QB4W Fully Connected Subgraph

bench/qd8-f16-qb4w-gemm.cc

@@ -7,14 +7,63 @@
 //   Specification: test/qd8-f16-qb4w-gemm-minmax.yaml
 //   Generator: tools/generate-gemm-test.py
 
-#include <benchmark/benchmark.h>
-#include "bench/gemm-benchmark.h"
-#include "bench/utils.h"
-
+#include <xnnpack/common.h>
 #include <xnnpack/isa-checks.h>
 #include <xnnpack/gemm.h>
 #include <xnnpack/microfnptr.h>
 #include <xnnpack/microparams-init.h>
+#include <xnnpack/pack.h>
+
+#include <benchmark/benchmark.h>
+#include "bench/gemm-benchmark.h"
+#include "bench/utils.h"
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  static void qd8_f16_qb4w_gemm_minmax_ukernel_1x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f16_qb4w_gemm_minmax_ukernel_1x8c8__avx2,
+      xnn_init_f16_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/1, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f16_qb4w_gemm_minmax_ukernel_1x8c8__avx2)
+
+  static void qd8_f16_qb4w_gemm_minmax_ukernel_2x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f16_qb4w_gemm_minmax_ukernel_2x8c8__avx2,
+      xnn_init_f16_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/2, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f16_qb4w_gemm_minmax_ukernel_2x8c8__avx2)
+
+  static void qd8_f16_qb4w_gemm_minmax_ukernel_3x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f16_qb4w_gemm_minmax_ukernel_3x8c8__avx2,
+      xnn_init_f16_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/3, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f16_qb4w_gemm_minmax_ukernel_3x8c8__avx2)
+
+  static void qd8_f16_qb4w_gemm_minmax_ukernel_4x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f16_qb4w_gemm_minmax_ukernel_4x8c8__avx2,
+      xnn_init_f16_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/4, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f16_qb4w_gemm_minmax_ukernel_4x8c8__avx2)
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
 
 
 static void qd8_f16_qb4w_gemm_minmax_ukernel_1x2__scalar(benchmark::State& state, const char* net) {

bench/qd8-f32-qb4w-gemm.cc

@@ -7,14 +7,63 @@
 //   Specification: test/qd8-f32-qb4w-gemm-minmax.yaml
 //   Generator: tools/generate-gemm-test.py
 
-#include <benchmark/benchmark.h>
-#include "bench/gemm-benchmark.h"
-#include "bench/utils.h"
-
+#include <xnnpack/common.h>
 #include <xnnpack/isa-checks.h>
 #include <xnnpack/gemm.h>
 #include <xnnpack/microfnptr.h>
 #include <xnnpack/microparams-init.h>
+#include <xnnpack/pack.h>
+
+#include <benchmark/benchmark.h>
+#include "bench/gemm-benchmark.h"
+#include "bench/utils.h"
+
+
+#if XNN_ARCH_X86 || XNN_ARCH_X86_64
+  static void qd8_f32_qb4w_gemm_minmax_ukernel_1x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x8c8__avx2,
+      xnn_init_f32_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/1, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f32_qb4w_gemm_minmax_ukernel_1x8c8__avx2)
+
+  static void qd8_f32_qb4w_gemm_minmax_ukernel_2x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x8c8__avx2,
+      xnn_init_f32_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/2, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f32_qb4w_gemm_minmax_ukernel_2x8c8__avx2)
+
+  static void qd8_f32_qb4w_gemm_minmax_ukernel_3x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x8c8__avx2,
+      xnn_init_f32_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/3, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f32_qb4w_gemm_minmax_ukernel_3x8c8__avx2)
+
+  static void qd8_f32_qb4w_gemm_minmax_ukernel_4x8c8__avx2(benchmark::State& state, const char* net) {
+    GEMMBenchmark(state,
+      xnn_qd8_f32_qb4w_gemm_minmax_ukernel_4x8c8__avx2,
+      xnn_init_f32_qb4w_minmax_avx_params,
+      xnn_pack_qs8_qb4w_gemm_goi_w,
+      /*mr=*/4, /*nr=*/8, /*kr=*/8, /*sr=*/1,
+      benchmark::utils::CheckAVX2);
+  }
+
+  BENCHMARK_GEMM_BL(qd8_f32_qb4w_gemm_minmax_ukernel_4x8c8__avx2)
+#endif  // XNN_ARCH_X86 || XNN_ARCH_X86_64
 
