Conor McCarthy : vkd3d-shader/spirv: Introduce HALF and UINT16 types for minimum precision.

[Alexandre Julliard]

Module: vkd3d
Branch: master
Commit: 066ea75945a570616ba072e834af170a09d19f37
URL:    https://gitlab.winehq.org/wine/vkd3d/-/commit/066ea75945a570616ba072e834af17...

Author: Conor McCarthy <cmccarthy(a)codeweavers.com>
Date:   Wed Dec 13 15:40:48 2023 +1000

vkd3d-shader/spirv: Introduce HALF and UINT16 types for minimum precision.

Minimum precision types must always be implemented as 32-bit to match how
reduced precision works in SPIR-V.

---

 libs/vkd3d-shader/d3d_asm.c              |  2 +
 libs/vkd3d-shader/spirv.c                | 84 ++++++++++++++++++++++++++++++--
 libs/vkd3d-shader/vkd3d_shader_private.h |  8 ++-
 3 files changed, 87 insertions(+), 7 deletions(-)

diff --git a/libs/vkd3d-shader/d3d_asm.c b/libs/vkd3d-shader/d3d_asm.c
index 2aabd5b1..7b30d260 100644
--- a/libs/vkd3d-shader/d3d_asm.c
+++ b/libs/vkd3d-shader/d3d_asm.c
@@ -649,6 +649,8 @@ static void shader_dump_data_type(struct vkd3d_d3d_asm_compiler *compiler, enum
         [VKD3D_DATA_UINT8    ] = "uint8",
         [VKD3D_DATA_UINT64   ] = "uint64",
         [VKD3D_DATA_BOOL     ] = "bool",
+        [VKD3D_DATA_UINT16   ] = "uint16",
+        [VKD3D_DATA_HALF     ] = "half",
     };
 
     const char *name;
diff --git a/libs/vkd3d-shader/spirv.c b/libs/vkd3d-shader/spirv.c
index 329b24ba..56c89f76 100644
--- a/libs/vkd3d-shader/spirv.c
+++ b/libs/vkd3d-shader/spirv.c
@@ -225,7 +225,7 @@ enum vkd3d_shader_input_sysval_semantic vkd3d_siv_from_sysval_indexed(enum vkd3d
 
 static bool data_type_is_floating_point(enum vkd3d_data_type data_type)
 {
-    return data_type == VKD3D_DATA_FLOAT || data_type == VKD3D_DATA_DOUBLE;
+    return data_type == VKD3D_DATA_HALF || data_type == VKD3D_DATA_FLOAT || data_type == VKD3D_DATA_DOUBLE;
 }
 
 #define VKD3D_SPIRV_VERSION 0x00010000
@@ -1843,6 +1843,7 @@ static uint32_t vkd3d_spirv_get_type_id_for_data_type(struct vkd3d_spirv_builder
     {
         switch (data_type)
         {
+            case VKD3D_DATA_HALF: /* Minimum precision. TODO: native 16-bit */
             case VKD3D_DATA_FLOAT:
             case VKD3D_DATA_SNORM:
             case VKD3D_DATA_UNORM:
@@ -1850,6 +1851,7 @@ static uint32_t vkd3d_spirv_get_type_id_for_data_type(struct vkd3d_spirv_builder
                 break;
             case VKD3D_DATA_INT:
             case VKD3D_DATA_UINT:
+            case VKD3D_DATA_UINT16: /* Minimum precision. TODO: native 16-bit */
                 return vkd3d_spirv_get_op_type_int(builder, 32, data_type == VKD3D_DATA_INT);
                 break;
             case VKD3D_DATA_DOUBLE:
@@ -3758,6 +3760,70 @@ static uint32_t spirv_compiler_emit_bool_to_double(struct spirv_compiler *compil
     return vkd3d_spirv_build_op_select(builder, type_id, val_id, true_id, false_id);
 }
 
+/* Based on the implementation in the OpenGL Mathematics library. */
+static uint32_t half_to_float(uint16_t value)
+{
+    uint32_t s = (value & 0x8000u) << 16;
+    uint32_t e = (value >> 10) & 0x1fu;
+    uint32_t m = value & 0x3ffu;
+
+    if (!e)
+    {
+        if (!m)
+        {
+            /* Plus or minus zero */
+            return s;
+        }
+        else
+        {
+            /* Denormalized number -- renormalize it */
+
+            while (!(m & 0x400u))
+            {
+                m <<= 1;
+                --e;
+            }
+
+            ++e;
+            m &= ~0x400u;
+        }
+    }
+    else if (e == 31u)
+    {
+        /* Positive or negative infinity for zero 'm'.
+         * Nan for non-zero 'm' -- preserve sign and significand bits */
+        return s | 0x7f800000u | (m << 13);
+    }
+
+    /* Normalized number */
+    e += 127u - 15u;
+    m <<= 13;
+
+    /* Assemble s, e and m. */
+    return s | (e << 23) | m;
+}
+
+static uint32_t convert_raw_constant32(enum vkd3d_data_type data_type, unsigned int uint_value)
+{
+    int16_t i;
+
+    /* TODO: native 16-bit support. */
+    if (data_type != VKD3D_DATA_UINT16 && data_type != VKD3D_DATA_HALF)
+        return uint_value;
+
+    if (data_type == VKD3D_DATA_HALF)
+        return half_to_float(uint_value);
+
+    /* Values in DXIL have no signedness, so it is ambiguous whether 16-bit constants should or
+     * should not be sign-extended when 16-bit execution is not supported. The AMD RX 580 Windows
+     * driver has no 16-bit support, and sign-extends all 16-bit constant ints to 32 bits. These
+     * results differ from SM 5. The RX 6750 XT supports 16-bit execution, so constants are not
+     * extended, and results match SM 5. It seems best to replicate the sign-extension, and if
+     * execution is 16-bit, the values will be truncated. */
+    i = uint_value;
+    return (int32_t)i;
+}
+
 static uint32_t spirv_compiler_emit_load_constant(struct spirv_compiler *compiler,
         const struct vkd3d_shader_register *reg, uint32_t swizzle, uint32_t write_mask)
 {
@@ -3770,14 +3836,15 @@ static uint32_t spirv_compiler_emit_load_constant(struct spirv_compiler *compile
     if (reg->dimension == VSIR_DIMENSION_SCALAR)
     {
         for (i = 0; i < component_count; ++i)
-            values[i] = *reg->u.immconst_u32;
+            values[i] = convert_raw_constant32(reg->data_type, reg->u.immconst_u32[0]);
     }
     else
     {
         for (i = 0, j = 0; i < VKD3D_VEC4_SIZE; ++i)
         {
             if (write_mask & (VKD3DSP_WRITEMASK_0 << i))
-                values[j++] = reg->u.immconst_u32[vsir_swizzle_get_component(swizzle, i)];
+                values[j++] = convert_raw_constant32(reg->data_type,
+                        reg->u.immconst_u32[vsir_swizzle_get_component(swizzle, i)]);
         }
     }
 
@@ -3921,6 +3988,13 @@ static uint32_t spirv_compiler_emit_constant_array(struct spirv_compiler *compil
 
     switch (icb->data_type)
     {
+        case VKD3D_DATA_HALF:
+        case VKD3D_DATA_UINT16:
+            /* Scalar only. */
+            for (i = 0; i < element_count; ++i)
+                elements[i] = vkd3d_spirv_get_op_constant(builder, elem_type_id,
+                        convert_raw_constant32(icb->data_type, icb->data[i]));
+            break;
         case VKD3D_DATA_FLOAT:
         case VKD3D_DATA_INT:
         case VKD3D_DATA_UINT:
@@ -6908,7 +6982,7 @@ static void spirv_compiler_emit_bool_cast(struct spirv_compiler *compiler,
     assert(src->reg.data_type == VKD3D_DATA_BOOL && dst->reg.data_type != VKD3D_DATA_BOOL);
 
     val_id = spirv_compiler_emit_load_src(compiler, src, dst->write_mask);
-    if (dst->reg.data_type == VKD3D_DATA_FLOAT)
+    if (dst->reg.data_type == VKD3D_DATA_HALF || dst->reg.data_type == VKD3D_DATA_FLOAT)
     {
         val_id = spirv_compiler_emit_bool_to_float(compiler, 1, val_id, instruction->handler_idx == VKD3DSIH_ITOF);
     }
@@ -6917,7 +6991,7 @@ static void spirv_compiler_emit_bool_cast(struct spirv_compiler *compiler,
         /* ITOD is not supported. Frontends which emit bool casts must use ITOF for double. */
         val_id = spirv_compiler_emit_bool_to_double(compiler, 1, val_id, instruction->handler_idx == VKD3DSIH_ITOF);
     }
-    else if (dst->reg.data_type == VKD3D_DATA_UINT)
+    else if (dst->reg.data_type == VKD3D_DATA_UINT16 || dst->reg.data_type == VKD3D_DATA_UINT)
     {
         val_id = spirv_compiler_emit_bool_to_int(compiler, 1, val_id, instruction->handler_idx == VKD3DSIH_ITOI);
     }
diff --git a/libs/vkd3d-shader/vkd3d_shader_private.h b/libs/vkd3d-shader/vkd3d_shader_private.h
index a666544c..d6f9d016 100644
--- a/libs/vkd3d-shader/vkd3d_shader_private.h
+++ b/libs/vkd3d-shader/vkd3d_shader_private.h
@@ -623,14 +623,16 @@ enum vkd3d_data_type
     VKD3D_DATA_UINT8,
     VKD3D_DATA_UINT64,
     VKD3D_DATA_BOOL,
+    VKD3D_DATA_UINT16,
+    VKD3D_DATA_HALF,
 
     VKD3D_DATA_COUNT,
 };
 
 static inline bool data_type_is_integer(enum vkd3d_data_type data_type)
 {
-    return data_type == VKD3D_DATA_INT || data_type == VKD3D_DATA_UINT8 || data_type == VKD3D_DATA_UINT
-            || data_type == VKD3D_DATA_UINT64;
+    return data_type == VKD3D_DATA_INT || data_type == VKD3D_DATA_UINT8 || data_type == VKD3D_DATA_UINT16
+            || data_type == VKD3D_DATA_UINT || data_type == VKD3D_DATA_UINT64;
 }
 
 static inline bool data_type_is_bool(enum vkd3d_data_type data_type)
@@ -1531,10 +1533,12 @@ static inline enum vkd3d_shader_component_type vkd3d_component_type_from_data_ty
 {
     switch (data_type)
     {
+        case VKD3D_DATA_HALF: /* Minimum precision. TODO: native 16-bit */
         case VKD3D_DATA_FLOAT:
         case VKD3D_DATA_UNORM:
         case VKD3D_DATA_SNORM:
             return VKD3D_SHADER_COMPONENT_FLOAT;
+        case VKD3D_DATA_UINT16: /* Minimum precision. TODO: native 16-bit */
         case VKD3D_DATA_UINT:
             return VKD3D_SHADER_COMPONENT_UINT;
         case VKD3D_DATA_INT: