1 files changed, 53 insertions, 0 deletions
diff --git a/host/lib/convert/sse2_fc32_to_sc16.cpp b/host/lib/convert/sse2_fc32_to_sc16.cpp
index 2d1f853b9..d02868d80 100644
--- a/host/lib/convert/sse2_fc32_to_sc16.cpp
+++ b/host/lib/convert/sse2_fc32_to_sc16.cpp
@@ -120,3 +120,56 @@ DECLARE_CONVERTER(fc32, 1, sc16_item32_be, 1, PRIORITY_SIMD)
     // convert any remaining samples
     xx_to_item32_sc16<uhd::htonx>(input + i, output + i, nsamps - i, scale_factor);
 }
+
+DECLARE_CONVERTER(fc32, 1, sc16_chdr, 1, PRIORITY_SIMD)
+{
+    const fc32_t* input = reinterpret_cast<const fc32_t*>(inputs[0]);
+    sc16_t* output      = reinterpret_cast<sc16_t*>(outputs[0]);
+
+    const __m128 scalar = _mm_set_ps1(float(scale_factor));
+
+// this macro converts values faster by using SSE intrinsics to convert 4 values at a time
+#define convert_fc32_1_to_item32_1_guts(_al_)                                  \
+    for (; i + 3 < nsamps; i += 4) {                                           \
+        /* load from input */                                                  \
+        __m128 tmplo =                                                         \
+            _mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 0)); \
+        __m128 tmphi =                                                         \
+            _mm_load##_al_##ps(reinterpret_cast<const float*>(input + i + 2)); \
+                                                                               \
+        /* convert and scale */                                                \
+        __m128i tmpilo = _mm_cvtps_epi32(_mm_mul_ps(tmplo, scalar));           \
+        __m128i tmpihi = _mm_cvtps_epi32(_mm_mul_ps(tmphi, scalar));           \
+                                                                               \
+        /* pack from 32 bit integers to 16 bit */                              \
+        __m128i tmpi = _mm_packs_epi32(tmpilo, tmpihi);                        \
+                                                                               \
+        /* store to output */                                                  \
+        _mm_storeu_si128(reinterpret_cast<__m128i*>(output + i), tmpi);        \
+    }
+
+    size_t i = 0;
+
+    // need to dispatch according to alignment for fastest conversion
+    switch (size_t(input) & 0xf) {
+        case 0x0:
+            // the data is 16-byte aligned, so do the fast processing of the bulk of the
+            // samples
+            convert_fc32_1_to_item32_1_guts(_) break;
+        case 0x8:
+            // the first sample is 8-byte aligned - process it to align the remainder of
+            // the samples to 16-bytes
+            xx_to_chdr_sc16(input, output, 1, scale_factor);
+            i++;
+            // do faster processing of the bulk of the samples now that we are 16-byte
+            // aligned
+            convert_fc32_1_to_item32_1_guts(_) break;
+        default:
+            // we are not 8 or 16-byte aligned, so do fast processing with the unaligned
+            // load
+            convert_fc32_1_to_item32_1_guts(u_)
+    }
+
+    // convert any remaining samples
+    xx_to_chdr_sc16(input + i, output + i, nsamps - i, scale_factor);
+}