static void fs_lower_opcode_derivative(struct toy_compiler *tc, struct toy_inst *inst) { struct toy_dst dst[4]; struct toy_src src[4]; int i; tdst_transpose(inst->dst, dst); tsrc_transpose(inst->src[0], src); /* * Every four fragments are from a 2x2 subspan, with * * fragment 1 on the top-left, * fragment 2 on the top-right, * fragment 3 on the bottom-left, * fragment 4 on the bottom-right. * * DDX should thus produce * * dst = src.yyww - src.xxzz * * and DDY should produce * * dst = src.zzww - src.xxyy * * But since we are in BRW_ALIGN_1, swizzling does not work and we have to * play with the region parameters. */ if (inst->opcode == TOY_OPCODE_DDX) { for (i = 0; i < 4; i++) { struct toy_src left, right; left = tsrc_rect(src[i], TOY_RECT_220); right = tsrc_offset(left, 0, 1); tc_ADD(tc, dst[i], right, tsrc_negate(left)); } } else { for (i = 0; i < 4; i++) { struct toy_src top, bottom; /* approximate with dst = src.zzzz - src.xxxx */ top = tsrc_rect(src[i], TOY_RECT_440); bottom = tsrc_offset(top, 0, 2); tc_ADD(tc, dst[i], bottom, tsrc_negate(top)); } } tc_discard_inst(tc, inst); }
static void fs_lower_opcode_kil(struct toy_compiler *tc, struct toy_inst *inst) { struct toy_dst pixel_mask_dst; struct toy_src f0, pixel_mask; struct toy_inst *tmp; /* lower half of r1.7:ud */ pixel_mask_dst = tdst_uw(tdst(TOY_FILE_GRF, 1, 7 * 4)); pixel_mask = tsrc_rect(tsrc_from(pixel_mask_dst), TOY_RECT_010); f0 = tsrc_rect(tsrc_uw(tsrc(TOY_FILE_ARF, BRW_ARF_FLAG, 0)), TOY_RECT_010); /* KILP or KIL */ if (tsrc_is_null(inst->src[0])) { struct toy_src dummy = tsrc_uw(tsrc(TOY_FILE_GRF, 0, 0)); struct toy_dst f0_dst = tdst_uw(tdst(TOY_FILE_ARF, BRW_ARF_FLAG, 0)); /* create a mask that masks out all pixels */ tmp = tc_MOV(tc, f0_dst, tsrc_rect(tsrc_imm_uw(0xffff), TOY_RECT_010)); tmp->exec_size = BRW_EXECUTE_1; tmp->mask_ctrl = BRW_MASK_DISABLE; tc_CMP(tc, tdst_null(), dummy, dummy, BRW_CONDITIONAL_NEQ); /* swapping the two src operands breaks glBitmap()!? */ tmp = tc_AND(tc, pixel_mask_dst, f0, pixel_mask); tmp->exec_size = BRW_EXECUTE_1; tmp->mask_ctrl = BRW_MASK_DISABLE; } else { struct toy_src src[4]; int i; tsrc_transpose(inst->src[0], src); /* mask out killed pixels */ for (i = 0; i < 4; i++) { tc_CMP(tc, tdst_null(), src[i], tsrc_imm_f(0.0f), BRW_CONDITIONAL_GE); /* swapping the two src operands breaks glBitmap()!? */ tmp = tc_AND(tc, pixel_mask_dst, f0, pixel_mask); tmp->exec_size = BRW_EXECUTE_1; tmp->mask_ctrl = BRW_MASK_DISABLE; } } tc_discard_inst(tc, inst); }
