Updated according to feedback from CNugteren

36 files changed, 82 insertions, 79 deletions

diff --git a/Adreno_730_tuning_results.zip b/Adreno_730_tuning_results.zip
deleted file mode 100644
index 6fa29af5..00000000
--- a/Adreno_730_tuning_results.zip
+++ /dev/null
diff --git a/Adreno_740_tuning_results.zip b/Adreno_740_tuning_results.zip
deleted file mode 100644
index 6a5622b8..00000000
--- a/Adreno_740_tuning_results.zip
+++ /dev/null
diff --git a/CHANGELOG b/CHANGELOG
index 7088fb49..824f4520 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -1,5 +1,10 @@
 Development version (next version)
 - Fixes two small issues in the plotting script
+- Modifications to improve performance on Qualcomm Adreno GPUs:
+  * Unique database entries for specific Adreno devices
+  * Toggle OpenCL kernel compilation options for Adreno
+  * New preprocessor directive RELAX_WORKGROUP_SIZE
+- Fixed a bug in handling of #undef in CLBlast loop unrolling and array-to-register mapping functions
 
 Version 1.5.3
 - Fix a correctness issue with DGEMM on SM 7.5 Turing GPUs
diff --git a/README.md b/README.md
index 834d6221..cbecc606 100644
--- a/README.md
+++ b/README.md
@@ -101,8 +101,6 @@ Known performance related issues:
 
 * Severe performance issues with Beignet v1.3.0 due to missing support for local memory. Please downgrade to v1.2.1 or upgrade to v1.3.1 or newer.
 
-* Performance issues on Qualcomm Adreno GPUs.
-
 Other known issues:
 
 * Routines returning an integer are currently not properly tested for half-precision FP16: IHAMAX/IHAMIN/IHMAX/IHMIN
diff --git a/src/kernels/level1/xamax.opencl b/src/kernels/level1/xamax.opencl
index 3600b9d2..06a6773b 100644
--- a/src/kernels/level1/xamax.opencl
+++ b/src/kernels/level1/xamax.opencl
@@ -32,7 +32,7 @@ R"(
 // The main reduction kernel, performing the loading and the majority of the operation
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
 #endif
 void Xamax(const int n,
@@ -102,7 +102,7 @@ void Xamax(const int n,
 // be launched with a single workgroup only.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
 #endif
 void XamaxEpilogue(const __global singlereal* restrict maxgm,
diff --git a/src/kernels/level1/xasum.opencl b/src/kernels/level1/xasum.opencl
index 875221f4..683c6fad 100644
--- a/src/kernels/level1/xasum.opencl
+++ b/src/kernels/level1/xasum.opencl
@@ -32,7 +32,7 @@ R"(
 // The main reduction kernel, performing the loading and the majority of the operation
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
 #endif
 void Xasum(const int n,
@@ -79,7 +79,7 @@ void Xasum(const int n,
 // be launched with a single workgroup only.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
 #endif
 void XasumEpilogue(const __global real* restrict input,
diff --git a/src/kernels/level1/xaxpy.opencl b/src/kernels/level1/xaxpy.opencl
index b20ad200..a106ed01 100644
--- a/src/kernels/level1/xaxpy.opencl
+++ b/src/kernels/level1/xaxpy.opencl
@@ -24,7 +24,7 @@ R"(
 // Full version of the kernel with offsets and strided accesses
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void Xaxpy(const int n, const real_arg arg_alpha,
@@ -43,7 +43,7 @@ void Xaxpy(const int n, const real_arg arg_alpha,
 // assumes that 'n' is dividable by 'VW' and 'WPT'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XaxpyFaster(const int n, const real_arg arg_alpha,
@@ -67,7 +67,7 @@ void XaxpyFaster(const int n, const real_arg arg_alpha,
 // dividable by 'VW', 'WGS' and 'WPT'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XaxpyFastest(const int n, const real_arg arg_alpha,
@@ -89,7 +89,7 @@ void XaxpyFastest(const int n, const real_arg arg_alpha,
 // Full version of the kernel with offsets and strided accesses: batched version
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XaxpyBatched(const int n, const __constant real_arg* arg_alphas,
diff --git a/src/kernels/level1/xcopy.opencl b/src/kernels/level1/xcopy.opencl
index 174bf0c6..493197af 100644
--- a/src/kernels/level1/xcopy.opencl
+++ b/src/kernels/level1/xcopy.opencl
@@ -24,7 +24,7 @@ R"(
 // Full version of the kernel with offsets and strided accesses
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void Xcopy(const int n,
@@ -43,7 +43,7 @@ void Xcopy(const int n,
 // dividable by 'VW', 'WGS' and 'WPT'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XcopyFast(const int n,
diff --git a/src/kernels/level1/xdot.opencl b/src/kernels/level1/xdot.opencl
index e14b6306..64f6eb9d 100644
--- a/src/kernels/level1/xdot.opencl
+++ b/src/kernels/level1/xdot.opencl
@@ -32,7 +32,7 @@ R"(
 // The main reduction kernel, performing the multiplication and the majority of the sum operation
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
 #endif
 void Xdot(const int n,
@@ -78,7 +78,7 @@ void Xdot(const int n,
 // be launched with a single workgroup only.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
 #endif
 void XdotEpilogue(const __global real* restrict input,
diff --git a/src/kernels/level1/xhad.opencl b/src/kernels/level1/xhad.opencl
index aee98f91..47bb5170 100644
--- a/src/kernels/level1/xhad.opencl
+++ b/src/kernels/level1/xhad.opencl
@@ -68,7 +68,7 @@ INLINE_FUNC realV MultiplyVectorVector(realV cvec, const realV aval, const realV
 // Full version of the kernel with offsets and strided accesses
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void Xhad(const int n, const real_arg arg_alpha, const real_arg arg_beta,
@@ -96,7 +96,7 @@ void Xhad(const int n, const real_arg arg_alpha, const real_arg arg_beta,
 // assumes that 'n' is dividable by 'VW' and 'WPT'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XhadFaster(const int n, const real_arg arg_alpha, const real_arg arg_beta,
@@ -127,7 +127,7 @@ void XhadFaster(const int n, const real_arg arg_alpha, const real_arg arg_beta,
 // dividable by 'VW', 'WGS' and 'WPT'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XhadFastest(const int n, const real_arg arg_alpha, const real_arg arg_beta,
diff --git a/src/kernels/level1/xnrm2.opencl b/src/kernels/level1/xnrm2.opencl
index fb45effb..36ea49b4 100644
--- a/src/kernels/level1/xnrm2.opencl
+++ b/src/kernels/level1/xnrm2.opencl
@@ -32,7 +32,7 @@ R"(
 // The main reduction kernel, performing the multiplication and the majority of the operation
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
 #endif
 void Xnrm2(const int n,
@@ -77,7 +77,7 @@ void Xnrm2(const int n,
 // be launched with a single workgroup only.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
 #endif
 void Xnrm2Epilogue(const __global real* restrict input,
diff --git a/src/kernels/level1/xscal.opencl b/src/kernels/level1/xscal.opencl
index 19ca9135..e4260c7c 100644
--- a/src/kernels/level1/xscal.opencl
+++ b/src/kernels/level1/xscal.opencl
@@ -24,7 +24,7 @@ R"(
 // Full version of the kernel with offsets and strided accesses
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void Xscal(const int n, const real_arg arg_alpha,
@@ -46,7 +46,7 @@ void Xscal(const int n, const real_arg arg_alpha,
 // dividable by 'VW', 'WGS' and 'WPT'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XscalFast(const int n, const real_arg arg_alpha,
diff --git a/src/kernels/level1/xswap.opencl b/src/kernels/level1/xswap.opencl
index a2b44de3..2d384423 100644
--- a/src/kernels/level1/xswap.opencl
+++ b/src/kernels/level1/xswap.opencl
@@ -24,7 +24,7 @@ R"(
 // Full version of the kernel with offsets and strided accesses
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void Xswap(const int n,
@@ -45,7 +45,7 @@ void Xswap(const int n,
 // dividable by 'VW', 'WGS' and 'WPT'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS, 1, 1)))
 #endif
 void XswapFast(const int n,
diff --git a/src/kernels/level2/xgemv.opencl b/src/kernels/level2/xgemv.opencl
index 16711aa4..15912a60 100644
--- a/src/kernels/level2/xgemv.opencl
+++ b/src/kernels/level2/xgemv.opencl
@@ -212,7 +212,7 @@ INLINE_FUNC real LoadMatrixA(const __global real* restrict agm, const int x, con
 // Full version of the kernel
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
 #endif
 void Xgemv(const int m, const int n,
diff --git a/src/kernels/level2/xgemv_fast.opencl b/src/kernels/level2/xgemv_fast.opencl
index 853d3d28..46087af7 100644
--- a/src/kernels/level2/xgemv_fast.opencl
+++ b/src/kernels/level2/xgemv_fast.opencl
@@ -90,7 +90,7 @@ INLINE_FUNC realVF LoadMatrixAVF(const __global realVF* restrict agm, const int
 // --> 'do_conjugate' is 0
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
 #endif
 void XgemvFast(const int m, const int n,
@@ -197,7 +197,7 @@ void XgemvFast(const int m, const int n,
 // --> 'do_conjugate' is 0
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS3, 1, 1)))
 #endif
 void XgemvFastRot(const int m, const int n,
diff --git a/src/kernels/level2/xger.opencl b/src/kernels/level2/xger.opencl
index 3620c66a..b1fe8447 100644
--- a/src/kernels/level2/xger.opencl
+++ b/src/kernels/level2/xger.opencl
@@ -20,7 +20,7 @@ R"(
 // Regular version of the rank-1 matrix update kernel (GER, GERU, GERC)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
 #endif
 void Xger(const int max1, const int max2,
diff --git a/src/kernels/level2/xher.opencl b/src/kernels/level2/xher.opencl
index 9e460cd4..eac8c10a 100644
--- a/src/kernels/level2/xher.opencl
+++ b/src/kernels/level2/xher.opencl
@@ -20,7 +20,7 @@ R"(
 // Symmetric version of the rank-1 matrix update kernel (HER, HPR, SYR, SPR)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
 #endif
 void Xher(const int n,
diff --git a/src/kernels/level2/xher2.opencl b/src/kernels/level2/xher2.opencl
index c3e85c15..9e7f3c6c 100644
--- a/src/kernels/level2/xher2.opencl
+++ b/src/kernels/level2/xher2.opencl
@@ -20,7 +20,7 @@ R"(
 // Symmetric version of the rank-2 matrix update kernel (HER2, HPR2, SYR2, SPR2)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
 #endif
 void Xher2(const int n,
diff --git a/src/kernels/level2/xtrsv.opencl b/src/kernels/level2/xtrsv.opencl
index e3b5418c..7677377e 100644
--- a/src/kernels/level2/xtrsv.opencl
+++ b/src/kernels/level2/xtrsv.opencl
@@ -41,7 +41,7 @@ void FillVector(const int n, const int inc, const int offset,
 
