#include "spirv_to_nir_private.h"
#include "nir_vla.h"
-nir_ssa_def *
-vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
+static struct vtn_ssa_value *
+vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
+ const struct glsl_type *type)
{
- struct vtn_value *val = vtn_untyped_value(b, value_id);
- switch (val->value_type) {
- case vtn_value_type_constant: {
- assert(glsl_type_is_vector_or_scalar(val->type));
- unsigned num_components = glsl_get_vector_elements(val->type);
- nir_load_const_instr *load =
- nir_load_const_instr_create(b->shader, num_components);
+ struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
+
+ if (entry)
+ return entry->data;
+
+ struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
+ val->type = type;
+
+ switch (glsl_get_base_type(type)) {
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ if (glsl_type_is_vector_or_scalar(type)) {
+ unsigned num_components = glsl_get_vector_elements(val->type);
+ nir_load_const_instr *load =
+ nir_load_const_instr_create(b->shader, num_components);
+
+ for (unsigned i = 0; i < num_components; i++)
+ load->value.u[i] = constant->value.u[i];
+
+ nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
+ val->def = &load->def;
+ } else {
+ assert(glsl_type_is_matrix(type));
+ unsigned rows = glsl_get_vector_elements(val->type);
+ unsigned columns = glsl_get_matrix_columns(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
- for (unsigned i = 0; i < num_components; i++)
- load->value.u[0] = val->constant->value.u[0];
+ for (unsigned i = 0; i < columns; i++) {
+ struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
+ col_val->type = glsl_get_column_type(val->type);
+ nir_load_const_instr *load =
+ nir_load_const_instr_create(b->shader, rows);
- nir_builder_instr_insert(&b->nb, &load->instr);
- return &load->def;
+ for (unsigned j = 0; j < rows; j++)
+ load->value.u[j] = constant->value.u[rows * i + j];
+
+ nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
+ col_val->def = &load->def;
+
+ val->elems[i] = col_val;
+ }
+ }
+ break;
+
+ case GLSL_TYPE_ARRAY: {
+ unsigned elems = glsl_get_length(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ const struct glsl_type *elem_type = glsl_get_array_element(val->type);
+ for (unsigned i = 0; i < elems; i++)
+ val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
+ elem_type);
+ break;
+ }
+
+ case GLSL_TYPE_STRUCT: {
+ unsigned elems = glsl_get_length(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++) {
+ const struct glsl_type *elem_type =
+ glsl_get_struct_field(val->type, i);
+ val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
+ elem_type);
+ }
+ break;
}
+ default:
+ unreachable("bad constant type");
+ }
+
+ return val;
+}
+
+struct vtn_ssa_value *
+vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
+{
+ struct vtn_value *val = vtn_untyped_value(b, value_id);
+ switch (val->value_type) {
+ case vtn_value_type_constant:
+ return vtn_const_ssa_value(b, val->constant, val->type);
+
case vtn_value_type_ssa:
return val->ssa;
default:
case SpvDecorationFPFastMathMode:
case SpvDecorationLinkageAttributes:
case SpvDecorationSpecId:
+ break;
default:
unreachable("Unhandled variable decoration");
}
}
+static struct vtn_ssa_value *
+_vtn_variable_load(struct vtn_builder *b,
+ nir_deref_var *src_deref, nir_deref *src_deref_tail)
+{
+ struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
+ val->type = src_deref_tail->type;
+
+ /* The deref tail may contain a deref to select a component of a vector (in
+ * other words, it might not be an actual tail) so we have to save it away
+ * here since we overwrite it later.
+ */
+ nir_deref *old_child = src_deref_tail->child;
+
+ if (glsl_type_is_vector_or_scalar(val->type)) {
+ nir_intrinsic_instr *load =
+ nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_var);
+ load->variables[0] =
+ nir_deref_as_var(nir_copy_deref(load, &src_deref->deref));
+ load->num_components = glsl_get_vector_elements(val->type);
+ nir_ssa_dest_init(&load->instr, &load->dest, load->num_components, NULL);
+
+ nir_builder_instr_insert(&b->nb, &load->instr);
+
+ if (src_deref->var->data.mode == nir_var_uniform &&
+ glsl_get_base_type(val->type) == GLSL_TYPE_BOOL) {
+ /* Uniform boolean loads need to be fixed up since they're defined
+ * to be zero/nonzero rather than NIR_FALSE/NIR_TRUE.