 
 static void qd8_f32_qb4w_gemm_minmax_ukernel_1x2__scalar(benchmark::State& state, const char* net) {

cmake/gen/avx2_microkernels.cmake

@@ -329,6 +329,10 @@ SET(ALL_AVX2_MICROKERNEL_SRCS
   src/math/gen/f32-tanh-avx2-expm1minus-rr1-p6h5ts-div.c
   src/math/gen/f32-tanh-avx2-expm1minus-rr1-p6h5ts-nr1.c
   src/math/gen/f32-tanh-avx2-expm1minus-rr1-p6h5ts-nr1adj.c
+  src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-1x8c8-minmax-avx2.c
+  src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-2x8c8-minmax-avx2.c
+  src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-3x8c8-minmax-avx2.c
+  src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-4x8c8-minmax-avx2.c
   src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x8c8-minmax-avx2.c
   src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x8c8-minmax-avx2.c
   src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-3x8c8-minmax-avx2.c
@@ -341,6 +345,10 @@ SET(ALL_AVX2_MICROKERNEL_SRCS
   src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-2x8c8-minmax-avx2.c
   src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-3x8c8-minmax-avx2.c
   src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x8c8-minmax-avx2.c
+  src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x8c8-minmax-avx2.c
+  src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x8c8-minmax-avx2.c
+  src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x8c8-minmax-avx2.c
+  src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x8c8-minmax-avx2.c
   src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x8c8-minmax-avx2.c
   src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x8c8-minmax-avx2.c
   src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x8c8-minmax-avx2.c

gen/avx2_microkernels.bzl

@@ -325,6 +325,10 @@ ALL_AVX2_MICROKERNEL_SRCS = [
     "src/math/gen/f32-tanh-avx2-expm1minus-rr1-p6h5ts-div.c",
     "src/math/gen/f32-tanh-avx2-expm1minus-rr1-p6h5ts-nr1.c",
     "src/math/gen/f32-tanh-avx2-expm1minus-rr1-p6h5ts-nr1adj.c",
+    "src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-1x8c8-minmax-avx2.c",
+    "src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-2x8c8-minmax-avx2.c",
+    "src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-3x8c8-minmax-avx2.c",
+    "src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-4x8c8-minmax-avx2.c",
     "src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-1x8c8-minmax-avx2.c",
     "src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x8c8-minmax-avx2.c",
     "src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-3x8c8-minmax-avx2.c",
@@ -337,6 +341,10 @@ ALL_AVX2_MICROKERNEL_SRCS = [
     "src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-2x8c8-minmax-avx2.c",
     "src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-3x8c8-minmax-avx2.c",
     "src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x8c8-minmax-avx2.c",
+    "src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x8c8-minmax-avx2.c",
+    "src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x8c8-minmax-avx2.c",
+    "src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x8c8-minmax-avx2.c",
+    "src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x8c8-minmax-avx2.c",
     "src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-1x8c8-minmax-avx2.c",
     "src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x8c8-minmax-avx2.c",
     "src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x8c8-minmax-avx2.c",

include/xnnpack.h

@@ -264,6 +264,9 @@ enum xnn_datatype {
   /// Dynamically quantized 8-bit signed integers packed with their per-row
   /// quantization parameters.
   xnn_datatype_qpint8 = 10,
+    /// Quantized 4-bit signed integer with shared per-channel-block quantization
+    /// parametes.
+  xnn_datatype_qbint4 = 11,
 };
 
 /// Define a tensor-type Value and add it to a Subgraph.
@@ -407,6 +410,24 @@ enum xnn_status xnn_define_channelwise_quantized_tensor_value_v2(
   uint32_t flags,
   uint32_t* id_out);
 
+/// Define a blockwise quantized tensor-type Value and add it to a Subgraph.
+/// @param block_size - size of a block in the tensor with blockwise quantization parameters. Block is defined as number of input channel element per output channel.
+///                     For Fully connected operators with 2d filters of size [output_channels, input_channels], expecting number of scale values to be,
+///                     = output_channels * (input_channels / block_size).
+enum xnn_status xnn_define_blockwise_quantized_tensor_value(
+  xnn_subgraph_t subgraph,
+  enum xnn_datatype datatype,
+  int32_t zero_point,
+  const float* scale,
+  size_t num_dims,
+  size_t channel_dim,
+  size_t block_size,
+  const size_t* dims,
+  const void* data,
+  uint32_t external_id,
+  uint32_t flags,
+  uint32_t* id_out);
+
 /// Define a dynamically quantized tensor-type Value and add it to a Subgraph.
 ///
 /// @param subgraph - a Subgraph object that will own the created Value.
@@ -4317,6 +4338,90 @@ enum xnn_status xnn_reshape_fully_connected_nc_qd8_f32_qc4w(
   size_t batch_size,
   pthreadpool_t threadpool);
 
