d3d12: Add d3d12mipmapping element

Adding a new element for texture conversion from single mip level
texture to mipmapping enabled RGBA texture

Part-of: <https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/7555>

11 files changed, 1965 insertions, 0 deletions

diff --git a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.cpp b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.cpp
index 73fffc48a4..ff82947f96 100644
--- a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.cpp
+++ b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.cpp
@@ -41,12 +41,14 @@ using namespace Microsoft::WRL;
 #include "converter_hlsl_cs.h"
 #include "plugin_hlsl_ps.h"
 #include "plugin_hlsl_vs.h"
+#include "plugin_hlsl_cs.h"
 #else
 static std::unordered_map<std::string, std::pair<const BYTE *, SIZE_T>> g_converter_ps_table;
 static std::unordered_map<std::string, std::pair<const BYTE *, SIZE_T>> g_converter_vs_table;
 static std::unordered_map<std::string, std::pair<const BYTE *, SIZE_T>> g_converter_cs_table;
 static std::unordered_map<std::string, std::pair<const BYTE *, SIZE_T>> g_plugin_ps_table;
 static std::unordered_map<std::string, std::pair<const BYTE *, SIZE_T>> g_plugin_vs_table;
+static std::unordered_map<std::string, std::pair<const BYTE *, SIZE_T>> g_plugin_cs_table;
 #endif
 
 static std::vector<std::pair<std::string, ID3DBlob *>> g_compiled_blobs;
@@ -81,6 +83,10 @@ static const ShaderItem g_vs_map[] = {
   {GST_D3D_PLUGIN_VS_POS, BUILD_SOURCE (VSMain_pos)},
 };
 
+static const ShaderItem g_cs_map[] = {
+  {GST_D3D_PLUGIN_CS_MIP_GEN, BUILD_SOURCE (CSMain_mipgen)},
+};
+
 #undef BUILD_SOURCE
 
 static const gchar * g_sm_map[] = {
@@ -196,6 +202,60 @@ gst_d3d_plugin_shader_get_ps_blob (GstD3DPluginPS type,
 }
 
 gboolean
+gst_d3d_plugin_shader_get_cs_blob (GstD3DPluginCS type,
+    GstD3DShaderModel shader_model, GstD3DShaderByteCode * byte_code)
+{
+  g_return_val_if_fail (type < GST_D3D_PLUGIN_CS_LAST, FALSE);
+  g_return_val_if_fail (shader_model < GST_D3D_SM_LAST, FALSE);
+  g_return_val_if_fail (byte_code, FALSE);
+
+  static std::mutex cache_lock;
+
+  auto shader_name = std::string (g_cs_map[type].name) + "_" +
+      std::string (g_sm_map[shader_model]);
+
+  std::lock_guard <std::mutex> lk (cache_lock);
+  auto it = g_plugin_cs_table.find (shader_name);
+  if (it != g_plugin_cs_table.end ()) {
+    byte_code->byte_code = it->second.first;
+    byte_code->byte_code_len = it->second.second;
+
+    return TRUE;
+  }
+
+  auto target = std::string ("cs_") + g_sm_map[shader_model];
+
+  ID3DBlob *blob = nullptr;
+  ComPtr<ID3DBlob> error_msg;
+
+  auto hr = gst_d3d_compile (g_cs_map[type].source, g_cs_map[type].source_size,
+      nullptr, nullptr, nullptr, "ENTRY_POINT", target.c_str (), 0, 0,
+      &blob, &error_msg);
+  if (FAILED (hr)) {
+    const gchar *err = nullptr;
+    if (error_msg)
+      err = (const gchar *) error_msg->GetBufferPointer ();
+
+    GST_ERROR ("Couldn't compile code, hr: 0x%x, error detail: %s, "
+        "source code: \n%s", (guint) hr, GST_STR_NULL (err),
+        g_cs_map[type].source);
+    return FALSE;
+  }
+
+  byte_code->byte_code = blob->GetBufferPointer ();
+  byte_code->byte_code_len = blob->GetBufferSize ();
+
+  g_plugin_cs_table[shader_name] = { (const BYTE *) blob->GetBufferPointer (),
+      blob->GetBufferSize ()};
+
+  std::lock_guard <std::mutex> blk (g_blob_lock);
+  g_compiled_blobs.push_back ({ shader_name, blob });
+
+  return TRUE;
+}
+
+
+gboolean
 gst_d3d_converter_shader_get_vs_blob (GstD3DShaderModel shader_model,
     GstD3DShaderByteCode * byte_code)
 {
diff --git a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.h b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.h
index e6958e41eb..1b7c474ec4 100644
--- a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.h
+++ b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/gstd3dshadercache.h
@@ -49,6 +49,12 @@ typedef enum
   GST_D3D_PLUGIN_VS_LAST,
 } GstD3DPluginVS;
 
+typedef enum
+{
+  GST_D3D_PLUGIN_CS_MIP_GEN,
+
+  GST_D3D_PLUGIN_CS_LAST,
+} GstD3DPluginCS;
 
 typedef enum
 {
@@ -100,6 +106,11 @@ gboolean gst_d3d_plugin_shader_get_ps_blob (GstD3DPluginPS type,
                                             GstD3DShaderByteCode * byte_code);
 
 GST_D3D_SHADER_API
+gboolean gst_d3d_plugin_shader_get_cs_blob (GstD3DPluginCS type,
+                                            GstD3DShaderModel shader_model,
+                                            GstD3DShaderByteCode * byte_code);
+
+GST_D3D_SHADER_API
 gboolean gst_d3d_converter_shader_get_vs_blob (GstD3DShaderModel shader_model,
                                                GstD3DShaderByteCode * byte_code);
 
