From: Wim Taymans Date: Thu, 26 Oct 2000 19:07:27 +0000 (+0000) Subject: Slightly faster yuv to rgb conversion. X-Git-Tag: RELEASE-0_1_0-SLIPSTREAM~289 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=31cae4b447fa5dc8507dc1ed23c2d9bb4d9fec32;p=platform%2Fupstream%2Fgstreamer.git Slightly faster yuv to rgb conversion. Original commit message from CVS: Slightly faster yuv to rgb conversion. --- diff --git a/libs/colorspace/Makefile.am b/libs/colorspace/Makefile.am index 12090a1..a6257df 100644 --- a/libs/colorspace/Makefile.am +++ b/libs/colorspace/Makefile.am @@ -2,14 +2,14 @@ filterdir = $(libdir)/gst filter_LTLIBRARIES = libgstcolorspace.la -libgstcolorspace_la_SOURCES = gstcolorspace.c rgb2rgb.c yuv2rgb.c +libgstcolorspace_la_SOURCES = gstcolorspace.c rgb2rgb.c yuv2rgb.c yuv2rgb_mmx16.s libgstcolorspaceincludedir = $(includedir)/gst/libs/gstcolorspace libgstcolorspaceinclude_HEADERS = gstcolorspace.h noinst_HEADERS = yuv2rgb.h -CFLAGS += -Wall -O2 -fomit-frame-pointer -funroll-all-loops -finline-functions -ffast-math +CFLAGS += -Wall -O2 -fomit-frame-pointer -finline-functions -ffast-math INCLUDES = $(GLIB_CFLAGS) $(GTK_CFLAGS) -I$(top_srcdir) -I$(top_srcdir)/include LDADD = $(GLIB_LIBS) $(GTK_LIBS) $(top_srcdir)/gst/libgst.la diff --git a/libs/colorspace/yuv2rgb.c b/libs/colorspace/yuv2rgb.c index f6dfb06..0fe9deb 100644 --- a/libs/colorspace/yuv2rgb.c +++ b/libs/colorspace/yuv2rgb.c @@ -34,6 +34,8 @@ #include "yuv2rgb.h" + +static void gst_colorspace_yuv420P_to_bgr16_mmx(GstColorSpaceConverter *space, unsigned char *src, unsigned char *dest); static void gst_colorspace_yuv420P_to_rgb32(GstColorSpaceConverter *space, unsigned char *src, unsigned char *dest); static void gst_colorspace_yuv420P_to_bgr32(GstColorSpaceConverter *space, unsigned char *src, unsigned char *dest); static void gst_colorspace_yuv420P_to_bgr32_mmx(GstColorSpaceConverter *space, unsigned char *src, unsigned char *dest); @@ -67,7 +69,7 @@ static void gst_colorspace_yuv_to_bgr32_mmx(GstColorSpaceYUVTables *tables, unsigned char *cb, unsigned char *out, int cols, int rows); -static void gst_colorspace_yuv_to_bgr16_mmx(GstColorSpaceYUVTables *tables, +extern void gst_colorspace_yuv_to_bgr16_mmx(GstColorSpaceYUVTables *tables, unsigned char *lum, unsigned char *cr, unsigned char *cb, @@ -213,8 +215,6 @@ static void gst_colorspace_yuv420P_to_rgb16(GstColorSpaceConverter *space, unsig } #ifdef HAVE_LIBMMX -static mmx_t MMX16_redmask = (mmx_t)(long long)0xf800f800f800f800LL; //dd 07c00 7c00h, 07c007c00h -static mmx_t MMX16_grnmask = (mmx_t)(long long)0x07e007e007e007e0LL; //dd 003e0 03e0h, 003e003e0h static void gst_colorspace_yuv420P_to_bgr32_mmx(GstColorSpaceConverter *space, unsigned char *src, unsigned char *dest) { int size; @@ -663,193 +663,17 @@ gst_colorspace_yuv_to_rgb32(tables, lum, cb, cr, out, rows, cols) } #ifdef HAVE_LIBMMX -static mmx_t MMX_10w = (mmx_t)(long long)0x000D000D000D000DLL; //dd 00080 0080h, 000800080h static mmx_t MMX_80w = (mmx_t)(long long)0x0080008000800080LL; //dd 00080 0080h, 000800080h static mmx_t MMX_00FFw = (mmx_t)(long long)0x00ff00ff00ff00ffLL; //dd 000FF 00FFh, 000FF00FFh static mmx_t MMX_FF00w = (mmx_t)(long long)0xff00ff00ff00ff00LL; //dd 000FF 00FFh, 000FF00FFh -static mmx_t MMX16_Vredcoeff = (mmx_t)(long long)0x0066006600660066LL; //dd 00066 0066h, 000660066h -static mmx_t MMX16_Ublucoeff = (mmx_t)(long long)0x0081008100810081LL; //dd 00081 0081h, 000810081h -static mmx_t MMX16_Ugrncoeff = (mmx_t)(long long)0xffe8ffe8ffe8ffe8LL; //dd 0FFE7 FFE7h, 0FFE7FFE7h -static mmx_t MMX16_Vgrncoeff = (mmx_t)(long long)0xffcdffcdffcdffcdLL; //dd 0FFCC FFCCh, 0FFCCFFCCh - -static mmx_t MMX16_Ycoeff = (mmx_t)(long long)0x004a004a004a004aLL; //dd 0004A 004Ah, 0004A004Ah - - static mmx_t MMX32_Vredcoeff = (mmx_t)(long long)0x0059005900590059LL; static mmx_t MMX32_Ubluecoeff = (mmx_t)(long long)0x0072007200720072LL; static mmx_t MMX32_Ugrncoeff = (mmx_t)(long long)0xffeaffeaffeaffeaLL; static mmx_t MMX32_Vgrncoeff = (mmx_t)(long long)0xffd2ffd2ffd2ffd2LL; static void -gst_colorspace_yuv_to_bgr16_mmx(tables, lum, cr, cb, out, rows, cols) - GstColorSpaceYUVTables *tables; - unsigned char *lum; - unsigned char *cr; - unsigned char *cb; - unsigned char *out; - int cols, rows; - -{ - unsigned short *row1 = (unsigned short* )out; // 32 bit target - int cols8 = cols>>3; - - int y, x; - - DEBUG("gst_colorspace_yuv420P_to_bgr16_mmx %p %p %p\n", lum, cr, cb); - - for (y=rows>>1; y; y--) { - for (x=cols8; x; x--) { - - movd_m2r(*(mmx_t *)cr, mm0); // 4 Cr 0 0 0 0 u3 u2 u1 u0 - pxor_r2r(mm7, mm7); - movd_m2r(*(mmx_t *)cb, mm1); // 4 Cb 0 0 0 0 v3 v2 v1 v0 - punpcklbw_r2r(mm7, mm0); // 4 W cb 0 u3 0 u2 0 u1 0 u0 - punpcklbw_r2r(mm7, mm1); // 4 W cr 0 v3 0 v2 0 v1 0 v0 - psubw_m2r(MMX_80w, mm0); - psubw_m2r(MMX_80w, mm1); - movq_r2r(mm0, mm2); // Cb 0 u3 0 u2 0 u1 0 u0 - movq_r2r(mm1, mm3); // Cr - pmullw_m2r(MMX16_Ugrncoeff, mm2); // Cb2green 0 R3 0 R2 0 R1 0 R0 - movq_m2r(*(mmx_t *)lum, mm6); // L1 l7 L6 L5 L4 L3 L2 L1 L0 - pmullw_m2r(MMX16_Ublucoeff, mm0); // Cb2blue - pand_m2r(MMX_00FFw, mm6); // L1 00 L6 00 L4 00 L2 00 L0 - pmullw_m2r(MMX16_Vgrncoeff, mm3); // Cr2green - movq_m2r(*(mmx_t *)lum, mm7); // L2 - pmullw_m2r(MMX16_Vredcoeff, mm1); // Cr2red - psubw_m2r(MMX_10w, mm6); - psrlw_i2r(8, mm7); // L2 00 L7 00 L5 00 L3 00 L1 - pmullw_m2r(MMX16_Ycoeff, mm6); // lum1 - psubw_m2r(MMX_10w, mm7); - paddw_r2r(mm3, mm2); // Cb2green + Cr2green == green - pmullw_m2r(MMX16_Ycoeff, mm7); // lum2 - - movq_r2r(mm6, mm4); // lum1 - paddw_r2r(mm0, mm6); // lum1 +blue 00 B6 00 B4 00 B2 00 B0 - movq_r2r(mm4, mm5); // lum1 - paddw_r2r(mm1, mm4); // lum1 +red 00 R6 00 R4 00 R2 00 R0 - paddw_r2r(mm2, mm5); // lum1 +green 00 G6 00 G4 00 G2 00 G0 - psraw_i2r(6, mm4); // R1 0 .. 64 - movq_r2r(mm7, mm3); // lum2 00 L7 00 L5 00 L3 00 L1 - psraw_i2r(6, mm5); // G1 - .. + - paddw_r2r(mm0, mm7); // Lum2 +blue 00 B7 00 B5 00 B3 00 B1 - psraw_i2r(6, mm6); // B1 0 .. 64 - packuswb_r2r(mm4, mm4); // R1 R1 - packuswb_r2r(mm5, mm5); // G1 G1 - packuswb_r2r(mm6, mm6); // B1 B1 - punpcklbw_r2r(mm4, mm4); - punpcklbw_r2r(mm5, mm5); - - pand_m2r(MMX16_redmask, mm4); - psllw_i2r(3, mm5); // GREEN 1 - punpcklbw_r2r(mm6, mm6); - pand_m2r(MMX16_grnmask, mm5); - pand_m2r(MMX16_redmask, mm6); - por_r2r(mm5, mm4); // - psrlw_i2r(11, mm6); // BLUE 1 - movq_r2r(mm3, mm5); // lum2 - paddw_r2r(mm1, mm3); // lum2 +red 00 R7 00 R5 00 R3 00 R1 - paddw_r2r(mm2, mm5); // lum2 +green 00 G7 00 G5 00 G3 00 G1 - psraw_i2r(6, mm3); // R2 - por_r2r(mm6, mm4); // MM4 - psraw_i2r(6, mm5); // G2 - movq_m2r(*(mmx_t *)(lum+cols), mm6); // L3 load lum2 - psraw_i2r(6, mm7); - packuswb_r2r(mm3, mm3); - packuswb_r2r(mm5, mm5); - packuswb_r2r(mm7, mm7); - pand_m2r(MMX_00FFw, mm6); // L3 - psubw_m2r(MMX_10w, mm6); - punpcklbw_r2r(mm3, mm3); - // "psubw MMX_10w, %%mm6\n" // L3 - punpcklbw_r2r(mm5, mm5); - pmullw_m2r(MMX16_Ycoeff, mm6); // lum3 - punpcklbw_r2r(mm7, mm7); - psllw_i2r(3, mm5); // GREEN 2 - pand_m2r(MMX16_redmask, mm7); - pand_m2r(MMX16_redmask, mm3); - psrlw_i2r(11, mm7); // BLUE 2 - pand_m2r(MMX16_grnmask, mm5); - por_r2r(mm7, mm3); - movq_m2r(*(mmx_t *)(lum+cols), mm7); // L4 load lum2 - por_r2r(mm5, mm3); // - psrlw_i2r(8, mm7); // L4 - movq_r2r(mm4, mm5); - // "psubw MMX_10w, %%mm7\n" // L4 - punpcklwd_r2r(mm3, mm4); - psubw_m2r(MMX_10w, mm7); - pmullw_m2r(MMX16_Ycoeff, mm7); // lum4 - punpckhwd_r2r(mm3, mm5); - - movq_r2m(mm4, *(row1)); // write row1 - movq_r2m(mm5, *(row1+4)); // write row1 - - movq_r2r(mm6, mm4); // Lum3 - paddw_r2r(mm0, mm6); // Lum3 +blue - - movq_r2r(mm4, mm5); // Lum3 - paddw_r2r(mm1, mm4); // Lum3 +red - paddw_r2r(mm2, mm5); // Lum3 +green - psraw_i2r(6, mm4); - movq_r2r(mm7, mm3); // Lum4 - psraw_i2r(6, mm5); - paddw_r2r(mm0, mm7); // Lum4 +blue - psraw_i2r(6, mm6); // Lum3 +blue - movq_r2r(mm3, mm0); // Lum4 - packuswb_r2r(mm4, mm4); - paddw_r2r(mm1, mm3); // Lum4 +red - packuswb_r2r(mm5, mm5); - paddw_r2r(mm2, mm0); // Lum4 +green - packuswb_r2r(mm6, mm6); - punpcklbw_r2r(mm4, mm4); - punpcklbw_r2r(mm5, mm5); - punpcklbw_r2r(mm6, mm6); - psllw_i2r(3, mm5); // GREEN 3 - pand_m2r(MMX16_redmask, mm4); - psraw_i2r(6, mm3); // psr 6 - psraw_i2r(6, mm0); - pand_m2r(MMX16_redmask, mm6); // BLUE - pand_m2r(MMX16_grnmask, mm5); - psrlw_i2r(11, mm6); // BLUE 3 - por_r2r(mm5, mm4); - psraw_i2r(6, mm7); - por_r2r(mm6, mm4); - packuswb_r2r(mm3, mm3); - packuswb_r2r(mm0, mm0); - packuswb_r2r(mm7, mm7); - punpcklbw_r2r(mm3, mm3); - punpcklbw_r2r(mm0, mm0); - punpcklbw_r2r(mm7, mm7); - pand_m2r(MMX16_redmask, mm3); - pand_m2r(MMX16_redmask, mm7); // BLUE - psllw_i2r(3, mm0); // GREEN 4 - psrlw_i2r(11, mm7); - pand_m2r(MMX16_grnmask, mm0); - por_r2r(mm7, mm3); - por_r2r(mm0, mm3); - - movq_r2r(mm4, mm5); - - punpcklwd_r2r(mm3, mm4); - punpckhwd_r2r(mm3, mm5); - - movq_r2m(mm4, *(row1+cols)); - movq_r2m(mm5, *(row1+cols+4)); - - lum+=8; - cr+=4; - cb+=4; - row1 +=8; - } - lum += cols; - row1 += cols; - } - - emms(); - -} -static void gst_colorspace_yuv_to_bgr32_mmx(tables, lum, cr, cb, out, rows, cols) GstColorSpaceYUVTables *tables; unsigned char *lum; diff --git a/libs/colorspace/yuv2rgb_mmx16.s b/libs/colorspace/yuv2rgb_mmx16.s new file mode 100644 index 0000000..8f7b1f7 --- /dev/null +++ b/libs/colorspace/yuv2rgb_mmx16.s @@ -0,0 +1,188 @@ +.globl mmx_80w +.data + .align 4 + .type mmx_80w,@object + .size mmx_80w,8 +mmx_80w: + .long 8388736 + .long 8388736 +.globl mmx_10w + .align 4 + .type mmx_10w,@object + .size mmx_10w,8 +mmx_10w: + .long 269488144 + .long 269488144 +.globl mmx_00ffw + .align 4 + .type mmx_00ffw,@object + .size mmx_00ffw,8 +mmx_00ffw: + .long 16711935 + .long 16711935 +.globl mmx_Y_coeff + .align 4 + .type mmx_Y_coeff,@object + .size mmx_Y_coeff,8 +mmx_Y_coeff: + .long 624895295 + .long 624895295 +.globl mmx_U_green + .align 4 + .type mmx_U_green,@object + .size mmx_U_green,8 +mmx_U_green: + .long -209849475 + .long -209849475 +.globl mmx_U_blue + .align 4 + .type mmx_U_blue,@object + .size mmx_U_blue,8 +mmx_U_blue: + .long 1083392147 + .long 1083392147 +.globl mmx_V_red + .align 4 + .type mmx_V_red,@object + .size mmx_V_red,8 +mmx_V_red: + .long 856830738 + .long 856830738 +.globl mmx_V_green + .align 4 + .type mmx_V_green,@object + .size mmx_V_green,8 +mmx_V_green: + .long -436410884 + .long -436410884 +.globl mmx_redmask + .align 4 + .type mmx_redmask,@object + .size mmx_redmask,8 +mmx_redmask: + .long -117901064 + .long -117901064 +.globl mmx_grnmask + .align 4 + .type mmx_grnmask,@object + .size mmx_grnmask,8 +mmx_grnmask: + .long -50529028 + .long -50529028 +.text + .align 4 +.globl gst_colorspace_yuv_to_bgr16_mmx + .type gst_colorspace_yuv_to_bgr16_mmx,@function +gst_colorspace_yuv_to_bgr16_mmx: + subl $8,%esp + pushl %ebp + pushl %edi + pushl %esi + movl 28(%esp),%edi + movl 32(%esp),%ecx + movl 36(%esp),%edx + movl $1,%ebp + movl 48(%esp),%esi + sarl $1,%esi + movl %esi,16(%esp) + + pxor %mm4, %mm4 # zero mm4 + + movl %esi,12(%esp) + sarl $2,12(%esp) + + movl 40(%esp),%esi + + .p2align 4,,7 +.L68: + + movd (%ecx), %mm0 # Load 4 Cb 00 00 00 00 00 u3 u2 u1 u0 + movd (%edx), %mm1 # Load 4 Cr 00 00 00 00 00 v2 v1 v0 + movq (%edi), %mm6 # Load 8 Y Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 + + movl 12(%esp),%eax + + .p2align 4,,7 +.L74: + punpcklbw %mm4, %mm0 # scatter 4 Cb 00 u3 00 u2 00 u1 00 u0 + punpcklbw %mm4, %mm1 # scatter 4 Cr 00 v3 00 v2 00 v1 00 v0 + psubsw mmx_80w, %mm0 # Cb -= 128 + psubsw mmx_80w, %mm1 # Cr -= 128 + psllw $3, %mm0 # Promote precision + psllw $3, %mm1 # Promote precision + movq %mm0, %mm2 # Copy 4 Cb 00 u3 00 u2 00 u1 00 u0 + movq %mm1, %mm3 # Copy 4 Cr 00 v3 00 v2 00 v1 00 v0 + pmulhw mmx_U_green, %mm2 # Mul Cb with green coeff -> Cb green + pmulhw mmx_V_green, %mm3 # Mul Cr with green coeff -> Cr green + pmulhw mmx_U_blue, %mm0 # Mul Cb -> Cblue 00 b3 00 b2 00 b1 00 b0 + pmulhw mmx_V_red, %mm1 # Mul Cr -> Cred 00 r3 00 r2 00 r1 00 r0 + paddsw %mm3, %mm2 # Cb green + Cr green -> Cgreen + psubusb mmx_10w, %mm6 # Y -= 16 + movq %mm6, %mm7 # Copy 8 Y Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 + pand mmx_00ffw, %mm6 # get Y even 00 Y6 00 Y4 00 Y2 00 Y0 + psrlw $8, %mm7 # get Y odd 00 Y7 00 Y5 00 Y3 00 Y1 + psllw $3, %mm6 # Promote precision + psllw $3, %mm7 # Promote precision + pmulhw mmx_Y_coeff, %mm6 # Mul 4 Y even 00 y6 00 y4 00 y2 00 y0 + pmulhw mmx_Y_coeff, %mm7 # Mul 4 Y odd 00 y7 00 y5 00 y3 00 y1 + movq %mm0, %mm3 # Copy Cblue + movq %mm1, %mm4 # Copy Cred + movq %mm2, %mm5 # Copy Cgreen + paddsw %mm6, %mm0 # Y even + Cblue 00 B6 00 B4 00 B2 00 B0 + paddsw %mm7, %mm3 # Y odd + Cblue 00 B7 00 B5 00 B3 00 B1 + paddsw %mm6, %mm1 # Y even + Cred 00 R6 00 R4 00 R2 00 R0 + paddsw %mm7, %mm4 # Y odd + Cred 00 R7 00 R5 00 R3 00 R1 + paddsw %mm6, %mm2 # Y even + Cgreen 00 G6 00 G4 00 G2 00 G0 + paddsw %mm7, %mm5 # Y odd + Cgreen 00 G7 00 G5 00 G3 00 G1 + packuswb %mm0, %mm0 # B6 B4 B2 B0 | B6 B4 B2 B0 + packuswb %mm1, %mm1 # R6 R4 R2 R0 | R6 R4 R2 R0 + packuswb %mm2, %mm2 # G6 G4 G2 G0 | G6 G4 G2 G0 + packuswb %mm3, %mm3 # B7 B5 B3 B1 | B7 B5 B3 B1 + packuswb %mm4, %mm4 # R7 R5 R3 R1 | R7 R5 R3 R1 + packuswb %mm5, %mm5 # G7 G5 G3 G1 | G7 G5 G3 G1 + punpcklbw %mm3, %mm0 # B7 B6 B5 B4 B3 B2 B1 B0 + punpcklbw %mm4, %mm1 # R7 R6 R5 R4 R3 R2 R1 R0 + punpcklbw %mm5, %mm2 # G7 G6 G5 G4 G3 G2 G1 G0 + pand mmx_redmask, %mm0 # b7b6b5b4 b3_0_0_0 b7b6b5b4 b3_0_0_0 + pand mmx_grnmask, %mm2 # g7g6g5g4 g3g2_0_0 g7g6g5g4 g3g2_0_0 + pand mmx_redmask, %mm1 # r7r6r5r4 r3_0_0_0 r7r6r5r4 r3_0_0_0 + psrlw $3,%mm0 #0_0_0_b7 b6b5b4b3 0_0_0_b7 b6b5b4b3 + pxor %mm4, %mm4 # zero mm4 + movq %mm0, %mm5 # Copy B7-B0 + movq %mm2, %mm7 # Copy G7-G0 + punpcklbw %mm4, %mm2 # 0_0_0_0 0_0_0_0 g7g6g5g4 g3g2_0_0 + punpcklbw %mm1, %mm0 # r7r6r5r4 r3_0_0_0 0_0_0_b7 b6b5b4b3 + psllw $3,%mm2 # 0_0_0_0 0_g7g6g5 g4g3g2_0 0_0_0_0 + por %mm2, %mm0 # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3 + movq 8(%edi), %mm6 # Load 8 Y Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0 + movq %mm0, (%esi) # store pixel 0-3 + punpckhbw %mm4, %mm7 # 0_0_0_0 0_0_0_0 g7g6g5g4 g3g2_0_0 + punpckhbw %mm1, %mm5 # r7r6r5r4 r3_0_0_0 0_0_0_b7 b6b5b4b3 + psllw $3,%mm7 # 0_0_0_0 0_g7g6g5 g4g3g2_0 0_0_0_0 + movd 4(%ecx), %mm0 # Load 4 Cb 00 00 00 00 u3 u2 u1 u0 + por %mm7, %mm5 # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3 + movd 4(%edx), %mm1 # Load 4 Cr 00 00 00 00 v3 v2 v1 v0 + movq %mm5, 8(%esi) # store pixel 4-7 + + addl $8,%edi + addl $4,%ecx + addl $4,%edx + addl $16,%esi + decl %eax + jnz .L74 +.L72: + xorl $1,%ebp + jne .L76 + subl 16(%esp),%ecx + subl 16(%esp),%edx +.L76: + subl $1,44(%esp) + jnz .L68 + + emms + + popl %esi + popl %edi + popl %ebp + addl $8,%esp + ret