Add: 32bpp SSE2 blitter palette animator (#6795)

Create a new blitter mode: 32bpp-sse2-anim, which is 32bpp-anim + this.
32bpp-sse2-anim is now used by default where 32bpp-anim would have been.
Also use this with the 32bpp-sse4-anim blitter.

See issue #6469.

8 files changed, 214 insertions, 44 deletions

diff --git a/src/blitter/32bpp_anim.cpp b/src/blitter/32bpp_anim.cpp
index 532ad23cc..578d85f18 100644
--- a/src/blitter/32bpp_anim.cpp
+++ b/src/blitter/32bpp_anim.cpp
@@ -22,7 +22,7 @@ static FBlitter_32bppAnim iFBlitter_32bppAnim;
 
 Blitter_32bppAnim::~Blitter_32bppAnim()
 {
-	free(this->anim_buf);
+	free(this->anim_alloc);
 }
 
 template <BlitterMode mode>
@@ -39,8 +39,7 @@ inline void Blitter_32bppAnim::Draw(const Blitter::BlitterParams *bp, ZoomLevel
 	}
 
 	Colour *dst = (Colour *)bp->dst + bp->top * bp->pitch + bp->left;
-	assert(_screen.pitch == this->anim_buf_pitch); // precondition for translating 'bp->dst' into an 'anim_buf' offset below.
-	uint16 *anim = this->anim_buf + ((uint32 *)bp->dst - (uint32 *)_screen.dst_ptr) + bp->top * this->anim_buf_pitch + bp->left;
+	uint16 *anim = this->anim_buf + this->ScreenToAnimOffset((uint32 *)bp->dst) + bp->top * this->anim_buf_pitch + bp->left;
 
 	const byte *remap = bp->remap; // store so we don't have to access it via bp everytime
 
@@ -280,8 +279,7 @@ void Blitter_32bppAnim::DrawColourMappingRect(void *dst, int width, int height,
 	}
 
 	Colour *udst = (Colour *)dst;
-	assert(_screen.pitch == this->anim_buf_pitch); // precondition for translating 'dst' into an 'anim_buf' offset below.
-	uint16 *anim = this->anim_buf + ((uint32 *)dst - (uint32 *)_screen.dst_ptr);
+	uint16 *anim = this->anim_buf + this->ScreenToAnimOffset((uint32 *)dst);
 
 	if (pal == PALETTE_TO_TRANSPARENT) {
 		do {
@@ -319,8 +317,8 @@ void Blitter_32bppAnim::SetPixel(void *video, int x, int y, uint8 colour)
 
 	/* Set the colour in the anim-buffer too, if we are rendering to the screen */
 	if (_screen_disable_anim) return;
-	assert(_screen.pitch == this->anim_buf_pitch); // precondition for translating 'video' into an 'anim_buf' offset below.
-	this->anim_buf[((uint32 *)video - (uint32 *)_screen.dst_ptr) + x + y * this->anim_buf_pitch] = colour | (DEFAULT_BRIGHTNESS << 8);
+
+	this->anim_buf[this->ScreenToAnimOffset((uint32 *)video) + x + y * this->anim_buf_pitch] = colour | (DEFAULT_BRIGHTNESS << 8);
 }
 
 void Blitter_32bppAnim::DrawRect(void *video, int width, int height, uint8 colour)
@@ -332,8 +330,7 @@ void Blitter_32bppAnim::DrawRect(void *video, int width, int height, uint8 colou
 	}
 
 	Colour colour32 = LookupColourInPalette(colour);
-	assert(_screen.pitch == this->anim_buf_pitch); // precondition for translating 'video' into an 'anim_buf' offset below.
-	uint16 *anim_line = ((uint32 *)video - (uint32 *)_screen.dst_ptr) + this->anim_buf;
+	uint16 *anim_line = this->ScreenToAnimOffset((uint32 *)video) + this->anim_buf;
 
 	do {
 		Colour *dst = (Colour *)video;
@@ -357,8 +354,7 @@ void Blitter_32bppAnim::CopyFromBuffer(void *video, const void *src, int width,
 	assert(video >= _screen.dst_ptr && video <= (uint32 *)_screen.dst_ptr + _screen.width + _screen.height * _screen.pitch);
 	Colour *dst = (Colour *)video;
 	const uint32 *usrc = (const uint32 *)src;
-	assert(_screen.pitch == this->anim_buf_pitch); // precondition for translating 'video' into an 'anim_buf' offset below.
-	uint16 *anim_line = ((uint32 *)video - (uint32 *)_screen.dst_ptr) + this->anim_buf;
+	uint16 *anim_line = this->ScreenToAnimOffset((uint32 *)video) + this->anim_buf;
 
 	for (; height > 0; height--) {
 		/* We need to keep those for palette animation. */
@@ -401,8 +397,7 @@ void Blitter_32bppAnim::CopyToBuffer(const void *video, void *dst, int width, in
 
 	if (this->anim_buf == NULL) return;
 
