Improve rounding number to the next power of 2.

Improve rounding number to the next power of 2.

Author: mateuszchudyk

visa/VisaToG4/TranslateSendLdStLsc.cpp

@@ -225,15 +225,9 @@ static int alignUp(int a, int n) { return n + a - 1 - ((n + a - 1) % a); }
 static int lscBlock2dComputeDataRegs(LSC_OP op,
                                      LSC_DATA_SHAPE_BLOCK2D dataShape2d,
                                      int BYTES_PER_REG, int dataSizeBits) {
-  auto roundUpToPowerOf2 = [](int n) {
-    while (n & (n - 1))
-      n++;
-    return n;
-  };
-
   bool transpose = dataShape2d.order == LSC_DATA_ORDER_TRANSPOSE;
   int grfRowPitchElems =
-      roundUpToPowerOf2(!transpose ? dataShape2d.width : dataShape2d.height);
+      RoundUpToPowerOf2(!transpose ? dataShape2d.width : dataShape2d.height);
   int blockRows = !transpose ? dataShape2d.height : dataShape2d.width;
   int elemsPerGrf = 8 * BYTES_PER_REG / dataSizeBits;
   // alignUp needed for padding between blocks; each block pads out to

visa/common.h

@@ -184,4 +184,29 @@ constexpr Type AlignUp(const Type value, const size_t alignment) {
   return common - (common % alignment);
 }
 
+//
+// The function returns:
+//   - 0 if 'value' = 0,
+//   - the lowest power of two greater of equal to 'value'.
+//
+// The main idea of the algorithm is to set all lower bits for each set bit of 'value'.
+//
+constexpr static uint32_t RoundUpToPowerOf2(uint32_t value) {
+  // Without decreasing, the result would be:
+  // - 1 for 'value' = 0
+  // - the next power of two if 'value' is a power of two.
+  value--;
+
+  value |= value >> 1;  // Duplicate each set bit to the 1st bit on the right, so each group of 1s has length >= 2 or ends on the first bit.
+  value |= value >> 2;  // Duplicate each set bit to the 2nd bit on the right, so each group of 1s has length >= 4 or ends on the first bit.
+  value |= value >> 4;  // ..
+  value |= value >> 8;  // ..
+  value |= value >> 16; // Duplicate each set bit to the 16th bit on the right, so each group of 1s has length >= 32 or ends on the first bit.
+
+  // Increase by one to get a power of two.
+  value++;
+
+  return value;
+}
+
 #endif //_COMMON_H_