1 /*
2  * Copyright (C) 2018-2021 Intel Corporation
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  */
7 
8 #include "driver_diagnostics.h"
9 
10 #include "shared/source/helpers/debug_helpers.h"
11 
12 namespace NEO {
13 
DriverDiagnostics(cl_diagnostics_verbose_level level)14 DriverDiagnostics::DriverDiagnostics(cl_diagnostics_verbose_level level) {
15     this->verboseLevel = level;
16 }
17 
validFlags(cl_diagnostics_verbose_level flags) const18 bool DriverDiagnostics::validFlags(cl_diagnostics_verbose_level flags) const {
19     return !!(verboseLevel & flags);
20 }
21 
22 const char *const DriverDiagnostics::hintFormat[] = {
23     "Performance hint: clCreateBuffer with pointer %p and size %u doesn't meet alignment restrictions. Size should be aligned to %u bytes and pointer should be aligned to %u. Buffer is not sharing the same physical memory with CPU.", //CL_BUFFER_DOESNT_MEET_ALIGNMENT_RESTRICTIONS
24     "Performance hint: clCreateBuffer with pointer %p and size %u meets alignment restrictions and buffer will share the same physical memory with CPU.",                                                                                 //CL_BUFFER_MEETS_ALIGNMENT_RESTRICTIONS
25     "Performance hint: clCreateBuffer needs to allocate memory for buffer. For subsequent operations the buffer will share the same physical memory with CPU.",                                                                           //CL_BUFFER_NEEDS_ALLOCATE_MEMORY
26     "Performance hint: clCreateImage with pointer %p meets alignment restrictions and image will share the same physical memory with CPU.",                                                                                               //CL_IMAGE_MEETS_ALIGNMENT_RESTRICTIONS
27     "Performance hint: Driver calls internal clFlush on the command queue each time 1 command is enqueued.",                                                                                                                              //DRIVER_CALLS_INTERNAL_CL_FLUSH
28     "Performance hint: Profiling adds overhead on all enqueue commands with events.",                                                                                                                                                     //PROFILING_ENABLED
29     "Performance hint: Profiled kernels will be executed with disabled preemption.",                                                                                                                                                      //PROFILING_ENABLED_WITH_DISABLED_PREEMPTION
30     "Performance hint: Subbuffer created from buffer %p shares the same memory with buffer.",                                                                                                                                             //SUBBUFFER_SHARES_MEMORY
31     "Performance hint: clSVMAlloc with pointer %p and size %u meets alignment restrictions.",                                                                                                                                             //CL_SVM_ALLOC_MEETS_ALIGNMENT_RESTRICTIONS
32     "Performance hint: clEnqueueReadBuffer call on a buffer %p with pointer %p will require driver to copy the data.Consider using clEnqueueMapBuffer with buffer that shares the same physical memory with CPU.",                        //CL_ENQUEUE_READ_BUFFER_REQUIRES_COPY_DATA
33     "Performance hint: clEnqueueReadBuffer call on a buffer %p with pointer %p will not require any data copy as the buffer shares the same physical memory with CPU.",                                                                   //CL_ENQUEUE_READ_BUFFER_DOESNT_REQUIRE_COPY_DATA
34     "Performance hint: Pointer %p and size %u passed to clEnqueueReadBuffer doesn't meet alignment restrictions. Size should be aligned to %u bytes and pointer should be aligned to %u. Driver needs to disable L3 caching.",            //CL_ENQUEUE_READ_BUFFER_DOESNT_MEET_ALIGNMENT_RESTRICTIONS
35     "Performance hint: clEnqueueReadBufferRect call on a buffer %p with pointer %p will require driver to copy the data.Consider using clEnqueueMapBuffer with buffer that shares the same physical memory with CPU.",                    //CL_ENQUEUE_READ_BUFFER_RECT_REQUIRES_COPY_DATA
36     "Performance hint: clEnqueueReadBufferRect call on a buffer %p with pointer %p will not require any data copy as the buffer shares the same physical memory with CPU.",                                                               //CL_ENQUEUE_READ_BUFFER_RECT_DOESNT_REQUIRES_COPY_DATA
37     "Performance hint: Pointer %p and size %u passed to clEnqueueReadBufferRect doesn't meet alignment restrictions. Size should be aligned to %u bytes and pointer should be aligned to %u. Driver needs to disable L3 caching.",        //CL_ENQUEUE_READ_BUFFER_RECT_DOESNT_MEET_ALIGNMENT_RESTRICTIONS
