* \param ifragZ_reg register containing integer fragment Z values (in)
* \param ifbZ_reg register containing integer frame buffer Z values (in/out)
* \param zmask_reg register containing result of Z test/comparison (out)
+ *
+ * Returns true if the Z-buffer needs to be updated.
*/
-static void
-gen_depth_test(const struct pipe_depth_stencil_alpha_state *dsa,
- struct spe_function *f,
+static boolean
+gen_depth_test(struct spe_function *f,
+ const struct pipe_depth_stencil_alpha_state *dsa,
int mask_reg, int ifragZ_reg, int ifbZ_reg, int zmask_reg)
{
+ /* NOTE: we use clgt below, not cgt, because we want to compare _unsigned_
+ * quantities. This only makes a difference for 32-bit Z values though.
+ */
ASSERT(dsa->depth.enabled);
switch (dsa->depth.func) {
case PIPE_FUNC_GREATER:
/* zmask = (ifragZ > ref) */
- spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg);
+ spe_clgt(f, zmask_reg, ifragZ_reg, ifbZ_reg);
/* mask = (mask & zmask) */
spe_and(f, mask_reg, mask_reg, zmask_reg);
break;
case PIPE_FUNC_LESS:
/* zmask = (ref > ifragZ) */
- spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg);
+ spe_clgt(f, zmask_reg, ifbZ_reg, ifragZ_reg);
/* mask = (mask & zmask) */
spe_and(f, mask_reg, mask_reg, zmask_reg);
break;
case PIPE_FUNC_LEQUAL:
/* zmask = (ifragZ > ref) */
- spe_cgt(f, zmask_reg, ifragZ_reg, ifbZ_reg);
+ spe_clgt(f, zmask_reg, ifragZ_reg, ifbZ_reg);
/* mask = (mask & ~zmask) */
spe_andc(f, mask_reg, mask_reg, zmask_reg);
break;
case PIPE_FUNC_GEQUAL:
/* zmask = (ref > ifragZ) */
- spe_cgt(f, zmask_reg, ifbZ_reg, ifragZ_reg);
+ spe_clgt(f, zmask_reg, ifbZ_reg, ifragZ_reg);
/* mask = (mask & ~zmask) */
spe_andc(f, mask_reg, mask_reg, zmask_reg);
break;
* framebufferZ = (ztest_passed ? fragmentZ : framebufferZ;
*/
spe_selb(f, ifbZ_reg, ifbZ_reg, ifragZ_reg, mask_reg);
+ return true;
}
+
+ return false;
}
spe_release_register(f, amask_reg);
}
-/* This is a convenient and oft-used sequence. It chooses
- * the smaller of each element of reg1 and reg2, and combines them
- * into the result register, as follows:
- *
- * The Float Compare Greater Than (fcgt) instruction will put
- * 1s into compare_reg where reg1 > reg2, and 0s where reg1 <= reg2.
- *
- * Then the Select Bits (selb) instruction will take bits from
- * reg1 where compare_reg is 0, and from reg2 where compare_reg is
- * 1. Ergo, result_reg will have the bits from reg1 where reg1 <= reg2,
- * and the bits from reg2 where reg1 > reg2, which is exactly the
- * MIN operation.
+/* This pair of functions is used inline to allocate and deallocate
+ * optional constant registers. Once a constant is discovered to be
+ * needed, we will likely need it again, so we don't want to deallocate
+ * it and have to allocate and load it again unnecessarily.
*/
-#define FLOAT_VECTOR_MIN(f, result_reg, reg1, reg2) {\
- int compare_reg = spe_allocate_available_register(f); \
- spe_fcgt(f, compare_reg, reg1, reg2); \
- spe_selb(f, result_reg, reg1, reg2, compare_reg); \
- spe_release_register(f, compare_reg); \
+static inline void
+setup_optional_register(struct spe_function *f, boolean *is_already_set, unsigned int *r)
+{
+ if (*is_already_set) return;
+ *r = spe_allocate_available_register(f);
+ *is_already_set = true;
}
-/* The FLOAT_VECTOR_MAX sequence is similar to the FLOAT_VECTOR_MIN
- * sequence above, except that the registers specified when selecting
- * bits are reversed.
- */
-#define FLOAT_VECTOR_MAX(f, result_reg, reg1, reg2) {\
- int compare_reg = spe_allocate_available_register(f); \
- spe_fcgt(f, compare_reg, reg1, reg2); \
- spe_selb(f, result_reg, reg2, reg1, compare_reg); \
- spe_release_register(f, compare_reg); \
+static inline void
+release_optional_register(struct spe_function *f, boolean *is_already_set, unsigned int r)
+{
+ if (!*is_already_set) return;
+ spe_release_register(f, r);
+ *is_already_set = false;
+}
+
+static inline void
+setup_const_register(struct spe_function *f, boolean *is_already_set, unsigned int *r, float value)
+{
+ if (*is_already_set) return;
+ setup_optional_register(f, is_already_set, r);
+ spe_load_float(f, *r, value);
+}
+
+static inline void
+release_const_register(struct spe_function *f, boolean *is_already_set, unsigned int r)
+{
+ release_optional_register(f, is_already_set, r);
}
/**
int tmp_reg = spe_allocate_available_register(f);
- /* These values might or might not eventually get put into
- * registers. We avoid allocating them and setting them until
- * they're actually needed; then we avoid setting them more than
- * once, and release them at the end of code generation.
+ /* Optional constant registers we might or might not end up using;
+ * if we do use them, make sure we only allocate them once by
+ * keeping a flag on each one.
*/
- boolean one_reg_set = false;
- int one_reg;
-#define SET_ONE_REG_IF_UNSET(f) if (!one_reg_set) {\
- one_reg = spe_allocate_available_register(f); \
- spe_load_float(f, one_reg, 1.0f); \
- one_reg_set = true; \
-}
-#define RELEASE_ONE_REG_IF_USED(f) if (one_reg_set) {\
- spe_release_register(f, one_reg); \
-}
-
- boolean const_color_set = false;
- int constR_reg, constG_reg, constB_reg;
-#define SET_CONST_COLOR_IF_UNSET(f, blend_color) if (!const_color_set) {\
- constR_reg = spe_allocate_available_register(f); \
- constG_reg = spe_allocate_available_register(f); \
- constG_reg = spe_allocate_available_register(f); \
- spe_load_float(f, constR_reg, blend_color->color[0]); \
- spe_load_float(f, constG_reg, blend_color->color[1]); \
- spe_load_float(f, constB_reg, blend_color->color[2]); \
- const_color_set = true;\
-}
-#define RELEASE_CONST_COLOR_IF_USED(f) if (const_color_set) {\
- spe_release_register(f, constR_reg); \
- spe_release_register(f, constG_reg); \
- spe_release_register(f, constB_reg); \
-}
-
- boolean const_alpha_set = false;
- int constA_reg;
-#define SET_CONST_ALPHA_IF_UNSET(f, blend_color) if (!const_alpha_set) {\
- constA_reg = spe_allocate_available_register(f); \
- spe_load_float(f, constA_reg, blend_color->color[3]); \
- const_alpha_set = true; \
-}
-#define RELEASE_CONST_ALPHA_IF_USED(f) if (const_alpha_set) {\
- spe_release_register(f, constA_reg); \
-}
-
- /* Real code starts here */
+ boolean one_reg_set = false;
+ unsigned int one_reg;
+ boolean constR_reg_set = false, constG_reg_set = false,
+ constB_reg_set = false, constA_reg_set = false;
+ unsigned int constR_reg, constG_reg, constB_reg, constA_reg;
ASSERT(blend->blend_enable);
spe_release_register(f, mask_reg);
}
-
/*
* Compute Src RGB terms. We're actually looking for the value
- * of (the appropriate RGB factors) * (the incoming source RGB color).
+ * of (the appropriate RGB factors) * (the incoming source RGB color),
+ * because in some cases (like PIPE_BLENDFACTOR_ONE and
+ * PIPE_BLENDFACTOR_ZERO) we can avoid doing unnecessary math.
*/
switch (blend->rgb_src_factor) {
case PIPE_BLENDFACTOR_ONE:
spe_fm(f, term1B_reg, fragB_reg, fragA_reg);
break;
case PIPE_BLENDFACTOR_INV_SRC_COLOR:
- /* factors = (1-R,1-G,1-B), so term = (R*(1-R), G*(1-G), B*(1-B)) */
- /* we'll need the optional constant {1,1,1,1} register */
- SET_ONE_REG_IF_UNSET(f)
- /* tmp = 1 - R */
- spe_fs(f, tmp_reg, one_reg, fragR_reg);
- /* term = R * tmp */
- spe_fm(f, term1R_reg, fragR_reg, tmp_reg);
- /* repeat for G and B */
- spe_fs(f, tmp_reg, one_reg, fragG_reg);
- spe_fm(f, term1G_reg, fragG_reg, tmp_reg);
- spe_fs(f, tmp_reg, one_reg, fragB_reg);
- spe_fm(f, term1B_reg, fragB_reg, tmp_reg);
+ /* factors = (1-R,1-G,1-B), so term = (R*(1-R), G*(1-G), B*(1-B))
+ * or in other words term = (R-R*R, G-G*G, B-B*B)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term1R_reg, fragR_reg, fragR_reg, fragR_reg);
+ spe_fnms(f, term1G_reg, fragG_reg, fragG_reg, fragG_reg);
+ spe_fnms(f, term1B_reg, fragB_reg, fragB_reg, fragB_reg);
break;
case PIPE_BLENDFACTOR_DST_COLOR:
/* factors = (Rfb,Gfb,Bfb), so term = (R*Rfb, G*Gfb, B*Bfb) */
spe_fm(f, term1B_reg, fragB_reg, fbB_reg);
break;
case PIPE_BLENDFACTOR_INV_DST_COLOR:
- /* factors = (1-Rfb,1-Gfb,1-Bfb), so term = (R*(1-Rfb),G*(1-Gfb),B*(1-Bfb)) */
- /* we'll need the optional constant {1,1,1,1} register */
- SET_ONE_REG_IF_UNSET(f)
- /* tmp = 1 - Rfb */
- spe_fs(f, tmp_reg, one_reg, fbR_reg);
- /* term = R * tmp */
- spe_fm(f, term1R_reg, fragR_reg, tmp_reg);
- /* repeat for G and B */
- spe_fs(f, tmp_reg, one_reg, fbG_reg);
- spe_fm(f, term1G_reg, fragG_reg, tmp_reg);
- spe_fs(f, tmp_reg, one_reg, fbB_reg);
- spe_fm(f, term1B_reg, fragB_reg, tmp_reg);
+ /* factors = (1-Rfb,1-Gfb,1-Bfb), so term = (R*(1-Rfb),G*(1-Gfb),B*(1-Bfb))
+ * or term = (R-R*Rfb, G-G*Gfb, B-B*Bfb)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term1R_reg, fragR_reg, fbR_reg, fragR_reg);
+ spe_fnms(f, term1G_reg, fragG_reg, fbG_reg, fragG_reg);
+ spe_fnms(f, term1B_reg, fragB_reg, fbB_reg, fragB_reg);
break;
case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
