Re: [PATCH vkd3d 1/3] vkd3d: Deal correctly with SM 5.1 register spaces.

28 Feb 2020

I'm not particularly qualified to review this code, but from suggestions 
from the vkd3d maintainer, and a few observations of my own...
On 2/20/20 5:18 AM, Hans-Kristian Arntzen wrote:
...
Resource index is found in idx[0] in SM 5.0, but idx[1] when using SM
5.1, and register space is encoded reparately. An rb_tree keeps track of
the internal resource index idx[0] and can map that to space/binding as
required when emitting SPIR-V.
For this to work, we must also make UAV counters register space aware.
In earlier implementation, UAV counter mask was assumed to correlate 1:1
with register_index, which breaks on SM 5.1.
Perhaps this should be broken out into a separate patch, then? If 
there's a reason why that doesn't make sense, could that be more clearly 
communicated?
Maybe the difference in resource index (i.e. [0] vs [1]) could also be 
broken out into a separate patch.
...
Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no
include/vkd3d_shader.h                   |  16 ++
  libs/vkd3d-shader/dxbc.c                 |  29 ++++
  libs/vkd3d-shader/spirv.c                | 194 ++++++++++++++++++++---
  libs/vkd3d-shader/vkd3d_shader_private.h |   5 +
  libs/vkd3d/command.c                     |  32 ++--
  libs/vkd3d/state.c                       |  65 +++++---
  libs/vkd3d/vkd3d_private.h               |   1 +
  7 files changed, 288 insertions(+), 54 deletions(-)

diff --git a/include/vkd3d_shader.h b/include/vkd3d_shader.h
index 6b4d3f5..ec52e26 100644
--- a/include/vkd3d_shader.h
+++ b/include/vkd3d_shader.h
@@ -35,6 +35,7 @@ enum vkd3d_shader_structure_type
      VKD3D_SHADER_STRUCTURE_TYPE_SCAN_INFO,
      VKD3D_SHADER_STRUCTURE_TYPE_TRANSFORM_FEEDBACK_INFO,
      VKD3D_SHADER_STRUCTURE_TYPE_DOMAIN_SHADER_COMPILE_ARGUMENTS,

VKD3D_SHADER_STRUCTURE_TYPE_EFFECTIVE_UAV_COUNTER_BINDING_INFO,
VKD3D_FORCE_32_BIT_ENUM(VKD3D_SHADER_STRUCTURE_TYPE),
};


@@ -138,6 +139,7 @@ struct vkd3d_shader_parameter
  struct vkd3d_shader_resource_binding
  {
      enum vkd3d_shader_descriptor_type type;

unsigned int register_space;
 unsigned int register_index;
 enum vkd3d_shader_visibility shader_visibility;
 unsigned int flags; /* vkd3d_shader_binding_flags */

@@ -159,8 +161,10 @@ struct vkd3d_shader_combined_resource_sampler
struct vkd3d_shader_uav_counter_binding
  {

unsigned int register_space;
 unsigned int register_index; /* u# */
 enum vkd3d_shader_visibility shader_visibility;

unsigned int counter_index;
struct vkd3d_shader_descriptor_binding binding;
 unsigned int offset;


@@ -168,6 +172,7 @@ struct vkd3d_shader_uav_counter_binding
struct vkd3d_shader_push_constant_buffer
  {

unsigned int register_space;
 unsigned int register_index;
 enum vkd3d_shader_visibility shader_visibility;

@@ -215,6 +220,17 @@ struct vkd3d_shader_transform_feedback_info
      unsigned int buffer_stride_count;
  };
+/* Extends vkd3d_shader_interface_info. */
+struct vkd3d_shader_effective_uav_counter_binding_info
+{

enum vkd3d_shader_structure_type type;
const void *next;

unsigned int *uav_register_spaces;
unsigned int *uav_register_bindings;
unsigned int uav_counter_count;

+};

enum vkd3d_shader_target
{
    VKD3D_SHADER_TARGET_NONE,

diff --git a/libs/vkd3d-shader/dxbc.c b/libs/vkd3d-shader/dxbc.c
index 98c51e4..b3f53ab 100644
--- a/libs/vkd3d-shader/dxbc.c
+++ b/libs/vkd3d-shader/dxbc.c
@@ -624,6 +624,10 @@ static void shader_sm4_read_dcl_resource(struct vkd3d_shader_instruction *ins,
          ins->flags = (opcode_token & VKD3D_SM5_UAV_FLAGS_MASK) >> VKD3D_SM5_UAV_FLAGS_SHIFT;
  shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.semantic.register_space);


if (shader_is_sm_5_1(priv))

   ins->declaration.semantic.register_index = ins->declaration.semantic.reg.reg.idx[1].offset;


else

   ins->declaration.semantic.register_index = ins->declaration.semantic.reg.reg.idx[0].offset;

}
static void shader_sm4_read_dcl_constant_buffer(struct vkd3d_shader_instruction *ins,


@@ -647,9 +651,12 @@ static void shader_sm4_read_dcl_constant_buffer(struct vkd3d_shader_instruction
              return;
          }

   ins->declaration.cb.register_index = ins->declaration.cb.src.reg.idx[1].offset;
    ins->declaration.cb.size = *tokens++;
    shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.cb.register_space);
}


else

   ins->declaration.cb.register_index = ins->declaration.cb.src.reg.idx[0].offset;

