spirv: Implement SPV_KHR_non_semantic_info
[mesa.git] / src / compiler / spirv / spirv_to_nir.c
index 2de9bdf33893974688979231fd2cef08d6df6d99..29fef9fad6632cb51719b460aa46891d1ac12d07 100644 (file)
@@ -371,6 +371,14 @@ vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
    return w;
 }
 
+static bool
+vtn_handle_non_semantic_instruction(struct vtn_builder *b, SpvOp ext_opcode,
+                                    const uint32_t *w, unsigned count)
+{
+   /* Do nothing. */
+   return true;
+}
+
 static void
 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
                      const uint32_t *w, unsigned count)
@@ -392,6 +400,8 @@ vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
          val->ext_handler = vtn_handle_amd_shader_trinary_minmax_instruction;
       } else if (strcmp(ext, "OpenCL.std") == 0) {
          val->ext_handler = vtn_handle_opencl_instruction;
+      } else if (strstr(ext, "NonSemantic.") == ext) {
+         val->ext_handler = vtn_handle_non_semantic_instruction;
       } else {
          vtn_fail("Unsupported extension: %s", ext);
       }
@@ -1351,6 +1361,7 @@ vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
             case SpvStorageClassFunction:
             case SpvStorageClassWorkgroup:
             case SpvStorageClassCrossWorkgroup:
+            case SpvStorageClassUniformConstant:
                val->type->stride = align(glsl_get_cl_size(val->type->deref->type),
                                          glsl_get_cl_alignment(val->type->deref->type));
                break;
@@ -1954,12 +1965,22 @@ vtn_split_barrier_semantics(struct vtn_builder *b,
    *before = SpvMemorySemanticsMaskNone;
    *after = SpvMemorySemanticsMaskNone;
 
-   const SpvMemorySemanticsMask order_semantics =
+   SpvMemorySemanticsMask order_semantics =
       semantics & (SpvMemorySemanticsAcquireMask |
                    SpvMemorySemanticsReleaseMask |
                    SpvMemorySemanticsAcquireReleaseMask |
                    SpvMemorySemanticsSequentiallyConsistentMask);
 
+   if (util_bitcount(order_semantics) > 1) {
+      /* Old GLSLang versions incorrectly set all the ordering bits.  This was
+       * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
+       * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
+       */
+      vtn_warn("Multiple memory ordering semantics specified, "
+               "assuming AcquireRelease.");
+      order_semantics = SpvMemorySemanticsAcquireReleaseMask;
+   }
+
    const SpvMemorySemanticsMask av_vis_semantics =
       semantics & (SpvMemorySemanticsMakeAvailableMask |
                    SpvMemorySemanticsMakeVisibleMask);
@@ -1979,9 +2000,6 @@ vtn_split_barrier_semantics(struct vtn_builder *b,
    if (other_semantics)
       vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics);
 
-   vtn_fail_if(util_bitcount(order_semantics) > 1,
-               "Multiple memory ordering bits specified");
-
    /* SequentiallyConsistent is treated as AcquireRelease. */
 
    /* The RELEASE barrier happens BEFORE the operation, and it is usually
@@ -2016,10 +2034,24 @@ vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope,
                                SpvMemorySemanticsMask semantics)
 {
    nir_memory_semantics nir_semantics = 0;
-   switch (semantics & (SpvMemorySemanticsAcquireMask |
-                        SpvMemorySemanticsReleaseMask |
-                        SpvMemorySemanticsAcquireReleaseMask |
-                        SpvMemorySemanticsSequentiallyConsistentMask)) {
+
+   SpvMemorySemanticsMask order_semantics =
+      semantics & (SpvMemorySemanticsAcquireMask |
+                   SpvMemorySemanticsReleaseMask |
+                   SpvMemorySemanticsAcquireReleaseMask |
+                   SpvMemorySemanticsSequentiallyConsistentMask);
+
+   if (util_bitcount(order_semantics) > 1) {
+      /* Old GLSLang versions incorrectly set all the ordering bits.  This was
+       * fixed in c51287d744fb6e7e9ccc09f6f8451e6c64b1dad6 of glslang repo,
+       * and it is in GLSLang since revision "SPIRV99.1321" (from Jul-2016).
+       */
+      vtn_warn("Multiple memory ordering semantics bits specified, "
+               "assuming AcquireRelease.");
+      order_semantics = SpvMemorySemanticsAcquireReleaseMask;
+   }
+
+   switch (order_semantics) {
    case 0:
       /* Not an ordering barrier. */
       break;
@@ -2039,7 +2071,7 @@ vtn_emit_scoped_memory_barrier(struct vtn_builder *b, SpvScope scope,
       break;
 
    default:
-      vtn_fail("Multiple memory ordering bits specified");
+      unreachable("Invalid memory order semantics");
    }
 
    if (semantics & SpvMemorySemanticsMakeAvailableMask) {
@@ -2181,6 +2213,42 @@ vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
    return src;
 }
 
