whispercpp 1.3.4 → 1.3.6

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

@@ -28,7 +28,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
 #endif
 
-#include "types.comp"
+#include "types.glsl"
 
 #ifndef LOAD_VEC_A
 #define LOAD_VEC_A 1
@@ -68,6 +68,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
+layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
 #endif
 
 layout (push_constant) uniform parameter
@@ -89,6 +90,8 @@ layout (push_constant) uniform parameter
     uint nbi1;
     uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
     uint k_split;
     uint ne02;
     uint ne12;
@@ -100,7 +103,6 @@ layout (push_constant) uniform parameter
 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 = 16;  // Assumed to be 32 if working with a quant
 layout (constant_id = 4) const uint WM = 32;
 layout (constant_id = 5) const uint WN = 32;
 layout (constant_id = 6) const uint WMITER = 2;
@@ -109,6 +111,14 @@ 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;
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#define BK 32
+#define BK_STEP 4
+#else
+layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
+#define BK_STEP 2
+#endif
+
 #ifdef COOPMAT
 #define SHMEM_STRIDE (BK / 2 + 4)
 #else
@@ -120,92 +130,28 @@ shared FLOAT_TYPE_VEC2 buf_b[BN * SHMEM_STRIDE];
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)
 
-#ifdef MUL_MAT_ID
-shared u16vec2 row_ids[BN];
-uint _ne1;
-
-#ifdef MUL_MAT_ID_USE_SUBGROUPS
-shared uvec4 ballots_sh[NUM_WARPS];
-
-void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
-    _ne1 = 0;
-    uint num_elements = p.nei1 * p.nei0;
-    uint nei0shift = findLSB(p.nei0);
-
-    uint ids[16];
-    uint iter = 0;
-
-    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
-        // prefetch up to 16 elements
-        if (iter == 0) {
-            [[unroll]] for (uint k = 0; k < 16; ++k) {
-                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
-                bool in_range = i < num_elements;
-                uint ii1;
-                if (nei0_is_pow2) {
-                    ii1 = i >> nei0shift;
-                } else {
-                    ii1 = i / p.nei0;
-                }
-                uint ii0 = i - ii1 * p.nei0;
-                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-            }
-        }
-        uint i = j + gl_LocalInvocationIndex;
-        bool in_range = i < num_elements;
-        uint ii1;
-        if (nei0_is_pow2) {
-            ii1 = i >> nei0shift;
-        } else {
-            ii1 = i / p.nei0;
-        }
-        uint ii0 = i - ii1 * p.nei0;
-        uint id = ids[iter++];
-        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-
-        ballots_sh[gl_SubgroupID] = ballot;
-        barrier();
-
-        uint subgroup_base = 0;
-        uint total = 0;
-        for (uint k = 0; k < gl_NumSubgroups; ++k) {
-            if (k == gl_SubgroupID) {
-                subgroup_base = total;
-            }
-            total += subgroupBallotBitCount(ballots_sh[k]);
-        }
-        barrier();
-
-        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
-        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
-            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
-        }
-        _ne1 += total;
-        iter &= 15;
-        if (_ne1 >= (ic + 1) * BN) {
-            break;
-        }
-    }
-    barrier();
-}
-#endif // MUL_MAT_ID_USE_SUBGROUPS
-#endif // MUL_MAT_ID
-
 #ifdef COOPMAT
 shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
 
-#include "mul_mm_funcs.comp"
+#include "mul_mm_id_funcs.glsl"
+#include "mul_mm_funcs.glsl"
 
 void main() {
+    const uint ic = gl_WorkGroupID.y;
+
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_WorkGroupID.z;
+    if (ic * BN >= data_expert_count[expert_idx]) {
+        return;
+    }
+#endif
 #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;
+#ifndef MUL_MAT_ID
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;
 
     const uint i13 = batch_idx / p.ne12;
     const uint i12 = batch_idx % p.ne12;
@@ -219,7 +165,6 @@ void main() {
     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;
@@ -291,13 +236,13 @@ void main() {
     const uint end_k = min(p.K, (ik + 1) * p.k_split);
 #endif
 
