vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

* vulkan: initial support for IQ3_S

* vulkan: initial support for IQ3_XXS

* vulkan: initial support for IQ2_XXS

* vulkan: initial support for IQ2_XS

* vulkan: optimize Q3_K by removing branches

* vulkan: implement dequantize variants for coopmat2

* vulkan: initial support for IQ2_S

* vulkan: vertically realign code

* port failing dequant callbacks from mul_mm

* Fix array length mismatches

* vulkan: avoid using workgroup size before it is referenced

* tests: increase timeout for Vulkan llvmpipe backend

---------

Co-authored-by: Jeff Bolz <redacted>

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index cd8422f8a266c6d0da7af595b4df44f76ebb76af..7eaf9c46070bffa74f041523045e7ce00aeb987c 100644 (file)
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -346,7 +346,8 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          # This is using llvmpipe and runs slower than other backends
+          ctest -L main --verbose --timeout 1800
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 6c7e606507bb37454f83e2782d47ebbf0994de21..9ca3959abf128c1cc4171241ef070c2fb4f0cc2a 100644 (file)
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1616,6 +1616,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         //CREATE_FA(GGML_TYPE_Q4_K, q4_k)
         //CREATE_FA(GGML_TYPE_Q5_K, q5_k)
         //CREATE_FA(GGML_TYPE_Q6_K, q6_k)
+        //CREATE_FA(GGML_TYPE_IQ2_XXS, iq2_xxs)
+        //CREATE_FA(GGML_TYPE_IQ2_XS, iq2_xs)
+        //CREATE_FA(GGML_TYPE_IQ2_S, iq2_s)
+        //CREATE_FA(GGML_TYPE_IQ3_XXS, iq3_xxs)
+        //CREATE_FA(GGML_TYPE_IQ3_S, iq3_s)
         CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl)
 #undef CREATE_FA
 
@@ -1644,7 +1649,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f16,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f16,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
 
         CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
@@ -1657,7 +1667,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
 #undef CREATE_MM
 #undef CREATE_MM2
     } else
@@ -1705,7 +1720,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         } else {
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -1718,7 +1738,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
             CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         }
 
         // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
@@ -1739,7 +1764,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             } else {
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -1752,7 +1782,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
                 CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             }
         }
 #undef CREATE_MM2
@@ -1796,7 +1831,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc,  matmul_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc,   matmul_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f16acc, matmul_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f16acc,   matmul_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
         // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
         if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
@@ -1815,7 +1855,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         }
 #undef CREATE_MM2
 #undef CREATE_MM
@@ -1851,7 +1896,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f32acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f32acc,  matmul_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f32acc,   matmul_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f32acc, matmul_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc,   matmul_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc,  matmul_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
 
         // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
         if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
@@ -1870,7 +1920,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
             CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc,  matmul_id_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc,   matmul_id_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         }
 #undef CREATE_MM
     }
@@ -1901,7 +1956,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f32_f32_len,  mul_mat_vec_iq2_xs_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f32_f32_len,   mul_mat_vec_iq2_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f32_f32_len,   mul_mat_vec_iq3_s_f32_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f32_f32_len,  mul_mat_vec_iq4_nl_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
 
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
@@ -1915,7 +1975,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i],  "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq2_xs_f16_f32_len,  mul_mat_vec_iq2_xs_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i],   "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq2_s_f16_f32_len,   mul_mat_vec_iq2_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i],   "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1),   mul_mat_vec_iq3_s_f16_f32_len,   mul_mat_vec_iq3_s_f16_f32_data,   "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i],  "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1),  mul_mat_vec_iq4_nl_f16_f32_len,  mul_mat_vec_iq4_nl_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
@@ -1930,7 +1995,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS],  "mul_mat_vec_id_iq2_xs_f32",  mul_mat_vec_id_iq2_xs_f32_len,  mul_mat_vec_id_iq2_xs_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S],   "mul_mat_vec_id_iq2_s_f32",   mul_mat_vec_id_iq2_s_f32_len,   mul_mat_vec_id_iq2_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S],   "mul_mat_vec_id_iq3_s_f32",   mul_mat_vec_id_iq3_s_f32_len,   mul_mat_vec_id_iq3_s_f32_data,   "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL],  "mul_mat_vec_id_iq4_nl_f32",  mul_mat_vec_id_iq4_nl_f32_len,  mul_mat_vec_id_iq4_nl_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1944,7 +2014,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_k", dequant_q4_k_len, dequant_q4_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_k", dequant_q5_k_len, dequant_q5_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q6_K], "dequant_q6_k", dequant_q6_k_len, dequant_q6_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_NL], "dequant_iq4_nl", dequant_iq4_nl_len, dequant_iq4_nl_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_XXS], "dequant_iq2_xxs", dequant_iq2_xxs_len, dequant_iq2_xxs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_XS],  "dequant_iq2_xs",  dequant_iq2_xs_len,  dequant_iq2_xs_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_S],   "dequant_iq2_s",   dequant_iq2_s_len,   dequant_iq2_s_data,   "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ3_XXS], "dequant_iq3_xxs", dequant_iq3_xxs_len, dequant_iq3_xxs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ3_S],   "dequant_iq3_s",   dequant_iq3_s_len,   dequant_iq3_s_data,   "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_NL],  "dequant_iq4_nl",  dequant_iq4_nl_len,  dequant_iq4_nl_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
 
     // get_rows
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32",  get_rows_f32_len,  get_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
@@ -1954,7 +2029,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_1], "get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl", get_rows_iq4_nl_len, get_rows_iq4_nl_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_XXS], "get_rows_iq2_xxs", get_rows_iq2_xxs_len, get_rows_iq2_xxs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_XS],  "get_rows_iq2_xs",  get_rows_iq2_xs_len,  get_rows_iq2_xs_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_S],   "get_rows_iq2_s",   get_rows_iq2_s_len,   get_rows_iq2_s_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ3_XXS], "get_rows_iq3_xxs", get_rows_iq3_xxs_len, get_rows_iq3_xxs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ3_S],   "get_rows_iq3_s",   get_rows_iq3_s_len,   get_rows_iq3_s_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl",  get_rows_iq4_nl_len,  get_rows_iq4_nl_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32",  get_rows_f32_f32_len,  get_rows_f32_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32",  get_rows_f16_f32_len,  get_rows_f16_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
@@ -1963,7 +2043,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_1], "get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl_f32", get_rows_iq4_nl_f32_len, get_rows_iq4_nl_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_XXS], "get_rows_iq2_xxs_f32", get_rows_iq2_xxs_f32_len, get_rows_iq2_xxs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_XS],  "get_rows_iq2_xs_f32",  get_rows_iq2_xs_f32_len,  get_rows_iq2_xs_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_S],   "get_rows_iq2_s_f32",   get_rows_iq2_s_f32_len,   get_rows_iq2_s_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ3_XXS], "get_rows_iq3_xxs_f32", get_rows_iq3_xxs_f32_len, get_rows_iq3_xxs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ3_S],   "get_rows_iq3_s_f32",   get_rows_iq3_s_f32_len,   get_rows_iq3_s_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl_f32",  get_rows_iq4_nl_f32_len,  get_rows_iq4_nl_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
 
@@ -2890,6 +2975,11 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_NL:
             break;
         default:
@@ -2938,6 +3028,11 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_NL:
             break;
         default:
@@ -2969,6 +3064,11 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_NL:
             break;
         default:
@@ -3012,6 +3112,11 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_NL:
             break;
         default:
@@ -3038,6 +3143,11 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ4_NL:
             break;
         default:
@@ -7907,6 +8017,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_Q4_K:
                     case GGML_TYPE_Q5_K:
                     case GGML_TYPE_Q6_K:
+                    case GGML_TYPE_IQ2_XXS:
+                    case GGML_TYPE_IQ2_XS:
+                    case GGML_TYPE_IQ2_S:
+                    case GGML_TYPE_IQ3_XXS:
+                    case GGML_TYPE_IQ3_S:
                     case GGML_TYPE_IQ4_NL:
                         break;
                     default:
@@ -7975,6 +8090,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 //case GGML_TYPE_Q4_K:
                 //case GGML_TYPE_Q5_K:
                 //case GGML_TYPE_Q6_K:
+                //case GGML_TYPE_IQ2_XXS:
+                //case GGML_TYPE_IQ2_XS:
+                //case GGML_TYPE_IQ2_S:
+                //case GGML_TYPE_IQ3_XXS:
+                //case GGML_TYPE_IQ3_S:
                 case GGML_TYPE_IQ4_NL:
                     break;
                 default:
@@ -7992,6 +8112,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_IQ2_XXS:
+                    case GGML_TYPE_IQ2_XS:
+                    case GGML_TYPE_IQ2_S:
+                    case GGML_TYPE_IQ3_XXS:
+                    case GGML_TYPE_IQ3_S:
                     case GGML_TYPE_IQ4_NL:
                         return true;
                     default:

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp
index c09bf496bf3f75f33e7330e56b93b22c0cb89402..aeae5400dfcb8752a35c9edaa2d27371d8f1d13a 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp
@@ -12,8 +12,8 @@ layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
 #endif
 
 void main() {
-#if defined(DATA_A_IQ4_NL)
-    init_iq4nl_shmem();
+#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
+    init_iq_shmem(gl_WorkGroupSize);
     if (gl_LocalInvocationIndex.x != 0) {
         return;
     }

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp
index ccf5b980a8a59ed9011b72b31b248b84957de853..d4b068e61866a451c7861b7b7be1d3893c4e1c53 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp
@@ -217,8 +217,8 @@ void quantize(uint dst_idx, uint src_idx)
 #endif
 
 void main() {
-#if defined(DATA_A_IQ4_NL)
-    init_iq4nl_shmem();
+#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
+    init_iq_shmem(gl_WorkGroupSize);
     if (gl_LocalInvocationIndex.x != 0) {
         return;
     }

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
index 91bb8f8db610e185c6d8ff320ac13fe8b5f4581f..ee68775317b8339cabf538a119c0fef693bd4ba7 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
@@ -88,6 +88,222 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_IQ2_XXS)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = (iqs / 8) % 4;
+    const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8];
+    // Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
+    const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2],
+        data_a_packed16[a_offset + ib].qs[4 * ib32 + 3]));
+    const float db = 0.25 * (0.5 + (signs >> 28));
+    const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
+    // Add parity bit
+    const uint sign8 = sign7 | (bitCount(sign7) << 7);
+    const uint sign = sign8 >> (iqs % 8);
+    const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4)));
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    return db * vec2(
+        grid.x * (sign0 ? -1.0 : 1.0),
+        grid.y * (sign1 ? -1.0 : 1.0)
+    );
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = (iqs / 8) % 4;
+    const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8];
+    // Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
+    const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2],
+        data_a_packed16[a_offset + ib].qs[4 * ib32 + 3]));
+    const float db = 0.25 * (0.5 + (signs >> 28));
+    const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
+    // Add parity bit
+    const uint sign8 = sign7 | (bitCount(sign7) << 7);
+    const uint sign = sign8 >> (iqs % 8);
+    const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4)));
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    bool sign2 = (sign & 4) != 0;
+    bool sign3 = (sign & 8) != 0;
+    return db * vec4(
+        grid.x * (sign0 ? -1.0 : 1.0),
+        grid.y * (sign1 ? -1.0 : 1.0),
+        grid.z * (sign2 ? -1.0 : 1.0),
+        grid.w * (sign3 ? -1.0 : 1.0)
+    );
+}
+#endif
+
+#if defined(DATA_A_IQ2_XS)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint scale = (data_a[a_offset + ib].scales[iqs / 32] >> (4 * ((iqs / 16) & 1))) & 0xf;
+    const uint qs = data_a[a_offset + ib].qs[iqs / 8];
+    const float db = 0.25 * (0.5 + scale);
+    const uint sign7 = qs >> 9;
+    // Add parity bit
+    const uint sign8 = sign7 | (bitCount(sign7) << 7);
+    const uint sign = sign8 >> (iqs % 8);
+    const u8vec4 grid = unpack8(iq2xs_grid[qs & 511][(iqs % 8) / 4] >> (8 * (iqs % 4)));
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    return db * vec2(
+        grid.x * (sign0 ? -1.0 : 1.0),
+        grid.y * (sign1 ? -1.0 : 1.0)
+    );
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint scale = (data_a[a_offset + ib].scales[iqs / 32] >> (4 * ((iqs / 16) & 1))) & 0xf;
+    const uint qs = data_a[a_offset + ib].qs[iqs / 8];
+    const float db = 0.25 * (0.5 + scale);
+    const uint sign7 = qs >> 9;
+    // Add parity bit
+    const uint sign8 = sign7 | (bitCount(sign7) << 7);
+    const uint sign = sign8 >> (iqs % 8);
+    const u8vec4 grid = unpack8(iq2xs_grid[qs & 511][(iqs % 8) / 4] >> (8 * (iqs % 4)));
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    bool sign2 = (sign & 4) != 0;
+    bool sign3 = (sign & 8) != 0;
+    return db * vec4(
+        grid.x * (sign0 ? -1.0 : 1.0),
+        grid.y * (sign1 ? -1.0 : 1.0),
+        grid.z * (sign2 ? -1.0 : 1.0),
+        grid.w * (sign3 ? -1.0 : 1.0)
+    );
+}
+#endif
+
+#if defined(DATA_A_IQ2_S)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = iqs / 8;
+
+    const uint scale = (data_a[a_offset + ib].scales[ib32] >> (4 * ((iqs / 16) & 1))) & 0xf;
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const uint qh = data_a[a_offset + ib].qh[ib32];
+    const uint qhshift = 2 * (ib8 % 4);
+    const uint sign = data_a[a_offset + ib].qs[QUANT_K / 8 + ib8] >> (iqs % 8);
+
+    const float db = 0.25 * (0.5 + scale);
+    const u8vec4 grid = unpack8(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(iqs % 8) / 4]);
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    return db * vec2(
+        grid[iqs % 4] * (sign0 ? -1.0 : 1.0),
+        grid[(iqs % 4) + 1] * (sign1 ? -1.0 : 1.0)
+    );
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = iqs / 8;
+
+    const uint scale = (data_a[a_offset + ib].scales[ib32] >> (4 * ((iqs / 16) & 1))) & 0xf;
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const uint qh = data_a[a_offset + ib].qh[ib32];
+    const uint qhshift = 2 * (ib8 % 4);
+    const uint sign = data_a[a_offset + ib].qs[QUANT_K / 8 + ib8] >> (iqs % 8);
+
+    const float db = 0.25 * (0.5 + scale);
+    const u8vec4 grid = unpack8(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(iqs % 8) / 4]);
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    bool sign2 = (sign & 4) != 0;
+    bool sign3 = (sign & 8) != 0;
+    return db * vec4(
+        grid.x * (sign0 ? -1.0 : 1.0),
+        grid.y * (sign1 ? -1.0 : 1.0),
+        grid.z * (sign2 ? -1.0 : 1.0),
+        grid.w * (sign3 ? -1.0 : 1.0)
+    );
+}
+#endif
+
+#if defined(DATA_A_IQ3_XXS)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint ib4 = iqs / 4;
+    const uint ib32 = iqs / 32;
+    const uint is = QUANT_K / 4 + 4 * ib32;
+    const uint qs = data_a[a_offset + ib].qs[ib4];
+    // Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
+    const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[is / 2],
+        data_a_packed16[a_offset + ib].qs[is / 2 + 1]));
+    const float db = 0.5 * (0.5 + (signs >> 28));
+    const uint sign7 = bitfieldExtract(signs, 7 * (int(ib4 / 2) % 4), 7);
+    // Add parity bit
+    const uint sign8 = sign7 | (bitCount(sign7) << 7);
+    const uint sign = sign8 >> (iqs % 8);
+    const u8vec4 grid = unpack8(iq3xxs_grid[qs] >> (8 * (iqs % 4)));
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    return db * vec2(
+        grid.x * (sign0 ? -1.0 : 1.0),
+        grid.y * (sign1 ? -1.0 : 1.0)
+    );
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib4 = iqs / 4;
+    const uint ib32 = iqs / 32;
+    const uint is = QUANT_K / 4 + 4 * ib32;
+    const uint qs = data_a[a_offset + ib].qs[ib4];
+    const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[is / 2],
+        data_a_packed16[a_offset + ib].qs[is / 2 + 1]));
+    const float db = 0.5 * (0.5 + (signs >> 28));
+    const uint sign7 = bitfieldExtract(signs, 7 * (int(ib4 / 2) % 4), 7);
+    // Add parity bit
+    const uint sign8 = sign7 | (bitCount(sign7) << 7);
+    const uint sign = sign8 >> (iqs % 8);
+    const u8vec4 grid = unpack8(iq3xxs_grid[qs]);
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    bool sign2 = (sign & 4) != 0;
+    bool sign3 = (sign & 8) != 0;
+    return db * vec4(
+        grid.x * (sign0 ? -1.0 : 1.0),
+        grid.y * (sign1 ? -1.0 : 1.0),
+        grid.z * (sign2 ? -1.0 : 1.0),
+        grid.w * (sign3 ? -1.0 : 1.0)
+    );
+}
+#endif
+
+#if defined(DATA_A_IQ3_S)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint qs = data_a[a_offset + ib].qs[iqs / 4];
+    const uint qh = data_a[a_offset + ib].qh[iqs / 32];
+    const uint sign = data_a[a_offset + ib].signs[iqs / 8] >> (iqs % 8);
+    const uint scale = data_a[a_offset + ib].scales[iqs / 64];
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    const float db = 1 + 2 * ((scale >> (4 * ((iqs / 32) & 1))) & 0xf);
+    const uint32_t grid = iq3s_grid[qs | ((qh << (8 - ((iqs / 4) % 8))) & 256)] >> (8 * (iqs % 4));
+    return db * vec2(
+        int(grid & 0xFF) * (sign0 ? -1.0 : 1.0),
+        int((grid >> 8) & 0xFF) * (sign1 ? -1.0 : 1.0)
+    );
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib4 = iqs / 4;
+    const uint ib32 = iqs / 32;
+    const uint qs = data_a[a_offset + ib].qs[ib4];
+    const uint qh = data_a[a_offset + ib].qh[ib32];
+    const uint sign = data_a[a_offset + ib].signs[iqs / 8] >> (iqs % 8);
+    const uint scale = data_a[a_offset + ib].scales[ib32 / 2];
+    bool sign0 = (sign & 1) != 0;
+    bool sign1 = (sign & 2) != 0;
+    bool sign2 = (sign & 4) != 0;
+    bool sign3 = (sign & 8) != 0;
+    const float db = 1 + 2 * ((scale >> (4 * (ib32 & 1))) & 0xf);
+    const uint32_t grid = iq3s_grid[qs | ((qh << (8 - ib4 % 8)) & 256)] >> (8 * (iqs % 4));
+    return db * vec4(
+        int(grid & 0xFF) * (sign0 ? -1.0 : 1.0),
+        int((grid >> 8) & 0xFF) * (sign1 ? -1.0 : 1.0),
+        int((grid >> 16) & 0xFF) * (sign2 ? -1.0 : 1.0),
+        int((grid >> 24) & 0xFF) * (sign3 ? -1.0 : 1.0)
+    );
+}
+#endif
+
 #if defined(DATA_A_IQ4_NL)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
@@ -105,7 +321,7 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif
 
