Polyfill vcvtnq_s32_f32 for AArch32 GCC

PiperOrigin-RevId: 380031679
google · Jun 17, 2021 · 16d79ed · 16d79ed
1 parent 0ae35f2
commit 16d79ed
Show file tree

Hide file tree

Showing 19 changed files with 32 additions and 0 deletions.
diff --git a/src/qs8-gemm/MRxNRc4-neondot.c.in b/src/qs8-gemm/MRxNRc4-neondot.c.in
@@ -12,6 +12,8 @@ $assert REQUANTIZATION in ["GEMMLOWP", "FP32"]
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+$if REQUANTIZATION == "FP32":
+  #include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/1x16c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/1x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/1x8c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/1x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/4x16c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/4x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/4x8c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/4x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/6x16c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/6x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/6x8c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/6x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/8x16c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/8x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-gemm/gen/8x8c4-minmax-fp32-neondot.c b/src/qs8-gemm/gen/8x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/gemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/MRxNRc4-neondot.c.in b/src/qs8-igemm/MRxNRc4-neondot.c.in
@@ -12,6 +12,8 @@ $assert REQUANTIZATION in ["GEMMLOWP", "FP32"]
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+$if REQUANTIZATION == "FP32":
+  #include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/1x16c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/1x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/1x8c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/1x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/4x16c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/4x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/4x8c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/4x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/6x16c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/6x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/6x8c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/6x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/8x16c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/8x16c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/qs8-igemm/gen/8x8c4-minmax-fp32-neondot.c b/src/qs8-igemm/gen/8x8c4-minmax-fp32-neondot.c
@@ -12,6 +12,7 @@
 #include <arm_neon.h>
 
 #include <xnnpack/igemm.h>
+#include <xnnpack/intrinsics-polyfill.h>
 #include <xnnpack/math.h>
 
 

diff --git a/src/xnnpack/intrinsics-polyfill.h b/src/xnnpack/intrinsics-polyfill.h
@@ -128,4 +128,16 @@ __m512i _mm512_set_epi8(
 }
 #endif  // GCC pre-9
 
+// AArch32 GCC, see
+// - https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71233
+// - https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95399
+#if XNN_ARCH_ARM && defined(__GNUC__) && !defined(__clang__) && defined(__ARM_NEON__)
+#include <arm_neon.h>
+
+static XNN_INTRINSIC
+int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+  return vcvtq_s32_f32(vrndnq_f32(v));
+}
+#endif  // AArch32 GCC
+
 #endif  // __AVX512F__