device tuning

Author: 0cc4m

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1498,7 +1498,7 @@ class vk_perf_logger {
         }
         if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
             const uint64_t m     = node->src[0]->ne[1];
-            const uint64_t n     = node->ne[1];
+            const uint64_t n     = (node->op == GGML_OP_MUL_MAT) ? node->ne[1] : node->ne[2];
             const uint64_t k     = node->src[1]->ne[0];
             const uint64_t batch = node->src[1]->ne[2] * node->src[1]->ne[3];
             std::string    name  = ggml_op_name(node->op);
@@ -6581,23 +6581,36 @@ static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_
         return true;
     }
 
+    // Quantization overhead is not worth it for small k
     switch (device->vendor_id) {
     case VK_VENDOR_ID_NVIDIA:
+        if (k <= 4096) {
+            return false;
+        }
+
         switch (src0_type) {
-        case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_Q8_0:
             return device->architecture == vk_device_architecture::NVIDIA_PRE_TURING;
         default:
             return true;
         }
     case VK_VENDOR_ID_AMD:
+        if (k < 2048) {
+            return false;
+        }
+
         switch (src0_type) {
         case GGML_TYPE_Q8_0:
             return device->architecture == vk_device_architecture::AMD_GCN;
         default:
             return true;
         }
     case VK_VENDOR_ID_INTEL:
+        if (k < 2048) {
+            return false;
+        }
+
         switch (src0_type) {
         // From tests on A770 Linux, may need more tuning
         case GGML_TYPE_Q4_0:
@@ -6611,7 +6624,6 @@ static bool ggml_vk_should_use_mmvq(const vk_device& device, uint32_t m, uint32_
     }
 
     GGML_UNUSED(m);
-    GGML_UNUSED(k);
 }
 
 static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
@@ -7288,7 +7300,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0 && ne21 >= 8;
+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;
 
     // Check for mmq first
     vk_matmul_pipeline mmp = quantize_y ? ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, GGML_TYPE_Q8_1, (ggml_prec)dst->op_params[0]) : nullptr;
@@ -7559,7 +7571,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
 
     const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
-    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0 && ggml_vk_should_use_mmvq(ctx->device, ne01, ne11, ne10, src0->type);
+    bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0 && ggml_vk_should_use_mmvq(ctx->device, ne01, ne12, ne10, src0->type);
 
     vk_pipeline to_fp16_vk_0 = nullptr;
     vk_pipeline to_fp16_vk_1 = nullptr;