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
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
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
vs_lower_opcode_tgsi_sv(struct vs_compile_context *vcc,
struct toy_dst dst, int dim, int idx)
{
struct toy_compiler *tc = &vcc->tc;
const struct toy_tgsi *tgsi = &vcc->tgsi;
int slot;
assert(!dim);
slot = toy_tgsi_find_system_value(tgsi, idx);
if (slot < 0)
return;
switch (tgsi->system_values[slot].semantic_name) {
case TGSI_SEMANTIC_INSTANCEID:
case TGSI_SEMANTIC_VERTEXID:
/*
* In 3DSTATE_VERTEX_ELEMENTS, we prepend an extra vertex element for
* the generated IDs, with VID in the X channel and IID in the Y
* channel.
*/
{
const int grf = vcc->first_vue_grf;
const struct toy_src src = tsrc(TOY_FILE_GRF, grf, 0);
const enum toy_swizzle swizzle =
(tgsi->system_values[slot].semantic_name ==
TGSI_SEMANTIC_INSTANCEID) ? TOY_SWIZZLE_Y : TOY_SWIZZLE_X;
tc_MOV(tc, tdst_d(dst), tsrc_d(tsrc_swizzle1(src, swizzle)));
}
break;
case TGSI_SEMANTIC_PRIMID:
default:
tc_fail(tc, "unhandled system value");
tc_MOV(tc, dst, tsrc_imm_d(0));
break;
}
}
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 bool
gs_compile_passthrough(struct gs_compile_context *gcc)
{
struct toy_compiler *tc = &gcc->tc;
struct ilo_shader *sh = gcc->shader;
gcc->is_static = true;
gcc->static_data.total_vertices = gcc->in_vue_count;
gcc->static_data.total_prims = 1;
gcc->static_data.last_vertex[0] = 1 << (gcc->in_vue_count - 1);
gs_init_vars(gcc);
gs_ff_sync(gcc, tdst_d(gcc->vars.tmp), tsrc_imm_d(gcc->static_data.total_prims));
gs_COPY1(tc, gcc->vars.urb_write_header, 0, tsrc_from(tdst_d(gcc->vars.tmp)), 0);
if (gcc->write_so)
gs_COPY4(tc, gcc->vars.so_index, 0, tsrc_from(tdst_d(gcc->vars.tmp)), 1);
{
int vert, attr;
for (vert = 0; vert < gcc->out_vue_min_count; vert++) {
for (attr = 0; attr < gcc->shader->out.count; attr++) {
tc_MOV(tc, tdst_from(gcc->vars.tgsi_outs[attr]),
tsrc_offset(gcc->payload.vues[vert], attr / 2, (attr % 2) * 4));
}
gs_lower_opcode_emit(gcc, NULL);
}
gs_lower_opcode_endprim(gcc, NULL);
}
if (!gcc->write_vue)
gs_discard(gcc);
gs_lower_virtual_opcodes(gcc);
toy_compiler_legalize_for_ra(tc);
toy_compiler_optimize(tc);
toy_compiler_allocate_registers(tc,
gcc->first_free_grf,
gcc->last_free_grf,
1);
toy_compiler_legalize_for_asm(tc);
if (tc->fail) {
ilo_err("failed to translate GS TGSI tokens: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_GS) {
int i;
ilo_printf("VUE count %d, VUE size %d\n",
gcc->in_vue_count, gcc->in_vue_size);
ilo_printf("%srasterizer discard\n",
(gcc->variant->u.gs.rasterizer_discard) ? "" : "no ");
for (i = 0; i < gcc->so_info->num_outputs; i++) {
ilo_printf("SO[%d] = OUT[%d]\n", i,
gcc->so_info->output[i].register_index);
}
ilo_printf("legalized instructions:\n");
toy_compiler_dump(tc);
ilo_printf("\n");
}
sh->kernel = toy_compiler_assemble(tc, &sh->kernel_size);
if (!sh->kernel) {
ilo_err("failed to compile GS: %s\n", tc->reason);
return false;
}
if (ilo_debug & ILO_DEBUG_GS) {
ilo_printf("disassembly:\n");
toy_compiler_disassemble(tc->dev, sh->kernel, sh->kernel_size, false);
ilo_printf("\n");
}
return true;
}
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);
}
/**
* 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;
}
/**
* Emit instructions to move sampling parameters to the message registers.
