libfreerdp-rfx: SSE2 Optimizations for iDWT Decode (round 2)

1 files changed, 79 insertions, 32 deletions

diff --git a/libfreerdp-rfx/sse/rfx_sse2.c b/libfreerdp-rfx/sse/rfx_sse2.c
index 642d725..b94ac75 100644
--- a/libfreerdp-rfx/sse/rfx_sse2.c
+++ b/libfreerdp-rfx/sse/rfx_sse2.c
@@ -128,13 +128,13 @@ void rfx_quantization_decode_SSE2(sint16 * buffer, const uint32 * quantization_v
 	rfx_quantization_decode_block_SSE2(buffer + 3840, 64, quantization_values[2]); // HL3
 	rfx_quantization_decode_block_SSE2(buffer + 3904, 64, quantization_values[1]); // LH3
 	rfx_quantization_decode_block_SSE2(buffer + 3868, 64, quantization_values[3]); // HH3
-	rfx_quantization_decode_block_SSE2(buffer + 4032, 64, quantization_values[0]); // LL3	
+	rfx_quantization_decode_block_SSE2(buffer + 4032, 64, quantization_values[0]); // LL3
 }
 
 static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 rfx_dwt_2d_decode_block_horiz_SSE2(sint16 * l, sint16 * h, sint16 * dst, int subband_width)
 {
-	int y, n;		
+	int y, n;
 	sint16 * l_ptr = l;
 	sint16 * h_ptr = h;
 	sint16 * dst_ptr = dst;
@@ -166,7 +166,7 @@ rfx_dwt_2d_decode_block_horiz_SSE2(sint16 * l, sint16 * h, sint16 * dst, int sub
 			
 			l_ptr+=8;
 			h_ptr+=8;
-		}		
+		}
 		l_ptr -= subband_width;
 		h_ptr -= subband_width;
 		
@@ -206,19 +206,87 @@ rfx_dwt_2d_decode_block_horiz_SSE2(sint16 * l, sint16 * h, sint16 * dst, int sub
 }
 
 static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+rfx_dwt_2d_decode_block_vert_SSE2(sint16 * l, sint16 * h, sint16 * dst, int subband_width)
+{
+	int x, n;
+	sint16 * l_ptr = l;
+	sint16 * h_ptr = h;
+	sint16 * dst_ptr = dst;
+	
+	int total_width = subband_width * subband_width;
+
+	/* Even coefficients */
+	for (n = 0; n < subband_width; n++)
+	{
+		for (x = 0; x < total_width; x+=8)
+		{
+			// dst[2n] = l[n] - ((h[n-1] + h[n] + 1) >> 1);
+			
+			__m128i l_n = _mm_load_si128((__m128i*) l_ptr);
+			__m128i h_n = _mm_load_si128((__m128i*) h_ptr);
+			
+			__m128i tmp_n = _mm_add_epi16(h_n, _mm_set1_epi16(1));;
+			if (n == 0)
+				tmp_n = _mm_add_epi16(tmp_n, h_n);
+			else
+			{
+				__m128i h_n_m = _mm_loadu_si128((__m128i*) (h_ptr - total_width));
+				tmp_n = _mm_add_epi16(tmp_n, h_n_m);
+			}
+			tmp_n = _mm_srai_epi16(tmp_n, 1);
+			
+			__m128i dst_n = _mm_sub_epi16(l_n, tmp_n);
+			_mm_store_si128((__m128i*) dst_ptr, dst_n);
+			
+			l_ptr+=8;
+			h_ptr+=8;
+			dst_ptr+=8;
+		}
+		dst_ptr+=total_width;
+	}
+	
+	h_ptr = h;
+	dst_ptr = dst + total_width;
+	
+	/* Odd coefficients */
+	for (n = 0; n < subband_width; n++)
+	{
+		for (x = 0; x < total_width; x+=8)
+		{
+			// dst[2n + 1] = (h[n] << 1) + ((dst[2n] + dst[2n + 2]) >> 1);
+			
+			__m128i h_n = _mm_load_si128((__m128i*) h_ptr);
+			__m128i dst_n_m = _mm_load_si128((__m128i*) (dst_ptr - total_width));
+			h_n = _mm_slli_epi16(h_n, 1);
+			
+			__m128i tmp_n = dst_n_m;
+			if (n == subband_width - 1)
+				tmp_n = _mm_add_epi16(tmp_n, dst_n_m);
+			else
+			{
+				__m128i dst_n_p = _mm_loadu_si128((__m128i*) (dst_ptr + total_width));
+				tmp_n = _mm_add_epi16(tmp_n, dst_n_p);
+			}
+			tmp_n = _mm_srai_epi16(tmp_n, 1);
+			
+			__m128i dst_n = _mm_add_epi16(tmp_n, h_n);
+			_mm_store_si128((__m128i*) dst_ptr, dst_n);
+
+			h_ptr+=8;
+			dst_ptr+=8;
+		}
+		dst_ptr+=total_width;
+	}
+}
+
+static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 rfx_dwt_2d_decode_block_SSE2(sint16 * buffer, sint16 * idwt, int subband_width)
 {
-	sint16 * dst, * l, * h;
-	sint16 * l_dst, * h_dst;
 	sint16 * hl, * lh, * hh, * ll;
-	int total_width;
-	int x, y;
-	int n;
+	sint16 * l_dst, * h_dst;
 
 	_mm_prefetch_buffer((char *) idwt, subband_width * 4 * sizeof(sint16));
 
-	total_width = subband_width + subband_width;
-
 	/* Inverse DWT in horizontal direction, results in 2 sub-bands in L, H order in tmp buffer idwt. */
 	/* The 4 sub-bands are stored in HL(0), LH(1), HH(2), LL(3) order. */
 	/* The lower part L uses LL(3) and HL(0). */
@@ -237,28 +305,7 @@ rfx_dwt_2d_decode_block_SSE2(sint16 * buffer, sint16 * idwt, int subband_width)
 	rfx_dwt_2d_decode_block_horiz_SSE2(lh, hh, h_dst, subband_width);
 
 	/* Inverse DWT in vertical direction, results are stored in original buffer. */
-	for (x = 0; x < total_width; x++)
-	{
-		/* Even coefficients */
-		for (n = 0; n < subband_width; n++)
-		{
-			y = n << 1;
-			dst = buffer + y * total_width + x;
-			l = idwt + n * total_width + x;
-			h = l + subband_width * total_width;
-			dst[0] = *l - (((n > 0 ? *(h - total_width) : *h) + (*h) + 1) >> 1);
-		}
-
-		/* Odd coefficients */
-		for (n = 0; n < subband_width; n++)
-		{
-			y = n << 1;
-			dst = buffer + y * total_width + x;
-			l = idwt + n * total_width + x;
-			h = l + subband_width * total_width;
-			dst[total_width] = (*h << 1) + ((dst[0] + dst[n < subband_width - 1 ? 2 * total_width : 0]) >> 1);
-		}
-	}
+	rfx_dwt_2d_decode_block_vert_SSE2(l_dst, h_dst, buffer, subband_width);
 }
 
 void rfx_dwt_2d_decode_SSE2(sint16 * buffer, sint16 * dwt_buffer_8, sint16 * dwt_buffer_16, sint16 * dwt_buffer_32)