Skip to content

Commit

Permalink
Merge pull request ROCm#52 from sampath1117/sr/opt_brightness_temp
Browse files Browse the repository at this point in the history 
[Updated] AVX optimizations for Brightness kernel - HOST
  • Loading branch information
@r-abishek
r-abishek authored Feb 18, 2022
2 parents 9927957 + 32be913 commit 9902690
Show file tree
Hide file tree
Showing 2 changed files with 245 additions and 324 deletions.
121 changes: 0 additions & 121 deletions src/include/cpu/rpp_cpu_simd.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -848,39 +848,6 @@ inline RppStatus rpp_store16_f32_to_i8_avx(Rpp8s *dstPtr, __m256 *p)
return RPP_SUCCESS;
}

inline RppStatus rpp_load16_u8_to_f32_avx(Rpp8u *srcPtr, __m256 *p)
{
__m128i px[2];
__m128i pxZero = _mm_setzero_si128();

px[0] = _mm_loadu_si128((__m128i *)srcPtr); /* load pixels 0-15 */
px[1] = _mm_unpackhi_epi8(px[0], pxZero); /* pixels 8-15 */
px[0] = _mm_unpacklo_epi8(px[0], pxZero); /* pixels 0-7 */
p[0] = _mm256_cvtepi32_ps(_mm256_setr_m128i(_mm_unpacklo_epi16(px[0], pxZero),_mm_unpackhi_epi16(px[0], pxZero)));/* pixels 0-7 */
p[1] = _mm256_cvtepi32_ps(_mm256_setr_m128i(_mm_unpacklo_epi16(px[1], pxZero),_mm_unpackhi_epi16(px[1], pxZero)));/* pixels 8-25 */

return RPP_SUCCESS;
}

inline RppStatus rpp_store16_f32_to_u8_avx(Rpp8u *dstPtr, __m256 *p)
{
__m256i pxCvt[2];
__m128i px[4];

pxCvt[0] = _mm256_cvtps_epi32(p[0]); /* pixels 0-7 */
pxCvt[1] = _mm256_cvtps_epi32(p[1]); /* pixels 8-15 */
px[0] = _mm256_extracti128_si256(pxCvt[0],0); /* pixels 0-3 */
px[1] = _mm256_extracti128_si256(pxCvt[0],1); /* pixels 4-7 */
px[2] = _mm256_extracti128_si256(pxCvt[1],0); /* pixels 8-11 */
px[3] = _mm256_extracti128_si256(pxCvt[1],1); /* pixels 12-15 */
px[0] = _mm_packus_epi32(px[0], px[1]); /* pixels 0-7 */
px[1] = _mm_packus_epi32(px[2], px[3]); /* pixels 8-15 */
px[2] = _mm_packus_epi16(px[0], px[1]); /* pixels 8-15 */
_mm_storeu_si128((__m128i *)dstPtr, px[2]); /* store pixels 0-15 */

return RPP_SUCCESS;
}

inline RppStatus rpp_load24_f32pkd3_to_f32pln3_avx(Rpp32f *srcPtr, __m256 *p)
{
__m128 p128[8];
Expand Down Expand Up @@ -942,20 +909,6 @@ inline RppStatus rpp_store24_f32pln3_to_f32pkd3_avx(Rpp32f *dstPtr, __m256 *p)
return RPP_SUCCESS;
}

inline RppStatus rpp_load8_f32_to_f32_avx(Rpp32f *srcPtr, __m256 *p)
{
p[0] = _mm256_loadu_ps(srcPtr);

return RPP_SUCCESS;
}

inline RppStatus rpp_store8_f32_to_f32_avx(Rpp32f *dstPtr, __m256 *p)
{
_mm256_storeu_ps(dstPtr, p[0]);

return RPP_SUCCESS;
}

inline RppStatus rpp_load16_f32_to_f32_avx(Rpp32f *srcPtr, __m256 *p)
{
p[0] = _mm256_loadu_ps(srcPtr);
Expand Down Expand Up @@ -1092,80 +1045,6 @@ inline RppStatus rpp_store48_f32pln3_to_i8pkd3_avx(Rpp8s *dstPtr, __m256 *p)
return RPP_SUCCESS;
}

inline RppStatus rpp_load16_i8_to_f32_avx(Rpp8s *srcPtr, __m256 *p)
{
__m128i px[2];
__m128i pxZero = _mm_setzero_si128();
__m128i pxConvertI8 = _mm_set1_epi8((char)128);

px[0] = _mm_loadu_si128((__m128i *)srcPtr); /* load pixels 0-15 */
px[0] = _mm_add_epi8(px[0], pxConvertI8); /* convert to u8 for px0 compute */
px[1] = _mm_unpackhi_epi8(px[0], pxZero); /* pixels 8-15 */
px[0] = _mm_unpacklo_epi8(px[0], pxZero); /* pixels 0-7 */
p[0] = _mm256_cvtepi32_ps(_mm256_setr_m128i(_mm_unpacklo_epi16(px[0], pxZero),_mm_unpackhi_epi16(px[0], pxZero))); /* pixels 0-7 */
p[1] = _mm256_cvtepi32_ps(_mm256_setr_m128i(_mm_unpacklo_epi16(px[1], pxZero),_mm_unpackhi_epi16(px[1], pxZero))); /* pixels 8-15 */

return RPP_SUCCESS;
}

inline RppStatus rpp_store16_f32_to_i8_avx(Rpp8s *dstPtr, __m256 *p)
{
__m256i pxCvt[2];
__m128i px[4];
__m128i pxConvertI8 = _mm_set1_epi8((char)128);

pxCvt[0] = _mm256_cvtps_epi32(p[0]); /* pixels 0-7 */
pxCvt[1] = _mm256_cvtps_epi32(p[1]); /* pixels 8-15 */
px[0] = _mm256_extracti128_si256(pxCvt[0],0); /* pixels 0-3 */
px[1] = _mm256_extracti128_si256(pxCvt[0],1); /* pixels 4-7 */
px[2] = _mm256_extracti128_si256(pxCvt[1],0); /* pixels 8-11 */
px[3] = _mm256_extracti128_si256(pxCvt[1],1); /* pixels 12-15 */
px[0] = _mm_packus_epi32(px[0], px[1]); /* pixels 0-7 */
px[1] = _mm_packus_epi32(px[2], px[3]); /* pixels 8-15 */
px[2] = _mm_packus_epi16(px[0], px[1]); /* pixels 0-15 */
px[2] = _mm_sub_epi8(px[2], pxConvertI8); /* convert back to i8 for px0 store */
_mm_storeu_si128((__m128i *)dstPtr, px[2]); /* store pixels 0-15 */

return RPP_SUCCESS;
}

inline RppStatus rpp_load16_i8_to_f32_avx(Rpp8s *srcPtr, __m256 *p)
{
__m128i px[2];
__m128i pxZero = _mm_setzero_si128();
__m128i pxConvertI8 = _mm_set1_epi8((char)128);

px[0] = _mm_loadu_si128((__m128i *)srcPtr); /* load pixels 0-15 */
px[0] = _mm_add_epi8(px[0], pxConvertI8); /* convert to u8 for px0 compute */
px[1] = _mm_unpackhi_epi8(px[0], pxZero); /* pixels 8-15 */
px[0] = _mm_unpacklo_epi8(px[0], pxZero); /* pixels 0-7 */
p[0] = _mm256_cvtepi32_ps(_mm256_setr_m128i(_mm_unpacklo_epi16(px[0], pxZero),_mm_unpackhi_epi16(px[0], pxZero))); /* pixels 0-7 */
p[1] = _mm256_cvtepi32_ps(_mm256_setr_m128i(_mm_unpacklo_epi16(px[1], pxZero),_mm_unpackhi_epi16(px[1], pxZero))); /* pixels 8-15 */

return RPP_SUCCESS;
}

inline RppStatus rpp_store16_f32_to_i8_avx(Rpp8s *dstPtr, __m256 *p)
{
__m256i pxCvt[2];
__m128i px[4];
__m128i pxConvertI8 = _mm_set1_epi8((char)128);

pxCvt[0] = _mm256_cvtps_epi32(p[0]); /* pixels 0-7 */
pxCvt[1] = _mm256_cvtps_epi32(p[1]); /* pixels 8-15 */
px[0] = _mm256_extracti128_si256(pxCvt[0],0); /* pixels 0-3 */
px[1] = _mm256_extracti128_si256(pxCvt[0],1); /* pixels 4-7 */
px[2] = _mm256_extracti128_si256(pxCvt[1],0); /* pixels 8-11 */
px[3] = _mm256_extracti128_si256(pxCvt[1],1); /* pixels 12-15 */
px[0] = _mm_packus_epi32(px[0], px[1]); /* pixels 0-7 */
px[1] = _mm_packus_epi32(px[2], px[3]); /* pixels 8-15 */
px[2] = _mm_packus_epi16(px[0], px[1]); /* pixels 0-15 */
px[2] = _mm_sub_epi8(px[2], pxConvertI8); /* convert back to i8 for px0 store */
_mm_storeu_si128((__m128i *)dstPtr, px[2]); /* store pixels 0-15 */

return RPP_SUCCESS;
}

inline RppStatus rpp_normalize48_avx(__m256 *p)
{
p[0] = _mm256_mul_ps(p[0], avx_p1op255);
Expand Down
Loading

0 comments on commit 9902690

Please sign in to comment.