#include "nir/nir_deref.h"
#include "spirv_info.h"
+#include "util/u_math.h"
+
#include <stdio.h>
void
} else if ((strcmp(ext, "SPV_AMD_shader_trinary_minmax") == 0)
&& (b->options && b->options->caps.trinary_minmax)) {
val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
+ } else if (strcmp(ext, "OpenCL.std") == 0) {
+ val->ext_handler = vtn_handle_opencl_instruction;
} else {
vtn_fail("Unsupported extension: %s", ext);
}
case SpvOpMemberDecorate:
case SpvOpDecorateStringGOOGLE:
case SpvOpMemberDecorateStringGOOGLE:
- case SpvOpExecutionMode: {
+ case SpvOpExecutionMode:
+ case SpvOpExecutionModeId: {
struct vtn_value *val = vtn_untyped_value(b, target);
struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
"Member argument of OpMemberDecorate too large");
break;
case SpvOpExecutionMode:
+ case SpvOpExecutionModeId:
dec->scope = VTN_DEC_EXECUTION_MODE;
break;
default:
struct vtn_type *type;
};
+/**
+ * Returns true if the given type contains a struct decorated Block or
+ * BufferBlock
+ */
+bool
+vtn_type_contains_block(struct vtn_builder *b, struct vtn_type *type)
+{
+ switch (type->base_type) {
+ case vtn_base_type_array:
+ return vtn_type_contains_block(b, type->array_element);
+ case vtn_base_type_struct:
+ if (type->block || type->buffer_block)
+ return true;
+ for (unsigned i = 0; i < type->length; i++) {
+ if (vtn_type_contains_block(b, type->members[i]))
+ return true;
+ }
+ return false;
+ default:
+ return false;
+ }
+}
+
/** Returns true if two types are "compatible", i.e. you can do an OpLoad,
* OpStore, or OpCopyMemory between them without breaking anything.
* Technically, the SPIR-V rules require the exact same type ID but this lets
type->access |= access;
}
+static void
+array_stride_decoration_cb(struct vtn_builder *b,
+ struct vtn_value *val, int member,
+ const struct vtn_decoration *dec, void *void_ctx)
+{
+ struct vtn_type *type = val->type;
+
+ if (dec->decoration == SpvDecorationArrayStride) {
+ vtn_fail_if(dec->literals[0] == 0, "ArrayStride must be non-zero");
+ type->stride = dec->literals[0];
+ }
+}
+
static void
struct_member_decoration_cb(struct vtn_builder *b,
struct vtn_value *val, int member,
break;
case SpvDecorationOffset:
ctx->type->offsets[member] = dec->literals[0];
+ ctx->fields[member].offset = dec->literals[0];
break;
case SpvDecorationMatrixStride:
/* Handled as a second pass */
break;
case SpvDecorationCPacked:
+ if (b->shader->info.stage != MESA_SHADER_KERNEL)
+ vtn_warn("Decoration only allowed for CL-style kernels: %s",
+ spirv_decoration_to_string(dec->decoration));
+ else
+ ctx->type->packed = true;
+ break;
+
case SpvDecorationSaturatedConversion:
case SpvDecorationFuncParamAttr:
case SpvDecorationFPRoundingMode:
case SpvDecorationFPFastMathMode:
case SpvDecorationAlignment:
- vtn_warn("Decoration only allowed for CL-style kernels: %s",
- spirv_decoration_to_string(dec->decoration));
+ if (b->shader->info.stage != MESA_SHADER_KERNEL) {
+ vtn_warn("Decoration only allowed for CL-style kernels: %s",
+ spirv_decoration_to_string(dec->decoration));
+ }
break;
case SpvDecorationHlslSemanticGOOGLE:
}
}
+/** Chases the array type all the way down to the tail and rewrites the
+ * glsl_types to be based off the tail's glsl_type.
+ */
+static void
+vtn_array_type_rewrite_glsl_type(struct vtn_type *type)
+{
+ if (type->base_type != vtn_base_type_array)
+ return;
+
+ vtn_array_type_rewrite_glsl_type(type->array_element);
+
+ type->type = glsl_array_type(type->array_element->type,
+ type->length, type->stride);
+}
+
/* Matrix strides are handled as a separate pass because we need to know
* whether the matrix is row-major or not first.
*/
vtn_fail_if(member < 0,
"The MatrixStride decoration is only allowed on members "
"of OpTypeStruct");
+ vtn_fail_if(dec->literals[0] == 0, "MatrixStride must be non-zero");
struct member_decoration_ctx *ctx = void_ctx;
mat_type->array_element = vtn_type_copy(b, mat_type->array_element);
mat_type->stride = mat_type->array_element->stride;
mat_type->array_element->stride = dec->literals[0];
+
+ mat_type->type = glsl_explicit_matrix_type(mat_type->type,
+ dec->literals[0], true);
+ mat_type->array_element->type = glsl_get_column_type(mat_type->type);
} else {
vtn_assert(mat_type->array_element->stride > 0);
mat_type->stride = dec->literals[0];
+
+ mat_type->type = glsl_explicit_matrix_type(mat_type->type,
+ dec->literals[0], false);
}
+
+ /* Now that we've replaced the glsl_type with a properly strided matrix
+ * type, rewrite the member type so that it's an array of the proper kind
+ * of glsl_type.
+ */
+ vtn_array_type_rewrite_glsl_type(ctx->type->members[member]);
+ ctx->fields[member].type = ctx->type->members[member]->type;
+}
+
+static void
+struct_block_decoration_cb(struct vtn_builder *b,
+ struct vtn_value *val, int member,
+ const struct vtn_decoration *dec, void *ctx)
+{
+ if (member != -1)
+ return;
+
+ struct vtn_type *type = val->type;
+ if (dec->decoration == SpvDecorationBlock)
+ type->block = true;
+ else if (dec->decoration == SpvDecorationBufferBlock)
+ type->buffer_block = true;
}
static void
switch (dec->decoration) {
case SpvDecorationArrayStride:
- vtn_assert(type->base_type == vtn_base_type_matrix ||
- type->base_type == vtn_base_type_array ||
+ vtn_assert(type->base_type == vtn_base_type_array ||
type->base_type == vtn_base_type_pointer);
- type->stride = dec->literals[0];
break;
case SpvDecorationBlock:
vtn_assert(type->base_type == vtn_base_type_struct);
- type->block = true;
+ vtn_assert(type->block);
break;
case SpvDecorationBufferBlock:
vtn_assert(type->base_type == vtn_base_type_struct);
- type->buffer_block = true;
+ vtn_assert(type->buffer_block);
break;
case SpvDecorationGLSLShared:
case SpvDecorationGLSLPacked:
break;
case SpvDecorationCPacked:
+ if (b->shader->info.stage != MESA_SHADER_KERNEL)
+ vtn_warn("Decoration only allowed for CL-style kernels: %s",
+ spirv_decoration_to_string(dec->decoration));
+ else
+ type->packed = true;
+ break;
+
case SpvDecorationSaturatedConversion:
case SpvDecorationFuncParamAttr:
case SpvDecorationFPRoundingMode:
{
switch (type->base_type) {
case vtn_base_type_scalar: {
- uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
+ uint32_t comp_size = glsl_type_is_boolean(type->type)
+ ? 4 : glsl_get_bit_size(type->type) / 8;
*size_out = comp_size;
*align_out = comp_size;
return type;
}
case vtn_base_type_vector: {
- uint32_t comp_size = glsl_get_bit_size(type->type) / 8;
+ uint32_t comp_size = glsl_type_is_boolean(type->type)
+ ? 4 : glsl_get_bit_size(type->type) / 8;
unsigned align_comps = type->length == 3 ? 4 : type->length;
*size_out = comp_size * type->length,
*align_out = comp_size * align_comps;
vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
{
- struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
+ struct vtn_value *val = NULL;
- val->type = rzalloc(b, struct vtn_type);
- val->type->id = w[1];
+ /* In order to properly handle forward declarations, we have to defer
+ * allocation for pointer types.
+ */
+ if (opcode != SpvOpTypePointer && opcode != SpvOpTypeForwardPointer) {
+ val = vtn_push_value(b, w[1], vtn_value_type_type);
+ vtn_fail_if(val->type != NULL,
+ "Only pointers can have forward declarations");
+ val->type = rzalloc(b, struct vtn_type);
+ val->type->id = w[1];
+ }
switch (opcode) {
case SpvOpTypeVoid:
val->type->base_type = vtn_base_type_vector;
val->type->type = glsl_vector_type(glsl_get_base_type(base->type), elems);
val->type->length = elems;
- val->type->stride = glsl_get_bit_size(base->type) / 8;
+ val->type->stride = glsl_type_is_boolean(val->type->type)
+ ? 4 : glsl_get_bit_size(base->type) / 8;
val->type->array_element = base;
break;
}
}
val->type->base_type = vtn_base_type_array;
- val->type->type = glsl_array_type(array_element->type, val->type->length);
val->type->array_element = array_element;
- val->type->stride = 0;
+ if (b->shader->info.stage == MESA_SHADER_KERNEL)
+ val->type->stride = glsl_get_cl_size(array_element->type);
+
+ vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL);
+ val->type->type = glsl_array_type(array_element->type, val->type->length,
+ val->type->stride);
break;
}
val->type->length = num_fields;
val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
val->type->offsets = ralloc_array(b, unsigned, num_fields);
+ val->type->packed = false;
NIR_VLA(struct glsl_struct_field, fields, count);
for (unsigned i = 0; i < num_fields; i++) {
.type = val->type->members[i]->type,
.name = ralloc_asprintf(b, "field%d", i),
.location = -1,
+ .offset = -1,
};
}
+ if (b->shader->info.stage == MESA_SHADER_KERNEL) {
+ unsigned offset = 0;
+ for (unsigned i = 0; i < num_fields; i++) {
+ offset = align(offset, glsl_get_cl_alignment(fields[i].type));
+ fields[i].offset = offset;
+ offset += glsl_get_cl_size(fields[i].type);
+ }
+ }
+
struct member_decoration_ctx ctx = {
.num_fields = num_fields,
.fields = fields,
vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
vtn_foreach_decoration(b, val, struct_member_matrix_stride_cb, &ctx);
- const char *name = val->name ? val->name : "struct";
+ vtn_foreach_decoration(b, val, struct_block_decoration_cb, NULL);
- val->type->type = glsl_struct_type(fields, num_fields, name);
+ const char *name = val->name;
+
+ if (val->type->block || val->type->buffer_block) {
+ /* Packing will be ignored since types coming from SPIR-V are
+ * explicitly laid out.
+ */
+ val->type->type = glsl_interface_type(fields, num_fields,
+ /* packing */ 0, false,
+ name ? name : "block");
+ } else {
+ val->type->type = glsl_struct_type(fields, num_fields,
+ name ? name : "struct", false);
+ }
break;
}
break;
}
- case SpvOpTypePointer: {
+ case SpvOpTypePointer:
+ case SpvOpTypeForwardPointer: {
+ /* We can't blindly push the value because it might be a forward
+ * declaration.
+ */
+ val = vtn_untyped_value(b, w[1]);
+
SpvStorageClass storage_class = w[2];
- struct vtn_type *deref_type =
- vtn_value(b, w[3], vtn_value_type_type)->type;
- val->type->base_type = vtn_base_type_pointer;
- val->type->storage_class = storage_class;
- val->type->deref = deref_type;
+ if (val->value_type == vtn_value_type_invalid) {
+ val->value_type = vtn_value_type_type;
+ val->type = rzalloc(b, struct vtn_type);
+ val->type->id = w[1];
+ val->type->base_type = vtn_base_type_pointer;
+ val->type->storage_class = storage_class;
- if (storage_class == SpvStorageClassUniform ||
- storage_class == SpvStorageClassStorageBuffer) {
/* These can actually be stored to nir_variables and used as SSA
* values so they need a real glsl_type.
*/
- val->type->type = glsl_vector_type(GLSL_TYPE_UINT, 2);
+ switch (storage_class) {
+ case SpvStorageClassUniform:
+ val->type->type = b->options->ubo_ptr_type;
+ break;
+ case SpvStorageClassStorageBuffer:
+ val->type->type = b->options->ssbo_ptr_type;
+ break;
+ case SpvStorageClassPhysicalStorageBufferEXT:
+ val->type->type = b->options->phys_ssbo_ptr_type;
+ break;
+ case SpvStorageClassPushConstant:
+ val->type->type = b->options->push_const_ptr_type;
+ break;
+ case SpvStorageClassWorkgroup:
+ val->type->type = b->options->shared_ptr_type;
+ break;
+ case SpvStorageClassCrossWorkgroup:
+ val->type->type = b->options->global_ptr_type;
+ break;
+ case SpvStorageClassFunction:
+ if (b->physical_ptrs)
+ val->type->type = b->options->temp_ptr_type;
+ break;
+ default:
+ /* In this case, no variable pointers are allowed so all deref
+ * chains are complete back to the variable and it doesn't matter
+ * what type gets used so we leave it NULL.
+ */
+ break;
+ }
+ } else {
+ vtn_fail_if(val->type->storage_class != storage_class,
+ "The storage classes of an OpTypePointer and any "
+ "OpTypeForwardPointers that provide forward "
+ "declarations of it must match.");
}
- if (storage_class == SpvStorageClassPushConstant) {
- /* These can actually be stored to nir_variables and used as SSA
- * values so they need a real glsl_type.
- */
- val->type->type = glsl_uint_type();
- }
+ if (opcode == SpvOpTypePointer) {
+ vtn_fail_if(val->type->deref != NULL,
+ "While OpTypeForwardPointer can be used to provide a "
+ "forward declaration of a pointer, OpTypePointer can "
+ "only be used once for a given id.");
- if (storage_class == SpvStorageClassWorkgroup &&
- b->options->lower_workgroup_access_to_offsets) {
- uint32_t size, align;
- val->type->deref = vtn_type_layout_std430(b, val->type->deref,
- &size, &align);
- val->type->length = size;
- val->type->align = align;
- /* These can actually be stored to nir_variables and used as SSA
- * values so they need a real glsl_type.
- */
- val->type->type = glsl_uint_type();
+ val->type->deref = vtn_value(b, w[3], vtn_value_type_type)->type;
+
+ vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL);
+
+ if (b->physical_ptrs) {
+ switch (storage_class) {
+ case SpvStorageClassFunction:
+ case SpvStorageClassWorkgroup:
+ case SpvStorageClassCrossWorkgroup:
+ val->type->stride = align(glsl_get_cl_size(val->type->deref->type),
+ glsl_get_cl_alignment(val->type->deref->type));
+ break;
+ default:
+ break;
+ }
+ }
+
+ if (storage_class == SpvStorageClassWorkgroup &&
+ b->options->lower_workgroup_access_to_offsets) {
+ uint32_t size, align;
+ val->type->deref = vtn_type_layout_std430(b, val->type->deref,
+ &size, &align);
+ val->type->length = size;
+ val->type->align = align;
+ }
}
break;
}
vtn_fail("Invalid SPIR-V image dimensionality");
}
- bool is_shadow = w[4];
+ /* w[4]: as per Vulkan spec "Validation Rules within a Module",
+ * The “Depth” operand of OpTypeImage is ignored.