-#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(float(data_a[a_offset + ib].d), 0);
 }

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
index 175e31fa76e4ba8afe471cc73b981570326576bf..974efd3f9a693a59cd126f9eb440d00588c5dc18 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
@@ -301,6 +301,160 @@ float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2
     return ret;
 }
 
+#if defined(DATA_A_IQ2_XXS)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS {
+   block_iq2_xxs block;
+};
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS_packed16 {
+   block_iq2_xxs_packed16 block;
+};
+
+float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    decodeBufIQ2_XXS_packed16 bl16 = decodeBufIQ2_XXS_packed16(bl);
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+
+    const uint ib32 = (idx & 0xE0) >> 5; // 0..7
+    const uint ib8 = (idx & 0x18) >> 3;  // 0..3
+    const uint iqs = 8 * ib32 + ib8;
+
+    const uint8_t qs = bl.block.qs[iqs];
+    const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3]));
+
+    const float16_t dscale = bl.block.d * 0.25hf * (0.5hf + float16_t(signscale >> 28));
+    uint sign = bitfieldExtract(signscale, 7 * int(ib8), 7);
+    sign |= bitCount(sign) << 7;
+
+    const uint8_t g = unpack8(iq2xxs_grid[qs][(idx & 4) >> 2])[idx & 3];
+
+    float16_t ret = dscale * float16_t(g) * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
+
+    return ret;
+}
+#endif
+
+#if defined(DATA_A_IQ2_XS)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XS {
+   block_iq2_xs block;
+};
+
+float16_t dequantFuncIQ2_XS(const in decodeBufIQ2_XS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+
+    const uint is = (idx & 0xE0) >> 5;     // 0..8
+    const uint sshift = (idx & 0x10) >> 2; // 0,4
+    const uint iqs = (idx & 0xF8) >> 3;    // 0..63
+
+    const uint16_t qs = bl.block.qs[iqs];
+    const float16_t dscale = bl.block.d * 0.25hf * (0.5hf + float16_t((bl.block.scales[is] >> sshift) & 0xF));
+
+    uint sign = uint(qs >> 9);
+    sign |= bitCount(sign) << 7;
+    const uint8_t g = unpack8(iq2xs_grid[qs & 0x1FF][(idx & 4) >> 2])[idx & 3];
+
+    float16_t ret = dscale * float16_t(g) * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
+    return ret;
+}
+#endif
+
+#if defined(DATA_A_IQ2_S)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_S {
+   block_iq2_s block;
+};
+
+float16_t dequantFuncIQ2_S(const in decodeBufIQ2_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    uint idx = coordInBlock[1];
+    uint lsb = idx & 1;
+    idx /= 2;
+
+    const uint ib8 = (idx % 128) / 4; // 0..31
+    const uint ib32 = ib8 / 4;        // 0..7
+
+    const uint scale = (bl.block.scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
+    const uint qs = bl.block.qs[ib8];
+    const uint qh = bl.block.qh[ib32];
+    const uint qhshift = 2 * (ib8 % 4);
+    const uint sign = bl.block.qs[QUANT_K / 8 + ib8] >> (2 * (idx % 4));
+
+    const float d = float(bl.block.d);
+    const float db = d * 0.25 * (0.5 + scale);
+    const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
+    const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
+    const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
+    return float16_t(v[lsb]);
+}
+#endif
+
+#if defined(DATA_A_IQ3_XXS)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS {
+   block_iq3_xxs block;
+};
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS_packed16 {
+   block_iq3_xxs_packed16 block;
+};
+
+float16_t dequantFuncIQ3_XXS(const in decodeBufIQ3_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    uint idx = coordInBlock[1];
+    uint lsb = idx & 1;
+    idx /= 2;
+
+    const uint iqs = (idx % 128) / 2;           // 0..63
+    const uint is = QUANT_K / 4 + 4 * (iqs / 8); // 8 values
+
+    const float d = float(bl.block.d);
+    const uint qs = bl.block.qs[iqs];
+    const uint signs = pack32(u8vec4(
+        bl.block.qs[is+0],
+        bl.block.qs[is+1],
+        bl.block.qs[is+2],
+        bl.block.qs[is+3]
+    ));
+    const float db = d * 0.5 * (0.5 + (signs >> 28));
+    const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
+    const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
+    const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
+    const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
+    const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+    return float16_t(v[lsb]);
+}
+#endif
+
+#if defined(DATA_A_IQ3_S)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_S {
+   block_iq3_s block;
+};
+
+float16_t dequantFuncIQ3_S(const in decodeBufIQ3_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    uint idx = coordInBlock[1];
+    uint lsb = idx & 1;
+    idx /= 2;
+
+    const uint iqs = (idx % 128) / 2;           // 0..63
+    const uint iqh = iqs / 8;
+
+    const float d = float(bl.block.d);
+    const uint qs = bl.block.qs[iqs];
+    const uint qh = bl.block.qh[iqh];
+    const int8_t sign = int8_t(bl.block.signs[iqs / 2] >> (2 * (idx % 4)));
+    const uint scale = bl.block.scales[iqs / 16];
+    const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
+    const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
+    const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
+    const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+
+    return float16_t(v[lsb]);
+}
+#endif
+
+
 #if defined(DATA_A_IQ4_NL)
 layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ4_NL {
    block_iq4_nl block;
@@ -340,6 +494,16 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 #define dequantFuncA dequantFuncQ5_K
 #elif defined(DATA_A_Q6_K)
 #define dequantFuncA dequantFuncQ6_K
+#elif defined(DATA_A_IQ2_XXS)
+#define dequantFuncA dequantFuncIQ2_XXS
+#elif defined(DATA_A_IQ2_XS)
+#define dequantFuncA dequantFuncIQ2_XS
+#elif defined(DATA_A_IQ2_S)
+#define dequantFuncA dequantFuncIQ2_S
+#elif defined(DATA_A_IQ3_XXS)
+#define dequantFuncA dequantFuncIQ3_XXS
+#elif defined(DATA_A_IQ3_S)
+#define dequantFuncA dequantFuncIQ3_S
 #elif defined(DATA_A_IQ4_NL)
 #define dequantFuncA dequantFuncIQ4_NL
 #endif

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp
new file mode 100644 (file)
index 0000000..48f6b65
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp
@@ -0,0 +1,44 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq2_s data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 subblock (32 values with 2 scales)
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint ib32 = gl_LocalInvocationID.x % 8;
+    const uint b_idx = 256 * ib + 32 * ib32;
+
+    const float d = float(data_a[ib].d);
+    const vec2 scale = vec2(data_a[ib].scales[ib32] & 0xf, data_a[ib].scales[ib32] >> 4);
+    const vec2 db = d * (0.5 + scale) * 0.25;
+
+    uint qh = data_a[ib].qh[ib32];
+    [[unroll]] for (uint l = 0; l < 4; ++l) {
+        uint qs = data_a[ib].qs[4 * ib32 + l];
+        const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l];
+        qs |= (qh << (8 - 2 * l)) & 0x300;
+        const uvec2 grid = iq2s_grid[qs & 511];
+        const u8vec4 grid0 = unpack8(grid.x);
+        const u8vec4 grid1 = unpack8(grid.y);
+        data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 1] = D_TYPE(db[l/2] * grid0.y * ((sign & 2) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 2] = D_TYPE(db[l/2] * grid0.z * ((sign & 4) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 3] = D_TYPE(db[l/2] * grid0.w * ((sign & 8) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 4] = D_TYPE(db[l/2] * grid1.x * ((sign & 16) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 5] = D_TYPE(db[l/2] * grid1.y * ((sign & 32) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 6] = D_TYPE(db[l/2] * grid1.z * ((sign & 64) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 7] = D_TYPE(db[l/2] * grid1.w * ((sign & 128) != 0 ? -1.0 : 1.0));
+    }
+}

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp
new file mode 100644 (file)
index 0000000..a08331c
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp
@@ -0,0 +1,43 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq2_xs data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 subblock (32 values with 2 scales)
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint ib32 = gl_LocalInvocationID.x % 8;
+    const uint b_idx = 256 * ib + 32 * ib32;
+
+    const float d = float(data_a[ib].d);
+    const vec2 scale = vec2(data_a[ib].scales[ib32] & 0xf, data_a[ib].scales[ib32] >> 4);
+    const vec2 db = d * (0.5 + scale) * 0.25;
+
+    [[unroll]] for (uint l = 0; l < 4; ++l) {
+        uint16_t qs = data_a[ib].qs[4 * ib32 + l];
+        const uint sign7 = qs >> 9;
+        const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit
+        const uvec2 grid = iq2xs_grid[qs & 511];
+        const u8vec4 grid0 = unpack8(grid.x);
+        const u8vec4 grid1 = unpack8(grid.y);
+        data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 1] = D_TYPE(db[l/2] * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 2] = D_TYPE(db[l/2] * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 3] = D_TYPE(db[l/2] * grid0.w * ((sign8 & 8) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 4] = D_TYPE(db[l/2] * grid1.x * ((sign8 & 16) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 5] = D_TYPE(db[l/2] * grid1.y * ((sign8 & 32) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 6] = D_TYPE(db[l/2] * grid1.z * ((sign8 & 64) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 7] = D_TYPE(db[l/2] * grid1.w * ((sign8 & 128) != 0 ? -1.0 : 1.0));
+    }
+}

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp
new file mode 100644 (file)
index 0000000..e370690
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp
@@ -0,0 +1,48 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq2_xxs data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 scale block (32 values)
+    // Each block is described by 4 lattice indices, 4x7 sign bits and 4 scale bits
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint is = gl_LocalInvocationID.x % 8;
+    const uint b_idx = 256 * ib + 32 * is;
+
+    const float d = float(data_a[ib].d);
+    uint signscale = pack32(u8vec4(
+        data_a[ib].qs[8*is + 4],
+        data_a[ib].qs[8*is + 5],
+        data_a[ib].qs[8*is + 6],
+        data_a[ib].qs[8*is + 7]
+    ));
+    const float db = d * (0.5 + (signscale >> 28)) * 0.25;
+
+    [[unroll]] for (uint l = 0; l < 4; ++l) {
+        const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
+        const uint sign8 = sign7 | (bitCount(sign7) << 7); // parity bit
+        const uvec2 grid = iq2xxs_grid[data_a[ib].qs[8 * is + l]];
+        const u8vec4 grid0 = unpack8(grid.x);
+        const u8vec4 grid1 = unpack8(grid.y);
+        data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 3] = D_TYPE(db * grid0.w * ((sign8 & 8) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 4] = D_TYPE(db * grid1.x * ((sign8 & 16) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 5] = D_TYPE(db * grid1.y * ((sign8 & 32) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 6] = D_TYPE(db * grid1.z * ((sign8 & 64) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 7] = D_TYPE(db * grid1.w * ((sign8 & 128) != 0 ? -1.0 : 1.0));
+    }
+}

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp
new file mode 100644 (file)
index 0000000..c3f4bca
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp
@@ -0,0 +1,39 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq3_s data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 scale nibble.
+    // Each block contains 4 scale bytes (8 scales) for 256 output values.
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint is = gl_LocalInvocationID.x % 8;
+    const uint b_idx = 256 * ib + 32 * is;
+
+    const float d = float(data_a[ib].d);
+    const float db = d * (1 + 2 * ((data_a[ib].scales[is] >> (4 * (is % 2))) & 0xf));
+
+    // We must produce 32 values using 4 sign bytes, 1 qh byte, 8 qs bytes.
+    uint qh = data_a[ib].qh[is];
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        uint qs = data_a[ib].qs[8 * is + l];
+        uint gidx = qs | ((qh << (8 - l)) & 256);
+        uint8_t signs = data_a[ib].signs[8 * is + l / 2] >> (4 * (l & 1));
+        u8vec4 grid = unpack8(iq3s_grid[gidx]);
+        data_b[b_idx + 4 * l + 0] = D_TYPE(db * grid.x * ((signs & 1) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 4 * l + 1] = D_TYPE(db * grid.y * ((signs & 2) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 4 * l + 2] = D_TYPE(db * grid.z * ((signs & 4) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 4 * l + 3] = D_TYPE(db * grid.w * ((signs & 8) != 0 ? -1.0 : 1.0));
+    }
+}

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp
new file mode 100644 (file)
index 0000000..a92b829
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp
@@ -0,0 +1,49 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq3_xxs data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 scale block (32 values)
+    // 8 threads handle 1 superblock
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint is = gl_LocalInvocationID.x % 8;
+    const uint b_idx = 256 * ib + 32 * is;
+    const uint s_idx = QUANT_K / 4 + 4 * is;
+
+    const float d = float(data_a[ib].d);
+    uint signscale = pack32(u8vec4(
+        data_a[ib].qs[s_idx + 0],
+        data_a[ib].qs[s_idx + 1],
+        data_a[ib].qs[s_idx + 2],
+        data_a[ib].qs[s_idx + 3]
+    ));
+    const float db = d * (0.5 + (signscale >> 28)) * 0.5;
+
+    [[unroll]] for (uint l = 0; l < 4; ++l) {
+        const uint sign7 = bitfieldExtract(signscale, 7 * int(l), 7);
+        // Restore parity bit.
+        const uint sign8 = sign7 | (bitCount(sign7) << 7);
+        const u8vec4 grid0 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l]]);
+        const u8vec4 grid1 = unpack8(iq3xxs_grid[data_a[ib].qs[8 * is + 2 * l + 1]]);
+        data_b[b_idx + 8 * l + 0] = D_TYPE(db * grid0.x * ((sign8 & 1) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 1] = D_TYPE(db * grid0.y * ((sign8 & 2) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 2] = D_TYPE(db * grid0.z * ((sign8 & 4) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 3] = D_TYPE(db * grid0.w * ((sign8 & 8) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 4] = D_TYPE(db * grid1.x * ((sign8 & 16) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 5] = D_TYPE(db * grid1.y * ((sign8 & 32) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 6] = D_TYPE(db * grid1.z * ((sign8 & 64) != 0 ? -1.0 : 1.0));
+        data_b[b_idx + 8 * l + 7] = D_TYPE(db * grid1.w * ((sign8 & 128) != 0 ? -1.0 : 1.0));
+    }
+}

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
index 8de14fc03f102957c1538d448791148b37bf298c..46d9ad15ebafc5f4c045feb7b0094e6eb58bb1e6 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
 
-    init_iq4nl_shmem();
+    init_iq_shmem(gl_WorkGroupSize);
 
     const uint tid = gl_LocalInvocationID.x % 64;
     const uint il  = tid/32;

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
index 3735d0dbbae1d71bb7bd1d2794c3410af23c2976..043a53023889c4f8da84cb98461b5164c2783101 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -104,8 +104,8 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 #endif
 
 void main() {
-#if defined(DATA_A_IQ4_NL)
-    init_iq4nl_shmem();
+#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
+    init_iq_shmem(gl_WorkGroupSize);
 #endif
 
     const uint32_t N = p.N;

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
index 1426fde6597b679a4ebfebc9a2b84bd28d4ae17c..09dc43d8dc3b86d92237fb7cdfc91357a47e8eaa 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
@@ -12,8 +12,8 @@ void main() {
     const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
     const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
-#if defined(DATA_A_IQ4_NL)
-    init_iq4nl_shmem();
+#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
+    init_iq_shmem(gl_WorkGroupSize);
 #endif
 
     if (i00 >= p.ne00) {

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
index 53902858de7dad5a887b27e16c553625ebd5a8fc..48156e7bab6cdd3a2e0c93f6f4c7fdac5636573c 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
@@ -133,8 +133,8 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 void main() {
     const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
 
-#if defined(DATA_A_IQ4_NL)
-    init_iq4nl_shmem();
+#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
+    init_iq_shmem(gl_WorkGroupSize);
 #endif
 
     // do NUM_ROWS at a time, unless there aren't enough remaining rows

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index 48122cbef906ebb92e59322df8c602594541df3d..d0559aac8ec9228947c999e991612ce9a171aaff 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -95,8 +95,8 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
 
 void main() {
-#if defined(DATA_A_IQ4_NL)
-    init_iq4nl_shmem();
+#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
+    init_iq_shmem(gl_WorkGroupSize);
 #endif
 
 #ifdef MUL_MAT_ID
@@ -343,10 +343,8 @@ void main() {
             const uint qsshift = halfsplit * 2;          // 0,2,4,6
             const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
 
-            const int8_t us = int8_t(is <  4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) :
-                                    is <  8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) :
-                                    is < 12 ? (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) :
-                                            (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4));
+            const int8_t us = int8_t(((data_a[ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF)
+                                  | (((data_a[ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
             const float dl = float(data_a[ib].d) * float(us - 32);
 
             buf_a[buf_idx    ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)));
@@ -439,6 +437,118 @@ void main() {
 
             buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
             buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+#elif defined(DATA_A_IQ2_XXS)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint ib32 = (idx % 128) / 16;         // 0..7
+            const uint ib8 = (idx / 4) % 4;
+
+            const float d = float(data_a[ib].d);
+            const uint qs = data_a[ib].qs[8 * ib32 + ib8];
+            const uint signs = pack32(u8vec4(
+                data_a[ib].qs[8*ib32 + 4],
+                data_a[ib].qs[8*ib32 + 5],
+                data_a[ib].qs[8*ib32 + 6],
+                data_a[ib].qs[8*ib32 + 7]
+            ));
+            const float db = d * 0.25 * (0.5 + (signs >> 28));
+            const uint32_t sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
+            const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
+            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
+            const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1));
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_IQ2_XS)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint ib32 = (idx % 128) / 16;         // 0..7
+            const uint ib8 = (idx / 4) % 4;             // 0..3
+
+            const float d = float(data_a[ib].d);
+            const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
+            const float db = d * 0.25 * (0.5 + scale);
+            const uint qs = data_a[ib].qs[4 * ib32 + ib8];
+            const uint sign7 = qs >> 9;
+            const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
+            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
+            const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1));
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_IQ2_S)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;        // 2 values per idx
+            const uint ib8 = (idx % 128) / 4; // 0..31
+            const uint ib32 = ib8 / 4;        // 0..7
+
+            const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
+            const uint qs = data_a[ib].qs[ib8];
+            const uint qh = data_a[ib].qh[ib32];
+            const uint qhshift = 2 * (ib8 % 4);
+            const uint sign = data_a[ib].qs[QUANT_K / 8 + ib8] >> (2 * (idx % 4));
+
+            const float d = float(data_a[ib].d);
+            const float db = d * 0.25 * (0.5 + scale);
+            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
+            const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_IQ3_XXS)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint iqs = (idx % 128) / 2;           // 0..63
+            const uint is = QUANT_K / 4 + 4 * (iqs / 8); // 8 values
+
+            const float d = float(data_a[ib].d);
+            const uint qs = data_a[ib].qs[iqs];
+            const uint signs = pack32(u8vec4(
+                data_a[ib].qs[is+0],
+                data_a[ib].qs[is+1],
+                data_a[ib].qs[is+2],
+                data_a[ib].qs[is+3]
+            ));
+            const float db = d * 0.5 * (0.5 + (signs >> 28));
+            const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
+            const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
+            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
+            const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_IQ3_S)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint iqs = (idx % 128) / 2;           // 0..63
+            const uint iqh = iqs / 8;
+
+            const float d = float(data_a[ib].d);
+            const uint qs = data_a[ib].qs[iqs];
+            const uint qh = data_a[ib].qh[iqh];
+            const int8_t sign = int8_t(data_a[ib].signs[iqs / 2] >> (2 * (idx % 4)));
+            const uint scale = data_a[ib].scales[iqs / 16];
+            const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
+            const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
+            const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
+            const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_IQ4_NL)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
             const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a;

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
index 57f9e724513608a94a814c8544f0f7567205a192..27c5d68b3d94ef696cdd17d0d91541c4ec6df0fe 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
@@ -106,8 +106,8 @@ D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem
 #endif
 
 void main() {
-#if defined(DATA_A_IQ4_NL)
-    init_iq4nl_shmem();
+#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
+    init_iq_shmem(gl_WorkGroupSize);
 #endif
 
 #ifdef MUL_MAT_ID

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
index 1e35b665283d8ddc41e07b5a3206f77de93bb219..9e56a35300bafc8eb402973d316445fb9a9f013a 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
@@ -294,6 +294,738 @@ struct block_q6_K_packed16
 
 // IQuants
 
+#define QUANT_K_IQ2_XXS 256
+#define QUANT_R_IQ2_XXS 1
+
+struct block_iq2_xxs
+{
+    float16_t d;
+    uint8_t qs[QUANT_K_IQ2_XXS/4];
+};
+
+struct block_iq2_xxs_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K_IQ2_XXS/8];
+};
+
+#if defined(DATA_A_IQ2_XXS)
+
+const uvec2[256] iq2xxs_grid_const = {
+    uvec2(0x08080808, 0x08080808), uvec2(0x0808082b, 0x08080808), uvec2(0x08081919, 0x08080808), uvec2(0x08082b08, 0x08080808),
+    uvec2(0x08082b2b, 0x08080808), uvec2(0x08190819, 0x08080808), uvec2(0x08191908, 0x08080808), uvec2(0x082b0808, 0x08080808),
+    uvec2(0x082b082b, 0x08080808), uvec2(0x082b2b08, 0x08080808), uvec2(0x082b2b2b, 0x08080808), uvec2(0x19080819, 0x08080808),
+    uvec2(0x19081908, 0x08080808), uvec2(0x19190808, 0x08080808), uvec2(0x19192b08, 0x08080808), uvec2(0x192b0819, 0x08080808),
+    uvec2(0x192b1908, 0x08080808), uvec2(0x2b080808, 0x08080808), uvec2(0x2b08082b, 0x08080808), uvec2(0x2b082b2b, 0x08080808),
+    uvec2(0x2b2b082b, 0x08080808), uvec2(0x08080819, 0x08080819), uvec2(0x08081908, 0x08080819), uvec2(0x08190808, 0x08080819),
+    uvec2(0x08191919, 0x08080819), uvec2(0x19080808, 0x08080819), uvec2(0x2b081908, 0x08080819), uvec2(0x2b192b08, 0x08080819),
+    uvec2(0x08080808, 0x0808082b), uvec2(0x0808082b, 0x0808082b), uvec2(0x082b082b, 0x0808082b), uvec2(0x2b08082b, 0x0808082b),
+    uvec2(0x08080819, 0x08081908), uvec2(0x08081908, 0x08081908), uvec2(0x08190808, 0x08081908), uvec2(0x082b0819, 0x08081908),
+    uvec2(0x082b1908, 0x08081908), uvec2(0x19080808, 0x08081908), uvec2(0x1908082b, 0x08081908), uvec2(0x19082b08, 0x08081908),
+    uvec2(0x192b0808, 0x08081908), uvec2(0x2b080819, 0x08081908), uvec2(0x2b081908, 0x08081908), uvec2(0x2b190808, 0x08081908),
+    uvec2(0x2b2b1908, 0x08081908), uvec2(0x08080808, 0x08081919), uvec2(0x0808082b, 0x08081919), uvec2(0x08082b08, 0x08081919),
+    uvec2(0x082b0808, 0x08081919), uvec2(0x1908192b, 0x08081919), uvec2(0x192b2b19, 0x08081919), uvec2(0x2b080808, 0x08081919),
+    uvec2(0x2b190819, 0x08081919), uvec2(0x08082b19, 0x0808192b), uvec2(0x08190808, 0x0808192b), uvec2(0x19080808, 0x0808192b),
+    uvec2(0x2b081908, 0x0808192b), uvec2(0x2b2b1908, 0x0808192b), uvec2(0x08080808, 0x08082b08), uvec2(0x08081919, 0x08082b08),
+    uvec2(0x08082b08, 0x08082b08), uvec2(0x08191908, 0x08082b08), uvec2(0x082b2b08, 0x08082b08), uvec2(0x19080819, 0x08082b08),
+    uvec2(0x19081908, 0x08082b08), uvec2(0x19190808, 0x08082b08), uvec2(0x1919082b, 0x08082b08), uvec2(0x2b082b08, 0x08082b08),
+    uvec2(0x08081908, 0x08082b19), uvec2(0x19080808, 0x08082b19), uvec2(0x0808082b, 0x08082b2b), uvec2(0x08191908, 0x08082b2b),
+    uvec2(0x08080819, 0x08190808), uvec2(0x08081908, 0x08190808), uvec2(0x08190808, 0x08190808), uvec2(0x082b0819, 0x08190808),
+    uvec2(0x19080808, 0x08190808), uvec2(0x192b0808, 0x08190808), uvec2(0x2b081908, 0x08190808), uvec2(0x2b190808, 0x08190808),
+    uvec2(0x2b191919, 0x08190808), uvec2(0x08080808, 0x08190819), uvec2(0x08082b08, 0x08190819), uvec2(0x082b0808, 0x08190819),
+    uvec2(0x19190808, 0x08190819), uvec2(0x19192b2b, 0x08190819), uvec2(0x2b080808, 0x08190819), uvec2(0x082b1908, 0x0819082b),
+    uvec2(0x19081919, 0x0819082b), uvec2(0x08080808, 0x08191908), uvec2(0x08082b08, 0x08191908), uvec2(0x082b0808, 0x08191908),
+    uvec2(0x082b1919, 0x08191908), uvec2(0x19082b19, 0x08191908), uvec2(0x2b080808, 0x08191908), uvec2(0x08192b08, 0x08191919),
+    uvec2(0x192b082b, 0x08191919), uvec2(0x08080808, 0x0819192b), uvec2(0x0819192b, 0x0819192b), uvec2(0x08080819, 0x08192b08),
+    uvec2(0x08081908, 0x08192b08), uvec2(0x08190808, 0x08192b08), uvec2(0x19080808, 0x08192b08), uvec2(0x2b080819, 0x08192b08),
+    uvec2(0x08080808, 0x08192b19), uvec2(0x08081919, 0x08192b19), uvec2(0x2b2b0808, 0x08192b19), uvec2(0x19190819, 0x08192b2b),
+    uvec2(0x08080808, 0x082b0808), uvec2(0x0808082b, 0x082b0808), uvec2(0x08082b2b, 0x082b0808), uvec2(0x19081908, 0x082b0808),
+    uvec2(0x192b0819, 0x082b0808), uvec2(0x2b080808, 0x082b0808), uvec2(0x2b08082b, 0x082b0808), uvec2(0x082b2b19, 0x082b0819),
+    uvec2(0x19082b08, 0x082b0819), uvec2(0x08080808, 0x082b082b), uvec2(0x0808082b, 0x082b082b), uvec2(0x08080819, 0x082b1908),
+    uvec2(0x08081908, 0x082b1908), uvec2(0x08190808, 0x082b1908), uvec2(0x19080808, 0x082b1908), uvec2(0x1919192b, 0x082b1908),
+    uvec2(0x08080808, 0x082b1919), uvec2(0x19080819, 0x082b1919), uvec2(0x192b1908, 0x082b1919), uvec2(0x2b190808, 0x082b192b),
+    uvec2(0x08082b08, 0x082b2b08), uvec2(0x082b0808, 0x082b2b08), uvec2(0x2b191908, 0x082b2b08), uvec2(0x19081908, 0x082b2b2b),
+    uvec2(0x08080819, 0x19080808), uvec2(0x08081908, 0x19080808), uvec2(0x08190808, 0x19080808), uvec2(0x08192b08, 0x19080808),
+    uvec2(0x082b0819, 0x19080808), uvec2(0x082b1908, 0x19080808), uvec2(0x19080808, 0x19080808), uvec2(0x19082b08, 0x19080808),
+    uvec2(0x1919192b, 0x19080808), uvec2(0x192b0808, 0x19080808), uvec2(0x2b080819, 0x19080808), uvec2(0x2b081908, 0x19080808),
+    uvec2(0x2b190808, 0x19080808), uvec2(0x08080808, 0x19080819), uvec2(0x082b0808, 0x19080819), uvec2(0x192b0819, 0x19080819),
+    uvec2(0x2b080808, 0x19080819), uvec2(0x2b081919, 0x19080819), uvec2(0x08080819, 0x1908082b), uvec2(0x08190808, 0x1908082b),
+    uvec2(0x19082b08, 0x1908082b), uvec2(0x1919192b, 0x1908082b), uvec2(0x192b2b08, 0x1908082b), uvec2(0x08080808, 0x19081908),
+    uvec2(0x08082b08, 0x19081908), uvec2(0x082b0808, 0x19081908), uvec2(0x2b080808, 0x19081908), uvec2(0x2b192b19, 0x19081908),
+    uvec2(0x0819082b, 0x19081919), uvec2(0x082b1908, 0x19081919), uvec2(0x08080808, 0x1908192b), uvec2(0x08080819, 0x19082b08),
+    uvec2(0x08081908, 0x19082b08), uvec2(0x08190808, 0x19082b08), uvec2(0x19080808, 0x19082b08), uvec2(0x19081919, 0x19082b08),
+    uvec2(0x08080808, 0x19082b19), uvec2(0x19192b08, 0x19082b19), uvec2(0x192b0819, 0x19082b19), uvec2(0x2b08082b, 0x19082b19),
+    uvec2(0x19081919, 0x19082b2b), uvec2(0x2b190808, 0x19082b2b), uvec2(0x08080808, 0x19190808), uvec2(0x08082b08, 0x19190808),
+    uvec2(0x08190819, 0x19190808), uvec2(0x08192b19, 0x19190808), uvec2(0x082b0808, 0x19190808), uvec2(0x2b080808, 0x19190808),
+    uvec2(0x2b082b08, 0x19190808), uvec2(0x08081908, 0x19190819), uvec2(0x1908082b, 0x19190819), uvec2(0x2b2b1908, 0x19190819),
+    uvec2(0x2b190819, 0x1919082b), uvec2(0x2b190808, 0x19191908), uvec2(0x2b19082b, 0x19191908), uvec2(0x08082b2b, 0x19191919),
+    uvec2(0x08080819, 0x1919192b), uvec2(0x19191908, 0x1919192b), uvec2(0x08080808, 0x19192b08), uvec2(0x08190819, 0x19192b08),
+    uvec2(0x08192b19, 0x19192b08), uvec2(0x192b1908, 0x19192b08), uvec2(0x19080808, 0x19192b19), uvec2(0x08082b08, 0x19192b2b),
+    uvec2(0x08081908, 0x192b0808), uvec2(0x08190808, 0x192b0808), uvec2(0x19080808, 0x192b0808), uvec2(0x192b2b08, 0x192b0808),
+    uvec2(0x08080808, 0x192b0819), uvec2(0x19191919, 0x192b0819), uvec2(0x08192b08, 0x192b082b), uvec2(0x192b0808, 0x192b082b),
+    uvec2(0x08080808, 0x192b1908), uvec2(0x08081919, 0x192b1908), uvec2(0x08190808, 0x192b1919), uvec2(0x0819082b, 0x192b1919),
+    uvec2(0x2b081908, 0x192b1919), uvec2(0x1908082b, 0x192b2b08), uvec2(0x08080808, 0x2b080808), uvec2(0x0808082b, 0x2b080808),
+    uvec2(0x08082b2b, 0x2b080808), uvec2(0x19080819, 0x2b080808), uvec2(0x2b08082b, 0x2b080808), uvec2(0x08081908, 0x2b080819),
+    uvec2(0x08192b08, 0x2b080819), uvec2(0x19080808, 0x2b080819), uvec2(0x08190819, 0x2b08082b), uvec2(0x08080819, 0x2b081908),
+    uvec2(0x08081908, 0x2b081908), uvec2(0x08190808, 0x2b081908), uvec2(0x08191919, 0x2b081908), uvec2(0x19080808, 0x2b081908),
+    uvec2(0x192b0808, 0x2b081908), uvec2(0x08080808, 0x2b081919), uvec2(0x1908192b, 0x2b081919), uvec2(0x2b191908, 0x2b081919),
+    uvec2(0x08082b19, 0x2b08192b), uvec2(0x19080808, 0x2b08192b), uvec2(0x192b0808, 0x2b08192b), uvec2(0x0808082b, 0x2b082b08),
+    uvec2(0x08081908, 0x2b082b19), uvec2(0x08190819, 0x2b082b2b), uvec2(0x08081908, 0x2b190808), uvec2(0x08190808, 0x2b190808),
+    uvec2(0x082b1908, 0x2b190808), uvec2(0x19080808, 0x2b190808), uvec2(0x2b2b0819, 0x2b190808), uvec2(0x0819192b, 0x2b190819),
+    uvec2(0x2b080808, 0x2b190819), uvec2(0x19081919, 0x2b19082b), uvec2(0x08080808, 0x2b191908), uvec2(0x082b082b, 0x2b191908),
+    uvec2(0x19081908, 0x2b191908), uvec2(0x19190819, 0x2b191919), uvec2(0x2b080819, 0x2b192b08), uvec2(0x082b0808, 0x2b192b19),
+    uvec2(0x0808082b, 0x2b2b0808), uvec2(0x19190808, 0x2b2b0808), uvec2(0x2b081919, 0x2b2b0808), uvec2(0x08082b19, 0x2b2b0819),
+    uvec2(0x08080808, 0x2b2b082b), uvec2(0x08192b08, 0x2b2b1908), uvec2(0x19190808, 0x2b2b2b08), uvec2(0x08081908, 0x2b2b2b19)
+};
+
+shared uvec2 iq2xxs_grid[256];
+
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < iq2xxs_grid.length(); i += wgsize.x) {
+        iq2xxs_grid[i] = iq2xxs_grid_const[i];
+    }
+    barrier();
+}
+
+#define QUANT_K QUANT_K_IQ2_XXS
+#define QUANT_R QUANT_R_IQ2_XXS
+#define A_TYPE block_iq2_xxs
+#define A_TYPE_PACKED16 block_iq2_xxs_packed16
+#endif
+
+#define QUANT_K_IQ2_XS 256
+#define QUANT_R_IQ2_XS 1
+
+struct block_iq2_xs
+{
+    float16_t d;
+    uint16_t qs[QUANT_K_IQ2_XS/8];
+    uint8_t scales[QUANT_K_IQ2_XS/32];
+};
+
+struct block_iq2_xs_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K_IQ2_XS/8];
+    uint16_t scales[QUANT_K_IQ2_XS/64];
+};
+
+#if defined(DATA_A_IQ2_XS)
+
+const uvec2 iq2xs_grid_const[512] = {
+    uvec2(0x08080808, 0x08080808), uvec2(0x0808082b, 0x08080808), uvec2(0x08081919, 0x08080808), uvec2(0x08082b08, 0x08080808),
+    uvec2(0x08082b2b, 0x08080808), uvec2(0x08190819, 0x08080808), uvec2(0x08191908, 0x08080808), uvec2(0x0819192b, 0x08080808),
+    uvec2(0x08192b19, 0x08080808), uvec2(0x082b0808, 0x08080808), uvec2(0x082b082b, 0x08080808), uvec2(0x082b1919, 0x08080808),
+    uvec2(0x082b2b08, 0x08080808), uvec2(0x19080819, 0x08080808), uvec2(0x19081908, 0x08080808), uvec2(0x1908192b, 0x08080808),
+    uvec2(0x19082b19, 0x08080808), uvec2(0x19190808, 0x08080808), uvec2(0x1919082b, 0x08080808), uvec2(0x19191919, 0x08080808),
+    uvec2(0x19192b08, 0x08080808), uvec2(0x192b0819, 0x08080808), uvec2(0x192b1908, 0x08080808), uvec2(0x2b080808, 0x08080808),
+    uvec2(0x2b08082b, 0x08080808), uvec2(0x2b081919, 0x08080808), uvec2(0x2b082b08, 0x08080808), uvec2(0x2b190819, 0x08080808),
+    uvec2(0x2b191908, 0x08080808), uvec2(0x2b192b19, 0x08080808), uvec2(0x2b2b0808, 0x08080808), uvec2(0x08080819, 0x08080819),
+    uvec2(0x08081908, 0x08080819), uvec2(0x0808192b, 0x08080819), uvec2(0x08082b19, 0x08080819), uvec2(0x08190808, 0x08080819),
+    uvec2(0x0819082b, 0x08080819), uvec2(0x08191919, 0x08080819), uvec2(0x08192b08, 0x08080819), uvec2(0x08192b2b, 0x08080819),
+    uvec2(0x082b0819, 0x08080819), uvec2(0x082b1908, 0x08080819), uvec2(0x19080808, 0x08080819), uvec2(0x1908082b, 0x08080819),
+    uvec2(0x19081919, 0x08080819), uvec2(0x19082b08, 0x08080819), uvec2(0x19190819, 0x08080819), uvec2(0x19191908, 0x08080819),
+    uvec2(0x192b0808, 0x08080819), uvec2(0x192b2b08, 0x08080819), uvec2(0x2b080819, 0x08080819), uvec2(0x2b081908, 0x08080819),
+    uvec2(0x2b190808, 0x08080819), uvec2(0x08080808, 0x0808082b), uvec2(0x0808082b, 0x0808082b), uvec2(0x08081919, 0x0808082b),
+    uvec2(0x08082b08, 0x0808082b), uvec2(0x08190819, 0x0808082b), uvec2(0x08191908, 0x0808082b), uvec2(0x082b0808, 0x0808082b),
+    uvec2(0x19080819, 0x0808082b), uvec2(0x19081908, 0x0808082b), uvec2(0x19190808, 0x0808082b), uvec2(0x19191919, 0x0808082b),
+    uvec2(0x2b080808, 0x0808082b), uvec2(0x2b082b2b, 0x0808082b), uvec2(0x08080819, 0x08081908), uvec2(0x08081908, 0x08081908),
+    uvec2(0x0808192b, 0x08081908), uvec2(0x08082b19, 0x08081908), uvec2(0x08190808, 0x08081908), uvec2(0x0819082b, 0x08081908),
+    uvec2(0x08191919, 0x08081908), uvec2(0x08192b08, 0x08081908), uvec2(0x082b0819, 0x08081908), uvec2(0x082b1908, 0x08081908),
+    uvec2(0x19080808, 0x08081908), uvec2(0x1908082b, 0x08081908), uvec2(0x19081919, 0x08081908), uvec2(0x19082b08, 0x08081908),
+    uvec2(0x19190819, 0x08081908), uvec2(0x19191908, 0x08081908), uvec2(0x1919192b, 0x08081908), uvec2(0x192b0808, 0x08081908),
+    uvec2(0x2b080819, 0x08081908), uvec2(0x2b081908, 0x08081908), uvec2(0x2b190808, 0x08081908), uvec2(0x08080808, 0x08081919),
+    uvec2(0x0808082b, 0x08081919), uvec2(0x08081919, 0x08081919), uvec2(0x08082b08, 0x08081919), uvec2(0x08190819, 0x08081919),
+    uvec2(0x08191908, 0x08081919), uvec2(0x082b0808, 0x08081919), uvec2(0x19080819, 0x08081919), uvec2(0x19081908, 0x08081919),
+    uvec2(0x19190808, 0x08081919), uvec2(0x192b0819, 0x08081919), uvec2(0x2b080808, 0x08081919), uvec2(0x08080819, 0x0808192b),
+    uvec2(0x08081908, 0x0808192b), uvec2(0x08190808, 0x0808192b), uvec2(0x082b192b, 0x0808192b), uvec2(0x19080808, 0x0808192b),
+    uvec2(0x1908082b, 0x0808192b), uvec2(0x2b081908, 0x0808192b), uvec2(0x08080808, 0x08082b08), uvec2(0x0808082b, 0x08082b08),
+    uvec2(0x08081919, 0x08082b08), uvec2(0x08082b08, 0x08082b08), uvec2(0x08082b2b, 0x08082b08), uvec2(0x08190819, 0x08082b08),
+    uvec2(0x08191908, 0x08082b08), uvec2(0x082b0808, 0x08082b08), uvec2(0x082b1919, 0x08082b08), uvec2(0x19080819, 0x08082b08),
+    uvec2(0x19081908, 0x08082b08), uvec2(0x19190808, 0x08082b08), uvec2(0x19192b08, 0x08082b08), uvec2(0x2b080808, 0x08082b08),
+    uvec2(0x2b2b0808, 0x08082b08), uvec2(0x2b2b2b2b, 0x08082b08), uvec2(0x08080819, 0x08082b19), uvec2(0x08081908, 0x08082b19),
+    uvec2(0x08190808, 0x08082b19), uvec2(0x19080808, 0x08082b19), uvec2(0x2b080819, 0x08082b19), uvec2(0x2b082b19, 0x08082b19),
+    uvec2(0x08080808, 0x08082b2b), uvec2(0x082b0808, 0x08082b2b), uvec2(0x082b2b08, 0x08082b2b), uvec2(0x2b19192b, 0x08082b2b),
+    uvec2(0x2b2b0808, 0x08082b2b), uvec2(0x08080819, 0x08190808), uvec2(0x08081908, 0x08190808), uvec2(0x0808192b, 0x08190808),
+    uvec2(0x08082b19, 0x08190808), uvec2(0x08190808, 0x08190808), uvec2(0x0819082b, 0x08190808), uvec2(0x08191919, 0x08190808),
+    uvec2(0x08192b08, 0x08190808), uvec2(0x082b0819, 0x08190808), uvec2(0x082b1908, 0x08190808), uvec2(0x19080808, 0x08190808),
+    uvec2(0x1908082b, 0x08190808), uvec2(0x19081919, 0x08190808), uvec2(0x19082b08, 0x08190808), uvec2(0x19190819, 0x08190808),
+    uvec2(0x19191908, 0x08190808), uvec2(0x192b0808, 0x08190808), uvec2(0x192b2b2b, 0x08190808), uvec2(0x2b080819, 0x08190808),
+    uvec2(0x2b081908, 0x08190808), uvec2(0x2b190808, 0x08190808), uvec2(0x08080808, 0x08190819), uvec2(0x0808082b, 0x08190819),
+    uvec2(0x08081919, 0x08190819), uvec2(0x08082b08, 0x08190819), uvec2(0x08190819, 0x08190819), uvec2(0x08191908, 0x08190819),
+    uvec2(0x082b0808, 0x08190819), uvec2(0x19080819, 0x08190819), uvec2(0x19081908, 0x08190819), uvec2(0x19190808, 0x08190819),
+    uvec2(0x2b080808, 0x08190819), uvec2(0x2b191908, 0x08190819), uvec2(0x2b19192b, 0x08190819), uvec2(0x08080819, 0x0819082b),
+    uvec2(0x08081908, 0x0819082b), uvec2(0x0808192b, 0x0819082b), uvec2(0x08190808, 0x0819082b), uvec2(0x19080808, 0x0819082b),
+    uvec2(0x192b0808, 0x0819082b), uvec2(0x08080808, 0x08191908), uvec2(0x0808082b, 0x08191908), uvec2(0x08081919, 0x08191908),
+    uvec2(0x08082b08, 0x08191908), uvec2(0x08190819, 0x08191908), uvec2(0x08191908, 0x08191908), uvec2(0x082b0808, 0x08191908),
+    uvec2(0x19080819, 0x08191908), uvec2(0x19081908, 0x08191908), uvec2(0x19082b19, 0x08191908), uvec2(0x19190808, 0x08191908),
+    uvec2(0x192b1908, 0x08191908), uvec2(0x2b080808, 0x08191908), uvec2(0x08080819, 0x08191919), uvec2(0x08081908, 0x08191919),
+    uvec2(0x08190808, 0x08191919), uvec2(0x19080808, 0x08191919), uvec2(0x08080808, 0x0819192b), uvec2(0x08191908, 0x0819192b),
+    uvec2(0x19082b19, 0x0819192b), uvec2(0x08080819, 0x08192b08), uvec2(0x08081908, 0x08192b08), uvec2(0x08190808, 0x08192b08),
+    uvec2(0x0819082b, 0x08192b08), uvec2(0x19080808, 0x08192b08), uvec2(0x19191908, 0x08192b08), uvec2(0x2b08192b, 0x08192b08),
+    uvec2(0x08080808, 0x08192b19), uvec2(0x08081919, 0x08192b19), uvec2(0x192b192b, 0x08192b19), uvec2(0x19190819, 0x08192b2b),
+    uvec2(0x2b2b2b19, 0x08192b2b), uvec2(0x08080808, 0x082b0808), uvec2(0x0808082b, 0x082b0808), uvec2(0x08081919, 0x082b0808),
+    uvec2(0x08082b08, 0x082b0808), uvec2(0x08082b2b, 0x082b0808), uvec2(0x08190819, 0x082b0808), uvec2(0x08191908, 0x082b0808),
+    uvec2(0x082b0808, 0x082b0808), uvec2(0x19080819, 0x082b0808), uvec2(0x19081908, 0x082b0808), uvec2(0x19190808, 0x082b0808),
+    uvec2(0x2b080808, 0x082b0808), uvec2(0x2b2b0808, 0x082b0808), uvec2(0x08080819, 0x082b0819), uvec2(0x08081908, 0x082b0819),
+    uvec2(0x08190808, 0x082b0819), uvec2(0x19080808, 0x082b0819), uvec2(0x19082b08, 0x082b0819), uvec2(0x192b1919, 0x082b0819),
+    uvec2(0x08080808, 0x082b082b), uvec2(0x082b082b, 0x082b082b), uvec2(0x2b080808, 0x082b082b), uvec2(0x2b2b2b08, 0x082b082b),
+    uvec2(0x08080819, 0x082b1908), uvec2(0x08081908, 0x082b1908), uvec2(0x08190808, 0x082b1908), uvec2(0x082b2b19, 0x082b1908),
