/** Emits the interpolation for the varying inputs. */
void
fs_visitor::emit_interpolation_setup_gen6()
{
struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
fs_builder abld = bld.annotate("compute pixel centers");
if (devinfo->gen >= 8 || dispatch_width == 8) {
/* The "Register Region Restrictions" page says for BDW (and newer,
* presumably):
*
* "When destination spans two registers, the source may be one or
* two registers. The destination elements must be evenly split
* between the two registers."
*
* Thus we can do a single add(16) in SIMD8 or an add(32) in SIMD16 to
* compute our pixel centers.
*/
fs_reg int_pixel_xy(VGRF, alloc.allocate(dispatch_width / 8),
BRW_REGISTER_TYPE_UW);
const fs_builder dbld = abld.exec_all().group(dispatch_width * 2, 0);
dbld.ADD(int_pixel_xy,
fs_reg(stride(suboffset(g1_uw, 4), 1, 4, 0)),
fs_reg(brw_imm_v(0x11001010)));
this->pixel_x = vgrf(glsl_type::float_type);
this->pixel_y = vgrf(glsl_type::float_type);
abld.emit(FS_OPCODE_PIXEL_X, this->pixel_x, int_pixel_xy);
abld.emit(FS_OPCODE_PIXEL_Y, this->pixel_y, int_pixel_xy);
} else {
/* The "Register Region Restrictions" page says for SNB, IVB, HSW:
*
* "When destination spans two registers, the source MUST span two
* registers."
*
* Since the GRF source of the ADD will only read a single register, we
* must do two separate ADDs in SIMD16.
*/
fs_reg int_pixel_x = vgrf(glsl_type::uint_type);
fs_reg int_pixel_y = vgrf(glsl_type::uint_type);
int_pixel_x.type = BRW_REGISTER_TYPE_UW;
int_pixel_y.type = BRW_REGISTER_TYPE_UW;
abld.ADD(int_pixel_x,
fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
fs_reg(brw_imm_v(0x10101010)));
abld.ADD(int_pixel_y,
fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
fs_reg(brw_imm_v(0x11001100)));
/* As of gen6, we can no longer mix float and int sources. We have
* to turn the integer pixel centers into floats for their actual
* use.
*/
this->pixel_x = vgrf(glsl_type::float_type);
this->pixel_y = vgrf(glsl_type::float_type);
abld.MOV(this->pixel_x, int_pixel_x);
abld.MOV(this->pixel_y, int_pixel_y);
}
abld = bld.annotate("compute pos.w");
this->pixel_w = fs_reg(brw_vec8_grf(payload.source_w_reg, 0));
this->wpos_w = vgrf(glsl_type::float_type);
abld.emit(SHADER_OPCODE_RCP, this->wpos_w, this->pixel_w);
for (int i = 0; i < BRW_WM_BARYCENTRIC_INTERP_MODE_COUNT; ++i) {
uint8_t reg = payload.barycentric_coord_reg[i];
this->delta_xy[i] = fs_reg(brw_vec16_grf(reg, 0));
}
}
/** Emits the interpolation for the varying inputs. */
void
fs_visitor::emit_interpolation_setup_gen6()
{
struct brw_reg g1_uw = retype(brw_vec1_grf(1, 0), BRW_REGISTER_TYPE_UW);
fs_builder abld = bld.annotate("compute pixel centers");
if (devinfo->gen >= 8 || dispatch_width == 8) {
/* The "Register Region Restrictions" page says for BDW (and newer,
* presumably):
*
* "When destination spans two registers, the source may be one or
* two registers. The destination elements must be evenly split
* between the two registers."
*
* Thus we can do a single add(16) in SIMD8 or an add(32) in SIMD16 to
* compute our pixel centers.
*/
fs_reg int_pixel_xy(VGRF, alloc.allocate(dispatch_width / 8),
BRW_REGISTER_TYPE_UW);
const fs_builder dbld = abld.exec_all().group(dispatch_width * 2, 0);
dbld.ADD(int_pixel_xy,
fs_reg(stride(suboffset(g1_uw, 4), 1, 4, 0)),
fs_reg(brw_imm_v(0x11001010)));
this->pixel_x = vgrf(glsl_type::float_type);
this->pixel_y = vgrf(glsl_type::float_type);
abld.emit(FS_OPCODE_PIXEL_X, this->pixel_x, int_pixel_xy);
abld.emit(FS_OPCODE_PIXEL_Y, this->pixel_y, int_pixel_xy);
} else {
/* The "Register Region Restrictions" page says for SNB, IVB, HSW:
*
* "When destination spans two registers, the source MUST span two
* registers."
*
* Since the GRF source of the ADD will only read a single register, we
* must do two separate ADDs in SIMD16.
*/
fs_reg int_pixel_x = vgrf(glsl_type::uint_type);
fs_reg int_pixel_y = vgrf(glsl_type::uint_type);
int_pixel_x.type = BRW_REGISTER_TYPE_UW;
int_pixel_y.type = BRW_REGISTER_TYPE_UW;
abld.ADD(int_pixel_x,
fs_reg(stride(suboffset(g1_uw, 4), 2, 4, 0)),
fs_reg(brw_imm_v(0x10101010)));
abld.ADD(int_pixel_y,
fs_reg(stride(suboffset(g1_uw, 5), 2, 4, 0)),
fs_reg(brw_imm_v(0x11001100)));
/* As of gen6, we can no longer mix float and int sources. We have
* to turn the integer pixel centers into floats for their actual
* use.
*/
this->pixel_x = vgrf(glsl_type::float_type);
this->pixel_y = vgrf(glsl_type::float_type);
abld.MOV(this->pixel_x, int_pixel_x);
abld.MOV(this->pixel_y, int_pixel_y);
}
abld = bld.annotate("compute pos.w");
this->pixel_w = fs_reg(brw_vec8_grf(payload.source_w_reg, 0));
this->wpos_w = vgrf(glsl_type::float_type);
abld.emit(SHADER_OPCODE_RCP, this->wpos_w, this->pixel_w);
struct brw_wm_prog_data *wm_prog_data = brw_wm_prog_data(prog_data);
uint32_t centroid_modes = wm_prog_data->barycentric_interp_modes &
(1 << BRW_BARYCENTRIC_PERSPECTIVE_CENTROID |
1 << BRW_BARYCENTRIC_NONPERSPECTIVE_CENTROID);
for (int i = 0; i < BRW_BARYCENTRIC_MODE_COUNT; ++i) {
uint8_t reg = payload.barycentric_coord_reg[i];
this->delta_xy[i] = fs_reg(brw_vec16_grf(reg, 0));
if (devinfo->needs_unlit_centroid_workaround &&
(centroid_modes & (1 << i))) {
/* 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);
uint8_t pixel_reg = payload.barycentric_coord_reg[i - 1];
set_predicate_inv(BRW_PREDICATE_NORMAL, true,
bld.half(0).MOV(brw_vec8_grf(reg, 0),
brw_vec8_grf(pixel_reg, 0)));
set_predicate_inv(BRW_PREDICATE_NORMAL, true,
bld.half(0).MOV(brw_vec8_grf(reg + 1, 0),
brw_vec8_grf(pixel_reg + 1, 0)));
if (dispatch_width == 16) {
set_predicate_inv(BRW_PREDICATE_NORMAL, true,
bld.half(1).MOV(brw_vec8_grf(reg + 2, 0),
brw_vec8_grf(pixel_reg + 2, 0)));
set_predicate_inv(BRW_PREDICATE_NORMAL, true,
bld.half(1).MOV(brw_vec8_grf(reg + 3, 0),
brw_vec8_grf(pixel_reg + 3, 0)));
}
assert(dispatch_width != 32); /* not implemented yet */
}
}
}
