ilo: support push constant model in shaders

Source constants from URB constant data when the constant data can fit in the
PCB.

6 files changed, 143 insertions, 12 deletions

diff --git a/src/gallium/drivers/ilo/ilo_shader.c b/src/gallium/drivers/ilo/ilo_shader.c
index b1a11a1391b..0c785201d1a 100644
--- a/src/gallium/drivers/ilo/ilo_shader.c
+++ b/src/gallium/drivers/ilo/ilo_shader.c
@@ -273,6 +273,13 @@ ilo_shader_variant_init(struct ilo_shader_variant *variant,
       break;
    }
 
+   /* use PCB unless constant buffer 0 is not in user buffer  */
+   if ((ilo->cbuf[info->type].enabled_mask & 0x1) &&
+       !ilo->cbuf[info->type].cso[0].user_buffer)
+      variant->use_pcb = false;
+   else
+      variant->use_pcb = true;
+
    num_views = ilo->view[info->type].count;
    assert(info->num_samplers <= num_views);
 
@@ -341,6 +348,8 @@ ilo_shader_variant_guess(struct ilo_shader_variant *variant,
       break;
    }
 
+   variant->use_pcb = true;
+
    variant->num_sampler_views = info->num_samplers;
    for (i = 0; i < info->num_samplers; i++) {
       if (info->shadow_samplers & (1 << i)) {
@@ -747,7 +756,8 @@ ilo_shader_create_vs(const struct ilo_dev_info *dev,
 
    /* states used in ilo_shader_variant_init() */
    shader->info.non_orthogonal_states = ILO_DIRTY_VIEW_VS |
-                                        ILO_DIRTY_RASTERIZER;
+                                        ILO_DIRTY_RASTERIZER |
+                                        ILO_DIRTY_CBUF;
 
    return shader;
 }
@@ -764,7 +774,8 @@ ilo_shader_create_gs(const struct ilo_dev_info *dev,
    /* states used in ilo_shader_variant_init() */
    shader->info.non_orthogonal_states = ILO_DIRTY_VIEW_GS |
                                         ILO_DIRTY_VS |
-                                        ILO_DIRTY_RASTERIZER;
+                                        ILO_DIRTY_RASTERIZER |
+                                        ILO_DIRTY_CBUF;
 
    return shader;
 }
@@ -781,7 +792,8 @@ ilo_shader_create_fs(const struct ilo_dev_info *dev,
    /* states used in ilo_shader_variant_init() */
    shader->info.non_orthogonal_states = ILO_DIRTY_VIEW_FS |
                                         ILO_DIRTY_RASTERIZER |
-                                        ILO_DIRTY_FB;
+                                        ILO_DIRTY_FB |
+                                        ILO_DIRTY_CBUF;
 
    return shader;
 }
@@ -1061,10 +1073,10 @@ ilo_shader_get_kernel_param(const struct ilo_shader_state *shader,
       val = kernel->in.start_grf;
       break;
    case ILO_KERNEL_SKIP_CBUF0_UPLOAD:
-      val = false;
+      val = kernel->skip_cbuf0_upload;
       break;
    case ILO_KERNEL_PCB_CBUF0_SIZE:
-      val = 0;
+      val = kernel->pcb.cbuf0_size;
       break;
 
    case ILO_KERNEL_VS_INPUT_INSTANCEID:
diff --git a/src/gallium/drivers/ilo/shader/ilo_shader_fs.c b/src/gallium/drivers/ilo/shader/ilo_shader_fs.c
index 36a308744c6..48d5721631c 100644
--- a/src/gallium/drivers/ilo/shader/ilo_shader_fs.c
+++ b/src/gallium/drivers/ilo/shader/ilo_shader_fs.c
@@ -306,6 +306,32 @@ fs_lower_opcode_tgsi_indirect_const(struct fs_compile_context *fcc,
    }
 }
 
+static bool
+fs_lower_opcode_tgsi_const_pcb(struct fs_compile_context *fcc,
+                               struct toy_dst dst, int dim,
+                               struct toy_src idx)
+{
+   const int grf = fcc->first_const_grf + idx.val32 / 2;
+   const int grf_subreg = (idx.val32 & 1) * 16;
+   struct toy_src src;
+   struct toy_dst real_dst[4];
+   int i;
+
+   if (!fcc->variant->use_pcb || dim != 0 || idx.file != TOY_FILE_IMM ||
+       grf >= fcc->first_attr_grf)
+      return false;
+
+   src = tsrc_rect(tsrc(TOY_FILE_GRF, grf, grf_subreg), TOY_RECT_010);
+
+   tdst_transpose(dst, real_dst);
+   for (i = 0; i < 4; i++) {
+      /* cast to type D to make sure these are raw moves */
+      tc_MOV(&fcc->tc, tdst_d(real_dst[i]), tsrc_d(tsrc_offset(src, 0, i)));
+   }
+
+   return true;
+}
+
 static void
 fs_lower_opcode_tgsi_const_gen6(struct fs_compile_context *fcc,
                                 struct toy_dst dst, int dim, struct toy_src idx)
@@ -322,6 +348,9 @@ fs_lower_opcode_tgsi_const_gen6(struct fs_compile_context *fcc,
    struct toy_dst tmp, real_dst[4];
    int i;
 
+   if (fs_lower_opcode_tgsi_const_pcb(fcc, dst, dim, idx))
+      return;
+
    /* set message header */
    inst = tc_MOV(tc, header, r0);
    inst->mask_ctrl = BRW_MASK_DISABLE;
@@ -365,6 +394,9 @@ fs_lower_opcode_tgsi_const_gen7(struct fs_compile_context *fcc,
    struct toy_dst tmp, real_dst[4];
    int i;
 
+   if (fs_lower_opcode_tgsi_const_pcb(fcc, dst, dim, idx))
+      return;
+
    /*
     * In 4c1fdae0a01b3f92ec03b61aac1d3df500d51fc6, pull constant load was
     * changed from OWord Block Read to ld to increase performance in the
@@ -1743,8 +1775,28 @@ fs_setup(struct fs_compile_context *fcc,
    fs_setup_shader_in(fcc->shader, &fcc->tgsi, fcc->variant->u.fs.flatshade);
    fs_setup_shader_out(fcc->shader, &fcc->tgsi);
 
-   /* we do not make use of push constant buffers yet */
-   num_consts = 0;
+   if (fcc->variant->use_pcb && !fcc->tgsi.const_indirect) {
+      num_consts = (fcc->tgsi.const_count + 1) / 2;
+
+      /*
+       * From the Sandy Bridge PRM, volume 2 part 1, page 287:
+       *
+       *     "The sum of all four read length fields (each incremented to
+       *      represent the actual read length) must be less than or equal to
+       *      64"
+       *
+       * Since we are usually under a high register pressure, do not allow
+       * for more than 8.
+       */
+      if (num_consts > 8)
+         num_consts = 0;
+   }
+   else {
+      num_consts = 0;
+   }
+
+   fcc->shader->skip_cbuf0_upload = (!fcc->tgsi.const_count || num_consts);
+   fcc->shader->pcb.cbuf0_size = num_consts * (sizeof(float) * 8);
 
    fcc->first_const_grf = fs_setup_payloads(fcc);
    fcc->first_attr_grf = fcc->first_const_grf + num_consts;
diff --git a/src/gallium/drivers/ilo/shader/ilo_shader_internal.h b/src/gallium/drivers/ilo/shader/ilo_shader_internal.h
index 07e8ee2f683..8d4a6a18c6b 100644
--- a/src/gallium/drivers/ilo/shader/ilo_shader_internal.h
+++ b/src/gallium/drivers/ilo/shader/ilo_shader_internal.h
@@ -59,6 +59,8 @@ struct ilo_shader_variant {
       } fs;
    } u;
 
+   bool use_pcb;
+
    int num_sampler_views;
    struct {
       unsigned r:3;
@@ -102,6 +104,8 @@ struct ilo_shader {
       bool has_pos;
    } out;
 
+   bool skip_cbuf0_upload;
+
    bool has_kill;
    bool dispatch_16;
 
@@ -124,6 +128,7 @@ struct ilo_shader {
 
    /* what does the push constant buffer consist of? */
    struct {
+      int cbuf0_size;
       int clip_state_size;
    } pcb;
 
diff --git a/src/gallium/drivers/ilo/shader/ilo_shader_vs.c b/src/gallium/drivers/ilo/shader/ilo_shader_vs.c
index dc166d7cc48..b5b44b57796 100644
--- a/src/gallium/drivers/ilo/shader/ilo_shader_vs.c
+++ b/src/gallium/drivers/ilo/shader/ilo_shader_vs.c
@@ -47,6 +47,7 @@ struct vs_compile_context {
 
    int num_grf_per_vrf;
    int first_const_grf;
+   int first_ucp_grf;
    int first_vue_grf;
    int first_free_grf;
    int last_free_grf;
@@ -79,6 +80,27 @@ vs_lower_opcode_tgsi_in(struct vs_compile_context *vcc,
    }
 }
 
+static bool
+vs_lower_opcode_tgsi_const_pcb(struct vs_compile_context *vcc,
+                               struct toy_dst dst, int dim,
+                               struct toy_src idx)
+{
+   const int i = idx.val32;
+   const int grf = vcc->first_const_grf + i / 2;
+   const int grf_subreg = (i & 1) * 16;
+   struct toy_src src;
+
+   if (!vcc->variant->use_pcb || dim != 0 || idx.file != TOY_FILE_IMM ||
+       grf >= vcc->first_ucp_grf)
+      return false;
+
+
+   src = tsrc_rect(tsrc(TOY_FILE_GRF, grf, grf_subreg), TOY_RECT_041);
+   tc_MOV(&vcc->tc, dst, src);
+
+   return true;
+}
+
 static void
 vs_lower_opcode_tgsi_const_gen6(struct vs_compile_context *vcc,
                                 struct toy_dst dst, int dim,
@@ -94,6 +116,9 @@ vs_lower_opcode_tgsi_const_gen6(struct vs_compile_context *vcc,
    struct toy_inst *inst;
    struct toy_src desc;
 
+   if (vs_lower_opcode_tgsi_const_pcb(vcc, dst, dim, idx))
+      return;
+
    /* set message header */
    inst = tc_MOV(tc, header, r0);
    inst->mask_ctrl = BRW_MASK_DISABLE;
@@ -121,6 +146,9 @@ vs_lower_opcode_tgsi_const_gen7(struct vs_compile_context *vcc,
       tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf, 0));
    struct toy_src desc;
 
+   if (vs_lower_opcode_tgsi_const_pcb(vcc, dst, dim, idx))
+      return;
+
    /*
     * In 259b65e2e7938de4aab323033cfe2b33369ddb07, pull constant load was
     * changed from OWord Dual Block Read to ld to increase performance in the
@@ -835,7 +863,7 @@ vs_collect_outputs(struct vs_compile_context *vcc, struct toy_src *outs)
                }
 
                for (j = first_ucp; j <= last_ucp; j++) {
-                  const int plane_grf = vcc->first_const_grf + j / 2;
+                  const int plane_grf = vcc->first_ucp_grf + j / 2;
                   const int plane_subreg = (j & 1) * 16;
                   const struct toy_src plane = tsrc_rect(tsrc(TOY_FILE_GRF,
                            plane_grf, plane_subreg), TOY_RECT_041);
@@ -1199,12 +1227,34 @@ vs_setup(struct vs_compile_context *vcc,
    vs_setup_shader_out(vcc->shader, &vcc->tgsi,
          (vcc->variant->u.vs.num_ucps > 0), vcc->output_map);
 
