diff --git a/algo/base/gpu/DeviceImage.h b/algo/base/gpu/DeviceImage.h
index 14a84da25e53aa41658544e5b84ce84a9b1cb910..523310d98647b5ba4242d060a2521dfce9cd927c 100644
--- a/algo/base/gpu/DeviceImage.h
+++ b/algo/base/gpu/DeviceImage.h
@@ -7,7 +7,7 @@
namespace cbm::algo
{
- struct GPUReco {
+ struct GPUReco : xpu::device_image {
};
} // namespace cbm::algo
diff --git a/algo/base/gpu/xpu_legacy.h b/algo/base/gpu/xpu_legacy.h
new file mode 100644
index 0000000000000000000000000000000000000000..861f15f2fe75c08d03d210f914c8ec4a9fd7f694
--- /dev/null
+++ b/algo/base/gpu/xpu_legacy.h
@@ -0,0 +1,103 @@
+/* Copyright (C) 2022 FIAS Frankfurt Institute for Advanced Studies, Frankfurt / Main
+ SPDX-License-Identifier: GPL-3.0-only
+ Authors: Felix Weiglhofer [committer]*/
+#ifndef CORE_COMPAT_XPU_LEGACY_H
+#define CORE_COMPAT_XPU_LEGACY_H
+
+#include <xpu/host.h>
+
+namespace xpu {
+
+inline constexpr auto host_to_device = xpu::h2d;
+inline constexpr auto device_to_host = xpu::d2h;
+
+template<typename T>
+class hd_buffer {
+
+public:
+ hd_buffer() = default;
+ hd_buffer(size_t size) : m_buffer(size, xpu::buf_io) {}
+
+ T *h() { return xpu::h_view(m_buffer).begin(); }
+ T *d() { return m_buffer.get(); }
+
+ xpu::buffer<T> &underlying() { return m_buffer; }
+
+private:
+ xpu::buffer<T> m_buffer;
+
+};
+
+template<typename T>
+class d_buffer {
+
+public:
+ d_buffer() = default;
+ d_buffer(size_t size) : m_buffer(size, xpu::buf_device) {}
+
+ T *d() { return m_buffer.get(); }
+
+ xpu::buffer<T> &underlying() { return m_buffer; }
+
+private:
+ xpu::buffer<T> m_buffer;
+
+};
+
+
+template<typename T>
+void copy(hd_buffer<T> &buf, direction dir) {
+ static xpu::queue _Q;
+ _Q.copy(buf.underlying(), dir);
+ _Q.wait();
+}
+
+enum class side {
+ host,
+ device
+};
+
+template<typename T, side S>
+struct cmem_io {
+ using type = T *;
+};
+
+template<typename T>
+struct cmem_io<T, side::host> {
+ using type = hd_buffer<T>;
+};
+
+template<typename T, side S>
+using cmem_io_t = typename cmem_io<T, S>::type;
+
+template<typename T, side S>
+struct cmem_device {
+ using type = T *;
+};
+
+template<typename T>
+struct cmem_device<T, side::host> {
+ using type = d_buffer<T>;
+};
+
+template<typename T, side S>
+using cmem_device_t = typename cmem_device<T, S>::type;
+
+} // namespace xpu
+
+#define XPU_BLOCK_SIZE_1D(...)
+
+#define XPU_EXPORT_KERNEL(Image, Kernel, ...) XPU_EXPORT_KERNEL_II(Image, Kernel, xpu::no_smem, 64, ##__VA_ARGS__)
+
+#define XPU_EXPORT_KERNEL_II(Image, Kernel, SMEM, BlockSize, ...) \
+ struct Kernel : xpu::kernel<Image> { \
+ using block_size = xpu::block_size<BlockSize>; \
+ using context = xpu::kernel_context<SMEM>; \
+ XPU_D void operator()(context &ctx, ##__VA_ARGS__); \
+ }
+
+#define XPU_KERNEL(Kernel, smemIgnored, ...) \
+ XPU_EXPORT(Kernel); \
+ XPU_D void Kernel::operator()(context &ctx, ##__VA_ARGS__)
+
+#endif
diff --git a/algo/detectors/sts/UnpackStsXpu.cxx b/algo/detectors/sts/UnpackStsXpu.cxx
index 7c9a5556359113b226e4906c21ffc66212403ee0..ab0dfcf45d32a27c15080b2e29f0d12653d9e46b 100644
--- a/algo/detectors/sts/UnpackStsXpu.cxx
+++ b/algo/detectors/sts/UnpackStsXpu.cxx
@@ -15,11 +15,83 @@
using std::unique_ptr;
using std::vector;
-XPU_BLOCK_SIZE_1D(cbm::algo::UnpackStsXpu::Unpack, 32);
+XPU_KERNEL(cbm::algo::UnpackK, xpu::no_smem, UnpackStsXpuPar* params, UnpackStsXpuElinkPar* elinkParams,
+ stsxyter::Message* content, uint64_t* msMessCount, uint64_t* msMessOffset, uint64_t* msStartTime,
+ uint32_t* msCompIdx, CbmStsDigi* digisOut, const uint64_t currentTsTime, int NElems)
+ {
+ int id = ctx.block_idx_x() * ctx.block_dim_x() + ctx.thread_idx_x();
+ if (id >= NElems || msMessCount[id] < 2) return; // exit if out of bounds or too few messages
+
+ UnpackStsXpuMonitorData monitor; //Monitor data, currently not stored. TO DO: Implement!
