whispercpp 1.3.4 → 1.3.5

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp

@@ -2,9 +2,9 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
-#include "types.comp"
-#include "generic_binary_head.comp"
-#include "dequant_funcs.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
+#include "dequant_funcs.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{glu_head.comp → glu_head.glsl}

@@ -1,6 +1,6 @@
 #extension GL_EXT_shader_16bit_storage : require
 
-#include "rte.comp"
+#include "rte.glsl"
 
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -3,9 +3,8 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_control_flow_attributes : require
 
-#include "rte.comp"
-
-#include "types.comp"
+#include "rte.glsl"
+#include "types.glsl"
 
 layout (push_constant) uniform parameter
 {
@@ -20,6 +19,7 @@ layout (push_constant) uniform parameter
     int s0; int s1;
     int p0; int p1;
     int d0; int d1;
+    uint batch_IC;
 } p;
 
 layout(constant_id = 0) const uint BLOCK_SIZE = 32;
@@ -35,12 +35,12 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 layout (buffer_reference) buffer D_ptr {D_TYPE d;};
 #endif
 
-void main() {
+void im2col(const uint y, const uint z) {
     const uint gidx = gl_GlobalInvocationID.x;
 
-    const uint oh = gl_GlobalInvocationID.y;
-    const uint batch = gl_GlobalInvocationID.z / p.IC;
-    const uint ic = gl_GlobalInvocationID.z % p.IC;
+    const uint oh = y;
+    const uint batch = z / p.IC;
+    const uint ic = z % p.IC;
 
     const uint src_base = ic * p.offset_delta + batch * p.batch_offset;
     const BDA_OFFSET_T dst_base = ((BDA_OFFSET_T(batch) * p.OH + oh) * p.OW) * p.CHW + BDA_OFFSET_T(ic) * (p.KW * p.KH);
@@ -102,3 +102,15 @@ void main() {
 #endif
     }
 }
+
+void main() {
+    uint y = gl_GlobalInvocationID.y;
+    while (y < p.OH) {
+        uint z = gl_GlobalInvocationID.z;
+        while (z < p.batch_IC) {
+            im2col(y, z);
+            z += gl_NumWorkGroups.z;
+        }
+        y += gl_NumWorkGroups.y;
+    }
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp

@@ -4,9 +4,8 @@
 #extension GL_EXT_control_flow_attributes : require
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
-#include "rte.comp"
-
-#include "types.comp"
+#include "rte.glsl"
+#include "types.glsl"
 
 layout (push_constant) uniform parameter
 {

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "generic_head.comp"
-#include "types.comp"
+#include "generic_head.glsl"
+#include "types.glsl"
 
 #extension GL_EXT_control_flow_attributes : enable
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp

@@ -0,0 +1,18 @@
+#version 450
+
+#include "rte.glsl"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const float val = float(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(log(val));
+}

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp

@@ -1,7 +1,7 @@
 #version 450
 
-#include "types.comp"
-#include "generic_binary_head.comp"
+#include "types.glsl"
+#include "generic_binary_head.glsl"
 
 const uint num_threads = 256;
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -2,7 +2,8 @@
 
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
-#include "mul_mat_vec_base.comp"
+#include "mul_mat_vec_base.glsl"
+#include "dequant_funcs.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
@@ -86,7 +87,6 @@ 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;
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{mul_mat_vec_base.comp → mul_mat_vec_base.glsl}

@@ -11,28 +11,7 @@
 #define EXPERT_COUNT 8
 #endif
 
-#include "types.comp"
-
-#ifndef MMQ
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-#else
-layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
-#endif
-
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-#ifdef B_TYPE_VEC2
-layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
-#endif
-#ifdef B_TYPE_VEC4
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
-#endif
-
-layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
-#ifdef MUL_MAT_ID
-layout (binding = 3) readonly buffer IDS {int data_ids[];};
-#endif
-
-#include "dequant_funcs.comp"
+#include "mul_mat_vec_iface.glsl"
 
 layout (push_constant) uniform parameter
 {
@@ -45,6 +24,8 @@ layout (push_constant) uniform parameter
     uint batch_stride_b;
     uint batch_stride_d;
 
+    uint fusion_flags;
+
 #ifdef MUL_MAT_ID
     uint nei0;
     uint ne11;
@@ -56,6 +37,10 @@ layout (push_constant) uniform parameter
 #endif
 } p;
 
+#ifdef MUL_MAT_ID
+uint expert_id;
+#endif
+
 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
     const uint expert_idx = gl_GlobalInvocationID.y;
@@ -75,14 +60,14 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
         batch_idx_a = i03 * p.ne02 + i02;
     }
 #else
-    const uint expert_id = data_ids[expert_idx];
+    expert_id = data_ids[expert_idx];
 #endif
 
     a_offset =
 #ifdef MUL_MAT_ID
-            expert_id * p.batch_stride_a;
+            expert_id * (p.batch_stride_a / QUANT_K);
 #else
-            batch_idx_a * p.batch_stride_a;
+            batch_idx_a * (p.batch_stride_a / QUANT_K);
 #endif
     b_offset =
 #ifdef MUL_MAT_ID
@@ -113,6 +98,26 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+#ifdef MUL_MAT_ID
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                }
+#else
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse1[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+#endif
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -148,6 +153,26 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                 [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
                     temp[j][n] += tmpsh[j][n][s];
                 }
+#ifdef MUL_MAT_ID
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                }
+#else
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse0[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                    temp[j][n] += FLOAT_TYPE(data_fuse1[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+#endif
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -173,6 +198,26 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+#ifdef MUL_MAT_ID
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
+                    const uint expert_idx = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                }
+#else
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_fuse0[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+                if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_fuse1[j*p.batch_stride_d + d_offset + first_row + n]);
+                }
+#endif
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
             }
         }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl

@@ -0,0 +1,35 @@
+#include "types.glsl"
+
+#define MAT_VEC_FUSION_FLAGS_BIAS0 0x1
+#define MAT_VEC_FUSION_FLAGS_BIAS1 0x2
+#define MAT_VEC_FUSION_FLAGS_SCALE0 0x4
+#define MAT_VEC_FUSION_FLAGS_SCALE1 0x8
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+#if defined(A_TYPE_VEC4)
+layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
+#endif
+#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
+
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+#ifdef B_TYPE_VEC2
+layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+#endif
+#ifdef B_TYPE_VEC4
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#endif
+
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+layout (binding = 3) readonly buffer Fuse0 {D_TYPE data_fuse0[];};
+layout (binding = 4) readonly buffer Fuse1 {D_TYPE data_fuse1[];};
+
+#ifdef MUL_MAT_ID
+layout (binding = 5) readonly buffer IDS {int data_ids[];};
+#endif
+

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp

@@ -1,41 +1,91 @@
 #version 450
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
-#include "mul_mat_vec_base.comp"
+#include "mul_mat_vec_base.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 
-void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i,
+                               const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    // Compute starting index in matrix B for this superblock
     const uint y_idx = i * QUANT_K + 32 * ib32;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    // Precompute indices for quantization lookup tables
+    const uint qh_base = 2 * ib32;
+    const uint qs_base = 4 * ib32;
+    const uint sc_index = ib32 / 2;
+    const uint sc_shift = 6 * (ib32 & 1);
+
+    // Loop over rows in the superblock
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        // Load per-block scales and shift for quantization
         const uint16_t[4] scales = data_a[ibi].scales;
         const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
         const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
+        const uint sc = data_a[ibi].scales[sc_index] >> sc_shift;
+
+        // Temporary caches for decoding
+        FLOAT_TYPE dl_cache[4];
+        uint16_t gvf_cache[4];
+        float delta_cache[4];
 
-        const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1));
+        // Precompute the multiplier and lookup values for 4 sub-blocks
         [[unroll]] for (uint l = 0; l < 4; ++l) {
-            const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1));
-            const uint qs = data_a[ibi].qs[4 * ib32 + l];
-            const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
-            const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1);
-
-            const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
-                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
-
-                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-                [[unroll]] for (int k = 0; k < 4; ++k) {
-                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
-                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
-                }
-                temp[j][n] = fma(dl, sum, temp[j][n]);
+            dl_cache[l] = FLOAT_TYPE(d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1));
+            const uint qh = data_a[ibi].qh[qh_base + l / 2] >> (4 * (l & 1));
+            const uint qs = data_a[ibi].qs[qs_base + l];
+            gvf_cache[l] = iq1s_grid[qs | ((qh & 7) << 8)];
+            delta_cache[l] = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+        }
+
+        // Loop over columns of the output
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            // Compute base index for matrix B
+            const uint base_b_idx = (j * p.batch_stride_b + b_offset + y_idx) / 4;
+            vec4 b_vals[8];
+
+            // Load 8 vec4 values from matrix B
+            [[unroll]] for (int idx = 0; idx < 8; ++idx) {
+                b_vals[idx] = vec4(data_b_v4[base_b_idx + idx]);
+            }
+
+            FLOAT_TYPE col_sum = FLOAT_TYPE(0.0);
+
+            // Loop over sub-blocks
+            [[unroll]] for (uint l = 0; l < 4; ++l) {
+                const uint16_t grid = gvf_cache[l];
+                const float dl = dl_cache[l];
+
+                // Decode 8 2-bit fbits from gvf_cache
+                float f0 = float(bitfieldExtract(grid, 0, 2));
+                float f1 = float(bitfieldExtract(grid, 2, 2));
+                float f2 = float(bitfieldExtract(grid, 4, 2));
+                float f3 = float(bitfieldExtract(grid, 6, 2));
+                float f4 = float(bitfieldExtract(grid, 8, 2));
+                float f5 = float(bitfieldExtract(grid, 10, 2));
+                float f6 = float(bitfieldExtract(grid, 12, 2));
+                float f7 = float(bitfieldExtract(grid, 14, 2));
+
+                // Pack into vec4 for vectorized FMA
+                const vec4 fbits_v0 = vec4(f0, f1, f2, f3);
+                const vec4 fbits_v1 = vec4(f4, f5, f6, f7);
+                const vec4 delta_v = vec4(delta_cache[l]);
+
+                // Vectorized fused multiply-add
+                vec4 sum_v = fma(b_vals[2*l + 0], fbits_v0 + delta_v, vec4(0.0));
+                sum_v      = fma(b_vals[2*l + 1], fbits_v1 + delta_v, sum_v);
+
+                // Horizontal add to get scalar sum
+                FLOAT_TYPE sum = sum_v.x + sum_v.y + sum_v.z + sum_v.w;
+
+                // Accumulate to column sum
+                col_sum = fma(dl, sum, col_sum);
             }
+            // Write result to temporary buffer
+            temp[j][n] += col_sum;
         }
         ibi += num_blocks_per_row;
     }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

@@ -1,40 +1,56 @@
 #version 450
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
-#include "mul_mat_vec_base.comp"
+#include "mul_mat_vec_base.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 
-void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
-    const uint y_idx = i * QUANT_K + 32 * ib32;
-
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const float d = float(data_a[ibi].d);
-        const uint qh = data_a[ibi].qh[ib32];
-        const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
-        const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
-
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i,
+                     const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx_base = i * QUANT_K + 32 * ib32;
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        const uint base_b_idx = (j * p.batch_stride_b + b_offset + y_idx_base) / 4;
         [[unroll]] for (uint l = 0; l < 4; ++l) {
-            const uint qs = data_a[ibi].qs[4 * ib32 + l];
-            const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
-            const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
-
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
-                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
-
-                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-                [[unroll]] for (int k = 0; k < 4; ++k) {
-                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
-                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
-                }
+            const vec4 b_val_0 = vec4(data_b_v4[base_b_idx + 2 * l]);
+            const vec4 b_val_1 = vec4(data_b_v4[base_b_idx + 2 * l + 1]);
+
+            // index for data_a
+            uint ibi = a_offset + first_row * num_blocks_per_row + i;
+
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                const float d = float(data_a[ibi].d);
+                const uint qh = data_a[ibi].qh[ib32];
+
+                const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+                const uint qs = data_a[ibi].qs[4 * ib32 + l];
+                const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
+                const uint16_t grid = uint16_t(iq1s_grid[qs | (idxhi << 8)]);
+
+                const float delta_val = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+                const vec4 delta_v = vec4(delta_val);
+                const vec4 fbits0 = vec4(
+                    float(bitfieldExtract(grid, 0, 2)),
+                    float(bitfieldExtract(grid, 2, 2)),
+                    float(bitfieldExtract(grid, 4, 2)),
+                    float(bitfieldExtract(grid, 6, 2))
+                );
+                const vec4 fbits1 = vec4(
+                    float(bitfieldExtract(grid, 8, 2)),
+                    float(bitfieldExtract(grid, 10, 2)),
+                    float(bitfieldExtract(grid, 12, 2)),
+                    float(bitfieldExtract(grid, 14, 2))
+                );
+
+                vec4 sum_v = fma(b_val_0, fbits0 + delta_v, vec4(0.0));
+                sum_v      = fma(b_val_1, fbits1 + delta_v, sum_v);
+                FLOAT_TYPE sum = dot(sum_v, vec4(1.0));
+
                 temp[j][n] = fma(dl, sum, temp[j][n]);
+                ibi += num_blocks_per_row;
             }
         }
-        ibi += num_blocks_per_row;
     }
 }
 

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp

@@ -1,7 +1,7 @@
 #version 450
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
-#include "mul_mat_vec_base.comp"
+#include "mul_mat_vec_base.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
@@ -12,7 +12,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint nibble_shift = 4 * (itid & 1);
     const uint ib32 = itid / 2; // 0..7
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;