A simple Vulkan compute sample

VkComputeSample

// This is free and unencumbered software released into the public domain.
//
// Anyone is free to copy, modify, publish, use, compile, sell, or
// distribute this software, either in source code form or as a compiled
// binary, for any purpose, commercial or non-commercial, and by any
// means.
//
// In jurisdictions that recognize copyright laws, the author or authors
// of this software dedicate any and all copyright interest in the
// software to the public domain. We make this dedication for the benefit
// of the public at large and to the detriment of our heirs and
// successors. We intend this dedication to be an overt act of
// relinquishment in perpetuity of all present and future rights to this
// software under copyright law.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
// IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
// OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
// OTHER DEALINGS IN THE SOFTWARE.
//
// For more information, please refer to <http://unlicense.org/>

#include "vulkan.h"

#include <stdio.h>
#include <stdlib.h>

#define BAIL_ON_BAD_RESULT(result) \
  if (VK_SUCCESS != (result)) { fprintf(stderr, "Failure at %u %s\n", __LINE__, __FILE__); exit(-1); }

VkResult vkGetBestTransferQueueNPH(VkPhysicalDevice physicalDevice, uint32_t* queueFamilyIndex) {
  uint32_t queueFamilyPropertiesCount = 0;
  vkGetPhysicalDeviceQueueFamilyProperties(physicalDevice, &queueFamilyPropertiesCount, 0);

  VkQueueFamilyProperties* const queueFamilyProperties = (VkQueueFamilyProperties*)_alloca(
    sizeof(VkQueueFamilyProperties) * queueFamilyPropertiesCount);

  vkGetPhysicalDeviceQueueFamilyProperties(physicalDevice, &queueFamilyPropertiesCount, queueFamilyProperties);

  // first try and find a queue that has just the transfer bit set
  for (uint32_t i = 0; i < queueFamilyPropertiesCount; i++) {
    // mask out the sparse binding bit that we aren't caring about (yet!)
    const VkQueueFlags maskedFlags = (~VK_QUEUE_SPARSE_BINDING_BIT & queueFamilyProperties[i].queueFlags);

    if (!((VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT) & maskedFlags) &&
      (VK_QUEUE_TRANSFER_BIT & maskedFlags)) {
      *queueFamilyIndex = i;
      return VK_SUCCESS;
    }
  }

  // otherwise we'll prefer using a compute-only queue,
  // remember that having compute on the queue implicitly enables transfer!
  for (uint32_t i = 0; i < queueFamilyPropertiesCount; i++) {
    // mask out the sparse binding bit that we aren't caring about (yet!)
    const VkQueueFlags maskedFlags = (~VK_QUEUE_SPARSE_BINDING_BIT & queueFamilyProperties[i].queueFlags);

    if (!(VK_QUEUE_GRAPHICS_BIT & maskedFlags) && (VK_QUEUE_COMPUTE_BIT & maskedFlags)) {
      *queueFamilyIndex = i;
      return VK_SUCCESS;
    }
  }

  // lastly get any queue that'll work for us (graphics, compute or transfer bit set)
  for (uint32_t i = 0; i < queueFamilyPropertiesCount; i++) {
    // mask out the sparse binding bit that we aren't caring about (yet!)
    const VkQueueFlags maskedFlags = (~VK_QUEUE_SPARSE_BINDING_BIT & queueFamilyProperties[i].queueFlags);

    if ((VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT) & maskedFlags) {
      *queueFamilyIndex = i;
      return VK_SUCCESS;
    }
  }

  return VK_ERROR_INITIALIZATION_FAILED;
}

VkResult vkGetBestComputeQueueNPH(VkPhysicalDevice physicalDevice, uint32_t* queueFamilyIndex) {
  uint32_t queueFamilyPropertiesCount = 0;
  vkGetPhysicalDeviceQueueFamilyProperties(physicalDevice, &queueFamilyPropertiesCount, 0);

