X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fcompiler%2Fspirv%2Fspirv_to_nir.c;h=425fbcbbe72092fcee454c47a3be803d2c58e19e;hb=10c4a7cf59733ae2058a76b880ea0767a59dad4f;hp=5aac137471074c98a8483054402a590f751ea079;hpb=b4eff831803b0cb9500ae5607fe7620533b28a3e;p=mesa.git diff --git a/src/compiler/spirv/spirv_to_nir.c b/src/compiler/spirv/spirv_to_nir.c index 5aac1374710..425fbcbbe72 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 @@ -162,14 +163,6 @@ _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) { @@ -236,31 +229,19 @@ vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant, nir_load_const_instr *load = nir_load_const_instr_create(b->shader, num_components, bit_size); - memcpy(load->value, constant->values[0], + 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 rows = glsl_get_vector_elements(val->type); unsigned columns = glsl_get_matrix_columns(val->type); val->elems = ralloc_array(b, struct vtn_ssa_value *, columns); - - for (unsigned i = 0; i < columns; i++) { - struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value); - col_val->type = glsl_get_column_type(val->type); - nir_load_const_instr *load = - nir_load_const_instr_create(b->shader, rows, bit_size); - - memcpy(load->value, constant->values[i], - sizeof(nir_const_value) * load->def.num_components); - - nir_instr_insert_before_cf_list(&b->nb.impl->body, &load->instr); - col_val->def = &load->def; - - val->elems[i] = col_val; - } + 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; } @@ -383,6 +364,14 @@ 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) @@ -394,13 +383,21 @@ vtn_handle_extension(struct vtn_builder *b, SpvOp opcode, if (strcmp(ext, "GLSL.std.450") == 0) { val->ext_handler = vtn_handle_glsl450_instruction; } else if ((strcmp(ext, "SPV_AMD_gcn_shader") == 0) - && (b->options && b->options->caps.gcn_shader)) { + && (b->options && b->options->caps.amd_gcn_shader)) { val->ext_handler = vtn_handle_amd_gcn_shader_instruction; + } else if ((strcmp(ext, "SPV_AMD_shader_ballot") == 0) + && (b->options && b->options->caps.amd_shader_ballot)) { + val->ext_handler = vtn_handle_amd_shader_ballot_instruction; } else if ((strcmp(ext, "SPV_AMD_shader_trinary_minmax") == 0) - && (b->options && b->options->caps.trinary_minmax)) { + && (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); } @@ -731,14 +728,21 @@ array_stride_decoration_cb(struct vtn_builder *b, struct vtn_type *type = val->type; if (dec->decoration == SpvDecorationArrayStride) { - vtn_fail_if(dec->operands[0] == 0, "ArrayStride must be non-zero"); - type->stride = dec->operands[0]; + if (vtn_type_contains_block(b, type)) { + vtn_warn("The ArrayStride decoration cannot be applied to an array " + "type which contains a structure type decorated Block " + "or BufferBlock"); + /* Ignore the decoration */ + } else { + vtn_fail_if(dec->operands[0] == 0, "ArrayStride must be non-zero"); + type->stride = dec->operands[0]; + } } } 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; @@ -751,6 +755,7 @@ struct_member_decoration_cb(struct vtn_builder *b, switch (dec->decoration) { case SpvDecorationRelaxedPrecision: case SpvDecorationUniform: + case SpvDecorationUniformId: break; /* FIXME: Do nothing with this for now. */ case SpvDecorationNonWritable: vtn_handle_access_qualifier(b, ctx->type, member, ACCESS_NON_WRITEABLE); @@ -770,6 +775,9 @@ 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; @@ -777,8 +785,7 @@ 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]; @@ -829,7 +836,7 @@ 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: @@ -852,6 +859,7 @@ struct_member_decoration_cb(struct vtn_builder *b, break; case SpvDecorationUserSemantic: + case SpvDecorationUserTypeGOOGLE: /* User semantic decorations can safely be ignored by the driver. */ break; @@ -880,7 +888,7 @@ 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) { @@ -937,7 +945,7 @@ 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; @@ -975,11 +983,13 @@ type_decoration_cb(struct vtn_builder *b, case SpvDecorationPatch: case SpvDecorationCentroid: case SpvDecorationSample: + case SpvDecorationExplicitInterpAMD: case SpvDecorationVolatile: case SpvDecorationCoherent: case SpvDecorationNonWritable: case SpvDecorationNonReadable: case SpvDecorationUniform: + case SpvDecorationUniformId: case SpvDecorationLocation: case SpvDecorationComponent: case SpvDecorationOffset: @@ -1031,6 +1041,10 @@ 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); } @@ -1040,111 +1054,52 @@ 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); } } -static struct vtn_type * -vtn_type_layout_std430(struct vtn_builder *b, struct vtn_type *type, - uint32_t *size_out, uint32_t *align_out) -{ - switch (type->base_type) { - case vtn_base_type_scalar: { - uint32_t comp_size = glsl_type_is_boolean(type->type) - ? 4 : glsl_get_bit_size(type->type) / 8; - *size_out = comp_size; - *align_out = comp_size; - return type; - } - - case vtn_base_type_vector: { - uint32_t comp_size = glsl_type_is_boolean(type->type) - ? 4 : glsl_get_bit_size(type->type) / 8; - unsigned align_comps = type->length == 3 ? 4 : type->length; - *size_out = comp_size * type->length, - *align_out = comp_size * align_comps; - return type; - } - - case vtn_base_type_matrix: - case vtn_base_type_array: { - /* We're going to add an array stride */ - type = vtn_type_copy(b, type); - uint32_t elem_size, elem_align; - type->array_element = vtn_type_layout_std430(b, type->array_element, - &elem_size, &elem_align); - type->stride = vtn_align_u32(elem_size, elem_align); - *size_out = type->stride * type->length; - *align_out = elem_align; - return type; - } - - case vtn_base_type_struct: { - /* We're going to add member offsets */ - type = vtn_type_copy(b, type); - uint32_t offset = 0; - uint32_t align = 0; - for (unsigned i = 0; i < type->length; i++) { - uint32_t mem_size, mem_align; - type->members[i] = vtn_type_layout_std430(b, type->members[i], - &mem_size, &mem_align); - offset = vtn_align_u32(offset, mem_align); - type->offsets[i] = offset; - offset += mem_size; - align = MAX2(align, mem_align); - } - *size_out = offset; - *align_out = align; - return type; - } - - default: - unreachable("Invalid SPIR-V type for std430"); - } -} - static void vtn_handle_type(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) @@ -1266,8 +1221,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, /* A length of 0 is used to denote unsized arrays */ val->type->length = 0; } else { - val->type->length = - vtn_value(b, w[3], vtn_value_type_constant)->constant->values[0][0].u32; + val->type->length = vtn_constant_uint(b, w[3]); } val->type->base_type = vtn_base_type_array; @@ -1391,13 +1345,27 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, val->type->deref = vtn_value(b, w[3], vtn_value_type_type)->type; - vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL); + /* Only certain storage classes use ArrayStride. The others (in + * particular Workgroup) are expected to be laid out by the driver. + */ + switch (storage_class) { + case SpvStorageClassUniform: + case SpvStorageClassPushConstant: + case SpvStorageClassStorageBuffer: + case SpvStorageClassPhysicalStorageBuffer: + vtn_foreach_decoration(b, val, array_stride_decoration_cb, NULL); + break; + default: + /* Nothing to do. */ + break; + } if (b->physical_ptrs) { switch (storage_class) { case SpvStorageClassFunction: case SpvStorageClassWorkgroup: case SpvStorageClassCrossWorkgroup: + case SpvStorageClassUniformConstant: val->type->stride = align(glsl_get_cl_size(val->type->deref->type), glsl_get_cl_alignment(val->type->deref->type)); break; @@ -1405,15 +1373,6 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode, break; } } - - if (storage_class == SpvStorageClassWorkgroup && - b->options->lower_workgroup_access_to_offsets) { - uint32_t size, align; - val->type->deref = vtn_type_layout_std430(b, val->type->deref, - &size, &align); - val->type->length = size; - val->type->align = align; - } } break; } @@ -1538,7 +1497,7 @@ vtn_null_constant(struct vtn_builder *b, struct vtn_type *type) nir_address_format addr_format = vtn_mode_to_address_format(b, mode); const nir_const_value *null_value = nir_address_format_null_value(addr_format); - memcpy(c->values[0], null_value, + memcpy(c->values, null_value, sizeof(nir_const_value) * nir_address_format_num_components(addr_format)); break; } @@ -1577,55 +1536,29 @@ 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 || @@ -1651,63 +1584,45 @@ 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][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)); int bit_size = glsl_get_bit_size(val->type->type); switch (bit_size) { case 64: - val->constant->values[0][0].u64 = vtn_u64_literal(&w[3]); + val->constant->values[0].u64 = vtn_u64_literal(&w[3]); break; case 32: - val->constant->values[0][0].u32 = w[3]; + val->constant->values[0].u32 = w[3]; break; case 16: - val->constant->values[0][0].u16 = w[3]; + val->constant->values[0].u16 = w[3]; break; case 8: - val->constant->values[0][0].u8 = w[3]; + val->constant->values[0].u8 = w[3]; break; 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][0].u64 = - get_specialization64(b, val, vtn_u64_literal(&w[3])); - break; - case 32: - val->constant->values[0][0].u32 = get_specialization(b, val, w[3]); - break; - case 16: - val->constant->values[0][0].u16 = get_specialization(b, val, w[3]); - break; - case 8: - val->constant->values[0][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; } @@ -1737,20 +1652,11 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, case vtn_base_type_vector: { assert(glsl_type_is_vector(val->type->type)); for (unsigned i = 0; i < elem_count; i++) - val->constant->values[0][i] = elems[i]->values[0][0]; + val->constant->values[i] = elems[i]->values[0]; break; } case vtn_base_type_matrix: - assert(glsl_type_is_matrix(val->type->type)); - for (unsigned i = 0; i < elem_count; i++) { - unsigned components = - glsl_get_components(glsl_get_column_type(val->type->type)); - memcpy(val->constant->values[i], elems[i]->values, - sizeof(nir_const_value) * components); - } - break; - case vtn_base_type_struct: case vtn_base_type_array: ralloc_steal(val->constant, elems); @@ -1766,7 +1672,9 @@ 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]]; @@ -1789,48 +1697,30 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, vtn_assert(bit_size == bit_size0 && bit_size == bit_size1); (void)bit_size0; (void)bit_size1; - if (bit_size == 64) { - uint64_t u64[8]; - if (v0->value_type == vtn_value_type_constant) { - for (unsigned i = 0; i < len0; i++) - u64[i] = v0->constant->values[0][i].u64; - } - if (v1->value_type == vtn_value_type_constant) { - for (unsigned i = 0; i < len1; i++) - u64[len0 + i] = v1->constant->values[0][i].u64; - } + nir_const_value undef = { .u64 = 0xdeadbeefdeadbeef }; + nir_const_value combined[NIR_MAX_VEC_COMPONENTS * 2]; - for (unsigned i = 0, j = 0; i < count - 6; i++, j++) { - uint32_t comp = w[i + 6]; - /* If component is not used, set the value to a known constant - * to detect if it is wrongly used. - */ - if (comp == (uint32_t)-1) - val->constant->values[0][j].u64 = 0xdeadbeefdeadbeef; - else - val->constant->values[0][j].u64 = u64[comp]; - } - } else { - /* This is for both 32-bit and 16-bit values */ - uint32_t u32[8]; - if (v0->value_type == vtn_value_type_constant) { - for (unsigned i = 0; i < len0; i++) - u32[i] = v0->constant->values[0][i].u32; - } - if (v1->value_type == vtn_value_type_constant) { - for (unsigned i = 0; i < len1; i++) - u32[len0 + i] = v1->constant->values[0][i].u32; - } + if (v0->value_type == vtn_value_type_constant) { + for (unsigned i = 0; i < len0; i++) + combined[i] = v0->constant->values[i]; + } + if (v1->value_type == vtn_value_type_constant) { + for (unsigned i = 0; i < len1; i++) + combined[len0 + i] = v1->constant->values[i]; + } - for (unsigned i = 0, j = 0; i < count - 6; i++, j++) { - uint32_t comp = w[i + 6]; + for (unsigned i = 0, j = 0; i < count - 6; i++, j++) { + uint32_t comp = w[i + 6]; + if (comp == (uint32_t)-1) { /* If component is not used, set the value to a known constant * to detect if it is wrongly used. */ - if (comp == (uint32_t)-1) - val->constant->values[0][j].u32 = 0xdeadbeef; - else - val->constant->values[0][j].u32 = u32[comp]; + val->constant->values[j] = undef; + } else { + vtn_fail_if(comp >= len0 + len1, + "All Component literals must either be FFFFFFFF " + "or in [0, N - 1] (inclusive)."); + val->constant->values[j] = combined[comp]; } } break; @@ -1854,7 +1744,6 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, } int elem = -1; - int col = 0; const struct vtn_type *type = comp->type; for (unsigned i = deref_start; i < count; i++) { vtn_fail_if(w[i] > type->length, @@ -1869,12 +1758,6 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, break; case vtn_base_type_matrix: - assert(col == 0 && elem == -1); - col = w[i]; - elem = 0; - type = type->array_element; - break; - case vtn_base_type_array: c = &(*c)->elements[w[i]]; type = type->array_element; @@ -1897,7 +1780,7 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, } else { unsigned num_components = type->length; for (unsigned i = 0; i < num_components; i++) - val->constant->values[0][i] = (*c)->values[col][elem + i]; + val->constant->values[i] = (*c)->values[elem + i]; } } else { struct vtn_value *insert = @@ -1908,7 +1791,7 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, } else { unsigned num_components = type->length; for (unsigned i = 0; i < num_components; i++) - (*c)->values[col][elem + i] = insert->constant->values[0][i]; + (*c)->values[elem + i] = insert->constant->values[i]; } } break; @@ -1954,8 +1837,13 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, if (!nir_alu_type_get_type_size(nir_op_infos[op].input_types[i])) bit_size = glsl_get_bit_size(src_val->type->type); + unsigned src_comps = nir_op_infos[op].input_sizes[i] ? + nir_op_infos[op].input_sizes[i] : + num_components; + unsigned j = swap ? 1 - i : i; - memcpy(src[j], src_val->constant->values[0], sizeof(src[j])); + for (unsigned c = 0; c < src_comps; c++) + src[j][c] = src_val->constant->values[c]; } /* fix up fixed size sources */ @@ -1981,7 +1869,9 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode, nir_const_value *srcs[3] = { src[0], src[1], src[2], }; - nir_eval_const_opcode(op, val->constant->values[0], num_components, bit_size, srcs); + nir_eval_const_opcode(op, val->constant->values, + num_components, bit_size, srcs, + b->shader->info.float_controls_execution_mode); break; } /* default */ } @@ -2004,6 +1894,234 @@ 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 SpvStorageClassPhysicalStorageBuffer: + return SpvMemorySemanticsUniformMemoryMask; + case SpvStorageClassWorkgroup: + return SpvMemorySemanticsWorkgroupMemoryMask; + default: + return SpvMemorySemanticsMaskNone; + } +} + +static void +vtn_split_barrier_semantics(struct vtn_builder *b, + SpvMemorySemanticsMask semantics, + SpvMemorySemanticsMask *before, + SpvMemorySemanticsMask *after) +{ + /* For memory semantics embedded in operations, we split them into up to + * two barriers, to be added before and after the operation. This is less + * strict than if we propagated until the final backend stage, but still + * result in correct execution. + * + * A further improvement could be pipe this information (and use!) into the + * next compiler layers, at the expense of making the handling of barriers + * more complicated. + */ + + *before = SpvMemorySemanticsMaskNone; + *after = SpvMemorySemanticsMaskNone; + + SpvMemorySemanticsMask order_semantics = + semantics & (SpvMemorySemanticsAcquireMask | + SpvMemorySemanticsReleaseMask | + SpvMemorySemanticsAcquireReleaseMask | + SpvMemorySemanticsSequentiallyConsistentMask); + + if (util_bitcount(order_semantics) > 1) { + /* Old GLSLang versions incorrectly set all the ordering bits. This was + * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo, + * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016). + */ + vtn_warn("Multiple memory ordering semantics specified, " + "assuming AcquireRelease."); + order_semantics = SpvMemorySemanticsAcquireReleaseMask; + } + + const SpvMemorySemanticsMask av_vis_semantics = + semantics & (SpvMemorySemanticsMakeAvailableMask | + SpvMemorySemanticsMakeVisibleMask); + + const SpvMemorySemanticsMask storage_semantics = + semantics & (SpvMemorySemanticsUniformMemoryMask | + SpvMemorySemanticsSubgroupMemoryMask | + SpvMemorySemanticsWorkgroupMemoryMask | + SpvMemorySemanticsCrossWorkgroupMemoryMask | + SpvMemorySemanticsAtomicCounterMemoryMask | + SpvMemorySemanticsImageMemoryMask | + SpvMemorySemanticsOutputMemoryMask); + + const SpvMemorySemanticsMask other_semantics = + semantics & ~(order_semantics | av_vis_semantics | storage_semantics); + + if (other_semantics) + vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics); + + /* SequentiallyConsistent is treated as AcquireRelease. */ + + /* The RELEASE barrier happens BEFORE the operation, and it is usually + * associated with a Store. All the write operations with a matching + * semantics will not be reordered after the Store. + */ + if (order_semantics & (SpvMemorySemanticsReleaseMask | + SpvMemorySemanticsAcquireReleaseMask | + SpvMemorySemanticsSequentiallyConsistentMask)) { + *before |= SpvMemorySemanticsReleaseMask | storage_semantics; + } + + /* The ACQUIRE barrier happens AFTER the operation, and it is usually + * associated with a Load. All the operations with a matching semantics + * will not be reordered before the Load. + */ + if (order_semantics & (SpvMemorySemanticsAcquireMask | + SpvMemorySemanticsAcquireReleaseMask | + SpvMemorySemanticsSequentiallyConsistentMask)) { + *after |= SpvMemorySemanticsAcquireMask | storage_semantics; + } + + if (av_vis_semantics & SpvMemorySemanticsMakeVisibleMask) + *before |= SpvMemorySemanticsMakeVisibleMask | storage_semantics; + + if (av_vis_semantics & SpvMemorySemanticsMakeAvailableMask) + *after |= SpvMemorySemanticsMakeAvailableMask | storage_semantics; +} + +static void +vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope, + SpvMemorySemanticsMask semantics) +{ + nir_memory_semantics nir_semantics = 0; + + SpvMemorySemanticsMask order_semantics = + semantics & (SpvMemorySemanticsAcquireMask | + SpvMemorySemanticsReleaseMask | + SpvMemorySemanticsAcquireReleaseMask | + SpvMemorySemanticsSequentiallyConsistentMask); + + if (util_bitcount(order_semantics) > 1) { + /* Old GLSLang versions incorrectly set all the ordering bits. This was + * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo, + * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016). + */ + vtn_warn("Multiple memory ordering semantics bits specified, " + "assuming AcquireRelease."); + order_semantics = SpvMemorySemanticsAcquireReleaseMask; + } + + switch (order_semantics) { + case 0: + /* Not an ordering barrier. */ + break; + + case SpvMemorySemanticsAcquireMask: + nir_semantics = NIR_MEMORY_ACQUIRE; + break; + + case SpvMemorySemanticsReleaseMask: + nir_semantics = NIR_MEMORY_RELEASE; + break; + + case SpvMemorySemanticsSequentiallyConsistentMask: + /* Fall through. Treated as AcquireRelease in Vulkan. */ + case SpvMemorySemanticsAcquireReleaseMask: + nir_semantics = NIR_MEMORY_ACQUIRE | NIR_MEMORY_RELEASE; + break; + + default: + unreachable("Invalid memory order semantics"); + } + + if (semantics & SpvMemorySemanticsMakeAvailableMask) { + vtn_fail_if(!b->options->caps.vk_memory_model, + "To use MakeAvailable memory semantics the VulkanMemoryModel " + "capability must be declared."); + nir_semantics |= NIR_MEMORY_MAKE_AVAILABLE; + } + + if (semantics & SpvMemorySemanticsMakeVisibleMask) { + vtn_fail_if(!b->options->caps.vk_memory_model, + "To use MakeVisible memory semantics the VulkanMemoryModel " + "capability must be declared."); + nir_semantics |= NIR_MEMORY_MAKE_VISIBLE; + } + + /* Vulkan Environment for SPIR-V says "SubgroupMemory, CrossWorkgroupMemory, + * and AtomicCounterMemory are ignored". + */ + semantics &= ~(SpvMemorySemanticsSubgroupMemoryMask | + SpvMemorySemanticsCrossWorkgroupMemoryMask | + SpvMemorySemanticsAtomicCounterMemoryMask); + + /* TODO: Consider adding nir_var_mem_image mode to NIR so it can be used + * for SpvMemorySemanticsImageMemoryMask. + */ + + nir_variable_mode modes = 0; + if (semantics & (SpvMemorySemanticsUniformMemoryMask | + 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) { + modes |= nir_var_shader_out; + } + + /* No barrier to add. */ + if (nir_semantics == 0 || modes == 0) + return; + + nir_scope nir_scope; + switch (scope) { + case SpvScopeDevice: + vtn_fail_if(b->options->caps.vk_memory_model && + !b->options->caps.vk_memory_model_device_scope, + "If the Vulkan memory model is declared and any instruction " + "uses Device scope, the VulkanMemoryModelDeviceScope " + "capability must be declared."); + nir_scope = NIR_SCOPE_DEVICE; + break; + + case SpvScopeQueueFamily: + vtn_fail_if(!b->options->caps.vk_memory_model, + "To use Queue Family scope, the VulkanMemoryModel capability " + "must be declared."); + nir_scope = NIR_SCOPE_QUEUE_FAMILY; + break; + + case SpvScopeWorkgroup: + nir_scope = NIR_SCOPE_WORKGROUP; + break; + + case SpvScopeSubgroup: + nir_scope = NIR_SCOPE_SUBGROUP; + break; + + case SpvScopeInvocation: + nir_scope = NIR_SCOPE_INVOCATION; + break; + + default: + 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); + + nir_intrinsic_set_memory_modes(intrin, modes); + nir_intrinsic_set_memory_scope(intrin, nir_scope); + nir_builder_instr_insert(&b->nb, &intrin->instr); +} + struct vtn_ssa_value * vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type) { @@ -2058,6 +2176,42 @@ vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type) return src; } +static uint32_t +image_operand_arg(struct vtn_builder *b, const uint32_t *w, uint32_t count, + uint32_t mask_idx, SpvImageOperandsMask op) +{ + static const SpvImageOperandsMask ops_with_arg = + SpvImageOperandsBiasMask | + SpvImageOperandsLodMask | + SpvImageOperandsGradMask | + SpvImageOperandsConstOffsetMask | + SpvImageOperandsOffsetMask | + SpvImageOperandsConstOffsetsMask | + SpvImageOperandsSampleMask | + SpvImageOperandsMinLodMask | + SpvImageOperandsMakeTexelAvailableMask | + SpvImageOperandsMakeTexelVisibleMask; + + assert(util_bitcount(op) == 1); + assert(w[mask_idx] & op); + assert(op & ops_with_arg); + + uint32_t idx = util_bitcount(w[mask_idx] & (op - 1) & ops_with_arg) + 1; + + /* Adjust indices for operands with two arguments. */ + static const SpvImageOperandsMask ops_with_two_args = + SpvImageOperandsGradMask; + idx += util_bitcount(w[mask_idx] & (op - 1) & ops_with_two_args); + + idx += mask_idx; + + vtn_fail_if(idx + (op & ops_with_two_args ? 1 : 0) >= count, + "Image op claims to have %s but does not enough " + "following operands", spirv_imageoperands_to_string(op)); + + return idx; +} + static void vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) @@ -2066,42 +2220,42 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, 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->type = - vtn_value(b, w[1], vtn_value_type_type)->type; 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; return; } else if (opcode == SpvOpImage) { - struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer); struct vtn_value *src_val = vtn_untyped_value(b, w[3]); if (src_val->value_type == vtn_value_type_sampled_image) { - val->pointer = src_val->sampled_image->image; + vtn_push_value_pointer(b, w[2], src_val->sampled_image->image); } else { vtn_assert(src_val->value_type == vtn_value_type_pointer); - val->pointer = src_val->pointer; + vtn_push_value_pointer(b, w[2], src_val->pointer); } return; } struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type; - struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); - struct vtn_sampled_image sampled; + struct vtn_pointer *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) { - sampled = *sampled_val->sampled_image; + image = sampled_val->sampled_image->image; + sampler = sampled_val->sampled_image->sampler; } else { vtn_assert(sampled_val->value_type == vtn_value_type_pointer); - sampled.type = sampled_val->pointer->type; - sampled.image = NULL; - sampled.sampler = sampled_val->pointer; + image = sampled_val->pointer; } - const struct glsl_type *image_type = sampled.type->type; + 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); + nir_alu_type dest_type = nir_type_invalid; /* Figure out the base texture operation */ nir_texop texop; @@ -2121,7 +2275,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, break; case SpvOpImageFetch: - if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) { + if (sampler_dim == GLSL_SAMPLER_DIM_MS) { texop = nir_texop_txf_ms; } else { texop = nir_texop_txf; @@ -2136,18 +2290,30 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, case SpvOpImageQuerySizeLod: case SpvOpImageQuerySize: texop = nir_texop_txs; + dest_type = nir_type_int; break; case SpvOpImageQueryLod: texop = nir_texop_lod; + dest_type = nir_type_float; break; case SpvOpImageQueryLevels: texop = nir_texop_query_levels; + dest_type = nir_type_int; break; case SpvOpImageQuerySamples: texop = nir_texop_texture_samples; + dest_type = nir_type_int; + break; + + case SpvOpFragmentFetchAMD: + texop = nir_texop_fragment_fetch; + break; + + case SpvOpFragmentMaskFetchAMD: + texop = nir_texop_fragment_mask_fetch; break; default: @@ -2157,11 +2323,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, nir_tex_src srcs[10]; /* 10 should be enough */ nir_tex_src *p = srcs; - nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler); - nir_deref_instr *texture = - sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler; - - p->src = nir_src_for_ssa(&texture->dest.ssa); + p->src = nir_src_for_ssa(&image_deref->dest.ssa); p->src_type = nir_tex_src_texture_deref; p++; @@ -2172,8 +2334,10 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, case nir_texop_txd: case nir_texop_tg4: case nir_texop_lod: - /* These operations require a sampler */ - p->src = nir_src_for_ssa(&sampler->dest.ssa); + 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_type = nir_tex_src_sampler_deref; p++; break; @@ -2183,6 +2347,8 @@ 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: @@ -2190,6 +2356,8 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, break; case nir_texop_txf_ms_mcs: vtn_fail("unexpected nir_texop_txf_ms_mcs"); + case nir_texop_tex_prefetch: + vtn_fail("unexpected nir_texop_tex_prefetch"); } unsigned idx = 4; @@ -2208,25 +2376,11 @@ 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++; @@ -2275,8 +2429,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, case SpvOpImageGather: /* This has a component as its next source */ - gather_component = - vtn_value(b, w[idx++], vtn_value_type_constant)->constant->values[0][0].u32; + gather_component = vtn_constant_uint(b, w[idx++]); break; default: @@ -2287,54 +2440,84 @@ 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++]; + uint32_t operands = w[idx]; if (operands & SpvImageOperandsBiasMask) { - vtn_assert(texop == nir_texop_tex); - texop = nir_texop_txb; - (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias); + 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); } if (operands & SpvImageOperandsLodMask) { vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf || - texop == nir_texop_txs); - (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod); + 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); } if (operands & SpvImageOperandsGradMask) { vtn_assert(texop == nir_texop_txl); texop = nir_texop_txd; - (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx); - (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy); + uint32_t arg = image_operand_arg(b, w, count, idx, + SpvImageOperandsGradMask); + (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ddx); + (*p++) = vtn_tex_src(b, w[arg + 1], nir_tex_src_ddy); } - if (operands & SpvImageOperandsOffsetMask || - operands & SpvImageOperandsConstOffsetMask) - (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset); + vtn_fail_if(util_bitcount(operands & (SpvImageOperandsConstOffsetsMask | + SpvImageOperandsOffsetMask | + SpvImageOperandsConstOffsetMask)) > 1, + "At most one of the ConstOffset, Offset, and ConstOffsets " + "image