From aaa687ca984b18bd1ea499c92285b490fd78e2a3 Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Mon, 28 Mar 2016 23:00:44 +0200
Subject: Added preliminary support for the xNRM2 routines

---
 src/kernels/level1/xnrm2.opencl | 120 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 120 insertions(+)
 create mode 100644 src/kernels/level1/xnrm2.opencl

(limited to 'src/kernels/level1')

diff --git a/src/kernels/level1/xnrm2.opencl b/src/kernels/level1/xnrm2.opencl
new file mode 100644
index 00000000..c50d7d63
--- /dev/null
+++ b/src/kernels/level1/xnrm2.opencl
@@ -0,0 +1,120 @@
+
+// =================================================================================================
+// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
+// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
+// width of 100 characters per line.
+//
+// Author(s):
+//   Cedric Nugteren <www.cedricnugteren.nl>
+//
+// This file contains the Xnrm2 kernel. It implements a dot-product computation using reduction
+// kernels. Reduction is split in two parts. In the first (main) kernel the X and Y vectors are
+// multiplied, followed by a per-thread and a per-workgroup reduction. The second (epilogue) kernel
+// is executed with a single workgroup only, computing the final result.
+//
+// =================================================================================================
+
+// Enables loading of this file using the C++ pre-processor's #include (C++11 standard raw string
+// literal). Comment-out this line for syntax-highlighting when developing.
+R"(
+
+// Parameters set by the tuner or by the database. Here they are given a basic default value in case
+// this kernel file is used outside of the CLBlast library.
+#ifndef WGS1
+  #define WGS1 64     // The local work-group size of the main kernel
+#endif
+#ifndef WGS2
+  #define WGS2 64     // The local work-group size of the epilogue kernel
+#endif
+
+// =================================================================================================
+
+// The main reduction kernel, performing the multiplication and the majority of the sum operation
+__attribute__((reqd_work_group_size(WGS1, 1, 1)))
+__kernel void Xnrm2(const int n,
+                    const __global real* restrict xgm, const int x_offset, const int x_inc,
+                    __global real* output) {
+  __local real lm[WGS1];
+  const int lid = get_local_id(0);
+  const int wgid = get_group_id(0);
+  const int num_groups = get_num_groups(0);
+
+  // Performs multiplication and the first steps of the reduction
+  real acc;
+  SetToZero(acc);
+  int id = wgid*WGS1 + lid;
+  while (id < n) {
+    real x1 = xgm[id*x_inc + x_offset];
+    real x2 = x1;
+    COMPLEX_CONJUGATE(x2);
+    MultiplyAdd(acc, x1, x2);
+    id += WGS1*num_groups;
+  }
+  lm[lid] = acc;
+  barrier(CLK_LOCAL_MEM_FENCE);
+
+  // Performs reduction in local memory
+  #pragma unroll
+  for (int s=WGS1/2; s>0; s=s>>1) {
+    if (lid < s) {
+      Add(lm[lid], lm[lid], lm[lid + s]);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+  }
+
+  // Stores the per-workgroup result
+  if (lid == 0) {
+    output[wgid] = lm[0];
+  }
+}
+
+// =================================================================================================
+
+// Computes the square root
+inline real SquareRoot(const real z) {
+  #if PRECISION == 3232 || PRECISION == 6464
+    double r = sqrt(z.x * z.x + z.y * z.y);
+    real zpr; zpr.x = z.x + r; zpr.y = z.y;
+    double zprabs = sqrt(zpr.x * zpr.x + zpr.y + zpr.y);
+    real result;
+    result.x = sqrt(r) * zpr.x / zprabs;
+    result.y = sqrt(r) * zpr.y / zprabs;
+    return result;
+  #else
+    return sqrt(z);
+  #endif
+}
+
+// The epilogue reduction kernel, performing the final bit of the sum operation. This kernel has to
+// be launched with a single workgroup only.
+__attribute__((reqd_work_group_size(WGS2, 1, 1)))
+__kernel void Xnrm2Epilogue(const __global real* restrict input,
+                            __global real* nrm2, const int nrm2_offset) {
+  __local real lm[WGS2];
+  const int lid = get_local_id(0);
+
+  // Performs the first step of the reduction while loading the data
+  Add(lm[lid], input[lid], input[lid + WGS2]);
+  barrier(CLK_LOCAL_MEM_FENCE);
+
+  // Performs reduction in local memory
+  #pragma unroll
+  for (int s=WGS2/2; s>0; s=s>>1) {
+    if (lid < s) {
+      Add(lm[lid], lm[lid], lm[lid + s]);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+  }
+
+  // Computes the square root and stores the final result
+  if (lid == 0) {
+    nrm2[nrm2_offset] = SquareRoot(lm[0]);
+  }
+}
+
+// =================================================================================================
+
+// End of the C++11 raw string literal
+)"
+
+// =================================================================================================
-- 
cgit v1.2.3