/**
* Initialize the instruction template, from which tc_add() initializes the
* newly added instructions.
*/
static void
tc_init_inst_templ(struct toy_compiler *tc)
{
struct toy_inst *templ = &tc->templ;
int i;
templ->opcode = GEN6_OPCODE_NOP;
templ->access_mode = GEN6_ALIGN_1;
templ->mask_ctrl = GEN6_MASKCTRL_NORMAL;
templ->dep_ctrl = GEN6_DEPCTRL_NORMAL;
templ->qtr_ctrl = GEN6_QTRCTRL_1Q;
templ->thread_ctrl = GEN6_THREADCTRL_NORMAL;
templ->pred_ctrl = GEN6_PREDCTRL_NONE;
templ->pred_inv = false;
templ->exec_size = GEN6_EXECSIZE_1;
templ->cond_modifier = GEN6_COND_NORMAL;
templ->acc_wr_ctrl = false;
templ->saturate = false;
templ->marker = false;
templ->dst = tdst_null();
for (i = 0; i < Elements(templ->src); i++)
templ->src[i] = tsrc_null();
for (i = 0; i < Elements(templ->tex.offsets); i++)
templ->tex.offsets[i] = tsrc_null();
list_inithead(&templ->list);
}
/**
* Set up message registers and return the message descriptor for sampling.
*/
static struct toy_src
vs_prepare_tgsi_sampling(struct vs_compile_context *vcc,
const struct toy_inst *inst,
int base_mrf, unsigned *ret_sampler_index)
{
struct toy_compiler *tc = &vcc->tc;
unsigned simd_mode, msg_type, msg_len, sampler_index, binding_table_index;
struct toy_src coords, ddx, ddy, bias_or_lod, ref_or_si;
int num_coords, ref_pos, num_derivs;
int sampler_src;
simd_mode = GEN6_MSG_SAMPLER_SIMD4X2;
coords = inst->src[0];
ddx = tsrc_null();
ddy = tsrc_null();
bias_or_lod = tsrc_null();
ref_or_si = tsrc_null();
num_derivs = 0;
sampler_src = 1;
num_coords = tgsi_util_get_texture_coord_dim(inst->tex.target, &ref_pos);
/* extract the parameters */
switch (inst->opcode) {
case TOY_OPCODE_TGSI_TXD:
if (ref_pos >= 0) {
assert(ref_pos < 4);
msg_type = GEN7_MSG_SAMPLER_SAMPLE_D_C;
ref_or_si = tsrc_swizzle1(coords, ref_pos);
if (ilo_dev_gen(tc->dev) < ILO_GEN(7.5))
tc_fail(tc, "TXD with shadow sampler not supported");
}
else {
msg_type = GEN6_MSG_SAMPLER_SAMPLE_D;
}
ddx = inst->src[1];
ddy = inst->src[2];
num_derivs = num_coords;
sampler_src = 3;
break;
case TOY_OPCODE_TGSI_TXL:
if (ref_pos >= 0) {
assert(ref_pos < 3);
msg_type = GEN6_MSG_SAMPLER_SAMPLE_L_C;
ref_or_si = tsrc_swizzle1(coords, ref_pos);
}
else {
msg_type = GEN6_MSG_SAMPLER_SAMPLE_L;
}
bias_or_lod = tsrc_swizzle1(coords, TOY_SWIZZLE_W);
break;
case TOY_OPCODE_TGSI_TXF:
msg_type = GEN6_MSG_SAMPLER_LD;
switch (inst->tex.target) {
case TGSI_TEXTURE_2D_MSAA:
case TGSI_TEXTURE_2D_ARRAY_MSAA:
assert(ref_pos >= 0 && ref_pos < 4);
/* lod is always 0 */
