ggml : mul mat tweaks (#2372)

* ggml : mul mat wip

ggml-ci

* ggml : alternative thread distribution for mul_mat

ggml-ci

* ggml : mul_mat block tiling attempt

* ggml : mul_mat threads yield

ggml-ci

diff --git a/ggml.c b/ggml.c
index c97c2b63f346b409b0b8a5f4455bff78d7a92028..beb7f464167d53b03a98e25f49ff97c666599e41 100644 (file)
--- a/ggml.c
+++ b/ggml.c
@@ -10731,71 +10731,95 @@ static void ggml_compute_forward_mul_mat(
         return;
     }
 
-    // parallelize by src0 rows
-    const int64_t dr = (ne01 + nth - 1)/nth;
+    const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+    const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
 
-    const int64_t ir10 = dr*ith;
-    const int64_t ir11 = MIN(ir10 + dr, ne01);
+    const int64_t nr0 = ne01;           // src0 rows
+    const int64_t nr1 = ne11*ne12*ne13; // src1 rows
 
-    // src1 rows
-    const int64_t nr1 = ne11*ne12*ne13;
+    //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
 
-    const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-    const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
+    // distribute the thread work across the inner or outer loop based on which one is larger
 
-    for (int64_t ir1 = 0; ir1 < nr1; ++ir1) {
-        const int64_t i13 = (ir1/(ne12*ne11));
-        const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
-        const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
-
-        const int64_t ir0 = (ir1/ne11)%(ne02*ne03);
-        const int64_t i03 = (ir0/(ne02));
-        // Hack for "Falcon multi-query-attention key stutter" / alternative to ggml_repeat2.
-        // See https://github.com/ggerganov/llama.cpp/issues/1602#issuecomment-1606087470:
-        // GG: this is likely the correct way to broadcast, though need some more thought
-        //     therefore leaving the comments to remind us for now
-        const int64_t i02 = (i12 / (ne12 / ne02));
-        // Original from PR/224 (and also essential/correct for non-broadcast matmuls in Falcon)
-        // const int64_t i02 = (ir0 - i03*ne02);
-
-        const int64_t i1 = i11;
-        const int64_t i2 = i12;
-        const int64_t i3 = i13;
-
-        const char * src0_row = (const char *) src0->data + (  0 + i02*nb02 + i03*nb03     );
-
-        // 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 + i13*ne12*ne11)*row_size
-             : (i11*nb11 + i12*nb12 + i13*nb13));
-
-        float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
-
-        for (int64_t ir = ir10; ir < ir11; ++ir) {
-            vec_dot(ne00, &dst_col[ir], src0_row + ir*nb01, src1_col);
-        }
+    const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+    const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
+
+    const int64_t ith0 = ith % nth0;
+    const int64_t ith1 = ith / nth0;
+
+    const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
+    const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
+
+    const int64_t ir010 = dr0*ith0;
+    const int64_t ir011 = MIN(ir010 + dr0, nr0);
+
+    const int64_t ir110 = dr1*ith1;
+    const int64_t ir111 = MIN(ir110 + dr1, nr1);
+
+    //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+
+    // threads with no work simply yield (not sure if it helps)
+    if (ir010 >= ir011 || ir110 >= ir111) {
+        sched_yield();
+        return;
     }
 
-    //int64_t t1 = ggml_time_us();
-    //static int64_t acc = 0;
-    //acc += t1 - t0;
-    //if (t1 - t0 > 10) {
-    //    printf("\n");
-    //    printf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
-    //    printf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
-    //    printf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
+    assert(ne12 % ne02 == 0);
+    assert(ne13 % ne03 == 0);
 
-    //    printf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
-    //}
-}
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
 
+    // block-tiling attempt
+    const int64_t blck_0 = 16;
+    const int64_t blck_1 = 16;
 
-// ggml_compute_forward_out_prod
+    // attempt to reduce false-sharing (does not seem to make a difference)
+    float tmp[16];
+
+    for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
+        for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
+            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
+                const int64_t i13 = (ir1/(ne12*ne11));
+                const int64_t i12 = (ir1 - i13*ne12*ne11)/ne11;
+                const int64_t i11 = (ir1 - i13*ne12*ne11 - i12*ne11);
+
+                // broadcast src0 into src1
+                const int64_t i03 = i13/r3;
+                const int64_t i02 = i12/r2;
+
+                const int64_t i1 = i11;
+                const int64_t i2 = i12;
+                const int64_t i3 = i13;
+
+                const char * src0_row = (const char *) src0->data + (0 + i02*nb02 + i03*nb03);
+
+                // 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 + i13*ne12*ne11)*row_size
+                     : (i11*nb11 + i12*nb12 + i13*nb13));
 
+                float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+
+                //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], src0_row + ir0*nb01, src1_col);
+                }
+                memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
+            }
+        }
+    }
+}
+
+// ggml_compute_forward_out_prod
 
 static void ggml_compute_forward_out_prod_f32(
         const struct ggml_compute_params * params,