+ */
bool is_array = w[5];
bool multisampled = w[6];
unsigned sampled = w[7];
glsl_get_base_type(sampled_type->type);
if (sampled == 1) {
val->type->sampled = true;
- val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
+ val->type->type = glsl_sampler_type(dim, false, is_array,
sampled_base_type);
} else if (sampled == 2) {
- vtn_assert(!is_shadow);
val->type->sampled = false;
val->type->type = glsl_image_type(dim, is_array, sampled_base_type);
} else {
}
vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
+
+ if (val->type->base_type == vtn_base_type_struct &&
+ (val->type->block || val->type->buffer_block)) {
+ for (unsigned i = 0; i < val->type->length; i++) {
+ vtn_fail_if(vtn_type_contains_block(b, val->type->members[i]),
+ "Block and BufferBlock decorations cannot decorate a "
+ "structure type that is nested at any level inside "
+ "another structure type decorated with Block or "
+ "BufferBlock.");
+ }
+ }
}
static nir_constant *
return;
vtn_assert(val->type->type == glsl_vector_type(GLSL_TYPE_UINT, 3));
-
- b->shader->info.cs.local_size[0] = val->constant->values[0].u32[0];
- b->shader->info.cs.local_size[1] = val->constant->values[0].u32[1];
- b->shader->info.cs.local_size[2] = val->constant->values[0].u32[2];
+ b->workgroup_size_builtin = val;
}
static void
opcode == SpvOpSpecConstantFalse)
int_val = get_specialization(b, val, int_val);
- val->constant->values[0].u32[0] = int_val ? NIR_TRUE : NIR_FALSE;
+ val->constant->values[0].b[0] = int_val != 0;
break;
}
case 8:
val->constant->values[0].u8[i] = elems[i]->values[0].u8[0];
break;
+ case 1:
+ val->constant->values[0].b[i] = elems[i]->values[0].b[0];
+ break;
default:
vtn_fail("Invalid SpvOpConstantComposite bit size");
}
case 8:
val->constant->values[0].u8[i] = (*c)->values[col].u8[elem + i];
break;
+ case 1:
+ val->constant->values[0].b[i] = (*c)->values[col].b[elem + i];
+ break;
default:
vtn_fail("Invalid SpvOpCompositeExtract bit size");
}
case 8:
(*c)->values[col].u8[elem + i] = insert->constant->values[0].u8[i];
break;
+ case 1:
+ (*c)->values[col].b[elem + i] = insert->constant->values[0].b[i];
+ break;
default:
vtn_fail("Invalid SpvOpCompositeInsert bit size");
}
child_type = glsl_get_array_element(type);
break;
case GLSL_TYPE_STRUCT:
+ case GLSL_TYPE_INTERFACE:
child_type = glsl_get_struct_field(type, i);
break;
default:
case nir_texop_txl:
case nir_texop_txd:
case nir_texop_tg4:
+ case nir_texop_lod:
/* These operations require a sampler */
p->src = nir_src_for_ssa(&sampler->dest.ssa);
p->src_type = nir_tex_src_sampler_deref;
case nir_texop_txf:
case nir_texop_txf_ms:
case nir_texop_txs:
- case nir_texop_lod:
case nir_texop_query_levels:
case nir_texop_texture_samples:
case nir_texop_samples_identical:
(*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
/* Now we need to handle some number of optional arguments */
- const struct vtn_ssa_value *gather_offsets = NULL;
+ struct vtn_value *gather_offsets = NULL;
if (idx < count) {
uint32_t operands = w[idx++];
(*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
if (operands & SpvImageOperandsConstOffsetsMask) {
- nir_tex_src none = {0};
- gather_offsets = vtn_ssa_value(b, w[idx++]);
- (*p++) = none;
+ vtn_assert(texop == nir_texop_tg4);
+ gather_offsets = vtn_value(b, w[idx++], vtn_value_type_constant);
}
if (operands & SpvImageOperandsSampleMask) {
texop = nir_texop_txf_ms;
(*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
}
+
+ if (operands & SpvImageOperandsMinLodMask) {
+ vtn_assert(texop == nir_texop_tex ||
+ texop == nir_texop_txb ||
+ texop == nir_texop_txd);
+ (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_min_lod);
+ }
}
/* We should have now consumed exactly all of the arguments */
vtn_assert(idx == count);
vtn_assert(glsl_get_vector_elements(ret_type->type) ==
nir_tex_instr_dest_size(instr));
- nir_ssa_def *def;
- nir_instr *instruction;
if (gather_offsets) {
- vtn_assert(glsl_get_base_type(gather_offsets->type) == GLSL_TYPE_ARRAY);
- vtn_assert(glsl_get_length(gather_offsets->type) == 4);
- nir_tex_instr *instrs[4] = {instr, NULL, NULL, NULL};
-
- /* Copy the current instruction 4x */
- for (uint32_t i = 1; i < 4; i++) {
- instrs[i] = nir_tex_instr_create(b->shader, instr->num_srcs);
- instrs[i]->op = instr->op;
- instrs[i]->coord_components = instr->coord_components;
- instrs[i]->sampler_dim = instr->sampler_dim;
- instrs[i]->is_array = instr->is_array;
- instrs[i]->is_shadow = instr->is_shadow;
- instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
- instrs[i]->component = instr->component;
- instrs[i]->dest_type = instr->dest_type;
-
- memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
-
- nir_ssa_dest_init(&instrs[i]->instr, &instrs[i]->dest,
- nir_tex_instr_dest_size(instr), 32, NULL);
- }
-
- /* Fill in the last argument with the offset from the passed in offsets
- * and insert the instruction into the stream.
- */
- for (uint32_t i = 0; i < 4; i++) {
- nir_tex_src src;
- src.src = nir_src_for_ssa(gather_offsets->elems[i]->def);
- src.src_type = nir_tex_src_offset;
- instrs[i]->src[instrs[i]->num_srcs - 1] = src;
- nir_builder_instr_insert(&b->nb, &instrs[i]->instr);
- }
-
- /* Combine the results of the 4 instructions by taking their .w
- * components
- */
- nir_alu_instr *vec4 = nir_alu_instr_create(b->shader, nir_op_vec4);
- nir_ssa_dest_init(&vec4->instr, &vec4->dest.dest, 4, 32, NULL);
- vec4->dest.write_mask = 0xf;
+ vtn_fail_if(gather_offsets->type->base_type != vtn_base_type_array ||
+ gather_offsets->type->length != 4,
+ "ConstOffsets must be an array of size four of vectors "
+ "of two integer components");
+
+ struct vtn_type *vec_type = gather_offsets->type->array_element;
+ vtn_fail_if(vec_type->base_type != vtn_base_type_vector ||
+ vec_type->length != 2 ||
+ !glsl_type_is_integer(vec_type->type),
+ "ConstOffsets must be an array of size four of vectors "
+ "of two integer components");
+
+ unsigned bit_size = glsl_get_bit_size(vec_type->type);
for (uint32_t i = 0; i < 4; i++) {
- vec4->src[i].src = nir_src_for_ssa(&instrs[i]->dest.ssa);
- vec4->src[i].swizzle[0] = 3;
+ const nir_const_value *cvec =
+ &gather_offsets->constant->elements[i]->values[0];
+ for (uint32_t j = 0; j < 2; j++) {
+ switch (bit_size) {
+ case 8: instr->tg4_offsets[i][j] = cvec->i8[j]; break;
+ case 16: instr->tg4_offsets[i][j] = cvec->i16[j]; break;
+ case 32: instr->tg4_offsets[i][j] = cvec->i32[j]; break;
+ case 64: instr->tg4_offsets[i][j] = cvec->i64[j]; break;
+ default:
+ vtn_fail("Unsupported bit size");
+ }
+ }
}
- def = &vec4->dest.dest.ssa;
- instruction = &vec4->instr;
- } else {
- def = &instr->dest.ssa;
- instruction = &instr->instr;
}
val->ssa = vtn_create_ssa_value(b, ret_type->type);
- val->ssa->def = def;
+ val->ssa->def = &instr->dest.ssa;
- nir_builder_instr_insert(&b->nb, instruction);
+ nir_builder_instr_insert(&b->nb, &instr->instr);
}
static void
unreachable("Invalid SPIR-V atomic");
}
- } else if (ptr->mode == vtn_variable_mode_workgroup &&
- !b->options->lower_workgroup_access_to_offsets) {
- nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
- const struct glsl_type *deref_type = deref->type;
- nir_intrinsic_op op = get_deref_nir_atomic_op(b, opcode);
+ } else if (vtn_pointer_uses_ssa_offset(b, ptr)) {
+ nir_ssa_def *offset, *index;
+ offset = vtn_pointer_to_offset(b, ptr, &index);
+
+ nir_intrinsic_op op;
+ if (ptr->mode == vtn_variable_mode_ssbo) {
+ op = get_ssbo_nir_atomic_op(b, opcode);
+ } else {
+ vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
+ b->options->lower_workgroup_access_to_offsets);
+ op = get_shared_nir_atomic_op(b, opcode);
+ }
+
atomic = nir_intrinsic_instr_create(b->nb.shader, op);
- atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
+ int src = 0;
switch (opcode) {
case SpvOpAtomicLoad:
- atomic->num_components = glsl_get_vector_elements(deref_type);
+ atomic->num_components = glsl_get_vector_elements(ptr->type->type);
+ nir_intrinsic_set_align(atomic, 4, 0);
+ if (ptr->mode == vtn_variable_mode_ssbo)
+ atomic->src[src++] = nir_src_for_ssa(index);
+ atomic->src[src++] = nir_src_for_ssa(offset);
break;
case SpvOpAtomicStore:
- atomic->num_components = glsl_get_vector_elements(deref_type);
+ atomic->num_components = glsl_get_vector_elements(ptr->type->type);
nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
- atomic->src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
+ nir_intrinsic_set_align(atomic, 4, 0);
+ atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
+ if (ptr->mode == vtn_variable_mode_ssbo)
+ atomic->src[src++] = nir_src_for_ssa(index);
+ atomic->src[src++] = nir_src_for_ssa(offset);
break;
case SpvOpAtomicExchange:
case SpvOpAtomicAnd:
case SpvOpAtomicOr:
case SpvOpAtomicXor:
- fill_common_atomic_sources(b, opcode, w, &atomic->src[1]);
+ if (ptr->mode == vtn_variable_mode_ssbo)
+ atomic->src[src++] = nir_src_for_ssa(index);
+ atomic->src[src++] = nir_src_for_ssa(offset);
+ fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
break;
default:
vtn_fail("Invalid SPIR-V atomic");
-
}
} else {
- nir_ssa_def *offset, *index;
- offset = vtn_pointer_to_offset(b, ptr, &index);
-
- nir_intrinsic_op op;
- if (ptr->mode == vtn_variable_mode_ssbo) {
- op = get_ssbo_nir_atomic_op(b, opcode);
- } else {
- vtn_assert(ptr->mode == vtn_variable_mode_workgroup &&
- b->options->lower_workgroup_access_to_offsets);
- op = get_shared_nir_atomic_op(b, opcode);
- }
-
+ nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
+ const struct glsl_type *deref_type = deref->type;
+ nir_intrinsic_op op = get_deref_nir_atomic_op(b, opcode);
atomic = nir_intrinsic_instr_create(b->nb.shader, op);
+ atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
- int src = 0;
switch (opcode) {
case SpvOpAtomicLoad:
- atomic->num_components = glsl_get_vector_elements(ptr->type->type);
- nir_intrinsic_set_align(atomic, 4, 0);
- if (ptr->mode == vtn_variable_mode_ssbo)
- atomic->src[src++] = nir_src_for_ssa(index);
- atomic->src[src++] = nir_src_for_ssa(offset);
+ atomic->num_components = glsl_get_vector_elements(deref_type);
break;
case SpvOpAtomicStore:
- atomic->num_components = glsl_get_vector_elements(ptr->type->type);
+ atomic->num_components = glsl_get_vector_elements(deref_type);
nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
- nir_intrinsic_set_align(atomic, 4, 0);
- atomic->src[src++] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
- if (ptr->mode == vtn_variable_mode_ssbo)
- atomic->src[src++] = nir_src_for_ssa(index);
- atomic->src[src++] = nir_src_for_ssa(offset);
+ atomic->src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
break;
case SpvOpAtomicExchange:
case SpvOpAtomicAnd:
case SpvOpAtomicOr:
case SpvOpAtomicXor:
- if (ptr->mode == vtn_variable_mode_ssbo)
- atomic->src[src++] = nir_src_for_ssa(index);
- atomic->src[src++] = nir_src_for_ssa(offset);
- fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
+ fill_common_atomic_sources(b, opcode, w, &atomic->src[1]);
break;
default:
return &vec->dest.dest.ssa;
}
+static nir_ssa_def *
+nir_ieq_imm(nir_builder *b, nir_ssa_def *x, uint64_t i)
+{
+ return nir_ieq(b, x, nir_imm_intN_t(b, i, x->bit_size));
+}
+
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;
+ return nir_vector_extract(&b->nb, src, nir_i2i(&b->nb, index, 32));
}
nir_ssa_def *
{
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)),
+ dest = nir_bcsel(&b->nb, nir_ieq_imm(&b->nb, index, i),
vtn_vector_insert(b, src, insert, i), dest);
return dest;
switch (opcode) {
case SpvOpEmitStreamVertex:
case SpvOpEndStreamPrimitive: {
- unsigned stream = vtn_constant_value(b, w[1])->values[0].u32[0];
+ unsigned stream = vtn_constant_uint(b, w[1]);
nir_intrinsic_set_stream_id(intrin, stream);
break;
}
}
case SpvOpMemoryBarrier: {
- SpvScope scope = vtn_constant_value(b, w[1])->values[0].u32[0];
- SpvMemorySemanticsMask semantics =
- vtn_constant_value(b, w[2])->values[0].u32[0];
+ SpvScope scope = vtn_constant_uint(b, w[1]);
+ SpvMemorySemanticsMask semantics = vtn_constant_uint(b, w[2]);
vtn_emit_memory_barrier(b, scope, semantics);
return;
}
case SpvOpControlBarrier: {
- SpvScope execution_scope =
- vtn_constant_value(b, w[1])->values[0].u32[0];
+ SpvScope execution_scope = vtn_constant_uint(b, w[1]);
if (execution_scope == SpvScopeWorkgroup)
vtn_emit_barrier(b, nir_intrinsic_barrier);
- SpvScope memory_scope =
- vtn_constant_value(b, w[2])->values[0].u32[0];
- SpvMemorySemanticsMask memory_semantics =
- vtn_constant_value(b, w[3])->values[0].u32[0];
+ SpvScope memory_scope = vtn_constant_uint(b, w[2]);
+ SpvMemorySemanticsMask memory_semantics = vtn_constant_uint(b, w[3]);
vtn_emit_memory_barrier(b, memory_scope, memory_semantics);
break;
}
return MESA_SHADER_FRAGMENT;
case SpvExecutionModelGLCompute:
return MESA_SHADER_COMPUTE;
+ case SpvExecutionModelKernel:
+ return MESA_SHADER_KERNEL;
default:
vtn_fail("Unsupported execution model");
}
case SpvCapabilityVector16:
case SpvCapabilityFloat16Buffer:
case SpvCapabilityFloat16:
- case SpvCapabilityInt64Atomics:
- case SpvCapabilityStorageImageMultisample:
- case SpvCapabilityInt8:
case SpvCapabilitySparseResidency:
- case SpvCapabilityMinLod:
vtn_warn("Unsupported SPIR-V capability: %s",
spirv_capability_to_string(cap));
break;
+ case SpvCapabilityMinLod:
+ spv_check_supported(min_lod, cap);
+ break;
+
case SpvCapabilityAtomicStorage:
spv_check_supported(atomic_storage, cap);
break;
spv_check_supported(geometry_streams, cap);
break;
+ case SpvCapabilityInt64Atomics:
+ spv_check_supported(int64_atomics, cap);
+ break;
+
+ case SpvCapabilityInt8:
+ spv_check_supported(int8, cap);
+ break;
+
+ case SpvCapabilityStorageImageMultisample:
+ spv_check_supported(storage_image_ms, cap);
+ break;
+
case SpvCapabilityAddresses:
+ spv_check_supported(address, cap);
+ break;
+
case SpvCapabilityKernel:
+ spv_check_supported(kernel, cap);
+ break;
+
case SpvCapabilityImageBasic:
case SpvCapabilityImageReadWrite:
case SpvCapabilityImageMipmap:
case SpvCapabilityVariablePointersStorageBuffer:
case SpvCapabilityVariablePointers:
spv_check_supported(variable_pointers, cap);
+ b->variable_pointers = true;
break;
case SpvCapabilityStorageUniformBufferBlock16:
spv_check_supported(post_depth_coverage, cap);
break;
+ case SpvCapabilityPhysicalStorageBufferAddressesEXT:
+ spv_check_supported(physical_storage_buffer_address, cap);
+ break;
+
default:
vtn_fail("Unhandled capability");
