Lines Matching refs:depth
274 AccumulatorType* accum_ptr, int depth) { in Run()
439 [depth] "+r"(depth) in Run()
463 AccumulatorType* accum_ptr, int depth) { in Run()
555 [depth] "+r"(depth) in Run()
581 AccumulatorType* accum_ptr, int depth) { in Run()
583 std::size_t run_depth = depth; in Run()
828 AccumulatorType* accum_ptr, int depth) { in Run()
892 [depth] "+r"(depth) in Run()
913 AccumulatorType* accum_ptr, int depth) { in Run()
978 [depth] "+r"(depth) in Run()
1000 AccumulatorType* accum_ptr, int depth) { in Run()
1065 [depth] "+r"(depth) in Run()
1086 AccumulatorType* accum_ptr, int depth) { in Run()
1150 [depth] "+r"(depth) in Run()
1172 AccumulatorType* accum_ptr, int depth) { in Run()
1296 [depth] "+r"(depth) in Run()
1315 AccumulatorType* accum_ptr, int depth) { in Run()
1483 [depth] "+r"(depth) in Run()
1504 AccumulatorType* accum_ptr, int depth) { in Run()
1613 [depth] "+r"(depth) in Run()
1636 AccumulatorType* accum_ptr, int depth) { in Run()
1708 [depth] "+r"(depth) in Run()
1733 AccumulatorType* accum_ptr, int depth) { in Run()
1974 [depth] "+r"(depth) in Run()
1996 AccumulatorType* accum_ptr, int depth) { in Run()
2218 [depth] "+r"(depth) in Run()
2252 AccumulatorType* accum_ptr, int depth) { in Run()
2254 std::size_t run_depth = depth; in Run()
2515 AccumulatorType* accum_ptr, int depth) { in Run()
2616 [depth] "+r"(depth) in Run()
2637 AccumulatorType* accum_ptr, int depth) { in Run()
2746 [depth] "+r"(depth) in Run()
2767 AccumulatorType* accum_ptr, int depth) { in Run()
2868 [depth] "+r"(depth) in Run()
2888 AccumulatorType* accum_ptr, int depth) { in Run()
2997 [depth] "+r"(depth) in Run()
3018 AccumulatorType* accum_ptr, int depth) { in Run()
3194 [depth] "+r"(depth) in Run()
3226 AccumulatorType* accum_ptr, int depth) { in Run()
3233 for (int d = 0; d < depth; d += 16) { in Run()
3281 AccumulatorType* accum_ptr, int depth) { in Run()
3288 for (int d = 0; d < depth; d += 2) { in Run()
3347 AccumulatorType* accum_ptr, int depth) { in Run()
3354 for (int d = 0; d < depth; d++) { in Run()
3403 AccumulatorType* accum_ptr, int depth) { in Run()
3404 const int depth_cells = static_cast<int>(depth / Format::kDepth); in Run()
3525 void test_kernel(int depth, const char* kernel_name) { in test_kernel() argument
3534 CacheLineAlignedBuffer<OperandType> lhs(kLhsWidth * depth); in test_kernel()
3535 CacheLineAlignedBuffer<OperandType> rhs(kRhsWidth * depth); in test_kernel()
3547 ReferenceKernel::Run(lhs.data(), rhs.data(), accum_reference.data(), depth); in test_kernel()
3548 Kernel::Run(lhs.data(), rhs.data(), accum.data(), depth); in test_kernel()
3556 << "Wrong accumulator for depth=" << depth << ", " in test_kernel()
3561 if (depth <= 16) { in test_kernel()
3563 PrintMatrix(kLhsWidth, depth, 1, kLhsWidth, lhs.data()); in test_kernel()
3565 PrintMatrix(depth, kRhsWidth, kRhsWidth, 1, rhs.data()); in test_kernel()
3581 int ops(int depth) { in ops() argument
3583 return 2 * Kernel::Format::Lhs::kWidth * Kernel::Format::Rhs::kWidth * depth; in ops()
3644 double benchmark(int depth) { in benchmark() argument
3652 CacheLineAlignedBuffer<OperandType> lhs(Kernel::Format::Lhs::kWidth * depth); in benchmark()
3653 CacheLineAlignedBuffer<OperandType> rhs(Kernel::Format::Rhs::kWidth * depth); in benchmark()
3660 Kernel::Run(lhs.data(), rhs.data(), accum.data(), depth); in benchmark()
3665 return iters_at_a_time * ops<Kernel>(depth) / elapsed; in benchmark()
3678 for (int depth = kKernelDepth; depth <= 1024; depth += kKernelDepth) { in benchmark_and_print_results() local
3679 test_kernel<Kernel>(depth, kernel_name); in benchmark_and_print_results()
3683 for (int depth = kKernelDepth; in benchmark_and_print_results() local
3684 depth <= BenchmarkDepthToFitInCache<Kernel>(); depth *= 2) { in benchmark_and_print_results()
3685 std::cout << kernel_name << "," << depth << "," in benchmark_and_print_results()
3686 << benchmark<Kernel>(depth) * 1e-9f << std::endl; in benchmark_and_print_results()
3689 const int depth = BenchmarkDepthToFitInCache<Kernel>(); in benchmark_and_print_results() local
3690 std::cout << kernel_name << "," << benchmark<Kernel>(depth) * 1e-9f in benchmark_and_print_results()