Store vertex attributes in SOA style: continuous region of vertex buffer per attribute.

Removed structures that specified different vertex format.
This commit is contained in:
Artem Kharytoniuk 2017-04-14 19:58:06 +03:00
parent 047be0bd46
commit eac342cdd9
6 changed files with 178 additions and 153 deletions

View File

@ -1063,7 +1063,13 @@ const void *RB_SwapBuffers( const void *data ) {
// VULKAN
extern FILE* vk_log_file;
vulkan_demo->end_frame();
if (r_logFile->integer)
fprintf(vk_log_file, "end_frame (vb_size %d, ib_size %d)\n",
vk.xyz_elements * 16 + vk.color_st_elements * 20,
(int)vk.index_buffer_offset);
vkCmdEndRenderPass(vk.command_buffer);
VkResult result = vkEndCommandBuffer(vk.command_buffer);
check_vk_result(result, "vkEndCommandBuffer");

View File

@ -791,6 +791,9 @@ static void ComputeTexCoords( shaderStage_t *pStage ) {
*/
static void RB_IterateStagesGeneric( shaderCommands_t *input )
{
// VULKAN
vk_bind_resources_shared_between_stages(input->numPasses);
for ( int stage = 0; stage < MAX_SHADER_STAGES; stage++ )
{
shaderStage_t *pStage = tess.xstages[stage];
@ -838,7 +841,9 @@ static void RB_IterateStagesGeneric( shaderCommands_t *input )
}
// VULKAN
vk_draw(pStage->vk_pipeline, multitexture);
vk_bind_stage_specific_resources(pStage->vk_pipeline, multitexture);
vkCmdDrawIndexed(vk.command_buffer, tess.numIndexes, 1, 0, 0, 0);
glState.vk_dirty_attachments = true;
// allow skipping out to show just lightmaps during development
if ( r_lightmap->integer && ( pStage->bundle[0].isLightmap || pStage->bundle[1].isLightmap ) )
@ -846,6 +851,9 @@ static void RB_IterateStagesGeneric( shaderCommands_t *input )
break;
}
}
// VULKAN
vk.xyz_elements += tess.numVertexes;
}

