cell_state_per_fragment.h File Reference

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Functions
void	cell_generate_depth_stencil_test (struct cell_depth_stencil_alpha_state *cdsa)
void	cell_generate_alpha_blend (struct cell_blend_state *cb)
	Generate code to perform alpha blending on the SPE.
void	cell_generate_logic_op (struct spe_function *f, const struct pipe_blend_state *blend, struct pipe_surface *surf)
	Generate code to perform color conversion and logic op.

---

Function Documentation

void cell_generate_alpha_blend

(

struct cell_blend_state *

cb

)

Generate code to perform alpha blending on the SPE.

Definition at line 959 of file cell_state_per_fragment.c.

References pipe_blend_state::alpha_dst_factor, pipe_blend_state::alpha_func, pipe_blend_state::alpha_src_factor, cell_blend_state::base, pipe_blend_state::blend_enable, cell_blend_state::code, pipe_blend_state::colormask, emit_alpha_factor_calculation(), emit_blend_calculation(), emit_color_factor_calculation(), PIPE_BLEND_ADD, PIPE_BLEND_MAX, PIPE_BLEND_MIN, PIPE_BLENDFACTOR_CONST_COLOR, PIPE_BLENDFACTOR_DST_COLOR, PIPE_BLENDFACTOR_ONE, PIPE_BLENDFACTOR_SRC1_COLOR, PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE, PIPE_BLENDFACTOR_SRC_COLOR, PIPE_BLENDFACTOR_ZERO, pipe_blend_state::rgb_dst_factor, pipe_blend_state::rgb_func, pipe_blend_state::rgb_src_factor, spe_allocate_register(), spe_bi(), spe_init_func(), SPE_INST_SIZE, and spe_function::store.

{
   struct pipe_blend_state *const b = &cb->base;
   struct spe_function *const f = &cb->code;

   /* This code generates a maximum of 3 (source alpha factor)
    * + 3 (destination alpha factor) + (3 * 6) (source color factor)
    * + (3 * 6) (destination color factor) + (4 * 2) (blend equation)
    * + 4 (fragment mask) + 1 (return) = 55 instlructions.  Round up to 64 to
    * make it a happy power-of-two.
    */
   spe_init_func(f, SPE_INST_SIZE * 64);


   const int frag[4] = {
      spe_allocate_register(f, 3),
      spe_allocate_register(f, 4),
      spe_allocate_register(f, 5),
      spe_allocate_register(f, 6),
   };
   const int pixel[4] = {
      spe_allocate_register(f, 7),
      spe_allocate_register(f, 8),
      spe_allocate_register(f, 9),
      spe_allocate_register(f, 10),
   };
   const int const_color[4] = {
      spe_allocate_register(f, 11),
      spe_allocate_register(f, 12),
      spe_allocate_register(f, 13),
      spe_allocate_register(f, 14),
   };
   unsigned func[4];
   unsigned sF[4];
   unsigned dF[4];
   unsigned i;
   int src_factor[4];
   int dst_factor[4];


   /* Does the selected blend mode make use of the source / destination
    * color (RGB) blend factors?
    */
   boolean need_color_factor = b->blend_enable
       && (b->rgb_func != PIPE_BLEND_MIN)
       && (b->rgb_func != PIPE_BLEND_MAX);

   /* Does the selected blend mode make use of the source / destination
    * alpha blend factors?
    */
   boolean need_alpha_factor = b->blend_enable
       && (b->alpha_func != PIPE_BLEND_MIN)
       && (b->alpha_func != PIPE_BLEND_MAX);


   if (b->blend_enable) {
      sF[0] = b->rgb_src_factor;
      sF[1] = sF[0];
      sF[2] = sF[0];
      switch (b->alpha_src_factor & 0x0f) {
      case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
         sF[3] = PIPE_BLENDFACTOR_ONE;
         break;
      case PIPE_BLENDFACTOR_SRC_COLOR:
      case PIPE_BLENDFACTOR_DST_COLOR:
      case PIPE_BLENDFACTOR_CONST_COLOR:
      case PIPE_BLENDFACTOR_SRC1_COLOR:
         sF[3] = b->alpha_src_factor + 1;
         break;
      default:
         sF[3] = b->alpha_src_factor;
      }