operands can be used on a given instruction."); + + if (operands & SpvImageOperandsOffsetMask) { + uint32_t arg = image_operand_arg(b, w, count, idx, + SpvImageOperandsOffsetMask); + (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset); + } + + if (operands & SpvImageOperandsConstOffsetMask) { + uint32_t arg = image_operand_arg(b, w, count, idx, + SpvImageOperandsConstOffsetMask); + (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset); + } if (operands & SpvImageOperandsConstOffsetsMask) { vtn_assert(texop == nir_texop_tg4); - gather_offsets = vtn_value(b, w[idx++], vtn_value_type_constant); + uint32_t arg = image_operand_arg(b, w, count, idx, + SpvImageOperandsConstOffsetsMask); + gather_offsets = vtn_value(b, w[arg], vtn_value_type_constant); } if (operands & SpvImageOperandsSampleMask) { vtn_assert(texop == nir_texop_txf_ms); + uint32_t arg = image_operand_arg(b, w, count, idx, + SpvImageOperandsSampleMask); texop = nir_texop_txf_ms; - (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index); + (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ms_index); } if (operands & SpvImageOperandsMinLodMask) { vtn_assert(texop == nir_texop_tex || texop == nir_texop_txb || texop == nir_texop_txd); - (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_min_lod); + uint32_t arg = image_operand_arg(b, w, count, idx, + SpvImageOperandsMinLodMask); + (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_min_lod); } } - /* We should have now consumed exactly all of the arguments */ - vtn_assert(idx == count); nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs); instr->op = texop; @@ -2349,21 +2532,26 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, is_shadow && glsl_get_components(ret_type->type) == 1; instr->component = gather_component; - if (sampled.image && (sampled.image->access & ACCESS_NON_UNIFORM)) + if (image && (image->access & ACCESS_NON_UNIFORM)) instr->texture_non_uniform = true; - if (sampled.sampler && (sampled.sampler->access & ACCESS_NON_UNIFORM)) + if (sampler && (sampler->access & ACCESS_NON_UNIFORM)) instr->sampler_non_uniform = true; - switch (glsl_get_sampler_result_type(image_type)) { - case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break; - case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break; - case GLSL_TYPE_UINT: instr->dest_type = nir_type_uint; break; - case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break; - default: - vtn_fail("Invalid base type for sampler result"); + /* for non-query ops, get dest_type from sampler type */ + if (dest_type == nir_type_invalid) { + switch (glsl_get_sampler_result_type(image_type)) { + 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; + case GLSL_TYPE_BOOL: dest_type = nir_type_bool; break; + default: + vtn_fail("Invalid base type for sampler result"); + } } + instr->dest_type = dest_type; + nir_ssa_dest_init(&instr->instr, &instr->dest, nir_tex_instr_dest_size(instr), 32, NULL); @@ -2386,7 +2574,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, unsigned bit_size = glsl_get_bit_size(vec_type->type); for (uint32_t i = 0; i < 4; i++) { const nir_const_value *cvec = - gather_offsets->constant->elements[i]->values[0]; + gather_offsets->constant->elements[i]->values; for (uint32_t j = 0; j < 2; j++) { switch (bit_size) { case 8: instr->tg4_offsets[i][j] = cvec[j].i8; break; @@ -2400,8 +2588,9 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode, } } - val->ssa = vtn_create_ssa_value(b, ret_type->type); - val->ssa->def = &instr->dest.ssa; + 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); } @@ -2486,10 +2675,13 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, val->image->image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; val->image->coord = get_image_coord(b, w[4]); val->image->sample = vtn_ssa_value(b, w[5])->def; + val->image->lod = nir_imm_int(&b->nb, 0); return; } struct vtn_image_pointer image; + SpvScope scope = SpvScopeInvocation; + SpvMemorySemanticsMask semantics = 0; switch (opcode) { case SpvOpAtomicExchange: @@ -2508,43 +2700,98 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicOr: case SpvOpAtomicXor: image = *vtn_value(b, w[3], vtn_value_type_image_pointer)->image; + scope = vtn_constant_uint(b, w[4]); + semantics = vtn_constant_uint(b, w[5]); break; case SpvOpAtomicStore: image = *vtn_value(b, w[1], vtn_value_type_image_pointer)->image; + scope = vtn_constant_uint(b, w[2]); + semantics = vtn_constant_uint(b, w[3]); break; case SpvOpImageQuerySize: image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; image.coord = NULL; image.sample = NULL; + image.lod = NULL; break; - case SpvOpImageRead: + case SpvOpImageRead: { image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; image.coord = get_image_coord(b, w[4]); - if (count > 5 && (w[5] & SpvImageOperandsSampleMask)) { - vtn_assert(w[5] == SpvImageOperandsSampleMask); - image.sample = vtn_ssa_value(b, w[6])->def; + const SpvImageOperandsMask operands = + count > 5 ? w[5] : SpvImageOperandsMaskNone; + + if (operands & SpvImageOperandsSampleMask) { + uint32_t arg = image_operand_arg(b, w, count, 5, + SpvImageOperandsSampleMask); + image.sample = vtn_ssa_value(b, w[arg])->def; } else { image.sample = nir_ssa_undef(&b->nb, 1, 32); } + + if (operands & SpvImageOperandsMakeTexelVisibleMask) { + vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0, + "MakeTexelVisible requires NonPrivateTexel to also be set."); + uint32_t arg = image_operand_arg(b, w, count, 5, + SpvImageOperandsMakeTexelVisibleMask); + semantics = SpvMemorySemanticsMakeVisibleMask; + scope = vtn_constant_uint(b, w[arg]); + } + + if (operands & SpvImageOperandsLodMask) { + uint32_t arg = image_operand_arg(b, w, count, 5, + SpvImageOperandsLodMask); + image.lod = vtn_ssa_value(b, w[arg])->def; + } else { + image.lod = nir_imm_int(&b->nb, 0); + } + + /* TODO: Volatile. */ + break; + } - case SpvOpImageWrite: + case SpvOpImageWrite: { image.image = vtn_value(b, w[1], vtn_value_type_pointer)->pointer; image.coord = get_image_coord(b, w[2]); /* texel = w[3] */ - if (count > 4 && (w[4] & SpvImageOperandsSampleMask)) { - vtn_assert(w[4] == SpvImageOperandsSampleMask); - image.sample = vtn_ssa_value(b, w[5])->def; + const SpvImageOperandsMask operands = + count > 4 ? w[4] : SpvImageOperandsMaskNone; + + if (operands & SpvImageOperandsSampleMask) { + uint32_t arg = image_operand_arg(b, w, count, 4, + SpvImageOperandsSampleMask); + image.sample = vtn_ssa_value(b, w[arg])->def; } else { image.sample = nir_ssa_undef(&b->nb, 1, 32); } + + if (operands & SpvImageOperandsMakeTexelAvailableMask) { + vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0, + "MakeTexelAvailable requires NonPrivateTexel to also be set."); + uint32_t arg = image_operand_arg(b, w, count, 4, + SpvImageOperandsMakeTexelAvailableMask); + semantics = SpvMemorySemanticsMakeAvailableMask; + scope = vtn_constant_uint(b, w[arg]); + } + + if (operands & SpvImageOperandsLodMask) { + uint32_t arg = image_operand_arg(b, w, count, 4, + SpvImageOperandsLodMask); + image.lod = vtn_ssa_value(b, w[arg])->def; + } else { + image.lod = nir_imm_int(&b->nb, 0); + } + + /* TODO: Volatile. */ + break; + } default: vtn_fail_with_opcode("Invalid image opcode", opcode); @@ -2565,10 +2812,10 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, OP(AtomicIDecrement, atomic_add) OP(AtomicIAdd, atomic_add) OP(AtomicISub, atomic_add) - OP(AtomicSMin, atomic_min) - OP(AtomicUMin, atomic_min) - OP(AtomicSMax, atomic_max) - OP(AtomicUMax, atomic_max) + OP(AtomicSMin, atomic_imin) + OP(AtomicUMin, atomic_umin) + OP(AtomicSMax, atomic_imax) + OP(AtomicUMax, atomic_umax) OP(AtomicAnd, atomic_and) OP(AtomicOr, atomic_or) OP(AtomicXor, atomic_xor) @@ -2591,10 +2838,20 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, intrin->src[2] = nir_src_for_ssa(image.sample); } + nir_intrinsic_set_access(intrin, image.image->access); + switch (opcode) { case SpvOpAtomicLoad: case SpvOpImageQuerySize: case SpvOpImageRead: + if (opcode == SpvOpImageRead || opcode == SpvOpAtomicLoad) { + /* 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: { @@ -2604,6 +2861,12 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, assert(op == nir_intrinsic_image_deref_store); intrin->num_components = 4; intrin->src[3] = nir_src_for_ssa(expand_to_vec4(&b->nb, value)); + /* 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); break; } @@ -2628,8 +2891,17 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, vtn_fail_with_opcode("Invalid image opcode", opcode); } + /* Image operations implicitly have the Image storage memory semantics. */ + semantics |= SpvMemorySemanticsImageMemoryMask; + + SpvMemorySemanticsMask before_semantics; + SpvMemorySemanticsMask after_semantics; + vtn_split_barrier_semantics(b, semantics, &before_semantics, &after_semantics); + + if (before_semantics) + vtn_emit_memory_barrier(b, scope, before_semantics); + if (opcode != SpvOpImageWrite && opcode != SpvOpAtomicStore) { - struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; unsigned dest_components = glsl_get_vector_elements(type->type); @@ -2646,11 +2918,15 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, if (intrin->num_components != dest_components) result = nir_channels(&b->nb, result, (1 << dest_components) - 1); - val->ssa = vtn_create_ssa_value(b, type->type); + struct vtn_value *val = + vtn_push_ssa(b, w[2], type, vtn_create_ssa_value(b, type->type)); val->ssa->def = result; } else { nir_builder_instr_insert(&b->nb, &intrin->instr); } + + if (after_semantics) + vtn_emit_memory_barrier(b, scope, after_semantics); } static nir_intrinsic_op @@ -2706,34 +2982,7 @@ 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"); - } -} - -static nir_intrinsic_op -get_shared_nir_atomic_op(struct vtn_builder *b, SpvOp opcode) -{ - switch (opcode) { - case SpvOpAtomicLoad: return nir_intrinsic_load_shared; - case SpvOpAtomicStore: return nir_intrinsic_store_shared; -#define OP(S, N) case SpvOp##S: return nir_intrinsic_shared_##N; - OP(AtomicExchange, atomic_exchange) - OP(AtomicCompareExchange, atomic_comp_swap) - OP(AtomicCompareExchangeWeak, atomic_comp_swap) - OP(AtomicIIncrement, atomic_add) - OP(AtomicIDecrement, atomic_add) - OP(AtomicIAdd, atomic_add) - OP(AtomicISub, atomic_add) - OP(AtomicSMin, atomic_imin) - OP(AtomicUMin, atomic_umin) - OP(AtomicSMax, atomic_imax) - OP(AtomicUMax, atomic_umax) - OP(AtomicAnd, atomic_and) - OP(AtomicOr, atomic_or) - OP(AtomicXor, atomic_xor) -#undef OP - default: - vtn_fail_with_opcode("Invalid shared atomic", opcode); + vtn_fail("Invalid uniform atomic"); } } @@ -2769,11 +3018,14 @@ 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; + switch (opcode) { case SpvOpAtomicLoad: case SpvOpAtomicExchange: @@ -2791,21 +3043,20 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, case SpvOpAtomicOr: case SpvOpAtomicXor: 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]); break; case SpvOpAtomicStore: ptr = vtn_value(b, w[1], vtn_value_type_pointer)->pointer; + scope = vtn_constant_uint(b, w[2]); + semantics = vtn_constant_uint(b, w[3]); break; default: vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode); } - /* - SpvScope scope = w[4]; - SpvMemorySemanticsMask semantics = w[5]; - */ - /* uniform as "atomic counter uniform" */ if (ptr->mode == vtn_variable_mode_uniform) { nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr); @@ -2856,15 +3107,9 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, nir_ssa_def *offset, *index; offset = vtn_pointer_to_offset(b, ptr, &index); - nir_intrinsic_op op; - if (ptr->mode == vtn_variable_mode_ssbo) { - op = get_ssbo_nir_atomic_op(b, opcode); - } else { - vtn_assert(ptr->mode == vtn_variable_mode_workgroup && - b->options->lower_workgroup_access_to_offsets); - op = get_shared_nir_atomic_op(b, opcode); - } + assert(ptr->mode == vtn_variable_mode_ssbo); + nir_intrinsic_op op = get_ssbo_nir_atomic_op(b, opcode); atomic = nir_intrinsic_instr_create(b->nb.shader, op); int src = 0; @@ -2950,6 +3195,18 @@ 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); + + SpvMemorySemanticsMask before_semantics; + SpvMemorySemanticsMask after_semantics; + vtn_split_barrier_semantics(b, semantics, &before_semantics, &after_semantics); + + if (before_semantics) + 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; @@ -2957,13 +3214,16 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode, glsl_get_vector_elements(type->type), glsl_get_bit_size(type->type), NULL); - struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); - val->ssa = rzalloc(b, struct vtn_ssa_value); - val->ssa->def = &atomic->dest.ssa; - val->ssa->type = type->type; + 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); } nir_builder_instr_insert(&b->nb, &atomic->instr); + + if (after_semantics) + vtn_emit_memory_barrier(b, scope, after_semantics); } static nir_alu_instr * @@ -3008,58 +3268,6 @@ 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, @@ -3154,17 +3362,29 @@ 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; } @@ -3179,6 +3399,9 @@ 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. @@ -3186,9 +3409,11 @@ vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src, 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]); + 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]]; } } @@ -3200,66 +3425,68 @@ static void vtn_handle_composite(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) { - struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); - const struct glsl_type *type = - vtn_value(b, w[1], vtn_value_type_type)->type->type; - val->ssa = vtn_create_ssa_value(b, type); + struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; + struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, type->type); switch (opcode) { case SpvOpVectorExtractDynamic: - val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def, - vtn_ssa_value(b, w[4])->def); + ssa->def = nir_vector_extract(&b->nb, vtn_ssa_value(b, w[3])->def, + vtn_ssa_value(b, w[4])->def); break; case SpvOpVectorInsertDynamic: - val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def, - vtn_ssa_value(b, w[4])->def, - vtn_ssa_value(b, w[5])->def); + ssa->def = nir_vector_insert(&b->nb, vtn_ssa_value(b, w[3])->def, + vtn_ssa_value(b, w[4])->def, + vtn_ssa_value(b, w[5])->def); break; case SpvOpVectorShuffle: - val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type), - vtn_ssa_value(b, w[3])->def, - vtn_ssa_value(b, w[4])->def, - w + 5); + 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, + w + 5); break; case SpvOpCompositeConstruct: { unsigned elems = count - 3; assume(elems >= 1); - if (glsl_type_is_vector_or_scalar(type)) { + 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; - val->ssa->def = - vtn_vector_construct(b, glsl_get_vector_elements(type), + ssa->def = + vtn_vector_construct(b, glsl_get_vector_elements(type->type), elems, srcs); } else { - val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems); + ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems); for (unsigned i = 0; i < elems; i++) - val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]); + ssa->elems[i] = vtn_ssa_value(b, w[3 + i]); } break; } case SpvOpCompositeExtract: - val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]), - w + 4, count - 4); + ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]), + w + 4, count - 4); break; case SpvOpCompositeInsert: - val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]), - vtn_ssa_value(b, w[3]), - w + 5, count - 5); + ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]), + vtn_ssa_value(b, w[3]), + w + 5, count - 5); break; case SpvOpCopyLogical: - case SpvOpCopyObject: - val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3])); + 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); } static void @@ -3269,15 +3496,21 @@ vtn_emit_barrier(struct vtn_builder *b, nir_intrinsic_op op) nir_builder_instr_insert(&b->nb, &intrin->instr); } -static void +void vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope, SpvMemorySemanticsMask semantics) { + if (b->shader->options->use_scoped_memory_barrier) { + vtn_emit_scoped_memory_barrier(b, scope, semantics); + return; + } + static const SpvMemorySemanticsMask all_memory_semantics = SpvMemorySemanticsUniformMemoryMask | SpvMemorySemanticsWorkgroupMemoryMask | SpvMemorySemanticsAtomicCounterMemoryMask | - SpvMemorySemanticsImageMemoryMask; + SpvMemorySemanticsImageMemoryMask | + SpvMemorySemanticsOutputMemoryMask; /* If we're not actually doing a memory barrier, bail */ if (!(semantics & all_memory_semantics)) @@ -3297,9 +3530,14 @@ 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) { - vtn_emit_barrier(b, nir_intrinsic_memory_barrier); - return; + /* Map the GLSL memoryBarrier() construct to the corresponding NIR one. */ + static const SpvMemorySemanticsMask glsl_memory_barrier = + SpvMemorySemanticsUniformMemoryMask | + SpvMemorySemanticsWorkgroupMemoryMask | + SpvMemorySemanticsImageMemoryMask; + if ((semantics & glsl_memory_barrier) == glsl_memory_barrier) { + vtn_emit_barrier(b, nir_intrinsic_memory_barrier); + semantics &= ~(glsl_memory_barrier | SpvMemorySemanticsAtomicCounterMemoryMask); } /* Issue a bunch of more specific barriers */ @@ -3319,6 +3557,10 @@ vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope, 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;; } @@ -3327,7 +3569,7 @@ vtn_emit_memory_barrier(struct vtn_builder *b, SpvScope scope, 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: @@ -3376,12 +3618,46 @@ vtn_handle_barrier(struct vtn_builder *b, SpvOp opcode, case SpvOpControlBarrier: { SpvScope execution_scope = vtn_constant_uint(b, w[1]); - if (execution_scope == SpvScopeWorkgroup) - vtn_emit_barrier(b, nir_intrinsic_barrier); - SpvScope memory_scope = vtn_constant_uint(b, w[2]); SpvMemorySemanticsMask memory_semantics = vtn_constant_uint(b, w[3]); + + /* 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; + } + vtn_emit_memory_barrier(b, memory_scope, memory_semantics); + + if (execution_scope == SpvScopeWorkgroup) + vtn_emit_barrier(b, nir_intrinsic_control_barrier); break; } @@ -3551,10 +3827,10 @@ 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", @@ -3610,7 +3886,6 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, case SpvCapabilityImageReadWrite: case SpvCapabilityImageMipmap: case SpvCapabilityPipes: - case SpvCapabilityGroups: case SpvCapabilityDeviceEnqueue: case SpvCapabilityLiteralSampler: case SpvCapabilityGenericPointer: @@ -3651,6 +3926,7 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, spv_check_supported(subgroup_basic, cap); break; + case SpvCapabilitySubgroupVoteKHR: case SpvCapabilityGroupNonUniformVote: spv_check_supported(subgroup_vote, cap); break; @@ -3674,6 +3950,10 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, spv_check_supported(subgroup_arithmetic, cap); break; + case SpvCapabilityGroups: + spv_check_supported(amd_shader_ballot, cap); + break; + case SpvCapabilityVariablePointersStorageBuffer: case SpvCapabilityVariablePointers: spv_check_supported(variable_pointers, cap); @@ -3687,6 +3967,8 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, spv_check_supported(storage_16bit, cap); break; + case SpvCapabilityShaderLayer: + case SpvCapabilityShaderViewportIndex: case SpvCapabilityShaderViewportIndexLayerEXT: spv_check_supported(shader_viewport_index_layer, cap); break; @@ -3729,7 +4011,15 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, spv_check_supported(post_depth_coverage, cap); break; - case SpvCapabilityPhysicalStorageBufferAddressesEXT: + case SpvCapabilityDenormFlushToZero: + case SpvCapabilityDenormPreserve: + case SpvCapabilitySignedZeroInfNanPreserve: + case SpvCapabilityRoundingModeRTE: + case SpvCapabilityRoundingModeRTZ: + spv_check_supported(float_controls, cap); + break; + + case SpvCapabilityPhysicalStorageBufferAddresses: spv_check_supported(physical_storage_buffer_address, cap); break; @@ -3742,6 +4032,46 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, spv_check_supported(float16, cap); break; + case SpvCapabilityFragmentShaderSampleInterlockEXT: + spv_check_supported(fragment_shader_sample_interlock, cap); + break; + + case SpvCapabilityFragmentShaderPixelInterlockEXT: + spv_check_supported(fragment_shader_pixel_interlock, cap); + break; + + case SpvCapabilityDemoteToHelperInvocationEXT: + spv_check_supported(demote_to_helper_invocation, cap); + break; + + case SpvCapabilityShaderClockKHR: + spv_check_supported(shader_clock, cap); + break; + + case SpvCapabilityVulkanMemoryModel: + spv_check_supported(vk_memory_model, cap); + break; + + case SpvCapabilityVulkanMemoryModelDeviceScope: + 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; + default: vtn_fail("Unhandled capability: %s (%u)", spirv_capability_to_string(cap), cap); @@ -3774,15 +4104,14 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, b->options->temp_addr_format = nir_address_format_64bit_global; 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->shader->info.cs.ptr_size = 0; 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)", @@ -3790,9 +4119,20 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, break; } - vtn_assert(w[2] == SpvMemoryModelSimple || - w[2] == SpvMemoryModelGLSL450 || - w[2] == SpvMemoryModelOpenCL); + switch (w[2]) { + case SpvMemoryModelSimple: + case SpvMemoryModelGLSL450: + case SpvMemoryModelOpenCL: + break; + case SpvMemoryModelVulkan: + vtn_fail_if(!b->options->caps.vk_memory_model, + "Vulkan memory model is unsupported by this driver"); + break; + default: + vtn_fail("Unsupported memory model: %s", + spirv_memorymodel_to_string(w[2])); + break; + } break; case SpvOpEntryPoint: @@ -3825,6 +4165,17 @@ 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 */ } @@ -3834,7 +4185,7 @@ 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); @@ -3998,6 +4349,34 @@ vtn_handle_execution_mode(struct vtn_builder *b, struct vtn_value *entry_point, b->shader->info.cs.derivative_group = DERIVATIVE_GROUP_LINEAR; break; + case SpvExecutionModePixelInterlockOrderedEXT: + vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT); + b->shader->info.fs.pixel_interlock_ordered = true; + break; + + case SpvExecutionModePixelInterlockUnorderedEXT: + vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT); + b->shader->info.fs.pixel_interlock_unordered = true; + break; + + case SpvExecutionModeSampleInterlockOrderedEXT: + vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT); + b->shader->info.fs.sample_interlock_ordered = true; + break; + + case SpvExecutionModeSampleInterlockUnorderedEXT: + vtn_assert(b->shader->info.stage == MESA_SHADER_FRAGMENT); + b->shader->info.fs.sample_interlock_unordered = true; + break; + + case SpvExecutionModeDenormPreserve: + case SpvExecutionModeDenormFlushToZero: + case SpvExecutionModeSignedZeroInfNanPreserve: + case SpvExecutionModeRoundingModeRTE: + case SpvExecutionModeRoundingModeRTZ: + /* Already handled in vtn_handle_rounding_mode_in_execution_mode() */ + break; + default: vtn_fail("Unhandled execution mode: %s (%u)", spirv_executionmode_to_string(mode->exec_mode), @@ -4005,6 +4384,63 @@ 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_assert(b->entry_point == entry_point); + + 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; +} + static bool vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) @@ -4078,6 +4514,14 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode, 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 */ } @@ -4173,8 +4617,9 @@ vtn_handle_ptr(struct vtn_builder *b, SpvOp opcode, "%s operands must have the same storage class", spirv_op_to_string(opcode)); - const struct glsl_type *type = - vtn_value(b, w[1], vtn_value_type_type)->type->type; + struct vtn_type *vtn_type = + vtn_value(b, w[1], vtn_value_type_type)->type; + 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)); @@ -4212,9 +4657,9 @@ vtn_handle_ptr(struct vtn_builder *b, SpvOp opcode, unreachable("Invalid ptr operation"); } - struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); - val->ssa = vtn_create_ssa_value(b, type); - val->ssa->def = def; + 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); } static bool @@ -4297,6 +4742,11 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, break; } + case SpvOpFragmentMaskFetchAMD: + case SpvOpFragmentFetchAMD: + vtn_handle_texture(b, opcode, w, count); + break; + case SpvOpAtomicLoad: case SpvOpAtomicExchange: case SpvOpAtomicCompareExchange: @@ -4433,6 +4883,20 @@ 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; @@ -4494,6 +4958,31 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, case SpvOpGroupNonUniformLogicalXor: case SpvOpGroupNonUniformQuadBroadcast: case SpvOpGroupNonUniformQuadSwap: + case SpvOpGroupAll: + case SpvOpGroupAny: + case SpvOpGroupBroadcast: + case SpvOpGroupIAdd: + case SpvOpGroupFAdd: + case SpvOpGroupFMin: + case SpvOpGroupUMin: + case SpvOpGroupSMin: + case SpvOpGroupFMax: + case SpvOpGroupUMax: + case SpvOpGroupSMax: + case SpvOpSubgroupBallotKHR: + case SpvOpSubgroupFirstInvocationKHR: + case SpvOpSubgroupReadInvocationKHR: + case SpvOpSubgroupAllKHR: + case SpvOpSubgroupAnyKHR: + case SpvOpSubgroupAllEqualKHR: + case SpvOpGroupIAddNonUniformAMD: + case SpvOpGroupFAddNonUniformAMD: + case SpvOpGroupFMinNonUniformAMD: + case SpvOpGroupUMinNonUniformAMD: + case SpvOpGroupSMinNonUniformAMD: + case SpvOpGroupFMaxNonUniformAMD: + case SpvOpGroupUMaxNonUniformAMD: + case SpvOpGroupSMaxNonUniformAMD: vtn_handle_subgroup(b, opcode, w, count); break; @@ -4503,6 +4992,84 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode, vtn_handle_ptr(b, opcode, w, count); break; + case SpvOpBeginInvocationInterlockEXT: + vtn_emit_barrier(b, nir_intrinsic_begin_invocation_interlock); + break; + + case SpvOpEndInvocationInterlockEXT: + vtn_emit_barrier(b, nir_intrinsic_end_invocation_interlock); + break; + + case SpvOpDemoteToHelperInvocationEXT: { + nir_intrinsic_instr *intrin = + nir_intrinsic_instr_create(b->shader, nir_intrinsic_demote); + nir_builder_instr_insert(&b->nb, &intrin->instr); + break; + } + + case SpvOpIsHelperInvocationEXT: { + nir_intrinsic_instr *intrin = + nir_intrinsic_instr_create(b->shader, nir_intrinsic_is_helper_invocation); + 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); + break; + } + + case SpvOpReadClockKHR: { + 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. + */ + 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; + const struct glsl_type *dest_type = type->type; + nir_ssa_def *result; + + if (glsl_type_is_vector(dest_type)) { + assert(dest_type == glsl_vector_type(GLSL_TYPE_UINT, 2)); + result = &intrin->dest.ssa; + } else { + assert(glsl_type_is_scalar(dest_type)); + assert(glsl_get_base_type(dest_type) == GLSL_TYPE_UINT64); + 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; + break; + } + + case SpvOpLifetimeStart: + case SpvOpLifetimeStop: + break; + default: vtn_fail_with_opcode("Unhandled opcode", opcode); } @@ -4526,7 +5093,7 @@ 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; @@ -4556,6 +5123,13 @@ vtn_create_builder(const uint32_t *words, size_t word_count, */ b->wa_glslang_179 = (generator_id == 8 && generator_version == 1); + /* 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_cs_barrier = (generator_id == 8 && generator_version < 3); + /* words[2] == generator magic */ unsigned value_id_bound = words[3]; if (words[4] != 0) { @@ -4673,7 +5247,15 @@ spirv_to_nir(const uint32_t *words, size_t word_count, } /* Set shader info defaults */ - b->shader->info.gs.invocations = 1; + 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. + */ + vtn_foreach_execution_mode(b, b->entry_point, + vtn_handle_rounding_mode_in_execution_mode, NULL); b->specializations = spec; b->num_specializations = num_spec; @@ -4691,7 +5273,7 @@ spirv_to_nir(const uint32_t *words, size_t word_count, glsl_vector_type(GLSL_TYPE_UINT, 3)); nir_const_value *const_size = - b->workgroup_size_builtin->constant->values[0]; + b->workgroup_size_builtin->constant->values; b->shader->info.cs.local_size[0] = const_size[0].u32; b->shader->info.cs.local_size[1] = const_size[1].u32; @@ -4709,7 +5291,8 @@ 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); @@ -4739,7 +5322,7 @@ 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);