@novastera-oss/llamarn 0.0.1-alpha.4

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

@@ -0,0 +1,442 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+
+#extension GL_EXT_integer_dot_product : require
+
+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
+
+#ifdef COOPMAT
+#extension GL_KHR_cooperative_matrix : enable
+#extension GL_KHR_memory_scope_semantics : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
+
+#ifdef MUL_MAT_ID
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#endif
+
+#include "types.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];};
+#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 {block_q8_1_packed32 data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+#ifdef MUL_MAT_ID
+layout (binding = 3) readonly buffer IDS {int data_ids[];};
+#endif
+
+layout (push_constant) uniform parameter
+{
+    uint M;
+    uint N;
+    uint K;
+    uint stride_a;
+    uint stride_b;
+    uint stride_d;
+
+    uint batch_stride_a;
+    uint batch_stride_b;
+    uint batch_stride_d;
+
+#ifdef MUL_MAT_ID
+    uint nei0;
+    uint nei1;
+    uint nbi1;
+    uint ne11;
+#else
+    uint k_split;
+    uint ne02;
+    uint ne12;
+    uint broadcast2;
+    uint broadcast3;
+#endif
+} p;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 64;
+layout (constant_id = 1) const uint BM = 64;
+layout (constant_id = 2) const uint BN = 64;
+// layout (constant_id = 3) const uint BK = 32;
+layout (constant_id = 4) const uint WM = 32;
+layout (constant_id = 5) const uint WN = 32;
+layout (constant_id = 6) const uint WMITER = 2;
+layout (constant_id = 7) const uint TM = 4;
+layout (constant_id = 8) const uint TN = 2;
+layout (constant_id = 9) const uint TK = 1;  // Only needed for coopmat
+layout (constant_id = 10) const uint WARP = 32;
+
+#define BK 32
+
+#ifdef COOPMAT
+#define SHMEM_STRIDE (BK / 4 + 4)
+#else
+#define SHMEM_STRIDE (BK / 4 + 1)
+#endif
+
+shared int32_t buf_a_qs[BM * SHMEM_STRIDE];
+
+#ifndef COOPMAT
+#if QUANT_AUXF == 1
+shared FLOAT_TYPE buf_a_dm[BM];
+#else
+shared FLOAT_TYPE_VEC2 buf_a_dm[BM];
+#endif
+#endif
+
+shared int32_t buf_b_qs[BN * SHMEM_STRIDE];
+#ifndef COOPMAT
+shared FLOAT_TYPE_VEC2 buf_b_ds[BN];
+#endif
+
+#define LOAD_VEC_A (4 * QUANT_R)
+#define LOAD_VEC_B 4
+
+#ifdef MUL_MAT_ID
+shared u16vec2 row_ids[4096];
+#endif // MUL_MAT_ID
+
+#define NUM_WARPS (BLOCK_SIZE / WARP)
+
+#ifdef COOPMAT
+shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
+#endif
+
+#include "mul_mmq_funcs.comp"
+
+void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.z;
+#else
+    const uint batch_idx = gl_GlobalInvocationID.z;
+
+    const uint i13 = batch_idx / p.ne12;
+    const uint i12 = batch_idx % p.ne12;
+
+    const uint i03 = i13 / p.broadcast3;
+    const uint i02 = i12 / p.broadcast2;
+
+    const uint batch_idx_a = i03 * p.ne02 + i02;
+#endif
+
+    const uint blocks_m = (p.M + BM - 1) / BM;
+    const uint ir = gl_WorkGroupID.x % blocks_m;
+    const uint ik = gl_WorkGroupID.x / blocks_m;
+    const uint ic = gl_WorkGroupID.y;
+
+    const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
+    const uint WSUBM = WM / WMITER;
+    const uint WSUBN = WN / WNITER;
+
+#ifdef COOPMAT
+    const uint warp_i = gl_SubgroupID;
+
+    const uint tiw = gl_SubgroupInvocationID;
+
+    const uint cms_per_row = WM / TM;
+    const uint cms_per_col = WN / TN;
+
+    const uint storestride = WARP / TM;
+    const uint store_r = tiw % TM;
+    const uint store_c = tiw / TM;
+#else
+    const uint warp_i = gl_LocalInvocationID.x / WARP;
+
+    const uint tiw = gl_LocalInvocationID.x % WARP;
+
+    const uint tiwr = tiw % (WSUBM / TM);
+    const uint tiwc = tiw / (WSUBM / TM);
+#endif
+
+    const uint warp_r = warp_i % (BM / WM);
+    const uint warp_c = warp_i / (BM / WM);
+
+    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
+    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
+    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
+    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
+
+    const uint loadstride_a = BLOCK_SIZE * LOAD_VEC_A / BK;
+    const uint loadstride_b = BLOCK_SIZE * LOAD_VEC_B / BK;
+
+#ifdef MUL_MAT_ID
+    uint _ne1 = 0;
+    for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
+        for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
+            if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
+                row_ids[_ne1] = u16vec2(ii0, ii1);
+                _ne1++;
+            }
+        }
+    }
+
+    barrier();
+
+    // Workgroup has no work
+    if (ic * BN >= _ne1) return;
+#endif
+
+#ifdef MUL_MAT_ID
+    const uint start_k = 0;
+    const uint end_k = p.K;
+#else
+    const uint start_k = ik * p.k_split;