-    uint pos_a = (
+    uint pos_a =
 #ifdef MUL_MAT_ID
-        expert_idx * p.batch_stride_a +
+        expert_idx * (p.batch_stride_a / LOAD_VEC_A) +
 #else
-        batch_idx_a * p.batch_stride_a +
+        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A) +
 #endif
-        ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
+        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
 #ifdef MUL_MAT_ID
     uint pos_b = 0;
 #else
@@ -313,12 +258,17 @@ void main() {
         sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
     }
 #else
-    ACC_TYPE sums[WMITER * TM * WNITER * TN];
+    ACC_TYPE_VEC2 sums[WMITER * TM * WNITER * TN/2];
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+    FLOAT_TYPE_VEC4 cache_a[WMITER * TM];
+    FLOAT_TYPE_VEC4 cache_b;
+#else
     FLOAT_TYPE_VEC2 cache_a[WMITER * TM];
-    FLOAT_TYPE_VEC2 cache_b[TN];
+    FLOAT_TYPE_VEC2 cache_b;
+#endif
 
-    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
-        sums[i] = ACC_TYPE(0.0f);
+    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
+        sums[i] = ACC_TYPE_VEC2(0.0f, 0.0f);
     }
 #endif
 
@@ -353,27 +303,46 @@ void main() {
             }
         }
 #else
-        [[unroll]] for (uint i = 0; i < BK / 2; i++) {
+        [[unroll]] for (uint i = 0; i < BK / BK_STEP; i++) {
             // Load from shared into cache
             [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                 [[unroll]] for (uint j = 0; j < TM; j++) {
+                #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                    cache_a[wsir * TM + j].xy = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i    ];
+                    cache_a[wsir * TM + j].zw = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i + 1];
+                #else
                     cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i];
+                #endif
                 }
             }
-            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
-                [[unroll]] for (uint j = 0; j < TN; j++) {
-                    cache_b[j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i];
-                }
 
-                [[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 sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
-                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr].x), ACC_TYPE(cache_b[cc].x), fma(ACC_TYPE(cache_a[wsir * TM + cr].y), ACC_TYPE(cache_b[cc].y), sums[sums_idx]));
+            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+                [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                    cache_b.xy = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i    ];
+                    cache_b.zw = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i + 1];
+                #else
+                    cache_b = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + i];
+                #endif
+
+                    [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+                        [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
+                            // [WNITER][TN][WMITER][TM / 2] -> [wsic][cc][wsir][cr]
+                            const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
+                        #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                            sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y),
+                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].w), ACC_TYPE(cache_b.w), sums[sums_idx].x))));
+                            sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y),
+                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].w), ACC_TYPE(cache_b.w), sums[sums_idx].y))));
+                        #else
+                            sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y), sums[sums_idx].x));
+                            sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y), sums[sums_idx].y));
+                        #endif
                         }
                     }
                 }
             }
+
         }
 #endif
 
@@ -388,8 +357,9 @@ void main() {
         }
     }
 #else
-    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
-        sums[i] = clamp(sums[i], -ACC_TYPE_MAX, ACC_TYPE_MAX);
+    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
+        sums[i].x = clamp(sums[i].x, -ACC_TYPE_MAX, ACC_TYPE_MAX);
+        sums[i].y = clamp(sums[i].y, -ACC_TYPE_MAX, ACC_TYPE_MAX);
     }
 #endif
 #endif
@@ -398,7 +368,7 @@ void main() {
     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;
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif
 
 #ifdef COOPMAT
@@ -407,6 +377,7 @@ void main() {
         [[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);
 
+            barrier();
             [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                 const uint row_i = dc + cm_col * TN + col + store_c;
                 if (row_i >= _ne1) break;
@@ -417,6 +388,7 @@ void main() {
                     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]);
                 }
             }
+            barrier();
         }
     }
 #else
@@ -434,18 +406,22 @@ void main() {
                 // 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);
 
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
                 [[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]);
                 }
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
             } 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);
 
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
                 [[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]);
                     }
                 }
+                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
             }
         }
     }
@@ -463,14 +439,21 @@ void main() {
 
                 const u16vec2 row_idx = row_ids[row_i - ic * BN];
 #endif // MUL_MAT_ID
-                [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+                [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
+                    const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
 #ifdef MUL_MAT_ID
-                    if (dr_warp + cr < p.M) {
-                        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]);
+                    if (dr_warp + 2 * cr < p.M) {
+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + 2 * cr] = D_TYPE(sums[sums_idx].x);
+                    }
+                    if (dr_warp + 2 * cr + 1 < p.M) {
+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + 2 * cr + 1] = D_TYPE(sums[sums_idx].y);
                     }
 #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]);
+                    if (dr_warp + 2 * cr < p.M && dc_warp + cc < p.N) {
+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + 2 * cr] = D_TYPE(sums[sums_idx].x);
+                    }
+                    if (dr_warp + 2 * cr + 1 < p.M && dc_warp + cc < p.N) {
+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + 2 * cr + 1] = D_TYPE(sums[sums_idx].y);
                     }
 #endif // MUL_MAT_ID
                 }

data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -18,8 +18,8 @@
 #extension GL_EXT_bfloat16 : enable
 #endif
 
