wip

Compute two result elements per workgroup (for Q{4,5}_{0,1}). This reuses
the B loads across the rows and also reuses some addressing calculations.
This required manually partially unrolling the loop, since the compiler
is less willing to unroll outer loops.

Add bounds-checking on the last iteration of the loop. I think this was at
least partly broken before.

Optimize the Q4_K shader to vectorize most loads and reduce the number of
bit twiddling instructions.

CMakePresets.json

@@ -24,11 +24,12 @@
             "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
         }
     },
-    { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
-    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
-    { "name": "reldbg",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
-    { "name": "static",  "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
-    { "name": "sycl_f16",  "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
+    { "name": "debug",    "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
+    { "name": "release",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
+    { "name": "reldbg",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "static",   "hidden": true, "cacheVariables": { "GGML_STATIC":      "ON" } },
+    { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16":    "ON" } },
+    { "name": "vulkan",   "hidden": true, "cacheVariables": { "GGML_VULKAN":      "ON" } },
 
     {
         "name": "arm64-windows-msvc", "hidden": true,
@@ -57,25 +58,28 @@
         }
     },
 
-    { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
-    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg" ] },
-    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg", "static" ] },
+    { "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
+    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
+    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
 
-    { "name": "arm64-apple-clang-debug"  , "inherits": [ "base", "arm64-apple-clang",  "debug"   ] },
-    { "name": "arm64-apple-clang-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg"   ] },
-    { "name": "arm64-apple-clang+static-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },
+    { "name": "arm64-apple-clang-debug", "inherits": [ "base", "arm64-apple-clang", "debug" ] },
+    { "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
+    { "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },
 
-    { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
+    { "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
     { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
 
-    { "name": "x64-windows-msvc-debug"  , "inherits": [ "base", "debug"   ] },
+    { "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] },
     { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
 
-    { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
+    { "name": "x64-windows-sycl-debug", "inherits": [ "sycl-base", "debug" ] },
     { "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
     { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
-    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
+    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] },
+
+    { "name": "x64-windows-vulkan-debug", "inherits": [ "base", "vulkan", "debug" ] },
+    { "name": "x64-windows-vulkan-release", "inherits": [ "base", "vulkan", "release" ] }
   ]
 }

examples/llama-bench/llama-bench.cpp

@@ -1337,7 +1337,14 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
         for (int i = 1; i < n_tokens; i++) {
             tokens[i] = std::rand() % n_vocab;
         }
-        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens));
+        auto batch = llama_batch_get_one(tokens.data(), n_tokens);
+        int8_t logits[512];
+        for (int i = 0; i < n_tokens; i++) {
+            logits[i] = 1;
+        }
+        batch.logits = logits;
+
+        llama_decode(ctx, batch);
         n_processed += n_tokens;
     }
 

ggml/src/ggml-aarch64.c

@@ -8,19 +8,42 @@
 
 #define UNUSED GGML_UNUSED
 
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x4 out;
 
     for (int i = 0; i < 4; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 2; i++) {
-        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 2 / blck_size_interleave;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    if (blck_size_interleave == 8) {
+        const uint64_t xor_mask = 0x8888888888888888ULL;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint64_t elems;
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        const uint32_t xor_mask = 0x88888888;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint32_t elems;
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
     }
 
     return out;
@@ -30,19 +53,25 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
 // returns an interleaved block_q4_0x8
 // in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
 // first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x8 out;
 
     for (int i = 0; i < 8; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 4; i++) {
-        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 4 / blck_size_interleave;
+    const uint64_t xor_mask = 0x8888888888888888ULL;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
     }
 
     return out;
@@ -71,11 +100,11 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds
             }
 
             if (nrows_interleaved == 8) {
-                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
+                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave);
                 out_ptr = (block_q4_0x8 *) out_ptr + 1;
             }
             else if (nrows_interleaved == 4) {
-                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
+                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave);
                 out_ptr = (block_q4_0x4 *) out_ptr + 1;
             }
         }

ggml/src/ggml-cpu/ggml-cpu-aarch64.c

@@ -3387,19 +3387,42 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 }
 
 // FIXME: this code is duplicated from ggml-aarch64.c
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x4 out;
 
     for (int i = 0; i < 4; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 2; i++) {
-        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 2 / blck_size_interleave;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    if (blck_size_interleave == 8) {
+        const uint64_t xor_mask = 0x8888888888888888ULL;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint64_t elems;
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        const uint32_t xor_mask = 0x88888888;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint32_t elems;
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
     }
 
     return out;
@@ -3409,19 +3432,25 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
 // returns an interleaved block_q4_0x8
 // in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
 // first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x8 out;
 
     for (int i = 0; i < 8; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 4; i++) {
-        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 4 / blck_size_interleave;
+    const uint64_t xor_mask = 0x8888888888888888ULL;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
     }
 
     return out;
@@ -3449,7 +3478,7 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
             for (int i = 0; i < nrows_interleaved; i++) {
                 dst_tmp[i] = src[x + i * nblocks];
             }
-            *dst++ = make_block_q4_0x4(dst_tmp, interleave_block, 0x88);
+            *dst++ = make_block_q4_0x4(dst_tmp, interleave_block);
         }
         src += nrows_interleaved * nblocks;
     }
@@ -3480,7 +3509,7 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block,
             for (int i  = 0; i < nrows_interleaved; i++ ) {
                 dst_tmp[i] = src[x + i * nblocks];
             }
-            *dst++ = make_block_q4_0x8(dst_tmp, interleave_block, 0x88);
+            *dst++ = make_block_q4_0x8(dst_tmp, interleave_block);
         }
         src += nrows_interleaved * nblocks;
     }

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

@@ -1365,47 +1365,48 @@ static void ggml_vk_load_shaders(vk_device& device) {
     }
 
     // mul mat vec
-    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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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);
-    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), {1, 1, 1}, { device->subgroup_size }, 1);
+    // 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_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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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);
-    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), {1, 1, 1}, { device->subgroup_size }, 1);
+    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_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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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), {1, 1, 1}, { device->subgroup_size }, 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);
-    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), {1, 1, 1}, { device->subgroup_size }, 1);
+    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);
 
     // 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);

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

@@ -3,54 +3,107 @@
 #ifdef FLOAT16
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#extension GL_EXT_null_initializer : enable
 
 #include "mul_mat_vec_base.comp"
 
 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;
 
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
+uint a_offset, b_offset, d_offset, y_offset;
 
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-    const uint tid = gl_LocalInvocationID.x;
+shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
-    // There are not enough cols to use all threads
-    if (tid >= p.ncols) {
-        return;
+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 iqs = (col%QUANT_K)/QUANT_R; // quant index
+    const uint iybs = col - col%QUANT_K; // y block start index
+
+    // 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
+    // probably skip fetching the second element for F16/F32, but as of now we
+    // still do.
+    const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
+
+    FLOAT_TYPE b0 = 0, b1 = 0;
+    b0 = FLOAT_TYPE(data_b[b_offset + iybs + iqs]);
+    if (!OOB) {
+        b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
     }
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib = ((first_row + n)*p.ncols + col)/QUANT_K; // block index
 
-    const uint block_size = min(p.ncols, BLOCK_SIZE);
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
-
-    tmp[tid] = FLOAT_TYPE(0.0f);
-
-    [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
-        const uint col = i*block_size + 2*tid;
-        const uint ib = (row*p.ncols + col)/QUANT_K; // block index
-        const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
-        const uint iybs = col - col%QUANT_K; // y block start index
-
-        vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
+        const vec2 v = dequantize(ib, iqs, a_offset);
 
         // matrix multiplication
-        tmp[tid] = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp[tid]));
+        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
+        if (!OOB) {
+            temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
+        }
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    const uint tid = gl_LocalInvocationID.x;
+
+    get_offsets(a_offset, b_offset, d_offset);
+    a_offset /= QUANT_K;
+
+    y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+
+    FLOAT_TYPE temp[NUM_ROWS] = {};
+
+    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 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;
+        }
+    }
+    while (i < num_iters) {
+        iter(temp, first_row, num_rows, tid, i, true);
+        i += 2;
     }
 
     // sum up partial sums and write back result
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        tmpsh[n][tid] = temp[n];
+    }
     barrier();
-    [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
         if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[n][tid] += tmpsh[n][tid + s];
+            }
         }
         barrier();
     }
     if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
+        }
+    }
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

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

@@ -1,11 +1,34 @@
 #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;
 
 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;
 
@@ -31,79 +54,81 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*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 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);
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(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));
+        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];
+        uvec4 scale0 = uvec4(unpack8(scale0_u32));
+        uvec4 scale4 = uvec4(unpack8(scale4_u32));
+        uvec4 scale8 = uvec4(unpack8(scale8_u32));
 
-#if K_QUANTS_PER_ITERATION == 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 +  2] & 0xf);
-        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset +  3] & 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 +  2]  >> 4);
-        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset +  3]  >> 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 + 66] & 0xf);
-        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] & 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 + 66]  >> 4);
-        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67]  >> 4);
+        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));
 
-        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx]),      q4_0,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]),  q4_1,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]),  q4_2,  FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) *  q4_3)));
-        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_4,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), q4_5,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), q4_6,  FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7)));
-        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx]),      q4_8,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]),  q4_9,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]),  q4_10, FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) *  q4_11)));
-        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_12, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), q4_13, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), q4_14, FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15)));
+        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_lo4 = qs0_u32 & 0x0F0F0F0F;
+        uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
+        uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
+
+        uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
+        uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
+        uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
+        uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
+
+        const uint32_t q4_0  = qs0_lo4.x;
+        const uint32_t q4_1  = qs0_lo4.y;
+        const uint32_t q4_2  = qs0_lo4.z;
+        const uint32_t q4_3  = qs0_lo4.w;
+        const uint32_t q4_4  = qs0_hi4.x;
+        const uint32_t q4_5  = qs0_hi4.y;
+        const uint32_t q4_6  = qs0_hi4.z;
+        const uint32_t q4_7  = qs0_hi4.w;
+        const uint32_t q4_8  = qs64_lo4.x;
+        const uint32_t q4_9  = qs64_lo4.y;
+        const uint32_t q4_10 = qs64_lo4.z;
+        const uint32_t q4_11 = qs64_lo4.w;
+        const uint32_t q4_12 = qs64_hi4.x;
+        const uint32_t q4_13 = qs64_hi4.y;
+        const uint32_t q4_14 = qs64_hi4.z;
+        const uint32_t q4_15 = qs64_hi4.w;
+
+        B_TYPE_VEC4 by10 =  data_b_v4[(b_offset + y1_idx) / 4];
+        B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
+        B_TYPE_VEC4 by20 =  data_b_v4[(b_offset + y2_idx) / 4];
+        B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
+
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
         const FLOAT_TYPE smin =
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), sc7,
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), sc7,
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 3]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 35]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 3]), sc6,     FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * 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]));
-#else
-        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     ]  >> 4);
-        const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
-        const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
-        const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
-        const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
-        const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
-
-        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), q4_0, FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
-        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_2, FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
-        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), q4_4, FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
-        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
-        const FLOAT_TYPE smin =
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
-          + fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7)))))));
-
-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) +
-                        sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, (fma(sx, FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f), fma(sy, FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f),
-                       fma(sz, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)), fma(sw, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))))))), fma(-dmin, smin, tmp[tmp_idx]));
-#endif
+            fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
+            fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
+            fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
+            fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * 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/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"}, {"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"}, {"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_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_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"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"}}));
 
         // Dequant shaders
         if (tname != "f16") {

scripts/compare-llama-bench.py

@@ -303,14 +303,11 @@ else:
 
     show = []
     # Show CPU and/or GPU by default even if the hardware for all results is the same:
-    if "gpu_blas" not in properties_different and "n_gpu_layers" not in properties_different:
-        gpu_blas = bool(rows_full[0][KEY_PROPERTIES.index("gpu_blas")])
+    if "n_gpu_layers" not in properties_different:
         ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")])
 
-        if not gpu_blas or ngl != 99 and "cpu_info" not in properties_different:
+        if ngl != 99 and "cpu_info" not in properties_different:
             show.append("cpu_info")
-        if gpu_blas and "gpu_info" not in properties_different:
-            show.append("gpu_info")
 
     show += properties_different
 

src/llama.cpp

@@ -2907,9 +2907,15 @@ struct llama_model {
     // for quantize-stats only
     std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
-    int64_t t_load_us = 0;
+    int64_t t_load_us  = 0;
     int64_t t_start_us = 0;
 
+    // total number of parameters in the model
+    uint64_t n_elements = 0;
+
+    // total size of all the tensors in the model in bytes
+    size_t  n_bytes     = 0;
+
     // keep track of loaded lora adapters
     std::set<struct llama_lora_adapter *> lora_adapters;
 
@@ -4275,8 +4281,8 @@ struct llama_model_loader {
     int n_tensors = 0;
     int n_created = 0;
 
-    int64_t n_elements = 0;
-    size_t  n_bytes    = 0;
+    uint64_t n_elements = 0;
+    size_t  n_bytes     = 0;
 
     bool use_mmap = false;
     bool check_tensors;
@@ -5344,6 +5350,11 @@ static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
     }
 }
 
+static void llm_load_stats(llama_model_loader & ml, llama_model & model) {
+    model.n_elements = ml.n_elements;
+    model.n_bytes = ml.n_bytes;
+}
+
 static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
     model.arch = ml.get_arch();
     if (model.arch == LLM_ARCH_UNKNOWN) {
@@ -9256,6 +9267,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
             throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
         }
 
+        llm_load_stats(ml, model);
         llm_load_print_meta(ml, model);
 
         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
@@ -18601,6 +18613,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     llama_model model;
     llm_load_arch(ml, model);
     llm_load_hparams(ml, model);
+    llm_load_stats(ml, model);
 
     struct quantize_state_internal qs(model, params);
 
@@ -19953,19 +19966,11 @@ int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t bu
 }
 
 uint64_t llama_model_size(const struct llama_model * model) {
-    uint64_t size = 0;
-    for (const auto & it : model->tensors_by_name) {
-        size += ggml_nbytes(it.second);
-    }
-    return size;
+    return model->n_bytes;
 }
 
 uint64_t llama_model_n_params(const struct llama_model * model) {
-    uint64_t nparams = 0;
-    for (const auto & it : model->tensors_by_name) {
-        nparams += ggml_nelements(it.second);
-    }
-    return nparams;
+    return model->n_elements;
 }
 
 struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name) {