}
static void shader_sm4_read_dcl_sampler(struct vkd3d_shader_instruction *ins,


@@ -663,6 +670,10 @@ static void shader_sm4_read_dcl_sampler(struct vkd3d_shader_instruction *ins,
          FIXME("Unhandled sampler mode %#x.\n", ins->flags);
      shader_sm4_read_src_param(priv, &tokens, end, VKD3D_DATA_SAMPLER, &ins->declaration.sampler.src);
      shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.sampler.register_space);

if (shader_is_sm_5_1(priv))

   ins->declaration.sampler.register_index = ins->declaration.sampler.src.reg.idx[1].offset;


else

   ins->declaration.sampler.register_index = ins->declaration.sampler.src.reg.idx[0].offset;

}
static void shader_sm4_read_dcl_index_range(struct vkd3d_shader_instruction *ins,


@@ -863,6 +874,10 @@ static void shader_sm5_read_dcl_uav_raw(struct vkd3d_shader_instruction *ins,
      shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_UAV, &ins->declaration.raw_resource.dst);
      ins->flags = (opcode_token & VKD3D_SM5_UAV_FLAGS_MASK) >> VKD3D_SM5_UAV_FLAGS_SHIFT;
      shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.raw_resource.register_space);

if (shader_is_sm_5_1(priv))

   ins->declaration.raw_resource.register_index = ins->declaration.raw_resource.dst.reg.idx[1].offset;


else

   ins->declaration.raw_resource.register_index = ins->declaration.raw_resource.dst.reg.idx[0].offset;

}
static void shader_sm5_read_dcl_uav_structured(struct vkd3d_shader_instruction *ins,


@@ -874,9 +889,14 @@ static void shader_sm5_read_dcl_uav_structured(struct vkd3d_shader_instruction *
      shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_UAV, &ins->declaration.structured_resource.reg);
      ins->flags = (opcode_token & VKD3D_SM5_UAV_FLAGS_MASK) >> VKD3D_SM5_UAV_FLAGS_SHIFT;
      ins->declaration.structured_resource.byte_stride = *tokens;

tokens++;
 if (ins->declaration.structured_resource.byte_stride % 4)
     FIXME("Byte stride %u is not multiple of 4.\n", ins->declaration.structured_resource.byte_stride);
 shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.structured_resource.register_space);

if (shader_is_sm_5_1(priv))

   ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[1].offset;


else

   ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[0].offset;

}
static void shader_sm5_read_dcl_tgsm_raw(struct vkd3d_shader_instruction *ins,


@@ -909,9 +929,14 @@ static void shader_sm5_read_dcl_resource_structured(struct vkd3d_shader_instruct
  shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_RESOURCE, &ins->declaration.structured_resource.reg);
  ins->declaration.structured_resource.byte_stride = *tokens;


tokens++;
 if (ins->declaration.structured_resource.byte_stride % 4)
     FIXME("Byte stride %u is not multiple of 4.\n", ins->declaration.structured_resource.byte_stride);
 shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.structured_resource.register_space);

if (shader_is_sm_5_1(priv))

   ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[1].offset;


else

   ins->declaration.structured_resource.register_index = ins->declaration.structured_resource.reg.reg.idx[0].offset;

}
static void shader_sm5_read_dcl_resource_raw(struct vkd3d_shader_instruction *ins,


@@ -922,6 +947,10 @@ static void shader_sm5_read_dcl_resource_raw(struct vkd3d_shader_instruction *in
  shader_sm4_read_dst_param(priv, &tokens, end, VKD3D_DATA_RESOURCE, &ins->declaration.dst);
  shader_sm4_read_register_space(priv, &tokens, end, &ins->declaration.raw_resource.register_space);


if (shader_is_sm_5_1(priv))

   ins->declaration.raw_resource.register_index = ins->declaration.dst.reg.idx[1].offset;


else

   ins->declaration.raw_resource.register_index = ins->declaration.dst.reg.idx[0].offset;

}
static void shader_sm5_read_sync(struct vkd3d_shader_instruction *ins,


diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c
index 3d88be9..f286f8b 100644
--- a/libs/vkd3d-shader/spirv.c
+++ b/libs/vkd3d-shader/spirv.c
@@ -1890,6 +1890,20 @@ struct vkd3d_symbol
      } info;
  };
+struct vkd3d_sm51_symbol_key
+{

enum vkd3d_shader_descriptor_type descriptor_type;
unsigned int idx;

+};



+struct vkd3d_sm51_symbol
+{

struct rb_entry entry;
struct vkd3d_sm51_symbol_key key;
unsigned int register_space;
unsigned int resource_idx;

+};

static int vkd3d_symbol_compare(const void *key, const struct rb_entry *entry)
{
    const struct vkd3d_symbol *a = key;

@@ -1900,6 +1914,13 @@ static int vkd3d_symbol_compare(const void *key, const struct rb_entry *entry)
      return memcmp(&a->key, &b->key, sizeof(a->key));
  }
+static int vkd3d_sm51_symbol_compare(const void *key, const struct rb_entry *entry)
+{

const struct vkd3d_sm51_symbol_key *a = key;
const struct vkd3d_sm51_symbol *b = RB_ENTRY_VALUE(entry, const struct vkd3d_sm51_symbol, entry);
return memcmp(a, &b->key, sizeof(*a));

+}

static void vkd3d_symbol_free(struct rb_entry *entry, void *context)
{
    struct vkd3d_symbol *s = RB_ENTRY_VALUE(entry, struct vkd3d_symbol, entry);

@@ -1907,6 +1928,13 @@ static void vkd3d_symbol_free(struct rb_entry *entry, void *context)
      vkd3d_free(s);
  }
+static void vkd3d_sm51_symbol_free(struct rb_entry *entry, void *context)
+{

struct vkd3d_sm51_symbol *s = RB_ENTRY_VALUE(entry, struct vkd3d_sm51_symbol, entry);

vkd3d_free(s);

+}

static void vkd3d_symbol_make_register(struct vkd3d_symbol *symbol,
        const struct vkd3d_shader_register *reg)
{

@@ -2052,6 +2080,7 @@ struct vkd3d_hull_shader_variables
struct vkd3d_dxbc_compiler
  {

struct vkd3d_shader_version shader_version;
 struct vkd3d_spirv_builder spirv_builder;
uint32_t options;


@@ -2062,6 +2091,8 @@ struct vkd3d_dxbc_compiler
      struct vkd3d_hull_shader_variables hs;
      uint32_t sample_positions_id;

struct rb_tree sm51_resource_table;

enum vkd3d_shader_type shader_type;

unsigned int branch_id;



@@ -2107,6 +2138,11 @@ struct vkd3d_dxbc_compiler
      size_t spec_constants_size;
  };
+static bool shader_is_sm_5_1(const struct vkd3d_dxbc_compiler *compiler)
+{

return (compiler->shader_version.major * 100 + compiler->shader_version.minor) >= 501;

+}

static bool is_control_point_phase(const struct vkd3d_shader_phase *phase)
{
    return phase && phase->type == VKD3DSIH_HS_CONTROL_POINT_PHASE;

@@ -2131,6 +2167,8 @@ struct vkd3d_dxbc_compiler *vkd3d_dxbc_compiler_create(const struct vkd3d_shader
  memset(compiler, 0, sizeof(*compiler));



compiler->shader_version = *shader_version;
max_element_count = max(output_signature->element_count, patch_constant_signature->element_count);
if (!(compiler->output_info = vkd3d_calloc(max_element_count, sizeof(*compiler->output_info))))
{



@@ -2142,6 +2180,7 @@ struct vkd3d_dxbc_compiler *vkd3d_dxbc_compiler_create(const struct vkd3d_shader
      compiler->options = compiler_options;
  rb_init(&compiler->symbol_table, vkd3d_symbol_compare);


rb_init(&compiler->sm51_resource_table, vkd3d_sm51_symbol_compare);
compiler->shader_type = shader_version->type;


@@ -2227,9 +2266,10 @@ static bool vkd3d_dxbc_compiler_check_shader_visibility(const struct vkd3d_dxbc_
  }
static struct vkd3d_push_constant_buffer_binding *vkd3d_dxbc_compiler_find_push_constant_buffer(

   const struct vkd3d_dxbc_compiler *compiler, const struct vkd3d_shader_register *reg)




   const struct vkd3d_dxbc_compiler *compiler, const struct vkd3d_shader_constant_buffer *cb)

{


unsigned int reg_idx = reg->idx[0].offset;


unsigned int reg_idx = cb->register_index;

unsigned int reg_space = cb->register_space;
 unsigned int i;
for (i = 0; i < compiler->shader_interface.push_constant_buffer_count; ++i)


@@ -2239,7 +2279,7 @@ static struct vkd3d_push_constant_buffer_binding *vkd3d_dxbc_compiler_find_push_
          if (!vkd3d_dxbc_compiler_check_shader_visibility(compiler, current->pc.shader_visibility))
              continue;

   if (current->pc.register_index == reg_idx)




   if (current->pc.register_index == reg_idx && current->pc.register_space == reg_space)
        return current;
}




@@ -2274,6 +2314,49 @@ static bool vkd3d_dxbc_compiler_has_combined_sampler(const struct vkd3d_dxbc_com
      return false;
  }
+static bool vkd3d_get_binding_info_for_register(

   struct vkd3d_dxbc_compiler *compiler,


   const struct vkd3d_shader_register *reg,


   unsigned int *reg_space, unsigned int *reg_binding)



+{

const struct vkd3d_sm51_symbol *symbol;
struct vkd3d_sm51_symbol_key key;
const struct rb_entry *entry;

if (shader_is_sm_5_1(compiler))
{
   key.descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_UNKNOWN;


   if (reg->type == VKD3DSPR_CONSTBUFFER)


       key.descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_CBV;


   else if (reg->type == VKD3DSPR_RESOURCE)


       key.descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_SRV;


   else if (reg->type == VKD3DSPR_UAV)


       key.descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_UAV;


   else if (reg->type == VKD3DSPR_SAMPLER)


       key.descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_SAMPLER;


   else


       FIXME("Unhandled register type %#x.\n", reg->type);



   key.idx = reg->idx[0].offset;


   entry = rb_get(&compiler->sm51_resource_table, &key);


   if (entry)


   {


       symbol = RB_ENTRY_VALUE(entry, const struct vkd3d_sm51_symbol, entry);


       *reg_space = symbol->register_space;


       *reg_binding = symbol->resource_idx;


       return true;


   }


   else


       return false;


}
else
{
   *reg_space = 0;


   *reg_binding = reg->idx[0].offset;


   return true;


}

+}

static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor_binding(
        struct vkd3d_dxbc_compiler *compiler, const struct vkd3d_shader_register *reg,
        enum vkd3d_shader_resource_type resource_type, bool is_uav_counter)

@@ -2282,8 +2365,9 @@ static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor
      enum vkd3d_shader_descriptor_type descriptor_type;
      enum vkd3d_shader_binding_flag resource_type_flag;
      struct vkd3d_shader_descriptor_binding binding;

unsigned int reg_idx = reg->idx[0].offset;
 unsigned int i;


unsigned int reg_space = 0;

unsigned int reg_idx = 0;
descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_UNKNOWN;
 if (reg->type == VKD3DSPR_CONSTBUFFER)


As long as we're here, it's my understanding that variable declarations 
should be sorted in "reverse Christmas tree" order.
...
@@ -2300,6 +2384,11 @@ static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor
      resource_type_flag = resource_type == VKD3D_SHADER_RESOURCE_BUFFER
              ? VKD3D_SHADER_BINDING_FLAG_BUFFER : VKD3D_SHADER_BINDING_FLAG_IMAGE;

if (!vkd3d_get_binding_info_for_register(compiler, reg, &reg_space, &reg_idx))
{
   ERR("Failed to find binding for resource type %#x.\n", reg->type);


}
if (is_uav_counter)
{
    assert(descriptor_type == VKD3D_SHADER_DESCRIPTOR_TYPE_UAV);



@@ -2313,8 +2402,19 @@ static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor
              if (current->offset)
                  FIXME("Atomic counter offsets are not supported yet.\n");

       if (current->register_index == reg_idx)




       /* Do not use space/binding, but just the plain index here, since that's how the UAV counter mask is exposed. */


       if (current->counter_index == reg->idx[0].offset)


       {


           /* Let pipeline know what the actual space/bindings for the counter are. */


           const struct vkd3d_shader_effective_uav_counter_binding_info *binding_info =


                   vkd3d_find_struct(shader_interface->next, EFFECTIVE_UAV_COUNTER_BINDING_INFO);


           if (binding_info && current->counter_index < binding_info->uav_counter_count)


           {


               binding_info->uav_register_spaces[current->counter_index] = reg_space;


               binding_info->uav_register_bindings[current->counter_index] = reg_idx;


           }
            return current->binding;


       }
    }
    if (shader_interface->uav_counter_count)
        FIXME("Could not find descriptor binding for UAV counter %u.\n", reg_idx);



@@ -2331,7 +2431,7 @@ static struct vkd3d_shader_descriptor_binding vkd3d_dxbc_compiler_get_descriptor
              if (!vkd3d_dxbc_compiler_check_shader_visibility(compiler, current->shader_visibility))
                  continue;

       if (current->type == descriptor_type && current->register_index == reg_idx)




       if (current->type == descriptor_type && current->register_index == reg_idx && current->register_space == reg_space)
            return current->binding;
    }
    if (shader_interface->binding_count)



@@ -2828,7 +2928,8 @@ static void vkd3d_dxbc_compiler_emit_dereference_register(struct vkd3d_dxbc_comp
      {
          assert(!reg->idx[0].rel_addr);
          indexes[index_count++] = vkd3d_dxbc_compiler_get_constant_uint(compiler, register_info->member_idx);

   indexes[index_count++] = vkd3d_dxbc_compiler_emit_register_addressing(compiler, &reg->idx[1]);




   indexes[index_count++] = vkd3d_dxbc_compiler_emit_register_addressing(compiler,


           &reg->idx[shader_is_sm_5_1(compiler) ? 2 : 1]);
}
else if (reg->type == VKD3DSPR_IMMCONSTBUFFER)
{



@@ -2838,6 +2939,11 @@ static void vkd3d_dxbc_compiler_emit_dereference_register(struct vkd3d_dxbc_comp
      {
          indexes[index_count++] = vkd3d_dxbc_compiler_emit_register_addressing(compiler, &reg->idx[1]);
      }

else if (reg->type == VKD3DSPR_SAMPLER)
{
   /* SM 5.1 will have an index here referring to something which we throw away. */


   index_count = 0;


}
 else if (register_info->is_aggregate)
 {
     struct vkd3d_shader_register_index reg_idx = reg->idx[0];

@@ -4914,7 +5020,8 @@ static void vkd3d_dxbc_compiler_emit_push_constant_buffers(struct vkd3d_dxbc_com
          if (!cb->reg.type)
              continue;

   cb_size = cb->reg.idx[1].offset;




   cb_size = (cb->pc.size + 15) / 16;


    length_id = vkd3d_dxbc_compiler_get_constant_uint(compiler, cb_size);
    member_ids[j]  = vkd3d_spirv_build_op_type_array(builder, vec4_id, length_id);
    vkd3d_spirv_build_op_decorate1(builder, member_ids[j], SpvDecorationArrayStride, 16);



@@ -4965,10 +5072,19 @@ static void vkd3d_dxbc_compiler_emit_dcl_constant_buffer(struct vkd3d_dxbc_compi
  assert(!(instruction->flags & ~VKD3DSI_INDEXED_DYNAMIC));



if (cb->register_space)
   FIXME("Unhandled register space %u.\n", cb->register_space);




if (shader_is_sm_5_1(compiler))
{
   struct vkd3d_sm51_symbol *sym;


   sym = vkd3d_calloc(1, sizeof(*sym));


   sym->key.idx = reg->idx[0].offset;


   sym->key.descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_CBV;


   sym->register_space = instruction->declaration.cb.register_space;


   sym->resource_idx = instruction->declaration.cb.register_index;


   if (rb_put(&compiler->sm51_resource_table, &sym->key, &sym->entry) == -1)


       vkd3d_free(sym);


}


if ((push_cb = vkd3d_dxbc_compiler_find_push_constant_buffer(compiler, reg)))


if ((push_cb = vkd3d_dxbc_compiler_find_push_constant_buffer(compiler, cb)))
 {
     /* Push constant buffers are handled in
      * vkd3d_dxbc_compiler_emit_push_constant_buffers().

@@ -5050,8 +5166,17 @@ static void vkd3d_dxbc_compiler_emit_dcl_sampler(struct vkd3d_dxbc_compiler *com
      uint32_t type_id, ptr_type_id, var_id;
      struct vkd3d_symbol reg_symbol;

if (instruction->declaration.sampler.register_space)
   FIXME("Unhandled register space %u.\n", instruction->declaration.sampler.register_space);




if (shader_is_sm_5_1(compiler))

{

   struct vkd3d_sm51_symbol *sym;


   sym = vkd3d_calloc(1, sizeof(*sym));


   sym->key.idx = reg->idx[0].offset;


   sym->key.descriptor_type = VKD3D_SHADER_DESCRIPTOR_TYPE_SAMPLER;


   sym->register_space = instruction->declaration.sampler.register_space;


   sym->resource_idx = instruction->declaration.sampler.register_index;


   if (rb_put(&compiler->sm51_resource_table, &sym->key, &sym->entry) == -1)


       vkd3d_free(sym);


}
if (vkd3d_dxbc_compiler_has_combined_sampler(compiler, NULL, reg))
     return;


@@ -5272,8 +5397,18 @@ static void vkd3d_dxbc_compiler_emit_dcl_resource(struct vkd3d_dxbc_compiler *co
  {
      const struct vkd3d_shader_semantic *semantic = &instruction->declaration.semantic;

if (semantic->register_space)
   FIXME("Unhandled register space %u.\n", semantic->register_space);




if (shader_is_sm_5_1(compiler))
{
   struct vkd3d_sm51_symbol *sym;


   sym = vkd3d_calloc(1, sizeof(*sym));


   sym->key.idx = semantic->reg.reg.idx[0].offset;


   sym->key.descriptor_type = semantic->reg.reg.type == VKD3DSPR_UAV ? VKD3D_SHADER_DESCRIPTOR_TYPE_UAV : VKD3D_SHADER_DESCRIPTOR_TYPE_SRV;


   sym->register_space = semantic->register_space;


   sym->resource_idx = semantic->register_index;


   if (rb_put(&compiler->sm51_resource_table, &sym->key, &sym->entry) == -1)


       vkd3d_free(sym);


}
if (instruction->flags)
    FIXME("Unhandled UAV flags %#x.\n", instruction->flags);




@@ -5286,8 +5421,18 @@ static void vkd3d_dxbc_compiler_emit_dcl_resource_raw(struct vkd3d_dxbc_compiler
  {
      const struct vkd3d_shader_raw_resource *resource = &instruction->declaration.raw_resource;

if (resource->register_space)
   FIXME("Unhandled register space %u.\n", resource->register_space);




if (shader_is_sm_5_1(compiler))
{
   struct vkd3d_sm51_symbol *sym;


   sym = vkd3d_calloc(1, sizeof(*sym));


   sym->key.idx = resource->dst.reg.idx[0].offset;


   sym->key.descriptor_type = resource->dst.reg.type == VKD3DSPR_UAV ? VKD3D_SHADER_DESCRIPTOR_TYPE_UAV : VKD3D_SHADER_DESCRIPTOR_TYPE_SRV;


   sym->register_space = resource->register_space;


   sym->resource_idx = resource->register_index;


   if (rb_put(&compiler->sm51_resource_table, &sym->key, &sym->entry) == -1)


       vkd3d_free(sym);


}
if (instruction->flags)
    FIXME("Unhandled UAV flags %#x.\n", instruction->flags);




@@ -5302,8 +5447,18 @@ static void vkd3d_dxbc_compiler_emit_dcl_resource_structured(struct vkd3d_dxbc_c
      const struct vkd3d_shader_register *reg = &resource->reg.reg;
      unsigned int stride = resource->byte_stride;

if (resource->register_space)
   FIXME("Unhandled register space %u.\n", resource->register_space);




if (shader_is_sm_5_1(compiler))
{
   struct vkd3d_sm51_symbol *sym;


   sym = vkd3d_calloc(1, sizeof(*sym));


   sym->key.idx = resource->reg.reg.idx[0].offset;


   sym->key.descriptor_type = resource->reg.reg.type == VKD3DSPR_UAV ? VKD3D_SHADER_DESCRIPTOR_TYPE_UAV : VKD3D_SHADER_DESCRIPTOR_TYPE_SRV;


   sym->register_space = resource->register_space;


   sym->resource_idx = resource->register_index;


   if (rb_put(&compiler->sm51_resource_table, &sym->key, &sym->entry) == -1)


       vkd3d_free(sym);


}
if (instruction->flags)
    FIXME("Unhandled UAV flags %#x.\n", instruction->flags);




@@ -8717,6 +8872,7 @@ void vkd3d_dxbc_compiler_destroy(struct vkd3d_dxbc_compiler *compiler)
      vkd3d_spirv_builder_free(&compiler->spirv_builder);
  rb_destroy(&compiler->symbol_table, vkd3d_symbol_free, NULL);


rb_destroy(&compiler->sm51_resource_table, vkd3d_sm51_symbol_free, NULL);
vkd3d_free(compiler->shader_phases);
 vkd3d_free(compiler->spec_constants);


diff --git a/libs/vkd3d-shader/vkd3d_shader_private.h b/libs/vkd3d-shader/vkd3d_shader_private.h
index 100d515..135b48b 100644
--- a/libs/vkd3d-shader/vkd3d_shader_private.h
+++ b/libs/vkd3d-shader/vkd3d_shader_private.h
@@ -615,6 +615,7 @@ struct vkd3d_shader_semantic
      enum vkd3d_data_type resource_data_type;
      struct vkd3d_shader_dst_param reg;
      unsigned int register_space;

unsigned int register_index;
};
enum vkd3d_shader_input_sysval_semantic


@@ -662,6 +663,7 @@ struct vkd3d_shader_register_semantic
  struct vkd3d_shader_sampler
  {
      struct vkd3d_shader_src_param src;

unsigned int register_index;
 unsigned int register_space;
};

