wip

tessellate_ipu/core/vertex/tile_hessenberg_vertex.cpp

@@ -22,17 +22,17 @@ static constexpr size_t MIN_ALIGN = 8;
 class HessenbergCorrectionVectorVertex : public MultiVertex {
  public:
   using T = float;
-  Input<Vector<T, poplar::VectorLayout::SPAN, MIN_ALIGN>> Rcol;      // (N,) R column.
-  Input<Vector<T, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>> sdiag;  // (N,) R diag. sign.
-  Input<Vector<int, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
-    cidx;
+  Input<Vector<T, poplar::VectorLayout::SPAN, MIN_ALIGN>>
+      Rcol;  // (N,) R column.
+  Input<Vector<T, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
+      sdiag;  // (N,) R diag. sign.
+  Input<Vector<int, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>> cidx;
 
   Output<Vector<T, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
       v;  // (N,) QR correction vector (not normalized)
   Output<Vector<T, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
       vrescale;  // (1,) QR correction vector rescaling (2 / norm)
 
-
   // Use static variables as easy sync. mechanisms between worker threads.
   // see: https://graphcore.slack.com/archives/C013LPHPX61/p1647443852989259
   static T shared_partial_sqnorms[6];
@@ -142,23 +142,20 @@ class [[poplar::constraint(
       scale1;  // (1,) first scaling factor.
   Input<Vector<T, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
       scale2;  // (1,) 2nd scaling factor.
-  Input<Vector<int, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
-      start_idx_;
+  Input<Vector<int, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>> start_idx_;
 
   Input<Vector<IndexType, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
       worker_offsets;  // (7,) threads work size + 1.
 
-
-
   bool compute(unsigned wid) {
     // Always assuming size % 2 == 0
     constexpr unsigned ptr_step = 1;
     const IndexType wstart = worker_offsets[wid];
     const IndexType wend = worker_offsets[wid + 1];
     const IndexType wsize = wend - wstart;
 
-    IndexType start_idx = start_idx_[0];  // X start idx. Must be a multiple of 4 (for bank
-                        // alignment aspects).
+    IndexType start_idx = start_idx_[0];  // X start idx. Must be a multiple of
+                                          // 4 (for bank alignment aspects).
 
     // Set the $TAS register with the proper scale.
     const T s = -scale1[0] * scale2[0];

tessellate_ipu/core/vertex/tile_small_dot.cpp

@@ -0,0 +1,43 @@
+// Copyright (c) 2023 Graphcore Ltd. All rights reserved.
+#include "tile_small_dot.hpp"
+
+#include <poplar/HalfFloat.hpp>
+#include <poplar/Vertex.hpp>
+
+// class [[poplar::constraint("elem(*x) != elem(*y)")]] DotProduct1dVertex
+//     : public MultiVertex {
+class Rotation2dVertex : public MultiVertex {
+ public:
+  using T = float;
+  using T2 = float2;
+  // Using `uint16` seems to be generating more efficient loops?
+  using IndexType = unsigned short;
+
+  static constexpr size_t MIN_ALIGN = 8;
+
+  Input<Vector<T, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
+      inrow0;  // (N,) first input row vector
+  Input<Vector<T, poplar::VectorLayout::ONE_PTR, MIN_ALIGN>>
+      inrow1;  // (N,) second input row vector
+
+  Input<Vector<T, poplar::VectorLayout::ONE_PTR>>
+      cs;  // (2,) rotation cosinus/sinus values
+
+  Input<Vector<IndexType, poplar::VectorLayout::ONE_PTR>>
+      worker_offsets;  // (7,) number threads + 1.
+
+  Output<Vector<T, poplar::VectorLayout::ONE_PTR>>
+      outrow0;  // (N,) first input row vector
+  Output<Vector<T, poplar::VectorLayout::ONE_PTR>>
+      outrow1;  // (N,) first input row vector
+
+  bool compute(unsigned wid) {
+    // Always assuming size % 2 == 0
+    const IndexType wstart = worker_offsets[wid];
+    const IndexType wend = worker_offsets[wid + 1];
+    const IndexType wsize = wend - wstart;
+
+
+    return true;
+  }
+};

tessellate_ipu/core/vertex/tile_small_dot.hpp

@@ -0,0 +1,5 @@
+// Copyright (c) 2023 Graphcore Ltd. All rights reserved.
+// #include <poplar/HalfFloat.hpp>
+// #include <poplar/Vertex.hpp>
+
+#include "intrinsics_utils.hpp"

tessellate_ipu/lax/tile_lax_small_dot.py

@@ -17,6 +17,7 @@
     tile_map_remove_ipu_attributes,
 )
 from tessellate_ipu.core.tile_interpreter_vertex_utils import make_num_elements_per_worker
+
 # from tessellate_ipu.lib.pytessellate_ipu_core import (  # noqa: F401
 #     IpuVertexAttributeI32,
 #     ipuGetTransformedInRowStride,