diff options
Diffstat (limited to 'src/gallium/auxiliary/draw/draw_vs_aos.c')
| -rw-r--r-- | src/gallium/auxiliary/draw/draw_vs_aos.c | 42 |
1 files changed, 28 insertions, 14 deletions
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c index 875ecb92db..1fb69ef81a 100644 --- a/src/gallium/auxiliary/draw/draw_vs_aos.c +++ b/src/gallium/auxiliary/draw/draw_vs_aos.c @@ -32,7 +32,7 @@ #include "util/u_memory.h" #include "util/u_math.h" #include "pipe/p_shader_tokens.h" -#include "pipe/p_debug.h" +#include "util/u_debug.h" #include "tgsi/tgsi_parse.h" #include "tgsi/tgsi_util.h" #include "tgsi/tgsi_exec.h" @@ -884,7 +884,7 @@ static void set_fpu_round_nearest( struct aos_compilation *cp ) } } - +#if 0 static void x87_emit_ex2( struct aos_compilation *cp ) { struct x86_reg st0 = x86_make_reg(file_x87, 0); @@ -907,13 +907,17 @@ static void x87_emit_ex2( struct aos_compilation *cp ) assert( stack == cp->func->x87_stack); } +#endif +#if 0 static void PIPE_CDECL print_reg( const char *msg, const float *reg ) { debug_printf("%s: %f %f %f %f\n", msg, reg[0], reg[1], reg[2], reg[3]); } +#endif +#if 0 static void emit_print( struct aos_compilation *cp, const char *message, /* must point to a static string! */ unsigned file, @@ -965,6 +969,7 @@ static void emit_print( struct aos_compilation *cp, /* Done... */ } +#endif /** * The traditional instructions. All operate on internal registers @@ -1103,7 +1108,7 @@ static boolean emit_LG2( struct aos_compilation *cp, const struct tgsi_full_inst return TRUE; } - +#if 0 static boolean emit_EX2( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) { x87_fld_src(cp, &op->FullSrcRegisters[0], 0); @@ -1111,6 +1116,7 @@ static boolean emit_EX2( struct aos_compilation *cp, const struct tgsi_full_inst x87_fstp_dest4(cp, &op->FullDstRegisters[0]); return TRUE; } +#endif static boolean emit_FLR( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) @@ -1566,7 +1572,6 @@ static boolean emit_RCP( struct aos_compilation *cp, const struct tgsi_full_inst */ static boolean emit_RSQ( struct aos_compilation *cp, const struct tgsi_full_instruction *op ) { - if (0) { struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]); struct x86_reg r = aos_get_xmm_reg(cp); @@ -1575,21 +1580,30 @@ static boolean emit_RSQ( struct aos_compilation *cp, const struct tgsi_full_inst return TRUE; } else { - struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]); - struct x86_reg r = aos_get_xmm_reg(cp); + struct x86_reg arg0 = fetch_src(cp, &op->FullSrcRegisters[0]); + struct x86_reg r = aos_get_xmm_reg(cp); struct x86_reg neg_half = get_reg_ptr( cp, AOS_FILE_INTERNAL, IMM_RSQ ); struct x86_reg one_point_five = x86_make_disp( neg_half, 4 ); struct x86_reg src = get_xmm_writable( cp, arg0 ); - - sse_rsqrtss( cp->func, r, src ); /* rsqrtss(a) */ - sse_mulss( cp->func, src, neg_half ); /* -.5 * a */ - sse_mulss( cp->func, src, r ); /* -.5 * a * r */ - sse_mulss( cp->func, src, r ); /* -.5 * a * r * r */ - sse_addss( cp->func, src, one_point_five ); /* 1.5 - .5 * a * r * r */ - sse_mulss( cp->func, r, src ); /* r * (1.5 - .5 * a * r * r) */ + struct x86_reg neg = aos_get_internal(cp, IMM_NEGS); + struct x86_reg tmp = aos_get_xmm_reg(cp); + + sse_movaps(cp->func, tmp, src); + sse_mulps(cp->func, tmp, neg); + sse_maxps(cp->func, tmp, src); + + sse_rsqrtss( cp->func, r, tmp ); /* rsqrtss(a) */ + sse_mulss( cp->func, tmp, neg_half ); /* -.5 * a */ + sse_mulss( cp->func, tmp, r ); /* -.5 * a * r */ + sse_mulss( cp->func, tmp, r ); /* -.5 * a * r * r */ + sse_addss( cp->func, tmp, one_point_five ); /* 1.5 - .5 * a * r * r */ + sse_mulss( cp->func, r, tmp ); /* r * (1.5 - .5 * a * r * r) */ store_scalar_dest(cp, &op->FullDstRegisters[0], r); + + aos_release_xmm_reg(cp, tmp.idx); + return TRUE; } } @@ -1877,7 +1891,7 @@ static boolean note_immediate( struct aos_compilation *cp, unsigned pos = cp->num_immediates++; unsigned j; - for (j = 0; j < imm->Immediate.Size; j++) { + for (j = 0; j < imm->Immediate.NrTokens - 1; j++) { cp->vaos->machine->immediate[pos][j] = imm->u.ImmediateFloat32[j].Float; } |