Speed up Punica compilation (#2632)

.buildkite/test-template.j2

@@ -5,7 +5,7 @@
 steps:
   - label: ":docker: build image"
     commands:
-      - "docker build --tag {{ docker_image }} --target test --progress plain ."
+      - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
       - "docker push {{ docker_image }}"
     env:
       DOCKER_BUILDKIT: "1"

csrc/punica/bgmv/bgmv_all.cu

-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_half, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_half, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_half, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_half, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_bfloat16, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_bfloat16, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, float, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, float, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_bfloat16)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, float, nv_half)
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, float, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_bf16_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_half)

csrc/punica/bgmv/bgmv_bf16_fp16_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_fp16_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_half)

csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_fp32_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_half)

csrc/punica/bgmv/bgmv_fp16_bf16_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp16_bf16_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_half)

csrc/punica/bgmv/bgmv_fp16_fp16_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_half, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_half, nv_half)

csrc/punica/bgmv/bgmv_fp16_fp32_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, float, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, float, nv_half)

csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_bfloat16, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp32_bf16_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_bfloat16, nv_half)

csrc/punica/bgmv/bgmv_fp32_fp16_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_half, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_half, nv_half)

csrc/punica/bgmv/bgmv_fp32_fp32_bf16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, float, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp32_fp32_fp16.cu

+#include "bgmv_config.h"
+#include "bgmv_impl.cuh"
+
+FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, float, nv_half)