libfreerdp-rfx: add SSE2 for DWT encoding.

6 files changed, 155 insertions, 1 deletions

diff --git a/include/freerdp/rfx.h b/include/freerdp/rfx.h
index 7b2fede..932f5f5 100644
--- a/include/freerdp/rfx.h
+++ b/include/freerdp/rfx.h
@@ -167,6 +167,7 @@ struct _RFX_CONTEXT
 	void (* quantization_decode)(sint16 * buffer, const uint32 * quantization_values);
 	void (* quantization_encode)(sint16 * buffer, const uint32 * quantization_values);
 	void (* dwt_2d_decode)(sint16 * buffer, sint16 * dwt_buffer);
+	void (* dwt_2d_encode)(sint16 * buffer, sint16 * dwt_buffer);
 
 	/* profiler definitions */
 	PROFILER_DEFINE(prof_rfx_decode_rgb);
diff --git a/libfreerdp-rfx/librfx.c b/libfreerdp-rfx/librfx.c
index 1693bfb..5d9b5a6 100644
--- a/libfreerdp-rfx/librfx.c
+++ b/libfreerdp-rfx/librfx.c
@@ -146,6 +146,7 @@ rfx_context_new(void)
 	context->quantization_decode = rfx_quantization_decode;	
 	context->quantization_encode = rfx_quantization_encode;	
 	context->dwt_2d_decode = rfx_dwt_2d_decode;
+	context->dwt_2d_encode = rfx_dwt_2d_encode;
 
 	/* detect and enable SIMD CPU acceleration */
 	RFX_INIT_SIMD(context);
diff --git a/libfreerdp-rfx/rfx_encode.c b/libfreerdp-rfx/rfx_encode.c
index ddb9d2d..cd20200 100644
--- a/libfreerdp-rfx/rfx_encode.c
+++ b/libfreerdp-rfx/rfx_encode.c
@@ -132,7 +132,7 @@ rfx_encode_component(RFX_CONTEXT * context, const uint32 * quantization_values,
 	PROFILER_ENTER(context->prof_rfx_encode_component);
 
 	PROFILER_ENTER(context->prof_rfx_dwt_2d_encode);
-		rfx_dwt_2d_encode(data, context->dwt_buffer);
+		context->dwt_2d_encode(data, context->dwt_buffer);
 	PROFILER_EXIT(context->prof_rfx_dwt_2d_encode);
 
 	PROFILER_ENTER(context->prof_rfx_quantization_encode);
diff --git a/libfreerdp-rfx/sse/rfx_sse.c b/libfreerdp-rfx/sse/rfx_sse.c
index 0407323..76a632d 100644
--- a/libfreerdp-rfx/sse/rfx_sse.c
+++ b/libfreerdp-rfx/sse/rfx_sse.c
@@ -33,10 +33,12 @@ void rfx_init_sse(RFX_CONTEXT * context)
 		IF_PROFILER(context->prof_rfx_quantization_decode->name = "rfx_quantization_decode_SSE2");
 		IF_PROFILER(context->prof_rfx_quantization_encode->name = "rfx_quantization_encode_SSE2");
 		IF_PROFILER(context->prof_rfx_dwt_2d_decode->name = "rfx_dwt_2d_decode_SSE2");
+		IF_PROFILER(context->prof_rfx_dwt_2d_encode->name = "rfx_dwt_2d_encode_SSE2");
 		
