static void
gs_write_vue(struct gs_compile_context *gcc,
struct toy_dst dst, struct toy_src msg_header,
const struct toy_src *outs, int num_outs,
bool eot)
{
struct toy_compiler *tc = &gcc->tc;
struct toy_dst mrf_header;
struct toy_src desc;
int sent = 0;
mrf_header = tdst_d(tdst(TOY_FILE_MRF, gcc->first_free_mrf, 0));
gs_COPY8(tc, mrf_header, msg_header);
while (sent < num_outs) {
int mrf = gcc->first_free_mrf + 1;
const int mrf_avail = gcc->last_free_mrf - mrf + 1;
int msg_len, num_entries, i;
bool complete;
num_entries = (num_outs - sent + 1) / 2;
complete = true;
if (num_entries > mrf_avail) {
num_entries = mrf_avail;
complete = false;
}
for (i = 0; i < num_entries; i++) {
gs_COPY4(tc, tdst(TOY_FILE_MRF, mrf + i / 2, 0), 0,
outs[sent + 2 * i], 0);
if (sent + i * 2 + 1 < gcc->shader->out.count) {
gs_COPY4(tc, tdst(TOY_FILE_MRF, mrf + i / 2, 0), 4,
outs[sent + 2 * i + 1], 0);
}
mrf++;
}
/* do not forget the header */
msg_len = num_entries + 1;
if (complete) {
desc = tsrc_imm_mdesc_urb(tc,
eot, msg_len, !eot, true, true, !eot,
false, sent, 0);
}
else {
desc = tsrc_imm_mdesc_urb(tc,
false, msg_len, 0, false, true, false,
false, sent, 0);
}
tc_add2(tc, TOY_OPCODE_URB_WRITE,
(complete) ? dst : tdst_null(), tsrc_from(mrf_header), desc);
sent += num_entries * 2;
}
}
static void
gs_discard(struct gs_compile_context *gcc)
{
struct toy_compiler *tc = &gcc->tc;
struct toy_dst mrf_header;
struct toy_src desc;
mrf_header = tdst_d(tdst(TOY_FILE_MRF, gcc->first_free_mrf, 0));
gs_COPY8(tc, mrf_header, tsrc_from(gcc->vars.urb_write_header));
desc = tsrc_imm_mdesc_urb(tc,
true, 1, 0, true, false, false,
false, 0, 0);
tc_add2(tc, TOY_OPCODE_URB_WRITE,
tdst_null(), tsrc_from(mrf_header), desc);
}
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));
}
}
/**
* 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;
}
}
/**
* Emit instructions to write the color buffers (and the depth buffer).
*/
static void
fs_write_fb(struct fs_compile_context *fcc)
{
struct toy_compiler *tc = &fcc->tc;
int base_mrf = fcc->first_free_mrf;
const struct toy_dst header = tdst_ud(tdst(TOY_FILE_MRF, base_mrf, 0));
bool header_present = false;
struct toy_src desc;
unsigned msg_type, ctrl;
int color_slots[ILO_MAX_DRAW_BUFFERS], num_cbufs;
int pos_slot = -1, cbuf, i;
for (i = 0; i < Elements(color_slots); i++)
color_slots[i] = -1;
for (i = 0; i < fcc->tgsi.num_outputs; i++) {
if (fcc->tgsi.outputs[i].semantic_name == TGSI_SEMANTIC_COLOR) {
assert(fcc->tgsi.outputs[i].semantic_index < Elements(color_slots));
color_slots[fcc->tgsi.outputs[i].semantic_index] = i;
}
else if (fcc->tgsi.outputs[i].semantic_name == TGSI_SEMANTIC_POSITION) {
pos_slot = i;
}
}
num_cbufs = fcc->variant->u.fs.num_cbufs;
/* still need to send EOT (and probably depth) */
if (!num_cbufs)
