diff --git a/src/amd/vulkan/radv_cmd_buffer.h b/src/amd/vulkan/radv_cmd_buffer.h
index 044aac9b20e..03c9bfe58d6 100644
--- a/src/amd/vulkan/radv_cmd_buffer.h
+++ b/src/amd/vulkan/radv_cmd_buffer.h
@@ -486,6 +486,8 @@ struct radv_cmd_state {
    bool uses_vrs_coarse_shading;
    bool uses_dynamic_patch_control_points;
    bool uses_fbfetch_output;
+
+   uint64_t shader_query_buf_va; /* GFX12+ */
 };
 
 struct radv_enc_state {
diff --git a/src/amd/vulkan/radv_constants.h b/src/amd/vulkan/radv_constants.h
index ebad1580761..6398ce1ae29 100644
--- a/src/amd/vulkan/radv_constants.h
+++ b/src/amd/vulkan/radv_constants.h
@@ -151,6 +151,11 @@
 #define RADV_SHADER_QUERY_MS_PRIM_GEN_OFFSET      60
 #define RADV_SHADER_QUERY_TS_INVOCATION_OFFSET    64
 
+/* Size of the shader query buffer for generated/written primitive queries
+ * using SSBO atomics on GFX12.
+ */
+#define RADV_SHADER_QUERY_BUF_SIZE (RADV_SHADER_QUERY_PRIM_XFB_OFFSET(3) - RADV_SHADER_QUERY_PRIM_GEN_OFFSET(0) + 4)
+
 /* Number of samples for line smooth lowering (hw requirement). */
 #define RADV_NUM_SMOOTH_AA_SAMPLES 4
 
diff --git a/src/amd/vulkan/radv_query.c b/src/amd/vulkan/radv_query.c
index 69eb8574947..b688be7e294 100644
--- a/src/amd/vulkan/radv_query.c
+++ b/src/amd/vulkan/radv_query.c
@@ -961,6 +961,28 @@ emit_sample_streamout(struct radv_cmd_buffer *cmd_buffer, uint64_t va, uint32_t
    radeon_emit(cs, va >> 32);
 }
 
+static void
+radv_alloc_shader_query_buf(struct radv_cmd_buffer *cmd_buffer)
+{
+   const struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
+   const struct radv_physical_device *pdev = radv_device_physical(device);
+   unsigned offset;
+   void *ptr;
+
+   assert(pdev->info.gfx_level >= GFX12);
+
+   if (cmd_buffer->state.shader_query_buf_va)
+      return;
+
+   if (!radv_cmd_buffer_upload_alloc_aligned(cmd_buffer, RADV_SHADER_QUERY_BUF_SIZE, 64, &offset, &ptr))
+      return;
+
+   memset(ptr, 0, RADV_SHADER_QUERY_BUF_SIZE);
+
+   cmd_buffer->state.shader_query_buf_va = radv_buffer_get_va(cmd_buffer->upload.upload_bo);
+   cmd_buffer->state.shader_query_buf_va += offset;
+}
+
 static void
 radv_begin_tfb_query(struct radv_cmd_buffer *cmd_buffer, uint64_t va, uint32_t index)
 {
@@ -1953,6 +1975,10 @@ radv_create_query_pool(struct radv_device *device, const VkQueryPoolCreateInfo *
    /* The number of task shader invocations needs to be queried on ACE. */
    pool->uses_ace = (pool->vk.pipeline_statistics & VK_QUERY_PIPELINE_STATISTIC_TASK_SHADER_INVOCATIONS_BIT_EXT);
 
+   pool->uses_shader_query_buf =
+      pdev->info.gfx_level >= GFX12 && (pool->vk.query_type == VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT ||
+                                        pool->vk.query_type == VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT);
+
    switch (pCreateInfo->queryType) {
    case VK_QUERY_TYPE_OCCLUSION:
       pool->stride = 16 * pdev->info.max_render_backends;
@@ -2701,6 +2727,9 @@ radv_CmdBeginQueryIndexedEXT(VkCommandBuffer commandBuffer, VkQueryPool queryPoo
       radv_cs_add_buffer(device->ws, cmd_buffer->gang.cs, pool->bo);
    }
 
+   if (pool->uses_shader_query_buf)
+      radv_alloc_shader_query_buf(cmd_buffer);
+
    emit_begin_query(cmd_buffer, pool, va, pool->vk.query_type, flags, index);
 }
 
diff --git a/src/amd/vulkan/radv_query.h b/src/amd/vulkan/radv_query.h
index 84f4dcd600d..b428bc0abbf 100644
--- a/src/amd/vulkan/radv_query.h
+++ b/src/amd/vulkan/radv_query.h
@@ -26,6 +26,7 @@ struct radv_query_pool {
    char *ptr;
    bool uses_emulated_queries;
    bool uses_ace; /* For task shader invocations on GFX10.3+ */
+   bool uses_shader_query_buf; /* For generated/written primitives on GFX12+ */
 };
 
 VK_DEFINE_NONDISP_HANDLE_CASTS(radv_query_pool, vk.base, VkQueryPool, VK_OBJECT_TYPE_QUERY_POOL)