X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=Pico%2FDraw.s;h=3a3da814d7753e105d274d853a60e59ec4046ab4;hb=03e4f2a349247334666c87abe3a908df72d23051;hp=cb1a6fa7964d44cf064bb5711a69e324d11b5ac5;hpb=6d7acf9eff33cdde5e3eac44a193448ac0cbf541;p=picodrive.git diff --git a/Pico/Draw.s b/Pico/Draw.s index cb1a6fa..3a3da81 100644 --- a/Pico/Draw.s +++ b/Pico/Draw.s @@ -6,6 +6,7 @@ @ (c) Copyright 2007, Grazvydas "notaz" Ignotas @ All Rights Reserved +.include "port_config.s" .extern Pico .extern PicoOpt @@ -69,7 +70,7 @@ streqb r4, [r1,#\offs] .endm -@ TileNorm (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: register with helper pattern 0xf, touches r3 high bits +@ TileNormShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: register with helper pattern 0xf, touches r3 high bits .macro TileNormShHP TilePixelShHP 12, 0 @ #0x0000f000 TilePixelShHP 8, 1 @ #0x00000f00 @@ -81,7 +82,7 @@ TilePixelShHP 16, 7 @ #0x000f0000 .endm -@ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf +@ TileFlipShHP (r1=pdest, r2=pixels8, r3=pal) r4: scratch, pat: register with helper pattern 0xf .macro TileFlipShHP TilePixelShHP 16, 0 @ #0x000f0000 TilePixelShHP 20, 1 @ #0x00f00000 @@ -211,9 +212,9 @@ @ int cells; // 0x14 @ }; -@ int DrawLayer(int plane, int *hcache, int maxcells, int sh) +@ void DrawLayer(int plane_sh, int *hcache, int cellskip, int maxcells); -.global DrawLayer @ int plane, int *hcache, int maxcells, int sh +.global DrawLayer DrawLayer: stmfd sp!, {r4-r11,lr} @@ -221,10 +222,11 @@ DrawLayer: ldr r11, =(Pico+0x22228) @ Pico.video mov r8, #1 - ldrb r7, [r11, #16] @ ??hh??ww + ldrb r7, [r11, #16] @ ??vv??hh mov r6, r1 @ hcache - orr r9, r2, r3, lsl #31 @ r9=maxcells|(sh<<31) + orr r9, r3, r0, lsl #30 + orr r9, r9, r2, lsl #8 @ r9=sh[31]|cellskip[15:8]|maxcells[7:0] (tmp) mov r1, r7, lsl #4 orr r1, r1, #0x00ff @@ -244,7 +246,7 @@ DrawLayer: sub r5, r5, #1 @ r5=xmask @ Find name table: - tst r0, r0 + ands r0, r0, #1 ldreqb r12, [r11, #2] ldrneb r12, [r11, #4] @@ -306,17 +308,28 @@ DrawLayer: tst r9, #1<<31 mov r3, #0 - orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|ty) + orrne r10,r10, #1<<23 @ r10=(cells<<24|sh<<23|hi_not_empty<<22|had_output<<21|ty) movne r3, #0x40 @ default to shadowed pal on sh mode - mvn r9, #0 @ r9=prevcode=-1 - cmp r7, #8 addne r10,r10, #0x01000000 @ we will loop cells+1 times if there is scroll + and r9, r9, #0xff00 + add r8, r8, r9, lsr #8 @ tilex+=cellskip + add r7, r7, r9, lsr #5 @ dx+=cellskip<<3; + sub r10,r10,r9, lsl #16 @ cells-=cellskip + @ cache some stuff to avoid mem access +.if OVERRIDE_HIGHCOL + ldr r11,=HighCol + mov r0, #0xf + ldr r11,[r11] +.else ldr r11,=HighCol