static void fs_lower_opcode_tgsi_const_gen7(struct fs_compile_context *fcc, struct toy_dst dst, int dim, struct toy_src idx) { struct toy_compiler *tc = &fcc->tc; const struct toy_dst offset = tdst_ud(tdst(TOY_FILE_MRF, fcc->first_free_mrf, 0)); struct toy_src desc; struct toy_inst *inst; struct toy_dst tmp, real_dst[4]; int i; /* * In 4c1fdae0a01b3f92ec03b61aac1d3df500d51fc6, pull constant load was * changed from OWord Block Read to ld to increase performance in the * classic driver. Since we use the constant cache instead of the data * cache, I wonder if we still want to follow the classic driver. */ /* set offset */ inst = tc_MOV(tc, offset, tsrc_rect(idx, TOY_RECT_010)); inst->exec_size = BRW_EXECUTE_8; inst->mask_ctrl = BRW_MASK_DISABLE; desc = tsrc_imm_mdesc_sampler(tc, 1, 1, false, BRW_SAMPLER_SIMD_MODE_SIMD4X2, GEN5_SAMPLER_MESSAGE_SAMPLE_LD, 0, ILO_WM_CONST_SURFACE(dim)); tmp = tc_alloc_tmp(tc); inst = tc_SEND(tc, tmp, tsrc_from(offset), desc, BRW_SFID_SAMPLER); inst->exec_size = BRW_EXECUTE_8; inst->mask_ctrl = BRW_MASK_DISABLE; tdst_transpose(dst, real_dst); for (i = 0; i < 4; i++) { const struct toy_src src = tsrc_offset(tsrc_rect(tsrc_from(tmp), TOY_RECT_010), 0, i); /* cast to type D to make sure these are raw moves */ tc_MOV(tc, tdst_d(real_dst[i]), tsrc_d(src)); } }
static void gs_lower_opcode_emit_vue_static(struct gs_compile_context *gcc) { struct toy_compiler *tc = &gcc->tc; struct toy_inst *inst2; bool eot; eot = (gcc->static_data.num_vertices == gcc->static_data.total_vertices); gcc->vars.prim_end = ((gcc->static_data.last_vertex[(gcc->static_data.num_vertices - 1) / 32] & 1 << ((gcc->static_data.num_vertices - 1) % 32)) != 0); if (eot && gcc->write_so) { inst2 = tc_OR(tc, tdst_offset(gcc->vars.urb_write_header, 0, 2), tsrc_from(gcc->vars.so_written), tsrc_imm_d(gcc->vars.prim_type << 2 | gcc->vars.prim_start << 1 | gcc->vars.prim_end)); inst2->exec_size = GEN6_EXECSIZE_1; inst2->src[0] = tsrc_rect(inst2->src[0], TOY_RECT_010); inst2->src[1] = tsrc_rect(inst2->src[1], TOY_RECT_010); } else { gs_COPY1(tc, gcc->vars.urb_write_header, 2, tsrc_imm_d(gcc->vars.prim_type << 2 | gcc->vars.prim_start << 1 | gcc->vars.prim_end), 0); } gs_write_vue(gcc, tdst_d(gcc->vars.tmp), tsrc_from(gcc->vars.urb_write_header), gcc->vars.tgsi_outs, gcc->shader->out.count, eot); if (!eot) { gs_COPY1(tc, gcc->vars.urb_write_header, 0, tsrc_from(tdst_d(gcc->vars.tmp)), 0); } gcc->vars.prim_start = gcc->vars.prim_end; gcc->vars.prim_end = false; }
static void gs_COPY1(struct toy_compiler *tc, struct toy_dst dst, int dst_ch, struct toy_src src, int src_ch) { struct toy_inst *inst; inst = tc_MOV(tc, tdst_offset(dst, 0, dst_ch), tsrc_rect(tsrc_offset(src, 0, src_ch), TOY_RECT_010)); inst->exec_size = GEN6_EXECSIZE_1; inst->mask_ctrl = GEN6_MASKCTRL_NOMASK; }
static void fs_lower_opcode_tgsi_const_gen6(struct fs_compile_context *fcc, struct toy_dst dst, int dim, struct toy_src idx) { const struct toy_dst header = tdst_ud(tdst(TOY_FILE_MRF, fcc->first_free_mrf, 0)); const struct toy_dst global_offset = tdst_ud(tdst(TOY_FILE_MRF, fcc->first_free_mrf, 2 * 4)); const struct toy_src r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0)); struct toy_compiler *tc = &fcc->tc; unsigned msg_type, msg_ctrl, msg_len; struct toy_inst *inst; struct toy_src desc; struct toy_dst tmp, real_dst[4]; int