}
break;
case SpvOpMemoryModel:
- vtn_assert(w[1] == SpvAddressingModelLogical);
+ switch (w[1]) {
+ case SpvAddressingModelPhysical32:
+ vtn_fail_if(b->shader->info.stage != MESA_SHADER_KERNEL,
+ "AddressingModelPhysical32 only supported for kernels");
+ b->shader->info.cs.ptr_size = 32;
+ b->physical_ptrs = true;
+ b->options->shared_ptr_type = glsl_uint_type();
+ b->options->global_ptr_type = glsl_uint_type();
+ b->options->temp_ptr_type = glsl_uint_type();
+ break;
+ case SpvAddressingModelPhysical64:
+ vtn_fail_if(b->shader->info.stage != MESA_SHADER_KERNEL,
+ "AddressingModelPhysical64 only supported for kernels");
+ b->shader->info.cs.ptr_size = 64;
+ b->physical_ptrs = true;
+ b->options->shared_ptr_type = glsl_uint64_t_type();
+ b->options->global_ptr_type = glsl_uint64_t_type();
+ b->options->temp_ptr_type = glsl_uint64_t_type();
+ break;
+ case SpvAddressingModelLogical:
+ vtn_fail_if(b->shader->info.stage >= MESA_SHADER_STAGES,
+ "AddressingModelLogical only supported for shaders");
+ b->shader->info.cs.ptr_size = 0;
+ b->physical_ptrs = false;
+ break;
+ case SpvAddressingModelPhysicalStorageBuffer64EXT:
+ vtn_fail_if(!b->options ||
+ !b->options->caps.physical_storage_buffer_address,
+ "AddressingModelPhysicalStorageBuffer64EXT not supported");
+ break;
+ default:
+ vtn_fail("Unknown addressing model");
+ break;
+ }
+
vtn_assert(w[2] == SpvMemoryModelSimple ||
- w[2] == SpvMemoryModelGLSL450);
+ w[2] == SpvMemoryModelGLSL450 ||
+ w[2] == SpvMemoryModelOpenCL);
break;
case SpvOpEntryPoint:
break;
case SpvOpExecutionMode:
+ case SpvOpExecutionModeId:
case SpvOpDecorationGroup:
case SpvOpDecorate:
case SpvOpMemberDecorate:
switch(mode->exec_mode) {
case SpvExecutionModeOriginUpperLeft:
case SpvExecutionModeOriginLowerLeft:
- b->origin_upper_left =
+ vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.origin_upper_left =
(mode->exec_mode == SpvExecutionModeOriginUpperLeft);
break;
break;
case SpvExecutionModeLocalSize:
- vtn_assert(b->shader->info.stage == MESA_SHADER_COMPUTE);
+ vtn_assert(gl_shader_stage_is_compute(b->shader->info.stage));
b->shader->info.cs.local_size[0] = mode->literals[0];
b->shader->info.cs.local_size[1] = mode->literals[1];
b->shader->info.cs.local_size[2] = mode->literals[2];
break;
+
+ case SpvExecutionModeLocalSizeId:
+ b->shader->info.cs.local_size[0] = vtn_constant_uint(b, mode->literals[0]);
+ b->shader->info.cs.local_size[1] = vtn_constant_uint(b, mode->literals[1]);
+ b->shader->info.cs.local_size[2] = vtn_constant_uint(b, mode->literals[2]);
+ break;
+
case SpvExecutionModeLocalSizeHint:
+ case SpvExecutionModeLocalSizeHintId:
break; /* Nothing to do with this */
case SpvExecutionModeOutputVertices:
break;
case SpvExecutionModePixelCenterInteger:
- b->pixel_center_integer = true;
+ vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.pixel_center_integer = true;
break;
case SpvExecutionModeXfb:
break;
case SpvExecutionModeVecTypeHint:
- case SpvExecutionModeContractionOff:
break; /* OpenCL */
+ case SpvExecutionModeContractionOff:
+ if (b->shader->info.stage != MESA_SHADER_KERNEL)
+ vtn_warn("ExectionMode only allowed for CL-style kernels: %s",
+ spirv_executionmode_to_string(mode->exec_mode));
+ else
+ b->exact = true;
+ break;
+
case SpvExecutionModeStencilRefReplacingEXT:
vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
break;
case SpvOpTypeStruct:
case SpvOpTypeOpaque:
case SpvOpTypePointer:
+ case SpvOpTypeForwardPointer:
case SpvOpTypeFunction:
case SpvOpTypeEvent:
case SpvOpTypeDeviceEvent:
case SpvOpAccessChain:
case SpvOpPtrAccessChain:
case SpvOpInBoundsAccessChain:
+ case SpvOpInBoundsPtrAccessChain:
case SpvOpArrayLength:
+ case SpvOpConvertPtrToU:
+ case SpvOpConvertUToPtr:
vtn_handle_variables(b, opcode, w, count);
break;
case SpvOpSConvert:
case SpvOpFConvert:
case SpvOpQuantizeToF16:
- case SpvOpConvertPtrToU:
- case SpvOpConvertUToPtr:
case SpvOpPtrCastToGeneric:
case SpvOpGenericCastToPtr:
case SpvOpBitcast:
{
/* Initialize the vtn_builder object */
struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
+ struct spirv_to_nir_options *dup_options =
+ ralloc(b, struct spirv_to_nir_options);
+ *dup_options = *options;
+
b->spirv = words;
b->spirv_word_count = word_count;
b->file = NULL;
exec_list_make_empty(&b->functions);
b->entry_point_stage = stage;
b->entry_point_name = entry_point_name;
- b->options = options;
+ b->options = dup_options;
/*
* Handle the SPIR-V header (first 5 dwords).
goto fail;
}
+ uint16_t generator_id = words[2] >> 16;
+ uint16_t generator_version = words[2];
+
+ /* The first GLSLang version bump actually 1.5 years after #179 was fixed
+ * but this should at least let us shut the workaround off for modern
+ * versions of GLSLang.
+ */
+ b->wa_glslang_179 = (generator_id == 8 && generator_version == 1);
+
/* words[2] == generator magic */
unsigned value_id_bound = words[3];
if (words[4] != 0) {
return NULL;
}
+static nir_function *
+vtn_emit_kernel_entry_point_wrapper(struct vtn_builder *b,
+ nir_function *entry_point)
+{
+ vtn_assert(entry_point == b->entry_point->func->impl->function);
+ vtn_fail_if(!entry_point->name, "entry points are required to have a name");
+ const char *func_name =
+ ralloc_asprintf(b->shader, "__wrapped_%s", entry_point->name);
+
+ /* we shouldn't have any inputs yet */
+ vtn_assert(!entry_point->shader->num_inputs);
+ vtn_assert(b->shader->info.stage == MESA_SHADER_KERNEL);
+
+ nir_function *main_entry_point = nir_function_create(b->shader, func_name);
+ main_entry_point->impl = nir_function_impl_create(main_entry_point);
+ nir_builder_init(&b->nb, main_entry_point->impl);
+ b->nb.cursor = nir_after_cf_list(&main_entry_point->impl->body);
+ b->func_param_idx = 0;
+
+ nir_call_instr *call = nir_call_instr_create(b->nb.shader, entry_point);
+
+ for (unsigned i = 0; i < entry_point->num_params; ++i) {
+ struct vtn_type *param_type = b->entry_point->func->type->params[i];
+
+ /* consider all pointers to function memory to be parameters passed
+ * by value
+ */
+ bool is_by_val = param_type->base_type == vtn_base_type_pointer &&
+ param_type->storage_class == SpvStorageClassFunction;
+
+ /* input variable */
+ nir_variable *in_var = rzalloc(b->nb.shader, nir_variable);
+ in_var->data.mode = nir_var_shader_in;
+ in_var->data.read_only = true;
+ in_var->data.location = i;
+
+ if (is_by_val)
+ in_var->type = param_type->deref->type;
+ else
+ in_var->type = param_type->type;
+
+ nir_shader_add_variable(b->nb.shader, in_var);
+ b->nb.shader->num_inputs++;
+
+ /* we have to copy the entire variable into function memory */
+ if (is_by_val) {
+ nir_variable *copy_var =
+ nir_local_variable_create(main_entry_point->impl, in_var->type,
+ "copy_in");
+ nir_copy_var(&b->nb, copy_var, in_var);
+ call->params[i] =
+ nir_src_for_ssa(&nir_build_deref_var(&b->nb, copy_var)->dest.ssa);
+ } else {
+ call->params[i] = nir_src_for_ssa(nir_load_var(&b->nb, in_var));
+ }
+ }
+
+ nir_builder_instr_insert(&b->nb, &call->instr);
+
+ return main_entry_point;
+}
+
nir_function *
spirv_to_nir(const uint32_t *words, size_t word_count,
struct nir_spirv_specialization *spec, unsigned num_spec,
/* Skip the SPIR-V header, handled at vtn_create_builder */
words+= 5;
+ b->shader = nir_shader_create(b, stage, nir_options, NULL);
+
/* Handle all the preamble instructions */
words = vtn_foreach_instruction(b, words, word_end,
vtn_handle_preamble_instruction);
return NULL;
}
- b->shader = nir_shader_create(b, stage, nir_options, NULL);
-
/* Set shader info defaults */
b->shader->info.gs.invocations = 1;
- /* Parse execution modes */
- vtn_foreach_execution_mode(b, b->entry_point,
- vtn_handle_execution_mode, NULL);
-
b->specializations = spec;
b->num_specializations = num_spec;
words = vtn_foreach_instruction(b, words, word_end,
vtn_handle_variable_or_type_instruction);
+ /* Parse execution modes */
+ vtn_foreach_execution_mode(b, b->entry_point,
+ vtn_handle_execution_mode, NULL);
+
+ if (b->workgroup_size_builtin) {
+ vtn_assert(b->workgroup_size_builtin->type->type ==
+ glsl_vector_type(GLSL_TYPE_UINT, 3));
+
+ nir_const_value *const_size =
+ &b->workgroup_size_builtin->constant->values[0];
+
+ b->shader->info.cs.local_size[0] = const_size->u32[0];
+ b->shader->info.cs.local_size[1] = const_size->u32[1];
+ b->shader->info.cs.local_size[2] = const_size->u32[2];
+ }
+
/* Set types on all vtn_values */
vtn_foreach_instruction(b, words, word_end, vtn_set_instruction_result_type);
progress = false;
foreach_list_typed(struct vtn_function, func, node, &b->functions) {
if (func->referenced && !func->emitted) {
- b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ b->const_table = _mesa_pointer_hash_table_create(b);
vtn_function_emit(b, func, vtn_handle_body_instruction);
progress = true;
}
} while (progress);
+ vtn_assert(b->entry_point->value_type == vtn_value_type_function);
+ nir_function *entry_point = b->entry_point->func->impl->function;
+ vtn_assert(entry_point);
+
+ /* post process entry_points with input params */
+ if (entry_point->num_params && b->shader->info.stage == MESA_SHADER_KERNEL)
+ entry_point = vtn_emit_kernel_entry_point_wrapper(b, entry_point);
+
+ entry_point->is_entrypoint = true;
+
+ /* When multiple shader stages exist in the same SPIR-V module, we
+ * generate input and output variables for every stage, in the same
+ * NIR program. These dead variables can be invalid NIR. For example,
+ * TCS outputs must be per-vertex arrays (or decorated 'patch'), while
+ * VS output variables wouldn't be.
+ *
+ * To ensure we have valid NIR, we eliminate any dead inputs and outputs
+ * right away. In order to do so, we must lower any constant initializers
+ * on outputs so nir_remove_dead_variables sees that they're written to.
+ */
+ nir_lower_constant_initializers(b->shader, nir_var_shader_out);
+ nir_remove_dead_variables(b->shader,
+ nir_var_shader_in | nir_var_shader_out);
+
/* We sometimes generate bogus derefs that, while never used, give the
* validator a bit of heartburn. Run dead code to get rid of them.
*/
nir_opt_dce(b->shader);
- vtn_assert(b->entry_point->value_type == vtn_value_type_function);
- nir_function *entry_point = b->entry_point->func->impl->function;
- vtn_assert(entry_point);
-
/* Unparent the shader from the vtn_builder before we delete the builder */
ralloc_steal(NULL, b->shader);
projects / mesa.git / blobdiff