diff --git a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/CSMain_mipgen.hlsl b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/CSMain_mipgen.hlsl
new file mode 100644
index 0000000000..11734ce281
--- /dev/null
+++ b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/CSMain_mipgen.hlsl
@@ -0,0 +1,421 @@
+/**
+ * MIT License
+ *
+ * Copyright (c) 2018 Jeremiah van Oosten
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+/* Source: https://github.com/jpvanoosten/LearningDirectX12 */
+
+#ifdef BUILDING_HLSL
+
+#define BLOCK_SIZE 8
+
+ // When reducing the size of a texture, it could be that downscaling the texture
+ // will result in a less than exactly 50% (1/2) of the original texture size.
+ // This happens if either the width, or the height (or both) dimensions of the texture
+ // are odd. For example, downscaling a 5x3 texture will result in a 2x1 texture which
+ // has a 60% reduction in the texture width and 66% reduction in the height.
+ // When this happens, we need to take more samples from the source texture to
+ // determine the pixel value in the destination texture.
+
+#define WIDTH_HEIGHT_EVEN 0     // Both the width and the height of the texture are even.
+#define WIDTH_ODD_HEIGHT_EVEN 1 // The texture width is odd and the height is even.
+#define WIDTH_EVEN_HEIGHT_ODD 2 // The texture width is even and teh height is odd.
+#define WIDTH_HEIGHT_ODD 3      // Both the width and height of the texture are odd.
+
+struct ComputeShaderInput
+{
+    uint3 GroupID           : SV_GroupID;           // 3D index of the thread group in the dispatch.
+    uint3 GroupThreadID     : SV_GroupThreadID;     // 3D index of local thread ID in a thread group.
+    uint3 DispatchThreadID  : SV_DispatchThreadID;  // 3D index of global thread ID in the dispatch.
+    uint  GroupIndex        : SV_GroupIndex;        // Flattened local index of the thread within a thread group.
+};
+
+cbuffer GenerateMipsCB : register( b0 )
+{
+    uint SrcMipLevel;   // Texture level of source mip
+    uint NumMipLevels;  // Number of OutMips to write: [1-4]
+    uint SrcDimension;  // Width and height of the source texture are even or odd.
+    uint padding;
+    float2 TexelSize;   // 1.0 / OutMip1.Dimensions
+}
+
+// Source mip map.
+Texture2D<float4> SrcMip : register( t0 );
+
+// Write up to 4 mip map levels.
+RWTexture2D<float4> OutMip1 : register( u0 );
+RWTexture2D<float4> OutMip2 : register( u1 );
+RWTexture2D<float4> OutMip3 : register( u2 );
+RWTexture2D<float4> OutMip4 : register( u3 );
+
+// Linear clamp sampler.
+SamplerState LinearClampSampler : register( s0 );
+
+// The reason for separating channels is to reduce bank conflicts in the
+// local data memory controller.  A large stride will cause more threads
+// to collide on the same memory bank.
+groupshared float gs_R[64];
+groupshared float gs_G[64];
+groupshared float gs_B[64];
+groupshared float gs_A[64];
+
+void StoreColor( uint Index, float4 Color )
+{
+    gs_R[Index] = Color.r;
+    gs_G[Index] = Color.g;
+    gs_B[Index] = Color.b;
+    gs_A[Index] = Color.a;
+}
+
+float4 LoadColor( uint Index )
+{
+    return float4( gs_R[Index], gs_G[Index], gs_B[Index], gs_A[Index] );
+}
+
+[numthreads( BLOCK_SIZE, BLOCK_SIZE, 1 )]
+void ENTRY_POINT( ComputeShaderInput IN )
+{
+    float4 Src1 = (float4)0;
+
+    // One bilinear sample is insufficient when scaling down by more than 2x.
+    // You will slightly undersample in the case where the source dimension
+    // is odd.  This is why it's a really good idea to only generate mips on
+    // power-of-two sized textures.  Trying to handle the undersampling case
+    // will force this shader to be slower and more complicated as it will
+    // have to take more source texture samples.
+
+    // Determine the path to use based on the dimension of the
+    // source texture.
+    // 0b00(0): Both width and height are even.
+    // 0b01(1): Width is odd, height is even.
+    // 0b10(2): Width is even, height is odd.
+    // 0b11(3): Both width and height are odd.
+    switch ( SrcDimension )
+    {
+        case WIDTH_HEIGHT_EVEN:
+        {
+            float2 UV = TexelSize * ( IN.DispatchThreadID.xy + 0.5 );
+
+            Src1 = SrcMip.SampleLevel( LinearClampSampler, UV, SrcMipLevel );
+        }
+        break;
+        case WIDTH_ODD_HEIGHT_EVEN:
+        {
+            // > 2:1 in X dimension
+            // Use 2 bilinear samples to guarantee we don't undersample when downsizing by more than 2x
+            // horizontally.
+            float2 UV1 = TexelSize * ( IN.DispatchThreadID.xy + float2( 0.25, 0.5 ) );
+            float2 Off = TexelSize * float2( 0.5, 0.0 );
+
+            Src1 = 0.5 * ( SrcMip.SampleLevel( LinearClampSampler, UV1, SrcMipLevel ) +
+                           SrcMip.SampleLevel( LinearClampSampler, UV1 + Off, SrcMipLevel ) );
+        }
+        break;
+        case WIDTH_EVEN_HEIGHT_ODD:
+        {
+            // > 2:1 in Y dimension
+            // Use 2 bilinear samples to guarantee we don't undersample when downsizing by more than 2x
+            // vertically.
+            float2 UV1 = TexelSize * ( IN.DispatchThreadID.xy + float2( 0.5, 0.25 ) );
+            float2 Off = TexelSize * float2( 0.0, 0.5 );
+
+            Src1 = 0.5 * ( SrcMip.SampleLevel( LinearClampSampler, UV1, SrcMipLevel ) +
+                           SrcMip.SampleLevel( LinearClampSampler, UV1 + Off, SrcMipLevel ) );
+        }
+        break;
+        case WIDTH_HEIGHT_ODD:
+        {
+            // > 2:1 in in both dimensions
+            // Use 4 bilinear samples to guarantee we don't undersample when downsizing by more than 2x
+            // in both directions.
+            float2 UV1 = TexelSize * ( IN.DispatchThreadID.xy + float2( 0.25, 0.25 ) );
+            float2 Off = TexelSize * 0.5;
+
+            Src1 =  SrcMip.SampleLevel( LinearClampSampler, UV1, SrcMipLevel );
+            Src1 += SrcMip.SampleLevel( LinearClampSampler, UV1 + float2( Off.x, 0.0   ), SrcMipLevel );
+            Src1 += SrcMip.SampleLevel( LinearClampSampler, UV1 + float2( 0.0,   Off.y ), SrcMipLevel );
+            Src1 += SrcMip.SampleLevel( LinearClampSampler, UV1 + float2( Off.x, Off.y ), SrcMipLevel );
+            Src1 *= 0.25;
+        }
+        break;
+    }
+
+    OutMip1[IN.DispatchThreadID.xy] = Src1;
+
+    // A scalar (constant) branch can exit all threads coherently.
+    if ( NumMipLevels == 1 )
+        return;
+
+    // Without lane swizzle operations, the only way to share data with other
+    // threads is through LDS.
+    StoreColor( IN.GroupIndex, Src1 );
+
+    // This guarantees all LDS writes are complete and that all threads have
+    // executed all instructions so far (and therefore have issued their LDS
+    // write instructions.)
+    GroupMemoryBarrierWithGroupSync();
+
+    // With low three bits for X and high three bits for Y, this bit mask
+    // (binary: 001001) checks that X and Y are even.
+    if ( ( IN.GroupIndex & 0x9 ) == 0 )
+    {
+        float4 Src2 = LoadColor( IN.GroupIndex + 0x01 );
+        float4 Src3 = LoadColor( IN.GroupIndex + 0x08 );
+        float4 Src4 = LoadColor( IN.GroupIndex + 0x09 );
+        Src1 = 0.25 * ( Src1 + Src2 + Src3 + Src4 );
+
+        OutMip2[IN.DispatchThreadID.xy / 2] = Src1;
+        StoreColor( IN.GroupIndex, Src1 );
+    }
+
+    if ( NumMipLevels == 2 )
+        return;
+
+    GroupMemoryBarrierWithGroupSync();
+
+    // This bit mask (binary: 011011) checks that X and Y are multiples of four.
+    if ( ( IN.GroupIndex & 0x1B ) == 0 )
+    {
+        float4 Src2 = LoadColor( IN.GroupIndex + 0x02 );
+        float4 Src3 = LoadColor( IN.GroupIndex + 0x10 );
+        float4 Src4 = LoadColor( IN.GroupIndex + 0x12 );
+        Src1 = 0.25 * ( Src1 + Src2 + Src3 + Src4 );
+
+        OutMip3[IN.DispatchThreadID.xy / 4] = Src1;
+        StoreColor( IN.GroupIndex, Src1 );
+    }
+
+    if ( NumMipLevels == 3 )
+        return;
+
+    GroupMemoryBarrierWithGroupSync();
+
+    // This bit mask would be 111111 (X & Y multiples of 8), but only one
+    // thread fits that criteria.
+    if ( IN.GroupIndex == 0 )
+    {
+        float4 Src2 = LoadColor( IN.GroupIndex + 0x04 );
+        float4 Src3 = LoadColor( IN.GroupIndex + 0x20 );
+        float4 Src4 = LoadColor( IN.GroupIndex + 0x24 );
+        Src1 = 0.25 * ( Src1 + Src2 + Src3 + Src4 );
+
+        OutMip4[IN.DispatchThreadID.xy / 8] = Src1;
+    }
+}
+#else
+static const char str_CSMain_mipgen[] =
+"#define BLOCK_SIZE 8\n"
+"\n"
+" // When reducing the size of a texture, it could be that downscaling the texture\n"
+" // will result in a less than exactly 50% (1/2) of the original texture size.\n"
+" // This happens if either the width, or the height (or both) dimensions of the texture\n"
+" // are odd. For example, downscaling a 5x3 texture will result in a 2x1 texture which\n"
+" // has a 60% reduction in the texture width and 66% reduction in the height.\n"
+" // When this happens, we need to take more samples from the source texture to\n"
+" // determine the pixel value in the destination texture.\n"
+"\n"
+"#define WIDTH_HEIGHT_EVEN 0     // Both the width and the height of the texture are even.\n"
+"#define WIDTH_ODD_HEIGHT_EVEN 1 // The texture width is odd and the height is even.\n"
+"#define WIDTH_EVEN_HEIGHT_ODD 2 // The texture width is even and teh height is odd.\n"
+"#define WIDTH_HEIGHT_ODD 3      // Both the width and height of the texture are odd.\n"
+"\n"
+"struct ComputeShaderInput\n"
+"{\n"
+"    uint3 GroupID           : SV_GroupID;           // 3D index of the thread group in the dispatch.\n"
+"    uint3 GroupThreadID     : SV_GroupThreadID;     // 3D index of local thread ID in a thread group.\n"
+"    uint3 DispatchThreadID  : SV_DispatchThreadID;  // 3D index of global thread ID in the dispatch.\n"
+"    uint  GroupIndex        : SV_GroupIndex;        // Flattened local index of the thread within a thread group.\n"
+"};\n"
+"\n"
+"cbuffer GenerateMipsCB : register( b0 )\n"
+"{\n"
+"    uint SrcMipLevel;   // Texture level of source mip\n"
+"    uint NumMipLevels;  // Number of OutMips to write: [1-4]\n"
+"    uint SrcDimension;  // Width and height of the source texture are even or odd.\n"
+"    uint padding;\n"
+"    float2 TexelSize;   // 1.0 / OutMip1.Dimensions\n"
+"}\n"
+"\n"
+"// Source mip map.\n"
+"Texture2D<float4> SrcMip : register( t0 );\n"
+"\n"
+"// Write up to 4 mip map levels.\n"
+"RWTexture2D<float4> OutMip1 : register( u0 );\n"
+"RWTexture2D<float4> OutMip2 : register( u1 );\n"
+"RWTexture2D<float4> OutMip3 : register( u2 );\n"
+"RWTexture2D<float4> OutMip4 : register( u3 );\n"
+"\n"
+"// Linear clamp sampler.\n"
+"SamplerState LinearClampSampler : register( s0 );\n"
+"\n"
+"// The reason for separating channels is to reduce bank conflicts in the\n"
+"// local data memory controller.  A large stride will cause more threads\n"
+"// to collide on the same memory bank.\n"
+"groupshared float gs_R[64];\n"
+"groupshared float gs_G[64];\n"
+"groupshared float gs_B[64];\n"
+"groupshared float gs_A[64];\n"
+"\n"
+"void StoreColor( uint Index, float4 Color )\n"
+"{\n"
+"    gs_R[Index] = Color.r;\n"
+"    gs_G[Index] = Color.g;\n"
+"    gs_B[Index] = Color.b;\n"
+"    gs_A[Index] = Color.a;\n"
+"}\n"
+"\n"
+"float4 LoadColor( uint Index )\n"
+"{\n"
+"    return float4( gs_R[Index], gs_G[Index], gs_B[Index], gs_A[Index] );\n"
+"}\n"
+"\n"
+"[numthreads( BLOCK_SIZE, BLOCK_SIZE, 1 )]\n"
+"void ENTRY_POINT( ComputeShaderInput IN )\n"
+"{\n"
+"    float4 Src1 = (float4)0;\n"
+"\n"
+"    // One bilinear sample is insufficient when scaling down by more than 2x.\n"
+"    // You will slightly undersample in the case where the source dimension\n"
+"    // is odd.  This is why it's a really good idea to only generate mips on\n"
+"    // power-of-two sized textures.  Trying to handle the undersampling case\n"
+"    // will force this shader to be slower and more complicated as it will\n"
+"    // have to take more source texture samples.\n"
+"\n"
+"    // Determine the path to use based on the dimension of the\n"
+"    // source texture.\n"
+"    // 0b00(0): Both width and height are even.\n"
+"    // 0b01(1): Width is odd, height is even.\n"
+"    // 0b10(2): Width is even, height is odd.\n"
+"    // 0b11(3): Both width and height are odd.\n"
+"    switch ( SrcDimension )\n"
+"    {\n"
+"        case WIDTH_HEIGHT_EVEN:\n"
+"        {\n"
+"            float2 UV = TexelSize * ( IN.DispatchThreadID.xy + 0.5 );\n"
+"\n"
+"            Src1 = SrcMip.SampleLevel( LinearClampSampler, UV, SrcMipLevel );\n"
+"        }\n"
+"        break;\n"
+"        case WIDTH_ODD_HEIGHT_EVEN:\n"
+"        {\n"
+"            // > 2:1 in X dimension\n"
+"            // Use 2 bilinear samples to guarantee we don't undersample when downsizing by more than 2x\n"
+"            // horizontally.\n"
+"            float2 UV1 = TexelSize * ( IN.DispatchThreadID.xy + float2( 0.25, 0.5 ) );\n"
+"            float2 Off = TexelSize * float2( 0.5, 0.0 );\n"
+"\n"
+"            Src1 = 0.5 * ( SrcMip.SampleLevel( LinearClampSampler, UV1, SrcMipLevel ) +\n"
+"                           SrcMip.SampleLevel( LinearClampSampler, UV1 + Off, SrcMipLevel ) );\n"
+"        }\n"
+"        break;\n"
+"        case WIDTH_EVEN_HEIGHT_ODD:\n"
+"        {\n"
+"            // > 2:1 in Y dimension\n"
+"            // Use 2 bilinear samples to guarantee we don't undersample when downsizing by more than 2x\n"
+"            // vertically.\n"
+"            float2 UV1 = TexelSize * ( IN.DispatchThreadID.xy + float2( 0.5, 0.25 ) );\n"
+"            float2 Off = TexelSize * float2( 0.0, 0.5 );\n"
+"\n"
+"            Src1 = 0.5 * ( SrcMip.SampleLevel( LinearClampSampler, UV1, SrcMipLevel ) +\n"
+"                           SrcMip.SampleLevel( LinearClampSampler, UV1 + Off, SrcMipLevel ) );\n"
+"        }\n"
+"        break;\n"
+"        case WIDTH_HEIGHT_ODD:\n"
+"        {\n"
+"            // > 2:1 in in both dimensions\n"
+"            // Use 4 bilinear samples to guarantee we don't undersample when downsizing by more than 2x\n"
+"            // in both directions.\n"
+"            float2 UV1 = TexelSize * ( IN.DispatchThreadID.xy + float2( 0.25, 0.25 ) );\n"
+"            float2 Off = TexelSize * 0.5;\n"
+"\n"
+"            Src1 =  SrcMip.SampleLevel( LinearClampSampler, UV1, SrcMipLevel );\n"
+"            Src1 += SrcMip.SampleLevel( LinearClampSampler, UV1 + float2( Off.x, 0.0   ), SrcMipLevel );\n"
+"            Src1 += SrcMip.SampleLevel( LinearClampSampler, UV1 + float2( 0.0,   Off.y ), SrcMipLevel );\n"
+"            Src1 += SrcMip.SampleLevel( LinearClampSampler, UV1 + float2( Off.x, Off.y ), SrcMipLevel );\n"
+"            Src1 *= 0.25;\n"
+"        }\n"
+"        break;\n"
+"    }\n"
+"\n"
+"    OutMip1[IN.DispatchThreadID.xy] = Src1;\n"
+"\n"
+"    // A scalar (constant) branch can exit all threads coherently.\n"
+"    if ( NumMipLevels == 1 )\n"
+"        return;\n"
+"\n"
+"    // Without lane swizzle operations, the only way to share data with other\n"
+"    // threads is through LDS.\n"
+"    StoreColor( IN.GroupIndex, Src1 );\n"
+"\n"
+"    // This guarantees all LDS writes are complete and that all threads have\n"
+"    // executed all instructions so far (and therefore have issued their LDS\n"
+"    // write instructions.)\n"
+"    GroupMemoryBarrierWithGroupSync();\n"
+"\n"
+"    // With low three bits for X and high three bits for Y, this bit mask\n"
+"    // (binary: 001001) checks that X and Y are even.\n"
+"    if ( ( IN.GroupIndex & 0x9 ) == 0 )\n"
+"    {\n"
+"        float4 Src2 = LoadColor( IN.GroupIndex + 0x01 );\n"
+"        float4 Src3 = LoadColor( IN.GroupIndex + 0x08 );\n"
+"        float4 Src4 = LoadColor( IN.GroupIndex + 0x09 );\n"
+"        Src1 = 0.25 * ( Src1 + Src2 + Src3 + Src4 );\n"
+"\n"
+"        OutMip2[IN.DispatchThreadID.xy / 2] = Src1;\n"
+"        StoreColor( IN.GroupIndex, Src1 );\n"
+"    }\n"
+"\n"
+"    if ( NumMipLevels == 2 )\n"
+"        return;\n"
+"\n"
+"    GroupMemoryBarrierWithGroupSync();\n"
+"\n"
+"    // This bit mask (binary: 011011) checks that X and Y are multiples of four.\n"
+"    if ( ( IN.GroupIndex & 0x1B ) == 0 )\n"
+"    {\n"
+"        float4 Src2 = LoadColor( IN.GroupIndex + 0x02 );\n"
+"        float4 Src3 = LoadColor( IN.GroupIndex + 0x10 );\n"
+"        float4 Src4 = LoadColor( IN.GroupIndex + 0x12 );\n"
+"        Src1 = 0.25 * ( Src1 + Src2 + Src3 + Src4 );\n"
+"\n"
+"        OutMip3[IN.DispatchThreadID.xy / 4] = Src1;\n"
+"        StoreColor( IN.GroupIndex, Src1 );\n"
+"    }\n"
+"\n"
+"    if ( NumMipLevels == 3 )\n"
+"        return;\n"
+"\n"
+"    GroupMemoryBarrierWithGroupSync();\n"
+"\n"
+"    // This bit mask would be 111111 (X & Y multiples of 8), but only one\n"
+"    // thread fits that criteria.\n"
+"    if ( IN.GroupIndex == 0 )\n"
+"    {\n"
+"        float4 Src2 = LoadColor( IN.GroupIndex + 0x04 );\n"
+"        float4 Src3 = LoadColor( IN.GroupIndex + 0x20 );\n"
+"        float4 Src4 = LoadColor( IN.GroupIndex + 0x24 );\n"
+"        Src1 = 0.25 * ( Src1 + Src2 + Src3 + Src4 );\n"
+"\n"
+"        OutMip4[IN.DispatchThreadID.xy / 8] = Src1;\n"
+"    }\n"
+"}\n";
+#endif
diff --git a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/hlsl.h b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/hlsl.h
index 67a81deeaa..843759a4a4 100644
--- a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/hlsl.h
+++ b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/hlsl.h
@@ -30,3 +30,4 @@
 #include "VSMain_color.hlsl"
 #include "VSMain_coord.hlsl"
 #include "VSMain_pos.hlsl"
+#include "CSMain_mipgen.hlsl"
diff --git a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/meson.build b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/meson.build
index fd23812fd9..c9b1797a1c 100644
--- a/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/meson.build
+++ b/subprojects/gst-plugins-bad/gst-libs/gst/d3dshader/plugin-hlsl/meson.build
@@ -10,6 +10,7 @@ hlsl_sources = [
   ['VSMain_color', 'vs'],
   ['VSMain_coord', 'vs'],
   ['VSMain_pos', 'vs'],
+  ['CSMain_mipgen', 'cs'],
 ]
 