+ */
+ val->def = nir_ine(&b->nb, &load->dest.ssa, nir_imm_int(&b->nb, 0));
+ } else {
+ val->def = &load->dest.ssa;
+ }
+ } else if (glsl_get_base_type(val->type) == GLSL_TYPE_ARRAY ||
+ glsl_type_is_matrix(val->type)) {
+ unsigned elems = glsl_get_length(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+
+ nir_deref_array *deref = nir_deref_array_create(b);
+ deref->deref_array_type = nir_deref_array_type_direct;
+ deref->deref.type = glsl_get_array_element(val->type);
+ src_deref_tail->child = &deref->deref;
+ for (unsigned i = 0; i < elems; i++) {
+ deref->base_offset = i;
+ val->elems[i] = _vtn_variable_load(b, src_deref, &deref->deref);
+ }
+ } else {
+ assert(glsl_get_base_type(val->type) == GLSL_TYPE_STRUCT);
+ unsigned elems = glsl_get_length(val->type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+
+ nir_deref_struct *deref = nir_deref_struct_create(b, 0);
+ src_deref_tail->child = &deref->deref;
+ for (unsigned i = 0; i < elems; i++) {
+ deref->index = i;
+ deref->deref.type = glsl_get_struct_field(val->type, i);
+ val->elems[i] = _vtn_variable_load(b, src_deref, &deref->deref);
+ }
+ }
+
+ src_deref_tail->child = old_child;
+
+ return val;
+}
+
+static void
+_vtn_variable_store(struct vtn_builder *b, nir_deref_var *dest_deref,
+ nir_deref *dest_deref_tail, struct vtn_ssa_value *src)
+{
+ nir_deref *old_child = dest_deref_tail->child;
+
+ if (glsl_type_is_vector_or_scalar(src->type)) {
+ nir_intrinsic_instr *store =
+ nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_var);
+ store->variables[0] =
+ nir_deref_as_var(nir_copy_deref(store, &dest_deref->deref));
+ store->src[0] = nir_src_for_ssa(src->def);
+
+ nir_builder_instr_insert(&b->nb, &store->instr);
+ } else if (glsl_get_base_type(src->type) == GLSL_TYPE_ARRAY ||
+ glsl_type_is_matrix(src->type)) {
+ unsigned elems = glsl_get_length(src->type);
+
+ nir_deref_array *deref = nir_deref_array_create(b);
+ deref->deref_array_type = nir_deref_array_type_direct;
+ deref->deref.type = glsl_get_array_element(src->type);
+ dest_deref_tail->child = &deref->deref;
+ for (unsigned i = 0; i < elems; i++) {
+ deref->base_offset = i;
+ _vtn_variable_store(b, dest_deref, &deref->deref, src->elems[i]);
+ }
+ } else {
+ assert(glsl_get_base_type(src->type) == GLSL_TYPE_STRUCT);
+ unsigned elems = glsl_get_length(src->type);
+
+ nir_deref_struct *deref = nir_deref_struct_create(b, 0);
+ dest_deref_tail->child = &deref->deref;
+ for (unsigned i = 0; i < elems; i++) {
+ deref->index = i;
+ deref->deref.type = glsl_get_struct_field(src->type, i);
+ _vtn_variable_store(b, dest_deref, &deref->deref, src->elems[i]);
+ }
+ }
+
+ dest_deref_tail->child = old_child;
+}
+
+/*
+ * Gets the NIR-level deref tail, which may have as a child an array deref
+ * selecting which component due to OpAccessChain supporting per-component
+ * indexing in SPIR-V.
+ */
+
+static nir_deref *
+get_deref_tail(nir_deref_var *deref)
+{
+ nir_deref *cur = &deref->deref;
+ while (!glsl_type_is_vector_or_scalar(cur->type) && cur->child)
+ cur = cur->child;
+
+ return cur;
+}
+
+static nir_ssa_def *vtn_vector_extract(struct vtn_builder *b,
+ nir_ssa_def *src, unsigned index);
+
+static nir_ssa_def *vtn_vector_extract_dynamic(struct vtn_builder *b,
+ nir_ssa_def *src,
+ nir_ssa_def *index);
+
+static struct vtn_ssa_value *
+vtn_variable_load(struct vtn_builder *b, nir_deref_var *src)
+{
+ nir_deref *src_tail = get_deref_tail(src);
+ struct vtn_ssa_value *val = _vtn_variable_load(b, src, src_tail);
+
+ if (src_tail->child) {
+ nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
+ assert(vec_deref->deref.child == NULL);
+ val->type = vec_deref->deref.type;
+ if (vec_deref->deref_array_type == nir_deref_array_type_direct)
+ val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
+ else
+ val->def = vtn_vector_extract_dynamic(b, val->def,
+ vec_deref->indirect.ssa);
+ }
+
+ return val;
+}
+
+static nir_ssa_def * vtn_vector_insert(struct vtn_builder *b,
+ nir_ssa_def *src, nir_ssa_def *insert,
+ unsigned index);
+
+static nir_ssa_def * vtn_vector_insert_dynamic(struct vtn_builder *b,
+ nir_ssa_def *src,
+ nir_ssa_def *insert,
+ nir_ssa_def *index);
+static void
+vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
+ nir_deref_var *dest)
+{
+ nir_deref *dest_tail = get_deref_tail(dest);
+ if (dest_tail->child) {
+ struct vtn_ssa_value *val = _vtn_variable_load(b, dest, dest_tail);
+ nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
+ assert(deref->deref.child == NULL);
+ if (deref->deref_array_type == nir_deref_array_type_direct)
+ val->def = vtn_vector_insert(b, val->def, src->def,
+ deref->base_offset);
+ else
+ val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
+ deref->indirect.ssa);
+ _vtn_variable_store(b, dest, dest_tail, val);
+ } else {
+ _vtn_variable_store(b, dest, dest_tail, src);
+ }
+}
+
+static void
+vtn_variable_copy(struct vtn_builder *b, nir_deref_var *src,
+ nir_deref_var *dest)
+{
+ nir_deref *src_tail = get_deref_tail(src);
+
+ if (src_tail->child) {
+ assert(get_deref_tail(dest)->child);
+ struct vtn_ssa_value *val = vtn_variable_load(b, src);
+ vtn_variable_store(b, val, dest);
+ } else {
+ nir_intrinsic_instr *copy =
+ nir_intrinsic_instr_create(b->shader, nir_intrinsic_copy_var);
+ copy->variables[0] = nir_deref_as_var(nir_copy_deref(copy, &dest->deref));
+ copy->variables[1] = nir_deref_as_var(nir_copy_deref(copy, &src->deref));
+
+ nir_builder_instr_insert(&b->nb, ©->instr);
+ }
+}
+
static void
vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
assert(idx_val->value_type == vtn_value_type_ssa);
deref_arr->deref_array_type = nir_deref_array_type_indirect;
deref_arr->base_offset = 0;
- deref_arr->indirect = nir_src_for_ssa(vtn_ssa_value(b, w[1]));
+ deref_arr->indirect =
+ nir_src_for_ssa(vtn_ssa_value(b, w[1])->def);
}
tail->child = &deref_arr->deref;
break;
nir_deref_var *dest = vtn_value(b, w[1], vtn_value_type_deref)->deref;
nir_deref_var *src = vtn_value(b, w[2], vtn_value_type_deref)->deref;
- nir_intrinsic_instr *copy =
- nir_intrinsic_instr_create(b->shader, nir_intrinsic_copy_var);
- copy->variables[0] = nir_deref_as_var(nir_copy_deref(copy, &dest->deref));
- copy->variables[1] = nir_deref_as_var(nir_copy_deref(copy, &src->deref));
-
- nir_builder_instr_insert(&b->nb, ©->instr);
+ vtn_variable_copy(b, src, dest);
break;
}
return;
}
- assert(glsl_type_is_vector_or_scalar(src_type));
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
-
- nir_intrinsic_instr *load =
- nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_var);
- load->variables[0] = nir_deref_as_var(nir_copy_deref(load, &src->deref));
- load->num_components = glsl_get_vector_elements(src_type);
- nir_ssa_dest_init(&load->instr, &load->dest, load->num_components,
- val->name);
-
- nir_builder_instr_insert(&b->nb, &load->instr);
- val->type = src_type;
-
- if (src->var->data.mode == nir_var_uniform &&
- glsl_get_base_type(src_type) == GLSL_TYPE_BOOL) {
- /* Uniform boolean loads need to be fixed up since they're defined
- * to be zero/nonzero rather than NIR_FALSE/NIR_TRUE.
- */
- val->ssa = nir_ine(&b->nb, &load->dest.ssa, nir_imm_int(&b->nb, 0));
- } else {
- val->ssa = &load->dest.ssa;
- }
+ val->ssa = vtn_variable_load(b, src);
break;
}
case SpvOpStore: {
nir_deref_var *dest = vtn_value(b, w[1], vtn_value_type_deref)->deref;
- const struct glsl_type *dest_type = nir_deref_tail(&dest->deref)->type;
- struct vtn_value *src_val = vtn_untyped_value(b, w[2]);
- if (src_val->value_type == vtn_value_type_ssa) {
- assert(glsl_type_is_vector_or_scalar(dest_type));
- nir_intrinsic_instr *store =
- nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_var);
- store->src[0] = nir_src_for_ssa(src_val->ssa);
- store->variables[0] = nir_deref_as_var(nir_copy_deref(store, &dest->deref));
- store->num_components = glsl_get_vector_elements(dest_type);
-
- nir_builder_instr_insert(&b->nb, &store->instr);
- } else {
- assert(src_val->value_type == vtn_value_type_constant);
-
- nir_variable *const_tmp = rzalloc(b->shader, nir_variable);
- const_tmp->type = dest_type;
- const_tmp->name = "const_temp";
- const_tmp->data.mode = nir_var_local;
- const_tmp->data.read_only = true;
- exec_list_push_tail(&b->impl->locals, &const_tmp->node);
-
- nir_intrinsic_instr *copy =
- nir_intrinsic_instr_create(b->shader, nir_intrinsic_copy_var);
- copy->variables[0] = nir_deref_as_var(nir_copy_deref(copy, &dest->deref));
- copy->variables[1] = nir_deref_var_create(copy, const_tmp);
-
- nir_builder_instr_insert(&b->nb, ©->instr);
- }
+ struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
+ vtn_variable_store(b, src, dest);
break;
}
vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
{
nir_tex_src src;
- src.src = nir_src_for_ssa(vtn_value(b, index, vtn_value_type_ssa)->ssa);
+ src.src = nir_src_for_ssa(vtn_value(b, index, vtn_value_type_ssa)->ssa->def);
src.src_type = type;
return src;
}
case SpvOpTextureGather:
case SpvOpTextureGatherOffset:
case SpvOpTextureGatherOffsets:
- case SpvOpTextureQueryLod:
+ case SpvOpTextureQueryLod: {
/* All these types have the coordinate as their first real argument */
- coord_components = glsl_get_vector_elements(b->values[w[4]].type);
- p->src = nir_src_for_ssa(vtn_ssa_value(b, w[4]));
+ struct vtn_ssa_value *coord = vtn_ssa_value(b, w[4]);
+ coord_components = glsl_get_vector_elements(coord->type);
+ p->src = nir_src_for_ssa(coord->def);
p->src_type = nir_tex_src_coord;
p++;
break;
+ }
+
default:
break;
}
instr->sampler = nir_deref_as_var(nir_copy_deref(instr, &sampler->deref));
nir_ssa_dest_init(&instr->instr, &instr->dest, 4, NULL);
- val->ssa = &instr->dest.ssa;
+ val->ssa->def = &instr->dest.ssa;
+ val->ssa->type = val->type;
nir_builder_instr_insert(&b->nb, &instr->instr);
}
+static struct vtn_ssa_value *
+vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
+{
+ struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
+ val->type = type;
+
+ if (!glsl_type_is_vector_or_scalar(type)) {
+ unsigned elems = glsl_get_length(type);
+ val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++) {
+ const struct glsl_type *child_type;
+
+ switch (glsl_get_base_type(type)) {
+ case GLSL_TYPE_INT:
+ case GLSL_TYPE_UINT:
+ case GLSL_TYPE_BOOL:
+ case GLSL_TYPE_FLOAT:
+ case GLSL_TYPE_DOUBLE:
+ child_type = glsl_get_column_type(type);
+ break;
+ case GLSL_TYPE_ARRAY:
+ child_type = glsl_get_array_element(type);
+ break;
+ case GLSL_TYPE_STRUCT:
+ child_type = glsl_get_struct_field(type, i);
+ break;
+ default:
+ unreachable("unkown base type");
+ }
+
+ val->elems[i] = vtn_create_ssa_value(b, child_type);
+ }
+ }
+
+ return val;
+}
+
+static nir_alu_instr *
+create_vec(void *mem_ctx, unsigned num_components)
+{
+ nir_op op;
+ switch (num_components) {
+ case 1: op = nir_op_fmov; break;
+ case 2: op = nir_op_vec2; break;
+ case 3: op = nir_op_vec3; break;
+ case 4: op = nir_op_vec4; break;
+ default: unreachable("bad vector size");
+ }
+
+ nir_alu_instr *vec = nir_alu_instr_create(mem_ctx, op);
+ nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components, NULL);
+
+ return vec;
+}
+
+static struct vtn_ssa_value *
+vtn_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
+{
+ if (src->transposed)
+ return src->transposed;
+
+ struct vtn_ssa_value *dest =
+ vtn_create_ssa_value(b, glsl_transposed_type(src->type));
+
+ for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
+ nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type));
+ if (glsl_type_is_vector_or_scalar(src->type)) {
+ vec->src[0].src = nir_src_for_ssa(src->def);
+ vec->src[0].swizzle[0] = i;
+ } else {
+ for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
+ vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
+ vec->src[j].swizzle[0] = i;
+ }
+ }
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+ dest->elems[i]->def = &vec->dest.dest.ssa;
+ }
+
+ dest->transposed = src;
+
+ return dest;
+}
+
+/*
+ * Normally, column vectors in SPIR-V correspond to a single NIR SSA
+ * definition. But for matrix multiplies, we want to do one routine for
+ * multiplying a matrix by a matrix and then pretend that vectors are matrices
+ * with one column. So we "wrap" these things, and unwrap the result before we
+ * send it off.