-	assert(_screen.pitch == this->anim_buf_pitch); // precondition for translating 'video' into an 'anim_buf' offset below.
-	const uint16 *anim_line = ((const uint32 *)video - (uint32 *)_screen.dst_ptr) + this->anim_buf;
+	const uint16 *anim_line = this->ScreenToAnimOffset((const uint32 *)video) + this->anim_buf;
 
 	for (; height > 0; height--) {
 		memcpy(udst, src, width * sizeof(uint32));
@@ -485,18 +480,22 @@ void Blitter_32bppAnim::PaletteAnimate(const Palette &palette)
 	Colour *dst = (Colour *)_screen.dst_ptr;
 
 	/* Let's walk the anim buffer and try to find the pixels */
+	const int width = this->anim_buf_width;
+	const int pitch_offset = _screen.pitch - width;
+	const int anim_pitch_offset = this->anim_buf_pitch - width;
 	for (int y = this->anim_buf_height; y != 0 ; y--) {
-		for (int x = this->anim_buf_width; x != 0 ; x--) {
-			uint colour = GB(*anim, 0, 8);
+		for (int x = width; x != 0 ; x--) {
+			uint16 value = *anim;
+			uint8 colour = GB(value, 0, 8);
 			if (colour >= PALETTE_ANIM_START) {
 				/* Update this pixel */
-				*dst = this->AdjustBrightness(LookupColourInPalette(colour), GB(*anim, 8, 8));
+				*dst = this->AdjustBrightness(LookupColourInPalette(colour), GB(value, 8, 8));
 			}
 			dst++;
 			anim++;
 		}
-		dst += _screen.pitch - this->anim_buf_width;
-		anim += this->anim_buf_pitch - this->anim_buf_width;
+		dst += pitch_offset;
+		anim += anim_pitch_offset;
 	}
 
 	/* Make sure the backend redraws the whole screen */
@@ -513,10 +512,13 @@ void Blitter_32bppAnim::PostResize()
 	if (_screen.width != this->anim_buf_width || _screen.height != this->anim_buf_height ||
 			_screen.pitch != this->anim_buf_pitch) {
 		/* The size of the screen changed; we can assume we can wipe all data from our buffer */
-		free(this->anim_buf);
+		free(this->anim_alloc);
 		this->anim_buf_width = _screen.width;
 		this->anim_buf_height = _screen.height;
-		this->anim_buf_pitch = _screen.pitch;
-		this->anim_buf = CallocT<uint16>(this->anim_buf_height * this->anim_buf_pitch);
+		this->anim_buf_pitch = (_screen.width + 7) & ~7;
+		this->anim_alloc = CallocT<uint16>(this->anim_buf_pitch * this->anim_buf_height + 8);
+
+		/* align buffer to next 16 byte boundary */
+		this->anim_buf = reinterpret_cast<uint16 *>((reinterpret_cast<uintptr_t>(this->anim_alloc) + 0xF) & (~0xF));
 	}
 }
diff --git a/src/blitter/32bpp_anim.hpp b/src/blitter/32bpp_anim.hpp
index 1b35c1766..da33ec95b 100644
--- a/src/blitter/32bpp_anim.hpp
+++ b/src/blitter/32bpp_anim.hpp
@@ -18,14 +18,16 @@
 class Blitter_32bppAnim : public Blitter_32bppOptimized {
 protected:
 	uint16 *anim_buf;    ///< In this buffer we keep track of the 8bpp indexes so we can do palette animation
+	void *anim_alloc;    ///< The raw allocated buffer, not necessarily aligned correctly
 	int anim_buf_width;  ///< The width of the animation buffer.
 	int anim_buf_height; ///< The height of the animation buffer.
-	int anim_buf_pitch;  ///< The pitch of the animation buffer.
+	int anim_buf_pitch;  ///< The pitch of the animation buffer (width rounded up to 16 byte boundary).
 	Palette palette;     ///< The current palette.
 
 public:
 	Blitter_32bppAnim() :
 		anim_buf(NULL),
+		anim_alloc(NULL),
 		anim_buf_width(0),
 		anim_buf_height(0),
 		anim_buf_pitch(0)
@@ -58,6 +60,15 @@ public:
 		return this->palette.palette[index];
 	}
 
+	inline int ScreenToAnimOffset(const uint32 *video)
+	{
+		int raw_offset = video - (const uint32 *)_screen.dst_ptr;
+		if (_screen.pitch == this->anim_buf_pitch) return raw_offset;
+		int lines = raw_offset / _screen.pitch;
+		int across = raw_offset % _screen.pitch;
+		return across + (lines * this->anim_buf_pitch);
+	}
+
 	template <BlitterMode mode> void Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom);
 };
 
diff --git a/src/blitter/32bpp_anim_sse2.cpp b/src/blitter/32bpp_anim_sse2.cpp
new file mode 100644
index 000000000..d5fa4268a
--- /dev/null
+++ b/src/blitter/32bpp_anim_sse2.cpp
@@ -0,0 +1,100 @@
+/* $Id$ */
+
+/*
+ * 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_anim.cpp Implementation of a partially SSSE2 32bpp blitter with animation support. */
+
+#ifdef WITH_SSE
+
+#include "../stdafx.h"
+#include "../video/video_driver.hpp"
+#include "32bpp_anim_sse2.hpp"
+#include "32bpp_sse_func.hpp"
+
+#include "../safeguards.h"
+
+/** Instantiation of the partially SSSE2 32bpp with animation blitter factory. */
+static FBlitter_32bppSSE2_Anim iFBlitter_32bppSSE2_Anim;
+
+void Blitter_32bppSSE2_Anim::PaletteAnimate(const Palette &palette)
+{
+	assert(!_screen_disable_anim);
+
+	this->palette = palette;
+	/* If first_dirty is 0, it is for 8bpp indication to send the new
+	 *  palette. However, only the animation colours might possibly change.
+	 *  Especially when going between toyland and non-toyland. */
+	assert(this->palette.first_dirty == PALETTE_ANIM_START || this->palette.first_dirty == 0);
+
+	const uint16 *anim = this->anim_buf;
+	Colour *dst = (Colour *)_screen.dst_ptr;
+
+	bool screen_dirty = false;
+
+	/* Let's walk the anim buffer and try to find the pixels */
+	const int width = this->anim_buf_width;
+	const int screen_pitch = _screen.pitch;
+	const int anim_pitch = this->anim_buf_pitch;
+	__m128i anim_cmp = _mm_set1_epi16(PALETTE_ANIM_START - 1);
+	__m128i brightness_cmp = _mm_set1_epi16(Blitter_32bppBase::DEFAULT_BRIGHTNESS);
+	__m128i colour_mask = _mm_set1_epi16(0xFF);
+	for (int y = this->anim_buf_height; y != 0 ; y--) {
+		Colour *next_dst_ln = dst + screen_pitch;
+		const uint16 *next_anim_ln = anim + anim_pitch;
+		int x = width;
+		while (x > 0) {
+			__m128i data = _mm_load_si128((const __m128i *) anim);
+
+			/* low bytes only, shifted into high positions */
+			__m128i colour_data = _mm_and_si128(data, colour_mask);
+
+			/* test if any colour >= PALETTE_ANIM_START */
+			int colour_cmp_result = _mm_movemask_epi8(_mm_cmpgt_epi16(colour_data, anim_cmp));
+			if (colour_cmp_result) {
+				/* test if any brightness is unexpected */
+				if (x < 8 || colour_cmp_result != 0xFFFF ||
+						_mm_movemask_epi8(_mm_cmpeq_epi16(_mm_srli_epi16(data, 8), brightness_cmp)) != 0xFFFF) {
+					/* slow path: < 8 pixels left or unexpected brightnesses */
+					for (int z = min<int>(x, 8); z != 0 ; z--) {
+						int value = _mm_extract_epi16(data, 0);
+						uint8 colour = GB(value, 0, 8);
+						if (colour >= PALETTE_ANIM_START) {
+							/* Update this pixel */
+							*dst = AdjustBrightneSSE(LookupColourInPalette(colour), GB(value, 8, 8));
+							screen_dirty = true;
+						}
+						data = _mm_srli_si128(data, 2);
+						dst++;
+					}
+				} else {
+					/* medium path: 8 pixels to animate all of expected brightnesses */
+					for (int z = 0; z < 8; z++) {
+						*dst = LookupColourInPalette(_mm_extract_epi16(colour_data, 0));
+						colour_data = _mm_srli_si128(colour_data, 2);
+						dst++;
+					}
+					screen_dirty = true;
+				}
+			} else {
+				/* fast path, no animation */
+				dst += 8;
+			}
+			anim += 8;
+			x -= 8;
+		}
+		dst = next_dst_ln;
+		anim = next_anim_ln;
+	}
+
+	if (screen_dirty) {
+		/* Make sure the backend redraws the whole screen */
+		VideoDriver::GetInstance()->MakeDirty(0, 0, _screen.width, _screen.height);
+	}
+}
+
+#endif /* WITH_SSE */
diff --git a/src/blitter/32bpp_anim_sse2.hpp b/src/blitter/32bpp_anim_sse2.hpp
new file mode 100644
index 000000000..0d4a5f1e6
--- /dev/null
+++ b/src/blitter/32bpp_anim_sse2.hpp
@@ -0,0 +1,43 @@
+/* $Id$ */
+
+/*
+ * 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_anim.hpp A partially SSE2 32 bpp blitter with animation support. */
+
+#ifndef BLITTER_32BPP_SSE2_ANIM_HPP
+#define BLITTER_32BPP_SSE2_ANIM_HPP
+
+#ifdef WITH_SSE
+
+#ifndef SSE_VERSION
+#define SSE_VERSION 2
+#endif
+
+#ifndef FULL_ANIMATION
+#define FULL_ANIMATION 1
+#endif
+
+#include "32bpp_anim.hpp"
+#include "32bpp_sse2.hpp"
+
+/** A partially 32 bpp blitter with palette animation. */
+class Blitter_32bppSSE2_Anim : public Blitter_32bppAnim {
+public:
+	/* virtual */ void PaletteAnimate(const Palette &palette);
+	/* virtual */ const char *GetName() { return "32bpp-sse2-anim"; }
+};
+
+/** Factory for the partially 32bpp blitter with animation. */
+class FBlitter_32bppSSE2_Anim : public BlitterFactory {
+public:
+	FBlitter_32bppSSE2_Anim() : BlitterFactory("32bpp-sse2-anim", "32bpp partially SSE2 Animation Blitter (palette animation)", HasCPUIDFlag(1, 3, 26)) {}
+	/* virtual */ Blitter *CreateInstance() { return new Blitter_32bppSSE2_Anim(); }
+};
+
+#endif /* WITH_SSE */
+#endif /* BLITTER_32BPP_ANIM_HPP */
diff --git a/src/blitter/32bpp_anim_sse4.cpp b/src/blitter/32bpp_anim_sse4.cpp
index 987cb0c6a..219fb466b 100644
--- a/src/blitter/32bpp_anim_sse4.cpp
+++ b/src/blitter/32bpp_anim_sse4.cpp
@@ -35,8 +35,7 @@ inline void Blitter_32bppSSE4_Anim::Draw(const Blitter::BlitterParams *bp, ZoomL
 {
 	const byte * const remap = bp->remap;
 	Colour *dst_line = (Colour *) bp->dst + bp->top * bp->pitch + bp->left;
-	assert(_screen.pitch == this->anim_buf_pitch); // precondition for translating 'bp->dst' into an 'anim_buf' offset below.
-	uint16 *anim_line = this->anim_buf + ((uint32 *)bp->dst - (uint32 *)_screen.dst_ptr) + bp->top * this->anim_buf_pitch + bp->left;
+	uint16 *anim_line = this->anim_buf + this->ScreenToAnimOffset((uint32 *)bp->dst) + bp->top * this->anim_buf_pitch + bp->left;
 	int effective_width = bp->width;
 
 	/* Find where to start reading in the source sprite. */
diff --git a/src/blitter/32bpp_anim_sse4.hpp b/src/blitter/32bpp_anim_sse4.hpp
index e2d4cfc23..9d9ad5dd9 100644
--- a/src/blitter/32bpp_anim_sse4.hpp
+++ b/src/blitter/32bpp_anim_sse4.hpp
@@ -23,13 +23,14 @@
 #endif
 