      dF[0] = b->rgb_dst_factor;
      dF[1] = dF[0];
      dF[2] = dF[0];
      switch (b->alpha_dst_factor & 0x0f) {
      case PIPE_BLENDFACTOR_SRC_COLOR:
      case PIPE_BLENDFACTOR_DST_COLOR:
      case PIPE_BLENDFACTOR_CONST_COLOR:
      case PIPE_BLENDFACTOR_SRC1_COLOR:
         dF[3] = b->alpha_dst_factor + 1;
         break;
      default:
         dF[3] = b->alpha_dst_factor;
      }

      func[0] = b->rgb_func;
      func[1] = func[0];
      func[2] = func[0];
      func[3] = b->alpha_func;
   } else {
      sF[0] = PIPE_BLENDFACTOR_ONE;
      sF[1] = PIPE_BLENDFACTOR_ONE;
      sF[2] = PIPE_BLENDFACTOR_ONE;
      sF[3] = PIPE_BLENDFACTOR_ONE;
      dF[0] = PIPE_BLENDFACTOR_ZERO;
      dF[1] = PIPE_BLENDFACTOR_ZERO;
      dF[2] = PIPE_BLENDFACTOR_ZERO;
      dF[3] = PIPE_BLENDFACTOR_ZERO;

      func[0] = PIPE_BLEND_ADD;
      func[1] = PIPE_BLEND_ADD;
      func[2] = PIPE_BLEND_ADD;
      func[3] = PIPE_BLEND_ADD;
   }


   /* If alpha writing is enabled and the alpha blend mode requires use of
    * the alpha factor, calculate the alpha factor.
    */
   if (((b->colormask & 8) != 0) && need_alpha_factor) {
      src_factor[3] = emit_alpha_factor_calculation(f, sF[3], const_color[3],
                                                    frag[3], pixel[3]);

      /* If the alpha destination blend factor is the same as the alpha source
       * blend factor, re-use the previously calculated value.
       */
      dst_factor[3] = (dF[3] == sF[3])
          ? src_factor[3]
          : emit_alpha_factor_calculation(f, dF[3], const_color[3],
                                          frag[3], pixel[3]);
   }


   if (sF[0] == sF[3]) {
      src_factor[0] = src_factor[3];
      src_factor[1] = src_factor[3];
      src_factor[2] = src_factor[3];
   } else if (sF[0] == dF[3]) {
      src_factor[0] = dst_factor[3];
      src_factor[1] = dst_factor[3];
      src_factor[2] = dst_factor[3];
   } else if (need_color_factor) {
      emit_color_factor_calculation(f,
                                    b->rgb_src_factor,
                                    b->colormask,
                                    frag, pixel, const_color, src_factor);
   }


   if (dF[0] == sF[3]) {
      dst_factor[0] = src_factor[3];
      dst_factor[1] = src_factor[3];
      dst_factor[2] = src_factor[3];
   } else if (dF[0] == dF[3]) {
      dst_factor[0] = dst_factor[3];
      dst_factor[1] = dst_factor[3];
      dst_factor[2] = dst_factor[3];
   } else if (dF[0] == sF[0]) {
      dst_factor[0] = src_factor[0];
      dst_factor[1] = src_factor[1];
      dst_factor[2] = src_factor[2];
   } else if (need_color_factor) {
      emit_color_factor_calculation(f,
                                    b->rgb_dst_factor,
                                    b->colormask,
                                    frag, pixel, const_color, dst_factor);
   }



   for (i = 0; i < 4; ++i) {
      if ((b->colormask & (1U << i)) != 0) {
         emit_blend_calculation(f,
                                func[i], sF[i], dF[i],
                                frag[i], src_factor[i],
                                pixel[i], dst_factor[i]);
      }
   }

   spe_bi(f, 0, 0, 0);

#if 0
   {
      const uint32_t *p = f->store;

      printf("# %u instructions\n", f->csr - f->store);
      printf("# blend (%sabled)\n",
             (cb->base.blend_enable) ? "en" : "dis");
      printf("#    RGB func / sf / df: %u %u %u\n",
             cb->base.rgb_func,
             cb->base.rgb_src_factor,
             cb->base.rgb_dst_factor);
      printf("#    ALP func / sf / df: %u %u %u\n",
             cb->base.alpha_func,
             cb->base.alpha_src_factor,
             cb->base.alpha_dst_factor);

      printf("\t.text\n");
      for (/* empty */; p < f->csr; p++) {
         printf("\t.long\t0x%04x\n", *p);
      }
      fflush(stdout);
   }
#endif
}

void cell_generate_depth_stencil_test

(

struct cell_depth_stencil_alpha_state *

cdsa

)

Definition at line 466 of file cell_state_per_fragment.c.

