1 /*
2  * Copyright (C) 2018-2021 Intel Corporation
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  */
7 
8 #include "shared/source/helpers/basic_math.h"
9 #include "shared/source/helpers/string.h"
10 #include "shared/test/common/helpers/unit_test_helper.h"
11 #include "shared/test/common/test_macros/test.h"
12 
13 #include <algorithm>
14 #include <cstdint>
15 using namespace NEO;
16 
17 union GRF {
18     float fRegs[8];
19     uint32_t dwRegs[8];
20     uint16_t wRegs[16];
21 };
22 
23 // clang-format off
24 uint16_t gFlattenedIDsSimD8[] =
25     {
26         0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
27         0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
28         0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
29         0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
30         0x20, 0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
31         0x28, 0x29, 0x2a, 0x2b, 0x2c, 0x2d, 0x2e, 0x2f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
32         0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
33         0x38, 0x39, 0x3a, 0x3b, 0x3c, 0x3d, 0x3e, 0x3f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
34         0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
35         0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
36         0x50, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
37         0x58, 0x59, 0x5a, 0x5b, 0x5c, 0x5d, 0x5e, 0x5f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
38         0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
39         0x68, 0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
40         0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
41         0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
42         0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
43         0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
44         0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
45         0x98, 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
46         0xa0, 0xa1, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
47         0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
48         0xb0, 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
49         0xb8, 0xb9, 0xba, 0xbb, 0xbc, 0xbd, 0xbe, 0xbf, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
50         0xc0, 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc7, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
51         0xc8, 0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
52         0xd0, 0xd1, 0xd2, 0xd3, 0xd4, 0xd5, 0xd6, 0xd7, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
53         0xd8, 0xd9, 0xda, 0xdb, 0xdc, 0xdd, 0xde, 0xdf, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
54         0xe0, 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
55         0xe8, 0xe9, 0xea, 0xeb, 0xec, 0xed, 0xee, 0xef, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
56         0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
57         0xf8, 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
58 };
59 
60 uint16_t gFlattenedIDsSimD1632[] =
61     {
62         0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
63         0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
64         0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
65         0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f,
66         0x20, 0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27,
67         0x28, 0x29, 0x2a, 0x2b, 0x2c, 0x2d, 0x2e, 0x2f,
68         0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x36, 0x37,
69         0x38, 0x39, 0x3a, 0x3b, 0x3c, 0x3d, 0x3e, 0x3f,
70         0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47,
71         0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f,
72         0x50, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57,
73         0x58, 0x59, 0x5a, 0x5b, 0x5c, 0x5d, 0x5e, 0x5f,
74         0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67,
75         0x68, 0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f,
76         0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77,
77         0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f,
78         0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
79         0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f,
80         0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97,
81         0x98, 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f,
82         0xa0, 0xa1, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa7,
83         0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf,
84         0xb0, 0xb1, 0xb2, 0xb3, 0xb4, 0xb5, 0xb6, 0xb7,
85         0xb8, 0xb9, 0xba, 0xbb, 0xbc, 0xbd, 0xbe, 0xbf,
86         0xc0, 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc7,
87         0xc8, 0xc9, 0xca, 0xcb, 0xcc, 0xcd, 0xce, 0xcf,
88         0xd0, 0xd1, 0xd2, 0xd3, 0xd4, 0xd5, 0xd6, 0xd7,
89         0xd8, 0xd9, 0xda, 0xdb, 0xdc, 0xdd, 0xde, 0xdf,
90         0xe0, 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7,
91         0xe8, 0xe9, 0xea, 0xeb, 0xec, 0xed, 0xee, 0xef,
92         0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7,
93         0xf8, 0xf9, 0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff,
94 };
95 // clang-format on
96 
getSizeFlattenedIDs(uint32_t simd,uint32_t lwsX,uint32_t lwsY,uint32_t lwsZ)97 size_t getSizeFlattenedIDs(uint32_t simd, uint32_t lwsX, uint32_t lwsY, uint32_t lwsZ) {
98     auto lws = lwsX * lwsY * lwsZ;
99     lws = std::min(lws, 256u);
100 
101     auto iDsPerGrf = std::min(simd, 16u);
102     auto numGRFs = Math::divideAndRoundUp(lws, iDsPerGrf);
103 
104     return numGRFs * sizeof(GRF);
105 }
106 
generateFlattenedIDs(void * buffer,uint32_t simd,uint32_t lwsX,uint32_t lwsY,uint32_t lwsZ)107 void generateFlattenedIDs(void *buffer, uint32_t simd, uint32_t lwsX, uint32_t lwsY, uint32_t lwsZ) {
108     auto pSrc = simd == 8 ? gFlattenedIDsSimD8 : gFlattenedIDsSimD1632;
109     auto copySize = getSizeFlattenedIDs(simd, lwsX, lwsY, lwsZ);
110 
111     memcpy_s(buffer, copySize, pSrc, copySize);
112 }
113 
114 struct FlattenedIDFixture : ::testing::TestWithParam<std::tuple<int, int, int, int>> {
SetUpFlattenedIDFixture115     void SetUp() override {
116         simd = std::get<0>(GetParam());
117         localWorkSizeX = std::get<1>(GetParam());
118         localWorkSizeY = std::get<2>(GetParam());
119         localWorkSizeZ = std::get<3>(GetParam());
120 
121         localWorkSize = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
122         if (localWorkSize > 256) {
123             localWorkSizeY = std::min(256 / localWorkSizeX, localWorkSizeY);
124             localWorkSizeZ = std::min(256 / (localWorkSizeX * localWorkSizeY), localWorkSizeZ);
125             localWorkSize = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
126         }
127 
128         memset(buffer, 0xff, sizeof(buffer));
129     }
130 
validateIDWithinLimitsFlattenedIDFixture131     void validateIDWithinLimits(uint32_t simd, uint32_t lwsX, uint32_t lwsY, uint32_t lwsZ, bool useFullRow) {
132         auto numWorkItems = lwsX * lwsY * lwsZ;
133         auto idsPerThread = simd;
134 
135         // As per BackEnd HLD, SIMD32 has 32 flattenedIDs per channel.  SIMD8/16 has up to 16 flattenedIDs.
136         auto skipPerThread = (simd == 32 || useFullRow) ? 32 : 16;
137 
138         auto pBuffer = buffer;
139         size_t itemIndex = 0;
140         auto numWorkItemsRemaining = numWorkItems;
141         while (numWorkItemsRemaining > 0) {
142             EXPECT_LT(pBuffer[itemIndex], numWorkItems);
143             ++itemIndex;
144             if (idsPerThread == itemIndex) {
145                 pBuffer += skipPerThread;
146                 itemIndex = 0;
147             }
148             --numWorkItemsRemaining;
149         }
150     }
151 
validateAllWorkItemsCoveredFlattenedIDFixture152     void validateAllWorkItemsCovered(uint32_t simd, uint32_t lwsX, uint32_t lwsY, uint32_t lwsZ, bool useFullRow) {
153         auto numWorkItems = lwsX * lwsY * lwsZ;
154         auto idsPerThread = simd;
155 
156         // As per BackEnd HLD, SIMD32 has 32 localIDs per channel.  SIMD8/16 has up to 16 localIDs.