-#include "types.comp"
-#include "utils.comp"
+#include "types.glsl"
+#include "utils.glsl"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
@@ -53,6 +53,8 @@ layout (push_constant) uniform parameter
     uint nbi1;
     uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
     uint k_split;
     uint ne02;
     uint ne12;
@@ -71,7 +73,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #if QUANT_K > 1
 #define DECODEFUNCA , dequantFuncA
 
-#include "dequant_funcs_cm2.comp"
+#include "dequant_funcs_cm2.glsl"
 
 #else
 #define DECODEFUNCA
@@ -92,6 +94,7 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
+layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
 
 shared u16vec4 row_ids[BN];
 
@@ -107,11 +110,7 @@ B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const i
 {
     const uint row_i = blockCoords[0];
 
-    if (row_i >= _ne1) {
-        return B_TYPE(0.0);
-    }
-
-    const u16vec4 row_idx = row_ids[row_i & (BN - 1)];
+    const u16vec4 row_idx = row_ids[row_i];
     B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]];
 
     return ret;
@@ -138,6 +137,8 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
     uint ids[16];
     uint iter = 0;
 
+    uint expert_count = data_expert_count[expert_idx];
+
     for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
         // prefetch up to 16 elements
         if (iter == 0) {
@@ -166,7 +167,9 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
         uint id = ids[iter++];
         uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
 
-        ballots_sh[gl_SubgroupID] = ballot;
+        if (gl_SubgroupInvocationID == 0) {
+            ballots_sh[gl_SubgroupID] = ballot;
+        }
         barrier();
 
         uint subgroup_base = 0;
@@ -185,7 +188,7 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
         }
         _ne1 += total;
         iter &= 15;
-        if (_ne1 >= (ic + 1) * BN) {
+        if (_ne1 >= (ic + 1) * BN || _ne1 == expert_count) {
             break;
         }
     }
@@ -194,16 +197,29 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
 #endif
 
 void main() {
+    const uint tid = gl_LocalInvocationIndex;
+    const uint ic = gl_WorkGroupID.y;
+
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_WorkGroupID.z;
+    if (ic * BN >= data_expert_count[expert_idx]) {
+        return;
+    }
+    // initialize to row 0 so we don't need to bounds check
+    if (tid < BN) {
+        row_ids[tid] = u16vec4(0);
+    }
+#if !defined(NEEDS_INIT_IQ_SHMEM)
+    barrier();
+#endif
+#endif
+
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-    const uint tid = gl_LocalInvocationIndex;
-
-#ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
-#else
-    const uint batch_idx = gl_GlobalInvocationID.z;
+#ifndef MUL_MAT_ID
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;
 
     const uint i13 = batch_idx / p.ne12;
     const uint i12 = batch_idx % p.ne12;
@@ -217,7 +233,6 @@ void main() {
     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;
 
 #ifdef MUL_MAT_ID
     if (bitCount(p.nei0) == 1) {
@@ -239,12 +254,12 @@ void main() {
 #endif
 
 #ifdef MUL_MAT_ID
-    uint pos_a = (expert_idx * p.batch_stride_a) / QUANT_K;
+    uint pos_a = expert_idx * (p.batch_stride_a / QUANT_K);
     uint pos_b = 0;
 #else
-    uint pos_a = (batch_idx_a * p.batch_stride_a) / QUANT_K;
+    uint pos_a = batch_idx_a * (p.batch_stride_a / QUANT_K);
     uint pos_b = batch_idx * p.batch_stride_b;
-    uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif
 
     uint stride_a = p.stride_a / QUANT_K;
@@ -482,7 +497,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 } else {
@@ -490,7 +505,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover4, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
@@ -526,7 +541,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 } else {
@@ -534,7 +549,7 @@ void main() {
                     coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
-                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
+                    coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BNover2, block_k, BK), tensorViewTranspose, decodeFuncB);
 
                     sum = coopMatMulAdd(mat_a, mat_b, sum);
                 }
@@ -571,7 +586,7 @@ void main() {
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
 #else
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif
@@ -583,7 +598,7 @@ void main() {
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, 0, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
 #else
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif