CUDA: mul_mat_v support for batch sizes > 1 (#14262)

* CUDA: mul_mat_v support for batch sizes > 1

* use 64 bit math for initial offset calculation

Author: JohannesGaessler

ggml/src/ggml-cuda/common.cuh

@@ -262,6 +262,10 @@ static bool fp16_mma_hardware_available(const int cc) {
         GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
 }
 
+static bool bf16_mma_hardware_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE;
+}
+
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
 static bool new_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1943,16 +1943,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
 
     bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
     bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
     bool use_mul_mat_q     = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
-    bool any_gpus_with_slow_fp16   = false;
-    bool any_gpus_without_fp16_mma = false;
+    bool any_gpus_with_slow_fp16 = false;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1963,16 +1961,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            const int cc              = ggml_cuda_info().devices[id].cc;
-            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
-            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+            const int cc            = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
-        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
-        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
-        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
     // debug helpers
@@ -1983,7 +1981,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    if (!split && use_mul_mat_vec) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
         ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);