static nir_shader *
build_resolve_fragment_shader(struct radv_device *dev, bool is_integer, int samples)
{
nir_builder b;
char name[64];
const struct glsl_type *vec2 = glsl_vector_type(GLSL_TYPE_FLOAT, 2);
const struct glsl_type *vec4 = glsl_vec4_type();
const struct glsl_type *sampler_type = glsl_sampler_type(GLSL_SAMPLER_DIM_MS,
false,
false,
GLSL_TYPE_FLOAT);
snprintf(name, 64, "meta_resolve_fs-%d-%s", samples, is_integer ? "int" : "float");
nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
b.shader->info.name = ralloc_strdup(b.shader, name);
nir_variable *input_img = nir_variable_create(b.shader, nir_var_uniform,
sampler_type, "s_tex");
input_img->data.descriptor_set = 0;
input_img->data.binding = 0;
nir_variable *fs_pos_in = nir_variable_create(b.shader, nir_var_shader_in, vec2, "fs_pos_in");
fs_pos_in->data.location = VARYING_SLOT_POS;
nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
vec4, "f_color");
color_out->data.location = FRAG_RESULT_DATA0;
nir_ssa_def *pos_in = nir_load_var(&b, fs_pos_in);
nir_intrinsic_instr *src_offset = nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_push_constant);
nir_intrinsic_set_base(src_offset, 0);
nir_intrinsic_set_range(src_offset, 8);
src_offset->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
src_offset->num_components = 2;
nir_ssa_dest_init(&src_offset->instr, &src_offset->dest, 2, 32, "src_offset");
nir_builder_instr_insert(&b, &src_offset->instr);
nir_ssa_def *pos_int = nir_f2i32(&b, pos_in);
nir_ssa_def *img_coord = nir_channels(&b, nir_iadd(&b, pos_int, &src_offset->dest.ssa), 0x3);
nir_variable *color = nir_local_variable_create(b.impl, glsl_vec4_type(), "color");
radv_meta_build_resolve_shader_core(&b, is_integer, samples, input_img,
color, img_coord);
nir_ssa_def *outval = nir_load_var(&b, color);
nir_store_var(&b, color_out, outval, 0xf);
return b.shader;
}
static struct vtn_ssa_value *
matrix_multiply(struct vtn_builder *b,
struct vtn_ssa_value *_src0, struct vtn_ssa_value *_src1)
{
struct vtn_ssa_value *src0 = wrap_matrix(b, _src0);
struct vtn_ssa_value *src1 = wrap_matrix(b, _src1);
struct vtn_ssa_value *src0_transpose = wrap_matrix(b, _src0->transposed);
struct vtn_ssa_value *src1_transpose = wrap_matrix(b, _src1->transposed);
unsigned src0_rows = glsl_get_vector_elements(src0->type);
unsigned src0_columns = glsl_get_matrix_columns(src0->type);
unsigned src1_columns = glsl_get_matrix_columns(src1->type);
const struct glsl_type *dest_type;
if (src1_columns > 1) {
dest_type = glsl_matrix_type(glsl_get_base_type(src0->type),
src0_rows, src1_columns);
} else {
dest_type = glsl_vector_type(glsl_get_base_type(src0->type), src0_rows);
}
struct vtn_ssa_value *dest = vtn_create_ssa_value(b, dest_type);
dest = wrap_matrix(b, dest);
bool transpose_result = false;
if (src0_transpose && src1_transpose) {
/* transpose(A) * transpose(B) = transpose(B * A) */
src1 = src0_transpose;
src0 = src1_transpose;
src0_transpose = NULL;
src1_transpose = NULL;
transpose_result = true;
}
if (src0_transpose && !src1_transpose &&
glsl_get_base_type(src0->type) == GLSL_TYPE_FLOAT) {
/* We already have the rows of src0 and the columns of src1 available,
* so we can just take the dot product of each row with each column to
* get the result.
*/
for (unsigned i = 0; i < src1_columns; i++) {
nir_ssa_def *vec_src[4];
for (unsigned j = 0; j < src0_rows; j++) {
vec_src[j] = nir_fdot(&b->nb, src0_transpose->elems[j]->def,
src1->elems[i]->def);
}
dest->elems[i]->def = nir_vec(&b->nb, vec_src, src0_rows);
}
} else {
/* We don't handle the case where src1 is transposed but not src0, since
* the general case only uses individual components of src1 so the
* optimizer should chew through the transpose we emitted for src1.