 shader_model = '5_0'
@@ -58,8 +59,19 @@ plugin_vs_collection = custom_target('plugin_hlsl_vs',
         '--output', '@OUTPUT@'
     ])
 
+plugin_cs_collection = custom_target('plugin_hlsl_cs',
+    input : plugin_hlsl_precompiled,
+    output : 'plugin_hlsl_cs.h',
+    command : [header_collector,
+        '--input', meson.current_build_dir(),
+        '--prefix', 'CSMain_',
+        '--name', 'g_plugin_cs_table',
+        '--output', '@OUTPUT@'
+    ])
+
 hlsl_precompiled += [
   plugin_hlsl_precompiled,
   plugin_ps_collection,
   plugin_vs_collection,
+  plugin_cs_collection,
 ]
diff --git a/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipgen.cpp b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipgen.cpp
new file mode 100644
index 0000000000..a5555f0a8c
--- /dev/null
+++ b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipgen.cpp
@@ -0,0 +1,366 @@
+/* GStreamer
+ * Copyright (C) 2024 Seungha Yang <seungha@centricular.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
+ * Boston, MA 02110-1301, USA.
+ */
+
+/*
+ * Copyright(c) 2018 Jeremiah van Oosten
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files(the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and / or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions :
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+/* Reference: https://github.com/jpvanoosten/LearningDirectX12 */
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "gstd3d12mipgen.h"
+#include <gst/d3d12/gstd3d12-private.h>
+#include <gst/d3dshader/gstd3dshader.h>
+#include <directx/d3dx12.h>
+#include <wrl.h>
+#include <algorithm>
+
+#define _XM_NO_INTRINSICS_
+#include <DirectXMath.h>
+
+GST_DEBUG_CATEGORY_STATIC (gst_d3d12_mip_gen_debug);
+#define GST_CAT_DEFAULT gst_d3d12_mip_gen_debug
+
+/* *INDENT-OFF* */
+using namespace Microsoft::WRL;
+using namespace DirectX;
+/* *INDENT-ON* */
+
+struct GenerateMipsCB
+{
+  UINT SrcMipLevel;
+  UINT NumMipLevels;
+  UINT SrcDimension;
+  UINT padding;
+  XMFLOAT2 TexelSize;
+};
+
+struct GstD3D12MipGenPrivate
+{
+  ~GstD3D12MipGenPrivate ()
+  {
+    pso = nullptr;
+    rs = nullptr;
+    gst_clear_object (&desc_pool);
+    gst_clear_object (&device);
+  }
+
+  GstD3D12Device *device = nullptr;
+  GstD3D12DescriptorPool *desc_pool = nullptr;
+  ComPtr < ID3D12PipelineState > pso;
+  ComPtr < ID3D12RootSignature > rs;
+  guint desc_inc_size;
+};
+
+struct _GstD3D12MipGen
+{
+  GstObject parent;
+
+  GstD3D12MipGenPrivate *priv;
+};
+/* *INDENT-ON* */
+
+static void gst_d3d12_mip_gen_finalize (GObject * object);
+
+#define gst_d3d12_mip_gen_parent_class parent_class
+G_DEFINE_TYPE (GstD3D12MipGen, gst_d3d12_mip_gen, GST_TYPE_OBJECT);
+
+static void
+gst_d3d12_mip_gen_class_init (GstD3D12MipGenClass * klass)
+{
+  auto object_class = G_OBJECT_CLASS (klass);
+
+  object_class->finalize = gst_d3d12_mip_gen_finalize;
+
+  GST_DEBUG_CATEGORY_INIT (gst_d3d12_mip_gen_debug,
+      "d3d12mipgen", 0, "d3d12mipgen");
+}
+
+static void
+gst_d3d12_mip_gen_init (GstD3D12MipGen * self)
+{
+  self->priv = new GstD3D12MipGenPrivate ();
+}
+
+static void
+gst_d3d12_mip_gen_finalize (GObject * object)
+{
+  auto self = GST_D3D12_MIP_GEN (object);
+
+  delete self->priv;
+
+  G_OBJECT_CLASS (parent_class)->finalize (object);
+}
+
+GstD3D12MipGen *
+gst_d3d12_mip_gen_new (GstD3D12Device * device)
+{
+  g_return_val_if_fail (GST_IS_D3D12_DEVICE (device), nullptr);
+
+  D3D12_VERSIONED_ROOT_SIGNATURE_DESC rs_desc = { };
+  CD3DX12_ROOT_PARAMETER root_params[3];
+  CD3DX12_DESCRIPTOR_RANGE range_srv;
+  CD3DX12_DESCRIPTOR_RANGE range_uav;
+  D3D12_STATIC_SAMPLER_DESC sampler_desc = { };
+
+  auto self = (GstD3D12MipGen *) g_object_new (GST_TYPE_D3D12_MIP_GEN, nullptr);
+  gst_object_ref_sink (self);
+
+  auto priv = self->priv;
+  priv->device = (GstD3D12Device *) gst_object_ref (device);
+
+  sampler_desc.Filter = D3D12_FILTER_MIN_MAG_LINEAR_MIP_POINT;
+  sampler_desc.AddressU = D3D12_TEXTURE_ADDRESS_MODE_CLAMP;
+  sampler_desc.AddressV = D3D12_TEXTURE_ADDRESS_MODE_CLAMP;
+  sampler_desc.AddressW = D3D12_TEXTURE_ADDRESS_MODE_CLAMP;
+  sampler_desc.MipLODBias = 0;
+  sampler_desc.MaxAnisotropy = 1;
+  sampler_desc.ComparisonFunc = D3D12_COMPARISON_FUNC_ALWAYS;
+  sampler_desc.BorderColor = D3D12_STATIC_BORDER_COLOR_OPAQUE_BLACK;
+  sampler_desc.MinLOD = 0;
+  sampler_desc.MaxLOD = D3D12_FLOAT32_MAX;
+  sampler_desc.ShaderRegister = 0;
+  sampler_desc.RegisterSpace = 0;
+  sampler_desc.ShaderVisibility = D3D12_SHADER_VISIBILITY_ALL;
+
+  root_params[0].InitAsConstants (6, 0, 0);
+
+  range_srv.Init (D3D12_DESCRIPTOR_RANGE_TYPE_SRV, 1, 0);
+  root_params[1].InitAsDescriptorTable (1, &range_srv);
+
+  range_uav.Init (D3D12_DESCRIPTOR_RANGE_TYPE_UAV, 4, 0);
+  root_params[2].InitAsDescriptorTable (1, &range_uav);
+
+  CD3DX12_VERSIONED_ROOT_SIGNATURE_DESC::Init_1_0 (rs_desc, 3, root_params,
+      1, &sampler_desc, D3D12_ROOT_SIGNATURE_FLAG_NONE);
+
+  ComPtr < ID3DBlob > rs_blob;
+  ComPtr < ID3DBlob > error_blob;
+  auto hr = D3DX12SerializeVersionedRootSignature (&rs_desc,
+      D3D_ROOT_SIGNATURE_VERSION_1, &rs_blob, &error_blob);
+
+  if (!gst_d3d12_result (hr, device)) {
+    const gchar *error_msg = nullptr;
+    if (error_blob)
+      error_msg = (const gchar *) error_blob->GetBufferPointer ();
+
+    GST_ERROR_OBJECT (self,
+        "Couldn't serialize root signature, hr: 0x%x, error detail: %s",
+        (guint) hr, GST_STR_NULL (error_msg));
+    gst_object_unref (self);
+    return nullptr;
+  }
+
+  auto device_handle = gst_d3d12_device_get_device_handle (device);
+  hr = device_handle->CreateRootSignature (0, rs_blob->GetBufferPointer (),
+      rs_blob->GetBufferSize (), IID_PPV_ARGS (&priv->rs));
+  if (!gst_d3d12_result (hr, device)) {
+    GST_ERROR_OBJECT (self, "Couldn't create root signature");
+    gst_object_unref (self);
+    return nullptr;
+  }
+
+  GstD3DShaderByteCode byte_code;
+  if (!gst_d3d_plugin_shader_get_cs_blob (GST_D3D_PLUGIN_CS_MIP_GEN,
+          GST_D3D_SM_5_0, &byte_code)) {
+    GST_ERROR_OBJECT (self, "Couldn't get shader byte code");
+    gst_object_unref (self);
+    return nullptr;
+  }
+
+  D3D12_COMPUTE_PIPELINE_STATE_DESC pso_desc = { };
+  pso_desc.pRootSignature = priv->rs.Get ();
+  pso_desc.CS.pShaderBytecode = byte_code.byte_code;
+  pso_desc.CS.BytecodeLength = byte_code.byte_code_len;
+  hr = device_handle->CreateComputePipelineState (&pso_desc,
+      IID_PPV_ARGS (&priv->pso));
+  if (!gst_d3d12_result (hr, device)) {
+    GST_ERROR_OBJECT (self, "Couldn't create PSO");
+    gst_object_unref (self);
+    return nullptr;
+  }
+
+  D3D12_DESCRIPTOR_HEAP_DESC desc_heap_desc = { };
+  desc_heap_desc.Type = D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV;
+  desc_heap_desc.NumDescriptors = 5;
+  desc_heap_desc.Flags = D3D12_DESCRIPTOR_HEAP_FLAG_SHADER_VISIBLE;
+
+  priv->desc_pool = gst_d3d12_descriptor_pool_new (device_handle,
+      &desc_heap_desc);
+  if (!priv->desc_pool) {
+    GST_ERROR_OBJECT (self, "Couldn't create descriptor pool");
+    gst_object_unref (self);
+    return nullptr;
+  }
+
+  priv->desc_inc_size = device_handle->GetDescriptorHandleIncrementSize
+      (D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV);
+
+  return self;
+}
+
+gboolean
+gst_d3d12_mip_gen_execute (GstD3D12MipGen * gen, ID3D12Resource * resource,
+    GstD3D12FenceData * fence_data, ID3D12GraphicsCommandList * cl)
+{
+  g_return_val_if_fail (GST_IS_D3D12_MIP_GEN (gen), FALSE);
+  g_return_val_if_fail (resource, FALSE);
+  g_return_val_if_fail (fence_data, FALSE);
+  g_return_val_if_fail (cl, FALSE);
+
+  auto desc = GetDesc (resource);
+
+  if (desc.MipLevels == 1) {
+    GST_LOG_OBJECT (gen, "Single mip level texture");
+    return TRUE;
+  }
+
+  if ((desc.Flags & D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS) !=
+      D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS ||
+      (desc.Flags & D3D12_RESOURCE_FLAG_DENY_SHADER_RESOURCE) ==
+      D3D12_RESOURCE_FLAG_DENY_SHADER_RESOURCE) {
+    GST_WARNING_OBJECT (gen, "Resource flag is incompatible");
+    return FALSE;
+  }
+
+  auto priv = gen->priv;
+  auto device = gst_d3d12_device_get_device_handle (priv->device);
+
+  cl->SetComputeRootSignature (priv->rs.Get ());
+  cl->SetPipelineState (priv->pso.Get ());
+
+  D3D12_SHADER_RESOURCE_VIEW_DESC srv_desc = { };
+  srv_desc.Format = desc.Format;
+  srv_desc.Shader4ComponentMapping = D3D12_DEFAULT_SHADER_4_COMPONENT_MAPPING;
+  srv_desc.ViewDimension = D3D12_SRV_DIMENSION_TEXTURE2D;
+  srv_desc.Texture2D.MipLevels = desc.MipLevels;
+
+  for (guint srcMip = 0; srcMip < desc.MipLevels - 1;) {
+    guint64 srcWidth = desc.Width >> srcMip;
+    guint srcHeight = desc.Height >> srcMip;
+    guint dstWidth = static_cast < guint > (srcWidth >> 1);
+    guint dstHeight = srcHeight >> 1;
+    GenerateMipsCB cbuf;
+
+    // 0b00(0): Both width and height are even.
+    // 0b01(1): Width is odd, height is even.
+    // 0b10(2): Width is even, height is odd.
+    // 0b11(3): Both width and height are odd.
+    cbuf.SrcDimension = (srcHeight & 1) << 1 | (srcWidth & 1);
+
+    // How many mipmap levels to compute this pass (max 4 mips per pass)
+    DWORD mipCount;
+
+    // The number of times we can half the size of the texture and get
+    // exactly a 50% reduction in size.
+    // A 1 bit in the width or height indicates an odd dimension.
+    // The case where either the width or the height is exactly 1 is handled
+    // as a special case (as the dimension does not require reduction).
+    _BitScanForward (&mipCount, (dstWidth == 1 ? dstHeight : dstWidth) |
+        (dstHeight == 1 ? dstWidth : dstHeight));
+    // Maximum number of mips to generate is 4.
+    mipCount = std::min < DWORD > (4, mipCount + 1);
+    // Clamp to total number of mips left over.
+    mipCount = (srcMip + mipCount) >= desc.MipLevels ?
+        desc.MipLevels - srcMip - 1 : mipCount;
+
+    // Dimensions should not reduce to 0.
