test(mgb/opr): add testcase for conv bias int4

GitOrigin-RevId: e3fff5e30b8be8398bfc2a96ea3753624a8e7161

dnn/src/fallback/conv_bias/algos.cpp

@@ -187,7 +187,9 @@ bool ConvBiasImpl::AlgoNaive::usable(
         const NCBKernSizeParam& param,
         AlgoSelectionStrategy /*algo_selection_strategy*/) const {
     MIDOUT_BEGIN(megdnn_fallback_naive, 0) {
-        return param.filter_meta.format == param::ConvBias::Format::NCHW;
+        auto algo_data_type = param.deduce_algo_data_type();
+        return param.filter_meta.format == param::ConvBias::Format::NCHW &&
+               contain_data_type(get_algo_type().data_type, algo_data_type);
     }
     MIDOUT_END();
     return false;

dnn/src/fallback/conv_bias/opr_impl.cpp

@@ -342,7 +342,8 @@ ConvBiasImpl::NCBKernSizeParam ConvBiasImpl::make_ncb_kern_size_param(
         param().format == Param::Format::NCHW4 ||
         param().format == Param::Format::NCHW44 ||
         param().format == Param::Format::NCHW44_DOT ||
-        param().format == Param::Format::NCHW) {
+        param().format == Param::Format::NCHW || 
+        param().format == Param::Format::NCHW64) {
         spatial_pos = 2;
     } else if (param().format == Param::Format::NHWC) {
         spatial_pos = 1;

dnn/src/naive/conv_bias/opr_impl.cpp

@@ -123,6 +123,7 @@ void forward_bias<dt_quint4, dt_quint4, dt_qint32, dt_qint32>(
         auto param = layout.dtype.param<dtype::Quantized4Asymm>();
         ret.dtype = dtype::Quantized8Asymm(param.scale, param.zero_point);
         ret.format = TensorFormat(ret.dtype);
+        ret.init_contiguous_stride();
         return ret;
     };
     TensorND new_src = {workspace_ptr, convert_layout(src.layout)};
@@ -147,6 +148,7 @@ void forward_bias<dt_qint4, dt_qint4, dt_qint32, dt_qint32>(
         auto param = layout.dtype.param<dtype::QuantizedS4>();
         ret.dtype = dtype::QuantizedS8(param.scale);
         ret.format = TensorFormat(ret.dtype);
+        ret.init_contiguous_stride();
         return ret;
     };
     TensorND new_src = {workspace_ptr, convert_layout(src.layout)};

dnn/src/naive/lowbit_utils.cpp

@@ -16,10 +16,20 @@
 void megdnn::naive::uint4_to_uint8(const TensorND& in, const TensorND& out) {
     auto in_ptr = static_cast<uint8_t*>(in.raw_ptr) + in.layout.span().low_byte;
     auto out_ptr = out.compatible_ptr<uint8_t>() + out.layout.span().low_byte;
-    for (size_t i = 0; i < in.layout.span().dist_elem(); i += 2) {
-        uint8_t val = in_ptr[i / 2];
-        out_ptr[i] = val & 0xF;
-        out_ptr[i + 1] = (val >> 4) & 0xF;
+    const auto& ly = in.layout;
+    auto dim_in = ly.shape[ly.ndim - 1];
+    auto elems = ly.total_nr_elems();
+    auto dim_out = elems / dim_in;
+    auto stride_out = div_ceil(dim_in, 2_z);
+    for (size_t i = 0; i < dim_out; ++i) {
+        for (size_t j = 0; j < dim_in; j += 2) {
+            uint8_t val = in_ptr[j / 2];
+            out_ptr[j] = val & 0xF;
+            if (j + 1 < dim_in)
+                out_ptr[j + 1] = (val >> 4) & 0xF;
+        }
+        in_ptr += stride_out;
+        out_ptr += dim_in;
     }
 }
 