+enum xnn_status xnn_create_fully_connected_nc_qd8_f32_qb4w(
+  size_t input_channels,
+  size_t output_channels,
+  size_t input_stride,
+  size_t output_stride,
+  size_t block_size,
+  uint8_t kernel_zero_point,
+  const float* kernel_scale,
+  const void* kernel,
+  const float* bias,
+  float output_min,
+  float output_max,
+  uint32_t flags,
+  xnn_code_cache_t code_cache,
+  xnn_weights_cache_t weights_cache,
+  xnn_operator_t* fully_connected_op_out);
+
+enum xnn_status xnn_setup_fully_connected_nc_qd8_f32_qb4w(
+  xnn_operator_t fully_connected_op,
+  const int8_t* input,
+  float* output,
+  const struct xnn_dynamic_quantization_params* quantization_params);
+
+enum xnn_status xnn_reshape_fully_connected_nc_qd8_f32_qb4w(
+  xnn_operator_t fully_connected_op,
+  size_t batch_size,
+  pthreadpool_t threadpool);
+
+enum xnn_status xnn_create_fully_connected_nc_qd8_f16_qb4w(
+  size_t input_channels,
+  size_t output_channels,
+  size_t input_stride,
+  size_t output_stride,
+  size_t block_size,
+  uint8_t kernel_zero_point,
+  const float* kernel_scale,
+  const void* kernel,
+  const float* bias,
+  float output_min,
+  float output_max,
+  uint32_t flags,
+  xnn_code_cache_t code_cache,
+  xnn_weights_cache_t weights_cache,
+  xnn_operator_t* fully_connected_op_out);
+
+enum xnn_status xnn_setup_fully_connected_nc_qd8_f16_qb4w(
+  xnn_operator_t fully_connected_op,
+  const int8_t* input,
+  void* output,
+  const struct xnn_dynamic_quantization_params* quantization_params);
+
+enum xnn_status xnn_reshape_fully_connected_nc_qd8_f16_qb4w(
+  xnn_operator_t fully_connected_op,
+  size_t batch_size,
+  pthreadpool_t threadpool);
+
+enum xnn_status xnn_create_fully_connected_nc_qd8_f32_qb4w(
+  size_t input_channels,
+  size_t output_channels,
+  size_t input_stride,
+  size_t output_stride,
+  size_t block_size,
+  uint8_t kernel_zero_point,
+  const float* kernel_scale,
+  const void* kernel,
+  const float* bias,
+  float output_min,
+  float output_max,
+  uint32_t flags,
+  xnn_code_cache_t code_cache,
+  xnn_weights_cache_t weights_cache,
+  xnn_operator_t* fully_connected_op_out);
+
+enum xnn_status xnn_setup_fully_connected_nc_qd8_f32_qb4w(
+  xnn_operator_t fully_connected_op,
+  const int8_t* input,
+  float* output,
+  const struct xnn_dynamic_quantization_params* quantization_params);
+
+enum xnn_status xnn_reshape_fully_connected_nc_qd8_f32_qb4w(
+  xnn_operator_t fully_connected_op,
+  size_t batch_size,
+  pthreadpool_t threadpool);
+
 enum xnn_status xnn_create_fully_connected_nc_qd8_f16_qc8w(
   size_t input_channels,
   size_t output_channels,

scripts/generate-qs8-gemm.sh

@@ -983,6 +983,16 @@ tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=2 -D DATATYPE=QC4_F32 -D AVX=2 -
 tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=3 -D DATATYPE=QC4_F32 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x8c8-minmax-avx2.c &
 tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=4 -D DATATYPE=QC4_F32 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-4x8c8-minmax-avx2.c &
 