*/
for (unsigned i = 0; i < src1_columns; i++) {
/* dest[i] = sum(src0[j] * src1[i][j] for all j) */
dest->elems[i]->def =
nir_fmul(&b->nb, src0->elems[0]->def,
nir_channel(&b->nb, src1->elems[i]->def, 0));
for (unsigned j = 1; j < src0_columns; j++) {
dest->elems[i]->def =
nir_fadd(&b->nb, dest->elems[i]->def,
nir_fmul(&b->nb, src0->elems[j]->def,
nir_channel(&b->nb, src1->elems[i]->def, j)));
}
}
}
dest = unwrap_matrix(dest);
if (transpose_result)
dest = vtn_ssa_transpose(b, dest);
return dest;
}
void
vtn_handle_subgroup(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
{
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
val->ssa = vtn_create_ssa_value(b, val->type->type);
switch (opcode) {
case SpvOpGroupNonUniformElect: {
vtn_fail_if(val->type->type != glsl_bool_type(),
"OpGroupNonUniformElect must return a Bool");
nir_intrinsic_instr *elect =
nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_elect);
nir_ssa_dest_init_for_type(&elect->instr, &elect->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &elect->instr);
val->ssa->def = &elect->dest.ssa;
break;
}
case SpvOpGroupNonUniformBallot: {
vtn_fail_if(val->type->type != glsl_vector_type(GLSL_TYPE_UINT, 4),
"OpGroupNonUniformBallot must return a uvec4");
nir_intrinsic_instr *ballot =
nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_ballot);
ballot->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
nir_ssa_dest_init(&ballot->instr, &ballot->dest, 4, 32, NULL);
ballot->num_components = 4;
nir_builder_instr_insert(&b->nb, &ballot->instr);
val->ssa->def = &ballot->dest.ssa;
break;
}
case SpvOpGroupNonUniformInverseBallot: {
/* This one is just a BallotBitfieldExtract with subgroup invocation.
* We could add a NIR intrinsic but it's easier to just lower it on the
* spot.
*/
nir_intrinsic_instr *intrin =
nir_intrinsic_instr_create(b->nb.shader,
nir_intrinsic_ballot_bitfield_extract);
intrin->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
intrin->src[1] = nir_src_for_ssa(nir_load_subgroup_invocation(&b->nb));
nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &intrin->instr);
val->ssa->def = &intrin->dest.ssa;
break;
}
case SpvOpGroupNonUniformBallotBitExtract:
case SpvOpGroupNonUniformBallotBitCount:
case SpvOpGroupNonUniformBallotFindLSB:
case SpvOpGroupNonUniformBallotFindMSB: {
nir_ssa_def *src0, *src1 = NULL;
nir_intrinsic_op op;
switch (opcode) {
case SpvOpGroupNonUniformBallotBitExtract:
op = nir_intrinsic_ballot_bitfield_extract;
src0 = vtn_ssa_value(b, w[4])->def;
src1 = vtn_ssa_value(b, w[5])->def;
break;
case SpvOpGroupNonUniformBallotBitCount:
switch ((SpvGroupOperation)w[4]) {
case SpvGroupOperationReduce:
op = nir_intrinsic_ballot_bit_count_reduce;
break;
case SpvGroupOperationInclusiveScan:
op = nir_intrinsic_ballot_bit_count_inclusive;
break;
case SpvGroupOperationExclusiveScan:
op = nir_intrinsic_ballot_bit_count_exclusive;
break;
default:
unreachable("Invalid group operation");
}
src0 = vtn_ssa_value(b, w[5])->def;
break;
case SpvOpGroupNonUniformBallotFindLSB:
op = nir_intrinsic_ballot_find_lsb;
src0 = vtn_ssa_value(b, w[4])->def;
break;
case SpvOpGroupNonUniformBallotFindMSB:
op = nir_intrinsic_ballot_find_msb;
src0 = vtn_ssa_value(b, w[4])->def;
break;
default:
unreachable("Unhandled opcode");
}
nir_intrinsic_instr *intrin =
nir_intrinsic_instr_create(b->nb.shader, op);
intrin->src[0] = nir_src_for_ssa(src0);
if (src1)
intrin->src[1] = nir_src_for_ssa(src1);
nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &intrin->instr);
val->ssa->def = &intrin->dest.ssa;
break;
}
case SpvOpGroupNonUniformBroadcastFirst:
vtn_build_subgroup_instr(b, nir_intrinsic_read_first_invocation,
val->ssa, vtn_ssa_value(b, w[4]), NULL, 0, 0);
break;
case SpvOpGroupNonUniformBroadcast:
vtn_build_subgroup_instr(b, nir_intrinsic_read_invocation,
val->ssa, vtn_ssa_value(b, w[4]),
vtn_ssa_value(b, w[5])->def, 0, 0);
break;
case SpvOpGroupNonUniformAll:
case SpvOpGroupNonUniformAny:
case SpvOpGroupNonUniformAllEqual: {
vtn_fail_if(val->type->type != glsl_bool_type(),
"OpGroupNonUniform(All|Any|AllEqual) must return a bool");
nir_intrinsic_op op;
switch (opcode) {
case SpvOpGroupNonUniformAll:
op = nir_intrinsic_vote_all;
break;
case SpvOpGroupNonUniformAny:
op = nir_intrinsic_vote_any;
break;
case SpvOpGroupNonUniformAllEqual: {
switch (glsl_get_base_type(val->type->type)) {
case GLSL_TYPE_FLOAT:
case GLSL_TYPE_DOUBLE:
op = nir_intrinsic_vote_feq;
break;
case GLSL_TYPE_UINT:
case GLSL_TYPE_INT:
case GLSL_TYPE_UINT64:
case GLSL_TYPE_INT64:
case GLSL_TYPE_BOOL:
op = nir_intrinsic_vote_ieq;
break;
default:
unreachable("Unhandled type");
}
break;
}
default:
unreachable("Unhandled opcode");
}
nir_ssa_def *src0 = vtn_ssa_value(b, w[4])->def;
nir_intrinsic_instr *intrin =
nir_intrinsic_instr_create(b->nb.shader, op);
intrin->num_components = src0->num_components;
intrin->src[0] = nir_src_for_ssa(src0);
nir_ssa_dest_init_for_type(&intrin->instr, &intrin->dest,
val->type->type, NULL);
nir_builder_instr_insert(&b->nb, &intrin->instr);
val->ssa->def = &intrin->dest.ssa;
break;
}
case SpvOpGroupNonUniformShuffle:
case SpvOpGroupNonUniformShuffleXor:
case SpvOpGroupNonUniformShuffleUp:
case SpvOpGroupNonUniformShuffleDown: {
nir_intrinsic_op op;
switch (opcode) {
case SpvOpGroupNonUniformShuffle:
op = nir_intrinsic_shuffle;
break;
case SpvOpGroupNonUniformShuffleXor:
op = nir_intrinsic_shuffle_xor;
break;
case SpvOpGroupNonUniformShuffleUp:
op = nir_intrinsic_shuffle_up;
break;
case SpvOpGroupNonUniformShuffleDown:
op = nir_intrinsic_shuffle_down;
break;
default:
unreachable("Invalid opcode");
}
vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
vtn_ssa_value(b, w[5])->def, 0, 0);
break;
}
case SpvOpGroupNonUniformQuadBroadcast:
vtn_build_subgroup_instr(b, nir_intrinsic_quad_broadcast,
val->ssa, vtn_ssa_value(b, w[4]),
vtn_ssa_value(b, w[5])->def, 0, 0);
break;
case SpvOpGroupNonUniformQuadSwap: {
unsigned direction = vtn_constant_uint(b, w[5]);
nir_intrinsic_op op;
switch (direction) {
case 0:
op = nir_intrinsic_quad_swap_horizontal;
break;
case 1:
op = nir_intrinsic_quad_swap_vertical;
break;
case 2:
op = nir_intrinsic_quad_swap_diagonal;
break;
default:
vtn_fail("Invalid constant value in OpGroupNonUniformQuadSwap");
}
vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[4]),
NULL, 0, 0);
break;
}
case SpvOpGroupNonUniformIAdd:
case SpvOpGroupNonUniformFAdd:
case SpvOpGroupNonUniformIMul:
case SpvOpGroupNonUniformFMul:
case SpvOpGroupNonUniformSMin:
case SpvOpGroupNonUniformUMin:
case SpvOpGroupNonUniformFMin:
case SpvOpGroupNonUniformSMax:
case SpvOpGroupNonUniformUMax:
case SpvOpGroupNonUniformFMax:
case SpvOpGroupNonUniformBitwiseAnd:
case SpvOpGroupNonUniformBitwiseOr:
case SpvOpGroupNonUniformBitwiseXor:
case SpvOpGroupNonUniformLogicalAnd:
case SpvOpGroupNonUniformLogicalOr:
case SpvOpGroupNonUniformLogicalXor: {
nir_op reduction_op;
switch (opcode) {
case SpvOpGroupNonUniformIAdd:
reduction_op = nir_op_iadd;
break;
case SpvOpGroupNonUniformFAdd:
reduction_op = nir_op_fadd;
break;
case SpvOpGroupNonUniformIMul:
reduction_op = nir_op_imul;
break;
case SpvOpGroupNonUniformFMul:
reduction_op = nir_op_fmul;
break;
case SpvOpGroupNonUniformSMin:
reduction_op = nir_op_imin;
break;
case SpvOpGroupNonUniformUMin:
reduction_op = nir_op_umin;
break;
case SpvOpGroupNonUniformFMin:
reduction_op = nir_op_fmin;
break;
case SpvOpGroupNonUniformSMax:
reduction_op = nir_op_imax;
break;
case SpvOpGroupNonUniformUMax:
reduction_op = nir_op_umax;
break;
case SpvOpGroupNonUniformFMax:
reduction_op = nir_op_fmax;
break;
case SpvOpGroupNonUniformBitwiseAnd:
case SpvOpGroupNonUniformLogicalAnd:
reduction_op = nir_op_iand;
break;
case SpvOpGroupNonUniformBitwiseOr:
case SpvOpGroupNonUniformLogicalOr:
reduction_op = nir_op_ior;
break;
case SpvOpGroupNonUniformBitwiseXor:
case SpvOpGroupNonUniformLogicalXor:
reduction_op = nir_op_ixor;
break;
default:
unreachable("Invalid reduction operation");
}
nir_intrinsic_op op;
unsigned cluster_size = 0;
switch ((SpvGroupOperation)w[4]) {
case SpvGroupOperationReduce:
op = nir_intrinsic_reduce;
break;
case SpvGroupOperationInclusiveScan:
op = nir_intrinsic_inclusive_scan;
break;
case SpvGroupOperationExclusiveScan:
op = nir_intrinsic_exclusive_scan;
break;
case SpvGroupOperationClusteredReduce:
op = nir_intrinsic_reduce;
assert(count == 7);
cluster_size = vtn_constant_uint(b, w[6]);
break;
default:
unreachable("Invalid group operation");
}
vtn_build_subgroup_instr(b, op, val->ssa, vtn_ssa_value(b, w[5]),
NULL, reduction_op, cluster_size);
break;
}
default:
unreachable("Invalid SPIR-V opcode");
}
}
static void
_vtn_block_load_store(struct vtn_builder *b, nir_intrinsic_op op, bool load,
nir_ssa_def *index, nir_ssa_def *offset,
struct vtn_access_chain *chain, unsigned chain_idx,
struct vtn_type *type, struct vtn_ssa_value **inout)
{
if (chain && chain_idx >= chain->length)
chain = NULL;
if (load && chain == NULL && *inout == NULL)
*inout = vtn_create_ssa_value(b, type->type);
enum glsl_base_type base_type = glsl_get_base_type(type->type);
switch (base_type) {
case GLSL_TYPE_UINT:
case GLSL_TYPE_INT:
case GLSL_TYPE_FLOAT:
case GLSL_TYPE_BOOL:
/* This is where things get interesting. At this point, we've hit
* a vector, a scalar, or a matrix.
*/
if (glsl_type_is_matrix(type->type)) {
if (chain == NULL) {
/* Loading the whole matrix */
struct vtn_ssa_value *transpose;
unsigned num_ops, vec_width;
if (type->row_major) {
num_ops = glsl_get_vector_elements(type->type);
vec_width = glsl_get_matrix_columns(type->type);
if (load) {
const struct glsl_type *transpose_type =
glsl_matrix_type(base_type, vec_width, num_ops);
*inout = vtn_create_ssa_value(b, transpose_type);
} else {
transpose = vtn_ssa_transpose(b, *inout);
inout = &transpose;
}
} else {
num_ops = glsl_get_matrix_columns(type->type);
vec_width = glsl_get_vector_elements(type->type);
}
for (unsigned i = 0; i < num_ops; i++) {
nir_ssa_def *elem_offset =
nir_iadd(&b->nb, offset,
nir_imm_int(&b->nb, i * type->stride));
_vtn_load_store_tail(b, op, load, index, elem_offset,
&(*inout)->elems[i],
glsl_vector_type(base_type, vec_width));
}
if (load && type->row_major)
*inout = vtn_ssa_transpose(b, *inout);
} else if (type->row_major) {
/* Row-major but with an access chiain. */
nir_ssa_def *col_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx],
type->array_element->stride);
offset = nir_iadd(&b->nb, offset, col_offset);
if (chain_idx + 1 < chain->length) {
/* Picking off a single element */
nir_ssa_def *row_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx + 1],
type->stride);
offset = nir_iadd(&b->nb, offset, row_offset);
if (load)
*inout = vtn_create_ssa_value(b, glsl_scalar_type(base_type));
_vtn_load_store_tail(b, op, load, index, offset, inout,
glsl_scalar_type(base_type));
} else {
/* Grabbing a column; picking one element off each row */
unsigned num_comps = glsl_get_vector_elements(type->type);
const struct glsl_type *column_type =
glsl_get_column_type(type->type);
nir_ssa_def *comps[4];
for (unsigned i = 0; i < num_comps; i++) {
nir_ssa_def *elem_offset =
nir_iadd(&b->nb, offset,
nir_imm_int(&b->nb, i * type->stride));
struct vtn_ssa_value *comp, temp_val;
if (!load) {
temp_val.def = nir_channel(&b->nb, (*inout)->def, i);
temp_val.type = glsl_scalar_type(base_type);
}
comp = &temp_val;
_vtn_load_store_tail(b, op, load, index, elem_offset,
&comp, glsl_scalar_type(base_type));
comps[i] = comp->def;
}
if (load) {
if (*inout == NULL)
*inout = vtn_create_ssa_value(b, column_type);
(*inout)->def = nir_vec(&b->nb, comps, num_comps);
}
}
} else {
/* Column-major with a deref. Fall through to array case. */
nir_ssa_def *col_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
offset = nir_iadd(&b->nb, offset, col_offset);
_vtn_block_load_store(b, op, load, index, offset,
chain, chain_idx + 1,
type->array_element, inout);
}
} else if (chain == NULL) {
/* Single whole vector */
assert(glsl_type_is_vector_or_scalar(type->type));
_vtn_load_store_tail(b, op, load, index, offset, inout, type->type);
} else {
/* Single component of a vector. Fall through to array case. */
nir_ssa_def *elem_offset =
vtn_access_link_as_ssa(b, chain->link[chain_idx], type->stride);
offset = nir_iadd(&b->nb, offset, elem_offset);
_vtn_block_load_store(b, op, load, index, offset, NULL, 0,
type->array_element, inout);
}
return;
case GLSL_TYPE_ARRAY: {
unsigned elems = glsl_get_length(type->type);
for (unsigned i = 0; i < elems; i++) {
nir_ssa_def *elem_off =
nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, i * type->stride));
_vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
type->array_element, &(*inout)->elems[i]);
}
return;
}
case GLSL_TYPE_STRUCT: {
unsigned elems = glsl_get_length(type->type);
for (unsigned i = 0; i < elems; i++) {
nir_ssa_def *elem_off =
nir_iadd(&b->nb, offset, nir_imm_int(&b->nb, type->offsets[i]));
_vtn_block_load_store(b, op, load, index, elem_off, NULL, 0,
type->members[i], &(*inout)->elems[i]);
}
return;
}
default:
unreachable("Invalid block member type");
}
}