- /* factors = (1-A,1-A,1-A), so term = (R*(1-A),G*(1-A),B*(1-A)) */
- /* we'll need the optional constant {1,1,1,1} register */
- SET_ONE_REG_IF_UNSET(f)
- /* tmp = 1 - A */
- spe_fs(f, tmp_reg, one_reg, fragA_reg);
- /* term = R * tmp */
- spe_fm(f, term1R_reg, fragR_reg, tmp_reg);
- /* repeat for G and B with the same (1-A) factor */
- spe_fm(f, term1G_reg, fragG_reg, tmp_reg);
- spe_fm(f, term1B_reg, fragB_reg, tmp_reg);
+ /* factors = (1-A,1-A,1-A), so term = (R*(1-A),G*(1-A),B*(1-A))
+ * or term = (R-R*A,G-G*A,B-B*A)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term1R_reg, fragR_reg, fragA_reg, fragR_reg);
+ spe_fnms(f, term1G_reg, fragG_reg, fragA_reg, fragG_reg);
+ spe_fnms(f, term1B_reg, fragB_reg, fragA_reg, fragB_reg);
break;
case PIPE_BLENDFACTOR_DST_ALPHA:
/* factors = (Afb, Afb, Afb), so term = (R*Afb, G*Afb, B*Afb) */
spe_fm(f, term1B_reg, fragB_reg, fbA_reg);
break;
case PIPE_BLENDFACTOR_INV_DST_ALPHA:
- /* factors = (1-Afb, 1-Afb, 1-Afb), so term = (R*(1-Afb),G*(1-Afb),B*(1-Afb)) */
- /* we'll need the optional constant {1,1,1,1} register */
- SET_ONE_REG_IF_UNSET(f)
- /* tmp = 1 - A */
- spe_fs(f, tmp_reg, one_reg, fbA_reg);
- /* term = R * tmp, G*tmp, and B*tmp */
- spe_fm(f, term1R_reg, fragR_reg, tmp_reg);
- spe_fm(f, term1G_reg, fragG_reg, tmp_reg);
- spe_fm(f, term1B_reg, fragB_reg, tmp_reg);
+ /* factors = (1-Afb, 1-Afb, 1-Afb), so term = (R*(1-Afb),G*(1-Afb),B*(1-Afb))
+ * or term = (R-R*Afb,G-G*Afb,b-B*Afb)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term1R_reg, fragR_reg, fbA_reg, fragR_reg);
+ spe_fnms(f, term1G_reg, fragG_reg, fbA_reg, fragG_reg);
+ spe_fnms(f, term1B_reg, fragB_reg, fbA_reg, fragB_reg);
break;
case PIPE_BLENDFACTOR_CONST_COLOR:
- /* We'll need the optional blend color registers */
- SET_CONST_COLOR_IF_UNSET(f,blend_color)
+ /* We need the optional constant color registers */
+ setup_const_register(f, &constR_reg_set, &constR_reg, blend_color->color[0]);
+ setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]);
+ setup_const_register(f, &constB_reg_set, &constB_reg, blend_color->color[2]);
/* now, factor = (Rc,Gc,Bc), so term = (R*Rc,G*Gc,B*Bc) */
spe_fm(f, term1R_reg, fragR_reg, constR_reg);
spe_fm(f, term1G_reg, fragG_reg, constG_reg);
break;
case PIPE_BLENDFACTOR_CONST_ALPHA:
/* we'll need the optional constant alpha register */
- SET_CONST_ALPHA_IF_UNSET(f, blend_color)
+ setup_const_register(f, &constA_reg_set, &constA_reg, blend_color->color[3]);
/* factor = (Ac,Ac,Ac), so term = (R*Ac,G*Ac,B*Ac) */
spe_fm(f, term1R_reg, fragR_reg, constA_reg);
spe_fm(f, term1G_reg, fragG_reg, constA_reg);
spe_fm(f, term1B_reg, fragB_reg, constA_reg);
break;
case PIPE_BLENDFACTOR_INV_CONST_COLOR:
- /* We need both the optional {1,1,1,1} register, and the optional
- * constant color registers
+ /* We need the optional constant color registers */
+ setup_const_register(f, &constR_reg_set, &constR_reg, blend_color->color[0]);
+ setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]);
+ setup_const_register(f, &constB_reg_set, &constB_reg, blend_color->color[2]);
+ /* factor = (1-Rc,1-Gc,1-Bc), so term = (R*(1-Rc),G*(1-Gc),B*(1-Bc))
+ * or term = (R-R*Rc, G-G*Gc, B-B*Bc)
+ * fnms(a,b,c,d) computes a = d - b*c
*/
- SET_ONE_REG_IF_UNSET(f)
- SET_CONST_COLOR_IF_UNSET(f, blend_color)
- /* factor = (1-Rc,1-Gc,1-Bc), so term = (R*(1-Rc),G*(1-Gc),B*(1-Bc)) */
- spe_fs(f, tmp_reg, one_reg, constR_reg);
- spe_fm(f, term1R_reg, fragR_reg, tmp_reg);
- spe_fs(f, tmp_reg, one_reg, constG_reg);
- spe_fm(f, term1G_reg, fragG_reg, tmp_reg);
- spe_fs(f, tmp_reg, one_reg, constB_reg);
- spe_fm(f, term1B_reg, fragB_reg, tmp_reg);
+ spe_fnms(f, term1R_reg, fragR_reg, constR_reg, fragR_reg);
+ spe_fnms(f, term1G_reg, fragG_reg, constG_reg, fragG_reg);
+ spe_fnms(f, term1B_reg, fragB_reg, constB_reg, fragB_reg);
break;
case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
- /* We need the optional {1,1,1,1} register and the optional
- * constant alpha register
+ /* We need the optional constant color registers */
+ setup_const_register(f, &constR_reg_set, &constR_reg, blend_color->color[0]);
+ setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]);
+ setup_const_register(f, &constB_reg_set, &constB_reg, blend_color->color[2]);
+ /* factor = (1-Ac,1-Ac,1-Ac), so term = (R*(1-Ac),G*(1-Ac),B*(1-Ac))
+ * or term = (R-R*Ac,G-G*Ac,B-B*Ac)
+ * fnms(a,b,c,d) computes a = d - b*c
*/
- SET_ONE_REG_IF_UNSET(f)
- SET_CONST_ALPHA_IF_UNSET(f, blend_color)
- /* factor = (1-Ac,1-Ac,1-Ac), so term = (R*(1-Ac),G*(1-Ac),B*(1-Ac)) */
- spe_fs(f, tmp_reg, one_reg, constA_reg);
- spe_fm(f, term1R_reg, fragR_reg, tmp_reg);
- spe_fm(f, term1G_reg, fragG_reg, tmp_reg);
- spe_fm(f, term1B_reg, fragB_reg, tmp_reg);
+ spe_fnms(f, term1R_reg, fragR_reg, constA_reg, fragR_reg);
+ spe_fnms(f, term1G_reg, fragG_reg, constA_reg, fragG_reg);
+ spe_fnms(f, term1B_reg, fragB_reg, constA_reg, fragB_reg);
break;
case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
/* We'll need the optional {1,1,1,1} register */
- SET_ONE_REG_IF_UNSET(f)
+ setup_const_register(f, &one_reg_set, &one_reg, 1.0f);
/* factor = (min(A,1-Afb),min(A,1-Afb),min(A,1-Afb)), so
* term = (R*min(A,1-Afb), G*min(A,1-Afb), B*min(A,1-Afb))
+ * We could expand the term (as a*min(b,c) == min(a*b,a*c)
+ * as long as a is positive), but then we'd have to do three
+ * spe_float_min() functions instead of one, so this is simpler.
*/
/* tmp = 1 - Afb */
spe_fs(f, tmp_reg, one_reg, fbA_reg);
/* tmp = min(A,tmp) */
- FLOAT_VECTOR_MIN(f, tmp_reg, fragA_reg, tmp_reg)
+ spe_float_min(f, tmp_reg, fragA_reg, tmp_reg);
/* term = R*tmp */
spe_fm(f, term1R_reg, fragR_reg, tmp_reg);
spe_fm(f, term1G_reg, fragG_reg, tmp_reg);
spe_fm(f, term1B_reg, fragB_reg, tmp_reg);
break;
- /* non-OpenGL cases? */
+ /* These are special D3D cases involving a second color output
+ * from the fragment shader. I'm not sure we can support them
+ * yet... XXX
+ */
case PIPE_BLENDFACTOR_SRC1_COLOR:
case PIPE_BLENDFACTOR_SRC1_ALPHA:
case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
}
/*
- * Compute Src Alpha term
+ * Compute Src Alpha term. Like the above, we're looking for
+ * the full term A*factor, not just the factor itself, because
+ * in many cases we can avoid doing unnecessary multiplies.
*/
switch (blend->alpha_src_factor) {
+ case PIPE_BLENDFACTOR_ZERO:
+ /* factor = 0, so term = 0 */
+ spe_load_float(f, term1A_reg, 0.0f);
+ break;
+
+ case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: /* fall through */
case PIPE_BLENDFACTOR_ONE:
+ /* factor = 1, so term = A */
spe_move(f, term1A_reg, fragA_reg);
break;
+
case PIPE_BLENDFACTOR_SRC_COLOR:
+ /* factor = A, so term = A*A */
spe_fm(f, term1A_reg, fragA_reg, fragA_reg);
break;
case PIPE_BLENDFACTOR_SRC_ALPHA:
spe_fm(f, term1A_reg, fragA_reg, fragA_reg);
break;
- /* XXX more cases */
+
+ case PIPE_BLENDFACTOR_INV_SRC_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_INV_SRC_COLOR:
+ /* factor = 1-A, so term = A*(1-A) = A-A*A */
+ /* fnms(a,b,c,d) computes a = d - b*c */
+ spe_fnms(f, term1A_reg, fragA_reg, fragA_reg, fragA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_DST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ /* factor = Afb, so term = A*Afb */
+ spe_fm(f, term1A_reg, fragA_reg, fbA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_INV_DST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_INV_DST_COLOR:
+ /* factor = 1-Afb, so term = A*(1-Afb) = A - A*Afb */
+ /* fnms(a,b,c,d) computes a = d - b*c */
+ spe_fnms(f, term1A_reg, fragA_reg, fbA_reg, fragA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_CONST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ /* We need the optional constA_reg register */
+ setup_const_register(f, &constA_reg_set, &constA_reg, blend_color->color[3]);
+ /* factor = Ac, so term = A*Ac */
+ spe_fm(f, term1A_reg, fragA_reg, constA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_INV_CONST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_INV_CONST_COLOR:
+ /* We need the optional constA_reg register */
+ setup_const_register(f, &constA_reg_set, &constA_reg, blend_color->color[3]);
+ /* factor = 1-Ac, so term = A*(1-Ac) = A-A*Ac */
+ /* fnms(a,b,c,d) computes a = d - b*c */
+ spe_fnms(f, term1A_reg, fragA_reg, constA_reg, fragA_reg);
+ break;
+
+ /* These are special D3D cases involving a second color output
+ * from the fragment shader. I'm not sure we can support them
+ * yet... XXX
+ */
+ case PIPE_BLENDFACTOR_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_SRC1_ALPHA:
+ case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
default:
ASSERT(0);
}
/*
- * Compute Dest RGB terms
+ * Compute Dest RGB term. Like the above, we're looking for
+ * the full term (Rfb,Gfb,Bfb)*(factor), not just the factor itself, because
+ * in many cases we can avoid doing unnecessary multiplies.
