_node_remove(&move[i]->node);
+ if (cursor_node) {
+ exec_node_insert_after(cursor_node, &move[i]->node);
+ } else {
+ exec_list_push_head(&top->instr_list, &move[i]->node);
+ }
+ cursor_node = &move[i]->node;
+ }
+ }
+ }
+ }
+ nir_metadata_preserve(f->impl, (nir_metadata)
+ ((unsigned) nir_metadata_block_index |
+ (unsigned) nir_metadata_dominance));
+ }
+}
+
+/**
+ * Demote per-sample barycentric intrinsics to centroid.
+ *
+ * Useful when rendering to a non-multisampled buffer.
+ */
+static void
+demote_sample_qualifiers(nir_shader *nir)
+{
+ nir_foreach_function(f, nir) {
+ if (!f->impl)
+ continue;
+
+ nir_builder b;
+ nir_builder_init(&b, f->impl);
+
+ nir_foreach_block(block, f->impl) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ if (intrin->intrinsic != nir_intrinsic_load_barycentric_sample &&
+ intrin->intrinsic != nir_intrinsic_load_barycentric_at_sample)
+ continue;
+
+ b.cursor = nir_before_instr(instr);
+ nir_ssa_def *centroid =
+ nir_load_barycentric(&b, nir_intrinsic_load_barycentric_centroid,
+ nir_intrinsic_interp_mode(intrin));
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
+ nir_src_for_ssa(centroid));
+ nir_instr_remove(instr);
+ }
+ }
+
+ nir_metadata_preserve(f->impl, (nir_metadata)
+ ((unsigned) nir_metadata_block_index |
+ (unsigned) nir_metadata_dominance));
+ }
+}
+
const unsigned *
brw_compile_fs(const struct brw_compiler *compiler, void *log_data,
void *mem_ctx,
nir_shader *shader = nir_shader_clone(mem_ctx, src_shader);
shader = brw_nir_apply_sampler_key(shader, compiler->devinfo, &key->tex,
true);
- brw_nir_lower_fs_inputs(shader);
+ brw_nir_lower_fs_inputs(shader, compiler->devinfo, key);
brw_nir_lower_fs_outputs(shader);
+ if (!key->multisample_fbo)
+ NIR_PASS_V(shader, demote_sample_qualifiers);
+ NIR_PASS_V(shader, move_interpolation_to_top);
shader = brw_postprocess_nir(shader, compiler->devinfo, true);
/* key->alpha_test_func means simulating alpha testing via discards,
* so the shader definitely kills pixels.
*/
- prog_data->uses_kill = shader->info.fs.uses_discard || key->alpha_test_func;
+ prog_data->uses_kill = shader->info->fs.uses_discard ||
+ key->alpha_test_func;
prog_data->uses_omask = key->multisample_fbo &&
- shader->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK);
+ shader->info->outputs_written & BITFIELD64_BIT(FRAG_RESULT_SAMPLE_MASK);
prog_data->computed_depth_mode = computed_depth_mode(shader);
prog_data->computed_stencil =
- shader->info.outputs_written & BITFIELD64_BIT(FRAG_RESULT_STENCIL);
+ shader->info->outputs_written & BITFIELD64_BIT(FRAG_RESULT_STENCIL);
prog_data->persample_dispatch =
key->multisample_fbo &&
(key->persample_interp ||
- (shader->info.system_values_read & (SYSTEM_BIT_SAMPLE_ID |
- SYSTEM_BIT_SAMPLE_POS)) ||
- shader->info.fs.uses_sample_qualifier);
+ (shader->info->system_values_read & (SYSTEM_BIT_SAMPLE_ID |
+ SYSTEM_BIT_SAMPLE_POS)) ||
+ shader->info->fs.uses_sample_qualifier ||
+ shader->info->outputs_read);
- prog_data->early_fragment_tests = shader->info.fs.early_fragment_tests;
+ prog_data->early_fragment_tests = shader->info->fs.early_fragment_tests;
prog_data->barycentric_interp_modes =
- brw_compute_barycentric_interp_modes(compiler->devinfo,
- key->flat_shade,
- key->persample_interp,
- shader);
+ brw_compute_barycentric_interp_modes(compiler->devinfo, shader);
cfg_t *simd8_cfg = NULL, *simd16_cfg = NULL;
uint8_t simd8_grf_start = 0, simd16_grf_start = 0;
* because it relies on prog_data->urb_setup which is computed in
* fs_visitor::calculate_urb_setup().
*/
- brw_compute_flat_inputs(prog_data, key->flat_shade, shader);
+ brw_compute_flat_inputs(prog_data, shader);
fs_generator g(compiler, log_data, mem_ctx, (void *) key, &prog_data->base,
v8.promoted_constants, v8.runtime_check_aads_emit,
if (unlikely(INTEL_DEBUG & DEBUG_WM)) {
g.enable_debug(ralloc_asprintf(mem_ctx, "%s fragment shader %s",
- shader->info.label ? shader->info.label :
- "unnamed",
- shader->info.name));
+ shader->info->label ?
+ shader->info->label : "unnamed",
+ shader->info->name));
}
if (simd8_cfg) {
return g.get_assembly(final_assembly_size);
}
-fs_reg *
-fs_visitor::emit_cs_local_invocation_id_setup()
-{
- assert(stage == MESA_SHADER_COMPUTE);
-
- fs_reg *reg = new(this->mem_ctx) fs_reg(vgrf(glsl_type::uvec3_type));
-
- struct brw_reg src =
- brw_vec8_grf(payload.local_invocation_id_reg, 0);
- src = retype(src, BRW_REGISTER_TYPE_UD);
- bld.MOV(*reg, src);
- src.nr += dispatch_width / 8;
- bld.MOV(offset(*reg, bld, 1), src);
- src.nr += dispatch_width / 8;
- bld.MOV(offset(*reg, bld, 2), src);
-
- return reg;
-}
-
fs_reg *
fs_visitor::emit_cs_work_group_id_setup()
{
return reg;
}
+static void
+fill_push_const_block_info(struct brw_push_const_block *block, unsigned dwords)
+{
+ block->dwords = dwords;
+ block->regs = DIV_ROUND_UP(dwords, 8);
+ block->size = block->regs * 32;
+}
+
+static void
+cs_fill_push_const_info(const struct gen_device_info *devinfo,
+ struct brw_cs_prog_data *cs_prog_data)
+{
+ const struct brw_stage_prog_data *prog_data = &cs_prog_data->base;
+ bool fill_thread_id =
+ cs_prog_data->thread_local_id_index >= 0 &&
+ cs_prog_data->thread_local_id_index < (int)prog_data->nr_params;
+ bool cross_thread_supported = devinfo->gen > 7 || devinfo->is_haswell;
+
+ /* The thread ID should be stored in the last param dword */
+ assert(prog_data->nr_params > 0 || !fill_thread_id);
+ assert(!fill_thread_id ||
+ cs_prog_data->thread_local_id_index ==
+ (int)prog_data->nr_params - 1);
+
+ unsigned cross_thread_dwords, per_thread_dwords;
+ if (!cross_thread_supported) {
+ cross_thread_dwords = 0u;
+ per_thread_dwords = prog_data->nr_params;
+ } else if (fill_thread_id) {
+ /* Fill all but the last register with cross-thread payload */
+ cross_thread_dwords = 8 * (cs_prog_data->thread_local_id_index / 8);
+ per_thread_dwords = prog_data->nr_params - cross_thread_dwords;
+ assert(per_thread_dwords > 0 && per_thread_dwords <= 8);
+ } else {
+ /* Fill all data using cross-thread payload */
+ cross_thread_dwords = prog_data->nr_params;
+ per_thread_dwords = 0u;
+ }
+
+ fill_push_const_block_info(&cs_prog_data->push.cross_thread, cross_thread_dwords);
+ fill_push_const_block_info(&cs_prog_data->push.per_thread, per_thread_dwords);
+
+ unsigned total_dwords =
+ (cs_prog_data->push.per_thread.size * cs_prog_data->threads +
+ cs_prog_data->push.cross_thread.size) / 4;
+ fill_push_const_block_info(&cs_prog_data->push.total, total_dwords);
+
+ assert(cs_prog_data->push.cross_thread.dwords % 8 == 0 ||
+ cs_prog_data->push.per_thread.size == 0);
+ assert(cs_prog_data->push.cross_thread.dwords +
+ cs_prog_data->push.per_thread.dwords ==
+ prog_data->nr_params);
+}
+
+static void
+cs_set_simd_size(struct brw_cs_prog_data *cs_prog_data, unsigned size)
+{
+ cs_prog_data->simd_size = size;
+ unsigned group_size = cs_prog_data->local_size[0] *
+ cs_prog_data->local_size[1] * cs_prog_data->local_size[2];
+ cs_prog_data->threads = (group_size + size - 1) / size;
+}
+
const unsigned *
brw_compile_cs(const struct brw_compiler *compiler, void *log_data,
void *mem_ctx,
true);
brw_nir_lower_cs_shared(shader);
prog_data->base.total_shared += shader->num_shared;
+
+ /* Now that we cloned the nir_shader, we can update num_uniforms based on
+ * the thread_local_id_index.