References pipe_depth_stencil_alpha_state::alpha, cell_depth_stencil_alpha_state::base, cell_depth_stencil_alpha_state::code, pipe_depth_stencil_alpha_state::depth, emit_alpha_test(), emit_depth_test(), emit_stencil_test(), pipe_alpha_state::enabled, pipe_depth_state::enabled, pipe_stencil_state::enabled, pipe_stencil_state::fail_op, pipe_stencil_state::func, pipe_depth_state::func, pipe_alpha_state::func, pipe_alpha_state::ref, pipe_stencil_state::ref_value, spe_allocate_available_register(), spe_allocate_register(), spe_and(), spe_andc(), spe_bi(), spe_init_func(), SPE_INST_SIZE, spe_or(), spe_release_register(), spe_selb(), pipe_depth_stencil_alpha_state::stencil, spe_function::store, pipe_stencil_state::value_mask, pipe_stencil_state::write_mask, pipe_depth_state::writemask, pipe_stencil_state::zfail_op, and pipe_stencil_state::zpass_op.

{
   struct pipe_depth_stencil_alpha_state *const dsa = &cdsa->base;
   struct spe_function *const f = &cdsa->code;

   /* This code generates a maximum of 6 (alpha test) + 3 (depth test)
    * + 25 (front stencil) + 25 (back stencil) + 4 = 63 instructions.  Round
    * up to 64 to make it a happy power-of-two.
    */
   spe_init_func(f, SPE_INST_SIZE * 64);


   /* Allocate registers for the function's input parameters.  Cleverly (and
    * clever code is usually dangerous, but I couldn't resist) the generated
    * function returns a structure.  Returned structures start with register
    * 3, and the structure fields are ordered to match up exactly with the
    * input parameters.
    */
   int mask = spe_allocate_register(f, 3);
   int depth = spe_allocate_register(f, 4);
   int stencil = spe_allocate_register(f, 5);
   int zvals = spe_allocate_register(f, 6);
   int frag_a = spe_allocate_register(f, 7);
   int facing = spe_allocate_register(f, 8);

   int depth_mask = spe_allocate_available_register(f);

   boolean depth_complement;


   emit_alpha_test(dsa, f, mask, frag_a);

   depth_complement = emit_depth_test(dsa, f, depth_mask, depth, zvals);

   if (dsa->stencil[0].enabled) {
      const int front_depth_pass = spe_allocate_available_register(f);
      int front_stencil = emit_stencil_test(dsa, 0, f, mask,
                                            depth_mask, depth_complement,
                                            stencil, front_depth_pass);

      if (dsa->stencil[1].enabled) {
         const int back_depth_pass = spe_allocate_available_register(f);
         int back_stencil = emit_stencil_test(dsa, 1, f, mask,
                                              depth_mask,  depth_complement,
                                              stencil, back_depth_pass);

         /* If the front facing stencil value and the back facing stencil
          * value are stored in the same register, there is no need to select
          * a value based on the facing.  This can happen if the stencil value
          * was not modified due to the write masks being zero, the stencil
          * operations being KEEP, etc.
          */
         if (front_stencil != back_stencil) {
            spe_selb(f, stencil, back_stencil, front_stencil, facing);
         }

         if (back_stencil != stencil) {
            spe_release_register(f, back_stencil);
         }

         if (front_stencil != stencil) {
            spe_release_register(f, front_stencil);
         }

         spe_selb(f, mask, back_depth_pass, front_depth_pass, facing);

         spe_release_register(f, back_depth_pass);
      } else {
         if (front_stencil != stencil) {
            spe_or(f, stencil, front_stencil, front_stencil);
            spe_release_register(f, front_stencil);
         }
         spe_or(f, mask, front_depth_pass, front_depth_pass);
      }

      spe_release_register(f, front_depth_pass);
   } else if (dsa->depth.enabled) {
      if (depth_complement) {
         spe_andc(f, mask, mask, depth_mask);
      } else {
         spe_and(f, mask, mask, depth_mask);
      }
   }

   if (dsa->depth.writemask) {
         spe_selb(f, depth, depth, zvals, mask);
   }