num_cbufs = 1;
/* we need the header to specify the pixel mask or render target */
if (fcc->tgsi.uses_kill || num_cbufs > 1) {
const struct toy_src r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0));
struct toy_inst *inst;
inst = tc_MOV(tc, header, r0);
inst->mask_ctrl = BRW_MASK_DISABLE;
base_mrf += fcc->num_grf_per_vrf;
/* this is a two-register header */
if (fcc->dispatch_mode == GEN6_WM_8_DISPATCH_ENABLE) {
inst = tc_MOV(tc, tdst_offset(header, 1, 0), tsrc_offset(r0, 1, 0));
inst->mask_ctrl = BRW_MASK_DISABLE;
base_mrf += fcc->num_grf_per_vrf;
}
header_present = true;
}
for (cbuf = 0; cbuf < num_cbufs; cbuf++) {
const int slot =
color_slots[(fcc->tgsi.props.fs_color0_writes_all_cbufs) ? 0 : cbuf];
int mrf = base_mrf, vrf;
struct toy_src src[4];
if (slot >= 0) {
const unsigned undefined_mask =
fcc->tgsi.outputs[slot].undefined_mask;
const int index = fcc->tgsi.outputs[slot].index;
vrf = toy_tgsi_get_vrf(&fcc->tgsi, TGSI_FILE_OUTPUT, 0, index);
if (vrf >= 0) {
const struct toy_src tmp = tsrc(TOY_FILE_VRF, vrf, 0);
tsrc_transpose(tmp, src);
}
else {
/* use (0, 0, 0, 0) */
tsrc_transpose(tsrc_imm_f(0.0f), src);
}
for (i = 0; i < 4; i++) {
const struct toy_dst dst = tdst(TOY_FILE_MRF, mrf, 0);
if (undefined_mask & (1 << i))
src[i] = tsrc_imm_f(0.0f);
tc_MOV(tc, dst, src[i]);
mrf += fcc->num_grf_per_vrf;
}
}
else {
/* use (0, 0, 0, 0) */
for (i = 0; i < 4; i++) {
const struct toy_dst dst = tdst(TOY_FILE_MRF, mrf, 0);
tc_MOV(tc, dst, tsrc_imm_f(0.0f));
mrf += fcc->num_grf_per_vrf;
}
}
/* select BLEND_STATE[rt] */
if (cbuf > 0) {
struct toy_inst *inst;
inst = tc_MOV(tc, tdst_offset(header, 0, 2), tsrc_imm_ud(cbuf));
inst->mask_ctrl = BRW_MASK_DISABLE;
inst->exec_size = BRW_EXECUTE_1;
inst->src[0].rect = TOY_RECT_010;
}
if (cbuf == 0 && pos_slot >= 0) {
const int index = fcc->tgsi.outputs[pos_slot].index;
const struct toy_dst dst = tdst(TOY_FILE_MRF, mrf, 0);
struct toy_src src[4];
int vrf;
vrf = toy_tgsi_get_vrf(&fcc->tgsi, TGSI_FILE_OUTPUT, 0, index);
if (vrf >= 0) {
const struct toy_src tmp = tsrc(TOY_FILE_VRF, vrf, 0);
tsrc_transpose(tmp, src);
}
else {
/* use (0, 0, 0, 0) */
tsrc_transpose(tsrc_imm_f(0.0f), src);
}
/* only Z */
tc_MOV(tc, dst, src[2]);
mrf += fcc->num_grf_per_vrf;
}
msg_type = (fcc->dispatch_mode == GEN6_WM_16_DISPATCH_ENABLE) ?
BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD16_SINGLE_SOURCE :
BRW_DATAPORT_RENDER_TARGET_WRITE_SIMD8_SINGLE_SOURCE_SUBSPAN01;
ctrl = (cbuf == num_cbufs - 1) << 12 |
msg_type << 8;
desc = tsrc_imm_mdesc_data_port(tc, cbuf == num_cbufs - 1,
mrf - fcc->first_free_mrf, 0,
header_present, false,
GEN6_DATAPORT_WRITE_MESSAGE_RENDER_TARGET_WRITE,
ctrl, ILO_WM_DRAW_SURFACE(cbuf));
tc_add2(tc, TOY_OPCODE_FB_WRITE, tdst_null(),
tsrc(TOY_FILE_MRF, fcc->first_free_mrf, 0), desc);
}
}
