ggml-cpu : add chunking support to mul_mat_id (#11666)

* ggml-cpu : add chunking support to mul_mat_id

* allocate chunk counter in wdata
parallelize src1 quantization by column to allows parallelization even when there is only one row

* disable for arm

* cleanup

* better way to disable for arm

* fix uninitialized counter when using 1 thread only

* revert test-backend-ops changes

ggml/src/ggml-cpu/ggml-cpu.c

@@ -7,10 +7,8 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
-#include "ggml-quants.h"
 #include "ggml-cpu-quants.h"
 #include "ggml-threading.h"
-#include "amx/amx.h"
 #include "ggml.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
@@ -1291,7 +1289,7 @@ struct ggml_threadpool {
     atomic_int n_graph;       // incremented when there is work to be done (i.e each graph)
     atomic_int GGML_CACHE_ALIGN n_barrier;
     atomic_int GGML_CACHE_ALIGN n_barrier_passed;
-    atomic_int current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads.
+    atomic_int GGML_CACHE_ALIGN current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads.
 
     // these are atomic as an annotation for thread-sanitizer
     atomic_bool stop;         // Used for stopping the threadpool altogether
@@ -7490,6 +7488,7 @@ UseGgmlGemm1:;
     if (src1->type != vec_dot_type) {
         char * wdata = params->wdata;
 
+        const size_t nbw0 = ggml_type_size(vec_dot_type);
         const size_t nbw1 = ggml_row_size(vec_dot_type, ne10);
         const size_t nbw2 = nbw1*ne11;
         const size_t nbw3 = nbw2*ne12;
@@ -7497,6 +7496,7 @@ UseGgmlGemm1:;
         assert(params->wsize >= ne13*nbw3);
         GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
+    #if 0
         for (int64_t i13 = 0; i13 < ne13; ++i13) {
             for (int64_t i12 = 0; i12 < ne12; ++i12) {
                 for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
@@ -7506,6 +7506,20 @@ UseGgmlGemm1:;
                 }
             }
         }
+    #else
+        for (int64_t i13 = 0; i13 < ne13; ++i13) {
+            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                for (int64_t i11 = 0; i11 < ne11; ++i11) {
+                    size_t bs = ggml_blck_size(vec_dot_type);
+                    int64_t ne10_block_start = (ith * ne10/bs) / nth;
+                    int64_t ne10_block_end   = ((ith + 1) * ne10/bs) / nth;
+                    from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10),
+                               (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0),
+                               (ne10_block_end - ne10_block_start) * bs);
+                }
+            }
+        }
+    #endif
     }
 
     if (ith == 0) {
@@ -7593,7 +7607,6 @@ UseGgmlGemm2:;
         if ((nr0 % 2 != 0) || (ne11 % 2 != 0) || ((ir0_end - ir0_start) % 2 != 0) || ((ir1_end - ir1_start) % 2 != 0)) {
             num_rows_per_vec_dot = 1;
         }
-
         ggml_compute_forward_mul_mat_one_chunk(params, dst, src0->type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
 
         if (nth >= nchunk0 * nchunk1) {
@@ -7606,6 +7619,84 @@ UseGgmlGemm2:;
 
