(svn r26213) -Add: specialised non-animated SSE4 blitter (MJP)

With 32bpp base set about 40% faster than 32bpp-optimized, or about 10% for 8bpp base sets in the Draw function. Respectively about 8 and 1% of total run time

2 files changed, 394 insertions, 0 deletions

diff --git a/src/blitter/32bpp_sse4.cpp b/src/blitter/32bpp_sse4.cpp
new file mode 100644
index 000000000..7243f754b
--- /dev/null
+++ b/src/blitter/32bpp_sse4.cpp
@@ -0,0 +1,327 @@
+/* $Id$ */
+
+/*
+ * This file is part of OpenTTD.
+ * OpenTTD is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, version 2.
+ * OpenTTD is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+ * See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with OpenTTD. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+/** @file 32bpp_sse4.cpp Implementation of the SSE4 32 bpp blitter. */
+
+#ifdef WITH_SSE
+
+#include "../stdafx.h"
+#include "../zoom_func.h"
+#include "../settings_type.h"
+#include "32bpp_sse4.hpp"
+
+/** Instantiation of the SSE4 32bpp blitter factory. */
+static FBlitter_32bppSSE4 iFBlitter_32bppSSE4;
+
+#if defined(__GNUC__)
+	#pragma GCC diagnostic push
+	#pragma GCC diagnostic ignored "-Wunused-variable"
+#endif
+/**
+ * Draws a sprite to a (screen) buffer. It is templated to allow faster operation.
+ *
+ * @tparam mode blitter mode
+ * @param bp further blitting parameters
+ * @param zoom zoom level at which we are drawing
+ */
+template <BlitterMode mode, Blitter_32bppSSE2::ReadMode read_mode, Blitter_32bppSSE2::BlockType bt_last>
+inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom)
+{
+	const byte * const remap = bp->remap;
+	Colour *dst_line = (Colour *) bp->dst + bp->top * bp->pitch + bp->left;
+	int effective_width = bp->width;
+
+	/* Find where to start reading in the source sprite. */
+	const SpriteData * const sd = (const SpriteData *) bp->sprite;
+	const SpriteInfo * const si = &sd->infos[zoom];
+	const MapValue *src_mv_line = (const MapValue *) &sd->data[si->mv_offset] + bp->skip_top * si->sprite_width;
+	const Colour *src_rgba_line = (const Colour *) ((const byte *) &sd->data[si->sprite_offset] + bp->skip_top * si->sprite_line_size);
+
+	if (read_mode != RM_WITH_MARGIN) {
+		src_rgba_line += bp->skip_left;
+		src_mv_line += bp->skip_left;
+	}
+
+	/* Load these variables into register before loop. */
+	const __m128i a_cm        = ALPHA_CONTROL_MASK;
+	const __m128i pack_low_cm = PACK_LOW_CONTROL_MASK;
+	const __m128i briAB_cm    = BRIGHTNESS_LOW_CONTROL_MASK;
+	const __m128i div_cleaner = BRIGHTNESS_DIV_CLEANER;
+	const __m128i ob_check    = OVERBRIGHT_PRESENCE_MASK;
+	const __m128i ob_mask     = OVERBRIGHT_VALUE_MASK;
+	const __m128i ob_cm       = OVERBRIGHT_CONTROL_MASK;
+	const __m128i tr_nom_base = TRANSPARENT_NOM_BASE;
+
+	for (int y = bp->height; y != 0; y--) {
+		const Colour *src = src_rgba_line + META_LENGTH;
+		Colour *dst = dst_line;
+		const MapValue *src_mv = src_mv_line;
+
+		switch (mode) {
+			default: {
+				switch (read_mode) {
+					case RM_WITH_MARGIN: {
+						src += src_rgba_line[0].data;
+						dst += src_rgba_line[0].data;
+						const int width_diff = si->sprite_width - bp->width;
+						effective_width = bp->width - (int) src_rgba_line[0].data;
+						const int delta_diff = (int) src_rgba_line[1].data - width_diff;
+						const int new_width = effective_width - (delta_diff & ~1);
+						effective_width = delta_diff > 0 ? new_width : effective_width;
+						if (effective_width <= 0) break;
+						/* FALLTHROUGH */
+					}
+
+					case RM_WITH_SKIP: {
+						__m128i srcABCD = _mm_loadu_si128((const __m128i*) src);
+						__m128i dstABCD = _mm_loadu_si128((__m128i*) dst);
+						for (uint x = (uint) effective_width / 2; x > 0; x--) {
+							ALPHA_BLEND_2(pack_low_cm);
+							srcABCD = _mm_blend_epi16(srcABCD, dstABCD, 0xF0);
+							Colour *old_dst = dst;
+							src += 2;
+							dst += 2;
+							/* It is VERY important to read next data before it gets invalidated in cpu cache.
+							 * And PEXTR latency is a real problem here.
+							 */
+							dstABCD = _mm_loadu_si128((__m128i*) dst);
+							_mm_storeu_si128((__m128i *) old_dst, srcABCD);
+							srcABCD = _mm_loadu_si128((const __m128i*) src);
+						}
+						if (bt_last == BT_ODD) {
+							ALPHA_BLEND_2(pack_low_cm);
+							*dst = (Colour) EXTR32(srcABCD, 0);
+						}
+						break;
+					}
+
+					default: NOT_REACHED();
+				}
+				break;
+			}
+
+			case BM_COLOUR_REMAP: {
+				switch (read_mode) {
+					case RM_WITH_MARGIN: {
+						src += src_rgba_line[0].data;
+						src_mv += src_rgba_line[0].data;
+						dst += src_rgba_line[0].data;
+						const int width_diff = si->sprite_width - bp->width;
+						effective_width = bp->width - (int) src_rgba_line[0].data;
+						const int delta_diff = (int) src_rgba_line[1].data - width_diff;
+						const int nd = effective_width - delta_diff;
+						effective_width = delta_diff > 0 ? nd : effective_width;
+						if (effective_width <= 0) break;
+						/* FALLTHROUGH */
+					}
+
+					case RM_WITH_SKIP: {
+						__m128i srcABCD = _mm_loadu_si128((const __m128i*) src);
+						__m128i dstABCD = _mm_loadu_si128((__m128i*) dst);
+						uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv));
+
+						for (uint x = (uint) effective_width / 2; x > 0; x--) {
+							/* Remap colours. */
+							if (mvX2 & 0x00FF00FF) {
+								/* Written so the compiler uses CMOV. */
+								const Colour src0 = src[0];
+								const uint m0 = (byte) mvX2;
+								const uint r0 = remap[m0];
+								const Colour c0map = (this->LookupColourInPalette(r0).data & 0x00FFFFFF) | (src0.data & 0xFF000000);
+								Colour c0 = 0; // Use alpha of 0 to keep dst as is.
+								c0 = r0 == 0 ? c0 : c0map;
+								c0 = m0 != 0 ? c0 : src0;
+								INSR32(c0.data, srcABCD, 0);
+
+								const Colour src1 = src[1];
+								const uint m1 = (byte) (mvX2 >> 16);
+								const uint r1 = remap[m1];
+								const Colour c1map = (this->LookupColourInPalette(r1).data & 0x00FFFFFF) | (src1.data & 0xFF000000);
+								Colour c1 = 0;
+								c1 = r1 == 0 ? c1 : c1map;
+								c1 = m1 != 0 ? c1 : src1;
+								INSR32(c1.data, srcABCD, 1);
+
+								if ((mvX2 & 0xFF00FF00) != 0x80008000) {
+									ADJUST_BRIGHTNESS_2(srcABCD, mvX2);
+								}
+							}
+
+							/* Blend colours. */
+							ALPHA_BLEND_2(pack_low_cm);
+							srcABCD = _mm_blend_epi16(srcABCD, dstABCD, 0xF0);
+							Colour *old_dst = dst;
+							dst += 2;
+							src += 2;
+							src_mv += 2;
+							dstABCD = _mm_loadu_si128((__m128i*) dst);
+							_mm_storeu_si128((__m128i *) old_dst, srcABCD);
+							mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv));
+							srcABCD = _mm_loadu_si128((const __m128i*) src);
+						}
+
+						if (effective_width & 1) {
+							/* In case the m-channel is zero, do not remap this pixel in any way. */
+							if ((byte) mvX2 == 0) {
+								if (src->a < 255) {
+									ALPHA_BLEND_2(pack_low_cm);
+									(*dst).data = EXTR32(srcABCD, 0);
+								} else
+									*dst = *src;
+							} else {
+								const uint r = remap[(byte) mvX2];
+								if (r != 0) {
+									Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), (byte) (mvX2 >> 8));
+									if (src->a == 255) {
+										*dst = remapped_colour;
+									} else {
+										remapped_colour.a = src->a;
+										INSR32(remapped_colour.data, srcABCD, 0);
+										ALPHA_BLEND_2(pack_low_cm);
+										(*dst).data = EXTR32(srcABCD, 0);
+									}
+								}
+							}
+						}
+						break;
+					}
+
+					default: NOT_REACHED();
+				}
+				src_mv_line += si->sprite_width;
+				break;
+			}
+
+			case BM_TRANSPARENT: {
+				/* Make the current colour a bit more black, so it looks like this image is transparent.
+				 * rgb = rgb * ((256/4) * 4 - (alpha/4)) / ((256/4) * 4)
+				 */
+				__m128i srcABCD = _mm_loadu_si128((const __m128i*) src);
+				__m128i dstABCD = _mm_loadu_si128((__m128i*) dst);
+				for (uint x = (uint) bp->width / 2; x > 0; x--) {
+					__m128i srcAB = _mm_unpacklo_epi8(srcABCD, _mm_setzero_si128());
+					__m128i dstAB = _mm_unpacklo_epi8(dstABCD, _mm_setzero_si128());
+					__m128i dstCD = _mm_unpackhi_epi8(dstABCD, _mm_setzero_si128());
+					__m128i alphaAB = _mm_shuffle_epi8(srcAB, a_cm);
+					alphaAB = _mm_srli_epi16(alphaAB, 2); // Reduce to 64 levels of shades so the max value fits in 16 bits.
+					__m128i nom = _mm_sub_epi16(tr_nom_base, alphaAB);
+					dstAB = _mm_mullo_epi16(dstAB, nom);
+					dstAB = _mm_srli_epi16(dstAB, 8);
+					dstAB = _mm_packus_epi16(dstAB, dstCD);
+					Colour *old_dst = dst;
+					src += 2;
+					dst += 2;
+					dstABCD = _mm_loadu_si128((__m128i*) dst);
+					_mm_storeu_si128((__m128i *) old_dst, dstAB);
+					srcABCD = _mm_loadu_si128((const __m128i*) src);
+				}
+				if (bp->width & 1) {
+					__m128i srcAB = _mm_unpacklo_epi8(srcABCD, _mm_setzero_si128());
+					__m128i dstAB = _mm_unpacklo_epi8(dstABCD, _mm_setzero_si128());
+					__m128i alphaAB = _mm_shuffle_epi8(srcAB, a_cm);
+					alphaAB = _mm_srli_epi16(alphaAB, 2);
+					__m128i nom = _mm_sub_epi16(tr_nom_base, alphaAB);
+					dstAB = _mm_mullo_epi16(dstAB, nom);
+					dstAB = _mm_srli_epi16(dstAB, 8);
+					dstAB = _mm_packus_epi16(dstAB, dstAB);
+					(*dst).data = EXTR32(dstAB, 0);
+				}
+
+				break;
+			}
+		}
+
+		src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size);
+		dst_line += bp->pitch;
+	}
+}
+#if defined(__GNUC__)
+	#pragma GCC diagnostic pop
+#endif
+
+/**
+ * Draws a sprite to a (screen) buffer. Calls adequate templated function.
