Merge pull request #1385 from yingcong-wu/yc/new-api-suggestgroupsize

Implement urKernelGetSuggestedLocalWorkSize

include/ur_api.h

@@ -224,6 +224,7 @@ typedef enum ur_function_t {
     UR_FUNCTION_COMMAND_BUFFER_COMMAND_GET_INFO_EXP = 222,                     ///< Enumerator for ::urCommandBufferCommandGetInfoExp
     UR_FUNCTION_ENQUEUE_TIMESTAMP_RECORDING_EXP = 223,                         ///< Enumerator for ::urEnqueueTimestampRecordingExp
     UR_FUNCTION_ENQUEUE_KERNEL_LAUNCH_CUSTOM_EXP = 224,                        ///< Enumerator for ::urEnqueueKernelLaunchCustomExp
+    UR_FUNCTION_KERNEL_GET_SUGGESTED_LOCAL_WORK_SIZE = 225,                    ///< Enumerator for ::urKernelGetSuggestedLocalWorkSize
     /// @cond
     UR_FUNCTION_FORCE_UINT32 = 0x7fffffff
     /// @endcond
@@ -5230,6 +5231,43 @@ urKernelCreateWithNativeHandle(
     ur_kernel_handle_t *phKernel                      ///< [out] pointer to the handle of the kernel object created.
 );
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Get the suggested local work size for a kernel.
+///
+/// @details
+///     - Query a suggested local work size for a kernel given a global size for
+///       each dimension.
+///     - The application may call this function from simultaneous threads for
+///       the same context.
+///
+/// @returns
+///     - ::UR_RESULT_SUCCESS
+///     - ::UR_RESULT_ERROR_UNINITIALIZED
+///     - ::UR_RESULT_ERROR_DEVICE_LOST
+///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
+///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
+///         + `NULL == hKernel`
+///         + `NULL == hQueue`
+///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
+///         + `NULL == pGlobalWorkOffset`
+///         + `NULL == pGlobalWorkSize`
+///         + `NULL == pSuggestedLocalWorkSize`
+///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
+UR_APIEXPORT ur_result_t UR_APICALL
+urKernelGetSuggestedLocalWorkSize(
+    ur_kernel_handle_t hKernel,      ///< [in] handle of the kernel
+    ur_queue_handle_t hQueue,        ///< [in] handle of the queue object
+    uint32_t numWorkDim,             ///< [in] number of dimensions, from 1 to 3, to specify the global
+                                     ///< and work-group work-items
+    const size_t *pGlobalWorkOffset, ///< [in] pointer to an array of numWorkDim unsigned values that specify
+                                     ///< the offset used to calculate the global ID of a work-item
+    const size_t *pGlobalWorkSize,   ///< [in] pointer to an array of numWorkDim unsigned values that specify
+                                     ///< the number of global work-items in workDim that will execute the
+                                     ///< kernel function
+    size_t *pSuggestedLocalWorkSize  ///< [out] pointer to an array of numWorkDim unsigned values that specify
+                                     ///< suggested local work size that will contain the result of the query
+);
+
 #if !defined(__GNUC__)
 #pragma endregion
 #endif
@@ -9943,6 +9981,19 @@ typedef struct ur_kernel_create_with_native_handle_params_t {
     ur_kernel_handle_t **pphKernel;
 } ur_kernel_create_with_native_handle_params_t;
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Function parameters for urKernelGetSuggestedLocalWorkSize
+/// @details Each entry is a pointer to the parameter passed to the function;
+///     allowing the callback the ability to modify the parameter's value
+typedef struct ur_kernel_get_suggested_local_work_size_params_t {
+    ur_kernel_handle_t *phKernel;
+    ur_queue_handle_t *phQueue;
+    uint32_t *pnumWorkDim;
+    const size_t **ppGlobalWorkOffset;
+    const size_t **ppGlobalWorkSize;
+    size_t **ppSuggestedLocalWorkSize;
+} ur_kernel_get_suggested_local_work_size_params_t;
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Function parameters for urKernelSetArgValue
 /// @details Each entry is a pointer to the parameter passed to the function;

include/ur_ddi.h

@@ -535,6 +535,16 @@ typedef ur_result_t(UR_APICALL *ur_pfnKernelCreateWithNativeHandle_t)(
     const ur_kernel_native_properties_t *,
     ur_kernel_handle_t *);
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Function-pointer for urKernelGetSuggestedLocalWorkSize
+typedef ur_result_t(UR_APICALL *ur_pfnKernelGetSuggestedLocalWorkSize_t)(
+    ur_kernel_handle_t,
+    ur_queue_handle_t,
+    uint32_t,
+    const size_t *,
+    const size_t *,
+    size_t *);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Function-pointer for urKernelSetArgValue
 typedef ur_result_t(UR_APICALL *ur_pfnKernelSetArgValue_t)(
@@ -603,6 +613,7 @@ typedef struct ur_kernel_dditable_t {
     ur_pfnKernelRelease_t pfnRelease;
     ur_pfnKernelGetNativeHandle_t pfnGetNativeHandle;
     ur_pfnKernelCreateWithNativeHandle_t pfnCreateWithNativeHandle;
+    ur_pfnKernelGetSuggestedLocalWorkSize_t pfnGetSuggestedLocalWorkSize;
     ur_pfnKernelSetArgValue_t pfnSetArgValue;
     ur_pfnKernelSetArgLocal_t pfnSetArgLocal;
     ur_pfnKernelSetArgPointer_t pfnSetArgPointer;

include/ur_print.h

@@ -1442,6 +1442,14 @@ UR_APIEXPORT ur_result_t UR_APICALL urPrintKernelGetNativeHandleParams(const str
 ///         - `buff_size < out_size`
 UR_APIEXPORT ur_result_t UR_APICALL urPrintKernelCreateWithNativeHandleParams(const struct ur_kernel_create_with_native_handle_params_t *params, char *buffer, const size_t buff_size, size_t *out_size);
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Print ur_kernel_get_suggested_local_work_size_params_t struct
+/// @returns
+///     - ::UR_RESULT_SUCCESS
+///     - ::UR_RESULT_ERROR_INVALID_SIZE
+///         - `buff_size < out_size`
+UR_APIEXPORT ur_result_t UR_APICALL urPrintKernelGetSuggestedLocalWorkSizeParams(const struct ur_kernel_get_suggested_local_work_size_params_t *params, char *buffer, const size_t buff_size, size_t *out_size);
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Print ur_kernel_set_arg_value_params_t struct
 /// @returns

include/ur_print.hpp

@@ -929,6 +929,9 @@ inline std::ostream &operator<<(std::ostream &os, enum ur_function_t value) {
     case UR_FUNCTION_ENQUEUE_KERNEL_LAUNCH_CUSTOM_EXP:
         os << "UR_FUNCTION_ENQUEUE_KERNEL_LAUNCH_CUSTOM_EXP";
         break;
+    case UR_FUNCTION_KERNEL_GET_SUGGESTED_LOCAL_WORK_SIZE:
+        os << "UR_FUNCTION_KERNEL_GET_SUGGESTED_LOCAL_WORK_SIZE";
+        break;
     default:
         os << "unknown enumerator";
         break;
@@ -11462,6 +11465,49 @@ inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct
     return os;
 }
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Print operator for the ur_kernel_get_suggested_local_work_size_params_t type
+/// @returns
+///     std::ostream &
+inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct ur_kernel_get_suggested_local_work_size_params_t *params) {
+
+    os << ".hKernel = ";
+
+    ur::details::printPtr(os,
+                          *(params->phKernel));
+
+    os << ", ";
+    os << ".hQueue = ";
+
+    ur::details::printPtr(os,
+                          *(params->phQueue));
+
+    os << ", ";
+    os << ".numWorkDim = ";
+
+    os << *(params->pnumWorkDim);
+
+    os << ", ";
+    os << ".pGlobalWorkOffset = ";
+
+    ur::details::printPtr(os,
+                          *(params->ppGlobalWorkOffset));
+
+    os << ", ";
+    os << ".pGlobalWorkSize = ";
+
+    ur::details::printPtr(os,
+                          *(params->ppGlobalWorkSize));
+
+    os << ", ";
+    os << ".pSuggestedLocalWorkSize = ";
+
+    ur::details::printPtr(os,
+                          *(params->ppSuggestedLocalWorkSize));
+
+    return os;
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief Print operator for the ur_kernel_set_arg_value_params_t type
 /// @returns
@@ -17143,6 +17189,9 @@ inline ur_result_t UR_APICALL printFunctionParams(std::ostream &os, ur_function_
     case UR_FUNCTION_KERNEL_CREATE_WITH_NATIVE_HANDLE: {
         os << (const struct ur_kernel_create_with_native_handle_params_t *)params;
     } break;
+    case UR_FUNCTION_KERNEL_GET_SUGGESTED_LOCAL_WORK_SIZE: {
+        os << (const struct ur_kernel_get_suggested_local_work_size_params_t *)params;
+    } break;
     case UR_FUNCTION_KERNEL_SET_ARG_VALUE: {
         os << (const struct ur_kernel_set_arg_value_params_t *)params;
     } break;

scripts/core/kernel.yml

@@ -534,3 +534,44 @@ params:
 returns:
     - $X_RESULT_ERROR_UNSUPPORTED_FEATURE:
         - "If the adapter has no underlying equivalent handle."
+--- #--------------------------------------------------------------------------
+type: function
+desc: "Get the suggested local work size for a kernel."
+class: $xKernel
+name: GetSuggestedLocalWorkSize
+ordinal: "0"
+details:
+    - "Query a suggested local work size for a kernel given a global size for each dimension."
+    - "The application may call this function from simultaneous threads for the same context."
+params:
+    - type: $x_kernel_handle_t
+      name: hKernel
+      desc: |
+            [in] handle of the kernel
+    - type: $x_queue_handle_t
+      name: hQueue
+      desc: |
+            [in] handle of the queue object
+    - type: uint32_t
+      name: numWorkDim
+      desc: |
+            [in] number of dimensions, from 1 to 3, to specify the global
+            and work-group work-items
+    - type: const size_t*
+      name: pGlobalWorkOffset
+      desc: |
+            [in] pointer to an array of numWorkDim unsigned values that specify
+            the offset used to calculate the global ID of a work-item
+    - type: const size_t*
+      name: pGlobalWorkSize      
+      desc: |
+            [in] pointer to an array of numWorkDim unsigned values that specify
+            the number of global work-items in workDim that will execute the 
+            kernel function
+    - type: size_t*
+      name: pSuggestedLocalWorkSize
+      desc: |
+            [out] pointer to an array of numWorkDim unsigned values that specify
+            suggested local work size that will contain the result of the query
+returns:
+    - $X_RESULT_ERROR_UNSUPPORTED_FEATURE

scripts/core/registry.yml

@@ -586,6 +586,9 @@ etors:
 - name: ENQUEUE_KERNEL_LAUNCH_CUSTOM_EXP
   desc: Enumerator for $xEnqueueKernelLaunchCustomExp
   value: '224'
+- name: KERNEL_GET_SUGGESTED_LOCAL_WORK_SIZE
+  desc: Enumerator for $xKernelGetSuggestedLocalWorkSize
+  value: '225'
 ---
 type: enum
 desc: Defines structure types

source/adapters/cuda/enqueue.hpp

@@ -17,6 +17,10 @@ ur_result_t enqueueEventsWait(ur_queue_handle_t CommandQueue, CUstream Stream,
                               uint32_t NumEventsInWaitList,
                               const ur_event_handle_t *EventWaitList);
 
+void guessLocalWorkSize(ur_device_handle_t Device, size_t *ThreadsPerBlock,
+                        const size_t *GlobalWorkSize, const uint32_t WorkDim,
+                        ur_kernel_handle_t Kernel);
+
 bool hasExceededMaxRegistersPerBlock(ur_device_handle_t Device,
                                      ur_kernel_handle_t Kernel,
                                      size_t BlockSize);

source/adapters/cuda/kernel.cpp

@@ -9,7 +9,9 @@
 //===----------------------------------------------------------------------===//
 
 #include "kernel.hpp"
+#include "enqueue.hpp"
 #include "memory.hpp"
+#include "queue.hpp"
 #include "sampler.hpp"
 
 UR_APIEXPORT ur_result_t UR_APICALL
@@ -380,3 +382,30 @@ urKernelSetArgSampler(ur_kernel_handle_t hKernel, uint32_t argIndex,
   }
   return Result;
 }
+
+UR_APIEXPORT ur_result_t UR_APICALL urKernelGetSuggestedLocalWorkSize(
+    ur_kernel_handle_t hKernel, ur_queue_handle_t hQueue, uint32_t workDim,
+    [[maybe_unused]] const size_t *pGlobalWorkOffset,
+    const size_t *pGlobalWorkSize, size_t *pSuggestedLocalWorkSize) {
+  // Preconditions
+  UR_ASSERT(hQueue->getContext() == hKernel->getContext(),
+            UR_RESULT_ERROR_INVALID_KERNEL);
+  UR_ASSERT(workDim > 0, UR_RESULT_ERROR_INVALID_WORK_DIMENSION);
+  UR_ASSERT(workDim < 4, UR_RESULT_ERROR_INVALID_WORK_DIMENSION);
+  UR_ASSERT(pSuggestedLocalWorkSize != nullptr,
+            UR_RESULT_ERROR_INVALID_NULL_POINTER);
+
+  ur_device_handle_t Device = hQueue->Device;
+  ur_result_t Result = UR_RESULT_SUCCESS;
+  size_t ThreadsPerBlock[3] = {};
+
+  // Set the active context here as guessLocalWorkSize needs an active context
+  ScopedContext Active(Device);
+
+  guessLocalWorkSize(Device, ThreadsPerBlock, pGlobalWorkSize, workDim,
+                     hKernel);
+
+  std::copy(ThreadsPerBlock, ThreadsPerBlock + workDim,
+            pSuggestedLocalWorkSize);
+  return Result;
+}

source/adapters/cuda/ur_interface_loader.cpp

@@ -125,6 +125,7 @@ UR_DLLEXPORT ur_result_t UR_APICALL urGetKernelProcAddrTable(
   pDdiTable->pfnSetArgValue = urKernelSetArgValue;
   pDdiTable->pfnSetExecInfo = urKernelSetExecInfo;
   pDdiTable->pfnSetSpecializationConstants = nullptr;
+  pDdiTable->pfnGetSuggestedLocalWorkSize = urKernelGetSuggestedLocalWorkSize;
   return UR_RESULT_SUCCESS;
 }
 

source/adapters/hip/enqueue.hpp

@@ -30,3 +30,7 @@ void setCopyRectParams(ur_rect_region_t Region, const void *SrcPtr,
                        const hipMemoryType DstType, ur_rect_offset_t DstOffset,
                        size_t DstRowPitch, size_t DstSlicePitch,
                        hipMemcpy3DParms &Params);
+
+void guessLocalWorkSize(ur_device_handle_t Device, size_t *ThreadsPerBlock,
+                        const size_t *GlobalWorkSize, const uint32_t WorkDim,
+                        const size_t MaxThreadsPerBlock[3]);

source/adapters/hip/kernel.cpp

@@ -9,6 +9,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "kernel.hpp"
+#include "enqueue.hpp"
 #include "memory.hpp"
 #include "sampler.hpp"
 
@@ -349,3 +350,31 @@ UR_APIEXPORT ur_result_t UR_APICALL urKernelSetSpecializationConstants(
     [[maybe_unused]] const ur_specialization_constant_info_t *pSpecConstants) {
   return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
 }
+
+UR_APIEXPORT ur_result_t UR_APICALL urKernelGetSuggestedLocalWorkSize(
+    [[maybe_unused]] ur_kernel_handle_t hKernel, ur_queue_handle_t hQueue,
+    uint32_t workDim, [[maybe_unused]] const size_t *pGlobalWorkOffset,
+    const size_t *pGlobalWorkSize, size_t *pSuggestedLocalWorkSize) {
+  UR_ASSERT(hQueue->getContext() == hKernel->getContext(),
+            UR_RESULT_ERROR_INVALID_QUEUE);
+  UR_ASSERT(workDim > 0, UR_RESULT_ERROR_INVALID_WORK_DIMENSION);
+  UR_ASSERT(workDim < 4, UR_RESULT_ERROR_INVALID_WORK_DIMENSION);
+  UR_ASSERT(pSuggestedLocalWorkSize != nullptr,
+            UR_RESULT_ERROR_INVALID_NULL_POINTER);
+
+  size_t MaxThreadsPerBlock[3];
+  size_t ThreadsPerBlock[3] = {32u, 1u, 1u};
+
+  MaxThreadsPerBlock[0] = hQueue->Device->getMaxBlockDimX();
+  MaxThreadsPerBlock[1] = hQueue->Device->getMaxBlockDimY();
+  MaxThreadsPerBlock[2] = hQueue->Device->getMaxBlockDimZ();
+
+  ur_device_handle_t Device = hQueue->getDevice();
+  ScopedContext Active(Device);
+
+  guessLocalWorkSize(Device, ThreadsPerBlock, pGlobalWorkSize, workDim,
+                     MaxThreadsPerBlock);
+  std::copy(ThreadsPerBlock, ThreadsPerBlock + workDim,
+            pSuggestedLocalWorkSize);
+  return UR_RESULT_SUCCESS;
+}

source/adapters/hip/ur_interface_loader.cpp

@@ -125,6 +125,7 @@ UR_DLLEXPORT ur_result_t UR_APICALL urGetKernelProcAddrTable(
   pDdiTable->pfnSetArgValue = urKernelSetArgValue;
   pDdiTable->pfnSetExecInfo = urKernelSetExecInfo;
   pDdiTable->pfnSetSpecializationConstants = urKernelSetSpecializationConstants;
+  pDdiTable->pfnGetSuggestedLocalWorkSize = urKernelGetSuggestedLocalWorkSize;
   return UR_RESULT_SUCCESS;
 }