 // ggml_compute_forward_mul_mat_id
 
+#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ids->ne[0]*ids->ne[1] + (i1)]
+
+struct mmid_row_mapping {
+    int32_t i1;
+    int32_t i2;
+};
+
+static void ggml_compute_forward_mul_mat_id_one_chunk(
+    struct ggml_tensor * dst,
+    const struct ggml_tensor * src0,
+    const struct ggml_tensor * src1,
+    const struct ggml_tensor * ids,
+    const int64_t cur_a,
+    const int64_t ir0_start,
+    const int64_t ir0_end,
+    const int64_t ir1_start,
+    const int64_t ir1_end,
+    const char * src0_cur,
+    const struct mmid_row_mapping * matrix_rows,
+    const size_t row_size,
+    const bool src1_cont,
+    const void * wdata) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const enum ggml_type type = src0->type;
+
+    ggml_vec_dot_t    const vec_dot      = type_traits_cpu[type].vec_dot;
+    enum ggml_type    const vec_dot_type = type_traits_cpu[type].vec_dot_type;
+
+    const int64_t blck_0 = 16;
+    const int64_t blck_1 = 16;
+
+    float tmp[16];
+
+    for (int64_t iir1 = ir1_start; iir1 < ir1_end; iir1 += blck_1) {
+        for (int64_t iir0 = ir0_start; iir0 < ir0_end; iir0 += blck_0) {
+            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir1_end; ++ir1) {
+                const int64_t _i12 = ir1; // logical row index for this expert
+
+                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, _i12);
+                const int id       = row_mapping.i1; // selected expert index
+
+                const int64_t  i11 = id % ne11;
+                const int64_t  i12 = row_mapping.i2; // row index in src1
+
+                const int64_t  i1 = id;  // selected expert index
+                const int64_t  i2 = i12; // row
+
+                // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+                //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+                //       the original src1 data pointer, so we should index using the indices directly
+                // TODO: this is a bit of a hack, we should probably have a better way to handle this
+                const char * src1_col = (const char *) wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                    ? (i11      + i12*ne11)*row_size
+                    : (i11*nb11 + i12*nb12));
+
+                float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2));
+
+                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ++ir0) {
+                    vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_cur + ir0*nb01, 0, src1_col, 0, 1);
+                }
+
+                memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir0_end) - iir0)*sizeof(float));
+            }
+        }
+    }
+}
+
+static void * incr_ptr_aligned(void ** p, size_t size, size_t align) {
+
+    void * ptr = *p;
+    ptr = (void *) GGML_PAD((uintptr_t) ptr, align);
+    *p = (void *) ((char *) ptr + size);
+    return ptr;
+}
+
 static void ggml_compute_forward_mul_mat_id(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
@@ -7623,7 +7714,6 @@ static void ggml_compute_forward_mul_mat_id(
 
     const bool src1_cont = ggml_is_contiguous(src1);
 
-    ggml_vec_dot_t    const vec_dot         = type_traits_cpu[type].vec_dot;
     enum ggml_type    const vec_dot_type    = type_traits_cpu[type].vec_dot_type;
     ggml_from_float_t const from_float      = type_traits_cpu[vec_dot_type].from_float;
 
@@ -7641,21 +7731,27 @@ static void ggml_compute_forward_mul_mat_id(
     const int n_ids = ids->ne[0]; // n_expert_used
     const int n_as  = ne02;       // n_expert
 
-    char * wdata_src1_end = (src1->type == vec_dot_type) ?
+    void * wdata_cur = params->wdata;
-            (char *) params->wdata :
-            (char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t));
 
-    struct mmid_row_mapping {
+    if (src1->type != vec_dot_type) {
-        int32_t i1;
+        incr_ptr_aligned(&wdata_cur, ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t));
-        int32_t i2;
+    }
-    };
 
-    int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
+    int64_t * matrix_row_counts = // [n_as]
-    struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *)(matrix_row_counts + n_as); // [n_as][ne11]
+        incr_ptr_aligned(&wdata_cur, n_as*sizeof(int64_t), sizeof(int64_t));
+
+    struct mmid_row_mapping * matrix_rows = // [n_as][ids->ne[0]*ids->ne[1]]
+        incr_ptr_aligned(&wdata_cur, n_as*ids->ne[0]*ids->ne[1]*sizeof(struct mmid_row_mapping), sizeof(int64_t));
+
+    char (*atomic_current_chunk)[CACHE_LINE_SIZE] = // [n_as]
+        incr_ptr_aligned(&wdata_cur, CACHE_LINE_SIZE * n_as, CACHE_LINE_SIZE);
+
+    GGML_ASSERT(params->wsize >= (size_t)((char *) wdata_cur - (char *) params->wdata));
 
     if (src1->type != vec_dot_type) {
         char * wdata = params->wdata;
 