 #include "32bpp_anim.hpp"
+#include "32bpp_anim_sse2.hpp"
 #include "32bpp_sse4.hpp"
 
 #undef MARGIN_NORMAL_THRESHOLD
 #define MARGIN_NORMAL_THRESHOLD 4
 
 /** The SSE4 32 bpp blitter with palette animation. */
-class Blitter_32bppSSE4_Anim FINAL : public Blitter_32bppAnim, public Blitter_32bppSSE_Base {
+class Blitter_32bppSSE4_Anim FINAL : public Blitter_32bppSSE2_Anim, public Blitter_32bppSSE_Base {
 private:
 
 public:
diff --git a/src/blitter/32bpp_base.cpp b/src/blitter/32bpp_base.cpp
index 26dd2f037..c396e4541 100644
--- a/src/blitter/32bpp_base.cpp
+++ b/src/blitter/32bpp_base.cpp
@@ -143,6 +143,36 @@ void Blitter_32bppBase::PaletteAnimate(const Palette &palette)
 	/* By default, 32bpp doesn't have palette animation */
 }
 
+Colour Blitter_32bppBase::ReallyAdjustBrightness(Colour colour, uint8 brightness)
+{
+	assert(DEFAULT_BRIGHTNESS == 1 << 7);
+
+	uint64 combined = (((uint64) colour.r) << 32) | (((uint64) colour.g) << 16) | ((uint64) colour.b);
+	combined *= brightness;
+
+	uint16 r = GB(combined, 39, 9);
+	uint16 g = GB(combined, 23, 9);
+	uint16 b = GB(combined, 7, 9);
+
+	if ((combined & 0x800080008000L) == 0L) {
+		return Colour(r, g, b, colour.a);
+	}
+
+	uint16 ob = 0;
+	/* Sum overbright */
+	if (r > 255) ob += r - 255;
+	if (g > 255) ob += g - 255;
+	if (b > 255) ob += b - 255;
+
+	/* Reduce overbright strength */
+	ob /= 2;
+	return Colour(
+		r >= 255 ? 255 : min(r + ob * (255 - r) / 256, 255),
+		g >= 255 ? 255 : min(g + ob * (255 - g) / 256, 255),
+		b >= 255 ? 255 : min(b + ob * (255 - b) / 256, 255),
+		colour.a);
+}
+
 Blitter::PaletteAnimation Blitter_32bppBase::UsePaletteAnimation()
 {
 	return Blitter::PALETTE_ANIMATION_NONE;
diff --git a/src/blitter/32bpp_base.hpp b/src/blitter/32bpp_base.hpp
index 26c3dee3f..9b7627170 100644
--- a/src/blitter/32bpp_base.hpp
+++ b/src/blitter/32bpp_base.hpp
@@ -146,30 +146,14 @@ public:
 
 	static const int DEFAULT_BRIGHTNESS = 128;
 
+	static Colour ReallyAdjustBrightness(Colour colour, uint8 brightness);
+
 	static inline Colour AdjustBrightness(Colour colour, uint8 brightness)
 	{
 		/* Shortcut for normal brightness */
 		if (brightness == DEFAULT_BRIGHTNESS) return colour;
 
-		uint16 ob = 0;
-		uint16 r = colour.r * brightness / DEFAULT_BRIGHTNESS;
-		uint16 g = colour.g * brightness / DEFAULT_BRIGHTNESS;
-		uint16 b = colour.b * brightness / DEFAULT_BRIGHTNESS;
-
-		/* Sum overbright */
-		if (r > 255) ob += r - 255;
-		if (g > 255) ob += g - 255;
-		if (b > 255) ob += b - 255;
-
-		if (ob == 0) return Colour(r, g, b, colour.a);
-
-		/* Reduce overbright strength */
-		ob /= 2;
-		return Colour(
-		                     r >= 255 ? 255 : min(r + ob * (255 - r) / 256, 255),
-		                     g >= 255 ? 255 : min(g + ob * (255 - g) / 256, 255),
-		                     b >= 255 ? 255 : min(b + ob * (255 - b) / 256, 255),
-		                     colour.a);
+		return ReallyAdjustBrightness(colour, brightness);
 	}
 };