Mesa (master): intel/compiler: Do interpolateAtOffset coordinate scaling in NIR

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 97ebb896afb6801d1fcd69556583975411379998
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=97ebb896afb6801d1fcd69556583975411379998

Author: Kenneth Graunke <kenneth at whitecape.org>
Date:   Tue Aug  4 19:01:13 2020 -0700

intel/compiler: Do interpolateAtOffset coordinate scaling in NIR

In our source languages, interpolateAtOffset() takes a floating point
offset in the range [-0.5, +0.5].  However, the hardware takes integer
valued offsets in the range [-8, 7], in units of 1/16th of a pixel.

So, we need to multiply and clamp the coordinates.  We were doing this
in the FS backend, but with the advent of IBC, I'd like to avoid doing
it twice.  This patch instead moves the lowering to NIR so we can reuse
it across both backends.

v2: Use nir_shader_instructions_pass (suggested by Eric Anholt).

Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira at intel.com>
Reviewed-by: Matt Turner <mattst88 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6193>

---

 src/intel/compiler/brw_fs_nir.cpp | 34 +++---------------------------
 src/intel/compiler/brw_nir.c      | 44 +++++++++++++++++++++++++++++++++++++++
 2 files changed, 47 insertions(+), 31 deletions(-)

diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index 56a0b2942bf..018be4588e8 100644
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -3629,8 +3629,8 @@ fs_visitor::nir_emit_fs_intrinsic(const fs_builder &bld,
 
       if (const_offset) {
          assert(nir_src_bit_size(instr->src[0]) == 32);
-         unsigned off_x = MIN2((int)(const_offset[0].f32 * 16), 7) & 0xf;
-         unsigned off_y = MIN2((int)(const_offset[1].f32 * 16), 7) & 0xf;
+         unsigned off_x = const_offset[0].u32 & 0xf;
+         unsigned off_y = const_offset[1].u32 & 0xf;
 
          emit_pixel_interpolater_send(bld,
                                       FS_OPCODE_INTERPOLATE_AT_SHARED_OFFSET,
@@ -3639,35 +3639,7 @@ fs_visitor::nir_emit_fs_intrinsic(const fs_builder &bld,
                                       brw_imm_ud(off_x | (off_y << 4)),
                                       interpolation);
       } else {
-         fs_reg src = vgrf(glsl_type::ivec2_type);
-         fs_reg offset_src = retype(get_nir_src(instr->src[0]),
-                                    BRW_REGISTER_TYPE_F);
-         for (int i = 0; i < 2; i++) {
-            fs_reg temp = vgrf(glsl_type::float_type);
-            bld.MUL(temp, offset(offset_src, bld, i), brw_imm_f(16.0f));
-            fs_reg itemp = vgrf(glsl_type::int_type);
-            /* float to int */
-            bld.MOV(itemp, temp);
-
-            /* Clamp the upper end of the range to +7/16.
-             * ARB_gpu_shader5 requires that we support a maximum offset
-             * of +0.5, which isn't representable in a S0.4 value -- if
-             * we didn't clamp it, we'd end up with -8/16, which is the
-             * opposite of what the shader author wanted.
-             *
-             * This is legal due to ARB_gpu_shader5's quantization
-             * rules:
-             *
-             * "Not all values of <offset> may be supported; x and y
-             * offsets may be rounded to fixed-point values with the
-             * number of fraction bits given by the
-             * implementation-dependent constant
-             * FRAGMENT_INTERPOLATION_OFFSET_BITS"
-             */
-            set_condmod(BRW_CONDITIONAL_L,
-                        bld.SEL(offset(src, bld, i), itemp, brw_imm_d(7)));
-         }
-
+         fs_reg src = retype(get_nir_src(instr->src[0]), BRW_REGISTER_TYPE_D);
          const enum opcode opcode = FS_OPCODE_INTERPOLATE_AT_PER_SLOT_OFFSET;
          emit_pixel_interpolater_send(bld,
                                       opcode,
diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
index b83a8800063..4589ddd9501 100644
--- a/src/intel/compiler/brw_nir.c
+++ b/src/intel/compiler/brw_nir.c
@@ -364,6 +364,45 @@ brw_nir_lower_tes_inputs(nir_shader *nir, const struct brw_vue_map *vue_map)
    }
 }
 
+/**
+ * Convert interpolateAtOffset() offsets from [-0.5, +0.5] floating point
+ * offsets to integer [-8, +7] offsets (in units of 1/16th of a pixel).
+ *
+ * We clamp to +7/16 on the upper end of the range, since +0.5 isn't
+ * representable in a S0.4 value; a naive conversion would give us -8/16,
+ * which is the opposite of what was intended.
+ *
+ * This is allowed by GL_ARB_gpu_shader5's quantization rules:
+ *
+ *    "Not all values of <offset> may be supported; x and y offsets may
+ *     be rounded to fixed-point values with the number of fraction bits
+ *     given by the implementation-dependent constant
+ *     FRAGMENT_INTERPOLATION_OFFSET_BITS."
+ */
+static bool
+lower_barycentric_at_offset(nir_builder *b, nir_instr *instr, void *data)
+{
+   if (instr->type != nir_instr_type_intrinsic)
+      return false;
+
+   nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+   if (intrin->intrinsic != nir_intrinsic_load_barycentric_at_offset)
+      return false;
+
+   b->cursor = nir_before_instr(instr);
+
+   assert(intrin->src[0].ssa);
+   nir_ssa_def *offset =
+      nir_imin(b, nir_imm_int(b, 7),
+               nir_f2i32(b, nir_fmul(b, nir_imm_float(b, 16),
+                                     intrin->src[0].ssa)));
+
+   nir_instr_rewrite_src(instr, &intrin->src[0], nir_src_for_ssa(offset));
+
+   return true;
+}
+
 void
 brw_nir_lower_fs_inputs(nir_shader *nir,
                         const struct gen_device_info *devinfo,
@@ -404,6 +443,11 @@ brw_nir_lower_fs_inputs(nir_shader *nir,
    if (devinfo->gen >= 11)
       nir_lower_interpolation(nir, ~0);
 
+   nir_shader_instructions_pass(nir, lower_barycentric_at_offset,
+                                nir_metadata_block_index |
+                                nir_metadata_dominance,
+                                NULL);
+
    /* This pass needs actual constants */
    nir_opt_constant_folding(nir);