Add AVX2 and SSE2 fastpaths to the DeckLink PCI driver.

diff --git a/decklink_capture.cpp b/decklink_capture.cpp
index c1b228dd43ead8afbe9d52cef3d3a1721489075f..2bc5fc9d77cdab7c5e83a4d5e9db7d2af379cb44 100644 (file)
--- a/decklink_capture.cpp
+++ b/decklink_capture.cpp
@@ -6,6 +6,9 @@
 #include <stdlib.h>
 #include <string.h>
 #include <cstddef>
+#ifdef __SSE2__
+#include <immintrin.h>
+#endif
 
 #include <DeckLinkAPI.h>
 #include <DeckLinkAPIConfiguration.h>
@@ -21,7 +24,6 @@ using namespace std::placeholders;
 namespace {
 
 // TODO: Support stride.
-// TODO: Support AVX2 (adapt from bmusb).
 void memcpy_interleaved(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
 {
        assert(n % 2 == 0);
@@ -34,6 +36,101 @@ void memcpy_interleaved(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size
        }
 }
 
+#ifdef __SSE2__
+
+// Returns the number of bytes consumed.
+size_t memcpy_interleaved_fastpath(uint8_t *dest1, uint8_t *dest2, const uint8_t *src, size_t n)
+{
+       const uint8_t *limit = src + n;
+       size_t consumed = 0;
+
+       // Align end to 32 bytes.
+       limit = (const uint8_t *)(intptr_t(limit) & ~31);
+
+       if (src >= limit) {
+               return 0;
+       }
+
+       // Process [0,31] bytes, such that start gets aligned to 32 bytes.
+       const uint8_t *aligned_src = (const uint8_t *)(intptr_t(src + 31) & ~31);
+       if (aligned_src != src) {
+               size_t n2 = aligned_src - src;
+               memcpy_interleaved(dest1, dest2, src, n2);
+               dest1 += n2 / 2;
+               dest2 += n2 / 2;
+               if (n2 % 1) {
+                       swap(dest1, dest2);
+               }
+               src = aligned_src;
+               consumed += n2;
+       }
+
+       // Make the length a multiple of 64.
+       if (((limit - src) % 64) != 0) {
+               limit -= 32;
+       }
+       assert(((limit - src) % 64) == 0);
+
+#if __AVX2__
+       const __restrict __m256i *in = (const __m256i *)src;
+       __restrict __m256i *out1 = (__m256i *)dest1;
+       __restrict __m256i *out2 = (__m256i *)dest2;
+
+       __m256i shuffle_cw = _mm256_set_epi8(
+               15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0,
+               15, 13, 11, 9, 7, 5, 3, 1, 14, 12, 10, 8, 6, 4, 2, 0);
+       while (in < (const __m256i *)limit) {
+               // Note: For brevity, comments show lanes as if they were 2x64-bit (they're actually 2x128).
+               __m256i data1 = _mm256_stream_load_si256(in);         // AaBbCcDd EeFfGgHh
+               __m256i data2 = _mm256_stream_load_si256(in + 1);     // IiJjKkLl MmNnOoPp
+
+               data1 = _mm256_shuffle_epi8(data1, shuffle_cw);       // ABCDabcd EFGHefgh
+               data2 = _mm256_shuffle_epi8(data2, shuffle_cw);       // IJKLijkl MNOPmnop
+       
+               data1 = _mm256_permute4x64_epi64(data1, 0b11011000);  // ABCDEFGH abcdefgh
+               data2 = _mm256_permute4x64_epi64(data2, 0b11011000);  // IJKLMNOP ijklmnop
+
+               __m256i lo = _mm256_permute2x128_si256(data1, data2, 0b00100000);
+               __m256i hi = _mm256_permute2x128_si256(data1, data2, 0b00110001);
+
+               _mm256_storeu_si256(out1, lo);
+               _mm256_storeu_si256(out2, hi);
+
+               in += 2;
+               ++out1;
+               ++out2;
+               consumed += 64;
+       }
+#else
+       const __restrict __m128i *in = (const __m128i *)src;
+       __restrict __m128i *out1 = (__m128i *)dest1;
+       __restrict __m128i *out2 = (__m128i *)dest2;
+
+       __m128i mask_lower_byte = _mm_set1_epi16(0x00ff);
+       while (in < (const __m128i *)limit) {
+               __m128i data1 = _mm_load_si128(in);
+               __m128i data2 = _mm_load_si128(in + 1);
+               __m128i data1_lo = _mm_and_si128(data1, mask_lower_byte);
+               __m128i data2_lo = _mm_and_si128(data2, mask_lower_byte);
+               __m128i data1_hi = _mm_srli_epi16(data1, 8);
+               __m128i data2_hi = _mm_srli_epi16(data2, 8);
+               __m128i lo = _mm_packus_epi16(data1_lo, data2_lo);
+               _mm_storeu_si128(out1, lo);
+               __m128i hi = _mm_packus_epi16(data1_hi, data2_hi);
+               _mm_storeu_si128(out2, hi);
+
+               in += 2;
+               ++out1;
+               ++out2;
+               consumed += 32;
+       }
+#endif
+
+       return consumed;
+}
+
+#endif  // __SSE2__
+
 }  // namespace
 
 DeckLinkCapture::DeckLinkCapture(IDeckLink *card, int card_index)
@@ -167,9 +264,25 @@ HRESULT STDMETHODCALLTYPE DeckLinkCapture::VideoInputFrameArrived(
                        const uint8_t *frame_bytes;
                        video_frame->GetBytes((void **)&frame_bytes);
 
-                       memcpy_interleaved(current_video_frame.data, current_video_frame.data2,
-                               frame_bytes, width * height * 2);
-                       current_video_frame.len += width * height * 2;
+                       uint8_t *data = current_video_frame.data;
+                       uint8_t *data2 = current_video_frame.data2;
+                       size_t num_bytes = width * height * 2;
+#ifdef __SSE2__
+                       size_t consumed = memcpy_interleaved_fastpath(data, data2, frame_bytes, num_bytes);
+                       frame_bytes += consumed;
+                       data += consumed / 2;
+                       data2 += consumed / 2;
+                       if (num_bytes % 2) {
+                               swap(data, data2);
+                       }
+                       current_video_frame.len += consumed;
+                       num_bytes -= consumed;
+#endif
+
+                       if (num_bytes > 0) {
+                               memcpy_interleaved(data, data2, frame_bytes, num_bytes);
+                       }
+                       current_video_frame.len += num_bytes;
 
                        video_format.width = width;
                        video_format.height = height;