nv50: LIT - sort-of

*somehow* we have the exact same bug here as on nv4x, the difference being
on nv4x the hw actually has a LIT opcode.. NVIDIA doesn't have the bug on
either arch FWIW.

diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
index fabd016..186eea5 100644 (file)
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -13,7 +13,6 @@
 #define NV50_SU_MAX_TEMP 64
 
 /* ARL
- * LIT - other buggery
  *
  * MSB - Like MAD, but MUL+SUB
  *     - Fuck it off, introduce a way to negate args for ops that
@@ -23,6 +22,8 @@
  *     - Maybe even relax restrictions a bit, can't do P_RESULT + P_IMMD,
  *       but can emit to P_TEMP first - then MOV later. NVIDIA does this
  *
+ * Hmmm.. what happens if we have src1+src2 both consts.. ouch !
+ *
  * Verify half-insns work where expected - and force disable them where they
  * don't work - MUL has it forcibly disabled atm as it fixes POW..
  */
@@ -164,7 +165,7 @@ alloc_immd(struct nv50_pc *pc, float f)
        struct nv50_reg *r = CALLOC_STRUCT(nv50_reg);
        unsigned hw;
 
-       hw = ctor_immd(pc, f, 0, 0, 0);
+       hw = ctor_immd(pc, f, 0, 0, 0) * 4;
        r->type = P_IMMD;
        r->hw = hw;
        r->index = -1;
@@ -657,6 +658,34 @@ emit_abs(struct nv50_pc *pc, struct nv50_reg *dst, struct nv50_reg *src)
        emit(pc, inst);
 }
 
+static void
+emit_lit(struct nv50_pc *pc, struct nv50_reg **dst, struct nv50_reg **src)
+{
+       struct nv50_reg *one = alloc_immd(pc, 1.0);
+       struct nv50_reg *zero = alloc_immd(pc, 0.0);
+       struct nv50_reg *neg128 = alloc_immd(pc, -127.999999);
+       struct nv50_reg *pos128 = alloc_immd(pc,  127.999999);
+       struct nv50_reg *tmp[4];
+
+       emit_mov(pc, dst[0], one);
+       emit_mov(pc, dst[3], one);
+
+       tmp[0] = temp_temp(pc);
+       emit_minmax(pc, 4, dst[1], src[0], zero);
+       set_pred_wr(pc, 1, 0, &pc->p->insns[pc->p->insns_nr - 2]);
+
+       tmp[1] = temp_temp(pc);
+       emit_minmax(pc, 4, tmp[1], src[1], zero);
+
+       tmp[3] = temp_temp(pc);
+       emit_minmax(pc, 4, tmp[3], src[3], neg128);
+       emit_minmax(pc, 5, tmp[3], tmp[3], pos128);
+
+       emit_pow(pc, dst[2], tmp[1], tmp[3]);
+       emit_mov(pc, dst[2], zero);
+       set_pred(pc, 3, 0, &pc->p->insns[pc->p->insns_nr - 2]);
+}
+
 static struct nv50_reg *
 tgsi_dst(struct nv50_pc *pc, int c, const struct tgsi_full_dst_register *dst)
 {
@@ -861,6 +890,9 @@ nv50_program_tx_insn(struct nv50_pc *pc, const union tgsi_full_token *tok)
                }
                free_temp(pc, temp);
                break;
+       case TGSI_OPCODE_LIT:
+               emit_lit(pc, &dst[0], &src[0][0]);
+               break;
        case TGSI_OPCODE_LG2:
                for (c = 0; c < 4; c++) {
                        if (!(mask & (1 << c)))

diff --git a/src/gallium/drivers/nv50/nv50_vbo.c b/src/gallium/drivers/nv50/nv50_vbo.c
index f086c76..badbef5 100644 (file)
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -35,7 +35,6 @@ nv50_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start,
        struct nv50_context *nv50 = nv50_context(pipe);
 
        nv50_state_validate(nv50);
-       NOUVEAU_ERR("unimplemented\n");
 
        BEGIN_RING(tesla, 0x142c, 1);
        OUT_RING  (0);