+    uvec2(0x19080808, 0x082b1908), uvec2(0x08080808, 0x082b1919), uvec2(0x19080819, 0x082b1919), uvec2(0x1919082b, 0x082b1919),
+    uvec2(0x2b192b19, 0x082b1919), uvec2(0x08080819, 0x082b192b), uvec2(0x08192b2b, 0x082b192b), uvec2(0x2b2b192b, 0x082b192b),
+    uvec2(0x08080808, 0x082b2b08), uvec2(0x08082b08, 0x082b2b08), uvec2(0x08082b2b, 0x082b2b08), uvec2(0x082b0808, 0x082b2b08),
+    uvec2(0x19191919, 0x082b2b08), uvec2(0x2b082b08, 0x082b2b08), uvec2(0x2b2b082b, 0x082b2b08), uvec2(0x192b2b08, 0x082b2b19),
+    uvec2(0x2b190808, 0x082b2b19), uvec2(0x08082b08, 0x082b2b2b), uvec2(0x082b0808, 0x082b2b2b), uvec2(0x2b08082b, 0x082b2b2b),
+    uvec2(0x2b082b08, 0x082b2b2b), uvec2(0x2b082b2b, 0x082b2b2b), uvec2(0x08080819, 0x19080808), uvec2(0x08081908, 0x19080808),
+    uvec2(0x0808192b, 0x19080808), uvec2(0x08082b19, 0x19080808), uvec2(0x08190808, 0x19080808), uvec2(0x0819082b, 0x19080808),
+    uvec2(0x08191919, 0x19080808), uvec2(0x08192b08, 0x19080808), uvec2(0x082b0819, 0x19080808), uvec2(0x082b1908, 0x19080808),
+    uvec2(0x19080808, 0x19080808), uvec2(0x1908082b, 0x19080808), uvec2(0x19081919, 0x19080808), uvec2(0x19082b08, 0x19080808),
+    uvec2(0x19082b2b, 0x19080808), uvec2(0x19190819, 0x19080808), uvec2(0x19191908, 0x19080808), uvec2(0x192b0808, 0x19080808),
+    uvec2(0x192b1919, 0x19080808), uvec2(0x2b080819, 0x19080808), uvec2(0x2b081908, 0x19080808), uvec2(0x2b190808, 0x19080808),
+    uvec2(0x08080808, 0x19080819), uvec2(0x0808082b, 0x19080819), uvec2(0x08081919, 0x19080819), uvec2(0x08082b08, 0x19080819),
+    uvec2(0x08190819, 0x19080819), uvec2(0x08191908, 0x19080819), uvec2(0x082b0808, 0x19080819), uvec2(0x19080819, 0x19080819),
+    uvec2(0x19081908, 0x19080819), uvec2(0x19190808, 0x19080819), uvec2(0x2b080808, 0x19080819), uvec2(0x2b081919, 0x19080819),
+    uvec2(0x2b2b082b, 0x19080819), uvec2(0x08080819, 0x1908082b), uvec2(0x08081908, 0x1908082b), uvec2(0x08190808, 0x1908082b),
+    uvec2(0x0819082b, 0x1908082b), uvec2(0x082b2b19, 0x1908082b), uvec2(0x19080808, 0x1908082b), uvec2(0x08080808, 0x19081908),
+    uvec2(0x0808082b, 0x19081908), uvec2(0x08081919, 0x19081908), uvec2(0x08082b08, 0x19081908), uvec2(0x08190819, 0x19081908),
+    uvec2(0x08191908, 0x19081908), uvec2(0x08192b19, 0x19081908), uvec2(0x082b0808, 0x19081908), uvec2(0x19080819, 0x19081908),
+    uvec2(0x19081908, 0x19081908), uvec2(0x19190808, 0x19081908), uvec2(0x2b080808, 0x19081908), uvec2(0x2b191908, 0x19081908),
+    uvec2(0x08080819, 0x19081919), uvec2(0x08081908, 0x19081919), uvec2(0x08190808, 0x19081919), uvec2(0x082b1908, 0x19081919),
+    uvec2(0x19080808, 0x19081919), uvec2(0x2b192b2b, 0x19081919), uvec2(0x08080808, 0x1908192b), uvec2(0x08082b2b, 0x1908192b),
+    uvec2(0x19081908, 0x1908192b), uvec2(0x19190808, 0x1908192b), uvec2(0x08080819, 0x19082b08), uvec2(0x08081908, 0x19082b08),
+    uvec2(0x08190808, 0x19082b08), uvec2(0x19080808, 0x19082b08), uvec2(0x19081919, 0x19082b08), uvec2(0x19191908, 0x19082b08),
+    uvec2(0x192b082b, 0x19082b08), uvec2(0x08080808, 0x19082b19), uvec2(0x08190819, 0x19082b19), uvec2(0x19081908, 0x19082b19),
+    uvec2(0x19190808, 0x19082b19), uvec2(0x192b2b19, 0x19082b19), uvec2(0x08081908, 0x19082b2b), uvec2(0x08080808, 0x19190808),
+    uvec2(0x0808082b, 0x19190808), uvec2(0x08081919, 0x19190808), uvec2(0x08082b08, 0x19190808), uvec2(0x08190819, 0x19190808),
+    uvec2(0x08191908, 0x19190808), uvec2(0x082b0808, 0x19190808), uvec2(0x082b2b08, 0x19190808), uvec2(0x19080819, 0x19190808),
+    uvec2(0x19081908, 0x19190808), uvec2(0x19190808, 0x19190808), uvec2(0x2b080808, 0x19190808), uvec2(0x08080819, 0x19190819),
+    uvec2(0x08081908, 0x19190819), uvec2(0x08190808, 0x19190819), uvec2(0x08191919, 0x19190819), uvec2(0x19080808, 0x19190819),
+    uvec2(0x1908082b, 0x19190819), uvec2(0x08080808, 0x1919082b), uvec2(0x19081908, 0x1919082b), uvec2(0x2b2b2b2b, 0x1919082b),
+    uvec2(0x08080819, 0x19191908), uvec2(0x08081908, 0x19191908), uvec2(0x08190808, 0x19191908), uvec2(0x082b0819, 0x19191908),
+    uvec2(0x19080808, 0x19191908), uvec2(0x192b0808, 0x19191908), uvec2(0x2b080819, 0x19191908), uvec2(0x2b2b0819, 0x19191908),
+    uvec2(0x08080808, 0x19191919), uvec2(0x08082b08, 0x19191919), uvec2(0x2b080808, 0x19191919), uvec2(0x2b082b08, 0x19191919),
+    uvec2(0x082b0819, 0x1919192b), uvec2(0x192b2b08, 0x1919192b), uvec2(0x2b2b0819, 0x1919192b), uvec2(0x08080808, 0x19192b08),
+    uvec2(0x08191908, 0x19192b08), uvec2(0x19080819, 0x19192b08), uvec2(0x19190808, 0x19192b08), uvec2(0x2b192b19, 0x19192b08),
+    uvec2(0x08192b2b, 0x19192b19), uvec2(0x19080808, 0x19192b19), uvec2(0x1908082b, 0x19192b19), uvec2(0x2b081919, 0x19192b2b),
+    uvec2(0x08080819, 0x192b0808), uvec2(0x08081908, 0x192b0808), uvec2(0x08190808, 0x192b0808), uvec2(0x19080808, 0x192b0808),
+    uvec2(0x19191908, 0x192b0808), uvec2(0x192b082b, 0x192b0808), uvec2(0x2b08192b, 0x192b0808), uvec2(0x2b2b2b19, 0x192b0808),
+    uvec2(0x08080808, 0x192b0819), uvec2(0x082b1908, 0x192b082b), uvec2(0x19082b2b, 0x192b082b), uvec2(0x2b19082b, 0x192b082b),
+    uvec2(0x08080808, 0x192b1908), uvec2(0x0819192b, 0x192b1908), uvec2(0x08190808, 0x192b1919), uvec2(0x19080808, 0x192b1919),
+    uvec2(0x19081919, 0x192b1919), uvec2(0x2b2b1908, 0x192b1919), uvec2(0x08080819, 0x192b2b08), uvec2(0x192b2b2b, 0x192b2b08),
+    uvec2(0x082b1919, 0x192b2b19), uvec2(0x0808192b, 0x192b2b2b), uvec2(0x19191908, 0x192b2b2b), uvec2(0x192b082b, 0x192b2b2b),
+    uvec2(0x08080808, 0x2b080808), uvec2(0x0808082b, 0x2b080808), uvec2(0x08081919, 0x2b080808), uvec2(0x08082b08, 0x2b080808),
+    uvec2(0x08190819, 0x2b080808), uvec2(0x08191908, 0x2b080808), uvec2(0x082b0808, 0x2b080808), uvec2(0x082b2b2b, 0x2b080808),
+    uvec2(0x19080819, 0x2b080808), uvec2(0x19081908, 0x2b080808), uvec2(0x19190808, 0x2b080808), uvec2(0x2b080808, 0x2b080808),
+    uvec2(0x2b08082b, 0x2b080808), uvec2(0x2b2b2b08, 0x2b080808), uvec2(0x2b2b2b2b, 0x2b080808), uvec2(0x08080819, 0x2b080819),
+    uvec2(0x08081908, 0x2b080819), uvec2(0x0808192b, 0x2b080819), uvec2(0x08190808, 0x2b080819), uvec2(0x19080808, 0x2b080819),
+    uvec2(0x19190819, 0x2b080819), uvec2(0x19192b19, 0x2b080819), uvec2(0x08080808, 0x2b08082b), uvec2(0x082b0808, 0x2b08082b),
+    uvec2(0x2b080808, 0x2b08082b), uvec2(0x2b08082b, 0x2b08082b), uvec2(0x2b2b0808, 0x2b08082b), uvec2(0x2b2b2b08, 0x2b08082b),
+    uvec2(0x08080819, 0x2b081908), uvec2(0x08081908, 0x2b081908), uvec2(0x08190808, 0x2b081908), uvec2(0x0819082b, 0x2b081908),
+    uvec2(0x08191919, 0x2b081908), uvec2(0x19080808, 0x2b081908), uvec2(0x192b0808, 0x2b081908), uvec2(0x2b082b19, 0x2b081908),
+    uvec2(0x08080808, 0x2b081919), uvec2(0x19081908, 0x2b081919), uvec2(0x2b2b1919, 0x2b081919), uvec2(0x08192b08, 0x2b08192b),
+    uvec2(0x192b2b2b, 0x2b08192b), uvec2(0x08080808, 0x2b082b08), uvec2(0x08082b08, 0x2b082b08), uvec2(0x082b1919, 0x2b082b08),
+    uvec2(0x19192b2b, 0x2b082b08), uvec2(0x2b080808, 0x2b082b08), uvec2(0x2b08082b, 0x2b082b08), uvec2(0x2b2b2b08, 0x2b082b08),
+    uvec2(0x0808192b, 0x2b082b19), uvec2(0x082b082b, 0x2b082b2b), uvec2(0x2b080808, 0x2b082b2b), uvec2(0x2b082b08, 0x2b082b2b),
+    uvec2(0x2b19192b, 0x2b082b2b), uvec2(0x2b2b2b08, 0x2b082b2b), uvec2(0x08080819, 0x2b190808), uvec2(0x08081908, 0x2b190808),
+    uvec2(0x08190808, 0x2b190808), uvec2(0x19080808, 0x2b190808), uvec2(0x1919192b, 0x2b190808), uvec2(0x2b081908, 0x2b190808),
+    uvec2(0x08080808, 0x2b190819), uvec2(0x082b082b, 0x2b190819), uvec2(0x192b1908, 0x2b190819), uvec2(0x1919192b, 0x2b19082b),
+    uvec2(0x2b082b19, 0x2b19082b), uvec2(0x08080808, 0x2b191908), uvec2(0x08081919, 0x2b191908), uvec2(0x19081908, 0x2b191908),
+    uvec2(0x19190808, 0x2b191908), uvec2(0x19192b08, 0x2b191908), uvec2(0x082b2b19, 0x2b191919), uvec2(0x2b190808, 0x2b191919),
+    uvec2(0x2b19082b, 0x2b191919), uvec2(0x19080819, 0x2b19192b), uvec2(0x19190819, 0x2b192b08), uvec2(0x2b2b192b, 0x2b192b08),
+    uvec2(0x19082b19, 0x2b192b19), uvec2(0x08191919, 0x2b192b2b), uvec2(0x192b0808, 0x2b192b2b), uvec2(0x08080808, 0x2b2b0808),
+    uvec2(0x0808082b, 0x2b2b0808), uvec2(0x08082b08, 0x2b2b0808), uvec2(0x08082b2b, 0x2b2b0808), uvec2(0x082b0808, 0x2b2b0808),
+    uvec2(0x082b2b2b, 0x2b2b0808), uvec2(0x2b2b0808, 0x2b2b0808), uvec2(0x19190819, 0x2b2b0819), uvec2(0x19192b19, 0x2b2b0819),
+    uvec2(0x2b2b192b, 0x2b2b0819), uvec2(0x08080808, 0x2b2b082b), uvec2(0x0808082b, 0x2b2b082b), uvec2(0x08082b08, 0x2b2b082b),
+    uvec2(0x082b2b2b, 0x2b2b082b), uvec2(0x2b080808, 0x2b2b082b), uvec2(0x2b2b0808, 0x2b2b082b), uvec2(0x19080808, 0x2b2b1908),
+    uvec2(0x2b191919, 0x2b2b1908), uvec2(0x192b1919, 0x2b2b192b), uvec2(0x2b192b08, 0x2b2b192b), uvec2(0x08082b2b, 0x2b2b2b08),
+    uvec2(0x082b0808, 0x2b2b2b08), uvec2(0x082b082b, 0x2b2b2b08), uvec2(0x082b2b08, 0x2b2b2b08), uvec2(0x2b2b0808, 0x2b2b2b08),
+    uvec2(0x2b2b2b08, 0x2b2b2b08), uvec2(0x08081908, 0x2b2b2b19), uvec2(0x2b081908, 0x2b2b2b19), uvec2(0x2b08192b, 0x2b2b2b19),
+    uvec2(0x082b2b08, 0x2b2b2b2b), uvec2(0x082b2b2b, 0x2b2b2b2b), uvec2(0x2b190819, 0x2b2b2b2b), uvec2(0x2b2b2b2b, 0x2b2b2b2b),
+};
+
+shared uvec2 iq2xs_grid[512];
+
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < iq2xs_grid.length(); i += wgsize.x) {
+        iq2xs_grid[i] = iq2xs_grid_const[i];
+    }
+    barrier();
+}
+
+#define QUANT_K QUANT_K_IQ2_XS
+#define QUANT_R QUANT_R_IQ2_XS
+#define A_TYPE block_iq2_xs
+#define A_TYPE_PACKED16 block_iq2_xs_packed16
+#endif
+
+#define QUANT_K_IQ2_S 256
+#define QUANT_R_IQ2_S 1
+
+struct block_iq2_s
+{
+    float16_t d;
+    uint8_t qs[QUANT_K_IQ2_S/4];
+    uint8_t qh[QUANT_K_IQ2_S/32];
+    uint8_t scales[QUANT_K_IQ2_S/32];
+};
+
+#if defined(DATA_A_IQ2_S)
+
+const uvec2 iq2s_grid_const[1024] = {
+    uvec2(0x08080808, 0x08080808), uvec2(0x0808082b, 0x08080808), uvec2(0x08081919, 0x08080808), uvec2(0x08082b08, 0x08080808),
+    uvec2(0x08082b2b, 0x08080808), uvec2(0x08190819, 0x08080808), uvec2(0x08191908, 0x08080808), uvec2(0x0819192b, 0x08080808),
+    uvec2(0x08192b19, 0x08080808), uvec2(0x082b0808, 0x08080808), uvec2(0x082b082b, 0x08080808), uvec2(0x082b1919, 0x08080808),
+    uvec2(0x082b2b08, 0x08080808), uvec2(0x19080819, 0x08080808), uvec2(0x19081908, 0x08080808), uvec2(0x1908192b, 0x08080808),
+    uvec2(0x19082b19, 0x08080808), uvec2(0x19190808, 0x08080808), uvec2(0x1919082b, 0x08080808), uvec2(0x19191919, 0x08080808),
+    uvec2(0x19192b08, 0x08080808), uvec2(0x192b0819, 0x08080808), uvec2(0x192b1908, 0x08080808), uvec2(0x192b192b, 0x08080808),
+    uvec2(0x192b2b19, 0x08080808), uvec2(0x2b080808, 0x08080808), uvec2(0x2b08082b, 0x08080808), uvec2(0x2b081919, 0x08080808),
+    uvec2(0x2b082b08, 0x08080808), uvec2(0x2b190819, 0x08080808), uvec2(0x2b191908, 0x08080808), uvec2(0x2b2b0808, 0x08080808),
+    uvec2(0x2b2b1919, 0x08080808), uvec2(0x2b2b2b2b, 0x08080808), uvec2(0x08080819, 0x08080819), uvec2(0x08081908, 0x08080819),
+    uvec2(0x0808192b, 0x08080819), uvec2(0x08082b19, 0x08080819), uvec2(0x08190808, 0x08080819), uvec2(0x0819082b, 0x08080819),
+    uvec2(0x08191919, 0x08080819), uvec2(0x08192b08, 0x08080819), uvec2(0x082b0819, 0x08080819), uvec2(0x082b1908, 0x08080819),
+    uvec2(0x19080808, 0x08080819), uvec2(0x1908082b, 0x08080819), uvec2(0x19081919, 0x08080819), uvec2(0x19082b08, 0x08080819),
+    uvec2(0x19190819, 0x08080819), uvec2(0x19191908, 0x08080819), uvec2(0x1919192b, 0x08080819), uvec2(0x19192b19, 0x08080819),
+    uvec2(0x192b0808, 0x08080819), uvec2(0x192b1919, 0x08080819), uvec2(0x192b2b08, 0x08080819), uvec2(0x2b080819, 0x08080819),
+    uvec2(0x2b081908, 0x08080819), uvec2(0x2b190808, 0x08080819), uvec2(0x2b19082b, 0x08080819), uvec2(0x2b191919, 0x08080819),
+    uvec2(0x2b2b0819, 0x08080819), uvec2(0x2b2b1908, 0x08080819), uvec2(0x08080808, 0x0808082b), uvec2(0x0808082b, 0x0808082b),
+    uvec2(0x08081919, 0x0808082b), uvec2(0x08082b08, 0x0808082b), uvec2(0x08190819, 0x0808082b), uvec2(0x08191908, 0x0808082b),
+    uvec2(0x082b0808, 0x0808082b), uvec2(0x082b2b2b, 0x0808082b), uvec2(0x19080819, 0x0808082b), uvec2(0x19081908, 0x0808082b),
+    uvec2(0x1908192b, 0x0808082b), uvec2(0x19082b19, 0x0808082b), uvec2(0x19190808, 0x0808082b), uvec2(0x19191919, 0x0808082b),
+    uvec2(0x2b080808, 0x0808082b), uvec2(0x2b081919, 0x0808082b), uvec2(0x2b082b2b, 0x0808082b), uvec2(0x2b191908, 0x0808082b),
+    uvec2(0x2b2b082b, 0x0808082b), uvec2(0x08080819, 0x08081908), uvec2(0x08081908, 0x08081908), uvec2(0x0808192b, 0x08081908),
+    uvec2(0x08082b19, 0x08081908), uvec2(0x08190808, 0x08081908), uvec2(0x0819082b, 0x08081908), uvec2(0x08191919, 0x08081908),
+    uvec2(0x08192b08, 0x08081908), uvec2(0x082b0819, 0x08081908), uvec2(0x082b1908, 0x08081908), uvec2(0x082b192b, 0x08081908),
+    uvec2(0x082b2b19, 0x08081908), uvec2(0x19080808, 0x08081908), uvec2(0x1908082b, 0x08081908), uvec2(0x19081919, 0x08081908),
+    uvec2(0x19082b08, 0x08081908), uvec2(0x19082b2b, 0x08081908), uvec2(0x19190819, 0x08081908), uvec2(0x19191908, 0x08081908),
+    uvec2(0x1919192b, 0x08081908), uvec2(0x19192b19, 0x08081908), uvec2(0x192b0808, 0x08081908), uvec2(0x192b082b, 0x08081908),
+    uvec2(0x192b1919, 0x08081908), uvec2(0x2b080819, 0x08081908), uvec2(0x2b081908, 0x08081908), uvec2(0x2b08192b, 0x08081908),
+    uvec2(0x2b082b19, 0x08081908), uvec2(0x2b190808, 0x08081908), uvec2(0x2b191919, 0x08081908), uvec2(0x2b192b08, 0x08081908),
+    uvec2(0x2b2b0819, 0x08081908), uvec2(0x2b2b1908, 0x08081908), uvec2(0x08080808, 0x08081919), uvec2(0x0808082b, 0x08081919),
+    uvec2(0x08081919, 0x08081919), uvec2(0x08082b08, 0x08081919), uvec2(0x08082b2b, 0x08081919), uvec2(0x08190819, 0x08081919),
+    uvec2(0x08191908, 0x08081919), uvec2(0x0819192b, 0x08081919), uvec2(0x08192b19, 0x08081919), uvec2(0x082b0808, 0x08081919),
+    uvec2(0x082b1919, 0x08081919), uvec2(0x082b2b08, 0x08081919), uvec2(0x19080819, 0x08081919), uvec2(0x19081908, 0x08081919),
+    uvec2(0x1908192b, 0x08081919), uvec2(0x19082b19, 0x08081919), uvec2(0x19190808, 0x08081919), uvec2(0x1919082b, 0x08081919),
+    uvec2(0x19191919, 0x08081919), uvec2(0x19192b08, 0x08081919), uvec2(0x192b0819, 0x08081919), uvec2(0x192b1908, 0x08081919),
+    uvec2(0x2b080808, 0x08081919), uvec2(0x2b08082b, 0x08081919), uvec2(0x2b081919, 0x08081919), uvec2(0x2b082b08, 0x08081919),
+    uvec2(0x2b190819, 0x08081919), uvec2(0x2b191908, 0x08081919), uvec2(0x2b2b0808, 0x08081919), uvec2(0x08080819, 0x0808192b),
+    uvec2(0x08081908, 0x0808192b), uvec2(0x0808192b, 0x0808192b), uvec2(0x08082b19, 0x0808192b), uvec2(0x08190808, 0x0808192b),
+    uvec2(0x08191919, 0x0808192b), uvec2(0x19080808, 0x0808192b), uvec2(0x19081919, 0x0808192b), uvec2(0x19082b08, 0x0808192b),
+    uvec2(0x19190819, 0x0808192b), uvec2(0x19191908, 0x0808192b), uvec2(0x192b0808, 0x0808192b), uvec2(0x2b080819, 0x0808192b),
+    uvec2(0x2b081908, 0x0808192b), uvec2(0x2b190808, 0x0808192b), uvec2(0x08080808, 0x08082b08), uvec2(0x0808082b, 0x08082b08),
+    uvec2(0x08081919, 0x08082b08), uvec2(0x08082b08, 0x08082b08), uvec2(0x08190819, 0x08082b08), uvec2(0x08191908, 0x08082b08),
+    uvec2(0x0819192b, 0x08082b08), uvec2(0x08192b19, 0x08082b08), uvec2(0x082b0808, 0x08082b08), uvec2(0x082b1919, 0x08082b08),
+    uvec2(0x082b2b2b, 0x08082b08), uvec2(0x19080819, 0x08082b08), uvec2(0x19081908, 0x08082b08), uvec2(0x1908192b, 0x08082b08),
+    uvec2(0x19082b19, 0x08082b08), uvec2(0x19190808, 0x08082b08), uvec2(0x1919082b, 0x08082b08), uvec2(0x19191919, 0x08082b08),
+    uvec2(0x19192b08, 0x08082b08), uvec2(0x192b0819, 0x08082b08), uvec2(0x192b1908, 0x08082b08), uvec2(0x2b080808, 0x08082b08),
+    uvec2(0x2b081919, 0x08082b08), uvec2(0x2b191908, 0x08082b08), uvec2(0x2b2b2b2b, 0x08082b08), uvec2(0x08080819, 0x08082b19),
+    uvec2(0x08081908, 0x08082b19), uvec2(0x08190808, 0x08082b19), uvec2(0x0819082b, 0x08082b19), uvec2(0x08191919, 0x08082b19),
+    uvec2(0x08192b08, 0x08082b19), uvec2(0x082b0819, 0x08082b19), uvec2(0x19080808, 0x08082b19), uvec2(0x19081919, 0x08082b19),
+    uvec2(0x19082b08, 0x08082b19), uvec2(0x19190819, 0x08082b19), uvec2(0x19191908, 0x08082b19), uvec2(0x192b0808, 0x08082b19),
+    uvec2(0x2b080819, 0x08082b19), uvec2(0x2b190808, 0x08082b19), uvec2(0x08080808, 0x08082b2b), uvec2(0x08190819, 0x08082b2b),
+    uvec2(0x08191908, 0x08082b2b), uvec2(0x082b082b, 0x08082b2b), uvec2(0x082b2b08, 0x08082b2b), uvec2(0x082b2b2b, 0x08082b2b),
+    uvec2(0x19190808, 0x08082b2b), uvec2(0x2b192b19, 0x08082b2b), uvec2(0x08080819, 0x08190808), uvec2(0x08081908, 0x08190808),
+    uvec2(0x0808192b, 0x08190808), uvec2(0x08082b19, 0x08190808), uvec2(0x08190808, 0x08190808), uvec2(0x0819082b, 0x08190808),
+    uvec2(0x08191919, 0x08190808), uvec2(0x08192b08, 0x08190808), uvec2(0x082b0819, 0x08190808), uvec2(0x082b1908, 0x08190808),
+    uvec2(0x082b192b, 0x08190808), uvec2(0x19080808, 0x08190808), uvec2(0x1908082b, 0x08190808), uvec2(0x19081919, 0x08190808),
