add support for image1d_buffer (#609)

* WIP texel buffers

Change-Id: I4eef91af393b5487bbc208bdc96d43daa4c5c462

* add support for image1d_buffer

remove 1DBUFFER mention in init_vulkan_image
limit CL_DEVICE_IMAGE_MAX_BUFFER_SIZE WITH CL_DEVICE_MAX_MEM_ALLOC_SIZE

* kpet feedback

* fix formatting

* update clspv

* use bigger padding in test

---------

Co-authored-by: Kévin Petit <kpet@free.fr>

src/device.cpp

@@ -1042,6 +1042,7 @@ bool cvk_device::supports_capability(spv::Capability capability) const {
  case spv::CapabilityImage1D:
  case spv::CapabilityImageQuery:
  case spv::CapabilityImageBuffer:
+ case spv::CapabilitySampledBuffer:
  return true;
  // Optional capabilities:
  case spv::CapabilityFloat16:

src/device.hpp

@@ -248,6 +248,11 @@ struct cvk_device : public _cl_device_id,
  return maxAllocSz;
  }
 
+ size_t image_max_buffer_size() const {
+ return std::min((uint64_t)vulkan_limits().maxTexelBufferElements,
+ max_mem_alloc_size());
+ }
+
  cl_uint mem_base_addr_align() const {
  // The OpenCL spec requires at least 1024 bits (long16's alignment)
  uint32_t required_by_vulkan_impl =

src/kernel.cpp

@@ -119,7 +119,9 @@ cl_int cvk_kernel::set_arg(cl_uint index, size_t size, const void* value) {
  // if the argument is an image, we need to set its metadata
  // (channel_order/channel_data_type).
  if (arg.kind == kernel_argument_kind::sampled_image ||
- arg.kind == kernel_argument_kind::storage_image) {
+ arg.kind == kernel_argument_kind::storage_image ||
+ arg.kind == kernel_argument_kind::storage_texel_buffer ||
+ arg.kind == kernel_argument_kind::uniform_texel_buffer) {
  set_image_metadata(index, value);
  }
 
@@ -155,9 +157,11 @@ bool cvk_kernel_argument_values::setup_descriptor_sets() {
  std::vector<VkWriteDescriptorSet> descriptor_writes;
  std::vector<VkDescriptorBufferInfo> buffer_info;
  std::vector<VkDescriptorImageInfo> image_info;
+ std::vector<VkBufferView> buffer_views;
  descriptor_writes.reserve(max_descriptor_writes);
  buffer_info.reserve(max_descriptor_writes);
  image_info.reserve(max_descriptor_writes);
+ buffer_views.reserve(max_descriptor_writes);
 
  // Setup module-scope variables
  if (program->module_constant_data_buffer() != nullptr &&
@@ -324,6 +328,36 @@ bool cvk_kernel_argument_values::setup_descriptor_sets() {
  descriptor_writes.push_back(writeDescriptorSet);
  break;
  }
+ case kernel_argument_kind::storage_texel_buffer:
+ case kernel_argument_kind::uniform_texel_buffer: {
+ auto image = static_cast<cvk_image*>(get_arg_value(arg));
+ bool uniform =
+ arg.kind == kernel_argument_kind::uniform_texel_buffer;
+ auto view = image->vulkan_buffer_view();
+ buffer_views.push_back(view);
+
+ cvk_debug_fn("buffer view %p @ set = %u, binding = %u", view,
+ arg.descriptorSet, arg.binding);
+
+ VkDescriptorType dtype =
+ uniform ? VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER
+ : VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
+
+ VkWriteDescriptorSet writeDescriptorSet = {
+ VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
+ nullptr,
+ ds[arg.descriptorSet],
+ arg.binding, // dstBinding
+ 0, // dstArrayElement
+ 1, // descriptorCount
+ dtype,
+ nullptr, // pImageInfo
+ nullptr, // pBufferInfo
+ &buffer_views.back(), // pTexelBufferView
+ };
+ descriptor_writes.push_back(writeDescriptorSet);
+ break;
+ }
  case kernel_argument_kind::pod: // skip POD arguments
  case kernel_argument_kind::pod_ubo:
  case kernel_argument_kind::pod_pushconstant:

src/memory.cpp

@@ -243,22 +243,33 @@ bool cvk_sampler::init() {
 VkFormatFeatureFlags
 cvk_image::required_format_feature_flags_for(cl_mem_object_type type,
  cl_mem_flags flags) {
- UNUSED(type); // TODO will be required for 1D buffer images
- // All images require TRANSFER_SRC, TRANSFER_DST
+ // 1Dbuffer requires
+ // RW / RaW: STORAGE_TEXEL_BUFFER
+ // RO: UNIFORM_TEXEL_BUFFER
+ // All other images require TRANSFER_SRC, TRANSFER_DST
  // read-only: SAMPLED_IMAGE, SAMPLED_IMAGE_FILTER_LINEAR
  // write-only: STORAGE_IMAGE
  // read-write: STORAGE_IMAGE, SAMPLED_IMAGE, SAMPLED_IMAGE_FILTER_LINEAR
  // read-and-write: STORAGE_IMAGE
  VkFormatFeatureFlags format_feature_flags = 0;
- format_feature_flags =
- VK_FORMAT_FEATURE_TRANSFER_SRC_BIT | VK_FORMAT_FEATURE_TRANSFER_DST_BIT;
-
+ if (type != CL_MEM_OBJECT_IMAGE1D_BUFFER) {
+ format_feature_flags = VK_FORMAT_FEATURE_TRANSFER_SRC_BIT |
+ VK_FORMAT_FEATURE_TRANSFER_DST_BIT;
+ }
  VkFormatFeatureFlags format_feature_flags_RO;
- format_feature_flags_RO = VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
- VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
-
+ if (type == CL_MEM_OBJECT_IMAGE1D_BUFFER) {
+ format_feature_flags_RO = VK_FORMAT_FEATURE_UNIFORM_TEXEL_BUFFER_BIT;
+ } else {
+ format_feature_flags_RO =
+ VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
+ VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
+ }
  VkFormatFeatureFlags format_feature_flags_WO;
- format_feature_flags_WO = VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;
+ if (type == CL_MEM_OBJECT_IMAGE1D_BUFFER) {
+ format_feature_flags_WO = VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_BIT;
+ } else {
+ format_feature_flags_WO = VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT;
+ }
 
  if (flags & (CL_MEM_KERNEL_READ_AND_WRITE | CL_MEM_WRITE_ONLY)) {
  format_feature_flags |= format_feature_flags_WO;
@@ -292,7 +303,7 @@ cl_image_format_to_vulkan_format(const cl_image_format& clfmt,
  VkComponentMapping* components_sampled,
  VkComponentMapping* components_storage);
 
-bool cvk_image::init() {
+bool cvk_image::init_vulkan_image() {
  // Translate image type and size
  VkImageType image_type;
  VkImageViewType view_type;
@@ -320,7 +331,6 @@ bool cvk_image::init() {
  size_t host_ptr_size = 0;
 
  switch (m_desc.image_type) {
- case CL_MEM_OBJECT_IMAGE1D_BUFFER:
  case CL_MEM_OBJECT_IMAGE1D:
  image_type = VK_IMAGE_TYPE_1D;
  view_type = VK_IMAGE_VIEW_TYPE_1D;
@@ -397,29 +407,24 @@ bool cvk_image::init() {
  return false;
  }
 
- if (m_desc.image_type == CL_MEM_OBJECT_IMAGE1D_BUFFER) {
- auto buffer = static_cast<cvk_mem*>(m_desc.buffer);
- m_memory = buffer->memory();
- buffer->retain();
- } else {
- // Select memory type
- cvk_device::allocation_parameters params =
- device->select_memory_for(m_image);
- if (params.memory_type_index == VK_MAX_MEMORY_TYPES) {
- cvk_error_fn("Could not get memory type!");
- return false;
- }
+ CVK_ASSERT(m_desc.image_type != CL_MEM_OBJECT_IMAGE1D_BUFFER);
+ // Select memory type
+ cvk_device::allocation_parameters params =
+ device->select_memory_for(m_image);
+ if (params.memory_type_index == VK_MAX_MEMORY_TYPES) {
+ cvk_error_fn("Could not get memory type!");
+ return false;
+ }
 
-  // Allocate memory
-  m_memory = std::make_unique<cvk_memory_allocation>(
-  vkdev, params.size, params.memory_type_index);
+ // Allocate memory
+ m_memory = std::make_unique<cvk_memory_allocation>(
+ vkdev, params.size, params.memory_type_index);
 
-  res = m_memory->allocate(device->uses_physical_addressing());
+ res = m_memory->allocate(device->uses_physical_addressing());
 