+        const size_t nbw0 = ggml_type_size(vec_dot_type);
         const size_t nbw1 = ggml_row_size(vec_dot_type, ne10);
         const size_t nbw2 = nbw1*ne11;
         const size_t nbw3 = nbw2*ne12;
@@ -7663,19 +7759,32 @@ static void ggml_compute_forward_mul_mat_id(
         assert(params->wsize >= ne13*nbw3);
         GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
+#if 0
         for (int64_t i13 = 0; i13 < ne13; ++i13) {
-            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+            for (int64_t i12 = ith; i12 < ne12; i12 += nth) {
-                for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
+                for (int64_t i11 = 0; i11 < ne11; ++i11) {
                     from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
                                (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
                                ne10);
                 }
             }
         }
+#else
+        for (int64_t i13 = 0; i13 < ne13; ++i13) {
+            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                for (int64_t i11 = 0; i11 < ne11; ++i11) {
+                    size_t bs = ggml_blck_size(vec_dot_type);
+                    int64_t ne10_block_start = (ith * ne10/bs) / nth;
+                    int64_t ne10_block_end   = ((ith + 1) * ne10/bs) / nth;
+                    from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10),
+                               (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0),
+                               (ne10_block_end - ne10_block_start) * bs);
+                }
+            }
+        }
+#endif
     }
 
-#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne12 + (i1)]
-
     if (ith == 0) {
         // initialize matrix_row_counts
         memset(matrix_row_counts, 0, n_as*sizeof(int64_t));
@@ -7693,9 +7802,14 @@ static void ggml_compute_forward_mul_mat_id(
         }
     }
 
+    // reset current_chunk
+    for (int cur_a = ith; cur_a < n_as; cur_a += nth) {
+        atomic_int * current_chunk_ctr = (atomic_int *)(atomic_current_chunk + cur_a);
+        *current_chunk_ctr = nth;
+    }
+
     ggml_barrier(params->threadpool);
 
-    // compute each matrix multiplication in sequence
     for (int cur_a = 0; cur_a < n_as; ++cur_a) {
         const int64_t cne1 = matrix_row_counts[cur_a];
 
@@ -7703,84 +7817,64 @@ static void ggml_compute_forward_mul_mat_id(
             continue;
         }
 
-        const char * src0_cur = (const char *) src0->data + cur_a*nb02;
+        const char * src0_cur = (const char *) src0->data + cur_a * nb02;
-
+        const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-        const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-        const int64_t nr0 = ne01; // src0 rows
+        const int64_t nr0 = ne01;
-        const int64_t nr1 = cne1; // src1 rows
+        const int64_t nr1 = cne1;
 
-        // distribute the thread work across the inner or outer loop based on which one is larger
+        int chunk_size = 16;
+        if (nr0 == 1 || nr1 == 1) {
+            chunk_size = 64;
+        }
 
-        const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+#if defined(__aarch64__)
-        const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
+        // disable for ARM
+        const bool disable_chunking = true;
+#else
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+#endif // defined(__aarch64__)
 
-        const int64_t ith0 = ith % nth0;
+        int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
-        const int64_t ith1 = ith / nth0;
+        int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
 
-        const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
+        if (nchunk0 * nchunk1 < nth * 4 || disable_chunking) {
-        const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
+            nchunk0 = nr0 > nr1 ? nth : 1;
+            nchunk1 = nr0 > nr1 ? 1 : nth;
+        }
 
-        const int64_t ir010 = dr0*ith0;
+        const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
-        const int64_t ir011 = MIN(ir010 + dr0, nr0);
+        const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
 
-        const int64_t ir110 = dr1*ith1;
+        int current_chunk = ith;
-        const int64_t ir111 = MIN(ir110 + dr1, nr1);
 