+    uvec2(0x19082b08, 0x08190808), uvec2(0x19190819, 0x08190808), uvec2(0x19191908, 0x08190808), uvec2(0x1919192b, 0x08190808),
+    uvec2(0x19192b19, 0x08190808), uvec2(0x192b0808, 0x08190808), uvec2(0x192b082b, 0x08190808), uvec2(0x192b1919, 0x08190808),
+    uvec2(0x192b2b08, 0x08190808), uvec2(0x2b080819, 0x08190808), uvec2(0x2b081908, 0x08190808), uvec2(0x2b08192b, 0x08190808),
+    uvec2(0x2b190808, 0x08190808), uvec2(0x2b191919, 0x08190808), uvec2(0x2b192b08, 0x08190808), uvec2(0x2b2b0819, 0x08190808),
+    uvec2(0x2b2b1908, 0x08190808), uvec2(0x08080808, 0x08190819), uvec2(0x0808082b, 0x08190819), uvec2(0x08081919, 0x08190819),
+    uvec2(0x08082b08, 0x08190819), uvec2(0x08082b2b, 0x08190819), uvec2(0x08190819, 0x08190819), uvec2(0x08191908, 0x08190819),
+    uvec2(0x0819192b, 0x08190819), uvec2(0x08192b19, 0x08190819), uvec2(0x082b0808, 0x08190819), uvec2(0x082b082b, 0x08190819),
+    uvec2(0x082b1919, 0x08190819), uvec2(0x082b2b08, 0x08190819), uvec2(0x19080819, 0x08190819), uvec2(0x19081908, 0x08190819),
+    uvec2(0x1908192b, 0x08190819), uvec2(0x19082b19, 0x08190819), uvec2(0x19190808, 0x08190819), uvec2(0x1919082b, 0x08190819),
+    uvec2(0x19191919, 0x08190819), uvec2(0x19192b08, 0x08190819), uvec2(0x192b0819, 0x08190819), uvec2(0x192b1908, 0x08190819),
+    uvec2(0x2b080808, 0x08190819), uvec2(0x2b08082b, 0x08190819), uvec2(0x2b081919, 0x08190819), uvec2(0x2b082b08, 0x08190819),
+    uvec2(0x2b190819, 0x08190819), uvec2(0x2b191908, 0x08190819), uvec2(0x08080819, 0x0819082b), uvec2(0x08081908, 0x0819082b),
+    uvec2(0x08082b19, 0x0819082b), uvec2(0x08190808, 0x0819082b), uvec2(0x08191919, 0x0819082b), uvec2(0x082b0819, 0x0819082b),
+    uvec2(0x082b1908, 0x0819082b), uvec2(0x19080808, 0x0819082b), uvec2(0x19081919, 0x0819082b), uvec2(0x19190819, 0x0819082b),
+    uvec2(0x19191908, 0x0819082b), uvec2(0x2b080819, 0x0819082b), uvec2(0x2b081908, 0x0819082b), uvec2(0x2b190808, 0x0819082b),
+    uvec2(0x08080808, 0x08191908), uvec2(0x0808082b, 0x08191908), uvec2(0x08081919, 0x08191908), uvec2(0x08082b08, 0x08191908),
+    uvec2(0x08190819, 0x08191908), uvec2(0x08191908, 0x08191908), uvec2(0x0819192b, 0x08191908), uvec2(0x08192b19, 0x08191908),
+    uvec2(0x082b0808, 0x08191908), uvec2(0x082b1919, 0x08191908), uvec2(0x082b2b08, 0x08191908), uvec2(0x19080819, 0x08191908),
+    uvec2(0x19081908, 0x08191908), uvec2(0x1908192b, 0x08191908), uvec2(0x19082b19, 0x08191908), uvec2(0x19190808, 0x08191908),
+    uvec2(0x1919082b, 0x08191908), uvec2(0x19191919, 0x08191908), uvec2(0x19192b08, 0x08191908), uvec2(0x192b0819, 0x08191908),
+    uvec2(0x192b1908, 0x08191908), uvec2(0x2b080808, 0x08191908), uvec2(0x2b08082b, 0x08191908), uvec2(0x2b081919, 0x08191908),
+    uvec2(0x2b082b08, 0x08191908), uvec2(0x2b190819, 0x08191908), uvec2(0x2b191908, 0x08191908), uvec2(0x2b2b0808, 0x08191908),
+    uvec2(0x08080819, 0x08191919), uvec2(0x08081908, 0x08191919), uvec2(0x0808192b, 0x08191919), uvec2(0x08082b19, 0x08191919),
+    uvec2(0x08190808, 0x08191919), uvec2(0x0819082b, 0x08191919), uvec2(0x08191919, 0x08191919), uvec2(0x08192b08, 0x08191919),
+    uvec2(0x082b0819, 0x08191919), uvec2(0x082b1908, 0x08191919), uvec2(0x19080808, 0x08191919), uvec2(0x1908082b, 0x08191919),
+    uvec2(0x19081919, 0x08191919), uvec2(0x19082b08, 0x08191919), uvec2(0x19190819, 0x08191919), uvec2(0x19191908, 0x08191919),
+    uvec2(0x192b0808, 0x08191919), uvec2(0x2b080819, 0x08191919), uvec2(0x2b081908, 0x08191919), uvec2(0x2b190808, 0x08191919),
+    uvec2(0x08080808, 0x0819192b), uvec2(0x08081919, 0x0819192b), uvec2(0x08082b08, 0x0819192b), uvec2(0x08190819, 0x0819192b),
+    uvec2(0x08191908, 0x0819192b), uvec2(0x082b0808, 0x0819192b), uvec2(0x19080819, 0x0819192b), uvec2(0x19081908, 0x0819192b),
+    uvec2(0x19190808, 0x0819192b), uvec2(0x2b080808, 0x0819192b), uvec2(0x2b2b2b2b, 0x0819192b), uvec2(0x08080819, 0x08192b08),
+    uvec2(0x08081908, 0x08192b08), uvec2(0x0808192b, 0x08192b08), uvec2(0x08082b19, 0x08192b08), uvec2(0x08190808, 0x08192b08),
+    uvec2(0x08191919, 0x08192b08), uvec2(0x08192b08, 0x08192b08), uvec2(0x082b0819, 0x08192b08), uvec2(0x19080808, 0x08192b08),
+    uvec2(0x1908082b, 0x08192b08), uvec2(0x19081919, 0x08192b08), uvec2(0x19082b08, 0x08192b08), uvec2(0x19190819, 0x08192b08),
+    uvec2(0x19191908, 0x08192b08), uvec2(0x192b0808, 0x08192b08), uvec2(0x2b080819, 0x08192b08), uvec2(0x2b081908, 0x08192b08),
+    uvec2(0x08080808, 0x08192b19), uvec2(0x0808082b, 0x08192b19), uvec2(0x08081919, 0x08192b19), uvec2(0x08082b08, 0x08192b19),
+    uvec2(0x08190819, 0x08192b19), uvec2(0x08191908, 0x08192b19), uvec2(0x082b0808, 0x08192b19), uvec2(0x19080819, 0x08192b19),
+    uvec2(0x19081908, 0x08192b19), uvec2(0x19190808, 0x08192b19), uvec2(0x192b2b19, 0x08192b19), uvec2(0x2b2b082b, 0x08192b19),
+    uvec2(0x08081908, 0x08192b2b), uvec2(0x08190808, 0x08192b2b), uvec2(0x19080808, 0x08192b2b), uvec2(0x1919192b, 0x08192b2b),
+    uvec2(0x08080808, 0x082b0808), uvec2(0x0808082b, 0x082b0808), uvec2(0x08081919, 0x082b0808), uvec2(0x08082b08, 0x082b0808),
+    uvec2(0x08190819, 0x082b0808), uvec2(0x08191908, 0x082b0808), uvec2(0x0819192b, 0x082b0808), uvec2(0x08192b19, 0x082b0808),
+    uvec2(0x082b0808, 0x082b0808), uvec2(0x082b1919, 0x082b0808), uvec2(0x082b2b2b, 0x082b0808), uvec2(0x19080819, 0x082b0808),
+    uvec2(0x19081908, 0x082b0808), uvec2(0x19190808, 0x082b0808), uvec2(0x1919082b, 0x082b0808), uvec2(0x19191919, 0x082b0808),
+    uvec2(0x192b1908, 0x082b0808), uvec2(0x2b080808, 0x082b0808), uvec2(0x2b082b2b, 0x082b0808), uvec2(0x2b191908, 0x082b0808),
+    uvec2(0x2b2b2b2b, 0x082b0808), uvec2(0x08080819, 0x082b0819), uvec2(0x08081908, 0x082b0819), uvec2(0x08190808, 0x082b0819),
+    uvec2(0x0819082b, 0x082b0819), uvec2(0x08191919, 0x082b0819), uvec2(0x082b0819, 0x082b0819), uvec2(0x19080808, 0x082b0819),
+    uvec2(0x1908082b, 0x082b0819), uvec2(0x19081919, 0x082b0819), uvec2(0x19190819, 0x082b0819), uvec2(0x19191908, 0x082b0819),
+    uvec2(0x192b0808, 0x082b0819), uvec2(0x2b080819, 0x082b0819), uvec2(0x2b081908, 0x082b0819), uvec2(0x2b190808, 0x082b0819),
+    uvec2(0x08080808, 0x082b082b), uvec2(0x08082b2b, 0x082b082b), uvec2(0x082b082b, 0x082b082b), uvec2(0x082b2b08, 0x082b082b),
+    uvec2(0x082b2b2b, 0x082b082b), uvec2(0x19081908, 0x082b082b), uvec2(0x19190808, 0x082b082b), uvec2(0x2b082b08, 0x082b082b),
+    uvec2(0x2b082b2b, 0x082b082b), uvec2(0x2b2b2b08, 0x082b082b), uvec2(0x08080819, 0x082b1908), uvec2(0x08081908, 0x082b1908),
+    uvec2(0x0808192b, 0x082b1908), uvec2(0x08082b19, 0x082b1908), uvec2(0x08190808, 0x082b1908), uvec2(0x08191919, 0x082b1908),
+    uvec2(0x08192b08, 0x082b1908), uvec2(0x082b0819, 0x082b1908), uvec2(0x082b1908, 0x082b1908), uvec2(0x19080808, 0x082b1908),
+    uvec2(0x1908082b, 0x082b1908), uvec2(0x19081919, 0x082b1908), uvec2(0x19082b08, 0x082b1908), uvec2(0x19190819, 0x082b1908),
+    uvec2(0x19191908, 0x082b1908), uvec2(0x192b0808, 0x082b1908), uvec2(0x2b080819, 0x082b1908), uvec2(0x2b081908, 0x082b1908),
+    uvec2(0x2b190808, 0x082b1908), uvec2(0x08080808, 0x082b1919), uvec2(0x08081919, 0x082b1919), uvec2(0x08082b08, 0x082b1919),
+    uvec2(0x08190819, 0x082b1919), uvec2(0x08191908, 0x082b1919), uvec2(0x082b0808, 0x082b1919), uvec2(0x19080819, 0x082b1919),
+    uvec2(0x19081908, 0x082b1919), uvec2(0x19190808, 0x082b1919), uvec2(0x192b192b, 0x082b1919), uvec2(0x2b080808, 0x082b1919),
+    uvec2(0x08080819, 0x082b192b), uvec2(0x08081908, 0x082b192b), uvec2(0x08190808, 0x082b192b), uvec2(0x19080808, 0x082b192b),
+    uvec2(0x19192b19, 0x082b192b), uvec2(0x08080808, 0x082b2b08), uvec2(0x08081919, 0x082b2b08), uvec2(0x08190819, 0x082b2b08),
+    uvec2(0x08191908, 0x082b2b08), uvec2(0x19080819, 0x082b2b08), uvec2(0x19081908, 0x082b2b08), uvec2(0x19190808, 0x082b2b08),
+    uvec2(0x2b082b2b, 0x082b2b08), uvec2(0x2b2b2b2b, 0x082b2b08), uvec2(0x08080819, 0x082b2b19), uvec2(0x08081908, 0x082b2b19),
+    uvec2(0x08190808, 0x082b2b19), uvec2(0x2b191919, 0x082b2b19), uvec2(0x08082b2b, 0x082b2b2b), uvec2(0x082b082b, 0x082b2b2b),
+    uvec2(0x192b1908, 0x082b2b2b), uvec2(0x2b082b08, 0x082b2b2b), uvec2(0x2b082b2b, 0x082b2b2b), uvec2(0x08080819, 0x19080808),
+    uvec2(0x08081908, 0x19080808), uvec2(0x0808192b, 0x19080808), uvec2(0x08082b19, 0x19080808), uvec2(0x08190808, 0x19080808),
+    uvec2(0x0819082b, 0x19080808), uvec2(0x08191919, 0x19080808), uvec2(0x08192b08, 0x19080808), uvec2(0x08192b2b, 0x19080808),
+    uvec2(0x082b0819, 0x19080808), uvec2(0x082b1908, 0x19080808), uvec2(0x082b192b, 0x19080808), uvec2(0x19080808, 0x19080808),
+    uvec2(0x1908082b, 0x19080808), uvec2(0x19081919, 0x19080808), uvec2(0x19082b08, 0x19080808), uvec2(0x19082b2b, 0x19080808),
+    uvec2(0x19190819, 0x19080808), uvec2(0x19191908, 0x19080808), uvec2(0x1919192b, 0x19080808), uvec2(0x19192b19, 0x19080808),
+    uvec2(0x192b0808, 0x19080808), uvec2(0x192b082b, 0x19080808), uvec2(0x192b1919, 0x19080808), uvec2(0x2b080819, 0x19080808),
+    uvec2(0x2b081908, 0x19080808), uvec2(0x2b190808, 0x19080808), uvec2(0x2b191919, 0x19080808), uvec2(0x2b192b08, 0x19080808),
+    uvec2(0x2b2b0819, 0x19080808), uvec2(0x2b2b1908, 0x19080808), uvec2(0x08080808, 0x19080819), uvec2(0x0808082b, 0x19080819),
+    uvec2(0x08081919, 0x19080819), uvec2(0x08082b08, 0x19080819), uvec2(0x08190819, 0x19080819), uvec2(0x08191908, 0x19080819),
+    uvec2(0x0819192b, 0x19080819), uvec2(0x08192b19, 0x19080819), uvec2(0x082b0808, 0x19080819), uvec2(0x082b082b, 0x19080819),
+    uvec2(0x082b1919, 0x19080819), uvec2(0x19080819, 0x19080819), uvec2(0x19081908, 0x19080819), uvec2(0x1908192b, 0x19080819),
+    uvec2(0x19082b19, 0x19080819), uvec2(0x19190808, 0x19080819), uvec2(0x1919082b, 0x19080819), uvec2(0x19191919, 0x19080819),
+    uvec2(0x19192b08, 0x19080819), uvec2(0x192b0819, 0x19080819), uvec2(0x192b1908, 0x19080819), uvec2(0x2b080808, 0x19080819),
+    uvec2(0x2b08082b, 0x19080819), uvec2(0x2b081919, 0x19080819), uvec2(0x2b082b08, 0x19080819), uvec2(0x2b190819, 0x19080819),
+    uvec2(0x2b191908, 0x19080819), uvec2(0x2b2b0808, 0x19080819), uvec2(0x08080819, 0x1908082b), uvec2(0x08081908, 0x1908082b),
+    uvec2(0x08190808, 0x1908082b), uvec2(0x0819082b, 0x1908082b), uvec2(0x08191919, 0x1908082b), uvec2(0x08192b08, 0x1908082b),
+    uvec2(0x082b1908, 0x1908082b), uvec2(0x19080808, 0x1908082b), uvec2(0x19081919, 0x1908082b), uvec2(0x19082b08, 0x1908082b),
+    uvec2(0x19190819, 0x1908082b), uvec2(0x19191908, 0x1908082b), uvec2(0x192b0808, 0x1908082b), uvec2(0x2b080819, 0x1908082b),
+    uvec2(0x2b081908, 0x1908082b), uvec2(0x08080808, 0x19081908), uvec2(0x0808082b, 0x19081908), uvec2(0x08081919, 0x19081908),
+    uvec2(0x08082b08, 0x19081908), uvec2(0x08082b2b, 0x19081908), uvec2(0x08190819, 0x19081908), uvec2(0x08191908, 0x19081908),
+    uvec2(0x0819192b, 0x19081908), uvec2(0x08192b19, 0x19081908), uvec2(0x082b0808, 0x19081908), uvec2(0x082b082b, 0x19081908),
+    uvec2(0x082b1919, 0x19081908), uvec2(0x082b2b08, 0x19081908), uvec2(0x19080819, 0x19081908), uvec2(0x19081908, 0x19081908),
+    uvec2(0x1908192b, 0x19081908), uvec2(0x19082b19, 0x19081908), uvec2(0x19190808, 0x19081908), uvec2(0x1919082b, 0x19081908),
+    uvec2(0x19191919, 0x19081908), uvec2(0x19192b08, 0x19081908), uvec2(0x192b0819, 0x19081908), uvec2(0x192b1908, 0x19081908),
+    uvec2(0x2b080808, 0x19081908), uvec2(0x2b08082b, 0x19081908), uvec2(0x2b081919, 0x19081908), uvec2(0x2b082b08, 0x19081908),
+    uvec2(0x2b190819, 0x19081908), uvec2(0x2b191908, 0x19081908), uvec2(0x2b2b0808, 0x19081908), uvec2(0x08080819, 0x19081919),
+    uvec2(0x08081908, 0x19081919), uvec2(0x0808192b, 0x19081919), uvec2(0x08082b19, 0x19081919), uvec2(0x08190808, 0x19081919),
+    uvec2(0x0819082b, 0x19081919), uvec2(0x08191919, 0x19081919), uvec2(0x08192b08, 0x19081919), uvec2(0x082b0819, 0x19081919),
+    uvec2(0x082b1908, 0x19081919), uvec2(0x19080808, 0x19081919), uvec2(0x1908082b, 0x19081919), uvec2(0x19081919, 0x19081919),
+    uvec2(0x19082b08, 0x19081919), uvec2(0x19190819, 0x19081919), uvec2(0x19191908, 0x19081919), uvec2(0x192b0808, 0x19081919),
+    uvec2(0x192b2b2b, 0x19081919), uvec2(0x2b080819, 0x19081919), uvec2(0x2b081908, 0x19081919), uvec2(0x2b190808, 0x19081919),
+    uvec2(0x08080808, 0x1908192b), uvec2(0x0808082b, 0x1908192b), uvec2(0x08081919, 0x1908192b), uvec2(0x08082b08, 0x1908192b),
+    uvec2(0x08190819, 0x1908192b), uvec2(0x08191908, 0x1908192b), uvec2(0x082b0808, 0x1908192b), uvec2(0x19080819, 0x1908192b),
+    uvec2(0x19081908, 0x1908192b), uvec2(0x19190808, 0x1908192b), uvec2(0x2b080808, 0x1908192b), uvec2(0x2b2b1919, 0x1908192b),
+    uvec2(0x08080819, 0x19082b08), uvec2(0x08081908, 0x19082b08), uvec2(0x08082b19, 0x19082b08), uvec2(0x08190808, 0x19082b08),
+    uvec2(0x0819082b, 0x19082b08), uvec2(0x08191919, 0x19082b08), uvec2(0x08192b08, 0x19082b08), uvec2(0x082b0819, 0x19082b08),
+    uvec2(0x082b1908, 0x19082b08), uvec2(0x19080808, 0x19082b08), uvec2(0x1908082b, 0x19082b08), uvec2(0x19081919, 0x19082b08),
+    uvec2(0x19082b08, 0x19082b08), uvec2(0x19190819, 0x19082b08), uvec2(0x19191908, 0x19082b08), uvec2(0x192b0808, 0x19082b08),
+    uvec2(0x2b081908, 0x19082b08), uvec2(0x2b190808, 0x19082b08), uvec2(0x08080808, 0x19082b19), uvec2(0x0808082b, 0x19082b19),
+    uvec2(0x08081919, 0x19082b19), uvec2(0x08082b08, 0x19082b19), uvec2(0x08190819, 0x19082b19), uvec2(0x08191908, 0x19082b19),
+    uvec2(0x082b0808, 0x19082b19), uvec2(0x19080819, 0x19082b19), uvec2(0x19081908, 0x19082b19), uvec2(0x19190808, 0x19082b19),
+    uvec2(0x2b080808, 0x19082b19), uvec2(0x2b19192b, 0x19082b19), uvec2(0x08080819, 0x19082b2b), uvec2(0x08081908, 0x19082b2b),
+    uvec2(0x08190808, 0x19082b2b), uvec2(0x19080808, 0x19082b2b), uvec2(0x08080808, 0x19190808), uvec2(0x0808082b, 0x19190808),
+    uvec2(0x08081919, 0x19190808), uvec2(0x08082b08, 0x19190808), uvec2(0x08190819, 0x19190808), uvec2(0x08191908, 0x19190808),
+    uvec2(0x0819192b, 0x19190808), uvec2(0x08192b19, 0x19190808), uvec2(0x082b0808, 0x19190808), uvec2(0x082b082b, 0x19190808),
+    uvec2(0x082b1919, 0x19190808), uvec2(0x082b2b08, 0x19190808), uvec2(0x19080819, 0x19190808), uvec2(0x19081908, 0x19190808),
+    uvec2(0x1908192b, 0x19190808), uvec2(0x19082b19, 0x19190808), uvec2(0x19190808, 0x19190808), uvec2(0x1919082b, 0x19190808),
+    uvec2(0x19191919, 0x19190808), uvec2(0x19192b08, 0x19190808), uvec2(0x192b0819, 0x19190808), uvec2(0x192b1908, 0x19190808),
+    uvec2(0x2b080808, 0x19190808), uvec2(0x2b08082b, 0x19190808), uvec2(0x2b081919, 0x19190808), uvec2(0x2b082b08, 0x19190808),
+    uvec2(0x2b190819, 0x19190808), uvec2(0x2b191908, 0x19190808), uvec2(0x08080819, 0x19190819), uvec2(0x08081908, 0x19190819),
+    uvec2(0x0808192b, 0x19190819), uvec2(0x08082b19, 0x19190819), uvec2(0x08190808, 0x19190819), uvec2(0x0819082b, 0x19190819),
+    uvec2(0x08191919, 0x19190819), uvec2(0x08192b08, 0x19190819), uvec2(0x082b0819, 0x19190819), uvec2(0x082b1908, 0x19190819),
+    uvec2(0x19080808, 0x19190819), uvec2(0x1908082b, 0x19190819), uvec2(0x19081919, 0x19190819), uvec2(0x19082b08, 0x19190819),
+    uvec2(0x19190819, 0x19190819), uvec2(0x19191908, 0x19190819), uvec2(0x192b0808, 0x19190819), uvec2(0x2b080819, 0x19190819),
+    uvec2(0x2b081908, 0x19190819), uvec2(0x2b190808, 0x19190819), uvec2(0x08080808, 0x1919082b), uvec2(0x08081919, 0x1919082b),
+    uvec2(0x08082b08, 0x1919082b), uvec2(0x08190819, 0x1919082b), uvec2(0x08191908, 0x1919082b), uvec2(0x082b0808, 0x1919082b),
+    uvec2(0x19080819, 0x1919082b), uvec2(0x19081908, 0x1919082b), uvec2(0x19190808, 0x1919082b), uvec2(0x192b2b19, 0x1919082b),
+    uvec2(0x2b080808, 0x1919082b), uvec2(0x08080819, 0x19191908), uvec2(0x08081908, 0x19191908), uvec2(0x0808192b, 0x19191908),
+    uvec2(0x08082b19, 0x19191908), uvec2(0x08190808, 0x19191908), uvec2(0x0819082b, 0x19191908), uvec2(0x08191919, 0x19191908),
+    uvec2(0x08192b08, 0x19191908), uvec2(0x082b0819, 0x19191908), uvec2(0x082b1908, 0x19191908), uvec2(0x19080808, 0x19191908),
+    uvec2(0x1908082b, 0x19191908), uvec2(0x19081919, 0x19191908), uvec2(0x19082b08, 0x19191908), uvec2(0x19190819, 0x19191908),
+    uvec2(0x19191908, 0x19191908), uvec2(0x192b0808, 0x19191908), uvec2(0x2b080819, 0x19191908), uvec2(0x2b081908, 0x19191908),
+    uvec2(0x2b190808, 0x19191908), uvec2(0x08080808, 0x19191919), uvec2(0x0808082b, 0x19191919), uvec2(0x08081919, 0x19191919),
+    uvec2(0x08082b08, 0x19191919), uvec2(0x08190819, 0x19191919), uvec2(0x08191908, 0x19191919), uvec2(0x082b0808, 0x19191919),
+    uvec2(0x19080819, 0x19191919), uvec2(0x19081908, 0x19191919), uvec2(0x19190808, 0x19191919), uvec2(0x2b080808, 0x19191919),
+    uvec2(0x08080819, 0x1919192b), uvec2(0x08081908, 0x1919192b), uvec2(0x08190808, 0x1919192b), uvec2(0x082b192b, 0x1919192b),
+    uvec2(0x19080808, 0x1919192b), uvec2(0x08080808, 0x19192b08), uvec2(0x0808082b, 0x19192b08), uvec2(0x08081919, 0x19192b08),
+    uvec2(0x08082b08, 0x19192b08), uvec2(0x08190819, 0x19192b08), uvec2(0x08191908, 0x19192b08), uvec2(0x082b0808, 0x19192b08),
+    uvec2(0x19080819, 0x19192b08), uvec2(0x19081908, 0x19192b08), uvec2(0x19190808, 0x19192b08), uvec2(0x19192b2b, 0x19192b08),
+    uvec2(0x2b080808, 0x19192b08), uvec2(0x08080819, 0x19192b19), uvec2(0x08081908, 0x19192b19), uvec2(0x08190808, 0x19192b19),
+    uvec2(0x19080808, 0x19192b19), uvec2(0x08080808, 0x19192b2b), uvec2(0x08192b19, 0x19192b2b), uvec2(0x2b081919, 0x19192b2b),
+    uvec2(0x2b2b2b08, 0x19192b2b), uvec2(0x08080819, 0x192b0808), uvec2(0x08081908, 0x192b0808), uvec2(0x0808192b, 0x192b0808),
+    uvec2(0x08190808, 0x192b0808), uvec2(0x0819082b, 0x192b0808), uvec2(0x08191919, 0x192b0808), uvec2(0x08192b08, 0x192b0808),
+    uvec2(0x082b0819, 0x192b0808), uvec2(0x082b1908, 0x192b0808), uvec2(0x19080808, 0x192b0808), uvec2(0x19081919, 0x192b0808),
+    uvec2(0x19082b08, 0x192b0808), uvec2(0x19190819, 0x192b0808), uvec2(0x19191908, 0x192b0808), uvec2(0x192b0808, 0x192b0808),
+    uvec2(0x2b081908, 0x192b0808), uvec2(0x2b190808, 0x192b0808), uvec2(0x08080808, 0x192b0819), uvec2(0x0808082b, 0x192b0819),
+    uvec2(0x08081919, 0x192b0819), uvec2(0x08082b08, 0x192b0819), uvec2(0x08190819, 0x192b0819), uvec2(0x08191908, 0x192b0819),
+    uvec2(0x082b0808, 0x192b0819), uvec2(0x19080819, 0x192b0819), uvec2(0x19081908, 0x192b0819), uvec2(0x19190808, 0x192b0819),
+    uvec2(0x2b080808, 0x192b0819), uvec2(0x2b192b19, 0x192b0819), uvec2(0x08081908, 0x192b082b), uvec2(0x08190808, 0x192b082b),
+    uvec2(0x19080808, 0x192b082b), uvec2(0x1919192b, 0x192b082b), uvec2(0x2b2b0819, 0x192b082b), uvec2(0x08080808, 0x192b1908),
+    uvec2(0x08081919, 0x192b1908), uvec2(0x08082b08, 0x192b1908), uvec2(0x08190819, 0x192b1908), uvec2(0x08191908, 0x192b1908),
+    uvec2(0x082b0808, 0x192b1908), uvec2(0x19080819, 0x192b1908), uvec2(0x19081908, 0x192b1908), uvec2(0x19190808, 0x192b1908),
+    uvec2(0x2b080808, 0x192b1908), uvec2(0x08080819, 0x192b1919), uvec2(0x08081908, 0x192b1919), uvec2(0x08190808, 0x192b1919),
+    uvec2(0x19080808, 0x192b1919), uvec2(0x19082b2b, 0x192b1919), uvec2(0x192b2b08, 0x192b1919), uvec2(0x2b19082b, 0x192b1919),
+    uvec2(0x08080808, 0x192b192b), uvec2(0x2b191908, 0x192b192b), uvec2(0x08080819, 0x192b2b08), uvec2(0x08081908, 0x192b2b08),
+    uvec2(0x08190808, 0x192b2b08), uvec2(0x192b1919, 0x192b2b08), uvec2(0x2b192b08, 0x192b2b08), uvec2(0x08080808, 0x192b2b19),
+    uvec2(0x082b2b2b, 0x192b2b19), uvec2(0x1908082b, 0x192b2b2b), uvec2(0x2b2b0819, 0x192b2b2b), uvec2(0x08080808, 0x2b080808),
+    uvec2(0x0808082b, 0x2b080808), uvec2(0x08081919, 0x2b080808), uvec2(0x08082b08, 0x2b080808), uvec2(0x08190819, 0x2b080808),
+    uvec2(0x08191908, 0x2b080808), uvec2(0x08192b19, 0x2b080808), uvec2(0x082b0808, 0x2b080808), uvec2(0x082b1919, 0x2b080808),
+    uvec2(0x19080819, 0x2b080808), uvec2(0x19081908, 0x2b080808), uvec2(0x19190808, 0x2b080808), uvec2(0x1919082b, 0x2b080808),
+    uvec2(0x19191919, 0x2b080808), uvec2(0x19192b08, 0x2b080808), uvec2(0x192b0819, 0x2b080808), uvec2(0x2b080808, 0x2b080808),
+    uvec2(0x2b081919, 0x2b080808), uvec2(0x2b190819, 0x2b080808), uvec2(0x2b191908, 0x2b080808), uvec2(0x08080819, 0x2b080819),
+    uvec2(0x08081908, 0x2b080819), uvec2(0x08082b19, 0x2b080819), uvec2(0x08190808, 0x2b080819), uvec2(0x0819082b, 0x2b080819),
+    uvec2(0x08191919, 0x2b080819), uvec2(0x08192b08, 0x2b080819), uvec2(0x082b0819, 0x2b080819), uvec2(0x082b1908, 0x2b080819),
+    uvec2(0x19080808, 0x2b080819), uvec2(0x1908082b, 0x2b080819), uvec2(0x19081919, 0x2b080819), uvec2(0x19082b08, 0x2b080819),
+    uvec2(0x19190819, 0x2b080819), uvec2(0x19191908, 0x2b080819), uvec2(0x2b080819, 0x2b080819), uvec2(0x2b081908, 0x2b080819),
+    uvec2(0x2b190808, 0x2b080819), uvec2(0x2b2b2b19, 0x2b080819), uvec2(0x08080808, 0x2b08082b), uvec2(0x08081919, 0x2b08082b),
+    uvec2(0x08082b2b, 0x2b08082b), uvec2(0x08190819, 0x2b08082b), uvec2(0x08191908, 0x2b08082b), uvec2(0x19080819, 0x2b08082b),
+    uvec2(0x19081908, 0x2b08082b), uvec2(0x19190808, 0x2b08082b), uvec2(0x08080819, 0x2b081908), uvec2(0x08081908, 0x2b081908),
+    uvec2(0x0808192b, 0x2b081908), uvec2(0x08082b19, 0x2b081908), uvec2(0x08190808, 0x2b081908), uvec2(0x0819082b, 0x2b081908),
+    uvec2(0x08191919, 0x2b081908), uvec2(0x08192b08, 0x2b081908), uvec2(0x082b0819, 0x2b081908), uvec2(0x19080808, 0x2b081908),
+    uvec2(0x1908082b, 0x2b081908), uvec2(0x19081919, 0x2b081908), uvec2(0x19082b08, 0x2b081908), uvec2(0x19190819, 0x2b081908),
+    uvec2(0x19191908, 0x2b081908), uvec2(0x192b0808, 0x2b081908), uvec2(0x2b080819, 0x2b081908), uvec2(0x2b081908, 0x2b081908),
+    uvec2(0x2b190808, 0x2b081908), uvec2(0x08080808, 0x2b081919), uvec2(0x0808082b, 0x2b081919), uvec2(0x08081919, 0x2b081919),
+    uvec2(0x08082b08, 0x2b081919), uvec2(0x08190819, 0x2b081919), uvec2(0x08191908, 0x2b081919), uvec2(0x082b0808, 0x2b081919),
+    uvec2(0x19080819, 0x2b081919), uvec2(0x19081908, 0x2b081919), uvec2(0x19190808, 0x2b081919), uvec2(0x2b080808, 0x2b081919),
+    uvec2(0x2b082b2b, 0x2b081919), uvec2(0x08080819, 0x2b08192b), uvec2(0x08081908, 0x2b08192b), uvec2(0x08190808, 0x2b08192b),
+    uvec2(0x082b2b19, 0x2b08192b), uvec2(0x19080808, 0x2b08192b), uvec2(0x08080808, 0x2b082b08), uvec2(0x08081919, 0x2b082b08),
+    uvec2(0x08190819, 0x2b082b08), uvec2(0x08191908, 0x2b082b08), uvec2(0x19080819, 0x2b082b08), uvec2(0x19081908, 0x2b082b08),
+    uvec2(0x19190808, 0x2b082b08), uvec2(0x2b2b082b, 0x2b082b08), uvec2(0x08080819, 0x2b082b19), uvec2(0x08081908, 0x2b082b19),
+    uvec2(0x19080808, 0x2b082b19), uvec2(0x192b1919, 0x2b082b19), uvec2(0x082b082b, 0x2b082b2b), uvec2(0x19192b08, 0x2b082b2b),
+    uvec2(0x19192b2b, 0x2b082b2b), uvec2(0x2b08082b, 0x2b082b2b), uvec2(0x2b2b082b, 0x2b082b2b), uvec2(0x08080819, 0x2b190808),
+    uvec2(0x08081908, 0x2b190808), uvec2(0x08082b19, 0x2b190808), uvec2(0x08190808, 0x2b190808), uvec2(0x0819082b, 0x2b190808),
+    uvec2(0x08191919, 0x2b190808), uvec2(0x08192b08, 0x2b190808), uvec2(0x082b1908, 0x2b190808), uvec2(0x19080808, 0x2b190808),
+    uvec2(0x1908082b, 0x2b190808), uvec2(0x19081919, 0x2b190808), uvec2(0x19082b08, 0x2b190808), uvec2(0x19190819, 0x2b190808),
+    uvec2(0x19191908, 0x2b190808), uvec2(0x192b0808, 0x2b190808), uvec2(0x2b080819, 0x2b190808), uvec2(0x2b081908, 0x2b190808),
+    uvec2(0x2b190808, 0x2b190808), uvec2(0x08080808, 0x2b190819), uvec2(0x08081919, 0x2b190819), uvec2(0x08190819, 0x2b190819),
+    uvec2(0x08191908, 0x2b190819), uvec2(0x19080819, 0x2b190819), uvec2(0x19081908, 0x2b190819), uvec2(0x19190808, 0x2b190819),
+    uvec2(0x19192b2b, 0x2b190819), uvec2(0x08080819, 0x2b19082b), uvec2(0x08081908, 0x2b19082b), uvec2(0x08190808, 0x2b19082b),
+    uvec2(0x19080808, 0x2b19082b), uvec2(0x2b2b192b, 0x2b19082b), uvec2(0x08080808, 0x2b191908), uvec2(0x0808082b, 0x2b191908),
+    uvec2(0x08081919, 0x2b191908), uvec2(0x08082b08, 0x2b191908), uvec2(0x08190819, 0x2b191908), uvec2(0x08191908, 0x2b191908),
+    uvec2(0x082b0808, 0x2b191908), uvec2(0x19080819, 0x2b191908), uvec2(0x19081908, 0x2b191908), uvec2(0x19190808, 0x2b191908),
+    uvec2(0x2b080808, 0x2b191908), uvec2(0x2b19192b, 0x2b191908), uvec2(0x08080819, 0x2b191919), uvec2(0x08081908, 0x2b191919),
+    uvec2(0x08190808, 0x2b191919), uvec2(0x19080808, 0x2b191919), uvec2(0x2b192b08, 0x2b191919), uvec2(0x2b2b0819, 0x2b191919),
+    uvec2(0x08080808, 0x2b19192b), uvec2(0x1908192b, 0x2b19192b), uvec2(0x192b1908, 0x2b19192b), uvec2(0x08080819, 0x2b192b08),
+    uvec2(0x08081908, 0x2b192b08), uvec2(0x08190808, 0x2b192b08), uvec2(0x082b192b, 0x2b192b08), uvec2(0x19080808, 0x2b192b08),
+    uvec2(0x2b2b2b19, 0x2b192b08), uvec2(0x08080808, 0x2b192b19), uvec2(0x19082b19, 0x2b192b19), uvec2(0x1919082b, 0x2b192b19),
+    uvec2(0x2b190808, 0x2b192b2b), uvec2(0x08080808, 0x2b2b0808), uvec2(0x08081919, 0x2b2b0808), uvec2(0x08082b2b, 0x2b2b0808),
+    uvec2(0x08191908, 0x2b2b0808), uvec2(0x082b082b, 0x2b2b0808), uvec2(0x082b2b2b, 0x2b2b0808), uvec2(0x19080819, 0x2b2b0808),
+    uvec2(0x19081908, 0x2b2b0808), uvec2(0x19190808, 0x2b2b0808), uvec2(0x2b2b082b, 0x2b2b0808), uvec2(0x2b2b2b2b, 0x2b2b0808),
+    uvec2(0x19080808, 0x2b2b0819), uvec2(0x192b1919, 0x2b2b0819), uvec2(0x0808082b, 0x2b2b082b), uvec2(0x08082b2b, 0x2b2b082b),
+    uvec2(0x082b082b, 0x2b2b082b), uvec2(0x082b2b08, 0x2b2b082b), uvec2(0x082b2b2b, 0x2b2b082b), uvec2(0x2b08082b, 0x2b2b082b),
+    uvec2(0x2b082b08, 0x2b2b082b), uvec2(0x2b082b2b, 0x2b2b082b), uvec2(0x2b2b2b08, 0x2b2b082b), uvec2(0x08080819, 0x2b2b1908),
+    uvec2(0x08081908, 0x2b2b1908), uvec2(0x08190808, 0x2b2b1908), uvec2(0x19080808, 0x2b2b1908), uvec2(0x2b082b19, 0x2b2b1908),
+    uvec2(0x2b2b1908, 0x2b2b1908), uvec2(0x08080808, 0x2b2b1919), uvec2(0x08192b19, 0x2b2b1919), uvec2(0x19190819, 0x2b2b192b),
+    uvec2(0x08082b2b, 0x2b2b2b08), uvec2(0x082b2b08, 0x2b2b2b08), uvec2(0x2b2b082b, 0x2b2b2b08), uvec2(0x19191908, 0x2b2b2b19),
+    uvec2(0x2b08192b, 0x2b2b2b19), uvec2(0x08082b08, 0x2b2b2b2b), uvec2(0x08082b2b, 0x2b2b2b2b), uvec2(0x082b0808, 0x2b2b2b2b),
+    uvec2(0x082b082b, 0x2b2b2b2b), uvec2(0x082b2b08, 0x2b2b2b2b), uvec2(0x2b082b08, 0x2b2b2b2b), uvec2(0x2b2b2b2b, 0x2b2b2b2b)
+};
+
+shared uvec2 iq2s_grid[1024];
+
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < iq2s_grid.length(); i += wgsize.x) {
+        iq2s_grid[i] = iq2s_grid_const[i];
+    }
+    barrier();
+}
+
+#define QUANT_K QUANT_K_IQ2_S
+#define QUANT_R QUANT_R_IQ2_S
+#define A_TYPE block_iq2_s
+#endif
+
+#define QUANT_K_IQ3_XXS 256
+#define QUANT_R_IQ3_XXS 1
+
+struct block_iq3_xxs
+{
+    float16_t d;
+    uint8_t qs[QUANT_K_IQ3_XXS/4 + QUANT_K_IQ3_XXS/8];
+};
+
+struct block_iq3_xxs_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K_IQ3_XXS/8 + QUANT_K_IQ3_XXS/16];
+};
+
+#if defined(DATA_A_IQ3_XXS)
+
+const uint32_t iq3xxs_grid_const[256] = {
+    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
+    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
+    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
+    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
+    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
+    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
+    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
+    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
+    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
+    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
+    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
+    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
+    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
+    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
+    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
+    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
+    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
+    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
+    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
+    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
+    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
+    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
+    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
+    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
+    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
+    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
+    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
+    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
+    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
+    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
+    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
+    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
+};
+
+shared uint32_t iq3xxs_grid[256];
+
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < iq3xxs_grid.length(); i += wgsize.x) {
+        iq3xxs_grid[i] = iq3xxs_grid_const[i];
+    }
+    barrier();
+}
+
+#define QUANT_K QUANT_K_IQ3_XXS
+#define QUANT_R QUANT_R_IQ3_XXS
+#define A_TYPE block_iq3_xxs
+#define A_TYPE_PACKED16 block_iq3_xxs_packed16
+#endif
+
+#define QUANT_K_IQ3_S 256
+#define QUANT_R_IQ3_S 1
+
+struct block_iq3_s
+{
+    float16_t d;
+    uint8_t qs[QUANT_K_IQ3_S/4];
+    uint8_t qh[QUANT_K_IQ3_S/32];
+    uint8_t signs[QUANT_K_IQ3_S/8];
+    uint8_t scales[QUANT_K_IQ3_S/64];
+};
+
+struct block_iq3_s_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K_IQ3_S/4/2];
+    uint16_t qh[QUANT_K_IQ3_S/32/2];
+    uint16_t signs[QUANT_K_IQ3_S/8/2];
+    uint16_t scales[QUANT_K_IQ3_S/64/2];
+};
+
+#if defined(DATA_A_IQ3_S)
+
+const uint32_t iq3s_grid_const[512] = {
+    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
+    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
+    0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
+    0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
+    0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
+    0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
+    0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
+    0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
+    0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
+    0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
+    0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
+    0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
+    0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
+    0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
+    0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
+    0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
+    0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
+    0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
+    0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
+    0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
+    0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
+    0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
+    0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
+    0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
+    0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
+    0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
+    0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
+    0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
+    0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
+    0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
+    0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
+    0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
+    0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
+    0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
+    0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
+    0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
+    0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
+    0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
+    0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
+    0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
+    0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
+    0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
+    0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
+    0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
+    0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
+    0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
+    0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
+    0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
+    0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
+    0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
+    0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
+    0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
+    0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
+    0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
+    0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
+    0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
+    0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
+    0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
+    0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
+    0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
+    0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
+    0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
+    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
+    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
+};
+
+shared uint32_t iq3s_grid[512];
+
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < iq3s_grid.length(); i += wgsize.x) {
+        iq3s_grid[i] = iq3s_grid_const[i];
+    }
+    barrier();
+}
+
+#define QUANT_K QUANT_K_IQ3_S
+#define QUANT_R QUANT_R_IQ3_S
+#define A_TYPE block_iq3_s
+#define A_TYPE_PACKED16 block_iq3_s_packed16
+#endif
+
 #define QUANT_K_IQ4_NL 32
 #define QUANT_R_IQ4_NL 2
 
