whispercpp 1.3.5 → 1.3.7

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

@@ -29,6 +29,7 @@
 #endif
 
 #include "types.glsl"
+#include "dot_product_funcs.glsl"
 
 #ifndef LOAD_VEC_A
 #define LOAD_VEC_A 1
@@ -90,6 +91,8 @@ 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;
@@ -123,8 +126,8 @@ layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working wit
 #define SHMEM_STRIDE (BK / 2 + 1)
 #endif
 
-shared FLOAT_TYPE_VEC2 buf_a[BM * SHMEM_STRIDE];
-shared FLOAT_TYPE_VEC2 buf_b[BN * SHMEM_STRIDE];
+shared FLOAT_TYPEV2 buf_a[BM * SHMEM_STRIDE];
+shared FLOAT_TYPEV2 buf_b[BN * SHMEM_STRIDE];
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)
 
@@ -139,7 +142,7 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
     if (ic * BN >= data_expert_count[expert_idx]) {
         return;
     }
@@ -149,7 +152,7 @@ void main() {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    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;
@@ -256,17 +259,17 @@ void main() {
         sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
     }
 #else
-    ACC_TYPE_VEC2 sums[WMITER * TM * WNITER * TN/2];
+    ACC_TYPEV2 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;
+    FLOAT_TYPEV4 cache_a[WMITER * TM];
+    FLOAT_TYPEV4 cache_b;
 #else
-    FLOAT_TYPE_VEC2 cache_a[WMITER * TM];
-    FLOAT_TYPE_VEC2 cache_b;
+    FLOAT_TYPEV2 cache_a[WMITER * TM];
+    FLOAT_TYPEV2 cache_b;
 #endif
 
     [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
-        sums[i] = ACC_TYPE_VEC2(0.0f, 0.0f);
+        sums[i] = ACC_TYPEV2(0.0f, 0.0f);
     }
 #endif
 
@@ -327,15 +330,8 @@ void main() {
                         [[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
+                            sums[sums_idx].x = dot_product(cache_a[wsir * TM + 2 * cr    ], cache_b, sums[sums_idx].x);
+                            sums[sums_idx].y = dot_product(cache_a[wsir * TM + 2 * cr + 1], cache_b, sums[sums_idx].y);
                         }
                     }
                 }
@@ -366,7 +362,7 @@ 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
@@ -375,6 +371,7 @@ 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;
@@ -385,6 +382,7 @@ 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
@@ -402,18 +400,22 @@ 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);
             }
         }
     }

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

@@ -11,6 +11,9 @@
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_NV_cooperative_matrix2 : enable
+#ifdef GGML_VULKAN_COOPMAT2_DECODE_VECTOR
+#extension GL_NV_cooperative_matrix_decode_vector : enable
+#endif
 #extension GL_EXT_buffer_reference : enable
 #extension GL_KHR_shader_subgroup_ballot : enable
 #extension GL_KHR_shader_subgroup_vote : enable
@@ -53,6 +56,8 @@ 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;
@@ -67,12 +72,17 @@ layout (push_constant) uniform parameter
 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 data_d[];};
+#if defined(MUL_MAT_ID) && defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+layout (binding = 1) readonly buffer B4 {B_TYPEV4 data_b_v4[];};
+#endif
 
 #if QUANT_K > 1
-#define DECODEFUNCA , dequantFuncA
-
 #include "dequant_funcs_cm2.glsl"
-
+#if defined(dequantFuncA_v) && defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+#define DECODEFUNCA , dequantFuncA, dequantFuncA_v
+#else
+#define DECODEFUNCA , dequantFuncA
+#endif
 #else
 #define DECODEFUNCA
 #endif
@@ -109,11 +119,33 @@ B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const i
     const uint row_i = blockCoords[0];
 
     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]];
+#if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+    // The decode-vector path gives B a K-dimension tensor-layout block size of BK.
+    const uint k = blockCoords[1] * BK + coordInBlock[1];
+#else
+    const uint k = blockCoords[1];
+#endif
+    B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + k];
 
     return ret;
 }
 
+#if defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+B_TYPEV4 decodeFuncB_v(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint row_i = blockCoords[0];
+
+    const u16vec4 row_idx = row_ids[row_i];
+    const uint k = blockCoords[1] * BK + coordInBlock[1];
+    const uint base = row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + k;
+
+    return data_b_v4[base >> 2];
+}
+#define DECODEFUNCB , decodeFuncB, decodeFuncB_v
+#else
+#define DECODEFUNCB , decodeFuncB
+#endif
+
 D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t ir, const in uint32_t ic)
 {
     uint dr = ir * BM + r;
@@ -165,7 +197,9 @@ 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;
@@ -197,7 +231,7 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
     if (ic * BN >= data_expert_count[expert_idx]) {
         return;
     }
@@ -215,7 +249,7 @@ void main() {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    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;
@@ -255,7 +289,7 @@ void main() {
 #else
     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;
@@ -281,6 +315,9 @@ void main() {
     tensorLayoutA = setTensorLayoutBlockSizeNV(tensorLayoutA, 1, QUANT_K);
     tensorLayoutAClamp = setTensorLayoutBlockSizeNV(tensorLayoutAClamp, 1, QUANT_K);
 #endif
+#if defined(MUL_MAT_ID) && defined(GGML_VULKAN_COOPMAT2_DECODE_VECTOR)
+    tensorLayoutB = setTensorLayoutBlockSizeNV(tensorLayoutB, 1, BK);
+#endif
 
     // Use end_k rather than p.K as the dimension because that's what
     // we need to bound check against when using split_k.
@@ -493,7 +530,7 @@ 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, 0, 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 {
@@ -501,7 +538,7 @@ 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, 0, 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);
                 }
@@ -537,7 +574,7 @@ 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, 0, 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 {
@@ -545,7 +582,7 @@ 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, 0, 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);
                 }
@@ -582,7 +619,7 @@ 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, 0, 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
@@ -594,7 +631,7 @@ 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, 0, 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