Skip to content

Commit

Permalink
Fix buffer alignment
Browse files Browse the repository at this point in the history
  • Loading branch information
christophe committed Sep 24, 2019
1 parent db5059a commit dd748b1
Showing 1 changed file with 47 additions and 33 deletions.
80 changes: 47 additions & 33 deletions src/miner.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -104,60 +104,61 @@ void initVulkanMiner(VulkanMiner &vulkanMiner,VkDevice vkDevice, CPUMiner cpuMin
vulkanMiner.scratchpadsSize2 = 64;
}

// Get memory alignment
VkDeviceMemory tmpMem = allocateGPUMemory(vulkanMiner.deviceId, vulkanMiner.vkDevice, 1024, true, true);
VkBuffer tmpBuf = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, tmpMem, 256, 0);
vulkanMiner.alignment = getBufferMemoryRequirements(vulkanMiner.vkDevice,tmpBuf);
vkDestroyBuffer(vulkanMiner.vkDevice,tmpBuf,nullptr);
vkFreeMemory(vkDevice , tmpMem, NULL);

// compute memory requirements
vulkanMiner.scratchpadsSize1 = alignBuffer(vulkanMiner.scratchpadsSize1,vulkanMiner.alignment);
vulkanMiner.scratchpadsSize2 = alignBuffer(vulkanMiner.scratchpadsSize2,vulkanMiner.alignment);
vulkanMiner.debugSize = alignBuffer(256 * sizeof(uint64_t),vulkanMiner.alignment);
vulkanMiner.local_memory_size = vulkanMiner.scratchpadsSize1 +
vulkanMiner.scratchpadsSize2 +
alignBuffer(vulkanMiner.stateSize * maxThreads,vulkanMiner.alignment) +
alignBuffer(4L * sizeof(int) * (maxThreads + 2L),vulkanMiner.alignment);
vulkanMiner.shared_memory_size = alignBuffer(sizeof(Params),vulkanMiner.alignment) +
alignBuffer(sizeof(GpuConstants),vulkanMiner.alignment) +
alignBuffer(vulkanMiner.inputsSize,vulkanMiner.alignment) +
alignBuffer(vulkanMiner.outputSize * sizeof(int64_t),vulkanMiner.alignment)+
alignBuffer(vulkanMiner.debugSize,vulkanMiner.alignment);
uint64_t memStateSize = alignBuffer(vulkanMiner.stateSize * maxThreads,vulkanMiner.alignment);
uint64_t memBranchSize = alignBuffer(4L * sizeof(int) * (maxThreads + 2L),vulkanMiner.alignment);
uint64_t memConstantSize = alignBuffer(sizeof(GpuConstants),vulkanMiner.alignment);
uint64_t memParamSize = alignBuffer(sizeof(Params),vulkanMiner.alignment);
uint64_t memInputSize = alignBuffer(vulkanMiner.inputsSize,vulkanMiner.alignment);
uint64_t memOutputSize = alignBuffer(vulkanMiner.outputSize * sizeof(int64_t),vulkanMiner.alignment);
vulkanMiner.local_memory_size = vulkanMiner.scratchpadsSize1 + vulkanMiner.scratchpadsSize2 + memStateSize +memBranchSize;
vulkanMiner.shared_memory_size = memParamSize + memConstantSize + memInputSize + memOutputSize+ vulkanMiner.debugSize;

vulkanMiner.gpuLocalMemory = allocateGPUMemory(deviceId, vulkanMiner.vkDevice, vulkanMiner.local_memory_size, true,true);
vulkanMiner.gpuSharedMemory = allocateGPUMemory(deviceId, vulkanMiner.vkDevice, vulkanMiner.shared_memory_size, false,true);
vulkanMiner.local_size_cn1 = local_size_cn1;

uint64_t o = 0;
// create the local buffers
// create the internal local buffers
vulkanMiner.gpu_scratchpadsBuffer1 = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuLocalMemory, vulkanMiner.scratchpadsSize1, o);
o += vulkanMiner.scratchpadsSize1;
vulkanMiner.gpu_scratchpadsBuffer2 = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuLocalMemory, vulkanMiner.scratchpadsSize2, o);
o += vulkanMiner.scratchpadsSize2;
vulkanMiner.gpu_statesBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuLocalMemory, vulkanMiner.stateSize * maxThreads, o);
o += alignBuffer(vulkanMiner.stateSize * maxThreads,vulkanMiner.alignment);
vulkanMiner.gpu_branchesBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuLocalMemory, 4 * sizeof(int) * (maxThreads + 2), o);
o += alignBuffer(4 * sizeof(int) * (maxThreads + 2),vulkanMiner.alignment);
vulkanMiner.gpu_statesBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuLocalMemory, memStateSize, o);
o += memStateSize;
vulkanMiner.gpu_branchesBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuLocalMemory, memBranchSize, o);
o += memBranchSize;

