1 of 62

Mobile Graphics,�Vulkan

Arman Papikyan

Lead Render Mobile Programmer at Ubisoft Ukraine

2 of 62

What is Graphics API (history)

  • Rendering on CPU

3 of 62

What is Graphics API (history)

  • Rendering on CPU,
  • First GPUs and fixed function shaders

4 of 62

What is Graphics API (history)

  • Rendering on CPU,
  • First GPUs and fixed function shaders,
  • Complex modern GPUs

5 of 62

What is Graphics API (history)

  • Rendering on CPU,
  • First GPUs and fixed function shaders,
  • Complex modern GPUs,
  • Standartized APIs – OpenGL

6 of 62

What is Graphics API (history)

  • Rendering on CPU,
  • First GPUs and fixed function shaders,
  • Complex modern GPUs,
  • Standartized APIs – OpenGL

- 1992

7 of 62

What is Graphics API (history)

  • Rendering on CPU,
  • First GPUs and fixed function shaders,
  • Complex modern GPUs,
  • Standartized APIs – OpenGL

- 1992

- 2003

8 of 62

What is Graphics API (history)

  • Rendering on CPU,
  • First GPUs and fixed function shaders,
  • Complex modern GPUs,
  • Standartized APIs – OpenGL

- 2016

- 1992

- 2003

9 of 62

  • GLSL compiler is slow

9

10 of 62

  • GLSL compiler is slow
  • Spikes ~2000 ms

10

11 of 62

  • GLSL compiler is slow
  • Spikes ~2000 ms
  • Prewarming hack

11

12 of 62

  • GLSL compiler is slow
  • Spikes ~2000 ms
  • Prewarming hack

​

12

  • Shader precompilation

13 of 62

  • GLSL compiler is slow
  • Spikes ~2000 ms
  • Prewarming hack

​

13

  • Shader precompilation
  • Safer and faster

14 of 62

  • GLSL compiler is slow
  • Spikes ~2000 ms
  • Prewarming hack

​

14

  • Shader precompilation
  • Safer and faster
  • Cache and store on disk

15 of 62

  • GLSL compiler is slow
  • Spikes ~2000 ms
  • Prewarming hack

​

15

  • Shader precompilation
  • Safer and faster
  • Cache and store on disk
  • Minimal error checking

16 of 62

  • GLSL compiler is slow
  • Spikes ~2000 ms
  • Prewarming hack

​

16

  • Shader precompilation
  • Safer and faster
  • Cache and store on disk
  • Minimal error checking
  • Rust / C++ -> Spir-V

17 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

�vkCmdBindVertexBuffers(commandBuffer, 0, 1, &_vertexBuffer, _offsets);

vkCmdBindIndexBuffer(commandBuffer, _indexBuffer, 0, VK_INDEX_TYPE_UINT32);

�vkCmdDrawIndexed(commandBuffer, _index_count, 1, 0, 0, 1);

�vkCmdEndRenderPass(commandBuffer);

VkResult result = vkEndCommandBuffer(commandBuffer);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

�glBindVertexArray(VAO);

�glDrawElements(GL_TRIANGLES, _index_count, GL_UNSIGNED_INT, 0);

> State Management

18 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

Set up the frame

> State Management

19 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

> State Management

Use shader

20 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

> State Management

Use Texture

21 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

�vkCmdBindVertexBuffers(commandBuffer, 0, 1, &_vertexBuffer, _offsets);

vkCmdBindIndexBuffer(commandBuffer, _indexBuffer, 0, VK_INDEX_TYPE_UINT32);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

�glBindVertexArray(VAO);

> State Management

Use vertex & index buffers

22 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

�vkCmdBindVertexBuffers(commandBuffer, 0, 1, &_vertexBuffer, _offsets);

vkCmdBindIndexBuffer(commandBuffer, _indexBuffer, 0, VK_INDEX_TYPE_UINT32);

�vkCmdDrawIndexed(commandBuffer, _index_count, 1, 0, 0, 1);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

�glBindVertexArray(VAO);

�glDrawElements(GL_TRIANGLES, _index_count, GL_UNSIGNED_INT, 0);

> State Management

Draw call

23 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

�vkCmdBindVertexBuffers(commandBuffer, 0, 1, &_vertexBuffer, _offsets);

vkCmdBindIndexBuffer(commandBuffer, _indexBuffer, 0, VK_INDEX_TYPE_UINT32);

�vkCmdDrawIndexed(commandBuffer, _index_count, 1, 0, 0, 1);

�vkCmdEndRenderPass(commandBuffer);

VkResult result = vkEndCommandBuffer(commandBuffer);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

�glBindVertexArray(VAO);

�glDrawElements(GL_TRIANGLES, _index_count, GL_UNSIGNED_INT, 0);

> State Management

Finish

24 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

�vkCmdBindVertexBuffers(commandBuffer, 0, 1, &_vertexBuffer, _offsets);

vkCmdBindIndexBuffer(commandBuffer, _indexBuffer, 0, VK_INDEX_TYPE_UINT32);

�vkCmdDrawIndexed(commandBuffer, _index_count, 1, 0, 0, 1);

�vkCmdEndRenderPass(commandBuffer);

VkResult result = vkEndCommandBuffer(commandBuffer);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

�glBindVertexArray(VAO);

�glDrawElements(GL_TRIANGLES, _index_count, GL_UNSIGNED_INT, 0);

> State Management

25 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

�vkCmdBindVertexBuffers(commandBuffer, 0, 1, &_vertexBuffer, _offsets);

vkCmdBindIndexBuffer(commandBuffer, _indexBuffer, 0, VK_INDEX_TYPE_UINT32);

�vkCmdDrawIndexed(commandBuffer, _index_count, 1, 0, 0, 1);

�vkCmdEndRenderPass(commandBuffer);

VkResult result = vkEndCommandBuffer(commandBuffer);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

�glBindVertexArray(VAO);

�glDrawElements(GL_TRIANGLES, _index_count, GL_UNSIGNED_INT, 0);

> State Management

26 of 62

vkBeginCommandBuffer(commandBuffer, &cmdBufInfo)

vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);

�vkCmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);

�vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &descriptorSet, 0, nullptr);

�vkCmdBindVertexBuffers(commandBuffer, 0, 1, &_vertexBuffer, _offsets);

vkCmdBindIndexBuffer(commandBuffer, _indexBuffer, 0, VK_INDEX_TYPE_UINT32);

�vkCmdDrawIndexed(commandBuffer, _index_count, 1, 0, 0, 1);

�vkCmdEndRenderPass(commandBuffer);

VkResult result = vkEndCommandBuffer(commandBuffer);

glClearColor(0.f, 0.f, 0.f, 1.f);

glClear(GL_COLOR_BUFFER_BIT);

�glUseProgram(shaderProgram);

�glActiveTexture(0);

glBindTexture(GL_TEXTURE_2D, tex);

glUniform1i(glGetUniformLocation(shaderProgram, "tex"), 0);

�glBindVertexArray(VAO);

�glDrawElements(GL_TRIANGLES, _index_count, GL_UNSIGNED_INT, 0);

> State Management

27 of 62

> Multithreading

28 of 62

> Multithreading

© Mobile Graphics Processing Fundamentals – 2021, ARM webinar

GPU

CPU

CPU

CPU

Shared Cache

System Cache

DRAM

Example S20 Snapdragon

​

1 x 2.84 GHz 100%

3 x 2.42 GHz 85%

4 x 1.80 GHz 63%

Small Core – Low Frequency

Efficiency

Big Core – High Frequency

Performance

29 of 62

> Multithreading

Power consumption relationship with performance is not linear

Transistor energy per operation ~ Voltage squared

© Mobile Graphics Processing Fundamentals – 2021, ARM webinar

30 of 62

Why is multithreading important?

  • Better utilization of multi-core HW,
  • Using smaller, power efficient cores,
  • Better work distribution,

​

  • Complexity + (more work ☹)

30

> Multithreading

31 of 62

Nvidia GeForce GTX 1080 Ti

​

Heap 10.87 GB

DEVICE_LOCAL

​

AMD Radeon RX Vega 64

​

Heap 7.75 GB

DEVICE_LOCAL

​

Heap 256 MB

DEVICE_LOCAL | HOST VISIBLE

Adreno / Mali / PowerVR..

​

HEAP – 0 MB

*no VRAM

31

© Memory Management in Vulkan – 2018, Jordan Logan, AMD

> Memory Management

32 of 62

Why do you need explicit memory management?

  • Choose HW that suits your needs,
  • Control the layout (suballocate), Predictable,
  • Subpasses allow tile-friendly use,

​

  • Complexity ++ (more work ☹)

32

> Memory Management

33 of 62

33

Rasterization modes:�

1. Immediate mode

Traditional GPU

​

2. Tiled mode

Mobile GPU

34 of 62

34

IMMEDIATE MODE

Render each triangle individually

for draw in renderPass:

for primitive in draw:

for vertex in primitive:

execute_vertex_shader(vertex)

if primitive not culled:

for fragment in primitive:

execute_fragment_shader(fragment)

© developer.arm.com/documentation/102662/0100/Immediate-Mode-GPUs

35 of 62

35

IMMEDIATE MODE

Render each triangle individually

for draw in renderPass:

for primitive in draw:

for vertex in primitive:

execute_vertex_shader(vertex)

if primitive not culled:

for fragment in primitive:

execute_fragment_shader(fragment)

© developer.arm.com/documentation/102662/0100/Immediate-Mode-GPUs

36 of 62

36

# Pass one

for draw in renderPass:

for primitive in draw:

for vertex in primitive:

execute_vertex_shader(vertex)

if primitive not culled:

append_tile_list(primitive)

1. Break into bins

TILED MODE

© developer.arm.com/documentation/102662/0100/Tile-based-GPUs

37 of 62

37

# Pass one

for draw in renderPass:

for primitive in draw:

for vertex in primitive:

execute_vertex_shader(vertex)

if primitive not culled:

append_tile_list(primitive)

1. Break into bins

TILED MODE

© developer.arm.com/documentation/102662/0100/Tile-based-GPUs

38 of 62

38

# Pass two

for tile in renderPass:

for primitive in tile:

for fragment in primitive:

execute_fragment_shader(fragment)

2. Render each tile

TILED MODE

© developer.arm.com/documentation/102662/0100/Tile-based-GPUs

39 of 62

39

# Pass one

for draw in renderPass:

for primitive in draw:

for vertex in primitive:

execute_vertex_shader(vertex)

if primitive not culled:

append_tile_list(primitive)

​

# Pass two

for tile in renderPass:

for primitive in tile:

for fragment in primitive:

execute_fragment_shader(fragment)

1. Break into bins

2. Render each tile

TILED MODE

© developer.arm.com/documentation/102662/0100/Tile-based-GPUs

40 of 62

Memory Bandwidth

  • Keeping the information on-chip memory,
  • Useful for post effects, VFX, etc,
  • Deferred rendering,
  • Reduces the bandwidth (less sys mem RW),
  • Available via Vulkan Subpasses,
  • GL alternative via extension – limited.

40

41 of 62

41

RESOURCE

Texture / Buffer / etc

R

W

R

W

Random Frame

​

Render scene => Color RT

Color RT => Anti-Alias RT

Anti-Alias RT => Display

RESOURCE

Texture / Buffer / etc

R

W

R

W

Lock

Lock

Random Frame

​

Render scene => Color RT

Lock

Color RT => Anti-Alias RT

Lock

Anti-Alias RT => Display

> Synchronization

42 of 62

42

> Synchronization

CPU

GPU

43 of 62

43

> Synchronization

CPU - Thread 1

CPU - Thread 2

GPU

44 of 62

44

> Synchronization

Fence: GPU - to - CPU

Semaphor: CPU - to - GPU

​

Barrier: GPU - to - GPU

​

Timeline Sempahor: All - to - All

45 of 62

45

> Synchronization

// Three dispatches that don’t have conflicting resource accesses�vkCmdDispatch( 1 );�vkCmdDispatch( 2 );�vkCmdDispatch( 3 );��// 4, 5, and 6 don’t share resources with 1, 2, and 3�// No reason for them to be blocked, so set an event to wait for later�vkCmdSetEvent( A, srcStageMask = COMPUTE );�vkCmdDispatch( 4 );�vkCmdDispatch( 5 );�vkCmdDispatch( 6 );��// 7 and 8 don’t use the same resources as 4, 5, and 6.  So use an event�vkCmdSetEvent( B, srcStageMask = COMPUTE );��// 7 and 8 need the results of 1, 2, and 3�// So we’ll wait for them by waiting on A�vkCmdWaitEvents( A, dstStageMask = COMPUTE );�vkCmdDispatch( 7 );�vkCmdDispatch( 8 );

�// 9 uses the same resources as 4, 5, and 6 so we wait.�// Also assumed is that 9 needs nothing from 7 and 8�vkCmdWaitEvents( B, dstStageMask = COMPUTE );

�vkCmdDispatch( 9 );

© khronos.org/blog/understanding-vulkan-synchronization

46 of 62

46

© github.com/KhronosGroup/Vulkan-ValidationLayers/blob/master/docs/synchronization_usage.md

> Synchronization

47 of 62

47

© khronos.org/blog/understanding-vulkan-synchronization

> Synchronization

VK_IMAGE_LAYOUT_UNDEFINED

VK_IMAGE_LAYOUT_GENERAL

VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL

VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL

VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL

VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL

VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL

VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL

VK_IMAGE_LAYOUT_PREINITIALIZED

VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL

VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL

VK_ACCESS_INDIRECT_COMMAND_READ_BIT

VK_ACCESS_INDEX_READ_BIT

VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT

VK_ACCESS_UNIFORM_READ_BIT

VK_ACCESS_INPUT_ATTACHMENT_READ_BIT

VK_ACCESS_SHADER_READ_BIT

VK_ACCESS_SHADER_WRITE_BIT

VK_ACCESS_COLOR_ATTACHMENT_READ_BIT

VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT

VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT

VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT

VK_ACCESS_TRANSFER_READ_BIT

VK_ACCESS_TRANSFER_WRITE_BIT

VK_ACCESS_HOST_READ_BIT

VK_ACCESS_HOST_WRITE_BIT

VK_ACCESS_MEMORY_READ_BIT

VK_ACCESS_MEMORY_WRITE_BIT

Stage

Layout

.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,

.dstAccessMask = VK_ACCESS_SHADER_READ_BIT,

.oldLayout = VK_IMAGE_LAYOUT_GENERAL,

.newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL

48 of 62

  • Manual synchronization is difficult,
  • GL Driver does all this by default,
  • Application could do it better.

48

> Synchronization

49 of 62

49

> Synchronization

// Three dispatches that don’t have conflicting resource accesses�vkCmdDispatch( 1 );�vkCmdDispatch( 2 );�vkCmdDispatch( 3 );��// 4, 5, and 6 don’t share resources with 1, 2, and 3�// No reason for them to be blocked, so set an event to wait for later�vkCmdSetEvent( A, srcStageMask = COMPUTE );�vkCmdDispatch( 4 );�vkCmdDispatch( 5 );�vkCmdDispatch( 6 );��// 7 and 8 don’t use the same resources as 4, 5, and 6.  So use an event�vkCmdSetEvent( B, srcStageMask = COMPUTE );��// 7 and 8 need the results of 1, 2, and 3�// So we’ll wait for them by waiting on A�vkCmdWaitEvents( A, dstStageMask = COMPUTE );�vkCmdDispatch( 7 );�vkCmdDispatch( 8 );

�// 9 uses the same resources as 4, 5, and 6 so we wait.�// Also assumed is that 9 needs nothing from 7 and 8�vkCmdWaitEvents( B, dstStageMask = COMPUTE );

�vkCmdDispatch( 9 );

© themaister.net/blog/2017/08/15/render-graphs-and-vulkan-a-deep-dive/

1

2

3

7

8

4

5

6

9

50 of 62

50

> Synchronization

© themaister.net/blog/2017/08/15/render-graphs-and-vulkan-a-deep-dive/

1

2

3

7

8

4

5

6

9

51 of 62

  • Manual synchronization is difficult,
  • GL Driver does all this by default,
  • Application could do it better,
  • Frame graph abstraction helps,
  • Frame graph knows more about your resources than the driver.

51

> Synchronization

52 of 62

- VK_KHR_acceleration_structure

- VK_KHR_ray_tracing_pipeline

- SPV_KHR_ray_tracing

- GLSL_EXT_ray_tracing

​

  • Not available on GL,
  • Not available on 99% mobile phones,
  • Supported on PC (RTX cards)

52

> Hardware Raytracing

53 of 62

Summary

  • Predictable shader compilation,
  • Minimal error checking,
  • State Management,
  • Multithreading,
  • Memory Management,
  • Synchronization,
  • Hardware RT

54 of 62

© ravbug.com/graphics/

55 of 62

GLES 2.0 2012+

GLES 3.1 2016+

​

​

Metal 2014+

Vulkan 2018+

​

​

​

Vulkan -> [MoltenVK] -> Metal

56 of 62

GLES 2.0 2012+ (96.5%)*

GLES 3.1 2016+ (90%)*

​

​

Metal 2014+ (up to 50%)*

Vulkan 2018+ (up to 30%)*

​

*Estimates

57 of 62

GLES 2.0 2012+ (96.5%)*

GLES 3.1 2016+ (90%)*

​

​

Metal 2014+ (up to 50%)*

Vulkan 2018+ (up to 30%)*

​

*Estimates

​

​

Vulkan supports all 2018+ mobile devices

Limited

58 of 62

References

  • Vulkan-tutorial.com,
  • Vulkan Best practices for mobile developers,
  • Youtube Lectures - Computer Graphics at TU Wien
  • Memory Management in Vulkan – 2018, Jordan Logan, AMD
  • Mobile Graphics Processing Fundamentals – 2021, ARM webinar
  • Developer.arm.com/documentation/102662/0100/Overview
  • Themaister.net/blog/2017/08/15/render-graphs-and-vulkan-a-deep-dive/

59 of 62

59

Thanks for your attention,

Questions?

60 of 62

60

Geometry

Pass

Draw

Material

Pass

Lighting

Pass

Geometry

Pass

Draw

Material + Lighting

Pass

61 of 62

61

© learnopengl.com/Advanced-Lighting/Deferred-Shading

62 of 62

62

© learnopengl.com/Advanced-Lighting/Deferred-Shading

O(num_geometry_fragments * num_lights)

O(screen_resolution * num_lights)