vkd3d-shader/hlsl: Implement the ByteAddressBuffer.Load*() methods.

Signed-off-by: Nikolay Sivov <nsivov@codeweavers.com>

Makefile.am

@@ -232,6 +232,7 @@ vkd3d_shader_tests = \
 	tests/hlsl/smoothstep.shader_test \
 	tests/hlsl/sqrt.shader_test \
 	tests/hlsl/srv-buffers.shader_test \
+	tests/hlsl/srv-byteaddressbuffer.shader_test \
 	tests/hlsl/state-block-function-syntax.shader_test \
 	tests/hlsl/state-block-syntax.shader_test \
 	tests/hlsl/static-initializer.shader_test \

libs/vkd3d-shader/hlsl.y

@@ -5603,6 +5603,55 @@ static bool raise_invalid_method_object_type(struct hlsl_ctx *ctx, const struct
     return false;
 }
 
+static bool add_raw_load_method_call(struct hlsl_ctx *ctx, struct hlsl_block *block, struct hlsl_ir_node *object,
+        const char *name, const struct parse_initializer *params, const struct vkd3d_shader_location *loc)
+{
+    struct hlsl_resource_load_params load_params = {.type = HLSL_RESOURCE_LOAD};
+    struct hlsl_ir_node *load;
+    unsigned int value_dim;
+
+    if (params->args_count != 1 && params->args_count != 2)
+    {
+        hlsl_error(ctx, loc, VKD3D_SHADER_ERROR_HLSL_WRONG_PARAMETER_COUNT,
+                "Wrong number of arguments to method 'Load': expected between 1 and 2, but got %u.",
+                params->args_count);
+        return false;
+    }
+
+    if (params->args_count == 2)
+    {
+        hlsl_fixme(ctx, loc, "Tiled resource status argument.");
+        return false;
+    }
+
+    if (params->args[0]->data_type->class != HLSL_CLASS_SCALAR)
+    {
+        hlsl_error(ctx, loc, VKD3D_SHADER_ERROR_HLSL_INVALID_TYPE, "Scalar address argument expected for '%s'.", name);
+        return false;
+    }
+
+    if (!strcmp(name, "Load"))
+        value_dim = 1;
+    else if (!strcmp(name, "Load2"))
+        value_dim = 2;
+    else if (!strcmp(name, "Load3"))
+        value_dim = 3;
+    else
+        value_dim = 4;
+
+    if (!(load_params.coords = add_implicit_conversion(ctx, block, params->args[0],
+            hlsl_get_scalar_type(ctx, HLSL_TYPE_UINT), loc)))
+        return false;
+
+    load_params.format = hlsl_get_vector_type(ctx, HLSL_TYPE_UINT, value_dim);
+    load_params.resource = object;
+
+    if (!(load = hlsl_new_resource_load(ctx, &load_params, loc)))
+        return false;
+    hlsl_block_add_instr(block, load);
+    return true;
+}
+
 static bool add_load_method_call(struct hlsl_ctx *ctx, struct hlsl_block *block, struct hlsl_ir_node *object,
         const char *name, const struct parse_initializer *params, const struct vkd3d_shader_location *loc)
 {
@@ -5612,6 +5661,9 @@ static bool add_load_method_call(struct hlsl_ctx *ctx, struct hlsl_block *block,
     struct hlsl_ir_node *load;
     bool multisampled;
 
+    if (object_type->sampler_dim == HLSL_SAMPLER_DIM_RAW_BUFFER)
+        return add_raw_load_method_call(ctx, block, object, name, params, loc);
+
     if (object_type->sampler_dim == HLSL_SAMPLER_DIM_STRUCTURED_BUFFER)
     {
         hlsl_fixme(ctx, loc, "Method '%s' for structured buffers.", name);
@@ -6264,7 +6316,10 @@ texture_methods[] =
 
     { "GetDimensions",      add_getdimensions_method_call, "00111111111110" },
 
-    { "Load",               add_load_method_call,          "00111011110110" },
+    { "Load",               add_load_method_call,          "00111011110111" },
+    { "Load2",              add_raw_load_method_call,      "00000000000001" },
+    { "Load3",              add_raw_load_method_call,      "00000000000001" },
+    { "Load4",              add_raw_load_method_call,      "00000000000001" },
 
