diff options
| -rw-r--r-- | src/blitter/32bpp_anim_sse4.cpp | 379 | ||||
| -rw-r--r-- | src/blitter/32bpp_sse2.cpp | 148 | ||||
| -rw-r--r-- | src/blitter/32bpp_sse4.cpp | 211 | ||||
| -rw-r--r-- | src/blitter/32bpp_ssse3.cpp | 212 |
4 files changed, 414 insertions, 536 deletions
diff --git a/src/blitter/32bpp_anim_sse4.cpp b/src/blitter/32bpp_anim_sse4.cpp index 3229338cd..e1072dd5f 100644 --- a/src/blitter/32bpp_anim_sse4.cpp +++ b/src/blitter/32bpp_anim_sse4.cpp @@ -30,12 +30,12 @@ IGNORE_UNINITIALIZED_WARNING_START template <BlitterMode mode, Blitter_32bppSSE2::ReadMode read_mode, Blitter_32bppSSE2::BlockType bt_last> inline void Blitter_32bppSSE4_Anim::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; uint16 *anim_line = this->anim_buf + ((uint32 *)bp->dst - (uint32 *)_screen.dst_ptr) + bp->top * this->anim_buf_width + bp->left; int effective_width = bp->width; /* Find where to start reading in the source sprite. */ - const byte * const remap = bp->remap; const Blitter_32bppSSE_Base::SpriteData * const sd = (const Blitter_32bppSSE_Base::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; @@ -45,6 +45,7 @@ inline void Blitter_32bppSSE4_Anim::Draw(const Blitter::BlitterParams *bp, ZoomL src_rgba_line += bp->skip_left; src_mv_line += bp->skip_left; } + const MapValue *src_mv = src_mv_line; /* Load these variables into register before loop. */ const __m128i a_cm = ALPHA_CONTROL_MASK; @@ -59,233 +60,202 @@ inline void Blitter_32bppSSE4_Anim::Draw(const Blitter::BlitterParams *bp, ZoomL for (int y = bp->height; y != 0; y--) { Colour *dst = dst_line; const Colour *src = src_rgba_line + META_LENGTH; - const MapValue *src_mv = src_mv_line; + if (mode != BM_TRANSPARENT) src_mv = src_mv_line; uint16 *anim = anim_line; - switch (mode) { - default: { - switch (read_mode) { - case RM_WITH_MARGIN: { - src += src_rgba_line[0].data; - dst += src_rgba_line[0].data; - src_mv += src_rgba_line[0].data; - anim += 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 */ - } + if (read_mode == RM_WITH_MARGIN) { + anim += src_rgba_line[0].data; + src += src_rgba_line[0].data; + dst += src_rgba_line[0].data; + if (mode != BM_TRANSPARENT) src_mv += 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; + effective_width = delta_diff > 0 ? new_width : effective_width; + if (effective_width <= 0) goto next_line; + } - case RM_WITH_SKIP: { - for (uint x = (uint) effective_width/2; x != 0; x--) { - uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); - __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); - __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); + switch (mode) { + default: + for (uint x = (uint) effective_width/2; x != 0; x--) { + uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); + __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); + __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - /* Remap colours. */ - const byte m0 = mvX2; - if (m0 >= PALETTE_ANIM_START) { - const Colour c0 = (this->LookupColourInPalette(m0).data & 0x00FFFFFF) | (src[0].data & 0xFF000000); - INSR32(AdjustBrightness(c0, (byte) (mvX2 >> 8)).data, srcABCD, 0); - } - const byte m1 = mvX2 >> 16; - if (m1 >= PALETTE_ANIM_START) { - const Colour c1 = (this->LookupColourInPalette(m1).data & 0x00FFFFFF) | (src[1].data & 0xFF000000); - INSR32(AdjustBrightness(c1, (byte) (mvX2 >> 24)).data, srcABCD, 1); - } + /* Remap colours. */ + const byte m0 = mvX2; + if (m0 >= PALETTE_ANIM_START) { + const Colour c0 = (this->LookupColourInPalette(m0).data & 0x00FFFFFF) | (src[0].data & 0xFF000000); + INSR32(AdjustBrightness(c0, (byte) (mvX2 >> 8)).data, srcABCD, 0); + } + const byte m1 = mvX2 >> 16; + if (m1 >= PALETTE_ANIM_START) { + const Colour c1 = (this->LookupColourInPalette(m1).data & 0x00FFFFFF) | (src[1].data & 0xFF000000); + INSR32(AdjustBrightness(c1, (byte) (mvX2 >> 