*/
switch (blend->rgb_dst_factor) {
case PIPE_BLENDFACTOR_ONE:
+ /* factors = (1,1,1), so term = (Rfb,Gfb,Bfb) */
spe_move(f, term2R_reg, fbR_reg);
spe_move(f, term2G_reg, fbG_reg);
spe_move(f, term2B_reg, fbB_reg);
break;
case PIPE_BLENDFACTOR_ZERO:
- spe_zero(f, term2R_reg);
- spe_zero(f, term2G_reg);
- spe_zero(f, term2B_reg);
+ /* factor s= (0,0,0), so term = (0,0,0) */
+ spe_load_float(f, term2R_reg, 0.0f);
+ spe_load_float(f, term2G_reg, 0.0f);
+ spe_load_float(f, term2B_reg, 0.0f);
break;
case PIPE_BLENDFACTOR_SRC_COLOR:
+ /* factors = (R,G,B), so term = (R*Rfb, G*Gfb, B*Bfb) */
spe_fm(f, term2R_reg, fbR_reg, fragR_reg);
spe_fm(f, term2G_reg, fbG_reg, fragG_reg);
spe_fm(f, term2B_reg, fbB_reg, fragB_reg);
break;
+ case PIPE_BLENDFACTOR_INV_SRC_COLOR:
+ /* factors = (1-R,1-G,1-B), so term = (Rfb*(1-R), Gfb*(1-G), Bfb*(1-B))
+ * or in other words term = (Rfb-Rfb*R, Gfb-Gfb*G, Bfb-Bfb*B)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term2R_reg, fragR_reg, fbR_reg, fbR_reg);
+ spe_fnms(f, term2G_reg, fragG_reg, fbG_reg, fbG_reg);
+ spe_fnms(f, term2B_reg, fragB_reg, fbB_reg, fbB_reg);
+ break;
case PIPE_BLENDFACTOR_SRC_ALPHA:
+ /* factors = (A,A,A), so term = (Rfb*A, Gfb*A, Bfb*A) */
spe_fm(f, term2R_reg, fbR_reg, fragA_reg);
spe_fm(f, term2G_reg, fbG_reg, fragA_reg);
spe_fm(f, term2B_reg, fbB_reg, fragA_reg);
break;
case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
-#if 0
- /* one = {1.0, 1.0, 1.0, 1.0} */
- if (!one_reg_set) {
- one_reg = spe_allocate_available_register(f);
- spe_load_float(f, one_reg, 1.0f);
- one_reg_set = true;
- }
- /* tmp = one - fragA */
- spe_fs(f, tmp_reg, one_reg, fragA_reg);
- /* term = fb * tmp */
- spe_fm(f, term2R_reg, fbR_reg, tmp_reg);
- spe_fm(f, term2G_reg, fbG_reg, tmp_reg);
- spe_fm(f, term2B_reg, fbB_reg, tmp_reg);
-#else
- /* Compute: term2x = fbx * (1.0 - fragA)
- * Which is: term2x = fbx - fbx * fragA
- * Use fnms t,a,b,c which computes t=c-a*b
- */
+ /* factors = (1-A,1-A,1-A) so term = (Rfb-Rfb*A,Gfb-Gfb*A,Bfb-Bfb*A) */
+ /* fnms(a,b,c,d) computes a = d - b*c */
spe_fnms(f, term2R_reg, fbR_reg, fragA_reg, fbR_reg);
spe_fnms(f, term2G_reg, fbG_reg, fragA_reg, fbG_reg);
spe_fnms(f, term2B_reg, fbB_reg, fragA_reg, fbB_reg);
-#endif
break;
- /* XXX more cases */
- // GL_ONE_MINUS_SRC_COLOR
- // GL_DST_COLOR
- // GL_ONE_MINUS_DST_COLOR
- // GL_DST_ALPHA
- // GL_CONSTANT_COLOR
- // GL_ONE_MINUS_CONSTANT_COLOR
- // GL_CONSTANT_ALPHA
- // GL_ONE_MINUS_CONSTANT_ALPHA
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ /* factors = (Rfb,Gfb,Bfb), so term = (Rfb*Rfb, Gfb*Gfb, Bfb*Bfb) */
+ spe_fm(f, term2R_reg, fbR_reg, fbR_reg);
+ spe_fm(f, term2G_reg, fbG_reg, fbG_reg);
+ spe_fm(f, term2B_reg, fbB_reg, fbB_reg);
+ break;
+ case PIPE_BLENDFACTOR_INV_DST_COLOR:
+ /* factors = (1-Rfb,1-Gfb,1-Bfb), so term = (Rfb*(1-Rfb),Gfb*(1-Gfb),Bfb*(1-Bfb))
+ * or term = (Rfb-Rfb*Rfb, Gfb-Gfb*Gfb, Bfb-Bfb*Bfb)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term2R_reg, fbR_reg, fbR_reg, fbR_reg);
+ spe_fnms(f, term2G_reg, fbG_reg, fbG_reg, fbG_reg);
+ spe_fnms(f, term2B_reg, fbB_reg, fbB_reg, fbB_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_DST_ALPHA:
+ /* factors = (Afb, Afb, Afb), so term = (Rfb*Afb, Gfb*Afb, Bfb*Afb) */
+ spe_fm(f, term2R_reg, fbR_reg, fbA_reg);
+ spe_fm(f, term2G_reg, fbG_reg, fbA_reg);
+ spe_fm(f, term2B_reg, fbB_reg, fbA_reg);
+ break;
+ case PIPE_BLENDFACTOR_INV_DST_ALPHA:
+ /* factors = (1-Afb, 1-Afb, 1-Afb), so term = (Rfb*(1-Afb),Gfb*(1-Afb),Bfb*(1-Afb))
+ * or term = (Rfb-Rfb*Afb,Gfb-Gfb*Afb,Bfb-Bfb*Afb)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term2R_reg, fbR_reg, fbA_reg, fbR_reg);
+ spe_fnms(f, term2G_reg, fbG_reg, fbA_reg, fbG_reg);
+ spe_fnms(f, term2B_reg, fbB_reg, fbA_reg, fbB_reg);
+ break;
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ /* We need the optional constant color registers */
+ setup_const_register(f, &constR_reg_set, &constR_reg, blend_color->color[0]);
+ setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]);
+ setup_const_register(f, &constB_reg_set, &constB_reg, blend_color->color[2]);
+ /* now, factor = (Rc,Gc,Bc), so term = (Rfb*Rc,Gfb*Gc,Bfb*Bc) */
+ spe_fm(f, term2R_reg, fbR_reg, constR_reg);
+ spe_fm(f, term2G_reg, fbG_reg, constG_reg);
+ spe_fm(f, term2B_reg, fbB_reg, constB_reg);
+ break;
+ case PIPE_BLENDFACTOR_CONST_ALPHA:
+ /* we'll need the optional constant alpha register */
+ setup_const_register(f, &constA_reg_set, &constA_reg, blend_color->color[3]);
+ /* factor = (Ac,Ac,Ac), so term = (Rfb*Ac,Gfb*Ac,Bfb*Ac) */
+ spe_fm(f, term2R_reg, fbR_reg, constA_reg);
+ spe_fm(f, term2G_reg, fbG_reg, constA_reg);
+ spe_fm(f, term2B_reg, fbB_reg, constA_reg);
+ break;
+ case PIPE_BLENDFACTOR_INV_CONST_COLOR:
+ /* We need the optional constant color registers */
+ setup_const_register(f, &constR_reg_set, &constR_reg, blend_color->color[0]);
+ setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]);
+ setup_const_register(f, &constB_reg_set, &constB_reg, blend_color->color[2]);
+ /* factor = (1-Rc,1-Gc,1-Bc), so term = (Rfb*(1-Rc),Gfb*(1-Gc),Bfb*(1-Bc))
+ * or term = (Rfb-Rfb*Rc, Gfb-Gfb*Gc, Bfb-Bfb*Bc)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term2R_reg, fbR_reg, constR_reg, fbR_reg);
+ spe_fnms(f, term2G_reg, fbG_reg, constG_reg, fbG_reg);
+ spe_fnms(f, term2B_reg, fbB_reg, constB_reg, fbB_reg);
+ break;
+ case PIPE_BLENDFACTOR_INV_CONST_ALPHA:
+ /* We need the optional constant color registers */
+ setup_const_register(f, &constR_reg_set, &constR_reg, blend_color->color[0]);
+ setup_const_register(f, &constG_reg_set, &constG_reg, blend_color->color[1]);
+ setup_const_register(f, &constB_reg_set, &constB_reg, blend_color->color[2]);
+ /* factor = (1-Ac,1-Ac,1-Ac), so term = (Rfb*(1-Ac),Gfb*(1-Ac),Bfb*(1-Ac))
+ * or term = (Rfb-Rfb*Ac,Gfb-Gfb*Ac,Bfb-Bfb*Ac)
+ * fnms(a,b,c,d) computes a = d - b*c
+ */
+ spe_fnms(f, term2R_reg, fbR_reg, constA_reg, fbR_reg);
+ spe_fnms(f, term2G_reg, fbG_reg, constA_reg, fbG_reg);
+ spe_fnms(f, term2B_reg, fbB_reg, constA_reg, fbB_reg);
+ break;
+ case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: /* not supported for dest RGB */
+ ASSERT(0);
+ break;
+
+ /* These are special D3D cases involving a second color output
+ * from the fragment shader. I'm not sure we can support them
+ * yet... XXX
+ */
+ case PIPE_BLENDFACTOR_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_SRC1_ALPHA:
+ case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
+
default:
ASSERT(0);
}
/*
- * Compute Dest Alpha term
+ * Compute Dest Alpha term. Like the above, we're looking for
+ * the full term Afb*factor, not just the factor itself, because
+ * in many cases we can avoid doing unnecessary multiplies.
*/
switch (blend->alpha_dst_factor) {
case PIPE_BLENDFACTOR_ONE:
+ /* factor = 1, so term = Afb */
spe_move(f, term2A_reg, fbA_reg);
break;
case PIPE_BLENDFACTOR_ZERO:
- spe_zero(f, term2A_reg);
+ /* factor = 0, so term = 0 */
+ spe_load_float(f, term2A_reg, 0.0f);
break;
- case PIPE_BLENDFACTOR_SRC_ALPHA:
+
+ case PIPE_BLENDFACTOR_SRC_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_SRC_COLOR:
+ /* factor = A, so term = Afb*A */
spe_fm(f, term2A_reg, fbA_reg, fragA_reg);
break;
- case PIPE_BLENDFACTOR_INV_SRC_ALPHA:
-#if 0
- /* one = {1.0, 1.0, 1.0, 1.0} */
- if (!one_reg_set) {
- one_reg = spe_allocate_available_register(f);
- spe_load_float(f, one_reg, 1.0f);
- one_reg_set = true;
- }
- /* tmp = one - fragA */
- spe_fs(f, tmp_reg, one_reg, fragA_reg);
- /* termA = fbA * tmp */
- spe_fm(f, term2A_reg, fbA_reg, tmp_reg);
-#else
- /* Compute: term2A = fbA * (1.0 - fragA)
- * Which is: term2A = fbA - fbA * fragA
- * Use fnms t,a,b,c which computes t=c-a*b
- */
+
+ case PIPE_BLENDFACTOR_INV_SRC_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_INV_SRC_COLOR:
+ /* factor = 1-A, so term = Afb*(1-A) = Afb-Afb*A */
+ /* fnms(a,b,c,d) computes a = d - b*c */
spe_fnms(f, term2A_reg, fbA_reg, fragA_reg, fbA_reg);
-#endif
break;
- /* XXX more cases */
- // GL_ONE_MINUS_SRC_COLOR
- // GL_DST_COLOR
- // GL_ONE_MINUS_DST_COLOR
- // GL_DST_ALPHA
- // GL_CONSTANT_COLOR
- // GL_ONE_MINUS_CONSTANT_COLOR
- // GL_CONSTANT_ALPHA
- // GL_ONE_MINUS_CONSTANT_ALPHA
+
+ case PIPE_BLENDFACTOR_DST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_DST_COLOR:
+ /* factor = Afb, so term = Afb*Afb */
+ spe_fm(f, term2A_reg, fbA_reg, fbA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_INV_DST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_INV_DST_COLOR:
+ /* factor = 1-Afb, so term = Afb*(1-Afb) = Afb - Afb*Afb */
+ /* fnms(a,b,c,d) computes a = d - b*c */
+ spe_fnms(f, term2A_reg, fbA_reg, fbA_reg, fbA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_CONST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_CONST_COLOR:
+ /* We need the optional constA_reg register */
+ setup_const_register(f, &constA_reg_set, &constA_reg, blend_color->color[3]);
+ /* factor = Ac, so term = Afb*Ac */
+ spe_fm(f, term2A_reg, fbA_reg, constA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_INV_CONST_ALPHA: /* fall through */
+ case PIPE_BLENDFACTOR_INV_CONST_COLOR:
+ /* We need the optional constA_reg register */
+ setup_const_register(f, &constA_reg_set, &constA_reg, blend_color->color[3]);
+ /* factor = 1-Ac, so term = Afb*(1-Ac) = Afb-Afb*Ac */
+ /* fnms(a,b,c,d) computes a = d - b*c */
+ spe_fnms(f, term2A_reg, fbA_reg, constA_reg, fbA_reg);
+ break;
+
+ case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE: /* not supported for dest alpha */
+ ASSERT(0);
+ break;
+
+ /* These are special D3D cases involving a second color output
+ * from the fragment shader. I'm not sure we can support them
+ * yet... XXX
+ */
+ case PIPE_BLENDFACTOR_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_SRC1_ALPHA:
+ case PIPE_BLENDFACTOR_INV_SRC1_COLOR:
+ case PIPE_BLENDFACTOR_INV_SRC1_ALPHA:
default:
ASSERT(0);
}
/*
- * Combine Src/Dest RGB terms
+ * Combine Src/Dest RGB terms as per the blend equation.