@@ -669,6 +671,7 @@ struct vkd3d_shader_constant_buffer
  {
      struct vkd3d_shader_src_param src;
      unsigned int size;

unsigned int register_index;
 unsigned int register_space;
};

@@ -676,12 +679,14 @@ struct vkd3d_shader_structured_resource
  {
      struct vkd3d_shader_dst_param reg;
      unsigned int byte_stride;

unsigned int register_index;
 unsigned int register_space;
};
struct vkd3d_shader_raw_resource
{
 struct vkd3d_shader_dst_param dst;

unsigned int register_index;
 unsigned int register_space;
};


diff --git a/libs/vkd3d/command.c b/libs/vkd3d/command.c
index 8a7ff66..7245802 100644
--- a/libs/vkd3d/command.c
+++ b/libs/vkd3d/command.c
@@ -2655,7 +2655,7 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list
      const struct d3d12_root_descriptor_table *descriptor_table;
      const struct d3d12_root_descriptor_table_range *range;
      VkDevice vk_device = list->device->vk_device;

unsigned int i, j, descriptor_count;


unsigned int i, j, k, descriptor_count;
 struct d3d12_desc *descriptor;
descriptor_table = root_signature_get_descriptor_table(root_signature, index);


@@ -2678,14 +2678,26 @@ static void d3d12_command_list_update_descriptor_table(struct d3d12_command_list
              unsigned int register_idx = range->base_register_idx + j;
          /* Track UAV counters. */


       if (range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV


               && register_idx < ARRAY_SIZE(bindings->vk_uav_counter_views))




       if (list->state->uav_counter_mask != 0 && range->descriptor_magic == VKD3D_DESCRIPTOR_MAGIC_UAV)
        {




           VkBufferView vk_counter_view = descriptor->magic == VKD3D_DESCRIPTOR_MAGIC_UAV


                   ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE;


           if (bindings->vk_uav_counter_views[register_idx] != vk_counter_view)


               bindings->uav_counter_dirty_mask |= 1u << register_idx;


           bindings->vk_uav_counter_views[register_idx] = vk_counter_view;




           const struct vkd3d_shader_uav_counter_binding *counter_bindings = list->state->uav_counters;


           for (k = 0; k < VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS; k++)


           {


               if (list->state->uav_counter_mask & (1u << k))


               {


                   if (counter_bindings->register_space == range->register_space &&


                       counter_bindings->register_index == register_idx)


                   {


                       VkBufferView vk_counter_view = descriptor->magic == VKD3D_DESCRIPTOR_MAGIC_UAV


                                                      ? descriptor->u.view->vk_counter_view : VK_NULL_HANDLE;


                       if (bindings->vk_uav_counter_views[k] != vk_counter_view)


                           bindings->uav_counter_dirty_mask |= 1u << k;


                       bindings->vk_uav_counter_views[k] = vk_counter_view;


                       break;


                   }


                   counter_bindings++;


               }


           }
        }

        if (!vk_write_descriptor_set_from_d3d12_desc(current_descriptor_write,



@@ -2841,7 +2853,7 @@ static void d3d12_command_list_update_uav_counter_descriptors(struct d3d12_comma
          const struct vkd3d_shader_uav_counter_binding *uav_counter = &state->uav_counters[i];
          const VkBufferView *vk_uav_counter_views = bindings->vk_uav_counter_views;

   assert(vk_uav_counter_views[uav_counter->register_index]);




   assert(vk_uav_counter_views[uav_counter->counter_index]);

    vk_descriptor_writes[i].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
    vk_descriptor_writes[i].pNext = NULL;



@@ -2852,7 +2864,7 @@ static void d3d12_command_list_update_uav_counter_descriptors(struct d3d12_comma
          vk_descriptor_writes[i].descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER;
          vk_descriptor_writes[i].pImageInfo = NULL;
          vk_descriptor_writes[i].pBufferInfo = NULL;

   vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->register_index];




   vk_descriptor_writes[i].pTexelBufferView = &vk_uav_counter_views[uav_counter->counter_index];
}

VK_CALL(vkUpdateDescriptorSets(vk_device, uav_counter_count, vk_descriptor_writes, 0, NULL));



diff --git a/libs/vkd3d/state.c b/libs/vkd3d/state.c
index e1f7da9..9add56b 100644
--- a/libs/vkd3d/state.c
+++ b/libs/vkd3d/state.c
@@ -309,12 +309,6 @@ static bool vk_binding_from_d3d12_descriptor_range(struct VkDescriptorSetLayoutB
              = vk_descriptor_type_from_d3d12_range_type(descriptor_range->RangeType, is_buffer);
      binding_desc->descriptorCount = 1;

if (descriptor_range->RegisterSpace)
{
   FIXME("Unhandled register space %u.\n", descriptor_range->RegisterSpace);


   return false;


}
binding_desc->stageFlags = stage_flags_from_visibility(shader_visibility);
binding_desc->pImmutableSamplers = NULL;