bias_or_lod = tsrc_imm_d(0);
ref_or_si = tsrc_swizzle1(coords, ref_pos);
break;
default:
bias_or_lod = tsrc_swizzle1(coords, TOY_SWIZZLE_W);
break;
}
/* offset the coordinates */
if (!tsrc_is_null(inst->tex.offsets[0])) {
struct toy_dst tmp;
tmp = tc_alloc_tmp(tc);
tc_ADD(tc, tmp, coords, inst->tex.offsets[0]);
coords = tsrc_from(tmp);
}
sampler_src = 1;
break;
case TOY_OPCODE_TGSI_TXQ:
msg_type = GEN6_MSG_SAMPLER_RESINFO;
num_coords = 0;
bias_or_lod = tsrc_swizzle1(coords, TOY_SWIZZLE_X);
break;
case TOY_OPCODE_TGSI_TXQ_LZ:
msg_type = GEN6_MSG_SAMPLER_RESINFO;
num_coords = 0;
sampler_src = 0;
break;
case TOY_OPCODE_TGSI_TXL2:
if (ref_pos >= 0) {
assert(ref_pos < 4);
msg_type = GEN6_MSG_SAMPLER_SAMPLE_L_C;
ref_or_si = tsrc_swizzle1(coords, ref_pos);
}
else {
msg_type = GEN6_MSG_SAMPLER_SAMPLE_L;
}
bias_or_lod = tsrc_swizzle1(inst->src[1], TOY_SWIZZLE_X);
sampler_src = 2;
break;
default:
assert(!"unhandled sampling opcode");
if (ret_sampler_index)
*ret_sampler_index = 0;
return tsrc_null();
break;
}
assert(inst->src[sampler_src].file == TOY_FILE_IMM);
sampler_index = inst->src[sampler_src].val32;
binding_table_index = vcc->shader->bt.tex_base + sampler_index;
/*
* From the Sandy Bridge PRM, volume 4 part 1, page 18:
*
* "Note that the (cube map) coordinates delivered to the sampling
* engine must already have been divided by the component with the
* largest absolute value."
*/
switch (inst->tex.target) {
case TGSI_TEXTURE_CUBE:
case TGSI_TEXTURE_SHADOWCUBE:
case TGSI_TEXTURE_CUBE_ARRAY:
case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
/* TXQ does not need coordinates */
if (num_coords >= 3) {
struct toy_dst tmp, max;
struct toy_src abs_coords[3];
unsigned i;
tmp = tc_alloc_tmp(tc);
max = tdst_writemask(tmp, TOY_WRITEMASK_W);
for (i = 0; i < 3; i++)
abs_coords[i] = tsrc_absolute(tsrc_swizzle1(coords, i));
tc_SEL(tc, max, abs_coords[0], abs_coords[0], GEN6_COND_GE);
tc_SEL(tc, max, tsrc_from(max), abs_coords[0], GEN6_COND_GE);
tc_INV(tc, max, tsrc_from(max));
for (i = 0; i < 3; i++)
tc_MUL(tc, tdst_writemask(tmp, 1 << i), coords, tsrc_from(max));
coords = tsrc_from(tmp);
}
break;
}
/* set up sampler parameters */
msg_len = vs_add_sampler_params(tc, msg_type, base_mrf,
coords, num_coords, bias_or_lod, ref_or_si, ddx, ddy, num_derivs);
/*
* From the Sandy Bridge PRM, volume 4 part 1, page 136:
*
* "The maximum message length allowed to the sampler is 11. This would
* disallow sample_d, sample_b_c, and sample_l_c with a SIMD Mode of
* SIMD16."
*/
if (msg_len > 11)
tc_fail(tc, "maximum length for messages to the sampler is 11");
if (ret_sampler_index)
*ret_sampler_index = sampler_index;
return tsrc_imm_mdesc_sampler(tc, msg_len, 1,
false, simd_mode, msg_type, sampler_index, binding_table_index);
}
static void
fs_lower_opcode_tgsi_sampling(struct fs_compile_context *fcc,
struct toy_inst *inst)
{
struct toy_compiler *tc = &fcc->tc;
struct toy_dst dst[4], tmp[4];
struct toy_src desc;
unsigned sampler_index;
int swizzles[4], i;
bool need_filter;
desc = fs_prepare_tgsi_sampling(tc, inst,
fcc->first_free_mrf,
fcc->variant->saturate_tex_coords,
&sampler_index);
switch (inst->opcode) {
case TOY_OPCODE_TGSI_TXF:
case TOY_OPCODE_TGSI_TXQ:
case TOY_OPCODE_TGSI_TXQ_LZ:
need_filter = false;
break;
default:
need_filter = true;
break;
}
toy_compiler_lower_to_send(tc, inst, false, BRW_SFID_SAMPLER);
inst->src[0] = tsrc(TOY_FILE_MRF, fcc->first_free_mrf, 0);
inst->src[1] = desc;
for (i = 2; i < Elements(inst->src); i++)
inst->src[i] = tsrc_null();
/* write to temps first */
tc_alloc_tmp4(tc, tmp);
tdst_transpose(inst->dst, dst);
inst->dst = tmp[0];
tc_move_inst(tc, inst);
if (need_filter) {
assert(sampler_index < fcc->variant->num_sampler_views);
swizzles[0] = fcc->variant->sampler_view_swizzles[sampler_index].r;
swizzles[1] = fcc->variant->sampler_view_swizzles[sampler_index].g;
swizzles[2] = fcc->variant->sampler_view_swizzles[sampler_index].b;
swizzles[3] = fcc->variant->sampler_view_swizzles[sampler_index].a;
}
else {
swizzles[0] = PIPE_SWIZZLE_RED;
swizzles[1] = PIPE_SWIZZLE_GREEN;
swizzles[2] = PIPE_SWIZZLE_BLUE;
swizzles[3] = PIPE_SWIZZLE_ALPHA;
}
/* swizzle the results */
for (i = 0; i < 4; i++) {
switch (swizzles[i]) {
case PIPE_SWIZZLE_ZERO:
tc_MOV(tc, dst[i], tsrc_imm_f(0.0f));
break;
case PIPE_SWIZZLE_ONE:
tc_MOV(tc, dst[i], tsrc_imm_f(1.0f));
break;
default:
tc_MOV(tc, dst[i], tsrc_from(tmp[swizzles[i]]));
break;
}
}
}
/**
* Set up message registers and return the message descriptor for sampling.
*/
static struct toy_src
fs_prepare_tgsi_sampling(struct toy_compiler *tc, const struct toy_inst *inst,
int base_mrf, const uint32_t *saturate_coords,
unsigned *ret_sampler_index)
{
unsigned simd_mode, msg_type, msg_len, sampler_index, binding_table_index;
struct toy_src coords[4], ddx[4], ddy[4], bias_or_lod, ref_or_si;
int num_coords, ref_pos, num_derivs;
int sampler_src, param_size, i;
switch (inst->exec_size) {
case BRW_EXECUTE_8:
simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD8;
param_size = 1;
break;
case BRW_EXECUTE_16:
simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
param_size = 2;
break;
default:
tc_fail(tc, "unsupported execute size for sampling");
return tsrc_null();
break;
}
num_coords = toy_tgsi_get_texture_coord_dim(inst->tex.target, &ref_pos);
tsrc_transpose(inst->src[0], coords);
bias_or_lod = tsrc_null();
ref_or_si = tsrc_null();
num_derivs = 0;
sampler_src = 1;
/*
* For TXD,
*
* src0 := (x, y, z, w)
* src1 := ddx
* src2 := ddy
* src3 := sampler
*
* For TEX2, TXB2, and TXL2,
*
* src0 := (x, y, z, w)
* src1 := (v or bias or lod, ...)
* src2 := sampler
*
* For TEX, TXB, TXL, and TXP,
*
* src0 := (x, y, z, w or bias or lod or projection)
* src1 := sampler
*
* For TXQ,
*
* src0 := (lod, ...)
* src1 := sampler
*
* For TXQ_LZ,
*
* src0 := sampler
*
* And for TXF,
*
* src0 := (x, y, z, w or lod)
* src1 := sampler
*
* State trackers should not generate opcode+texture combinations with
* which the two definitions conflict (e.g., TXB with SHADOW2DARRAY).