+    const uint end_k = min(p.K, (ik + 1) * p.k_split);
+#endif
+
+    uint pos_a_ib = (
+#ifdef MUL_MAT_ID
+        expert_idx * p.batch_stride_a +
+#else
+        batch_idx_a * p.batch_stride_a +
+#endif
+        ir * BM * p.stride_a + start_k) / BK;
+#ifdef MUL_MAT_ID
+    uint pos_b_ib = 0;
+#else
+    uint pos_b_ib = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / BK;
+#endif
+
+#ifdef COOPMAT
+    coopmat<int8_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
+    coopmat<int8_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+    coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_result;
+
+    coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> factors[cms_per_row * cms_per_col];
+
+    coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
+
+    [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
+        sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
+    }
+#else
+    int32_t cache_a_qs[WMITER * TM * BK / 4];
+
+    int32_t cache_b_qs[TN * BK / 4];
+
+    ACC_TYPE sums[WMITER * TM * WNITER * TN];
+
+    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
+        sums[i] = ACC_TYPE(0.0f);
+    }
+#endif
+
+#if QUANT_AUXF == 1
+    FLOAT_TYPE cache_a_dm[WMITER * TM];
+#else
+    FLOAT_TYPE_VEC2 cache_a_dm[WMITER * TM];
+#endif
+
+    FLOAT_TYPE_VEC2 cache_b_ds[TN];
+
+    for (uint block = start_k; block < end_k; block += BK) {
+        [[unroll]] for (uint l = 0; loadc_a + l < BM; l += loadstride_a) {
+            const uint ib = pos_a_ib + (loadc_a + l) * p.stride_a / BK;
+            const uint iqs = loadr_a;
+            const uint buf_ib = loadc_a + l;
+
+            if (iqs == 0) {
+#if QUANT_AUXF == 1
+                buf_a_dm[buf_ib] = get_d(ib);
+#else
+                buf_a_dm[buf_ib] = get_dm(ib);
+#endif
+            }
+#if QUANT_R == 1
+            buf_a_qs[buf_ib * SHMEM_STRIDE + iqs] = repack(ib, iqs);
+#else
+            const i32vec2 vals = repack(ib, iqs);
+            buf_a_qs[buf_ib * SHMEM_STRIDE + iqs    ] = vals.x;
+            buf_a_qs[buf_ib * SHMEM_STRIDE + iqs + 4] = vals.y;
+#endif
+        }
+        [[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) {
+#ifdef MUL_MAT_ID
+            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
+            const uint idx = pos_b_ib + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
+            const uint ib = idx / 8;
+            const uint iqs = idx & 0x7;
+#else
+            const uint ib = pos_b_ib + (loadc_b + l) * p.stride_b / BK;
+            const uint iqs = loadr_b;
+#endif
+
+            const uint buf_ib = loadc_b + l;
+
+            if (iqs == 0) {
+                buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib].ds);
+            }
+            buf_b_qs[buf_ib * SHMEM_STRIDE + iqs] = data_b[ib].qs[iqs];
+        }
+
+        barrier();
+
+        pos_a_ib += 1;
+        pos_b_ib += 1;
+
+#ifdef COOPMAT
+        [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+            const uint ib_a = warp_r * WM + cm_row * TM;
+            // Load from shared into cache
+            coopMatLoad(cache_a, buf_a_qs, ib_a * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);
+
+            // TODO: only cache values that are actually needed
+            [[unroll]] for (uint t_idx = 0; t_idx < TM; t_idx++) {
+                cache_a_dm[t_idx] = buf_a_dm[ib_a + t_idx];
+            }
+
+            [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                const uint ib_b = warp_c * WN + cm_col * TN;
+                coopMatLoad(cache_b, buf_b_qs, ib_b * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);
+
+                // TODO: only cache values that are actually needed
+                [[unroll]] for (uint t_idx = 0; t_idx < TN; t_idx++) {
+                    cache_b_dm[t_idx] = buf_b_d[ib_b + t_idx];
+                }
+
+                cm_result = coopmat<int32_t, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0);
+                cm_result = coopMatMulAdd(cache_a, cache_b, cm_result);
+
+                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
+                    coopmat_stage[warp_i * TM * TN + (store_c + col) * TM + store_r] = ACC_TYPE(float(cache_a_d[store_r]) * float(cache_b_d[store_c + col]));
+                }
+
+                coopMatLoad(factors, coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
+                sums[cm_col * cms_per_row + cm_row] += factors * coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(cm_result);
+            }
+        }
+#else
+        // Load from shared into cache
+        [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+            [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+                const uint ib = warp_r * WM + wsir * WSUBM + tiwr * TM + cr;
+                cache_a_dm[wsir * TM + cr] = buf_a_dm[ib];
+                [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
+                    cache_a_qs[(wsir * TM + cr) * (BK / 4) + idx_k] = buf_a_qs[ib * SHMEM_STRIDE + idx_k];
+                }
+            }
+        }
+
+        [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+            [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                const uint ib = warp_c * WN + wsic * WSUBN + tiwc * TN + cc;
+                cache_b_ds[cc] = buf_b_ds[ib];
+                [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
+                    cache_b_qs[cc * (BK / 4) + idx_k] = buf_b_qs[ib * SHMEM_STRIDE + idx_k];
+                }
+            }
+
+            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+                [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                    [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+                        const uint cache_a_idx = wsir * TM + cr;
+                        const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
+                        int32_t q_sum = 0;
+                        [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) {
+                            q_sum += dotPacked4x8EXT(cache_a_qs[cache_a_idx * (BK / 4) + idx_k],
+                                                     cache_b_qs[cc * (BK / 4) + idx_k]);
+                        }
+
+                        sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc]);
+                    }
+                }
+            }
+        }
+#endif
+
+        barrier();
+    }
+
+    const uint dr = ir * BM + warp_r * WM;
+    const uint dc = ic * BN + warp_c * WN;
+
+#ifndef MUL_MAT_ID
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+#endif
+
+#ifdef COOPMAT
+#ifdef MUL_MAT_ID
+    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+            coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
+
+            [[unroll]] for (uint col = 0; col < BN; col += storestride) {
+                const uint row_i = dc + cm_col * TN + col + store_c;
+                if (row_i >= _ne1) break;
+
+                const u16vec2 row_idx = row_ids[row_i];
+
+                data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+            }
+        }
+    }
+#else
+    const bool is_aligned = p.stride_d % 4 == 0;  // Assumption: D_TYPE == float
+
+    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+            const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N;
+
+            if (is_aligned && is_in_bounds) {
+                // Full coopMat is within bounds and stride_d is aligned with 16B
+                coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_dtype = coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(sums[cm_col * cms_per_row + cm_row]);
+                coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor);
+            } else if (is_in_bounds) {
+                // Full coopMat is within bounds, but stride_d is not aligned
+                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
+
+                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
+                    data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                }
+            } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) {
+                // Partial coopMat is within bounds
+                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
+
+                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
+                    if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) {
+                        data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                    }
+                }
+            }
+        }
+    }
+#endif // MUL_MAT_ID
+#else
+    [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+        [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+
+            const uint dr_warp = dr + wsir * WSUBM + tiwr * TM;
+            const uint dc_warp = dc + wsic * WSUBN + tiwc * TN;
+            [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+#ifdef MUL_MAT_ID
+                const uint row_i = dc_warp + cc;
+                if (row_i >= _ne1) break;
+
+                const u16vec2 row_idx = row_ids[row_i];
+#endif // MUL_MAT_ID
+                [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+#ifdef MUL_MAT_ID
+                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+#else
+                    if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    }
+#endif // MUL_MAT_ID
+                }
+            }
+        }
+    }
+#endif // COOPMAT
+}