@@ -495,12 +489,6 @@ static HRESULT d3d12_root_signature_init_push_constants(struct d3d12_root_signat
          if (p->ParameterType != D3D12_ROOT_PARAMETER_TYPE_32BIT_CONSTANTS)
              continue;

   if (p->u.Constants.RegisterSpace)


   {


       FIXME("Unhandled register space %u for parameter %u.\n", p->u.Constants.RegisterSpace, i);


       return E_NOTIMPL;


   }


    idx = push_constant_count == 1 ? 0 : p->ShaderVisibility;
    offset = push_constants_offset[idx];
    push_constants_offset[idx] += p->u.Constants.Num32BitValues * sizeof(uint32_t);



@@ -510,6 +498,7 @@ static HRESULT d3d12_root_signature_init_push_constants(struct d3d12_root_signat
                  ? push_constants[0].stageFlags : stage_flags_from_visibility(p->ShaderVisibility);
          root_constant->offset = offset;

   root_signature->root_constants[j].register_space = p->u.Constants.RegisterSpace;
    root_signature->root_constants[j].register_index = p->u.Constants.ShaderRegister;
    root_signature->root_constants[j].shader_visibility
            = vkd3d_shader_visibility_from_d3d12(p->ShaderVisibility);



@@ -533,7 +522,7 @@ struct vkd3d_descriptor_set_context
  };
static void d3d12_root_signature_append_vk_binding(struct d3d12_root_signature *root_signature,

   enum vkd3d_shader_descriptor_type descriptor_type, unsigned int register_idx,




   enum vkd3d_shader_descriptor_type descriptor_type, unsigned int register_space, unsigned int register_idx,
    bool buffer_descriptor, enum vkd3d_shader_visibility shader_visibility,
    struct vkd3d_descriptor_set_context *context)

{

@@ -541,6 +530,7 @@ static void d3d12_root_signature_append_vk_binding(struct d3d12_root_signature *
              = &root_signature->descriptor_mapping[context->descriptor_index++];
  mapping->type = descriptor_type;


mapping->register_space = register_space;
 mapping->register_index = register_idx;
 mapping->shader_visibility = shader_visibility;
 mapping->flags = buffer_descriptor ? VKD3D_SHADER_BINDING_FLAG_BUFFER : VKD3D_SHADER_BINDING_FLAG_IMAGE;

@@ -549,7 +539,7 @@ static void d3d12_root_signature_append_vk_binding(struct d3d12_root_signature *
  }
static uint32_t d3d12_root_signature_assign_vk_bindings(struct d3d12_root_signature *root_signature,

   enum vkd3d_shader_descriptor_type descriptor_type, unsigned int base_register_idx,




   enum vkd3d_shader_descriptor_type descriptor_type, unsigned int register_space, unsigned int base_register_idx,
    unsigned int binding_count, bool is_buffer_descriptor, bool duplicate_descriptors,
    enum vkd3d_shader_visibility shader_visibility, struct vkd3d_descriptor_set_context *context)

{

@@ -566,10 +556,10 @@ static uint32_t d3d12_root_signature_assign_vk_bindings(struct d3d12_root_signat
      {
          if (duplicate_descriptors)
              d3d12_root_signature_append_vk_binding(root_signature, descriptor_type,

               base_register_idx + i, true, shader_visibility, context);




               register_space, base_register_idx + i, true, shader_visibility, context);

    d3d12_root_signature_append_vk_binding(root_signature, descriptor_type,




           base_register_idx + i, is_buffer_descriptor, shader_visibility, context);




           register_space, base_register_idx + i, is_buffer_descriptor, shader_visibility, context);
}
return first_binding;

}

@@ -625,7 +615,7 @@ static HRESULT d3d12_root_signature_init_root_descriptor_tables(struct d3d12_roo
          vk_binding = d3d12_root_signature_assign_vk_bindings(root_signature,
                  vkd3d_descriptor_type_from_d3d12_range_type(range->RangeType),


               range->BaseShaderRegister, range->NumDescriptors, false, true,




               range->RegisterSpace, range->BaseShaderRegister, range->NumDescriptors, false, true,
                vkd3d_shader_visibility_from_d3d12(p->ShaderVisibility), context);

        /* Unroll descriptor range. */



@@ -658,6 +648,7 @@ static HRESULT d3d12_root_signature_init_root_descriptor_tables(struct d3d12_roo
              table->ranges[j].binding = vk_binding;
              table->ranges[j].descriptor_magic = vkd3d_descriptor_magic_from_d3d12(range->RangeType);
              table->ranges[j].base_register_idx = range->BaseShaderRegister;

       table->ranges[j].register_space = range->RegisterSpace;
    }
}




@@ -683,15 +674,9 @@ static HRESULT d3d12_root_signature_init_root_descriptors(struct d3d12_root_sign
      root_signature->push_descriptor_mask |= 1u << i;



   if (p->u.Descriptor.RegisterSpace)


   {


       FIXME("Unhandled register space %u for parameter %u.\n", p->u.Descriptor.RegisterSpace, i);


       return E_NOTIMPL;


   }


    cur_binding->binding = d3d12_root_signature_assign_vk_bindings(root_signature,
            vkd3d_descriptor_type_from_d3d12_root_parameter_type(p->ParameterType),


           p->u.Descriptor.ShaderRegister, 1, true, false,




           p->u.Descriptor.RegisterSpace, p->u.Descriptor.ShaderRegister, 1, true, false,
            vkd3d_shader_visibility_from_d3d12(p->ShaderVisibility), context);
    cur_binding->descriptorType = vk_descriptor_type_from_d3d12_root_parameter(p->ParameterType);
    cur_binding->descriptorCount = 1;