  VkQueueFamilyProperties* const queueFamilyProperties = (VkQueueFamilyProperties*)_alloca(
    sizeof(VkQueueFamilyProperties) * queueFamilyPropertiesCount);

  vkGetPhysicalDeviceQueueFamilyProperties(physicalDevice, &queueFamilyPropertiesCount, queueFamilyProperties);

  // first try and find a queue that has just the compute bit set
  for (uint32_t i = 0; i < queueFamilyPropertiesCount; i++) {
    // mask out the sparse binding bit that we aren't caring about (yet!) and the transfer bit
    const VkQueueFlags maskedFlags = (~(VK_QUEUE_TRANSFER_BIT | VK_QUEUE_SPARSE_BINDING_BIT) &
      queueFamilyProperties[i].queueFlags);

    if (!(VK_QUEUE_GRAPHICS_BIT & maskedFlags) && (VK_QUEUE_COMPUTE_BIT & maskedFlags)) {
      *queueFamilyIndex = i;
      return VK_SUCCESS;
    }
  }

  // lastly get any queue that'll work for us
  for (uint32_t i = 0; i < queueFamilyPropertiesCount; i++) {
    // mask out the sparse binding bit that we aren't caring about (yet!) and the transfer bit
    const VkQueueFlags maskedFlags = (~(VK_QUEUE_TRANSFER_BIT | VK_QUEUE_SPARSE_BINDING_BIT) &
      queueFamilyProperties[i].queueFlags);

    if (VK_QUEUE_COMPUTE_BIT & maskedFlags) {
      *queueFamilyIndex = i;
      return VK_SUCCESS;
    }
  }

  return VK_ERROR_INITIALIZATION_FAILED;
}

int main(int argc, const char * const argv[]) {
  (void)argc;
  (void)argv;

  const VkApplicationInfo applicationInfo = {
    VK_STRUCTURE_TYPE_APPLICATION_INFO,
    0,
    "VKComputeSample",
    0,
    "",
    0,
    VK_MAKE_VERSION(1, 0, 9)
  };
  
  const VkInstanceCreateInfo instanceCreateInfo = {
    VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
    0,
    0,
    &applicationInfo,
    0,
    0,
    0,
    0
  };
  
  VkInstance instance;
  BAIL_ON_BAD_RESULT(vkCreateInstance(&instanceCreateInfo, 0, &instance));
  
  uint32_t physicalDeviceCount = 0;
  BAIL_ON_BAD_RESULT(vkEnumeratePhysicalDevices(instance, &physicalDeviceCount, 0));

  VkPhysicalDevice* const physicalDevices = (VkPhysicalDevice*)malloc(
    sizeof(VkPhysicalDevice) * physicalDeviceCount);

  BAIL_ON_BAD_RESULT(vkEnumeratePhysicalDevices(instance, &physicalDeviceCount, physicalDevices));

  for (uint32_t i = 0; i < physicalDeviceCount; i++) {
    uint32_t queueFamilyIndex = 0;
    BAIL_ON_BAD_RESULT(vkGetBestComputeQueueNPH(physicalDevices[i], &queueFamilyIndex));

    const float queuePrioritory = 1.0f;
    const VkDeviceQueueCreateInfo deviceQueueCreateInfo = {
      VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
      0,
      0,
      queueFamilyIndex,
      1,
      &queuePrioritory
    };

    const VkDeviceCreateInfo deviceCreateInfo = {
      VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO,
      0,
      0,
      1,
      &deviceQueueCreateInfo,
      0,
      0,
      0,
      0,
      0
    };

    VkDevice device;
    BAIL_ON_BAD_RESULT(vkCreateDevice(physicalDevices[i], &deviceCreateInfo, 0, &device));

    VkPhysicalDeviceMemoryProperties properties;

    vkGetPhysicalDeviceMemoryProperties(physicalDevices[i], &properties);

    const int32_t bufferLength = 16384;

    const uint32_t bufferSize = sizeof(int32_t) * bufferLength;