- if (res != VK_SUCCESS) {
- cvk_error_fn("Could not allocate memory!");
- return false;
- }
+ if (res != VK_SUCCESS) {
+ cvk_error_fn("Could not allocate memory!");
+ return false;
  }
 
  // Bind the image to memory
@@ -485,6 +490,56 @@ bool cvk_image::init() {
  return true;
 }
 
+bool cvk_image::init_vulkan_texel_buffer() {
+ VkResult res;
+
+ auto device = m_context->device();
+ auto vkdev = device->vulkan_device();
+
+ VkFormat format;
+ VkComponentMapping components_sampled, components_storage;
+
+ auto success = cl_image_format_to_vulkan_format(
+ m_format, &format, &components_sampled, &components_storage);
+ if (!success) {
+ return false;
+ }
+
+ CVK_ASSERT(buffer());
+ CVK_ASSERT(buffer()->is_buffer_type());
+
+ auto vkbuf = static_cast<cvk_buffer*>(buffer())->vulkan_buffer();
+ auto offset = static_cast<cvk_buffer*>(buffer())->vulkan_buffer_offset();
+
+ VkBufferViewCreateInfo createInfo = {
+ VK_STRUCTURE_TYPE_BUFFER_VIEW_CREATE_INFO,
+ nullptr,
+ 0, // flags
+ vkbuf, // buffer
+ format, // format
+ offset, // offset
+ VK_WHOLE_SIZE // range
+ };
+
+ res = vkCreateBufferView(vkdev, &createInfo, nullptr, &m_buffer_view);
+ if (res != VK_SUCCESS) {
+ cvk_error_fn("Could not create buffer view");
+ return false;
+ }
+
+ buffer()->retain();
+
+ return true;
+}
+
+bool cvk_image::init() {
+ if (is_backed_by_buffer_view()) {
+ return init_vulkan_texel_buffer();
+ } else {
+ return init_vulkan_image();
+ }
+}
+
 void cvk_image::prepare_fill_pattern(const void* input_pattern,
  fill_pattern_array& pattern,
  size_t* size_ret) const {

src/memory.hpp

@@ -314,10 +314,13 @@ struct cvk_buffer : public cvk_mem {
  cvk_mem* create_subbuffer(cl_mem_flags, size_t origin, size_t size);
 
  VkBufferUsageFlags prepare_usage_flags() {
- VkBufferUsageFlags usage_flags = VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
- VK_BUFFER_USAGE_TRANSFER_DST_BIT |
- VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
- VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+ VkBufferUsageFlags usage_flags =
+ VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
+ VK_BUFFER_USAGE_TRANSFER_DST_BIT |
+ VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT |
+ VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
+ VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT |
+ VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT;
  if (m_context->device()->uses_physical_addressing()) {
  usage_flags |= VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT;
  }
@@ -485,10 +488,16 @@ struct cvk_image : public cvk_mem {
  /* FIXME parent_offset */ 0, std::move(properties),
  desc->image_type),
  m_desc(*desc), m_format(*format), m_image(VK_NULL_HANDLE),
- m_sampled_view(VK_NULL_HANDLE), m_storage_view(VK_NULL_HANDLE) {
+ m_sampled_view(VK_NULL_HANDLE), m_storage_view(VK_NULL_HANDLE),
+ m_buffer_view(VK_NULL_HANDLE) {
  // All images require asynchronous initialiation for the initial
- // layout transition (and copy/use host ptr init)
- m_init_tracker.set_state(cvk_mem_init_state::required);
+ // layout transition (and copy/use host ptr init) apart from
+ // those backed by a texel buffer
+ if (is_backed_by_buffer_view()) {
+ m_init_tracker.set_state(cvk_mem_init_state::completed);
+ } else {
+ m_init_tracker.set_state(cvk_mem_init_state::required);
+ }
  }
 
  ~cvk_image() {
@@ -502,6 +511,9 @@ struct cvk_image : public cvk_mem {
  if (m_storage_view != VK_NULL_HANDLE) {
  vkDestroyImageView(vkdev, m_storage_view, nullptr);
  }
+ if (m_buffer_view != VK_NULL_HANDLE) {
+ vkDestroyBufferView(vkdev, m_buffer_view, nullptr);
+ }
  if (buffer() != nullptr) {
  buffer()->release();
  }
@@ -530,9 +542,26 @@ struct cvk_image : public cvk_mem {
  const cl_image_format* format, void* host_ptr,
  std::vector<cl_mem_properties>&& properties);
 
- VkImage vulkan_image() const { return m_image; }
- VkImageView vulkan_sampled_view() const { return m_sampled_view; }
- VkImageView vulkan_storage_view() const { return m_storage_view; }
+ bool is_backed_by_buffer_view() const {
+ return type() == CL_MEM_OBJECT_IMAGE1D_BUFFER;
+ }
+
+ VkImage vulkan_image() const {
+ CVK_ASSERT(!is_backed_by_buffer_view());
+ return m_image;
+ }
+ VkImageView vulkan_sampled_view() const {
+ CVK_ASSERT(!is_backed_by_buffer_view());
+ return m_sampled_view;
+ }
+ VkImageView vulkan_storage_view() const {
+ CVK_ASSERT(!is_backed_by_buffer_view());
+ return m_storage_view;
+ }
+ VkBufferView vulkan_buffer_view() const {
+ CVK_ASSERT(is_backed_by_buffer_view());
+ return m_buffer_view;
+ }
  const cl_image_format& format() const { return m_format; }
  size_t element_size() const {
  switch (m_format.image_channel_data_type) {
@@ -693,6 +722,8 @@ struct cvk_image : public cvk_mem {
  size_t* size_ret) const;
 
 private:
+ bool init_vulkan_image();
+ bool init_vulkan_texel_buffer();
  bool init();
 
  size_t num_channels() const {
@@ -746,6 +777,7 @@ struct cvk_image : public cvk_mem {
  VkImage m_image;
  VkImageView m_sampled_view;
  VkImageView m_storage_view;
+ VkBufferView m_buffer_view;
  std::unordered_map<void*, std::list<cvk_image_mapping>> m_mappings;
  std::mutex m_mappings_lock;
  std::unique_ptr<cvk_buffer> m_init_data;

src/program.cpp

@@ -117,6 +117,10 @@ spv_result_t parse_reflection(void* user_data,
  return kernel_argument_kind::sampled_image;
  case NonSemanticClspvReflectionArgumentStorageImage:
  return kernel_argument_kind::storage_image;
+ case NonSemanticClspvReflectionArgumentStorageTexelBuffer:
+ return kernel_argument_kind::storage_texel_buffer;
+ case NonSemanticClspvReflectionArgumentUniformTexelBuffer:
+ return kernel_argument_kind::uniform_texel_buffer;
  case NonSemanticClspvReflectionArgumentSampler:
  return kernel_argument_kind::sampler;
  case NonSemanticClspvReflectionArgumentWorkgroup:
@@ -234,6 +238,8 @@ spv_result_t parse_reflection(void* user_data,
  case NonSemanticClspvReflectionArgumentUniform:
  case NonSemanticClspvReflectionArgumentSampledImage:
  case NonSemanticClspvReflectionArgumentStorageImage:
+ case NonSemanticClspvReflectionArgumentStorageTexelBuffer:
+ case NonSemanticClspvReflectionArgumentUniformTexelBuffer:
  case NonSemanticClspvReflectionArgumentSampler: {
  // These arguments have descriptor set, binding and an optional
  // arg info.
@@ -1743,6 +1749,12 @@ bool cvk_entry_point::build_descriptor_sets_layout_bindings_for_arguments(
  case kernel_argument_kind::storage_image:
  dt = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
  break;
+ case kernel_argument_kind::storage_texel_buffer:
+ dt = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
+ break;
+ case kernel_argument_kind::uniform_texel_buffer:
+ dt = VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER;
+ break;
  case kernel_argument_kind::sampler:
  dt = VK_DESCRIPTOR_TYPE_SAMPLER;
  break;

src/program.hpp

@@ -47,6 +47,8 @@ enum class kernel_argument_kind
  pointer_pushconstant,
  sampled_image,
  storage_image,
+ storage_texel_buffer,
+ uniform_texel_buffer,
  sampler,
  local,
  unused,
@@ -101,7 +103,9 @@ struct kernel_argument {
  return (kind == kernel_argument_kind::buffer) ||
  (kind == kernel_argument_kind::buffer_ubo) ||
  (kind == kernel_argument_kind::sampled_image) ||
- (kind == kernel_argument_kind::storage_image);
+ (kind == kernel_argument_kind::storage_image) ||
+ (kind == kernel_argument_kind::storage_texel_buffer) ||
+ (kind == kernel_argument_kind::uniform_texel_buffer);
  }
 
  bool is_unused() const { return kind == kernel_argument_kind::unused; }