@@ -728,7 +713,7 @@ static HRESULT d3d12_root_signature_init_static_samplers(struct d3d12_root_signa
              return hr;
      cur_binding->binding = d3d12_root_signature_assign_vk_bindings(root_signature,


           VKD3D_SHADER_DESCRIPTOR_TYPE_SAMPLER, s->ShaderRegister, 1, false, false,




           VKD3D_SHADER_DESCRIPTOR_TYPE_SAMPLER, s->RegisterSpace, s->ShaderRegister, 1, false, false,
            vkd3d_shader_visibility_from_d3d12(s->ShaderVisibility), context);
    cur_binding->descriptorType = VK_DESCRIPTOR_TYPE_SAMPLER;
    cur_binding->descriptorCount = 1;



@@ -1451,7 +1436,14 @@ static HRESULT d3d12_pipeline_state_init_compute_uav_counters(struct d3d12_pipel
          if (!(shader_info->uav_counter_mask & (1u << i)))
              continue;

   /* UAV counters will lookup Vulkan bindings based on the mask index directly.


    * We currently don't know the actual space/binding for this UAV,


    * but register_space/register_index are fixed up later after compilation is finished. */


   state->uav_counters[j].register_space = 0;
    state->uav_counters[j].register_index = i;



   state->uav_counters[j].counter_index = i;


    state->uav_counters[j].shader_visibility = VKD3D_SHADER_VISIBILITY_COMPUTE;
    state->uav_counters[j].binding.set = context.set_index;
    state->uav_counters[j].binding.binding = context.descriptor_binding;



@@ -1507,6 +1499,10 @@ static HRESULT d3d12_pipeline_state_init_compute(struct d3d12_pipeline_state *st
      struct vkd3d_shader_code dxbc;
      HRESULT hr;
      int ret;

unsigned int i, j;

unsigned int uav_counter_spaces[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS] = { 0 };

unsigned int uav_counter_bindings[VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS] = { 0 };

struct vkd3d_shader_effective_uav_counter_binding_info uav_binding_info = { VKD3D_SHADER_STRUCTURE_TYPE_EFFECTIVE_UAV_COUNTER_BINDING_INFO };
state->ID3D12PipelineState_iface.lpVtbl = &d3d12_pipeline_state_vtbl;
 state->refcount = 1;


@@ -1550,8 +1546,14 @@ static HRESULT d3d12_pipeline_state_init_compute(struct d3d12_pipeline_state *st
      shader_interface.uav_counters = state->uav_counters;
      shader_interface.uav_counter_count = vkd3d_popcount(state->uav_counter_mask);

shader_interface.next = &uav_binding_info;
uav_binding_info.uav_register_spaces = uav_counter_spaces;
uav_binding_info.uav_register_bindings = uav_counter_bindings;
uav_binding_info.uav_counter_count = VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS;
vk_pipeline_layout = state->vk_pipeline_layout
        ? state->vk_pipeline_layout : root_signature->vk_pipeline_layout;


if (FAILED(hr = vkd3d_create_compute_pipeline(device, &desc->CS, &shader_interface,
        vk_pipeline_layout, &state->u.compute.vk_pipeline)))
{



@@ -1575,6 +1577,17 @@ static HRESULT d3d12_pipeline_state_init_compute(struct d3d12_pipeline_state *st
          return hr;
      }

/* Map back to actual space/bindings for the UAV counter now that we know. */
for (i = 0, j = 0; i < VKD3D_SHADER_MAX_UNORDERED_ACCESS_VIEWS; i++)
{
   if (state->uav_counter_mask & (1u << i))


   {


       state->uav_counters[j].register_space = uav_counter_spaces[i];


       state->uav_counters[j].register_index = uav_counter_bindings[i];


       j++;


   }


}
state->vk_bind_point = VK_PIPELINE_BIND_POINT_COMPUTE;
d3d12_device_add_ref(state->device = device);




@@ -2911,6 +2924,7 @@ HRESULT vkd3d_uav_clear_state_init(struct vkd3d_uav_clear_state *state, struct d
  binding.type = VKD3D_SHADER_DESCRIPTOR_TYPE_UAV;
  binding.register_index = 0;


binding.register_space = 0;
 binding.shader_visibility = VKD3D_SHADER_VISIBILITY_COMPUTE;
 binding.binding.set = 0;
 binding.binding.binding = 0;

@@ -2919,6 +2933,7 @@ HRESULT vkd3d_uav_clear_state_init(struct vkd3d_uav_clear_state *state, struct d
      push_constant_range.offset = 0;
      push_constant_range.size = sizeof(struct vkd3d_uav_clear_args);

push_constant.register_space = 0;
 push_constant.register_index = 0;
 push_constant.shader_visibility = VKD3D_SHADER_VISIBILITY_COMPUTE;
 push_constant.offset = 0;

diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h
index 0c031d2..2309031 100644
--- a/libs/vkd3d/vkd3d_private.h
+++ b/libs/vkd3d/vkd3d_private.h
@@ -659,6 +659,7 @@ struct d3d12_root_descriptor_table_range
  uint32_t descriptor_magic;
  unsigned int base_register_idx;


unsigned int register_space;
};
struct d3d12_root_descriptor_table



    

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

2005

2004

2003

2002

2001

Re: [PATCH vkd3d 1/3] vkd3d: Deal correctly with SM 5.1 register spaces.

Signed-off-by: Hans-Kristian Arntzen post@arntzen-software.no