From c684ffa02d8d43ee04b99ee63ccd1adb66e81c1a Mon Sep 17 00:00:00 2001 From: Keith Whitwell Date: Thu, 22 May 2008 13:41:49 +0100 Subject: [PATCH] draw: clean up internal immediates in aos sse --- src/gallium/auxiliary/draw/draw_vs_aos.c | 64 +++++++++++++++------ src/gallium/auxiliary/draw/draw_vs_aos.h | 5 +- src/gallium/auxiliary/draw/draw_vs_aos_io.c | 10 ++-- 3 files changed, 55 insertions(+), 24 deletions(-) diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c index fde92c7226a..0b8600696a4 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos.c @@ -44,12 +44,6 @@ #ifdef PIPE_ARCH_X86 -#define DISASSEM 0 - - - - - static INLINE boolean eq( struct x86_reg a, struct x86_reg b ) { @@ -92,13 +86,6 @@ static struct x86_reg get_reg_ptr(struct aos_compilation *cp, } -struct x86_reg aos_get_internal( struct aos_compilation *cp, - unsigned imm ) -{ - return get_reg_ptr( cp, - AOS_FILE_INTERNAL, - imm ); -} #define X87_CW_EXCEPTION_INV_OP (1<<0) #define X87_CW_EXCEPTION_DENORM_OP (1<<1) @@ -123,6 +110,9 @@ static void init_internals( struct aos_machine *machine ) float inv = 1.0f/255.0f; float f255 = 255.0f; + ASSIGN_4V(machine->internal[IMM_SWZ], 1.0f, -1.0f, 0.0f, 1.0f); + *(unsigned *)&machine->internal[IMM_SWZ][3] = 0xffffffff; + ASSIGN_4V(machine->internal[IMM_ONES], 1.0f, 1.0f, 1.0f, 1.0f); ASSIGN_4V(machine->internal[IMM_NEGS], -1.0f, -1.0f, -1.0f, -1.0f); ASSIGN_4V(machine->internal[IMM_IDENTITY], 0.0f, 0.0f, 0.0f, 1.0f); @@ -337,6 +327,39 @@ struct x86_reg aos_get_shader_reg( struct aos_compilation *cp, +static struct x86_reg aos_get_shader_reg_xmm( struct aos_compilation *cp, + unsigned file, + unsigned idx ) +{ + struct x86_reg reg = aos_get_shader_reg( cp, file, idx ); + + if (reg.file != file_XMM) { + struct x86_reg tmp = aos_get_xmm_reg(cp); + sse_movups(cp->func, tmp, reg); + aos_adopt_xmm_reg( cp, tmp, file, idx, FALSE ); + reg = tmp; + } + + return reg; +} + + + +struct x86_reg aos_get_internal_xmm( struct aos_compilation *cp, + unsigned imm ) +{ + return aos_get_shader_reg_xmm( cp, AOS_FILE_INTERNAL, imm ); +} + + +struct x86_reg aos_get_internal( struct aos_compilation *cp, + unsigned imm ) +{ + return aos_get_shader_reg( cp, AOS_FILE_INTERNAL, imm ); +} + + + /* Emulate pshufd insn in regular SSE, if necessary: @@ -461,15 +484,15 @@ static struct x86_reg fetch_src( struct aos_compilation *cp, arg0 = dst; } - if (negs) { - struct x86_reg imm_negs = aos_get_internal(cp, IMM_NEGS); + if (negs && negs != 0xf) { + struct x86_reg imm_swz = aos_get_internal_xmm(cp, IMM_SWZ); struct x86_reg tmp = aos_get_xmm_reg(cp); /* Load 1,-1,0,0 * Use neg as arg to pshufd * Multiply */ - emit_pshufd(cp, tmp, imm_negs, + emit_pshufd(cp, tmp, imm_swz, SHUF((negs & 1) ? 1 : 0, (negs & 2) ? 1 : 0, (negs & 4) ? 1 : 0, @@ -479,12 +502,17 @@ static struct x86_reg fetch_src( struct aos_compilation *cp, aos_release_xmm_reg(cp, tmp.idx); arg0 = dst; } + else if (negs) { + struct x86_reg imm_negs = aos_get_internal_xmm(cp, IMM_NEGS); + sse_mulps(cp->func, dst, imm_negs); + arg0 = dst; + } + if (abs && abs != 0xf) { ERROR(cp, "unsupported partial abs"); } - - if (abs) { + else if (abs) { struct x86_reg neg = aos_get_internal(cp, IMM_NEGS); struct x86_reg tmp = aos_get_xmm_reg(cp); diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h index c2afd4e9a0d..efdc9a38f40 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.h +++ b/src/gallium/auxiliary/draw/draw_vs_aos.h @@ -142,13 +142,16 @@ boolean aos_emit_outputs( struct aos_compilation *cp ); #define IMM_ONES 0 /* 1, 1,1,1 */ -#define IMM_NEGS 1 /* 1,-1,0,0 */ +#define IMM_SWZ 1 /* 1,-1,0, 0xffffffff */ #define IMM_IDENTITY 2 /* 0, 0,0,1 */ #define IMM_INV_255 3 /* 1/255, 1/255, 1/255, 1/255 */ #define IMM_255 4 /* 255, 255, 255, 255 */ +#define IMM_NEGS 5 /* -1,-1,-1,-1 */ struct x86_reg aos_get_internal( struct aos_compilation *cp, unsigned imm ); +struct x86_reg aos_get_internal_xmm( struct aos_compilation *cp, + unsigned imm ); #define ERROR(cp, msg) \ diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c index 72b2b3d11d2..0dda9df97d6 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c @@ -54,7 +54,7 @@ static void emit_load_R32G32B32( struct aos_compilation *cp, struct x86_reg src_ptr ) { sse_movss(cp->func, data, x86_make_disp(src_ptr, 8)); - sse_shufps(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) ); + sse_shufps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) ); sse_shufps(cp->func, data, data, SHUF(Y,Z,X,W) ); sse_movlps(cp->func, data, src_ptr); } @@ -63,7 +63,7 @@ static void emit_load_R32G32( struct aos_compilation *cp, struct x86_reg data, struct x86_reg src_ptr ) { - sse_movups(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ) ); + sse_movups(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ) ); sse_movlps(cp->func, data, src_ptr); } @@ -73,7 +73,7 @@ static void emit_load_R32( struct aos_compilation *cp, struct x86_reg src_ptr ) { sse_movss(cp->func, data, src_ptr); - sse_orps(cp->func, data, aos_get_internal( cp, IMM_IDENTITY ) ); + sse_orps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ) ); } @@ -82,8 +82,8 @@ static void emit_load_R8G8B8A8_UNORM( struct aos_compilation *cp, struct x86_reg src_ptr ) { sse_movss(cp->func, data, src_ptr); - sse2_punpcklbw(cp->func, data, aos_get_internal( cp, IMM_IDENTITY )); - sse2_punpcklbw(cp->func, data, aos_get_internal( cp, IMM_IDENTITY )); + sse2_punpcklbw(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY )); + sse2_punpcklbw(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY )); sse2_cvtdq2ps(cp->func, data, data); sse_mulps(cp->func, data, aos_get_internal(cp, IMM_INV_255)); } -- 2.30.2