    // we are going to need two buffers from this one memory
    const VkDeviceSize memorySize = bufferSize * 2; 

    // set memoryTypeIndex to an invalid entry in the properties.memoryTypes array
    uint32_t memoryTypeIndex = VK_MAX_MEMORY_TYPES;

    for (uint32_t k = 0; k < properties.memoryTypeCount; k++) {
      if ((VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT & properties.memoryTypes[k].propertyFlags) &&
        (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT & properties.memoryTypes[k].propertyFlags) &&
        (memorySize < properties.memoryHeaps[properties.memoryTypes[k].heapIndex].size)) {
        memoryTypeIndex = k;
        break;
      }
    }

    BAIL_ON_BAD_RESULT(memoryTypeIndex == VK_MAX_MEMORY_TYPES ? VK_ERROR_OUT_OF_HOST_MEMORY : VK_SUCCESS);

    const VkMemoryAllocateInfo memoryAllocateInfo = {
      VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
      0,
      memorySize,
      memoryTypeIndex
    };

    VkDeviceMemory memory;
    BAIL_ON_BAD_RESULT(vkAllocateMemory(device, &memoryAllocateInfo, 0, &memory));

    int32_t *payload;
    BAIL_ON_BAD_RESULT(vkMapMemory(device, memory, 0, memorySize, 0, (void *)&payload));

    for (uint32_t k = 1; k < memorySize / sizeof(int32_t); k++) {
      payload[k] = rand();
    }

    vkUnmapMemory(device, memory);

    const VkBufferCreateInfo bufferCreateInfo = {
      VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
      0,
      0,
      bufferSize,
      VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
      VK_SHARING_MODE_EXCLUSIVE,
      1,
      &queueFamilyIndex
    };

    VkBuffer in_buffer;
    BAIL_ON_BAD_RESULT(vkCreateBuffer(device, &bufferCreateInfo, 0, &in_buffer));

    BAIL_ON_BAD_RESULT(vkBindBufferMemory(device, in_buffer, memory, 0));

    VkBuffer out_buffer;
    BAIL_ON_BAD_RESULT(vkCreateBuffer(device, &bufferCreateInfo, 0, &out_buffer));

    BAIL_ON_BAD_RESULT(vkBindBufferMemory(device, out_buffer, memory, bufferSize));

    enum {
      RESERVED_ID = 0,
      FUNC_ID,
      IN_ID,
      OUT_ID,
      GLOBAL_INVOCATION_ID,
      VOID_TYPE_ID,
      FUNC_TYPE_ID,
      INT_TYPE_ID,
      INT_ARRAY_TYPE_ID,
      STRUCT_ID,
      POINTER_TYPE_ID,
      ELEMENT_POINTER_TYPE_ID,
      INT_VECTOR_TYPE_ID,
      INT_VECTOR_POINTER_TYPE_ID,
      INT_POINTER_TYPE_ID,
      CONSTANT_ZERO_ID,
      CONSTANT_ARRAY_LENGTH_ID,
      LABEL_ID,
      IN_ELEMENT_ID,
      OUT_ELEMENT_ID,
      GLOBAL_INVOCATION_X_ID,
      GLOBAL_INVOCATION_X_PTR_ID,
      TEMP_LOADED_ID,
      BOUND
    };

