ggml : fix BLAS with unsupported types (llama/9775)

* ggml : do not use BLAS with types without to_float

* ggml : return pointer from ggml_internal_get_type_traits to avoid unnecessary copies

* ggml : rename ggml_internal_get_type_traits -> ggml_get_type_traits

it's not really internal if everybody uses it

diff --git a/include/ggml.h b/include/ggml.h
index e7678d07135ab7c6c47c8cf8b00a7fbd23748c06..4508da4fb3a41607fcf81e6bd939490af599cf25 100644 (file)
--- a/include/ggml.h
+++ b/include/ggml.h
@@ -2535,7 +2535,7 @@ extern "C" {
     typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
                                        const void * GGML_RESTRICT y, int nr, int nc);
 
-    typedef struct {
+    struct ggml_type_traits {
         const char             * type_name;
         int64_t                  blck_size;
         int64_t                  blck_size_interleave; // interleave elements in blocks
@@ -2551,9 +2551,9 @@ extern "C" {
         int64_t                  ncols; // number of columns to process simultaneously
         ggml_gemv_t              gemv;
         ggml_gemm_t              gemm;
-    } ggml_type_traits_t;
+    };
 
-    GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
+    GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);
 
 #ifdef  __cplusplus
 }

diff --git a/src/ggml-backend.cpp b/src/ggml-backend.cpp
index fbd49d13dda14247c29a87ae366a4411c4a1067f..627b4dbc7873213b2c2a923d6572ded397afdc6a 100644 (file)
--- a/src/ggml-backend.cpp
+++ b/src/ggml-backend.cpp
@@ -1177,7 +1177,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
                 op->type != GGML_TYPE_IQ1_S   &&
                 op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
-            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
+            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_get_type_traits(op->src[0]->type)->vec_dot_type;
         case GGML_OP_ROPE_BACK:
             return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
         case GGML_OP_IM2COL_BACK:

diff --git a/src/ggml-blas.cpp b/src/ggml-blas.cpp
index 0c6574de500a69e0d9252c3a15e7368f3dd28c2d..55f7245861105eada9dfa7777042159edb32a908 100644 (file)
--- a/src/ggml-blas.cpp
+++ b/src/ggml-blas.cpp
@@ -65,8 +65,8 @@ static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct gg
 
     // convert src0 to float
     if (type != GGML_TYPE_F32) {
-        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
-        ggml_to_float_t const to_float = type_traits.to_float;
+        const auto * type_traits = ggml_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits->to_float;
 
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -420,19 +420,21 @@ static bool ggml_backend_blas_device_supports_op(ggml_backend_dev_t dev, const s
             // TODO: find the optimal value
             const int64_t min_batch = 32;
 
-            return (ggml_is_contiguous(src0) &&
-                    ggml_is_contiguous(src1) &&
-                    src1->type == GGML_TYPE_F32 &&
-                    (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch));
+            return ggml_is_contiguous(src0) &&
+                   ggml_is_contiguous(src1) &&
+                   src1->type == GGML_TYPE_F32 &&
+                   (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
         }
 
         case GGML_OP_OUT_PROD:
-            return (op->src[0]->type == GGML_TYPE_F32 &&
-                    op->src[1]->type == GGML_TYPE_F32 &&
-                    ggml_is_matrix(src0) &&
-                    ggml_is_matrix(src1) &&
-                    ggml_is_contiguous(src0) &&
-                    (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
+            return op->src[0]->type == GGML_TYPE_F32 &&
+                   op->src[1]->type == GGML_TYPE_F32 &&
+                   ggml_is_matrix(src0) &&
+                   ggml_is_matrix(src1) &&
+                   ggml_is_contiguous(src0) &&
+                   (ggml_is_contiguous(src1) || ggml_is_transposed(src1)) &&
+                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
 
         default:
             return false;

diff --git a/src/ggml-vulkan.cpp b/src/ggml-vulkan.cpp
index 30bd376da61882aca8361f98a4333ae4933c3961..374c6ecd7ade5b0789ab1bdd0011603a06ee0618 100644 (file)
--- a/src/ggml-vulkan.cpp
+++ b/src/ggml-vulkan.cpp
@@ -5287,9 +5287,9 @@ static void ggml_vk_dequantize_data(const void * from, float * to, size_t ne, gg
         return;
     }
 
-    ggml_type_traits_t tt = ggml_internal_get_type_traits(quant);
+    const auto * tt = ggml_get_type_traits(quant);
 
-    ggml_to_float_t dequant_fn = tt.to_float;
+    ggml_to_float_t dequant_fn = tt->to_float;
 
     dequant_fn(from, to, ne);
 }

diff --git a/src/ggml.c b/src/ggml.c
index 03b832d0f2189ea224e8db91f89df60f1e7db81a..3f01092d9f59aebc47ea51c7ac2810d5d4dcd866 100644 (file)
--- a/src/ggml.c
+++ b/src/ggml.c
@@ -729,7 +729,7 @@ static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float *
 static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc);
 static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc);
 
-static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
+static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
     [GGML_TYPE_I8] = {
         .type_name                = "i8",
         .blck_size                = 1,
@@ -1151,9 +1151,9 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
 };
 
 // For internal test use
-ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
+const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type) {
     GGML_ASSERT(type < GGML_TYPE_COUNT);
-    return type_traits[type];
+    return &type_traits[type];
 }
 
 //

diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index fa26cc653aaf82e46712b1e0804758a7731e6df8..ee1a8877e1b0a997b7f5599b6c9d2fbe2534781e 100644 (file)
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -133,7 +133,7 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
     std::vector<uint8_t> buf(ggml_nbytes(t));
     ggml_backend_tensor_get(t, buf.data(), 0, ggml_nbytes(t));
 
-    ggml_type_traits_t tt = ggml_internal_get_type_traits(t->type);
+    const auto * tt = ggml_get_type_traits(t->type);
     size_t bs = ggml_blck_size(t->type);
     std::vector<float> vq(ggml_blck_size(t->type));
     bool quantized = ggml_is_quantized(t->type);
@@ -159,7 +159,7 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
                     } else if (t->type == GGML_TYPE_I8) {
                         tv.push_back((float)*(int8_t *) &buf[i]);
                     } else if (quantized) {
-                        tt.to_float(&buf[i], vq.data(), bs);
+                        tt->to_float(&buf[i], vq.data(), bs);
                         tv.insert(tv.end(), vq.begin(), vq.end());
                     } else {
                         GGML_ABORT("fatal error");

diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp
index ccf5721a3ab833425ada163569b9be34278fe405..d50417ba01f45bab3cb8f3b33541b1c30b2ddcb6 100644 (file)
--- a/tests/test-quantize-fns.cpp
+++ b/tests/test-quantize-fns.cpp
@@ -44,26 +44,26 @@ static float array_rmse(const float * a1, const float * a2, size_t n) {
 }
 
 // Total quantization error on test data
-static float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
+static float total_quantization_error(const ggml_type_traits * qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
 
-    qfns.from_float(test_data, tmp_q.data(), test_size);
-    qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
+    qfns->from_float(test_data, tmp_q.data(), test_size);
+    qfns->to_float(tmp_q.data(), tmp_out.data(), test_size);
     return array_rmse(test_data, tmp_out.data(), test_size);
 }
 
 // Total quantization error on test data
-static float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
+static float reference_quantization_error(const ggml_type_traits * qfns, size_t test_size, const float * test_data) {
     std::vector<uint8_t> tmp_q(2*test_size);
     std::vector<float> tmp_out(test_size);
     std::vector<float> tmp_out_ref(test_size);
 
-    qfns.from_float(test_data, tmp_q.data(), test_size);
-    qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
+    qfns->from_float(test_data, tmp_q.data(), test_size);
+    qfns->to_float(tmp_q.data(), tmp_out.data(), test_size);
 
-    qfns.from_float_ref(test_data, tmp_q.data(), test_size);
-    qfns.to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
+    qfns->from_float_ref(test_data, tmp_q.data(), test_size);
+    qfns->to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
 
     return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
 }
@@ -78,18 +78,18 @@ static float dot_product(const float * a1, const float * a2, size_t test_size) {
 
 // Total dot product error
 static float dot_product_error(
-    ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2
+    const ggml_type_traits * qfns, size_t test_size, const float * test_data1, const float *test_data2
 ) {
     std::vector<uint8_t> tmp_q1(2*test_size);
     std::vector<uint8_t> tmp_q2(2*test_size);
 
-    auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
+    const auto * vdot = ggml_get_type_traits(qfns->vec_dot_type);
 