+static uint32_t
+image_operand_arg(struct vtn_builder *b, const uint32_t *w, uint32_t count,
+                  uint32_t mask_idx, SpvImageOperandsMask op)
+{
+   static const SpvImageOperandsMask ops_with_arg =
+      SpvImageOperandsBiasMask |
+      SpvImageOperandsLodMask |
+      SpvImageOperandsGradMask |
+      SpvImageOperandsConstOffsetMask |
+      SpvImageOperandsOffsetMask |
+      SpvImageOperandsConstOffsetsMask |
+      SpvImageOperandsSampleMask |
+      SpvImageOperandsMinLodMask |
+      SpvImageOperandsMakeTexelAvailableMask |
+      SpvImageOperandsMakeTexelVisibleMask;
+
+   assert(util_bitcount(op) == 1);
+   assert(w[mask_idx] & op);
+   assert(op & ops_with_arg);
+
+   uint32_t idx = util_bitcount(w[mask_idx] & (op - 1) & ops_with_arg) + 1;
+
+   /* Adjust indices for operands with two arguments. */
+   static const SpvImageOperandsMask ops_with_two_args =
+      SpvImageOperandsGradMask;
+   idx += util_bitcount(w[mask_idx] & (op - 1) & ops_with_two_args);
+
+   idx += mask_idx;
+
+   vtn_fail_if(idx + (op & ops_with_two_args ? 1 : 0) >= count,
+               "Image op claims to have %s but does not enough "
+               "following operands", spirv_imageoperands_to_string(op));
+
+   return idx;
+}
+
 static void
 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
                    const uint32_t *w, unsigned count)
@@ -2189,8 +2257,6 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       struct vtn_value *val =
          vtn_push_value(b, w[2], vtn_value_type_sampled_image);
       val->sampled_image = ralloc(b, struct vtn_sampled_image);
-      val->sampled_image->type =
-         vtn_value(b, w[1], vtn_value_type_type)->type;
       val->sampled_image->image =
          vtn_value(b, w[3], vtn_value_type_pointer)->pointer;
       val->sampled_image->sampler =
@@ -2209,18 +2275,21 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
 
    struct vtn_type *ret_type = vtn_value(b, w[1], vtn_value_type_type)->type;
 
-   struct vtn_sampled_image sampled;
+   struct vtn_pointer *image = NULL, *sampler = NULL;
    struct vtn_value *sampled_val = vtn_untyped_value(b, w[3]);
    if (sampled_val->value_type == vtn_value_type_sampled_image) {
-      sampled = *sampled_val->sampled_image;
+      image = sampled_val->sampled_image->image;
+      sampler = sampled_val->sampled_image->sampler;
    } else {
       vtn_assert(sampled_val->value_type == vtn_value_type_pointer);
-      sampled.type = sampled_val->pointer->type;
-      sampled.image = NULL;
-      sampled.sampler = sampled_val->pointer;
+      image = sampled_val->pointer;
    }
 
-   const struct glsl_type *image_type = sampled.type->type;
+   nir_deref_instr *image_deref = vtn_pointer_to_deref(b, image);
+   nir_deref_instr *sampler_deref =
+      sampler ? vtn_pointer_to_deref(b, sampler) : NULL;
+
+   const struct glsl_type *image_type = sampled_val->type->type;
    const enum glsl_sampler_dim sampler_dim = glsl_get_sampler_dim(image_type);
    const bool is_array = glsl_sampler_type_is_array(image_type);
    nir_alu_type dest_type = nir_type_invalid;
@@ -2243,7 +2312,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       break;
 
    case SpvOpImageFetch:
-      if (glsl_get_sampler_dim(image_type) == GLSL_SAMPLER_DIM_MS) {
+      if (sampler_dim == GLSL_SAMPLER_DIM_MS) {
          texop = nir_texop_txf_ms;
       } else {
          texop = nir_texop_txf;
@@ -2283,11 +2352,7 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
    nir_tex_src srcs[10]; /* 10 should be enough */
    nir_tex_src *p = srcs;
 
-   nir_deref_instr *sampler = vtn_pointer_to_deref(b, sampled.sampler);
-   nir_deref_instr *texture =
-      sampled.image ? vtn_pointer_to_deref(b, sampled.image) : sampler;
-
-   p->src = nir_src_for_ssa(&texture->dest.ssa);
+   p->src = nir_src_for_ssa(&image_deref->dest.ssa);
    p->src_type = nir_tex_src_texture_deref;
    p++;
 
@@ -2298,8 +2363,10 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
    case nir_texop_txd:
    case nir_texop_tg4:
    case nir_texop_lod:
-      /* These operations require a sampler */
-      p->src = nir_src_for_ssa(&sampler->dest.ssa);
+      vtn_fail_if(sampler == NULL,
+                  "%s requires an image of type OpTypeSampledImage",
+                  spirv_op_to_string(opcode));
+      p->src = nir_src_for_ssa(&sampler_deref->dest.ssa);
       p->src_type = nir_tex_src_sampler_deref;
       p++;
       break;
@@ -2417,51 +2484,75 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
    /* Now we need to handle some number of optional arguments */
    struct vtn_value *gather_offsets = NULL;
    if (idx < count) {
-      uint32_t operands = w[idx++];
+      uint32_t operands = w[idx];
 
       if (operands & SpvImageOperandsBiasMask) {
          vtn_assert(texop == nir_texop_tex);
          texop = nir_texop_txb;
-         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsBiasMask);
+         (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_bias);
       }
 
       if (operands & SpvImageOperandsLodMask) {
          vtn_assert(texop == nir_texop_txl || texop == nir_texop_txf ||
                     texop == nir_texop_txs);
-         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsLodMask);
+         (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_lod);
       }
 
       if (operands & SpvImageOperandsGradMask) {
          vtn_assert(texop == nir_texop_txl);
          texop = nir_texop_txd;
-         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
-         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsGradMask);
+         (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ddx);
+         (*p++) = vtn_tex_src(b, w[arg + 1], nir_tex_src_ddy);
       }
 
-      if (operands & SpvImageOperandsOffsetMask ||
-          operands & SpvImageOperandsConstOffsetMask)
-         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
+      vtn_fail_if(util_bitcount(operands & (SpvImageOperandsConstOffsetsMask |
+                                            SpvImageOperandsOffsetMask |
+                                            SpvImageOperandsConstOffsetMask)) > 1,
+                  "At most one of the ConstOffset, Offset, and ConstOffsets "
+                  "image operands can be used on a given instruction.");
+
+      if (operands & SpvImageOperandsOffsetMask) {
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsOffsetMask);
+         (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset);
+      }
+
+      if (operands & SpvImageOperandsConstOffsetMask) {
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsConstOffsetMask);
+         (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_offset);
+      }
 
       if (operands & SpvImageOperandsConstOffsetsMask) {
          vtn_assert(texop == nir_texop_tg4);
-         gather_offsets = vtn_value(b, w[idx++], vtn_value_type_constant);
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsConstOffsetsMask);
+         gather_offsets = vtn_value(b, w[arg], vtn_value_type_constant);
       }
 
       if (operands & SpvImageOperandsSampleMask) {
          vtn_assert(texop == nir_texop_txf_ms);
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsSampleMask);
          texop = nir_texop_txf_ms;
-         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
+         (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_ms_index);
       }
 
       if (operands & SpvImageOperandsMinLodMask) {
          vtn_assert(texop == nir_texop_tex ||
                     texop == nir_texop_txb ||
                     texop == nir_texop_txd);
-         (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_min_lod);
+         uint32_t arg = image_operand_arg(b, w, count, idx,
+                                          SpvImageOperandsMinLodMask);
+         (*p++) = vtn_tex_src(b, w[arg], nir_tex_src_min_lod);
       }
    }
-   /* We should have now consumed exactly all of the arguments */
-   vtn_assert(idx == count);
 
    nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
    instr->op = texop;
@@ -2476,10 +2567,10 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       is_shadow && glsl_get_components(ret_type->type) == 1;
    instr->component = gather_component;
 
-   if (sampled.image && (sampled.image->access & ACCESS_NON_UNIFORM))
+   if (image && (image->access & ACCESS_NON_UNIFORM))
       instr->texture_non_uniform = true;
 
-   if (sampled.sampler && (sampled.sampler->access & ACCESS_NON_UNIFORM))
+   if (sampler && (sampler->access & ACCESS_NON_UNIFORM))
       instr->sampler_non_uniform = true;
 
    /* for non-query ops, get dest_type from sampler type */
@@ -2666,10 +2757,10 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       const SpvImageOperandsMask operands =
          count > 5 ? w[5] : SpvImageOperandsMaskNone;
 
-      int idx = 6;
       if (operands & SpvImageOperandsSampleMask) {
-         image.sample = vtn_ssa_value(b, w[idx])->def;
-         idx++;
+         uint32_t arg = image_operand_arg(b, w, count, 5,
+                                          SpvImageOperandsSampleMask);
+         image.sample = vtn_ssa_value(b, w[arg])->def;
       } else {
          image.sample = nir_ssa_undef(&b->nb, 1, 32);
       }
@@ -2677,9 +2768,10 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       if (operands & SpvImageOperandsMakeTexelVisibleMask) {
          vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0,
                      "MakeTexelVisible requires NonPrivateTexel to also be set.");
+         uint32_t arg = image_operand_arg(b, w, count, 5,
+                                          SpvImageOperandsMakeTexelVisibleMask);
          semantics = SpvMemorySemanticsMakeVisibleMask;
-         scope = vtn_constant_uint(b, w[idx]);
-         idx++;
+         scope = vtn_constant_uint(b, w[arg]);
       }
 
       /* TODO: Volatile. */
@@ -2696,10 +2788,10 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       const SpvImageOperandsMask operands =
          count > 4 ? w[4] : SpvImageOperandsMaskNone;
 
-      int idx = 5;
       if (operands & SpvImageOperandsSampleMask) {
-         image.sample = vtn_ssa_value(b, w[idx])->def;
-         idx++;
+         uint32_t arg = image_operand_arg(b, w, count, 4,
+                                          SpvImageOperandsSampleMask);
+         image.sample = vtn_ssa_value(b, w[arg])->def;
       } else {
          image.sample = nir_ssa_undef(&b->nb, 1, 32);
       }
@@ -2707,8 +2799,10 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       if (operands & SpvImageOperandsMakeTexelAvailableMask) {
          vtn_fail_if((operands & SpvImageOperandsNonPrivateTexelMask) == 0,
                      "MakeTexelAvailable requires NonPrivateTexel to also be set.");
+         uint32_t arg = image_operand_arg(b, w, count, 4,
+                                          SpvImageOperandsMakeTexelAvailableMask);
          semantics = SpvMemorySemanticsMakeAvailableMask;
-         scope = vtn_constant_uint(b, w[idx]);
+         scope = vtn_constant_uint(b, w[arg]);
       }
 
       /* TODO: Volatile. */
@@ -3988,7 +4082,6 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
       case SpvAddressingModelLogical:
          vtn_fail_if(b->shader->info.stage >= MESA_SHADER_STAGES,
                      "AddressingModelLogical only supported for shaders");
-         b->shader->info.cs.ptr_size = 0;
          b->physical_ptrs = false;
          break;
       case SpvAddressingModelPhysicalStorageBuffer64EXT:
@@ -4048,6 +4141,17 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
       vtn_handle_decoration(b, opcode, w, count);
       break;
 
+   case SpvOpExtInst: {
+      struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
+      if (val->ext_handler == vtn_handle_non_semantic_instruction) {
+         /* NonSemantic extended instructions are acceptable in preamble. */
+         vtn_handle_non_semantic_instruction(b, w[4], w, count);
+         return true;
+      } else {
+         return false; /* End of preamble. */
+      }
+   }
+
    default:
       return false; /* End of preamble */
    }
@@ -4386,6 +4490,14 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
       vtn_handle_variables(b, opcode, w, count);
       break;
 
+   case SpvOpExtInst: {
+      struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
+      /* NonSemantic extended instructions are acceptable in preamble, others
+       * will indicate the end of preamble.
+       */
+      return val->ext_handler == vtn_handle_non_semantic_instruction;
+   }
+
    default:
       return false; /* End of preamble */
    }
@@ -4898,6 +5010,10 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
       break;
    }
 
+   case SpvOpLifetimeStart:
+   case SpvOpLifetimeStop:
+      break;
+
    default:
       vtn_fail_with_opcode("Unhandled opcode", opcode);
    }
@@ -5068,7 +5184,8 @@ spirv_to_nir(const uint32_t *words, size_t word_count,
    }
 
    /* Set shader info defaults */
-   b->shader->info.gs.invocations = 1;
+   if (stage == MESA_SHADER_GEOMETRY)
+      b->shader->info.gs.invocations = 1;
 
    /* Parse rounding mode execution modes. This has to happen earlier than
     * other changes in the execution modes since they can affect, for example,