X-Git-Url: https://git.libre-soc.org/?p=mesa.git;a=blobdiff_plain;f=src%2Fcompiler%2Fspirv%2Fspirv_to_nir.c;h=dabb3c674bf504ddf06b3983a92fcca47e076b75;hp=aff11a2d57933a1f62a8ec9b69a2116f5f07605c;hb=2ba3ffa76c608fc6929b22d16731cc42202cacca;hpb=2402232c90ef362a6cf14ff1cf5518e6c25bf9f9 diff --git a/src/compiler/spirv/spirv_to_nir.c b/src/compiler/spirv/spirv_to_nir.c index aff11a2d579..dabb3c674bf 100644 --- a/src/compiler/spirv/spirv_to_nir.c +++ b/src/compiler/spirv/spirv_to_nir.c @@ -32,6 +32,7 @@ #include "nir/nir_deref.h" #include "spirv_info.h" +#include "util/format/u_format.h" #include "util/u_math.h" #include @@ -162,19 +163,11 @@ _vtn_fail(struct vtn_builder *b, const char *file, unsigned line, longjmp(b->fail_jump, 1); } -struct spec_constant_value { - bool is_double; - union { - uint32_t data32; - uint64_t data64; - }; -}; - static struct vtn_ssa_value * vtn_undef_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; + val->type = glsl_get_bare_type(type); if (glsl_type_is_vector_or_scalar(type)) { unsigned num_components = glsl_get_vector_elements(val->type); @@ -183,18 +176,12 @@ vtn_undef_ssa_value(struct vtn_builder *b, const struct glsl_type *type) } else { unsigned elems = glsl_get_length(val->type); val->elems = ralloc_array(b, struct vtn_ssa_value *, elems); - if (glsl_type_is_matrix(type)) { - const struct glsl_type *elem_type = - glsl_vector_type(glsl_get_base_type(type), - glsl_get_vector_elements(type)); - - for (unsigned i = 0; i < elems; i++) - val->elems[i] = vtn_undef_ssa_value(b, elem_type); - } else if (glsl_type_is_array(type)) { + if (glsl_type_is_array_or_matrix(type)) { const struct glsl_type *elem_type = glsl_get_array_element(type); for (unsigned i = 0; i < elems; i++) val->elems[i] = vtn_undef_ssa_value(b, elem_type); } else { + vtn_assert(glsl_type_is_struct_or_ifc(type)); for (unsigned i = 0; i < elems; i++) { const struct glsl_type *elem_type = glsl_get_struct_field(type, i); val->elems[i] = vtn_undef_ssa_value(b, elem_type); @@ -215,68 +202,36 @@ vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant, 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_INT16: - case GLSL_TYPE_UINT16: - case GLSL_TYPE_UINT8: - case GLSL_TYPE_INT8: - case GLSL_TYPE_INT64: - case GLSL_TYPE_UINT64: - case GLSL_TYPE_BOOL: - case GLSL_TYPE_FLOAT: - case GLSL_TYPE_FLOAT16: - case GLSL_TYPE_DOUBLE: { - int bit_size = glsl_get_bit_size(type); - 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, bit_size); - - memcpy(load->value, constant->values, - sizeof(nir_const_value) * load->def.num_components); - - nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr); - val->def = &load->def; - } else { - assert(glsl_type_is_matrix(type)); - unsigned columns = glsl_get_matrix_columns(val->type); - val->elems = ralloc_array(b, struct vtn_ssa_value *, columns); - const struct glsl_type *column_type = glsl_get_column_type(val->type); - for (unsigned i = 0; i < columns; i++) - val->elems[i] = vtn_const_ssa_value(b, constant->elements[i], - column_type); - } - break; - } + val->type = glsl_get_bare_type(type); - 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; - } + if (glsl_type_is_vector_or_scalar(type)) { + unsigned num_components = glsl_get_vector_elements(val->type); + unsigned bit_size = glsl_get_bit_size(type); + nir_load_const_instr *load = + nir_load_const_instr_create(b->shader, num_components, bit_size); - case GLSL_TYPE_STRUCT: { + memcpy(load->value, constant->values, + sizeof(nir_const_value) * num_components); + + nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr); + val->def = &load->def; + } else { 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); + if (glsl_type_is_array_or_matrix(type)) { + const struct glsl_type *elem_type = glsl_get_array_element(type); + for (unsigned i = 0; i < elems; i++) { + val->elems[i] = vtn_const_ssa_value(b, constant->elements[i], + elem_type); + } + } else { + vtn_assert(glsl_type_is_struct_or_ifc(type)); + for (unsigned i = 0; i < elems; i++) { + const struct glsl_type *elem_type = glsl_get_struct_field(type, i); + val->elems[i] = vtn_const_ssa_value(b, constant->elements[i], + elem_type); + } } - break; - } - - default: - vtn_fail("bad constant type"); } return val; @@ -308,6 +263,113 @@ vtn_ssa_value(struct vtn_builder *b, uint32_t value_id) } } +struct vtn_value * +vtn_push_ssa_value(struct vtn_builder *b, uint32_t value_id, + struct vtn_ssa_value *ssa) +{ + struct vtn_type *type = vtn_get_value_type(b, value_id); + + /* See vtn_create_ssa_value */ + vtn_fail_if(ssa->type != glsl_get_bare_type(type->type), + "Type mismatch for SPIR-V SSA value"); + + struct vtn_value *val; + if (type->base_type == vtn_base_type_pointer) { + val = vtn_push_pointer(b, value_id, vtn_pointer_from_ssa(b, ssa->def, type)); + } else { + /* Don't trip the value_type_ssa check in vtn_push_value */ + val = vtn_push_value(b, value_id, vtn_value_type_invalid); + val->value_type = vtn_value_type_ssa; + val->ssa = ssa; + } + + return val; +} + +nir_ssa_def * +vtn_get_nir_ssa(struct vtn_builder *b, uint32_t value_id) +{ + struct vtn_ssa_value *ssa = vtn_ssa_value(b, value_id); + vtn_fail_if(!glsl_type_is_vector_or_scalar(ssa->type), + "Expected a vector or scalar type"); + return ssa->def; +} + +struct vtn_value * +vtn_push_nir_ssa(struct vtn_builder *b, uint32_t value_id, nir_ssa_def *def) +{ + /* Types for all SPIR-V SSA values are set as part of a pre-pass so the + * type will be valid by the time we get here. + */ + struct vtn_type *type = vtn_get_value_type(b, value_id); + vtn_fail_if(def->num_components != glsl_get_vector_elements(type->type) || + def->bit_size != glsl_get_bit_size(type->type), + "Mismatch between NIR and SPIR-V type."); + struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, type->type); + ssa->def = def; + return vtn_push_ssa_value(b, value_id, ssa); +} + +static nir_deref_instr * +vtn_get_image(struct vtn_builder *b, uint32_t value_id) +{ + struct vtn_type *type = vtn_get_value_type(b, value_id); + vtn_assert(type->base_type == vtn_base_type_image); + return nir_build_deref_cast(&b->nb, vtn_get_nir_ssa(b, value_id), + nir_var_uniform, type->glsl_image, 0); +} + +static void +vtn_push_image(struct vtn_builder *b, uint32_t value_id, + nir_deref_instr *deref) +{ + struct vtn_type *type = vtn_get_value_type(b, value_id); + vtn_assert(type->base_type == vtn_base_type_image); + vtn_push_nir_ssa(b, value_id, &deref->dest.ssa); +} + +static nir_deref_instr * +vtn_get_sampler(struct vtn_builder *b, uint32_t value_id) +{ + struct vtn_type *type = vtn_get_value_type(b, value_id); + vtn_assert(type->base_type == vtn_base_type_sampler); + return nir_build_deref_cast(&b->nb, vtn_get_nir_ssa(b, value_id), + nir_var_uniform, glsl_bare_sampler_type(), 0); +} + +nir_ssa_def * +vtn_sampled_image_to_nir_ssa(struct vtn_builder *b, + struct vtn_sampled_image si) +{ + return nir_vec2(&b->nb, &si.image->dest.ssa, &si.sampler->dest.ssa); +} + +static void +vtn_push_sampled_image(struct vtn_builder *b, uint32_t value_id, + struct vtn_sampled_image si) +{ + struct vtn_type *type = vtn_get_value_type(b, value_id); + vtn_assert(type->base_type == vtn_base_type_sampled_image); + vtn_push_nir_ssa(b, value_id, vtn_sampled_image_to_nir_ssa(b, si)); +} + +static struct vtn_sampled_image +vtn_get_sampled_image(struct vtn_builder *b, uint32_t value_id) +{ + struct vtn_type *type = vtn_get_value_type(b, value_id); + vtn_assert(type->base_type == vtn_base_type_sampled_image); + nir_ssa_def *si_vec2 = vtn_get_nir_ssa(b, value_id); + + struct vtn_sampled_image si = { NULL, }; + si.image = nir_build_deref_cast(&b->nb, nir_channel(&b->nb, si_vec2, 0), + nir_var_uniform, + type->image->glsl_image, 0); + si.sampler = nir_build_deref_cast(&b->nb, nir_channel(&b->nb, si_vec2, 1), + nir_var_uniform, + glsl_bare_sampler_type(), 0); + return si; +} + static char * vtn_string_literal(struct vtn_builder *b, const uint32_t *words, unsigned word_count, unsigned *words_used) @@ -371,6 +433,14 @@ vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start, return w; } +static bool +vtn_handle_non_semantic_instruction(struct vtn_builder *b, SpvOp ext_opcode, + const uint32_t *w, unsigned count) +{ + /* Do nothing. */ + return true; +} + static void vtn_handle_extension(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) @@ -390,8 +460,13 @@ vtn_handle_extension(struct vtn_builder *b, SpvOp opcode, } else if ((strcmp(ext, "SPV_AMD_shader_trinary_minmax") == 0) && (b->options && b->options->caps.amd_trinary_minmax)) { val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction; + } else if ((strcmp(ext, "SPV_AMD_shader_explicit_vertex_parameter") == 0) + && (b->options && b->options->caps.amd_shader_explicit_vertex_parameter)) { + val->ext_handler = vtn_handle_amd_shader_explicit_vertex_parameter_instruction; } else if (strcmp(ext, "OpenCL.std") == 0) { val->ext_handler = vtn_handle_opencl_instruction; + } else if (strstr(ext, "NonSemantic.") == ext) { + val->ext_handler = vtn_handle_non_semantic_instruction; } else { vtn_fail("Unsupported extension: %s", ext); } @@ -687,6 +762,120 @@ vtn_type_copy(struct vtn_builder *b, struct vtn_type *src) return dest; } +static const struct glsl_type * +wrap_type_in_array(const struct glsl_type *type, + const struct glsl_type *array_type) +{ + if (!glsl_type_is_array(array_type)) + return type; + + const struct glsl_type *elem_type = + wrap_type_in_array(type, glsl_get_array_element(array_type)); + return glsl_array_type(elem_type, glsl_get_length(array_type), + glsl_get_explicit_stride(array_type)); +} + +static bool +vtn_type_needs_explicit_layout(struct vtn_builder *b, enum vtn_variable_mode mode) +{ + /* For OpenCL we never want to strip the info from the types, and it makes + * type comparisons easier in later stages. + */ + if (b->options->environment == NIR_SPIRV_OPENCL) + return true; + + switch (mode) { + case vtn_variable_mode_input: + case vtn_variable_mode_output: + /* Layout decorations kept because we need offsets for XFB arrays of + * blocks. + */ + return b->shader->info.has_transform_feedback_varyings; + + case vtn_variable_mode_ssbo: + case vtn_variable_mode_phys_ssbo: + case vtn_variable_mode_ubo: + return true; + + default: + return false; + } +} + +const struct glsl_type * +vtn_type_get_nir_type(struct vtn_builder *b, struct vtn_type *type, + enum vtn_variable_mode mode) +{ + if (mode == vtn_variable_mode_atomic_counter) { + vtn_fail_if(glsl_without_array(type->type) != glsl_uint_type(), + "Variables in the AtomicCounter storage class should be " + "(possibly arrays of arrays of) uint."); + return wrap_type_in_array(glsl_atomic_uint_type(), type->type); + } + + if (mode == vtn_variable_mode_uniform) { + switch (type->base_type) { + case vtn_base_type_array: { + const struct glsl_type *elem_type = + vtn_type_get_nir_type(b, type->array_element, mode); + + return glsl_array_type(elem_type, type->length, + glsl_get_explicit_stride(type->type)); + } + + case vtn_base_type_struct: { + bool need_new_struct = false; + const uint32_t num_fields = type->length; + NIR_VLA(struct glsl_struct_field, fields, num_fields); + for (unsigned i = 0; i < num_fields; i++) { + fields[i] = *glsl_get_struct_field_data(type->type, i); + const struct glsl_type *field_nir_type = + vtn_type_get_nir_type(b, type->members[i], mode); + if (fields[i].type != field_nir_type) { + fields[i].type = field_nir_type; + need_new_struct = true; + } + } + if (need_new_struct) { + if (glsl_type_is_interface(type->type)) { + return glsl_interface_type(fields, num_fields, + /* packing */ 0, false, + glsl_get_type_name(type->type)); + } else { + return glsl_struct_type(fields, num_fields, + glsl_get_type_name(type->type), + glsl_struct_type_is_packed(type->type)); + } + } else { + /* No changes, just pass it on */ + return type->type; + } + } + + case vtn_base_type_image: + return type->glsl_image; + + case vtn_base_type_sampler: + return glsl_bare_sampler_type(); + + case vtn_base_type_sampled_image: + return type->image->glsl_image; + + default: + return type->type; + } + } + + /* Layout decorations are allowed but ignored in certain conditions, + * to allow SPIR-V generators perform type deduplication. Discard + * unnecessary ones when passing to NIR. + */ + if (!vtn_type_needs_explicit_layout(b, mode)) + return glsl_get_bare_type(type->type); + + return type->type; +} + static struct vtn_type * mutable_matrix_member(struct vtn_builder *b, struct vtn_type *type, int member) { @@ -736,7 +925,7 @@ array_stride_decoration_cb(struct vtn_builder *b, static void struct_member_decoration_cb(struct vtn_builder *b, - struct vtn_value *val, int member, + UNUSED struct vtn_value *val, int member, const struct vtn_decoration *dec, void *void_ctx) { struct member_decoration_ctx *ctx = void_ctx; @@ -769,6 +958,9 @@ struct_member_decoration_cb(struct vtn_builder *b, case SpvDecorationFlat: ctx->fields[member].interpolation = INTERP_MODE_FLAT; break; + case SpvDecorationExplicitInterpAMD: + ctx->fields[member].interpolation = INTERP_MODE_EXPLICIT; + break; case SpvDecorationCentroid: ctx->fields[member].centroid = true; break; @@ -776,8 +968,7 @@ struct_member_decoration_cb(struct vtn_builder *b, ctx->fields[member].sample = true; break; case SpvDecorationStream: - /* Vulkan only allows one GS stream */ - vtn_assert(dec->operands[0] == 0); + /* This is handled later by var_decoration_cb in vtn_variables.c */ break; case SpvDecorationLocation: ctx->fields[member].location = dec->operands[0]; @@ -828,7 +1019,7 @@ struct_member_decoration_cb(struct vtn_builder *b, case SpvDecorationXfbBuffer: case SpvDecorationXfbStride: - vtn_warn("Vulkan does not have transform feedback"); + /* This is handled later by var_decoration_cb in vtn_variables.c */ break; case SpvDecorationCPacked: @@ -851,6 +1042,7 @@ struct_member_decoration_cb(struct vtn_builder *b, break; case SpvDecorationUserSemantic: + case SpvDecorationUserTypeGOOGLE: /* User semantic decorations can safely be ignored by the driver. */ break; @@ -879,7 +1071,7 @@ vtn_array_type_rewrite_glsl_type(struct vtn_type *type) */ static void struct_member_matrix_stride_cb(struct vtn_builder *b, - struct vtn_value *val, int member, + UNUSED struct vtn_value *val, int member, const struct vtn_decoration *dec, void *void_ctx) { @@ -936,7 +1128,7 @@ struct_block_decoration_cb(struct vtn_builder *b, static void type_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member, - const struct vtn_decoration *dec, void *ctx) + const struct vtn_decoration *dec, UNUSED void *ctx) { struct vtn_type *type = val->type; @@ -974,6 +1166,7 @@ type_decoration_cb(struct vtn_builder *b, case SpvDecorationPatch: case SpvDecorationCentroid: case SpvDecorationSample: + case SpvDecorationExplicitInterpAMD: case SpvDecorationVolatile: case SpvDecorationCoherent: case SpvDecorationNonWritable: @@ -1031,6 +1224,10 @@ type_decoration_cb(struct vtn_builder *b, spirv_decoration_to_string(dec->decoration)); break; + case SpvDecorationUserTypeGOOGLE: + /* User semantic decorations can safely be ignored by the driver. */ + break; + default: vtn_fail_with_decoration("Unhandled decoration", dec->decoration); } @@ -1040,46 +1237,46 @@ static unsigned translate_image_format(struct vtn_builder *b, SpvImageFormat format) { switch (format) { - case SpvImageFormatUnknown: return 0; /* GL_NONE */ - case SpvImageFormatRgba32f: return 0x8814; /* GL_RGBA32F */ - case SpvImageFormatRgba16f: return 0x881A; /* GL_RGBA16F */ - case SpvImageFormatR32f: return 0x822E; /* GL_R32F */ - case SpvImageFormatRgba8: return 0x8058; /* GL_RGBA8 */ - case SpvImageFormatRgba8Snorm: return 0x8F97; /* GL_RGBA8_SNORM */ - case SpvImageFormatRg32f: return 0x8230; /* GL_RG32F */ - case SpvImageFormatRg16f: return 0x822F; /* GL_RG16F */ - case SpvImageFormatR11fG11fB10f: return 0x8C3A; /* GL_R11F_G11F_B10F */ - case SpvImageFormatR16f: return 0x822D; /* GL_R16F */ - case SpvImageFormatRgba16: return 0x805B; /* GL_RGBA16 */ - case SpvImageFormatRgb10A2: return 0x8059; /* GL_RGB10_A2 */ - case SpvImageFormatRg16: return 0x822C; /* GL_RG16 */ - case SpvImageFormatRg8: return 0x822B; /* GL_RG8 */ - case SpvImageFormatR16: return 0x822A; /* GL_R16 */ - case SpvImageFormatR8: return 0x8229; /* GL_R8 */ - case SpvImageFormatRgba16Snorm: return 0x8F9B; /* GL_RGBA16_SNORM */ - case SpvImageFormatRg16Snorm: return 0x8F99; /* GL_RG16_SNORM */ - case SpvImageFormatRg8Snorm: return 0x8F95; /* GL_RG8_SNORM */ - case SpvImageFormatR16Snorm: return 0x8F98; /* GL_R16_SNORM */ - case SpvImageFormatR8Snorm: return 0x8F94; /* GL_R8_SNORM */ - case SpvImageFormatRgba32i: return 0x8D82; /* GL_RGBA32I */ - case SpvImageFormatRgba16i: return 0x8D88; /* GL_RGBA16I */ - case SpvImageFormatRgba8i: return 0x8D8E; /* GL_RGBA8I */ - case SpvImageFormatR32i: return 0x8235; /* GL_R32I */ - case SpvImageFormatRg32i: return 0x823B; /* GL_RG32I */ - case SpvImageFormatRg16i: return 0x8239; /* GL_RG16I */ - case SpvImageFormatRg8i: return 0x8237; /* GL_RG8I */ - case SpvImageFormatR16i: return 0x8233; /* GL_R16I */ - case SpvImageFormatR8i: return 0x8231; /* GL_R8I */ - case SpvImageFormatRgba32ui: return 0x8D70; /* GL_RGBA32UI */ - case SpvImageFormatRgba16ui: return 0x8D76; /* GL_RGBA16UI */ - case SpvImageFormatRgba8ui: return 0x8D7C; /* GL_RGBA8UI */ - case SpvImageFormatR32ui: return 0x8236; /* GL_R32UI */ - case SpvImageFormatRgb10a2ui: return 0x906F; /* GL_RGB10_A2UI */ - case SpvImageFormatRg32ui: return 0x823C; /* GL_RG32UI */ - case SpvImageFormatRg16ui: return 0x823A; /* GL_RG16UI */ - case SpvImageFormatRg8ui: return 0x8238; /* GL_RG8UI */ - case SpvImageFormatR16ui: return 0x8234; /* GL_R16UI */ - case SpvImageFormatR8ui: return 0x8232; /* GL_R8UI */ + case SpvImageFormatUnknown: return PIPE_FORMAT_NONE; + case SpvImageFormatRgba32f: return PIPE_FORMAT_R32G32B32A32_FLOAT; + case SpvImageFormatRgba16f: return PIPE_FORMAT_R16G16B16A16_FLOAT; + case SpvImageFormatR32f: return PIPE_FORMAT_R32_FLOAT; + case SpvImageFormatRgba8: return PIPE_FORMAT_R8G8B8A8_UNORM; + case SpvImageFormatRgba8Snorm: return PIPE_FORMAT_R8G8B8A8_SNORM; + case SpvImageFormatRg32f: return PIPE_FORMAT_R32G32_FLOAT; + case SpvImageFormatRg16f: return PIPE_FORMAT_R16G16_FLOAT; + case SpvImageFormatR11fG11fB10f: return PIPE_FORMAT_R11G11B10_FLOAT; + case SpvImageFormatR16f: return PIPE_FORMAT_R16_FLOAT; + case SpvImageFormatRgba16: return PIPE_FORMAT_R16G16B16A16_UNORM; + case SpvImageFormatRgb10A2: return PIPE_FORMAT_R10G10B10A2_UNORM; + case SpvImageFormatRg16: return PIPE_FORMAT_R16G16_UNORM; + case SpvImageFormatRg8: return PIPE_FORMAT_R8G8_UNORM; + case SpvImageFormatR16: return PIPE_FORMAT_R16_UNORM; + case SpvImageFormatR8: return PIPE_FORMAT_R8_UNORM; + case SpvImageFormatRgba16Snorm: return PIPE_FORMAT_R16G16B16A16_SNORM; + case SpvImageFormatRg16Snorm: return PIPE_FORMAT_R16G16_SNORM; + case SpvImageFormatRg8Snorm: return PIPE_FORMAT_R8G8_SNORM; + case SpvImageFormatR16Snorm: return PIPE_FORMAT_R16_SNORM; + case SpvImageFormatR8Snorm: return PIPE_FORMAT_R8_SNORM; + case SpvImageFormatRgba32i: return PIPE_FORMAT_R32G32B32A32_SINT; + case SpvImageFormatRgba16i: return PIPE_FORMAT_R16G16B16A16_SINT; + case SpvImageFormatRgba8i: return PIPE_FORMAT_R8G8B8A8_SINT; + case SpvImageFormatR32i: return PIPE_FORMAT_R32_SINT; + case SpvImageFormatRg32i: return PIPE_FORMAT_R32G32_SINT; + case SpvImageFormatRg16i: return PIPE_FORMAT_R16G16_SINT; + case SpvImageFormatRg8i: return PIPE_FORMAT_R8G8_SINT; + case SpvImageFormatR16i: return PIPE_FORMAT_R16_SINT; + case SpvImageFormatR8i: return PIPE_FORMAT_R8_SINT; + case SpvImageFormatRgba32ui: return PIPE_FORMAT_R32G32B32A32_UINT; + case SpvImageFormatRgba16ui: return PIPE_FORMAT_R16G16B16A16_UINT; + case SpvImageFormatRgba8ui: return PIPE_FORMAT_R8G8B8A8_UINT; + case SpvImageFormatR32ui: return PIPE_FORMAT_R32_UINT; + case SpvImageFormatRgb10a2ui: return PIPE_FORMAT_R10G10B10A2_UINT; + case SpvImageFormatRg32ui: return PIPE_FORMAT_R32G32_UINT; + case SpvImageFormatRg16ui: return PIPE_FORMAT_R16G16_UINT; + case SpvImageFormatRg8ui: return PIPE_FORMAT_R8G8_UINT; + case SpvImageFormatR16ui: return PIPE_FORMAT_R16_UINT; + case SpvImageFormatR8ui: return PIPE_FORMAT_R8_UINT; default: vtn_fail("Invalid image format: %s (%u)", spirv_imageformat_to_string(format), format); @@ -1158,7 +1355,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, } case SpvOpTypeVector: { - struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type; + struct vtn_type *base = vtn_get_type(b, w[2]); unsigned elems = w[3]; vtn_fail_if(base->base_type != vtn_base_type_scalar, @@ -1176,7 +1373,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, } case SpvOpTypeMatrix: { - struct vtn_type *base = vtn_value(b, w[2], vtn_value_type_type)->type; + struct vtn_type *base = vtn_get_type(b, w[2]); unsigned columns = w[3]; vtn_fail_if(base->base_type != vtn_base_type_vector, @@ -1200,8 +1397,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, case SpvOpTypeRuntimeArray: case SpvOpTypeArray: { - struct vtn_type *array_element = - vtn_value(b, w[2], vtn_value_type_type)->type; + struct vtn_type *array_element = vtn_get_type(b, w[2]); if (opcode == SpvOpTypeRuntimeArray) { /* A length of 0 is used to denote unsized arrays */ @@ -1231,8 +1427,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, NIR_VLA(struct glsl_struct_field, fields, count); for (unsigned i = 0; i < num_fields; i++) { - val->type->members[i] = - vtn_value(b, w[i + 2], vtn_value_type_type)->type; + val->type->members[i] = vtn_get_type(b, w[i + 2]); fields[i] = (struct glsl_struct_field) { .type = val->type->members[i]->type, .name = ralloc_asprintf(b, "field%d", i), @@ -1281,14 +1476,13 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, val->type->base_type = vtn_base_type_function; val->type->type = NULL; - val->type->return_type = vtn_value(b, w[2], vtn_value_type_type)->type; + val->type->return_type = vtn_get_type(b, w[2]); const unsigned num_params = count - 3; val->type->length = num_params; val->type->params = ralloc_array(b, struct vtn_type *, num_params); for (unsigned i = 0; i < count - 3; i++) { - val->type->params[i] = - vtn_value(b, w[i + 3], vtn_value_type_type)->type; + val->type->params[i] = vtn_get_type(b, w[i + 3]); } break; } @@ -1329,7 +1523,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, "forward declaration of a pointer, OpTypePointer can " "only be used once for a given id."); - val->type->deref = vtn_value(b, w[3], vtn_value_type_type)->type; + val->type->deref = vtn_get_type(b, w[3]); /* Only certain storage classes use ArrayStride. The others (in * particular Workgroup) are expected to be laid out by the driver. @@ -1338,7 +1532,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, case SpvStorageClassUniform: case SpvStorageClassPushConstant: case SpvStorageClassStorageBuffer: - case SpvStorageClassPhysicalStorageBufferEXT: + case SpvStorageClassPhysicalStorageBuffer: vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL); break; default: @@ -1366,12 +1560,23 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, case SpvOpTypeImage: { val->type->base_type = vtn_base_type_image; - const struct vtn_type *sampled_type = - vtn_value(b, w[2], vtn_value_type_type)->type; + /* Images are represented in NIR as a scalar SSA value that is the + * result of a deref instruction. An OpLoad on an OpTypeImage pointer + * from UniformConstant memory just takes the NIR deref from the pointer + * and turns it into an SSA value. + */ + val->type->type = nir_address_format_to_glsl_type( + vtn_mode_to_address_format(b, vtn_variable_mode_function)); - vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar || - glsl_get_bit_size(sampled_type->type) != 32, - "Sampled type of OpTypeImage must be a 32-bit scalar"); + const struct vtn_type *sampled_type = vtn_get_type(b, w[2]); + if (b->shader->info.stage == MESA_SHADER_KERNEL) { + vtn_fail_if(sampled_type->base_type != vtn_base_type_void, + "Sampled type of OpTypeImage must be void for kernels"); + } else { + vtn_fail_if(sampled_type->base_type != vtn_base_type_scalar || + glsl_get_bit_size(sampled_type->type) != 32, + "Sampled type of OpTypeImage must be a 32-bit scalar"); + } enum glsl_sampler_dim dim; switch ((SpvDim)w[3]) { @@ -1397,6 +1602,9 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, if (count > 9) val->type->access_qualifier = w[9]; + else if (b->shader->info.stage == MESA_SHADER_KERNEL) + /* Per the CL C spec: If no qualifier is provided, read_only is assumed. */ + val->type->access_qualifier = SpvAccessQualifierReadOnly; else val->type->access_qualifier = SpvAccessQualifierReadWrite; @@ -1414,32 +1622,53 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, enum glsl_base_type sampled_base_type = glsl_get_base_type(sampled_type->type); if (sampled == 1) { - val->type->sampled = true; - val->type->type = glsl_sampler_type(dim, false, is_array, - sampled_base_type); + val->type->glsl_image = glsl_sampler_type(dim, false, is_array, + sampled_base_type); } else if (sampled == 2) { - val->type->sampled = false; - val->type->type = glsl_image_type(dim, is_array, sampled_base_type); + val->type->glsl_image = glsl_image_type(dim, is_array, + sampled_base_type); + } else if (b->shader->info.stage == MESA_SHADER_KERNEL) { + val->type->glsl_image = glsl_image_type(dim, is_array, + GLSL_TYPE_VOID); } else { vtn_fail("We need to know if the image will be sampled"); } break; } - case SpvOpTypeSampledImage: + case SpvOpTypeSampledImage: { val->type->base_type = vtn_base_type_sampled_image; - val->type->image = vtn_value(b, w[2], vtn_value_type_type)->type; - val->type->type = val->type->image->type; + val->type->image = vtn_get_type(b, w[2]); + + /* Sampled images are represented NIR as a vec2 SSA value where each + * component is the result of a deref instruction. The first component + * is the image and the second is the sampler. An OpLoad on an + * OpTypeSampledImage pointer from UniformConstant memory just takes + * the NIR deref from the pointer and duplicates it to both vector + * components. + */ + nir_address_format addr_format = + vtn_mode_to_address_format(b, vtn_variable_mode_function); + assert(nir_address_format_num_components(addr_format) == 1); + unsigned bit_size = nir_address_format_bit_size(addr_format); + assert(bit_size == 32 || bit_size == 64); + + enum glsl_base_type base_type = + bit_size == 32 ? GLSL_TYPE_UINT : GLSL_TYPE_UINT64; + val->type->type = glsl_vector_type(base_type, 2); break; + } case SpvOpTypeSampler: - /* The actual sampler type here doesn't really matter. It gets - * thrown away the moment you combine it with an image. What really - * matters is that it's a sampler type as opposed to an integer type - * so the backend knows what to do. - */ val->type->base_type = vtn_base_type_sampler; - val->type->type = glsl_bare_sampler_type(); + + /* Samplers are represented in NIR as a scalar SSA value that is the + * result of a deref instruction. An OpLoad on an OpTypeSampler pointer + * from UniformConstant memory just takes the NIR deref from the pointer + * and turns it into an SSA value. + */ + val->type->type = nir_address_format_to_glsl_type( + vtn_mode_to_address_format(b, vtn_variable_mode_function)); break; case SpvOpTypeOpaque: @@ -1522,55 +1751,29 @@ vtn_null_constant(struct vtn_builder *b, struct vtn_type *type) } static void -spec_constant_decoration_cb(struct vtn_builder *b, struct vtn_value *v, - int member, const struct vtn_decoration *dec, - void *data) +spec_constant_decoration_cb(struct vtn_builder *b, UNUSED struct vtn_value *val, + ASSERTED int member, + const struct vtn_decoration *dec, void *data) { vtn_assert(member == -1); if (dec->decoration != SpvDecorationSpecId) return; - struct spec_constant_value *const_value = data; - + nir_const_value *value = data; for (unsigned i = 0; i < b->num_specializations; i++) { if (b->specializations[i].id == dec->operands[0]) { - if (const_value->is_double) - const_value->data64 = b->specializations[i].data64; - else - const_value->data32 = b->specializations[i].data32; + *value = b->specializations[i].value; return; } } } -static uint32_t -get_specialization(struct vtn_builder *b, struct vtn_value *val, - uint32_t const_value) -{ - struct spec_constant_value data; - data.is_double = false; - data.data32 = const_value; - vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data); - return data.data32; -} - -static uint64_t -get_specialization64(struct vtn_builder *b, struct vtn_value *val, - uint64_t const_value) -{ - struct spec_constant_value data; - data.is_double = true; - data.data64 = const_value; - vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &data); - return data.data64; -} - static void handle_workgroup_size_decoration_cb(struct vtn_builder *b, struct vtn_value *val, - int member, + ASSERTED int member, const struct vtn_decoration *dec, - void *data) + UNUSED void *data) { vtn_assert(member == -1); if (dec->decoration != SpvDecorationBuiltIn || @@ -1596,18 +1799,21 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, "Result type of %s must be OpTypeBool", spirv_op_to_string(opcode)); - uint32_t int_val = (opcode == SpvOpConstantTrue || - opcode == SpvOpSpecConstantTrue); + bool bval = (opcode == SpvOpConstantTrue || + opcode == SpvOpSpecConstantTrue); + + nir_const_value u32val = nir_const_value_for_uint(bval, 32); if (opcode == SpvOpSpecConstantTrue || opcode == SpvOpSpecConstantFalse) - int_val = get_specialization(b, val, int_val); + vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &u32val); - val->constant->values[0].b = int_val != 0; + val->constant->values[0].b = u32val.u32 != 0; break; } - case SpvOpConstant: { + case SpvOpConstant: + case SpvOpSpecConstant: { vtn_fail_if(val->type->base_type != vtn_base_type_scalar, "Result type of %s must be a scalar", spirv_op_to_string(opcode)); @@ -1628,31 +1834,10 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, default: vtn_fail("Unsupported SpvOpConstant bit size: %u", bit_size); } - break; - } - case SpvOpSpecConstant: { - vtn_fail_if(val->type->base_type != vtn_base_type_scalar, - "Result type of %s must be a scalar", - spirv_op_to_string(opcode)); - int bit_size = glsl_get_bit_size(val->type->type); - switch (bit_size) { - case 64: - val->constant->values[0].u64 = - get_specialization64(b, val, vtn_u64_literal(&w[3])); - break; - case 32: - val->constant->values[0].u32 = get_specialization(b, val, w[3]); - break; - case 16: - val->constant->values[0].u16 = get_specialization(b, val, w[3]); - break; - case 8: - val->constant->values[0].u8 = get_specialization(b, val, w[3]); - break; - default: - vtn_fail("Unsupported SpvOpSpecConstant bit size"); - } + if (opcode == SpvOpSpecConstant) + vtn_foreach_decoration(b, val, spec_constant_decoration_cb, + &val->constant->values[0]); break; } @@ -1702,7 +1887,9 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, } case SpvOpSpecConstantOp: { - SpvOp opcode = get_specialization(b, val, w[3]); + nir_const_value u32op = nir_const_value_for_uint(w[3], 32); + vtn_foreach_decoration(b, val, spec_constant_decoration_cb, &u32op); + SpvOp opcode = u32op.u32; switch (opcode) { case SpvOpVectorShuffle: { struct vtn_value *v0 = &b->values[w[4]]; @@ -1840,11 +2027,9 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, case SpvOpUConvert: /* We have a source in a conversion */ src_alu_type = - nir_get_nir_type_for_glsl_type( - vtn_value(b, w[4], vtn_value_type_constant)->type->type); + nir_get_nir_type_for_glsl_type(vtn_get_value_type(b, w[4])->type); /* We use the bitsize of the conversion source to evaluate the opcode later */ - bit_size = glsl_get_bit_size( - vtn_value(b, w[4], vtn_value_type_constant)->type->type); + bit_size = glsl_get_bit_size(vtn_get_value_type(b, w[4])->type); break; default: bit_size = glsl_get_bit_size(val->type->type); @@ -1910,10 +2095,6 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, val->constant = vtn_null_constant(b, val->type); break; - case SpvOpConstantSampler: - vtn_fail("OpConstantSampler requires Kernel Capability"); - break; - default: vtn_fail_with_opcode("Unhandled opcode", opcode); } @@ -1922,20 +2103,6 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL); } -SpvMemorySemanticsMask -vtn_storage_class_to_memory_semantics(SpvStorageClass sc) -{ - switch (sc) { - case SpvStorageClassStorageBuffer: - case SpvStorageClassPhysicalStorageBufferEXT: - return SpvMemorySemanticsUniformMemoryMask; - case SpvStorageClassWorkgroup: - return SpvMemorySemanticsWorkgroupMemoryMask; - default: - return SpvMemorySemanticsMaskNone; - } -} - static void vtn_split_barrier_semantics(struct vtn_builder *b, SpvMemorySemanticsMask semantics, @@ -1985,7 +2152,8 @@ vtn_split_barrier_semantics(struct vtn_builder *b, SpvMemorySemanticsOutputMemoryMask); const SpvMemorySemanticsMask other_semantics = - semantics & ~(order_semantics | av_vis_semantics | storage_semantics); + semantics & ~(order_semantics | av_vis_semantics | storage_semantics | + SpvMemorySemanticsVolatileMask); if (other_semantics) vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics); @@ -2019,9 +2187,9 @@ vtn_split_barrier_semantics(struct vtn_builder *b, *after |= SpvMemorySemanticsMakeAvailableMask | storage_semantics; } -static void -vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope, - SpvMemorySemanticsMask semantics) +static nir_memory_semantics +vtn_mem_semantics_to_nir_mem_semantics(struct vtn_builder *b, + SpvMemorySemanticsMask semantics) { nir_memory_semantics nir_semantics = 0; @@ -2078,6 +2246,13 @@ vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope, nir_semantics |= NIR_MEMORY_MAKE_VISIBLE; } + return nir_semantics; +} + +static nir_variable_mode +vtn_mem_sematics_to_nir_var_modes(struct vtn_builder *b, + SpvMemorySemanticsMask semantics) +{ /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory, * and AtomicCounterMemory are ignored". */ @@ -2091,21 +2266,24 @@ vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope, nir_variable_mode modes = 0; if (semantics & (SpvMemorySemanticsUniformMemoryMask | - SpvMemorySemanticsImageMemoryMask)) - modes |= nir_var_mem_ubo | nir_var_mem_ssbo | nir_var_uniform; + SpvMemorySemanticsImageMemoryMask)) { + modes |= nir_var_uniform | + nir_var_mem_ubo | + nir_var_mem_ssbo | + nir_var_mem_global; + } if (semantics & SpvMemorySemanticsWorkgroupMemoryMask) modes |= nir_var_mem_shared; if (semantics & SpvMemorySemanticsOutputMemoryMask) { - vtn_fail_if(!b->options->caps.vk_memory_model, - "To use Output memory semantics, the VulkanMemoryModel " - "capability must be declared."); modes |= nir_var_shader_out; } - /* No barrier to add. */ - if (nir_semantics == 0 || modes == 0) - return; + return modes; +} +static nir_scope +vtn_scope_to_nir_scope(struct vtn_builder *b, SpvScope scope) +{ nir_scope nir_scope; switch (scope) { case SpvScopeDevice: @@ -2140,54 +2318,75 @@ vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope, vtn_fail("Invalid memory scope"); } - nir_intrinsic_instr *intrin = - nir_intrinsic_instr_create(b->shader, nir_intrinsic_scoped_memory_barrier); - nir_intrinsic_set_memory_semantics(intrin, nir_semantics); + return nir_scope; +} - nir_intrinsic_set_memory_modes(intrin, modes); - nir_intrinsic_set_memory_scope(intrin, nir_scope); - nir_builder_instr_insert(&b->nb, &intrin->instr); +static void +vtn_emit_scoped_control_barrier(struct vtn_builder *b, SpvScope exec_scope, + SpvScope mem_scope, + SpvMemorySemanticsMask semantics) +{ + nir_memory_semantics nir_semantics = + vtn_mem_semantics_to_nir_mem_semantics(b, semantics); + nir_variable_mode modes = vtn_mem_sematics_to_nir_var_modes(b, semantics); + nir_scope nir_exec_scope = vtn_scope_to_nir_scope(b, exec_scope); + + /* Memory semantics is optional for OpControlBarrier. */ + nir_scope nir_mem_scope; + if (nir_semantics == 0 || modes == 0) + nir_mem_scope = NIR_SCOPE_NONE; + else + nir_mem_scope = vtn_scope_to_nir_scope(b, mem_scope); + + nir_scoped_barrier(&b->nb, nir_exec_scope, nir_mem_scope, nir_semantics, modes); +} + +static void +vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope, + SpvMemorySemanticsMask semantics) +{ + nir_variable_mode modes = vtn_mem_sematics_to_nir_var_modes(b, semantics); + nir_memory_semantics nir_semantics = + vtn_mem_semantics_to_nir_mem_semantics(b, semantics); + + /* No barrier to add. */ + if (nir_semantics == 0 || modes == 0) + return; + + nir_scope nir_mem_scope = vtn_scope_to_nir_scope(b, scope); + nir_scoped_barrier(&b->nb, NIR_SCOPE_NONE, nir_mem_scope, nir_semantics, modes); } struct vtn_ssa_value * vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type) { + /* Always use bare types for SSA values for a couple of reasons: + * + * 1. Code which emits deref chains should never listen to the explicit + * layout information on the SSA value if any exists. If we've + * accidentally been relying on this, we want to find those bugs. + * + * 2. We want to be able to quickly check that an SSA value being assigned + * to a SPIR-V value has the right type. Using bare types everywhere + * ensures that we can pointer-compare. + */ struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value); - val->type = type; + val->type = glsl_get_bare_type(type); + if (!glsl_type_is_vector_or_scalar(type)) { - unsigned elems = glsl_get_length(type); + 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 *child_type; - - switch (glsl_get_base_type(type)) { - case GLSL_TYPE_INT: - case GLSL_TYPE_UINT: - case GLSL_TYPE_INT16: - case GLSL_TYPE_UINT16: - case GLSL_TYPE_UINT8: - case GLSL_TYPE_INT8: - case GLSL_TYPE_INT64: - case GLSL_TYPE_UINT64: - case GLSL_TYPE_BOOL: - case GLSL_TYPE_FLOAT: - case GLSL_TYPE_FLOAT16: - 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: - case GLSL_TYPE_INTERFACE: - child_type = glsl_get_struct_field(type, i); - break; - default: - vtn_fail("unkown base type"); + if (glsl_type_is_array_or_matrix(type)) { + const struct glsl_type *elem_type = glsl_get_array_element(type); + for (unsigned i = 0; i < elems; i++) + val->elems[i] = vtn_create_ssa_value(b, elem_type); + } else { + vtn_assert(glsl_type_is_struct_or_ifc(type)); + for (unsigned i = 0; i < elems; i++) { + const struct glsl_type *elem_type = glsl_get_struct_field(type, i); + val->elems[i] = vtn_create_ssa_value(b, elem_type); } - - val->elems[i] = vtn_create_ssa_value(b, child_type); } } @@ -2198,7 +2397,7 @@ static nir_tex_src 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_ssa_value(b, index)->def); + src.src = nir_src_for_ssa(vtn_get_nir_ssa(b, index)); src.src_type = type; return src; } @@ -2239,49 +2438,53 @@ image_operand_arg(struct vtn_builder *b, const uint32_t *w, uint32_t count, return idx; } +static void +non_uniform_decoration_cb(struct vtn_builder *b, + struct vtn_value *val, int member, + const struct vtn_decoration *dec, void *void_ctx) +{ + enum gl_access_qualifier *access = void_ctx; + switch (dec->decoration) { + case SpvDecorationNonUniformEXT: + *access |= ACCESS_NON_UNIFORM; + break; + + default: + break; + } +} + static void vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) { + struct vtn_type *ret_type = vtn_get_type(b, w[1]); + if (opcode == SpvOpSampledImage) { - struct vtn_value *val = - vtn_push_value(b, w[2], vtn_value_type_sampled_image); - val->sampled_image = ralloc(b, struct vtn_sampled_image); - val->sampled_image->image = - vtn_value(b, w[3], vtn_value_type_pointer)->pointer; - val->sampled_image->sampler = - vtn_value(b, w[4], vtn_value_type_pointer)->pointer; + struct vtn_sampled_image si = { + .image = vtn_get_image(b, w[3]), + .sampler = vtn_get_sampler(b, w[4]), + }; + vtn_push_sampled_image(b, w[2], si); return; } else if (opcode == SpvOpImage) { - struct vtn_value *src_val = vtn_untyped_value(b, w[3]); - if (src_val->value_type == vtn_value_type_sampled_image) { - vtn_push_value_pointer(b, w[2], src_val->sampled_image->image); - } else { - vtn_assert(src_val->value_type == vtn_value_type_pointer); - vtn_push_value_pointer(b, w[2], src_val->pointer); - } + struct vtn_sampled_image si = vtn_get_sampled_image(b, w[3]); + vtn_push_image(b, w[2], si.image); return; } - struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type; - - struct vtn_pointer *image = NULL, *sampler = NULL; + nir_deref_instr *image = NULL, *sampler = NULL; struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]); - if (sampled_val->value_type == vtn_value_type_sampled_image) { - image = sampled_val->sampled_image->image; - sampler = sampled_val->sampled_image->sampler; + if (sampled_val->type->base_type == vtn_base_type_sampled_image) { + struct vtn_sampled_image si = vtn_get_sampled_image(b, w[3]); + image = si.image; + sampler = si.sampler; } else { - vtn_assert(sampled_val->value_type == vtn_value_type_pointer); - image = sampled_val->pointer; + image = vtn_get_image(b, w[3]); } - nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image); - nir_deref_instr *sampler_deref = - sampler ? vtn_pointer_to_deref(b, sampler) : NULL; - - const struct glsl_type *image_type = sampled_val->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 enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image->type); + const bool is_array = glsl_sampler_type_is_array(image->type); nir_alu_type dest_type = nir_type_invalid; /* Figure out the base texture operation */ @@ -2335,6 +2538,14 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, dest_type = nir_type_int; break; + case SpvOpFragmentFetchAMD: + texop = nir_texop_fragment_fetch; + break; + + case SpvOpFragmentMaskFetchAMD: + texop = nir_texop_fragment_mask_fetch; + break; + default: vtn_fail_with_opcode("Unhandled opcode", opcode); } @@ -2342,7 +2553,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, nir_tex_src srcs[10]; /* 10 should be enough */ nir_tex_src *p = srcs; - p->src = nir_src_for_ssa(&image_deref->dest.ssa); + p->src = nir_src_for_ssa(&image->dest.ssa); p->src_type = nir_tex_src_texture_deref; p++; @@ -2356,7 +2567,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, vtn_fail_if(sampler == NULL, "%s requires an image of type OpTypeSampledImage", spirv_op_to_string(opcode)); - p->src = nir_src_for_ssa(&sampler_deref->dest.ssa); + p->src = nir_src_for_ssa(&sampler->dest.ssa); p->src_type = nir_tex_src_sampler_deref; p++; break; @@ -2366,6 +2577,8 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, case nir_texop_query_levels: case nir_texop_texture_samples: case nir_texop_samples_identical: + case nir_texop_fragment_fetch: + case nir_texop_fragment_mask_fetch: /* These don't */ break; case nir_texop_txf_ms_fb: @@ -2393,32 +2606,30 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, case SpvOpImageFetch: case SpvOpImageGather: case SpvOpImageDrefGather: - case SpvOpImageQueryLod: { + case SpvOpImageQueryLod: + case SpvOpFragmentFetchAMD: + case SpvOpFragmentMaskFetchAMD: { /* All these types have the coordinate as their first real argument */ - switch (sampler_dim) { - case GLSL_SAMPLER_DIM_1D: - case GLSL_SAMPLER_DIM_BUF: - coord_components = 1; - break; - case GLSL_SAMPLER_DIM_2D: - case GLSL_SAMPLER_DIM_RECT: - case GLSL_SAMPLER_DIM_MS: - coord_components = 2; - break; - case GLSL_SAMPLER_DIM_3D: - case GLSL_SAMPLER_DIM_CUBE: - coord_components = 3; - break; - default: - vtn_fail("Invalid sampler type"); - } + coord_components = glsl_get_sampler_dim_coordinate_components(sampler_dim); if (is_array && texop != nir_texop_lod) coord_components++; - coord = vtn_ssa_value(b, w[idx++])->def; + struct vtn_ssa_value *coord_val = vtn_ssa_value(b, w[idx++]); + coord = coord_val->def; p->src = nir_src_for_ssa(nir_channels(&b->nb, coord, (1 << coord_components) - 1)); + + /* OpenCL allows integer sampling coordinates */ + if (glsl_type_is_integer(coord_val->type) && + opcode == SpvOpImageSampleExplicitLod) { + vtn_fail_if(b->shader->info.stage != MESA_SHADER_KERNEL, + "Unless the Kernel capability is being used, the coordinate parameter " + "OpImageSampleExplicitLod must be floating point."); + + p->src = nir_src_for_ssa(nir_i2f32(&b->nb, p->src.ssa)); + } + p->src_type = nir_tex_src_coord; p++; break; @@ -2471,14 +2682,20 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, if (opcode == SpvOpImageQuerySizeLod) (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod); + /* For OpFragmentFetchAMD, we always have a multisample index */ + if (opcode == SpvOpFragmentFetchAMD) + (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index); + /* Now we need to handle some number of optional arguments */ struct vtn_value *gather_offsets = NULL; if (idx < count) { uint32_t operands = w[idx]; if (operands & SpvImageOperandsBiasMask) { - vtn_assert(texop == nir_texop_tex); - texop = nir_texop_txb; + vtn_assert(texop == nir_texop_tex || + texop == nir_texop_tg4); + if (texop == nir_texop_tex) + texop = nir_texop_txb; uint32_t arg = image_operand_arg(b, w, count, idx, SpvImageOperandsBiasMask); (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_bias); @@ -2486,7 +2703,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, if (operands & SpvImageOperandsLodMask) { vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf || - texop == nir_texop_txs); + texop == nir_texop_txs || texop == nir_texop_tg4); uint32_t arg = image_operand_arg(b, w, count, idx, SpvImageOperandsLodMask); (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_lod); @@ -2557,15 +2774,40 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, is_shadow && glsl_get_components(ret_type->type) == 1; instr->component = gather_component; - if (image && (image->access & ACCESS_NON_UNIFORM)) + /* The Vulkan spec says: + * + * "If an instruction loads from or stores to a resource (including + * atomics and image instructions) and the resource descriptor being + * accessed is not dynamically uniform, then the operand corresponding + * to that resource (e.g. the pointer or sampled image operand) must be + * decorated with NonUniform." + * + * It's very careful to specify that the exact operand must be decorated + * NonUniform. The SPIR-V parser is not expected to chase through long + * chains to find the NonUniform decoration. It's either right there or we + * can assume it doesn't exist. + */ + enum gl_access_qualifier access = 0; + vtn_foreach_decoration(b, sampled_val, non_uniform_decoration_cb, &access); + + if (image && (access & ACCESS_NON_UNIFORM)) instr->texture_non_uniform = true; - if (sampler && (sampler->access & ACCESS_NON_UNIFORM)) + if (sampler && (access & ACCESS_NON_UNIFORM)) instr->sampler_non_uniform = true; - /* for non-query ops, get dest_type from sampler type */ + /* for non-query ops, get dest_type from SPIR-V return type */ if (dest_type == nir_type_invalid) { - switch (glsl_get_sampler_result_type(image_type)) { + /* the return type should match the image type, unless the image type is + * VOID (CL image), in which case the return type dictates the sampler + */ + enum glsl_base_type sampler_base = + glsl_get_sampler_result_type(image->type); + enum glsl_base_type ret_base = glsl_get_base_type(ret_type->type); + vtn_fail_if(sampler_base != ret_base && sampler_base != GLSL_TYPE_VOID, + "SPIR-V return type mismatches image type. This is only valid " + "for untyped images (OpenCL)."); + switch (ret_base) { case GLSL_TYPE_FLOAT: dest_type = nir_type_float; break; case GLSL_TYPE_INT: dest_type = nir_type_int; break; case GLSL_TYPE_UINT: dest_type = nir_type_uint; break; @@ -2613,11 +2855,9 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, } } - struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, ret_type->type); - ssa->def = &instr->dest.ssa; - vtn_push_ssa(b, w[2], ret_type, ssa); - nir_builder_instr_insert(&b->nb, &instr->instr); + + vtn_push_nir_ssa(b, w[2], &instr->dest.ssa); } static void @@ -2635,13 +2875,13 @@ fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicISub: src[0] = - nir_src_for_ssa(nir_ineg(&b->nb, vtn_ssa_value(b, w[6])->def)); + nir_src_for_ssa(nir_ineg(&b->nb, vtn_get_nir_ssa(b, w[6]))); break; case SpvOpAtomicCompareExchange: case SpvOpAtomicCompareExchangeWeak: - src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[8])->def); - src[1] = nir_src_for_ssa(vtn_ssa_value(b, w[7])->def); + src[0] = nir_src_for_ssa(vtn_get_nir_ssa(b, w[8])); + src[1] = nir_src_for_ssa(vtn_get_nir_ssa(b, w[7])); break; case SpvOpAtomicExchange: @@ -2653,7 +2893,8 @@ fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicAnd: case SpvOpAtomicOr: case SpvOpAtomicXor: - src[0] = nir_src_for_ssa(vtn_ssa_value(b, w[6])->def); + case SpvOpAtomicFAddEXT: + src[0] = nir_src_for_ssa(vtn_get_nir_ssa(b, w[6])); break; default: @@ -2664,15 +2905,14 @@ fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode, static nir_ssa_def * get_image_coord(struct vtn_builder *b, uint32_t value) { - struct vtn_ssa_value *coord = vtn_ssa_value(b, value); + nir_ssa_def *coord = vtn_get_nir_ssa(b, value); /* The image_load_store intrinsics assume a 4-dim coordinate */ - unsigned dim = glsl_get_vector_elements(coord->type); unsigned swizzle[4]; for (unsigned i = 0; i < 4; i++) - swizzle[i] = MIN2(i, dim - 1); + swizzle[i] = MIN2(i, coord->num_components - 1); - return nir_swizzle(&b->nb, coord->def, swizzle, 4); + return nir_swizzle(&b->nb, coord, swizzle, 4); } static nir_ssa_def * @@ -2697,9 +2937,10 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, vtn_push_value(b, w[2], vtn_value_type_image_pointer); val->image = ralloc(b, struct vtn_image_pointer); - val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; + val->image->image = vtn_nir_deref(b, w[3]); val->image->coord = get_image_coord(b, w[4]); - val->image->sample = vtn_ssa_value(b, w[5])->def; + val->image->sample = vtn_get_nir_ssa(b, w[5]); + val->image->lod = nir_imm_int(&b->nb, 0); return; } @@ -2707,6 +2948,9 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, SpvScope scope = SpvScopeInvocation; SpvMemorySemanticsMask semantics = 0; + enum gl_access_qualifier access = 0; + + struct vtn_value *res_val; switch (opcode) { case SpvOpAtomicExchange: case SpvOpAtomicCompareExchange: @@ -2723,25 +2967,50 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicAnd: case SpvOpAtomicOr: case SpvOpAtomicXor: - image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image; + case SpvOpAtomicFAddEXT: + res_val = vtn_value(b, w[3], vtn_value_type_image_pointer); + image = *res_val->image; scope = vtn_constant_uint(b, w[4]); semantics = vtn_constant_uint(b, w[5]); + access |= ACCESS_COHERENT; break; case SpvOpAtomicStore: - image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image; + res_val = vtn_value(b, w[1], vtn_value_type_image_pointer); + image = *res_val->image; scope = vtn_constant_uint(b, w[2]); semantics = vtn_constant_uint(b, w[3]); + access |= ACCESS_COHERENT; + break; + + case SpvOpImageQuerySizeLod: + res_val = vtn_untyped_value(b, w[3]); + image.image = vtn_get_image(b, w[3]); + image.coord = NULL; + image.sample = NULL; + image.lod = vtn_ssa_value(b, w[4])->def; break; case SpvOpImageQuerySize: - image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; + res_val = vtn_untyped_value(b, w[3]); + image.image = vtn_get_image(b, w[3]); + image.coord = NULL; + image.sample = NULL; + image.lod = NULL; + break; + + case SpvOpImageQueryFormat: + case SpvOpImageQueryOrder: + res_val = vtn_untyped_value(b, w[3]); + image.image = vtn_get_image(b, w[3]); image.coord = NULL; image.sample = NULL; + image.lod = NULL; break; case SpvOpImageRead: { - image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; + res_val = vtn_untyped_value(b, w[3]); + image.image = vtn_get_image(b, w[3]); image.coord = get_image_coord(b, w[4]); const SpvImageOperandsMask operands = @@ -2750,7 +3019,7 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, if (operands & SpvImageOperandsSampleMask) { uint32_t arg = image_operand_arg(b, w, count, 5, SpvImageOperandsSampleMask); - image.sample = vtn_ssa_value(b, w[arg])->def; + image.sample = vtn_get_nir_ssa(b, w[arg]); } else { image.sample = nir_ssa_undef(&b->nb, 1, 32); } @@ -2764,13 +3033,22 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, scope = vtn_constant_uint(b, w[arg]); } + if (operands & SpvImageOperandsLodMask) { + uint32_t arg = image_operand_arg(b, w, count, 5, + SpvImageOperandsLodMask); + image.lod = vtn_get_nir_ssa(b, w[arg]); + } else { + image.lod = nir_imm_int(&b->nb, 0); + } + /* TODO: Volatile. */ break; } case SpvOpImageWrite: { - image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer; + res_val = vtn_untyped_value(b, w[1]); + image.image = vtn_get_image(b, w[1]); image.coord = get_image_coord(b, w[2]); /* texel = w[3] */ @@ -2781,7 +3059,7 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, if (operands & SpvImageOperandsSampleMask) { uint32_t arg = image_operand_arg(b, w, count, 4, SpvImageOperandsSampleMask); - image.sample = vtn_ssa_value(b, w[arg])->def; + image.sample = vtn_get_nir_ssa(b, w[arg]); } else { image.sample = nir_ssa_undef(&b->nb, 1, 32); } @@ -2795,6 +3073,14 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, scope = vtn_constant_uint(b, w[arg]); } + if (operands & SpvImageOperandsLodMask) { + uint32_t arg = image_operand_arg(b, w, count, 4, + SpvImageOperandsLodMask); + image.lod = vtn_get_nir_ssa(b, w[arg]); + } else { + image.lod = nir_imm_int(&b->nb, 0); + } + /* TODO: Volatile. */ break; @@ -2804,10 +3090,14 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, vtn_fail_with_opcode("Invalid image opcode", opcode); } + if (semantics & SpvMemorySemanticsVolatileMask) + access |= ACCESS_VOLATILE; + nir_intrinsic_op op; switch (opcode) { #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break; OP(ImageQuerySize, size) + OP(ImageQuerySizeLod, size) OP(ImageRead, load) OP(ImageWrite, store) OP(AtomicLoad, load) @@ -2826,6 +3116,9 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, OP(AtomicAnd, atomic_and) OP(AtomicOr, atomic_or) OP(AtomicXor, atomic_xor) + OP(AtomicFAddEXT, atomic_fadd) + OP(ImageQueryFormat, format) + OP(ImageQueryOrder, order) #undef OP default: vtn_fail_with_opcode("Invalid image opcode", opcode); @@ -2833,33 +3126,76 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->shader, op); - nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image.image); - intrin->src[0] = nir_src_for_ssa(&image_deref->dest.ssa); + intrin->src[0] = nir_src_for_ssa(&image.image->dest.ssa); - /* ImageQuerySize doesn't take any extra parameters */ - if (opcode != SpvOpImageQuerySize) { + switch (opcode) { + case SpvOpImageQuerySize: + case SpvOpImageQuerySizeLod: + case SpvOpImageQueryFormat: + case SpvOpImageQueryOrder: + break; + default: /* The image coordinate is always 4 components but we may not have that * many. Swizzle to compensate. */ intrin->src[1] = nir_src_for_ssa(expand_to_vec4(&b->nb, image.coord)); intrin->src[2] = nir_src_for_ssa(image.sample); + break; } - nir_intrinsic_set_access(intrin, image.image->access); + /* The Vulkan spec says: + * + * "If an instruction loads from or stores to a resource (including + * atomics and image instructions) and the resource descriptor being + * accessed is not dynamically uniform, then the operand corresponding + * to that resource (e.g. the pointer or sampled image operand) must be + * decorated with NonUniform." + * + * It's very careful to specify that the exact operand must be decorated + * NonUniform. The SPIR-V parser is not expected to chase through long + * chains to find the NonUniform decoration. It's either right there or we + * can assume it doesn't exist. + */ + vtn_foreach_decoration(b, res_val, non_uniform_decoration_cb, &access); + nir_intrinsic_set_access(intrin, access); switch (opcode) { - case SpvOpAtomicLoad: + case SpvOpImageQueryFormat: + case SpvOpImageQueryOrder: + /* No additional sources */ + break; case SpvOpImageQuerySize: + intrin->src[1] = nir_src_for_ssa(nir_imm_int(&b->nb, 0)); + break; + case SpvOpImageQuerySizeLod: + intrin->src[1] = nir_src_for_ssa(image.lod); + break; + case SpvOpAtomicLoad: case SpvOpImageRead: + /* Only OpImageRead can support a lod parameter if + * SPV_AMD_shader_image_load_store_lod is used but the current NIR + * intrinsics definition for atomics requires us to set it for + * OpAtomicLoad. + */ + intrin->src[3] = nir_src_for_ssa(image.lod); break; case SpvOpAtomicStore: case SpvOpImageWrite: { const uint32_t value_id = opcode == SpvOpAtomicStore ? w[4] : w[3]; - nir_ssa_def *value = vtn_ssa_value(b, value_id)->def; + struct vtn_ssa_value *value = vtn_ssa_value(b, value_id); /* 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)); + intrin->src[3] = nir_src_for_ssa(expand_to_vec4(&b->nb, value->def)); + /* Only OpImageWrite can support a lod parameter if + * SPV_AMD_shader_image_load_store_lod is used but the current NIR + * intrinsics definition for atomics requires us to set it for + * OpAtomicStore. + */ + intrin->src[4] = nir_src_for_ssa(image.lod); + + if (opcode == SpvOpImageWrite) + nir_intrinsic_set_type(intrin, nir_get_nir_type_for_glsl_type(value->type)); break; } @@ -2877,6 +3213,7 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicAnd: case SpvOpAtomicOr: case SpvOpAtomicXor: + case SpvOpAtomicFAddEXT: fill_common_atomic_sources(b, opcode, w, &intrin->src[3]); break; @@ -2895,25 +3232,25 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, vtn_emit_memory_barrier(b, scope, before_semantics); if (opcode != SpvOpImageWrite && opcode != SpvOpAtomicStore) { - struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; + struct vtn_type *type = vtn_get_type(b, w[1]); unsigned dest_components = glsl_get_vector_elements(type->type); - intrin->num_components = nir_intrinsic_infos[op].dest_components; - if (intrin->num_components == 0) + if (nir_intrinsic_infos[op].dest_components == 0) intrin->num_components = dest_components; nir_ssa_dest_init(&intrin->instr, &intrin->dest, - intrin->num_components, 32, NULL); + nir_intrinsic_dest_components(intrin), 32, NULL); nir_builder_instr_insert(&b->nb, &intrin->instr); nir_ssa_def *result = &intrin->dest.ssa; - if (intrin->num_components != dest_components) + if (nir_intrinsic_dest_components(intrin) != dest_components) result = nir_channels(&b->nb, result, (1 << dest_components) - 1); - struct vtn_value *val = - vtn_push_ssa(b, w[2], type, vtn_create_ssa_value(b, type->type)); - val->ssa->def = result; + vtn_push_nir_ssa(b, w[2], result); + + if (opcode == SpvOpImageRead) + nir_intrinsic_set_type(intrin, nir_get_nir_type_for_glsl_type(type->type)); } else { nir_builder_instr_insert(&b->nb, &intrin->instr); } @@ -2943,6 +3280,7 @@ get_ssbo_nir_atomic_op(struct vtn_builder *b, SpvOp opcode) OP(AtomicAnd, atomic_and) OP(AtomicOr, atomic_or) OP(AtomicXor, atomic_xor) + OP(AtomicFAddEXT, atomic_fadd) #undef OP default: vtn_fail_with_opcode("Invalid SSBO atomic", opcode); @@ -2975,7 +3313,7 @@ get_uniform_nir_atomic_op(struct vtn_builder *b, SpvOp opcode) * only need to support GLSL Atomic Counters that are uints and don't * allow direct storage. */ - unreachable("Invalid uniform atomic"); + vtn_fail("Invalid uniform atomic"); } } @@ -3000,6 +3338,7 @@ get_deref_nir_atomic_op(struct vtn_builder *b, SpvOp opcode) OP(AtomicAnd, atomic_and) OP(AtomicOr, atomic_or) OP(AtomicXor, atomic_xor) + OP(AtomicFAddEXT, atomic_fadd) #undef OP default: vtn_fail_with_opcode("Invalid shared atomic", opcode); @@ -3011,13 +3350,14 @@ get_deref_nir_atomic_op(struct vtn_builder *b, SpvOp opcode) */ static void vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, - const uint32_t *w, unsigned count) + const uint32_t *w, UNUSED unsigned count) { struct vtn_pointer *ptr; nir_intrinsic_instr *atomic; SpvScope scope = SpvScopeInvocation; SpvMemorySemanticsMask semantics = 0; + enum gl_access_qualifier access = 0; switch (opcode) { case SpvOpAtomicLoad: @@ -3035,6 +3375,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicAnd: case SpvOpAtomicOr: case SpvOpAtomicXor: + case SpvOpAtomicFAddEXT: ptr = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; scope = vtn_constant_uint(b, w[4]); semantics = vtn_constant_uint(b, w[5]); @@ -3050,10 +3391,12 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode); } + if (semantics & SpvMemorySemanticsVolatileMask) + access |= ACCESS_VOLATILE; + /* uniform as "atomic counter uniform" */ - if (ptr->mode == vtn_variable_mode_uniform) { + if (ptr->mode == vtn_variable_mode_atomic_counter) { 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->src[0] = nir_src_for_ssa(&deref->dest.ssa); @@ -3065,14 +3408,6 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, 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); - break; - case SpvOpAtomicExchange: case SpvOpAtomicCompareExchange: case SpvOpAtomicCompareExchangeWeak: @@ -3105,6 +3440,8 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, nir_intrinsic_op op = get_ssbo_nir_atomic_op(b, opcode); atomic = nir_intrinsic_instr_create(b->nb.shader, op); + nir_intrinsic_set_access(atomic, access | ACCESS_COHERENT); + int src = 0; switch (opcode) { case SpvOpAtomicLoad: @@ -3119,7 +3456,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, 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); + atomic->src[src++] = nir_src_for_ssa(vtn_get_nir_ssa(b, w[4])); if (ptr->mode == vtn_variable_mode_ssbo) atomic->src[src++] = nir_src_for_ssa(index); atomic->src[src++] = nir_src_for_ssa(offset); @@ -3139,6 +3476,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicAnd: case SpvOpAtomicOr: case SpvOpAtomicXor: + case SpvOpAtomicFAddEXT: if (ptr->mode == vtn_variable_mode_ssbo) atomic->src[src++] = nir_src_for_ssa(index); atomic->src[src++] = nir_src_for_ssa(offset); @@ -3155,6 +3493,11 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, atomic = nir_intrinsic_instr_create(b->nb.shader, op); atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa); + if (ptr->mode != vtn_variable_mode_workgroup) + access |= ACCESS_COHERENT; + + nir_intrinsic_set_access(atomic, access); + switch (opcode) { case SpvOpAtomicLoad: atomic->num_components = glsl_get_vector_elements(deref_type); @@ -3163,7 +3506,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, 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); + atomic->src[1] = nir_src_for_ssa(vtn_get_nir_ssa(b, w[4])); break; case SpvOpAtomicExchange: @@ -3180,6 +3523,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicAnd: case SpvOpAtomicOr: case SpvOpAtomicXor: + case SpvOpAtomicFAddEXT: fill_common_atomic_sources(b, opcode, w, &atomic->src[1]); break; @@ -3191,7 +3535,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, /* Atomic ordering operations will implicitly apply to the atomic operation * storage class, so include that too. */ - semantics |= vtn_storage_class_to_memory_semantics(ptr->ptr_type->storage_class); + semantics |= vtn_mode_to_memory_semantics(ptr->mode); SpvMemorySemanticsMask before_semantics; SpvMemorySemanticsMask after_semantics; @@ -3201,16 +3545,13 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, vtn_emit_memory_barrier(b, scope, before_semantics); if (opcode != SpvOpAtomicStore) { - struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; + struct vtn_type *type = vtn_get_type(b, w[1]); nir_ssa_dest_init(&atomic->instr, &atomic->dest, glsl_get_vector_elements(type->type), glsl_get_bit_size(type->type), NULL); - struct vtn_ssa_value *ssa = rzalloc(b, struct vtn_ssa_value); - ssa->def = &atomic->dest.ssa; - ssa->type = type->type; - vtn_push_ssa(b, w[2], type, ssa); + vtn_push_nir_ssa(b, w[2], &atomic->dest.ssa); } nir_builder_instr_insert(&b->nb, &atomic->instr); @@ -3261,58 +3602,6 @@ vtn_ssa_transpose(struct vtn_builder *b, struct vtn_ssa_value *src) return dest; } -nir_ssa_def * -vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index) -{ - return nir_channel(&b->nb, src, index); -} - -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, src->num_components, - src->bit_size); - - 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 * -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) -{ - return nir_vector_extract(&b->nb, src, nir_i2i(&b->nb, index, 32)); -} - -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_imm(&b->nb, index, 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, @@ -3407,17 +3696,29 @@ vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src, struct vtn_ssa_value *cur = dest; unsigned i; for (i = 0; i < num_indices - 1; i++) { + /* If we got a vector here, that means the next index will be trying to + * dereference a scalar. + */ + vtn_fail_if(glsl_type_is_vector_or_scalar(cur->type), + "OpCompositeInsert has too many indices."); + vtn_fail_if(indices[i] >= glsl_get_length(cur->type), + "All indices in an OpCompositeInsert must be in-bounds"); cur = cur->elems[indices[i]]; } if (glsl_type_is_vector_or_scalar(cur->type)) { + vtn_fail_if(indices[i] >= glsl_get_vector_elements(cur->type), + "All indices in an OpCompositeInsert must be in-bounds"); + /* 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]); + cur->def = nir_vector_insert_imm(&b->nb, cur->def, insert->def, indices[i]); } else { + vtn_fail_if(indices[i] >= glsl_get_length(cur->type), + "All indices in an OpCompositeInsert must be in-bounds"); cur->elems[indices[i]] = insert; } @@ -3432,16 +3733,22 @@ vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src, for (unsigned i = 0; i < num_indices; i++) { if (glsl_type_is_vector_or_scalar(cur->type)) { vtn_assert(i == num_indices - 1); + vtn_fail_if(indices[i] >= glsl_get_vector_elements(cur->type), + "All indices in an OpCompositeExtract must be in-bounds"); + /* 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]); + const struct glsl_type *scalar_type = + glsl_scalar_type(glsl_get_base_type(cur->type)); + struct vtn_ssa_value *ret = vtn_create_ssa_value(b, scalar_type); + ret->def = nir_channel(&b->nb, cur->def, indices[i]); return ret; } else { + vtn_fail_if(indices[i] >= glsl_get_length(cur->type), + "All indices in an OpCompositeExtract must be in-bounds"); cur = cur->elems[indices[i]]; } } @@ -3453,25 +3760,25 @@ static void vtn_handle_composite(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) { - struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; + struct vtn_type *type = vtn_get_type(b, w[1]); struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, type->type); switch (opcode) { case SpvOpVectorExtractDynamic: - ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def, - vtn_ssa_value(b, w[4])->def); + ssa->def = nir_vector_extract(&b->nb, vtn_get_nir_ssa(b, w[3]), + vtn_get_nir_ssa(b, w[4])); break; case SpvOpVectorInsertDynamic: - 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); + ssa->def = nir_vector_insert(&b->nb, vtn_get_nir_ssa(b, w[3]), + vtn_get_nir_ssa(b, w[4]), + vtn_get_nir_ssa(b, w[5])); break; case SpvOpVectorShuffle: ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type->type), - vtn_ssa_value(b, w[3])->def, - vtn_ssa_value(b, w[4])->def, + vtn_get_nir_ssa(b, w[3]), + vtn_get_nir_ssa(b, w[4]), w + 5); break; @@ -3481,7 +3788,7 @@ vtn_handle_composite(struct vtn_builder *b, SpvOp opcode, if (glsl_type_is_vector_or_scalar(type->type)) { 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; + srcs[i] = vtn_get_nir_ssa(b, w[3 + i]); ssa->def = vtn_vector_construct(b, glsl_get_vector_elements(type->type), elems, srcs); @@ -3504,15 +3811,17 @@ vtn_handle_composite(struct vtn_builder *b, SpvOp opcode, break; case SpvOpCopyLogical: - case SpvOpCopyObject: ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3])); break; + case SpvOpCopyObject: + vtn_copy_value(b, w[3], w[2]); + return; default: vtn_fail_with_opcode("unknown composite operation", opcode); } - vtn_push_ssa(b, w[2], type, ssa); + vtn_push_ssa_value(b, w[2], ssa); } static void @@ -3526,7 +3835,7 @@ void vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope, SpvMemorySemanticsMask semantics) { - if (b->options->use_scoped_memory_barrier) { + if (b->shader->options->use_scoped_barrier) { vtn_emit_scoped_memory_barrier(b, scope, semantics); return; } @@ -3535,7 +3844,8 @@ vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope, SpvMemorySemanticsUniformMemoryMask | SpvMemorySemanticsWorkgroupMemoryMask | SpvMemorySemanticsAtomicCounterMemoryMask | - SpvMemorySemanticsImageMemoryMask; + SpvMemorySemanticsImageMemoryMask | + SpvMemorySemanticsOutputMemoryMask; /* If we're not actually doing a memory barrier, bail */ if (!(semantics & all_memory_semantics)) @@ -3555,37 +3865,49 @@ vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope, /* There's only two scopes thing left */ vtn_assert(scope == SpvScopeInvocation || scope == SpvScopeDevice); - if ((semantics & all_memory_semantics) == all_memory_semantics) { + /* Map the GLSL memoryBarrier() construct and any barriers with more than one + * semantic to the corresponding NIR one. + */ + if (util_bitcount(semantics & all_memory_semantics) > 1) { vtn_emit_barrier(b, nir_intrinsic_memory_barrier); + if (semantics & SpvMemorySemanticsOutputMemoryMask) { + /* GLSL memoryBarrier() (and the corresponding NIR one) doesn't include + * TCS outputs, so we have to emit it's own intrinsic for that. We + * then need to emit another memory_barrier to prevent moving + * non-output operations to before the tcs_patch barrier. + */ + vtn_emit_barrier(b, nir_intrinsic_memory_barrier_tcs_patch); + vtn_emit_barrier(b, nir_intrinsic_memory_barrier); + } return; } - /* Issue a bunch of more specific barriers */ - uint32_t bits = semantics; - while (bits) { - SpvMemorySemanticsMask semantic = 1 << u_bit_scan(&bits); - switch (semantic) { - case SpvMemorySemanticsUniformMemoryMask: - vtn_emit_barrier(b, nir_intrinsic_memory_barrier_buffer); - break; - case SpvMemorySemanticsWorkgroupMemoryMask: - vtn_emit_barrier(b, nir_intrinsic_memory_barrier_shared); - break; - case SpvMemorySemanticsAtomicCounterMemoryMask: - vtn_emit_barrier(b, nir_intrinsic_memory_barrier_atomic_counter); - break; - case SpvMemorySemanticsImageMemoryMask: - vtn_emit_barrier(b, nir_intrinsic_memory_barrier_image); - break; - default: - break;; - } + /* Issue a more specific barrier */ + switch (semantics & all_memory_semantics) { + case SpvMemorySemanticsUniformMemoryMask: + vtn_emit_barrier(b, nir_intrinsic_memory_barrier_buffer); + break; + case SpvMemorySemanticsWorkgroupMemoryMask: + vtn_emit_barrier(b, nir_intrinsic_memory_barrier_shared); + break; + case SpvMemorySemanticsAtomicCounterMemoryMask: + vtn_emit_barrier(b, nir_intrinsic_memory_barrier_atomic_counter); + break; + case SpvMemorySemanticsImageMemoryMask: + vtn_emit_barrier(b, nir_intrinsic_memory_barrier_image); + break; + case SpvMemorySemanticsOutputMemoryMask: + if (b->nb.shader->info.stage == MESA_SHADER_TESS_CTRL) + vtn_emit_barrier(b, nir_intrinsic_memory_barrier_tcs_patch); + break; + default: + break; } } static void vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode, - const uint32_t *w, unsigned count) + const uint32_t *w, UNUSED unsigned count) { switch (opcode) { case SpvOpEmitVertex: @@ -3633,13 +3955,52 @@ vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode, } case SpvOpControlBarrier: { + SpvScope execution_scope = vtn_constant_uint(b, w[1]); 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); - SpvScope execution_scope = vtn_constant_uint(b, w[1]); - if (execution_scope == SpvScopeWorkgroup) - vtn_emit_barrier(b, nir_intrinsic_barrier); + /* GLSLang, prior to commit 8297936dd6eb3, emitted OpControlBarrier with + * memory semantics of None for GLSL barrier(). + * And before that, prior to c3f1cdfa, emitted the OpControlBarrier with + * Device instead of Workgroup for execution scope. + */ + if (b->wa_glslang_cs_barrier && + b->nb.shader->info.stage == MESA_SHADER_COMPUTE && + (execution_scope == SpvScopeWorkgroup || + execution_scope == SpvScopeDevice) && + memory_semantics == SpvMemorySemanticsMaskNone) { + execution_scope = SpvScopeWorkgroup; + memory_scope = SpvScopeWorkgroup; + memory_semantics = SpvMemorySemanticsAcquireReleaseMask | + SpvMemorySemanticsWorkgroupMemoryMask; + } + + /* From the SPIR-V spec: + * + * "When used with the TessellationControl execution model, it also + * implicitly synchronizes the Output Storage Class: Writes to Output + * variables performed by any invocation executed prior to a + * OpControlBarrier will be visible to any other invocation after + * return from that OpControlBarrier." + */ + if (b->nb.shader->info.stage == MESA_SHADER_TESS_CTRL) { + memory_semantics &= ~(SpvMemorySemanticsAcquireMask | + SpvMemorySemanticsReleaseMask | + SpvMemorySemanticsAcquireReleaseMask | + SpvMemorySemanticsSequentiallyConsistentMask); + memory_semantics |= SpvMemorySemanticsAcquireReleaseMask | + SpvMemorySemanticsOutputMemoryMask; + } + + if (b->shader->options->use_scoped_barrier) { + vtn_emit_scoped_control_barrier(b, execution_scope, memory_scope, + memory_semantics); + } else { + vtn_emit_memory_barrier(b, memory_scope, memory_semantics); + + if (execution_scope == SpvScopeWorkgroup) + vtn_emit_barrier(b, nir_intrinsic_control_barrier); + } break; } @@ -3809,10 +4170,10 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, case SpvCapabilityInputAttachment: case SpvCapabilityImageGatherExtended: case SpvCapabilityStorageImageExtendedFormats: + case SpvCapabilityVector16: break; case SpvCapabilityLinkage: - case SpvCapabilityVector16: case SpvCapabilityFloat16Buffer: case SpvCapabilitySparseResidency: vtn_warn("Unsupported SPIR-V capability: %s", @@ -3865,11 +4226,17 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, break; case SpvCapabilityImageBasic: + spv_check_supported(kernel_image, cap); + break; + + case SpvCapabilityLiteralSampler: + spv_check_supported(literal_sampler, cap); + break; + case SpvCapabilityImageReadWrite: case SpvCapabilityImageMipmap: case SpvCapabilityPipes: case SpvCapabilityDeviceEnqueue: - case SpvCapabilityLiteralSampler: case SpvCapabilityGenericPointer: vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s", spirv_capability_to_string(cap)); @@ -4001,7 +4368,7 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, spv_check_supported(float_controls, cap); break; - case SpvCapabilityPhysicalStorageBufferAddressesEXT: + case SpvCapabilityPhysicalStorageBufferAddresses: spv_check_supported(physical_storage_buffer_address, cap); break; @@ -4038,6 +4405,30 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, spv_check_supported(vk_memory_model_device_scope, cap); break; + case SpvCapabilityImageReadWriteLodAMD: + spv_check_supported(amd_image_read_write_lod, cap); + break; + + case SpvCapabilityIntegerFunctions2INTEL: + spv_check_supported(integer_functions2, cap); + break; + + case SpvCapabilityFragmentMaskAMD: + spv_check_supported(amd_fragment_mask, cap); + break; + + case SpvCapabilityImageGatherBiasLodAMD: + spv_check_supported(amd_image_gather_bias_lod, cap); + break; + + case SpvCapabilityAtomicFloat32AddEXT: + spv_check_supported(float32_atomic_add, cap); + break; + + case SpvCapabilityAtomicFloat64AddEXT: + spv_check_supported(float64_atomic_add, cap); + break; + default: vtn_fail("Unhandled capability: %s (%u)", spirv_capability_to_string(cap), cap); @@ -4056,28 +4447,38 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, "AddressingModelPhysical32 only supported for kernels"); b->shader->info.cs.ptr_size = 32; b->physical_ptrs = true; - b->options->shared_addr_format = nir_address_format_32bit_global; - b->options->global_addr_format = nir_address_format_32bit_global; - b->options->temp_addr_format = nir_address_format_32bit_global; + assert(nir_address_format_bit_size(b->options->global_addr_format) == 32); + assert(nir_address_format_num_components(b->options->global_addr_format) == 1); + assert(nir_address_format_bit_size(b->options->shared_addr_format) == 32); + assert(nir_address_format_num_components(b->options->shared_addr_format) == 1); + if (!b->options->constant_as_global) { + assert(nir_address_format_bit_size(b->options->ubo_addr_format) == 32); + assert(nir_address_format_num_components(b->options->ubo_addr_format) == 1); + } 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_addr_format = nir_address_format_64bit_global; - b->options->global_addr_format = nir_address_format_64bit_global; - b->options->temp_addr_format = nir_address_format_64bit_global; + assert(nir_address_format_bit_size(b->options->global_addr_format) == 64); + assert(nir_address_format_num_components(b->options->global_addr_format) == 1); + assert(nir_address_format_bit_size(b->options->shared_addr_format) == 64); + assert(nir_address_format_num_components(b->options->shared_addr_format) == 1); + if (!b->options->constant_as_global) { + assert(nir_address_format_bit_size(b->options->ubo_addr_format) == 64); + assert(nir_address_format_num_components(b->options->ubo_addr_format) == 1); + } break; case SpvAddressingModelLogical: - vtn_fail_if(b->shader->info.stage >= MESA_SHADER_STAGES, + vtn_fail_if(b->shader->info.stage == MESA_SHADER_KERNEL, "AddressingModelLogical only supported for shaders"); b->physical_ptrs = false; break; - case SpvAddressingModelPhysicalStorageBuffer64EXT: + case SpvAddressingModelPhysicalStorageBuffer64: vtn_fail_if(!b->options || !b->options->caps.physical_storage_buffer_address, - "AddressingModelPhysicalStorageBuffer64EXT not supported"); + "AddressingModelPhysicalStorageBuffer64 not supported"); break; default: vtn_fail("Unknown addressing model: %s (%u)", @@ -4085,6 +4486,7 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, break; } + b->mem_model = w[2]; switch (w[2]) { case SpvMemoryModelSimple: case SpvMemoryModelGLSL450: @@ -4131,6 +4533,17 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, vtn_handle_decoration(b, opcode, w, count); break; + case SpvOpExtInst: { + struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension); + if (val->ext_handler == vtn_handle_non_semantic_instruction) { + /* NonSemantic extended instructions are acceptable in preamble. */ + vtn_handle_non_semantic_instruction(b, w[4], w, count); + return true; + } else { + return false; /* End of preamble. */ + } + } + default: return false; /* End of preamble */ } @@ -4140,7 +4553,7 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, static void vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point, - const struct vtn_decoration *mode, void *data) + const struct vtn_decoration *mode, UNUSED void *data) { vtn_assert(b->entry_point == entry_point); @@ -4191,14 +4604,7 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point, b->shader->info.cs.local_size[2] = mode->operands[2]; break; - case SpvExecutionModeLocalSizeId: - b->shader->info.cs.local_size[0] = vtn_constant_uint(b, mode->operands[0]); - b->shader->info.cs.local_size[1] = vtn_constant_uint(b, mode->operands[1]); - b->shader->info.cs.local_size[2] = vtn_constant_uint(b, mode->operands[2]); - break; - case SpvExecutionModeLocalSizeHint: - case SpvExecutionModeLocalSizeHintId: break; /* Nothing to do with this */ case SpvExecutionModeOutputVertices: @@ -4328,8 +4734,60 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point, case SpvExecutionModeDenormFlushToZero: case SpvExecutionModeSignedZeroInfNanPreserve: case SpvExecutionModeRoundingModeRTE: - case SpvExecutionModeRoundingModeRTZ: - /* Already handled in vtn_handle_rounding_mode_in_execution_mode() */ + case SpvExecutionModeRoundingModeRTZ: { + unsigned execution_mode = 0; + switch (mode->exec_mode) { + case SpvExecutionModeDenormPreserve: + switch (mode->operands[0]) { + case 16: execution_mode = FLOAT_CONTROLS_DENORM_PRESERVE_FP16; break; + case 32: execution_mode = FLOAT_CONTROLS_DENORM_PRESERVE_FP32; break; + case 64: execution_mode = FLOAT_CONTROLS_DENORM_PRESERVE_FP64; break; + default: vtn_fail("Floating point type not supported"); + } + break; + case SpvExecutionModeDenormFlushToZero: + switch (mode->operands[0]) { + case 16: execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16; break; + case 32: execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32; break; + case 64: execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64; break; + default: vtn_fail("Floating point type not supported"); + } + break; + case SpvExecutionModeSignedZeroInfNanPreserve: + switch (mode->operands[0]) { + case 16: execution_mode = FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16; break; + case 32: execution_mode = FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32; break; + case 64: execution_mode = FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64; break; + default: vtn_fail("Floating point type not supported"); + } + break; + case SpvExecutionModeRoundingModeRTE: + switch (mode->operands[0]) { + case 16: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16; break; + case 32: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32; break; + case 64: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64; break; + default: vtn_fail("Floating point type not supported"); + } + break; + case SpvExecutionModeRoundingModeRTZ: + switch (mode->operands[0]) { + case 16: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16; break; + case 32: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32; break; + case 64: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64; break; + default: vtn_fail("Floating point type not supported"); + } + break; + default: + break; + } + + b->shader->info.float_controls_execution_mode |= execution_mode; + break; + } + + case SpvExecutionModeLocalSizeId: + case SpvExecutionModeLocalSizeHintId: + /* Handled later by vtn_handle_execution_mode_id(). */ break; default: @@ -4340,60 +4798,28 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point, } static void -vtn_handle_rounding_mode_in_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point, - const struct vtn_decoration *mode, void *data) +vtn_handle_execution_mode_id(struct vtn_builder *b, struct vtn_value *entry_point, + const struct vtn_decoration *mode, UNUSED void *data) { - vtn_assert(b->entry_point == entry_point); - unsigned execution_mode = 0; + vtn_assert(b->entry_point == entry_point); - switch(mode->exec_mode) { - case SpvExecutionModeDenormPreserve: - switch (mode->operands[0]) { - case 16: execution_mode = FLOAT_CONTROLS_DENORM_PRESERVE_FP16; break; - case 32: execution_mode = FLOAT_CONTROLS_DENORM_PRESERVE_FP32; break; - case 64: execution_mode = FLOAT_CONTROLS_DENORM_PRESERVE_FP64; break; - default: vtn_fail("Floating point type not supported"); - } - break; - case SpvExecutionModeDenormFlushToZero: - switch (mode->operands[0]) { - case 16: execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16; break; - case 32: execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32; break; - case 64: execution_mode = FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64; break; - default: vtn_fail("Floating point type not supported"); - } - break; - case SpvExecutionModeSignedZeroInfNanPreserve: - switch (mode->operands[0]) { - case 16: execution_mode = FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16; break; - case 32: execution_mode = FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32; break; - case 64: execution_mode = FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64; break; - default: vtn_fail("Floating point type not supported"); - } - break; - case SpvExecutionModeRoundingModeRTE: - switch (mode->operands[0]) { - case 16: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16; break; - case 32: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32; break; - case 64: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64; break; - default: vtn_fail("Floating point type not supported"); - } + switch (mode->exec_mode) { + case SpvExecutionModeLocalSizeId: + b->shader->info.cs.local_size[0] = vtn_constant_uint(b, mode->operands[0]); + b->shader->info.cs.local_size[1] = vtn_constant_uint(b, mode->operands[1]); + b->shader->info.cs.local_size[2] = vtn_constant_uint(b, mode->operands[2]); break; - case SpvExecutionModeRoundingModeRTZ: - switch (mode->operands[0]) { - case 16: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16; break; - case 32: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32; break; - case 64: execution_mode = FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64; break; - default: vtn_fail("Floating point type not supported"); - } + + case SpvExecutionModeLocalSizeHintId: + /* Nothing to do with this hint. */ break; default: + /* Nothing to do. Literal execution modes already handled by + * vtn_handle_execution_mode(). */ break; } - - b->shader->info.float_controls_execution_mode |= execution_mode; } static bool @@ -4454,7 +4880,6 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpConstantFalse: case SpvOpConstant: case SpvOpConstantComposite: - case SpvOpConstantSampler: case SpvOpConstantNull: case SpvOpSpecConstantTrue: case SpvOpSpecConstantFalse: @@ -4466,9 +4891,18 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpUndef: case SpvOpVariable: + case SpvOpConstantSampler: vtn_handle_variables(b, opcode, w, count); break; + case SpvOpExtInst: { + struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension); + /* NonSemantic extended instructions are acceptable in preamble, others + * will indicate the end of preamble. + */ + return val->ext_handler == vtn_handle_non_semantic_instruction; + } + default: return false; /* End of preamble */ } @@ -4543,70 +4977,66 @@ vtn_handle_select(struct vtn_builder *b, SpvOp opcode, vtn_fail("Result type of OpSelect must be a scalar, composite, or pointer"); } - struct vtn_type *res_type = vtn_value(b, w[1], vtn_value_type_type)->type; - struct vtn_ssa_value *ssa = vtn_nir_select(b, - vtn_ssa_value(b, w[3]), vtn_ssa_value(b, w[4]), vtn_ssa_value(b, w[5])); - - vtn_push_ssa(b, w[2], res_type, ssa); + vtn_push_ssa_value(b, w[2], + vtn_nir_select(b, vtn_ssa_value(b, w[3]), + vtn_ssa_value(b, w[4]), + vtn_ssa_value(b, w[5]))); } static void vtn_handle_ptr(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) { - struct vtn_type *type1 = vtn_untyped_value(b, w[3])->type; - struct vtn_type *type2 = vtn_untyped_value(b, w[4])->type; - vtn_fail_if(type1->base_type != vtn_base_type_pointer || - type2->base_type != vtn_base_type_pointer, - "%s operands must have pointer types", - spirv_op_to_string(opcode)); - vtn_fail_if(type1->storage_class != type2->storage_class, - "%s operands must have the same storage class", - spirv_op_to_string(opcode)); + struct vtn_type *type1 = vtn_get_value_type(b, w[3]); + struct vtn_type *type2 = vtn_get_value_type(b, w[4]); + vtn_fail_if(type1->base_type != vtn_base_type_pointer || + type2->base_type != vtn_base_type_pointer, + "%s operands must have pointer types", + spirv_op_to_string(opcode)); + vtn_fail_if(type1->storage_class != type2->storage_class, + "%s operands must have the same storage class", + spirv_op_to_string(opcode)); - struct vtn_type *vtn_type = - vtn_value(b, w[1], vtn_value_type_type)->type; - const struct glsl_type *type = vtn_type->type; + struct vtn_type *vtn_type = vtn_get_type(b, w[1]); + const struct glsl_type *type = vtn_type->type; - nir_address_format addr_format = vtn_mode_to_address_format( - b, vtn_storage_class_to_mode(b, type1->storage_class, NULL, NULL)); + nir_address_format addr_format = vtn_mode_to_address_format( + b, vtn_storage_class_to_mode(b, type1->storage_class, NULL, NULL)); - nir_ssa_def *def; + nir_ssa_def *def; - switch (opcode) { - case SpvOpPtrDiff: { - /* OpPtrDiff returns the difference in number of elements (not byte offset). */ - unsigned elem_size, elem_align; - glsl_get_natural_size_align_bytes(type1->deref->type, - &elem_size, &elem_align); - - def = nir_build_addr_isub(&b->nb, - vtn_ssa_value(b, w[3])->def, - vtn_ssa_value(b, w[4])->def, - addr_format); - def = nir_idiv(&b->nb, def, nir_imm_intN_t(&b->nb, elem_size, def->bit_size)); - def = nir_i2i(&b->nb, def, glsl_get_bit_size(type)); - break; - } + switch (opcode) { + case SpvOpPtrDiff: { + /* OpPtrDiff returns the difference in number of elements (not byte offset). */ + unsigned elem_size, elem_align; + glsl_get_natural_size_align_bytes(type1->deref->type, + &elem_size, &elem_align); - case SpvOpPtrEqual: - case SpvOpPtrNotEqual: { - def = nir_build_addr_ieq(&b->nb, - vtn_ssa_value(b, w[3])->def, - vtn_ssa_value(b, w[4])->def, - addr_format); - if (opcode == SpvOpPtrNotEqual) - def = nir_inot(&b->nb, def); - break; - } + def = nir_build_addr_isub(&b->nb, + vtn_get_nir_ssa(b, w[3]), + vtn_get_nir_ssa(b, w[4]), + addr_format); + def = nir_idiv(&b->nb, def, nir_imm_intN_t(&b->nb, elem_size, def->bit_size)); + def = nir_i2i(&b->nb, def, glsl_get_bit_size(type)); + break; + } - default: - unreachable("Invalid ptr operation"); - } + case SpvOpPtrEqual: + case SpvOpPtrNotEqual: { + def = nir_build_addr_ieq(&b->nb, + vtn_get_nir_ssa(b, w[3]), + vtn_get_nir_ssa(b, w[4]), + addr_format); + if (opcode == SpvOpPtrNotEqual) + def = nir_inot(&b->nb, def); + break; + } + + default: + unreachable("Invalid ptr operation"); + } - struct vtn_ssa_value *ssa_value = vtn_create_ssa_value(b, type); - ssa_value->def = def; - vtn_push_ssa(b, w[2], vtn_type, ssa_value); + vtn_push_nir_ssa(b, w[2], def); } static bool @@ -4624,7 +5054,7 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpUndef: { struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_undef); - val->type = vtn_value(b, w[1], vtn_value_type_type)->type; + val->type = vtn_get_type(b, w[1]); break; } @@ -4664,7 +5094,6 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpImageFetch: case SpvOpImageGather: case SpvOpImageDrefGather: - case SpvOpImageQuerySizeLod: case SpvOpImageQueryLod: case SpvOpImageQueryLevels: case SpvOpImageQuerySamples: @@ -4674,21 +5103,29 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpImageRead: case SpvOpImageWrite: case SpvOpImageTexelPointer: + case SpvOpImageQueryFormat: + case SpvOpImageQueryOrder: vtn_handle_image(b, opcode, w, count); break; + case SpvOpImageQuerySizeLod: case SpvOpImageQuerySize: { - struct vtn_pointer *image = - vtn_value(b, w[3], vtn_value_type_pointer)->pointer; - if (glsl_type_is_image(image->type->type)) { + struct vtn_type *image_type = vtn_get_value_type(b, w[3]); + vtn_assert(image_type->base_type == vtn_base_type_image); + if (glsl_type_is_image(image_type->glsl_image)) { vtn_handle_image(b, opcode, w, count); } else { - vtn_assert(glsl_type_is_sampler(image->type->type)); + vtn_assert(glsl_type_is_sampler(image_type->glsl_image)); vtn_handle_texture(b, opcode, w, count); } break; } + case SpvOpFragmentMaskFetchAMD: + case SpvOpFragmentFetchAMD: + vtn_handle_texture(b, opcode, w, count); + break; + case SpvOpAtomicLoad: case SpvOpAtomicExchange: case SpvOpAtomicCompareExchange: @@ -4703,7 +5140,8 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicUMax: case SpvOpAtomicAnd: case SpvOpAtomicOr: - case SpvOpAtomicXor: { + case SpvOpAtomicXor: + case SpvOpAtomicFAddEXT: { struct vtn_value *pointer = vtn_untyped_value(b, w[3]); if (pointer->value_type == vtn_value_type_image_pointer) { vtn_handle_image(b, opcode, w, count); @@ -4825,6 +5263,20 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpVectorTimesMatrix: case SpvOpMatrixTimesVector: case SpvOpMatrixTimesMatrix: + case SpvOpUCountLeadingZerosINTEL: + case SpvOpUCountTrailingZerosINTEL: + case SpvOpAbsISubINTEL: + case SpvOpAbsUSubINTEL: + case SpvOpIAddSatINTEL: + case SpvOpUAddSatINTEL: + case SpvOpIAverageINTEL: + case SpvOpUAverageINTEL: + case SpvOpIAverageRoundedINTEL: + case SpvOpUAverageRoundedINTEL: + case SpvOpISubSatINTEL: + case SpvOpUSubSatINTEL: + case SpvOpIMul32x16INTEL: + case SpvOpUMul32x16INTEL: vtn_handle_alu(b, opcode, w, count); break; @@ -4941,17 +5393,24 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, nir_ssa_dest_init(&intrin->instr, &intrin->dest, 1, 1, NULL); nir_builder_instr_insert(&b->nb, &intrin->instr); - struct vtn_type *res_type = - vtn_value(b, w[1], vtn_value_type_type)->type; - struct vtn_ssa_value *val = vtn_create_ssa_value(b, res_type->type); - val->def = &intrin->dest.ssa; - - vtn_push_ssa(b, w[2], res_type, val); + vtn_push_nir_ssa(b, w[2], &intrin->dest.ssa); break; } case SpvOpReadClockKHR: { - assert(vtn_constant_uint(b, w[3]) == SpvScopeSubgroup); + SpvScope scope = vtn_constant_uint(b, w[3]); + nir_scope nir_scope; + + switch (scope) { + case SpvScopeDevice: + nir_scope = NIR_SCOPE_DEVICE; + break; + case SpvScopeSubgroup: + nir_scope = NIR_SCOPE_SUBGROUP; + break; + default: + vtn_fail("invalid read clock scope"); + } /* Operation supports two result types: uvec2 and uint64_t. The NIR * intrinsic gives uvec2, so pack the result for the other case. @@ -4959,9 +5418,10 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_shader_clock); nir_ssa_dest_init(&intrin->instr, &intrin->dest, 2, 32, NULL); + nir_intrinsic_set_memory_scope(intrin, nir_scope); nir_builder_instr_insert(&b->nb, &intrin->instr); - struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; + struct vtn_type *type = vtn_get_type(b, w[1]); const struct glsl_type *dest_type = type->type; nir_ssa_def *result; @@ -4974,13 +5434,14 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, result = nir_pack_64_2x32(&b->nb, &intrin->dest.ssa); } - struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); - val->type = type; - val->ssa = vtn_create_ssa_value(b, dest_type); - val->ssa->def = result; + vtn_push_nir_ssa(b, w[2], result); break; } + case SpvOpLifetimeStart: + case SpvOpLifetimeStop: + break; + default: vtn_fail_with_opcode("Unhandled opcode", opcode); } @@ -5004,7 +5465,7 @@ vtn_create_builder(const uint32_t *words, size_t word_count, b->file = NULL; b->line = -1; b->col = -1; - exec_list_make_empty(&b->functions); + list_inithead(&b->functions); b->entry_point_stage = stage; b->entry_point_name = entry_point_name; b->options = dup_options; @@ -5028,11 +5489,12 @@ vtn_create_builder(const uint32_t *words, size_t word_count, 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. + /* In GLSLang commit 8297936dd6eb3, their handling of barrier() was fixed + * to provide correct memory semantics on compute shader barrier() + * commands. Prior to that, we need to fix them up ourselves. This + * GLSLang fix caused them to bump to generator version 3. */ - b->wa_glslang_179 = (generator_id == 8 && generator_version == 1); + b->wa_glslang_cs_barrier = (generator_id == 8 && generator_version < 3); /* words[2] == generator magic */ unsigned value_id_bound = words[3]; @@ -5082,12 +5544,23 @@ vtn_emit_kernel_entry_point_wrapper(struct vtn_builder *b, /* input variable */ nir_variable *in_var = rzalloc(b->nb.shader, nir_variable); - in_var->data.mode = nir_var_shader_in; + in_var->data.mode = nir_var_uniform; in_var->data.read_only = true; in_var->data.location = i; + if (param_type->base_type == vtn_base_type_image) { + in_var->data.access = 0; + if (param_type->access_qualifier & SpvAccessQualifierReadOnly) + in_var->data.access |= ACCESS_NON_WRITEABLE; + if (param_type->access_qualifier & SpvAccessQualifierWriteOnly) + in_var->data.access |= ACCESS_NON_READABLE; + } if (is_by_val) in_var->type = param_type->deref->type; + else if (param_type->base_type == vtn_base_type_image) + in_var->type = param_type->glsl_image; + else if (param_type->base_type == vtn_base_type_sampler) + in_var->type = glsl_bare_sampler_type(); else in_var->type = param_type->type; @@ -5102,6 +5575,10 @@ vtn_emit_kernel_entry_point_wrapper(struct vtn_builder *b, 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 if (param_type->base_type == vtn_base_type_image || + param_type->base_type == vtn_base_type_sampler) { + /* Don't load the var, just pass a deref of it */ + call->params[i] = nir_src_for_ssa(&nir_build_deref_var(&b->nb, in_var)->dest.ssa); } else { call->params[i] = nir_src_for_ssa(nir_load_var(&b->nb, in_var)); } @@ -5150,16 +5627,17 @@ spirv_to_nir(const uint32_t *words, size_t word_count, return NULL; } + /* Ensure a sane address mode is being used for function temps */ + assert(nir_address_format_bit_size(b->options->temp_addr_format) == nir_get_ptr_bitsize(b->shader)); + assert(nir_address_format_num_components(b->options->temp_addr_format) == 1); + /* Set shader info defaults */ if (stage == MESA_SHADER_GEOMETRY) b->shader->info.gs.invocations = 1; - /* Parse rounding mode execution modes. This has to happen earlier than - * other changes in the execution modes since they can affect, for example, - * the result of the floating point constants. - */ + /* Parse execution modes. */ vtn_foreach_execution_mode(b, b->entry_point, - vtn_handle_rounding_mode_in_execution_mode, NULL); + vtn_handle_execution_mode, NULL); b->specializations = spec; b->num_specializations = num_spec; @@ -5168,9 +5646,11 @@ spirv_to_nir(const uint32_t *words, size_t word_count, words = vtn_foreach_instruction(b, words, word_end, vtn_handle_variable_or_type_instruction); - /* Parse execution modes */ + /* Parse execution modes that depend on IDs. Must happen after we have + * constants parsed. + */ vtn_foreach_execution_mode(b, b->entry_point, - vtn_handle_execution_mode, NULL); + vtn_handle_execution_mode_id, NULL); if (b->workgroup_size_builtin) { vtn_assert(b->workgroup_size_builtin->type->type == @@ -5195,7 +5675,8 @@ spirv_to_nir(const uint32_t *words, size_t word_count, bool progress; do { progress = false; - foreach_list_typed(struct vtn_function, func, node, &b->functions) { + vtn_foreach_cf_node(node, &b->functions) { + struct vtn_function *func = vtn_cf_node_as_function(node); if (func->referenced && !func->emitted) { b->const_table = _mesa_pointer_hash_table_create(b); @@ -5213,6 +5694,9 @@ spirv_to_nir(const uint32_t *words, size_t word_count, if (entry_point->num_params && b->shader->info.stage == MESA_SHADER_KERNEL) entry_point = vtn_emit_kernel_entry_point_wrapper(b, entry_point); + /* structurize the CFG */ + nir_lower_goto_ifs(b->shader); + entry_point->is_entrypoint = true; /* When multiple shader stages exist in the same SPIR-V module, we @@ -5225,9 +5709,9 @@ spirv_to_nir(const uint32_t *words, size_t word_count, * 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_lower_variable_initializers(b->shader, nir_var_shader_out); nir_remove_dead_variables(b->shader, - nir_var_shader_in | nir_var_shader_out); + nir_var_shader_in | nir_var_shader_out, NULL); /* We sometimes generate bogus derefs that, while never used, give the * validator a bit of heartburn. Run dead code to get rid of them.