X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fcompiler%2Fspirv%2Fspirv_to_nir.c;h=6ae1ea81188217f28e4b4d3e6cf35981c323d509;hb=6a154aea0d3375aa8469f28bb8a85e5ee79eef4a;hp=a4b09a404df13bdbb8576955829789a528ea60d1;hpb=37bfd854c71be8ace37080fbca36d0a1ef68fb8a;p=mesa.git diff --git a/src/compiler/spirv/spirv_to_nir.c b/src/compiler/spirv/spirv_to_nir.c index a4b09a404df..6ae1ea81188 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) { @@ -398,6 +391,9 @@ 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) { @@ -746,7 +742,7 @@ 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; @@ -779,6 +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; @@ -786,8 +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]; @@ -838,7 +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: @@ -889,7 +887,7 @@ 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) { @@ -946,7 +944,7 @@ 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; @@ -984,6 +982,7 @@ type_decoration_cb(struct vtn_builder *b, case SpvDecorationPatch: case SpvDecorationCentroid: case SpvDecorationSample: + case SpvDecorationExplicitInterpAMD: case SpvDecorationVolatile: case SpvDecorationCoherent: case SpvDecorationNonWritable: @@ -1050,46 +1049,46 @@ 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); @@ -1348,7 +1347,7 @@ 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: @@ -1532,55 +1531,29 @@ 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 || @@ -1606,18 +1579,21 @@ 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)); @@ -1638,31 +1614,10 @@ 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; } @@ -1712,7 +1667,9 @@ 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]]; @@ -1937,7 +1894,7 @@ vtn_storage_class_to_memory_semantics(SpvStorageClass sc) { switch (sc) { case SpvStorageClassStorageBuffer: - case SpvStorageClassPhysicalStorageBufferEXT: + case SpvStorageClassPhysicalStorageBuffer: return SpvMemorySemanticsUniformMemoryMask; case SpvStorageClassWorkgroup: return SpvMemorySemanticsWorkgroupMemoryMask; @@ -2101,14 +2058,15 @@ 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; } @@ -2345,6 +2303,14 @@ 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); } @@ -2376,6 +2342,8 @@ 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: @@ -2403,25 +2371,11 @@ 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++; @@ -2481,6 +2435,10 @@ 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) { @@ -2710,6 +2668,7 @@ 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; } @@ -2748,6 +2707,7 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode, image.image = vtn_value(b, w[3], vtn_value_type_pointer)->pointer; image.coord = NULL; image.sample = NULL; + image.lod = NULL; break; case SpvOpImageRead: { @@ -2774,6 +2734,14 @@ 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, 5, + SpvImageOperandsLodMask); + image.lod = vtn_ssa_value(b, w[arg])->def; + } else { + image.lod = nir_imm_int(&b->nb, 0); + } + /* TODO: Volatile. */ break; @@ -2805,6 +2773,14 @@ 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_ssa_value(b, w[arg])->def; + } else { + image.lod = nir_imm_int(&b->nb, 0); + } + /* TODO: Volatile. */ break; @@ -2861,6 +2837,14 @@ vtn_handle_image(struct vtn_builder *b, SpvOp 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: { @@ -2870,6 +2854,12 @@ 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; } @@ -2985,7 +2975,7 @@ 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"); } } @@ -3021,7 +3011,7 @@ 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; @@ -3271,58 +3261,6 @@ 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, @@ -3417,17 +3355,29 @@ 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; } @@ -3442,6 +3392,9 @@ 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. @@ -3449,9 +3402,11 @@ 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]]; } } @@ -3468,14 +3423,14 @@ vtn_handle_composite(struct vtn_builder *b, SpvOp opcode, 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_ssa_value(b, w[3])->def, + vtn_ssa_value(b, w[4])->def); 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_ssa_value(b, w[3])->def, + vtn_ssa_value(b, w[4])->def, + vtn_ssa_value(b, w[5])->def); break; case SpvOpVectorShuffle: @@ -3514,9 +3469,11 @@ 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); @@ -3536,7 +3493,7 @@ 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_memory_barrier) { vtn_emit_scoped_memory_barrier(b, scope, semantics); return; } @@ -3545,7 +3502,8 @@ 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)) @@ -3565,9 +3523,14 @@ 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 */ @@ -3587,6 +3550,10 @@ 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;; } @@ -3595,7 +3562,7 @@ 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: @@ -3643,13 +3610,47 @@ 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]); + + /* 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); - SpvScope execution_scope = vtn_constant_uint(b, w[1]); if (execution_scope == SpvScopeWorkgroup) - vtn_emit_barrier(b, nir_intrinsic_barrier); + vtn_emit_barrier(b, nir_intrinsic_control_barrier); break; } @@ -4011,7 +4012,7 @@ 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; @@ -4052,6 +4053,14 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode, 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; + default: vtn_fail("Unhandled capability: %s (%u)", spirv_capability_to_string(cap), cap); @@ -4084,14 +4093,14 @@ 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->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)", @@ -4165,7 +4174,7 @@ 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); @@ -4722,6 +4731,11 @@ 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: @@ -4858,6 +4872,20 @@ 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; @@ -5041,7 +5069,7 @@ 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; @@ -5071,6 +5099,13 @@ 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) { @@ -5232,7 +5267,8 @@ 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); @@ -5262,7 +5298,7 @@ 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);