#include "nir/nir_vla.h"
#include "nir/nir_control_flow.h"
#include "nir/nir_constant_expressions.h"
+#include "nir/nir_deref.h"
#include "spirv_info.h"
#include <stdio.h>
va_end(args);
}
+void
+_vtn_err(struct vtn_builder *b, const char *file, unsigned line,
+ const char *fmt, ...)
+{
+ va_list args;
+
+ va_start(args, fmt);
+ vtn_log_err(b, NIR_SPIRV_DEBUG_LEVEL_ERROR, "SPIR-V ERROR:\n",
+ file, line, fmt, args);
+ va_end(args);
+}
+
void
_vtn_fail(struct vtn_builder *b, const char *file, unsigned line,
const char *fmt, ...)
vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
{
+ const char *ext = (const char *)&w[2];
switch (opcode) {
case SpvOpExtInstImport: {
struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
- if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
+ if (strcmp(ext, "GLSL.std.450") == 0) {
val->ext_handler = vtn_handle_glsl450_instruction;
- } else if ((strcmp((const char *)&w[2], "SPV_AMD_gcn_shader") == 0)
+ } else if ((strcmp(ext, "SPV_AMD_gcn_shader") == 0)
&& (b->options && b->options->caps.gcn_shader)) {
val->ext_handler = vtn_handle_amd_gcn_shader_instruction;
- } else if ((strcmp((const char *)&w[2], "SPV_AMD_shader_trinary_minmax") == 0)
+ } 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 {
- vtn_fail("Unsupported extension");
+ vtn_fail("Unsupported extension: %s", ext);
}
break;
}
}
}
-static void
+void
vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
{
case SpvOpDecorate:
case SpvOpMemberDecorate:
+ case SpvOpDecorateStringGOOGLE:
+ case SpvOpMemberDecorateStringGOOGLE:
case SpvOpExecutionMode: {
struct vtn_value *val = vtn_untyped_value(b, target);
struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
switch (opcode) {
case SpvOpDecorate:
+ case SpvOpDecorateStringGOOGLE:
dec->scope = VTN_DEC_DECORATION;
break;
case SpvOpMemberDecorate:
+ case SpvOpMemberDecorateStringGOOGLE:
dec->scope = VTN_DEC_STRUCT_MEMBER0 + *(w++);
vtn_fail_if(dec->scope < VTN_DEC_STRUCT_MEMBER0, /* overflow */
"Member argument of OpMemberDecorate too large");
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
return type;
}
+static void
+vtn_handle_access_qualifier(struct vtn_builder *b, struct vtn_type *type,
+ int member, enum gl_access_qualifier access)
+{
+ type->members[member] = vtn_type_copy(b, type->members[member]);
+ type = type->members[member];
+
+ 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,
assert(member < ctx->num_fields);
switch (dec->decoration) {
+ case SpvDecorationRelaxedPrecision:
+ case SpvDecorationUniform:
+ break; /* FIXME: Do nothing with this for now. */
case SpvDecorationNonWritable:
+ vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_NON_WRITEABLE);
+ break;
case SpvDecorationNonReadable:
- case SpvDecorationRelaxedPrecision:
+ vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_NON_READABLE);
+ break;
case SpvDecorationVolatile:
+ vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_VOLATILE);
+ break;
case SpvDecorationCoherent:
- case SpvDecorationUniform:
- break; /* FIXME: Do nothing with this for now. */
+ vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_COHERENT);
+ break;
case SpvDecorationNoPerspective:
ctx->fields[member].interpolation = INTERP_MODE_NOPERSPECTIVE;
break;
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 */
spirv_decoration_to_string(dec->decoration));
break;
+ case SpvDecorationHlslSemanticGOOGLE:
+ /* HLSL semantic decorations can safely be ignored by the driver. */
+ break;
+
default:
vtn_fail("Unhandled decoration");
}
}
+/** 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
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);
case SpvDecorationNonWritable:
case SpvDecorationNonReadable:
case SpvDecorationUniform:
- case SpvDecorationStream:
case SpvDecorationLocation:
case SpvDecorationComponent:
case SpvDecorationOffset:
case SpvDecorationXfbBuffer:
case SpvDecorationXfbStride:
+ case SpvDecorationHlslSemanticGOOGLE:
vtn_warn("Decoration only allowed for struct members: %s",
spirv_decoration_to_string(dec->decoration));
break;
+ case SpvDecorationStream:
+ /* We don't need to do anything here, as stream is filled up when
+ * aplying the decoration to a variable, just check that if it is not a
+ * struct member, it should be a struct.
+ */
+ vtn_assert(type->base_type == vtn_base_type_struct);
+ break;
+
case SpvDecorationRelaxedPrecision:
case SpvDecorationSpecId:
case SpvDecorationInvariant:
}
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;
+
+ 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->storage_class = storage_class;
val->type->deref = deref_type;
- 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);
- }
-
- 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.
+ vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL);
+
+ /* These can actually be stored to nir_variables and used as SSA
+ * values so they need a real glsl_type.
+ */
+ 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 SpvStorageClassPushConstant:
+ val->type->type = b->options->push_const_ptr_type;
+ break;
+ case SpvStorageClassWorkgroup:
+ val->type->type = b->options->shared_ptr_type;
+ if (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;
+ 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.
*/
- val->type->type = glsl_uint_type();
+ break;
}
break;
}
}
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 *
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;
}
spirv_op_to_string(opcode), elem_count, val->type->length);
nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
- for (unsigned i = 0; i < elem_count; i++)
- elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
+ for (unsigned i = 0; i < elem_count; i++) {
+ struct vtn_value *val = vtn_untyped_value(b, w[i + 3]);
+
+ if (val->value_type == vtn_value_type_constant) {
+ elems[i] = val->constant;
+ } else {
+ vtn_fail_if(val->value_type != vtn_value_type_undef,
+ "only constants or undefs allowed for "
+ "SpvOpConstantComposite");
+ /* to make it easier, just insert a NULL constant for now */
+ elems[i] = vtn_null_constant(b, val->type->type);
+ }
+ }
switch (val->type->base_type) {
case vtn_base_type_vector: {
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");
}
nir_const_value src[4];
for (unsigned i = 0; i < count - 4; i++) {
- nir_constant *c =
- vtn_value(b, w[4 + i], vtn_value_type_constant)->constant;
+ struct vtn_value *src_val =
+ vtn_value(b, w[4 + i], vtn_value_type_constant);
+
+ /* If this is an unsized source, pull the bit size from the
+ * source; otherwise, we'll use the bit size from the destination.
+ */
+ if (!nir_alu_type_get_type_size(nir_op_infos[op].input_types[i]))
+ bit_size = glsl_get_bit_size(src_val->type->type);
unsigned j = swap ? 1 - i : i;
- src[j] = c->values[0];
+ src[j] = src_val->constant->values[0];
+ }
+
+ /* fix up fixed size sources */
+ switch (op) {
+ case nir_op_ishl:
+ case nir_op_ishr:
+ case nir_op_ushr: {
+ if (bit_size == 32)
+ break;
+ for (unsigned i = 0; i < num_components; ++i) {
+ switch (bit_size) {
+ case 64: src[1].u32[i] = src[1].u64[i]; break;
+ case 16: src[1].u32[i] = src[1].u16[i]; break;
+ case 8: src[1].u32[i] = src[1].u8[i]; break;
+ }
+ }
+ break;
+ }
+ default:
+ break;
}
val->constant->values[0] =
vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
}
-static void
-vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
- const uint32_t *w, unsigned count)
-{
- struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type;
- struct vtn_function *vtn_callee =
- vtn_value(b, w[3], vtn_value_type_function)->func;
- struct nir_function *callee = vtn_callee->impl->function;
-
- vtn_callee->referenced = true;
-
- nir_call_instr *call = nir_call_instr_create(b->nb.shader, callee);
- for (unsigned i = 0; i < call->num_params; i++) {
- unsigned arg_id = w[4 + i];
- struct vtn_value *arg = vtn_untyped_value(b, arg_id);
- if (arg->value_type == vtn_value_type_pointer &&
- arg->pointer->ptr_type->type == NULL) {
- nir_deref_var *d = vtn_pointer_to_deref(b, arg->pointer);
- call->params[i] = nir_deref_var_clone(d, call);
- } else {
- struct vtn_ssa_value *arg_ssa = vtn_ssa_value(b, arg_id);
-
- /* Make a temporary to store the argument in */
- nir_variable *tmp =
- nir_local_variable_create(b->nb.impl, arg_ssa->type, "arg_tmp");
- call->params[i] = nir_deref_var_create(call, tmp);
-
- vtn_local_store(b, arg_ssa, call->params[i]);
- }
- }
-
- nir_variable *out_tmp = NULL;
- vtn_assert(res_type->type == callee->return_type);
- if (!glsl_type_is_void(callee->return_type)) {
- out_tmp = nir_local_variable_create(b->nb.impl, callee->return_type,
- "out_tmp");
- call->return_deref = nir_deref_var_create(call, out_tmp);
- }
-
- nir_builder_instr_insert(&b->nb, &call->instr);
-
- if (glsl_type_is_void(callee->return_type)) {
- vtn_push_value(b, w[2], vtn_value_type_undef);
- } else {
- vtn_push_ssa(b, w[2], res_type, vtn_local_load(b, call->return_deref));
- }
-}
-
struct vtn_ssa_value *
vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
{
const struct glsl_type *image_type = sampled.type->type;
const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
const bool is_array = glsl_sampler_type_is_array(image_type);
- const bool is_shadow = glsl_sampler_type_is_shadow(image_type);
/* Figure out the base texture operation */
nir_texop texop;
vtn_fail("Unhandled opcode");
}
- nir_tex_src srcs[8]; /* 8 should be enough */
+ nir_tex_src srcs[10]; /* 10 should be enough */
nir_tex_src *p = srcs;
+ nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler);
+ nir_deref_instr *texture =
+ sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler;
+
+ p->src = nir_src_for_ssa(&texture->dest.ssa);
+ p->src_type = nir_tex_src_texture_deref;
+ p++;
+
+ switch (texop) {
+ case nir_texop_tex:
+ case nir_texop_txb:
+ case nir_texop_txl:
+ case nir_texop_txd:
+ case nir_texop_tg4:
+ /* These operations require a sampler */
+ p->src = nir_src_for_ssa(&sampler->dest.ssa);
+ p->src_type = nir_tex_src_sampler_deref;
+ p++;
+ break;
+ 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:
+ /* These don't */
+ break;
+ case nir_texop_txf_ms_mcs:
+ vtn_fail("unexpected nir_texop_txf_ms_mcs");
+ }
+
unsigned idx = 4;
struct nir_ssa_def *coord;
break;
}
+ bool is_shadow = false;
unsigned gather_component = 0;
switch (opcode) {
case SpvOpImageSampleDrefImplicitLod:
case SpvOpImageSampleProjDrefExplicitLod:
case SpvOpImageDrefGather:
/* These all have an explicit depth value as their next source */
+ is_shadow = true;
(*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparator);
break;
(*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++) = (nir_tex_src){};
+ (*p++) = none;
}
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_fail("Invalid base type for sampler result");
}
- nir_deref_var *sampler = vtn_pointer_to_deref(b, sampled.sampler);
- nir_deref_var *texture;
- if (sampled.image) {
- nir_deref_var *image = vtn_pointer_to_deref(b, sampled.image);
- texture = image;
- } else {
- texture = sampler;
- }
-
- instr->texture = nir_deref_var_clone(texture, instr);
-
- switch (instr->op) {
- case nir_texop_tex:
- case nir_texop_txb:
- case nir_texop_txl:
- case nir_texop_txd:
- case nir_texop_tg4:
- /* These operations require a sampler */
- instr->sampler = nir_deref_var_clone(sampler, instr);
- break;
- 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:
- /* These don't */
- instr->sampler = NULL;
- break;
- case nir_texop_txf_ms_mcs:
- vtn_fail("unexpected nir_texop_txf_ms_mcs");
- }
-
nir_ssa_dest_init(&instr->instr, &instr->dest,
nir_tex_instr_dest_size(instr), 32, NULL);
instrs[i]->is_new_style_shadow = instr->is_new_style_shadow;
instrs[i]->component = instr->component;
instrs[i]->dest_type = instr->dest_type;
- instrs[i]->texture = nir_deref_var_clone(texture, instrs[i]);
- instrs[i]->sampler = NULL;
memcpy(instrs[i]->src, srcs, instr->num_srcs * sizeof(*instr->src));
return nir_swizzle(&b->nb, coord->def, swizzle, 4, false);
}
+static nir_ssa_def *
+expand_to_vec4(nir_builder *b, nir_ssa_def *value)
+{
+ if (value->num_components == 4)
+ return value;
+
+ unsigned swiz[4];
+ for (unsigned i = 0; i < 4; i++)
+ swiz[i] = i < value->num_components ? i : 0;
+ return nir_swizzle(b, value, swiz, 4, false);
+}
+
static void
vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
nir_intrinsic_op op;
switch (opcode) {
-#define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_var_##N; break;
+#define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
OP(ImageQuerySize, size)
OP(ImageRead, load)
OP(ImageWrite, store)
nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op);
- nir_deref_var *image_deref = vtn_pointer_to_deref(b, image.image);
- intrin->variables[0] = nir_deref_var_clone(image_deref, intrin);
+ nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image.image);
+ intrin->src[0] = nir_src_for_ssa(&image_deref->dest.ssa);
/* ImageQuerySize doesn't take any extra parameters */
if (opcode != SpvOpImageQuerySize) {
/* The image coordinate is always 4 components but we may not have that
* many. Swizzle to compensate.
*/
- unsigned swiz[4];
- for (unsigned i = 0; i < 4; i++)
- swiz[i] = i < image.coord->num_components ? i : 0;
- intrin->src[0] = nir_src_for_ssa(nir_swizzle(&b->nb, image.coord,
- swiz, 4, false));
- intrin->src[1] = nir_src_for_ssa(image.sample);
+ intrin->src[1] = nir_src_for_ssa(expand_to_vec4(&b->nb, image.coord));
+ intrin->src[2] = nir_src_for_ssa(image.sample);
}
switch (opcode) {
case SpvOpImageRead:
break;
case SpvOpAtomicStore:
- intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
- break;
- case SpvOpImageWrite:
- intrin->src[2] = nir_src_for_ssa(vtn_ssa_value(b, w[3])->def);
+ case SpvOpImageWrite: {
+ const uint32_t value_id = opcode == SpvOpAtomicStore ? w[4] : w[3];
+ nir_ssa_def *value = vtn_ssa_value(b, value_id)->def;
+ /* nir_intrinsic_image_deref_store always takes a vec4 value */
+ assert(op == nir_intrinsic_image_deref_store);
+ intrin->num_components = 4;
+ intrin->src[3] = nir_src_for_ssa(expand_to_vec4(&b->nb, value));
break;
+ }
case SpvOpAtomicCompareExchange:
case SpvOpAtomicIIncrement:
case SpvOpAtomicAnd:
case SpvOpAtomicOr:
case SpvOpAtomicXor:
- fill_common_atomic_sources(b, opcode, w, &intrin->src[2]);
+ fill_common_atomic_sources(b, opcode, w, &intrin->src[3]);
break;
default:
vtn_fail("Invalid image opcode");
}
- if (opcode != SpvOpImageWrite) {
+ if (opcode != SpvOpImageWrite && opcode != SpvOpAtomicStore) {
struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type;
- unsigned dest_components =
- nir_intrinsic_infos[intrin->intrinsic].dest_components;
- if (intrin->intrinsic == nir_intrinsic_image_var_size) {
- dest_components = intrin->num_components =
- glsl_get_vector_elements(type->type);
- }
+ unsigned dest_components = glsl_get_vector_elements(type->type);
+ intrin->num_components = nir_intrinsic_infos[op].dest_components;
+ if (intrin->num_components == 0)
+ intrin->num_components = dest_components;
nir_ssa_dest_init(&intrin->instr, &intrin->dest,
- dest_components, 32, NULL);
+ intrin->num_components, 32, NULL);
nir_builder_instr_insert(&b->nb, &intrin->instr);
+ nir_ssa_def *result = &intrin->dest.ssa;
+ if (intrin->num_components != dest_components)
+ result = nir_channels(&b->nb, result, (1 << dest_components) - 1);
+
val->ssa = vtn_create_ssa_value(b, type->type);
- val->ssa->def = &intrin->dest.ssa;
+ val->ssa->def = result;
} else {
nir_builder_instr_insert(&b->nb, &intrin->instr);
}
}
}
+static nir_intrinsic_op
+get_uniform_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
+{
+ switch (opcode) {
+#define OP(S, N) case SpvOp##S: return nir_intrinsic_atomic_counter_ ##N;
+ OP(AtomicLoad, read_deref)
+ OP(AtomicExchange, exchange)
+ OP(AtomicCompareExchange, comp_swap)
+ OP(AtomicIIncrement, inc_deref)
+ OP(AtomicIDecrement, post_dec_deref)
+ OP(AtomicIAdd, add_deref)
+ OP(AtomicISub, add_deref)
+ OP(AtomicUMin, min_deref)
+ OP(AtomicUMax, max_deref)
+ OP(AtomicAnd, and_deref)
+ OP(AtomicOr, or_deref)
+ OP(AtomicXor, xor_deref)
+#undef OP
+ default:
+ /* We left the following out: AtomicStore, AtomicSMin and
+ * AtomicSmax. Right now there are not nir intrinsics for them. At this
+ * moment Atomic Counter support is needed for ARB_spirv support, so is
+ * only need to support GLSL Atomic Counters that are uints and don't
+ * allow direct storage.
+ */
+ unreachable("Invalid uniform atomic");
+ }
+}
+
static nir_intrinsic_op
get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
{
}
static nir_intrinsic_op
-get_var_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
+get_deref_nir_atomic_op(struct vtn_builder *b, SpvOp opcode)
{
switch (opcode) {
- case SpvOpAtomicLoad: return nir_intrinsic_load_var;
- case SpvOpAtomicStore: return nir_intrinsic_store_var;
-#define OP(S, N) case SpvOp##S: return nir_intrinsic_var_##N;
+ case SpvOpAtomicLoad: return nir_intrinsic_load_deref;
+ case SpvOpAtomicStore: return nir_intrinsic_store_deref;
+#define OP(S, N) case SpvOp##S: return nir_intrinsic_deref_##N;
OP(AtomicExchange, atomic_exchange)
OP(AtomicCompareExchange, atomic_comp_swap)
OP(AtomicIIncrement, atomic_add)
}
}
+/*
+ * Handles shared atomics, ssbo atomics and atomic counters.
+ */
static void
-vtn_handle_ssbo_or_shared_atomic(struct vtn_builder *b, SpvOp opcode,
- const uint32_t *w, unsigned count)
+vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
+ const uint32_t *w, unsigned count)
{
struct vtn_pointer *ptr;
nir_intrinsic_instr *atomic;
SpvMemorySemanticsMask semantics = w[5];
*/
- if (ptr->mode == vtn_variable_mode_workgroup &&
- !b->options->lower_workgroup_access_to_offsets) {
- nir_deref_var *deref = vtn_pointer_to_deref(b, ptr);
- const struct glsl_type *deref_type = nir_deref_tail(&deref->deref)->type;
- nir_intrinsic_op op = get_var_nir_atomic_op(b, opcode);
+ /* uniform as "atomic counter uniform" */
+ if (ptr->mode == vtn_variable_mode_uniform) {
+ nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
+ const struct glsl_type *deref_type = deref->type;
+ nir_intrinsic_op op = get_uniform_nir_atomic_op(b, opcode);
atomic = nir_intrinsic_instr_create(b->nb.shader, op);
- atomic->variables[0] = nir_deref_var_clone(deref, atomic);
+ atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
+
+ /* SSBO needs to initialize index/offset. In this case we don't need to,
+ * as that info is already stored on the ptr->var->var nir_variable (see
+ * vtn_create_variable)
+ */
switch (opcode) {
case SpvOpAtomicLoad:
case SpvOpAtomicStore:
atomic->num_components = glsl_get_vector_elements(deref_type);
nir_intrinsic_set_write_mask(atomic, (1 << atomic->num_components) - 1);
- atomic->src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[4])->def);
break;
case SpvOpAtomicExchange:
case SpvOpAtomicAnd:
case SpvOpAtomicOr:
case SpvOpAtomicXor:
- fill_common_atomic_sources(b, opcode, w, &atomic->src[0]);
+ /* Nothing: we don't need to call fill_common_atomic_sources here, as
+ * atomic counter uniforms doesn't have sources
+ */
break;
default:
- vtn_fail("Invalid SPIR-V atomic");
+ unreachable("Invalid SPIR-V atomic");
}
- } else {
+ } else if (vtn_pointer_uses_ssa_offset(b, ptr)) {
nir_ssa_def *offset, *index;
- offset = vtn_pointer_to_offset(b, ptr, &index, NULL);
+ offset = vtn_pointer_to_offset(b, ptr, &index);
nir_intrinsic_op op;
if (ptr->mode == vtn_variable_mode_ssbo) {
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);
case SpvOpAtomicStore:
atomic->num_components = glsl_get_vector_elements(ptr->type->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);
fill_common_atomic_sources(b, opcode, w, &atomic->src[src]);
break;
+ default:
+ vtn_fail("Invalid SPIR-V atomic");
+ }
+ } else {
+ 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);
+
+ switch (opcode) {
+ case SpvOpAtomicLoad:
+ atomic->num_components = glsl_get_vector_elements(deref_type);
+ break;
+
+ case SpvOpAtomicStore:
+ atomic->num_components = glsl_get_vector_elements(deref_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);
+ break;
+
+ case SpvOpAtomicExchange:
+ case SpvOpAtomicCompareExchange:
+ case SpvOpAtomicCompareExchangeWeak:
+ case SpvOpAtomicIIncrement:
+ case SpvOpAtomicIDecrement:
+ case SpvOpAtomicIAdd:
+ case SpvOpAtomicISub:
+ case SpvOpAtomicSMin:
+ case SpvOpAtomicUMin:
+ case SpvOpAtomicSMax:
+ case SpvOpAtomicUMax:
+ case SpvOpAtomicAnd:
+ case SpvOpAtomicOr:
+ case SpvOpAtomicXor:
+ fill_common_atomic_sources(b, opcode, w, &atomic->src[1]);
+ break;
+
default:
vtn_fail("Invalid SPIR-V atomic");
}
{
nir_op op;
switch (num_components) {
- case 1: op = nir_op_fmov; break;
+ case 1: op = nir_op_imov; break;
case 2: op = nir_op_vec2; break;
case 3: op = nir_op_vec3; break;
case 4: op = nir_op_vec4; break;
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);
+ return nir_channel(&b->nb, src, index);
}
nir_ssa_def *
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)),
+ dest = nir_bcsel(&b->nb, nir_ieq_imm(&b->nb, index, i),
vtn_vector_extract(b, src, i), dest);
return dest;
{
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;
unsigned elems = count - 3;
assume(elems >= 1);
if (glsl_type_is_vector_or_scalar(type)) {
- nir_ssa_def *srcs[4];
+ nir_ssa_def *srcs[NIR_MAX_VEC_COMPONENTS];
for (unsigned i = 0; i < elems; i++)
srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
val->ssa->def =
switch (opcode) {
case SpvOpEmitStreamVertex:
- case SpvOpEndStreamPrimitive:
- nir_intrinsic_set_stream_id(intrin, w[1]);
+ case SpvOpEndStreamPrimitive: {
+ unsigned stream = vtn_constant_value(b, w[1])->values[0].u32[0];
+ nir_intrinsic_set_stream_id(intrin, stream);
break;
+ }
+
default:
break;
}
return MESA_SHADER_FRAGMENT;
case SpvExecutionModelGLCompute:
return MESA_SHADER_COMPUTE;
+ case SpvExecutionModelKernel:
+ return MESA_SHADER_KERNEL;
default:
vtn_fail("Unsupported execution model");
}
spirv_capability_to_string(cap)); \
} while(0)
+
+void
+vtn_handle_entry_point(struct vtn_builder *b, const uint32_t *w,
+ unsigned count)
+{
+ struct vtn_value *entry_point = &b->values[w[2]];
+ /* Let this be a name label regardless */
+ unsigned name_words;
+ entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
+
+ if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
+ stage_for_execution_model(b, w[1]) != b->entry_point_stage)
+ return;
+
+ vtn_assert(b->entry_point == NULL);
+ b->entry_point = entry_point;
+}
+
static bool
vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
case SpvCapabilityStorageImageExtendedFormats:
break;
- case SpvCapabilityGeometryStreams:
case SpvCapabilityLinkage:
case SpvCapabilityVector16:
case SpvCapabilityFloat16Buffer:
case SpvCapabilityFloat16:
- case SpvCapabilityInt64Atomics:
- case SpvCapabilityAtomicStorage:
- case SpvCapabilityInt16:
- case SpvCapabilityStorageImageMultisample:
case SpvCapabilityInt8:
case SpvCapabilitySparseResidency:
- case SpvCapabilityMinLod:
- case SpvCapabilityTransformFeedback:
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;
+
case SpvCapabilityFloat64:
spv_check_supported(float64, cap);
break;
case SpvCapabilityInt64:
spv_check_supported(int64, cap);
break;
+ case SpvCapabilityInt16:
+ spv_check_supported(int16, cap);
+ break;
+
+ case SpvCapabilityTransformFeedback:
+ spv_check_supported(transform_feedback, cap);
+ break;
+
+ case SpvCapabilityGeometryStreams:
+ spv_check_supported(geometry_streams, cap);
+ break;
+
+ case SpvCapabilityInt64Atomics:
+ spv_check_supported(int64_atomics, cap);
+ break;
+
+ case SpvCapabilityStorageImageMultisample:
+ spv_check_supported(storage_image_ms, cap);
+ break;
case SpvCapabilityAddresses:
case SpvCapabilityKernel:
case SpvCapabilityGroupNonUniformQuad:
spv_check_supported(subgroup_quad, cap);
+ break;
case SpvCapabilityGroupNonUniformArithmetic:
case SpvCapabilityGroupNonUniformClustered:
spv_check_supported(subgroup_arithmetic, cap);
+ break;
case SpvCapabilityVariablePointersStorageBuffer:
case SpvCapabilityVariablePointers:
spv_check_supported(variable_pointers, cap);
+ b->variable_pointers = true;
break;
case SpvCapabilityStorageUniformBufferBlock16:
spv_check_supported(shader_viewport_index_layer, cap);
break;
+ case SpvCapabilityStorageBuffer8BitAccess:
+ case SpvCapabilityUniformAndStorageBuffer8BitAccess:
+ case SpvCapabilityStoragePushConstant8:
+ spv_check_supported(storage_8bit, cap);
+ break;
+
+ case SpvCapabilityInputAttachmentArrayDynamicIndexingEXT:
+ case SpvCapabilityUniformTexelBufferArrayDynamicIndexingEXT:
+ case SpvCapabilityStorageTexelBufferArrayDynamicIndexingEXT:
+ spv_check_supported(descriptor_array_dynamic_indexing, cap);
+ break;
+
+ case SpvCapabilityRuntimeDescriptorArrayEXT:
+ spv_check_supported(runtime_descriptor_array, cap);
+ break;
+
+ case SpvCapabilityStencilExportEXT:
+ spv_check_supported(stencil_export, cap);
+ break;
+
+ case SpvCapabilitySampleMaskPostDepthCoverage:
+ spv_check_supported(post_depth_coverage, cap);
+ break;
+
default:
vtn_fail("Unhandled capability");
}
w[2] == SpvMemoryModelGLSL450);
break;
- case SpvOpEntryPoint: {
- struct vtn_value *entry_point = &b->values[w[2]];
- /* Let this be a name label regardless */
- unsigned name_words;
- entry_point->name = vtn_string_literal(b, &w[3], count - 3, &name_words);
-
- if (strcmp(entry_point->name, b->entry_point_name) != 0 ||
- stage_for_execution_model(b, w[1]) != b->entry_point_stage)
- break;
-
- vtn_assert(b->entry_point == NULL);
- b->entry_point = entry_point;
+ case SpvOpEntryPoint:
+ vtn_handle_entry_point(b, w, count);
break;
- }
case SpvOpString:
vtn_push_value(b, w[1], vtn_value_type_string)->str =
case SpvOpMemberDecorate:
case SpvOpGroupDecorate:
case SpvOpGroupMemberDecorate:
+ case SpvOpDecorateStringGOOGLE:
+ case SpvOpMemberDecorateStringGOOGLE:
vtn_handle_decoration(b, opcode, w, count);
break;
b->shader->info.fs.early_fragment_tests = true;
break;
+ case SpvExecutionModePostDepthCoverage:
+ vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+ b->shader->info.fs.post_depth_coverage = true;
+ break;
+
case SpvExecutionModeInvocations:
vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
b->shader->info.gs.invocations = MAX2(1, mode->literals[0]);
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];
vtn_assert(b->shader->info.stage == MESA_SHADER_GEOMETRY);
b->shader->info.gs.vertices_in =
vertices_in_from_spv_execution_mode(b, mode->exec_mode);
+ b->shader->info.gs.input_primitive =
+ gl_primitive_from_spv_execution_mode(b, mode->exec_mode);
}
break;
break;
case SpvExecutionModeXfb:
- vtn_fail("Unhandled execution mode");
+ b->shader->info.has_transform_feedback_varyings = true;
break;
case SpvExecutionModeVecTypeHint:
case SpvExecutionModeContractionOff:
break; /* OpenCL */
+ case SpvExecutionModeStencilRefReplacingEXT:
+ vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT);
+ break;
+
default:
vtn_fail("Unhandled execution mode");
}
case SpvOpMemberDecorate:
case SpvOpGroupDecorate:
case SpvOpGroupMemberDecorate:
+ case SpvOpDecorateStringGOOGLE:
+ case SpvOpMemberDecorateStringGOOGLE:
vtn_fail("Invalid opcode types and variables section");
break;
case SpvOpImageQuerySize: {
struct vtn_pointer *image =
vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
- if (image->mode == vtn_variable_mode_image) {
+ if (glsl_type_is_image(image->type->type)) {
vtn_handle_image(b, opcode, w, count);
} else {
- vtn_assert(image->mode == vtn_variable_mode_sampler);
+ vtn_assert(glsl_type_is_sampler(image->type->type));
vtn_handle_texture(b, opcode, w, count);
}
break;
vtn_handle_image(b, opcode, w, count);
} else {
vtn_assert(pointer->value_type == vtn_value_type_pointer);
- vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
+ vtn_handle_atomics(b, opcode, w, count);
}
break;
}
vtn_handle_image(b, opcode, w, count);
} else {
vtn_assert(pointer->value_type == vtn_value_type_pointer);
- vtn_handle_ssbo_or_shared_atomic(b, opcode, w, count);
+ vtn_handle_atomics(b, opcode, w, count);
}
break;
}
b->entry_point_name = entry_point_name;
b->options = options;
- /* Handle the SPIR-V header (first 4 dwords) */
- vtn_assert(word_count > 5);
+ /*
+ * Handle the SPIR-V header (first 5 dwords).
+ * Can't use vtx_assert() as the setjmp(3) target isn't initialized yet.
+ */
+ if (word_count <= 5)
+ goto fail;
+
+ if (words[0] != SpvMagicNumber) {
+ vtn_err("words[0] was 0x%x, want 0x%x", words[0], SpvMagicNumber);
+ goto fail;
+ }
+ if (words[1] < 0x10000) {
+ vtn_err("words[1] was 0x%x, want >= 0x10000", words[1]);
+ 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);
- vtn_assert(words[0] == SpvMagicNumber);
- vtn_assert(words[1] >= 0x10000);
/* words[2] == generator magic */
unsigned value_id_bound = words[3];
- vtn_assert(words[4] == 0);
+ if (words[4] != 0) {
+ vtn_err("words[4] was %u, want 0", words[4]);
+ goto fail;
+ }
b->value_id_bound = value_id_bound;
b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
return b;
+ fail:
+ ralloc_free(b);
+ return NULL;
}
nir_function *
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);
+ /* 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);
ralloc_free(b);
+ entry_point->is_entrypoint = true;
return entry_point;
}