static void
vs_lower_opcode_tgsi_const_gen6(struct vs_compile_context *vcc,
struct toy_dst dst, int dim,
struct toy_src idx)
{
const struct toy_dst header =
tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf, 0));
const struct toy_dst block_offsets =
tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf + 1, 0));
const struct toy_src r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0));
struct toy_compiler *tc = &vcc->tc;
unsigned msg_type, msg_ctrl, msg_len;
struct toy_inst *inst;
struct toy_src desc;
if (vs_lower_opcode_tgsi_const_pcb(vcc, dst, dim, idx))
return;
/* set message header */
inst = tc_MOV(tc, header, r0);
inst->mask_ctrl = GEN6_MASKCTRL_NOMASK;
/* set block offsets */
tc_MOV(tc, block_offsets, idx);
msg_type = GEN6_MSG_DP_OWORD_DUAL_BLOCK_READ;
msg_ctrl = GEN6_MSG_DP_OWORD_DUAL_BLOCK_SIZE_1;
msg_len = 2;
desc = tsrc_imm_mdesc_data_port(tc, false, msg_len, 1, true, false,
msg_type, msg_ctrl, vcc->shader->bt.const_base + dim);
tc_SEND(tc, dst, tsrc_from(header), desc, vcc->const_cache);
}
static void
vs_lower_opcode_tgsi_const_gen7(struct vs_compile_context *vcc,
struct toy_dst dst, int dim,
struct toy_src idx)
{
struct toy_compiler *tc = &vcc->tc;
const struct toy_dst offset =
tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf, 0));
struct toy_src desc;
if (vs_lower_opcode_tgsi_const_pcb(vcc, dst, dim, idx))
return;
/*
* In 259b65e2e7938de4aab323033cfe2b33369ddb07, pull constant load was
* changed from OWord Dual 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 */
tc_MOV(tc, offset, idx);
desc = tsrc_imm_mdesc_sampler(tc, 1, 1, false,
GEN6_MSG_SAMPLER_SIMD4X2,
GEN6_MSG_SAMPLER_LD,
0,
vcc->shader->bt.const_base + dim);
tc_SEND(tc, dst, tsrc_from(offset), desc, GEN6_SFID_SAMPLER);
}
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 void
fs_lower_opcode_tgsi_imm(struct fs_compile_context *fcc,
struct toy_dst dst, int idx)
{
const uint32_t *imm;
struct toy_dst real_dst[4];
int ch;
imm = toy_tgsi_get_imm(&fcc->tgsi, idx, NULL);
tdst_transpose(dst, real_dst);
/* raw moves */
for (ch = 0; ch < 4; ch++)
tc_MOV(&fcc->tc, tdst_ud(real_dst[ch]), tsrc_imm_ud(imm[ch]));
}
static void
vs_lower_opcode_tgsi_imm(struct vs_compile_context *vcc,
struct toy_dst dst, int idx)
{
const uint32_t *imm;
int ch;
imm = toy_tgsi_get_imm(&vcc->tgsi, idx, NULL);
for (ch = 0; ch < 4; ch++) {
/* raw moves */
tc_MOV(&vcc->tc,
tdst_writemask(tdst_ud(dst), 1 << ch),
tsrc_imm_ud(imm[ch]));
}
}
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_tgsi_imm(struct gs_compile_context *gcc,
struct toy_dst dst, int idx)
{
const uint32_t *imm;
int ch;
imm = toy_tgsi_get_imm(&gcc->tgsi, idx, NULL);
for (ch = 0; ch < 4; ch++) {
struct toy_inst *inst;
/* raw moves */
inst = tc_MOV(&gcc->tc,
tdst_writemask(tdst_ud(dst), 1 << ch),
tsrc_imm_ud(imm[ch]));
inst->access_mode = GEN6_ALIGN_16;
}
}
static void
cs_dummy(struct cs_compile_context *ccc)
{
struct toy_compiler *tc = &ccc->tc;
struct toy_dst header;
struct toy_src r0, desc;
struct toy_inst *inst;
header = tdst_ud(tdst(TOY_FILE_MRF, ccc->first_free_mrf, 0));
r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0));
inst = tc_MOV(tc, header, r0);
inst->exec_size = GEN6_EXECSIZE_8;
inst->mask_ctrl = GEN6_MASKCTRL_NOMASK;
desc = tsrc_imm_mdesc(tc, true, 1, 0, true,
GEN6_MSG_TS_RESOURCE_SELECT_NO_DEREF |
GEN6_MSG_TS_REQUESTER_TYPE_ROOT |
GEN6_MSG_TS_OPCODE_DEREF);
tc_SEND(tc, tdst_null(), tsrc_from(header), desc, GEN6_SFID_SPAWNER);
}
/**
* 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);
}
}