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
 #endif
 void trsv_forward(int n,
@@ -93,7 +93,7 @@ void trsv_forward(int n,
 
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
 #endif
 void trsv_backward(int n,
diff --git a/src/kernels/level3/convert_hermitian.opencl b/src/kernels/level3/convert_hermitian.opencl
index b69be23d..4bb61f4e 100644
--- a/src/kernels/level3/convert_hermitian.opencl
+++ b/src/kernels/level3/convert_hermitian.opencl
@@ -23,7 +23,7 @@ R"(
 // stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void HermLowerToSquared(const int src_dim,
@@ -66,7 +66,7 @@ void HermLowerToSquared(const int src_dim,
 // Same as above, but now the matrix' data is stored in the upper-triangle
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void HermUpperToSquared(const int src_dim,
diff --git a/src/kernels/level3/convert_symmetric.opencl b/src/kernels/level3/convert_symmetric.opencl
index 2ce17f40..264bf9c5 100644
--- a/src/kernels/level3/convert_symmetric.opencl
+++ b/src/kernels/level3/convert_symmetric.opencl
@@ -22,7 +22,7 @@ R"(
 // stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void SymmLowerToSquared(const int src_dim,
@@ -59,7 +59,7 @@ void SymmLowerToSquared(const int src_dim,
 // Same as above, but now the matrix' data is stored in the upper-triangle
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void SymmUpperToSquared(const int src_dim,
diff --git a/src/kernels/level3/convert_triangular.opencl b/src/kernels/level3/convert_triangular.opencl
index 563f719f..092e1cf2 100644
--- a/src/kernels/level3/convert_triangular.opencl
+++ b/src/kernels/level3/convert_triangular.opencl
@@ -22,7 +22,7 @@ R"(
 // stored as the lower-triangle of the input matrix. This uses the padding kernel's parameters.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void TriaLowerToSquared(const int src_dim,
@@ -61,7 +61,7 @@ void TriaLowerToSquared(const int src_dim,
 // Same as above, but now the matrix' data is stored in the upper-triangle
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void TriaUpperToSquared(const int src_dim,
diff --git a/src/kernels/level3/copy_fast.opencl b/src/kernels/level3/copy_fast.opencl
index e1a815f5..04e4b2ab 100644
--- a/src/kernels/level3/copy_fast.opencl
+++ b/src/kernels/level3/copy_fast.opencl
@@ -37,7 +37,7 @@ R"(
 // COPY_VW. Also requires both matrices to be of the same dimensions and without offset.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
 #endif
 void CopyMatrixFast(const int ld,
diff --git a/src/kernels/level3/copy_pad.opencl b/src/kernels/level3/copy_pad.opencl
index 6335bd81..dca93b76 100644
--- a/src/kernels/level3/copy_pad.opencl
+++ b/src/kernels/level3/copy_pad.opencl
@@ -61,7 +61,7 @@ INLINE_FUNC void _CopyPadMatrix(const int src_one, const int src_two,
 // Interface to the above function
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void CopyPadMatrix(const int src_one, const int src_two,
@@ -124,7 +124,7 @@ INLINE_FUNC void _CopyMatrix(const int src_one, const int src_two,
 // Interface to the above function
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void CopyMatrix(const int src_one, const int src_two,
@@ -148,7 +148,7 @@ void CopyMatrix(const int src_one, const int src_two,
 // Batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void CopyPadMatrixBatched(const int src_one, const int src_two,
@@ -170,7 +170,7 @@ void CopyPadMatrixBatched(const int src_one, const int src_two,
 // Batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void CopyMatrixBatched(const int src_one, const int src_two,
@@ -195,7 +195,7 @@ void CopyMatrixBatched(const int src_one, const int src_two,
 // Strided-batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void CopyPadMatrixStridedBatched(const int src_one, const int src_two,
@@ -217,7 +217,7 @@ void CopyPadMatrixStridedBatched(const int src_one, const int src_two,
 // Strided-batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PAD_DIMX, PAD_DIMY, 1)))
 #endif
 void CopyMatrixStridedBatched(const int src_one, const int src_two,
diff --git a/src/kernels/level3/invert_diagonal_blocks_part1.opencl b/src/kernels/level3/invert_diagonal_blocks_part1.opencl
index 3df477d1..580f7b8b 100644
--- a/src/kernels/level3/invert_diagonal_blocks_part1.opencl
+++ b/src/kernels/level3/invert_diagonal_blocks_part1.opencl
@@ -84,7 +84,7 @@ R"(
 // Inverts a diagonal block of INTERNAL_BLOCK_SIZE by INTERNAL_BLOCK_SIZE elements in a larger matrix
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(INTERNAL_BLOCK_SIZE, 1, 1)))
 #endif
 void InvertDiagonalBlock(const int n, __global const real* restrict src, const int src_offset, const int src_ld,
diff --git a/src/kernels/level3/transpose_fast.opencl b/src/kernels/level3/transpose_fast.opencl
index e89984cc..9d1b7552 100644
--- a/src/kernels/level3/transpose_fast.opencl
+++ b/src/kernels/level3/transpose_fast.opencl
@@ -38,7 +38,7 @@ R"(
 // offset. A more general version is available in 'padtranspose.opencl'.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(TRA_DIM, TRA_DIM, 1)))
 #endif
 void TransposeMatrixFast(const int ld,
diff --git a/src/kernels/level3/transpose_pad.opencl b/src/kernels/level3/transpose_pad.opencl
index 31de0e62..3877a3d5 100644
--- a/src/kernels/level3/transpose_pad.opencl
+++ b/src/kernels/level3/transpose_pad.opencl
@@ -86,7 +86,7 @@ INLINE_FUNC void _TransposePadMatrix(LOCAL_PTR real* tile,
 // Interface to the above function
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
 #endif
 void TransposePadMatrix(const int src_one, const int src_two,
@@ -178,7 +178,7 @@ INLINE_FUNC void _TransposeMatrix(LOCAL_PTR real* tile,
 // Interface to the above function
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
 #endif
 void TransposeMatrix(const int src_one, const int src_two,
@@ -203,7 +203,7 @@ void TransposeMatrix(const int src_one, const int src_two,
 // Batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
 #endif
 void TransposePadMatrixBatched(const int src_one, const int src_two,
@@ -226,7 +226,7 @@ void TransposePadMatrixBatched(const int src_one, const int src_two,
 // Batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
 #endif
 void TransposeMatrixBatched(const int src_one, const int src_two,
@@ -252,7 +252,7 @@ void TransposeMatrixBatched(const int src_one, const int src_two,
 // Strided-batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
 #endif
 void TransposePadMatrixStridedBatched(const int src_one, const int src_two,
@@ -275,7 +275,7 @@ void TransposePadMatrixStridedBatched(const int src_one, const int src_two,
 // Strided-batched version of the above
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(PADTRA_TILE, PADTRA_TILE, 1)))
 #endif
 void TransposeMatrixStridedBatched(const int src_one, const int src_two,
diff --git a/src/kernels/level3/xgemm_batched.opencl b/src/kernels/level3/xgemm_batched.opencl
index e014b7a9..41d07d19 100644
--- a/src/kernels/level3/xgemm_batched.opencl
+++ b/src/kernels/level3/xgemm_batched.opencl
@@ -21,7 +21,7 @@ R"(
 
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
 #endif
 void XgemmBatched(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -68,7 +68,7 @@ void XgemmBatched(const int kSizeM, const int kSizeN, const int kSizeK,
 
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
 #endif
 void XgemmStridedBatched(const int kSizeM, const int kSizeN, const int kSizeK,
diff --git a/src/kernels/level3/xgemm_direct_batched.opencl b/src/kernels/level3/xgemm_direct_batched.opencl
index ec0b008b..102ae762 100644
--- a/src/kernels/level3/xgemm_direct_batched.opencl
+++ b/src/kernels/level3/xgemm_direct_batched.opencl
@@ -22,7 +22,7 @@ R"(
 // Direct version of the batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -47,7 +47,7 @@ void XgemmDirectBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the batched GEMM kernel with [A, B] = [non-transposed, transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -72,7 +72,7 @@ void XgemmDirectBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the batched GEMM kernel with [A, B] = [transposed, non-transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -97,7 +97,7 @@ void XgemmDirectBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the batched GEMM kernel with [A, B] = [transposed, transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -126,7 +126,7 @@ void XgemmDirectBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, non-transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectStridedBatchedNN(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -149,7 +149,7 @@ void XgemmDirectStridedBatchedNN(const int kSizeM, const int kSizeN, const int k
 // Direct version of the strided-batched GEMM kernel with [A, B] = [non-transposed, transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectStridedBatchedNT(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -172,7 +172,7 @@ void XgemmDirectStridedBatchedNT(const int kSizeM, const int kSizeN, const int k
 // Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, non-transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectStridedBatchedTN(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -195,7 +195,7 @@ void XgemmDirectStridedBatchedTN(const int kSizeM, const int kSizeN, const int k
 // Direct version of the strided-batched GEMM kernel with [A, B] = [transposed, transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectStridedBatchedTT(const int kSizeM, const int kSizeN, const int kSizeK,
diff --git a/src/kernels/level3/xgemm_direct_part3.opencl b/src/kernels/level3/xgemm_direct_part3.opencl
index 66b880e6..5508170e 100644
--- a/src/kernels/level3/xgemm_direct_part3.opencl
+++ b/src/kernels/level3/xgemm_direct_part3.opencl
@@ -220,7 +220,7 @@ INLINE_FUNC void XgemmDirect(const int kSizeM, const int kSizeN, const int kSize
 // Direct version of the GEMM kernel with [A, B] = [non-transposed, non-transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectNN(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -239,7 +239,7 @@ void XgemmDirectNN(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the GEMM kernel with [A, B] = [non-transposed, transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectNT(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -258,7 +258,7 @@ void XgemmDirectNT(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the GEMM kernel with [A, B] = [transposed, non-transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectTN(const int kSizeM, const int kSizeN, const int kSizeK,
@@ -277,7 +277,7 @@ void XgemmDirectTN(const int kSizeM, const int kSizeN, const int kSizeK,
 // Direct version of the GEMM kernel with [A, B] = [transposed, transposed]
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XgemmDirectTT(const int kSizeM, const int kSizeN, const int kSizeK,
diff --git a/src/kernels/level3/xgemm_part4.opencl b/src/kernels/level3/xgemm_part4.opencl
index a64e2efa..05524337 100644
--- a/src/kernels/level3/xgemm_part4.opencl
+++ b/src/kernels/level3/xgemm_part4.opencl
@@ -21,7 +21,7 @@ R"(
 // Main entry point of the kernel. This is the upper-triangular version.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
 #endif
 void XgemmUpper(const int kSizeN, const int kSizeK,
@@ -61,7 +61,7 @@ void XgemmUpper(const int kSizeN, const int kSizeK,
 // Main entry point of the kernel. This is the lower-triangular version.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
 #endif
 void XgemmLower(const int kSizeN, const int kSizeK,
@@ -105,7 +105,7 @@ void XgemmLower(const int kSizeN, const int kSizeK,
 // Main entry point of the kernel. This is the regular full version.
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMC, NDIMC, 1)))
 #endif
 void Xgemm(const int kSizeM, const int kSizeN, const int kSizeK,
diff --git a/src/kernels/levelx/col2im.opencl b/src/kernels/levelx/col2im.opencl
index fcc307c6..ab0ffbfa 100644
--- a/src/kernels/levelx/col2im.opencl
+++ b/src/kernels/levelx/col2im.opencl
@@ -94,7 +94,7 @@ INLINE_FUNC void Xcol2im(const int input_h, const int input_w, const int channel
 // Kernel flip version of the Xcol2im kernel (for convolution)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
 #endif
 void Xcol2imKernelFlip(const int input_h, const int input_w, const int channels,
@@ -119,7 +119,7 @@ void Xcol2imKernelFlip(const int input_h, const int input_w, const int channels,
 // Normal version of the Xcol2im kernel (for cross-correlation)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
 #endif
 void Xcol2imKernelNormal(const int input_h, const int input_w, const int channels,
diff --git a/src/kernels/levelx/im2col.opencl b/src/kernels/levelx/im2col.opencl
index 8324468e..59af38fc 100644
--- a/src/kernels/levelx/im2col.opencl
+++ b/src/kernels/levelx/im2col.opencl
@@ -76,7 +76,7 @@ INLINE_FUNC void Xim2col(const int input_h, const int input_w, const int channel
 // Kernel flip version of the Xim2col kernel (for convolution)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
 #endif
 void Xim2colKernelFlip(const int input_h, const int input_w, const int channels,
@@ -97,7 +97,7 @@ void Xim2colKernelFlip(const int input_h, const int input_w, const int channels,
 // Normal version of the Xim2col kernel (for cross-correlation)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(COPY_DIMX, COPY_DIMY, 1)))
 #endif
 void Xim2colKernelNormal(const int input_h, const int input_w, const int channels,
diff --git a/src/kernels/levelx/xconvgemm_part2.opencl b/src/kernels/levelx/xconvgemm_part2.opencl
index 79c40f59..38ddd7eb 100644
--- a/src/kernels/levelx/xconvgemm_part2.opencl
+++ b/src/kernels/levelx/xconvgemm_part2.opencl
@@ -25,7 +25,7 @@ R"(
 #if defined(CONVGEMM_WITH_IM2COL)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void Xconvgemm(const int num_patches, const int num_kernels, const int patch_size,
@@ -291,7 +291,7 @@ INLINE_FUNC void Xconvgemm(const int num_patches, const int num_kernels, const i
 #if !defined(CONVGEMM_WITH_IM2COL)
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XconvgemmFlip(const int num_patches, const int num_kernels, const int patch_size,
@@ -316,7 +316,7 @@ void XconvgemmFlip(const int num_patches, const int num_kernels, const int patch
 
 #if RELAX_WORKGROUP_SIZE == 1
   __kernel
-#elif
+#else
   __kernel __attribute__((reqd_work_group_size(MDIMCD, NDIMCD, 1)))
 #endif
 void XconvgemmNormal(const int num_patches, const int num_kernels, const int patch_size,
diff --git a/src/utilities/compile.cpp b/src/utilities/compile.cpp
index 7170c30a..59aa6107 100644
--- a/src/utilities/compile.cpp
+++ b/src/utilities/compile.cpp
@@ -43,7 +43,7 @@ std::shared_ptr<Program> CompileFromSource(
 
   // For specific devices, use the non-IEE754 compliant OpenCL mad() instruction. This can improve
   // performance, but might result in a reduced accuracy.
-  if ((device.IsAMD() && device.IsGPU()) || device.IsQualcomm()) {
+  if ((device.IsAMD() && device.IsGPU()) || (device.IsQualcomm() && device.IsGPU())) {
     header_string += "#define USE_CL_MAD 1\n";
   }
 
@@ -54,7 +54,7 @@ std::shared_ptr<Program> CompileFromSource(
 
   // For specific devices add a global synchronisation barrier to the GEMM kernel to optimize
   // performance through better cache behaviour
-  if ((device.IsARM() && device.IsGPU()) || device.IsQualcomm()) {
+  if ((device.IsARM() && device.IsGPU()) || (device.IsQualcomm() && device.IsGPU())) {
     header_string += "#define GLOBAL_MEM_FENCE 1\n";
   }
 
diff --git a/src/utilities/utilities.cpp b/src/utilities/utilities.cpp
index 32de2e2e..fbdcf9c2 100644
--- a/src/utilities/utilities.cpp
+++ b/src/utilities/utilities.cpp
@@ -463,7 +463,7 @@ std::string GetDeviceArchitecture(const Device& device) {
     else if (device.HasExtension(kKhronosAttributesAMD)) {
       device_architecture = device.Name(); // Name is architecture for AMD APP and AMD ROCm
     }
-    else if (device.IsQualcomm()) { // queries the Adreno GPU architecture version
+    else if ((device.IsQualcomm() && device.IsGPU())) { // queries the Adreno GPU architecture version
       device_architecture = device.AdrenoVersion();
     }
     // Note: no else - 'device_architecture' might be the empty string