    enum {
      INPUT = 1,
      UNIFORM = 2,
      BUFFER_BLOCK = 3,
      ARRAY_STRIDE = 6,
      BUILTIN = 11,
      BINDING = 33,
      OFFSET = 35,
      DESCRIPTOR_SET = 34,
      GLOBAL_INVOCATION = 28,
      OP_TYPE_VOID = 19,
      OP_TYPE_FUNCTION = 33,
      OP_TYPE_INT = 21,
      OP_TYPE_VECTOR = 23,
      OP_TYPE_ARRAY = 28,
      OP_TYPE_STRUCT = 30,
      OP_TYPE_POINTER = 32,
      OP_VARIABLE = 59,
      OP_DECORATE = 71,
      OP_MEMBER_DECORATE = 72,
      OP_FUNCTION = 54,
      OP_LABEL = 248,
      OP_ACCESS_CHAIN = 65,
      OP_CONSTANT = 43,
      OP_LOAD = 61,
      OP_STORE = 62,
      OP_RETURN = 253,
      OP_FUNCTION_END = 56,
      OP_CAPABILITY = 17,
      OP_MEMORY_MODEL = 14,
      OP_ENTRY_POINT = 15,
      OP_EXECUTION_MODE = 16,
      OP_COMPOSITE_EXTRACT = 81,
    };

    int32_t shader[] = {
      // first is the SPIR-V header
      0x07230203, // magic header ID
      0x00010000, // version 1.0.0
      0,          // generator (optional)
      BOUND,      // bound
      0,          // schema

      // OpCapability Shader
      (2 << 16) | OP_CAPABILITY, 1, 

      // OpMemoryModel Logical Simple
      (3 << 16) | OP_MEMORY_MODEL, 0, 0, 

      // OpEntryPoint GLCompute %FUNC_ID "f" %IN_ID %OUT_ID
      (4 << 16) | OP_ENTRY_POINT, 5, FUNC_ID, 0x00000066,

      // OpExecutionMode %FUNC_ID LocalSize 1 1 1
      (6 << 16) | OP_EXECUTION_MODE, FUNC_ID, 17, 1, 1, 1,

      // next declare decorations

      (3 << 16) | OP_DECORATE, STRUCT_ID, BUFFER_BLOCK, 

      (4 << 16) | OP_DECORATE, GLOBAL_INVOCATION_ID, BUILTIN, GLOBAL_INVOCATION,

      (4 << 16) | OP_DECORATE, IN_ID, DESCRIPTOR_SET, 0,

      (4 << 16) | OP_DECORATE, IN_ID, BINDING, 0,

      (4 << 16) | OP_DECORATE, OUT_ID, DESCRIPTOR_SET, 0,

      (4 << 16) | OP_DECORATE, OUT_ID, BINDING, 1,

      (4 << 16) | OP_DECORATE, INT_ARRAY_TYPE_ID, ARRAY_STRIDE, 4,

      (5 << 16) | OP_MEMBER_DECORATE, STRUCT_ID, 0, OFFSET, 0,

      // next declare types
      (2 << 16) | OP_TYPE_VOID, VOID_TYPE_ID,

      (3 << 16) | OP_TYPE_FUNCTION, FUNC_TYPE_ID, VOID_TYPE_ID,

      (4 << 16) | OP_TYPE_INT, INT_TYPE_ID, 32, 1,

      (4 << 16) | OP_CONSTANT, INT_TYPE_ID, CONSTANT_ARRAY_LENGTH_ID, bufferLength,

      (4 << 16) | OP_TYPE_ARRAY, INT_ARRAY_TYPE_ID, INT_TYPE_ID, CONSTANT_ARRAY_LENGTH_ID,

      (3 << 16) | OP_TYPE_STRUCT, STRUCT_ID, INT_ARRAY_TYPE_ID,

      (4 << 16) | OP_TYPE_POINTER, POINTER_TYPE_ID, UNIFORM, STRUCT_ID,

      (4 << 16) | OP_TYPE_POINTER, ELEMENT_POINTER_TYPE_ID, UNIFORM, INT_TYPE_ID,

      (4 << 16) | OP_TYPE_VECTOR, INT_VECTOR_TYPE_ID, INT_TYPE_ID, 3,

      (4 << 16) | OP_TYPE_POINTER, INT_VECTOR_POINTER_TYPE_ID, INPUT, INT_VECTOR_TYPE_ID,

      (4 << 16) | OP_TYPE_POINTER, INT_POINTER_TYPE_ID, INPUT, INT_TYPE_ID,

      // then declare constants
      (4 << 16) | OP_CONSTANT, INT_TYPE_ID, CONSTANT_ZERO_ID, 0,