   spe_bi(f, 0, 0, 0);  /* return from function call */


#if 0
   {
      const uint32_t *p = f->store;
      unsigned i;

      printf("# alpha (%sabled)\n",
             (dsa->alpha.enabled) ? "en" : "dis");
      printf("#    func: %u\n", dsa->alpha.func);
      printf("#    ref: %.2f\n", dsa->alpha.ref);

      printf("# depth (%sabled)\n",
             (dsa->depth.enabled) ? "en" : "dis");
      printf("#    func: %u\n", dsa->depth.func);

      for (i = 0; i < 2; i++) {
         printf("# %s stencil (%sabled)\n",
                (i == 0) ? "front" : "back",
                (dsa->stencil[i].enabled) ? "en" : "dis");

         printf("#    func: %u\n", dsa->stencil[i].func);
         printf("#    op (sf, zf, zp): %u %u %u\n",
                dsa->stencil[i].fail_op,
                dsa->stencil[i].zfail_op,
                dsa->stencil[i].zpass_op);
         printf("#    ref value / value mask / write mask: %02x %02x %02x\n",
                dsa->stencil[i].ref_value,
                dsa->stencil[i].value_mask,
                dsa->stencil[i].write_mask);
      }

      printf("\t.text\n");
      for (/* empty */; p < f->csr; p++) {
         printf("\t.long\t0x%04x\n", *p);
      }
      fflush(stdout);
   }
#endif
}

void cell_generate_logic_op	(	struct spe_function *	f,
		const struct pipe_blend_state *	blend,
		struct pipe_surface *	surf
	)

Generate code to perform color conversion and logic op.

Definition at line 1182 of file cell_state_per_fragment.c.

References assert, ASSERT, pipe_blend_state::colormask, pipe_surface::format, pipe_blend_state::logicop_enable, pipe_blend_state::logicop_func, PC_OFFSET(), PIPE_FORMAT_A8R8G8B8_UNORM, PIPE_FORMAT_B8G8R8A8_UNORM, PIPE_LOGICOP_AND, PIPE_LOGICOP_AND_INVERTED, PIPE_LOGICOP_AND_REVERSE, PIPE_LOGICOP_CLEAR, PIPE_LOGICOP_COPY, PIPE_LOGICOP_COPY_INVERTED, PIPE_LOGICOP_EQUIV, PIPE_LOGICOP_INVERT, PIPE_LOGICOP_NAND, PIPE_LOGICOP_NOOP, PIPE_LOGICOP_NOR, PIPE_LOGICOP_OR, PIPE_LOGICOP_OR_INVERTED, PIPE_LOGICOP_OR_REVERSE, PIPE_LOGICOP_SET, PIPE_LOGICOP_XOR, spe_a(), spe_allocate_available_register(), spe_allocate_register(), spe_and(), spe_andc(), spe_bi(), spe_cfltu(), spe_eqv(), spe_il(), spe_ilh(), spe_init_func(), SPE_INST_SIZE, spe_lqr(), spe_nand(), spe_nor(), spe_or(), spe_orc(), spe_selb(), spe_shufb(), spe_xor(), and spe_function::store.

{
   const unsigned logic_op = (blend->logicop_enable)
       ? blend->logicop_func : PIPE_LOGICOP_COPY;

   /* This code generates a maximum of 37 instructions.  An additional 32
    * bytes (equiv. to 8 instructions) are needed for data storage.  Round up
    * to 64 to make it a happy power-of-two.
    */
   spe_init_func(f, SPE_INST_SIZE * 64);


   /* Pixel colors in framebuffer format in AoS layout.
    */
   const int pixel[4] = {
      spe_allocate_register(f, 3),
      spe_allocate_register(f, 4),
      spe_allocate_register(f, 5),
      spe_allocate_register(f, 6),
   };

   /* Fragment colors stored as floats in SoA layout.
    */
   const int frag[4] = {
      spe_allocate_register(f, 7),
      spe_allocate_register(f, 8),
      spe_allocate_register(f, 9),
      spe_allocate_register(f, 10),
   };

   const int mask = spe_allocate_register(f, 11);


