Lines Matching refs:depth
289 AccumulatorType* accum_ptr, int depth) { in Run()
454 [depth] "+r"(depth) in Run()
478 AccumulatorType* accum_ptr, int depth) { in Run()
570 [depth] "+r"(depth) in Run()
596 AccumulatorType* accum_ptr, int depth) { in Run()
598 std::size_t run_depth = depth; in Run()
843 AccumulatorType* accum_ptr, int depth) { in Run()
907 [depth] "+r"(depth) in Run()
928 AccumulatorType* accum_ptr, int depth) { in Run()
993 [depth] "+r"(depth) in Run()
1015 AccumulatorType* accum_ptr, int depth) { in Run()
1080 [depth] "+r"(depth) in Run()
1101 AccumulatorType* accum_ptr, int depth) { in Run()
1165 [depth] "+r"(depth) in Run()
1187 AccumulatorType* accum_ptr, int depth) { in Run()
1311 [depth] "+r"(depth) in Run()
1330 AccumulatorType* accum_ptr, int depth) { in Run()
1498 [depth] "+r"(depth) in Run()
1519 AccumulatorType* accum_ptr, int depth) { in Run()
1628 [depth] "+r"(depth) in Run()
1651 AccumulatorType* accum_ptr, int depth) { in Run()
1723 [depth] "+r"(depth) in Run()
1748 AccumulatorType* accum_ptr, int depth) { in Run()
1989 [depth] "+r"(depth) in Run()
2011 AccumulatorType* accum_ptr, int depth) { in Run()
2233 [depth] "+r"(depth) in Run()
2267 AccumulatorType* accum_ptr, int depth) { in Run()
2269 std::size_t run_depth = depth; in Run()
2544 AccumulatorType* accum_ptr, int depth) { in Run()
2652 [depth] "+r"(depth) in Run()
2675 AccumulatorType* accum_ptr, int depth) { in Run()
2792 [depth] "+r"(depth) in Run()
2815 AccumulatorType* accum_ptr, int depth) { in Run()
2916 [depth] "+r"(depth) in Run()
2937 AccumulatorType* accum_ptr, int depth) { in Run()
3046 [depth] "+r"(depth) in Run()
3067 AccumulatorType* accum_ptr, int depth) { in Run()
3168 [depth] "+r"(depth) in Run()
3188 AccumulatorType* accum_ptr, int depth) { in Run()
3297 [depth] "+r"(depth) in Run()
3318 AccumulatorType* accum_ptr, int depth) { in Run()
3494 [depth] "+r"(depth) in Run()
3515 AccumulatorType* accum_ptr, int depth) { in Run()
3658 [depth] "+r"(depth) in Run()
3690 AccumulatorType* accum_ptr, int depth) { in Run()
3697 for (int d = 0; d < depth; d += 16) { in Run()
3745 AccumulatorType* accum_ptr, int depth) { in Run()
3752 for (int d = 0; d < depth; d += 2) { in Run()
3811 AccumulatorType* accum_ptr, int depth) { in Run()
3818 for (int d = 0; d < depth; d++) { in Run()
3887 AccumulatorType* accum_ptr, int depth) { in Run()
3897 while (depth > 0) { in Run()
3942 depth -= 2; in Run()
3971 AccumulatorType* accum_ptr, int depth) { in Run()
4105 [depth] "+r"(depth) in Run()
4134 AccumulatorType* accum_ptr, int depth) { in Run()
4260 [depth] "+r"(depth) in Run()
4289 AccumulatorType* accum_ptr, int depth) { in Run()
4299 while (depth > 0) { in Run()
4360 depth -= 2; in Run()
4389 AccumulatorType* accum_ptr, int depth) { in Run()
4589 [depth] "+r"(depth) in Run()
4619 AccumulatorType* accum_ptr, int depth) { in Run()
4808 [depth] "+r"(depth) in Run()
4835 AccumulatorType* accum_ptr, int depth) { in Run()
4836 const int depth_cells = static_cast<int>(depth / Format::kDepth); in Run()
4958 void test_kernel(int depth, const char* kernel_name) { in test_kernel() argument
4967 CacheLineAlignedBuffer<OperandType> lhs(kLhsWidth * depth); in test_kernel()
4968 CacheLineAlignedBuffer<OperandType> rhs(kRhsWidth * depth); in test_kernel()
4980 ReferenceKernel::Run(lhs.data(), rhs.data(), accum_reference.data(), depth); in test_kernel()
4981 Kernel::Run(lhs.data(), rhs.data(), accum.data(), depth); in test_kernel()
4989 << "Wrong accumulator for depth=" << depth << ", " in test_kernel()
4994 if (depth <= 16) { in test_kernel()
4996 PrintMatrix(kLhsWidth, depth, 1, kLhsWidth, lhs.data()); in test_kernel()
4998 PrintMatrix(depth, kRhsWidth, kRhsWidth, 1, rhs.data()); in test_kernel()
5014 int ops(int depth) { in ops() argument
5016 return 2 * Kernel::Format::Lhs::kWidth * Kernel::Format::Rhs::kWidth * depth; in ops()
5077 double benchmark(int depth) { in benchmark() argument
5085 CacheLineAlignedBuffer<OperandType> lhs(Kernel::Format::Lhs::kWidth * depth); in benchmark()
5086 CacheLineAlignedBuffer<OperandType> rhs(Kernel::Format::Rhs::kWidth * depth); in benchmark()
5093 Kernel::Run(lhs.data(), rhs.data(), accum.data(), depth); in benchmark()
5098 return iters_at_a_time * ops<Kernel>(depth) / elapsed; in benchmark()
5111 for (int depth = kKernelDepth; depth <= 1024; depth += kKernelDepth) { in benchmark_and_print_results() local
5112 test_kernel<Kernel>(depth, kernel_name); in benchmark_and_print_results()
5116 for (int depth = kKernelDepth; in benchmark_and_print_results() local
5117 depth <= BenchmarkDepthToFitInCache<Kernel>(); depth *= 2) { in benchmark_and_print_results()
5118 std::cout << kernel_name << "," << depth << "," in benchmark_and_print_results()
5119 << benchmark<Kernel>(depth) * 1e-9f << std::endl; in benchmark_and_print_results()
5122 const int depth = BenchmarkDepthToFitInCache<Kernel>(); in benchmark_and_print_results() local
5123 std::cout << kernel_name << "," << benchmark<Kernel>(depth) * 1e-9f in benchmark_and_print_results()