+
+ // --- Get message count and offset for this MS
+ const uint32_t numMessages = msMessCount[id];
+ const uint32_t messOffset = msMessOffset[id];
+
+ // --- Get starting position of this MS in message buffer
+ stsxyter::Message* message = &content[messOffset];
+
+ // --- Get starting position of this MS in digi buffer
+ CbmStsDigi* digis = &digisOut[messOffset];
+
+ // --- Get component index and unpack parameters of this MS
+ const uint32_t comp = msCompIdx[id];
+ const UnpackStsXpuPar& unpackPar = params[comp];
+
+ // --- Get starting position of elink parameters of this MS
+ UnpackStsXpuElinkPar* elinkPar = &elinkParams[unpackPar.fElinkOffset];
+
+ // --- Init counter for produced digis
+ uint64_t numDigis = 0;
+
+ // --- The first message in the MS is expected to be of type EPOCH and can be ignored.
+ if (message[0].GetMessType() != stsxyter::MessType::Epoch) {
+ monitor.fNumErrInvalidFirstMessage++;
+ msMessCount[id] = 0;
+ return;
+ }
+
+ // --- The second message must be of type ts_msb.
+ if (message[1].GetMessType() != stsxyter::MessType::TsMsb) {
+ monitor.fNumErrInvalidFirstMessage++;
+ msMessCount[id] = 0;
+ return;
+ }
+
+ // --- Current TS_MSB epoch cycle
+ uint64_t currentCycle = msStartTime[id] / UnpackStsXpu::fkCycleLength;
+
+ // --- Process first message (ts_msb)
+ uint32_t currentEpoch = 0; ///< Current epoch number within epoch cycle
+ uint64_t currentEpochTime = 0; ///< Current epoch time relative to timeslice in clock cycles
+ UnpackStsXpu::ProcessTsmsbMessage(message[1], currentEpoch, currentEpochTime, currentCycle, currentTsTime);
+
+ // --- Message loop
+ for (uint32_t messageNr = 2; messageNr < numMessages; messageNr++) {
+
+ // --- Action depending on message type
+ switch (message[messageNr].GetMessType()) {
+ case stsxyter::MessType::Hit: {
+ UnpackStsXpu::ProcessHitMessage(message[messageNr], digis, numDigis, unpackPar, elinkPar, monitor, currentEpochTime);
+ break;
+ }
+ case stsxyter::MessType::TsMsb: {
+ UnpackStsXpu::ProcessTsmsbMessage(message[messageNr], currentEpoch, currentEpochTime, currentCycle, currentTsTime);
+ break;
+ }
+ default: {
+ monitor.fNumNonHitOrTsbMessage++;
+ break;
+ }
+ }
+ }
+ // --- Store number of digis in buffer
+ msMessCount[id] = numDigis;
+ }
+
namespace cbm::algo
{
-
// ---- Algorithm execution ---------------------------------------------
UnpackStsXpu::resultType UnpackStsXpu::operator()(const fles::Timeslice* ts, StsReadoutConfig& config)
{
@@ -60,6 +132,8 @@ namespace cbm::algo
result.second.fNumErrInvalidMsSize++;
continue;
}
+ xpu::t_add_bytes(msDescr.size);
+ xpu::k_add_bytes<UnpackK>(msDescr.size);
msIdx.push_back(msDescr.idx);
