svml: SSE is good enough for native m128i and m128d types & functions

simd-everywhere · Apr 3, 2024 · 9982b27 · 9982b27
1 parent 8d95f83
commit 9982b27
Show file tree

Hide file tree

Showing 3 changed files with 24 additions and 24 deletions.
diff --git a/simde/arm/neon/types.h b/simde/arm/neon/types.h
@@ -432,7 +432,7 @@ typedef union {
 typedef union {
   SIMDE_ARM_NEON_DECLARE_VECTOR(simde_float64, values, 16);
 
-  #if defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SSE2_NATIVE) || defined(SIMDE_X86_SVML_NATIVE)
     __m128d m128d;
   #endif
 
@@ -697,7 +697,7 @@ typedef union {
     #define SIMDE_ARM_NEON_NEED_PORTABLE_F32X4
   #endif
 
-  #if defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SSE2_NATIVE) || defined(SIMDE_X86_SVML_NATIVE)
     typedef  __m128i  simde_int8x16_t;
     typedef  __m128i  simde_int16x8_t;
     typedef  __m128i  simde_int32x4_t;
@@ -1438,7 +1438,7 @@ typedef union {
   SIMDE_DEFINE_CONVERSION_FUNCTION_(simde_float32x4_to_m128,              __m128, simde_float32x4_t)
   SIMDE_DEFINE_CONVERSION_FUNCTION_(simde_float32x4_from_m128, simde_float32x4_t,            __m128)
 #endif
-#if defined(SIMDE_X86_SSE2_NATIVE)
+#if defined(SIMDE_X86_SSE2_NATIVE) || defined(SIMDE_X86_SVML_NATIVE)
   SIMDE_DEFINE_CONVERSION_FUNCTION_(simde_int8x16_to_m128i,               __m128i,   simde_int8x16_t)
   SIMDE_DEFINE_CONVERSION_FUNCTION_(simde_int16x8_to_m128i,               __m128i,   simde_int16x8_t)
   SIMDE_DEFINE_CONVERSION_FUNCTION_(simde_int32x4_to_m128i,               __m128i,   simde_int32x4_t)

diff --git a/simde/x86/sse2.h b/simde/x86/sse2.h
@@ -174,7 +174,7 @@ typedef union {
     SIMDE_ALIGN_TO_16 simde__m64_private m64_private[2];
     SIMDE_ALIGN_TO_16 simde__m64         m64[2];
 
-  #if defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SSE2_NATIVE) || defined(SIMDE_X86_SVML_NATIVE)
     SIMDE_ALIGN_TO_16 __m128d        n;
   #elif defined(SIMDE_ARM_NEON_A32V7_NATIVE)
     SIMDE_ALIGN_TO_16 int8x16_t      neon_i8;
@@ -226,7 +226,7 @@ typedef union {
   #endif
 } simde__m128d_private;
 
-#if defined(SIMDE_X86_SSE2_NATIVE)
+#if defined(SIMDE_X86_SSE2_NATIVE) || defined(SIMDE_X86_SVML_NATIVE)
   typedef __m128i simde__m128i;
   typedef __m128d simde__m128d;
 #elif defined(SIMDE_ARM_NEON_A32V7_NATIVE)
@@ -256,7 +256,7 @@ typedef union {
   typedef simde__m128d_private simde__m128d;
 #endif
 
-#if defined(SIMDE_X86_SSE2_ENABLE_NATIVE_ALIASES)
+#if defined(SIMDE_X86_SSE2_ENABLE_NATIVE_ALIASES) || defined(SIMDE_X86_SVML_ENABLE_NATIVE_ALIASES)
   typedef simde__m128i __m128i;
   typedef simde__m128d __m128d;
 #endif

diff --git a/simde/x86/svml.h b/simde/x86/svml.h
@@ -2675,7 +2675,7 @@ simde_mm512_mask_cosh_pd(simde__m512d src, simde__mmask8 k, simde__m512d a) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epi8 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epi8(a, b);
   #else
     simde__m128i_private
@@ -2705,7 +2705,7 @@ simde_mm_div_epi8 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epi16 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epi16(a, b);
   #else
     simde__m128i_private
@@ -2735,7 +2735,7 @@ simde_mm_div_epi16 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epi32 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epi32(a, b);
   #else
     simde__m128i_private
@@ -2768,7 +2768,7 @@ simde_mm_div_epi32 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epi64 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epi64(a, b);
   #else
     simde__m128i_private
@@ -2798,7 +2798,7 @@ simde_mm_div_epi64 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epu8 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epu8(a, b);
   #else
     simde__m128i_private
@@ -2828,7 +2828,7 @@ simde_mm_div_epu8 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epu16 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epu16(a, b);
   #else
     simde__m128i_private
@@ -2858,7 +2858,7 @@ simde_mm_div_epu16 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epu32 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epu32(a, b);
   #else
     simde__m128i_private
@@ -2891,7 +2891,7 @@ simde_mm_div_epu32 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_div_epu64 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_div_epu64(a, b);
   #else
     simde__m128i_private
@@ -5093,7 +5093,7 @@ simde_mm512_mask_cdfnorm_pd(simde__m512d src, simde__mmask8 k, simde__m512d a) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_idivrem_epi32 (simde__m128i* mem_addr, simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_idivrem_epi32(HEDLEY_REINTERPRET_CAST(__m128i*, mem_addr), a, b);
   #else
     simde__m128i r;
@@ -8902,7 +8902,7 @@ simde_mm256_csqrt_ps (simde__m256 a) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epi8 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epi8(a, b);
   #else
     simde__m128i_private
@@ -8930,7 +8930,7 @@ simde_mm_rem_epi8 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epi16 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epi16(a, b);
   #else
     simde__m128i_private
@@ -8958,7 +8958,7 @@ simde_mm_rem_epi16 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epi32 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epi32(a, b);
   #else
     simde__m128i_private
@@ -8989,7 +8989,7 @@ simde_mm_rem_epi32 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epi64 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epi64(a, b);
   #else
     simde__m128i_private
@@ -9017,7 +9017,7 @@ simde_mm_rem_epi64 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epu8 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epu8(a, b);
   #else
     simde__m128i_private
@@ -9045,7 +9045,7 @@ simde_mm_rem_epu8 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epu16 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epu16(a, b);
   #else
     simde__m128i_private
@@ -9073,7 +9073,7 @@ simde_mm_rem_epu16 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epu32 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epu32(a, b);
   #else
     simde__m128i_private
@@ -9104,7 +9104,7 @@ simde_mm_rem_epu32 (simde__m128i a, simde__m128i b) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_rem_epu64 (simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_rem_epu64(a, b);
   #else
     simde__m128i_private
@@ -12087,7 +12087,7 @@ simde_mm512_mask_trunc_pd(simde__m512d src, simde__mmask8 k, simde__m512d a) {
 SIMDE_FUNCTION_ATTRIBUTES
 simde__m128i
 simde_mm_udivrem_epi32 (simde__m128i * mem_addr, simde__m128i a, simde__m128i b) {
-  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE2_NATIVE)
+  #if defined(SIMDE_X86_SVML_NATIVE) && defined(SIMDE_X86_SSE_NATIVE)
     return _mm_udivrem_epi32(mem_addr, a, b);
   #else
     simde__m128i r;