@novastera-oss/llamarn 0.3.1 → 0.4.1

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -305,6 +305,27 @@ void main() {
         return;
     }
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+
+            float ms = 1.0f;
+            float vs = 1.0f;
+
+            if (sink > Mf[r]) {
+                ms = exp(Mf[r] - sink);
+
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                    Of[r][d] *= ms;
+                }
+            } else {
+                vs = exp(sink - Mf[r]);
+            }
+
+            Lf[r] = Lf[r]*ms + vs;
+        }
+    }
+
     float Lfrcp[Br];
     [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
         Lfrcp[r] = 1.0 / Lf[r];

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp

@@ -9,6 +9,10 @@ layout (constant_id = 4) const uint32_t HSV = 32;
 layout (constant_id = 5) const uint32_t Clamp = 0;
 layout (constant_id = 6) const uint32_t D_split = 16;
 
+// Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths
+const uint32_t HSK_pad = (HSK + 15) & ~15;
+const uint32_t HSV_pad = (HSV + 15) & ~15;
+
 layout (push_constant) uniform parameter {
     uint32_t N;
     uint32_t KV;
@@ -50,10 +54,13 @@ layout (push_constant) uniform parameter {
     uint32_t k_num;
 } p;
 
+#define SINK_ENABLE_BIT (1<<24)
 #define MASK_ENABLE_BIT (1<<16)
 #define N_LOG2_MASK 0xFFFF
 
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
+layout (binding = 4) readonly buffer S {float data_s[];};
+
+layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
 
 #if defined(A_TYPE_PACKED16)
 #define BINDING_IDX_K 0
@@ -111,6 +118,14 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
     return ACC_TYPE(pow(base, ACC_TYPE(exph)));
 }
 
+// Load the sink value, indexed by Q's dimension 2.
+ACC_TYPE perElemOpGetSink(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
+{
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
+
+    return ACC_TYPE(data_s[h]);
+}
+
 uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
          iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
          q_stride, k_stride, v_stride, m_stride;

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -46,14 +46,14 @@ const uint32_t MatBc = 16;
 shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
 shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
 
-const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
+const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
 // Avoid padding for hsk==256 to make it fit in 48KB shmem.
 const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
 shared ACC_TYPE sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
+const uint32_t kshstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -74,6 +74,21 @@ void main() {
 
 #define tile_row(r) (row_tid * rows_per_thread + (r))
 
+    // Zero-initialize shared memory for Q/K when HSK is not a multiple of 16 (HSK_pad > HSK).
+    if ((HSK % 16) != 0) {
+        [[unroll]] for (uint i = 0; i < Br * qstride; i += gl_WorkGroupSize.x) {
+            if (i + tid < Br * qstride) {
+                Qf[i + tid] = f16vec4(0);
+            }
+        }
+        [[unroll]] for (uint i = 0; i < Bc * kshstride; i += gl_WorkGroupSize.x) {
+            if (i + tid < Bc * kshstride) {
+                ksh[i + tid] = f16vec4(0);
+            }
+        }
+        barrier();
+    }
+
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
@@ -151,14 +166,14 @@ void main() {
         }
         barrier();
 
-        // K * Q^T -> S^T: Bc x HSK * HSK x Br -> Bc x Br
+        // K * Q^T -> S^T: Bc x HSK_pad * HSK_pad x Br -> Bc x Br
         // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
         // This is written transposed in order to allow for N being 8 if implementations need it
         coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
         coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
-        for (uint32_t d = 0; d < HSK / 16; ++d) {
+        for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
@@ -210,7 +225,7 @@ void main() {
 
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+                Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             }
         }
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -233,7 +248,7 @@ void main() {
                 vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
 #endif
                 [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                    Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf);
+                    Of[r][d] += ACC_TYPE(Pf[r]) * ACC_TYPEV4(Vf);
                 }
             }
         }
@@ -288,7 +303,7 @@ void main() {
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
-            Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+            Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
 
             barrier();
@@ -329,6 +344,27 @@ void main() {
         return;
     }
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+
+            float ms = 1.0f;
+            float vs = 1.0f;
+
+            if (sink > Mf[r]) {
+                ms = exp(Mf[r] - sink);
+
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                    Of[r][d] *= ACC_TYPE(ms);
+                }
+            } else {
+                vs = exp(sink - Mf[r]);
+            }
+
+            Lf[r] = Lf[r]*ms + vs;
+        }
+    }
+
     float Lfrcp[rows_per_thread];
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         Lfrcp[r] = 1.0 / Lf[r];
@@ -336,7 +372,7 @@ void main() {
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            Of[r][d] *= float16_t(Lfrcp[r]);
+            Of[r][d] *= ACC_TYPE(Lfrcp[r]);
         }
     }
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -104,16 +104,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA> Qf16;
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseAccumulator> Q;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA> Qf16;
 
     uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
-    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -140,10 +140,10 @@ void main() {
 
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, HSK, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK_pad, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK_pad), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
         if (p.logit_softcap != 0.0f) {
@@ -208,31 +208,31 @@ void main() {
         rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
         rowsum = coopMatMulAdd(P_A, One, rowsum);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV_pad, gl_MatrixUseB> V;
         uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
-        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV_pad) DECODEFUNC);
 
         L = eM*L + rowsum;
 
         // This is the "diagonal" matrix in the paper, but since we do componentwise
         // multiply rather than matrix multiply it has the diagonal element smeared
         // across the row
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> eMdiag;
 
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
         // multiply with fp16 accumulation, then add to O.
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(0);
         PV = coopMatMulAdd(P_A, V, PV);
 
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(PV);
     }
 
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
+        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
 
         uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
@@ -243,11 +243,39 @@ void main() {
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
+    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> S;
+        coopMatPerElementNV(S, S, perElemOpGetSink, iq2);
+
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> Mr;
+
+        // resize M by using smear/reduce
+        coopMatReduceNV(Mr, M, gl_CooperativeMatrixReduceRowNV, smearReduce);
+
+        // O, Ldiag, Mr all have the same type so all element locations match
+        [[unroll]] for (uint32_t i = 0; i < Ldiag.length(); ++i) {
+            ACC_TYPE sink = S[i];
+
+            ACC_TYPE ms = ACC_TYPE(1.0f);
+            ACC_TYPE vs = ACC_TYPE(1.0f);
+
+            if (sink > Mr[i]) {
+                ms = exp(Mr[i] - sink);
+
+                O[i] *= ms;
+            } else {
+                vs = exp(sink - Mr[i]);
+            }
+
+            Ldiag[i] = Ldiag[i]*ms + vs;
+        }
+    }
+
     [[unroll]]
     for (int k = 0; k < Ldiag.length(); ++k) {
         Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k];
@@ -257,7 +285,7 @@ void main() {
 
     uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
     } else {
@@ -267,6 +295,6 @@ void main() {
         // permute dimensions
         tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
 
-        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), tensorViewPermute);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV_pad), tensorViewPermute);
     }
 }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp

@@ -7,13 +7,15 @@ layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {float data_a[];};
-layout (binding = 1) writeonly buffer D {float data_d[];};
+layout (binding = 1) readonly buffer B {float data_s[];};
+layout (binding = 2) writeonly buffer D {float data_d[];};
 
 layout (push_constant) uniform parameter {
     uint D;
     uint N;
     uint ne3;
     uint k_num;
+    uint sinks;
 } p;
 
 shared float tmpsh[BLOCK_SIZE];
@@ -73,6 +75,22 @@ void main() {
     }
     L = tmpsh[0];
 
+    float sink;
+    if (p.sinks != 0) {
+        sink = data_s[n];
+
+        float ms = 1.0f;
+        float vs = 1.0f;
+
+        if (sink > m_max) {
+            ms = exp(m_max - sink);
+        } else {
+            vs = exp(sink - m_max);
+        }
+
+        L = L*ms + vs;
+    }
+
     L = 1.0 / L;
 
     // D dimension is split across workgroups in the y dimension
@@ -85,6 +103,13 @@ void main() {
             float m = data_a[m_offset + k * lm_stride];
             O += exp(m - m_max) * data_a[o_offset];
         }
+        if (p.sinks != 0) {
+            if (sink > m_max) {
+                float ms = 1.0f;
+                ms = exp(m_max - sink);
+                O *= ms;
+            }
+        }
         O *= L;
         data_d[iq3 * D * N + D * n + d] = O;
     }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp

@@ -2,6 +2,7 @@
 #extension GL_EXT_control_flow_attributes : require
 
 #include "rte.comp"
+#include "utils.comp"
 
 layout (push_constant) uniform parameter
 {
@@ -28,25 +29,9 @@ uint get_aoffset() { return p.misalign_offsets >> 16; }
 uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
 uint get_doffset() { return p.misalign_offsets & 0xFF; }
 
-// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
-uint fastmod(uint a, uint b) {
-    if ((b & (b-1)) == 0) {
-        return a & (b-1);
-    }
-    return a % b;
-}
-
-uint fastdiv(uint a, uint b) {
-    return (a < b) ? 0 : (a / b);
-}
 
 void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
-    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
 }
 
 uint src0_idx(uint i00, uint i01, uint i02, uint i03) {

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp

@@ -14,4 +14,6 @@ layout (push_constant) uniform parameter
     uint ne00;
     uint ne20;
     uint mode;
+    float alpha;
+    float limit;
 } p;

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp

@@ -1,6 +1,10 @@
 #extension GL_EXT_control_flow_attributes : enable
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
+#if USE_SUBGROUP_ADD
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#endif
 
 #ifdef MUL_MAT_ID
 #define EXPERT_COUNT 8
@@ -90,7 +94,38 @@ layout (constant_id = 2) const uint NUM_COLS = 1;
 
 shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
 
-void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    // subgroupAdd is probably faster on devices that support it,
+    // particularly when the workgroup has more than one subgroup
+#if USE_SUBGROUP_ADD
+    // sum up partial sums within a subgroup
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            temp[j][n] = subgroupAdd(temp[j][n]);
+        }
+    }
+
+    // Go through shared memory to sum partials across subgroups
+    if (gl_SubgroupInvocationID == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[j][n][gl_SubgroupID] = temp[j][n];
+            }
+        }
+    }
+    barrier();
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                temp[j][n] = FLOAT_TYPE(0);
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    temp[j][n] += tmpsh[j][n][s];
+                }
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
+            }
+        }
+    }
+#else
     // sum up partial sums and write back result
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -115,4 +150,5 @@ void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32
             }
         }
     }
+#endif
 }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp

@@ -26,6 +26,9 @@ layout (push_constant) uniform parameter
     uint ne12;
     uint b_offset;
     uint d_offset;
+    uint nb03;
+    uint nb13;
+    uint nb23;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -34,6 +37,7 @@ void main() {
     const uint tid       = gl_LocalInvocationID.x;
     const uint row_x     = gl_GlobalInvocationID.y;
     const uint channel   = gl_GlobalInvocationID.z;
+    const uint i3        = gl_WorkGroupID.x;
     const uint channel_x = channel / p.channel_x_divisor;
     const uint channel_y = channel % p.ne12;
 
@@ -41,7 +45,7 @@ void main() {
     const uint nrows_dst = p.nrows_x;
     const uint row_dst   = row_x;
 
-    const uint idst = channel*nrows_dst + row_dst;
+    const uint idst = i3*p.nb23 + channel*nrows_dst + row_dst;
 
     FLOAT_TYPE temp = 0.0f;
 
@@ -58,8 +62,8 @@ void main() {
 
                 const uint row_y = col_x;
 
-                const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-                const uint iy = channel_y*p.channel_stride_y + row_y;
+                const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+                const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
                 const vec4 av4 = vec4(data_a_v4[ix / 4]);
                 const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -74,8 +78,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const vec4 av4 = vec4(data_a_v4[ix / 4]);
             const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -91,8 +95,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);