   /* Short-circuit the noop and invert cases.
    */
   if ((logic_op == PIPE_LOGICOP_NOOP) || (blend->colormask == 0)) {
      spe_bi(f, 0, 0, 0);
      return;
   } else if (logic_op == PIPE_LOGICOP_INVERT) {
      spe_nor(f, pixel[0], pixel[0], pixel[0]);
      spe_nor(f, pixel[1], pixel[1], pixel[1]);
      spe_nor(f, pixel[2], pixel[2], pixel[2]);
      spe_nor(f, pixel[3], pixel[3], pixel[3]);
      spe_bi(f, 0, 0, 0);
      return;
   }


   const int tmp[4] = {
      spe_allocate_available_register(f),
      spe_allocate_available_register(f),
      spe_allocate_available_register(f),
      spe_allocate_available_register(f),
   };

   const int shuf_xpose_hi = spe_allocate_available_register(f);
   const int shuf_xpose_lo = spe_allocate_available_register(f);
   const int shuf_color = spe_allocate_available_register(f);


   /* Pointer to the begining of the function's private data area.
    */
   uint32_t *const data = ((uint32_t *) f->store) + (64 - 8);


   /* Convert fragment colors to framebuffer format in AoS layout.
    */
   switch (surf->format) {
   case PIPE_FORMAT_A8R8G8B8_UNORM:
      data[0] = 0x00010203;
      data[1] = 0x10111213;
      data[2] = 0x04050607;
      data[3] = 0x14151617;
      data[4] = 0x0c000408;
      data[5] = 0x80808080;
      data[6] = 0x80808080;
      data[7] = 0x80808080;
      break;
   case PIPE_FORMAT_B8G8R8A8_UNORM:
      data[0] = 0x03020100;
      data[1] = 0x13121110;
      data[2] = 0x07060504;
      data[3] = 0x17161514;
      data[4] = 0x0804000c;
      data[5] = 0x80808080;
      data[6] = 0x80808080;
      data[7] = 0x80808080;
      break;
   default:
      fprintf(stderr, "CELL: Bad pixel format in cell_generate_logic_op()");
      ASSERT(0);
   }

   spe_ilh(f, tmp[0], 0x0808);
   spe_lqr(f, shuf_xpose_hi, PC_OFFSET(f, data+0));
   spe_lqr(f, shuf_color, PC_OFFSET(f, data+4));
   spe_a(f, shuf_xpose_lo, shuf_xpose_hi, tmp[0]);

   spe_shufb(f, tmp[0], frag[0], frag[2], shuf_xpose_hi);
   spe_shufb(f, tmp[1], frag[0], frag[2], shuf_xpose_lo);
   spe_shufb(f, tmp[2], frag[1], frag[3], shuf_xpose_hi);
   spe_shufb(f, tmp[3], frag[1], frag[3], shuf_xpose_lo);

   spe_shufb(f, frag[0], tmp[0], tmp[2], shuf_xpose_hi);
   spe_shufb(f, frag[1], tmp[0], tmp[2], shuf_xpose_lo);
   spe_shufb(f, frag[2], tmp[1], tmp[3], shuf_xpose_hi);
   spe_shufb(f, frag[3], tmp[1], tmp[3], shuf_xpose_lo);

   spe_cfltu(f, frag[0], frag[0], 32);
   spe_cfltu(f, frag[1], frag[1], 32);
   spe_cfltu(f, frag[2], frag[2], 32);
   spe_cfltu(f, frag[3], frag[3], 32);

   spe_shufb(f, frag[0], frag[0], pixel[0], shuf_color);
   spe_shufb(f, frag[1], frag[1], pixel[1], shuf_color);
   spe_shufb(f, frag[2], frag[2], pixel[2], shuf_color);
   spe_shufb(f, frag[3], frag[3], pixel[3], shuf_color);