mov r0, #0xf +.endif + + mvn r9, #0 @ r9=prevcode=-1 add r1, r11, r7 @ r1=pdest @@ -342,6 +355,7 @@ DrawLayer: beq .DrawStrip_samecode @ we know stuff about this tile already mov r9, r7 @ remember code + orr r10, r10, #1<<21 @ seen non hi-prio tile movs r2, r9, lsl #20 @ if (code&0x1000) mov r2, r2, lsl #1 @@ -363,16 +377,17 @@ DrawLayer: beq .DrawStrip_SingleColor @ tileline singlecolor tst r9, #0x0800 - beq .DrawStrip_TileNorm + bne .DrawStrip_TileFlip @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern - TileFlip r0 - b .dsloop - .DrawStrip_TileNorm: TileNorm r0 b .dsloop +.DrawStrip_TileFlip: + TileFlip r0 + b .dsloop + .DrawStrip_SingleColor: and r4, r2, #0xf orr r4, r3, r4 @@ -386,18 +401,6 @@ DrawLayer: strneb r4, [r1], #1 @ have a remaining unaligned pixel? b .dsloop_subr1 -.DrawStrip_hiprio: - tst r10, #0x00c00000 - beq .DrawStrip_hiprio_maybempt - sub r0, r1, r11 - orr r7, r7, r0, lsl #16 - orr r7, r7, r10, lsl #25 @ (ty<<25) - tst r7, #0x1000 - eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26; - str r7, [r6], #4 @ cache hi priority tile - mov r0, #0xf - b .dsloop - .DrawStrip_hiprio_maybempt: cmp r7, r9 beq .dsloop @ must've been empty, otherwise we wouldn't get here @@ -409,13 +412,29 @@ DrawLayer: ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels mov r9, r7 @ remember code tst r2, r2 - orrne r10, r10, #1<<22 - bne .DrawStrip_hiprio + beq .dsloop + orr r10, r10, #1<<22 + +.DrawStrip_hiprio: + tst r10, #0x00c00000 + beq .DrawStrip_hiprio_maybempt + sub r0, r1, r11 + orr r7, r7, r0, lsl #16 + orr r7, r7, r10, lsl #25 @ (ty<<25) + tst r7, #0x1000 + eorne r7, r7, #7<<26 @ if(code&0x1000) cval^=7<<26; + str r7, [r6], #4 @ cache hi priority tile + mov r0, #0xf b .dsloop .dsloop_exit: + tst r10, #1<<21 @ seen non hi-prio tile + ldreq r1, =rendstatus mov r0, #0 + ldreq r2, [r1] str r0, [r6] @ terminate the cache list + orreq r2, r2, #0x40 @ had a layer with all hi-prio tiles + streq r2, [r1] ldmfd sp!, {r4-r11,lr} bx lr @@ -426,7 +445,7 @@ DrawLayer: rsb r8, r3, #0 mov r8, r8, lsr #3 @ r8=tilex=(-ts->hscroll)>>3 bic r8, r8, #0xff000000 - orr r8, r8, r5, lsl #25 @ r8=(xmask[31:25]|tilex[15:0]) + orr r8, r8, r5, lsl #25 @ r8=(xmask[31:25]|had_output[24]|tilex[23:0]) ldr r4, =Scanline orr r5, r1, r10, lsl #24 @@ -437,24 +456,34 @@ DrawLayer: add r7, r1, #1 @ r7=dx=((ts->hscroll-1)&7)+1 mov r10,r9, lsl #16 - tst r0, r0 + tst r0, #1 orrne r10,r10, #0x8000 tst r9, #1<<31 mov r3, #0 orr r10,r10, #0xff000000 @ will be adjusted on entering loop - orrne r10,r10, #1<<23 @ r10=(cells[31:24]|sh[23]|hi_not_empty[22]|cells_max[21:16]|plane[15]|ty[14:0]) + orrne r10,r10, #1<<23 @ r10=(cell[31:24]|sh[23]|hi_not_empty[22]|cells_max[21:16]|plane[15]|ty[14:0]) movne r3, #0x40 @ default to shadowed pal on sh mode - mvn r9, #0 @ r9=prevcode=-1 + cmp r7, #8 + subne r10,r10, #0x01000000 @ have hscroll, start with negative cell + + and r9, r9, #0xff00 + add r8, r8, r9, lsr #8 @ tilex+=cellskip + add r7, r7, r9, lsr #5 @ dx+=cellskip<<3; + add r10,r10,r9, lsl #16 @ cell+=cellskip @ cache some stuff to avoid mem access +.if OVERRIDE_HIGHCOL ldr r11,=HighCol mov r0, #0xf - add r1, r11, r7 @ r1=pdest - - cmp r7, #8 - subne r10,r10, #0x01000000 @ have hscroll, start with negative cell + ldr r11,[r11] +.else + ldr r11,=HighCol + mov r0, #0xf +.endif + mvn r9, #0 @ r9=prevcode=-1 + add r1, r11, r7 @ r1=pdest @ r4 & r7 are scratch in this loop .dsloop_vs_subr1: @@ -463,7 +492,7 @@ DrawLayer: add r10,r10, #0x01000000 and r4, r10, #0x003f0000 cmp r4, r10, asr #8 - ble .dsloop_exit + ble .dsloop_vs_exit @ calc offset and read tileline code to r7, also calc ty add r7, lr, #0x012000 @@ -500,6 +529,7 @@ DrawLayer: beq .DrawStrip_vs_samecode @ we know stuff about this tile already mov r9, r7 @ remember code + orr r8, r8, #1<<24 @ seen non hi-prio tile movs r2, r9, lsl #20 @ if (code&0x1000) mov r2, r2, lsl #1 @@ -521,16 +551,17 @@ DrawLayer: beq .DrawStrip_vs_SingleColor @ tileline singlecolor tst r9, #0x0800 - beq .DrawStrip_vs_TileNorm + bne .DrawStrip_vs_TileFlip @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern - TileFlip r0 - b .dsloop_vs - .DrawStrip_vs_TileNorm: TileNorm r0 b .dsloop_vs +.DrawStrip_vs_TileFlip: + TileFlip r0 + b .dsloop_vs + .DrawStrip_vs_SingleColor: and r4, r2, #0xf orr r4, r3, r4 @@ -571,6 +602,18 @@ DrawLayer: bne .DrawStrip_vs_hiprio b .dsloop_vs +.dsloop_vs_exit: + tst r8, #1<<24 @ seen non hi-prio tile + ldreq r1, =rendstatus + mov r0, #0 + ldreq r2, [r1] + str r0, [r6] @ terminate the cache list + orreq r2, r2, #0x40 @ had a layer with all hi-prio tiles + streq r2, [r1] + + ldmfd sp!, {r4-r11,lr} + bx lr + @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ @@ -610,10 +653,18 @@ DrawLayer: BackFill: stmfd sp!, {r4-r9,lr} +.if OVERRIDE_HIGHCOL + ldr lr, =HighCol + mov r0, r0, lsl #26 + ldr lr, [lr] + mov r0, r0, lsr #26 + add lr, lr, #8 +.else ldr lr, =(HighCol+8) - mov r0, r0, lsl #26 mov r0, r0, lsr #26 +.endif + orr r0, r0, r1, lsl #6 orr r0, r0, r0, lsl #8 orr r0, r0, r0, lsl #16 @@ -645,26 +696,34 @@ BackFill: @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ -.global DrawTilesFromCache @ int *hc, int sh +.global DrawTilesFromCache @ int *hc, int sh, int rlim DrawTilesFromCache: stmfd sp!, {r4-r8,r11,lr} - mvn r5, #0 @ r5=prevcode=-1 - mov r8, r1 - @ cache some stuff to avoid mem access +.if OVERRIDE_HIGHCOL + ldr r11,=HighCol + mov r12,#0xf + ldr r11,[r11] +.else ldr r11,=HighCol - ldr lr, =(Pico+0x10000) @ lr=Pico.vram mov r12,#0xf +.endif + ldr lr, =(Pico+0x10000) @ lr=Pico.vram + + mvn r5, #0 @ r5=prevcode=-1 + ands r8, r1, #1 + orr r8, r8, r2, lsl #1 + bne .dtfc_check_rendflags @ scratch: r4, r7 .dtfc_loop: ldr r6, [r0], #4 @ read code movs r1, r6, lsr #16 @ r1=dx; ldmeqfd sp!, {r4-r8,r11,pc} @ dx is never zero, this must be a terminator, return - bic r1, r1, #0xfe00 - add r1, r11, r1 @ r1=pdest + bic r4, r1, #0xfe00 + add r1, r11, r4 @ r1=pdest mov r7, r6, lsl #16 cmp r5, r7, lsr #16 @@ -682,7 +741,10 @@ DrawTilesFromCache: ldr r2, [lr, r2, lsl #1] @ pack=*(unsigned int *)(Pico.vram+addr); // Get 8 pixels .dtfc_samecode: - tst r8, r8 + rsbs r4, r4, r8, lsr #1 + bmi .dtfc_cut_tile + + tst r8, #1 bne .dtfc_shadow tst r2, r2 @@ -692,16 +754,17 @@ DrawTilesFromCache: beq .dtfc_SingleColor @ tileline singlecolor tst r5, #0x0800 - beq .dtfc_TileNorm + bne .dtfc_TileFlip @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern - TileFlip r12 - b .dtfc_loop - .dtfc_TileNorm: TileNorm r12 b .dtfc_loop +.dtfc_TileFlip: + TileFlip r12 + b .dtfc_loop + .dtfc_SingleColor: and r4, r2, #0xf orr r4, r3, r4 @@ -723,16 +786,17 @@ DrawTilesFromCache: beq .dtfc_SingleColor @ tileline singlecolor tst r5, #0x0800 - beq .dtfc_TileNormShHP + bne .dtfc_TileFlipShHP @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern - TileFlipShHP - b .dtfc_loop - .dtfc_TileNormShHP: TileNormShHP b .dtfc_loop +.dtfc_TileFlipShHP: + TileFlipShHP + b .dtfc_loop + .dtfc_shadow_blank: ldrb r4, [r1] @ 1ci ldrb r12,[r1,#1] @@ -769,6 +833,64 @@ DrawTilesFromCache: mov r12, #0xf b .dtfc_loop +.dtfc_cut_tile: + add r4, r4, #7 @ 0-6 + mov r4, r4, lsl #2 + mov r12,#0xf<<28 + mov r12,r12,asr r4 + mov r2, r2, ror #16 + tst r5, #0x0800 @ flipped? + mvnne r12,r12 + and r2, r2, r12 + mov r2, r2, ror #16 + mov r12,#0xf + tst r8, #1 + bne .dtfc_shadow + tst r2, r2 + beq .dtfc_loop + tst r5, #0x0800 + beq .dtfc_TileNorm + b .dtfc_TileFlip + +@ check if we have detected layer covered with hi-prio tiles: +.dtfc_check_rendflags: + ldr r1, =rendstatus + ldr r2, [r1] + tst r2, #0xc0 + beq .dtfc_loop + bic r8, r8, #1 @ sh/hi mode off + tst r2, #0x80 + bne .dtfc_loop @ already processed + orr r2, r2, #0x80 + str r2, [r1] + + add r1, r11,#8 + mov r3, #320/4 + mov r7, #0x80 + orr r7, r7, r7, lsl #8 + orr r7, r7, r7, lsl #16 + mov r6, #0x3f + orr r6, r6, r6, lsl #8 + orr r6, r6, r6, lsl #16 +.dtfc_loop_shprep: + subs r3, r3, #1 + bmi .dtfc_loop @ done + ldr r2, [r1] + tst r2, r7 + andeq r2, r2, r6 + streq r2, [r1], #4 + beq .dtfc_loop_shprep + tst r2, #0x80000000 + biceq r2, r2, #0xc0000000 + tst r2, #0x00800000 + biceq r2, r2, #0x00c00000 + tst r2, #0x00008000 + biceq r2, r2, #0x0000c000 + tst r2, #0x00000080 + biceq r2, r2, #0x000000c0 + str r2, [r1], #4 + b .dtfc_loop_shprep + .pool @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ @@ -780,11 +902,17 @@ DrawSpritesFromCache: stmfd sp!, {r4-r11,lr} @ cache some stuff to avoid mem access +.if OVERRIDE_HIGHCOL + ldr r11,=HighCol + mov r12,#0xf + ldr r11,[r11] +.else ldr r11,=HighCol + mov r12,#0xf +.endif ldr lr, =(Pico+0x10000) @ lr=Pico.vram mov r6, r1, lsl #31 orr r6, r6, #1<<30 - mov r12,#0xf mov r10, r0 @@ -845,16 +973,17 @@ DrawSpritesFromCache: beq .dsfc_SingleColor @ tileline singlecolor tst r9, #0x10000 - beq .dsfc_TileNorm + bne .dsfc_TileFlip @ TileFlip (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern - TileFlip r12 - b .dsfc_inloop - .dsfc_TileNorm: TileNorm r12 b .dsfc_inloop +.dsfc_TileFlip: + TileFlip r12 + b .dsfc_inloop + .dsfc_SingleColor: tst r0, #1 @ not aligned? and r4, r2, #0xf @@ -873,16 +1002,17 @@ DrawSpritesFromCache: beq .dsfc_singlec_sh tst r9, #0x10000 - beq .dsfc_TileNorm_sh + bne .dsfc_TileFlip_sh @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern - TileFlipSh - b .dsfc_inloop - .dsfc_TileNorm_sh: TileNormSh b .dsfc_inloop +.dsfc_TileFlip_sh: + TileFlipSh + b .dsfc_inloop + .dsfc_singlec_sh: cmp r2, #0xe0000000 bcc .dsfc_SingleColor @ normal singlecolor tileline (carry inverted in ARM) @@ -946,9 +1076,15 @@ DrawSprite: bne .dspr_cache @ if(code&0x8000) // high priority - cache it @ cache some stuff to avoid mem access +.if OVERRIDE_HIGHCOL + ldr r11,=HighCol + mov r12,#0xf + ldr r11,[r11] +.else ldr r11,=HighCol - ldr lr, =(Pico+0x10000) @ lr=Pico.vram mov r12,#0xf +.endif + ldr lr, =(Pico+0x10000) @ lr=Pico.vram mov r5, r5, lsl #4 @ delta<<=4; // Delta of address and r4, r9, #0x6000 @@ -989,17 +1125,18 @@ DrawSprite: beq .dspr_SingleColor @ tileline singlecolor tst r9, #0x0800 - beq .dspr_TileNorm + bne .dspr_TileFlip @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern - TileFlip r12 - b .dspr_loop - @ scratch: r4, r7 .dspr_TileNorm: TileNorm r12 b .dspr_loop +.dspr_TileFlip: + TileFlip r12 + b .dspr_loop + .dspr_SingleColor: and r4, r2, #0xf orr r4, r3, r4 @@ -1018,16 +1155,17 @@ DrawSprite: beq .dspr_singlec_sh tst r9, #0x0800 - beq .dspr_TileNorm_sh + bne .dspr_TileFlip_sh @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r12: helper pattern - TileFlipSh - b .dspr_loop - .dspr_TileNorm_sh: TileNormSh b .dspr_loop +.dspr_TileFlip_sh: + TileFlipSh + b .dspr_loop + .dspr_singlec_sh: cmp r2, #0xe0000000 bcc .dspr_SingleColor @ normal tileline @@ -1104,12 +1242,19 @@ DrawWindow: orr r6, r6, r3, lsl #8 @ shadow mode sub r8, r1, r0 - mov r8, r8, lsl #1 @ cells - - mvn r9, #0 @ r9=prevcode=-1 @ cache some stuff to avoid mem access +.if OVERRIDE_HIGHCOL + ldr r11,=HighCol + mov r8, r8, lsl #1 @ cells + ldr r11,[r11] + mvn r9, #0 @ r9=prevcode=-1 + add r11,r11,#8 +.else ldr r11,=(HighCol+8) + mov r8, r8, lsl #1 @ cells + mvn r9, #0 @ r9=prevcode=-1 +.endif add r1, r11, r0, lsl #4 @ r1=pdest mov r0, #0xf b .dwloop_enter @@ -1155,16 +1300,17 @@ DrawWindow: beq .dw_SingleColor @ tileline singlecolor tst r9, #0x0800 - beq .dw_TileNorm + bne .dw_TileFlip @ (r1=pdest, r2=pixels8, r3=pal) r4: scratch, r0: helper pattern - TileFlip r0 - b .dwloop - .dw_TileNorm: TileNorm r0 b .dwloop +.dw_TileFlip: + TileFlip r0 + b .dwloop + .dw_SingleColor: and r4, r0, r2 @ #0x0000000f orr r4, r3, r4 @@ -1309,9 +1455,17 @@ FinalizeLineBGR444: .fl_noshBGR444: +.if OVERRIDE_HIGHCOL + ldr r1, =HighCol + mov lr, #0xff + ldr r1, [r1] + mov lr, lr, lsl #1 + add r1, r1, #8 +.else ldr r1, =(HighCol+8) mov lr, #0xff mov lr, lr, lsl #1 +.endif .fl_loopBGR444: @@ -1381,14 +1535,16 @@ FinalizeLineBGR444: orr \reg, \reg, r3 @ add blue back .endm +.global vidConvCpyRGB565 + vidConvCpyRGB565: @ void *to, void *from, int pixels stmfd sp!, {r4-r9,lr} - mov r12, r2, lsr #3 @ repeats + mov r12, r2, lsr #3 @ repeats mov lr, #0x001c0000 orr lr, lr, #0x01c @ lr == pattern 0x001c001c mov r8, #0x00030000 - orr r8, r8, #0x003 @ lr == pattern 0x001c001c + orr r8, r8, #0x003 .loopRGB565: ldmia r1!, {r4-r7} @@ -1463,9 +1619,17 @@ FinalizeLineRGB555: sub r3, r3, #0x40*2 .fl_noshRGB555: +.if OVERRIDE_HIGHCOL + ldr r1, =HighCol + ldr r0, =DrawLineDest + ldr r1, [r1] + ldr r0, [r0] + add r1, r1, #8 +.else ldr r0, =DrawLineDest ldr r1, =(HighCol+8) ldr r0, [r0] +.endif ldrb r12, [r8, #12] mov lr, #0xff @@ -1483,8 +1647,14 @@ FinalizeLineRGB555: addeq r0, r0, #32*2 .fl_no32colRGB555: -.fl_loopRGB555: +.if UNALIGNED_DRAWLINEDEST + @ this is basically for Gizmondo, which has unaligned odd lines in the framebuffer + tst r0, #2 + bne .fl_RGB555u +.endif + +.fl_loopRGB555: ldr r12, [r1], #4 ldr r7, [r1], #4 @@ -1506,12 +1676,12 @@ FinalizeLineRGB555: ldrh r6, [r3, r6] and r12,lr, r7, lsr #15 ldrh r12,[r3, r12] + and r7, lr, r7, lsr #23 + ldrh r7, [r3, r7] orr r8, r8, r6, lsl #16 - and r6, lr, r7, lsr #23 - ldrh r6, [r3, r6] @ 1 cycle interlock here (r6) subs r2, r2, #1 - orr r12,r12, r6, lsl #16 + orr r12,r12, r7, lsl #16 stmia r0!, {r4,r5,r8,r12} bne .fl_loopRGB555 @@ -1525,6 +1695,11 @@ FinalizeLineRGB555: mov r9, #0x3900 @ f800 07e0 001f | e000 0780 001c | 3800 01e0 0007 orr r9, r9, #0x00e7 +.if UNALIGNED_DRAWLINEDEST + tst