// create the CPU shared buffers
o = 0;
vulkanMiner.gpu_params = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, sizeof(Params), o);
o += alignBuffer(sizeof(Params),vulkanMiner.alignment);
vulkanMiner.gpu_constants = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, sizeof(GpuConstants), o);
o += alignBuffer(sizeof(GpuConstants),vulkanMiner.alignment);
vulkanMiner.gpu_inputsBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, vulkanMiner.inputsSize, o);
o += alignBuffer(vulkanMiner.inputsSize,vulkanMiner.alignment);
vulkanMiner.gpu_outputBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, vulkanMiner.outputSize * sizeof(int64_t), o);
o += alignBuffer(vulkanMiner.outputSize * sizeof(int64_t),vulkanMiner.alignment);
vulkanMiner.gpu_params = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, memParamSize, o);
o += memParamSize;
vulkanMiner.gpu_constants = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, memConstantSize, o);
o += memConstantSize;
vulkanMiner.gpu_inputsBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, memInputSize, o);
o += memInputSize;
vulkanMiner.gpu_outputBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, memOutputSize, o);
o += memOutputSize;
vulkanMiner.gpu_debugBuffer = createBuffer(vulkanMiner.vkDevice, computeQueueFamillyIndex, vulkanMiner.gpuSharedMemory, vulkanMiner.debugSize, o);
o += alignBuffer(vulkanMiner.debugSize,vulkanMiner.alignment);
o += vulkanMiner.debugSize;

vulkanMiner.pipelineLayout = bindBuffers(vulkanMiner.vkDevice, vulkanMiner.descriptorSet,vulkanMiner.descriptorPool,vulkanMiner.descriptorSetLayout,
vulkanMiner.gpu_scratchpadsBuffer1, vulkanMiner.gpu_scratchpadsBuffer2, vulkanMiner.gpu_statesBuffer,
vulkanMiner.gpu_branchesBuffer, vulkanMiner.gpu_params, vulkanMiner.gpu_constants, vulkanMiner.gpu_inputsBuffer, vulkanMiner.gpu_outputBuffer, vulkanMiner.gpu_debugBuffer);

// Transfer constants to GPU
void *ptr = NULL;
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory, sizeof(Params), sizeof(GpuConstants), 0, (void ** )&ptr), "vkMapMemory");
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory, memParamSize, memConstantSize, 0, (void ** )&ptr), "vkMapMemory");
memcpy(ptr, (const void*) &gpuConstants, sizeof(GpuConstants));
vkUnmapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory);

Expand Down Expand Up @@ -230,8 +231,8 @@ static void getParams(VulkanMiner &vulkanMiner,Params &params ) {
params.mask = 0x1FFF0;
}
if (vulkanMiner.cpuMiner.type == AeonCrypto && getVariant() == K12_ALGO) {
// guessing vulkanMiner.alignment is 1 on AMD and >1 on Nvidia
params.threads = vulkanMiner.groups[getCurrentIndex()]*(vulkanMiner.alignment == 4 ? K12_LOCAL_SIZE_AMD : K12_LOCAL_SIZE_NV) *4096;
// guessing vulkanMiner.alignment is 4 or 16 on AMD and > 16 on Nvidia
params.threads = vulkanMiner.groups[getCurrentIndex()]*(vulkanMiner.alignment <= 16 ? K12_LOCAL_SIZE_AMD : K12_LOCAL_SIZE_NV) *4096;
}
params.chunk2 = vulkanMiner.chunk2;
}
Expand Down Expand Up @@ -291,8 +292,8 @@ void loadSPIRV(VulkanMiner &vulkanMiner) {
vulkanMiner.pipeline_cn6 = loadShader(vulkanMiner.vkDevice, vulkanMiner.pipelineLayout,vulkanMiner.shader_module, "spirv/cn6.spv");
vulkanMiner.pipeline_cn7 = loadShader(vulkanMiner.vkDevice, vulkanMiner.pipelineLayout,vulkanMiner.shader_module, "spirv/cn7.spv");
if (vulkanMiner.cpuMiner.type == AeonCrypto)
vulkanMiner.pipeline_k12 = loadShader(vulkanMiner.vkDevice, vulkanMiner.pipelineLayout,vulkanMiner.shader_module, vulkanMiner.alignment == 4 ? "spirv/k12_amd.spv" : "spirv/k12_nv.spv");
//shaderStats(vulkanMiner.vkDevice,vulkanMiner.pipeline_k12); exit(0);
vulkanMiner.pipeline_k12 = loadShader(vulkanMiner.vkDevice, vulkanMiner.pipelineLayout,vulkanMiner.shader_module, vulkanMiner.alignment <= 16 ? "spirv/k12_amd.spv" : "spirv/k12_nv.spv");
//shaderStats(vulkanMiner.vkDevice,vulkanMiner.pipeline_cn7); exit(0);
}

// Specific command buffer for K12 algo (SHA3)
Expand Down Expand Up @@ -324,6 +325,7 @@ static void createCommandBuffer(VulkanMiner &vulkanMiner) {
vkCmdBindDescriptorSets(vulkanMiner.vkCommandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, vulkanMiner.pipelineLayout, 0, 1, &vulkanMiner.descriptorSet, 0, 0);
vkCmdDispatch(vulkanMiner.vkCommandBuffer, 1, 1, 1);
vkCmdPipelineBarrier(vulkanMiner.vkCommandBuffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 1, &vulkanMiner.memoryBarrier, 0, nullptr, 0, nullptr);

vulkanMiner.nonce += vulkanMiner.threads[current_index]; // nonce is incremented during those buffer reset

vkCmdBindPipeline(vulkanMiner.vkCommandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, vulkanMiner.pipeline_cn0);
Expand Down Expand Up @@ -395,11 +397,12 @@ void minerIterate(VulkanMiner &vulkanMiner) {
} while (res == VK_TIMEOUT);
vkResetFences(vulkanMiner.vkDevice, 1, &vulkanMiner.drawFence);
vulkanMiner.nrResults = vulkanMiner.resultPtr[0] < vulkanMiner.outputSize ? vulkanMiner.resultPtr[0] : vulkanMiner.outputSize-1;

for (int i = 0; i < vulkanMiner.nrResults; i++) {
vulkanMiner.tmpResults[i] = vulkanMiner.resultPtr[i+1];
}
vulkanMiner.cnrSubmittedHeight = vulkanMiner.height;
memcpy(&vulkanMiner.originalInput,&vulkanMiner.input,vulkanMiner.inputLen);
memcpy(&vulkanMiner.originalInput,&vulkanMiner.input,MAX_BLOB_SIZE/2);
}

void reloadInput(VulkanMiner &vulkanMiner, int64_t nonce) {
Expand All @@ -410,7 +413,11 @@ void reloadInput(VulkanMiner &vulkanMiner, int64_t nonce) {

// transfer blob to GPU
char *ptr = NULL;
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory, sizeof(Params)+sizeof(GpuConstants), MAX_BLOB_SIZE/2, 0, (void **)&ptr),"vkMapMemory");
uint64_t memConstantSize = alignBuffer(sizeof(GpuConstants),vulkanMiner.alignment);
uint64_t memParamSize = alignBuffer(sizeof(Params),vulkanMiner.alignment);
uint64_t memInputSize = alignBuffer(vulkanMiner.inputsSize,vulkanMiner.alignment);
uint64_t tfxOrigin = memParamSize+memConstantSize;
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory, tfxOrigin, memInputSize, 0, (void **)&ptr),"vkMapMemory");
memcpy(ptr,(const void*)vulkanMiner.input,MAX_BLOB_SIZE/2);
vkUnmapMemory(vulkanMiner.vkDevice,vulkanMiner.gpuSharedMemory);

Expand All @@ -424,13 +431,20 @@ void sendMiningParameters(VulkanMiner &vulkanMiner) {
getParams(vulkanMiner,params);
char *ptr = NULL;
unmapMiningResults(vulkanMiner);
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory, 0, sizeof(Params), 0, (void **)&ptr),"vkMapMemory");
uint64_t memParamSize = alignBuffer(sizeof(Params),vulkanMiner.alignment);
uint64_t tfxSize = memParamSize;
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory, 0, tfxSize, 0, (void **)&ptr),"vkMapMemory");
memcpy(ptr,(const void*)&params,sizeof(Params));
vkUnmapMemory(vulkanMiner.vkDevice,vulkanMiner.gpuSharedMemory);
}

void mapMiningResults(VulkanMiner &vulkanMiner) {
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory,sizeof(Params)+sizeof(GpuConstants)+vulkanMiner.inputsSize, vulkanMiner.outputSize*sizeof(int), 0, (void **)&vulkanMiner.resultPtr),"vkMapMemory");
uint64_t memConstantSize = alignBuffer(sizeof(GpuConstants),vulkanMiner.alignment);
uint64_t memParamSize = alignBuffer(sizeof(Params),vulkanMiner.alignment);
uint64_t memInputSize = alignBuffer(vulkanMiner.inputsSize,vulkanMiner.alignment);
uint64_t memOutputSize = alignBuffer(vulkanMiner.outputSize * sizeof(int64_t),vulkanMiner.alignment);
uint64_t tfxOrigin = memParamSize+memConstantSize+memInputSize;
CHECK_RESULT_NORET(vkMapMemory(vulkanMiner.vkDevice, vulkanMiner.gpuSharedMemory,tfxOrigin, memOutputSize, 0, (void **)&vulkanMiner.resultPtr),"vkMapMemory");
}

void unmapMiningResults(VulkanMiner &vulkanMiner) {
Expand Down Expand Up @@ -634,7 +648,7 @@ void *MinerThread(void *args)
minerIterate(miner);
hashRates[miner.index] = 1e9*(float)miner.threads[getCurrentIndex()] / (float)(now() - t0);
if (getVariant() == K12_ALGO)
hashRates[miner.index] *= (miner.alignment == 4 ? K12_LOCAL_SIZE_AMD : K12_LOCAL_SIZE_NV)/16*4096; // KangarooTwelve
hashRates[miner.index] *= (miner.alignment <= 16 ? K12_LOCAL_SIZE_AMD : K12_LOCAL_SIZE_NV)/16*4096; // KangarooTwelve
t0 = now();
}
unmapMiningResults(miner);
Expand Down

0 comments on commit dd748b1

Please sign in to comment.