@@ -318,11 +1050,11 @@ const int8_t kvalues_iq4nl_const[16] = {
 
 shared FLOAT_TYPE kvalues_iq4nl[16];
 
-void init_iq4nl_shmem()
+void init_iq_shmem(uvec3 wgsize)
 {
     // copy the table into shared memory and sync
-    if (gl_LocalInvocationIndex.x < 16) {
-        kvalues_iq4nl[gl_LocalInvocationIndex.x] = FLOAT_TYPE(kvalues_iq4nl_const[gl_LocalInvocationIndex.x]);
+    for (uint i = gl_LocalInvocationIndex.x; i < kvalues_iq4nl.length(); i += wgsize.x) {
+        kvalues_iq4nl[i] = FLOAT_TYPE(kvalues_iq4nl_const[i]);
     }
     barrier();
 }

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index e9c6cb9d45da59872138ade3c73f60e66913f301..93ddbfadc5f9abd15ff26e9449ca077c94457e16 100644 (file)
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -55,6 +55,11 @@ const std::vector<std::string> type_names = {
     "q4_k",
     "q5_k",
     "q6_k",
+    "iq2_xxs",
+    "iq2_xs",
+    "iq2_s",
+    "iq3_xxs",
+    "iq3_s",
     "iq4_nl"
 };