vulkan: further optimize mul_mat_vec using larger loads (llama/10387)

* vulkan: Use pipeline_robustness to disable robustness in mul_mat_vec.

Add some early returns for nonexistent rows in mul_mat_vec shaders. These
can only be hit when dispatching a 2D grid of workgroups. Fix the logic
for the 2D grid of workgroups to round up.

Enable the pipeline robustness extension if it's available, and use it to
disable robustness for these pipelines. The instructions to do the bounds
checking contend for the same ALU resources as the bit twiddling dequant
instructions.

* vulkan: Add GLSL structure aliases for quant types to allow larger loads

In Vulkan it's not possible to cast pointer types, so instead you have to
declare an aliased binding for the memory with a different type. This
commit adds aliases for the quant formats using 16b ints, and in a few
places where the struct size is a multiple of 4 also using 32b ints.
Currently only q4_k's aliases are used, but others will be used in
subsequent commits.

* vulkan: use larger loads in q5_k and q6_k shaders.

Similar to the optimization I did in q4_k recently, this vectorizes some loads
and reduces the number of bit twiddling instructions.

* vulkan: use larger K step per iteration in mul_mat_vec.

Add vec4 dequantization functions, and use them to do K=8 per iteration in
mul_mat_vec. This uses 16b loads for the quant values and 128b loads for B
which helps reduce the load on the memory system.

The K_PER_ITER==2 logic is still there, just for F16/F32, and really only
because they support unaligned sizes.

Tweak the num_iters/unrolling logic to be simpler and catch a couple missed
unrolling opportunities.

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -158,6 +158,7 @@ struct vk_device_struct {
     std::string name;
     uint64_t max_memory_allocation_size;
     bool fp16;
+    bool pipeline_robustness;
     vk::Device device;
     uint32_t vendor_id;
     vk_queue compute_queue;
@@ -654,7 +655,7 @@ static uint32_t compile_count = 0;
 static std::mutex compile_count_mutex;
 static std::condition_variable compile_count_cond;
 
-static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align) {
+static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align, bool disable_robustness) {
     VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")");
     GGML_ASSERT(parameter_count > 0);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
@@ -724,6 +725,15 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         vk::PipelineCreateFlags(),
         pipeline_shader_create_info,
         pipeline->layout);
+
+    vk::PipelineRobustnessCreateInfoEXT rci;
+
+    if (device->pipeline_robustness && disable_robustness) {
+        rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
+        rci.uniformBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
+        compute_pipeline_create_info.setPNext(&rci);
+    }
+
     pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
 
     {
@@ -1261,7 +1271,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
 
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align) {
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align, bool disable_robustness = false) {
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1271,7 +1281,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             }
             compile_count++;
         }
-        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align));
+        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness));
     };
 
     if (device->fp16) {
@@ -1370,45 +1380,45 @@ static void ggml_vk_load_shaders(vk_device& device) {
     // computing two rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", 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, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", 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, 1, 1}, {device->subgroup_size, 2}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  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}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  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}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", 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, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", 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, 1, 1}, {device->subgroup_size}, 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);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    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), {1, 1, 1}, {device->subgroup_size}, 1);
-    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), {1, 1, 1}, {device->subgroup_size}, 1);
-    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), {1, 1, 1}, {device->subgroup_size}, 1);
-    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, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    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), {1, 1, 1}, {device->subgroup_size}, 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), {1, 1, 1}, {device->subgroup_size}, 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), {1, 1, 1}, {device->subgroup_size}, 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, 1, 1}, {device->subgroup_size, 2}, 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);
@@ -1591,12 +1601,15 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         bool fp16_storage = false;
         bool fp16_compute = false;
+        bool pipeline_robustness = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
                 fp16_storage = true;
             } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
                 fp16_compute = true;
+            } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) {
+                pipeline_robustness = true;
             }
         }
 
@@ -1642,10 +1655,22 @@ static vk_device ggml_vk_get_device(size_t idx) {
         vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
         vk11_features.pNext = &vk12_features;
 
+        VkPhysicalDevicePipelineRobustnessFeaturesEXT pl_robustness_features;
+        pl_robustness_features.pNext = nullptr;
+        pl_robustness_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_ROBUSTNESS_FEATURES_EXT;
+        pl_robustness_features.pipelineRobustness = VK_FALSE;
+
+        if (pipeline_robustness) {
+            vk12_features.pNext = &pl_robustness_features;
+            device_extensions.push_back("VK_EXT_pipeline_robustness");
+        }
+
         vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
 
         device->fp16 = device->fp16 && vk12_features.shaderFloat16;
 
+        device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
+
         if (!vk11_features.storageBuffer16BitAccess) {
             std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl;
             throw std::runtime_error("Unsupported device");
@@ -3190,7 +3215,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     if (ne01 > max_groups_x) {
         groups_z = 64;
-        groups_x /= groups_z;
+        groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
     // compute
@@ -3767,7 +3792,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
     if (ne01 > max_groups_x) {
         groups_z = 64;
-        groups_x /= groups_z;
+        groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
     // compute

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp

@@ -2,6 +2,15 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
 #endif
 
+#include "types.comp"
+
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 #if defined(DATA_A_F32)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
@@ -20,6 +29,11 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) - 8.0f) * d;
+}
 #endif
 
 #if defined(DATA_A_Q4_1)
@@ -29,6 +43,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return vec2(vui & 0xF, vui >> 4) * d + m;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const float m = float(data_a_packed16[a_offset + ib].m);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) * d + m;
+}
 #endif
 
 #if defined(DATA_A_Q5_0)
@@ -39,6 +59,14 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0];
+    const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return (vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) - 16.0f) * d;
+}
 #endif
 
 #if defined(DATA_A_Q5_1)
@@ -50,6 +78,15 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const float m = float(data_a_packed16[a_offset + ib].m);
+    const uint uint_qh = data_a_packed16[a_offset + ib].qh;
+    const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) * d + m;
+}
 #endif
 
 #if defined(DATA_A_Q8_0)
@@ -57,6 +94,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const float d = float(data_a[a_offset + ib].d);
     return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2];
+    uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1];
+    return vec4(int8_t(v0 & 0xFF), int8_t((v0 >> 8) & 0xFF), int8_t(v1 & 0xFF), int8_t((v1 >> 8) & 0xFF)) * d;
+}
 #endif
 
 #if defined(DATA_A_IQ4_NL)
@@ -65,4 +108,9 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[(vui >> 12) & 0xF]) * d;
+}
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -3,7 +3,7 @@
 #ifdef FLOAT16
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
-#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #include "mul_mat_vec_base.comp"
 
@@ -12,16 +12,48 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 layout (constant_id = 1) const uint NUM_ROWS = 1;
 
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#define K_PER_ITER 8
+#else
+#define K_PER_ITER 2
+#endif
+
+
 uint a_offset, b_offset, d_offset, y_offset;
 
 shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
 void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
 {
-    const uint col = i*BLOCK_SIZE + 2*tid;
+    const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
     const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
     const uint iybs = col - col%QUANT_K; // y block start index
 
+#if K_PER_ITER == 8
+#if QUANT_R == 2
+    B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
+    B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
+    FLOAT_TYPE b0 = FLOAT_TYPE(bv02.x);
+    FLOAT_TYPE b1 = FLOAT_TYPE(bv13.x);
+    FLOAT_TYPE b2 = FLOAT_TYPE(bv02.y);
+    FLOAT_TYPE b3 = FLOAT_TYPE(bv13.y);
+    FLOAT_TYPE b4 = FLOAT_TYPE(bv02.z);
+    FLOAT_TYPE b5 = FLOAT_TYPE(bv13.z);
+    FLOAT_TYPE b6 = FLOAT_TYPE(bv02.w);
+    FLOAT_TYPE b7 = FLOAT_TYPE(bv13.w);
+#else
+    B_TYPE_VEC4 bv0 = data_b_v4[(b_offset + iybs + iqs) / 4];
+    B_TYPE_VEC4 bv1 = data_b_v4[(b_offset + iybs + iqs) / 4 + 1];
+    FLOAT_TYPE b0 = FLOAT_TYPE(bv0.x);
+    FLOAT_TYPE b1 = FLOAT_TYPE(bv0.y);
+    FLOAT_TYPE b2 = FLOAT_TYPE(bv0.z);
+    FLOAT_TYPE b3 = FLOAT_TYPE(bv0.w);
+    FLOAT_TYPE b4 = FLOAT_TYPE(bv1.x);
+    FLOAT_TYPE b5 = FLOAT_TYPE(bv1.y);
+    FLOAT_TYPE b6 = FLOAT_TYPE(bv1.z);
+    FLOAT_TYPE b7 = FLOAT_TYPE(bv1.w);
+#endif
+#else
     // Check if the second of the pair of elements is OOB, and don't fetch B or
     // accumulate it. We still fetch a pair of elements for A, which is fine for
     // quantized formats since they'll be within the same block. We should
@@ -34,9 +66,24 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
     if (!OOB) {
         b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
     }
+#endif
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib = ((first_row + n)*p.ncols + col)/QUANT_K; // block index
 