i; /* set message header */ inst = tc_MOV(tc, header, r0); inst->mask_ctrl = BRW_MASK_DISABLE; /* set global offset */ inst = tc_MOV(tc, global_offset, idx); inst->mask_ctrl = BRW_MASK_DISABLE; inst->exec_size = BRW_EXECUTE_1; inst->src[0].rect = TOY_RECT_010; msg_type = BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ; msg_ctrl = BRW_DATAPORT_OWORD_BLOCK_1_OWORDLOW << 8; msg_len = 1; desc = tsrc_imm_mdesc_data_port(tc, false, msg_len, 1, true, false, msg_type, msg_ctrl, ILO_WM_CONST_SURFACE(dim)); tmp = tc_alloc_tmp(tc); tc_SEND(tc, tmp, tsrc_from(header), desc, fcc->const_cache); tdst_transpose(dst, real_dst); for (i = 0; i < 4; i++) { const struct toy_src src = tsrc_offset(tsrc_rect(tsrc_from(tmp), TOY_RECT_010), 0, i); /* cast to type D to make sure these are raw moves */ tc_MOV(tc, tdst_d(real_dst[i]), tsrc_d(src)); } }
static bool vs_lower_opcode_tgsi_const_pcb(struct vs_compile_context *vcc, struct toy_dst dst, int dim, struct toy_src idx) { const int i = idx.val32; const int grf = vcc->first_const_grf + i / 2; const int grf_subreg = (i & 1) * 16; struct toy_src src; if (!vcc->variant->use_pcb || dim != 0 || idx.file != TOY_FILE_IMM || grf >= vcc->first_ucp_grf) return false; src = tsrc_rect(tsrc(TOY_FILE_GRF, grf, grf_subreg), TOY_RECT_041); tc_MOV(&vcc->tc, dst, src); return true; }
static void fetch_face(struct fs_compile_context *fcc, struct toy_dst dst) { struct toy_compiler *tc = &fcc->tc; const struct toy_src r0 = tsrc_d(tsrc(TOY_FILE_GRF, 0, 0)); struct toy_dst tmp_f, tmp; struct toy_dst real_dst[4]; tdst_transpose(dst, real_dst); tmp_f = tc_alloc_tmp(tc); tmp = tdst_d(tmp_f); tc_SHR(tc, tmp, tsrc_rect(r0, TOY_RECT_010), tsrc_imm_d(15)); tc_AND(tc, tmp, tsrc_from(tmp), tsrc_imm_d(1)); tc_MOV(tc, tmp_f, tsrc_from(tmp)); /* convert to 1.0 and -1.0 */ tc_MUL(tc, tmp_f, tsrc_from(tmp_f), tsrc_imm_f(-2.0f)); tc_ADD(tc, real_dst[0], tsrc_from(tmp_f), tsrc_imm_f(1.0f)); tc_MOV(tc, real_dst[1], tsrc_imm_f(0.0f)); tc_MOV(tc, real_dst[2], tsrc_imm_f(0.0f)); tc_MOV(tc, real_dst[3], tsrc_imm_f(1.0f)); }
static void gs_lower_opcode_tgsi_in(struct gs_compile_context *gcc, struct toy_dst dst, int dim, int idx) { struct toy_compiler *tc = &gcc->tc; struct toy_src attr; int slot, reg = -1, subreg; slot = toy_tgsi_find_input(&gcc->tgsi, idx); if (slot >= 0) { int i; for (i = 0; i < gcc->variant->u.gs.num_inputs; i++) { if (gcc->variant->u.gs.semantic_names[i] == gcc->tgsi.inputs[slot].semantic_name && gcc->variant->u.gs.semantic_indices[i] == gcc->tgsi.inputs[slot].semantic_index) { reg = i / 2; subreg = (i % 2) * 4; break; } } } if (reg < 0) { tc_MOV(tc, dst, tsrc_imm_f(0.0f)); return; } /* fix vertex ordering for GEN6_3DPRIM_TRISTRIP_REVERSE */ if (gcc->in_vue_count == 3 && dim < 2) { struct toy_inst *inst; /* get PrimType */ inst = tc_AND(tc, tdst_d(gcc->vars.tmp), tsrc_offset(gcc->payload.header, 0, 2), tsrc_imm_d(0x1f)); inst->exec_size = GEN6_EXECSIZE_1; inst->src[0] = tsrc_rect(inst->src[0], TOY_RECT_010); inst->src[1] = tsrc_rect(inst->src[1], TOY_RECT_010); inst = tc_CMP(tc, tdst_null(), tsrc_from(tdst_d(gcc->vars.tmp)), tsrc_imm_d(GEN6_3DPRIM_TRISTRIP_REVERSE), GEN6_COND_NZ); inst->src[0] = tsrc_rect(inst->src[0], TOY_RECT_010); attr = tsrc_offset(gcc->payload.vues[dim], reg, subreg); inst = tc_MOV(tc, dst, attr); inst->pred_ctrl = GEN6_PREDCTRL_NORMAL; /* swap IN[0] and IN[1] for GEN6_3DPRIM_TRISTRIP_REVERSE */ dim = !dim; attr = tsrc_offset(gcc->payload.vues[dim], reg, subreg); inst = tc_MOV(tc, dst, attr); inst->pred_ctrl = GEN6_PREDCTRL_NORMAL; inst->pred_inv = true; } else { attr = tsrc_offset(gcc->payload.vues[dim], reg, subreg); tc_MOV(tc, dst, attr); } }
static void gs_lower_opcode_emit_so_static(struct gs_compile_context *gcc) { struct toy_compiler *tc = &gcc->tc; struct toy_inst *inst; int i, j; if (gcc->static_data.num_vertices_in_prim < gcc->out_vue_min_count) return; inst = tc_MOV(tc, tdst_w(gcc->vars.tmp), tsrc_imm_v(0x03020100)); inst->exec_size = GEN6_EXECSIZE_8; inst->mask_ctrl = GEN6_MASKCTRL_NOMASK; tc_ADD(tc, tdst_d(gcc->vars.tmp), tsrc_from(tdst_d(gcc->vars.tmp)), tsrc_rect(tsrc_from(gcc->vars.so_index), TOY_RECT_010)); tc_IF(tc, tdst_null(), tsrc_rect(tsrc_offset(tsrc_from(tdst_d(gcc->vars.tmp)), 0, gcc->out_vue_min_count - 1), TOY_RECT_010), tsrc_rect(tsrc_offset(gcc->payload.svbi, 0, 4), TOY_RECT_010), GEN6_COND_LE); { for (i = 0; i < gcc->out_vue_min_count; i++) { for (j = 0; j < gcc->so_info->num_outputs; j++) { const int idx = gcc->so_info->output[j].register_index; struct toy_src index, out; int binding_table_index; bool write_commit; index = tsrc_d(tsrc_offset(tsrc_from(gcc->vars.tmp), 0, i)); if (i == gcc->out_vue_min_count - 1) { out = gcc->vars.tgsi_outs[idx]; } else { /* gcc->vars.buffer_cur also points to the first vertex */ const int buf = (gcc->vars.buffer_cur + i) % gcc->vars.buffer_needed; out = tsrc_offset(tsrc_from(gcc->vars.buffers[buf]), idx, 0); } out = tsrc_offset(out, 0, gcc->so_info->output[j].start_component); /* * From the Sandy Bridge PRM, volume 4 part 2, page 19: * * "The Kernel must do a write commit on the last write to DAP * prior to a URB_WRITE with End of Thread." */ write_commit = (gcc->static_data.num_vertices == gcc->static_data.total_vertices && i == gcc->out_vue_min_count - 1 && j == gcc->so_info->num_outputs - 1); binding_table_index = gcc->shader->bt.gen6_so_base + j; gs_write_so(gcc, gcc->vars.tmp, index, out, write_commit, binding_table_index); /* * From the Sandy Bridge PRM, volume 4 part 1, page 168: * * "The write commit does not modify the destination register, but * merely clears the dependency associated with the destination * register. Thus, a simple "mov" instruction using the register as a * source is sufficient to wait for the write commit to occur." */ if (write_commit) tc_MOV(tc, gcc->vars.tmp, tsrc_from(gcc->vars.tmp)); } } /* SONumPrimsWritten occupies the higher word of m0.2 of URB_WRITE */ tc_ADD(tc, gcc->vars.so_written, tsrc_from(gcc->vars.so_written), tsrc_imm_d(1 << 16)); tc_ADD(tc, gcc->vars.so_index, tsrc_from(gcc->vars.so_index), tsrc_imm_d(gcc->out_vue_min_count)); } tc_ENDIF(tc); }
/** * Emit instructions to write the VUE. */ static void vs_write_vue(struct vs_compile_context *vcc) { struct toy_compiler *tc = &vcc->tc; struct toy_src outs[PIPE_MAX_SHADER_OUTPUTS]; struct toy_dst header; struct toy_src r0; struct toy_inst *inst; int sent_attrs, total_attrs; header = tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf, 0)); r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0)); inst = tc_MOV(tc, header, r0); inst->mask_ctrl = GEN6_MASKCTRL_NOMASK; if (ilo_dev_gen(tc->dev) >= ILO_GEN(7)) { inst = tc_OR(tc, tdst_offset(header, 0, 5), tsrc_rect(tsrc_offset(r0, 0, 5), TOY_RECT_010), tsrc_rect(tsrc_imm_ud(0xff00), TOY_RECT_010)); inst->exec_size = GEN6_EXECSIZE_1; inst->access_mode = GEN6_ALIGN_1; inst->mask_ctrl = GEN6_MASKCTRL_NOMASK; } total_attrs = vs_collect_outputs(vcc, outs); sent_attrs = 0; while (sent_attrs < total_attrs) { struct toy_src desc; int mrf = vcc->first_free_mrf + 1, avail_mrf_for_attrs; int num_attrs, msg_len, i; bool eot; num_attrs = total_attrs - sent_attrs; eot = true; /* see if we need another message */ avail_mrf_for_attrs = vcc->last_free_mrf - mrf + 1; if (num_attrs > avail_mrf_for_attrs) { /* * From the Sandy Bridge PRM, volume 4 part 2, page 22: * * "Offset. This field specifies a destination offset (in 256-bit * units) from the start of the URB entry(s), as referenced by * URB Return Handle n, at which the data (if any) will be * written." * * As we need to offset the following messages, we must make sure * this one writes an even number of attributes. */ num_attrs = avail_mrf_for_attrs & ~1; eot = false; } if (ilo_dev_gen(tc->dev) >= ILO_GEN(7)) { /* do not forget about the header */ msg_len = 1 + num_attrs; } else { /* * From the Sandy Bridge PRM, volume 4 part 2, page 26: * * "At least 256 bits per vertex (512 bits total, M1 & M2) must * be written. Writing only 128 bits per vertex (256 bits * total, M1 only) results in UNDEFINED operation." * * "[DevSNB] Interleave writes must be in multiples of 256 per * vertex." * * That is, we must write or appear to write an even number of * attributes, starting from two. */ if (num_attrs % 2 && num_attrs == avail_mrf_for_attrs) { num_attrs--; eot = false; } msg_len = 1 + align(num_attrs, 2); } for (i = 0; i < num_attrs; i++) tc_MOV(tc, tdst(TOY_FILE_MRF, mrf++, 0), outs[sent_attrs + i]); assert(sent_attrs % 2 == 0); desc = tsrc_imm_mdesc_urb(tc, eot, msg_len, 0, eot, true, false, true, sent_attrs / 2, 0); tc_add2(tc, TOY_OPCODE_URB_WRITE, tdst_null(), tsrc_from(header), desc); sent_attrs += num_attrs; } }
/** * Collect the toy registers to be written to the VUE. */ static int vs_collect_outputs(struct vs_compile_context *vcc, struct toy_src *outs) { const struct toy_tgsi *tgsi = &vcc->tgsi; unsigned i; for (i = 0; i < vcc->shader->out.count; i++) { const int slot = vcc->output_map[i]; const int vrf = (slot >= 0) ? toy_tgsi_get_vrf(tgsi, TGSI_FILE_OUTPUT, 0, tgsi->outputs[slot].index) : -1; struct toy_src src; if (vrf >= 0) { struct toy_dst dst; dst = tdst(TOY_FILE_VRF, vrf, 0); src = tsrc_from(dst); if (i == 0) { /* PSIZE is at channel W */ tc_MOV(&vcc->tc, tdst_writemask(dst, TOY_WRITEMASK_W), tsrc_swizzle1(src, TOY_SWIZZLE_X)); /* the other channels are for the header */ dst = tdst_d(dst); tc_MOV(&vcc->tc, tdst_writemask(dst, TOY_WRITEMASK_XYZ), tsrc_imm_d(0)); } else { /* initialize unused channels to 0.0f */ if (tgsi->outputs[slot].undefined_mask) { dst = tdst_writemask(dst, tgsi->outputs[slot].undefined_mask); tc_MOV(&vcc->tc, dst, tsrc_imm_f(0.0f)); } } } else { /* XXX this is too ugly */ if (vcc->shader->out.semantic_names[i] == TGSI_SEMANTIC_CLIPDIST && slot < 0) { /* ok, we need to compute clip distance */ int clipvert_slot = -1, clipvert_vrf, j; for (j = 0; j < tgsi->num_outputs; j++) { if (tgsi->outputs[j].semantic_name == TGSI_SEMANTIC_CLIPVERTEX) { clipvert_slot = j; break; } else if (tgsi->outputs[j].semantic_name == TGSI_SEMANTIC_POSITION) { /* remember pos, but keep looking */ clipvert_slot = j; } } clipvert_vrf = (clipvert_slot >= 0) ? toy_tgsi_get_vrf(tgsi, TGSI_FILE_OUTPUT, 0, tgsi->outputs[clipvert_slot].index) : -1; if (clipvert_vrf >= 0) { struct toy_dst tmp = tc_alloc_tmp(&vcc->tc); struct toy_src clipvert = tsrc(TOY_FILE_VRF, clipvert_vrf, 0); int first_ucp, last_ucp; if (vcc->shader->out.semantic_indices[i]) { first_ucp = 4; last_ucp = MIN2(7, vcc->variant->u.vs.num_ucps - 1); } else { first_ucp = 0; last_ucp = MIN2(3, vcc->variant->u.vs.num_ucps - 1); } for (j = first_ucp; j <= last_ucp; j++) { const int plane_grf = vcc->first_ucp_grf + j / 2; const int plane_subreg = (j & 1) * 16; const struct toy_src plane = tsrc_rect(tsrc(TOY_FILE_GRF, plane_grf, plane_subreg), TOY_RECT_041); const unsigned writemask = 1 << ((j >= 4) ? j - 4 : j); tc_DP4(&vcc->tc, tdst_writemask(tmp, writemask), clipvert, plane); } src = tsrc_from(tmp); } else { src = tsrc_imm_f(0.0f); } } else { src = (i == 0) ? tsrc_imm_d(0) : tsrc_imm_f(0.0f); } } outs[i] = src; } return i; }
static void fetch_position(struct fs_compile_context *fcc, struct toy_dst dst) { struct toy_compiler *tc = &fcc->tc; const struct toy_src src_z = tsrc(TOY_FILE_GRF, fcc->payloads[0].source_depth, 0); const struct toy_src src_w = tsrc(TOY_FILE_GRF, fcc->payloads[0].source_w, 0); const int fb_height = (fcc->variant->u.fs.fb_height) ? fcc->variant->u.fs.fb_height : 1; const bool origin_upper_left = (fcc->tgsi.props.fs_coord_origin == TGSI_FS_COORD_ORIGIN_UPPER_LEFT); const bool pixel_center_integer = (fcc->tgsi.props.fs_coord_pixel_center == TGSI_FS_COORD_PIXEL_CENTER_INTEGER); struct