*/
switch (blend->rgb_func) {
case PIPE_BLEND_ADD:
spe_fs(f, fragB_reg, term2B_reg, term1B_reg);
break;
case PIPE_BLEND_MIN:
- FLOAT_VECTOR_MIN(f, fragR_reg, term1R_reg, term2R_reg)
- FLOAT_VECTOR_MIN(f, fragG_reg, term1G_reg, term2G_reg)
- FLOAT_VECTOR_MIN(f, fragB_reg, term1B_reg, term2B_reg)
+ spe_float_min(f, fragR_reg, term1R_reg, term2R_reg);
+ spe_float_min(f, fragG_reg, term1G_reg, term2G_reg);
+ spe_float_min(f, fragB_reg, term1B_reg, term2B_reg);
break;
case PIPE_BLEND_MAX:
- FLOAT_VECTOR_MAX(f, fragR_reg, term1R_reg, term2R_reg)
- FLOAT_VECTOR_MAX(f, fragG_reg, term1G_reg, term2G_reg)
- FLOAT_VECTOR_MAX(f, fragB_reg, term1B_reg, term2B_reg)
+ spe_float_max(f, fragR_reg, term1R_reg, term2R_reg);
+ spe_float_max(f, fragG_reg, term1G_reg, term2G_reg);
+ spe_float_max(f, fragB_reg, term1B_reg, term2B_reg);
break;
default:
ASSERT(0);
spe_fs(f, fragA_reg, term2A_reg, term1A_reg);
break;
case PIPE_BLEND_MIN:
- FLOAT_VECTOR_MIN(f, fragA_reg, term1A_reg, term2A_reg)
+ spe_float_min(f, fragA_reg, term1A_reg, term2A_reg);
break;
case PIPE_BLEND_MAX:
- FLOAT_VECTOR_MAX(f, fragA_reg, term1A_reg, term2A_reg)
+ spe_float_max(f, fragA_reg, term1A_reg, term2A_reg);
break;
default:
ASSERT(0);
spe_release_register(f, tmp_reg);
/* Free any optional registers that actually got used */
- RELEASE_ONE_REG_IF_USED(f)
- RELEASE_CONST_COLOR_IF_USED(f)
- RELEASE_CONST_ALPHA_IF_USED(f)
+ release_const_register(f, &one_reg_set, one_reg);
+ release_const_register(f, &constR_reg_set, constR_reg);
+ release_const_register(f, &constG_reg_set, constG_reg);
+ release_const_register(f, &constB_reg_set, constB_reg);
+ release_const_register(f, &constA_reg_set, constA_reg);
}
struct spe_function *f,
int fragRGBA_reg, int fbRGBA_reg)
{
- /* XXX to-do */
- /* operate on 32-bit packed pixels, not float colors */
-}
-
-
-static void
-gen_colormask(uint colormask,
- struct spe_function *f,
- int fragRGBA_reg, int fbRGBA_reg)
-{
- /* XXX to-do */
- /* operate on 32-bit packed pixels, not float colors */
+ /* We've got four 32-bit RGBA packed pixels in each of
+ * fragRGBA_reg and fbRGBA_reg, not sets of floating-point
+ * reds, greens, blues, and alphas.
+ * */
+ ASSERT(blend->logicop_enable);
+
+ switch(blend->logicop_func) {
+ case PIPE_LOGICOP_CLEAR: /* 0 */
+ spe_zero(f, fragRGBA_reg);
+ break;
+ case PIPE_LOGICOP_NOR: /* ~(s | d) */
+ spe_nor(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_AND_INVERTED: /* ~s & d */
+ /* andc R, A, B computes R = A & ~B */
+ spe_andc(f, fragRGBA_reg, fbRGBA_reg, fragRGBA_reg);
+ break;
+ case PIPE_LOGICOP_COPY_INVERTED: /* ~s */
+ spe_complement(f, fragRGBA_reg, fragRGBA_reg);
+ break;
+ case PIPE_LOGICOP_AND_REVERSE: /* s & ~d */
+ /* andc R, A, B computes R = A & ~B */
+ spe_andc(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_INVERT: /* ~d */
+ /* Note that (A nor A) == ~(A|A) == ~A */
+ spe_nor(f, fragRGBA_reg, fbRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_XOR: /* s ^ d */
+ spe_xor(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_NAND: /* ~(s & d) */
+ spe_nand(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_AND: /* s & d */
+ spe_and(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_EQUIV: /* ~(s ^ d) */
+ spe_xor(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ spe_complement(f, fragRGBA_reg, fragRGBA_reg);
+ break;
+ case PIPE_LOGICOP_NOOP: /* d */
+ spe_move(f, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_OR_INVERTED: /* ~s | d */
+ /* orc R, A, B computes R = A | ~B */
+ spe_orc(f, fragRGBA_reg, fbRGBA_reg, fragRGBA_reg);
+ break;
+ case PIPE_LOGICOP_COPY: /* s */
+ break;
+ case PIPE_LOGICOP_OR_REVERSE: /* s | ~d */
+ /* orc R, A, B computes R = A | ~B */
+ spe_orc(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_OR: /* s | d */
+ spe_or(f, fragRGBA_reg, fragRGBA_reg, fbRGBA_reg);
+ break;
+ case PIPE_LOGICOP_SET: /* 1 */
+ spe_load_int(f, fragRGBA_reg, 0xffffffff);
+ break;
+ default:
+ ASSERT(0);
+ }
}
-
/**
- * Generate code to pack a quad of float colors into a four 32-bit integers.
+ * Generate code to pack a quad of float colors into four 32-bit integers.
*
* \param f SPE function to append instruction onto.
* \param color_format the dest color packing format
int r_reg, int g_reg, int b_reg, int a_reg,
int rgba_reg)
{
+ int rg_reg = spe_allocate_available_register(f);
+ int ba_reg = spe_allocate_available_register(f);
+
/* Convert float[4] in [0.0,1.0] to int[4] in [0,~0], with clamping */
spe_cfltu(f, r_reg, r_reg, 32);
spe_cfltu(f, g_reg, g_reg, 32);
spe_cfltu(f, b_reg, b_reg, 32);
spe_cfltu(f, a_reg, a_reg, 32);
- /* Shift the most significant bytes to least the significant positions.
+ /* Shift the most significant bytes to the least significant positions.
* I.e.: reg = reg >> 24
*/
spe_rotmi(f, r_reg, r_reg, -24);
* OR-ing all those together gives us four packed colors:
* RGBA = {0xffffffff, 0xaa114477, 0xbb225588, 0xcc336699}
*/
- spe_or(f, rgba_reg, r_reg, g_reg);
- spe_or(f, rgba_reg, rgba_reg, b_reg);
- spe_or(f, rgba_reg, rgba_reg, a_reg);
+ spe_or(f, rg_reg, r_reg, g_reg);
+ spe_or(f, ba_reg, a_reg, b_reg);
+ spe_or(f, rgba_reg, rg_reg, ba_reg);
+
+ spe_release_register(f, rg_reg);
+ spe_release_register(f, ba_reg);
+}
+
+static void
+gen_colormask(struct spe_function *f,
+ uint colormask,
+ enum pipe_format color_format,
+ int fragRGBA_reg, int fbRGBA_reg)
+{
+ /* We've got four 32-bit RGBA packed pixels in each of
+ * fragRGBA_reg and fbRGBA_reg, not sets of floating-point
+ * reds, greens, blues, and alphas. Further, the pixels
+ * are packed according to the given color format, not
+ * necessarily RGBA...
+ */
+ unsigned int r_mask;
+ unsigned int g_mask;
+ unsigned int b_mask;
+ unsigned int a_mask;
+
+ /* Calculate exactly where the bits for any particular color
+ * end up, so we can mask them correctly.
+ */
+ switch(color_format) {
+ case PIPE_FORMAT_A8R8G8B8_UNORM:
+ /* ARGB */
+ a_mask = 0xff000000;
+ r_mask = 0x00ff0000;
+ g_mask = 0x0000ff00;
+ b_mask = 0x000000ff;
+ break;
+ case PIPE_FORMAT_B8G8R8A8_UNORM:
+ /* BGRA */
+ b_mask = 0xff000000;
+ g_mask = 0x00ff0000;
+ r_mask = 0x0000ff00;
+ a_mask = 0x000000ff;
+ break;
+ default:
+ ASSERT(0);
+ }
+
+ /* For each R, G, B, and A component we're supposed to mask out,
+ * clear its bits. Then our mask operation later will work
+ * as expected.
+ */
+ if (!(colormask & PIPE_MASK_R)) {
+ r_mask = 0;
+ }
+ if (!(colormask & PIPE_MASK_G)) {
+ g_mask = 0;
+ }
+ if (!(colormask & PIPE_MASK_B)) {
+ b_mask = 0;
+ }
+ if (!(colormask & PIPE_MASK_A)) {
+ a_mask = 0;
+ }
+
+ /* Get a temporary register to hold the mask that will be applied to the fragment */
+ int colormask_reg = spe_allocate_available_register(f);
+
+ /* The actual mask we're going to use is an OR of the remaining R, G, B, and A
+ * masks. Load the result value into our temporary register.
+ */
+ spe_load_uint(f, colormask_reg, r_mask | g_mask | b_mask | a_mask);
+
+ /* Use the mask register to select between the fragment color
+ * values and the frame buffer color values. Wherever the
+ * mask has a 0 bit, the current frame buffer color should override
+ * the fragment color. Wherever the mask has a 1 bit, the
+ * fragment color should persevere. The Select Bits (selb rt, rA, rB, rM)
+ * instruction will select bits from its first operand rA wherever the
+ * the mask bits rM are 0, and from its second operand rB wherever the
+ * mask bits rM are 1. That means that the frame buffer color is the
+ * first operand, and the fragment color the second.