+ */
+ assert(prog_data->thread_local_id_index >= 0);
+ shader->num_uniforms =
+ MAX2(shader->num_uniforms,
+ (unsigned)4 * (prog_data->thread_local_id_index + 1));
+
+ brw_nir_lower_intrinsics(shader, &prog_data->base);
shader = brw_postprocess_nir(shader, compiler->devinfo, true);
- prog_data->local_size[0] = shader->info.cs.local_size[0];
- prog_data->local_size[1] = shader->info.cs.local_size[1];
- prog_data->local_size[2] = shader->info.cs.local_size[2];
+ prog_data->local_size[0] = shader->info->cs.local_size[0];
+ prog_data->local_size[1] = shader->info->cs.local_size[1];
+ prog_data->local_size[2] = shader->info->cs.local_size[2];
unsigned local_workgroup_size =
- shader->info.cs.local_size[0] * shader->info.cs.local_size[1] *
- shader->info.cs.local_size[2];
+ shader->info->cs.local_size[0] * shader->info->cs.local_size[1] *
+ shader->info->cs.local_size[2];
unsigned max_cs_threads = compiler->devinfo->max_cs_threads;
unsigned simd_required = DIV_ROUND_UP(local_workgroup_size, max_cs_threads);
fail_msg = v8.fail_msg;
} else {
cfg = v8.cfg;
- prog_data->simd_size = 8;
+ cs_set_simd_size(prog_data, 8);
+ cs_fill_push_const_info(compiler->devinfo, prog_data);
prog_data->base.dispatch_grf_start_reg = v8.payload.num_regs;
}
}
}
} else {
cfg = v16.cfg;
- prog_data->simd_size = 16;
+ cs_set_simd_size(prog_data, 16);
+ cs_fill_push_const_info(compiler->devinfo, prog_data);
prog_data->dispatch_grf_start_reg_16 = v16.payload.num_regs;
}
}
}
} else {
cfg = v32.cfg;
- prog_data->simd_size = 32;
+ cs_set_simd_size(prog_data, 32);
+ cs_fill_push_const_info(compiler->devinfo, prog_data);
}
}
MESA_SHADER_COMPUTE);
if (INTEL_DEBUG & DEBUG_CS) {
char *name = ralloc_asprintf(mem_ctx, "%s compute shader %s",
- shader->info.label ? shader->info.label :
+ shader->info->label ? shader->info->label :
"unnamed",
- shader->info.name);
+ shader->info->name);
g.enable_debug(name);
}
return g.get_assembly(final_assembly_size);
}
-void
-brw_cs_fill_local_id_payload(const struct brw_cs_prog_data *prog_data,
- void *buffer, uint32_t threads, uint32_t stride)
+/**
+ * Test the dispatch mask packing assumptions of
+ * brw_stage_has_packed_dispatch(). Call this from e.g. the top of
+ * fs_visitor::emit_nir_code() to cause a GPU hang if any shader invocation is
+ * executed with an unexpected dispatch mask.
+ */
+static UNUSED void
+brw_fs_test_dispatch_packing(const fs_builder &bld)
{
- if (prog_data->local_invocation_id_regs == 0)
- return;
+ const gl_shader_stage stage = bld.shader->stage;
- /* 'stride' should be an integer number of registers, that is, a multiple
- * of 32 bytes.
- */
- assert(stride % 32 == 0);
-
- unsigned x = 0, y = 0, z = 0;
- for (unsigned t = 0; t < threads; t++) {
- uint32_t *param = (uint32_t *) buffer + stride * t / 4;
-
- for (unsigned i = 0; i < prog_data->simd_size; i++) {
- param[0 * prog_data->simd_size + i] = x;
- param[1 * prog_data->simd_size + i] = y;
- param[2 * prog_data->simd_size + i] = z;
-
- x++;
- if (x == prog_data->local_size[0]) {
- x = 0;
- y++;
- if (y == prog_data->local_size[1]) {
- y = 0;
- z++;
- if (z == prog_data->local_size[2])
- z = 0;
- }
- }
- }
+ if (brw_stage_has_packed_dispatch(bld.shader->devinfo, stage,
+ bld.shader->stage_prog_data)) {
+ const fs_builder ubld = bld.exec_all().group(1, 0);
+ const fs_reg tmp = component(bld.vgrf(BRW_REGISTER_TYPE_UD), 0);
+ const fs_reg mask = (stage == MESA_SHADER_FRAGMENT ? brw_vmask_reg() :
+ brw_dmask_reg());
+
+ ubld.ADD(tmp, mask, brw_imm_ud(1));
+ ubld.AND(tmp, mask, tmp);
+
+ /* This will loop forever if the dispatch mask doesn't have the expected
+ * form '2^n-1', in which case tmp will be non-zero.
+ */
+ bld.emit(BRW_OPCODE_DO);
+ bld.CMP(bld.null_reg_ud(), tmp, brw_imm_ud(0), BRW_CONDITIONAL_NZ);
+ set_predicate(BRW_PREDICATE_NORMAL, bld.emit(BRW_OPCODE_WHILE));
}
}