toy_src subspan_x, subspan_y; struct toy_dst tmp, tmp_uw; struct toy_dst real_dst[4]; tdst_transpose(dst, real_dst); subspan_x = tsrc_uw(tsrc(TOY_FILE_GRF, 1, 2 * 4)); subspan_x = tsrc_rect(subspan_x, TOY_RECT_240); subspan_y = tsrc_offset(subspan_x, 0, 1); tmp_uw = tdst_uw(tc_alloc_tmp(tc)); tmp = tc_alloc_tmp(tc); /* X */ tc_ADD(tc, tmp_uw, subspan_x, tsrc_imm_v(0x10101010)); tc_MOV(tc, tmp, tsrc_from(tmp_uw)); if (pixel_center_integer) tc_MOV(tc, real_dst[0], tsrc_from(tmp)); else tc_ADD(tc, real_dst[0], tsrc_from(tmp), tsrc_imm_f(0.5f)); /* Y */ tc_ADD(tc, tmp_uw, subspan_y, tsrc_imm_v(0x11001100)); tc_MOV(tc, tmp, tsrc_from(tmp_uw)); if (origin_upper_left && pixel_center_integer) { tc_MOV(tc, real_dst[1], tsrc_from(tmp)); } else { struct toy_src y = tsrc_from(tmp); float offset = 0.0f; if (!pixel_center_integer) offset += 0.5f; if (!origin_upper_left) { offset += (float) (fb_height - 1); y = tsrc_negate(y); } tc_ADD(tc, real_dst[1], y, tsrc_imm_f(offset)); } /* Z and W */ tc_MOV(tc, real_dst[2], src_z); tc_INV(tc, real_dst[3], src_w); }
static void fetch_attr(struct fs_compile_context *fcc, struct toy_dst dst, int slot) { struct toy_compiler *tc = &fcc->tc; struct toy_dst real_dst[4]; bool is_const = false; int grf, mode, ch; tdst_transpose(dst, real_dst); grf = fcc->first_attr_grf + slot * 2; switch (fcc->tgsi.inputs[slot].interp) { case TGSI_INTERPOLATE_CONSTANT: is_const = true; break; case TGSI_INTERPOLATE_LINEAR: if (fcc->tgsi.inputs[slot].centroid) mode = BRW_WM_NONPERSPECTIVE_CENTROID_BARYCENTRIC; else mode = BRW_WM_NONPERSPECTIVE_PIXEL_BARYCENTRIC; break; case TGSI_INTERPOLATE_COLOR: if (fcc->variant->u.fs.flatshade) { is_const = true; break; } /* fall through */ case TGSI_INTERPOLATE_PERSPECTIVE: if (fcc->tgsi.inputs[slot].centroid) mode = BRW_WM_PERSPECTIVE_CENTROID_BARYCENTRIC; else mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC; break; default: assert(!"unexpected FS interpolation"); mode = BRW_WM_PERSPECTIVE_PIXEL_BARYCENTRIC; break; } if (is_const) { struct toy_src a0[4]; a0[0] = tsrc(TOY_FILE_GRF, grf + 0, 3 * 4); a0[1] = tsrc(TOY_FILE_GRF, grf + 0, 7 * 4); a0[2] = tsrc(TOY_FILE_GRF, grf + 1, 3 * 4); a0[3] = tsrc(TOY_FILE_GRF, grf + 1, 7 * 4); for (ch = 0; ch < 4; ch++) tc_MOV(tc, real_dst[ch], tsrc_rect(a0[ch], TOY_RECT_010)); } else { struct toy_src attr[4], uv; attr[0] = tsrc(TOY_FILE_GRF, grf + 0, 0); attr[1] = tsrc(TOY_FILE_GRF, grf + 0, 4 * 4); attr[2] = tsrc(TOY_FILE_GRF, grf + 1, 0); attr[3] = tsrc(TOY_FILE_GRF, grf + 1, 4 * 4); uv = tsrc(TOY_FILE_GRF, fcc->payloads[0].barycentric_interps[mode], 0); for (ch = 0; ch < 4; ch++) { tc_add2(tc, BRW_OPCODE_PLN, real_dst[ch], tsrc_rect(attr[ch], TOY_RECT_010), uv); } } if (fcc->tgsi.inputs[slot].semantic_name == TGSI_SEMANTIC_FOG) { tc_MOV(tc, real_dst[1], tsrc_imm_f(0.0f)); tc_MOV(tc, real_dst[2], tsrc_imm_f(0.0f)); tc_MOV(tc, real_dst[3], tsrc_imm_f(1.0f)); } }