+#if K_PER_ITER == 8
+        const vec4 v = dequantize4(ib, iqs, a_offset);
+        const vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+
+        // matrix multiplication
+        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v.z), b2, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v.w), b3, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.x), b4, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.y), b5, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.z), b6, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.w), b7, temp[n]);
+#else
         const vec2 v = dequantize(ib, iqs, a_offset);
 
         // matrix multiplication
@@ -44,6 +91,7 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
         if (!OOB) {
             temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
         }
+#endif
     }
 }
 
@@ -61,22 +109,33 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
         temp[i] = FLOAT_TYPE(0);
     }
 
-    const int unroll_count = 8;
-
-    const uint num_iters = (p.ncols >= 2*tid) ? ((p.ncols - 2*tid + BLOCK_SIZE - 1) / BLOCK_SIZE) : 0;
-    const uint unrolled_iters = num_iters & ~(2*unroll_count - 1);
+    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
+    if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
+        num_iters++;
+    }
+    int unroll_count = 4;
+    uint unrolled_iters = num_iters & ~(unroll_count - 1);
 
     uint i = 0;
     while (i < unrolled_iters) {
         // Manually partially unroll the loop
         [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
-            iter(temp, first_row, num_rows, tid, i, false);
-            i += 2;
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
+        }
+    }
+    unroll_count = 2;
+    unrolled_iters = num_iters & ~(unroll_count - 1);
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
         }
     }
     while (i < num_iters) {
-        iter(temp, first_row, num_rows, tid, i, true);
-        i += 2;
+        iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true);
+        i++;
     }
 
     // sum up partial sums and write back result
@@ -106,6 +165,9 @@ void main() {
     if (first_row + NUM_ROWS <= p.stride_d) {
         compute_outputs(first_row, NUM_ROWS);
     } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
         compute_outputs(first_row, p.stride_d - first_row);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp

@@ -12,6 +12,9 @@
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -9,6 +9,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -9,6 +9,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -8,30 +8,14 @@ layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 
 shared FLOAT_TYPE tmp[32];
 
-// Declare aliased versions of A and B bindings that can use 16b/32b loads for
-// the quantized values, and vec4 loads for B.
-struct block_q4_K_u32
-{
-    f16vec2 d;
-    uint32_t scales[3*QUANT_K/64/4];
-    uint32_t qs[QUANT_K/2/4];
-};
-
-struct block_q4_K_u16
-{
-    f16vec2 d;
-    uint16_t scales[3*QUANT_K/64/2];
-    uint16_t qs[QUANT_K/2/2];
-};
-
-layout (binding = 0) readonly buffer A_u32 {block_q4_K_u32 data_a_u32[];};
-layout (binding = 0) readonly buffer A_u16 {block_q4_K_u16 data_a_u16[];};
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
-
 // This shader assumes K_QUANTS_PER_ITERATION == 2 for alignment of loads
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -64,9 +48,9 @@ void main() {
         const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
         const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
 
-        uint32_t scale0_u32 = data_a_u16[ib0 + i].scales[v_im    ];
-        uint32_t scale4_u32 = data_a_u16[ib0 + i].scales[v_im + 2];
-        uint32_t scale8_u32 = data_a_u16[ib0 + i].scales[v_im + 4];
+        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
         uvec4 scale0 = uvec4(unpack8(scale0_u32));
         uvec4 scale4 = uvec4(unpack8(scale4_u32));
         uvec4 scale8 = uvec4(unpack8(scale8_u32));
@@ -80,8 +64,8 @@ void main() {
         const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
         const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
 