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

@@ -0,0 +1,99 @@
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+
+#include "types.comp"
+
+// Each iqs value maps to a 32-bit integer
+
+#if defined(DATA_A_Q4_0)
+i32vec2 repack(uint ib, uint iqs) {
+    // Use 2-byte loads since a q4_0 block (18 bytes) is not divisible by 4
+    const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2    ],
+                                   data_a[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    return i32vec2( vui       & 0x0F0F0F0F,
+                   (vui >> 4) & 0x0F0F0F0F);
+}
+
+ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
+    return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y));
+}
+#endif
+
+#if defined(DATA_A_Q4_1)
+i32vec2 repack(uint ib, uint iqs) {
+    // Use 4-byte loads since a q4_1 block (20 bytes) is divisible by 4
+    const uint32_t vui = data_a_packed32[ib].qs[iqs];
+    return i32vec2( vui       & 0x0F0F0F0F,
+                   (vui >> 4) & 0x0F0F0F0F);
+}
+
+ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
+    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
+}
+#endif
+
+#if defined(DATA_A_Q5_0)
+i32vec2 repack(uint ib, uint iqs) {
+    // Use 2-byte loads since a q5_0 block (22 bytes) is not divisible by 4
+    const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2    ],
+                                   data_a[ib].qs[iqs * 2 + 1]);
+    const uint32_t vui = pack32(quants);
+    const int32_t qh = int32_t((uint32_t(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]) >> (4 * iqs));
+    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
+                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
+
+    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
+                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
+
+    return i32vec2(v0, v1);
+}
+
+ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
+    return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y));
+}
+#endif
+
+#if defined(DATA_A_Q5_1)
+i32vec2 repack(uint ib, uint iqs) {
+    // Use 4-byte loads since a q5_1 block (24 bytes) is divisible by 4
+    const uint32_t vui = data_a_packed32[ib].qs[iqs];
+    const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs));
+    const int32_t v0 = int32_t(vui & 0x0F0F0F0F)
+                     | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28)
+
+    const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F)
+                     | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28)
+
+    return i32vec2(v0, v1);
+}
+
+ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) {
+    return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y);
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+int32_t repack(uint ib, uint iqs) {
+    // Use 2-byte loads since a q8_0 block (34 bytes) is not divisible by 4
+    return pack32(i16vec2(data_a[ib].qs[iqs * 2    ],
+                          data_a[ib].qs[iqs * 2 + 1]));
+}
+
+ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) {
+    return ACC_TYPE(float(q_sum) * da * dsb.x);
+}
+#endif
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+FLOAT_TYPE get_d(uint ib) {
+    return FLOAT_TYPE(data_a[ib].d);
+}
+#endif
+
+#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
+FLOAT_TYPE_VEC2 get_dm(uint ib) {
+    return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm);
+}
+#endif

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

@@ -0,0 +1,44 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared vec2 sum[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    sum[tid] = vec2(0.0f, 0.0f);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const float xi = float(data_a[row*p.KX + col]);
+        sum[tid].x += xi;
+        sum[tid].y += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum[tid] += sum[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = sum[0].x / p.KX;
+    const float var = sum[0].y / p.KX - mean * mean;
+    const float inv_std = inversesqrt(var + p.param1);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
+    }
+}

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

@@ -0,0 +1,42 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) buffer X {A_TYPE x[];};
+layout (binding = 1) readonly buffer G {A_TYPE grad[];};
+layout (binding = 2) buffer GM {A_TYPE gradm[];};
+layout (binding = 3) buffer GV {A_TYPE gradv[];};
+layout (binding = 4) readonly buffer P {float params[7];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float alpha  = params[0];
+    const float beta1  = params[1];
+    const float beta2  = params[2];
+    const float eps    = params[3];
+    const float wd     = params[4];
+    const float beta1h = params[5];
+    const float beta2h = params[6];
+
+    const float gi = grad[i];
+    const float gmi = gradm[i]*beta1 +    gi*(1.0f - beta1);
+    const float gvi = gradv[i]*beta2 + gi*gi*(1.0f - beta2);
+
+    gradm[i] = gmi;
+    gradv[i] = gvi;
+
+    const float mh =      gmi*beta1h;
+    const float vh = sqrt(gvi*beta2h) + eps;
+
+    x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
+}

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

@@ -0,0 +1,28 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
+    const uint i2 = (idx - i3_offset) / (p.ne11*p.ne10);
+    const uint i2_offset = i2*p.ne11*p.ne10;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne10;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint dst_idx = i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0*p.nb10;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+    data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : 0.0f);
+}