whispercpp 1.3.4 → 1.3.6

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;

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.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;
 
@@ -11,36 +11,54 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * 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;
+    // Precompute db multiplication factors
+    float db_vals[NUM_ROWS];
     [[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;
-        const float db = d * (0.5 + scale) * 0.25;
-
+        const uint scale_raw = data_a[ibi].scales[ib32];
+        const uint scale = (scale_raw >> nibble_shift) & 0xF;
+        // Merge constant calculations d * (0.5 + scale) * 0.25 = d*0.125 + d*scale*0.25
+        db_vals[n] = d * (0.125f + float(scale) * 0.25f);
+        ibi += num_blocks_per_row;
+    }
+    ibi = a_offset + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        // Preload grid and sign data for all l values
+        vec4 grid0_vals[2], grid1_vals[2];
+        uint sign_vals[2], sign7_vals[2];
         [[unroll]] for (uint l = 0; l < 2; ++l) {
             const uint qs = data_a[ibi].qs[2 * itid + l];
-            const uint sign = qs >> 9;
-            const uint sign7 = bitCount(sign);
-            const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x));
-            const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y));
-
-            [[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 =
-                      fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
-                      fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
-                      fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
-                      fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
-                      fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
-                      fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
-                      fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
-                      fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
-                      FLOAT_TYPE(0.0)))))))));
-                temp[j][n] = fma(db, sum, temp[j][n]);
+            sign_vals[l] = qs >> 9;
+            sign7_vals[l] = bitCount(sign_vals[l]);
+            const uvec2 grid_data = iq2xs_grid[qs & 511];
+            grid0_vals[l] = vec4(unpack8(grid_data.x));
+            grid1_vals[l] = vec4(unpack8(grid_data.y));
+        }
+        // Preload B data for all j columns (reduce repeated index calculations)
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+            [[unroll]] for (uint l = 0; l < 2; ++l) {
+                const uint sign = sign_vals[l];
+                const uint sign7 = sign7_vals[l];
+                const vec4 grid0 = grid0_vals[l];
+                const vec4 grid1 = grid1_vals[l];
+                // Precompute indices
+                const uint b_idx = (j * p.batch_stride_b + b_offset + y_idx) / 4 + 2 * l;
+                const vec4 b0 = vec4(data_b_v4[b_idx + 0]);
+                const vec4 b4 = vec4(data_b_v4[b_idx + 1]);
+                sum +=
+                    fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
+                    fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
+                    fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
+                    fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w),
+                    fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x),
+                    fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
+                    fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
+                    fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 &  1) != 0 ? -grid1.w : grid1.w),
+                    FLOAT_TYPE(0.0)))))))));
             }
+            temp[j][n] = fma(FLOAT_TYPE(db_vals[n]), sum, temp[j][n]);
         }
         ibi += num_blocks_per_row;
     }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.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;
 
@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     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 signscale = pack32(u16vec2(

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_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;
 
@@ -10,7 +10,7 @@ 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;
+    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[ib32/2] >> (4 * (ib32 & 1))) & 0xF;

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.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;
 
@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     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 signscale = pack32(u16vec2(

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

@@ -8,12 +8,7 @@
 
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
-
-layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#include "mul_mat_vec_iface.glsl"
 
 layout (push_constant) uniform parameter
 {
@@ -29,6 +24,7 @@ layout (push_constant) uniform parameter
     uint nb03;
     uint nb13;
     uint nb23;
+    uint fusion_flags;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -117,6 +113,12 @@ void main() {
     }
 
     if (tid == 0) {
-        dst[idst] = tmp[0];
+        if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+            tmp[0] += FLOAT_TYPE(data_fuse0[idst]);
+        }
+        if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+            tmp[0] += FLOAT_TYPE(data_fuse1[idst]);
+        }
+        data_d[idst] = tmp[0];
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp

@@ -10,12 +10,7 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
-
-layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+#include "mul_mat_vec_iface.glsl"
 
 layout(constant_id = 0) const int BLOCK_SIZE = 32;
 // gqa_ratio is in the range [1,8]
@@ -29,6 +24,7 @@ layout (push_constant) uniform parameter
     uint nchannels_y;
     uint b_offset;
     uint d_offset;
+    uint fusion_flags;
 } p;
 
 #if !USE_SUBGROUP_ADD
@@ -148,7 +144,13 @@ void main() {
         [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
             // dst is not transposed and not permuted
             const uint idst = (channel + c)*nrows_dst + row_dst;
-            dst[idst] = temp[c];
+            if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
+                temp[c] += FLOAT_TYPE(data_fuse0[idst]);
+            }
+            if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS1) != 0) {
+                temp[c] += FLOAT_TYPE(data_fuse1[idst]);
+            }
+            data_d[idst] = temp[c];
         }
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.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;
 
@@ -15,7 +15,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads
@@ -41,9 +41,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
         const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
         const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
 
-        vec2 d = vec2(data_a[ib0 + i].d);
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        const FLOAT_TYPE_VEC2 dm = vec2(data_a[ib0 + i].dm);
 
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
@@ -75,7 +73,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
                        fma(FLOAT_TYPE(b96[l]),  sccache2[csel][ix][6 + 8*v_im],
                        fma(FLOAT_TYPE(b112[l]), sccache2[csel][ix][7 + 8*v_im], sum2))))))));
             }
-            temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
+            temp[j][n] = fma(dm.x, sum1, fma(-dm.y, sum2, temp[j][n]));
         }
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.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;
 
@@ -14,7 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -2,7 +2,7 @@
 
 #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;
 
@@ -13,10 +13,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
     const uint y2_idx = y1_idx + 128;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-        vec2 d = vec2(data_a[ib0 + i].d);
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
+        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
@@ -81,7 +79,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
                 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[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+            temp[j][n] = fma(dm.x, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dm.y, smin, temp[j][n]));
         }
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -2,7 +2,7 @@
 
 #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;
 
@@ -13,10 +13,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
     const uint y2_idx = y1_idx + 128;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-        vec2 d = vec2(data_a[ib0 + i].d);
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
+        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
@@ -113,7 +111,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
               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[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+            temp[j][n] = fma(dm.x, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dm.y, smin, temp[j][n]));
         }
     }
 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -2,7 +2,7 @@
 
 #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;
 
@@ -15,7 +15,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads