Fixed problems with block 2d operations for JointMatrix for some parameter configruations.

This patch addresses problems with block 2d load and stores emited by
JointMatrix extension for matrices with i8 element size.

Author: mbelicki

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

@@ -152,7 +152,7 @@ extern __constant int __JointMatrixLoadStoreOpt;
     int width = (sizeof (element_type)) * stride - 1; /* in bytes */ \
     int pitch = width; /* JointMatrices are expected to be contigunous in memory, without padding at the end of a row */ \
     int height = M - 1; /* row count */ \
-    long x = (offset - baseoffset) / (sizeof (element_type)); /* in elements */ \
+    long x = (offset - baseoffset) / (sizeof (contrib_type)); /* in elements */ \
     int2 coords = (int2)(x, 0); \
     OUT_VEC##M(u##contrib_type) DEFINE_BLOCK2D_RW_NAME(read, contrib_bitwidth, M, K)(long, int, int, int, int2); \
     OUT_VEC##M(u##contrib_type) res = DEFINE_BLOCK2D_RW_NAME(read, contrib_bitwidth, M, K)(baseoffset, width, height, pitch, coords); \
@@ -164,7 +164,7 @@ extern __constant int __JointMatrixLoadStoreOpt;
     int width = (sizeof (element_type)) * stride - 1; /* in bytes */ \
     int pitch = width; /* JointMatrices are expected to be contigunous in memory, without padding at the end of a row */ \
     int height = M - 1; /* row count */ \
-    long x = (offset - baseoffset) / (sizeof (element_type)); /* in elements */ \
+    long x = (offset - baseoffset) / (sizeof (contrib_type)); /* in elements */ \
     int2 coords = (int2)(x, 0); \
     void DEFINE_BLOCK2D_RW_NAME(write, contrib_bitwidth, M, K)(long, int, int, int, int2, OUT_VEC##M(u##contrib_type)); \
     OUT_VEC##M(u##contrib_type) val = VEC_TO_VEC##M(u##contrib_type, vec); \
@@ -304,7 +304,7 @@ DEFINE_LOAD(Accumulator_RowMajor, _SG16, int, 32, int, 32, 1, 16, 1x16, ROW_MAJO
 // set block_opt to false to disable block non-continous optimization per one built-in as a workaround
 #define DEFINE_STORE(layout, sg, element_type, elem_bitwidth, contrib_type, contrib_bitwidth, M, K, shape, order, us, stride_opt, block_opt) \
   INLINE void MANGLE_STORE_NAME(layout, sg, elem_bitwidth, shape) (char *mem, OUT_VEC##M(contrib_type) vec, int stride) { \
-      if (__JointMatrixLoadStoreOpt >= BLOCK2D_IMPL && (M == 2 || M == 4 || M == 8) && order == ROW_MAJOR) { \
+      if (__JointMatrixLoadStoreOpt >= BLOCK2D_IMPL && (M == 2 || M == 4 || M == 8) && order == ROW_MAJOR && elem_bitwidth > 8) { \
           IMPLEMENT_BLOCK2D_STORE##sg(element_type, contrib_type, contrib_bitwidth, M, K, vec) \
       } \
       if (__JointMatrixLoadStoreOpt >= VECTOR_CONT_IMPL && stride == stride_opt \
@@ -330,22 +330,19 @@ DEFINE_LOAD(Accumulator_RowMajor, _SG16, int, 32, int, 32, 1, 16, 1x16, ROW_MAJO
   }
 
 // TODO: investigate why intel_sub_group_block_write causes an assertion and enable blocked non-continuous optimization
-DEFINE_STORE(PackedA_RowMajor, , char, 8, int, 32, 8, 32, 8x32, ROW_MAJOR, , 32, false)
-
-// TODO: investigate why intel_sub_group_block_write causes an assertion and enable blocked non-continuous optimization
-DEFINE_STORE(PackedA_RowMajor, , short, 16, int, 32, 8, 16, 8x16, ROW_MAJOR, , 16, false)
-
-// TODO: investigate why intel_sub_group_block_write_us causes an assertion and enable blocked non-continuous optimization
-DEFINE_STORE(PackedA_RowMajor, _SG16, char, 8, short, 16, 8, 32, 8x32, ROW_MAJOR, _us, 32, false)
-
-// TODO: investigate why intel_sub_group_block_write_us causes an assertion and enable blocked non-continuous optimization
+DEFINE_STORE(PackedA_RowMajor,      , char,  8,  int,   32, 8, 32, 8x32, ROW_MAJOR,    , 32, false)
+DEFINE_STORE(PackedA_RowMajor,      , short, 16, int,   32, 8, 16, 8x16, ROW_MAJOR,    , 16, false)
+DEFINE_STORE(PackedA_RowMajor, _SG16, char,  8,  short, 16, 8, 32, 8x32, ROW_MAJOR, _us, 32, false)
 DEFINE_STORE(PackedA_RowMajor, _SG16, short, 16, short, 16, 8, 16, 8x16, ROW_MAJOR, _us, 16, false)
 
-DEFINE_STORE(PackedB_PackedB, , short, 16, int, 32, 8, 16, 16x8, ROW_MAJOR, , 16, true)
-DEFINE_STORE(PackedB_PackedB, , short, 16, int, 32, 8, 16, 16x16, ROW_MAJOR, , 32, true)
+DEFINE_STORE(PackedB_PackedB,      , short, 16, int, 32, 8, 16, 16x8,  ROW_MAJOR, , 16, true)
+DEFINE_STORE(PackedB_PackedB,      , short, 16, int, 32, 8, 16, 16x16, ROW_MAJOR, , 32, true)
+DEFINE_STORE(PackedB_PackedB, _SG16, short, 16, int, 32, 8, 16, 16x8,  ROW_MAJOR, , 16, true)
+DEFINE_STORE(PackedB_PackedB, _SG16, short, 16, int, 32, 8, 16, 16x16, ROW_MAJOR, , 32, true)
 
 // TODO: investigate why intel_sub_group_block_write causes an assertion and enable blocked non-continuous optimization
-DEFINE_STORE(PackedB_PackedB, , char, 8, int, 32, 8, 32, 32x8, ROW_MAJOR, , 16, false)
+DEFINE_STORE(PackedB_PackedB,      , char, 8, int, 32, 8, 32, 32x8, ROW_MAJOR, , 16, false)
+DEFINE_STORE(PackedB_PackedB, _SG16, char, 8, int, 32, 8, 32, 32x8, ROW_MAJOR, , 16, false)
 
 DEFINE_STORE(Accumulator_RowMajor, , int, 32, int, 32, 8, 8, 8x8, ROW_MAJOR, , 8, true)
 DEFINE_STORE(Accumulator_RowMajor, , int, 32, int, 32, 7, 8, 7x8, ROW_MAJOR, , 8, true)
@@ -356,7 +353,7 @@ DEFINE_STORE(Accumulator_RowMajor, , int, 32, int, 32, 3, 8, 3x8, ROW_MAJOR, , 8
 DEFINE_STORE(Accumulator_RowMajor, , int, 32, int, 32, 2, 8, 2x8, ROW_MAJOR, , 8, true)
 DEFINE_STORE(Accumulator_RowMajor, , int, 32, int, 32, 1, 8, 1x8, ROW_MAJOR, , 8, true)
 
-DEFINE_STORE(Accumulator_RowMajor, , int, 32, int, 32, 8, 16, 8x16, ROW_MAJOR, , 16, true)
+DEFINE_STORE(Accumulator_RowMajor,      , int, 32, int, 32, 8, 16, 8x16, ROW_MAJOR, , 16, true)
 
 DEFINE_STORE(Accumulator_RowMajor, _SG16, int, 32, int, 32, 8, 16, 8x16, ROW_MAJOR, , 16, true)
 DEFINE_STORE(Accumulator_RowMajor, _SG16, int, 32, int, 32, 7, 16, 7x16, ROW_MAJOR, , 16, true)
@@ -367,4 +364,5 @@ DEFINE_STORE(Accumulator_RowMajor, _SG16, int, 32, int, 32, 3, 16, 3x16, ROW_MAJ
 DEFINE_STORE(Accumulator_RowMajor, _SG16, int, 32, int, 32, 2, 16, 2x16, ROW_MAJOR, , 16, true)
 DEFINE_STORE(Accumulator_RowMajor, _SG16, int, 32, int, 32, 1, 16, 1x16, ROW_MAJOR, , 16, true)
 
-DEFINE_STORE(Accumulator_ColumnMajor, , int, 32, int, 32, 8, 8, 8x8, COL_MAJOR, , -1, false)
+DEFINE_STORE(Accumulator_ColumnMajor,      , int, 32, int, 32, 8, 8, 8x8, COL_MAJOR, , -1, false)
+DEFINE_STORE(Accumulator_ColumnMajor, _SG16, int, 32, int, 32, 8, 8, 8x8, COL_MAJOR, , -1, false)

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -20258,7 +20258,7 @@ void EmitPass::emitLSC2DBlockOperation(llvm::GenIntrinsicInst* inst)
 
     if (isRead == false)
     {
-        destination = GetSymbol(inst->getOperand(12));
+        destination = BroadcastIfUniform(GetSymbol(inst->getOperand(12)));
     }
 
     m_encoder->LSC_2DBlockMessage(