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=1 -D DATATYPE=QB4_F16 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-1x8c8-minmax-avx2.c &
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=2 -D DATATYPE=QB4_F16 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-2x8c8-minmax-avx2.c &
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=3 -D DATATYPE=QB4_F16 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-3x8c8-minmax-avx2.c &
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=4 -D DATATYPE=QB4_F16 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-4x8c8-minmax-avx2.c &
+
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=1 -D DATATYPE=QB4_F32 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-1x8c8-minmax-avx2.c &
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=2 -D DATATYPE=QB4_F32 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x8c8-minmax-avx2.c &
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=3 -D DATATYPE=QB4_F32 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x8c8-minmax-avx2.c &
+tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=4 -D DATATYPE=QB4_F32 -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=     -o src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-4x8c8-minmax-avx2.c &
+
 tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=1 -D DATATYPE=QC8     -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=FP32 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-1x8c8-minmax-fp32-avx2.c &
 tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=2 -D DATATYPE=QC8     -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=FP32 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-2x8c8-minmax-fp32-avx2.c &
 tools/xngen src/qs8-gemm/MRx8c8-avx2.c.in -D MR=3 -D DATATYPE=QC8     -D AVX=2 -D PREFETCH=0 -D REQUANTIZATION=FP32 -o src/qs8-qc8w-gemm/gen/qs8-qc8w-gemm-3x8c8-minmax-fp32-avx2.c &

scripts/generate-tests.sh

@@ -75,9 +75,11 @@ tools/generate-gemm-test.py --spec test/qu8-gemm-minmax-fp32.yaml --output-test
 tools/generate-gemm-test.py --spec test/qu8-gemm-minmax-rndnu.yaml --output-test test/qu8-gemm-minmax-rndnu.cc --output-test test/qu8-gemm-minmax-rndnu-2.cc --output-bench bench/qu8-gemm-rndnu.cc &
 
 tools/generate-gemm-test.py --spec test/qd8-f16-qc4w-gemm-minmax.yaml --output-test test/qd8-f16-qc4w-gemm-minmax.cc  --output-test test/qd8-f16-qc4w-gemm-minmax-2.cc  --output-test test/qd8-f16-qc4w-gemm-minmax-3.cc  --output-test test/qd8-f16-qc4w-gemm-minmax-4.cc --output-bench bench/qd8-f16-qc4w-gemm.cc &
+tools/generate-gemm-test.py --spec test/qd8-f16-qb4w-gemm-minmax.yaml --output-test test/qd8-f16-qb4w-gemm-minmax.cc --output-bench bench/qd8-f16-qb4w-gemm.cc &
 tools/generate-gemm-test.py --spec test/qd8-f16-qc8w-gemm-minmax.yaml --output-test test/qd8-f16-qc8w-gemm-minmax.cc --output-test test/qd8-f16-qc8w-gemm-minmax-2.cc --output-test test/qd8-f16-qc8w-gemm-minmax-3.cc --output-test test/qd8-f16-qc8w-gemm-minmax-4.cc --output-bench bench/qd8-f16-qc8w-gemm.cc &
 tools/generate-gemm-test.py --spec test/qd8-f32-qc8w-gemm-minmax.yaml --output-test test/qd8-f32-qc8w-gemm-minmax.cc  --output-test test/qd8-f32-qc8w-gemm-minmax-2.cc  --output-test test/qd8-f32-qc8w-gemm-minmax-3.cc  --output-test test/qd8-f32-qc8w-gemm-minmax-4.cc --output-bench bench/qd8-f32-qc8w-gemm.cc &
 tools/generate-gemm-test.py --spec test/qd8-f32-qc4w-gemm-minmax.yaml --output-test test/qd8-f32-qc4w-gemm-minmax.cc  --output-test test/qd8-f32-qc4w-gemm-minmax-2.cc  --output-test test/qd8-f32-qc4w-gemm-minmax-3.cc  --output-test test/qd8-f32-qc4w-gemm-minmax-4.cc --output-bench bench/qd8-f32-qc4w-gemm.cc &
+tools/generate-gemm-test.py --spec test/qd8-f32-qb4w-gemm-minmax.yaml --output-test test/qd8-f32-qb4w-gemm-minmax.cc --output-bench bench/qd8-f32-qb4w-gemm.cc &
 
 tools/generate-gemm-test.py --spec test/qp8-f32-qc4w-gemm-minmax.yaml --output-test test/qp8-f32-qc4w-gemm-minmax.cc --output-bench bench/qp8-f32-qc4w-gemm.cc &