-        // threads with no work simply yield (not sure if it helps)
+        atomic_int * current_chunk_ctr = (atomic_int *)(atomic_current_chunk + cur_a);
-        //if (ir010 >= ir011 || ir110 >= ir111) {
-        //    sched_yield();
-        //    continue;
-        //}
 
-        // block-tiling attempt
+        while (current_chunk < nchunk0 * nchunk1) {
-        const int64_t blck_0 = 16;
+            const int64_t ith0 = current_chunk % nchunk0;
-        const int64_t blck_1 = 16;
+            const int64_t ith1 = current_chunk / nchunk0;
 
-        // attempt to reduce false-sharing (does not seem to make a difference)
+            const int64_t ir0_start = dr0 * ith0;
-        float tmp[16];
+            const int64_t ir0_end = MIN(ir0_start + dr0, nr0);
 
-        for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
+            const int64_t ir1_start = dr1 * ith1;
-            for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
+            const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
-                for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                    const int64_t _i12 = ir1; // logical row index for this expert
 
-                    struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, _i12);
+            ggml_compute_forward_mul_mat_id_one_chunk(
-                    const int id       = row_mapping.i1; // selected expert index
+                dst, src0, src1, ids, cur_a,
+                ir0_start, ir0_end, ir1_start, ir1_end,
+                src0_cur, matrix_rows, row_size, src1_cont, wdata
+            );
 
-                    const int64_t  i11 = id % ne11;
+            if (nth >= nchunk0 * nchunk1) {
-                    const int64_t  i12 = row_mapping.i2; // row index in src1
+                break;
-
-                    const int64_t  i1 = id;  // selected expert index
-                    const int64_t  i2 = i12; // row
-
-                    // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
-                    //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
-                    //       the original src1 data pointer, so we should index using the indices directly
-                    // TODO: this is a bit of a hack, we should probably have a better way to handle this
-                    const char * src1_col = (const char *) wdata +
-                        (src1_cont || src1->type != vec_dot_type
-                        ? (i11      + i12*ne11)*row_size
-                        : (i11*nb11 + i12*nb12));
-
-                    float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2));
-
-                    //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
-                    //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
-                    //}
-
-                    for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
-                        vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_cur + ir0*nb01, 0, src1_col, 0, 1);
-                    }
-
-                    memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
-                }
             }
+
+            current_chunk = atomic_fetch_add_explicit(current_chunk_ctr, 1, memory_order_relaxed);
         }
     }
-
-#undef MMID_MATRIX_ROW
 }
 
 // ggml_compute_forward_out_prod
@@ -13713,14 +13807,19 @@ struct ggml_cplan ggml_graph_plan(
                         cur = 0;
                         const struct ggml_tensor * src0 = node->src[0];
                         const struct ggml_tensor * src1 = node->src[1];
+                        const struct ggml_tensor * ids = node->src[2];
                         const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
-                        if (src1->type != vec_dot_type) {
-                            cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
-                        }
                         const int n_as = src0->ne[2];
-                        cur += GGML_PAD(cur, sizeof(int64_t));       // align
+                        // src1
-                        cur += n_as * sizeof(int64_t);               // matrix_row_counts
+                        if (src1->type != vec_dot_type) {
-                        cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
+                            cur += ggml_row_size(vec_dot_type, ggml_nelements(src1)) + sizeof(int64_t);
+                        }
+                        // matrix_row_counts
+                        cur += n_as * sizeof(int64_t) + sizeof(int64_t);
+                        // matrix_rows
+                        cur += n_as*ids->ne[0]*ids->ne[1]*sizeof(struct mmid_row_mapping) + sizeof(int64_t);
+                        // atomic_current_chunk
+                        cur += CACHE_LINE_SIZE*n_as + CACHE_LINE_SIZE;
                     } break;
                 case GGML_OP_OUT_PROD:
                     {