-        uint32_t qs0_u32 = data_a_u32[ib0 + i].qs[q_offset / 4];
-        uint32_t qs64_u32 = data_a_u32[ib0 + i].qs[q_offset / 4 + 16];
+        uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
+        uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
 
         uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
         uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -1,5 +1,7 @@
 #version 450
 
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
 #include "mul_mat_vec_base.comp"
 
 layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
@@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -31,70 +37,106 @@ void main() {
     const uint8_t hm1 = uint8_t(1 << (2*v_im));
     const uint8_t hm2 = uint8_t(hm1 << 4);
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
-
-        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
-        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
-        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
-        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
-        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
-        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
-        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
-        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
-
-        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
-        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
-        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf);
-        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf);
-        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
-        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
-        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset + 16]  >> 4);
-        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset + 17]  >> 4);
-        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
-        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
-        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf);
-        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf);
-        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
-        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
-        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80]  >> 4);
-        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81]  >> 4);
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+        uvec4 scale0 = uvec4(unpack8(scale0_u32));
+        uvec4 scale4 = uvec4(unpack8(scale4_u32));
+        uvec4 scale8 = uvec4(unpack8(scale8_u32));
+
+        const uint32_t sc0 = (  scale0.x       & 0x3f);
+        const uint32_t sc1 = (  scale0.y       & 0x3f);
+        const uint32_t sc2 = (  scale4.x       & 0x3f);
+        const uint32_t sc3 = (  scale4.y       & 0x3f);
+        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+
+        uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+        uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
+
+        uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
+        uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
+
+        uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
+        uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
+        uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
+        uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
+
+        const uint32_t q4_0  = qs0_16_lo4.x;
+        const uint32_t q4_1  = qs0_16_lo4.y;
+        const uint32_t q4_2  = qs0_16_lo4.z;
+        const uint32_t q4_3  = qs0_16_lo4.w;
+        const uint32_t q4_4  = qs0_16_hi4.x;
+        const uint32_t q4_5  = qs0_16_hi4.y;
+        const uint32_t q4_6  = qs0_16_hi4.z;
+        const uint32_t q4_7  = qs0_16_hi4.w;
+        const uint32_t q4_8  = qs64_80_lo4.x;
+        const uint32_t q4_9  = qs64_80_lo4.y;
+        const uint32_t q4_10 = qs64_80_lo4.z;
+        const uint32_t q4_11 = qs64_80_lo4.w;
+        const uint32_t q4_12 = qs64_80_hi4.x;
+        const uint32_t q4_13 = qs64_80_hi4.y;
+        const uint32_t q4_14 = qs64_80_hi4.z;
+        const uint32_t q4_15 = qs64_80_hi4.w;
+
+        B_TYPE_VEC2 by10 =  data_b_v2[(b_offset + y1_idx) / 2];
+        B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
+        B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
+        B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24];
+        B_TYPE_VEC2 by20 =  data_b_v2[(b_offset + y2_idx) / 2];
+        B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8];
+        B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
+        B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
+
+        uint32_t qh0 = data_a_packed16[ib0 + i].qh[l0 / 2];
+        uint32_t qh1 = qh0 >> 8;
+        uint32_t qh16 = data_a_packed16[ib0 + i].qh[l0 / 2 + 8];
+        uint32_t qh17 = qh16 >> 8;
 