+    // This can happen if the width and height are not the same.
+    dstWidth = std::max < DWORD > (1, dstWidth);
+    dstHeight = std::max < DWORD > (1, dstHeight);
+
+    cbuf.SrcMipLevel = srcMip;
+    cbuf.NumMipLevels = mipCount;
+    cbuf.TexelSize.x = 1.0f / (float) dstWidth;
+    cbuf.TexelSize.y = 1.0f / (float) dstHeight;
+
+    if (srcMip != 0) {
+      D3D12_RESOURCE_BARRIER barrier =
+          CD3DX12_RESOURCE_BARRIER::Transition (resource,
+          D3D12_RESOURCE_STATE_UNORDERED_ACCESS,
+          D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE, srcMip);
+      cl->ResourceBarrier (1, &barrier);
+    }
+
+    GstD3D12Descriptor *desc_heap;
+    if (!gst_d3d12_descriptor_pool_acquire (priv->desc_pool, &desc_heap)) {
+      GST_ERROR_OBJECT (gen, "Couldn't acquire descriptor heap");
+      return FALSE;
+    }
+
+    gst_d3d12_fence_data_push (fence_data,
+        FENCE_NOTIFY_MINI_OBJECT (desc_heap));
+    auto desc_handle = gst_d3d12_descriptor_get_handle (desc_heap);
+    auto cpu_handle = CD3DX12_CPU_DESCRIPTOR_HANDLE
+        (GetCPUDescriptorHandleForHeapStart (desc_handle));
+
+    device->CreateShaderResourceView (resource, &srv_desc, cpu_handle);
+
+    for (guint mip = 0; mip < mipCount; mip++) {
+      D3D12_UNORDERED_ACCESS_VIEW_DESC uavDesc = { };
+      uavDesc.Format = desc.Format;
+      uavDesc.ViewDimension = D3D12_UAV_DIMENSION_TEXTURE2D;
+      uavDesc.Texture2D.MipSlice = srcMip + mip + 1;
+
+      cpu_handle.Offset (priv->desc_inc_size);
+      device->CreateUnorderedAccessView (resource,
+          nullptr, &uavDesc, cpu_handle);
+    }
+
+    auto gpu_handle = CD3DX12_GPU_DESCRIPTOR_HANDLE
+        (GetGPUDescriptorHandleForHeapStart (desc_handle));
+
+    ID3D12DescriptorHeap *heaps[] = { desc_handle };
+    cl->SetDescriptorHeaps (1, heaps);
+    cl->SetComputeRoot32BitConstants (0, 6, &cbuf, 0);
+    cl->SetComputeRootDescriptorTable (1, gpu_handle);
+    gpu_handle.Offset (priv->desc_inc_size);
+    cl->SetComputeRootDescriptorTable (2, gpu_handle);
+
+    cl->Dispatch ((dstWidth + 7) / 8, (dstHeight + 7) / 8, 1);
+
+    D3D12_RESOURCE_BARRIER barrier = CD3DX12_RESOURCE_BARRIER::UAV (resource);
+    cl->ResourceBarrier (1, &barrier);
+
+    srcMip += mipCount;
+  }
+
+  return TRUE;
+}
diff --git a/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipgen.h b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipgen.h
new file mode 100644
index 0000000000..9206457761
--- /dev/null
+++ b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipgen.h
@@ -0,0 +1,38 @@
+/* GStreamer
+ * Copyright (C) 2024 Seungha Yang <seungha@centricular.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
+ * Boston, MA 02110-1301, USA.
+ */
+
+#pragma once
+
+#include <gst/gst.h>
+#include <gst/d3d12/gstd3d12.h>
+
+G_BEGIN_DECLS
+
+#define GST_TYPE_D3D12_MIP_GEN (gst_d3d12_mip_gen_get_type())
+G_DECLARE_FINAL_TYPE (GstD3D12MipGen, gst_d3d12_mip_gen, GST, D3D12_MIP_GEN, GstObject);
+
+GstD3D12MipGen * gst_d3d12_mip_gen_new (GstD3D12Device * device);
+
+gboolean         gst_d3d12_mip_gen_execute (GstD3D12MipGen * gen,
+                                            ID3D12Resource * resource,
+                                            GstD3D12FenceData * fence_data,
+                                            ID3D12GraphicsCommandList * cl);
+
+G_END_DECLS
+
diff --git a/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipmapping.cpp b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipmapping.cpp
new file mode 100644
index 0000000000..843ef562bd
--- /dev/null
+++ b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipmapping.cpp
@@ -0,0 +1,1019 @@
+/* GStreamer
+ * Copyright (C) 2024 Seungha Yang <seungha@centricular.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
+ * Boston, MA 02110-1301, USA.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "gstd3d12mipmapping.h"
+#include "gstd3d12mipgen.h"
+#include "gstd3d12pluginutils.h"
+#include <directx/d3dx12.h>
+#include <mutex>
+#include <memory>
+#include <queue>
+#include <wrl.h>
+#include <atomic>
+
+/* *INDENT-OFF* */
+using namespace Microsoft::WRL;
+/* *INDENT-ON* */
+
+GST_DEBUG_CATEGORY_STATIC (gst_d3d12_mip_mapping_debug);
+#define GST_CAT_DEFAULT gst_d3d12_mip_mapping_debug
+
+static GstStaticPadTemplate sink_template = GST_STATIC_PAD_TEMPLATE ("sink",
+    GST_PAD_SINK,
+    GST_PAD_ALWAYS,
+    GST_STATIC_CAPS (GST_VIDEO_CAPS_MAKE_WITH_FEATURES
+        (GST_CAPS_FEATURE_MEMORY_D3D12_MEMORY, GST_D3D12_ALL_FORMATS) "; "
+        GST_VIDEO_CAPS_MAKE_WITH_FEATURES
+        (GST_CAPS_FEATURE_MEMORY_D3D12_MEMORY ","
+            GST_CAPS_FEATURE_META_GST_VIDEO_OVERLAY_COMPOSITION,
+            GST_D3D12_ALL_FORMATS)));
+
+static GstStaticPadTemplate src_template = GST_STATIC_PAD_TEMPLATE ("src",
+    GST_PAD_SRC,
+    GST_PAD_ALWAYS,
+    GST_STATIC_CAPS (GST_VIDEO_CAPS_MAKE_WITH_FEATURES
+        (GST_CAPS_FEATURE_MEMORY_D3D12_MEMORY, "RGBA") "; "
+        GST_VIDEO_CAPS_MAKE_WITH_FEATURES
+        (GST_CAPS_FEATURE_MEMORY_D3D12_MEMORY ","
+            GST_CAPS_FEATURE_META_GST_VIDEO_OVERLAY_COMPOSITION,
+            "RGBA")));
+
+enum
+{
+  PROP_0,
+  PROP_ASYNC_DEPTH,
+};
+
+#define DEFAULT_ASYNC_DEPTH 0
+
+/* *INDENT-OFF* */
+struct MipMappingContext
+{
+  MipMappingContext (GstD3D12Device * dev)
+  {
+    device = (GstD3D12Device *) gst_object_ref (dev);
+    auto device_handle = gst_d3d12_device_get_device_handle (device);
+    ca_pool = gst_d3d12_command_allocator_pool_new (device_handle,
+        D3D12_COMMAND_LIST_TYPE_DIRECT);
+  }
+
+   ~MipMappingContext ()
+  {
+    gst_d3d12_device_fence_wait (device, D3D12_COMMAND_LIST_TYPE_DIRECT,
+        fence_val);
+
+    gst_clear_object (&ca_pool);
+    gst_clear_object (&conv);
+    gst_clear_object (&gen);
+    gst_clear_object (&device);
+  }
+
+  GstD3D12Device *device = nullptr;
+  GstD3D12Converter *conv = nullptr;
+  GstD3D12MipGen *gen = nullptr;
+  ComPtr<ID3D12GraphicsCommandList> cl;
+  std::queue<guint64> scheduled;
+  GstD3D12CommandAllocatorPool *ca_pool;
+  guint64 fence_val = 0;
+};
+
+struct GstD3D12MipMappingPrivate
+{
+  GstD3D12MipMappingPrivate ()
+  {
+    fence_data_pool = gst_d3d12_fence_data_pool_new ();
+  }
+
+  ~GstD3D12MipMappingPrivate ()
+  {
+    gst_clear_object (&fence_data_pool);
+  }
+
+  std::unique_ptr < MipMappingContext > ctx;
+  GstD3D12FenceDataPool *fence_data_pool;
+  D3D12_BOX in_rect = { };
+  D3D12_BOX prev_in_rect = { };
+
+  std::atomic<guint> async_depth = { DEFAULT_ASYNC_DEPTH };
+
+  std::mutex lock;
+};
+/* *INDENT-ON* */
+
+struct _GstD3D12MipMapping
+{
+  GstD3D12BaseFilter parent;
+
+  GstD3D12MipMappingPrivate *priv;
+};
+
+static void gst_d3d12_mip_mapping_finalize (GObject * object);
+static void gst_d3d12_mip_mapping_set_property (GObject * object, guint prop_id,
+    const GValue * value, GParamSpec * pspec);
+static void gst_d3d12_mip_mapping_get_property (GObject * object, guint prop_id,
+    GValue * value, GParamSpec * pspec);
+static gboolean gst_d3d12_mip_mapping_stop (GstBaseTransform * trans);
+static GstCaps *gst_d3d12_mip_mapping_transform_caps (GstBaseTransform *
+    trans, GstPadDirection direction, GstCaps * caps, GstCaps * filter);
+static GstCaps *gst_d3d12_mip_mapping_fixate_caps (GstBaseTransform *
+    base, GstPadDirection direction, GstCaps * caps, GstCaps * othercaps);
+static gboolean gst_d3d12_mip_mapping_propose_allocation (GstBaseTransform *
+    trans, GstQuery * decide_query, GstQuery * query);
+static gboolean gst_d3d12_mip_mapping_decide_allocation (GstBaseTransform *
+    trans, GstQuery * query);
+static gboolean gst_d3d12_mip_mapping_transform_meta (GstBaseTransform * trans,
+    GstBuffer * outbuf, GstMeta * meta, GstBuffer * inbuf);
+static GstFlowReturn gst_d3d12_mip_mapping_transform (GstBaseTransform * trans,
+    GstBuffer * inbuf, GstBuffer * outbuf);
+static gboolean gst_d3d12_mip_mapping_set_info (GstD3D12BaseFilter * filter,
+    GstCaps * incaps, GstVideoInfo * in_info, GstCaps * outcaps,
+    GstVideoInfo * out_info);
+
+#define gst_d3d12_mip_mapping_parent_class parent_class
+G_DEFINE_TYPE (GstD3D12MipMapping, gst_d3d12_mip_mapping,
+    GST_TYPE_D3D12_BASE_FILTER);
+
+static void
+gst_d3d12_mip_mapping_class_init (GstD3D12MipMappingClass * klass)
+{
+  auto object_class = G_OBJECT_CLASS (klass);
+  auto element_class = GST_ELEMENT_CLASS (klass);
+  auto trans_class = GST_BASE_TRANSFORM_CLASS (klass);
+  auto filter_class = GST_D3D12_BASE_FILTER_CLASS (klass);
+
+  object_class->set_property = gst_d3d12_mip_mapping_set_property;
+  object_class->get_property = gst_d3d12_mip_mapping_get_property;
+  object_class->finalize = gst_d3d12_mip_mapping_finalize;
+
+  g_object_class_install_property (object_class, PROP_ASYNC_DEPTH,
+      g_param_spec_uint ("async-depth", "Async Depth",
+          "Number of in-flight GPU commands which can be scheduled without "
+          "synchronization (0 = unlimited)", 0, G_MAXINT, DEFAULT_ASYNC_DEPTH,
+          (GParamFlags) (GST_PARAM_MUTABLE_PLAYING |
+              G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS)));
+
+  gst_element_class_add_static_pad_template (element_class, &sink_template);
+  gst_element_class_add_static_pad_template (element_class, &src_template);
+
+  gst_element_class_set_static_metadata (element_class,
+      "Direct3D12 MipMapping",
+      "Filter/Converter/Video/Hardware",
+      "Generates RGBA MipMap texture from input",
+      "Seungha Yang <seungha@centricular.com>");
+
+  trans_class->passthrough_on_same_caps = FALSE;
+
+  trans_class->stop = GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_stop);
+  trans_class->transform_caps =
+      GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_transform_caps);
+  trans_class->fixate_caps =
+      GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_fixate_caps);
+  trans_class->propose_allocation =
+      GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_propose_allocation);
+  trans_class->decide_allocation =
+      GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_decide_allocation);
+  trans_class->transform_meta =
+      GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_transform_meta);
+  trans_class->transform = GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_transform);
+
+  filter_class->set_info = GST_DEBUG_FUNCPTR (gst_d3d12_mip_mapping_set_info);
+
+  gst_type_mark_as_plugin_api (GST_TYPE_D3D12_SAMPLING_METHOD,
+      (GstPluginAPIFlags) 0);
+
+  GST_DEBUG_CATEGORY_INIT (gst_d3d12_mip_mapping_debug, "d3d12convert", 0,
+      "d3d12convert");
+}
+
+static void
+gst_d3d12_mip_mapping_init (GstD3D12MipMapping * self)
+{
+  self->priv = new GstD3D12MipMappingPrivate ();
+}
+
+static void
+gst_d3d12_mip_mapping_finalize (GObject * object)
+{
+  auto self = GST_D3D12_MIP_MAPPING (object);
+
+  delete self->priv;
+
+  G_OBJECT_CLASS (parent_class)->finalize (object);
+}
+
+static void
+gst_d3d12_mip_mapping_set_property (GObject * object, guint prop_id,
+    const GValue * value, GParamSpec * pspec)
+{
+  auto self = GST_D3D12_MIP_MAPPING (object);
+  auto priv = self->priv;
+
+  switch (prop_id) {
+    case PROP_ASYNC_DEPTH:
+      priv->async_depth = g_value_get_uint (value);
+      break;
+    default:
+      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
+      break;
+  }
+}
+
+static void
+gst_d3d12_mip_mapping_get_property (GObject * object, guint prop_id,
+    GValue * value, GParamSpec * pspec)
+{
+  auto self = GST_D3D12_MIP_MAPPING (object);
+  auto priv = self->priv;
+
+  switch (prop_id) {
+    case PROP_ASYNC_DEPTH:
+      g_value_set_uint (value, priv->async_depth);
+      break;
+    default:
+      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
+      break;
+  }
+}
+
+static gboolean
+gst_d3d12_mip_mapping_stop (GstBaseTransform * trans)
+{
+  auto self = GST_D3D12_MIP_MAPPING (trans);
+  auto priv = self->priv;
+
+  priv->ctx = nullptr;
+
+  return GST_BASE_TRANSFORM_CLASS (parent_class)->stop (trans);
+}
+
+static GstCaps *
+gst_d3d12_mip_mapping_caps_remove_format_info (GstCaps * caps)
+{
+  GstStructure *st;
+  GstCapsFeatures *f;
+  gint i, n;
+  GstCaps *res;
+  GstCapsFeatures *feature =
+      gst_caps_features_from_string (GST_CAPS_FEATURE_MEMORY_D3D12_MEMORY);
+
+  res = gst_caps_new_empty ();
+
+  n = gst_caps_get_size (caps);
+  for (i = 0; i < n; i++) {
+    st = gst_caps_get_structure (caps, i);
+    f = gst_caps_get_features (caps, i);
+
+    /* If this is already expressed by the existing caps
+     * skip this structure */
+    if (i > 0 && gst_caps_is_subset_structure_full (res, st, f))
+      continue;
+
+    st = gst_structure_copy (st);
+    /* Only remove format info for the cases when we can actually convert */
+    if (!gst_caps_features_is_any (f)
+        && gst_caps_features_is_equal (f, feature)) {
+      gst_structure_remove_fields (st, "format", "colorimetry", "chroma-site",
+          NULL);
+    }
+
+    gst_caps_append_structure_full (res, st, gst_caps_features_copy (f));
+  }
+  gst_caps_features_free (feature);
+
+  return res;
+}
+
+static GstCaps *
+gst_d3d12_mip_mapping_transform_caps (GstBaseTransform *
+    trans, GstPadDirection direction, GstCaps * caps, GstCaps * filter)
+{
+  GstCaps *tmp, *tmp2;
+  GstCaps *result;
+
+  /* Get all possible caps that we can transform to */
+  tmp = gst_d3d12_mip_mapping_caps_remove_format_info (caps);
+
+  if (filter) {
+    tmp2 = gst_caps_intersect_full (filter, tmp, GST_CAPS_INTERSECT_FIRST);
+    gst_caps_unref (tmp);
+    tmp = tmp2;
+  }
+
+  result = tmp;
+
+  GST_DEBUG_OBJECT (trans, "transformed %" GST_PTR_FORMAT " into %"
+      GST_PTR_FORMAT, caps, result);
+
+  return result;
+}
+
+/*
+ * This is an incomplete matrix of in formats and a score for the prefered output
+ * format.
+ *
+ *         out: RGB24   RGB16  ARGB  AYUV  YUV444  YUV422 YUV420 YUV411 YUV410  PAL  GRAY
+ *  in
+ * RGB24          0      2       1     2     2       3      4      5      6      7    8
+ * RGB16          1      0       1     2     2       3      4      5      6      7    8
+ * ARGB           2      3       0     1     4       5      6      7      8      9    10
+ * AYUV           3      4       1     0     2       5      6      7      8      9    10
+ * YUV444         2      4       3     1     0       5      6      7      8      9    10
+ * YUV422         3      5       4     2     1       0      6      7      8      9    10
+ * YUV420         4      6       5     3     2       1      0      7      8      9    10
+ * YUV411         4      6       5     3     2       1      7      0      8      9    10
+ * YUV410         6      8       7     5     4       3      2      1      0      9    10
+ * PAL            1      3       2     6     4       6      7      8      9      0    10
+ * GRAY           1      4       3     2     1       5      6      7      8      9    0
+ *
+ * PAL or GRAY are never prefered, if we can we would convert to PAL instead
+ * of GRAY, though
+ * less subsampling is prefered and if any, preferably horizontal
+ * We would like to keep the alpha, even if we would need to to colorspace conversion
+ * or lose depth.
+ */
+#define SCORE_FORMAT_CHANGE       1
+#define SCORE_DEPTH_CHANGE        1
+#define SCORE_ALPHA_CHANGE        1
+#define SCORE_CHROMA_W_CHANGE     1
+#define SCORE_CHROMA_H_CHANGE     1
+#define SCORE_PALETTE_CHANGE      1
+
+#define SCORE_COLORSPACE_LOSS     2     /* RGB <-> YUV */
+#define SCORE_DEPTH_LOSS          4     /* change bit depth */
+#define SCORE_ALPHA_LOSS          8     /* lose the alpha channel */
+#define SCORE_CHROMA_W_LOSS      16     /* vertical subsample */
+#define SCORE_CHROMA_H_LOSS      32     /* horizontal subsample */
+#define SCORE_PALETTE_LOSS       64     /* convert to palette format */
+#define SCORE_COLOR_LOSS        128     /* convert to GRAY */
+
+#define COLORSPACE_MASK (GST_VIDEO_FORMAT_FLAG_YUV | \
+                         GST_VIDEO_FORMAT_FLAG_RGB | GST_VIDEO_FORMAT_FLAG_GRAY)
+#define ALPHA_MASK      (GST_VIDEO_FORMAT_FLAG_ALPHA)
+#define PALETTE_MASK    (GST_VIDEO_FORMAT_FLAG_PALETTE)
+
+/* calculate how much loss a conversion would be */
+static void
+score_value (GstBaseTransform * base, const GstVideoFormatInfo * in_info,
+    const GValue * val, gint * min_loss, const GstVideoFormatInfo ** out_info)
+{
+  const gchar *fname;
+  const GstVideoFormatInfo *t_info;
+  guint in_flags, t_flags;
+  gint loss;
+
+  fname = g_value_get_string (val);
+  t_info = gst_video_format_get_info (gst_video_format_from_string (fname));
+  if (!t_info || t_info->format == GST_VIDEO_FORMAT_UNKNOWN)
+    return;
+
+  /* accept input format immediately without loss */
+  if (in_info == t_info) {
+    *min_loss = 0;
+    *out_info = t_info;
+    return;
+  }
+
+  loss = SCORE_FORMAT_CHANGE;
+
+  in_flags = GST_VIDEO_FORMAT_INFO_FLAGS (in_info);
+  in_flags &= ~GST_VIDEO_FORMAT_FLAG_LE;
+  in_flags &= ~GST_VIDEO_FORMAT_FLAG_COMPLEX;
+  in_flags &= ~GST_VIDEO_FORMAT_FLAG_UNPACK;
+
+  t_flags = GST_VIDEO_FORMAT_INFO_FLAGS (t_info);
+  t_flags &= ~GST_VIDEO_FORMAT_FLAG_LE;
+  t_flags &= ~GST_VIDEO_FORMAT_FLAG_COMPLEX;
+  t_flags &= ~GST_VIDEO_FORMAT_FLAG_UNPACK;
+
+  if ((t_flags & PALETTE_MASK) != (in_flags & PALETTE_MASK)) {
+    loss += SCORE_PALETTE_CHANGE;
+    if (t_flags & PALETTE_MASK)
+      loss += SCORE_PALETTE_LOSS;
+  }
+
+  if ((t_flags & COLORSPACE_MASK) != (in_flags & COLORSPACE_MASK)) {
+    loss += SCORE_COLORSPACE_LOSS;
+    if (t_flags & GST_VIDEO_FORMAT_FLAG_GRAY)
+      loss += SCORE_COLOR_LOSS;
+  }
+
+  if ((t_flags & ALPHA_MASK) != (in_flags & ALPHA_MASK)) {
+    loss += SCORE_ALPHA_CHANGE;
+    if (in_flags & ALPHA_MASK)
+      loss += SCORE_ALPHA_LOSS;
+  }
+
+  if ((in_info->h_sub[1]) != (t_info->h_sub[1])) {
+    loss += SCORE_CHROMA_H_CHANGE;
+    if ((in_info->h_sub[1]) < (t_info->h_sub[1]))
+      loss += SCORE_CHROMA_H_LOSS;
+  }
+  if ((in_info->w_sub[1]) != (t_info->w_sub[1])) {
+    loss += SCORE_CHROMA_W_CHANGE;
+    if ((in_info->w_sub[1]) < (t_info->w_sub[1]))
+      loss += SCORE_CHROMA_W_LOSS;
+  }
+
+  if ((in_info->bits) != (t_info->bits)) {
+    loss += SCORE_DEPTH_CHANGE;
+    if ((in_info->bits) > (t_info->bits))
+      loss += SCORE_DEPTH_LOSS + (in_info->bits - t_info->bits);
+  }
+
+  GST_DEBUG_OBJECT (base, "score %s -> %s = %d",
+      GST_VIDEO_FORMAT_INFO_NAME (in_info),
+      GST_VIDEO_FORMAT_INFO_NAME (t_info), loss);
+
+  if (loss < *min_loss) {
+    GST_DEBUG_OBJECT (base, "found new best %d", loss);
+    *out_info = t_info;
+    *min_loss = loss;
+  }
+}
+
+static void
+gst_d3d12_mip_mapping_fixate_format (GstBaseTransform * trans, GstCaps * caps,
+    GstCaps * result)
+{
+  GstStructure *ins, *outs;
+  const gchar *in_format;
+  const GstVideoFormatInfo *in_info, *out_info = nullptr;
+  gint min_loss = G_MAXINT;
+  guint i, capslen;
+
+  ins = gst_caps_get_structure (caps, 0);
+  in_format = gst_structure_get_string (ins, "format");
+  if (!in_format) {
+    return;
+  }
+
+  GST_DEBUG_OBJECT (trans, "source format %s", in_format);
+
+  in_info =
+      gst_video_format_get_info (gst_video_format_from_string (in_format));
+  if (!in_info)
+    return;
+
+  outs = gst_caps_get_structure (result, 0);
+
+  capslen = gst_caps_get_size (result);
+  GST_DEBUG ("iterate %d structures", capslen);
+  for (i = 0; i < capslen; i++) {
+    GstStructure *tests;
+    const GValue *format;
+
+    tests = gst_caps_get_structure (result, i);
+    format = gst_structure_get_value (tests, "format");
+
+    /* should not happen */
+    if (format == nullptr)
+      continue;
+
+    if (GST_VALUE_HOLDS_LIST (format)) {
+      gint j, len;
+
+      len = gst_value_list_get_size (format);
+      GST_DEBUG_OBJECT (trans, "have %d formats", len);
+      for (j = 0; j < len; j++) {
+        const GValue *val;
+
+        val = gst_value_list_get_value (format, j);
+        if (G_VALUE_HOLDS_STRING (val)) {
+          score_value (trans, in_info, val, &min_loss, &out_info);