*/
switch (inst->opcode) {
case TOY_OPCODE_TGSI_TEX:
if (ref_pos >= 0) {
assert(ref_pos < 4);
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE;
ref_or_si = coords[ref_pos];
}
else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
}
break;
case TOY_OPCODE_TGSI_TXD:
if (ref_pos >= 0)
tc_fail(tc, "TXD with shadow sampler not supported");
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
tsrc_transpose(inst->src[1], ddx);
tsrc_transpose(inst->src[2], ddy);
num_derivs = num_coords;
sampler_src = 3;
break;
case TOY_OPCODE_TGSI_TXP:
if (ref_pos >= 0) {
assert(ref_pos < 3);
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE;
ref_or_si = coords[ref_pos];
}
else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
}
/* project the coordinates */
{
struct toy_dst tmp[4];
tc_alloc_tmp4(tc, tmp);
tc_INV(tc, tmp[3], coords[3]);
for (i = 0; i < num_coords && i < 3; i++) {
tc_MUL(tc, tmp[i], coords[i], tsrc_from(tmp[3]));
coords[i] = tsrc_from(tmp[i]);
}
if (ref_pos >= i) {
tc_MUL(tc, tmp[ref_pos], ref_or_si, tsrc_from(tmp[3]));
ref_or_si = tsrc_from(tmp[ref_pos]);
}
}
break;
case TOY_OPCODE_TGSI_TXB:
if (ref_pos >= 0) {
assert(ref_pos < 3);
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE;
ref_or_si = coords[ref_pos];
}
else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS;
}
bias_or_lod = coords[3];
break;
case TOY_OPCODE_TGSI_TXL:
if (ref_pos >= 0) {
assert(ref_pos < 3);
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
ref_or_si = coords[ref_pos];
}
else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
}
bias_or_lod = coords[3];
break;
case TOY_OPCODE_TGSI_TXF:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
switch (inst->tex.target) {
case TGSI_TEXTURE_2D_MSAA:
case TGSI_TEXTURE_2D_ARRAY_MSAA:
assert(ref_pos >= 0 && ref_pos < 4);
/* lod is always 0 */
bias_or_lod = tsrc_imm_d(0);
ref_or_si = coords[ref_pos];
break;
default:
bias_or_lod = coords[3];
break;
}
/* offset the coordinates */
if (!tsrc_is_null(inst->tex.offsets[0])) {
struct toy_dst tmp[4];
struct toy_src offsets[4];
tc_alloc_tmp4(tc, tmp);
tsrc_transpose(inst->tex.offsets[0], offsets);
for (i = 0; i < num_coords; i++) {
tc_ADD(tc, tmp[i], coords[i], offsets[i]);
coords[i] = tsrc_from(tmp[i]);
}
}
sampler_src = 1;
break;
case TOY_OPCODE_TGSI_TXQ:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
num_coords = 0;
bias_or_lod = coords[0];
break;
case TOY_OPCODE_TGSI_TXQ_LZ:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
num_coords = 0;
sampler_src = 0;
break;
case TOY_OPCODE_TGSI_TEX2:
if (ref_pos >= 0) {
assert(ref_pos < 5);
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_COMPARE;
if (ref_pos >= 4) {
struct toy_src src1[4];
tsrc_transpose(inst->src[1], src1);
ref_or_si = src1[ref_pos - 4];
}
else {
ref_or_si = coords[ref_pos];
}
}
else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE;
}
sampler_src = 2;
break;
case TOY_OPCODE_TGSI_TXB2:
if (ref_pos >= 0) {
assert(ref_pos < 4);
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS_COMPARE;
ref_or_si = coords[ref_pos];
}
else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_BIAS;
}
{
struct toy_src src1[4];
tsrc_transpose(inst->src[1], src1);
bias_or_lod = src1[0];
}
sampler_src = 2;
break;
case TOY_OPCODE_TGSI_TXL2:
if (ref_pos >= 0) {
assert(ref_pos < 4);
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
ref_or_si = coords[ref_pos];
}
else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
}
{
struct toy_src src1[4];
tsrc_transpose(inst->src[1], src1);
bias_or_lod = src1[0];
}
sampler_src = 2;
break;
default:
assert(!"unhandled sampling opcode");
return tsrc_null();
break;
}
assert(inst->src[sampler_src].file == TOY_FILE_IMM);
sampler_index = inst->src[sampler_src].val32;
binding_table_index = ILO_WM_TEXTURE_SURFACE(sampler_index);
/*
* From the Sandy Bridge PRM, volume 4 part 1, page 18:
*
* "Note that the (cube map) coordinates delivered to the sampling
* engine must already have been divided by the component with the
* largest absolute value."