View File

@ -6,14 +6,21 @@
#include "vk_demo.h"
#include <algorithm>
#include <chrono>
#include <cassert>
#include <iostream>
#include <string>
#include <vector>
static const int VERTEX_BUFFER_SIZE = 4 * 1024 * 1024;
const int VERTEX_CHUNK_SIZE = 512 * 1024;
const int XYZ_SIZE = 4 * VERTEX_CHUNK_SIZE;
const int COLOR_SIZE = 1 * VERTEX_CHUNK_SIZE;
const int ST0_SIZE = 2 * VERTEX_CHUNK_SIZE;
const int ST1_SIZE = 2 * VERTEX_CHUNK_SIZE;
const int XYZ_OFFSET = 0;
const int COLOR_OFFSET = XYZ_OFFSET + XYZ_SIZE;
const int ST0_OFFSET = COLOR_OFFSET + COLOR_SIZE;
const int ST1_OFFSET = ST0_OFFSET + ST0_SIZE;
static const int VERTEX_BUFFER_SIZE = XYZ_SIZE + COLOR_SIZE + ST0_SIZE + ST1_SIZE;
static const int INDEX_BUFFER_SIZE = 2 * 1024 * 1024;
static const std::vector<const char*> instance_extensions = {
@ -510,6 +517,11 @@ bool vk_initialize(HWND hwnd) {
vk.vertex_buffer_memory = get_allocator()->allocate_staging_memory(vk.vertex_buffer);
result = vkBindBufferMemory(vk.device, vk.vertex_buffer, vk.vertex_buffer_memory, 0);
check_vk_result(result, "vkBindBufferMemory");
void* data;
result = vkMapMemory(vk.device, vk.vertex_buffer_memory, 0, VERTEX_BUFFER_SIZE, 0, &data);
check_vk_result(result, "vkMapMemory");
vk.vertex_buffer_ptr = (byte*)data;
}
{
VkBufferCreateInfo desc;
@ -528,6 +540,11 @@ bool vk_initialize(HWND hwnd) {
vk.index_buffer_memory = get_allocator()->allocate_staging_memory(vk.index_buffer);
result = vkBindBufferMemory(vk.device, vk.index_buffer, vk.index_buffer_memory, 0);
check_vk_result(result, "vkBindBufferMemory");
void* data;
result = vkMapMemory(vk.device, vk.index_buffer_memory, 0, INDEX_BUFFER_SIZE, 0, &data);
check_vk_result(result, "vkMapMemory");
vk.index_buffer_ptr = (byte*)data;
}
} catch (const std::exception&) {
@ -718,28 +735,67 @@ static VkPipeline create_pipeline(const Vk_Pipeline_Desc& desc) {
if (desc.state_bits & GLS_ATEST_BITS)
shader_stages_state.back().pSpecializationInfo = &specialization_info;
//
// Vertex input
//
VkVertexInputBindingDescription bindings[4];
// xyz array
bindings[0].binding = 0;
bindings[0].stride = sizeof(vec4_t);
bindings[0].inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
// color array
bindings[1].binding = 1;
bindings[1].stride = sizeof(color4ub_t);
bindings[1].inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
// st0 array
bindings[2].binding = 2;
bindings[2].stride = sizeof(vec2_t);
bindings[2].inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
// st1 array
bindings[3].binding = 3;
bindings[3].stride = sizeof(vec2_t);
bindings[3].inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
VkVertexInputAttributeDescription attribs[4];
// xyz
attribs[0].location = 0;
attribs[0].binding = 0;
attribs[0].format = VK_FORMAT_R32G32B32A32_SFLOAT;
attribs[0].offset = 0;
// color
attribs[1].location = 1;
attribs[1].binding = 1;
attribs[1].format = VK_FORMAT_R8G8B8A8_UNORM;
attribs[1].offset = 0;
// st0
attribs[2].location = 2;
attribs[2].binding = 2;
attribs[2].format = VK_FORMAT_R32G32_SFLOAT;
attribs[2].offset = 0;
// st1
attribs[3].location = 3;
attribs[3].binding = 3;
attribs[3].format = VK_FORMAT_R32G32_SFLOAT;
attribs[3].offset = 0;
VkPipelineVertexInputStateCreateInfo vertex_input_state;
vertex_input_state.sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO;
vertex_input_state.pNext = nullptr;
vertex_input_state.flags = 0;
vertex_input_state.vertexBindingDescriptionCount = (desc.shader_type == Vk_Shader_Type::single_texture) ? 3 : 4;
vertex_input_state.pVertexBindingDescriptions = bindings;
vertex_input_state.vertexAttributeDescriptionCount = (desc.shader_type == Vk_Shader_Type::single_texture) ? 3 : 4;
vertex_input_state.pVertexAttributeDescriptions = attribs;
auto bindings = Vk_Vertex::get_bindings();
auto attribs = Vk_Vertex::get_attributes();
auto bindings2 = Vk_Vertex2::get_bindings();
auto attribs2 = Vk_Vertex2::get_attributes();
if (desc.shader_type == Vk_Shader_Type::single_texture) {
vertex_input_state.vertexBindingDescriptionCount = (uint32_t)bindings.size();
vertex_input_state.pVertexBindingDescriptions = bindings.data();
vertex_input_state.vertexAttributeDescriptionCount = (uint32_t)attribs.size();
vertex_input_state.pVertexAttributeDescriptions = attribs.data();
} else {
vertex_input_state.vertexBindingDescriptionCount = (uint32_t)bindings2.size();
vertex_input_state.pVertexBindingDescriptions = bindings2.data();
vertex_input_state.vertexAttributeDescriptionCount = (uint32_t)attribs2.size();
vertex_input_state.pVertexAttributeDescriptions = attribs2.data();
}
//
// Primitive assembly.
//
VkPipelineInputAssemblyStateCreateInfo input_assembly_state;
input_assembly_state.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO;
input_assembly_state.pNext = nullptr;
@ -747,6 +803,9 @@ static VkPipeline create_pipeline(const Vk_Pipeline_Desc& desc) {
input_assembly_state.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
input_assembly_state.primitiveRestartEnable = VK_FALSE;
//
// Viewport.
//
VkPipelineViewportStateCreateInfo viewport_state;
viewport_state.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO;
viewport_state.pNext = nullptr;
@ -756,6 +815,9 @@ static VkPipeline create_pipeline(const Vk_Pipeline_Desc& desc) {
viewport_state.scissorCount = 1;
viewport_state.pScissors = nullptr; // dynamic scissor state
//
// Rasterization.
//
VkPipelineRasterizationStateCreateInfo rasterization_state;
rasterization_state.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO;
rasterization_state.pNext = nullptr;
@ -1011,7 +1073,8 @@ void vk_destroy_resources() {
vkDeviceWaitIdle(vk.device);
vk_destroy_pipelines();
vk.vertex_buffer_offset = 0;
vk.xyz_elements = 0;
vk.color_st_elements = 0;
vk.index_buffer_offset = 0;
}
@ -1075,60 +1138,32 @@ void vk_get_mvp_transform(float mvp[16]) {
}
}
void vk_draw(VkPipeline pipeline, bool multitexture) {
void vk_bind_resources_shared_between_stages(int num_passes) {
extern FILE* vk_log_file;
if (r_logFile->integer)
fprintf(vk_log_file, "render_tess (%s, vert %d, inds %d)\n", multitexture ? "M" : "S", tess.numVertexes, tess.numIndexes);
fprintf(vk_log_file, "render_tess (passes %d, vert %d, inds %d)\n", num_passes, tess.numVertexes, tess.numIndexes);
auto vertex_size = multitexture ? sizeof(Vk_Vertex2) : sizeof(Vk_Vertex);
// xyz
{
if ((vk.xyz_elements + tess.numVertexes) * sizeof(vec4_t) > XYZ_SIZE)
ri.Error(ERR_DROP, "vulkan: vertex buffer overflow (xyz)\n");
// update vertex buffer
std::size_t vertexes_size = tess.numVertexes * vertex_size;
if (vk.vertex_buffer_offset + vertexes_size > VERTEX_BUFFER_SIZE)
ri.Error(ERR_DROP, "vk_draw: vertex buffer overflow\n");
void* data;
VkResult result = vkMapMemory(vk.device, vk.vertex_buffer_memory, vk.vertex_buffer_offset, vertexes_size, 0, &data);
check_vk_result(result, "vkMapMemory");
Timer t;
unsigned char* ptr = (unsigned char*)data;
for (int i = 0; i < tess.numVertexes; i++, ptr += vertex_size) {
Vk_Vertex* v = (Vk_Vertex*)ptr;
v->pos[0] = tess.xyz[i][0];
v->pos[1] = tess.xyz[i][1];
v->pos[2] = tess.xyz[i][2];
v->color[0] = tess.svars.colors[i][0];
v->color[1] = tess.svars.colors[i][1];
v->color[2] = tess.svars.colors[i][2];
v->color[3] = tess.svars.colors[i][3];
v->st[0] = tess.svars.texcoords[0][i][0];
v->st[1] = tess.svars.texcoords[0][i][1];
if (multitexture) {
auto v2 = (Vk_Vertex2*)ptr;
v2->st2[0] = tess.svars.texcoords[1][i][0];
v2->st2[1] = tess.svars.texcoords[1][i][1];
}
byte* dst = vk.vertex_buffer_ptr + XYZ_OFFSET + vk.xyz_elements * sizeof(vec4_t);
Com_Memcpy(dst, tess.xyz, tess.numVertexes * sizeof(vec4_t));
}
vulkan_demo->vertex_copy_time += t.Elapsed_Seconds();
vkUnmapMemory(vk.device, vk.vertex_buffer_memory);
vkCmdBindVertexBuffers(vk.command_buffer, 0, 1, &vk.vertex_buffer, &vk.vertex_buffer_offset);
vk.vertex_buffer_offset += vertexes_size;
std::size_t indexes_size = tess.numIndexes * sizeof(uint32_t);
// update index buffer
std::size_t indexes_size = tess.numIndexes * sizeof(uint32_t);
if (vk.index_buffer_offset + indexes_size > INDEX_BUFFER_SIZE)
ri.Error(ERR_DROP, "vk_draw: index buffer overflow\n");
{
if (vk.index_buffer_offset + indexes_size > INDEX_BUFFER_SIZE)
ri.Error(ERR_DROP, "vk_draw: index buffer overflow\n");
result = vkMapMemory(vk.device, vk.index_buffer_memory, vk.index_buffer_offset, indexes_size, 0, &data);
check_vk_result(result, "vkMapMemory");
uint32_t* ind = (uint32_t*)data;
for (int i = 0; i < tess.numIndexes; i++, ind++) {
*ind = tess.indexes[i];
byte* dst = vk.index_buffer_ptr + vk.index_buffer_offset;
Com_Memcpy(dst, tess.indexes, indexes_size);
}
vkUnmapMemory(vk.device, vk.index_buffer_memory);
// configure indexes stream
vkCmdBindIndexBuffer(vk.command_buffer, vk.index_buffer, vk.index_buffer_offset, VK_INDEX_TYPE_UINT32);
vk.index_buffer_offset += indexes_size;
@ -1136,15 +1171,63 @@ void vk_draw(VkPipeline pipeline, bool multitexture) {
float mvp[16];
vk_get_mvp_transform(mvp);
vkCmdPushConstants(vk.command_buffer, vk.pipeline_layout, VK_SHADER_STAGE_VERTEX_BIT, 0, 64, mvp);
}
void vk_bind_stage_specific_resources(VkPipeline pipeline, bool multitexture) {
//
// Specify color/st for each draw call since they are regenerated for each Q3 shader's stage.
// xyz are specified only once for all stages.
//
// color
{
if ((vk.color_st_elements + tess.numVertexes) * sizeof(color4ub_t) > COLOR_SIZE)
ri.Error(ERR_DROP, "vulkan: vertex buffer overflow (color)\n");
byte* dst = vk.vertex_buffer_ptr + COLOR_OFFSET + vk.color_st_elements * sizeof(color4ub_t);
Com_Memcpy(dst, tess.svars.colors, tess.numVertexes * sizeof(color4ub_t));
}
// st0
{
if ((vk.color_st_elements + tess.numVertexes) * sizeof(vec2_t) > ST0_SIZE)
ri.Error(ERR_DROP, "vulkan: vertex buffer overflow (st0)\n");
byte* dst = vk.vertex_buffer_ptr + ST0_OFFSET + vk.color_st_elements * sizeof(vec2_t);
Com_Memcpy(dst, tess.svars.texcoords[0], tess.numVertexes * sizeof(vec2_t));
}
// st1
if (multitexture) {
if ((vk.color_st_elements + tess.numVertexes) * sizeof(vec2_t) > ST1_SIZE)
ri.Error(ERR_DROP, "vulkan: vertex buffer overflow (st1)\n");
byte* dst = vk.vertex_buffer_ptr + ST1_OFFSET + vk.color_st_elements * sizeof(vec2_t);
Com_Memcpy(dst, tess.svars.texcoords[1], tess.numVertexes * sizeof(vec2_t));
}
// configure vertex data stream
VkBuffer bufs[4] = { vk.vertex_buffer, vk.vertex_buffer, vk.vertex_buffer, vk.vertex_buffer }; // turtles all the way down
VkDeviceSize offs[4] = {
XYZ_OFFSET + vk.xyz_elements * sizeof(vec4_t),
COLOR_OFFSET + vk.color_st_elements * sizeof(color4ub_t),
ST0_OFFSET + vk.color_st_elements * sizeof(vec2_t),
ST1_OFFSET + vk.color_st_elements * sizeof(vec2_t)
};
vkCmdBindVertexBuffers(vk.command_buffer, 0, multitexture ? 4 : 3, bufs, offs);
vk.color_st_elements += tess.numVertexes;
// bind descriptor sets
image_t* image = glState.vk_current_images[0];
image_t* image2 = glState.vk_current_images[1];
VkDescriptorSet sets[2] = { image->vk_descriptor_set, image2 ? image2->vk_descriptor_set : VkDescriptorSet() };
int set_count = multitexture ? 2 : 1;
vkCmdBindDescriptorSets(vk.command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, vk.pipeline_layout, 0, set_count, sets, 0, nullptr);
vkCmdBindDescriptorSets(vk.command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, vk.pipeline_layout, 0, multitexture ? 2 : 1, sets, 0, nullptr);
// bind pipeline
vkCmdBindPipeline(vk.command_buffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
// configure pipeline's dynamic state
VkRect2D r = vk_get_viewport_rect();
vkCmdSetScissor(vk.command_buffer, 0, 1, &r);
@ -1160,9 +1243,4 @@ void vk_draw(VkPipeline pipeline, bool multitexture) {
if (tess.shader->polygonOffset) {
vkCmdSetDepthBias(vk.command_buffer, r_offsetUnits->value, 0.0f, r_offsetFactor->value);
}
// Draw primitives
vkCmdDrawIndexed(vk.command_buffer, tess.numIndexes, 1, 0, 0, 0);
glState.vk_dirty_attachments = true;
}

View File

@ -18,7 +18,8 @@ void vk_destroy_resources();
VkRect2D vk_get_viewport_rect();
void vk_get_mvp_transform(float mvp[16]);
void vk_draw(VkPipeline pipeline, bool multitexture);
void vk_bind_resources_shared_between_stages(int num_passes);
void vk_bind_stage_specific_resources(VkPipeline pipeline, bool multitexture);
struct Vk_Staging_Buffer {
VkBuffer handle = VK_NULL_HANDLE;
@ -56,71 +57,6 @@ VkPipeline vk_find_pipeline(const Vk_Pipeline_Desc& desc);
VkDescriptorSet vk_create_descriptor_set(VkImageView image_view);
// Vertex formats
struct Vk_Vertex {
vec3_t pos;
unsigned char color[4];
vec2_t st;
static std::array<VkVertexInputBindingDescription, 1> get_bindings() {
VkVertexInputBindingDescription binding_desc;
binding_desc.binding = 0;
binding_desc.stride = sizeof(Vk_Vertex);
binding_desc.inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
return {binding_desc};
}
static std::array<VkVertexInputAttributeDescription, 3> get_attributes() {
VkVertexInputAttributeDescription position_attrib;
position_attrib.location = 0;
position_attrib.binding = 0;
position_attrib.format = VK_FORMAT_R32G32B32_SFLOAT;
position_attrib.offset = offsetof(struct Vk_Vertex, pos);
VkVertexInputAttributeDescription color_attrib;
color_attrib.location = 1;
color_attrib.binding = 0;
color_attrib.format = VK_FORMAT_R8G8B8A8_UNORM;
color_attrib.offset = offsetof(struct Vk_Vertex, color);
VkVertexInputAttributeDescription st_attrib;
st_attrib.location = 2;
st_attrib.binding = 0;
st_attrib.format = VK_FORMAT_R32G32_SFLOAT;
st_attrib.offset = offsetof(struct Vk_Vertex, st);
return {position_attrib, color_attrib, st_attrib};
}
};
struct Vk_Vertex2 : Vk_Vertex {
vec2_t st2;
static std::array<VkVertexInputBindingDescription, 1> get_bindings() {
VkVertexInputBindingDescription binding_desc;
binding_desc.binding = 0;
binding_desc.stride = sizeof(Vk_Vertex2);
binding_desc.inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
return {binding_desc};
}
static std::array<VkVertexInputAttributeDescription, 4> get_attributes() {
auto vk_vertex_attribs = Vk_Vertex::get_attributes();
std::array<VkVertexInputAttributeDescription, 4> result;
result[0] = vk_vertex_attribs[0];
result[1] = vk_vertex_attribs[1];
result[2] = vk_vertex_attribs[2];
VkVertexInputAttributeDescription st2_attrib;
st2_attrib.location = 3;
st2_attrib.binding = 0;
st2_attrib.format = VK_FORMAT_R32G32_SFLOAT;
st2_attrib.offset = offsetof(struct Vk_Vertex2, st2);
result[3] = st2_attrib;
return result;
}
};
// Shaders.
extern unsigned char single_texture_vert_spv[];
extern long long single_texture_vert_spv_size;
@ -170,10 +106,13 @@ struct Vulkan_Instance {
VkBuffer vertex_buffer = VK_NULL_HANDLE;
VkDeviceMemory vertex_buffer_memory = VK_NULL_HANDLE;
VkDeviceSize vertex_buffer_offset = 0;
byte* vertex_buffer_ptr = nullptr; // pointer to mapped vertex buffer
int xyz_elements = 0;
int color_st_elements = 0;
VkBuffer index_buffer = VK_NULL_HANDLE;
VkDeviceMemory index_buffer_memory = VK_NULL_HANDLE;
byte* index_buffer_ptr = nullptr; // pointer to mapped index buffer
VkDeviceSize index_buffer_offset = 0;
};
@ -183,4 +122,6 @@ struct Vulkan_Resources {
int num_pipelines = 0;
Vk_Pipeline_Desc pipeline_desc[MAX_VK_PIPELINES];
VkPipeline pipelines[MAX_VK_PIPELINES];
// TODO: put images here too.
};

View File

@ -128,15 +128,9 @@ void Vulkan_Demo::begin_frame() {
vkCmdBeginRenderPass(vk.command_buffer, &render_pass_begin_info, VK_SUBPASS_CONTENTS_INLINE);
vk.vertex_buffer_offset = 0;
vk.xyz_elements = 0;
vk.color_st_elements = 0;
vk.index_buffer_offset = 0;
glState.vk_dirty_attachments = false;
}
void Vulkan_Demo::end_frame() {
if (r_logFile->integer)
fprintf(vk_log_file, "end_frame (vb_size %d, ib_size %d, copy_time %d)\n", (int)vk.vertex_buffer_offset, (int)vk.index_buffer_offset, int(vertex_copy_time * 1000000000));
vertex_copy_time = 0.0f;
vkCmdEndRenderPass(vk.command_buffer);
}

View File

@ -13,7 +13,6 @@ public:
Vulkan_Demo(int window_width, int window_height);
void begin_frame();
void end_frame();
public:
VkImage create_texture(const uint8_t* pixels, int bytes_per_pixel, int width, int height, VkImageView& image_view);
@ -30,5 +29,4 @@ public:
VkSampler texture_image_sampler = VK_NULL_HANDLE;
uint32_t swapchain_image_index = -1;
double vertex_copy_time = 0.0f;
};