ggml : do not use BLAS with types without to_float

slaren · slaren · commit 0226d3d36b2f · 2024-10-07T21:58:30.000+02:00
ggml : return pointer from ggml_internal_get_type_traits to avoid unnecessary copies
diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp
@@ -314,9 +314,9 @@ struct lora_merge_ctx {
             // optionally dequantize it
             printf("%s :   + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
             auto nels = ggml_nelements(inp_base);
-            ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type);
+            const auto * qtype = ggml_internal_get_type_traits(base->type);
             std::vector<uint8_t> dequant_buf(nels * sizeof(float));
-            qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
+            qtype->to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
             ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
         } else {
             ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp
@@ -142,7 +142,7 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
 }
 
 static void test_roundtrip_on_chunk(
-    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits_t & qfns, bool use_reference,
+    const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, bool use_reference,
     float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
 ) {
     if (layer->type == GGML_TYPE_F16) {
@@ -166,7 +166,7 @@ static void test_roundtrip_on_chunk(
 
 // Run quantization function for a single layer and update error stats
 static void test_roundtrip_on_layer(
-    std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference,
+    std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, bool use_reference,
     const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
     std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
 ) {
@@ -371,8 +371,8 @@ int main(int argc, char ** argv) {
         if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
             continue;
         }
-        ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
-        if (qfns.from_float && qfns.to_float) {
+        const auto *  qfns = ggml_internal_get_type_traits(type);
+        if (qfns->from_float && qfns->to_float) {
             if (params.verbose) {
                 printf("testing %s ...\n",  ggml_type_name(type));
             }
@@ -393,7 +393,7 @@ int main(int argc, char ** argv) {
                 test_roundtrip_on_layer(
                         layer_name,
                         params.per_layer_stats,
-                        qfns,
+                        *qfns,
                         params.reference,
                         kv_tensor.second,
                         input_scratch,
diff --git a/ggml/include/ggml.h b/ggml/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_internal_get_type_traits(enum ggml_type type);
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
@@ -1181,7 +1181,7 @@ 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_internal_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/ggml/src/ggml-blas.cpp b/ggml/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_internal_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_internal_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_internal_get_type_traits(src0->type)->to_float != NULL);
 
         default:
             return false;
diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/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_internal_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/ggml/src/ggml.c b/ggml/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_internal_get_type_traits(enum ggml_type type) {
     GGML_ASSERT(type < GGML_TYPE_COUNT);
-    return type_traits[type];
+    return &type_traits[type];
 }
 
 //
diff --git a/pocs/vdot/q8dot.cpp b/pocs/vdot/q8dot.cpp
@@ -136,7 +136,7 @@ int main(int argc, char** argv) {
 
     auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
 
-    auto funcs = ggml_internal_get_type_traits(ggml_type);
+    const auto * funcs = ggml_internal_get_type_traits(ggml_type);
 
     Stat simple, ggml;
 
@@ -156,8 +156,8 @@ int main(int argc, char** argv) {
 
         t1 = std::chrono::high_resolution_clock::now();
         float fs;
-        if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
-        else funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
+        if (type == 0) funcs->vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
+        else funcs->vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
         t2 = std::chrono::high_resolution_clock::now();
         t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
         if (iloop > 3) ggml.addResult(fs, t);
diff --git a/pocs/vdot/vdot.cpp b/pocs/vdot/vdot.cpp
@@ -236,7 +236,7 @@ int main(int argc, char** argv) {
     int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64);
     int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64);
 
-    auto funcs = useQ4_1 ? ggml_internal_get_type_traits(GGML_TYPE_Q4_1) : ggml_internal_get_type_traits(GGML_TYPE_Q4_0);
+    const auto * funcs = useQ4_1 ? ggml_internal_get_type_traits(GGML_TYPE_Q4_1) : ggml_internal_get_type_traits(GGML_TYPE_Q4_0);
 
     std::vector<block_q4_0> q40;
     std::vector<block_q4_1> q41;
@@ -261,9 +261,9 @@ int main(int argc, char** argv) {
         // Note, we do not include this in the timing as in practical application
         // we already have the quantized model weights.
         if (useQ4_1) {
-            funcs.from_float(x1.data(), q41.data(), kVecSize);
+            funcs->from_float(x1.data(), q41.data(), kVecSize);
         } else {
-            funcs.from_float(x1.data(), q40.data(), kVecSize);
+            funcs->from_float(x1.data(), q40.data(), kVecSize);
         }
 
         // Now measure time the dot product needs using the "scalar" version above
@@ -282,10 +282,10 @@ int main(int argc, char** argv) {
             dot_q4_q8(kVecSize, &result, q40.data(), q8.data());
         }
         else {
-            auto vdot = ggml_internal_get_type_traits(funcs.vec_dot_type);
-            vdot.from_float(y1.data(), q8.data(), kVecSize);
-            if (useQ4_1) funcs.vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
-            else funcs.vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
+            const auto * vdot = ggml_internal_get_type_traits(funcs->vec_dot_type);
+            vdot->from_float(y1.data(), q8.data(), kVecSize);
+            if (useQ4_1) funcs->vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
+            else funcs->vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
         }
         sumq += result;
         t2 = std::chrono::high_resolution_clock::now();
diff --git a/src/llama.cpp b/src/llama.cpp
@@ -17872,10 +17872,9 @@ static void llama_tensor_dequantize_internal(
     }
     float * f32_output = (float *) output.data();
 
-    ggml_type_traits_t qtype;
+    const ggml_type_traits * qtype = ggml_internal_get_type_traits(tensor->type);
     if (ggml_is_quantized(tensor->type)) {
-        qtype = ggml_internal_get_type_traits(tensor->type);
-        if (qtype.to_float == NULL) {
+        if (qtype->to_float == NULL) {
             throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
         }
     } else if (tensor->type != GGML_TYPE_F16 &&
@@ -17889,7 +17888,7 @@ static void llama_tensor_dequantize_internal(
         } else if (tensor->type == GGML_TYPE_BF16) {
             ggml_bf16_to_fp32_row((ggml_bf16_t *)tensor->data, f32_output, nelements);
         } else if (ggml_is_quantized(tensor->type)) {
-            qtype.to_float(tensor->data, f32_output, nelements);
+            qtype->to_float(tensor->data, f32_output, nelements);
         } else {
             GGML_ABORT("fatal error"); // unreachable
         }
@@ -17925,7 +17924,7 @@ static void llama_tensor_dequantize_internal(
             } else if (typ == GGML_TYPE_BF16) {
                 ggml_bf16_to_fp32_row((ggml_bf16_t *)inbuf, outbuf, nels);
             } else {
-                qtype.to_float(inbuf, outbuf, nels);
+                qtype->to_float(inbuf, outbuf, nels);
             }
         };
         workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
diff --git 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_internal_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
@@ -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_internal_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_internal_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

Original file line number	Diff line number	Diff line change
`@@ -5287,9 +5287,9 @@ static void ggml_vk_dequantize_data(const void * from, float * to, size_t ne, gg`
`5287`	`5287`	`return;`
`5288`	`5288`	`}`
`5289`	`5289`
`5290`		`- ggml_type_traits_t tt = ggml_internal_get_type_traits(quant);`
	`5290`	`+ const auto * tt = ggml_internal_get_type_traits(quant);`
`5291`	`5291`
`5292`		`- ggml_to_float_t dequant_fn = tt.to_float;`
	`5292`	`+ ggml_to_float_t dequant_fn = tt->to_float;`
`5293`	`5293`
`5294`	`5294`	`dequant_fn(from, to, ne);`
`5295`	`5295`	`}`