+          if (min_loss == 0)
+            break;
+        }
+      }
+    } else if (G_VALUE_HOLDS_STRING (format)) {
+      score_value (trans, in_info, format, &min_loss, &out_info);
+    }
+  }
+  if (out_info)
+    gst_structure_set (outs, "format", G_TYPE_STRING,
+        GST_VIDEO_FORMAT_INFO_NAME (out_info), nullptr);
+}
+
+static gboolean
+subsampling_unchanged (GstVideoInfo * in_info, GstVideoInfo * out_info)
+{
+  guint i;
+  const GstVideoFormatInfo *in_format, *out_format;
+
+  if (GST_VIDEO_INFO_N_COMPONENTS (in_info) !=
+      GST_VIDEO_INFO_N_COMPONENTS (out_info))
+    return FALSE;
+
+  in_format = in_info->finfo;
+  out_format = out_info->finfo;
+
+  for (i = 0; i < GST_VIDEO_INFO_N_COMPONENTS (in_info); i++) {
+    if (GST_VIDEO_FORMAT_INFO_W_SUB (in_format,
+            i) != GST_VIDEO_FORMAT_INFO_W_SUB (out_format, i))
+      return FALSE;
+    if (GST_VIDEO_FORMAT_INFO_H_SUB (in_format,
+            i) != GST_VIDEO_FORMAT_INFO_H_SUB (out_format, i))
+      return FALSE;
+  }
+
+  return TRUE;
+}
+
+static void
+transfer_colorimetry_from_input (GstBaseTransform * trans, GstCaps * in_caps,
+    GstCaps * out_caps)
+{
+  GstStructure *out_caps_s = gst_caps_get_structure (out_caps, 0);
+  GstStructure *in_caps_s = gst_caps_get_structure (in_caps, 0);
+  gboolean have_colorimetry =
+      gst_structure_has_field (out_caps_s, "colorimetry");
+  gboolean have_chroma_site =
+      gst_structure_has_field (out_caps_s, "chroma-site");
+
+  /* If the output already has colorimetry and chroma-site, stop,
+   * otherwise try and transfer what we can from the input caps */
+  if (have_colorimetry && have_chroma_site)
+    return;
+
+  {
+    GstVideoInfo in_info, out_info;
+    const GValue *in_colorimetry =
+        gst_structure_get_value (in_caps_s, "colorimetry");
+
+    if (!gst_video_info_from_caps (&in_info, in_caps)) {
+      GST_WARNING_OBJECT (trans,
+          "Failed to convert sink pad caps to video info");
+      return;
+    }
+    if (!gst_video_info_from_caps (&out_info, out_caps)) {
+      GST_WARNING_OBJECT (trans,
+          "Failed to convert src pad caps to video info");
+      return;
+    }
+
+    if (!have_colorimetry && in_colorimetry != nullptr) {
+      if ((GST_VIDEO_INFO_IS_YUV (&out_info)
+              && GST_VIDEO_INFO_IS_YUV (&in_info))
+          || (GST_VIDEO_INFO_IS_RGB (&out_info)
+              && GST_VIDEO_INFO_IS_RGB (&in_info))
+          || (GST_VIDEO_INFO_IS_GRAY (&out_info)
+              && GST_VIDEO_INFO_IS_GRAY (&in_info))) {
+        /* Can transfer the colorimetry intact from the input if it has it */
+        gst_structure_set_value (out_caps_s, "colorimetry", in_colorimetry);
+      } else {
+        gchar *colorimetry_str;
+
+        /* Changing between YUV/RGB - forward primaries and transfer function, but use
+         * default range and matrix.
+         * the primaries is used for conversion between RGB and XYZ (CIE 1931 coordinate).
+         * the transfer function could be another reference (e.g., HDR)
+         */
+        out_info.colorimetry.primaries = in_info.colorimetry.primaries;
+        out_info.colorimetry.transfer = in_info.colorimetry.transfer;
+
+        colorimetry_str =
+            gst_video_colorimetry_to_string (&out_info.colorimetry);
+        gst_caps_set_simple (out_caps, "colorimetry", G_TYPE_STRING,
+            colorimetry_str, nullptr);
+        g_free (colorimetry_str);
+      }
+    }
+
+    /* Only YUV output needs chroma-site. If the input was also YUV and had the same chroma
+     * subsampling, transfer the siting. If the sub-sampling is changing, then the planes get
+     * scaled anyway so there's no real reason to prefer the input siting. */
+    if (!have_chroma_site && GST_VIDEO_INFO_IS_YUV (&out_info)) {
+      if (GST_VIDEO_INFO_IS_YUV (&in_info)) {
+        const GValue *in_chroma_site =
+            gst_structure_get_value (in_caps_s, "chroma-site");
+        if (in_chroma_site != nullptr
+            && subsampling_unchanged (&in_info, &out_info))
+          gst_structure_set_value (out_caps_s, "chroma-site", in_chroma_site);
+      }
+    }
+  }
+}
+
+static GstCaps *
+gst_d3d12_mip_mapping_get_fixed_format (GstBaseTransform * trans,
+    GstPadDirection direction, GstCaps * caps, GstCaps * othercaps)
+{
+  GstCaps *result;
+
+  result = gst_caps_intersect (othercaps, caps);
+  if (gst_caps_is_empty (result)) {
+    gst_caps_unref (result);
+    result = gst_caps_copy (othercaps);
+  }
+
+  gst_d3d12_mip_mapping_fixate_format (trans, caps, result);
+
+  /* fixate remaining fields */
+  result = gst_caps_fixate (result);
+
+  if (direction == GST_PAD_SINK) {
+    if (gst_caps_is_subset (caps, result)) {
+      gst_caps_replace (&result, caps);
+    } else {
+      /* Try and preserve input colorimetry / chroma information */
+      transfer_colorimetry_from_input (trans, caps, result);
+    }
+  }
+
+  return result;
+}
+
+static GstCaps *
+gst_d3d12_mip_mapping_fixate_caps (GstBaseTransform * trans,
+    GstPadDirection direction, GstCaps * caps, GstCaps * othercaps)
+{
+  GST_DEBUG_OBJECT (trans,
+      "trying to fixate othercaps %" GST_PTR_FORMAT " based on caps %"
+      GST_PTR_FORMAT, othercaps, caps);
+
+  auto format = gst_d3d12_mip_mapping_get_fixed_format (trans, direction, caps,
+      othercaps);
+  gst_caps_unref (othercaps);
+
+  if (gst_caps_is_empty (format)) {
+    GST_ERROR_OBJECT (trans, "Could not convert formats");
+  } else {
+    GST_DEBUG_OBJECT (trans, "fixated othercaps to %" GST_PTR_FORMAT, format);
+  }
+
+  return format;
+}
+
+static gboolean
+gst_d3d12_mip_mapping_propose_allocation (GstBaseTransform * trans,
+    GstQuery * decide_query, GstQuery * query)
+{
+  auto filter = GST_D3D12_BASE_FILTER (trans);
+  GstVideoInfo info;
+  GstBufferPool *pool = nullptr;
+  GstCaps *caps;
+  guint n_pools, i;
+  guint size;
+
+  if (!GST_BASE_TRANSFORM_CLASS (parent_class)->propose_allocation (trans,
+          decide_query, query)) {
+    return FALSE;
+  }
+
+  gst_query_parse_allocation (query, &caps, nullptr);
+
+  if (!caps)
+    return FALSE;
+
+  if (!gst_video_info_from_caps (&info, caps)) {
+    GST_ERROR_OBJECT (filter, "Invalid caps %" GST_PTR_FORMAT, caps);
+    return FALSE;
+  }
+
+  n_pools = gst_query_get_n_allocation_pools (query);
+  for (i = 0; i < n_pools; i++) {
+    gst_query_parse_nth_allocation_pool (query, i, &pool, nullptr, nullptr,
+        nullptr);
+    if (pool) {
+      if (!GST_IS_D3D12_BUFFER_POOL (pool)) {
+        gst_clear_object (&pool);
+      } else {
+        auto dpool = GST_D3D12_BUFFER_POOL (pool);
+        if (!gst_d3d12_device_is_equal (dpool->device, filter->device))
+          gst_clear_object (&pool);
+      }
+    }
+  }
+
+  if (!pool)
+    pool = gst_d3d12_buffer_pool_new (filter->device);
+
+  auto config = gst_buffer_pool_get_config (pool);
+  gst_buffer_pool_config_add_option (config, GST_BUFFER_POOL_OPTION_VIDEO_META);
+
+  auto d3d12_params =
+      gst_buffer_pool_config_get_d3d12_allocation_params (config);
+  if (!d3d12_params) {
+    d3d12_params = gst_d3d12_allocation_params_new (filter->device, &info,
+        GST_D3D12_ALLOCATION_FLAG_DEFAULT,
+        D3D12_RESOURCE_FLAG_ALLOW_SIMULTANEOUS_ACCESS, D3D12_HEAP_FLAG_NONE);
+  } else {
+    gst_d3d12_allocation_params_set_resource_flags (d3d12_params,
+        D3D12_RESOURCE_FLAG_ALLOW_SIMULTANEOUS_ACCESS);
+    gst_d3d12_allocation_params_unset_resource_flags (d3d12_params,
+        D3D12_RESOURCE_FLAG_DENY_SHADER_RESOURCE);
+  }
+
+  gst_buffer_pool_config_set_d3d12_allocation_params (config, d3d12_params);
+  gst_d3d12_allocation_params_free (d3d12_params);
+
+  /* size will be updated by d3d12 buffer pool */
+  gst_buffer_pool_config_set_params (config, caps, 0, 0, 0);
+
+  if (!gst_buffer_pool_set_config (pool, config)) {
+    GST_ERROR_OBJECT (filter, "failed to set config");
+    gst_object_unref (pool);
+    return FALSE;
+  }
+
+  gst_query_add_allocation_meta (query, GST_VIDEO_META_API_TYPE, nullptr);
+  gst_query_add_allocation_meta (query,
+      GST_VIDEO_OVERLAY_COMPOSITION_META_API_TYPE, nullptr);
+  gst_query_add_allocation_meta (query, GST_VIDEO_CROP_META_API_TYPE, nullptr);
+
+  /* d3d12 buffer pool will update buffer size based on allocated texture,
+   * get size from config again */
+  config = gst_buffer_pool_get_config (pool);
+  gst_buffer_pool_config_get_params (config, nullptr, &size, nullptr, nullptr);
+  gst_structure_free (config);
+
+  gst_query_add_allocation_pool (query, pool, size, 0, 0);
+
+  gst_object_unref (pool);
+
+  return TRUE;
+}
+
+static gboolean
+gst_d3d12_mip_mapping_decide_allocation (GstBaseTransform * trans,
+    GstQuery * query)
+{
+  auto filter = GST_D3D12_BASE_FILTER (trans);
+  GstCaps *outcaps = nullptr;
+  GstBufferPool *pool = nullptr;
+  guint size, min = 0, max = 0;
+  GstStructure *config;
+  gboolean update_pool = FALSE;
+  GstVideoInfo info;
+
+  gst_query_parse_allocation (query, &outcaps, nullptr);
+
+  if (!outcaps)
+    return FALSE;
+
+  if (!gst_video_info_from_caps (&info, outcaps)) {
+    GST_ERROR_OBJECT (filter, "Invalid caps %" GST_PTR_FORMAT, outcaps);
+    return FALSE;
+  }
+
+  size = GST_VIDEO_INFO_SIZE (&info);
+  if (gst_query_get_n_allocation_pools (query) > 0) {
+    gst_query_parse_nth_allocation_pool (query, 0, &pool, &size, &min, &max);
+    if (pool) {
+      if (!GST_IS_D3D12_BUFFER_POOL (pool)) {
+        gst_clear_object (&pool);
+      } else {
+        auto dpool = GST_D3D12_BUFFER_POOL (pool);
+        if (!gst_d3d12_device_is_equal (dpool->device, filter->device))
+          gst_clear_object (&pool);
+      }
+    }
+
+    update_pool = TRUE;
+  }
+
+  if (!pool)
+    pool = gst_d3d12_buffer_pool_new (filter->device);
+
+  config = gst_buffer_pool_get_config (pool);
+  gst_buffer_pool_config_add_option (config, GST_BUFFER_POOL_OPTION_VIDEO_META);
+
+  D3D12_RESOURCE_FLAGS resource_flags =
+      D3D12_RESOURCE_FLAG_ALLOW_SIMULTANEOUS_ACCESS |