         const FLOAT_TYPE sx =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx +  1]), (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by10.x), (q4_0 + (((qh0 & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by10.y), (q4_1 + (((qh1 & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by116.x), (q4_2 + (((qh16 & hm1) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by116.y) * (q4_3 + (((qh17 & hm1) != 0) ? 16 : 0)))));
         const FLOAT_TYPE sy =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by132.x), (q4_4 + (((qh0 & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by132.y), (q4_5 + (((qh1 & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by148.x), (q4_6 + (((qh16 & (hm1 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by148.y) * (q4_7 + (((qh17 & (hm1 << 1)) != 0) ? 16 : 0)))));
         const FLOAT_TYPE sz =
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx +  1]), (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by20.x), (q4_8  + (((qh0 & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by20.y), (q4_9  + (((qh1 & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by216.x), (q4_10 + (((qh16 & hm2) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by216.y) * (q4_11 + (((qh17 & hm2) != 0) ? 16 : 0)))));
         const FLOAT_TYPE sw =
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by232.x), (q4_12 + (((qh0 & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by232.y), (q4_13 + (((qh1 & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by248.x), (q4_14 + (((qh16 & (hm2 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by248.y) * (q4_15 + (((qh17 & (hm2 << 1)) != 0) ? 16 : 0)))));
         const FLOAT_TYPE smin =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2,
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3,
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6,
-              (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7)));
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
+          fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
+          fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
+          fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
+              (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
+        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -1,5 +1,7 @@
 #version 450
 
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
 #include "mul_mat_vec_base.comp"
 
 layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
@@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -36,41 +42,66 @@ void main() {
     const uint s_offset  =  8*v_im + is;
     const uint y_offset = 128*v_im + l0;
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
         const uint y_idx   = i * QUANT_K + y_offset;
 
         const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
-#if K_QUANTS_PER_ITERATION == 1
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
-#else
+        FLOAT_TYPE scales[4];
+        scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
+        scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
+        scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
+        scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
+
+        uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
+        uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
+
+        uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
+        uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
+        uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
+
+        uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
+        uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
+        uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
+        uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
+        uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
+
+        uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
+        uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
+        uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
+        uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
+
+        uvec4 q0 = uvec4(unpack8(q0_u32));
+        uvec4 q1 = uvec4(unpack8(q1_u32));
+        uvec4 q2 = uvec4(unpack8(q2_u32));
+        uvec4 q3 = uvec4(unpack8(q3_u32));
+
+        B_TYPE_VEC4 by0  = data_b_v4[(b_offset + y_idx) / 4];
+        B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
+        B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
+        B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
+
         FLOAT_TYPE sum = FLOAT_TYPE(0.0);
         [[unroll]] for (int l = 0; l < 4; ++l) {
-            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
+            sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
+                  fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
+                  fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
+                  fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
         }
-        tmp[16 * ix + tid] += sum;
-#endif
+        temp += sum * d;
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
-       }
+        }
         barrier();
     }
     if (tid == 0) {

ggml/src/ggml-vulkan/vulkan-shaders/types.comp

@@ -1,6 +1,8 @@
-#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
-#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
-#endif
+
+#if !defined(GGML_TYPES_COMP)
+#define GGML_TYPES_COMP
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #if defined(DATA_A_F32)
 #define QUANT_K 1
@@ -38,8 +40,14 @@ struct block_q4_0
     float16_t d;
     uint8_t qs[16];
 };
+struct block_q4_0_packed16
+{
+    float16_t d;
+    uint16_t qs[16/2];
+};
 
 #define A_TYPE block_q4_0
+#define A_TYPE_PACKED16 block_q4_0_packed16
 #endif
 
 #if defined(DATA_A_Q4_1)
@@ -54,7 +62,15 @@ struct block_q4_1
     uint8_t qs[16];
 };
 
+struct block_q4_1_packed16
+{
+    float16_t d;
+    float16_t m;
+    uint16_t qs[16/2];
+};
+
 #define A_TYPE block_q4_1
+#define A_TYPE_PACKED16 block_q4_1_packed16
 #endif
 
 #if defined(DATA_A_Q5_0)
@@ -70,7 +86,15 @@ struct block_q5_0
     uint8_t qs[16];
 };
 
+struct block_q5_0_packed16
+{
+    float16_t d;
+    uint16_t qh[2];
+    uint16_t qs[16/2];
+};
+
 #define A_TYPE block_q5_0
+#define A_TYPE_PACKED16 block_q5_0_packed16
 #endif
 
 #if defined(DATA_A_Q5_1)
@@ -87,7 +111,16 @@ struct block_q5_1
     uint8_t qs[16];
 };
 
+struct block_q5_1_packed16
+{
+    float16_t d;
+    float16_t m;
+    uint qh;
+    uint16_t qs[16/2];
+};
+
 #define A_TYPE block_q5_1
+#define A_TYPE_PACKED16 block_q5_1_packed16
 #endif
 
 #if defined(DATA_A_Q8_0)
@@ -100,8 +133,14 @@ struct block_q8_0
     float16_t d;
     int8_t qs[32];
 };
+struct block_q8_0_packed16
+{
+    float16_t d;
+    uint16_t qs[32/2];
+};
 
 #define A_TYPE block_q8_0
+#define A_TYPE_PACKED16 block_q8_0_packed16
 #endif
 
 // K-quants
@@ -116,7 +155,23 @@ struct block_q2_K
     f16vec2 d;
 };
 
+struct block_q2_K_packed16
+{
+    uint16_t scales[QUANT_K/16/2];
+    uint16_t qs[QUANT_K/4/2];
+    f16vec2 d;
+};
+
+struct block_q2_K_packed32
+{
+    uint32_t scales[QUANT_K/16/4];
+    uint32_t qs[QUANT_K/4/4];
+    f16vec2 d;
+};
+
 #define A_TYPE block_q2_K
+#define A_TYPE_PACKED16 block_q2_K_packed16
+#define A_TYPE_PACKED32 block_q2_K_packed32
 #endif
 
 #if defined(DATA_A_Q3_K)
@@ -131,7 +186,16 @@ struct block_q3_K
     float16_t d;
 };
 
+struct block_q3_K_packed16
+{
+    uint16_t hmask[QUANT_K/8/2];
+    uint16_t qs[QUANT_K/4/2];
+    uint16_t scales[12/2];
+    float16_t d;
+};
+
 #define A_TYPE block_q3_K
+#define A_TYPE_PACKED16 block_q3_K_packed16
 #endif
 
 #if defined(DATA_A_Q4_K)
@@ -145,7 +209,23 @@ struct block_q4_K
     uint8_t qs[QUANT_K/2];
 };
 
+struct block_q4_K_packed16
+{
+    f16vec2 d;
+    uint16_t scales[3*QUANT_K/64/2];
+    uint16_t qs[QUANT_K/2/2];
+};
+
+struct block_q4_K_packed32
+{
+    f16vec2 d;
+    uint32_t scales[3*QUANT_K/64/4];
+    uint32_t qs[QUANT_K/2/4];
+};
+
 #define A_TYPE block_q4_K
+#define A_TYPE_PACKED16 block_q4_K_packed16
+#define A_TYPE_PACKED32 block_q4_K_packed32
 #endif
 
 #if defined(DATA_A_Q5_K)
@@ -160,7 +240,16 @@ struct block_q5_K
     uint8_t qs[QUANT_K/2];
 };
 
+struct block_q5_K_packed16
+{
+    f16vec2 d;
+    uint16_t scales[12/2];
+    uint16_t qh[QUANT_K/8/2];
+    uint16_t qs[QUANT_K/2/2];
+};
+
 #define A_TYPE block_q5_K
+#define A_TYPE_PACKED16 block_q5_K_packed16
 #endif
 
 #if defined(DATA_A_Q6_K)
@@ -175,7 +264,16 @@ struct block_q6_K
     float16_t d;
 };
 
+struct block_q6_K_packed16
+{
+    uint16_t ql[QUANT_K/2/2];
+    uint16_t qh[QUANT_K/4/2];
+    int8_t scales[QUANT_K/16];
+    float16_t d;
+};
+
 #define A_TYPE block_q6_K
+#define A_TYPE_PACKED16 block_q6_K_packed16
 #endif
 
 // IQuants
@@ -191,10 +289,19 @@ struct block_iq4_nl
     uint8_t qs[QUANT_K/2];
 };
 
+struct block_iq4_nl_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K/2/2];
+};
+
 #define A_TYPE block_iq4_nl
+#define A_TYPE_PACKED16 block_iq4_nl_packed16
 
 const int8_t kvalues_iq4nl[16] = {
     int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
     int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113)
 };
 #endif
+
+#endif // !defined(GGML_TYPES_COMP)

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -317,10 +317,10 @@ void process_shaders() {
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
 
-        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
-        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
 
-        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
         if (tname != "f16") {