+ */
+
+static struct vtn_ssa_value *
+vtn_wrap_matrix(struct vtn_builder *b, struct vtn_ssa_value *val)
+{
+ if (val == NULL)
+ return NULL;
+
+ if (glsl_type_is_matrix(val->type))
+ return val;
+
+ struct vtn_ssa_value *dest = rzalloc(b, struct vtn_ssa_value);
+ dest->type = val->type;
+ dest->elems = ralloc_array(b, struct vtn_ssa_value *, 1);
+ dest->elems[0] = val;
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+vtn_unwrap_matrix(struct vtn_ssa_value *val)
+{
+ if (glsl_type_is_matrix(val->type))
+ return val;
+
+ return val->elems[0];
+}
+
+static struct vtn_ssa_value *
+vtn_matrix_multiply(struct vtn_builder *b,
+ struct vtn_ssa_value *_src0, struct vtn_ssa_value *_src1)
+{
+
+ struct vtn_ssa_value *src0 = vtn_wrap_matrix(b, _src0);
+ struct vtn_ssa_value *src1 = vtn_wrap_matrix(b, _src1);
+ struct vtn_ssa_value *src0_transpose = vtn_wrap_matrix(b, _src0->transposed);
+ struct vtn_ssa_value *src1_transpose = vtn_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);
+
+ struct vtn_ssa_value *dest =
+ vtn_create_ssa_value(b, glsl_matrix_type(glsl_get_base_type(src0->type),
+ src0_rows, src1_columns));
+
+ dest = vtn_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_alu_instr *vec = create_vec(b, src0_rows);
+ for (unsigned j = 0; j < src0_rows; j++) {
+ vec->src[j].src =
+ nir_src_for_ssa(nir_fdot(&b->nb, src0_transpose->elems[j]->def,
+ src1->elems[i]->def));
+ }
+
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+ dest->elems[i]->def = &vec->dest.dest.ssa;
+ }
+ } 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,
+ vtn_vector_extract(b, 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,
+ vtn_vector_extract(b,
+ src1->elems[i]->def, j)));
+ }
+ }
+ }
+
+ dest = vtn_unwrap_matrix(dest);
+
+ if (transpose_result)
+ dest = vtn_transpose(b, dest);
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+vtn_mat_times_scalar(struct vtn_builder *b,
+ struct vtn_ssa_value *mat,
+ nir_ssa_def *scalar)
+{
+ struct vtn_ssa_value *dest = vtn_create_ssa_value(b, mat->type);
+ for (unsigned i = 0; i < glsl_get_matrix_columns(mat->type); i++) {
+ if (glsl_get_base_type(mat->type) == GLSL_TYPE_FLOAT)
+ dest->elems[i]->def = nir_fmul(&b->nb, mat->elems[i]->def, scalar);
+ else
+ dest->elems[i]->def = nir_imul(&b->nb, mat->elems[i]->def, scalar);
+ }
+
+ return dest;
+}
+
static void
vtn_handle_matrix_alu(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
{
- unreachable("Matrix math not handled");
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
+
+ switch (opcode) {
+ case SpvOpTranspose: {
+ struct vtn_ssa_value *src = vtn_ssa_value(b, w[3]);
+ val->ssa = vtn_transpose(b, src);
+ break;
+ }
+
+ case SpvOpOuterProduct: {
+ struct vtn_ssa_value *src0 = vtn_ssa_value(b, w[3]);
+ struct vtn_ssa_value *src1 = vtn_ssa_value(b, w[4]);
+
+ val->ssa = vtn_matrix_multiply(b, src0, vtn_transpose(b, src1));
+ break;
+ }
+
+ case SpvOpMatrixTimesScalar: {
+ struct vtn_ssa_value *mat = vtn_ssa_value(b, w[3]);
+ struct vtn_ssa_value *scalar = vtn_ssa_value(b, w[4]);
+
+ if (mat->transposed) {
+ val->ssa = vtn_transpose(b, vtn_mat_times_scalar(b, mat->transposed,
+ scalar->def));
+ } else {
+ val->ssa = vtn_mat_times_scalar(b, mat, scalar->def);
+ }
+ break;
+ }
+
+ case SpvOpVectorTimesMatrix:
+ case SpvOpMatrixTimesVector:
+ case SpvOpMatrixTimesMatrix: {
+ struct vtn_ssa_value *src0 = vtn_ssa_value(b, w[3]);
+ struct vtn_ssa_value *src1 = vtn_ssa_value(b, w[4]);
+
+ val->ssa = vtn_matrix_multiply(b, src0, src1);
+ break;
+ }
+
+ default: unreachable("unknown matrix opcode");
+ }
}
static void
{
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
+ val->ssa = vtn_create_ssa_value(b, val->type);
/* Collect the various SSA sources */
unsigned num_inputs = count - 3;
nir_ssa_def *src[4];
for (unsigned i = 0; i < num_inputs; i++)
- src[i] = vtn_ssa_value(b, w[i + 3]);
+ src[i] = vtn_ssa_value(b, w[i + 3])->def;
/* Indicates that the first two arguments should be swapped. This is
* used for implementing greater-than and less-than-or-equal.
case SpvOpDPdxCoarse: op = nir_op_fddx_coarse; break;
case SpvOpDPdyCoarse: op = nir_op_fddy_coarse; break;
case SpvOpFwidth:
- val->ssa = nir_fadd(&b->nb,
- nir_fabs(&b->nb, nir_fddx(&b->nb, src[0])),
- nir_fabs(&b->nb, nir_fddx(&b->nb, src[1])));
+ val->ssa->def = nir_fadd(&b->nb,
+ nir_fabs(&b->nb, nir_fddx(&b->nb, src[0])),
+ nir_fabs(&b->nb, nir_fddx(&b->nb, src[1])));
return;
case SpvOpFwidthFine:
- val->ssa = nir_fadd(&b->nb,
- nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[0])),
- nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[1])));
+ val->ssa->def = nir_fadd(&b->nb,
+ nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[0])),
+ nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[1])));
return;
case SpvOpFwidthCoarse:
- val->ssa = nir_fadd(&b->nb,
- nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[0])),
- nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[1])));
+ val->ssa->def = nir_fadd(&b->nb,
+ nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[0])),
+ nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[1])));
return;
case SpvOpVectorTimesScalar:
/* The builder will take care of splatting for us. */
- val->ssa = nir_fmul(&b->nb, src[0], src[1]);
+ val->ssa->def = nir_fmul(&b->nb, src[0], src[1]);
return;
case SpvOpSRem:
nir_alu_instr *instr = nir_alu_instr_create(b->shader, op);
nir_ssa_dest_init(&instr->instr, &instr->dest.dest,
glsl_get_vector_elements(val->type), val->name);
- val->ssa = &instr->dest.dest.ssa;
+ val->ssa->def = &instr->dest.dest.ssa;
for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++)
instr->src[i].src = nir_src_for_ssa(src[i]);
nir_builder_instr_insert(&b->nb, &instr->instr);
}
+static nir_ssa_def *
+vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
+{
+ unsigned swiz[4] = { index };
+ return nir_swizzle(&b->nb, src, swiz, 1, true);
+}
+
+
+static nir_ssa_def *
+vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
+ unsigned index)
+{
+ nir_alu_instr *vec = create_vec(b->shader, src->num_components);
+
+ for (unsigned i = 0; i < src->num_components; i++) {
+ if (i == index) {
+ vec->src[i].src = nir_src_for_ssa(insert);
+ } else {
+ vec->src[i].src = nir_src_for_ssa(src);
+ vec->src[i].swizzle[0] = i;
+ }
+ }
+
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+
+ return &vec->dest.dest.ssa;
+}
+
+static nir_ssa_def *
+vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
+ nir_ssa_def *index)
+{
+ nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
+ for (unsigned i = 1; i < src->num_components; i++)
+ dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
+ vtn_vector_extract(b, src, i), dest);
+
+ return dest;
+}
+
+static nir_ssa_def *
+vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
+ nir_ssa_def *insert, nir_ssa_def *index)
+{
+ nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
+ for (unsigned i = 1; i < src->num_components; i++)
+ dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
+ vtn_vector_insert(b, src, insert, i), dest);
+
+ return dest;
+}
+
+static nir_ssa_def *
+vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
+ nir_ssa_def *src0, nir_ssa_def *src1,
+ const uint32_t *indices)
+{
+ nir_alu_instr *vec = create_vec(b->shader, num_components);
+
+ nir_ssa_undef_instr *undef = nir_ssa_undef_instr_create(b->shader, 1);
+ nir_builder_instr_insert(&b->nb, &undef->instr);
+
+ for (unsigned i = 0; i < num_components; i++) {
+ uint32_t index = indices[i];
+ if (index == 0xffffffff) {
+ vec->src[i].src = nir_src_for_ssa(&undef->def);
+ } else if (index < src0->num_components) {
+ vec->src[i].src = nir_src_for_ssa(src0);
+ vec->src[i].swizzle[0] = index;
+ } else {
+ vec->src[i].src = nir_src_for_ssa(src1);
+ vec->src[i].swizzle[0] = index - src0->num_components;
+ }
+ }
+
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+
+ return &vec->dest.dest.ssa;
+}
+
+/*
+ * Concatentates a number of vectors/scalars together to produce a vector
+ */
+static nir_ssa_def *
+vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
+ unsigned num_srcs, nir_ssa_def **srcs)
+{
+ nir_alu_instr *vec = create_vec(b->shader, num_components);
+
+ unsigned dest_idx = 0;
+ for (unsigned i = 0; i < num_srcs; i++) {
+ nir_ssa_def *src = srcs[i];
+ for (unsigned j = 0; j < src->num_components; j++) {
+ vec->src[dest_idx].src = nir_src_for_ssa(src);
+ vec->src[dest_idx].swizzle[0] = j;
+ dest_idx++;
+ }
+ }
+
+ nir_builder_instr_insert(&b->nb, &vec->instr);
+
+ return &vec->dest.dest.ssa;
+}
+
+static struct vtn_ssa_value *
+vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
+{
+ struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
+ dest->type = src->type;
+
+ if (glsl_type_is_vector_or_scalar(src->type)) {
+ dest->def = src->def;
+ } else {
+ unsigned elems = glsl_get_length(src->type);
+
+ dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++)
+ dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
+ }
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
+ struct vtn_ssa_value *insert, const uint32_t *indices,
+ unsigned num_indices)
+{
+ struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
+
+ struct vtn_ssa_value *cur = dest;
+ unsigned i;
+ for (i = 0; i < num_indices - 1; i++) {
+ cur = cur->elems[indices[i]];
+ }
+
+ if (glsl_type_is_vector_or_scalar(cur->type)) {
+ /* According to the SPIR-V spec, OpCompositeInsert may work down to
+ * the component granularity. In that case, the last index will be
+ * the index to insert the scalar into the vector.
+ */
+
+ cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
+ } else {
+ cur->elems[indices[i]] = insert;
+ }
+
+ return dest;
+}
+
+static struct vtn_ssa_value *
+vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
+ const uint32_t *indices, unsigned num_indices)
+{
+ struct vtn_ssa_value *cur = src;
+ for (unsigned i = 0; i < num_indices; i++) {
+ if (glsl_type_is_vector_or_scalar(cur->type)) {
+ assert(i == num_indices - 1);
+ /* According to the SPIR-V spec, OpCompositeExtract may work down to
+ * the component granularity. The last index will be the index of the
+ * vector to extract.
+ */
+
+ struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
+ ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
+ ret->def = vtn_vector_extract(b, cur->def, indices[i]);
+ return ret;
+ }
+ }
+
+ return cur;
+}
+
+static void
+vtn_handle_composite(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->type = vtn_value(b, w[1], vtn_value_type_type)->type;
+
+ switch (opcode) {
+ case SpvOpVectorExtractDynamic:
+ val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def);
+ break;
+
+ case SpvOpVectorInsertDynamic:
+ val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def,
+ vtn_ssa_value(b, w[5])->def);
+ break;
+
+ case SpvOpVectorShuffle:
+ val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(val->type),
+ vtn_ssa_value(b, w[3])->def,
+ vtn_ssa_value(b, w[4])->def,
+ w + 5);
+ break;
+
+ case SpvOpCompositeConstruct: {
+ val->ssa = rzalloc(b, struct vtn_ssa_value);
+ unsigned elems = count - 3;
+ if (glsl_type_is_vector_or_scalar(val->type)) {
+ nir_ssa_def *srcs[4];
+ for (unsigned i = 0; i < elems; i++)
+ srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
+ val->ssa->def =
+ vtn_vector_construct(b, glsl_get_vector_elements(val->type),
+ elems, srcs);
+ } else {
+ val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
+ for (unsigned i = 0; i < elems; i++)
+ val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
+ }
+ break;
+ }
+ case SpvOpCompositeExtract:
+ val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
+ w + 4, count - 4);
+ break;
+
+ case SpvOpCompositeInsert:
+ val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
+ vtn_ssa_value(b, w[3]),
+ w + 5, count - 5);
+ break;
+
+ case SpvOpCopyObject:
+ val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
+ break;
+
+ default:
+ unreachable("unknown composite operation");
+ }
+
+ val->ssa->type = val->type;
+}
+
+static void
+vtn_phi_node_init(struct vtn_builder *b, struct vtn_ssa_value *val)
+{
+ if (glsl_type_is_vector_or_scalar(val->type)) {
+ nir_phi_instr *phi = nir_phi_instr_create(b->shader);
+ nir_ssa_dest_init(&phi->instr, &phi->dest,
+ glsl_get_vector_elements(val->type), NULL);
+ exec_list_make_empty(&phi->srcs);
+ nir_builder_instr_insert(&b->nb, &phi->instr);
+ val->def = &phi->dest.ssa;
+ } else {
+ unsigned elems = glsl_get_length(val->type);
+ for (unsigned i = 0; i < elems; i++)
+ vtn_phi_node_init(b, val->elems[i]);
+ }
+}
+
+static struct vtn_ssa_value *
+vtn_phi_node_create(struct vtn_builder *b, const struct glsl_type *type)
+{
+ struct vtn_ssa_value *val = vtn_create_ssa_value(b, type);
+ vtn_phi_node_init(b, val);
+ return val;
+}
+
+static void
+vtn_handle_phi_first_pass(struct vtn_builder *b, const uint32_t *w)
+{
+ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
+ val->type = vtn_value(b, w[1], vtn_value_type_type)->type;
+ val->ssa = vtn_phi_node_create(b, val->type);
+}
+
+static void
+vtn_phi_node_add_src(struct vtn_ssa_value *phi, const nir_block *pred,
+ struct vtn_ssa_value *val)
+{
+ assert(phi->type == val->type);
+ if (glsl_type_is_vector_or_scalar(phi->type)) {
+ nir_phi_instr *phi_instr = nir_instr_as_phi(phi->def->parent_instr);
+ nir_phi_src *src = ralloc(phi_instr, nir_phi_src);
+ src->pred = (nir_block *) pred;
+ src->src = nir_src_for_ssa(val->def);
+ exec_list_push_tail(&phi_instr->srcs, &src->node);
+ } else {
+ unsigned elems = glsl_get_length(phi->type);
+ for (unsigned i = 0; i < elems; i++)
+ vtn_phi_node_add_src(phi->elems[i], pred, val->elems[i]);
+ }
+}
+
+static struct vtn_ssa_value *
+vtn_get_phi_node_src(struct vtn_builder *b, nir_block *block,
+ const struct glsl_type *type, const uint32_t *w,
+ unsigned count)
+{
+ struct hash_entry *entry = _mesa_hash_table_search(b->block_table, block);
+ if (entry) {
+ struct vtn_block *spv_block = entry->data;
+ for (unsigned off = 4; off < count; off += 2) {
+ if (spv_block == vtn_value(b, w[off], vtn_value_type_block)->block) {
+ return vtn_ssa_value(b, w[off - 1]);
+ }
+ }
+ }
+
+ nir_builder_insert_before_block(&b->nb, block);
+ struct vtn_ssa_value *phi = vtn_phi_node_create(b, type);
+
+ struct set_entry *entry2;
+ set_foreach(block->predecessors, entry2) {
+ nir_block *pred = (nir_block *) entry2->key;
+ struct vtn_ssa_value *val = vtn_get_phi_node_src(b, pred, type, w,
+ count);
+ vtn_phi_node_add_src(phi, pred, val);
+ }
+
+ return phi;