      // then declare variables
      (4 << 16) | OP_VARIABLE, POINTER_TYPE_ID, IN_ID, UNIFORM,

      (4 << 16) | OP_VARIABLE, POINTER_TYPE_ID, OUT_ID, UNIFORM,

      (4 << 16) | OP_VARIABLE, INT_VECTOR_POINTER_TYPE_ID, GLOBAL_INVOCATION_ID, INPUT,

      // then declare function
      (5 << 16) | OP_FUNCTION, VOID_TYPE_ID, FUNC_ID, 0, FUNC_TYPE_ID,

      (2 << 16) | OP_LABEL, LABEL_ID,

      (5 << 16) | OP_ACCESS_CHAIN, INT_POINTER_TYPE_ID, GLOBAL_INVOCATION_X_PTR_ID, GLOBAL_INVOCATION_ID, CONSTANT_ZERO_ID,

      (4 << 16) | OP_LOAD, INT_TYPE_ID, GLOBAL_INVOCATION_X_ID, GLOBAL_INVOCATION_X_PTR_ID,

      (6 << 16) | OP_ACCESS_CHAIN, ELEMENT_POINTER_TYPE_ID, IN_ELEMENT_ID, IN_ID, CONSTANT_ZERO_ID, GLOBAL_INVOCATION_X_ID,

      (4 << 16) | OP_LOAD, INT_TYPE_ID, TEMP_LOADED_ID, IN_ELEMENT_ID,

      (6 << 16) | OP_ACCESS_CHAIN, ELEMENT_POINTER_TYPE_ID, OUT_ELEMENT_ID, OUT_ID, CONSTANT_ZERO_ID, GLOBAL_INVOCATION_X_ID,

      (3 << 16) | OP_STORE, OUT_ELEMENT_ID, TEMP_LOADED_ID,

      (1 << 16) | OP_RETURN,

      (1 << 16) | OP_FUNCTION_END,
    };

    VkShaderModuleCreateInfo shaderModuleCreateInfo = {
      VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
      0,
      0,
      sizeof(shader),
      shader
    };

    VkShaderModule shader_module;

    BAIL_ON_BAD_RESULT(vkCreateShaderModule(device, &shaderModuleCreateInfo, 0, &shader_module));

    VkDescriptorSetLayoutBinding descriptorSetLayoutBindings[2] = {
      {
        0,
        VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
        1,
        VK_SHADER_STAGE_COMPUTE_BIT,
        0
      },
      {
        1,
        VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
        1,
        VK_SHADER_STAGE_COMPUTE_BIT,
        0
      }
    };

    VkDescriptorSetLayoutCreateInfo descriptorSetLayoutCreateInfo = {
      VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
      0,
      0,
      2,
      descriptorSetLayoutBindings
    };

    VkDescriptorSetLayout descriptorSetLayout;
    BAIL_ON_BAD_RESULT(vkCreateDescriptorSetLayout(device, &descriptorSetLayoutCreateInfo, 0, &descriptorSetLayout));

    VkPipelineLayoutCreateInfo pipelineLayoutCreateInfo = {
      VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
      0,
      0,
      1,
      &descriptorSetLayout,
      0,
      0
    };

    VkPipelineLayout pipelineLayout;
    BAIL_ON_BAD_RESULT(vkCreatePipelineLayout(device, &pipelineLayoutCreateInfo, 0, &pipelineLayout));

    VkComputePipelineCreateInfo computePipelineCreateInfo = {
      VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
      0,
      0,
      {
        VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
        0,
        0,
        VK_SHADER_STAGE_COMPUTE_BIT,
        shader_module,
        "f",
        0
      },
      pipelineLayout,
      0,
      0
    };

    VkPipeline pipeline;
    BAIL_ON_BAD_RESULT(vkCreateComputePipelines(device, 0, 1, &computePipelineCreateInfo, 0, &pipeline));

    VkCommandPoolCreateInfo commandPoolCreateInfo = {
      VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO,
      0,
      0,
      queueFamilyIndex
    };

    VkDescriptorPoolSize descriptorPoolSize = {
      VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
      2
    };

    VkDescriptorPoolCreateInfo descriptorPoolCreateInfo = {
      VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
      0,
      0,
      1,
      1,
      &descriptorPoolSize
    };

    VkDescriptorPool descriptorPool;
    BAIL_ON_BAD_RESULT(vkCreateDescriptorPool(device, &descriptorPoolCreateInfo, 0, &descriptorPool));

    VkDescriptorSetAllocateInfo descriptorSetAllocateInfo = {
      VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
      0,
      descriptorPool,
      1,
      &descriptorSetLayout
    };

    VkDescriptorSet descriptorSet;
    BAIL_ON_BAD_RESULT(vkAllocateDescriptorSets(device, &descriptorSetAllocateInfo, &descriptorSet));

    VkDescriptorBufferInfo in_descriptorBufferInfo = {
      in_buffer,
      0,
      VK_WHOLE_SIZE
    };

    VkDescriptorBufferInfo out_descriptorBufferInfo = {
      out_buffer,
      0,
      VK_WHOLE_SIZE
    };

    VkWriteDescriptorSet writeDescriptorSet[2] = {
      {
        VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
        0,
        descriptorSet,
        0,
        0,
        1,
        VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
        0,
        &in_descriptorBufferInfo,
        0
      },
      {
        VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
        0,
        descriptorSet,
        1,
        0,
        1,
        VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
        0,
        &out_descriptorBufferInfo,
        0
      }
    };

    vkUpdateDescriptorSets(device, 2, writeDescriptorSet, 0, 0);

    VkCommandPool commandPool;
    BAIL_ON_BAD_RESULT(vkCreateCommandPool(device, &commandPoolCreateInfo, 0, &commandPool));

    VkCommandBufferAllocateInfo commandBufferAllocateInfo = {
      VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO,
      0,
      commandPool,
      VK_COMMAND_BUFFER_LEVEL_PRIMARY,
      1
    };

    VkCommandBuffer commandBuffer;
    BAIL_ON_BAD_RESULT(vkAllocateCommandBuffers(device, &commandBufferAllocateInfo, &commandBuffer));

    VkCommandBufferBeginInfo commandBufferBeginInfo = {
      VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO,
      0,
      VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT,
      0
    };

    BAIL_ON_BAD_RESULT(vkBeginCommandBuffer(commandBuffer, &commandBufferBeginInfo));

    vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);

    vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE,
      pipelineLayout, 0, 1, &descriptorSet, 0, 0);

    vkCmdDispatch(commandBuffer, bufferSize / sizeof(int32_t), 1, 1);

    BAIL_ON_BAD_RESULT(vkEndCommandBuffer(commandBuffer));

    VkQueue queue;
    vkGetDeviceQueue(device, queueFamilyIndex, 0, &queue);

    VkSubmitInfo submitInfo = {
      VK_STRUCTURE_TYPE_SUBMIT_INFO,
      0,
      0,
      0,
      0,
      1,
      &commandBuffer,
      0,
      0
    };

    BAIL_ON_BAD_RESULT(vkQueueSubmit(queue, 1, &submitInfo, 0));

    BAIL_ON_BAD_RESULT(vkQueueWaitIdle(queue));

    BAIL_ON_BAD_RESULT(vkMapMemory(device, memory, 0, memorySize, 0, (void *)&payload));

    for (uint32_t k = 0, e = bufferSize / sizeof(int32_t); k < e; k++) {
      BAIL_ON_BAD_RESULT(payload[k + e] == payload[k] ? VK_SUCCESS : VK_ERROR_OUT_OF_HOST_MEMORY);
    }
  }
}

Qualcomm 820 with Adreno 530 GPU:
from "vkCreateBuffer" to "vkQueueWaitIdle", ~ 7-9ms.
from "vkQueueSubmit" to "vkQueueWaitIdle", ~ 2ms

And I must release resources or it stuck at the end.