r0, #2 + bne .fl_32scale_RGB555u +.endif + .fl_loop32scale_RGB555: ldr r12, [r1], #4 ldr r7, [r1], #4 @@ -1581,6 +1756,121 @@ FinalizeLineRGB555: ldmfd sp!, {r4-r8,lr} bx lr +.if UNALIGNED_DRAWLINEDEST + @ unaligned versions of loops + @ warning: starts drawing 2bytes before dst + +.fl_RGB555u: + sub r0, r0, #2 @ initial adjustment + mov r8, #0 + +.fl_loopRGB555u: + ldr r12, [r1], #4 + ldr r7, [r1], #4 + + and r6, lr, r12,lsl #1 + ldrh r6, [r3, r6] + and r5, lr, r12,lsr #7 + ldrh r5, [r3, r5] + orr r4, r8, r6, lsl #16 + + and r6, lr, r12,lsr #15 + ldrh r6, [r3, r6] + and r8, lr, r12,lsr #23 + ldrh r8, [r3, r8] + orr r5, r5, r6, lsl #16 + + and r6, lr, r7, lsl #1 + ldrh r6, [r3, r6] + and r12,lr, r7, lsr #7 + ldrh r12,[r3, r12] + orr r6, r8, r6, lsl #16 + + and r8, lr, r7, lsr #15 + ldrh r8, [r3, r8] + and r7, lr, r7, lsr #23 + + subs r2, r2, #1 + orr r12,r12,r8, lsl #16 + ldrh r8, [r3, r7] + + stmia r0!, {r4,r5,r6,r12} + bne .fl_loopRGB555u + + strh r8, [r0], #2 + + ldmfd sp!, {r4-r8,lr} + bx lr + + +.fl_32scale_RGB555u: + sub r0, r0, #2 @ initial adjustment + mov r4, #0 + + @ r9 f800 07e0 001f | e000 0780 001c | 3800 01e0 0007 +.fl_loop32scale_RGB555u: + ldr r12, [r1], #4 + ldr r7, [r1], #4 + + and r6, lr, r12,lsl #1 + ldrh r6, [r3, r6] + and r5, lr, r12,lsr #7 + ldrh r5, [r3, r5] + and r6, r6, r9, lsl #2 + orr r4, r4, r6, lsl #16 @ r4 = pix_d -1, 0 + + and r5, r5, r9, lsl #2 + sub r8, r5, r5, lsr #2 @ r8 = 3/4 pix_s 1 + add r6, r8, r6, lsr #2 @ r6 = (1/4 pix_s 0) + (3/4 pix_s 1) + orr r5, r6, r5, lsl #15 + + and r6, lr, r12,lsr #15 + ldrh r6, [r3, r6] + and r12,lr, r12,lsr #23 + ldrh r12,[r3, r12] + and r6, r6, r9, lsl #2 + add r5, r5, r6, lsl #15 @ r5 = pix_d 1, 2 + + and r8, lr, r7, lsl #1 + ldrh r8, [r3, r8] + and r10,lr, r7, lsr #7 + ldrh r10,[r3, r10] + and r12,r12,r9, lsl #2 + sub r6, r6, r6, lsr #2 @ r6 = 3/4 pix_s 2 + add r6, r6, r12,lsr #2 + orr r6, r6, r12,lsl #16 @ r6 = pix_d 3, 4 + + and r8, r8, r9, lsl #2 + and r10,r10,r9, lsl #2 + sub r12,r10,r10,lsr #2 @ r12 = 3/4 pix_s 5 + orr r8, r8, r8, lsl #14 + add r8, r8, r12,lsl #16 @ r8 = pix_d 5, 6 + and r12,lr, r7, lsr #15 + ldrh r12,[r3, r12] + and r7, lr, r7, lsr #23 + ldrh r7, [r3, r7] + and r12,r12,r9, lsl #2 + add r10,r10,r12 + mov r10,r10, lsr #1 + sub r12,r12,r12,lsr #2 @ r12 = 3/4 pix_s 6 + orr r10,r10,r12,lsl #16 + and r7, r7, r9, lsl #2 + add r10,r10,r7, lsl #14 @ r10 = pix_d 7, 8 + + subs r2, r2, #1 + + stmia r0!, {r4,r5,r6,r8,r10} + mov r4, r7 + bne .fl_loop32scale_RGB555u + + strh r4, [r0], #2 + + ldmfd sp!, {r9,r10} + ldmfd sp!, {r4-r8,lr} + bx lr + +.endif @ UNALIGNED_DRAWLINEDEST + @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@