Merge changes Iee153445,Iee274471 am: 79df15ea88 am: 10f298fc41 am: 7cb5001398t_frc_odp_330442040t_frc_odp_330442000t_frc_ase_330444010android-wear-13.0.0-gpl_r3android-wear-13.0.0-gpl_r2android-wear-13.0.0-gpl_r1android-vts-13.0_r8android-vts-13.0_r7android-vts-13.0_r6android-vts-13.0_r5android-vts-13.0_r4android-vts-13.0_r3android-vts-13.0_r2android-t-qpr3-beta-3-gplandroid-t-qpr3-beta-1-gplandroid-t-qpr2-beta-3-gplandroid-t-qpr2-beta-2-gplandroid-t-qpr1-beta-3-gplandroid-t-qpr1-beta-1-gplandroid-cts-13.0_r8android-cts-13.0_r7android-cts-13.0_r6android-cts-13.0_r5android-cts-13.0_r4android-cts-13.0_r3android-cts-13.0_r2android-13.0.0_r83android-13.0.0_r82android-13.0.0_r81android-13.0.0_r80android-13.0.0_r79android-13.0.0_r78android-13.0.0_r77android-13.0.0_r76android-13.0.0_r75android-13.0.0_r74android-13.0.0_r73android-13.0.0_r72android-13.0.0_r71android-13.0.0_r70android-13.0.0_r69android-13.0.0_r68android-13.0.0_r67android-13.0.0_r66android-13.0.0_r65android-13.0.0_r64android-13.0.0_r63android-13.0.0_r62android-13.0.0_r61android-13.0.0_r60android-13.0.0_r59android-13.0.0_r58android-13.0.0_r57android-13.0.0_r56android-13.0.0_r55android-13.0.0_r54android-13.0.0_r53android-13.0.0_r52android-13.0.0_r51android-13.0.0_r50android-13.0.0_r49android-13.0.0_r48android-13.0.0_r47android-13.0.0_r46android-13.0.0_r45android-13.0.0_r44android-13.0.0_r43android-13.0.0_r42android-13.0.0_r41android-13.0.0_r40android-13.0.0_r39android-13.0.0_r38android-13.0.0_r37android-13.0.0_r36android-13.0.0_r35android-13.0.0_r34android-13.0.0_r33android-13.0.0_r32android-13.0.0_r30android-13.0.0_r29android-13.0.0_r28android-13.0.0_r27android-13.0.0_r24android-13.0.0_r23android-13.0.0_r22android-13.0.0_r21android-13.0.0_r20android-13.0.0_r19android-13.0.0_r18android-13.0.0_r17android-13.0.0_r16aml_go_odp_330912000aml_go_ads_330915100aml_go_ads_330915000aml_go_ads_330913000android13-tests-releaseandroid13-tests-devandroid13-qpr3-s9-releaseandroid13-qpr3-s8-releaseandroid13-qpr3-s7-releaseandroid13-qpr3-s6-releaseandroid13-qpr3-s5-releaseandroid13-qpr3-s4-releaseandroid13-qpr3-s3-releaseandroid13-qpr3-s2-releaseandroid13-qpr3-s14-releaseandroid13-qpr3-s13-releaseandroid13-qpr3-s12-releaseandroid13-qpr3-s11-releaseandroid13-qpr3-s10-releaseandroid13-qpr3-s1-releaseandroid13-qpr3-releaseandroid13-qpr3-c-s8-releaseandroid13-qpr3-c-s7-releaseandroid13-qpr3-c-s6-releaseandroid13-qpr3-c-s5-releaseandroid13-qpr3-c-s4-releaseandroid13-qpr3-c-s3-releaseandroid13-qpr3-c-s2-releaseandroid13-qpr3-c-s12-releaseandroid13-qpr3-c-s11-releaseandroid13-qpr3-c-s10-releaseandroid13-qpr3-c-s1-releaseandroid13-qpr2-s9-releaseandroid13-qpr2-s8-releaseandroid13-qpr2-s7-releaseandroid13-qpr2-s6-releaseandroid13-qpr2-s5-releaseandroid13-qpr2-s3-releaseandroid13-qpr2-s2-releaseandroid13-qpr2-s12-releaseandroid13-qpr2-s11-releaseandroid13-qpr2-s10-releaseandroid13-qpr2-s1-releaseandroid13-qpr2-releaseandroid13-qpr2-b-s1-releaseandroid13-qpr1-s8-releaseandroid13-qpr1-s7-releaseandroid13-qpr1-s6-releaseandroid13-qpr1-s5-releaseandroid13-qpr1-s4-releaseandroid13-qpr1-s3-releaseandroid13-qpr1-s2-releaseandroid13-qpr1-s1-releaseandroid13-qpr1-releaseandroid13-mainline-go-adservices-releaseandroid13-frc-odp-releaseandroid13-devandroid13-d4-s2-releaseandroid13-d4-s1-releaseandroid13-d4-releaseandroid13-d3-s1-releaseandroid13-d2-releaseandroid-wear-13.0.0-gpl_r1

Original change: https://android-review.googlesource.com/c/platform/external/eigen/+/1999079

Change-Id: I4c76dc5ddc7fb0ae9fc42436f28bd8bf9de50a97

1 files changed, 218 insertions, 0 deletions

diff --git a/unsupported/test/cxx11_tensor_contract_gpu.cu b/unsupported/test/cxx11_tensor_contract_gpu.cu
new file mode 100644
index 000000000..575bdc1f9
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_contract_gpu.cu
@@ -0,0 +1,218 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2014 Navdeep Jaitly <ndjaitly@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#define EIGEN_TEST_NO_LONGDOUBLE
+#define EIGEN_TEST_NO_COMPLEX
+
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE int
+#define EIGEN_USE_GPU
+
+#include "main.h"
+#include <unsupported/Eigen/CXX11/Tensor>
+
+#include <unsupported/Eigen/CXX11/src/Tensor/TensorGpuHipCudaDefines.h>
+
+using Eigen::Tensor;
+typedef Tensor<float, 1>::DimensionPair DimPair;
+
+template<int DataLayout>
+void test_gpu_contraction(int m_size, int k_size, int n_size)
+{
+  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
+  // with these dimensions, the output has 300 * 140 elements, which is
+  // more than 30 * 1024, which is the number of threads in blocks on
+  // a 15 SM GK110 GPU
+  Tensor<float, 2, DataLayout> t_left(m_size, k_size);
+  Tensor<float, 2, DataLayout> t_right(k_size, n_size);
+  Tensor<float, 2, DataLayout> t_result(m_size, n_size);
+  Tensor<float, 2, DataLayout> t_result_gpu(m_size, n_size);
+  Eigen::array<DimPair, 1> dims(DimPair(1, 0));
+
+  t_left.setRandom();
+  t_right.setRandom();
+
+  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
+  std::size_t t_right_bytes = t_right.size() * sizeof(float);
+  std::size_t t_result_bytes = t_result.size() * sizeof(float);
+
+  float* d_t_left;
+  float* d_t_right;
+  float* d_t_result;
+
+  gpuMalloc((void**)(&d_t_left), t_left_bytes);
+  gpuMalloc((void**)(&d_t_right), t_right_bytes);
+  gpuMalloc((void**)(&d_t_result), t_result_bytes);
+
+  gpuMemcpy(d_t_left, t_left.data(), t_left_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_t_right, t_right.data(), t_right_bytes, gpuMemcpyHostToDevice);
+
+  Eigen::GpuStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_left(d_t_left, Eigen::array<int, 2>(m_size, k_size));
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_right(d_t_right, Eigen::array<int, 2>(k_size, n_size));
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_result(d_t_result, Eigen::array<int, 2>(m_size, n_size));
+
+
+  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
+  t_result = t_left.contract(t_right, dims);
+
+  gpuMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, gpuMemcpyDeviceToHost);
+  for (DenseIndex i = 0; i < t_result.size(); i++) {
+    if (fabs(t_result(i) - t_result_gpu(i)) < 1e-4f) {
+      continue;
+    }
+    if (Eigen::internal::isApprox(t_result(i), t_result_gpu(i), 1e-4f)) {
+      continue;
+    }
+    std::cout << "mismatch detected at index " << i << ": " << t_result(i)
+              << " vs " <<  t_result_gpu(i) << std::endl;
+    assert(false);
+  }
+
+  gpuFree((void*)d_t_left);
+  gpuFree((void*)d_t_right);
+  gpuFree((void*)d_t_result);
+}
+
+
+template<int DataLayout>
+void test_scalar(int m_size, int k_size, int n_size)
+{
+  std::cout << "Testing for (" << m_size << "," << k_size << "," << n_size << ")" << std::endl;
+  // with these dimensions, the output has 300 * 140 elements, which is
+  // more than 30 * 1024, which is the number of threads in blocks on
+  // a 15 SM GK110 GPU
+  Tensor<float, 2, DataLayout> t_left(m_size, k_size);
+  Tensor<float, 2, DataLayout> t_right(k_size, n_size);
+  Tensor<float, 0, DataLayout> t_result;
+  Tensor<float, 0, DataLayout> t_result_gpu;
+  Eigen::array<DimPair, 2> dims(DimPair(0, 0), DimPair(1, 1));
+
+  t_left.setRandom();
+  t_right.setRandom();
+