-    qfns.from_float(test_data1, tmp_q1.data(), test_size);
-    vdot.from_float(test_data2, tmp_q2.data(), test_size);
+    qfns->from_float(test_data1, tmp_q1.data(), test_size);
+    vdot->from_float(test_data2, tmp_q2.data(), test_size);
 
     float result = INFINITY;
-    qfns.vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
+    qfns->vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
 
     const float dot_ref = dot_product(test_data1, test_data2, test_size);
 
@@ -131,10 +131,10 @@ int main(int argc, char * argv[]) {
 
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
         ggml_type type = (ggml_type) i;
-        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
+        const auto * qfns = ggml_get_type_traits(type);
 
         // deprecated - skip
-        if (qfns.blck_size == 0) {
+        if (qfns->blck_size == 0) {
             continue;
         }
 
@@ -143,7 +143,7 @@ int main(int argc, char * argv[]) {
         printf("Testing %s\n", ggml_type_name((ggml_type) i));
         ggml_quantize_init(ei);
 
-        if (qfns.from_float && qfns.to_float) {
+        if (qfns->from_float && qfns->to_float) {
             const float total_error = total_quantization_error(qfns, test_size, test_data.data());
             const float max_quantization_error =
                 type == GGML_TYPE_TQ1_0   ? MAX_QUANTIZATION_TOTAL_ERROR_TERNARY :

diff --git a/tests/test-quantize-perf.cpp b/tests/test-quantize-perf.cpp
index 24e066053219a1f45cba4861aa4978997b228610..bdbdd90a8d41aa167a79c6625d169f87704e1809 100644 (file)
--- a/tests/test-quantize-perf.cpp
+++ b/tests/test-quantize-perf.cpp
@@ -122,9 +122,9 @@ static void usage(char * argv[]) {
     printf("  --type TYPE           set test type as");
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
         ggml_type type = (ggml_type) i;
-        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
+        const auto * qfns = ggml_get_type_traits(type);
         if (ggml_type_name(type) != NULL) {
-            if (qfns.from_float && qfns.to_float) {
+            if (qfns->from_float && qfns->to_float) {
                 printf(" %s", ggml_type_name(type));
             }
         }
@@ -270,12 +270,12 @@ int main(int argc, char * argv[]) {
 
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
         ggml_type type = (ggml_type) i;
-        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
+        const auto * qfns = ggml_get_type_traits(type);
         if (!params.include_types.empty() && ggml_type_name(type) && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
             continue;
         }
 
-        if (qfns.from_float && qfns.to_float) {
+        if (qfns->from_float && qfns->to_float) {
             printf("%s\n", ggml_type_name(type));
 
             ggml_quantize_init(type);
@@ -285,7 +285,7 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        qfns.from_float_ref(test_data1, test_q1, size);
+                        qfns->from_float_ref(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -299,7 +299,7 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        qfns.from_float(test_data1, test_q1, size);
+                        qfns->from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -310,11 +310,11 @@ int main(int argc, char * argv[]) {
 
             if (params.op_dequantize_row_q) {
                 printf("  dequantize_row_q\n");
-                qfns.from_float(test_data1, test_q1, largest);
+                qfns->from_float(test_data1, test_q1, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        qfns.to_float(test_q1, test_out, size);
+                        qfns->to_float(test_q1, test_out, size);
                         return test_out[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -328,8 +328,8 @@ int main(int argc, char * argv[]) {
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
-                        auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
-                        vdot.from_float(test_data1, test_q1, size);
+                        const auto * vdot = ggml_get_type_traits(qfns->vec_dot_type);
+                        vdot->from_float(test_data1, test_q1, size);
                         return test_q1[0];
                     };
                     size_t quantized_size = ggml_row_size(type, size);
@@ -340,13 +340,13 @@ int main(int argc, char * argv[]) {
 
             if (params.op_vec_dot_q) {
                 printf("  vec_dot_q\n");
-                qfns.from_float(test_data1, test_q1, largest);
-                qfns.from_float(test_data2, test_q2, largest);
+                qfns->from_float(test_data1, test_q1, largest);
+                qfns->from_float(test_data2, test_q2, largest);
                 for (size_t size : params.test_sizes) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
                         float result;
-                        qfns.vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
+                        qfns->vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
                         return result;
                     };
                     size_t quantized_size = ggml_row_size(type, size);