+ *
+ * @param bp further blitting parameters
+ * @param mode blitter mode
+ * @param zoom zoom level at which we are drawing
+ */
+void Blitter_32bppSSE4::Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomLevel zoom)
+{
+	const BlockType bt_last = (BlockType) (bp->width & 1);
+	switch (mode) {
+		case BM_NORMAL: {
+			if (bp->skip_left != 0 || bp->width <= MARGIN_NORMAL_THRESHOLD) {
+				switch (bt_last) {
+					case BT_EVEN: Draw<BM_NORMAL, RM_WITH_SKIP, BT_EVEN>(bp, zoom); return;
+					case BT_ODD:  Draw<BM_NORMAL, RM_WITH_SKIP, BT_ODD>(bp, zoom); return;
+					default: NOT_REACHED();
+				}
+			} else {
+				switch (bt_last) {
+					case BT_EVEN: Draw<BM_NORMAL, RM_WITH_MARGIN, BT_EVEN>(bp, zoom); return;
+					case BT_ODD:  Draw<BM_NORMAL, RM_WITH_MARGIN, BT_ODD>(bp, zoom); return;
+					default: NOT_REACHED();
+				}
+			}
+			break;
+		}
+		case BM_COLOUR_REMAP:
+			if (bp->skip_left != 0 || bp->width <= MARGIN_REMAP_THRESHOLD) {
+				Draw<BM_COLOUR_REMAP, RM_WITH_SKIP, BT_NONE>(bp, zoom); return;
+			} else {
+				Draw<BM_COLOUR_REMAP, RM_WITH_MARGIN, BT_NONE>(bp, zoom); return;
+			}
+		case BM_TRANSPARENT:  Draw<BM_TRANSPARENT, RM_NONE, BT_NONE>(bp, zoom); return;
+		default: NOT_REACHED();
+	}
+}
+
+/** Same code as seen in 32bpp_sse2.cpp but some macros are not the same. */
+inline Colour Blitter_32bppSSE4::AdjustBrightness(Colour colour, uint8 brightness)
+{
+	/* Shortcut for normal brightness. */
+	if (brightness == DEFAULT_BRIGHTNESS) return colour;
+
+	return this->ReallyAdjustBrightness(colour, brightness);
+}
+
+Colour Blitter_32bppSSE4::ReallyAdjustBrightness(Colour colour, uint8 brightness)
+{
+	ALIGN(16) uint64 c16 = colour.b | (uint64) colour.g << 16 | (uint64) colour.r << 32;
+	c16 *= brightness;
+	uint64 c16_ob = c16; // Helps out of order execution.
+	c16 /= DEFAULT_BRIGHTNESS;
+	c16 &= 0x01FF01FF01FF;
+
+	/* Sum overbright (maximum for each rgb is 508, 9 bits, -255 is changed in -256 so we just have to take the 8 lower bits into account). */
+	c16_ob = (((c16_ob >> (8 + 7)) & 0x0100010001) * 0xFF) & c16;
+	uint64 ob = (uint16) c16_ob + (uint16) (c16_ob >> 16) + (uint16) (c16_ob >> 32);
+
+	const uint32 alpha32 = colour.data & 0xFF000000;
+	__m128i ret;
+	INSR64(c16, ret, 0);
+	if (ob != 0) {
+		/* Reduce overbright strength. */
+		ob /= 2;
+		__m128i ob128;
+		INSR64(ob | ob << 16 | ob << 32, ob128, 0);
+		__m128i white = OVERBRIGHT_VALUE_MASK;
+		__m128i c128 = ret;
+		ret = _mm_subs_epu16(white, c128); /* PSUBUSW,   (255 - rgb) */
+		ret = _mm_mullo_epi16(ret, ob128); /* PMULLW, ob*(255 - rgb) */
+		ret = _mm_srli_epi16(ret, 8);      /* PSRLW,  ob*(255 - rgb)/256 */
+		ret = _mm_add_epi16(ret, c128);    /* PADDW,  ob*(255 - rgb)/256 + rgb */
+	}
+
+	ret = _mm_packus_epi16(ret, ret);      /* PACKUSWB, saturate and pack. */
+	return alpha32 | EXTR32(ret, 0);
+}
+
+#endif /* WITH_SSE */
diff --git a/src/blitter/32bpp_sse4.hpp b/src/blitter/32bpp_sse4.hpp
new file mode 100644
index 000000000..1930404b6
--- /dev/null
+++ b/src/blitter/32bpp_sse4.hpp
@@ -0,0 +1,67 @@
+/* $Id$ */
+
+/*
+ * This file is part of OpenTTD.