+  std::size_t t_left_bytes = t_left.size()  * sizeof(float);
+  std::size_t t_right_bytes = t_right.size() * sizeof(float);
+  std::size_t t_result_bytes = sizeof(float);
+
+  float* d_t_left;
+  float* d_t_right;
+  float* d_t_result;
+
+  gpuMalloc((void**)(&d_t_left), t_left_bytes);
+  gpuMalloc((void**)(&d_t_right), t_right_bytes);
+  gpuMalloc((void**)(&d_t_result), t_result_bytes);
+
+  gpuMemcpy(d_t_left, t_left.data(), t_left_bytes, gpuMemcpyHostToDevice);
+  gpuMemcpy(d_t_right, t_right.data(), t_right_bytes, gpuMemcpyHostToDevice);
+
+  Eigen::GpuStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_left(d_t_left, m_size, k_size);
+  Eigen::TensorMap<Eigen::Tensor<float, 2, DataLayout> >
+      gpu_t_right(d_t_right, k_size, n_size);
+  Eigen::TensorMap<Eigen::Tensor<float, 0, DataLayout> >
+      gpu_t_result(d_t_result);
+
+  gpu_t_result.device(gpu_device) = gpu_t_left.contract(gpu_t_right, dims);
+  t_result = t_left.contract(t_right, dims);
+
+  gpuMemcpy(t_result_gpu.data(), d_t_result, t_result_bytes, gpuMemcpyDeviceToHost);
+  if (fabs(t_result() - t_result_gpu()) > 1e-4f &&
+      !Eigen::internal::isApprox(t_result(), t_result_gpu(), 1e-4f)) {
+    std::cout << "mismatch detected: " << t_result()
+              << " vs " <<  t_result_gpu() << std::endl;
+    assert(false);
+  }
+
+  gpuFree((void*)d_t_left);
+  gpuFree((void*)d_t_right);
+  gpuFree((void*)d_t_result);
+}
+
+
+template<int DataLayout>
+void test_gpu_contraction_m() {
+  for (int k = 32; k < 256; k++) {
+    test_gpu_contraction<ColMajor>(k, 128, 128);
+    test_gpu_contraction<RowMajor>(k, 128, 128);
+  }
+}
+
+template<int DataLayout>
+void test_gpu_contraction_k() {
+  for (int k = 32; k < 256; k++) {
+    test_gpu_contraction<ColMajor>(128, k, 128);
+    test_gpu_contraction<RowMajor>(128, k, 128);
+  }
+}
+
+template<int DataLayout>
+void test_gpu_contraction_n() {
+  for (int k = 32; k < 256; k++) {
+    test_gpu_contraction<ColMajor>(128, 128, k);
+    test_gpu_contraction<RowMajor>(128, 128, k);
+  }
+}
+
+
+template<int DataLayout>
+void test_gpu_contraction_sizes() {
+  int m_sizes[] = { 31,  39,   63,   64,   65,
+                   127, 129,  255,  257 , 511,
+                   512, 513, 1023, 1024, 1025};
+
+  int n_sizes[] = { 31,  39,   63,   64,   65,
+                   127, 129,  255,  257,  511,
+                   512, 513, 1023, 1024, 1025};
+
+  int k_sizes[] = {  31,   39,  63,  64,   65,
+                     95,   96, 127, 129,  255,
+                    257,  511, 512, 513, 1023,
+                   1024, 1025};
+
+  for (int i = 0; i < 15; i++) {
+    for (int j = 0; j < 15; j++) {
+      for (int k = 0; k < 17; k++) {
+        test_gpu_contraction<DataLayout>(m_sizes[i], n_sizes[j], k_sizes[k]);
+      }
+    }
+  }
+}
+
+EIGEN_DECLARE_TEST(cxx11_tensor_contract_gpu)
+{
+  CALL_SUBTEST_1(test_gpu_contraction<ColMajor>(128, 128, 128));
+  CALL_SUBTEST_1(test_gpu_contraction<RowMajor>(128, 128, 128));
+
+  CALL_SUBTEST_1(test_scalar<ColMajor>(128, 128, 128));
+  CALL_SUBTEST_1(test_scalar<RowMajor>(128, 128, 128));
+
+  CALL_SUBTEST_2(test_gpu_contraction_m<ColMajor>());
+  CALL_SUBTEST_3(test_gpu_contraction_m<RowMajor>());
+
+  CALL_SUBTEST_4(test_gpu_contraction_k<ColMajor>());
+  CALL_SUBTEST_5(test_gpu_contraction_k<RowMajor>());
+
+  CALL_SUBTEST_6(test_gpu_contraction_n<ColMajor>());
+  CALL_SUBTEST_7(test_gpu_contraction_n<RowMajor>());
+
+#if !defined(EIGEN_USE_HIP)
+// disable these subtests for HIP
+  CALL_SUBTEST_8(test_gpu_contraction_sizes<ColMajor>());
+  CALL_SUBTEST_9(test_gpu_contraction_sizes<RowMajor>());
+#endif	
+}