     { "Sample",             add_sample_method_call,        "00111111001000" },
     { "SampleBias",         add_sample_lod_method_call,    "00111111001000" },

libs/vkd3d-shader/tpf.c

@@ -4831,7 +4831,15 @@ static void write_sm4_dcl_textures(const struct tpf_compiler *tpf, const struct
         }
         else
         {
+            switch (component_type->sampler_dim)
+            {
+                case HLSL_SAMPLER_DIM_RAW_BUFFER:
+                    instr.opcode = VKD3D_SM5_OP_DCL_RESOURCE_RAW;
+                    break;
+                default:
                     instr.opcode = VKD3D_SM4_OP_DCL_RESOURCE;
+                    break;
+            }
         }
         instr.extra_bits |= (sm4_resource_dimension(component_type) << VKD3D_SM4_RESOURCE_TYPE_SHIFT);
 
@@ -5133,12 +5141,15 @@ static void write_sm4_ld(const struct tpf_compiler *tpf, const struct hlsl_ir_no
             && (resource_type->sampler_dim == HLSL_SAMPLER_DIM_2DMS || resource_type->sampler_dim == HLSL_SAMPLER_DIM_2DMSARRAY);
     bool uav = (hlsl_deref_get_regset(tpf->ctx, resource) == HLSL_REGSET_UAVS);
     const struct vkd3d_shader_version *version = &tpf->program->shader_version;
+    bool raw = resource_type->sampler_dim == HLSL_SAMPLER_DIM_RAW_BUFFER;
     unsigned int coords_writemask = VKD3DSP_WRITEMASK_ALL;
     struct sm4_instruction instr;
 
     memset(&instr, 0, sizeof(instr));
     if (uav)
         instr.opcode = VKD3D_SM5_OP_LD_UAV_TYPED;
+    else if (raw)
+        instr.opcode = VKD3D_SM5_OP_LD_RAW;
     else
         instr.opcode = multisampled ? VKD3D_SM4_OP_LD2DMS : VKD3D_SM4_OP_LD;
 
@@ -6227,6 +6238,7 @@ static void tpf_write_shdr(struct tpf_compiler *tpf, struct hlsl_ir_function_dec
     const struct hlsl_buffer *cbuffer;
     struct hlsl_ctx *ctx = tpf->ctx;
     size_t token_count_position;
+    uint32_t global_flags = 0;
 
     static const uint16_t shader_types[VKD3D_SHADER_TYPE_COUNT] =
     {
@@ -6248,6 +6260,27 @@ static void tpf_write_shdr(struct tpf_compiler *tpf, struct hlsl_ir_function_dec
     put_u32(&buffer, vkd3d_make_u32((version->major << 4) | version->minor, shader_types[version->type]));
     token_count_position = put_u32(&buffer, 0);
 
+    if (version->major == 4)
+    {
+        for (i = 0; i < extern_resources_count; ++i)
+        {
+            const struct extern_resource *resource = &extern_resources[i];
+            const struct hlsl_type *type = resource->component_type;
+
+            if (type && type->class == HLSL_CLASS_TEXTURE && type->sampler_dim == HLSL_SAMPLER_DIM_RAW_BUFFER)
+            {
+                global_flags |= VKD3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
+                break;
+            }
+        }
+    }
+
+    if (entry_func->early_depth_test && vkd3d_shader_ver_ge(version, 5, 0))
+        global_flags |= VKD3DSGF_FORCE_EARLY_DEPTH_STENCIL;
+
+    if (global_flags)
+        write_sm4_dcl_global_flags(tpf, global_flags);
+
     if (version->type == VKD3D_SHADER_TYPE_HULL)
     {
         tpf_write_hs_decls(tpf);
@@ -6282,9 +6315,6 @@ static void tpf_write_shdr(struct tpf_compiler *tpf, struct hlsl_ir_function_dec
             write_sm4_dcl_textures(tpf, resource, true);
     }
 
-    if (entry_func->early_depth_test && vkd3d_shader_ver_ge(version, 5, 0))
-        write_sm4_dcl_global_flags(tpf, VKD3DSGF_FORCE_EARLY_DEPTH_STENCIL);
-
     if (version->type == VKD3D_SHADER_TYPE_HULL)
         tpf_write_hs_control_point_phase(tpf);
 

tests/hlsl/srv-byteaddressbuffer.shader_test

@@ -0,0 +1,56 @@
+[require]
+shader model >= 4.0
+
+[rtv 0]
+format r32g32b32a32-uint
+size (2d, 32, 32)
+
+[srv 0]
+format r32-typeless
+size (raw_buffer, 4)
+
+1 2 3 4
+
+[pixel shader]
+ByteAddressBuffer t;
+uint offset;
+uint count;
+
+uint4 main() : sv_target
+{
+    if (count == 1)
+        return uint4(t.Load(offset), 0xa, 0xb, 0xc);
+    if (count == 2)
+        return uint4(t.Load2(offset), 0xd, 0xe);
+    if (count == 3)
+        return uint4(t.Load3(offset), 0xf);
+    if (count == 4)
+        return t.Load4(offset);
+    return 0;
+}
+
+[test]
+uniform 0 uint4 0 1 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x1, 0xa, 0xb, 0xc)
+uniform 0 uint4 4 1 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x2, 0xa, 0xb, 0xc)
+uniform 0 uint4 0 2 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x1, 0x2, 0xd, 0xe)
+uniform 0 uint4 4 2 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x2, 0x3, 0xd, 0xe)
+uniform 0 uint4 0 3 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x1, 0x2, 0x3, 0xf)
+uniform 0 uint4 4 3 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x2, 0x3, 0x4, 0xf)
+uniform 0 uint4 0 4 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x1, 0x2, 0x3, 0x4)
+uniform 0 uint4 4 4 0 0
+todo(msl | glsl) draw quad
+probe (0, 0) rgbaui(0x2, 0x3, 0x4, 0x0)

tests/shader_runner_d3d11.c

@@ -478,9 +478,19 @@ static void init_resource_srv_buffer(struct d3d11_shader_runner *runner, struct
     resource->resource = (ID3D11Resource *)resource->buffer;
 
     srv_desc.Format = params->desc.format;
+    if (params->is_raw)
+    {
+        srv_desc.ViewDimension = D3D11_SRV_DIMENSION_BUFFEREX;
+        srv_desc.BufferEx.FirstElement = 0;
+        srv_desc.BufferEx.NumElements = params->data_size / params->desc.texel_size;
+        srv_desc.BufferEx.Flags = D3D11_BUFFEREX_SRV_FLAG_RAW;
+    }
+    else
+    {
         srv_desc.ViewDimension = D3D11_SRV_DIMENSION_BUFFER;
         srv_desc.Buffer.FirstElement = 0;
         srv_desc.Buffer.NumElements = params->data_size / params->desc.texel_size;
+    }
     hr = ID3D11Device_CreateShaderResourceView(device, resource->resource, &srv_desc, &resource->srv);
     ok(hr == S_OK, "Failed to create view, hr %#lx.\n", hr);
 }

tests/shader_runner_d3d12.c

@@ -159,6 +159,7 @@ static struct resource *d3d12_runner_create_resource(struct shader_runner *r, co
                 srv_desc.ViewDimension = D3D12_SRV_DIMENSION_BUFFER;
                 srv_desc.Shader4ComponentMapping = D3D12_DEFAULT_SHADER_4_COMPONENT_MAPPING;
                 srv_desc.Buffer.NumElements = params->desc.width * params->desc.height;
+                srv_desc.Buffer.Flags = params->is_raw ? D3D12_BUFFER_SRV_FLAG_RAW : 0;
 
                 ID3D12Device_CreateShaderResourceView(device, resource->resource,
                         &srv_desc, get_cpu_descriptor_handle(test_context, runner->heap, resource->r.desc.slot));