+ * OpenTTD is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, version 2.
+ * OpenTTD is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
+ * See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with OpenTTD. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+/** @file 32bpp_sse4.hpp SSE4 32 bpp blitter. */
+
+#ifndef BLITTER_32BPP_SSE4_HPP
+#define BLITTER_32BPP_SSE4_HPP
+
+#ifdef WITH_SSE
+
+#include "32bpp_ssse3.hpp"
+#include "smmintrin.h"
+
+#undef EXTR32
+#define EXTR32(from, rank) _mm_extract_epi32((*(um128i*) &from).m128i, rank)
+#undef INSR32
+#define INSR32(val, into, rank) (*(um128i*) &into).m128i = _mm_insert_epi32((*(um128i*) &into).m128i, val, rank)
+
+#if defined(__GNUC__)
+	#pragma GCC diagnostic push
+	#pragma GCC diagnostic ignored "-Wuninitialized"
+	#pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
+#endif
+#ifdef _SQ64
+	#undef INSR64
+	#define INSR64(val, into, rank) (*(um128i*) &into).m128i = _mm_insert_epi64((*(um128i*) &into).m128i, val, rank)
+#else
+	typedef union { uint64 u64; struct _u32 { uint32 low, high; } u32; } u6432;
+	#undef INSR64
+	#define INSR64(val, into, rank) { \
+		u6432 v; \
+		v.u64 = val; \
+		(*(um128i*) &into).m128i = _mm_insert_epi32((*(um128i*) &into).m128i, v.u32.low, (rank)*2); \
+		(*(um128i*) &into).m128i = _mm_insert_epi32((*(um128i*) &into).m128i, v.u32.high, (rank)*2 + 1); \
+	}
+#endif
+#if defined(__GNUC__)
+	#pragma GCC diagnostic pop
+#endif
+
+/** The SSE4 32 bpp blitter (without palette animation). */
+class Blitter_32bppSSE4 : public Blitter_32bppSSSE3 {
+public:
+	Colour AdjustBrightness(Colour colour, uint8 brightness);
+	Colour ReallyAdjustBrightness(Colour colour, uint8 brightness);
+
+	/* virtual */ void Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomLevel zoom);
+	template <BlitterMode mode, ReadMode read_mode, BlockType bt_last>
+	void Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom);
+	/* virtual */ const char *GetName() { return "32bpp-sse4"; }
+};
+
+/** Factory for the SSE4 32 bpp blitter (without palette animation). */
+class FBlitter_32bppSSE4: public BlitterFactory {
+public:
+	FBlitter_32bppSSE4() : BlitterFactory("32bpp-sse4", "32bpp SSE4 Blitter (no palette animation)", HasCPUIDFlag(1, 2, 19)) {}
+	/* virtual */ Blitter *CreateInstance() { return new Blitter_32bppSSE4(); }
+};
+
+#endif /* WITH_SSE */
+#endif /* BLITTER_32BPP_SSE4_HPP */