@@ -27,11 +37,23 @@ void megdnn::naive::uint8_to_uint4(const TensorND& in, const TensorND& out) {
     auto in_ptr = static_cast<uint8_t*>(in.raw_ptr) + in.layout.span().low_byte;
     auto out_ptr =
             static_cast<uint8_t*>(out.raw_ptr) + out.layout.span().low_byte;
-    for (size_t i = 0; i < out.layout.span().dist_elem(); i += 2) {
-        uint8_t a = in_ptr[i], b = in_ptr[i + 1];
-        a = std::min(a, DTypeTrait<dtype::Quantized4Asymm>::max());
-        b = std::min(b, DTypeTrait<dtype::Quantized4Asymm>::max());
-        out_ptr[i / 2] = a + (b << 4);
+    const auto& ly = in.layout;
+    auto dim_in = ly.shape[ly.ndim - 1];
+    auto elems = ly.total_nr_elems();
+    auto dim_out = elems / dim_in;
+    auto stride_out = div_ceil(dim_in, 2_z);
+    for (size_t i = 0; i < dim_out; ++i) {
+        for (size_t j = 0; j < dim_in; j += 2) {
+            uint8_t a = in_ptr[j];
+            uint8_t b = 0;
+            if (j + 1 < dim_in)
+                b = in_ptr[j + 1];
+            a = std::min(a, DTypeTrait<dtype::Quantized4Asymm>::max());
+            b = std::min(b, DTypeTrait<dtype::Quantized4Asymm>::max());
+            out_ptr[j / 2] = a + (b << 4);
+        }
+        in_ptr += dim_in;
+        out_ptr += stride_out;
     }
 }
 
@@ -40,13 +62,21 @@ void megdnn::naive::int4_to_int8(const TensorND& in, const TensorND& out) {
     auto in_ptr = static_cast<int8_t*>(in.raw_ptr) + in.layout.span().low_byte;
     auto out_ptr =
             static_cast<int8_t*>(out.raw_ptr) + out.layout.span().low_byte;
-
-    megdnn_assert(in.layout.span().dist_elem() % 2 == 0);
-    for (size_t i = 0; i < in.layout.span().dist_elem(); i += 2) {
-        int8_t cur = in_ptr[i / 2];
-        out_ptr[i] = cur << 4;
-        out_ptr[i] = out_ptr[i] >> 4;
-        out_ptr[i + 1] = cur >> 4;
+    const auto& ly = in.layout;
+    auto dim_in = ly.shape[ly.ndim - 1];
+    auto elems = ly.total_nr_elems();
+    auto dim_out = elems / dim_in;
+    auto stride_out = div_ceil(dim_in, 2_z);
+    for (size_t i = 0; i < dim_out; ++i) {
+        for (size_t j = 0; j < dim_in; j += 2) {
+            int8_t cur = in_ptr[j / 2];
+            out_ptr[j] = cur << 4;
+            out_ptr[j] = out_ptr[j] >> 4;
+            if (j + 1 < dim_in)
+                out_ptr[j + 1] = cur >> 4;
+        }
+        in_ptr += stride_out;
+        out_ptr += dim_in;
     }
 }
 
@@ -54,12 +84,24 @@ void megdnn::naive::int8_to_int4(const TensorND& in, const TensorND& out) {
     auto in_ptr = static_cast<int8_t*>(in.raw_ptr) + in.layout.span().low_byte;
     auto out_ptr =
             static_cast<int8_t*>(out.raw_ptr) + out.layout.span().low_byte;
-    for (size_t i = 0; i < out.layout.span().dist_elem(); i += 2) {
-        int8_t a = in_ptr[i], b = in_ptr[i + 1];
-        a = std::min(a, DTypeTrait<dtype::QuantizedS4>::max());
-        a = std::max(a, DTypeTrait<dtype::QuantizedS4>::min());
-        b = std::min(b, DTypeTrait<dtype::QuantizedS4>::max());
-        b = std::max(b, DTypeTrait<dtype::QuantizedS4>::min());
-        out_ptr[i / 2] = (a & 0xF) | (b << 4);
+    const auto& ly = in.layout;
+    auto dim_in = ly.shape[ly.ndim - 1];
+    auto elems = ly.total_nr_elems();
+    auto dim_out = elems / dim_in;
+    auto stride_out = div_ceil(dim_in, 2_z);
+    for (size_t i = 0; i < dim_out; ++i) {
+        for (size_t j = 0; j < dim_in; j += 2) {
+            int8_t a = in_ptr[j];
+            int8_t b = 0;
+            if (j + 1 < dim_in)
+                b = in_ptr[j + 1];
+            a = std::min(a, DTypeTrait<dtype::QuantizedS4>::max());
+            a = std::max(a, DTypeTrait<dtype::QuantizedS4>::min());
+            b = std::min(b, DTypeTrait<dtype::QuantizedS4>::max());
+            b = std::max(b, DTypeTrait<dtype::QuantizedS4>::min());
+            out_ptr[j / 2] = (a & 0xF) | (b << 4);
+        }
+        in_ptr += dim_in;
+        out_ptr += stride_out;
     }
 }