-   /* fit each pair of user clip planes into a register */
-   num_consts = (vcc->variant->u.vs.num_ucps + 1) / 2;
+   if (vcc->variant->use_pcb && !vcc->tgsi.const_indirect) {
+      num_consts = (vcc->tgsi.const_count + 1) / 2;
+
+      /*
+       * From the Sandy Bridge PRM, volume 2 part 1, page 138:
+       *
+       *     "The sum of all four read length fields (each incremented to
+       *      represent the actual read length) must be less than or equal to
+       *      32"
+       */
+      if (num_consts > 32)
+         num_consts = 0;
+   }
+   else {
+      num_consts = 0;
+   }
+
+   vcc->shader->skip_cbuf0_upload = (!vcc->tgsi.const_count || num_consts);
+   vcc->shader->pcb.cbuf0_size = num_consts * (sizeof(float) * 8);
 
    /* r0 is reserved for payload header */
    vcc->first_const_grf = 1;
-   vcc->first_vue_grf = vcc->first_const_grf + num_consts;
+   vcc->first_ucp_grf = vcc->first_const_grf + num_consts;
+
+   /* fit each pair of user clip planes into a register */
+   vcc->first_vue_grf = vcc->first_ucp_grf +
+      (vcc->variant->u.vs.num_ucps + 1) / 2;
+
    vcc->first_free_grf = vcc->first_vue_grf + vcc->shader->in.count;
    vcc->last_free_grf = 127;
 
diff --git a/src/gallium/drivers/ilo/shader/toy_tgsi.c b/src/gallium/drivers/ilo/shader/toy_tgsi.c
index d5a3f2fe5af..bf1e37ef584 100644
--- a/src/gallium/drivers/ilo/shader/toy_tgsi.c
+++ b/src/gallium/drivers/ilo/shader/toy_tgsi.c
@@ -2244,6 +2244,12 @@ parse_instruction(struct toy_tgsi *tgsi,
       break;
    }
 
+   for (i = 0; i < tgsi_inst->Instruction.NumSrcRegs; i++) {
+      const struct tgsi_full_src_register *s = &tgsi_inst->Src[i];
+      if (s->Register.File == TGSI_FILE_CONSTANT && s->Register.Indirect)
+         tgsi->const_indirect = true;
+   }
+
    /* remember channels written */
    for (i = 0; i < tgsi_inst->Instruction.NumDstRegs; i++) {
       const struct tgsi_full_dst_register *d = &tgsi_inst->Dst[i];
@@ -2398,8 +2404,11 @@ parse_declaration(struct toy_tgsi *tgsi,
       /* immediates should be declared with TGSI_TOKEN_TYPE_IMMEDIATE */
       assert(!"unexpected immediate declaration");
       break;
-   case TGSI_FILE_NULL:
    case TGSI_FILE_CONSTANT:
+      if (tgsi->const_count <= decl->Range.Last)
+         tgsi->const_count = decl->Range.Last + 1;
+      break;
+   case TGSI_FILE_NULL:
    case TGSI_FILE_TEMPORARY:
    case TGSI_FILE_SAMPLER:
    case TGSI_FILE_PREDICATE:
diff --git a/src/gallium/drivers/ilo/shader/toy_tgsi.h b/src/gallium/drivers/ilo/shader/toy_tgsi.h
index 95fc897b7e0..38be9f4f891 100644
--- a/src/gallium/drivers/ilo/shader/toy_tgsi.h
+++ b/src/gallium/drivers/ilo/shader/toy_tgsi.h
@@ -91,6 +91,9 @@ struct toy_tgsi {
    } system_values[8];
    int num_system_values;
 
+   int const_count;
+   bool const_indirect;
+
    bool uses_kill;
 };