spirv: implement Volatile memory semantic
[mesa.git] / src / compiler / spirv / spirv_to_nir.c
index 059d4b77717393831e2f0d99f7782e590372c7e1..dabb3c674bf504ddf06b3983a92fcca47e076b75 100644 (file)
@@ -2095,10 +2095,6 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
       val->constant = vtn_null_constant(b, val->type);
       break;
 
-   case SpvOpConstantSampler:
-      vtn_fail("OpConstantSampler requires Kernel Capability");
-      break;
-
    default:
       vtn_fail_with_opcode("Unhandled opcode", opcode);
    }
@@ -2107,20 +2103,6 @@ vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
    vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
 }
 
-SpvMemorySemanticsMask
-vtn_storage_class_to_memory_semantics(SpvStorageClass sc)
-{
-   switch (sc) {
-   case SpvStorageClassStorageBuffer:
-   case SpvStorageClassPhysicalStorageBuffer:
-      return SpvMemorySemanticsUniformMemoryMask;
-   case SpvStorageClassWorkgroup:
-      return SpvMemorySemanticsWorkgroupMemoryMask;
-   default:
-      return SpvMemorySemanticsMaskNone;
-   }
-}
-
 static void
 vtn_split_barrier_semantics(struct vtn_builder *b,
                             SpvMemorySemanticsMask semantics,
@@ -2170,7 +2152,8 @@ vtn_split_barrier_semantics(struct vtn_builder *b,
                    SpvMemorySemanticsOutputMemoryMask);
 
    const SpvMemorySemanticsMask other_semantics =
-      semantics & ~(order_semantics | av_vis_semantics | storage_semantics);
+      semantics & ~(order_semantics | av_vis_semantics | storage_semantics |
+                    SpvMemorySemanticsVolatileMask);
 
    if (other_semantics)
       vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics);
@@ -2632,9 +2615,21 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
       if (is_array && texop != nir_texop_lod)
          coord_components++;
 
-      coord = vtn_get_nir_ssa(b, w[idx++]);
+      struct vtn_ssa_value *coord_val = vtn_ssa_value(b, w[idx++]);
+      coord = coord_val->def;
       p->src = nir_src_for_ssa(nir_channels(&b->nb, coord,
                                             (1 << coord_components) - 1));
+
+      /* OpenCL allows integer sampling coordinates */
+      if (glsl_type_is_integer(coord_val->type) &&
+          opcode == SpvOpImageSampleExplicitLod) {
+         vtn_fail_if(b->shader->info.stage != MESA_SHADER_KERNEL,
+                     "Unless the Kernel capability is being used, the coordinate parameter "
+                     "OpImageSampleExplicitLod must be floating point.");
+
+         p->src = nir_src_for_ssa(nir_i2f32(&b->nb, p->src.ssa));
+      }
+
       p->src_type = nir_tex_src_coord;
       p++;
       break;
@@ -2801,9 +2796,18 @@ vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
    if (sampler && (access & ACCESS_NON_UNIFORM))
       instr->sampler_non_uniform = true;
 
-   /* for non-query ops, get dest_type from sampler type */
+   /* for non-query ops, get dest_type from SPIR-V return type */
    if (dest_type == nir_type_invalid) {
-      switch (glsl_get_sampler_result_type(image->type)) {
+      /* the return type should match the image type, unless the image type is
+       * VOID (CL image), in which case the return type dictates the sampler
+       */
+      enum glsl_base_type sampler_base =
+         glsl_get_sampler_result_type(image->type);
+      enum glsl_base_type ret_base = glsl_get_base_type(ret_type->type);
+      vtn_fail_if(sampler_base != ret_base && sampler_base != GLSL_TYPE_VOID,
+                  "SPIR-V return type mismatches image type. This is only valid "
+                  "for untyped images (OpenCL).");
+      switch (ret_base) {
       case GLSL_TYPE_FLOAT:   dest_type = nir_type_float;   break;
       case GLSL_TYPE_INT:     dest_type = nir_type_int;     break;
       case GLSL_TYPE_UINT:    dest_type = nir_type_uint;    break;
@@ -2979,6 +2983,14 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       access |= ACCESS_COHERENT;
       break;
 
+   case SpvOpImageQuerySizeLod:
+      res_val = vtn_untyped_value(b, w[3]);
+      image.image = vtn_get_image(b, w[3]);
+      image.coord = NULL;
+      image.sample = NULL;
+      image.lod = vtn_ssa_value(b, w[4])->def;
+      break;
+
    case SpvOpImageQuerySize:
       res_val = vtn_untyped_value(b, w[3]);
       image.image = vtn_get_image(b, w[3]);
@@ -2987,6 +2999,15 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       image.lod = NULL;
       break;
 
+   case SpvOpImageQueryFormat:
+   case SpvOpImageQueryOrder:
+      res_val = vtn_untyped_value(b, w[3]);
+      image.image = vtn_get_image(b, w[3]);
+      image.coord = NULL;
+      image.sample = NULL;
+      image.lod = NULL;
+      break;
+
    case SpvOpImageRead: {
       res_val = vtn_untyped_value(b, w[3]);
       image.image = vtn_get_image(b, w[3]);
@@ -3069,10 +3090,14 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
       vtn_fail_with_opcode("Invalid image opcode", opcode);
    }
 
+   if (semantics & SpvMemorySemanticsVolatileMask)
+      access |= ACCESS_VOLATILE;
+
    nir_intrinsic_op op;
    switch (opcode) {
 #define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
    OP(ImageQuerySize,            size)
+   OP(ImageQuerySizeLod,         size)
    OP(ImageRead,                 load)
    OP(ImageWrite,                store)
    OP(AtomicLoad,                load)
@@ -3092,6 +3117,8 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
    OP(AtomicOr,                  atomic_or)
    OP(AtomicXor,                 atomic_xor)
    OP(AtomicFAddEXT,             atomic_fadd)
+   OP(ImageQueryFormat,          format)
+   OP(ImageQueryOrder,           order)
 #undef OP
    default:
       vtn_fail_with_opcode("Invalid image opcode", opcode);
@@ -3101,15 +3128,19 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
 
    intrin->src[0] = nir_src_for_ssa(&image.image->dest.ssa);
 
-   if (opcode == SpvOpImageQuerySize) {
-      /* ImageQuerySize only has an LOD which is currently always 0 */
-      intrin->src[1] = nir_src_for_ssa(nir_imm_int(&b->nb, 0));
-   } else {
+   switch (opcode) {
+   case SpvOpImageQuerySize:
+   case SpvOpImageQuerySizeLod:
+   case SpvOpImageQueryFormat:
+   case SpvOpImageQueryOrder:
+      break;
+   default:
       /* The image coordinate is always 4 components but we may not have that
        * many.  Swizzle to compensate.
        */
       intrin->src[1] = nir_src_for_ssa(expand_to_vec4(&b->nb, image.coord));
       intrin->src[2] = nir_src_for_ssa(image.sample);
+      break;
    }
 
    /* The Vulkan spec says:
@@ -3129,17 +3160,24 @@ vtn_handle_image(struct vtn_builder *b, SpvOp opcode,
    nir_intrinsic_set_access(intrin, access);
 
    switch (opcode) {
-   case SpvOpAtomicLoad:
+   case SpvOpImageQueryFormat:
+   case SpvOpImageQueryOrder:
+      /* No additional sources */
+      break;
    case SpvOpImageQuerySize:
+      intrin->src[1] = nir_src_for_ssa(nir_imm_int(&b->nb, 0));
+      break;
+   case SpvOpImageQuerySizeLod:
+      intrin->src[1] = nir_src_for_ssa(image.lod);
+      break;
+   case SpvOpAtomicLoad:
    case SpvOpImageRead:
-      if (opcode == SpvOpImageRead || opcode == SpvOpAtomicLoad) {
-         /* Only OpImageRead can support a lod parameter if
-          * SPV_AMD_shader_image_load_store_lod is used but the current NIR
-          * intrinsics definition for atomics requires us to set it for
-          * OpAtomicLoad.
-          */
-         intrin->src[3] = nir_src_for_ssa(image.lod);
-      }
+      /* Only OpImageRead can support a lod parameter if
+      * SPV_AMD_shader_image_load_store_lod is used but the current NIR
+      * intrinsics definition for atomics requires us to set it for
+      * OpAtomicLoad.
+      */
+      intrin->src[3] = nir_src_for_ssa(image.lod);
       break;
    case SpvOpAtomicStore:
    case SpvOpImageWrite: {
@@ -3319,6 +3357,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
 
    SpvScope scope = SpvScopeInvocation;
    SpvMemorySemanticsMask semantics = 0;
+   enum gl_access_qualifier access = 0;
 
    switch (opcode) {
    case SpvOpAtomicLoad:
@@ -3352,6 +3391,9 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
       vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode);
    }
 
+   if (semantics & SpvMemorySemanticsVolatileMask)
+      access |= ACCESS_VOLATILE;
+
    /* uniform as "atomic counter uniform" */
    if (ptr->mode == vtn_variable_mode_atomic_counter) {
       nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
@@ -3398,7 +3440,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
       nir_intrinsic_op op  = get_ssbo_nir_atomic_op(b, opcode);
       atomic = nir_intrinsic_instr_create(b->nb.shader, op);
 
-      nir_intrinsic_set_access(atomic, ACCESS_COHERENT);
+      nir_intrinsic_set_access(atomic, access | ACCESS_COHERENT);
 
       int src = 0;
       switch (opcode) {
@@ -3452,7 +3494,9 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
       atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
 
       if (ptr->mode != vtn_variable_mode_workgroup)
-         nir_intrinsic_set_access(atomic, ACCESS_COHERENT);
+         access |= ACCESS_COHERENT;
+
+      nir_intrinsic_set_access(atomic, access);
 
       switch (opcode) {
       case SpvOpAtomicLoad:
@@ -3491,7 +3535,7 @@ vtn_handle_atomics(struct vtn_builder *b, SpvOp opcode,
    /* Atomic ordering operations will implicitly apply to the atomic operation
     * storage class, so include that too.
     */
-   semantics |= vtn_storage_class_to_memory_semantics(ptr->ptr_type->storage_class);
+   semantics |= vtn_mode_to_memory_semantics(ptr->mode);
 
    SpvMemorySemanticsMask before_semantics;
    SpvMemorySemanticsMask after_semantics;
@@ -4185,11 +4229,14 @@ vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
          spv_check_supported(kernel_image, cap);
          break;
 
+      case SpvCapabilityLiteralSampler:
+         spv_check_supported(literal_sampler, cap);
+         break;
+
       case SpvCapabilityImageReadWrite:
       case SpvCapabilityImageMipmap:
       case SpvCapabilityPipes:
       case SpvCapabilityDeviceEnqueue:
-      case SpvCapabilityLiteralSampler:
       case SpvCapabilityGenericPointer:
          vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
                   spirv_capability_to_string(cap));
@@ -4833,7 +4880,6 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpConstantFalse:
    case SpvOpConstant:
    case SpvOpConstantComposite:
-   case SpvOpConstantSampler:
    case SpvOpConstantNull:
    case SpvOpSpecConstantTrue:
    case SpvOpSpecConstantFalse:
@@ -4845,6 +4891,7 @@ vtn_handle_variable_or_type_instruction(struct vtn_builder *b, SpvOp opcode,
 
    case SpvOpUndef:
    case SpvOpVariable:
+   case SpvOpConstantSampler:
       vtn_handle_variables(b, opcode, w, count);
       break;
 
@@ -5047,7 +5094,6 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpImageFetch:
    case SpvOpImageGather:
    case SpvOpImageDrefGather:
-   case SpvOpImageQuerySizeLod:
    case SpvOpImageQueryLod:
    case SpvOpImageQueryLevels:
    case SpvOpImageQuerySamples:
@@ -5057,9 +5103,12 @@ vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
    case SpvOpImageRead:
    case SpvOpImageWrite:
    case SpvOpImageTexelPointer:
+   case SpvOpImageQueryFormat:
+   case SpvOpImageQueryOrder:
       vtn_handle_image(b, opcode, w, count);
       break;
 
+   case SpvOpImageQuerySizeLod:
    case SpvOpImageQuerySize: {
       struct vtn_type *image_type = vtn_get_value_type(b, w[3]);
       vtn_assert(image_type->base_type == vtn_base_type_image);