+ */
+ spe_selb(f, fragRGBA_reg, fbRGBA_reg, fragRGBA_reg, colormask_reg);
+
+ /* Release the temporary register and we're done */
+ spe_release_register(f, colormask_reg);
}
+/* This function is annoyingly similar to gen_depth_test(), above, except
+ * that instead of comparing two varying values (i.e. fragment and buffer),
+ * we're comparing a varying value with a static value. As such, we have
+ * access to the Compare Immediate instructions where we don't in
+ * gen_depth_test(), which is what makes us very different.
+ *
+ * The return value in the stencil_pass_reg is a bitmask of valid
+ * fragments that also passed the stencil test. The bitmask of valid
+ * fragments that failed would be found in (mask_reg & ~stencil_pass_reg).
+ */
+static void
+gen_stencil_test(struct spe_function *f, const struct pipe_stencil_state *state,
+ unsigned int mask_reg, unsigned int fbS_reg,
+ unsigned int stencil_pass_reg)
+{
+ /* Generate code that puts the set of passing fragments into the stencil_pass_reg
+ * register, taking into account whether each fragment was active to begin with.
+ */
+ switch (state->func) {
+ case PIPE_FUNC_EQUAL:
+ /* stencil_pass = mask & (s == reference) */
+ spe_compare_equal_uint(f, stencil_pass_reg, fbS_reg, state->ref_value);
+ spe_and(f, stencil_pass_reg, mask_reg, stencil_pass_reg);
+ break;
+
+ case PIPE_FUNC_NOTEQUAL:
+ /* stencil_pass = mask & ~(s == reference) */
+ spe_compare_equal_uint(f, stencil_pass_reg, fbS_reg, state->ref_value);
+ spe_andc(f, stencil_pass_reg, mask_reg, stencil_pass_reg);
+ break;
+ case PIPE_FUNC_GREATER:
+ /* stencil_pass = mask & (s > reference) */
+ spe_compare_greater_uint(f, stencil_pass_reg, fbS_reg, state->ref_value);
+ spe_and(f, stencil_pass_reg, mask_reg, stencil_pass_reg);
+ break;
+
+ case PIPE_FUNC_LESS: {
+ /* stencil_pass = mask & (reference > s) */
+ /* There's no convenient Compare Less Than Immediate instruction, so
+ * we'll have to do this one the harder way, by loading a register and
+ * comparing directly. Compare Logical Greater Than Word (clgt)
+ * treats its operands as unsigned - no sign extension.
+ */
+ unsigned int tmp_reg = spe_allocate_available_register(f);
+ spe_load_uint(f, tmp_reg, state->ref_value);
+ spe_clgt(f, stencil_pass_reg, tmp_reg, fbS_reg);
+ spe_and(f, stencil_pass_reg, mask_reg, stencil_pass_reg);
+ spe_release_register(f, tmp_reg);
+ break;
+ }
+
+ case PIPE_FUNC_LEQUAL:
+ /* stencil_pass = mask & (s <= reference) = mask & ~(s > reference) */
+ spe_compare_greater_uint(f, stencil_pass_reg, fbS_reg, state->ref_value);
+ spe_andc(f, stencil_pass_reg, mask_reg, stencil_pass_reg);
+ break;
+
+ case PIPE_FUNC_GEQUAL: {
+ /* stencil_pass = mask & (s >= reference) = mask & ~(reference > s) */
+ /* As above, we have to do this by loading a register */
+ unsigned int tmp_reg = spe_allocate_available_register(f);
+ spe_load_uint(f, tmp_reg, state->ref_value);
+ spe_clgt(f, stencil_pass_reg, tmp_reg, fbS_reg);
+ spe_andc(f, stencil_pass_reg, mask_reg, stencil_pass_reg);
+ spe_release_register(f, tmp_reg);
+ break;
+ }
+
+ case PIPE_FUNC_NEVER:
+ /* stencil_pass = mask & 0 = 0 */
+ spe_load_uint(f, stencil_pass_reg, 0);
+ break;
+
+ case PIPE_FUNC_ALWAYS:
+ /* stencil_pass = mask & 1 = mask */
+ spe_move(f, stencil_pass_reg, mask_reg);
+ break;
+ }
+
+ /* The fragments that passed the stencil test are now in stencil_pass_reg.
+ * The fragments that failed would be (mask_reg & ~stencil_pass_reg).
+ */
+}
+
+/* This function generates code that calculates a set of new stencil values
+ * given the earlier values and the operation to apply. It does not
+ * apply any tests. It is intended to be called up to 3 times
+ * (for the stencil fail operation, for the stencil pass-z fail operation,
+ * and for the stencil pass-z pass operation) to collect up to three
+ * possible sets of values, and for the caller to combine them based
+ * on the result of the tests.
+ *
+ * stencil_max_value should be (2^n - 1) where n is the number of bits
+ * in the stencil buffer - in other words, it should be usable as a mask.
+ */
+static void
+gen_stencil_values(struct spe_function *f, unsigned int stencil_op,
+ unsigned int stencil_ref_value, unsigned int stencil_max_value,
+ unsigned int fbS_reg, unsigned int newS_reg)
+{
+ /* The code below assumes that newS_reg and fbS_reg are not the same
+ * register; if they can be, the calculations below will have to use
+ * an additional temporary register. For now, mark the assumption
+ * with an assertion that will fail if they are the same.
+ */
+ ASSERT(fbS_reg != newS_reg);
+
+ /* The code also assumes the the stencil_max_value is of the form
+ * 2^n-1 and can therefore be used as a mask for the valid bits in
+ * addition to a maximum. Make sure this is the case as well.
+ * The clever math below exploits the fact that incrementing a
+ * binary number serves to flip all the bits of a number starting at
+ * the LSB and continuing to (and including) the first zero bit
+ * found. That means that a number and its increment will always
+ * have at least one bit in common (the high order bit, if nothing
+ * else) *unless* the number is zero, *or* the number is of a form
+ * consisting of some number of 1s in the low-order bits followed
+ * by nothing but 0s in the high-order bits. The latter case
+ * implies it's of the form 2^n-1.
+ */
+ ASSERT(stencil_max_value > 0 && ((stencil_max_value + 1) & stencil_max_value) == 0);
+
+ switch(stencil_op) {
+ case PIPE_STENCIL_OP_KEEP:
+ /* newS = S */
+ spe_move(f, newS_reg, fbS_reg);
+ break;
+
+ case PIPE_STENCIL_OP_ZERO:
+ /* newS = 0 */
+ spe_zero(f, newS_reg);
+ break;
+
+ case PIPE_STENCIL_OP_REPLACE:
+ /* newS = stencil reference value */
+ spe_load_uint(f, newS_reg, stencil_ref_value);
+ break;
+
+ case PIPE_STENCIL_OP_INCR: {
+ /* newS = (s == max ? max : s + 1) */
+ unsigned int equals_reg = spe_allocate_available_register(f);
+
+ spe_compare_equal_uint(f, equals_reg, fbS_reg, stencil_max_value);
+ /* Add Word Immediate computes rT = rA + 10-bit signed immediate */
+ spe_ai(f, newS_reg, fbS_reg, 1);
+ /* Select from the current value or the new value based on the equality test */
+ spe_selb(f, newS_reg, fbS_reg, newS_reg, equals_reg);
+
+ spe_release_register(f, equals_reg);
+ break;
+ }
+ case PIPE_STENCIL_OP_DECR: {
+ /* newS = (s == 0 ? 0 : s - 1) */
+ unsigned int equals_reg = spe_allocate_available_register(f);
+
+ spe_compare_equal_uint(f, equals_reg, fbS_reg, 0);
+ /* Add Word Immediate with a (-1) value works */
+ spe_ai(f, newS_reg, fbS_reg, -1);
+ /* Select from the current value or the new value based on the equality test */
+ spe_selb(f, newS_reg, fbS_reg, newS_reg, equals_reg);
+
+ spe_release_register(f, equals_reg);
+ break;
+ }
+ case PIPE_STENCIL_OP_INCR_WRAP:
+ /* newS = (s == max ? 0 : s + 1), but since max is 2^n-1, we can
+ * do a normal add and mask off the correct bits
+ */
+ spe_ai(f, newS_reg, fbS_reg, 1);
+ spe_and_uint(f, newS_reg, newS_reg, stencil_max_value);
+ break;
+
+ case PIPE_STENCIL_OP_DECR_WRAP:
+ /* newS = (s == 0 ? max : s - 1), but we'll pull the same mask trick as above */
+ spe_ai(f, newS_reg, fbS_reg, -1);
+ spe_and_uint(f, newS_reg, newS_reg, stencil_max_value);
+ break;
+
+ case PIPE_STENCIL_OP_INVERT:
+ /* newS = ~s. We take advantage of the mask/max value to invert only
+ * the valid bits for the field so we don't have to do an extra "and".
+ */
+ spe_xor_uint(f, newS_reg, fbS_reg, stencil_max_value);
+ break;
+
+ default:
+ ASSERT(0);
+ }
+}
+
+
+/* This function generates code to get all the necessary possible
+ * stencil values. For each of the output registers (fail_reg,
+ * zfail_reg, and zpass_reg), it either allocates a new register
+ * and calculates a new set of values based on the stencil operation,
+ * or it reuses a register allocation and calculation done for an
+ * earlier (matching) operation, or it reuses the fbS_reg register
+ * (if the stencil operation is KEEP, which doesn't change the
+ * stencil buffer).
+ *
+ * Since this function allocates a variable number of registers,
+ * to avoid incurring complex logic to free them, they should
+ * be allocated after a spe_allocate_register_set() call
+ * and released by the corresponding spe_release_register_set() call.
+ */
+static void
+gen_get_stencil_values(struct spe_function *f, const struct pipe_depth_stencil_alpha_state *dsa,
+ unsigned int fbS_reg,
+ unsigned int *fail_reg, unsigned int *zfail_reg,
+ unsigned int *zpass_reg, unsigned int *back_fail_reg,
+ unsigned int *back_zfail_reg, unsigned int *back_zpass_reg)
+{
+ unsigned zfail_op, back_zfail_op;
+
+ /* Stenciling had better be enabled here */
+ ASSERT(dsa->stencil[0].enabled);
+
+ /* If the depth test is not enabled, it is treated as though it always
+ * passes. In particular, that means that the "zfail_op" (and the backfacing
+ * counterpart, if active) are not considered - a failing stencil test will
+ * trigger the "fail_op", and a passing stencil test will trigger the
+ * "zpass_op".
+ *
+ * By overriding the operations in this case to be PIPE_STENCIL_OP_KEEP,
+ * we keep them from being calculated.