 		context->decode_YCbCr_to_RGB = rfx_decode_YCbCr_to_RGB_SSE2;
 		context->encode_RGB_to_YCbCr = rfx_encode_RGB_to_YCbCr_SSE2;
 		context->quantization_decode = rfx_quantization_decode_SSE2;
 		context->quantization_encode = rfx_quantization_encode_SSE2;
 		context->dwt_2d_decode = rfx_dwt_2d_decode_SSE2;
+		context->dwt_2d_encode = rfx_dwt_2d_encode_SSE2;
 }
diff --git a/libfreerdp-rfx/sse/rfx_sse2.c b/libfreerdp-rfx/sse/rfx_sse2.c
index a215f86..cdaef30 100644
--- a/libfreerdp-rfx/sse/rfx_sse2.c
+++ b/libfreerdp-rfx/sse/rfx_sse2.c
@@ -423,3 +423,152 @@ rfx_dwt_2d_decode_SSE2(sint16 * buffer, sint16 * dwt_buffer)
 	rfx_dwt_2d_decode_block_SSE2(buffer + 3072, dwt_buffer, 16);
 	rfx_dwt_2d_decode_block_SSE2(buffer, dwt_buffer, 32);
 }
+
+static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+rfx_dwt_2d_encode_block_vert_SSE2(sint16 * src, sint16 * l, sint16 * h, int subband_width)
+{
+	int total_width;
+	int x;
+	int n;
+	__m128i src_2n;
+	__m128i src_2n_1;
+	__m128i src_2n_2;
+	__m128i h_n;
+	__m128i h_n_m;
+	__m128i l_n;
+
+	total_width = subband_width << 1;
+
+	for (n = 0; n < subband_width; n++)
+	{
+		for (x = 0; x < total_width; x += 8)
+		{
+			src_2n = _mm_load_si128((__m128i*) src);
+			src_2n_1 = _mm_load_si128((__m128i*) (src + total_width));
+			if (n < subband_width - 1)
+				src_2n_2 = _mm_load_si128((__m128i*) (src + 2 * total_width));
+			else
+				src_2n_2 = src_2n_1;
+
+			/* h[n] = (src[2n + 1] - ((src[2n] + src[2n + 2]) >> 1)) >> 1 */
+
+			h_n = _mm_add_epi16(src_2n, src_2n_2);
+			h_n = _mm_srai_epi16(h_n, 1);
+			h_n = _mm_sub_epi16(src_2n_1, h_n);
+			h_n = _mm_srai_epi16(h_n, 1);
+
+			_mm_store_si128((__m128i*) h, h_n);
+
+			if (n == 0)
+				h_n_m = h_n;
+			else
+				h_n_m = _mm_load_si128((__m128i*) (h - total_width));
+
+			/* l[n] = src[2n] + ((h[n - 1] + h[n]) >> 1) */
+
+			l_n = _mm_add_epi16(h_n_m, h_n);
+			l_n = _mm_srai_epi16(l_n, 1);
+			l_n = _mm_add_epi16(l_n, src_2n);
+
+			_mm_store_si128((__m128i*) l, l_n);
+
+			src += 8;
+			l += 8;
+			h += 8;
+		}
+		src += total_width;
+	}
+}
+
+static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+rfx_dwt_2d_encode_block_horiz_SSE2(sint16 * src, sint16 * l, sint16 * h, int subband_width)
+{
+	int y;
+	int n;
+	int first;
+	__m128i src_2n;
+	__m128i src_2n_1;
+	__m128i src_2n_2;
+	__m128i h_n;
+	__m128i h_n_m;
+	__m128i l_n;
+
+	for (y = 0; y < subband_width; y++)
+	{
+		for (n = 0; n < subband_width; n += 8)
+		{
+			/* The following 3 Set operations consumes more than half of the total DWT processing time! */
+			src_2n = _mm_set_epi16(src[14], src[12], src[10], src[8], src[6], src[4], src[2], src[0]);
+			src_2n_1 = _mm_set_epi16(src[15], src[13], src[11], src[9], src[7], src[5], src[3], src[1]);
+			src_2n_2 = _mm_set_epi16(n == subband_width - 8 ? src[15] : src[16],
+				src[14], src[12], src[10], src[8], src[6], src[4], src[2]);
+
+			/* h[n] = (src[2n + 1] - ((src[2n] + src[2n + 2]) >> 1)) >> 1 */
+
+			h_n = _mm_add_epi16(src_2n, src_2n_2);
+			h_n = _mm_srai_epi16(h_n, 1);
+			h_n = _mm_sub_epi16(src_2n_1, h_n);
+			h_n = _mm_srai_epi16(h_n, 1);
+
+			_mm_store_si128((__m128i*) h, h_n);
+
+			h_n_m = _mm_loadu_si128((__m128i*) (h - 1));
+			if (n == 0)
+			{
+				first = _mm_extract_epi16(h_n_m, 1);
+				h_n_m = _mm_insert_epi16(h_n_m, first, 0);
+			}
+
+			/* l[n] = src[2n] + ((h[n - 1] + h[n]) >> 1) */
+
+			l_n = _mm_add_epi16(h_n_m, h_n);
+			l_n = _mm_srai_epi16(l_n, 1);
+			l_n = _mm_add_epi16(l_n, src_2n);
+
+			_mm_store_si128((__m128i*) l, l_n);
+
+			src += 16;
+			l += 8;
+			h += 8;
+		}
+	}
+}
+
+static __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+rfx_dwt_2d_encode_block_SSE2(sint16 * buffer, sint16 * dwt, int subband_width)
+{
+	sint16 * hl, * lh, * hh, * ll;
+	sint16 * l_src, * h_src;
+
+	_mm_prefetch_buffer((char *) dwt, subband_width * 4 * sizeof(sint16));
+
+	/* DWT in vertical direction, results in 2 sub-bands in L, H order in tmp buffer dwt. */
+
+	l_src = dwt;
+	h_src = dwt + subband_width * subband_width * 2;
+
+	rfx_dwt_2d_encode_block_vert_SSE2(buffer, l_src, h_src, subband_width);
+
+	/* DWT in horizontal direction, results in 4 sub-bands in HL(0), LH(1), HH(2), LL(3) order, stored in original buffer. */
+	/* The lower part L generates LL(3) and HL(0). */
+	/* The higher part H generates LH(1) and HH(2). */
+
+	ll = buffer + subband_width * subband_width * 3;
+	hl = buffer;
+
+	lh = buffer + subband_width * subband_width;
+	hh = buffer + subband_width * subband_width * 2;
+
+	rfx_dwt_2d_encode_block_horiz_SSE2(l_src, ll, hl, subband_width);
+	rfx_dwt_2d_encode_block_horiz_SSE2(h_src, lh, hh, subband_width);
+}
+
+void
+rfx_dwt_2d_encode_SSE2(sint16 * buffer, sint16 * dwt_buffer)
+{
+	_mm_prefetch_buffer((char *) buffer, 4096 * sizeof(sint16));
+	
+	rfx_dwt_2d_encode_block_SSE2(buffer, dwt_buffer, 32);
+	rfx_dwt_2d_encode_block_SSE2(buffer + 3072, dwt_buffer, 16);
+	rfx_dwt_2d_encode_block_SSE2(buffer + 3840, dwt_buffer, 8);
+}
diff --git a/libfreerdp-rfx/sse/rfx_sse2.h b/libfreerdp-rfx/sse/rfx_sse2.h
index 0fcf604..85921da 100644
--- a/libfreerdp-rfx/sse/rfx_sse2.h
+++ b/libfreerdp-rfx/sse/rfx_sse2.h
@@ -27,5 +27,6 @@ void rfx_encode_RGB_to_YCbCr_SSE2(sint16 * y_r_buffer, sint16 * cb_g_buffer, sin
 void rfx_quantization_decode_SSE2(sint16 * buffer, const uint32 * quantization_values);
 void rfx_quantization_encode_SSE2(sint16 * buffer, const uint32 * quantization_values);
 void rfx_dwt_2d_decode_SSE2(sint16 * buffer, sint16 * dwt_buffer);
+void rfx_dwt_2d_encode_SSE2(sint16 * buffer, sint16 * dwt_buffer);
 
 #endif /* __RFX_SSE2_H */