i965/fs: Implement texelFetch() on Gen4.
authorKenneth Graunke <kenneth@whitecape.org>
Tue, 6 Sep 2011 23:39:01 +0000 (16:39 -0700)
committerKenneth Graunke <kenneth@whitecape.org>
Mon, 19 Sep 2011 22:30:54 +0000 (15:30 -0700)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
src/mesa/drivers/dri/i965/brw_fs_emit.cpp
src/mesa/drivers/dri/i965/brw_fs_visitor.cpp

index 906c15821e00dab73f76799d2673e4485c294811..f742e84e1c6f204b7a7a650695c6bfbc56fdbb52 100644 (file)
@@ -295,6 +295,11 @@ fs_visitor::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src)
         assert(inst->mlen == 7 || inst->mlen == 10);
         msg_type = BRW_SAMPLER_MESSAGE_SIMD8_SAMPLE_GRADIENTS;
         break;
+      case FS_OPCODE_TXF:
+        assert(inst->mlen == 9);
+        msg_type = BRW_SAMPLER_MESSAGE_SIMD16_LD;
+        simd_mode = BRW_SAMPLER_SIMD_MODE_SIMD16;
+        break;
       case FS_OPCODE_TXS:
         assert(inst->mlen == 3);
         msg_type = BRW_SAMPLER_MESSAGE_SIMD16_RESINFO;
index 2c8d16035863d61dc105da445a53a7a6bb58048e..36042319be14cef11b4e66fb1e2a9fba96e6c807 100644 (file)
@@ -678,17 +678,25 @@ fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
        * instructions.  We'll need to do SIMD16 here.
        */
       simd16 = true;
-      assert(ir->op == ir_txb || ir->op == ir_txl);
+      assert(ir->op == ir_txb || ir->op == ir_txl || ir->op == ir_txf);
 
       for (int i = 0; i < ir->coordinate->type->vector_elements; i++) {
         fs_inst *inst = emit(BRW_OPCODE_MOV, fs_reg(MRF,
-                                                    base_mrf + mlen + i * 2),
+                                                    base_mrf + mlen + i * 2,
+                                                    coordinate.type),
                              coordinate);
         if (i < 3 && c->key.gl_clamp_mask[i] & (1 << sampler))
            inst->saturate = true;
         coordinate.reg_offset++;
       }
 
+      /* Initialize the rest of u/v/r with 0.0.  Empirically, this seems to
+       * be necessary for TXF (ld), but seems wise to do for all messages.
+       */
+      for (int i = ir->coordinate->type->vector_elements; i < 3; i++) {
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen + i * 2), fs_reg(0.0f));
+      }
+
       /* lod/bias appears after u/v/r. */
       mlen += 6;
 
@@ -698,7 +706,8 @@ fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
         mlen++;
       } else {
         ir->lod_info.lod->accept(this);
-        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen), this->result);
+        emit(BRW_OPCODE_MOV, fs_reg(MRF, base_mrf + mlen, this->result.type),
+                             this->result);
         mlen++;
       }
 
@@ -737,7 +746,7 @@ fs_visitor::emit_texture_gen4(ir_texture *ir, fs_reg dst, fs_reg coordinate,
       inst = emit(FS_OPCODE_TXS, dst);
       break;
    case ir_txf:
-      assert(!"GLSL 1.30 features unsupported");
+      inst = emit(FS_OPCODE_TXF, dst);
       break;
    }
    inst->base_mrf = base_mrf;