+      D3D12_RESOURCE_FLAG_ALLOW_UNORDERED_ACCESS |
+      D3D12_RESOURCE_FLAG_ALLOW_RENDER_TARGET;
+
+  auto d3d12_params = gst_d3d12_allocation_params_new (filter->device, &info,
+      GST_D3D12_ALLOCATION_FLAG_DEFAULT, resource_flags,
+      D3D12_HEAP_FLAG_SHARED);
+
+  /* Auto generate mip maps */
+  gst_d3d12_allocation_params_set_mip_levels (d3d12_params, 0);
+
+  gst_buffer_pool_config_set_d3d12_allocation_params (config, d3d12_params);
+  gst_d3d12_allocation_params_free (d3d12_params);
+
+  gst_buffer_pool_config_set_params (config, outcaps, size, min, max);
+  gst_buffer_pool_set_config (pool, config);
+
+  /* d3d12 buffer pool will update buffer size based on allocated texture,
+   * get size from config again */
+  config = gst_buffer_pool_get_config (pool);
+  gst_buffer_pool_config_get_params (config, nullptr, &size, nullptr, nullptr);
+  gst_structure_free (config);
+
+  if (update_pool)
+    gst_query_set_nth_allocation_pool (query, 0, pool, size, min, max);
+  else
+    gst_query_add_allocation_pool (query, pool, size, min, max);
+
+  gst_object_unref (pool);
+
+  return GST_BASE_TRANSFORM_CLASS (parent_class)->decide_allocation (trans,
+      query);
+}
+
+static gboolean
+gst_d3d12_mip_mapping_set_info (GstD3D12BaseFilter * filter,
+    GstCaps * incaps, GstVideoInfo * in_info, GstCaps * outcaps,
+    GstVideoInfo * out_info)
+{
+  auto self = GST_D3D12_MIP_MAPPING (filter);
+  auto priv = self->priv;
+
+  priv->ctx = nullptr;
+
+  GST_DEBUG_OBJECT (self, "Setup convert with format %s -> %s",
+      gst_video_format_to_string (GST_VIDEO_INFO_FORMAT (in_info)),
+      gst_video_format_to_string (GST_VIDEO_INFO_FORMAT (out_info)));
+
+  /* if present, these must match */
+  if (in_info->interlace_mode != out_info->interlace_mode) {
+    GST_ERROR_OBJECT (self, "input and output formats do not match");
+    return FALSE;
+  }
+
+  auto ctx = std::make_unique < MipMappingContext > (filter->device);
+
+  ctx->conv = gst_d3d12_converter_new (filter->device, nullptr, in_info,
+      out_info, nullptr, nullptr, nullptr);
+  if (!ctx->conv) {
+    GST_ERROR_OBJECT (self, "Couldn't create converter");
+    return FALSE;
+  }
+
+  ctx->gen = gst_d3d12_mip_gen_new (filter->device);
+  if (!ctx->gen) {
+    GST_ERROR_OBJECT (self, "Couldn't create mip generator");
+    return FALSE;
+  }
+
+  priv->in_rect = CD3DX12_BOX (0, 0,
+      GST_VIDEO_INFO_WIDTH (in_info), GST_VIDEO_INFO_HEIGHT (in_info));
+  priv->prev_in_rect = priv->in_rect;
+
+  priv->ctx = std::move (ctx);
+
+  return TRUE;
+}
+
+static gboolean
+gst_d3d12_mip_mapping_transform_meta (GstBaseTransform * trans,
+    GstBuffer * outbuf, GstMeta * meta, GstBuffer * inbuf)
+{
+  if (meta->info->api == GST_VIDEO_CROP_META_API_TYPE)
+    return FALSE;
+
+  return GST_BASE_TRANSFORM_CLASS (parent_class)->transform_meta (trans,
+      outbuf, meta, inbuf);
+}
+
+static GstFlowReturn
+gst_d3d12_mip_mapping_transform (GstBaseTransform * trans, GstBuffer * inbuf,
+    GstBuffer * outbuf)
+{
+  auto self = GST_D3D12_MIP_MAPPING (trans);
+  auto priv = self->priv;
+  D3D12_BOX in_rect;
+
+  auto crop_meta = gst_buffer_get_video_crop_meta (inbuf);
+  if (crop_meta) {
+    GST_LOG_OBJECT (self, "Have crop rect, x:y:w:h = %d:%d:%d:%d",
+        crop_meta->x, crop_meta->y, crop_meta->width, crop_meta->height);
+
+    in_rect = CD3DX12_BOX (crop_meta->x, crop_meta->y,
+        crop_meta->x + crop_meta->width, crop_meta->y + crop_meta->height);
+  } else {
+    in_rect = priv->in_rect;
+  }
+
+  if (in_rect != priv->in_rect) {
+    priv->prev_in_rect = in_rect;
+    g_object_set (priv->ctx->conv, "src-x", (gint) in_rect.left,
+        "src-y", (gint) in_rect.top,
+        "src-width", (gint) in_rect.right - in_rect.left,
+        "src-height", (gint) in_rect.bottom - in_rect.top, nullptr);
+  }
+
+  GstD3D12CommandAllocator *gst_ca;
+  if (!gst_d3d12_command_allocator_pool_acquire (priv->ctx->ca_pool, &gst_ca)) {
+    GST_ERROR_OBJECT (self, "Couldn't acquire command allocator");
+    return GST_FLOW_ERROR;
+  }
+
+  auto ca = gst_d3d12_command_allocator_get_handle (gst_ca);
+  auto hr = ca->Reset ();
+  if (!gst_d3d12_result (hr, priv->ctx->device)) {
+    GST_ERROR_OBJECT (self, "Couldn't reset command allocator");
+    gst_d3d12_command_allocator_unref (gst_ca);
+    return GST_FLOW_ERROR;
+  }
+
+  if (!priv->ctx->cl) {
+    auto device = gst_d3d12_device_get_device_handle (priv->ctx->device);
+    hr = device->CreateCommandList (0, D3D12_COMMAND_LIST_TYPE_DIRECT,
+        ca, nullptr, IID_PPV_ARGS (&priv->ctx->cl));
+    if (!gst_d3d12_result (hr, priv->ctx->device)) {
+      GST_ERROR_OBJECT (self, "Couldn't create command list");
+      gst_d3d12_command_allocator_unref (gst_ca);
+      return GST_FLOW_ERROR;
+    }
+  } else {
+    hr = priv->ctx->cl->Reset (ca, nullptr);
+    if (!gst_d3d12_result (hr, priv->ctx->device)) {
+      GST_ERROR_OBJECT (self, "Couldn't reset command list");
+      gst_d3d12_command_allocator_unref (gst_ca);
+      return GST_FLOW_ERROR;
+    }
+  }
+
+  GstD3D12FenceData *fence_data;
+  gst_d3d12_fence_data_pool_acquire (priv->fence_data_pool, &fence_data);
+  gst_d3d12_fence_data_push (fence_data, FENCE_NOTIFY_MINI_OBJECT (gst_ca));
+
+  auto cq = gst_d3d12_device_get_command_queue (priv->ctx->device,
+      D3D12_COMMAND_LIST_TYPE_DIRECT);
+  auto fence = gst_d3d12_command_queue_get_fence_handle (cq);
+  if (!gst_d3d12_converter_convert_buffer (priv->ctx->conv,
+          inbuf, outbuf, fence_data, priv->ctx->cl.Get (), TRUE)) {
+    GST_ERROR_OBJECT (self, "Couldn't build command list");
+    gst_d3d12_fence_data_unref (fence_data);
+    return GST_FLOW_ERROR;
+  }
+
+  auto dmem = (GstD3D12Memory *) gst_buffer_peek_memory (outbuf, 0);
+  auto tex = gst_d3d12_memory_get_resource_handle (dmem);
+
+  D3D12_RESOURCE_BARRIER barrier = CD3DX12_RESOURCE_BARRIER::Transition
+      (tex, D3D12_RESOURCE_STATE_RENDER_TARGET,
+      D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE, 0);
+  priv->ctx->cl->ResourceBarrier (1, &barrier);
+
+  if (!gst_d3d12_mip_gen_execute (priv->ctx->gen, tex, fence_data,
+          priv->ctx->cl.Get ())) {
+    GST_ERROR_OBJECT (self, "Couldn't build mip gen command");
+    gst_d3d12_fence_data_unref (fence_data);
+    return GST_FLOW_ERROR;
+  }
+
+  hr = priv->ctx->cl->Close ();
+  if (!gst_d3d12_result (hr, priv->ctx->device)) {
+    GST_ERROR_OBJECT (self, "Couldn't close command list");
+    gst_d3d12_fence_data_unref (fence_data);
+    return GST_FLOW_ERROR;
+  }
+
+  ID3D12CommandList *cmd_list[] = { priv->ctx->cl.Get () };
+
+  hr = gst_d3d12_command_queue_execute_command_lists (cq,
+      1, cmd_list, &priv->ctx->fence_val);
+  if (!gst_d3d12_result (hr, priv->ctx->device)) {
+    GST_ERROR_OBJECT (self, "Couldn't execute command list");
+    gst_d3d12_fence_data_unref (fence_data);
+    return GST_FLOW_ERROR;
+  }
+
+  gst_d3d12_buffer_set_fence (outbuf, fence, priv->ctx->fence_val, FALSE);
+  gst_d3d12_command_queue_set_notify (cq, priv->ctx->fence_val,
+      FENCE_NOTIFY_MINI_OBJECT (fence_data));
+
+  priv->ctx->scheduled.push (priv->ctx->fence_val);
+
+  auto completed = gst_d3d12_device_get_completed_value (priv->ctx->device,
+      D3D12_COMMAND_LIST_TYPE_DIRECT);
+  while (!priv->ctx->scheduled.empty ()) {
+    if (priv->ctx->scheduled.front () > completed)
+      break;
+
+    priv->ctx->scheduled.pop ();
+  }
+
+  auto async_depth = priv->async_depth.load ();
+  if (async_depth > 0 && priv->ctx->scheduled.size () > async_depth) {
+    auto fence_to_wait = priv->ctx->scheduled.front ();
+    priv->ctx->scheduled.pop ();
+    gst_d3d12_device_fence_wait (priv->ctx->device,
+        D3D12_COMMAND_LIST_TYPE_DIRECT, fence_to_wait);
+  }
+
+  return GST_FLOW_OK;
+}
diff --git a/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipmapping.h b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipmapping.h
new file mode 100644
index 0000000000..668711d88e
--- /dev/null
+++ b/subprojects/gst-plugins-bad/sys/d3d12/gstd3d12mipmapping.h
@@ -0,0 +1,32 @@
+/* GStreamer
+ * Copyright (C) 2024 Seungha Yang <seungha@centricular.com>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
+ * Boston, MA 02110-1301, USA.
+ */
+
+#pragma once
+
+#include <gst/gst.h>
+#include "gstd3d12basefilter.h"
+
+G_BEGIN_DECLS
+
+#define GST_TYPE_D3D12_MIP_MAPPING (gst_d3d12_mip_mapping_get_type())
+G_DECLARE_FINAL_TYPE (GstD3D12MipMapping, gst_d3d12_mip_mapping,
+    GST, D3D12_MIP_MAPPING, GstD3D12BaseFilter)
+
+G_END_DECLS
+
diff --git a/subprojects/gst-plugins-bad/sys/d3d12/meson.build b/subprojects/gst-plugins-bad/sys/d3d12/meson.build
index bf0e09e46e..594c41bf72 100644
--- a/subprojects/gst-plugins-bad/sys/d3d12/meson.build
+++ b/subprojects/gst-plugins-bad/sys/d3d12/meson.build
@@ -18,6 +18,8 @@ d3d12_sources = [
   'gstd3d12ipcsink.cpp',
   'gstd3d12ipcsrc.cpp',
   'gstd3d12mpeg2dec.cpp',
+  'gstd3d12mipgen.cpp',
+  'gstd3d12mipmapping.cpp',
   'gstd3d12overlaycompositor.cpp',
   'gstd3d12pluginutils.cpp',
   'gstd3d12screencapture.cpp',
diff --git a/subprojects/gst-plugins-bad/sys/d3d12/plugin.cpp b/subprojects/gst-plugins-bad/sys/d3d12/plugin.cpp
index 00299adf87..e597399fcd 100644
--- a/subprojects/gst-plugins-bad/sys/d3d12/plugin.cpp
+++ b/subprojects/gst-plugins-bad/sys/d3d12/plugin.cpp
@@ -49,6 +49,7 @@
 #include "gstd3d12ipcsrc.h"
 #include "gstd3d12ipcsink.h"
 #include "gstd3d12swapchainsink.h"
+#include "gstd3d12mipmapping.h"
 #include <windows.h>
 #include <versionhelpers.h>
 #include <wrl.h>
@@ -181,6 +182,8 @@ plugin_init (GstPlugin * plugin)
       "d3d12ipcsink", GST_RANK_NONE, GST_TYPE_D3D12_IPC_SINK);
   gst_element_register (plugin,
       "d3d12swapchainsink", GST_RANK_NONE, GST_TYPE_D3D12_SWAPCHAIN_SINK);
+  gst_element_register (plugin,
+      "d3d12mipmapping", GST_RANK_NONE, GST_TYPE_D3D12_MIP_MAPPING);
 
   g_object_set_data_full (G_OBJECT (plugin),
       "plugin-d3d12-shutdown", (gpointer) "shutdown-data",