[SYCL][ESIMD] - change in barrier and linear example (#142)

Gang Chen · web-flow · commit 8e2ab8fa9976 · 2021-02-15T09:24:16.000+03:00
Signed-off-by: Gang Y Chen &lt;gang.y.chen@intel.com&gt;
diff --git a/SYCL/ESIMD/histogram_256_slm.cpp b/SYCL/ESIMD/histogram_256_slm.cpp
@@ -37,7 +37,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t *input_ptr, uint32_t *output,
   slm_offset *= sizeof(int);
   simd<uint, 16> slm_data = 0;
   slm_store<uint, 16>(slm_data, slm_offset);
-  esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);
   esimd_barrier();
 
   // Each thread handles NUM_BLOCKSxBLOCK_WIDTH pixel blocks
@@ -56,7 +55,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t *input_ptr, uint32_t *output,
     }
     start_off += BLOCK_WIDTH;
   }
-  esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);
   esimd_barrier();
 
   // Update global sum by atomically adding each local histogram
diff --git a/SYCL/ESIMD/histogram_256_slm_spec.cpp b/SYCL/ESIMD/histogram_256_slm_spec.cpp
@@ -38,7 +38,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t *input_ptr, uint32_t *output,
   slm_offset *= sizeof(int);
   simd<uint, 16> slm_data = 0;
   slm_store<uint, 16>(slm_data, slm_offset);
-  esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);
   esimd_barrier();
 
   // Each thread handles NUM_BLOCKSxBLOCK_WIDTH pixel blocks
@@ -57,7 +56,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t *input_ptr, uint32_t *output,
     }
     start_off += BLOCK_WIDTH;
   }
-  esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);
   esimd_barrier();
 
   // Update global sum by atomically adding each local histogram
diff --git a/SYCL/ESIMD/linear/linear.cpp b/SYCL/ESIMD/linear/linear.cpp
@@ -104,17 +104,15 @@ int main(int argc, char *argv[]) {
             in = media_block_load<unsigned char, 8, 32>(accInput, h_pos * 24,
                                                         v_pos * 6);
 
-            simd<float, 8 * 32> vin_float = vin;
-            auto in_float = vin_float.format<float, 8, 32>();
-            m = in_float.select<6, 1, 24, 1>(1, 3);
-            m += in_float.select<6, 1, 24, 1>(0, 0);
-            m += in_float.select<6, 1, 24, 1>(0, 3);
-            m += in_float.select<6, 1, 24, 1>(0, 6);
-            m += in_float.select<6, 1, 24, 1>(1, 0);
-            m += in_float.select<6, 1, 24, 1>(1, 6);
-            m += in_float.select<6, 1, 24, 1>(2, 0);
-            m += in_float.select<6, 1, 24, 1>(2, 3);
-            m += in_float.select<6, 1, 24, 1>(2, 6);
+            m = in.select<6, 1, 24, 1>(1, 3);
+            m += in.select<6, 1, 24, 1>(0, 0);
+            m += in.select<6, 1, 24, 1>(0, 3);
+            m += in.select<6, 1, 24, 1>(0, 6);
+            m += in.select<6, 1, 24, 1>(1, 0);
+            m += in.select<6, 1, 24, 1>(1, 6);
+            m += in.select<6, 1, 24, 1>(2, 0);
+            m += in.select<6, 1, 24, 1>(2, 3);
+            m += in.select<6, 1, 24, 1>(2, 6);
             m = m * 0.111f;
 
             vout = vm;
diff --git a/SYCL/ESIMD/slm_barrier.cpp b/SYCL/ESIMD/slm_barrier.cpp
@@ -68,7 +68,6 @@ void load_to_slm(uint grpSize, uint localId, uint slmOffset, char *addr,
     vOffsets += (grpSize * 256);
   }
 
-  esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);
   esimd_barrier();
 }
 

Original file line number	Diff line number	Diff line change
`@@ -37,7 +37,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t input_ptr, uint32_t output,`
`37`	`37`	`slm_offset *= sizeof(int);`
`38`	`38`	`simd<uint, 16> slm_data = 0;`
`39`	`39`	`slm_store<uint, 16>(slm_data, slm_offset);`
`40`		`- esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);`
`41`	`40`	`esimd_barrier();`
`42`	`41`
`43`	`42`	`// Each thread handles NUM_BLOCKSxBLOCK_WIDTH pixel blocks`
`@@ -56,7 +55,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t input_ptr, uint32_t output,`
`56`	`55`	`}`
`57`	`56`	`start_off += BLOCK_WIDTH;`
`58`	`57`	`}`
`59`		`- esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);`
`60`	`58`	`esimd_barrier();`
`61`	`59`
`62`	`60`	`// Update global sum by atomically adding each local histogram`
Original file line number	Diff line number	Diff line change
`@@ -38,7 +38,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t input_ptr, uint32_t output,`
`38`	`38`	`slm_offset *= sizeof(int);`
`39`	`39`	`simd<uint, 16> slm_data = 0;`
`40`	`40`	`slm_store<uint, 16>(slm_data, slm_offset);`
`41`		`- esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);`
`42`	`41`	`esimd_barrier();`
`43`	`42`
`44`	`43`	`// Each thread handles NUM_BLOCKSxBLOCK_WIDTH pixel blocks`
`@@ -57,7 +56,6 @@ ESIMD_INLINE void histogram_atomic(const uint32_t input_ptr, uint32_t output,`
`57`	`56`	`}`
`58`	`57`	`start_off += BLOCK_WIDTH;`
`59`	`58`	`}`
`60`		`- esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);`
`61`	`59`	`esimd_barrier();`
`62`	`60`
`63`	`61`	`// Update global sum by atomically adding each local histogram`
Original file line number	Diff line number	Diff line change
`@@ -68,7 +68,6 @@ void load_to_slm(uint grpSize, uint localId, uint slmOffset, char *addr,`
`68`	`68`	`vOffsets += (grpSize * 256);`
`69`	`69`	`}`
`70`	`70`
`71`		`- esimd_fence(ESIMD_GLOBAL_COHERENT_FENCE);`
`72`	`71`	`esimd_barrier();`
`73`	`72`	`}`
`74`	`73`