157         auto skipPerThread = (simd == 32 || useFullRow) ? 32 : 16;
158 
159         // Initialize local ID hit table
160         uint32_t localIDHitTable[8];
161         memset(localIDHitTable, 0, sizeof(localIDHitTable));
162 
163         auto pBuffer = buffer;
164         size_t itemIndex = 0;
165         auto numWorkItemsRemaining = numWorkItems;
166         while (numWorkItemsRemaining > 0) {
167             // Flatten out the IDs
168             auto workItem = pBuffer[itemIndex];
169             ASSERT_LT(workItem, 256u);
170 
171             // Look up in the hit table
172             auto &hitItem = localIDHitTable[workItem / 32];
173             auto hitBit = 1 << (workItem % 32);
174 
175             // No double-hits
176             EXPECT_EQ(0u, hitItem & hitBit);
177 
178             // Set that work item as hit
179             hitItem |= hitBit;
180 
181             ++itemIndex;
182             if (idsPerThread == itemIndex) {
183                 pBuffer += skipPerThread;
184                 itemIndex = 0;
185             }
186             --numWorkItemsRemaining;
187         }
188 
189         // All entries in hit table should be in form of n^2 - 1
190         for (unsigned int i : localIDHitTable) {
191             EXPECT_EQ(0u, i & (i + 1));
192         }
193     }
194 
195     // Test parameters
196     uint32_t localWorkSize;
197     uint32_t localWorkSizeX;
198     uint32_t localWorkSizeY;
199     uint32_t localWorkSizeZ;
200     uint32_t simd;
201 
202     // Provide support for a max LWS of 256
203     // 32 threads @ SIMD8
204     // 16 lanes per thread (SIMD8 - only 8 used)
205     uint16_t buffer[32 * 16];
206 };
207 
HWTEST_P(FlattenedIDFixture,WhenGeneratingFlattenedIdsThenAllIdsWithinLimits)208 HWTEST_P(FlattenedIDFixture, WhenGeneratingFlattenedIdsThenAllIdsWithinLimits) {
209     generateFlattenedIDs(buffer, simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ);
210 
211     validateIDWithinLimits(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ, UnitTestHelper<FamilyType>::useFullRowForLocalIdsGeneration);
212 }
213 
HWTEST_P(FlattenedIDFixture,WhenGeneratingFlattenedIdsThenAllWorkItemsCovered)214 HWTEST_P(FlattenedIDFixture, WhenGeneratingFlattenedIdsThenAllWorkItemsCovered) {
215     generateFlattenedIDs(buffer, simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ);
216 
217     validateAllWorkItemsCovered(simd, localWorkSizeX, localWorkSizeY, localWorkSizeZ, UnitTestHelper<FamilyType>::useFullRowForLocalIdsGeneration);
218 }
219 
TEST_P(FlattenedIDFixture,WhenGeneratingFlattenedIdsThenLocalIdsAreCorrect)220 TEST_P(FlattenedIDFixture, WhenGeneratingFlattenedIdsThenLocalIdsAreCorrect) {
221     auto sizePerThreadData = getSizeFlattenedIDs(
222         simd,
223         localWorkSizeX,
224         localWorkSizeY,
225         localWorkSizeZ);
226 
227     // Should be multiple of GRFs
228     auto sizeGRF = sizeof(GRF);
229     EXPECT_EQ(0u, sizePerThreadData % sizeGRF);
230 
231     auto workItems = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
232     auto idsPerGRF = (simd == 8) ? 8 : 16;
233     auto numGRFsExpected = Math::divideAndRoundUp(workItems, idsPerGRF);
234     EXPECT_EQ(numGRFsExpected, sizePerThreadData / sizeGRF);
235 }
236 
237 #define SIMDParams ::testing::Values(8, 16, 32)
238 #if HEAVY_DUTY_TESTING
239 #define LWSXParams ::testing::Values(1, 7, 8, 9, 15, 16, 17, 31, 32, 33, 64, 128, 256)
240 #define LWSYParams ::testing::Values(1, 2, 3, 4, 5, 6, 7, 8)
241 #define LWSZParams ::testing::Values(1, 2, 3, 4)
242 #else
243 #define LWSXParams ::testing::Values(1, 7, 8, 9, 15, 16, 17, 31, 32, 33, 64, 128, 256)
244 #define LWSYParams ::testing::Values(1, 2, 4, 8)
245 #define LWSZParams ::testing::Values(1)
246 #endif
247 
248 INSTANTIATE_TEST_CASE_P(AllCombinations, FlattenedIDFixture, ::testing::Combine(SIMDParams, LWSXParams, LWSYParams, LWSZParams));
249 
250 // To debug a specific configuration replace the list of Values with specific values.
251 // NOTE: You'll need a unique test prefix
252 INSTANTIATE_TEST_CASE_P(SingleTest, FlattenedIDFixture,
253                         ::testing::Combine(
254                             ::testing::Values(32),  //SIMD
255                             ::testing::Values(5),   //LWSX
256                             ::testing::Values(6),   //LWSY
257                             ::testing::Values(7))); //LWSZ
258