Add some small arm neon optimizations (#1847)

* Remove unused includes Signed-off-by: Mark Reid <mindmark@gmail.com> * Use neon hardware support for f16 conversions Signed-off-by: Mark Reid <mindmark@gmail.com> * Add some small neon optimizations use blendv,floor and fma intrinsics were possible Signed-off-by: Mark Reid <mindmark@gmail.com> --------- Signed-off-by: Mark Reid <mindmark@gmail.com>
AcademySoftwareFoundation · Sep 14, 2023 · 14f0afa · 14f0afa
1 parent 16b3157
commit 14f0afa
Show file tree

Hide file tree

Showing 6 changed files with 61 additions and 7 deletions.
diff --git a/src/OpenColorIO/AVX.h b/src/OpenColorIO/AVX.h
@@ -9,7 +9,6 @@
 #if OCIO_USE_AVX
 
 #include <immintrin.h>
-#include <stdio.h>
 
 #include <OpenColorIO/OpenColorIO.h>
 #include "BitDepthUtils.h"

diff --git a/src/OpenColorIO/AVX2.h b/src/OpenColorIO/AVX2.h
@@ -9,7 +9,6 @@
 #if OCIO_USE_AVX2
 
 #include <immintrin.h>
-#include <stdio.h>
 
 #include <OpenColorIO/OpenColorIO.h>
 #include "BitDepthUtils.h"

diff --git a/src/OpenColorIO/CPUInfoConfig.h.in b/src/OpenColorIO/CPUInfoConfig.h.in
@@ -8,6 +8,8 @@
 // Relevant only for arm64 architecture.
 #if defined(__aarch64__)
     #cmakedefine01 OCIO_USE_SSE2NEON
+#else
+    #define OCIO_USE_SSE2NEON 0
 #endif
 
 // On the Apple platform, a universal build is created for both x86_64 and arm64 architectures. 

diff --git a/src/OpenColorIO/SSE2.h b/src/OpenColorIO/SSE2.h
@@ -18,8 +18,6 @@
     #endif
 #endif
 
-#include <stdio.h>
-
 #include <OpenColorIO/OpenColorIO.h>
 #include "BitDepthUtils.h"
 
@@ -76,6 +74,8 @@ static inline void sse2RGBATranspose_4x4(__m128 row0, __m128 row1, __m128 row2,
     out_a = _mm_movehl_ps(tmp3, tmp1);
 }
 
+#if !OCIO_USE_SSE2NEON
+
 static inline __m128i sse2_blendv(__m128i a, __m128i b, __m128i mask)
 {
     return _mm_xor_si128(_mm_and_si128(_mm_xor_si128(a, b), mask), a);
@@ -164,6 +164,8 @@ static inline __m128 sse2_cvtph_ps(__m128i a)
     return  _mm_or_ps(o, sign);
 }
 
+#endif
+
 // Note Packing functions perform no 0.0 - 1.0 normalization
 // but perform 0 - max value clamping for integer formats
 template<BitDepth BD> struct SSE2RGBAPack {};
@@ -290,21 +292,48 @@ struct SSE2RGBAPack<BIT_DEPTH_F16>
         __m128i rgba_00_01 = _mm_loadu_si128((const __m128i*)(in + 0));
         __m128i rgba_02_03 = _mm_loadu_si128((const __m128i*)(in + 8));
 
+#if OCIO_USE_SSE2NEON
+        // use neon hardware support for f16 to f32
+        __m128 rgba0 = vreinterpretq_m128_f32(
+            vcvt_f32_f16(vget_low_f16(vreinterpretq_f16_s64(vreinterpretq_s64_m128i(rgba_00_01))))
+        );
+        __m128 rgba1 = vreinterpretq_m128_f32(
+            vcvt_f32_f16(vget_high_f16(vreinterpretq_f16_s64(vreinterpretq_s64_m128i(rgba_00_01))))
+        );
+        __m128 rgba2 = vreinterpretq_m128_f32(
+            vcvt_f32_f16(vget_low_f16(vreinterpretq_f16_s64(vreinterpretq_s64_m128i(rgba_02_03))))
+        );
+        __m128 rgba3 = vreinterpretq_m128_f32(
+            vcvt_f32_f16(vget_high_f16(vreinterpretq_f16_s64(vreinterpretq_s64_m128i(rgba_02_03))))
+        );
+#else
         __m128 rgba0 = sse2_cvtph_ps(rgba_00_01);
         __m128 rgba1 = sse2_cvtph_ps(_mm_shuffle_epi32(rgba_00_01, _MM_SHUFFLE(1,0,3,2)));
         __m128 rgba2 = sse2_cvtph_ps(rgba_02_03);
         __m128 rgba3 = sse2_cvtph_ps(_mm_shuffle_epi32(rgba_02_03, _MM_SHUFFLE(1,0,3,2)));
-
+#endif
         sse2RGBATranspose_4x4(rgba0, rgba1, rgba2, rgba3, r, g, b, a);
     }
 
     static inline  void Store(half *out, __m128 r, __m128 g, __m128 b, __m128 a)
     {
         __m128 rgba0, rgba1, rgba2, rgba3;
-        __m128i rgba;
-
         sse2RGBATranspose_4x4(r, g, b, a, rgba0, rgba1, rgba2, rgba3);
 
