Merge pull request #276 from sampath1117/sr/remap_changes

Removed unnecessary inline helpers used in water kernel
r-abishek · Jun 4, 2024 · 14d5c8f · 14d5c8f
2 parents 69aed52 + d02a4c3
commit 14d5c8f
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 14 deletions.
diff --git a/src/include/cpu/rpp_cpu_simd.hpp b/src/include/cpu/rpp_cpu_simd.hpp
@@ -3085,12 +3085,6 @@ inline void rpp_store24_f32pln3_to_u8pkd3_avx(Rpp8u* dstPtr, __m256* p)
     _mm256_storeu_si256((__m256i *)(dstPtr), px1);          /* store the 24 U8 pixels in dst */
 }
 
-inline void rpp_store8_u8pln1_to_u8pln1_avx(Rpp8u* dstPtr, __m256i &p)
-{
-    __m128i pTemp = _mm256_castsi256_si128(p);
-    rpp_storeu_si64((__m128i *)(dstPtr), pTemp);
-}
-
 inline void rpp_store8_f32pln1_to_u8pln1_avx(Rpp8u* dstPtr, __m256 &p)
 {
     __m256i px1 = _mm256_permute4x64_epi64(_mm256_packus_epi32(_mm256_cvtps_epi32(p), avx_px0), _MM_SHUFFLE(3,1,2,0));
@@ -3256,12 +3250,6 @@ inline void rpp_store24_f32pln3_to_i8pkd3_avx(Rpp8s* dstPtr, __m256* p)
     _mm256_storeu_si256((__m256i *)(dstPtr), px1);          /* store the 12 U8 pixels in dst */
 }
 
-inline void rpp_store8_i8pln1_to_i8pln1_avx(Rpp8s* dstPtr, __m256i &p)
-{
-    __m128i pTemp = _mm256_castsi256_si128(p);
-    rpp_storeu_si64((__m128i *)(dstPtr), pTemp);
-}
-
 inline void rpp_store8_f32pln1_to_i8pln1_avx(Rpp8s* dstPtr, __m256 &p)
 {
     __m256i px1 = _mm256_permute4x64_epi64(_mm256_packus_epi32(_mm256_cvtps_epi32(p), avx_px0), _MM_SHUFFLE(3,1,2,0));

diff --git a/src/modules/cpu/kernel/water.hpp b/src/modules/cpu/kernel/water.hpp
@@ -281,7 +281,7 @@ RppStatus water_u8_u8_host_tensor(Rpp8u *srcPtr,
                     {
                         __m256i pRow;
                         rpp_simd_load(rpp_generic_nn_load_u8pln1_avx, srcPtrTempChn, srcLocArray, invalidLoad, pRow);
-                        rpp_simd_store(rpp_store8_u8pln1_to_u8pln1_avx, dstPtrTempChn, pRow);
+                        rpp_storeu_si64((__m128i *)(dstPtrTempChn), _mm256_castsi256_si128(pRow));
                         srcPtrTempChn += srcDescPtr->strides.cStride;
                         dstPtrTempChn += dstDescPtr->strides.cStride;
                     }
@@ -965,7 +965,7 @@ RppStatus water_i8_i8_host_tensor(Rpp8s *srcPtr,
                     {
                         __m256i pRow;
                         rpp_simd_load(rpp_generic_nn_load_i8pln1_avx, srcPtrTempChn, srcLocArray, invalidLoad, pRow);
-                        rpp_simd_store(rpp_store8_i8pln1_to_i8pln1_avx, dstPtrTempChn, pRow);
+                        rpp_storeu_si64((__m128i *)(dstPtrTempChn), _mm256_castsi256_si128(pRow));
                         srcPtrTempChn += srcDescPtr->strides.cStride;
                         dstPtrTempChn += dstDescPtr->strides.cStride;
                     }