+}
+
+static bool
+vtn_handle_phi_second_pass(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
+{
+ if (opcode == SpvOpLabel) {
+ b->block = vtn_value(b, w[1], vtn_value_type_block)->block;
+ return true;
+ }
+
+ if (opcode != SpvOpPhi)
+ return true;
+
+ struct vtn_ssa_value *phi = vtn_value(b, w[2], vtn_value_type_ssa)->ssa;
+
+ struct set_entry *entry;
+ set_foreach(b->block->block->predecessors, entry) {
+ nir_block *pred = (nir_block *) entry->key;
+
+ struct vtn_ssa_value *val = vtn_get_phi_node_src(b, pred, phi->type, w,
+ count);
+ vtn_phi_node_add_src(phi, pred, val);
+ }
+
+ return true;
+}
+
static bool
vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
}
case SpvOpFunctionEnd:
+ b->func->end = w;
b->func = NULL;
break;
vtn_handle_matrix_alu(b, opcode, w, count);
break;
+ case SpvOpVectorExtractDynamic:
+ case SpvOpVectorInsertDynamic:
+ case SpvOpVectorShuffle:
+ case SpvOpCompositeConstruct:
+ case SpvOpCompositeExtract:
+ case SpvOpCompositeInsert:
+ case SpvOpCopyObject:
+ vtn_handle_composite(b, opcode, w, count);
+ break;
+
+ case SpvOpPhi:
+ vtn_handle_phi_first_pass(b, w);
+ break;
+
default:
unreachable("Unhandled opcode");
}
{
struct vtn_block *block = start;
while (block != end_block) {
- if (block->block != NULL) {
- /* We've already visited this block once before so this is a
- * back-edge. Back-edges are only allowed to point to a loop
- * merge.
- */
- assert(block == cont_block);
- return;
- }
-
if (block->merge_op == SpvOpLoopMerge) {
/* This is the jump into a loop. */
- cont_block = block;
- break_block = vtn_value(b, block->merge_block_id,
- vtn_value_type_block)->block;
+ struct vtn_block *new_cont_block = block;
+ struct vtn_block *new_break_block =
+ vtn_value(b, block->merge_block_id, vtn_value_type_block)->block;
nir_loop *loop = nir_loop_create(b->shader);
nir_cf_node_insert_end(b->nb.cf_node_list, &loop->cf_node);
block->merge_op = SpvOpNop;
nir_builder_insert_after_cf_list(&b->nb, &loop->body);
- vtn_walk_blocks(b, block, break_block, cont_block, NULL);
+ vtn_walk_blocks(b, block, new_break_block, new_cont_block, NULL);
nir_builder_insert_after_cf_list(&b->nb, old_list);
- block = break_block;
+ block = new_break_block;
continue;
}
vtn_foreach_instruction(b, block->label, block->branch,
vtn_handle_body_instruction);
+ nir_cf_node *cur_cf_node =
+ exec_node_data(nir_cf_node, exec_list_get_tail(b->nb.cf_node_list),
+ node);
+ nir_block *cur_block = nir_cf_node_as_block(cur_cf_node);
+ _mesa_hash_table_insert(b->block_table, cur_block, block);
+
switch (branch_op) {
case SpvOpBranch: {
struct vtn_block *branch_block =
return;
} else if (branch_block == end_block) {
+ /* We're branching to the merge block of an if, since for loops
+ * and functions end_block == NULL, so we're done here.
+ */
return;
} else {
+ /* We're branching to another block, and according to the rules,
+ * we can only branch to another block with one predecessor (so
+ * we're the only one jumping to it) so we can just process it
+ * next.
+ */
block = branch_block;
continue;
}
vtn_value(b, w[3], vtn_value_type_block)->block;
nir_if *if_stmt = nir_if_create(b->shader);
- if_stmt->condition = nir_src_for_ssa(vtn_ssa_value(b, w[1]));
+ if_stmt->condition = nir_src_for_ssa(vtn_ssa_value(b, w[1])->def);
nir_cf_node_insert_end(b->nb.cf_node_list, &if_stmt->cf_node);
if (then_block == break_block) {
&jump->instr);
block = then_block;
} else {
- /* Conventional if statement */
+ /* According to the rules we're branching to two blocks that don't
+ * have any other predecessors, so we can handle this as a
+ * conventional if.
+ */
assert(block->merge_op == SpvOpSelectionMerge);
struct vtn_block *merge_block =
vtn_value(b, block->merge_block_id, vtn_value_type_block)->block;
foreach_list_typed(struct vtn_function, func, node, &b->functions) {
b->impl = nir_function_impl_create(func->overload);
+ b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+ b->block_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
nir_builder_init(&b->nb, b->impl);
nir_builder_insert_after_cf_list(&b->nb, &b->impl->body);
vtn_walk_blocks(b, func->start_block, NULL, NULL, NULL);
+ vtn_foreach_instruction(b, func->start_block->label, func->end,
+ vtn_handle_phi_second_pass);
}
ralloc_free(b);
projects / platform / upstream / mesa.git / commitdiff