implemented changes to boost Adreno performance according to https://jira-dc.qualcomm.com/jira/browse/OSR-8731

6 files changed, 40 insertions, 8 deletions

diff --git a/src/kernels/level2/xgemv.opencl b/src/kernels/level2/xgemv.opencl
index ba29aba6..16711aa4 100644
--- a/src/kernels/level2/xgemv.opencl
+++ b/src/kernels/level2/xgemv.opencl
@@ -210,7 +210,11 @@ INLINE_FUNC real LoadMatrixA(const __global real* restrict agm, const int x, con
 // =================================================================================================
 
 // Full version of the kernel
-__kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(WGS1, 1, 1)))
+#endif
 void Xgemv(const int m, const int n,
                     const real_arg arg_alpha,
                     const real_arg arg_beta,
diff --git a/src/kernels/level2/xgemv_fast.opencl b/src/kernels/level2/xgemv_fast.opencl
index 45ceb36c..853d3d28 100644
--- a/src/kernels/level2/xgemv_fast.opencl
+++ b/src/kernels/level2/xgemv_fast.opencl
@@ -88,7 +88,11 @@ INLINE_FUNC realVF LoadMatrixAVF(const __global realVF* restrict agm, const int
 // --> 'a_ld' is a multiple of VW2
 // --> 'a_rotated' is 0
 // --> 'do_conjugate' is 0
-__kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(WGS2, 1, 1)))
+#endif
 void XgemvFast(const int m, const int n,
                const real_arg arg_alpha,
                const real_arg arg_beta,
@@ -191,7 +195,11 @@ void XgemvFast(const int m, const int n,
 // --> 'a_ld' is a multiple of VW3
 // --> 'a_rotated' is 1
 // --> 'do_conjugate' is 0
-__kernel __attribute__((reqd_work_group_size(WGS3, 1, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(WGS3, 1, 1)))
+#endif
 void XgemvFastRot(const int m, const int n,
                   const real_arg arg_alpha,
                   const real_arg arg_beta,
diff --git a/src/kernels/level2/xger.opencl b/src/kernels/level2/xger.opencl
index ca6071cd..3620c66a 100644
--- a/src/kernels/level2/xger.opencl
+++ b/src/kernels/level2/xger.opencl
@@ -18,7 +18,11 @@ R"(
 // =================================================================================================
 
 // Regular version of the rank-1 matrix update kernel (GER, GERU, GERC)
-__kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
+#endif
 void Xger(const int max1, const int max2,
           const real_arg arg_alpha,
           const __global real* restrict xgm, const int x_offset, const int x_inc,
diff --git a/src/kernels/level2/xher.opencl b/src/kernels/level2/xher.opencl
index 8a57bdfc..9e460cd4 100644
--- a/src/kernels/level2/xher.opencl
+++ b/src/kernels/level2/xher.opencl
@@ -18,7 +18,11 @@ R"(
 // =================================================================================================
 
 // Symmetric version of the rank-1 matrix update kernel (HER, HPR, SYR, SPR)
-__kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
+#endif
 void Xher(const int n,
           const real_arg arg_alpha,
           const __global real* restrict xgm, const int x_offset, const int x_inc,
diff --git a/src/kernels/level2/xher2.opencl b/src/kernels/level2/xher2.opencl
index 73305149..c3e85c15 100644
--- a/src/kernels/level2/xher2.opencl
+++ b/src/kernels/level2/xher2.opencl
@@ -18,7 +18,11 @@ R"(
 // =================================================================================================
 
 // Symmetric version of the rank-2 matrix update kernel (HER2, HPR2, SYR2, SPR2)
-__kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(WGS1, WGS2, 1)))
+#endif
 void Xher2(const int n,
            const real_arg arg_alpha,
            const __global real* restrict xgm, const int x_offset, const int x_inc,
diff --git a/src/kernels/level2/xtrsv.opencl b/src/kernels/level2/xtrsv.opencl
index e7b6ae79..e3b5418c 100644
--- a/src/kernels/level2/xtrsv.opencl
+++ b/src/kernels/level2/xtrsv.opencl
@@ -39,7 +39,11 @@ void FillVector(const int n, const int inc, const int offset,
 
 // =================================================================================================
 
-__kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
+#endif
 void trsv_forward(int n,
                   const __global real *A, const int a_offset, int a_ld,
                   __global real *b, const int b_offset, int b_inc,
@@ -87,7 +91,11 @@ void trsv_forward(int n,
   }
 }
 
-__kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
+#if RELAX_WORKGROUP_SIZE == 1
+  __kernel
+#elif
+  __kernel __attribute__((reqd_work_group_size(TRSV_BLOCK_SIZE, 1, 1)))
+#endif
 void trsv_backward(int n,
                    const __global real *A, const int a_offset, int a_ld,
                    __global real *b, const int b_offset, int b_inc,