intel/fs: Support min_lod parameters on texture instructions

We have to lower some shadow instructions because they don't exist in
hardware and we have to lower txb+offset+clamp because the message gets
too big and we run into the sampler message length limit of 11 regs.

Acked-by: Ian Romanick <[email protected]>

4 files changed, 31 insertions, 2 deletions

diff --git a/src/intel/compiler/brw_eu_defines.h b/src/intel/compiler/brw_eu_defines.h
index 52957882b10..affe977835b 100644
--- a/src/intel/compiler/brw_eu_defines.h
+++ b/src/intel/compiler/brw_eu_defines.h
@@ -811,6 +811,8 @@ enum tex_logical_srcs {
    TEX_LOGICAL_SRC_LOD,
    /** dPdy if the operation takes explicit derivatives */
    TEX_LOGICAL_SRC_LOD2,
+   /** Min LOD */
+   TEX_LOGICAL_SRC_MIN_LOD,
    /** Sample index */
    TEX_LOGICAL_SRC_SAMPLE_INDEX,
    /** MCS data */
diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index 713a6c7f40a..3125e5feb1d 100644
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -4472,6 +4472,7 @@ lower_sampler_logical_send_gen7(const fs_builder &bld, fs_inst *inst, opcode op,
                                 const fs_reg &coordinate,
                                 const fs_reg &shadow_c,
                                 fs_reg lod, const fs_reg &lod2,
+                                const fs_reg &min_lod,
                                 const fs_reg &sample_index,
                                 const fs_reg &mcs,
                                 const fs_reg &surface,
@@ -4682,6 +4683,15 @@ lower_sampler_logical_send_gen7(const fs_builder &bld, fs_inst *inst, opcode op,
          bld.MOV(sources[length++], offset(coordinate, bld, i));
    }
 
+   if (min_lod.file != BAD_FILE) {
+      /* Account for all of the missing coordinate sources */
+      length += 4 - coord_components;
+      if (op == SHADER_OPCODE_TXD)
+         length += (3 - grad_components) * 2;
+
+      bld.MOV(sources[length++], min_lod);
+   }
+
    int mlen;
    if (reg_width == 2)
       mlen = length * reg_width - header_size;
@@ -4713,6 +4723,7 @@ lower_sampler_logical_send(const fs_builder &bld, fs_inst *inst, opcode op)
    const fs_reg &shadow_c = inst->src[TEX_LOGICAL_SRC_SHADOW_C];
    const fs_reg &lod = inst->src[TEX_LOGICAL_SRC_LOD];
    const fs_reg &lod2 = inst->src[TEX_LOGICAL_SRC_LOD2];
+   const fs_reg &min_lod = inst->src[TEX_LOGICAL_SRC_MIN_LOD];
    const fs_reg &sample_index = inst->src[TEX_LOGICAL_SRC_SAMPLE_INDEX];
    const fs_reg &mcs = inst->src[TEX_LOGICAL_SRC_MCS];
    const fs_reg &surface = inst->src[TEX_LOGICAL_SRC_SURFACE];
@@ -4725,7 +4736,8 @@ lower_sampler_logical_send(const fs_builder &bld, fs_inst *inst, opcode op)
 
    if (devinfo->gen >= 7) {
       lower_sampler_logical_send_gen7(bld, inst, op, coordinate,
-                                      shadow_c, lod, lod2, sample_index,
+                                      shadow_c, lod, lod2, min_lod,
+                                      sample_index,
                                       mcs, surface, sampler, tg4_offset,
                                       coord_components, grad_components);
    } else if (devinfo->gen >= 5) {
@@ -5262,6 +5274,14 @@ static unsigned
 get_sampler_lowered_simd_width(const struct gen_device_info *devinfo,
                                const fs_inst *inst)
 {
+   /* If we have a min_lod parameter on anything other than a simple sample
+    * message, it will push it over 5 arguments and we have to fall back to
+    * SIMD8.
+    */
+   if (inst->opcode != SHADER_OPCODE_TEX &&
+       inst->components_read(TEX_LOGICAL_SRC_MIN_LOD))
+      return 8;
+
    /* Calculate the number of coordinate components that have to be present
     * assuming that additional arguments follow the texel coordinates in the
     * message payload.  On IVB+ there is no need for padding, on ILK-SNB we
diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index 165c70c7c29..d5a05aacdf5 100644
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -2968,7 +2968,7 @@ fs_visitor::emit_non_coherent_fb_read(const fs_builder &bld, const fs_reg &dst,
 
    /* Emit the instruction. */
    const fs_reg srcs[] = { coords, fs_reg(), brw_imm_ud(0), fs_reg(),
-                           sample, mcs,
+                           fs_reg(), sample, mcs,
                            brw_imm_ud(surface), brw_imm_ud(0),
                            fs_reg(), brw_imm_ud(3), brw_imm_ud(0) };
    STATIC_ASSERT(ARRAY_SIZE(srcs) == TEX_LOGICAL_NUM_SRCS);
@@ -4792,6 +4792,10 @@ fs_visitor::nir_emit_texture(const fs_builder &bld, nir_tex_instr *instr)
             break;
          }
          break;
+      case nir_tex_src_min_lod:
+         srcs[TEX_LOGICAL_SRC_MIN_LOD] =
+            retype(get_nir_src_imm(instr->src[i].src), BRW_REGISTER_TYPE_F);
+         break;
       case nir_tex_src_ms_index:
          srcs[TEX_LOGICAL_SRC_SAMPLE_INDEX] = retype(src, BRW_REGISTER_TYPE_UD);
          break;
diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
index aa6788b9fe5..2723ab6d4fb 100644
--- a/src/intel/compiler/brw_nir.c
+++ b/src/intel/compiler/brw_nir.c
@@ -656,6 +656,9 @@ brw_preprocess_nir(const struct brw_compiler *compiler, nir_shader *nir)
       .lower_txf_offset = true,
       .lower_rect_offset = true,
       .lower_txd_cube_map = true,
+      .lower_txb_shadow_clamp = true,
+      .lower_txd_shadow_clamp = true,
+      .lower_txd_offset_clamp = true,
    };
 
    OPT(nir_lower_tex, &tex_options);