... cleanup at the end of the loop
vkFreeCommandBuffers(device, commandPool, 1, &commandBuffer);
vkDestroyCommandPool(device, commandPool, NULL);
vkDestroyDescriptorPool(device, descriptorPool, NULL);
vkDestroyPipeline(device, pipeline, NULL);
vkDestroyPipelineLayout(device, pipelineLayout, NULL);
vkDestroyDescriptorSetLayout(device, descriptorSetLayout, NULL);
vkDestroyShaderModule(device, shader_module, NULL);
vkDestroyBuffer(device, out_buffer, NULL);
vkDestroyBuffer(device, in_buffer, NULL);
vkDestroyDevice(device, NULL);

Anyone tried the performance from "vkQueueSubmit" to "vkQueueWaitIdle" on Arm Mali G-71/G-72 GPU

A few errors compiling (e.g. line 219: (void *) should be (void **)).
Also don't forget to destroy all the Vulkan objects created.
Despite this, thank you for putting this code up. There is not much Vulkan examples, even more so, focused purely on the computing part of Vulkan. This code has enabled me to find a trivial 1 character mistake in my code, which caused my compute program to not run at all.
Thumbs up!

oh GOSH! why not use OpenCL?.. its much much better, at least it has C-like code! i dont understand that shader[] variable, seems you have to be an assembler and linker in one head, in order to write such program. That is not intended for average user and regular use. I am very disappointed with it, for use as a "compute api".. example is not simple, but obfuscated and complicated at first glance

As per RomanPolek's comment, I also had to put some casts in and also change the data type of 'shader[]' to uint32_t to get it to compile.

However, for some reason only the very first element of the payload array gets copied over! I verified this by replacing rand() with an array of known values and reading the memory in debug mode. The value in payload[0] makes it over to the second buffer but nothing else does, thus causing line 607 to bail out. Has anyone else experienced this? Can anyone give advice on how I could go about debugging this?

thanks

oh GOSH! why not use OpenCL?.. its much much better, at least it has C-like code! i dont understand that shader[] variable, seems you have to be an assembler and linker in one head, in order to write such program. That is not intended for average user and regular use. I am very disappointed with it, for use as a "compute api".. example is not simple, but obfuscated and complicated at first glance

@xakepp35 You can write compute shader in GLSL and then compile to SPIR-V.

As gpoulton points out, Vulkan requires VkShaderModuleCreateInfo.pCode to be of type uint32_t.
I must add, that even after changing the type of the shader-code, i get the following error:

UNASSIGNED-CoreValidation-Shader-InconsistentSpirv(ERROR / SPEC): msgNum: 0 - SPIR-V module not valid: Interface variable id <4> is used by entry point 'f' id <1>, but is not listed as an interface
%gl_GlobalInvocationID = OpVariable %_ptr_Input_v3int Input

As per RomanPolek's comment, I also had to put some casts in and also change the data type of 'shader[]' to uint32_t to get it to compile.

However, for some reason only the very first element of the payload array gets copied over! I verified this by replacing rand() with an array of known values and reading the memory in debug mode. The value in payload[0] makes it over to the second buffer but nothing else does, thus causing line 607 to bail out. Has anyone else experienced this? Can anyone give advice on how I could go about debugging this?

thanks

I'm getting this same result, has anybody solved it?

Thanks for your code sample
Can you explain me what is the NPH suffix and what it means ?
I can't find any explanation on the web...
Thanks!

the NPH is just my initials - Neil Paul Henning. Since C doesn't have any namespaces I decided to namespace the methods!

This is great. I think learning Vulkan through the compute side first makes a lot more sense, rather than going through the very long graphics setup.

UNASSIGNED-CoreValidation-Shader-InconsistentSpirv(ERROR / SPEC): msgNum: 0 - SPIR-V module not valid: Interface variable id <4> is used by entry point 'f' id <1>, but is not listed as an interface
%gl_GlobalInvocationID = OpVariable %_ptr_Input_v3int Input

the module validates fine for me after changing line 325 to:
(4 << 16) | OP_ENTRY_POINT, 5, FUNC_ID, 0x66, GLOBAL_INVOCATION_ID

UNASSIGNED-CoreValidation-Shader-InconsistentSpirv(ERROR / SPEC): msgNum: 0 - SPIR-V module not valid: Interface variable id <4> is used by entry point 'f' id <1>, but is not listed as an interface
%gl_GlobalInvocationID = OpVariable %_ptr_Input_v3int Input

the module validates fine for me after changing line 325 to:
(4 << 16) | OP_ENTRY_POINT, 5, FUNC_ID, 0x66, GLOBAL_INVOCATION_ID

It works with me with (5 << 16) | OP_ENTRY_POINT, 5, FUNC_ID, 0x66, GLOBAL_INVOCATION_ID.

Seemingly (x << 16) means "the current instruction consists of x uint32_t values."

I measured times with the following code:

        auto tm0 = std::chrono::high_resolution_clock::now();
        BAIL_ON_BAD_RESULT(vkQueueSubmit(queue, 1, &submitInfo, 0));
        BAIL_ON_BAD_RESULT(vkQueueWaitIdle(queue));
        auto tm1 = std::chrono::high_resolution_clock::now();
        BAIL_ON_BAD_RESULT(vkMapMemory(device, memory, 0, memorySize, 0, (void **)&payload));
        auto tm2 = std::chrono::high_resolution_clock::now();

        for (uint32_t k = 0, e = bufferSize / sizeof(int32_t); k < e; k++) {
            BAIL_ON_BAD_RESULT(payload[k + e] == payload[k] ? VK_SUCCESS : VK_ERROR_OUT_OF_HOST_MEMORY);
        }
        auto tm3 = std::chrono::high_resolution_clock::now();
        using milliseconds = std::chrono::duration<double, std::milli>;
        milliseconds tmProcess = tm1-tm0;
        milliseconds tmMap     = tm2-tm1;
        milliseconds tmRead    = tm3-tm2;
        printf("Times (ms):\n    process: %f\n    map    : %f\n    read    : %f",
            tmProcess.count(), tmMap.count(), tmRead.count());

NVidia Geforce RTX 2080 super, Windows 10, Core I9-10980HK.

Times (ms):
    process: 0.315900
    map    : 0.000300
    read    : 3.107100

Reading back the memory takes most of the time because a memory without VK_MEMORY_PROPERTY_HOST_CACHED_BIT was being selected. If I change memory selection code like this:

        for (uint32_t k = 0; k < properties.memoryTypeCount; k++) {
            if ((VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT & properties.memoryTypes[k].propertyFlags) &&
                (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT & properties.memoryTypes[k].propertyFlags) &&
                (VK_MEMORY_PROPERTY_HOST_CACHED_BIT & properties.memoryTypes[k].propertyFlags) &&
                (memorySize < properties.memoryHeaps[properties.memoryTypes[k].heapIndex].size)) {
                memoryTypeIndex = k;
                break;
            }
        }

then I get

Times (ms):
    process: 0.327000
    map    : 0.000500
    read    : 0.010500

Thank you to everyone who participates because it enriches the development of NPH.

VkQueueFamilyProperties* const queueFamilyProperties = (VkQueueFamilyProperties*)_alloca(
sizeof(VkQueueFamilyProperties) * queueFamilyPropertiesCount);

in Visual Studio Comunity advise me to change: _alloca for _malloca

@xakepp35 You can write compute shader in GLSL and then compile to SPIR-V.

Or alternatively, write an OpenCL kernel and then compile to SPIR-V using clspv