issue/808: v0.1.0增加QY编译选项

README.md

@@ -50,6 +50,7 @@ python scripts/install.py [XMAKE_CONFIG_FLAGS]
 | `--metax-gpu=[y\|n]`     | 是否编译沐曦 GPU 接口实现         | n
 | `--moore-gpu=[y\|n]`     | 是否编译摩尔线程 GPU 接口实现     | n
 | `--iluvatar-gpu=[y\|n]`  | 是否编译沐曦 GPU 接口实现         | n
+| `--qy-gpu=[y\|n]`        | 是否编QY GPU 接口实现             | n
 | `--hygon-dcu=[y\|n]`     | 是否编译海光 DCU 接口实现         | n
 | `--kunlun-xpu=[y\|n]`    | 是否编译昆仑 XPU 接口实现         | n
 | `--ninetoothed=[y\|n]`   | 是否编译九齿实现                 | n

include/infinicore.h

@@ -46,6 +46,7 @@ typedef enum {
     INFINI_DEVICE_ILUVATAR = 6,
     INFINI_DEVICE_KUNLUN = 7,
     INFINI_DEVICE_HYGON = 8,
+    INFINI_DEVICE_QY = 9,
     INFINI_DEVICE_TYPE_COUNT
 } infiniDevice_t;
 

src/infiniccl-test/main.cpp

@@ -12,7 +12,7 @@ void printUsage() {
     std::cout << "infiniccl-test --<device>" << std::endl
               << std::endl;
     std::cout << "  --<device>" << std::endl;
-    std::cout << "    Specify the device type --(nvidia|cambricon|ascend|metax|moore|iluvatar|kunlun|hygon)." << std::endl
+    std::cout << "    Specify the device type --(nvidia|qy|cambricon|ascend|metax|moore|iluvatar|kunlun|hygon)." << std::endl
               << std::endl;
     std::cout << "The program will run tests on all visible devices of the specified device type."
               << " Use Environmental Variables such as CUDA_VSIBLE_DEVICES to limit visible device IDs.";
@@ -45,6 +45,7 @@ ParsedArgs parseArgs(int argc, char *argv[]) {
         else PARSE_DEVICE("--iluvatar", INFINI_DEVICE_ILUVATAR)
         else PARSE_DEVICE("--kunlun", INFINI_DEVICE_KUNLUN)
         else PARSE_DEVICE("--hygon", INFINI_DEVICE_HYGON)
+        else PARSE_DEVICE("--qy", INFINI_DEVICE_QY)
         else {
             printUsage();
         }

src/infiniccl/cuda/infiniccl_cuda.h

@@ -4,7 +4,7 @@
 #include "../infiniccl_impl.h"
 
 // Windows does not support CUDA
-#if (defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API)) && defined(ENABLE_CCL) && !defined(_WIN32)
+#if (defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API) || defined(ENABLE_QY_API)) && defined(ENABLE_CCL) && !defined(_WIN32)
 INFINICCL_DEVICE_API_IMPL(cuda)
 #else
 INFINICCL_DEVICE_API_NOOP(cuda)

src/infiniccl/infiniccl.cc

@@ -26,6 +26,7 @@ __C infiniStatus_t infinicclCommInitAll(
         COMM_INIT_ALL(INFINI_DEVICE_METAX, metax);
         COMM_INIT_ALL(INFINI_DEVICE_MOORE, moore);
         COMM_INIT_ALL(INFINI_DEVICE_KUNLUN, kunlun);
+        COMM_INIT_ALL(INFINI_DEVICE_QY, cuda);
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
     }
@@ -51,6 +52,7 @@ __C infiniStatus_t infinicclCommDestroy(infinicclComm_t comm) {
         COMM_DESTROY(INFINI_DEVICE_METAX, metax);
         COMM_DESTROY(INFINI_DEVICE_MOORE, moore);
         COMM_DESTROY(INFINI_DEVICE_KUNLUN, kunlun);
+        COMM_DESTROY(INFINI_DEVICE_QY, cuda);
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
     }
@@ -83,6 +85,7 @@ __C infiniStatus_t infinicclAllReduce(
         ALL_REDUCE(INFINI_DEVICE_METAX, metax);
         ALL_REDUCE(INFINI_DEVICE_MOORE, moore);
         ALL_REDUCE(INFINI_DEVICE_KUNLUN, kunlun);
+        ALL_REDUCE(INFINI_DEVICE_QY, cuda);
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;