   /* If logic op is enabled, perform the requested logical operation on the
    * converted fragment colors and the pixel colors.
    */
   switch (logic_op) {
   case PIPE_LOGICOP_CLEAR:
      spe_il(f, frag[0], 0);
      spe_il(f, frag[1], 0);
      spe_il(f, frag[2], 0);
      spe_il(f, frag[3], 0);
      break;
   case PIPE_LOGICOP_NOR:
      spe_nor(f, frag[0], frag[0], pixel[0]);
      spe_nor(f, frag[1], frag[1], pixel[1]);
      spe_nor(f, frag[2], frag[2], pixel[2]);
      spe_nor(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_AND_INVERTED:
      spe_andc(f, frag[0], pixel[0], frag[0]);
      spe_andc(f, frag[1], pixel[1], frag[1]);
      spe_andc(f, frag[2], pixel[2], frag[2]);
      spe_andc(f, frag[3], pixel[3], frag[3]);
      break;
   case PIPE_LOGICOP_COPY_INVERTED:
      spe_nor(f, frag[0], frag[0], frag[0]);
      spe_nor(f, frag[1], frag[1], frag[1]);
      spe_nor(f, frag[2], frag[2], frag[2]);
      spe_nor(f, frag[3], frag[3], frag[3]);
      break;
   case PIPE_LOGICOP_AND_REVERSE:
      spe_andc(f, frag[0], frag[0], pixel[0]);
      spe_andc(f, frag[1], frag[1], pixel[1]);
      spe_andc(f, frag[2], frag[2], pixel[2]);
      spe_andc(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_XOR:
      spe_xor(f, frag[0], frag[0], pixel[0]);
      spe_xor(f, frag[1], frag[1], pixel[1]);
      spe_xor(f, frag[2], frag[2], pixel[2]);
      spe_xor(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_NAND:
      spe_nand(f, frag[0], frag[0], pixel[0]);
      spe_nand(f, frag[1], frag[1], pixel[1]);
      spe_nand(f, frag[2], frag[2], pixel[2]);
      spe_nand(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_AND:
      spe_and(f, frag[0], frag[0], pixel[0]);
      spe_and(f, frag[1], frag[1], pixel[1]);
      spe_and(f, frag[2], frag[2], pixel[2]);
      spe_and(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_EQUIV:
      spe_eqv(f, frag[0], frag[0], pixel[0]);
      spe_eqv(f, frag[1], frag[1], pixel[1]);
      spe_eqv(f, frag[2], frag[2], pixel[2]);
      spe_eqv(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_OR_INVERTED:
      spe_orc(f, frag[0], pixel[0], frag[0]);
      spe_orc(f, frag[1], pixel[1], frag[1]);
      spe_orc(f, frag[2], pixel[2], frag[2]);
      spe_orc(f, frag[3], pixel[3], frag[3]);
      break;
   case PIPE_LOGICOP_COPY:
      break;
   case PIPE_LOGICOP_OR_REVERSE:
      spe_orc(f, frag[0], frag[0], pixel[0]);
      spe_orc(f, frag[1], frag[1], pixel[1]);
      spe_orc(f, frag[2], frag[2], pixel[2]);
      spe_orc(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_OR:
      spe_or(f, frag[0], frag[0], pixel[0]);
      spe_or(f, frag[1], frag[1], pixel[1]);
      spe_or(f, frag[2], frag[2], pixel[2]);
      spe_or(f, frag[3], frag[3], pixel[3]);
      break;
   case PIPE_LOGICOP_SET:
      spe_il(f, frag[0], ~0);
      spe_il(f, frag[1], ~0);
      spe_il(f, frag[2], ~0);
      spe_il(f, frag[3], ~0);
      break;

   /* These two cases are short-circuited above.
    */
   case PIPE_LOGICOP_INVERT:
   case PIPE_LOGICOP_NOOP:
   default:
      assert(0);
   }


   /* Apply fragment mask.
    */
   spe_ilh(f, tmp[0], 0x0000);
   spe_ilh(f, tmp[1], 0x0404);
   spe_ilh(f, tmp[2], 0x0808);
   spe_ilh(f, tmp[3], 0x0c0c);

   spe_shufb(f, tmp[0], mask, mask, tmp[0]);
   spe_shufb(f, tmp[1], mask, mask, tmp[1]);
   spe_shufb(f, tmp[2], mask, mask, tmp[2]);
   spe_shufb(f, tmp[3], mask, mask, tmp[3]);

   spe_selb(f, pixel[0], pixel[0], frag[0], tmp[0]);
   spe_selb(f, pixel[1], pixel[1], frag[1], tmp[1]);
   spe_selb(f, pixel[2], pixel[2], frag[2], tmp[2]);
   spe_selb(f, pixel[3], pixel[3], frag[3], tmp[3]);

   spe_bi(f, 0, 0, 0);

#if 0
   {
      const uint32_t *p = f->store;
      unsigned i;

      printf("# %u instructions\n", f->csr - f->store);

      printf("\t.text\n");
      for (i = 0; i < 64; i++) {
         printf("\t.long\t0x%04x\n", p[i]);
      }
      fflush(stdout);
   }
#endif
}

---