Add get_coord formula for get_coord for 16 and 32 bit datatypes.

Platforms:
All

Keywords:
Feature

Related-to: GSD-11139
Resolves:

Author: ggojska

IGC/BiFModule/Languages/OpenCL/PreRelease/Matrix/IBiF_matrix.cl

@@ -794,7 +794,7 @@ The storage in memory will be: 0 0 1 1 2 2 ... 7 7
 // R - number of rows
 // C - number of columns
 // VF - VNNI Factor
-#define DEFINE_GET_COORD(layout, sg, elem_bitwidth, contrib_bitwidth, R, C, VF) \
+#define DEFINE_GET_COORD_ROWPACKED(layout, sg, elem_bitwidth, contrib_bitwidth, R, C, VF) \
   INLINE int2 MANGLE_GETCOORD_NAME(layout, sg, elem_bitwidth, R, C) (int index) { \
     int sg_size = get_sub_group_size(); \
     int wi_id = get_sub_group_local_id(); \
@@ -807,32 +807,67 @@ The storage in memory will be: 0 0 1 1 2 2 ... 7 7
     return result; \
   }
 
-// ------ PVC -------
-// layout, sg, elem_bitwidth, contrib_bitwidth, R, C, VF
-//int8
-DEFINE_GET_COORD(PackedA, _SG16, 8, 16, 8, 32, 1)
-DEFINE_GET_COORD(PackedB, _SG16, 8, 32, 32, 16, 4)
+#define DEFINE_GET_COORD(layout, sg, elem_bitwidth, R, C, slices) \
+  INLINE int2 MANGLE_GETCOORD_NAME(layout, sg, elem_bitwidth, R, C) (int index) { \
+    int sg_size = get_sub_group_size(); \
+    int wi_id = get_sub_group_local_id(); \
+    int elems_per_slice = (R * C / sg_size) / slices; \
+    int slice_cols = (C / slices); \
+    int sg_cols_per_wi = slice_cols / sg_size; \
+    int row = (index % elems_per_slice) / sg_cols_per_wi; \
+    int col = wi_id + ((index % elems_per_slice) % sg_cols_per_wi) * sg_size + (index / elems_per_slice * slice_cols); \
+    int2 result = (int2)(row, col); \
+    return result; \
+  }
 
-//bfloat16
-DEFINE_GET_COORD(PackedA, _SG16, 16, 16, 8, 16, 1)
-DEFINE_GET_COORD(PackedA, _SG16, 16, 16, 16, 16, 1)
-DEFINE_GET_COORD(PackedB, _SG16, 16, 32, 16, 16, 2)
+// ------ PVC -------
+// DEFINE_GET_COORD_ROWPACKED(layout, sg, elem_bitwidth, contrib_bitwidth, R, C, VF)
+// DEFINE_GET_COORD(layout, sg, elem_bitwidth, R, C, slices)
+// int8
+DEFINE_GET_COORD_ROWPACKED(PackedA, _SG16, 8, 16, 8, 32, 1)
+DEFINE_GET_COORD(PackedB, _SG16, 8, 32, 16, 1)
+
+// 16bit A
+DEFINE_GET_COORD(PackedA, _SG16, 16, 1, 16, 1)
+DEFINE_GET_COORD(PackedA, _SG16, 16, 8, 16, 1)
+DEFINE_GET_COORD(PackedA, _SG16, 16, 16, 16, 1)
+DEFINE_GET_COORD(PackedA, _SG16, 16, 1, 32, 1)
+DEFINE_GET_COORD(PackedA, _SG16, 16, 32, 16, 1)
+DEFINE_GET_COORD(PackedA, _SG16, 16, 32, 32, 2)
+
+// 16bit PackedB
+DEFINE_GET_COORD(PackedB, _SG16, 16, 16, 16, 1)
+DEFINE_GET_COORD(PackedB, _SG16, 16, 16, 64, 4)
+DEFINE_GET_COORD(PackedB, _SG16, 16, 32, 64, 4)
+
+// 16bit Row_major B
+DEFINE_GET_COORD(PackedB_RowMajor, _SG16, 16, 16, 16, 1)
+DEFINE_GET_COORD(PackedB_RowMajor, _SG16, 16, 16, 64, 4)
+DEFINE_GET_COORD(PackedB_RowMajor, _SG16, 16, 32, 64, 4)
 
 // Accumulator
-DEFINE_GET_COORD(Accumulator, _SG16, 32, 32, 8, 16, 1)
-DEFINE_GET_COORD(Accumulator, _SG16, 32, 32, 16, 16, 1)
+DEFINE_GET_COORD(Accumulator, _SG16, 32, 8, 16, 1)
+DEFINE_GET_COORD(Accumulator, _SG16, 32, 16, 16, 1)
+DEFINE_GET_COORD(Accumulator, _SG16, 32, 32, 64, 4)
+DEFINE_GET_COORD(Accumulator, _SG16, 32, 1, 64, 4)
+
+// Accumulator 16bit
+DEFINE_GET_COORD(Accumulator, _SG16, 16, 8, 16, 1)
+DEFINE_GET_COORD(Accumulator, _SG16, 16, 16, 16, 1)
+DEFINE_GET_COORD(Accumulator, _SG16, 16, 32, 64, 2)
+DEFINE_GET_COORD(Accumulator, _SG16, 16, 1, 64, 4)
 
 // --------- XMX8 ------------
 //int8
-DEFINE_GET_COORD(PackedA, , 8, 32, 8, 32, 1)
-DEFINE_GET_COORD(PackedB, , 8, 32, 32, 8, 4)
+DEFINE_GET_COORD_ROWPACKED(PackedA, , 8, 32, 8, 32, 1)
+DEFINE_GET_COORD_ROWPACKED(PackedB, , 8, 32, 32, 8, 4)
 
 //bfloat16
-DEFINE_GET_COORD(PackedA, , 16, 32, 8, 16, 1)
-DEFINE_GET_COORD(PackedB, , 16, 32, 16, 8, 2)
+DEFINE_GET_COORD_ROWPACKED(PackedA, , 16, 32, 8, 16, 1)
+DEFINE_GET_COORD_ROWPACKED(PackedB, , 16, 32, 16, 8, 2)
 
 // Accumulator
-DEFINE_GET_COORD(Accumulator, , 32, 32, 8, 8, 1)
+DEFINE_GET_COORD_ROWPACKED(Accumulator, , 32, 32, 8, 8, 1)
 
 /* experimental large slice support: */
 

IGC/BiFModule/Languages/OpenCL/PreRelease/Matrix/IBiF_matrix_generator.cpp

@@ -1058,7 +1058,12 @@ ImplementLargeLoadBase(MatrixSpec spec, AddrSpace addr, int numLoads, bool isChe
             s += "__private char *dst1 = dst + WiRowsPerLoad * ContribByteWidth;\n";
             // Prepare mem (source) pointers
             s += "char *mem0 = mem;\n";
-            s += "char *mem1 = mem + 16 * ElemByteWidth;\n";
+            // 32bit Accumulator 32x64 is sliced into 4* 32x16
+            // 16bit Accumulator 32x64 is sliced into 2* 32x32
+            if (spec.BitWidth == 16 && spec.Layout == Layout_Accumulator_RowMajor)
+                s += "char *mem1 = mem + 32 * ElemByteWidth;\n";
+            else
+                s += "char *mem1 = mem + 16 * ElemByteWidth;\n";
             // Call load sub-functions
             s += "LoadFunc(dst0, mem0, stride, cacheOpt);\n";
             s += "LoadFunc(dst1, mem1, stride, cacheOpt);\n";