*/
static int
vs_add_sampler_params(struct toy_compiler *tc, int msg_type, int base_mrf,
struct toy_src coords, int num_coords,
struct toy_src bias_or_lod, struct toy_src ref_or_si,
struct toy_src ddx, struct toy_src ddy, int num_derivs)
{
const unsigned coords_writemask = (1 << num_coords) - 1;
struct toy_dst m[3];
int num_params, i;
assert(num_coords <= 4);
assert(num_derivs <= 3 && num_derivs <= num_coords);
for (i = 0; i < Elements(m); i++)
m[i] = tdst(TOY_FILE_MRF, base_mrf + i, 0);
switch (msg_type) {
case GEN6_MSG_SAMPLER_SAMPLE_L:
tc_MOV(tc, tdst_writemask(m[0], coords_writemask), coords);
tc_MOV(tc, tdst_writemask(m[1], TOY_WRITEMASK_X), bias_or_lod);
num_params = 5;
break;
case GEN6_MSG_SAMPLER_SAMPLE_D:
tc_MOV(tc, tdst_writemask(m[0], coords_writemask), coords);
tc_MOV(tc, tdst_writemask(m[1], TOY_WRITEMASK_XZ),
tsrc_swizzle(ddx, 0, 0, 1, 1));
tc_MOV(tc, tdst_writemask(m[1], TOY_WRITEMASK_YW),
tsrc_swizzle(ddy, 0, 0, 1, 1));
if (num_derivs > 2) {
tc_MOV(tc, tdst_writemask(m[2], TOY_WRITEMASK_X),
tsrc_swizzle1(ddx, 2));
tc_MOV(tc, tdst_writemask(m[2], TOY_WRITEMASK_Y),
tsrc_swizzle1(ddy, 2));
}
num_params = 4 + num_derivs * 2;
break;
case GEN6_MSG_SAMPLER_SAMPLE_L_C:
tc_MOV(tc, tdst_writemask(m[0], coords_writemask), coords);
tc_MOV(tc, tdst_writemask(m[1], TOY_WRITEMASK_X), ref_or_si);
tc_MOV(tc, tdst_writemask(m[1], TOY_WRITEMASK_Y), bias_or_lod);
num_params = 6;
break;
case GEN6_MSG_SAMPLER_LD:
assert(num_coords <= 3);
tc_MOV(tc, tdst_writemask(tdst_d(m[0]), coords_writemask), coords);
tc_MOV(tc, tdst_writemask(tdst_d(m[0]), TOY_WRITEMASK_W), bias_or_lod);
if (ilo_dev_gen(tc->dev) >= ILO_GEN(7)) {
num_params = 4;
}
else {
tc_MOV(tc, tdst_writemask(tdst_d(m[1]), TOY_WRITEMASK_X), ref_or_si);
num_params = 5;
}
break;
case GEN6_MSG_SAMPLER_RESINFO:
tc_MOV(tc, tdst_writemask(tdst_d(m[0]), TOY_WRITEMASK_X), bias_or_lod);
num_params = 1;
break;
default:
tc_fail(tc, "unknown sampler opcode");
num_params = 0;
break;
}
return (num_params + 3) / 4;
}
static int
fs_add_sampler_params_gen7(struct toy_compiler *tc, int msg_type,
int base_mrf, int param_size,
struct toy_src *coords, int num_coords,
struct toy_src bias_or_lod, struct toy_src ref_or_si,
struct toy_src *ddx, struct toy_src *ddy,
int num_derivs)
{
int num_params, i;
assert(num_coords <= 4);
assert(num_derivs <= 3 && num_derivs <= num_coords);
#define SAMPLER_PARAM(p) (tdst(TOY_FILE_MRF, base_mrf + (p) * param_size, 0))
switch (msg_type) {
case GEN5_SAMPLER_MESSAGE_SAMPLE:
for (i = 0; i < num_coords; i++)
tc_MOV(tc, SAMPLER_PARAM(i), coords[i]);
num_params = num_coords;
break;
case GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS:
case GEN5_SAMPLER_MESSAGE_SAMPLE_LOD:
tc_MOV(tc, SAMPLER_PARAM(0), bias_or_lod);
for (i = 0; i < num_coords; i++)
tc_MOV(tc, SAMPLER_PARAM(1 + i), coords[i]);
num_params = 1 + num_coords;
break;
case GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE:
tc_MOV(tc, SAMPLER_PARAM(0), ref_or_si);
for (i = 0; i < num_coords; i++)
tc_MOV(tc, SAMPLER_PARAM(1 + i), coords[i]);
num_params = 1 + num_coords;
break;
case GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS:
for (i = 0; i < num_coords; i++) {
tc_MOV(tc, SAMPLER_PARAM(i * 3), coords[i]);
if (i < num_derivs) {
tc_MOV(tc, SAMPLER_PARAM(i * 3 + 1), ddx[i]);
tc_MOV(tc, SAMPLER_PARAM(i * 3 + 2), ddy[i]);
}
}
num_params = num_coords * 3 - ((num_coords > num_derivs) ? 2 : 0);
break;
case GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE:
case GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE:
tc_MOV(tc, SAMPLER_PARAM(0), ref_or_si);
tc_MOV(tc, SAMPLER_PARAM(1), bias_or_lod);
for (i = 0; i < num_coords; i++)
tc_MOV(tc, SAMPLER_PARAM(2 + i), coords[i]);
num_params = 2 + num_coords;
break;
case GEN5_SAMPLER_MESSAGE_SAMPLE_LD:
assert(num_coords >= 1 && num_coords <= 3);
tc_MOV(tc, tdst_d(SAMPLER_PARAM(0)), coords[0]);
tc_MOV(tc, tdst_d(SAMPLER_PARAM(1)), bias_or_lod);
for (i = 1; i < num_coords; i++)
tc_MOV(tc, tdst_d(SAMPLER_PARAM(1 + i)), coords[i]);
num_params = 1 + num_coords;
break;
case GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO:
tc_MOV(tc, tdst_d(SAMPLER_PARAM(0)), bias_or_lod);
num_params = 1;
break;
default:
tc_fail(tc, "unknown sampler opcode");
num_params = 0;
break;
}
#undef SAMPLER_PARAM
return num_params * param_size;
}