24)).data, srcABCD, 1); + } - /* Update anim buffer. */ - const byte a0 = src[0].a; - const byte a1 = src[1].a; - uint32 anim01 = 0; - if (a0 == 255) { - if (a1 == 255) { - *(uint32*) anim = mvX2; - goto bmno_full_opacity; - } - anim01 = (uint16) mvX2; - } else if (a0 == 0) { - if (a1 == 0) { - goto bmno_full_transparency; - } else { - if (a1 == 255) anim[1] = (uint16) (mvX2 >> 16); - goto bmno_alpha_blend; - } - } - if (a1 > 0) { - if (a1 == 255) anim01 |= mvX2 & 0xFFFF0000; - *(uint32*) anim = anim01; - } else { - anim[0] = (uint16) anim01; - } + /* Update anim buffer. */ + const byte a0 = src[0].a; + const byte a1 = src[1].a; + uint32 anim01 = 0; + if (a0 == 255) { + if (a1 == 255) { + *(uint32*) anim = mvX2; + goto bmno_full_opacity; + } + anim01 = (uint16) mvX2; + } else if (a0 == 0) { + if (a1 == 0) { + goto bmno_full_transparency; + } else { + if (a1 == 255) anim[1] = (uint16) (mvX2 >> 16); + goto bmno_alpha_blend; + } + } + if (a1 > 0) { + if (a1 == 255) anim01 |= mvX2 & 0xFFFF0000; + *(uint32*) anim = anim01; + } else { + anim[0] = (uint16) anim01; + } - /* Blend colours. */ + /* Blend colours. */ bmno_alpha_blend: - ALPHA_BLEND_2(); + ALPHA_BLEND_2(); bmno_full_opacity: - _mm_storel_epi64((__m128i *) dst, srcABCD); + _mm_storel_epi64((__m128i *) dst, srcABCD); bmno_full_transparency: - src_mv += 2; - src += 2; - anim += 2; - dst += 2; - } + src_mv += 2; + src += 2; + anim += 2; + dst += 2; + } - if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { - if (src->a == 0) { - } else if (src->a == 255) { - *anim = *(const uint16*) src_mv; - *dst = (src_mv->m >= PALETTE_ANIM_START) ? AdjustBrightness(LookupColourInPalette(src_mv->m), src_mv->v) : *src; - } else { - *anim = 0; - __m128i srcABCD; - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - if (src_mv->m >= PALETTE_ANIM_START) { - Colour colour = AdjustBrightness(LookupColourInPalette(src_mv->m), src_mv->v); - colour.a = src->a; - srcABCD = _mm_cvtsi32_si128(colour.data); - } else { - srcABCD = _mm_cvtsi32_si128(src->data); - } - ALPHA_BLEND_2(); - dst->data = _mm_cvtsi128_si32(srcABCD); - } + if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { + if (src->a == 0) { + } else if (src->a == 255) { + *anim = *(const uint16*) src_mv; + *dst = (src_mv->m >= PALETTE_ANIM_START) ? AdjustBrightness(LookupColourInPalette(src_mv->m), src_mv->v) : *src; + } else { + *anim = 0; + __m128i srcABCD; + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + if (src_mv->m >= PALETTE_ANIM_START) { + Colour colour = AdjustBrightness(LookupColourInPalette(src_mv->m), src_mv->v); + colour.a = src->a; + srcABCD = _mm_cvtsi32_si128(colour.data); + } else { + srcABCD = _mm_cvtsi32_si128(src->data); } - break; + ALPHA_BLEND_2(); + dst->data = _mm_cvtsi128_si32(srcABCD); } - - 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; - anim += 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; - effective_width = delta_diff > 0 ? new_width : effective_width; - if (effective_width <= 0) break; - /* FALLTHROUGH */ - } - case RM_WITH_SKIP: { - for (uint x = (uint) effective_width / 2; x != 0; x--) { - uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); - __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); - __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - - /* Remap colours. */ - const uint m0 = (byte) mvX2; - const uint r0 = remap[m0]; - const uint m1 = (byte) (mvX2 >> 16); - const uint r1 = remap[m1]; - if (mvX2 & 0x00FF00FF) { - /* Written so the compiler uses CMOV. */ - const Colour src0 = src[0]; - const Colour c0map = (this->LookupColourInPalette(r0).data & 0x00FFFFFF) | (src0.data & 0xFF000000); - Colour c0 = dst[0]; - c0 = r0 == 0 ? c0 : c0map; - c0 = m0 != 0 ? c0 : src0; - INSR32(c0.data, srcABCD, 0); - - const Colour src1 = src[1]; - const Colour c1map = (this->LookupColourInPalette(r1).data & 0x00FFFFFF) | (src1.data & 0xFF000000); - Colour c1 = dst[1]; - c1 = r1 == 0 ? c1 : c1map; - c1 = m1 != 0 ? c1 : src1; - INSR32(c1.data, srcABCD, 1); + case BM_COLOUR_REMAP: + for (uint x = (uint) effective_width / 2; x != 0; x--) { + uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); + __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); + __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - if ((mvX2 & 0xFF00FF00) != 0x80008000) { - ADJUST_BRIGHTNESS_2(srcABCD, mvX2); - } - } + /* Remap colours. */ + const uint m0 = (byte) mvX2; + const uint r0 = remap[m0]; + const uint m1 = (byte) (mvX2 >> 16); + const uint r1 = remap[m1]; + if (mvX2 & 0x00FF00FF) { + /* Written so the compiler uses CMOV. */ + const Colour src0 = src[0]; + const Colour c0map = (this->LookupColourInPalette(r0).data & 0x00FFFFFF) | (src0.data & 0xFF000000); + Colour c0 = dst[0]; + c0 = r0 == 0 ? c0 : c0map; + c0 = m0 != 0 ? c0 : src0; + INSR32(c0.data, srcABCD, 0); + + const Colour src1 = src[1]; + const Colour c1map = (this->LookupColourInPalette(r1).data & 0x00FFFFFF) | (src1.data & 0xFF000000); + Colour c1 = dst[1]; + c1 = r1 == 0 ? c1 : c1map; + c1 = m1 != 0 ? c1 : src1; + INSR32(c1.data, srcABCD, 1); + + if ((mvX2 & 0xFF00FF00) != 0x80008000) { + ADJUST_BRIGHTNESS_2(srcABCD, mvX2); + } + } - /* Update anim buffer. */ - const byte a0 = src[0].a; - const byte a1 = src[1].a; - uint32 anim01 = mvX2 & 0xFF00FF00; - if (a0 == 255) { - anim01 |= r0; - if (a1 == 255) { - *(uint32*) anim = anim01 | (r1 << 16); - goto bmcr_full_opacity; - } - } else if (a0 == 0) { - if (a1 == 0) { - goto bmcr_full_transparency; - } else { - if (a1 == 255) { - anim[1] = r1 | (anim01 >> 16); - } - goto bmcr_alpha_blend; - } - } - if (a1 > 0) { - if (a1 == 255) anim01 |= r1 << 16; - *(uint32*) anim = anim01; - } else { - anim[0] = (uint16) anim01; + /* Update anim buffer. */ + const byte a0 = src[0].a; + const byte a1 = src[1].a; + uint32 anim01 = mvX2 & 0xFF00FF00; + if (a0 == 255) { + anim01 |= r0; + if (a1 == 255) { + *(uint32*) anim = anim01 | (r1 << 16); + goto bmcr_full_opacity; + } + } else if (a0 == 0) { + if (a1 == 0) { + goto bmcr_full_transparency; + } else { + if (a1 == 255) { + anim[1] = r1 | (anim01 >> 16); } + goto bmcr_alpha_blend; + } + } + if (a1 > 0) { + if (a1 == 255) anim01 |= r1 << 16; + *(uint32*) anim = anim01; + } else { + anim[0] = (uint16) anim01; + } - /* Blend colours. */ + /* Blend colours. */ bmcr_alpha_blend: - ALPHA_BLEND_2(); + ALPHA_BLEND_2(); bmcr_full_opacity: - _mm_storel_epi64((__m128i *) dst, srcABCD); + _mm_storel_epi64((__m128i *) dst, srcABCD); bmcr_full_transparency: - src_mv += 2; - dst += 2; - src += 2; - anim += 2; - } + src_mv += 2; + dst += 2; + src += 2; + anim += 2; + } - if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { - /* In case the m-channel is zero, do not remap this pixel in any way. */ - __m128i srcABCD; - if (src->a == 0) break; - if (src_mv->m) { - const uint r = remap[src_mv->m]; - *anim = (src->a == 255) ? r | ((uint16) src_mv->v << 8 ) : 0; - if (r != 0) { - Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); - if (src->a == 255) { - *dst = remapped_colour; - } else { - remapped_colour.a = src->a; - srcABCD = _mm_cvtsi32_si128(remapped_colour.data); - goto bmcr_alpha_blend_single; - } - } + if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { + /* In case the m-channel is zero, do not remap this pixel in any way. */ + __m128i srcABCD; + if (src->a == 0) break; + if (src_mv->m) { + const uint r = remap[src_mv->m]; + *anim = (src->a == 255) ? r | ((uint16) src_mv->v << 8 ) : 0; + if (r != 0) { + Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); + if (src->a == 255) { + *dst = remapped_colour; } else { - *anim = 0; - srcABCD = _mm_cvtsi32_si128(src->data); - if (src->a < 255) { -bmcr_alpha_blend_single: - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - ALPHA_BLEND_2(); - } - dst->data = _mm_cvtsi128_si32(srcABCD); + remapped_colour.a = src->a; + srcABCD = _mm_cvtsi32_si128(remapped_colour.data); + goto bmcr_alpha_blend_single; } } - break; + } else { + *anim = 0; + srcABCD = _mm_cvtsi32_si128(src->data); + if (src->a < 255) { +bmcr_alpha_blend_single: + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + ALPHA_BLEND_2(); + } + dst->data = _mm_cvtsi128_si32(srcABCD); } - - default: NOT_REACHED(); } break; - } - case BM_TRANSPARENT: { + case BM_TRANSPARENT: /* Make the current colour a bit more black, so it looks like this image is transparent. */ for (uint x = (uint) bp->width / 2; x > 0; x--) { __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); @@ -298,6 +268,7 @@ bmcr_alpha_blend_single: if (src[-2].a) anim[-2] = 0; if (src[-1].a) anim[-1] = 0; } + if ((bt_last == BT_NONE && bp->width & 1) || bt_last == BT_ODD) { __m128i srcABCD = _mm_cvtsi32_si128(src->data); __m128i dstABCD = _mm_cvtsi32_si128(dst->data); @@ -306,10 +277,10 @@ bmcr_alpha_blend_single: if (src[0].a) anim[0] = 0; } break; - } } - src_mv_line += si->sprite_width; +next_line: + if (mode != BM_TRANSPARENT) src_mv_line += si->sprite_width; src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size); dst_line += bp->pitch; anim_line += this->anim_buf_width; @@ -328,19 +299,15 @@ void Blitter_32bppSSE4_Anim::Draw(Blitter::BlitterParams *bp, BlitterMode mode, { switch (mode) { case BM_NORMAL: { - const BlockType bt_last = (BlockType) (bp->width & 1); if (bp->skip_left != 0 || bp->width <= MARGIN_NORMAL_THRESHOLD) { + const BlockType bt_last = (BlockType) (bp->width & 1); 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(); - } + Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE>(bp, zoom); return; } break; } diff --git a/src/blitter/32bpp_sse2.cpp b/src/blitter/32bpp_sse2.cpp index a5d8a1987..ddaa352e2 100644 --- a/src/blitter/32bpp_sse2.cpp +++ b/src/blitter/32bpp_sse2.cpp @@ -30,6 +30,7 @@ IGNORE_UNINITIALIZED_WARNING_START template <BlitterMode mode, Blitter_32bppSSE2::ReadMode read_mode, Blitter_32bppSSE2::BlockType bt_last> inline void Blitter_32bppSSE2::Draw(const Blitter::BlitterParams *bp, ZoomLevel zoom) { + const byte *remap = bp->remap; Colour *dst_line = (Colour *) bp->dst + bp->top * bp->pitch + bp->left; int effective_width = bp->width; @@ -43,6 +44,7 @@ inline void Blitter_32bppSSE2::Draw(const Blitter::BlitterParams *bp, ZoomLevel src_rgba_line += bp->skip_left; src_mv_line += bp->skip_left; } + const MapValue *src_mv = src_mv_line; /* Load these variables into register before loop. */ const __m128i clear_hi = CLEAR_HIGH_BYTE_MASK; @@ -51,99 +53,71 @@ inline void Blitter_32bppSSE2::Draw(const Blitter::BlitterParams *bp, ZoomLevel for (int y = bp->height; y != 0; y--) { Colour *dst = dst_line; const Colour *src = src_rgba_line + META_LENGTH; - const MapValue *src_mv = src_mv_line; + if (mode == BM_COLOUR_REMAP) src_mv = src_mv_line; + + if (read_mode == RM_WITH_MARGIN) { + src += src_rgba_line[0].data; + dst += src_rgba_line[0].data; + if (mode == BM_COLOUR_REMAP) src_mv += 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; + effective_width = delta_diff > 0 ? new_width : effective_width; + if (effective_width <= 0) goto next_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: { - for (uint x = (uint) effective_width / 2; x > 0; x--) { - __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); - __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - ALPHA_BLEND_2(); - _mm_storel_epi64((__m128i*) dst, srcABCD); - src += 2; - dst += 2; - } - if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { - __m128i srcABCD = _mm_cvtsi32_si128(src->data); - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - ALPHA_BLEND_2(); - dst->data = _mm_cvtsi128_si32(srcABCD); - } - break; - } + default: + for (uint x = (uint) effective_width / 2; x > 0; x--) { + __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); + __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); + ALPHA_BLEND_2(); + _mm_storel_epi64((__m128i*) dst, srcABCD); + src += 2; + dst += 2; + } - default: NOT_REACHED(); + if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { + __m128i srcABCD = _mm_cvtsi32_si128(src->data); + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + ALPHA_BLEND_2(); + dst->data = _mm_cvtsi128_si32(srcABCD); } 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 new_width = effective_width - delta_diff; - effective_width = delta_diff > 0 ? new_width : effective_width; - if (effective_width <= 0) break; - /* FALLTHROUGH */ - } - case RM_WITH_SKIP: { - const byte *remap = bp->remap; - for (uint x = (uint) effective_width; x != 0; x--) { - /* In case the m-channel is zero, do not remap this pixel in any way. */ - __m128i srcABCD; - if (src_mv->m) { - const uint r = remap[src_mv->m]; - if (r != 0) { - Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); - if (src->a == 255) { - *dst = remapped_colour; - } else { - remapped_colour.a = src->a; - srcABCD = _mm_cvtsi32_si128(remapped_colour.data); - goto bmcr_alpha_blend_single; - } - } + case BM_COLOUR_REMAP: + for (uint x = (uint) effective_width; x != 0; x--) { + /* In case the m-channel is zero, do not remap this pixel in any way. */ + __m128i srcABCD; + if (src_mv->m) { + const uint r = remap[src_mv->m]; + if (r != 0) { + Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); + if (src->a == 255) { + *dst = remapped_colour; } else { - srcABCD = _mm_cvtsi32_si128(src->data); - if (src->a < 255) { -bmcr_alpha_blend_single: - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - ALPHA_BLEND_2(); - } - dst->data = _mm_cvtsi128_si32(srcABCD); + remapped_colour.a = src->a; + srcABCD = _mm_cvtsi32_si128(remapped_colour.data); + goto bmcr_alpha_blend_single; } - src_mv++; - dst++; - src++; } - break; + } else { + srcABCD = _mm_cvtsi32_si128(src->data); + if (src->a < 255) { +bmcr_alpha_blend_single: + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + ALPHA_BLEND_2(); + } + dst->data = _mm_cvtsi128_si32(srcABCD); } - - default: NOT_REACHED(); + src_mv++; + dst++; + src++; } - src_mv_line += si->sprite_width; break; - } - case BM_TRANSPARENT: { + + case BM_TRANSPARENT: /* Make the current colour a bit more black, so it looks like this image is transparent. */ for (uint x = (uint) bp->width / 2; x > 0; x--) { __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); @@ -153,6 +127,7 @@ bmcr_alpha_blend_single: src += 2; dst += 2; } + if ((bt_last == BT_NONE && bp->width & 1) || bt_last == BT_ODD) { __m128i srcABCD = _mm_cvtsi32_si128(src->data); __m128i dstABCD = _mm_cvtsi32_si128(dst->data); @@ -160,9 +135,10 @@ bmcr_alpha_blend_single: dst->data = _mm_cvtsi128_si32(dstAB); } break; - } } +next_line: + if (mode == BM_COLOUR_REMAP) src_mv_line += si->sprite_width; src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size); dst_line += bp->pitch; } @@ -180,19 +156,15 @@ void Blitter_32bppSSE2::Draw(Blitter::BlitterParams *bp, BlitterMode mode, ZoomL { switch (mode) { case BM_NORMAL: { - const BlockType bt_last = (BlockType) (bp->width & 1); if (bp->skip_left != 0 || bp->width <= MARGIN_NORMAL_THRESHOLD) { + const BlockType bt_last = (BlockType) (bp->width & 1); 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(); - } + Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE>(bp, zoom); return; } break; } diff --git a/src/blitter/32bpp_sse4.cpp b/src/blitter/32bpp_sse4.cpp index d43cd8cd9..6bdb9e950 100644 --- a/src/blitter/32bpp_sse4.cpp +++ b/src/blitter/32bpp_sse4.cpp @@ -44,6 +44,7 @@ inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel src_rgba_line += bp->skip_left; src_mv_line += bp->skip_left; } + const