src/infiniop-test/src/main.cpp

@@ -22,7 +22,7 @@ void printUsage() {
     std::cout << "    Path to the test gguf file" << std::endl
               << std::endl;
     std::cout << "  --<device>[:id]" << std::endl;
-    std::cout << "    (Optional) Specify the device type --(cpu|nvidia|cambricon|ascend|metax|moore|iluvatar|kunlun|hygon) and device ID (optional). CPU by default." << std::endl
+    std::cout << "    (Optional) Specify the device type --(cpu|nvidia|qy|cambricon|ascend|metax|moore|iluvatar|kunlun|hygon) and device ID (optional). CPU by default." << std::endl
               << std::endl;
     std::cout << "  --warmup <warmups>" << std::endl;
     std::cout << "    (Optional) Number of warmups to perform before timing. Default to 0." << std::endl
@@ -79,6 +79,7 @@ ParsedArgs parseArgs(int argc, char *argv[]) {
             PARSE_DEVICE("--iluvatar", INFINI_DEVICE_ILUVATAR)
             PARSE_DEVICE("--kunlun", INFINI_DEVICE_KUNLUN)
             PARSE_DEVICE("--hygon", INFINI_DEVICE_HYGON)
+            PARSE_DEVICE("--qy", INFINI_DEVICE_QY)
             else if (arg == "--warmup" && i + 1 < argc) {
                 args.warmups = std::stoi(argv[++i]);
             }

src/infiniop/devices/handle.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/cpu_handle.h"
 #endif
-#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API) || defined(ENABLE_QY_API)
 #include "nvidia/nvidia_handle.h"
 #endif
 #ifdef ENABLE_CAMBRICON_API
@@ -44,6 +44,9 @@ __C infiniStatus_t infiniopCreateHandle(infiniopHandle_t *handle_ptr) {
 #ifdef ENABLE_NVIDIA_API
         CREATE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, qy);
+#endif
 #ifdef ENABLE_ILUVATAR_API
         CREATE(INFINI_DEVICE_ILUVATAR, iluvatar);
 #endif
@@ -87,6 +90,9 @@ __C infiniStatus_t infiniopDestroyHandle(infiniopHandle_t handle) {
 #ifdef ENABLE_NVIDIA_API
         DELETE(INFINI_DEVICE_NVIDIA, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        DELETE(INFINI_DEVICE_QY, qy);
+#endif
 #ifdef ENABLE_ILUVATAR_API
         DELETE(INFINI_DEVICE_ILUVATAR, iluvatar);
 #endif

src/infiniop/devices/nvidia/nvidia_common.cu

@@ -65,8 +65,10 @@ cudnnDataType_t getCudnnDtype(infiniDtype_t dt) {
         return CUDNN_DATA_HALF;
     case INFINI_DTYPE_F32:
         return CUDNN_DATA_FLOAT;
+#ifndef ENABLE_QY_API
     case INFINI_DTYPE_F64:
         return CUDNN_DATA_DOUBLE;
+#endif
     case INFINI_DTYPE_BF16:
         return CUDNN_DATA_BFLOAT16;
     case INFINI_DTYPE_I8:
@@ -116,4 +118,16 @@ infiniStatus_t Handle::create(InfiniopHandle **handle_ptr, int device_id) {
 
 } // namespace hygon
 
+namespace qy {
+
+Handle::Handle(int device_id)
+    : nvidia::Handle(INFINI_DEVICE_QY, device_id) {}
+
+infiniStatus_t Handle::create(InfiniopHandle **handle_ptr, int device_id) {
+    *handle_ptr = new Handle(device_id);
+    return INFINI_STATUS_SUCCESS;
+}
+
+} // namespace qy
+
 } // namespace device

src/infiniop/devices/nvidia/nvidia_handle.h

@@ -46,6 +46,17 @@ struct Handle : public nvidia::Handle {
 
 } // namespace hygon
 
+namespace qy {
+
+struct Handle : public nvidia::Handle {
+    Handle(int device_id);
+
+public:
+    static infiniStatus_t create(InfiniopHandle **handle_ptr, int device_id);
+};
+
+} // namespace qy
+
 } // namespace device
 
 #endif // __INFINIOP_CUDA_HANDLE_H__

src/infiniop/devices/nvidia/nvidia_kernel_common.cuh

@@ -50,7 +50,7 @@ exp_(const float val) {
     return expf(val);
 }
 
-#if !defined(ENABLE_ILUVATAR_API) && !defined(ENABLE_HYGON_API)
+#if !defined(ENABLE_ILUVATAR_API) && !defined(ENABLE_QY_API) && !defined(ENABLE_HYGON_API)
 __forceinline__ __device__ long double
 exp_(const long double val) {
     return expl(val);

src/infiniop/ops/add/operator.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/add_cpu.h"
 #endif
-#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API)
 #include "nvidia/add_nvidia.cuh"
 #endif
 #ifdef ENABLE_METAX_API
@@ -45,6 +45,9 @@ __C infiniStatus_t infiniopCreateAddDescriptor(
 #ifdef ENABLE_ILUVATAR_API
         CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CREATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -79,6 +82,9 @@ __C infiniStatus_t infiniopGetAddWorkspaceSize(infiniopAddDescriptor_t desc, siz
 #ifdef ENABLE_ILUVATAR_API
         GET(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         GET(INFINI_DEVICE_METAX, metax);
 #endif
@@ -121,6 +127,9 @@ __C infiniStatus_t infiniopAdd(
 #ifdef ENABLE_ILUVATAR_API
         CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CALCULATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -157,6 +166,9 @@ infiniopDestroyAddDescriptor(infiniopAddDescriptor_t desc) {
 #ifdef ENABLE_ILUVATAR_API
         DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        DELETE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         DELETE(INFINI_DEVICE_METAX, metax);
 #endif

src/infiniop/ops/causal_softmax/operator.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/causal_softmax_cpu.h"
 #endif
-#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_HYGON_API) || defined(ENABLE_QY_API)
 #include "nvidia/causal_softmax_nvidia.cuh"
 #endif
 #ifdef ENABLE_METAX_API
@@ -48,6 +48,9 @@ __C infiniStatus_t infiniopCreateCausalSoftmaxDescriptor(
 #ifdef ENABLE_ILUVATAR_API
         CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_HYGON_API
         CREATE(INFINI_DEVICE_HYGON, nvidia);
 #endif
@@ -87,6 +90,9 @@ __C infiniStatus_t infiniopGetCausalSoftmaxWorkspaceSize(infiniopCausalSoftmaxDe
 #ifdef ENABLE_ILUVATAR_API
         GET(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_HYGON_API
         GET(INFINI_DEVICE_HYGON, nvidia);
 #endif
@@ -131,6 +137,9 @@ __C infiniStatus_t infiniopCausalSoftmax(
 #ifdef ENABLE_ILUVATAR_API
         CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_HYGON_API
         CALCULATE(INFINI_DEVICE_HYGON, nvidia);
 #endif
@@ -170,6 +179,9 @@ __C infiniStatus_t infiniopDestroyCausalSoftmaxDescriptor(infiniopCausalSoftmaxD
 #ifdef ENABLE_ILUVATAR_API
         DESTROY(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        DESTROY(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_HYGON_API
         DESTROY(INFINI_DEVICE_HYGON, nvidia);
 #endif

src/infiniop/ops/clip/operator.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/clip_cpu.h"
 #endif
-#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API)
 #include "nvidia/clip_nvidia.cuh"
 #endif
 #ifdef ENABLE_METAX_API
@@ -42,6 +42,9 @@ __C infiniStatus_t infiniopCreateClipDescriptor(
 #ifdef ENABLE_ILUVATAR_API
         CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CREATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -73,6 +76,9 @@ __C infiniStatus_t infiniopGetClipWorkspaceSize(infiniopClipDescriptor_t desc, s
 #ifdef ENABLE_ILUVATAR_API
         GET(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         GET(INFINI_DEVICE_METAX, metax)
 #endif
@@ -112,6 +118,9 @@ __C infiniStatus_t infiniopClip(
 #ifdef ENABLE_ILUVATAR_API
         CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         CALCULATE(INFINI_DEVICE_METAX, metax);
 #endif
@@ -145,6 +154,9 @@ infiniopDestroyClipDescriptor(infiniopClipDescriptor_t desc) {
 #ifdef ENABLE_ILUVATAR_API
         DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        DELETE(INFINI_DEVICE_QY, nvidia);
+#endif
 #ifdef ENABLE_METAX_API
         DELETE(INFINI_DEVICE_METAX, metax);
 #endif

src/infiniop/ops/conv/operator.cc

@@ -5,7 +5,7 @@
 #ifdef ENABLE_CPU_API
 #include "cpu/conv_cpu.h"
 #endif
-#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API)
+#if defined(ENABLE_NVIDIA_API) || defined(ENABLE_ILUVATAR_API) || defined(ENABLE_QY_API)
 #include "nvidia/conv_nvidia.cuh"
 #endif
 
@@ -42,6 +42,9 @@ __C __export infiniStatus_t infiniopCreateConvDescriptor(infiniopHandle_t handle
 #ifdef ENABLE_ILUVATAR_API
         CREATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CREATE(INFINI_DEVICE_QY, nvidia);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
@@ -70,6 +73,9 @@ infiniopGetConvWorkspaceSize(
 #ifdef ENABLE_ILUVATAR_API
         GET(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        GET(INFINI_DEVICE_QY, nvidia);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
@@ -106,6 +112,9 @@ __C infiniStatus_t infiniopConv(
 #ifdef ENABLE_ILUVATAR_API
         CALCULATE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        CALCULATE(INFINI_DEVICE_QY, nvidia);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;
@@ -130,6 +139,9 @@ infiniopDestroyConvDescriptor(infiniopConvDescriptor_t desc) {
 #ifdef ENABLE_ILUVATAR_API
         DELETE(INFINI_DEVICE_ILUVATAR, nvidia);
 #endif
+#ifdef ENABLE_QY_API
+        DELETE(INFINI_DEVICE_QY, nvidia);
+#endif
 
     default:
         return INFINI_STATUS_DEVICE_TYPE_NOT_SUPPORTED;