Continue implementation
This commit is contained in:
0cc4m
parent
4a96d0eb7f
commit
88d4ec05a8
1 changed files with 128 additions and 172 deletions
300
ggml-vulkan.cpp
300
ggml-vulkan.cpp
|
|
@ -3,23 +3,20 @@
|
||||||
#include <vulkan/vulkan.hpp>
|
#include <vulkan/vulkan.hpp>
|
||||||
#include "external/vk_mem_alloc.h"
|
#include "external/vk_mem_alloc.h"
|
||||||
|
|
||||||
#include <iostream>
|
#include <atomic>
|
||||||
#include <fstream>
|
#include <fstream>
|
||||||
|
#include <iostream>
|
||||||
|
#include <limits>
|
||||||
|
|
||||||
#include "ggml.h"
|
#include "ggml.h"
|
||||||
|
|
||||||
// static cl_platform_id platform;
|
#define VK_API_VERSION VK_API_VERSION_1_2
|
||||||
// static cl_device_id device;
|
|
||||||
// static cl_context context;
|
|
||||||
// static cl_command_queue queue;
|
|
||||||
// static cl_program program;
|
|
||||||
// static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q4_2, kernel_q5_0, kernel_q5_1, kernel_q8_0;
|
|
||||||
// static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c;
|
|
||||||
// static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0;
|
|
||||||
|
|
||||||
vk::Instance instance;
|
vk::Instance vk_instance;
|
||||||
vk::PhysicalDevice physical_device;
|
uint32_t vk_compute_queue_family_index;
|
||||||
|
vk::PhysicalDevice vk_physical_device;
|
||||||
vk::Device vk_device;
|
vk::Device vk_device;
|
||||||
|
vk::DescriptorSetLayout vk_pipeline_matmul_dsl;
|
||||||
vk::Pipeline vk_pipeline_matmul;
|
vk::Pipeline vk_pipeline_matmul;
|
||||||
VmaAllocation vk_buffer_qa_alloc, vk_buffer_a_alloc, vk_buffer_b_alloc, vk_buffer_c_alloc;
|
VmaAllocation vk_buffer_qa_alloc, vk_buffer_a_alloc, vk_buffer_b_alloc, vk_buffer_c_alloc;
|
||||||
vk::Buffer vk_buffer_qa, vk_buffer_a, vk_buffer_b, vk_buffer_c;
|
vk::Buffer vk_buffer_qa, vk_buffer_a, vk_buffer_b, vk_buffer_c;
|
||||||
|
|
@ -27,29 +24,27 @@ vk::Buffer vk_buffer_qa, vk_buffer_a, vk_buffer_b, vk_buffer_c;
|
||||||
void ggml_vk_init(void) {
|
void ggml_vk_init(void) {
|
||||||
char* GGML_VULKAN_DEVICE = getenv("GGML_VULKAN_DEVICE");
|
char* GGML_VULKAN_DEVICE = getenv("GGML_VULKAN_DEVICE");
|
||||||
int dev_num = (GGML_VULKAN_DEVICE == NULL ? 0 : atoi(GGML_VULKAN_DEVICE));
|
int dev_num = (GGML_VULKAN_DEVICE == NULL ? 0 : atoi(GGML_VULKAN_DEVICE));
|
||||||
printf("\nInitializing Vulkan...");
|
|
||||||
printf("\nAttempting to use: Device=%d\n", dev_num);
|
|
||||||
|
|
||||||
vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, VK_API_VERSION_1_2 };
|
vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, VK_API_VERSION };
|
||||||
const std::vector<const char*> layers = { "VK_LAYER_KHRONOS_validation" };
|
const std::vector<const char*> layers = { "VK_LAYER_KHRONOS_validation" };
|
||||||
vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags(), &app_info, layers.size(), layers.data());
|
vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags(), &app_info, layers.size(), layers.data());
|
||||||
instance = vk::createInstance(instance_create_info);
|
vk_instance = vk::createInstance(instance_create_info);
|
||||||
|
|
||||||
physical_device = instance.enumeratePhysicalDevices()[dev_num];
|
vk_physical_device = vk_instance.enumeratePhysicalDevices()[dev_num];
|
||||||
vk::PhysicalDeviceProperties device_props = physical_device.getProperties();
|
vk::PhysicalDeviceProperties device_props = vk_physical_device.getProperties();
|
||||||
std::cout << "Picked: " << device_props.deviceName << std::endl;
|
std::cout << "ggml_vulkan: Using " << device_props.deviceName << std::endl;
|
||||||
|
|
||||||
std::vector<vk::QueueFamilyProperties> queue_family_props = physical_device.getQueueFamilyProperties();
|
std::vector<vk::QueueFamilyProperties> queue_family_props = vk_physical_device.getQueueFamilyProperties();
|
||||||
auto prop_it = std::find_if(queue_family_props.begin(), queue_family_props.end(), [](const vk::QueueFamilyProperties& prop)
|
auto prop_it = std::find_if(queue_family_props.begin(), queue_family_props.end(), [](const vk::QueueFamilyProperties& prop)
|
||||||
{
|
{
|
||||||
return prop.queueFlags & vk::QueueFlagBits::eCompute;
|
return prop.queueFlags & vk::QueueFlagBits::eCompute;
|
||||||
});
|
});
|
||||||
const uint32_t compute_queue_family_index = std::distance(queue_family_props.begin(), prop_it);
|
vk_compute_queue_family_index = std::distance(queue_family_props.begin(), prop_it);
|
||||||
|
|
||||||
const float queue_priority = 1.0f;
|
const float queue_priority = 1.0f;
|
||||||
vk::DeviceQueueCreateInfo device_queue_create_info(vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, &queue_priority);
|
vk::DeviceQueueCreateInfo device_queue_create_info(vk::DeviceQueueCreateFlags(), vk_compute_queue_family_index, 1, &queue_priority);
|
||||||
vk::DeviceCreateInfo device_create_info(vk::DeviceCreateFlags(), device_queue_create_info);
|
vk::DeviceCreateInfo device_create_info(vk::DeviceCreateFlags(), device_queue_create_info);
|
||||||
vk_device = physical_device.createDevice(device_create_info);
|
vk_device = vk_physical_device.createDevice(device_create_info);
|
||||||
|
|
||||||
std::vector<char> matmul_shader_contents;
|
std::vector<char> matmul_shader_contents;
|
||||||
if (std::ifstream shader_file{ "ggml-vulkan-matmul.spv", std::ios::binary | std::ios::ate }) {
|
if (std::ifstream shader_file{ "ggml-vulkan-matmul.spv", std::ios::binary | std::ios::ate }) {
|
||||||
|
|
@ -74,9 +69,9 @@ void ggml_vk_init(void) {
|
||||||
vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
|
vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
|
||||||
vk::DescriptorSetLayoutCreateFlags(),
|
vk::DescriptorSetLayoutCreateFlags(),
|
||||||
descriptor_set_layout_binding);
|
descriptor_set_layout_binding);
|
||||||
vk::DescriptorSetLayout descriptor_set_layout = vk_device.createDescriptorSetLayout(descriptor_set_layout_create_info);
|
vk_pipeline_matmul_dsl = vk_device.createDescriptorSetLayout(descriptor_set_layout_create_info);
|
||||||
|
|
||||||
vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), descriptor_set_layout);
|
vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), vk_pipeline_matmul_dsl);
|
||||||
vk::PipelineLayout pipeline_layout = vk_device.createPipelineLayout(pipeline_layout_create_info);
|
vk::PipelineLayout pipeline_layout = vk_device.createPipelineLayout(pipeline_layout_create_info);
|
||||||
vk::PipelineCache pipeline_cache = vk_device.createPipelineCache(vk::PipelineCacheCreateInfo());
|
vk::PipelineCache pipeline_cache = vk_device.createPipelineCache(vk::PipelineCacheCreateInfo());
|
||||||
|
|
||||||
|
|
@ -92,151 +87,112 @@ void ggml_vk_init(void) {
|
||||||
vk_pipeline_matmul = vk_device.createComputePipeline(pipeline_cache, compute_pipeline_create_info).value;
|
vk_pipeline_matmul = vk_device.createComputePipeline(pipeline_cache, compute_pipeline_create_info).value;
|
||||||
}
|
}
|
||||||
|
|
||||||
// static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) {
|
// buffer pool for vulkan
|
||||||
// if (req_size <= *cur_size) {
|
#define MAX_VK_BUFFERS 256
|
||||||
// return;
|
|
||||||
// }
|
struct scoped_spin_lock {
|
||||||
//
|
std::atomic_flag& lock;
|
||||||
// // Reallocate buffer with enough space
|
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
|
||||||
// if (*cur_size > 0) {
|
while (lock.test_and_set(std::memory_order_acquire)) {
|
||||||
// clReleaseMemObject(*buf);
|
; // spin
|
||||||
// }
|
}
|
||||||
// cl_int err;
|
}
|
||||||
// *buf = clCreateBuffer(context, flags, req_size, NULL, &err);
|
~scoped_spin_lock() {
|
||||||
// *cur_size = req_size;
|
lock.clear(std::memory_order_release);
|
||||||
// CL_CHECK(err, "clCreateBuffer");
|
}
|
||||||
// }
|
scoped_spin_lock(const scoped_spin_lock&) = delete;
|
||||||
//
|
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
|
||||||
// void ggml_cl_sgemm_wrapper(
|
};
|
||||||
// const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b,
|
|
||||||
// const int m, const int n, const int k,
|
struct vk_buffer {
|
||||||
// const float alpha, const void *host_a, const int lda,
|
vk::Buffer buffer;
|
||||||
// const float *host_b, const int ldb, const float beta,
|
size_t size = 0;
|
||||||
// float *host_c, const int ldc, const int btype) {
|
};
|
||||||
// cl_int err = 0;
|
|
||||||
//
|
static vk_buffer g_vk_buffer_pool[MAX_VK_BUFFERS];
|
||||||
// cl_kernel kernel;
|
static std::atomic_flag g_vk_pool_lock = ATOMIC_FLAG_INIT;
|
||||||
// size_t global = n * k, local, size_qb;
|
|
||||||
// bool dequant;
|
static vk::Buffer ggml_vk_pool_malloc(size_t size, size_t * actual_size) {
|
||||||
// cl_block_q5_0* cl_host_b;
|
scoped_spin_lock lock(g_vk_pool_lock);
|
||||||
//
|
|
||||||
// switch (btype) {
|
int best_i = -1;
|
||||||
// case GGML_TYPE_F32:
|
size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
|
||||||
// dequant = false;
|
int worst_i = -1;
|
||||||
// break;
|
size_t worst_size = 0; //largest unused buffer seen so far
|
||||||
// case GGML_TYPE_Q4_0:
|
for (int i = 0; i < MAX_VK_BUFFERS; ++i) {
|
||||||
// dequant = true;
|
vk_buffer &b = g_vk_buffer_pool[i];
|
||||||
// kernel = kernel_q4_0;
|
if (b.size > 0 && b.size >= size && b.size < best_size)
|
||||||
// local = 16;
|
{
|
||||||
// size_qb = global * (sizeof(float) + local) / 32;
|
best_i = i;
|
||||||
// break;
|
best_size = b.size;
|
||||||
// case GGML_TYPE_Q4_1:
|
}
|
||||||
// dequant = true;
|
if (b.size > 0 && b.size > worst_size)
|
||||||
// kernel = kernel_q4_1;
|
{
|
||||||
// local = 16;
|
worst_i = i;
|
||||||
// size_qb = global * (sizeof(float) * 2 + local) / 32;
|
worst_size = b.size;
|
||||||
// break;
|
}
|
||||||
// case GGML_TYPE_Q4_2:
|
}
|
||||||
// dequant = true;
|
if(best_i!=-1) //found the smallest buffer that fits our needs
|
||||||
// kernel = kernel_q4_2;
|
{
|
||||||
// local = 8;
|
vk_buffer& b = g_vk_buffer_pool[best_i];
|
||||||
// size_qb = global * (sizeof(ggml_fp16_t) + local) / 16;
|
vk::Buffer buffer = b.buffer;
|
||||||
// break;
|
*actual_size = b.size;
|
||||||
// case GGML_TYPE_Q5_0:
|
b.size = 0;
|
||||||
// dequant = true;
|
return buffer;
|
||||||
// kernel = kernel_q5_0;
|
}
|
||||||
// local = 16;
|
if(worst_i!=-1) //no buffer that fits our needs, resize largest one to save memory
|
||||||
// // For some reason OpenCL seems to be incapable of working with structs of size 22.
|
{
|
||||||
// // 20 and 24 bytes are fine. Workaround to do the fp16 to fp32 step on CPU...
|
vk_buffer& b = g_vk_buffer_pool[worst_i];
|
||||||
// // TODO Find the reason, fix and remove workaround.
|
vk::Buffer buffer = b.buffer;
|
||||||
// const block_q5_0* b = (const block_q5_0*) host_b;
|
b.size = 0;
|
||||||
// cl_host_b = (cl_block_q5_0*) malloc(sizeof(cl_block_q5_0) * global / 32);
|
// vkReleaseMemObject(buffer);
|
||||||
// for (size_t i = 0; i < global / 32; i++) {
|
}
|
||||||
// cl_host_b[i].d = ggml_fp16_to_fp32(b[i].d);
|
vk::Buffer buffer;
|
||||||
// memcpy(&cl_host_b[i].qh, b[i].qh, sizeof(uint32_t));
|
|
||||||
// memcpy(&cl_host_b[i].qs, b[i].qs, QK5_0 / 2);
|
vk::BufferCreateInfo buffer_create_info{
|
||||||
// }
|
vk::BufferCreateFlags(),
|
||||||
// host_b = (const float*) cl_host_b;
|
size,
|
||||||
// size_qb = global * (sizeof(float) + sizeof(uint32_t) + local) / 32;
|
vk::BufferUsageFlagBits::eStorageBuffer,
|
||||||
// break;
|
vk::SharingMode::eExclusive,
|
||||||
// case GGML_TYPE_Q5_1:
|
1,
|
||||||
// dequant = true;
|
&vk_compute_queue_family_index
|
||||||
// kernel = kernel_q5_1;
|
};
|
||||||
// local = 16;
|
|
||||||
// size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
|
VmaAllocatorCreateInfo allocator_info = {};
|
||||||
// break;
|
allocator_info.vulkanApiVersion = VK_API_VERSION;
|
||||||
// case GGML_TYPE_Q8_0:
|
allocator_info.physicalDevice = vk_physical_device;
|
||||||
// dequant = true;
|
allocator_info.device = vk_device;
|
||||||
// kernel = kernel_q8_0;
|
allocator_info.instance = vk_instance;
|
||||||
// local = 32;
|
|
||||||
// size_qb = global * (sizeof(float) + local) / 32;
|
VmaAllocator allocator;
|
||||||
// break;
|
vmaCreateAllocator(&allocator_info, &allocator);
|
||||||
// default:
|
|
||||||
// fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype);
|
VmaAllocationCreateInfo allocation_info = {};
|
||||||
// abort();
|
allocation_info.usage = VMA_MEMORY_USAGE_CPU_TO_GPU;
|
||||||
// }
|
|
||||||
//
|
VmaAllocation buffer_allocation;
|
||||||
// const size_t size_a = m * k * sizeof(float);
|
vmaCreateBuffer(allocator,
|
||||||
// const size_t size_b = n * k * sizeof(float);
|
&static_cast<VkBufferCreateInfo>(buffer_create_info),
|
||||||
// const size_t size_c = m * n * sizeof(float);
|
&allocation_info,
|
||||||
//
|
&static_cast<VkBuffer>(buffer),
|
||||||
// // Prepare buffers
|
&buffer_allocation,
|
||||||
// ggml_cl_malloc(size_a, &cl_size_a, CL_MEM_READ_ONLY, &cl_buffer_a);
|
nullptr);
|
||||||
// if (dequant) {
|
|
||||||
// ggml_cl_malloc(size_qb, &cl_size_qb, CL_MEM_READ_ONLY, &cl_buffer_qb);
|
*actual_size = size;
|
||||||
// }
|
return buffer;
|
||||||
// ggml_cl_malloc(size_b, &cl_size_b, CL_MEM_READ_WRITE, &cl_buffer_b);
|
}
|
||||||
// ggml_cl_malloc(size_c, &cl_size_c, CL_MEM_WRITE_ONLY, &cl_buffer_c);
|
|
||||||
//
|
static void ggml_vk_pool_free(vk::Buffer buffer, size_t size) {
|
||||||
// cl_event ev_a, ev_qb, ev_b;
|
scoped_spin_lock lock(g_vk_pool_lock);
|
||||||
//
|
|
||||||
// if (dequant) {
|
for (int i = 0; i < MAX_VK_BUFFERS; ++i) {
|
||||||
// err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb);
|
vk_buffer& b = g_vk_buffer_pool[i];
|
||||||
// err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b);
|
if (b.size == 0) {
|
||||||
// CL_CHECK(err, "clSetKernelArg");
|
b.buffer = buffer;
|
||||||
// err = clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
|
b.size = size;
|
||||||
// CL_CHECK(err, "clEnqueueWriteBuffer qb");
|
return;
|
||||||
// } else {
|
}
|
||||||
// err = clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
|
}
|
||||||
// CL_CHECK(err, "clEnqueueWriteBuffer b");
|
fprintf(stderr, "WARNING: vk buffer pool full, increase MAX_VK_BUFFERS\n");
|
||||||
// }
|
vkReleaseMemObject(mem);
|
||||||
//
|
}
|
||||||
// err = clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a);
|
|
||||||
// CL_CHECK(err, "clEnqueueWriteBuffer a");
|
|
||||||
// if (dequant) {
|
|
||||||
// err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b);
|
|
||||||
// CL_CHECK(err, "clEnqueueNDRangeKernel");
|
|
||||||
// clReleaseEvent(ev_qb);
|
|
||||||
// }
|
|
||||||
// clWaitForEvents(1, &ev_a);
|
|
||||||
// clWaitForEvents(1, &ev_b);
|
|
||||||
// clReleaseEvent(ev_a);
|
|
||||||
// clReleaseEvent(ev_b);
|
|
||||||
//
|
|
||||||
// cl_event ev_sgemm;
|
|
||||||
// CLBlastStatusCode status = CLBlastSgemm((CLBlastLayout)order,
|
|
||||||
// (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
|
|
||||||
// m, n, k,
|
|
||||||
// alpha,
|
|
||||||
// cl_buffer_a, 0, lda,
|
|
||||||
// cl_buffer_b, 0, ldb,
|
|
||||||
// beta,
|
|
||||||
// cl_buffer_c, 0, ldc,
|
|
||||||
// &queue, &ev_sgemm);
|
|
||||||
//
|
|
||||||
// if (status != CLBlastSuccess) {
|
|
||||||
// fprintf(stderr, "Error: CLBlast SGEMM %d\n", status);
|
|
||||||
// abort();
|
|
||||||
// }
|
|
||||||
//
|
|
||||||
// cl_event ev_c;
|
|
||||||
// clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c);
|
|
||||||
//
|
|
||||||
// // Wait for completion
|
|
||||||
// clWaitForEvents(1, &ev_c);
|
|
||||||
// clReleaseEvent(ev_sgemm);
|
|
||||||
// clReleaseEvent(ev_c);
|
|
||||||
// if (btype == GGML_TYPE_Q5_0) {
|
|
||||||
// free((void*) cl_host_b);
|
|
||||||
// }
|
|
||||||
// }
|
|
||||||
|
|
|
||||||
Loading…
Add table
Add a link
Reference in a new issue