MapValue *src_mv = src_mv_line; /* Load these variables into register before loop. */ const __m128i a_cm = ALPHA_CONTROL_MASK; @@ -58,136 +59,107 @@ inline void Blitter_32bppSSE4::Draw(const Blitter::BlitterParams *bp, ZoomLevel 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; + if (mode == BM_COLOUR_REMAP) src_mv = src_mv_line; + + if (read_mode == RM_WITH_MARGIN) { + src += src_rgba_line[0].data; + dst += src_rgba_line[0].data; + if (mode == BM_COLOUR_REMAP) src_mv += 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; + effective_width = delta_diff > 0 ? new_width : effective_width; + if (effective_width <= 0) goto next_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: { - for (uint x = (uint) effective_width / 2; x > 0; x--) { - __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); - __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - ALPHA_BLEND_2(); - _mm_storel_epi64((__m128i*) dst, srcABCD); - src += 2; - dst += 2; - } - if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { - __m128i srcABCD = _mm_cvtsi32_si128(src->data); - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - ALPHA_BLEND_2(); - dst->data = _mm_cvtsi128_si32(srcABCD); - } - break; - } + default: + for (uint x = (uint) effective_width / 2; x > 0; x--) { + __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); + __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); + ALPHA_BLEND_2(); + _mm_storel_epi64((__m128i*) dst, srcABCD); + src += 2; + dst += 2; + } - default: NOT_REACHED(); + if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { + __m128i srcABCD = _mm_cvtsi32_si128(src->data); + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + ALPHA_BLEND_2(); + dst->data = _mm_cvtsi128_si32(srcABCD); } 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 new_width = effective_width - delta_diff; - effective_width = delta_diff > 0 ? new_width : effective_width; - if (effective_width <= 0) break; - /* FALLTHROUGH */ + case BM_COLOUR_REMAP: + for (uint x = (uint) effective_width / 2; x > 0; x--) { + __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); + __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); + uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); + + /* 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); + } } - case RM_WITH_SKIP: { - for (uint x = (uint) effective_width / 2; x > 0; x--) { - __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); - __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); - - /* 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); + /* Blend colours. */ + ALPHA_BLEND_2(); + _mm_storel_epi64((__m128i *) dst, srcABCD); + dst += 2; + src += 2; + src_mv += 2; + } - if ((mvX2 & 0xFF00FF00) != 0x80008000) { - ADJUST_BRIGHTNESS_2(srcABCD, mvX2); - } + if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { + /* In case the m-channel is zero, do not remap this pixel in any way. */ + __m128i srcABCD; + if (src_mv->m) { + const uint r = remap[src_mv->m]; + if (r != 0) { + Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); + if (src->a == 255) { + *dst = remapped_colour; + } else { + remapped_colour.a = src->a; + srcABCD = _mm_cvtsi32_si128(remapped_colour.data); + goto bmcr_alpha_blend_single; } - - /* Blend colours. */ - ALPHA_BLEND_2(); - _mm_storel_epi64((__m128i *) dst, srcABCD); - dst += 2; - src += 2; - src_mv += 2; } - - if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { - /* In case the m-channel is zero, do not remap this pixel in any way. */ - __m128i srcABCD; - if (src_mv->m) { - const uint r = remap[src_mv->m]; - if (r != 0) { - Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); - if (src->a == 255) { - *dst = remapped_colour; - } else { - remapped_colour.a = src->a; - srcABCD = _mm_cvtsi32_si128(remapped_colour.data); - goto bmcr_alpha_blend_single; - } - } - } else { - srcABCD = _mm_cvtsi32_si128(src->data); - if (src->a < 255) { + } else { + srcABCD = _mm_cvtsi32_si128(src->data); + if (src->a < 255) { bmcr_alpha_blend_single: - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - ALPHA_BLEND_2(); - } - dst->data = _mm_cvtsi128_si32(srcABCD); - } + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + ALPHA_BLEND_2(); } - break; + dst->data = _mm_cvtsi128_si32(srcABCD); } - - default: NOT_REACHED(); } - src_mv_line += si->sprite_width; break; - } - case BM_TRANSPARENT: { + case BM_TRANSPARENT: /* Make the current colour a bit more black, so it looks like this image is transparent. */ for (uint x = (uint) bp->width / 2; x > 0; x--) { __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); @@ -197,17 +169,18 @@ bmcr_alpha_blend_single: src += 2; dst += 2; } + if ((bt_last == BT_NONE && bp->width & 1) || bt_last == BT_ODD) { __m128i srcABCD = _mm_cvtsi32_si128(src->data); __m128i dstABCD = _mm_cvtsi32_si128(dst->data); DARKEN_2(); dst->data = _mm_cvtsi128_si32(dstAB); } - break; - } } +next_line: + if (mode == BM_COLOUR_REMAP) src_mv_line += si->sprite_width; src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size); dst_line += bp->pitch; } @@ -223,21 +196,17 @@ IGNORE_UNINITIALIZED_WARNING_STOP */ 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) { + const BlockType bt_last = (BlockType) (bp->width & 1); 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(); - } + Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE>(bp, zoom); return; } break; } diff --git a/src/blitter/32bpp_ssse3.cpp b/src/blitter/32bpp_ssse3.cpp index 32a981b1b..32f073d45 100644 --- a/src/blitter/32bpp_ssse3.cpp +++ b/src/blitter/32bpp_ssse3.cpp @@ -44,6 +44,7 @@ inline void Blitter_32bppSSSE3::Draw(const Blitter::BlitterParams *bp, ZoomLevel src_rgba_line += bp->skip_left; src_mv_line += bp->skip_left; } + const MapValue *src_mv = src_mv_line; /* Load these variables into register before loop. */ const __m128i a_cm = ALPHA_CONTROL_MASK; @@ -58,136 +59,107 @@ inline void Blitter_32bppSSSE3::Draw(const Blitter::BlitterParams *bp, ZoomLevel for (int y = bp->height; y != 0; y--) { Colour *dst = dst_line; const Colour *src = src_rgba_line + META_LENGTH; - const MapValue *src_mv = src_mv_line; + if (mode == BM_COLOUR_REMAP) src_mv = src_mv_line; + + if (read_mode == RM_WITH_MARGIN) { + src += src_rgba_line[0].data; + dst += src_rgba_line[0].data; + if (mode == BM_COLOUR_REMAP) src_mv += 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; + effective_width = delta_diff > 0 ? new_width : effective_width; + if (effective_width <= 0) goto next_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: { - for (uint x = (uint) effective_width / 2; x > 0; x--) { - __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); - __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - ALPHA_BLEND_2(); - _mm_storel_epi64((__m128i*) dst, srcABCD); - src += 2; - dst += 2; - } - if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { - __m128i srcABCD = _mm_cvtsi32_si128(src->data); - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - ALPHA_BLEND_2(); - dst->data = _mm_cvtsi128_si32(srcABCD); - } - break; - } + default: + for (uint x = (uint) effective_width / 2; x > 0; x--) { + __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); + __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); + ALPHA_BLEND_2(); + _mm_storel_epi64((__m128i*) dst, srcABCD); + src += 2; + dst += 2; + } - default: NOT_REACHED(); + if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { + __m128i srcABCD = _mm_cvtsi32_si128(src->data); + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + ALPHA_BLEND_2(); + dst->data = _mm_cvtsi128_si32(srcABCD); } 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 new_width = effective_width - delta_diff; - effective_width = delta_diff > 0 ? new_width : effective_width; - if (effective_width <= 0) break; - /* FALLTHROUGH */ + case BM_COLOUR_REMAP: + for (uint x = (uint) effective_width / 2; x > 0; x--) { + __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); + __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); + uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); + + /* 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; + srcABCD = _mm_cvtsi32_si128(c0.data); + + 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); + } } - case RM_WITH_SKIP: { - for (uint x = (uint) effective_width / 2; x > 0; x--) { - __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); - __m128i dstABCD = _mm_loadl_epi64((__m128i*) dst); - uint32 mvX2 = *((uint32 *) const_cast<MapValue *>(src_mv)); - - /* 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; - srcABCD = _mm_cvtsi32_si128(c0.data); - - 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(); - _mm_storel_epi64((__m128i *) dst, srcABCD); - dst += 2; - src += 2; - src_mv += 2; - } + /* Blend colours. */ + ALPHA_BLEND_2(); + _mm_storel_epi64((__m128i *) dst, srcABCD); + dst += 2; + src += 2; + src_mv += 2; + } - if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { - /* In case the m-channel is zero, do not remap this pixel in any way. */ - __m128i srcABCD; - if (src_mv->m) { - const uint r = remap[src_mv->m]; - if (r != 0) { - Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); - if (src->a == 255) { - *dst = remapped_colour; - } else { - remapped_colour.a = src->a; - srcABCD = _mm_cvtsi32_si128(remapped_colour.data); - goto bmcr_alpha_blend_single; - } - } + if ((bt_last == BT_NONE && effective_width & 1) || bt_last == BT_ODD) { + /* In case the m-channel is zero, do not remap this pixel in any way. */ + __m128i srcABCD; + if (src_mv->m) { + const uint r = remap[src_mv->m]; + if (r != 0) { + Colour remapped_colour = AdjustBrightness(this->LookupColourInPalette(r), src_mv->v); + if (src->a == 255) { + *dst = remapped_colour; } else { - srcABCD = _mm_cvtsi32_si128(src->data); - if (src->a < 255) { -bmcr_alpha_blend_single: - __m128i dstABCD = _mm_cvtsi32_si128(dst->data); - ALPHA_BLEND_2(); - } - dst->data = _mm_cvtsi128_si32(srcABCD); + remapped_colour.a = src->a; + srcABCD = _mm_cvtsi32_si128(remapped_colour.data); + goto bmcr_alpha_blend_single; } } - break; + } else { + srcABCD = _mm_cvtsi32_si128(src->data); + if (src->a < 255) { +bmcr_alpha_blend_single: + __m128i dstABCD = _mm_cvtsi32_si128(dst->data); + ALPHA_BLEND_2(); + } + dst->data = _mm_cvtsi128_si32(srcABCD); } - - default: NOT_REACHED(); } - src_mv_line += si->sprite_width; break; - } - case BM_TRANSPARENT: { + case BM_TRANSPARENT: /* Make the current colour a bit more black, so it looks like this image is transparent. */ for (uint x = (uint) bp->width / 2; x > 0; x--) { __m128i srcABCD = _mm_loadl_epi64((const __m128i*) src); @@ -197,6 +169,7 @@ bmcr_alpha_blend_single: src += 2; dst += 2; } + if ((bt_last == BT_NONE && bp->width & 1) || bt_last == BT_ODD) { __m128i srcABCD = _mm_cvtsi32_si128(src->data); __m128i dstABCD = _mm_cvtsi32_si128(dst->data); @@ -204,9 +177,10 @@ bmcr_alpha_blend_single: dst->data = _mm_cvtsi128_si32(dstAB); } break; - } } +next_line: + if (mode == BM_COLOUR_REMAP) src_mv_line += si->sprite_width; src_rgba_line = (const Colour*) ((const byte*) src_rgba_line + si->sprite_line_size); dst_line += bp->pitch; } @@ -224,19 +198,15 @@ void Blitter_32bppSSSE3::Draw(Blitter::BlitterParams *bp, BlitterMode mode, Zoom { switch (mode) { case BM_NORMAL: { - const BlockType bt_last = (BlockType) (bp->width & 1); if (bp->skip_left != 0 || bp->width <= MARGIN_NORMAL_THRESHOLD) { + const BlockType bt_last = (BlockType) (bp->width & 1); 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(); - } + Draw<BM_NORMAL, RM_WITH_MARGIN, BT_NONE>(bp, zoom); return; } break; } |