compIdx.push_back(comp);
messCount.push_back(numMessages);
@@ -101,7 +175,7 @@ namespace cbm::algo
const uint64_t currentTsTime = ts->start_time() / epochLengthInNs;
// --- Do unpacking for each microslice
- xpu::run_kernel<Unpack>(xpu::grid::n_threads(numMs), fParams.d(), fElinkParams.d(), tsContent.d(), msMessCount.d(),
+ xpu::run_kernel<UnpackK>(xpu::n_threads(numMs), fParams.d(), fElinkParams.d(), tsContent.d(), msMessCount.d(),
msMessOffset.d(), msStartTime.d(), msCompIdx.d(), digisOut.d(), currentTsTime, numMs);
// --- Copy results back to host (only two buffers are modified on device)
@@ -109,6 +183,7 @@ namespace cbm::algo
xpu::copy(digisOut, xpu::device_to_host);
// --- Store digis TO DO: make Kernel for this, needs a way to sum arrays in XPU first
+ xpu::push_timer("Store digis");
for (uint64_t i = 0; i < numMs; i++) {
uint64_t offset = msMessOffset.h()[i];
uint64_t numDigis = msMessCount.h()[i];
@@ -116,85 +191,11 @@ namespace cbm::algo
result.first.push_back(digisOut.h()[offset + j]);
}
}
+ xpu::pop_timer();
return result;
}
- XPU_KERNEL(UnpackStsXpu::Unpack, xpu::no_smem, UnpackStsXpuPar* params, UnpackStsXpuElinkPar* elinkParams,
- stsxyter::Message* content, uint64_t* msMessCount, uint64_t* msMessOffset, uint64_t* msStartTime,
- uint32_t* msCompIdx, CbmStsDigi* digisOut, const uint64_t currentTsTime, int NElems)
- {
- int id = xpu::block_idx::x() * xpu::block_dim::x() + xpu::thread_idx::x();
- if (id >= NElems || msMessCount[id] < 2) return; // exit if out of bounds or too few messages
-
- UnpackStsXpuMonitorData monitor; //Monitor data, currently not stored. TO DO: Implement!
-
- // --- Get message count and offset for this MS
- const uint32_t numMessages = msMessCount[id];
- const uint32_t messOffset = msMessOffset[id];
-
- // --- Get starting position of this MS in message buffer
- stsxyter::Message* message = &content[messOffset];
-
- // --- Get starting position of this MS in digi buffer
- CbmStsDigi* digis = &digisOut[messOffset];
-
- // --- Get component index and unpack parameters of this MS
- const uint32_t comp = msCompIdx[id];
- const UnpackStsXpuPar& unpackPar = params[comp];
-
- // --- Get starting position of elink parameters of this MS
- UnpackStsXpuElinkPar* elinkPar = &elinkParams[unpackPar.fElinkOffset];
-
- // --- Init counter for produced digis
- uint64_t numDigis = 0;
-
- // --- The first message in the MS is expected to be of type EPOCH and can be ignored.
- if (message[0].GetMessType() != stsxyter::MessType::Epoch) {
- monitor.fNumErrInvalidFirstMessage++;
- msMessCount[id] = 0;
- return;
- }
-
- // --- The second message must be of type ts_msb.
- if (message[1].GetMessType() != stsxyter::MessType::TsMsb) {
- monitor.fNumErrInvalidFirstMessage++;
- msMessCount[id] = 0;
- return;
- }
-
- // --- Current TS_MSB epoch cycle
- uint64_t currentCycle = msStartTime[id] / fkCycleLength;
-
- // --- Process first message (ts_msb)
- uint32_t currentEpoch = 0; ///< Current epoch number within epoch cycle
- uint64_t currentEpochTime = 0; ///< Current epoch time relative to timeslice in clock cycles
- ProcessTsmsbMessage(message[1], currentEpoch, currentEpochTime, currentCycle, currentTsTime);
-
- // --- Message loop
- for (uint32_t messageNr = 2; messageNr < numMessages; messageNr++) {
-
- // --- Action depending on message type
- switch (message[messageNr].GetMessType()) {
- case stsxyter::MessType::Hit: {
- ProcessHitMessage(message[messageNr], digis, numDigis, unpackPar, elinkPar, monitor, currentEpochTime);
- break;
- }
- case stsxyter::MessType::TsMsb: {
- ProcessTsmsbMessage(message[messageNr], currentEpoch, currentEpochTime, currentCycle, currentTsTime);
- break;
- }
- default: {
- monitor.fNumNonHitOrTsbMessage++;
- break;
- }
- }
- }
- // --- Store number of digis in buffer
- msMessCount[id] = numDigis;
- }
-
-
// ----- Process hit message --------------------------------------------
XPU_D inline void UnpackStsXpu::ProcessHitMessage(const stsxyter::Message& message, CbmStsDigi* digis,
uint64_t& numDigis, const UnpackStsXpuPar& unpackPar,
diff --git a/algo/detectors/sts/UnpackStsXpu.h b/algo/detectors/sts/UnpackStsXpu.h
index 1e88b7f5ea08800f15ba89ed5a54991bbdf54e93..f9da48db38e4d02a827c838f80833a62c6dc9a62 100644
--- a/algo/detectors/sts/UnpackStsXpu.h
+++ b/algo/detectors/sts/UnpackStsXpu.h
@@ -5,7 +5,10 @@
#ifndef CBM_ALGO_UNPACKSTSXPU_H
#define CBM_ALGO_UNPACKSTSXPU_H 1
+
#include "CbmStsDigi.h"
+#include "gpu/DeviceImage.h"
+#include "gpu/xpu_legacy.h"
#include "MicrosliceDescriptor.hpp"
#include "Timeslice.hpp"
@@ -21,7 +24,6 @@
#include "StsReadoutConfig.h"
#include "StsXyterMessage.h"
-#include "gpu/DeviceImage.h"
namespace cbm::algo
@@ -73,6 +75,9 @@ namespace cbm::algo
}
};
+ XPU_EXPORT_KERNEL(GPUReco, UnpackK, UnpackStsXpuPar* params, UnpackStsXpuElinkPar* elinkParams,
+ stsxyter::Message* content, uint64_t* msMessCount, uint64_t* msMessOffset, uint64_t* msStartTime,
+ uint32_t* msCompIdx, CbmStsDigi* digisOut, const uint64_t currentTsTime, int NElems);
/** @class UnpackStsXpu
** @author Pierre-Alain Loizeau <p.-a.loizeau@gsi.de>
@@ -101,16 +106,6 @@ namespace cbm::algo
**/
resultType operator()(const fles::Timeslice* ts, StsReadoutConfig& config);
-
- struct StsXpuUnpack {
- }; // Identifier used by xpu to find where kernels are located
-
-
- // Run unpacker for each microslice
- XPU_EXPORT_KERNEL(GPUReco, Unpack, UnpackStsXpuPar* params, UnpackStsXpuElinkPar* elinkParams,
- stsxyter::Message* content, uint64_t* msMessCount, uint64_t* msMessOffset, uint64_t* msStartTime,
- uint32_t* msCompIdx, CbmStsDigi* digisOut, const uint64_t currentTsTime, int NElems);
-
//Stores parameter structs for all elinks
xpu::hd_buffer<UnpackStsXpuElinkPar> fElinkParams;
@@ -120,6 +115,8 @@ namespace cbm::algo
private: // methods
+ friend struct UnpackK;
+
/** @brief Process a hit message
** @param message SMX message (32-bit word)
** @param digi buffer
diff --git a/external/.gitignore b/external/.gitignore
index 732dd15370edb746477f1fc066877aaa4e0a8f25..74cb20f77bcff3c16dfba60a8bf1e0c0ac4feccd 100644
--- a/external/.gitignore
+++ b/external/.gitignore
@@ -11,5 +11,6 @@ jsroot
googletest
yaml-cpp/
xpu/
+xpu-dev
GSL
bba
diff --git a/external/CMakeLists.txt b/external/CMakeLists.txt
index abeed2bf6461baa432cec83c724ce4f90431a22e..0238ffa7b666d50d1c23af2b2655cc32b1983b03 100644
--- a/external/CMakeLists.txt
+++ b/external/CMakeLists.txt
@@ -52,7 +52,7 @@ if(DOWNLOAD_EXTERNALS)
if (NOT ${CBM_XPU_DEV})
download_project_if_needed(PROJECT xpu
GIT_REPOSITORY "https://github.com/fweig/xpu.git"
- GIT_TAG "dd7d1d7e4b4d71079b9cfadbd662a264288308b0" # v0.7.6
+ GIT_TAG "d142d9ac7135488925b5b94f36da22ec55d9271c" # v0.9.3
SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/xpu
CONFIGURE_COMMAND ""
BUILD_COMMAND ""