+ */
+ if (dsa->depth.enabled) {
+ zfail_op = dsa->stencil[0].zfail_op;
+ back_zfail_op = dsa->stencil[1].zfail_op;
+ }
+ else {
+ zfail_op = PIPE_STENCIL_OP_KEEP;
+ back_zfail_op = PIPE_STENCIL_OP_KEEP;
+ }
+
+ /* One-sided or front-facing stencil */
+ if (dsa->stencil[0].fail_op == PIPE_STENCIL_OP_KEEP) {
+ *fail_reg = fbS_reg;
+ }
+ else {
+ *fail_reg = spe_allocate_available_register(f);
+ gen_stencil_values(f, dsa->stencil[0].fail_op, dsa->stencil[0].ref_value,
+ 0xff, fbS_reg, *fail_reg);
+ }
+
+ if (zfail_op == PIPE_STENCIL_OP_KEEP) {
+ *zfail_reg = fbS_reg;
+ }
+ else if (zfail_op == dsa->stencil[0].fail_op) {
+ *zfail_reg = *fail_reg;
+ }
+ else {
+ *zfail_reg = spe_allocate_available_register(f);
+ gen_stencil_values(f, dsa->stencil[0].zfail_op, dsa->stencil[0].ref_value,
+ 0xff, fbS_reg, *zfail_reg);
+ }
+
+ if (dsa->stencil[0].zpass_op == PIPE_STENCIL_OP_KEEP) {
+ *zpass_reg = fbS_reg;
+ }
+ else if (dsa->stencil[0].zpass_op == dsa->stencil[0].fail_op) {
+ *zpass_reg = *fail_reg;
+ }
+ else if (dsa->stencil[0].zpass_op == zfail_op) {
+ *zpass_reg = *zfail_reg;
+ }
+ else {
+ *zpass_reg = spe_allocate_available_register(f);
+ gen_stencil_values(f, dsa->stencil[0].zpass_op, dsa->stencil[0].ref_value,
+ 0xff, fbS_reg, *zpass_reg);
+ }
+
+ /* If two-sided stencil is enabled, we have more work to do. */
+ if (!dsa->stencil[1].enabled) {
+ /* This just flags that the registers need not be deallocated later */
+ *back_fail_reg = fbS_reg;
+ *back_zfail_reg = fbS_reg;
+ *back_zpass_reg = fbS_reg;
+ }
+ else {
+ /* Same calculations as above, but for the back stencil */
+ if (dsa->stencil[1].fail_op == PIPE_STENCIL_OP_KEEP) {
+ *back_fail_reg = fbS_reg;
+ }
+ else if (dsa->stencil[1].fail_op == dsa->stencil[0].fail_op) {
+ *back_fail_reg = *fail_reg;
+ }
+ else if (dsa->stencil[1].fail_op == zfail_op) {
+ *back_fail_reg = *zfail_reg;
+ }
+ else if (dsa->stencil[1].fail_op == dsa->stencil[0].zpass_op) {
+ *back_fail_reg = *zpass_reg;
+ }
+ else {
+ *back_fail_reg = spe_allocate_available_register(f);
+ gen_stencil_values(f, dsa->stencil[1].fail_op, dsa->stencil[1].ref_value,
+ 0xff, fbS_reg, *back_fail_reg);
+ }
+
+ if (back_zfail_op == PIPE_STENCIL_OP_KEEP) {
+ *back_zfail_reg = fbS_reg;
+ }
+ else if (back_zfail_op == dsa->stencil[0].fail_op) {
+ *back_zfail_reg = *fail_reg;
+ }
+ else if (back_zfail_op == zfail_op) {
+ *back_zfail_reg = *zfail_reg;
+ }
+ else if (back_zfail_op == dsa->stencil[0].zpass_op) {
+ *back_zfail_reg = *zpass_reg;
+ }
+ else if (back_zfail_op == dsa->stencil[1].fail_op) {
+ *back_zfail_reg = *back_fail_reg;
+ }
+ else {
+ *back_zfail_reg = spe_allocate_available_register(f);
+ gen_stencil_values(f, dsa->stencil[1].zfail_op, dsa->stencil[1].ref_value,
+ 0xff, fbS_reg, *back_zfail_reg);
+ }
+
+ if (dsa->stencil[1].zpass_op == PIPE_STENCIL_OP_KEEP) {
+ *back_zpass_reg = fbS_reg;
+ }
+ else if (dsa->stencil[1].zpass_op == dsa->stencil[0].fail_op) {
+ *back_zpass_reg = *fail_reg;
+ }
+ else if (dsa->stencil[1].zpass_op == zfail_op) {
+ *back_zpass_reg = *zfail_reg;
+ }
+ else if (dsa->stencil[1].zpass_op == dsa->stencil[0].zpass_op) {
+ *back_zpass_reg = *zpass_reg;
+ }
+ else if (dsa->stencil[1].zpass_op == dsa->stencil[1].fail_op) {
+ *back_zpass_reg = *back_fail_reg;
+ }
+ else if (dsa->stencil[1].zpass_op == back_zfail_op) {
+ *back_zpass_reg = *back_zfail_reg;
+ }
+ else {
+ *back_zfail_reg = spe_allocate_available_register(f);
+ gen_stencil_values(f, dsa->stencil[1].zpass_op, dsa->stencil[1].ref_value,
+ 0xff, fbS_reg, *back_zpass_reg);
+ }
+ } /* End of calculations for back-facing stencil */
+}
+
+/* Note that fbZ_reg may *not* be set on entry, if in fact
+ * the depth test is not enabled. This function must not use
+ * the register if depth is not enabled.
+ */
+static boolean
+gen_stencil_depth_test(struct spe_function *f,
+ const struct pipe_depth_stencil_alpha_state *dsa,
+ const int const facing_reg,
+ const int mask_reg, const int fragZ_reg,
+ const int fbZ_reg, const int fbS_reg)
+{
+ /* True if we've generated code that could require writeback to the
+ * depth and/or stencil buffers
+ */
+ boolean modified_buffers = false;
+
+ boolean need_to_calculate_stencil_values;
+ boolean need_to_writemask_stencil_values;
+
+ /* Registers. We may or may not actually allocate these, depending
+ * on whether the state values indicate that we need them.
+ */
+ unsigned int stencil_pass_reg, stencil_fail_reg;
+ unsigned int stencil_fail_values, stencil_pass_depth_fail_values, stencil_pass_depth_pass_values;
+ unsigned int stencil_writemask_reg;
+ unsigned int zmask_reg;
+ unsigned int newS_reg;
+
+ /* Stenciling is quite complex: up to six different configurable stencil
+ * operations/calculations can be required (three each for front-facing
+ * and back-facing fragments). Many of those operations will likely
+ * be identical, so there's good reason to try to avoid calculating
+ * the same values more than once (which unfortunately makes the code less
+ * straightforward).
+ *
+ * To make register management easier, we start a new
+ * register set; we can release all the registers in the set at
+ * once, and avoid having to keep track of exactly which registers
+ * we allocate. We can still allocate and free registers as
+ * desired (if we know we no longer need a register), but we don't
+ * have to spend the complexity to track the more difficult variant
+ * register usage scenarios.
+ */
+ spe_comment(f, 0, "Allocating stencil register set");
+ spe_allocate_register_set(f);
+
+ /* Calculate the writemask. If the writemask is trivial (either
+ * all 0s, meaning that we don't need to calculate any stencil values
+ * because they're not going to change the stencil anyway, or all 1s,
+ * meaning that we have to calculate the stencil values but do not
+ * need to mask them), we can avoid generating code. Don't forget
+ * that we need to consider backfacing stencil, if enabled.
+ */
+ if (dsa->stencil[0].write_mask == 0x0 && (!dsa->stencil[1].enabled || dsa->stencil[1].write_mask == 0x00)) {
+ /* Trivial: don't need to calculate stencil values, and don't need to
+ * write them back to the framebuffer.
+ */
+ need_to_calculate_stencil_values = false;
+ need_to_writemask_stencil_values = false;
+ }
+ else if (dsa->stencil[0].write_mask == 0xff && (!dsa->stencil[1].enabled || dsa->stencil[1].write_mask == 0xff)) {
+ /* Still trivial, but a little less so. We need to write the stencil
+ * values, but we don't need to mask them.
+ */
+ need_to_calculate_stencil_values = true;
+ need_to_writemask_stencil_values = false;
+ }
+ else {
+ /* The general case: calculate, mask, and write */
+ need_to_calculate_stencil_values = true;
+ need_to_writemask_stencil_values = true;
+
+ /* While we're here, generate code that calculates what the
+ * writemask should be. If backface stenciling is enabled,
+ * and the backface writemask is not the same as the frontface
+ * writemask, we'll have to generate code that merges the
+ * two masks into a single effective mask based on fragment facing.
+ */
+ spe_comment(f, 0, "Computing stencil writemask");
+ stencil_writemask_reg = spe_allocate_available_register(f);
+ spe_load_uint(f, stencil_writemask_reg, dsa->stencil[0].write_mask);
+ if (dsa->stencil[1].enabled && dsa->stencil[0].write_mask != dsa->stencil[1].write_mask) {
+ unsigned int back_write_mask_reg = spe_allocate_available_register(f);
+ spe_comment(f, 0, "Resolving two-sided stencil writemask");
+ spe_load_uint(f, back_write_mask_reg, dsa->stencil[1].write_mask);
+ spe_selb(f, stencil_writemask_reg, stencil_writemask_reg, back_write_mask_reg, facing_reg);
+ spe_release_register(f, back_write_mask_reg);
+ }
+ }
+
+ /* At least one-sided stenciling must be on. Generate code that
+ * runs the stencil test on the basic/front-facing stencil, leaving
+ * the mask of passing stencil bits in stencil_pass_reg. This mask will
+ * be used both to mask the set of active pixels, and also to
+ * determine how the stencil buffer changes.
+ *
+ * This test will *not* change the value in mask_reg (because we don't
+ * yet know whether to apply the two-sided stencil or one-sided stencil).
+ */
+ spe_comment(f, 0, "Running basic stencil test");
+ stencil_pass_reg = spe_allocate_available_register(f);
+ gen_stencil_test(f, &dsa->stencil[0], mask_reg, fbS_reg, stencil_pass_reg);
+
+ /* If two-sided stenciling is on, generate code to run the stencil
+ * test on the backfacing stencil as well, and combine the two results
+ * into the one correct result based on facing.
+ */
+ if (dsa->stencil[1].enabled) {
+ unsigned int temp_reg = spe_allocate_available_register(f);
+ spe_comment(f, 0, "Running backface stencil test");
+ gen_stencil_test(f, &dsa->stencil[1], mask_reg, fbS_reg, temp_reg);
+ spe_selb(f, stencil_pass_reg, stencil_pass_reg, temp_reg, facing_reg);
+ spe_release_register(f, temp_reg);
+ }
+
+ /* Generate code that, given the mask of valid fragments and the
+ * mask of valid fragments that passed the stencil test, computes
+ * the mask of valid fragments that failed the stencil test. We
+ * have to do this before we run a depth test (because the
+ * depth test should not be performed on fragments that failed the
+ * stencil test, and because the depth test will update the
+ * mask of valid fragments based on the results of the depth test).
+ */
+ spe_comment(f, 0, "Computing stencil fail mask and updating fragment mask");
+ stencil_fail_reg = spe_allocate_available_register(f);
+ spe_andc(f, stencil_fail_reg, mask_reg, stencil_pass_reg);
+ /* Now remove the stenciled-out pixels from the valid fragment mask,
+ * so we can later use the valid fragment mask in the depth test.
+ */
+ spe_and(f, mask_reg, mask_reg, stencil_pass_reg);
+
+ /* We may not need to calculate stencil values, if the writemask is off */
+ if (need_to_calculate_stencil_values) {
+ unsigned int back_stencil_fail_values, back_stencil_pass_depth_fail_values, back_stencil_pass_depth_pass_values;
+ unsigned int front_stencil_fail_values, front_stencil_pass_depth_fail_values, front_stencil_pass_depth_pass_values;
+
+ /* Generate code that calculates exactly which stencil values we need,
+ * without calculating the same value twice (say, if two different
+ * stencil ops have the same value). This code will work for one-sided
+ * and two-sided stenciling (so that we take into account that operations
+ * may match between front and back stencils), and will also take into
+ * account whether the depth test is enabled (if the depth test is off,
+ * we don't need any of the zfail results, because the depth test always
+ * is considered to pass if it is disabled). Any register value that
+ * does not need to be calculated will come back with the same value
+ * that's in fbS_reg.
+ *
+ * This function will allocate a variant number of registers that
+ * will be released as part of the register set.
+ */
+ spe_comment(f, 0, "Computing stencil values");
+ gen_get_stencil_values(f, dsa, fbS_reg,
+ &front_stencil_fail_values, &front_stencil_pass_depth_fail_values,
+ &front_stencil_pass_depth_pass_values, &back_stencil_fail_values,
+ &back_stencil_pass_depth_fail_values, &back_stencil_pass_depth_pass_values);
+
+ /* Tricky, tricky, tricky - the things we do to create optimal
+ * code...
+ *
+ * The various stencil values registers may overlap with each other
+ * and with fbS_reg arbitrarily (as any particular operation is
+ * only calculated once and stored in one register, no matter
+ * how many times it is used). So we can't change the values
+ * within those registers directly - if we change a value in a
+ * register that's being referenced by two different calculations,
+ * we've just unwittingly changed the second value as well...
+ *
+ * Avoid this by allocating new registers to hold the results
+ * (there may be 2, if the depth test is off, or 3, if it is on).
+ * These will be released as part of the register set.
+ */
+ if (!dsa->stencil[1].enabled) {
+ /* The easy case: if two-sided stenciling is *not* enabled, we
+ * just use the front-sided values.
+ */
+ stencil_fail_values = front_stencil_fail_values;
+ stencil_pass_depth_fail_values = front_stencil_pass_depth_fail_values;
+ stencil_pass_depth_pass_values = front_stencil_pass_depth_pass_values;
+ }
+ else { /* two-sided stencil enabled */
+ spe_comment(f, 0, "Resolving backface stencil values");
+ /* Allocate new registers for the needed merged values */
+ stencil_fail_values = spe_allocate_available_register(f);
+ spe_selb(f, stencil_fail_values, front_stencil_fail_values, back_stencil_fail_values, facing_reg);
+ if (dsa->depth.enabled) {
+ stencil_pass_depth_fail_values = spe_allocate_available_register(f);
+ spe_selb(f, stencil_pass_depth_fail_values, front_stencil_pass_depth_fail_values, back_stencil_pass_depth_fail_values, facing_reg);
+ }
+ else {
+ stencil_pass_depth_fail_values = fbS_reg;
+ }
+ stencil_pass_depth_pass_values = spe_allocate_available_register(f);
+ spe_selb(f, stencil_pass_depth_pass_values, front_stencil_pass_depth_pass_values, back_stencil_pass_depth_pass_values, facing_reg);
+ }
+ }
+
+ /* We now have all the stencil values we need. We also need
+ * the results of the depth test to figure out which
+ * stencil values will become the new stencil values. (Even if
+ * we aren't actually calculating stencil values, we need to apply
+ * the depth test if it's enabled.)
+ *
+ * The code generated by gen_depth_test() returns the results of the
+ * test in the given register, but also alters the mask_reg based
+ * on the results of the test.
+ */
+ if (dsa->depth.enabled) {
+ spe_comment(f, 0, "Running stencil depth test");
+ zmask_reg = spe_allocate_available_register(f);
+ modified_buffers |= gen_depth_test(f, dsa, mask_reg, fragZ_reg, fbZ_reg, zmask_reg);
+ }
+
+ if (need_to_calculate_stencil_values) {
+
+ /* If we need to writemask the stencil values before going into
+ * the stencil buffer, we'll have to use a new register to
+ * hold the new values. If not, we can just keep using the
+ * current register.
+ */
+ if (need_to_writemask_stencil_values) {
+ newS_reg = spe_allocate_available_register(f);
+ spe_comment(f, 0, "Saving current stencil values for writemasking");
+ spe_move(f, newS_reg, fbS_reg);
+ }
+ else {
+ newS_reg = fbS_reg;
+ }
+
+ /* Merge in the selected stencil fail values */
+ if (stencil_fail_values != fbS_reg) {
+ spe_comment(f, 0, "Loading stencil fail values");
+ spe_selb(f, newS_reg, newS_reg, stencil_fail_values, stencil_fail_reg);
+ modified_buffers = true;
+ }
+
+ /* Same for the stencil pass/depth fail values. If this calculation
+ * is not needed (say, if depth test is off), then the
+ * stencil_pass_depth_fail_values register will be equal to fbS_reg
+ * and we'll skip the calculation.
+ */
+ if (stencil_pass_depth_fail_values != fbS_reg) {
+ /* We don't actually have a stencil pass/depth fail mask yet.
+ * Calculate it here from the stencil passing mask and the
+ * depth passing mask. Note that zmask_reg *must* have been
+ * set above if we're here.
+ */
+ unsigned int stencil_pass_depth_fail_mask = spe_allocate_available_register(f);
+ spe_comment(f, 0, "Loading stencil pass/depth fail values");
+ spe_andc(f, stencil_pass_depth_fail_mask, stencil_pass_reg, zmask_reg);
+
+ spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_fail_values, stencil_pass_depth_fail_mask);
+
+ spe_release_register(f, stencil_pass_depth_fail_mask);
+ modified_buffers = true;
+ }
+
+ /* Same for the stencil pass/depth pass mask. Note that we
+ * *can* get here with zmask_reg being unset (if the depth
+ * test is off but the stencil test is on). In this case,
+ * we assume the depth test passes, and don't need to mask
+ * the stencil pass mask with the Z mask.
+ */
+ if (stencil_pass_depth_pass_values != fbS_reg) {
+ if (dsa->depth.enabled) {
+ unsigned int stencil_pass_depth_pass_mask = spe_allocate_available_register(f);
+ /* We'll need a separate register */
+ spe_comment(f, 0, "Loading stencil pass/depth pass values");
+ spe_and(f, stencil_pass_depth_pass_mask, stencil_pass_reg, zmask_reg);
+ spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_pass_values, stencil_pass_depth_pass_mask);
+ spe_release_register(f, stencil_pass_depth_pass_mask);
+ }
+ else {
+ /* We can use the same stencil-pass register */
+ spe_comment(f, 0, "Loading stencil pass values");
+ spe_selb(f, newS_reg, newS_reg, stencil_pass_depth_pass_values, stencil_pass_reg);
+ }
+ modified_buffers = true;
+ }
+
+ /* Almost done. If we need to writemask, do it now, leaving the
+ * results in the fbS_reg register passed in. If we don't need
+ * to writemask, then the results are *already* in the fbS_reg,
+ * so there's nothing more to do.
+ */
+
+ if (need_to_writemask_stencil_values && modified_buffers) {
+ /* The Select Bytes command makes a fine writemask. Where
+ * the mask is 0, the first (original) values are retained,
+ * effectively masking out changes. Where the mask is 1, the
+ * second (new) values are retained, incorporating changes.
+ */
+ spe_comment(f, 0, "Writemasking new stencil values");
+ spe_selb(f, fbS_reg, fbS_reg, newS_reg, stencil_writemask_reg);
+ }
+
+ } /* done calculating stencil values */
+
+ /* The stencil and/or depth values have been applied, and the
+ * mask_reg, fbS_reg, and fbZ_reg values have been updated.
+ * We're all done, except that we've allocated a fair number
+ * of registers that we didn't bother tracking. Release all
+ * those registers as part of the register set, and go home.
+ */
+ spe_comment(f, 0, "Releasing stencil register set");
+ spe_release_register_set(f);
+
+ /* Return true if we could have modified the stencil and/or
+ * depth buffers.
+ */
+ return modified_buffers;
+}
/**
void
cell_gen_fragment_function(struct cell_context *cell, struct spe_function *f)
{
- const struct pipe_depth_stencil_alpha_state *dsa =
- &cell->depth_stencil->base;
- const struct pipe_blend_state *blend = &cell->blend->base;
+ const struct pipe_depth_stencil_alpha_state *dsa = cell->depth_stencil;
+ const struct pipe_blend_state *blend = cell->blend;
const struct pipe_blend_color *blend_color = &cell->blend_color;
const enum pipe_format color_format = cell->framebuffer.cbufs[0]->format;
const int fragB_reg = 10; /* vector float */
const int fragA_reg = 11; /* vector float */
const int mask_reg = 12; /* vector uint */
+ const int facing_reg = 13; /* uint */
/* offset of quad from start of tile
* XXX assuming 4-byte pixels for color AND Z/stencil!!!!
spe_allocate_register(f, fragB_reg);
spe_allocate_register(f, fragA_reg);
spe_allocate_register(f, mask_reg);
+ spe_allocate_register(f, facing_reg);
quad_offset_reg = spe_allocate_available_register(f);
fbRGBA_reg = spe_allocate_available_register(f);
ASSERT(TILE_SIZE == 32);
- spe_rotmi(f, x2_reg, x_reg, -1); /* x2 = x / 2 */
+ spe_comment(f, 0, "Compute quad offset within tile");
spe_rotmi(f, y2_reg, y_reg, -1); /* y2 = y / 2 */
+ spe_rotmi(f, x2_reg, x_reg, -1); /* x2 = x / 2 */
spe_shli(f, y2_reg, y2_reg, 4); /* y2 *= 16 */
spe_a(f, quad_offset_reg, y2_reg, x2_reg); /* offset = y2 + x2 */
spe_shli(f, quad_offset_reg, quad_offset_reg, 4); /* offset *= 16 */
spe_release_register(f, y2_reg);
}
-
if (dsa->alpha.enabled) {
gen_alpha_test(dsa, f, mask_reg, fragA_reg);
}
+ /* If we need the stencil buffers (because one- or two-sided stencil is
+ * enabled) or the depth buffer (because the depth test is enabled),
+ * go grab them. Note that if either one- or two-sided stencil is
+ * enabled, dsa->stencil[0].enabled will be true.
+ */
if (dsa->depth.enabled || dsa->stencil[0].enabled) {
const enum pipe_format zs_format = cell->framebuffer.zsbuf->format;
boolean write_depth_stencil;
- int fbZ_reg = spe_allocate_available_register(f); /* Z values */
- int fbS_reg = spe_allocate_available_register(f); /* Stencil values */
+ /* We may or may not need to allocate a register for Z or stencil values */
+ boolean fbS_reg_set = false, fbZ_reg_set = false;
+ unsigned int fbS_reg, fbZ_reg = 0;
+
+ spe_comment(f, 0, "Fetching Z/stencil quad from tile");
/* fetch quad of depth/stencil values from tile at (x,y) */
/* Load: fbZS_reg = memory[depth_tile_reg + offset_reg] */
+ /* XXX Not sure this is allowed if we've only got a 16-bit Z buffer... */
spe_lqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg);
- if (dsa->depth.enabled) {
- /* Extract Z bits from fbZS_reg into fbZ_reg */
- if (zs_format == PIPE_FORMAT_S8Z24_UNORM ||
- zs_format == PIPE_FORMAT_X8Z24_UNORM) {
- int mask_reg = spe_allocate_available_register(f);
- spe_fsmbi(f, mask_reg, 0x7777); /* mask[0,1,2,3] = 0x00ffffff */
- spe_and(f, fbZ_reg, fbZS_reg, mask_reg); /* fbZ = fbZS & mask */
- spe_release_register(f, mask_reg);
- /* OK, fbZ_reg has four 24-bit Z values now */
- }
- else {
- /* XXX handle other z/stencil formats */
- ASSERT(0);
- }
-
- /* Convert fragZ values from float[4] to uint[4] */
- if (zs_format == PIPE_FORMAT_S8Z24_UNORM ||
- zs_format == PIPE_FORMAT_X8Z24_UNORM ||
- zs_format == PIPE_FORMAT_Z24S8_UNORM ||
- zs_format == PIPE_FORMAT_Z24X8_UNORM) {
- /* 24-bit Z values */
- int scale_reg = spe_allocate_available_register(f);
-
- /* scale_reg[0,1,2,3] = float(2^24-1) */
- spe_load_float(f, scale_reg, (float) 0xffffff);
-
- /* XXX these two instructions might be combined */
- spe_fm(f, fragZ_reg, fragZ_reg, scale_reg); /* fragZ *= scale */
- spe_cfltu(f, fragZ_reg, fragZ_reg, 0); /* fragZ = (int) fragZ */
-
- spe_release_register(f, scale_reg);
- }
- else {
- /* XXX handle 16-bit Z format */
- ASSERT(0);
- }
+ /* From the Z/stencil buffer format, pull out the bits we need for
+ * Z and/or stencil. We'll also convert the incoming fragment Z
+ * value in fragZ_reg from a floating point value in [0.0..1.0] to
+ * an unsigned integer value with the appropriate resolution.
+ */
+ switch(zs_format) {
+
+ case PIPE_FORMAT_S8Z24_UNORM: /* fall through */
+ case PIPE_FORMAT_X8Z24_UNORM:
+ if (dsa->depth.enabled) {
+ /* We need the Z part at least */
+ setup_optional_register(f, &fbZ_reg_set, &fbZ_reg);
+ /* four 24-bit Z values in the low-order bits */
+ spe_and_uint(f, fbZ_reg, fbZS_reg, 0x00ffffff);
+
+ /* Incoming fragZ_reg value is a float in 0.0...1.0; convert
+ * to a 24-bit unsigned integer
+ */
+ spe_cfltu(f, fragZ_reg, fragZ_reg, 32);
+ spe_rotmi(f, fragZ_reg, fragZ_reg, -8);
+ }
+ if (dsa->stencil[0].enabled) {
+ setup_optional_register(f, &fbS_reg_set, &fbS_reg);
+ /* four 8-bit Z values in the high-order bits */
+ spe_rotmi(f, fbS_reg, fbZS_reg, -24);
+ }
+ break;
+
+ case PIPE_FORMAT_Z24S8_UNORM: /* fall through */
+ case PIPE_FORMAT_Z24X8_UNORM:
+ if (dsa->depth.enabled) {
+ setup_optional_register(f, &fbZ_reg_set, &fbZ_reg);
+ /* shift by 8 to get the upper 24-bit values */
+ spe_rotmi(f, fbS_reg, fbZS_reg, -8);
+
+ /* Incoming fragZ_reg value is a float in 0.0...1.0; convert
+ * to a 24-bit unsigned integer
+ */
+ spe_cfltu(f, fragZ_reg, fragZ_reg, 32);
+ spe_rotmi(f, fragZ_reg, fragZ_reg, -8);
+ }
+ if (dsa->stencil[0].enabled) {
+ setup_optional_register(f, &fbS_reg_set, &fbS_reg);
+ /* 8-bit stencil in the low-order bits - mask them out */
+ spe_and_uint(f, fbS_reg, fbZS_reg, 0x000000ff);
+ }
+ break;
+
+ case PIPE_FORMAT_Z32_UNORM:
+ if (dsa->depth.enabled) {
+ setup_optional_register(f, &fbZ_reg_set, &fbZ_reg);
+ /* Copy over 4 32-bit values */
+ spe_move(f, fbZ_reg, fbZS_reg);
+
+ /* Incoming fragZ_reg value is a float in 0.0...1.0; convert
+ * to a 32-bit unsigned integer
+ */
+ spe_cfltu(f, fragZ_reg, fragZ_reg, 32);
+ }
+ /* No stencil, so can't do anything there */
+ break;
+
+ case PIPE_FORMAT_Z16_UNORM:
+ if (dsa->depth.enabled) {
+ /* XXX Not sure this is correct, but it was here before, so we're
+ * going with it for now
+ */
+ setup_optional_register(f, &fbZ_reg_set, &fbZ_reg);
+ /* Copy over 4 32-bit values */
+ spe_move(f, fbZ_reg, fbZS_reg);
+
+ /* Incoming fragZ_reg value is a float in 0.0...1.0; convert
+ * to a 16-bit unsigned integer
+ */
+ spe_cfltu(f, fragZ_reg, fragZ_reg, 32);
+ spe_rotmi(f, fragZ_reg, fragZ_reg, -16);
+ }
+ /* No stencil */
+ break;
+
+ default:
+ ASSERT(0); /* invalid format */
}
+ /* If stencil is enabled, use the stencil-specific code
+ * generator to generate both the stencil and depth (if needed)
+ * tests. Otherwise, if only depth is enabled, generate
+ * a quick depth test. The test generators themselves will
+ * report back whether the depth/stencil buffer has to be
+ * written back.
+ */
if (dsa->stencil[0].enabled) {
- /* Extract Stencil bit sfrom fbZS_reg into fbS_reg */
- if (zs_format == PIPE_FORMAT_S8Z24_UNORM ||
- zs_format == PIPE_FORMAT_X8Z24_UNORM) {
- /* XXX extract with a shift */
- ASSERT(0);
- }
- else if (zs_format == PIPE_FORMAT_Z24S8_UNORM ||
- zs_format == PIPE_FORMAT_Z24X8_UNORM) {
- /* XXX extract with a mask */
- ASSERT(0);
- }
- }
-
+ /* This will perform the stencil and depth tests, and update
+ * the mask_reg, fbZ_reg, and fbS_reg as required by the
+ * tests.
+ */
+ ASSERT(fbS_reg_set);
+ spe_comment(f, 0, "Perform stencil test");
- if (dsa->stencil[0].enabled) {
- /* XXX this may involve depth testing too */
- // gen_stencil_test(dsa, f, ... );
- ASSERT(0);
+ /* Note that fbZ_reg may not be set on entry, if stenciling
+ * is enabled but there's no Z-buffer. The
+ * gen_stencil_depth_test() function must ignore the
+ * fbZ_reg register if depth is not enabled.
+ */
+ write_depth_stencil = gen_stencil_depth_test(f, dsa, facing_reg, mask_reg, fragZ_reg, fbZ_reg, fbS_reg);
}
else if (dsa->depth.enabled) {
int zmask_reg = spe_allocate_available_register(f);
- gen_depth_test(dsa, f, mask_reg, fragZ_reg, fbZ_reg, zmask_reg);
+ ASSERT(fbZ_reg_set);
+ spe_comment(f, 0, "Perform depth test");
+ write_depth_stencil = gen_depth_test(f, dsa, mask_reg, fragZ_reg, fbZ_reg, zmask_reg);
spe_release_register(f, zmask_reg);
}
-
- /* do we need to write Z and/or Stencil back into framebuffer? */
- write_depth_stencil = (dsa->depth.writemask |
- dsa->stencil[0].write_mask |
- dsa->stencil[1].write_mask);
+ else {
+ write_depth_stencil = false;
+ }
if (write_depth_stencil) {
/* Merge latest Z and Stencil values into fbZS_reg.
* fbZ_reg has four Z vals in bits [23..0] or bits [15..0].
* fbS_reg has four 8-bit Z values in bits [7..0].
*/
+ spe_comment(f, 0, "Store quad's depth/stencil values in tile");
if (zs_format == PIPE_FORMAT_S8Z24_UNORM ||
zs_format == PIPE_FORMAT_X8Z24_UNORM) {
- spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */
- spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */
+ if (fbS_reg_set && fbZ_reg_set) {
+ spe_shli(f, fbS_reg, fbS_reg, 24); /* fbS = fbS << 24 */
+ spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */
+ }
+ else if (fbS_reg_set) {
+ spe_shli(f, fbZS_reg, fbS_reg, 24); /* fbS = fbS << 24 */
+ }
+ else {
+ spe_move(f, fbZS_reg, fbZ_reg);
+ }
}
- else if (zs_format == PIPE_FORMAT_S8Z24_UNORM ||
- zs_format == PIPE_FORMAT_X8Z24_UNORM) {
- /* XXX to do */
- ASSERT(0);
+ else if (zs_format == PIPE_FORMAT_Z24S8_UNORM ||
+ zs_format == PIPE_FORMAT_Z24X8_UNORM) {
+ if (fbS_reg_set && fbZ_reg_set) {
+ spe_shli(f, fbZ_reg, fbZ_reg, 8); /* fbZ = fbZ << 8 */
+ spe_or(f, fbZS_reg, fbS_reg, fbZ_reg); /* fbZS = fbS | fbZ */
+ }
+ else if (fbS_reg_set) {
+ spe_move(f, fbZS_reg, fbS_reg);
+ }
+ else {
+ spe_shli(f, fbZ_reg, fbZ_reg, 8); /* fbZ = fbZ << 8 */
+ }
+ }
+ else if (zs_format == PIPE_FORMAT_Z32_UNORM) {
+ if (fbZ_reg_set) {
+ spe_move(f, fbZS_reg, fbZ_reg); /* fbZS = fbZ */
+ }
}
else if (zs_format == PIPE_FORMAT_Z16_UNORM) {
- /* XXX to do */
- ASSERT(0);
+ if (fbZ_reg_set) {
+ spe_move(f, fbZS_reg, fbZ_reg); /* fbZS = fbZ */
+ }
}
else if (zs_format == PIPE_FORMAT_S8_UNORM) {
- /* XXX to do */
- ASSERT(0);
+ ASSERT(0); /* XXX to do */
}
else {
- /* bad zs_format */
- ASSERT(0);
+ ASSERT(0); /* bad zs_format */
}
/* Store: memory[depth_tile_reg + quad_offset_reg] = fbZS */
spe_stqx(f, fbZS_reg, depth_tile_reg, quad_offset_reg);
}
- spe_release_register(f, fbZ_reg);
- spe_release_register(f, fbS_reg);
+ release_optional_register(f, &fbZ_reg_set, fbZ_reg);
+ release_optional_register(f, &fbS_reg_set, fbS_reg);
}
-
/* Get framebuffer quad/colors. We'll need these for blending,
* color masking, and to obey the quad/pixel mask.
* Load: fbRGBA_reg = memory[color_tile + quad_offset]
* Note: if mask={~0,~0,~0,~0} and we're not blending or colormasking
* we could skip this load.
*/
+ spe_comment(f, 0, "Fetch quad colors from tile");
spe_lqx(f, fbRGBA_reg, color_tile_reg, quad_offset_reg);
-
if (blend->blend_enable) {
+ spe_comment(f, 0, "Perform blending");
gen_blend(blend, blend_color, f, color_format,
fragR_reg, fragG_reg, fragB_reg, fragA_reg, fbRGBA_reg);
}
int rgba_reg = spe_allocate_available_register(f);
/* Pack four float colors as four 32-bit int colors */
+ spe_comment(f, 0, "Convert float quad colors to packed int framebuffer colors");
gen_pack_colors(f, color_format,
fragR_reg, fragG_reg, fragB_reg, fragA_reg,
rgba_reg);
if (blend->logicop_enable) {
+ spe_comment(f, 0, "Compute logic op");
gen_logicop(blend, f, rgba_reg, fbRGBA_reg);
}
- if (blend->colormask != 0xf) {
- gen_colormask(blend->colormask, f, rgba_reg, fbRGBA_reg);
+ if (blend->colormask != PIPE_MASK_RGBA) {
+ spe_comment(f, 0, "Compute color mask");
+ gen_colormask(f, blend->colormask, color_format, rgba_reg, fbRGBA_reg);
}
-
/* Mix fragment colors with framebuffer colors using the quad/pixel mask:
* if (mask[i])
* rgba[i] = rgba[i];
/* Store updated quad in tile:
* memory[color_tile + quad_offset] = rgba_reg;
*/
+ spe_comment(f, 0, "Store quad colors into color tile");
spe_stqx(f, rgba_reg, color_tile_reg, quad_offset_reg);
spe_release_register(f, rgba_reg);
spe_bi(f, SPE_REG_RA, 0, 0); /* return from function call */
-
spe_release_register(f, fbRGBA_reg);
spe_release_register(f, fbZS_reg);
spe_release_register(f, quad_offset_reg);
projects / mesa.git / blobdiff