Misc: improve conditions for vec_dot handling

Author: sw

ggml.c

@@ -8127,7 +8127,7 @@ static void ggml_compute_forward_mul_mat(
 #endif
 
     if (params->type == GGML_TASK_INIT) {
-        if (vec_dot_type != GGML_TYPE_F32) {
+        if (src1->type != vec_dot_type) {
             char * wdata = params->wdata;
             const size_t row_size = ne10*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
 
@@ -8159,7 +8159,7 @@ static void ggml_compute_forward_mul_mat(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    void * wdata = (vec_dot_type == GGML_TYPE_F32) ? src1->data : params->wdata;
+    void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
     const size_t row_size = ne00*GGML_TYPE_SIZE[vec_dot_type]/GGML_BLCK_SIZE[vec_dot_type];
 
     for (int ir = ir0; ir < ir1; ++ir) {
@@ -11041,6 +11041,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                         //printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks = %d\n", nr0, nr1, nr0*nr1, node->n_tasks);
 
                         size_t cur = 0;
+                        const enum ggml_type vec_dot_type = type_handling[node->src0->type].vec_dot_type;
 
                         if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
                             cur = 0;
@@ -11049,7 +11050,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                                 node->n_tasks = 1;
                             }
 #endif
-                        } else if (type_handling[node->src0->type].vec_dot && node->src1->type == GGML_TYPE_F32) {
+                        } else if (node->src1->type != vec_dot_type) {
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1;
@@ -11058,8 +11059,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                             } else
 #endif
                             {
-                                const enum ggml_type type_q = type_handling[node->src0->type].vec_dot_type;
-                                cur = GGML_TYPE_SIZE[type_q]*ggml_nelements(node->src1)/GGML_BLCK_SIZE[type_q];
+                                cur = GGML_TYPE_SIZE[vec_dot_type]*ggml_nelements(node->src1)/GGML_BLCK_SIZE[vec_dot_type];
                             }
                         } else {
                             GGML_ASSERT(false);

tests/test-quantize-perf.cpp

@@ -142,7 +142,7 @@ int main(int argc, char * argv[]) {
                 break;
             }
             std::string op {argv[i]};
-            if (op == "from_float_reference") {
+            if (op == "quantize_row_q_reference") {
                 params.op_quantize_row_q_reference = true;
             } else if (op == "quantize_row_q") {
                 params.op_quantize_row_q = true;
@@ -229,7 +229,7 @@ int main(int argc, char * argv[]) {
             printf("%s\n", ggml_type_name(type));
 
             if (params.op_quantize_row_q_reference) {
-                printf("  from_float_reference\n");
+                printf("  quantize_row_q_reference\n");
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void ) {