38     "Performance hint: clEnqueueWriteBuffer call on a buffer %p require driver to copy the data. Consider using clEnqueueMapBuffer with buffer that shares the same physical memory with CPU.",                                           //CL_ENQUEUE_WRITE_BUFFER_REQUIRES_COPY_DATA
39     "Performance hint: clEnqueueWriteBuffer call on a buffer %p with pointer %p will not require any data copy as the buffer shares the same physical memory with CPU.",                                                                  //CL_ENQUEUE_WRITE_BUFFER_DOESNT_REQUIRE_COPY_DATA
40     "Performance hint: clEnqueueWriteBufferRect call on a buffer %p require driver to copy the data. Consider using clEnqueueMapBuffer with buffer that shares the same physical memory with CPU.",                                       //CL_ENQUEUE_WRITE_BUFFER_RECT_REQUIRES_COPY_DATA
41     "Performance hint: clEnqueueWriteBufferRect call on a buffer %p will not require any data copy as the buffer shares the same physical memory with CPU.",                                                                              //CL_ENQUEUE_WRITE_BUFFER_RECT_DOESNT_REQUIRE_COPY_DATA
42     "Performance hint: Pointer %p and size %u passed to clEnqueueReadImage doesn't meet alignment restrictions. Size should be aligned to %u bytes and pointer should be aligned to %u. Driver needs to disable L3 caching.",             //CL_ENQUEUE_READ_IMAGE_DOESNT_MEET_ALIGNMENT_RESTRICTIONS
43     "Performance hint: clEnqueueReadImage call on an image %p will not require any data copy as the image shares the same physical memory with CPU.",                                                                                     //CL_ENQUEUE_READ_IMAGE_DOESNT_REQUIRES_COPY_DATA
44     "Performance hint: clEnqueueWriteImage call on an image %p require driver to copy the data.",                                                                                                                                         //CL_ENQUEUE_WRITE_IMAGE_REQUIRES_COPY_DATA
45     "Performance hint: clEnqueueWriteImage call on an image %p will not require any data copy as the image shares the same physical memory with CPU.",                                                                                    //CL_ENQUEUE_WRITE_IMAGE_DOESNT_REQUIRES_COPY_DATA
46     "Performance hint: clEnqueueMapBuffer call on a buffer %p will require driver to make a copy as buffer is not sharing the same physical memory with CPU.",                                                                            //CL_ENQUEUE_MAP_BUFFER_REQUIRES_COPY_DATA
47     "Performance hint: clEnqueueMapBuffer call on a buffer %p will not require any data copy as buffer shares the same physical memory with CPU.",                                                                                        //CL_ENQUEUE_MAP_BUFFER_DOESNT_REQUIRE_COPY_DATA
48     "Performance hint: clEnqueueMapImage call on an image %p will require driver to make a copy, as image is not sharing the same physical memory with CPU.",                                                                             //CL_ENQUEUE_MAP_IMAGE_REQUIRES_COPY_DATA
49     "Performance hint: clEnqueueMapImage call on an image %p will not require any data copy as image shares the same physical memory with CPU.",                                                                                          //CL_ENQUEUE_MAP_IMAGE_DOESNT_REQUIRE_COPY_DATA
50     "Performance hint: clEnqueueUnmapMemObject call with pointer %p will not require any data copy.",                                                                                                                                     //CL_ENQUEUE_UNMAP_MEM_OBJ_DOESNT_REQUIRE_COPY_DATA
51     "Performance hint: clEnqueueUnmapMemObject call with pointer %p will require driver to copy the data to memory object %p.",                                                                                                           //CL_ENQUEUE_UNMAP_MEM_OBJ_REQUIRES_COPY_DATA
52     "Performance hint: clEnqueueSVMMap call with pointer %p will not require any data copy.",                                                                                                                                             //CL_ENQUEUE_SVM_MAP_DOESNT_REQUIRE_COPY_DATA
53     "Performance hint: Printf detected in kernel %s, it may cause overhead.",                                                                                                                                                             //PRINTF_DETECTED_IN_KERNEL
54     "Performance hint: Null local workgroup size detected ( kernel name: %s ); following sizes will be used for execution : { %u, %u, %u }.",                                                                                             //NULL_LOCAL_WORKGROUP_SIZE
55     "Performance hint: Local workgroup sizes { %u, %u, %u } selected for this workload ( kernel name: %s ) may not be optimal, consider using following local workgroup size: { %u, %u, %u }.",                                           //BAD_LOCAL_WORKGROUP_SIZE
56     "Performance hint: Kernel %s register pressure is too high, spill fills will be generated, additional surface needs to be allocated of size %u, consider simplifying your kernel.",                                                   //REGISTER_PRESSURE_TOO_HIGH
57     "Performance hint: Kernel %s private memory usage is too high and exhausts register space, additional surface needs to be allocated of size %u, consider reducing amount of private memory used, avoid using private memory arrays.", //PRIVATE_MEMORY_USAGE_TOO_HIGH
58     "Performance hint: Kernel %s submission requires coherency with CPU; this will impact performance.",                                                                                                                                  //KERNEL_REQUIRES_COHERENCY
59     "Performance hint: Kernel %s requires aux translation on argument [%u] = \"%s\"",                                                                                                                                                     //KERNEL_ARGUMENT_AUX_TRANSLATION
60     "Performance hint: Kernel %s requires aux translation for allocation with pointer %p and size %u",                                                                                                                                    //KERNEL_ALLOCATION_AUX_TRANSLATION
61     "Performance hint: Buffer %p will use compressed memory.",                                                                                                                                                                            //BUFFER_IS_COMPRESSED
62     "Performance hint: Buffer %p will not use compressed memory.",                                                                                                                                                                        //BUFFER_IS_NOT_COMPRESSED
63     "Performance hint: Image %p will use compressed memory.",                                                                                                                                                                             //IMAGE_IS_COMPRESSED
64     "Performance hint: Image %p will not use compressed memory."};                                                                                                                                                                        //IMAGE_IS_NOT_COMPRESSED
65 
obtainHintForTransferOperation(cl_command_type commandType,bool transferRequired)66 PerformanceHints DriverDiagnostics::obtainHintForTransferOperation(cl_command_type commandType, bool transferRequired) {
67     PerformanceHints hint;
68     switch (commandType) {
69     case CL_COMMAND_MAP_BUFFER:
70         hint = transferRequired ? CL_ENQUEUE_MAP_BUFFER_REQUIRES_COPY_DATA : CL_ENQUEUE_MAP_BUFFER_DOESNT_REQUIRE_COPY_DATA;
71         break;
72     case CL_COMMAND_MAP_IMAGE:
73         hint = transferRequired ? CL_ENQUEUE_MAP_IMAGE_REQUIRES_COPY_DATA : CL_ENQUEUE_MAP_IMAGE_DOESNT_REQUIRE_COPY_DATA;
74         break;
75     case CL_COMMAND_UNMAP_MEM_OBJECT:
76         hint = transferRequired ? CL_ENQUEUE_UNMAP_MEM_OBJ_REQUIRES_COPY_DATA : CL_ENQUEUE_UNMAP_MEM_OBJ_DOESNT_REQUIRE_COPY_DATA;
77         break;
78     case CL_COMMAND_WRITE_BUFFER:
79         hint = transferRequired ? CL_ENQUEUE_WRITE_BUFFER_REQUIRES_COPY_DATA : CL_ENQUEUE_WRITE_BUFFER_DOESNT_REQUIRE_COPY_DATA;
80         break;
81     case CL_COMMAND_READ_BUFFER:
82         hint = transferRequired ? CL_ENQUEUE_READ_BUFFER_REQUIRES_COPY_DATA : CL_ENQUEUE_READ_BUFFER_DOESNT_REQUIRE_COPY_DATA;
83         break;
84     case CL_COMMAND_WRITE_BUFFER_RECT:
85         hint = transferRequired ? CL_ENQUEUE_WRITE_BUFFER_RECT_REQUIRES_COPY_DATA : CL_ENQUEUE_WRITE_BUFFER_RECT_DOESNT_REQUIRE_COPY_DATA;
86         break;
87     case CL_COMMAND_READ_BUFFER_RECT:
88         hint = transferRequired ? CL_ENQUEUE_READ_BUFFER_RECT_REQUIRES_COPY_DATA : CL_ENQUEUE_READ_BUFFER_RECT_DOESNT_REQUIRES_COPY_DATA;
89         break;
90     case CL_COMMAND_WRITE_IMAGE:
91         hint = transferRequired ? CL_ENQUEUE_WRITE_IMAGE_REQUIRES_COPY_DATA : CL_ENQUEUE_WRITE_IMAGE_DOESNT_REQUIRES_COPY_DATA;
92         break;
93     case CL_COMMAND_READ_IMAGE:
94         UNRECOVERABLE_IF(transferRequired)
95         hint = CL_ENQUEUE_READ_IMAGE_DOESNT_REQUIRES_COPY_DATA;
96         break;
97     default:
98         UNRECOVERABLE_IF(true);
99     }
100     return hint;
101 }
102 } // namespace NEO
103