+#if OCIO_USE_SSE2NEON
+        // use neon hardware support for f32 to f16
+        float16x8_t rgba;
+        float16x4_t rgba00_01 = vcvt_f16_f32(vreinterpretq_f32_m128(rgba0));
+        float16x4_t rgba03_03 = vcvt_f16_f32(vreinterpretq_f32_m128(rgba1));
+        float16x4_t rgba04_05 = vcvt_f16_f32(vreinterpretq_f32_m128(rgba2));
+        float16x4_t rgba06_07 = vcvt_f16_f32(vreinterpretq_f32_m128(rgba3));
+        rgba = vcombine_f16(rgba00_01, rgba03_03);
+        vst1q_f16((float16_t *)(out+0), rgba);
+
+        rgba = vcombine_f16(rgba04_05, rgba06_07);
+        vst1q_f16((float16_t *)(out+8), rgba);
+#else
+        __m128i rgba;
         __m128i rgba00_01 = sse2_cvtps_ph(rgba0);
         __m128i rgba02_03 = sse2_cvtps_ph(rgba1);
         __m128i rgba04_05 = sse2_cvtps_ph(rgba2);
@@ -315,6 +344,7 @@ struct SSE2RGBAPack<BIT_DEPTH_F16>
 
         rgba = _mm_xor_si128(rgba04_05, _mm_shuffle_epi32(rgba06_07, _MM_SHUFFLE(1,0,3,2)));
         _mm_storeu_si128((__m128i*)(out+8), rgba);
+#endif
     }
 };
 

diff --git a/src/OpenColorIO/ops/lut1d/Lut1DOpCPU_SSE2.cpp b/src/OpenColorIO/ops/lut1d/Lut1DOpCPU_SSE2.cpp
@@ -24,13 +24,23 @@ namespace {
 
 static inline __m128 fmadd_ps_sse2(__m128 a, __m128 b, __m128 c)
 {
+#if OCIO_USE_SSE2NEON
+    return vreinterpretq_m128_f32(
+        vfmaq_f32(vreinterpretq_f32_m128(c), vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))
+    );
+#else
     return  _mm_add_ps(_mm_mul_ps(a, b), c);
+#endif
 }
 
 static inline __m128 floor_ps_sse2(__m128 v)
 {
+#if OCIO_USE_SSE2NEON
+    return _mm_floor_ps(v);
+#else
     // NOTE: using truncate cvtt
     return _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
+#endif
 }
 
 

diff --git a/src/OpenColorIO/ops/lut3d/Lut3DOpCPU_SSE2.cpp b/src/OpenColorIO/ops/lut3d/Lut3DOpCPU_SSE2.cpp
@@ -38,18 +38,32 @@ struct rgbavec_sse2 {
 
 static inline __m128 floor_ps_sse2(__m128 v)
 {
+#if OCIO_USE_SSE2NEON
+    return _mm_floor_ps(v);
+#else
     // NOTE: using truncate cvtt
     return _mm_cvtepi32_ps(_mm_cvttps_epi32(v));
+#endif
 }
 
 static inline __m128 blendv_ps_sse2(__m128 a, __m128 b, __m128 mask)
 {
+#if OCIO_USE_SSE2NEON
+    return _mm_blendv_ps(a, b, mask);
+#else
     return _mm_xor_ps(_mm_and_ps(_mm_xor_ps(a, b), mask), a);
+#endif
 }
 
 static inline __m128 fmadd_ps_sse2(__m128 a, __m128 b, __m128 c)
 {
+#if OCIO_USE_SSE2NEON
+    return vreinterpretq_m128_f32(
+        vfmaq_f32(vreinterpretq_f32_m128(c), vreinterpretq_f32_m128(a), vreinterpretq_f32_m128(b))
+    );
+#else
     return  _mm_add_ps(_mm_mul_ps(a, b), c);
+#endif
 }
 
 static inline rgbavec_sse2 interp_tetrahedral_sse2(const Lut3DContextSSE2 &ctx, __m128 r, __m128 g, __m128 b, __m128 a)