From 8c3a0153b9dcdaf066de06a812028622f91dfa7e Mon Sep 17 00:00:00 2001
From: Xavier Arteaga <xavier@softwareradiosystems.com>
Date: Wed, 21 Nov 2018 18:06:14 +0100
Subject: [PATCH] Added missing AVX512 intrinsics and flags. Fixes #291.

---
 CMakeLists.txt                       |  4 ++--
 cmake/modules/FindSSE.cmake          |  2 +-
 lib/include/srslte/phy/utils/simd.h  | 16 ++++++++++++++--
 lib/src/phy/utils/test/vector_test.c | 26 ++++++++++++++++++++++++++
 4 files changed, 43 insertions(+), 5 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index ccc1ce54d..1d0f05568 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -306,8 +306,8 @@ if(CMAKE_C_COMPILER_ID MATCHES "GNU" OR CMAKE_C_COMPILER_ID MATCHES "Clang")
   endif (HAVE_FMA)
 
   if (HAVE_AVX512)
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512f -mavx512cd -DLV_HAVE_AVX512")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -mavx512cd -DLV_HAVE_AVX512")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx512f -mavx512cd -mavx512bw -mavx512dq -DLV_HAVE_AVX512")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -mavx512cd -mavx512bw -mavx512dq -DLV_HAVE_AVX512")
   endif(HAVE_AVX512)
 
   if(NOT ${CMAKE_BUILD_TYPE} STREQUAL "Debug")
diff --git a/cmake/modules/FindSSE.cmake b/cmake/modules/FindSSE.cmake
index e5101deff..24fc23662 100644
--- a/cmake/modules/FindSSE.cmake
+++ b/cmake/modules/FindSSE.cmake
@@ -142,7 +142,7 @@ if (ENABLE_SSE)
         # Check compiler for AVX intrinsics
         #
         if (CMAKE_COMPILER_IS_GNUCC OR (CMAKE_C_COMPILER_ID MATCHES "Clang") OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang"))
-            set(CMAKE_REQUIRED_FLAGS "-mavx512f")
+            set(CMAKE_REQUIRED_FLAGS "-mavx512f -mavx512cd -mavx512bw -mavx512dq -DLV_HAVE_AVX512")
             check_c_source_runs("
           #include <immintrin.h>
           int main()
diff --git a/lib/include/srslte/phy/utils/simd.h b/lib/include/srslte/phy/utils/simd.h
index 491c1f661..2529033c5 100644
--- a/lib/include/srslte/phy/utils/simd.h
+++ b/lib/include/srslte/phy/utils/simd.h
@@ -1341,7 +1341,13 @@ static inline simd_s_t srslte_simd_s_mul(simd_s_t a, simd_s_t b) {
 
 static inline simd_s_t srslte_simd_s_neg(simd_s_t a, simd_s_t b) {
 #ifdef LV_HAVE_AVX512
-#error sign instruction not available in avx512
+  __m256i a0 = _mm512_extracti64x4_epi64(a, 0);
+  __m256i a1 = _mm512_extracti64x4_epi64(a, 1);
+  __m256i b0 = _mm512_extracti64x4_epi64(b, 0);
+  __m256i b1 = _mm512_extracti64x4_epi64(b, 1);
+  __m256i r0 = _mm256_sign_epi16(a0, b0);
+  __m256i r1 = _mm256_sign_epi16(a1, b1);
+  return _mm512_inserti64x4(_mm512_broadcast_i64x4(r0), r1, 1);
 #else /* LV_HAVE_AVX512 */
 #ifdef LV_HAVE_AVX2
   return _mm256_sign_epi16(a, b);
@@ -1814,7 +1820,13 @@ static inline simd_s_t srslte_simd_b_sub(simd_s_t a, simd_s_t b) {
 
 static inline simd_s_t srslte_simd_b_neg(simd_b_t a, simd_b_t b) {
 #ifdef LV_HAVE_AVX512
-#error sign instruction not available in avx512
+  __m256i a0 = _mm512_extracti64x4_epi64(a, 0);
+  __m256i a1 = _mm512_extracti64x4_epi64(a, 1);
+  __m256i b0 = _mm512_extracti64x4_epi64(b, 0);
+  __m256i b1 = _mm512_extracti64x4_epi64(b, 1);
+  __m256i r0 = _mm256_sign_epi8(a0, b0);
+  __m256i r1 = _mm256_sign_epi8(a1, b1);
+  return _mm512_inserti64x4(_mm512_broadcast_i64x4(r0), r1, 1);
 #else /* LV_HAVE_AVX512 */
 #ifdef LV_HAVE_AVX2
   return _mm256_sign_epi8(a, b);
diff --git a/lib/src/phy/utils/test/vector_test.c b/lib/src/phy/utils/test/vector_test.c
index 44f8af1ca..03e33305b 100644
--- a/lib/src/phy/utils/test/vector_test.c
+++ b/lib/src/phy/utils/test/vector_test.c
@@ -226,6 +226,29 @@ TEST(srslte_vec_prod_sss,
          free(z);
 )
 
+TEST(srslte_vec_neg_sss,
+     MALLOC(int16_t, x);
+         MALLOC(int16_t, y);
+         MALLOC(int16_t, z);
+
+         int16_t gold = 0.0f;
+         for (int i = 0; i < block_size; i++) {
+           x[i] = RANDOM_S();
+           do { y[i] = RANDOM_S(); } while (!y[i]);
+         }
+
+         TEST_CALL(srslte_vec_neg_sss(x, y, z, block_size))
+
+         for (int i = 0; i < block_size; i++) {
+           gold = y[i] < 0 ? -x[i] : x[i];
+           mse += abs(gold - z[i]);
+         }
+
+         free(x);
+         free(y);
+         free(z);
+)
+
 TEST(srslte_vec_acc_cc,
      MALLOC(cf_t, x);
          cf_t z;
@@ -868,6 +891,9 @@ int main(int argc, char **argv) {
     passed[func_count][size_count] = test_srslte_vec_prod_sss(func_names[func_count], &timmings[func_count][size_count], block_size);
     func_count++;
 
+    passed[func_count][size_count] = test_srslte_vec_neg_sss(func_names[func_count], &timmings[func_count][size_count], block_size);
+    func_count++;
+
     passed[func_count][size_count] = test_srslte_vec_acc_cc(func_names[func_count], &timmings[func_count][size_count], block_size);
     func_count++;