*/
switch (inst->tex.target) {
case TGSI_TEXTURE_CUBE:
case TGSI_TEXTURE_SHADOWCUBE:
case TGSI_TEXTURE_CUBE_ARRAY:
case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
/* TXQ does not need coordinates */
if (num_coords >= 3) {
struct toy_dst tmp[4];
tc_alloc_tmp4(tc, tmp);
tc_SEL(tc, tmp[3], tsrc_absolute(coords[0]),
tsrc_absolute(coords[1]), BRW_CONDITIONAL_GE);
tc_SEL(tc, tmp[3], tsrc_from(tmp[3]),
tsrc_absolute(coords[2]), BRW_CONDITIONAL_GE);
tc_INV(tc, tmp[3], tsrc_from(tmp[3]));
for (i = 0; i < 3; i++) {
tc_MUL(tc, tmp[i], coords[i], tsrc_from(tmp[3]));
coords[i] = tsrc_from(tmp[i]);
}
}
break;
}
/*
* Saturate (s, t, r). saturate_coords is set for sampler and coordinate
* that uses linear filtering and PIPE_TEX_WRAP_CLAMP respectively. It is
* so that sampling outside the border gets the correct colors.
*/
for (i = 0; i < MIN2(num_coords, 3); i++) {
bool is_rect;
if (!(saturate_coords[i] & (1 << sampler_index)))
continue;
switch (inst->tex.target) {
case TGSI_TEXTURE_RECT:
case TGSI_TEXTURE_SHADOWRECT:
is_rect = true;
break;
default:
is_rect = false;
break;
}
if (is_rect) {
struct toy_src min, max;
struct toy_dst tmp;
tc_fail(tc, "GL_CLAMP with rectangle texture unsupported");
tmp = tc_alloc_tmp(tc);
/* saturate to [0, width] or [0, height] */
/* TODO TXQ? */
min = tsrc_imm_f(0.0f);
max = tsrc_imm_f(2048.0f);
tc_SEL(tc, tmp, coords[i], min, BRW_CONDITIONAL_G);
tc_SEL(tc, tmp, tsrc_from(tmp), max, BRW_CONDITIONAL_L);
coords[i] = tsrc_from(tmp);
}
else {
struct toy_dst tmp;
struct toy_inst *inst2;
tmp = tc_alloc_tmp(tc);
/* saturate to [0.0f, 1.0f] */
inst2 = tc_MOV(tc, tmp, coords[i]);
inst2->saturate = true;
coords[i] = tsrc_from(tmp);
}
}
/* set up sampler parameters */
if (tc->gen >= ILO_GEN(7)) {
msg_len = fs_add_sampler_params_gen7(tc, msg_type, base_mrf, param_size,
coords, num_coords, bias_or_lod, ref_or_si, ddx, ddy, num_derivs);
}
else {
msg_len = fs_add_sampler_params_gen6(tc, msg_type, base_mrf, param_size,
coords, num_coords, bias_or_lod, ref_or_si, ddx, ddy, num_derivs);
}
/*
* From the Sandy Bridge PRM, volume 4 part 1, page 136:
*
* "The maximum message length allowed to the sampler is 11. This would
* disallow sample_d, sample_b_c, and sample_l_c with a SIMD Mode of
* SIMD16."
*/
if (msg_len > 11)
tc_fail(tc, "maximum length for messages to the sampler is 11");
if (ret_sampler_index)
*ret_sampler_index = sampler_index;
return tsrc_imm_mdesc_sampler(tc, msg_len, 4 * param_size,
false, simd_mode, msg_type, sampler_index, binding_table_index);
}
